http://wiki.math.uwaterloo.ca/statwiki/api.php?action=feedcontributions&user=Wfisher&feedformat=atomstatwiki - User contributions [US]2022-01-23T04:17:20ZUser contributionsMediaWiki 1.28.3http://wiki.math.uwaterloo.ca/statwiki/index.php?title=CapsuleNets&diff=41732CapsuleNets2018-11-28T15:41:25Z<p>Wfisher: spelling</p>
<hr />
<div>The paper "Dynamic Routing Between Capsules" was written by three researchers at Google Brain: Sara Sabour, Nicholas Frosst, and Geoffrey E. Hinton. This paper was published and presented at the 31st Conference on Neural Information Processing Systems (NIPS 2017) in Long Beach, California. The same three researchers recently published a highly related paper "Matrix Capsules with EM Routing" for ICLR 2018.<br />
<br />
=Motivation=<br />
<br />
Ever since AlexNet eclipsed the performance of competing architectures in the 2012 ImageNet challenge, convolutional neural networks have maintained their dominance in computer vision applications. Despite the recent successes and innovations brought about by convolutional neural networks, some assumptions made in these networks are perhaps unwarranted and deficient. Using a novel neural network architecture, the authors create CapsuleNets, a network that they claim is able to learn image representations in a more robust, human-like manner. With only a 3 layer capsule network, they achieved near state-of-the-art results on MNIST.<br />
==Adversarial Examples==<br />
<br />
First discussed by Christian Szegedy et. al. in late 2013, adversarial examples have been heavily discussed by the deep learning community as a potential security threat to AI learning. Adversarial examples are defined as inputs that an attacker creates intentionally fool a machine learning model. An example of an adversarial example is shown below: <br />
<br />
[[File:adversarial_img_1.png |center]]<br />
To the human eye, the image appears to be a panda both before and after noise is injected into the image, whereas the trained ConvNet model discerns the noisy image as a Gibbon with almost 100% certainty. The fact that the network is unable to classify the above image as a panda after the epsilon perturbation leads to many potential security risks in AI dependent systems such as self-driving vehicles. Although various methods have been suggested to combat adversarial examples, robust defences are hard to construct due to the inherent difficulties in constructing theoretical models for the adversarial example crafting process. However, beyond the fact that these examples may serve as a security threat, it emphasizes that these convolutional neural networks do not learn image classification/object detection patterns the same way that a human would. Rather than identifying the core features of a panda such as: its eyes, mouth, nose, and the gradient changes in its black/white fur, the convolutional neural network seems to be learning image representations in a completely different manner. Deep learning researchers often attempt to model neural networks after human learning, and it is clear that further steps must be taken to robustify ConvNets against targeted noise perturbations.<br />
<br />
==Drawbacks of CNNs==<br />
Hinton claims that the key fault with traditional CNNs lies within the pooling function. Although pooling builds translational invariance into the network, it fails to preserve spatial relationships between objects. When we pool, we effectively reduce a kxk kernel of convolved cells into a scalar input. This results in a desired local invariance without inhibiting the network's ability to detect features, but causes valuable spatial information to be lost.<br />
<br />
In the example below, the network is able to detect the similar features (eyes, mouth, nose, etc) within both images, but fails to recognize that one image is a human face, while the other is a Picasso-esque due to the CNN's inability to encode spatial relationships after multiple pooling layers.<br />
<br />
<br />
[[File:Equivariance Face.png |center]]<br />
<br />
Conversely, we hope that a CNN can recognize that both of the following pictures contain a kitten. Unfortunately, when we feed the two images into a ResNet50 architecture, only the first image is correctly classified, while the second image is predicted to be a guinea pig.<br />
<br />
<br />
[[File:kitten.jpeg |center]]<br />
<br />
<br />
[[File:kitten-rotated-180.jpg |center]]<br />
<br />
For a more in depth discussion on the problems with ConvNets, please listen to Geoffrey Hinton's talk "What is wrong with convolutional neural nets?" given at MIT during the Brain & Cognitive Sciences - Fall Colloquium Series (December 4, 2014).<br />
<br />
==Intuition for Capsules==<br />
Human vision ignores irrelevant details by using a carefully determined sequence of fixation points to ensure that only a tiny fraction of the optic array is ever processed at the highest resolution. Hinton argues that our brains reason visual information by deconstructing it into a hierarchical representation which we then match to familiar patterns and relationships from memory. The key difference between this understanding and the functionality of CNNs is that recognition of an object should not depend on the angle from which it is viewed. <br />
<br />
To enforce rotational and translational equivariance, Capsule Networks store and preserve hierarchical pose relationships between objects. The core idea behind capsule theory is the explicit numerical representations of relative relationships between different objects within an image. Building these relationships into the Capsule Networks model, the network is able to recognize newly seen objects as a rotated view of a previously seen object. For example, the below image shows the Statue of Liberty under five different angles. If a person had only seen the Statue of Liberty from one angle, they would be able to ascertain that all five pictures below contain the same object (just from a different angle).<br />
<br />
[[File:Rotational Invariance.jpeg |center]]<br />
<br />
Building on this idea of hierarchical representation of spatial relationships between key entities within an image, the authors introduce Capsule Networks. Unlike traditional CNNs, Capsule Networks are better equipped to classify correctly under rotational invariance. Furthermore, the authors managed to achieve state of the art results on MNIST using a fraction of the training samples that alternative state of the art networks require.<br />
<br />
<br />
=Background, Notation, and Definitions=<br />
<br />
==What is a Capsule==<br />
"Each capsule learns to recognize an implicitly defined visual entity over a limited domain of viewing conditions and deformations and it outputs both the probability that the entity is present within its limited domain and a set of “instantiation parameters” that may include the precise pose, lighting and deformation of the visual entity relative to an implicitly defined canonical version of that entity. When the capsule is working properly, the probability of the visual entity being present is locally invariant — it does not change as the entity moves over the manifold of possible appearances within the limited domain covered by the capsule. The instantiation parameters, however, are “equivariant” — as the viewing conditions change and the entity moves over the appearance manifold, the instantiation parameters change by a corresponding amount because they are representing the intrinsic coordinates of the entity on the appearance manifold."<br />
<br />
In essence, capsules store object properties in a vector form; probability of detection is encoded as the vector's length, while spatial properties are encoded as the individual vector components. Thus, when a feature is present but the image captures it under a different angle, the probability of detection remains unchanged.<br />
<br />
A brief overview/understanding of capsules can be found in other papers from the author. To quote from [https://openreview.net/pdf?id=HJWLfGWRb this paper]:<br />
<br />
<blockquote><br />
A capsule network consists of several layers of capsules. The set of capsules in layer L is denoted<br />
as <math>\Omega_L</math>. Each capsule has a 4x4 pose matrix, <math>M</math>, and an activation probability, <math>a</math>. These are like the<br />
activities in a standard neural net: they depend on the current input and are not stored. In between<br />
each capsule i in layer L and each capsule j in layer L + 1 is a 4x4 trainable transformation matrix,<br />
<math>W_{ij}</math> . These <math>W_{ij}</math>'s (and two learned biases per capsule) are the only stored parameters and they<br />
are learned discriminatively. The pose matrix of capsule i is transformed by <math>W_{ij}</math> to cast a vote<br />
<math>V_{ij} = M_iW_{ij}</math> for the pose matrix of capsule j. The poses and activations of all the capsules in layer<br />
L + 1 are calculated by using a non-linear routing procedure which gets as input <math>V_{ij}</math> and <math>a_i</math> for all<br />
<math>i \in \Omega_L, j \in \Omega_{L+1}</math><br />
</blockquote><br />
<math></math><br />
<br />
==Notation==<br />
<br />
We want the length of the output vector of a capsule to represent the probability that the entity represented by the capsule is present in the current input. The paper performs a non-linear squashing operation to ensure that vector length falls between 0 and 1, with shorter vectors (less likely to exist entities) being shrunk towards 0. <br />
<br />
\begin{align} \mathbf{v}_j &= \frac{||\mathbf{s}_j||^2}{1+ ||\mathbf{s}_j||^2} \frac{\mathbf{s}_j}{||\mathbf{s}_j||} \end{align}<br />
<br />
where <math>\mathbf{v}_j</math> is the vector output of capsule <math>j</math> and <math>s_j</math> is its total input.<br />
<br />
For all but the first layer of capsules, the total input to a capsule <math>s_j</math> is a weighted sum over all “prediction vectors” <math>\hat{\mathbf{u}}_{j|i}</math> from the capsules in the layer below and is produced by multiplying the output <math>\mathbf{u}i</math> of a capsule in the layer below by a weight matrix <math>\mathbf{W}ij</math><br />
<br />
\begin{align}<br />
\mathbf{s}_j = \sum_i c_{ij}\hat{\mathbf{u}}_{j|i}, \hat{\mathbf{u}}_{j|i}= \mathbf{W}_{ij}\mathbf{u}_i<br />
\end{align}<br />
where the <math>c_{ij}</math> are coupling coefficients that are determined by the iterative dynamic routing process.<br />
<br />
The coupling coefficients between capsule <math>i</math> and all the capsules in the layer above sum to 1 and are determined by a “routing softmax” whose initial logits <math>b_{ij}</math> are the log prior probabilities that capsule <math>i</math> should be coupled to capsule <math>j</math>.<br />
<br />
\begin{align}<br />
c_{ij} = \frac{\exp(b_{ij})}{\sum_k \exp(b_{ik})}<br />
\end{align}<br />
<br />
=Network Training and Dynamic Routing=<br />
<br />
==Understanding Capsules==<br />
The notation can get somewhat confusing, so I will provide intuition behind the computational steps within a capsule. The following image is taken from naturomic's talk on Capsule Networks.<br />
<br />
[[File:CapsuleNets.jpeg|center|800px]]<br />
<br />
The above image illustrates the key mathematical operations happening within a capsule (and compares them to the structure of a neuron). Although the operations are rather straightforward, it's crucial to note that the capsule computes an affine transformation onto each input vector. The length of the input vectors <math>\mathbf{u}_{i}</math> represent the probability of entity <math>i</math> existing in a lower level. This vector is then reoriented with an affine transform using <math>\mathbf{W}_{ij}</math> matrices that encode spatial relationships between entity <math>\mathbf{u}_{i}</math> and other lower level features.<br />
<br />
We illustrate the intuition behind vector-to-vector matrix multiplication within capsules using the following example: if vectors <math>\mathbf{u}_{1}</math>, <math>\mathbf{u}_{2}</math>, and <math>\mathbf{u}_{3}</math> represent detection of eyes, nose, and mouth respectively, then after multiplication with trained weight matrices <math>\mathbf{W}_{ij}</math> (where j denotes existence of a face), we should get a general idea of the general location of the higher level feature (face), similar to the image below.<br />
<br />
[[File:Predictions.jpeg |center]]<br />
<br />
==Dynamic Routing==<br />
A capsule <math>i</math> in a lower-level layer needs to decide how to send its output vector to higher-level capsules <math>j</math>. This decision is made with probability proportional to <math>c_{ij}</math>. If there are <math>K</math> capsules in the level that capsule <math>i</math> routes to, then we know the following properties about <math>c_{ij}</math>: <math>\sum_{j=1}^M c_{ij} = 1, c_{ij} \geq 0</math><br />
<br />
In essence, the <math>\{c_{ij}\}_{j=1}^M</math> denotes a discrete probability distribution with respect to capsule <math>i</math>'s output location. Lower level capsules decide which higher level capsules to send vectors into by adjusting the corresponding routing weights <math>\{c_{ij}\}_{j=1}^M</math>. After a few iterations in training, numerous vectors will have already been sent to all higher level capsules. Based on the similarity between the current vector being routed and all vectors already sent into the higher level capsules, we decide which capsule to send the current vector into.<br />
[[File:Dynamic Routing.png|center|900px]]<br />
<br />
In the image above, we notice that a cluster of points similar to the current vector has already been routed into capsule K, while most points in capsule J are high dissimilar. It thus makes more sense to route the current observation into capsule K; we adjust the corresponding weight upwards during training.<br />
<br />
These weights are determined through the dynamic routing procedure:<br />
[[File:Routing Algo.png|900px]]<br />
<br />
<br />
Although dynamic routing is not the only manner in which we can encode relationships between capsules, the premise of the paper is to demonstrate the capabilities of capsules under a simple implementation. Since the paper's release in 2017, numerous alternative routing implementations have been released including an EM matrix routing algorithm by the same authors (ICLR 208).<br />
<br />
=Architecture=<br />
The capsule network architecture given by the authors has 11.36 million trainable parameters. The paper itself is not very detailed on exact implementation of each architectural layer, and hence it leaves some degree of ambiguity on coding various aspects of the original network. The capsule network has 6 overall layers, with the first three layers denoting components of the encoder, and the last 3 denoting components of the decoder.<br />
<br />
==Loss Function==<br />
[[File:Loss Function.png|900px]]<br />
<br />
The cost function looks very complicated, but can be broken down into intuitive components. Before diving into the equation, remember that the length of the vector denotes the probability of object existence. The left side of the equation denotes loss when the network classifies an observation correctly; the term becomes zero when classification is incorrect. To compute loss when the network correctly classifies the label, we subtract the vector norm from a fixed quantity <math>m^+ := 0.9</math>. On the other hand, when the network classifies a label incorrectly, we penalize the loss based on the network's confidence in the incorrect label; we compute the loss by subtracting <math>m^- := 0.1</math> from the vector norm.<br />
<br />
A graphical representation of loss function values under varying vector norms is given below.<br />
[[File:Loss function chart.png|900px]]<br />
<br />
==Encoder Layers==<br />
All experiments within this paper were conducted on the MNIST dataset, and thus the architecture is built to classify the corresponding dataset. For more complex datasets, the experiments were less promising. <br />
<br />
[[File:Architecture.png|center|900px]]<br />
<br />
The encoder layer takes in a 28x28 MNIST image, and learns a 16 dimensional representation of instantiation parameters.<br />
<br />
'''Layer 1: Convolution''': <br />
This layer is a standard convolution layer. Using kernels with size 9x9x1, a stride of 1, and a ReLU activation function, we detect the 2D features within the network.<br />
<br />
'''Layer 2: PrimaryCaps''': <br />
We represent the low level features detected during convolution as 32 primary capsules. Each capsule applies eight convolutional kernels with stride 2 to the output of the convolution layer, and feeds the corresponding transformed tensors into the DigiCaps layer.<br />
<br />
'''Layer 3: DigiCaps''': <br />
This layer contains 10 digit capsules, one for each digit. As explained in the dynamic routing procedure, each input vector from the PrimaryCaps layer has its own corresponding weight matrix <math>W_{ij}</math>. Using the routing coefficients <math>c_{ij}</math> and temporary coefficients <math>b_{ij}</math>, we train the DigiCaps layer to output a ten 16 dimensional vectors. The length of the <math>i^{th}</math> vector in this layer corresponds to the probability of detection of digit <math>i</math>.<br />
<br />
==Decoder Layers==<br />
The decoder layer aims to train the capsules to extract meaningful features for image detection/classification. During training, it takes the 16 layer instantiation vector of the correct (not predicted) DigiCaps layer, and attempts to recreate the 28x28 MNIST image as best as possible. Setting the loss function as reconstruction error (Euclidean distance between reconstructed image and original image), we tune the capsules to encode features that are meaningful within the actual image.<br />
<br />
[[File:Decoder.png|center|900px]]<br />
<br />
The layer consists of three fully connected layers, and transforms a 16x1 vector from the encoder layer into a 28x28 image.<br />
<br />
In addition to the digicaps loss function, we add reconstruction error as a form of regularization. We minimize the Euclidean distance between the outputs of the logistic units and the pixel intensities of the original and reconstructed images. We scale down this reconstruction loss by 0.0005 so that it does not dominate the margin loss during training. As illustrated below, reconstructions from the 16D output of the CapsNet are robust while keeping only important details.<br />
<br />
[[File:Reconstruction.png|center|900px]]<br />
<br />
=MNIST Experimental Results=<br />
<br />
==Accuracy==<br />
The paper tests on the MNIST dataset with 60K training examples, and 10K testing. Wan et al. [2013] achieves 0.21% test error with ensembling and augmenting the data with rotation and scaling. They achieve 0.39% without them. As shown in Table 1, the authors manage to achieve 0.25% test error with only a 3 layer network; the previous state of the art only beat this number with very deep networks. This example shows the importance of routing and reconstruction regularizer, which boosts the performance. On the other hand, while the accuracies are very high, the number of parameters is much smaller compared to the baseline model.<br />
<br />
[[File:Accuracies.png|center|900px]]<br />
<br />
==What Capsules Represent for MNIST==<br />
The following figure shows the digit representation under capsules. Each row shows the reconstruction when one of the 16 dimensions in the DigitCaps representation is tweaked by intervals of 0.05 in the range [−0.25, 0.25]. By tweaking the values, we notice how the reconstruction changes, and thus get a sense for what each dimension is representing. The authors found that some dimensions represent global properties of the digits, while other represent localized properties. <br />
[[File:CapsuleReps.png|center|900px]]<br />
<br />
One example the authors provide is: different dimensions are used for the length of the ascender of a 6 and the size of the loop. The variations include stroke thickness, skew and width, as well as digit-specific variations. The authors are able to show dimension representations using a decoder network by feeding a perturbed vector.<br />
<br />
==Robustness of CapsNet==<br />
The authors conclude that DigitCaps capsules learn more robust representations for each digit class than traditional CNNs. The trained CapsNet becomes moderately robust to small affine transformations in the test data.<br />
<br />
To compare the robustness of CapsNet to affine transformations against traditional CNNs, both models (CapsNet and a traditional CNN with MaxPooling and DropOut) were trained on a padded and translated MNIST training set, in which each example is an MNIST digit placed randomly on a black background of 40 × 40 pixels. The networks were then tested on the [http://www.cs.toronto.edu/~tijmen/affNIST/ affNIST] dataset (MNIST digits with random affine transformation). An under-trained CapsNet which achieved 99.23% accuracy on the MNIST test set achieved a corresponding 79% accuracy on the affnist test set. A traditional CNN achieved similar accuracy (99.22%) on the mnist test set, but only 66% on the affnist test set.<br />
<br />
=MultiMNIST & Other Experiments=<br />
<br />
==MultiMNIST==<br />
To evaluate the performance of the model on highly overlapping digits, the authors generate a 'MultiMNIST' dataset. In MultiMNIST, images are two overlaid MNIST digits of the same set(train or test) but different classes. The results indicate a classification error rate of 5%. Additionally, CapsNet can be used to segment the image into the two digits that compose it. Moreover, the model is able to deal with the overlaps and reconstruct digits correctly since each digit capsule can learn the style from the votes of PrimaryCapsules layer (Figure 5).<br />
<br />
There are some additional steps to generating the MultiMNIST dataset.<br />
<br />
1. Both images are shifted by up to 4 pixels in each direction resulting in a 36 × 36 image. Bounding boxes of digits in MNIST overlap by approximately 80%, so this is used to make both digits identifiable (since there is no RGB difference learnable by the network to separate the digits)<br />
<br />
2. The label becomes a vector of two numbers, representing the original digit and the randomly generated (and overlaid) digit.<br />
<br />
<br />
<br />
[[File:CapsuleNets MultiMNIST.PNG|600px|thumb|center|Figure 5: Sample reconstructions of a CapsNet with 3 routing iterations on MultiMNIST test dataset.<br />
The two reconstructed digits are overlayed in green and red as the lower image. The upper image<br />
shows the input image. L:(l1; l2) represents the label for the two digits in the image and R:(r1; r2)<br />
represents the two digits used for reconstruction. The two right most columns show two examples<br />
with wrong classification reconstructed from the label and from the prediction (P). In the (2; 8)<br />
example the model confuses 8 with a 7 and in (4; 9) it confuses 9 with 0. The other columns have<br />
correct classifications and show that the model accounts for all the pixels while being able to assign<br />
one pixel to two digits in extremely difficult scenarios (column 1 − 4). Note that in dataset generation<br />
the pixel values are clipped at 1. The two columns with the (*) mark show reconstructions from a<br />
digit that is neither the label nor the prediction. These columns suggests that the model is not just<br />
finding the best fit for all the digits in the image including the ones that do not exist. Therefore in case<br />
of (5; 0) it cannot reconstruct a 7 because it knows that there is a 5 and 0 that fit best and account for<br />
all the pixels. Also, in case of (8; 1) the loop of 8 has not triggered 0 because it is already accounted<br />
for by 8. Therefore it will not assign one pixel to two digits if one of them does not have any other<br />
support.]]<br />
<br />
=Critique=<br />
Although the network performs incredibly favourably in the author's experiments, it has a long way to go on more complex datasets. On CIFAR 10, the network achieved subpar results, and the experimental results seem to be worse when the problem becomes more complex. This is anticipated, since these networks are still in their early stage; later innovations might come in the upcoming decades/years.<br />
<br />
Hinton talks about CapsuleNets revolutionizing areas such as self-driving, but such groundbreaking innovations are far away from CIFAR10, and even further from MNIST. Only time can tell if CapsNets will live up to their hype.<br />
<br />
Capsules inherently segment images, and learn a lower dimensional embedding in a new manner, which makes them likely to perform well on segmentation and computer vision tasks once further research is done. <br />
<br />
Additionally these networks are more interpretable than CNNs, and have strong theoretical reasoning for why they could work. Naturally, it would be hard for a new architecture to beat the heavily researched/modified CNNs.<br />
<br />
* ([https://openreview.net/forum?id=HJWLfGWRb]) it's not fully clear how effective it can be performed / how scalable it is. Evaluation is performed on a small dataset for shape recognition. The approach will need to be tested on larger, more challenging datasets.<br />
<br />
=Future Work=<br />
The same authors [N. F. Geoffrey E Hinton, Sara Sabour] presented another paper "MATRIX CAPSULES WITH EM ROUTING" in ICLR 2018, which achieved better results than the work presented in this paper. They presented a novel capsule type, where each capsule has a logistic unit and a 4x4 pose matrix. This new type reduced number of errors by 45%, and performed better than standard CNN on white box adversarial attacks. <br />
Moreover, we may try to change the curvature and sensitivities to various factors by introducing new form of loss function. It may improve the performance of the model for more complicated data set which is one of the model's drawback.<br />
<br />
=References=<br />
#N. F. Geoffrey E Hinton, Sara Sabour. Matrix capsules with em routing. In International Conference on Learning Representations, 2018.<br />
#S. Sabour, N. Frosst, and G. E. Hinton, “Dynamic routing between capsules,” arXiv preprint arXiv:1710.09829v2, 2017<br />
# Hinton, G. E., Krizhevsky, A. and Wang, S. D. (2011), Transforming Auto-encoders <br />
#Geoffrey Hinton's talk: What is wrong with convolutional neural nets? - Talk given at MIT. Brain & Cognitive Sciences - Fall Colloquium Series. [https://www.youtube.com/watch?v=rTawFwUvnLE ]<br />
#Understanding Hinton’s Capsule Networks - Max Pechyonkin's series [https://medium.com/ai%C2%B3-theory-practice-business/understanding-hintons-capsule-networks-part-i-intuition-b4b559d1159b]</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=CapsuleNets&diff=41729CapsuleNets2018-11-28T15:40:01Z<p>Wfisher: Add a quote from another paper describing capsules (T)</p>
<hr />
<div>The paper "Dynamic Routing Between Capsules" was written by three researchers at Google Brain: Sara Sabour, Nicholas Frosst, and Geoffrey E. Hinton. This paper was published and presented at the 31st Conference on Neural Information Processing Systems (NIPS 2017) in Long Beach, California. The same three researchers recently published a highly related paper "Matrix Capsules with EM Routing" for ICLR 2018.<br />
<br />
=Motivation=<br />
<br />
Ever since AlexNet eclipsed the performance of competing architectures in the 2012 ImageNet challenge, convolutional neural networks have maintained their dominance in computer vision applications. Despite the recent successes and innovations brought about by convolutional neural networks, some assumptions made in these networks are perhaps unwarranted and deficient. Using a novel neural network architecture, the authors create CapsuleNets, a network that they claim is able to learn image representations in a more robust, human-like manner. With only a 3 layer capsule network, they achieved near state-of-the-art results on MNIST.<br />
==Adversarial Examples==<br />
<br />
First discussed by Christian Szegedy et. al. in late 2013, adversarial examples have been heavily discussed by the deep learning community as a potential security threat to AI learning. Adversarial examples are defined as inputs that an attacker creates intentionally fool a machine learning model. An example of an adversarial example is shown below: <br />
<br />
[[File:adversarial_img_1.png |center]]<br />
To the human eye, the image appears to be a panda both before and after noise is injected into the image, whereas the trained ConvNet model discerns the noisy image as a Gibbon with almost 100% certainty. The fact that the network is unable to classify the above image as a panda after the epsilon perturbation leads to many potential security risks in AI dependent systems such as self-driving vehicles. Although various methods have been suggested to combat adversarial examples, robust defences are hard to construct due to the inherent difficulties in constructing theoretical models for the adversarial example crafting process. However, beyond the fact that these examples may serve as a security threat, it emphasizes that these convolutional neural networks do not learn image classification/object detection patterns the same way that a human would. Rather than identifying the core features of a panda such as: its eyes, mouth, nose, and the gradient changes in its black/white fur, the convolutional neural network seems to be learning image representations in a completely different manner. Deep learning researchers often attempt to model neural networks after human learning, and it is clear that further steps must be taken to robustify ConvNets against targeted noise perturbations.<br />
<br />
==Drawbacks of CNNs==<br />
Hinton claims that the key fault with traditional CNNs lies within the pooling function. Although pooling builds translational invariance into the network, it fails to preserve spatial relationships between objects. When we pool, we effectively reduce a kxk kernel of convolved cells into a scalar input. This results in a desired local invariance without inhibiting the network's ability to detect features, but causes valuable spatial information to be lost.<br />
<br />
In the example below, the network is able to detect the similar features (eyes, mouth, nose, etc) within both images, but fails to recognize that one image is a human face, while the other is a Picasso-esque due to the CNN's inability to encode spatial relationships after multiple pooling layers.<br />
<br />
<br />
[[File:Equivariance Face.png |center]]<br />
<br />
Conversely, we hope that a CNN can recognize that both of the following pictures contain a kitten. Unfortunately, when we feed the two images into a ResNet50 architecture, only the first image is correctly classified, while the second image is predicted to be a guinea pig.<br />
<br />
<br />
[[File:kitten.jpeg |center]]<br />
<br />
<br />
[[File:kitten-rotated-180.jpg |center]]<br />
<br />
For a more in depth discussion on the problems with ConvNets, please listen to Geoffrey Hinton's talk "What is wrong with convolutional neural nets?" given at MIT during the Brain & Cognitive Sciences - Fall Colloquium Series (December 4, 2014).<br />
<br />
==Intuition for Capsules==<br />
Human vision ignores irrelevant details by using a carefully determined sequence of fixation points to ensure that only a tiny fraction of the optic array is ever processed at the highest resolution. Hinton argues that our brains reason visual information by deconstructing it into a hierarchical representation which we then match to familiar patterns and relationships from memory. The key difference between this understanding and the functionality of CNNs is that recognition of an object should not depend on the angle from which it is viewed. <br />
<br />
To enforce rotational and translational equivariance, Capsule Networks store and preserve hierarchical pose relationships between objects. The core idea behind capsule theory is the explicit numerical representations of relative relationships between different objects within an image. Building these relationships into the Capsule Networks model, the network is able to recognize newly seen objects as a rotated view of a previously seen object. For example, the below image shows the Statue of Liberty under five different angles. If a person had only seen the Statue of Liberty from one angle, they would be able to ascertain that all five pictures below contain the same object (just from a different angle).<br />
<br />
[[File:Rotational Invariance.jpeg |center]]<br />
<br />
Building on this idea of hierarchical representation of spatial relationships between key entities within an image, the authors introduce Capsule Networks. Unlike traditional CNNs, Capsule Networks are better equipped to classify correctly under rotational invariance. Furthermore, the authors managed to achieve state of the art results on MNIST using a fraction of the training samples that alternative state of the art networks require.<br />
<br />
<br />
=Background, Notation, and Definitions=<br />
<br />
==What is a Capsule==<br />
"Each capsule learns to recognize an implicitly defined visual entity over a limited domain of viewing conditions and deformations and it outputs both the probability that the entity is present within its limited domain and a set of “instantiation parameters” that may include the precise pose, lighting and deformation of the visual entity relative to an implicitly defined canonical version of that entity. When the capsule is working properly, the probability of the visual entity being present is locally invariant — it does not change as the entity moves over the manifold of possible appearances within the limited domain covered by the capsule. The instantiation parameters, however, are “equivariant” — as the viewing conditions change and the entity moves over the appearance manifold, the instantiation parameters change by a corresponding amount because they are representing the intrinsic coordinates of the entity on the appearance manifold."<br />
<br />
In essence, capsules store object properties in a vector form; probability of detection is encoded as the vector's length, while spatial properties are encoded as the individual vector components. Thus, when a feature is present but the image captures it under a different angle, the probability of detection remains unchanged.<br />
<br />
A breif overview/understanding of capsules can be found in other papers from the author. To quote from [https://openreview.net/pdf?id=HJWLfGWRb this paper]:<br />
<br />
<blockquote><br />
A capsule network consists of several layers of capsules. The set of capsules in layer L is denoted<br />
as <math>\Omega_L</math>. Each capsule has a 4x4 pose matrix, <math>M</math>, and an activation probability, <math>a</math>. These are like the<br />
activities in a standard neural net: they depend on the current input and are not stored. In between<br />
each capsule i in layer L and each capsule j in layer L + 1 is a 4x4 trainable transformation matrix,<br />
<math>W_{ij}</math> . These <math>W_{ij}</math>'s (and two learned biases per capsule) are the only stored parameters and they<br />
are learned discriminatively. The pose matrix of capsule i is transformed by <math>W_{ij}</math> to cast a vote<br />
<math>V_{ij} = M_iW_{ij}</math> for the pose matrix of capsule j. The poses and activations of all the capsules in layer<br />
L + 1 are calculated by using a non-linear routing procedure which gets as input <math>V_{ij}</math> and <math>a_i</math> for all<br />
<math>i \in \Omega_L, j \in \Omega_{L+1}</math><br />
</blockquote><br />
<math></math><br />
<br />
==Notation==<br />
<br />
We want the length of the output vector of a capsule to represent the probability that the entity represented by the capsule is present in the current input. The paper performs a non-linear squashing operation to ensure that vector length falls between 0 and 1, with shorter vectors (less likely to exist entities) being shrunk towards 0. <br />
<br />
\begin{align} \mathbf{v}_j &= \frac{||\mathbf{s}_j||^2}{1+ ||\mathbf{s}_j||^2} \frac{\mathbf{s}_j}{||\mathbf{s}_j||} \end{align}<br />
<br />
where <math>\mathbf{v}_j</math> is the vector output of capsule <math>j</math> and <math>s_j</math> is its total input.<br />
<br />
For all but the first layer of capsules, the total input to a capsule <math>s_j</math> is a weighted sum over all “prediction vectors” <math>\hat{\mathbf{u}}_{j|i}</math> from the capsules in the layer below and is produced by multiplying the output <math>\mathbf{u}i</math> of a capsule in the layer below by a weight matrix <math>\mathbf{W}ij</math><br />
<br />
\begin{align}<br />
\mathbf{s}_j = \sum_i c_{ij}\hat{\mathbf{u}}_{j|i}, \hat{\mathbf{u}}_{j|i}= \mathbf{W}_{ij}\mathbf{u}_i<br />
\end{align}<br />
where the <math>c_{ij}</math> are coupling coefficients that are determined by the iterative dynamic routing process.<br />
<br />
The coupling coefficients between capsule <math>i</math> and all the capsules in the layer above sum to 1 and are determined by a “routing softmax” whose initial logits <math>b_{ij}</math> are the log prior probabilities that capsule <math>i</math> should be coupled to capsule <math>j</math>.<br />
<br />
\begin{align}<br />
c_{ij} = \frac{\exp(b_{ij})}{\sum_k \exp(b_{ik})}<br />
\end{align}<br />
<br />
=Network Training and Dynamic Routing=<br />
<br />
==Understanding Capsules==<br />
The notation can get somewhat confusing, so I will provide intuition behind the computational steps within a capsule. The following image is taken from naturomic's talk on Capsule Networks.<br />
<br />
[[File:CapsuleNets.jpeg|center|800px]]<br />
<br />
The above image illustrates the key mathematical operations happening within a capsule (and compares them to the structure of a neuron). Although the operations are rather straightforward, it's crucial to note that the capsule computes an affine transformation onto each input vector. The length of the input vectors <math>\mathbf{u}_{i}</math> represent the probability of entity <math>i</math> existing in a lower level. This vector is then reoriented with an affine transform using <math>\mathbf{W}_{ij}</math> matrices that encode spatial relationships between entity <math>\mathbf{u}_{i}</math> and other lower level features.<br />
<br />
We illustrate the intuition behind vector-to-vector matrix multiplication within capsules using the following example: if vectors <math>\mathbf{u}_{1}</math>, <math>\mathbf{u}_{2}</math>, and <math>\mathbf{u}_{3}</math> represent detection of eyes, nose, and mouth respectively, then after multiplication with trained weight matrices <math>\mathbf{W}_{ij}</math> (where j denotes existence of a face), we should get a general idea of the general location of the higher level feature (face), similar to the image below.<br />
<br />
[[File:Predictions.jpeg |center]]<br />
<br />
==Dynamic Routing==<br />
A capsule <math>i</math> in a lower-level layer needs to decide how to send its output vector to higher-level capsules <math>j</math>. This decision is made with probability proportional to <math>c_{ij}</math>. If there are <math>K</math> capsules in the level that capsule <math>i</math> routes to, then we know the following properties about <math>c_{ij}</math>: <math>\sum_{j=1}^M c_{ij} = 1, c_{ij} \geq 0</math><br />
<br />
In essence, the <math>\{c_{ij}\}_{j=1}^M</math> denotes a discrete probability distribution with respect to capsule <math>i</math>'s output location. Lower level capsules decide which higher level capsules to send vectors into by adjusting the corresponding routing weights <math>\{c_{ij}\}_{j=1}^M</math>. After a few iterations in training, numerous vectors will have already been sent to all higher level capsules. Based on the similarity between the current vector being routed and all vectors already sent into the higher level capsules, we decide which capsule to send the current vector into.<br />
[[File:Dynamic Routing.png|center|900px]]<br />
<br />
In the image above, we notice that a cluster of points similar to the current vector has already been routed into capsule K, while most points in capsule J are high dissimilar. It thus makes more sense to route the current observation into capsule K; we adjust the corresponding weight upwards during training.<br />
<br />
These weights are determined through the dynamic routing procedure:<br />
[[File:Routing Algo.png|900px]]<br />
<br />
<br />
Although dynamic routing is not the only manner in which we can encode relationships between capsules, the premise of the paper is to demonstrate the capabilities of capsules under a simple implementation. Since the paper's release in 2017, numerous alternative routing implementations have been released including an EM matrix routing algorithm by the same authors (ICLR 208).<br />
<br />
=Architecture=<br />
The capsule network architecture given by the authors has 11.36 million trainable parameters. The paper itself is not very detailed on exact implementation of each architectural layer, and hence it leaves some degree of ambiguity on coding various aspects of the original network. The capsule network has 6 overall layers, with the first three layers denoting components of the encoder, and the last 3 denoting components of the decoder.<br />
<br />
==Loss Function==<br />
[[File:Loss Function.png|900px]]<br />
<br />
The cost function looks very complicated, but can be broken down into intuitive components. Before diving into the equation, remember that the length of the vector denotes the probability of object existence. The left side of the equation denotes loss when the network classifies an observation correctly; the term becomes zero when classification is incorrect. To compute loss when the network correctly classifies the label, we subtract the vector norm from a fixed quantity <math>m^+ := 0.9</math>. On the other hand, when the network classifies a label incorrectly, we penalize the loss based on the network's confidence in the incorrect label; we compute the loss by subtracting <math>m^- := 0.1</math> from the vector norm.<br />
<br />
A graphical representation of loss function values under varying vector norms is given below.<br />
[[File:Loss function chart.png|900px]]<br />
<br />
==Encoder Layers==<br />
All experiments within this paper were conducted on the MNIST dataset, and thus the architecture is built to classify the corresponding dataset. For more complex datasets, the experiments were less promising. <br />
<br />
[[File:Architecture.png|center|900px]]<br />
<br />
The encoder layer takes in a 28x28 MNIST image, and learns a 16 dimensional representation of instantiation parameters.<br />
<br />
'''Layer 1: Convolution''': <br />
This layer is a standard convolution layer. Using kernels with size 9x9x1, a stride of 1, and a ReLU activation function, we detect the 2D features within the network.<br />
<br />
'''Layer 2: PrimaryCaps''': <br />
We represent the low level features detected during convolution as 32 primary capsules. Each capsule applies eight convolutional kernels with stride 2 to the output of the convolution layer, and feeds the corresponding transformed tensors into the DigiCaps layer.<br />
<br />
'''Layer 3: DigiCaps''': <br />
This layer contains 10 digit capsules, one for each digit. As explained in the dynamic routing procedure, each input vector from the PrimaryCaps layer has its own corresponding weight matrix <math>W_{ij}</math>. Using the routing coefficients <math>c_{ij}</math> and temporary coefficients <math>b_{ij}</math>, we train the DigiCaps layer to output a ten 16 dimensional vectors. The length of the <math>i^{th}</math> vector in this layer corresponds to the probability of detection of digit <math>i</math>.<br />
<br />
==Decoder Layers==<br />
The decoder layer aims to train the capsules to extract meaningful features for image detection/classification. During training, it takes the 16 layer instantiation vector of the correct (not predicted) DigiCaps layer, and attempts to recreate the 28x28 MNIST image as best as possible. Setting the loss function as reconstruction error (Euclidean distance between reconstructed image and original image), we tune the capsules to encode features that are meaningful within the actual image.<br />
<br />
[[File:Decoder.png|center|900px]]<br />
<br />
The layer consists of three fully connected layers, and transforms a 16x1 vector from the encoder layer into a 28x28 image.<br />
<br />
In addition to the digicaps loss function, we add reconstruction error as a form of regularization. We minimize the Euclidean distance between the outputs of the logistic units and the pixel intensities of the original and reconstructed images. We scale down this reconstruction loss by 0.0005 so that it does not dominate the margin loss during training. As illustrated below, reconstructions from the 16D output of the CapsNet are robust while keeping only important details.<br />
<br />
[[File:Reconstruction.png|center|900px]]<br />
<br />
=MNIST Experimental Results=<br />
<br />
==Accuracy==<br />
The paper tests on the MNIST dataset with 60K training examples, and 10K testing. Wan et al. [2013] achieves 0.21% test error with ensembling and augmenting the data with rotation and scaling. They achieve 0.39% without them. As shown in Table 1, the authors manage to achieve 0.25% test error with only a 3 layer network; the previous state of the art only beat this number with very deep networks. This example shows the importance of routing and reconstruction regularizer, which boosts the performance. On the other hand, while the accuracies are very high, the number of parameters is much smaller compared to the baseline model.<br />
<br />
[[File:Accuracies.png|center|900px]]<br />
<br />
==What Capsules Represent for MNIST==<br />
The following figure shows the digit representation under capsules. Each row shows the reconstruction when one of the 16 dimensions in the DigitCaps representation is tweaked by intervals of 0.05 in the range [−0.25, 0.25]. By tweaking the values, we notice how the reconstruction changes, and thus get a sense for what each dimension is representing. The authors found that some dimensions represent global properties of the digits, while other represent localized properties. <br />
[[File:CapsuleReps.png|center|900px]]<br />
<br />
One example the authors provide is: different dimensions are used for the length of the ascender of a 6 and the size of the loop. The variations include stroke thickness, skew and width, as well as digit-specific variations. The authors are able to show dimension representations using a decoder network by feeding a perturbed vector.<br />
<br />
==Robustness of CapsNet==<br />
The authors conclude that DigitCaps capsules learn more robust representations for each digit class than traditional CNNs. The trained CapsNet becomes moderately robust to small affine transformations in the test data.<br />
<br />
To compare the robustness of CapsNet to affine transformations against traditional CNNs, both models (CapsNet and a traditional CNN with MaxPooling and DropOut) were trained on a padded and translated MNIST training set, in which each example is an MNIST digit placed randomly on a black background of 40 × 40 pixels. The networks were then tested on the [http://www.cs.toronto.edu/~tijmen/affNIST/ affNIST] dataset (MNIST digits with random affine transformation). An under-trained CapsNet which achieved 99.23% accuracy on the MNIST test set achieved a corresponding 79% accuracy on the affnist test set. A traditional CNN achieved similar accuracy (99.22%) on the mnist test set, but only 66% on the affnist test set.<br />
<br />
=MultiMNIST & Other Experiments=<br />
<br />
==MultiMNIST==<br />
To evaluate the performance of the model on highly overlapping digits, the authors generate a 'MultiMNIST' dataset. In MultiMNIST, images are two overlaid MNIST digits of the same set(train or test) but different classes. The results indicate a classification error rate of 5%. Additionally, CapsNet can be used to segment the image into the two digits that compose it. Moreover, the model is able to deal with the overlaps and reconstruct digits correctly since each digit capsule can learn the style from the votes of PrimaryCapsules layer (Figure 5).<br />
<br />
There are some additional steps to generating the MultiMNIST dataset.<br />
<br />
1. Both images are shifted by up to 4 pixels in each direction resulting in a 36 × 36 image. Bounding boxes of digits in MNIST overlap by approximately 80%, so this is used to make both digits identifiable (since there is no RGB difference learnable by the network to separate the digits)<br />
<br />
2. The label becomes a vector of two numbers, representing the original digit and the randomly generated (and overlaid) digit.<br />
<br />
<br />
<br />
[[File:CapsuleNets MultiMNIST.PNG|600px|thumb|center|Figure 5: Sample reconstructions of a CapsNet with 3 routing iterations on MultiMNIST test dataset.<br />
The two reconstructed digits are overlayed in green and red as the lower image. The upper image<br />
shows the input image. L:(l1; l2) represents the label for the two digits in the image and R:(r1; r2)<br />
represents the two digits used for reconstruction. The two right most columns show two examples<br />
with wrong classification reconstructed from the label and from the prediction (P). In the (2; 8)<br />
example the model confuses 8 with a 7 and in (4; 9) it confuses 9 with 0. The other columns have<br />
correct classifications and show that the model accounts for all the pixels while being able to assign<br />
one pixel to two digits in extremely difficult scenarios (column 1 − 4). Note that in dataset generation<br />
the pixel values are clipped at 1. The two columns with the (*) mark show reconstructions from a<br />
digit that is neither the label nor the prediction. These columns suggests that the model is not just<br />
finding the best fit for all the digits in the image including the ones that do not exist. Therefore in case<br />
of (5; 0) it cannot reconstruct a 7 because it knows that there is a 5 and 0 that fit best and account for<br />
all the pixels. Also, in case of (8; 1) the loop of 8 has not triggered 0 because it is already accounted<br />
for by 8. Therefore it will not assign one pixel to two digits if one of them does not have any other<br />
support.]]<br />
<br />
=Critique=<br />
Although the network performs incredibly favourably in the author's experiments, it has a long way to go on more complex datasets. On CIFAR 10, the network achieved subpar results, and the experimental results seem to be worse when the problem becomes more complex. This is anticipated, since these networks are still in their early stage; later innovations might come in the upcoming decades/years.<br />
<br />
Hinton talks about CapsuleNets revolutionizing areas such as self-driving, but such groundbreaking innovations are far away from CIFAR10, and even further from MNIST. Only time can tell if CapsNets will live up to their hype.<br />
<br />
Capsules inherently segment images, and learn a lower dimensional embedding in a new manner, which makes them likely to perform well on segmentation and computer vision tasks once further research is done. <br />
<br />
Additionally these networks are more interpretable than CNNs, and have strong theoretical reasoning for why they could work. Naturally, it would be hard for a new architecture to beat the heavily researched/modified CNNs.<br />
<br />
* ([https://openreview.net/forum?id=HJWLfGWRb]) it's not fully clear how effective it can be performed / how scalable it is. Evaluation is performed on a small dataset for shape recognition. The approach will need to be tested on larger, more challenging datasets.<br />
<br />
=Future Work=<br />
The same authors [N. F. Geoffrey E Hinton, Sara Sabour] presented another paper "MATRIX CAPSULES WITH EM ROUTING" in ICLR 2018, which achieved better results than the work presented in this paper. They presented a novel capsule type, where each capsule has a logistic unit and a 4x4 pose matrix. This new type reduced number of errors by 45%, and performed better than standard CNN on white box adversarial attacks. <br />
Moreover, we may try to change the curvature and sensitivities to various factors by introducing new form of loss function. It may improve the performance of the model for more complicated data set which is one of the model's drawback.<br />
<br />
=References=<br />
#N. F. Geoffrey E Hinton, Sara Sabour. Matrix capsules with em routing. In International Conference on Learning Representations, 2018.<br />
#S. Sabour, N. Frosst, and G. E. Hinton, “Dynamic routing between capsules,” arXiv preprint arXiv:1710.09829v2, 2017<br />
# Hinton, G. E., Krizhevsky, A. and Wang, S. D. (2011), Transforming Auto-encoders <br />
#Geoffrey Hinton's talk: What is wrong with convolutional neural nets? - Talk given at MIT. Brain & Cognitive Sciences - Fall Colloquium Series. [https://www.youtube.com/watch?v=rTawFwUvnLE ]<br />
#Understanding Hinton’s Capsule Networks - Max Pechyonkin's series [https://medium.com/ai%C2%B3-theory-practice-business/understanding-hintons-capsule-networks-part-i-intuition-b4b559d1159b]</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=stat946F18/Autoregressive_Convolutional_Neural_Networks_for_Asynchronous_Time_Series&diff=41726stat946F18/Autoregressive Convolutional Neural Networks for Asynchronous Time Series2018-11-28T15:19:29Z<p>Wfisher: Add description of AR (T)</p>
<hr />
<div>This page is a summary of the paper "[http://proceedings.mlr.press/v80/binkowski18a/binkowski18a.pdf Autoregressive Convolutional Neural Networks for Asynchronous Time Series]" by Mikołaj Binkowski, Gautier Marti, Philippe Donnat. It was published at ICML in 2018. The code for this paper is provided [https://github.com/mbinkowski/nntimeseries here].<br />
<br />
=Introduction=<br />
In this paper, the authors propose a deep convolutional network architecture called Significance-Offset Convolutional Neural Network for regression of multivariate asynchronous time series. The model is inspired by standard autoregressive(AR) models and gating systems used in recurrent neural networks. The model is evaluated on various time series data including:<br />
# Hedge fund proprietary dataset of over 2 million quotes for a credit derivative index, <br />
# An artificially generated noisy auto-regressive series, <br />
# A UCI household electricity consumption dataset. <br />
<br />
This paper focuses on time series with multi-variate and noisy signals, especially financial data. Financial time series is challenging to predict due to their low signal-to-noise ratio and heavy-tailed distributions. For example, the same signal (e.g. price of a stock) is obtained from different sources (e.g. financial news, an investment bank, financial analyst etc.) asynchronously. Each source may have a different bias or noise. (Figure 1) The investment bank with more clients can update their information more precisely than the investment bank with fewer clients, then the significance of each past observations may depend on other factors that change in time. Therefore, the traditional econometric models such as AR, VAR, VARMA[1] might not be sufficient. However, their relatively good performance could allow us to combine such linear econometric models with deep neural networks that can learn highly nonlinear relationships. This model is inspired by the gating mechanism which is successful in RNNs and Highway Networks.<br />
<br />
The time series forecasting problem can be expressed as a conditional probability distribution below,<br />
<div style="text-align: center;"><math>p(X_{t+d}|X_t,X_{t-1},...) = f(X_t,X_{t-1},...)</math></div><br />
Thus, we focus on modeling the predictors of future values of time series given their past values. <br />
The predictability of financial dataset still remains an open problem and is discussed in various publications [2].<br />
<br />
[[File:Junyi1.png | 500px|thumb|center|Figure 1: Quotes from four different market participants (sources) for the same credit default swaps (CDS) throughout one day. Each trader displays from time to time the prices for which he offers to buy (bid) and sell (ask) the underlying CDS. The filled area marks the difference between the best sell and buy offers (spread) at each time.]]<br />
<br />
The paper also provides empirical evidence that their model which combines linear models with deep learning models could perform better than just DL models like CNN, LSTMs and Phased LSTMs.<br />
<br />
=Related Work=<br />
===Time series forecasting===<br />
From recent proceedings in main machine learning venues i.e. ICML, NIPS, AISTATS, UAI, we can notice that time series are often forecast using Gaussian processes[3,4], especially for irregularly sampled time series[5]. Though still largely independent, combined models have started to appear, for example, the Gaussian Copula Process Volatility model[6]. For this paper, the authors use coupling AR models and neural networks to achieve such combined models.<br />
<br />
Although deep neural networks have been applied into many fields and produced satisfactory results, there still is little literature on deep learning for time series forecasting. More recently, the papers include Sirignano (2016)[7] that used 4-layer perceptrons in modeling price change distributions in Limit Order Books, and Borovykh et al. (2017)[8] who applied more recent WaveNet architecture to several short univariate and bivariate time-series (including financial ones). Heaton et al. (2016)[9] claimed to use autoencoders with a single hidden layer to compress multivariate financial data. Neil et al. (2016)[10] presented augmentation of LSTM architecture suitable for asynchronous series, which stimulates learning dependencies of different frequencies through time gate. <br />
<br />
In this paper, the authors examine the capabilities of several architectures (CNN, residual network, multi-layer LSTM, and phase LSTM) on AR-like artificial asynchronous and noisy time series, household electricity consumption dataset, and on real financial data from the credit default swap market with some inefficiencies.<br />
<br />
====AR Model====<br />
<br />
An autoregressive (AR) model describes the next value in a time-series as a combination of previous values, scaling factors, a bias, and noise [https://onlinecourses.science.psu.edu/stat501/node/358/ (source)]. For a p-th order (relating the current state to the p last states), the equation of the model is:<br />
<br />
<math> X_t = c + \sum_{i=1}^p \varphi_i X_{t-i}+ \varepsilon_t \,</math> [https://en.wikipedia.org/w/index.php?title=Autoregressive_model&action=edit&section=1 (equation source)]<br />
<br />
With parameters/coefficients <math>\varphi_i</math>, constant <math>c</math>, and noise <math>\varepsilon_t</math> This can be extended to vector form to create the VAR model mentioned in the paper.<br />
<br />
===Gating and weighting mechanisms===<br />
Gating mechanisms for neural networks has ability to overcome the problem of vanishing gradient, and can be expressed as <math display="inline">f(x)=c(x) \otimes \sigma(x)</math>, where <math>f</math> is the output function, <math>c</math> is a "candidate output" (a nonlinear function of <math>x</math>), <math>\otimes</math> is an element-wise matrix product, and <math>\sigma : \mathbb{R} \rightarrow [0,1] </math> is a sigmoid nonlinearity that controls the amount of output passed to the next layer. This composition of functions may lead to popular recurrent architecture such as LSTM and GRU[11].<br />
<br />
The idea of the gating system is aimed to weight outputs of the intermediate layers within neural networks, and is most closely related to softmax gating used in MuFuRu(Multi-Function Recurrent Unit)[12], i.e.<br />
<math display="inline"> f(x) = \sum_{l=1}^L p^l(x) \otimes f^l(x), p(x)=softmax(\widehat{p}(x)), </math>, where <math>(f^l)_{l=1}^L </math>are candidate outputs(composition operators in MuFuRu), <math>(\widehat{p}^l)_{l=1}^L </math>are linear functions of inputs. <br />
<br />
This idea is also successfully used in attention networks[13] such as image captioning and machine translation. In this paper, the method is similar as this. The difference is that modeling the functions as multi-layer CNNs. Another difference is that not using recurrent layers, which can enable the network to remember the parts of the sentence/image already translated/described.<br />
<br />
=Motivation=<br />
There are mainly five motivations they stated in the paper:<br />
#The forecasting problem in this paper has done almost independently by econometrics and machine learning communities. Unlike in machine learning, research in econometrics are more likely to explain variables rather than improving out-of-sample prediction power. These models tend to 'over-fit' on financial time series, their parameters are unstable and have poor performance on out-of-sample prediction.<br />
#Although Gaussian processes provide a useful theoretical framework that is able to handle asynchronous data, they often follow heavy-tailed distribution for financial datasets.<br />
#Predictions of autoregressive time series may involve highly nonlinear functions if sampled irregularly. For AR time series with higher order and have more past observations, the expectation of it <math display="inline">\mathbb{E}[X(t)|{X(t-m), m=1,...,M}]</math> may involve more complicated functions that in general may not allow closed-form expression.<br />
#In practice, the dimensions of multivariate time series are often observed separately and asynchronously, such series at fixed frequency may lead to lose information or enlarge the dataset, which is shown in Figure 2(a). Therefore, the core of the proposed architecture SOCNN represents separate dimensions as a single one with dimension and duration indicators as additional features(Figure 2(b)).<br />
#Given a series of pairs of consecutive input values and corresponding durations, <math display="inline"> x_n = (X(t_n),t_n-t_{n-1}) </math>. One may expect that LSTM may memorize the input values in each step and weight them at the output according to the duration, but this approach may lead to an imbalance between the needs for memory and for linearity. The weights that are assigned to the memorized observations potentially require several layers of nonlinearity to be computed properly, while past observations might just need to be memorized as they are.<br />
<br />
[[File:Junyi2.png | 550px|thumb|center|Figure 2: (a) Fixed sampling frequency and its drawbacks; keep- ing all available information leads to much more datapoints. (b) Proposed data representation for the asynchronous series. Consecutive observations are stored together as a single value series, regardless of which series they belong to; this information, however, is stored in indicator features, alongside durations between observations.]]<br />
<br />
<br />
=Model Architecture=<br />
Suppose there's a multivariate time series <math display="inline">(x_n)_{n=0}^{\infty} \subset \mathbb{R}^d </math>, we want to predict the conditional future values of a subset of elements of <math>x_n</math><br />
<div style="text-align: center;"><math>y_n = \mathbb{E} [x_n^I | {x_{n-m}, m=1,2,...}], </math></div><br />
where <math> I=\{i_1,i_2,...i_{d_I}\} \subset \{1,2,...,d\} </math> is a subset of features of <math>x_n</math>.<br />
<br />
Let <math> \textbf{x}_n^{-M} = (x_{n-m})_{m=1}^M </math>. <br />
<br />
The estimator of <math>y_n</math> can be expressed as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M [F(\textbf{x}_n^{-M}) \otimes \sigma(S(\textbf{x}_n^{-M}))].,_m ,</math></div><br />
The estimate is the summation of the columns of the matrix in bracket. Here<br />
#<math>F,S : \mathbb{R}^{d \times M} \rightarrow \mathbb{R}^{d_I \times M}</math> are neural networks. <br />
#* <math>S</math> is a fully convolutional network which is composed of convolutional layers only. <br />
#* <math display="inline">F(\textbf{x}_n^{-M}) = W \otimes [off(x_{n-m}) + x_{n-m}^I)]_{m=1}^M </math> <br />
#** <math> W \in \mathbb{R}^{d_I \times M}</math> <br />
#** <math> off: \mathbb{R}^d \rightarrow \mathbb{R}^{d_I} </math> is a multilayer perceptron.<br />
<br />
#<math>\sigma</math> is a normalized activation function independent at each row, i.e. <math display="inline"> \sigma ((a_1^T, ..., a_{d_I}^T)^T)=(\sigma(a_1)^T,..., \sigma(a_{d_I})^T)^T </math><br />
#* <math>a_{i} \in \mathbb{R}^{M}</math><br />
#* and <math>\sigma </math> is defined such that <math>\sigma(a)^{T} \mathbf{1}_{M}=1</math><br />
# <math>\otimes</math> is element-wise matrix multiplication.<br />
#<math>A.,_m</math> denotes the m-th column of a matrix A.<br />
<br />
Since <math>\sum_{m=1}^M W.,_m=W(1,1,...,1)^T</math> and <math>\sum_{m=1}^M S.,_m=S(1,1,...,1)^T</math>, we can express <math>\hat{y}_n</math> as:<br />
<div style="text-align: center;"><math>\hat{y}_n = \sum_{m=1}^M W.,_m \otimes (off(x_{n-m}) + x_{n-m}^I) \otimes \sigma(S.,_m(\textbf{x}_n^{-M}))</math></div><br />
This is the proposed network, Significance-Offset Convolutional Neural Network, <math>off</math> and <math>S</math> in the equation are corresponding to Offset and Significance in the name respectively.<br />
Figure 3 shows the scheme of network.<br />
<br />
[[File:Junyi3.png | 600px|thumb|center|Figure 3: A scheme of the proposed SOCNN architecture. The network preserves the time-dimension up to the top layer, while the number of features per timestep (filters) in the hidden layers is custom. The last convolutional layer, however, has the number of filters equal to dimension of the output. The Weighting frame shows how outputs from offset and significance networks are combined in accordance with Eq. of <math>\hat{y}_n</math>.]]<br />
<br />
The form of <math>\hat{y}_n</math> forced to separate the temporal dependence (obtained in weights <math>W_m</math>). S is determined by its filters which capture local dependencies and are independent of the relative position in time, the predictors <math>off(x_{n-m})</math> are completely independent of position in time. An adjusted single regressor for the target variable is provided by each past observation through the offset network. Since in asynchronous sampling procedure, consecutive values of x come from different signals and might be heterogeneous, therefore adjustment of offset network is important. In addition, significance network provides data-dependent weight for each regressor and sums them up in an autoregressive manner.<br />
<br />
===Relation to asynchronous data===<br />
One common problem of time series is that durations are varying between consecutive observations, the paper states two ways to solve this problem<br />
#Data preprocessing: aligning the observations at some fixed frequency e.g. duplicating and interpolating observations as shown in Figure 2(a). However, as mentioned in the figure, this approach will tend to loss of information and enlarge the size of the dataset and model complexity.<br />
#Add additional features: Treating the duration or time of the observations as additional features, it is the core of SOCNN, which is shown in Figure 2(b).<br />
<br />
===Loss function===<br />
The output of the offset network is series of separate predictors of changes between corresponding observations <math>x_{n-m}^I</math> and the target value<math>y_n</math>, this is the reason why we use auxiliary loss function, which equals to mean squared error of such intermediate predictions:<br />
<div style="text-align: center;"><math>L^{aux}(\textbf{x}_n^{-M}, y_n)=\frac{1}{M} \sum_{m=1}^M ||off(x_{n-m}) + x_{n-m}^I -y_n||^2 </math></div><br />
The total loss for the sample <math> \textbf{x}_n^{-M},y_n) </math> is then given by:<br />
<div style="text-align: center;"><math>L^{tot}(\textbf{x}_n^{-M}, y_n)=L^2(\widehat{y}_n, y_n)+\alpha L^{aux}(\textbf{x}_n^{-M}, y_n)</math></div><br />
where <math>\widehat{y}_n</math> was mentioned before, <math>\alpha \geq 0</math> is a constant.<br />
<br />
=Experiments=<br />
The paper evaluated SOCNN architecture on three datasets: artificially generated datasets, [https://archive.ics.uci.edu/ml/datasets/Individual+household+electric+power+consumption household electric power consumption dataset], and the financial dataset of bid/ask quotes sent by several market participants active in the credit derivatives market. Comparing its performance with simple CNN, single and multiplayer LSTM and 25-layer ResNet. Apart from the evaluation of the SOCNN architecture the paper also discusses the impact of network components such as: such as auxiliary<br />
loss and the depth of the offset sub-network. The code and datasets are available [https://github.com/mbinkowski/nntimeseries here]<br />
<br />
==Datasets==<br />
Artificial data: They generated 4 artificial series, <math> X_{K \times N}</math>, where <math>K \in \{16,64\} </math>. Therefore there is a synchronous and an asynchronous series for each K value.<br />
<br />
Electricity data: This UCI dataset contains 7 different features excluding date and time. The features include global active power, global reactive power, voltage, global intensity, sub-metering 1, sub-metering 2 and sub-metering 3, recorded every minute for 47 months. The data has been altered so that one observation contains only one value of 7 features, while durations between consecutive observations are ranged from 1 to 7 minutes. The goal is to predict all 7 features for the next time step.<br />
<br />
Non-anonymous quotes: The dataset contains 2.1 million quotes from 28 different sources from different market participants such as analysts, banks etc. Each quote is characterized by 31 features: the offered price, 28 indicators of the quoting source, the direction indicator (the quote refers to either a buy or a sell offer) and duration from the previous quote. For each source and direction, we want to predict the next quoted price from this given source and direction considering the last 60 quotes.<br />
<br />
==Training details==<br />
They applied grid search on some hyperparameters in order to get the significance of its components. The hyperparameters include the offset sub-network's depth and the auxiliary weight <math>\alpha</math>. For offset sub-network's depth, they use 1, 10,1 for artificial, electricity and quotes dataset respectively; and they compared the values of <math>\alpha</math> in {0,0.1,0.01}.<br />
<br />
They chose LeakyReLU as activation function for all networks:<br />
<div style="text-align: center;"><math>\sigma^{LeakyReLU}(x) = x</math> if <math>x\geq 0</math>, and <math>0.1x</math> otherwise </div><br />
They use the same number of layers, same stride and similar kernel size structure in CNN. In each trained CNN, they applied max pooling with the pool size of 2 every 2 convolutional layers.<br />
<br />
Table 1 presents the configuration of network hyperparameters used in comparison<br />
<br />
[[File:Junyi4.png | 400px|center|]]<br />
<br />
===Network Training===<br />
The training and validation data were sampled randomly from the first 80% of timesteps in each series, with ratio of 3 to 1. The remaining 20% of data was used as a test set.<br />
<br />
All models were trained using Adam optimizer because the authors found that its rate of convergence was much faster than standard Stochastic Gradient Descent in early tests.<br />
<br />
They used a batch size of 128 for artificial and electricity data, and 256 for quotes dataset, and applied batch normalization between each convolution and the following activation. <br />
<br />
At the beginning of each epoch, the training samples were randomly sampled. To prevent overfitting, they applied dropout and early stopping.<br />
<br />
Weights were initialized using the normalized uniform procedure proposed by Glorot & Bengio (2010).[14]<br />
<br />
The authors carried out the experiments on Tensorflow and Keras and used different GPU to optimize the model for different datasets.<br />
<br />
==Results==<br />
Table 2 shows all results performed from all datasets.<br />
[[File:Junyi5.png | 600px|center|]]<br />
We can see that SOCNN outperforms in all asynchronous artificial, electricity and quotes datasets. For synchronous data, LSTM might be slightly better, but SOCNN almost has the same results with LSTM. Phased LSTM and ResNet have performed really bad on artificial asynchronous dataset and quotes dataset respectively. Notice that having more than one layer of offset network would have negative impact on results. Also, the higher weights of auxiliary loss(<math>\alpha</math>considerably improved the test error on asynchronous dataset, see Table 3. However, for other datasets, its impact was negligible.<br />
[[File:Junyi6.png | 400px|center|]]<br />
In general, SOCNN has significantly lower variance of the test and validation errors, especially in the early stage of the training process and for quotes dataset. This effect can be seen in the learning curves for Asynchronous 64 artificial dataset presented in Figure 5.<br />
[[File:Junyi7.png | 500px|thumb|center|Figure 5: Learning curves with different auxiliary weights for SOCNN model trained on Asynchronous 64 dataset. The solid lines indicate the test error while the dashed lines indicate the training error.]]<br />
<br />
Finally, we want to test the robustness of the proposed model SOCNN, adding noise terms to asynchronous 16 dataset and check how these networks perform. The result is shown in Figure 6.<br />
[[File:Junyi8.png | 600px|thumb|center|Figure 6: Experiment comparing robustness of the considered networks for Asynchronous 16 dataset. The plots show how the error would change if an additional noise term was added to the input series. The dotted curves show the total significance and average absolute offset (not to scale) outputs for the noisy observations. Interestingly, the significance of the noisy observations increases with the magnitude of noise; i.e. noisy observations are far from being discarded by SOCNN.]]<br />
From Figure 6, the purple line and green line seems staying at the same position in training and testing process. SOCNN and single-layer LSTM are most robust compared to other networks, and least prone to overfitting.<br />
<br />
=Conclusion and Discussion=<br />
In this paper, the authors have proposed a new architecture called Significance-Offset Convolutional Neural Network, which combines AR-like weighting mechanism and convolutional neural network. This new architecture is designed for high-noise asynchronous time series and achieves outperformance in forecasting several asynchronous time series compared to popular convolutional and recurrent networks. <br />
<br />
The SOCNN can be extended further by adding intermediate weighting layers of the same type in the network structure. Another possible extension but needs further empirical studies is that we consider not just <math>1 \times 1</math> convolutional kernels on the offset sub-network. Also, this new architecture might be tested on other real-life datasets with relevant characteristics in the future, especially on econometric datasets and more generally for time series (stochastic processes) regression.<br />
<br />
=Critiques=<br />
#The paper is most likely an application paper, and the proposed new architecture shows improved performance over baselines in the asynchronous time series.<br />
#The quote data cannot be reached, only two datasets available.<br />
#The 'Significance' network was described as critical to the model in paper, but they did not show how the performance of SOCNN with respect to the significance network.<br />
#The transform of the original data to asynchronous data is not clear.<br />
#The experiments on the main application are not reproducible because the data is proprietary.<br />
#The way that train and test data were split is unclear. This could be important in the case of the financial data set.<br />
#Although the auxiliary loss function was mentioned as an important part, the advantages of it was not too clear in the paper. Maybe it is better that the paper describes a little more about its effectiveness.<br />
#It was not mentioned clearly in the paper whether the model training was done on a rolling basis for time series forecasting.<br />
#The noise term used in section 5's model robustness analysis uses evenly distributed noise (see Appendix B). While the analysis is a good start, analysis with different noise distributions would make the findings more generalizable.<br />
#The paper uses financial/economic data as one of its testing data set. Instead of comparing neural network models such as CNN which is known to work badly on time series data, it would be much better if the author compared to well-known econometric time series models such as GARCH and VAR.<br />
<br />
=References=<br />
[1] Hamilton, J. D. Time series analysis, volume 2. Princeton university press Princeton, 1994. <br />
<br />
[2] Fama, E. F. Efficient capital markets: A review of theory and empirical work. The journal of Finance, 25(2):383–417, 1970.<br />
<br />
[3] Petelin, D., Sˇindela ́ˇr, J., Pˇrikryl, J., and Kocijan, J. Financial modeling using gaussian process models. In Intelligent Data Acquisition and Advanced Computing Systems (IDAACS), 2011 IEEE 6th International Conference on, volume 2, pp. 672–677. IEEE, 2011.<br />
<br />
[4] Tobar, F., Bui, T. D., and Turner, R. E. Learning stationary time series using gaussian processes with nonparametric kernels. In Advances in Neural Information Processing Systems, pp. 3501–3509, 2015.<br />
<br />
[5] Hwang, Y., Tong, A., and Choi, J. Automatic construction of nonparametric relational regression models for multiple time series. In Proceedings of the 33rd International Conference on Machine Learning, 2016.<br />
<br />
[6] Wilson, A. and Ghahramani, Z. Copula processes. In Advances in Neural Information Processing Systems, pp. 2460–2468, 2010.<br />
<br />
[7] Sirignano, J. Extended abstract: Neural networks for limit order books, February 2016.<br />
<br />
[8] Borovykh, A., Bohte, S., and Oosterlee, C. W. Condi- tional time series forecasting with convolutional neural networks, March 2017.<br />
<br />
[9] Heaton, J. B., Polson, N. G., and Witte, J. H. Deep learn- ing in finance, February 2016.<br />
<br />
[10] Neil, D., Pfeiffer, M., and Liu, S.-C. Phased lstm: Acceler- ating recurrent network training for long or event-based sequences. In Advances In Neural Information Process- ing Systems, pp. 3882–3890, 2016.<br />
<br />
[11] Chung, J., Gulcehre, C., Cho, K., and Bengio, Y. Em- pirical evaluation of gated recurrent neural networks on sequence modeling, December 2014.<br />
<br />
[12] Weissenborn, D. and Rockta ̈schel, T. MuFuRU: The Multi-Function recurrent unit, June 2016.<br />
<br />
[13] Cho, K., Courville, A., and Bengio, Y. Describing multi- media content using attention-based Encoder–Decoder networks. IEEE Transactions on Multimedia, 17(11): 1875–1886, July 2015. ISSN 1520-9210.<br />
<br />
[14] Glorot, X. and Bengio, Y. Understanding the dif- ficulty of training deep feedforward neural net- works. In In Proceedings of the International Con- ference on Artificial Intelligence and Statistics (AIS- TATSaˆ10). Society for Artificial Intelligence and Statistics, 2010.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=policy_optimization_with_demonstrations&diff=41429policy optimization with demonstrations2018-11-26T18:01:27Z<p>Wfisher: Misc edits (E)</p>
<hr />
<div>= Introduction =<br />
<br />
The reinforcement learning (RL) method has made significant progress in a variety of applications, but the exploration problems regarding how to gain more experience from novel policies to improve long-term performance are still challenges, especially in environments where reward signals are sparse and rare. There are currently two ways to solve such exploration problems in RL: 1) Guide the agent to explore states that have never been seen. 2) Guide the agent to imitate a demonstration trajectory sampled from an expert policy to learn. When guiding the agent to imitate the expert behavior for learning, there are also two methods: putting the demonstration directly into the replay memory [1] [2] [3] or using the demonstration trajectory to pre-train the policy in a supervised manner [4]. However, neither of these methods takes full advantage of the demonstration data. To address this problem, a novel policy optimization method from demonstration (POfD) is proposed, which takes full advantage of the demonstration and there is no need to ensure that the expert policy is the optimal policy. In this paper, the authors evaluate the performance of POfD on Mujoco [5] in sparse-reward environments. The experiments results show that the performance of POfD is greatly improved compared with some strong baselines and even to the policy gradient method in dense-reward environments.<br />
<br />
==Intuition==<br />
The agent should imitate the demonstrated behavior when rewards are sparse and then explore new states on its own after acquiring sufficient skills, which is a dynamic intrinsic reward mechanism that can be reshaped in terms of the native rewards in RL. At present the state of the art exploration in Reinforcement learning is simply epsilon greedy which just makes random moves for a small percentage of times to explore unexplored moves. This is very naive and is one of the main reasons for the high sample complexity in RL. On the other hand, if there is an expert demonstrator who can guide exploration, the agent can make more guided and accurate exploratory moves.<br />
<br />
=Related Work =<br />
There are some related works in overcoming exploration difficulties by learning from demonstration [6] and imitation learning in RL.<br />
<br />
For learning from demonstration (LfD),<br />
# Most LfD methods adopt value-based RL algorithms, such as DQfD [2] which are applied into the discrete action spaces and DDPGfD [3] which extends this idea to the continuous spaces. But both of them under-utilize the demonstration data.<br />
# There are some methods based on policy iteration [7] [8], which shapes the value function by using demonstration data. But they get the bad performance when demonstration data is imperfect.<br />
# A hybrid framework [9] that learns the policy in which the probability of taking demonstrated actions is maximized is proposed, which considers less demonstration data.<br />
# A reward reshaping mechanism [10] that encourages taking actions close to the demonstrated ones is proposed. It is similar to the method in this paper, but there exists some differences as it is defined as a potential function based on multi-variate Gaussian to model the distribution of state-actions.<br />
All of the above methods require a lot of perfect demonstrations to get satisfactory performance, which is different from POfD in this paper.<br />
<br />
For imitation learning, <br />
# Inverse Reinforce Learning [11] problems are solved by alternating between fitting the reward function and selecting the policy [12] [13]. But it cannot be extended to big-scale problems.<br />
# Generative Adversarial Imitation Learning (GAIL) [14] uses a discriminator to distinguish whether a state-action pair is from the expert or the learned policy and it can be applied into the high-dimensional continuous control problems.<br />
<br />
Both of the above methods are effective for imitation learning, but cannot leverage the valuable feedback given by the environments and usually suffer from bad performance when the expert data is imperfect. That is different from POfD in this paper.<br />
<br />
There is also another idea in which an agent learns using hybrid imitation learning and reinforcement learning reward[23, 24]. However, unlike this paper, they did not provide some theoretical support for their method and only explained some intuitive explanations.<br />
<br />
=Background=<br />
<br />
==Preliminaries==<br />
Markov Decision Process (MDP) [15] is defined by a tuple <math>⟨S, A, P, r, \gamma⟩ </math>, where <math>S</math> is the state, <math>A </math> is the action, <math>P(s'|s,a)</math> is the transition distribution of taking action <math> a </math> at state <math>s </math>, <math> r(s,a) </math>is the reward function, and <math> \gamma </math> is discounted factor between 0 and 1. Policy <math> \pi(a|s) </math> is a mapping from state to action, the performance of <math> \pi </math> is usually evaluated by its expected discounted reward <math> \eta(\pi) </math>: <br />
\[\eta(\pi)=\mathbb{E}_{\pi}[r(s,a)]=\mathbb{E}_{(s_0,a_0,s_1,...)}[\sum_{t=0}^\infty\gamma^{t}r(s_t,a_t)] \]<br />
The value function is <math> V_{\pi}(s) =\mathbb{E}_{\pi}[r(·,·)|s_0=s] </math>, the action value function is <math> Q_{\pi}(s,a) =\mathbb{E}_{\pi}[r(·,·)|s_0=s,a_0=a] </math>, and the advantage function that reflects the expected additional reward after taking action a at state s is <math> A_{\pi}(s,a)=Q_{\pi}(s,a)-V_{\pi}(s)</math>.<br />
Then the authors define Occupancy measure, which is used to estimate the probability that state <math>s</math> and state action pairs <math>(s,a)</math> when executing a certain policy.<br />
[[File:def1.png|500px|center]]<br />
Then the performance of <math> \pi </math> can be rewritten to: <br />
[[File:equ2.png|500px|center]]<br />
At the same time, the authors propose a lemma: <br />
[[File:lemma1.png|500px|center]]<br />
<br />
==Problem Definition==<br />
Generally, RL tasks and environments do not provide a comprehensive reward and instead rely on sparse feedback indicating whether the goal is reached.<br />
<br />
In this paper, the authors aim to develop a method that can boost exploration by leveraging effectively the demonstrations <math>D^E </math>from the expert policy <math> \pi_E </math> and maximize <math> \eta(\pi) </math> in the sparse-reward environment. The authors define the demonstrations <math>D^E=\{\tau_1,\tau_2,...,\tau_N\} </math>, where the i-th trajectory <math>\tau_i=\{(s_0^i,a_0^i),(s_1^i,a_1^i),...,(s_T^i,a_T^i)\} </math> is generated from the expert policy. In addition, there is an assumption on the quality of the expert policy:<br />
[[File:asp1.png|500px|center]]<br />
Moreover, it is not necessary to ensure that the expert policy is advantageous over all the policies. This is because that POfD will learn a better policy than expert policy by exploring on its own in later learning stages.<br />
<br />
=Method=<br />
<br />
==Policy Optimization with Demonstration (POfD)==<br />
[[File:ff1.png|500px|center]]<br />
This method optimizes the policy by forcing the policy to explore in the nearby region of the expert policy that is specified by several demonstrated trajectories <math>D^E </math> (as shown in Fig.1) in order to avoid causing slow convergence or failure when the environment feedback is sparse. In addition, the authors encourage the policy π to explore by "following" the demonstrations <math>D^E </math>. Thus, a new learning objective is given:<br />
\[ \mathcal{L}(\pi_{\theta})=-\eta(\pi_{\theta})+\lambda_{1}D_{JS}(\pi_{\theta},\pi_{E})\]<br />
where <math>D_{JS}(\pi_{\theta},\pi_{E})</math> is Jensen-Shannon divergence between current policy <math>\pi_{\theta}</math> and the expert policy <math>\pi_{E}</math> , <math>\lambda_1</math> is a trading-off parameter, and <math>\theta</math> is policy parameter. According to Lemma 1, the authors use <math>D_{JS}(\rho_{\theta},\rho_{E})</math> to instead of <math>D_{JS}(\pi_{\theta},\pi_{E})</math>, because it is easier to optimize through adversarial training on demonstrations. The learning objective is: <br />
\[ \mathcal{L}(\pi_{\theta})=-\eta(\pi_{\theta})+\lambda_{1}D_{JS}(\rho_{\theta},\rho_{E})\]<br />
<br />
==Benefits of Exploration with Demonstrations==<br />
The authors introduce the benefits of POfD. Firstly, we consider the expression of expected return in policy gradient methods [16].<br />
\[ \eta(\pi)=\eta(\pi_{old})+\mathbb{E}_{\tau\sim\pi}[\sum_{t=0}^\infty\gamma^{t}A_{\pi_{old}}(s,a)]\]<br />
<math>\eta(\pi)</math>is the advantage over the policy πold in the previous iteration, so the expression can be rewritten by<br />
\[ \eta(\pi)=\eta(\pi_{old})+\sum_{s}\rho_{\pi}(s)\sum_{a}\pi(a|s)A_{\pi_{old}}(s,a)\]<br />
The local approximation to <math>\eta(\pi)</math> up to first order is usually as the surrogate learning objective to be optimized by policy gradient methods due to the difficulties brought by complex dependency of <math>\rho_{\pi}(s)</math> over <math> \pi </math>:<br />
\[ J_{\pi_{old}}(\pi)=\eta(\pi_{old})+\sum_{s}\rho_{\pi_{old}}(s)\sum_{a}\pi(a|s)A_{\pi_{old}}(s,a)\]<br />
The policy gradient methods improve <math>\eta(\pi)</math> monotonically by optimizing the above <math>J_{\pi_{old}}(\pi)</math> with a sufficiently small update step from <math>\pi_{old}</math> to <math>\pi</math> such that <math>D_{KL}^{max}(\pi, \pi_{old})</math> is bounded [16] [17] [18]. For POfD, it imposes a regularization <math>D_{JS}(\pi_{\theta}, \pi_{E})</math> in order to encourage explorations around regions demonstrated by the expert policy. Theorem 1 shows such benefits,<br />
[[File:them1.png|500px|center]]<br />
<br />
In fact, POfD brings another factor, <math>D_{J S}^{max}(\pi_{i}, \pi_{E})</math>, that would fully use the advantage <math>{\hat \delta}</math>and add improvements with a margin over pure policy gradient methods.<br />
<br />
==Optimization==<br />
<br />
For POfD, the authors choose to optimize the lower bound of learning objective rather than optimizing the Jensen-Shannon divergence. This optimization method is compatible with any policy gradient methods. Theorem 2 gives the lower bound of <math>D_{JS}(\rho_{\theta}, \rho_{E})</math>：<br />
[[File:them2.png|500px|center]]<br />
Thus, the occupancy measure matching objective can be written as:<br />
[[File:eqnlm.png|500px|center]]<br />
where <math> D(s,a)=\frac{1}{1+e^{-U(s,a)}}: S\times A \rightarrow (0,1)</math>, and its supremum ranging is like a discriminator for distinguishing whether the state-action pair is a current policy or an expert policy.<br />
To avoid overfitting, the authors add causal entropy <math>−H (\pi_{\theta}) </math> as the regularization term. Thus, the learning objective is: <br />
\[\min_{\theta}\mathcal{L}=-\eta(\pi_{\theta})-\lambda_{2}H(\pi_{\theta})+\lambda_{1} \sup_{{D\in(0,1)}^{S\times A}} \mathbb{E}_{\pi_{\theta}}[\log(D(s,a))]+\mathbb{E}_{\pi_{E}}[\log(1-D(s,a))]\]<br />
At this point, the problem has been like Generative Adversarial Networks (GANs) [19]. The difference is that the discriminative model D of GANs is well-trained but the expert policy of POfD is not optimal. Then suppose D is parameterized by w. If it is from an expert policy, <math>D_w</math>is toward 1, otherwise it is toward 0. Thus, the minimax learning objective is:<br />
\[\min_{\theta}\max_{w}\mathcal{L}=-\eta(\pi_{\theta})-\lambda_{2}H (\pi_{\theta})+\lambda_{1}( \mathbb{E}_{\pi_{\theta}}[\log(D_{w}(s,a))]+\mathbb{E}_{\pi_{E}}[\log(1-D_{w}(s,a))])\]<br />
The minimax learning objective can be rewritten by substituting the expression of <math> \eta(\pi) </math>:<br />
\[\min_{\theta}\max_{w}-\mathbb{E}_{\pi_{\theta}}[r'(s,a)]-\lambda_{2}H (\pi_{\theta})+\lambda_{1}\mathbb{E}_{\pi_{E}}[\log(1-D_{w}(s,a))]\]<br />
where <math> r'(s,a)=r(a,b)-\lambda_{1}\log(D_{w}(s,a))</math> is the reshaped reward function.<br />
The above objective can be optimized efficiently by alternately updating policy parameters θ and discriminator parameters w, then the gradient is given by:<br />
\[\mathbb{E}_{\pi}[\nabla_{w}\log(D_{w}(s,a))]+\mathbb{E}_{\pi_{E}}[\nabla_{w}\log(1-D_{w}(s,a))]\]<br />
Then, fixing the discriminator <math>D_w</math>, the reshaped policy gradient is:<br />
\[\nabla_{\theta}\mathbb{E}_{\pi_{\theta}}[r'(s,a)]=\mathbb{E}_{\pi_{\theta}}[\nabla_{\theta}\log\pi_{\theta}(a|s)Q'(s,a)]\]<br />
where <math>Q'(\bar{s},\bar{a})=\mathbb{E}_{\pi_{\theta}}[r'(s,a)|s_0=\bar{s},a_0=\bar{a}]</math>.<br />
<br />
At the end, Algorithm 1 gives the detailed process.<br />
[[File:pofd.png|500px|center]]<br />
<br />
=Discussion on Existing LfD Methods=<br />
<br />
==DQFD==<br />
DQFD [2] puts the demonstrations into a replay memory D and keeps them throughout the Q-learning process. The objective for DQFD is:<br />
\[J_{DQfD}={\hat{\mathbb{E}}}_{D}[(R_t(n)-Q_w(s_t,a_t))^2]+\alpha{\hat{\mathbb{E}}}_{D^E}[(R_t(n)-Q_w(s_t,a_t))^2]\]<br />
The second term can be rewritten as <math> {\hat{\mathbb{E}}}_{D^E}[(R_t(n)-Q_w(s_t,a_t))^2]={\hat{\mathbb{E}}}_{D^E}[(\hat{\rho}_E(s,a)-\rho_{\pi}(s,a))^{2}r^2(s,a)]</math>, which can be regarded as a regularization forcing current policy's occupancy measure to match the expert's empirical occupancy measure, weighted by the potential reward.<br />
<br />
==DDPGfD==<br />
DDPGfD [3] also puts the demonstrations into a replay memory D, but it is based on an actor-critic framework [21]. The objective for DDPGfD is the same as DQFD. Its policy gradient is:<br />
\[\nabla_{\theta}J_{DDPGfD}\approx \mathbb{E}_{s,a}[\nabla_{a}Q_w(s,a)\nabla_{\theta}\pi_{\theta}(s)], a=\pi_{\theta}(s) \]<br />
From this equation, policy is updated relying on learned Q-network <math>Q_w </math>rather than the demonstrations <math>D^{E} </math>. DDPGfD shares the same objective function for <math>Q_w </math> as DQfD, thus they have the same way of leveraging demonstrations, that is the demonstrations in DQfD and DDPGfD induce an occupancy measure matching regularization.<br />
<br />
=Experiments=<br />
<br />
==Goal==<br />
The authors aim at investigating 1) whether POfD can aid exploration by leveraging a few demonstrations, even though the demonstrations are imperfect. 2) whether POfD can succeed and achieve high empirical return, especially in environments where reward signals are sparse and rare. <br />
<br />
==Settings==<br />
The authors conduct the experiments on 8 physical control tasks, ranging from low-dimensional spaces to high-dimensional spaces and naturally sparse environments based on OpenAI Gym [20] and Mujoco [5]. Due to the uniqueness of the environments, the authors introduce 4 ways to sparsify their built-in dense rewards. TYPE1: a reward of +1 is given when the agent reaches the terminal state, and otherwisel 0. TYPE2: a reward of +1 is given when the agent survives for a while. TYPE3: a reward of +1 is given for every time the agent moves forward over a specific number of units in Mujoco environments. TYPE4: specially designed for InvertedDoublePendulum, a reward +1 is given when the second pole stays above a specific height of 0.89. The details are shown in Table 1. Moreover, only one single imperfect trajectory is used as the demonstrations in this paper. The authors collect the demonstrations by training an agent insufficiently by running TRPO in the corresponding dense environment. <br />
[[File:pofdt1.png|900px|center]]<br />
<br />
==Baselines==<br />
The authors compare POfD against 5 strong baselines:<br />
* training the policy with TRPO [17] in dense environments, which is called expert <br />
* training the policy with TRPO [17] in sparse environments<br />
* applying GAIL [14] to learn the policy from demonstrations<br />
* DQfD [2]<br />
* DDPGfD [3]<br />
<br />
==Results==<br />
Firstly, the authors test the performance of POfD in sparse control environments with discrete actions. From Table 1, POfD achieves performance comparable with the policy learned under dense environments. From Figure 2, only POfD successes to explore sufficiently and achieves great performance in both sparse environments. TRPO [17] and DQFD [2] fail to explore and GAIL [14] converges to the imperfect demonstration in MountainCar [22].<br />
<br />
[[File:pofdf2.png|500px|center]]<br />
<br />
Then, the authors test the performance of POfD under spares environments with continuous actions space. From Figure 3, POfD achieves expert-level performance in terms of accumulated rewards and surpasses other strong baselines training the policy with TRPO. By watching the learning process of different methods, we can see that TRPO consistently fails to explore the environments when the feedback is sparse, except for HalfCheetah. This may be because there is no terminal state in HalfCheetah, thus a random agent can perform reasonably well as long as the time horizon is sufficiently long. This is shown in Figure3 where the improvement of TRPO begins to show after 400 iterations. DDPGfD and GAIL have common drawback: during training process, they both converge to the imperfect demonstration data. For HalfCheetah, GAIL fails to converge and DDPGfD converges to an even worse point. This situation is expected because the policy and value networks tend to over-fit when having few data, so the training process of GAIL and DDPGfD is severely biased by the imperfect data. Finally, our proposed method can effectively explore the environment with the help of demonstration-based intrinsic reward reshaping, and succeeds consistently across different tasks both in terms of learning stability and convergence speed.<br />
[[File:pofdf3.png|900px|center]]<br />
<br />
The authors also implement a locomotion task <math>Humanoid</math>, which teaches a human-like robot to walk. The state space of dimension is 376, which is very hard to render. As a result, POfD still outperformed all three baselike methods, as they failed to learn policies in such a sparse reward environment.<br />
<br />
The reacher environment is a task that the target is to control a robot arm to touch an object. the location of the object is random for each instantiation. The authors select 15 random trajectories as demonstration data, and the performance of POfD is much better than the expert, while all other baseline methods failed.<br />
<br />
=Conclusion=<br />
In this paper, a method, POfD, is proposed that can acquire knowledge from a limited amount of imperfect demonstration data to aid exploration in environments with sparse feedback. It is compatible with any policy gradient methods. POfD induces implicit dynamic reward shaping and brings provable benefits for policy improvement. Moreover, the experiments results have shown the validity and effectivness of POfD in encouraging the agent to explore around the nearby region of the expert policy and learn better policies. The key contribution is that POfD helps the agent work with few and imperfect demonstrations in an environment with sparse rewards.<br />
<br />
=Critique=<br />
# A novel demonstration-based policy optimization method is proposed. In the process of policy optimization, POfD reshapes the reward function. This new reward function can guide the agent to imitate the expert behaviour when the reward is sparse and explore on its own when the reward value can be obtained, which can take full advantage of the demonstration data and there is no need to ensure that the expert policy is the optimal policy.<br />
# POfD can be combined with any policy gradient methods. Its performance surpasses five strong baselines and can be comparable to the agents trained in the dense-reward environment.<br />
# The paper is structured and the flow of ideas is easy to follow. For related work, the authors clearly explain similarities and differences among these related works.<br />
# This paper's scalability is demonstrated. The experiments environments are ranging from low-dimensional spaces to high-dimensional spaces and from discrete action spaces to continuous actions spaces. For future work, can it be realized in the real world?<br />
# There is a doubt that whether it is a correct method to use the trajectory that was insufficiently learned in dense-reward environment as the imperfect demonstration.<br />
# In this paper, the performance only is judged by the cumulative reward, can other evaluation terms be considered? For example, the convergence rate.<br />
<br />
=References=<br />
[1] Nair, A., McGrew, B., Andrychowicz, M., Zaremba, W., and Abbeel, P. Overcoming exploration in reinforcement learning with demonstrations. arXiv preprint arXiv:1709.10089, 2017.<br />
<br />
[2] Hester, T., Vecerik, M., Pietquin, O., Lanctot, M., Schaul, T., Piot, B., Sendonaris, A., Dulac-Arnold, G., Osband, I., Agapiou, J., et al. Learning from demonstrations for real world reinforcement learning. arXiv preprint arXiv:1704.03732, 2017.<br />
<br />
[3] Večerík, M., Hester, T., Scholz, J., Wang, F., Pietquin, O., Piot, B., Heess, N., Rotho ̈rl, T., Lampe, T., and Riedmiller, M. Leveraging demonstrations for deep reinforcement learning on robotics problems with sparse rewards. arXiv preprint arXiv:1707.08817, 2017.<br />
<br />
[4] Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., et al. Mastering the game of go with deep neural networks and tree search. nature, 529(7587):484–489, 2016.<br />
<br />
[5] Todorov, E., Erez, T., and Tassa, Y. Mujoco: A physics engine for model-based control. In Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ International Con- ference on, pp. 5026–5033. IEEE, 2012.<br />
<br />
[6] Schaal, S. Learning from demonstration. In Advances in neural information processing systems, pp. 1040–1046, 1997.<br />
<br />
[7] Kim, B., Farahmand, A.-m., Pineau, J., and Precup, D. Learning from limited demonstrations. In Advances in Neural Information Processing Systems, pp. 2859–2867, 2013.<br />
<br />
[8] Piot, B., Geist, M., and Pietquin, O. Boosted bellman resid- ual minimization handling expert demonstrations. In Joint European Conference on Machine Learning and Knowl- edge Discovery in Databases, pp. 549–564. Springer, 2014.<br />
<br />
[9] Aravind S. Lakshminarayanan, Sherjil Ozair, Y. B. Rein- forcement learning with few expert demonstrations. In NIPS workshop, 2016.<br />
<br />
[10] Brys, T., Harutyunyan, A., Suay, H. B., Chernova, S., Tay- lor, M. E., and Nowe ́, A. Reinforcement learning from demonstration through shaping. In IJCAI, pp. 3352–3358, 2015.<br />
<br />
[11] Ng, A. Y., Russell, S. J., et al. Algorithms for inverse reinforcement learning. In Icml, pp. 663–670, 2000.<br />
<br />
[12] Syed, U. and Schapire, R. E. A game-theoretic approach to apprenticeship learning. In Advances in neural informa- tion processing systems, pp. 1449–1456, 2008.<br />
<br />
[13] Syed, U., Bowling, M., and Schapire, R. E. Apprenticeship learning using linear programming. In Proceedings of the 25th international conference on Machine learning, pp. 1032–1039. ACM, 2008.<br />
<br />
[14] Ho, J. and Ermon, S. Generative adversarial imitation learn- ing. In Advances in Neural Information Processing Sys- tems, pp. 4565–4573, 2016.<br />
<br />
[15] Sutton, R. S. and Barto, A. G. Reinforcement learning: An introduction, volume 1. MIT press Cambridge, 1998.<br />
<br />
[16] Kakade, S. M. A natural policy gradient. In Advances in neural information processing systems, pp. 1531–1538, 2002.<br />
<br />
[17] Schulman, J., Levine, S., Abbeel, P., Jordan, M., and Moritz, P. Trust region policy optimization. In Proceedings of the 32nd International Conference on Machine Learning (ICML-15), pp. 1889–1897, 2015.<br />
<br />
[18] Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.<br />
<br />
[19] Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. Generative adversarial nets. In Advances in neural information processing systems, pp. 2672–2680, 2014.<br />
<br />
[20] Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J., Tang, J., and Zaremba, W. Openai gym, 2016.<br />
<br />
[21] Lillicrap, T. P., Hunt, J. J., Pritzel, A., Heess, N., Erez, T., Tassa, Y., Silver, D., and Wierstra, D. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015.<br />
<br />
[22] Moore, A. W. Efficient memory-based learning for robot control. 1990.<br />
<br />
[23] Zhu, Y., Wang, Z., Merel, J., Rusu, A., Erez, T., Cabi, S., Tunyasuvunakool, S., Kramar, J., Hadsell, R., de Freitas, N., et al. Reinforcement and imitation learning for diverse visuomotor skills. arXiv preprint arXiv:1802.09564, 2018.<br />
<br />
[24] Li, Y., Song, J., and Ermon, S. Infogail: Interpretable imitation learning from visual demonstrations. In Advances in Neural Information Processing Systems, pp. 3815–3825, 2017.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Visual_Reinforcement_Learning_with_Imagined_Goals&diff=41362Visual Reinforcement Learning with Imagined Goals2018-11-26T03:26:05Z<p>Wfisher: Add graph from results (T)</p>
<hr />
<div>Video and details of this work is available [https://sites.google.com/site/visualrlwithimaginedgoals/ here]<br />
<br />
=Introduction and Motivation=<br />
<br />
Humans are able to accomplish many tasks without any explicit or supervised training, simply by exploring their environment. We are able to set our own goals and learn from our experiences, and thus able to accomplish specific tasks without ever having been trained explicitly for them. It would be ideal if an autonomous agent can also set its own goals and learn from its environment.<br />
<br />
In the paper “Visual Reinforcement Learning with Imagined Goals”, the authors are able to devise such an unsupervised reinforcement learning system. They introduce a system that sets abstract goals and autonomously learns to achieve those goals. They then show that the system can use these autonomously learned skills to perform a variety of user-specified goals, such as pushing objects, grasping objects, and opening doors, without any additional learning. Lastly, they demonstrate that their method is efficient enough to work in the real world on a Sawyer robot. The robot learns to set and achieve goals with only images as the input to the system.<br />
<br />
The algorithm proposed by the authors is summarised below. A Variational Auto Encoder on the (left) is trained to learn a latent representation of images gathered during training time (center). These latent variables can then be used to train a policy on imagined goals (center), which can then be used for accomplishing user-specified goals (right).<br />
<br />
[[File: WF_Sec_11Nov25_01.png | 800px]]<br />
<br />
=Related Work =<br />
<br />
Many previous works on vision-based deep reinforcement learning for robotics studied a variety of behaviours such as grasping [1], pushing [2], navigation [3], and other manipulation tasks [4]. However, their assumptions on the models limit their suitability for training general-purpose robots. Some scholars proposed time-varying models which require episodic setups. There are also scholars that propose an approach that uses goal images, but it requires instrumented training simulations. There is no example that uses model-free RL that learns policies to train on real-world robotic systems without having ground-truth information. <br />
<br />
In this paper, the authors utilize a goal-conditioned value function to tackle more general tasks through goal relabeling, which improves sample efficiency. Specifically, they use a model-free Q-learning method that operates on raw state observations and actions.<br />
<br />
Unsupervised learning has been used in a number of prior works to acquire better representations of RL. In these methods, the learned representation is used as a substitute for the state for the policy. However, these methods require additional information, such as access to the ground truth reward function based on the true state during training time [5], expert trajectories [6], human demonstrations [7], or pre-trained object-detection features [8]. In contrast, the authors learn to generate goals and use the learned representation to get a reward function for those goals without any of these extra sources of supervision.<br />
<br />
=Goal-Conditioned Reinforcement Learning=<br />
<br />
The ultimate goal in reinforcement learning is to learn a policy, that when given a state and goal, can dictate the optimal action. In this paper, goals are not explicitly defined during training. If a goal is not explicitly defined, the agent must be able to generate a set of synthetic goals automatically. Thus, suppose we let an autonomous agent explore an environment with a random policy. After executing each action, state observations are collected and stored. These state observations are structured in the form of images. The agent can randomly select goals from the set of state observations, and can also randomly select initial states from the set of state observations.<br />
<br />
[[File:human-giving-goal.png|center|thumb|400px|The task: Make the world look like this image. [9]]]<br />
<br />
Now given a set of all possible states, a goal, and an initial state, a reinforcement learning framework can be used to find the optimal policy such that the value function is maximized. However, to implement such a framework, a reward function needs to be defined. One choice for the reward is the negative distance between the current state and the goal state, so that maximizing the reward corresponds to minimizing the distance to a goal state.<br />
<br />
In reinforcement learning, a goal-conditioned Q function can be used to find a single policy to maximize rewards and therefore reach goal states. A goal-conditioned Q function Q(s,a,g) tells us how good an action a is, given the current state s and goal g. For example, a Q function tells us, “How good is it to move my hand up (action a), if I’m holding a plate (state s) and want to put the plate on the table (goal g)?” Once this Q function is trained, a goal-conditioned policy can be obtained by performing the following optimization<br />
<br />
[[File:policy-extraction.png|center|600px]]<br />
<br />
which effectively says, “choose the best action according to this Q function.” By using this procedure, one can obtain a policy that maximizes the sum of rewards, i.e. reaches various goals.<br />
<br />
The reason why Q learning is popular is that in can be train in an off-policy manner. Therefore, the only things Q function needs are samples of state, action, next state, goal, and reward: (s,a,s′,g,r). This data can be collected by any policy and can be reused across multiples tasks. So a preliminary goal-conditioned Q-learning algorithm looks like this:<br />
<br />
[[File:ql.png|center|600px]]<br />
<br />
The main drawback in this training procedure is collecting data. In theory, one could learn to solve various tasks without even interacting with the world if more data are available. Unfortunately, it is difficult to learn an accurate model of the world, so sampling are usually used to get state-action-next-state data, (s,a,s′). However, if the reward function r(s,g) can be accessed, one can retroactively relabeled goals and recompute rewards. In this way, more data can be artificially generated given a single (s,a,s′) tuple. So, the training procedure can be modified like so:<br />
<br />
[[File:qlr.png|center|600px]]<br />
<br />
This goal resampling makes it possible to simultaneously learn how to reach multiple goals at once without needing more data from the environment. Thus, this simple modification can result in substantially faster learning. However, the method described above makes two major assumptions: (1) you have access to a reward function and (2) you have access to a goal sampling distribution p(g). When moving to vision-based tasks where goals are images, both of these assumptions introduce practical concerns.<br />
<br />
For one, a fundamental problem with this reward function is that it assumes that the distance between raw images will yield semantically useful information. Images are noisy. A large amount of information in an image that may not be related to the object we analyze. Thus, the distance between two images may not correlate with their semantic distance.<br />
<br />
Second, because the goals are images, a goal image distribution p(g) is needed so that one can sample goal images. Manually designing a distribution over goal images is a non-trivial task and image generation is still an active field of research. It would be ideal if the agent can autonomously imagine its own goals and learn how to reach them.<br />
<br />
=Variational Autoencoder (VAE)=<br />
An autoencoder is a type of machine learning model that can learn to extract a robust, space-efficient feature vector from an image. This generative model converts high-dimensional observations x, like images, into low-dimensional latent variables z, and vice versa. The model is trained so that the latent variables capture the underlying factors of variation in an image. A current image x and goal image xg can be converted into latent variables z and zg, respectively. These latent variables can then be used to represent ate the state and goal for the reinforcement learning algorithm. Learning Q functions and policies on top of this low-dimensional latent space rather than directly on images results in faster learning.<br />
<br />
[[File:robot-interpreting-scene.png|center|thumb|600px|The agent encodes the current image (x) and goal image (xg) into a latent space and use distances in that latent space for reward. [9]]]<br />
<br />
Using the latent variable representations for the images and goals also solves the problem of computing rewards. Instead of using pixel-wise error as our reward, the distance in the latent space is used as the reward to train the agent to reach a goal. The paper shows that this corresponds to rewarding reaching states that maximize the probability of the latent goal zg.<br />
<br />
This generative model is also important because it allows an agent to easily generate goals in the latent space. In particular, the authors design the generative model so that latent variables are sampled from the VAE prior. This sampling mechanism is used for two reasons: First, it provides a mechanism for an agent to set its own goals. The agent simply samples a value for the latent variable from the generative model, and tries to reach that latent goal. Second, this resampling mechanism is also used to relabel goals as mentioned above. Since the VAE prior is trained by real images, meaningful latent goals can be sampled from the latent variable prior. This will help the agent set its own goals and practice towards them if no goal is provided at test time.<br />
<br />
[[File:robot-imagining-goals.png|center|thumb|600px|Even without a human providing a goal, our agent can still generate its own goals, both for exploration and for goal relabeling. [9]]]<br />
<br />
The authors summarize the purpose of the latent variable representation of images as follows: (1) captures the underlying factors of a scene, (2) provides meaningful distances to optimize, and (3) provides an efficient goal sampling mechanism which can be used by the agent to generate its own goals. The overall method is called reinforcement learning with imagined goals (RIG) by the authors.<br />
The process involves starts with collecting data through a simple exploration policy. Possible alternative explorations could be employed here including off-the-shelf exploration bonuses or unsupervised reinforcement learning methods. Then, a VAE latent variable model is trained on state observations and fine-tuned during training. The latent variable model is used for multiple purposes: sampling a latent goal <math>zg</math> from the model and conditioning the policy on this goal. All states and goals are embedded using the model’s encoder and then used to train the goal-conditioned value function. The authors then resample goals from the prior and compute rewards in the latent space.<br />
<br />
=Algorithm=<br />
[[File:algorithm1.png|center|thumb|600px|]]<br />
<br />
The data is first collected via a simple exploration policy and then train a VAE latent variable model on state observations and then fine tune over the course of training. When the goal-conditioned value function is trained, the authors resample prior goals and compute rewards in the latent space.<br />
<br />
=Experiments=<br />
<br />
The authors evaluated their method against some prior algorithms and ablated versions of their approach on a suite of simulated and real-world tasks: Visual Reacher, Visual Pusher, and Visual Multi-Object Pusher. They compared their model with the following prior works: L&R, DSAE, HER, and Oracle. It is concluded that their approach substantially outperforms the previous methods and is close to the state-based "oracle" method in terms of efficiency and performance.<br />
<br />
The figure below shows the performance of different algorithms on this task. This involved a simulated environment with a Sawyer arm. The authors' algorithm was given only visual input, and the available controls were end-effector velocity. The plots show the distance to the goal state as a function of simulation steps. The oracle, as a baseline, was given true object location information, as opposed to visual pixel information.<br />
<br />
[[File:WF_Sec_11Nov_25_02.png|1000px]]<br />
<br />
<br />
They then investigated the effectiveness of distances in the VAE latent space for the Visual Pusher task. They observed that latent distance significantly outperforms the log probability and pixel mean-squared error. The resampling strategies are also varied while fixing other components of the algorithm to study the effect of relabeling strategy. In this experiment, the RIG, which is an equal mixture of the VAE and Future sampling strategies, performs best. Subsequently, learning with variable numbers of objects was studied by evaluating on a task where the environment, based on the Visual Multi-Object Pusher, randomly contains zero, one, or two objects during testing. The results show that their model can tackle this task successfully.<br />
<br />
Finally, the authors tested the RIG in a real-world robot for its ability to reach user-specified positions and push objects to desired locations, as indicated by a goal image. The robot is trained with access only to 84x84 RGB images and without access to joint angles or object positions. The robot first learns by settings its own goals in the latent space and autonomously practices reaching different positions without human involvement. After a reasonable amount of time of training, the robot is given a goal image. Because the robot has practiced reaching so many goals, it is able to reach this goal without additional training:<br />
<br />
[[File:reaching.JPG|center|thumb|600px|(Left) The robot setup is pictured. (Right) Test rollouts of the learned policy.]]<br />
<br />
The method for reaching only needs 10,000 samples and an hour of real-world interactions.<br />
<br />
They also used RIG to train a policy to push objects to target locations:<br />
<br />
[[File:pushing.JPG|center|thumb|600px|The robot pushing setup is<br />
pictured, with frames from test rollouts of the learned policy.]]<br />
<br />
The pushing task is more complicated and the method requires about 25,000 samples. Since the authors do not have the true position during training, so they used test episode returns as the VAE latent distance reward.<br />
<br />
=Conclusion & Future Work=<br />
<br />
In this paper, a new RL algorithm is proposed to efficiently solve goal-conditioned, vision-based tasks without any ground truth state information or reward functions. The author suggests that one could instead use other representations, such as language and demonstrations, to specify goals. Also, while the paper provides a mechanism to sample goals for autonomous exploration, one can combine the proposed method with existing work by choosing these goals in a more principled way, i.e. a procedure that is not only goal-oriented, but also information seeking or uncertainty aware, to perform even better exploration. Furthermore, combining the idea of this paper with methods from multitask learning and meta-learning is a promising path to create general-purpose agents that can continuously and efficiently acquire skill. Lastly, there are a variety of robot tasks whose state representation would be difficult to capture with sensors, such as manipulating deformable objects or handling scenes with variable number of objects. It is interesting to see whether the RIG can be scaled up to solve these tasks. [10] A new paper was published last week that built on the framework of goal conditioned Reinforcement Learning to extract state representations based on the actions required to reach them, which is abbreviated ARC for actionable representation for control.<br />
<br />
=Critique=<br />
1. This paper is novel because it uses visual data and trains in an unsupervised fashion. The algorithm has no access to a ground truth state or to a pre-defined reward function. It can perform well in a real-world environment with no explicit programming.<br />
<br />
2. From the videos, one major concern is that the output of robotic arm's position is not stable during training and test time. It is likely that the encoder reduces the image features too much so that the images in the latent space are too blury to be used goal images. It would be better if this can be investigated in future. It would be better, if a method is investigated with multiple data sources, and the agent is trained to choose that source which has more complete information. <br />
<br />
3. The algorithm seems to perform better when there is only one object in the images. For example, in Visual Multi-Object Pusher experiment, the relative positions of two pucks do not correspond well with the relative positions of two pucks in goal images. The same situation is also observed in Variable-object experiment. We may guess that the more information contain in a image, the less likely the robot will perform well. This limits the applicability of the current algorithm to solving real-world problems.<br />
<br />
4. The instability mentioned in #2 is even more apparent in the multi-object scenario, and appears to result from the model attempting to optimize on the position of both objects at the same time. Reducing the problem to a sequence of single-object targets may reduce the amount of time the robots spends moving between the multiple objects in the scene (which it currently does quite frequently). <br />
<br />
=References=<br />
1. Lerrel Pinto, Marcin Andrychowicz, Peter Welinder, Wojciech Zaremba, and Pieter Abbeel. Asymmetric<br />
Actor Critic for Image-Based Robot Learning. arXiv preprint arXiv:1710.06542, 2017.<br />
<br />
2. Pulkit Agrawal, Ashvin Nair, Pieter Abbeel, Jitendra Malik, and Sergey Levine. Learning to Poke by<br />
Poking: Experiential Learning of Intuitive Physics. In Advances in Neural Information Processing Systems<br />
(NIPS), 2016.<br />
<br />
3. Deepak Pathak, Parsa Mahmoudieh, Guanghao Luo, Pulkit Agrawal, Dian Chen, Yide Shentu, Evan<br />
Shelhamer, Jitendra Malik, Alexei A Efros, and Trevor Darrell. Zero-Shot Visual Imitation. In International<br />
Conference on Learning Representations (ICLR), 2018.<br />
<br />
4. Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David<br />
Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. In International<br />
Conference on Learning Representations (ICLR), 2016.<br />
<br />
5. Irina Higgins, Arka Pal, Andrei A Rusu, Loic Matthey, Christopher P Burgess, Alexander Pritzel, Matthew<br />
Botvinick, Charles Blundell, and Alexander Lerchner. Darla: Improving zero-shot transfer in reinforcement<br />
learning. International Conference on Machine Learning (ICML), 2017.<br />
<br />
6. Aravind Srinivas, Allan Jabri, Pieter Abbeel, Sergey Levine, and Chelsea Finn. Universal Planning<br />
Networks. In International Conference on Machine Learning (ICML), 2018.<br />
<br />
7. Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, and Sergey<br />
Levine. Time-contrastive networks: Self-supervised learning from video. arXiv preprint arXiv:1704.06888,<br />
2017.<br />
<br />
8. Alex Lee, Sergey Levine, and Pieter Abbeel. Learning Visual Servoing with Deep Features and Fitted<br />
Q-Iteration. In International Conference on Learning Representations (ICLR), 2017.<br />
<br />
9. Online source: https://bair.berkeley.edu/blog/2018/09/06/rig/<br />
<br />
10. https://arxiv.org/pdf/1811.07819.pdf</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=File:WF_Sec_11Nov_25_02.png&diff=41361File:WF Sec 11Nov 25 02.png2018-11-26T03:25:17Z<p>Wfisher: Source: https://arxiv.org/pdf/1807.04742.pdf</p>
<hr />
<div>Source: https://arxiv.org/pdf/1807.04742.pdf</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Visual_Reinforcement_Learning_with_Imagined_Goals&diff=41356Visual Reinforcement Learning with Imagined Goals2018-11-26T03:17:17Z<p>Wfisher: Add intro description of algorithm (T)</p>
<hr />
<div>Video and details of this work is available [https://sites.google.com/site/visualrlwithimaginedgoals/ here]<br />
<br />
=Introduction and Motivation=<br />
<br />
Humans are able to accomplish many tasks without any explicit or supervised training, simply by exploring their environment. We are able to set our own goals and learn from our experiences, and thus able to accomplish specific tasks without ever having been trained explicitly for them. It would be ideal if an autonomous agent can also set its own goals and learn from its environment.<br />
<br />
In the paper “Visual Reinforcement Learning with Imagined Goals”, the authors are able to devise such an unsupervised reinforcement learning system. They introduce a system that sets abstract goals and autonomously learns to achieve those goals. They then show that the system can use these autonomously learned skills to perform a variety of user-specified goals, such as pushing objects, grasping objects, and opening doors, without any additional learning. Lastly, they demonstrate that their method is efficient enough to work in the real world on a Sawyer robot. The robot learns to set and achieve goals with only images as the input to the system.<br />
<br />
The algorithm proposed by the authors is summarised below. A Variational Auto Encoder on the (left) is trained to learn a latent representation of images gathered during training time (center). These latent variables can then be used to train a policy on imagined goals (center), which can then be used for accomplishing user-specified goals (right).<br />
<br />
[[File: WF_Sec_11Nov25_01.png | 800px]]<br />
<br />
=Related Work =<br />
<br />
Many previous works on vision-based deep reinforcement learning for robotics studied a variety of behaviours such as grasping [1], pushing [2], navigation [3], and other manipulation tasks [4]. However, their assumptions on the models limit their suitability for training general-purpose robots. Some scholars proposed time-varying models which require episodic setups. There are also scholars that propose an approach that uses goal images, but it requires instrumented training simulations. There is no example that uses model-free RL that learns policies to train on real-world robotic systems without having ground-truth information. <br />
<br />
In this paper, the authors utilize a goal-conditioned value function to tackle more general tasks through goal relabeling, which improves sample efficiency. Specifically, they use a model-free Q-learning method that operates on raw state observations and actions.<br />
<br />
Unsupervised learning has been used in a number of prior works to acquire better representations of RL. In these methods, the learned representation is used as a substitute for the state for the policy. However, these methods require additional information, such as access to the ground truth reward function based on the true state during training time [5], expert trajectories [6], human demonstrations [7], or pre-trained object-detection features [8]. In contrast, the authors learn to generate goals and use the learned representation to get a reward function for those goals without any of these extra sources of supervision.<br />
<br />
=Goal-Conditioned Reinforcement Learning=<br />
<br />
The ultimate goal in reinforcement learning is to learn a policy, that when given a state and goal, can dictate the optimal action. In this paper, goals are not explicitly defined during training. If a goal is not explicitly defined, the agent must be able to generate a set of synthetic goals automatically. Thus, suppose we let an autonomous agent explore an environment with a random policy. After executing each action, state observations are collected and stored. These state observations are structured in the form of images. The agent can randomly select goals from the set of state observations, and can also randomly select initial states from the set of state observations.<br />
<br />
[[File:human-giving-goal.png|center|thumb|400px|The task: Make the world look like this image. [9]]]<br />
<br />
Now given a set of all possible states, a goal, and an initial state, a reinforcement learning framework can be used to find the optimal policy such that the value function is maximized. However, to implement such a framework, a reward function needs to be defined. One choice for the reward is the negative distance between the current state and the goal state, so that maximizing the reward corresponds to minimizing the distance to a goal state.<br />
<br />
In reinforcement learning, a goal-conditioned Q function can be used to find a single policy to maximize rewards and therefore reach goal states. A goal-conditioned Q function Q(s,a,g) tells us how good an action a is, given the current state s and goal g. For example, a Q function tells us, “How good is it to move my hand up (action a), if I’m holding a plate (state s) and want to put the plate on the table (goal g)?” Once this Q function is trained, a goal-conditioned policy can be obtained by performing the following optimization<br />
<br />
[[File:policy-extraction.png|center|600px]]<br />
<br />
which effectively says, “choose the best action according to this Q function.” By using this procedure, one can obtain a policy that maximizes the sum of rewards, i.e. reaches various goals.<br />
<br />
The reason why Q learning is popular is that in can be train in an off-policy manner. Therefore, the only things Q function needs are samples of state, action, next state, goal, and reward: (s,a,s′,g,r). This data can be collected by any policy and can be reused across multiples tasks. So a preliminary goal-conditioned Q-learning algorithm looks like this:<br />
<br />
[[File:ql.png|center|600px]]<br />
<br />
The main drawback in this training procedure is collecting data. In theory, one could learn to solve various tasks without even interacting with the world if more data are available. Unfortunately, it is difficult to learn an accurate model of the world, so sampling are usually used to get state-action-next-state data, (s,a,s′). However, if the reward function r(s,g) can be accessed, one can retroactively relabeled goals and recompute rewards. In this way, more data can be artificially generated given a single (s,a,s′) tuple. So, the training procedure can be modified like so:<br />
<br />
[[File:qlr.png|center|600px]]<br />
<br />
This goal resampling makes it possible to simultaneously learn how to reach multiple goals at once without needing more data from the environment. Thus, this simple modification can result in substantially faster learning. However, the method described above makes two major assumptions: (1) you have access to a reward function and (2) you have access to a goal sampling distribution p(g). When moving to vision-based tasks where goals are images, both of these assumptions introduce practical concerns.<br />
<br />
For one, a fundamental problem with this reward function is that it assumes that the distance between raw images will yield semantically useful information. Images are noisy. A large amount of information in an image that may not be related to the object we analyze. Thus, the distance between two images may not correlate with their semantic distance.<br />
<br />
Second, because the goals are images, a goal image distribution p(g) is needed so that one can sample goal images. Manually designing a distribution over goal images is a non-trivial task and image generation is still an active field of research. It would be ideal if the agent can autonomously imagine its own goals and learn how to reach them.<br />
<br />
=Variational Autoencoder (VAE)=<br />
An autoencoder is a type of machine learning model that can learn to extract a robust, space-efficient feature vector from an image. This generative model converts high-dimensional observations x, like images, into low-dimensional latent variables z, and vice versa. The model is trained so that the latent variables capture the underlying factors of variation in an image. A current image x and goal image xg can be converted into latent variables z and zg, respectively. These latent variables can then be used to represent ate the state and goal for the reinforcement learning algorithm. Learning Q functions and policies on top of this low-dimensional latent space rather than directly on images results in faster learning.<br />
<br />
[[File:robot-interpreting-scene.png|center|thumb|600px|The agent encodes the current image (x) and goal image (xg) into a latent space and use distances in that latent space for reward. [9]]]<br />
<br />
Using the latent variable representations for the images and goals also solves the problem of computing rewards. Instead of using pixel-wise error as our reward, the distance in the latent space is used as the reward to train the agent to reach a goal. The paper shows that this corresponds to rewarding reaching states that maximize the probability of the latent goal zg.<br />
<br />
This generative model is also important because it allows an agent to easily generate goals in the latent space. In particular, the authors design the generative model so that latent variables are sampled from the VAE prior. This sampling mechanism is used for two reasons: First, it provides a mechanism for an agent to set its own goals. The agent simply samples a value for the latent variable from the generative model, and tries to reach that latent goal. Second, this resampling mechanism is also used to relabel goals as mentioned above. Since the VAE prior is trained by real images, meaningful latent goals can be sampled from the latent variable prior. This will help the agent set its own goals and practice towards them if no goal is provided at test time.<br />
<br />
[[File:robot-imagining-goals.png|center|thumb|600px|Even without a human providing a goal, our agent can still generate its own goals, both for exploration and for goal relabeling. [9]]]<br />
<br />
The authors summarize the purpose of the latent variable representation of images as follows: (1) captures the underlying factors of a scene, (2) provides meaningful distances to optimize, and (3) provides an efficient goal sampling mechanism which can be used by the agent to generate its own goals. The overall method is called reinforcement learning with imagined goals (RIG) by the authors.<br />
The process involves starts with collecting data through a simple exploration policy. Possible alternative explorations could be employed here including off-the-shelf exploration bonuses or unsupervised reinforcement learning methods. Then, a VAE latent variable model is trained on state observations and fine-tuned during training. The latent variable model is used for multiple purposes: sampling a latent goal <math>zg</math> from the model and conditioning the policy on this goal. All states and goals are embedded using the model’s encoder and then used to train the goal-conditioned value function. The authors then resample goals from the prior and compute rewards in the latent space.<br />
<br />
=Algorithm=<br />
[[File:algorithm1.png|center|thumb|600px|]]<br />
<br />
The data is first collected via a simple exploration policy and then train a VAE latent variable model on state observations and then fine tune over the course of training. When the goal-conditioned value function is trained, the authors resample prior goals and compute rewards in the latent space.<br />
<br />
=Experiments=<br />
<br />
The authors evaluated their method against some prior algorithms and ablated versions of their approach on a suite of simulated and real-world tasks: Visual Reacher, Visual Pusher, and Visual Multi-Object Pusher. They compared their model with the following prior works: L&R, DSAE, HER, and Oracle. It is concluded that their approach substantially outperforms the previous methods and is close to the state-based "oracle" method in terms of efficiency and performance.<br />
<br />
They then investigated the effectiveness of distances in the VAE latent space for the Visual Pusher task. They observed that latent distance significantly outperforms the log probability and pixel mean-squared error. The resampling strategies are also varied while fixing other components of the algorithm to study the effect of relabeling strategy. In this experiment, the RIG, which is an equal mixture of the VAE and Future sampling strategies, performs best. Subsequently, learning with variable numbers of objects was studied by evaluating on a task where the environment, based on the Visual Multi-Object Pusher, randomly contains zero, one, or two objects during testing. The results show that their model can tackle this task successfully.<br />
<br />
Finally, the authors tested the RIG in a real-world robot for its ability to reach user-specified positions and push objects to desired locations, as indicated by a goal image. The robot is trained with access only to 84x84 RGB images and without access to joint angles or object positions. The robot first learns by settings its own goals in the latent space and autonomously practices reaching different positions without human involvement. After a reasonable amount of time of training, the robot is given a goal image. Because the robot has practiced reaching so many goals, it is able to reach this goal without additional training:<br />
<br />
[[File:reaching.JPG|center|thumb|600px|(Left) The robot setup is pictured. (Right) Test rollouts of the learned policy.]]<br />
<br />
The method for reaching only needs 10,000 samples and an hour of real-world interactions.<br />
<br />
They also used RIG to train a policy to push objects to target locations:<br />
<br />
[[File:pushing.JPG|center|thumb|600px|The robot pushing setup is<br />
pictured, with frames from test rollouts of the learned policy.]]<br />
<br />
The pushing task is more complicated and the method requires about 25,000 samples. Since the authors do not have the true position during training, so they used test episode returns as the VAE latent distance reward.<br />
<br />
=Conclusion & Future Work=<br />
<br />
In this paper, a new RL algorithm is proposed to efficiently solve goal-conditioned, vision-based tasks without any ground truth state information or reward functions. The author suggests that one could instead use other representations, such as language and demonstrations, to specify goals. Also, while the paper provides a mechanism to sample goals for autonomous exploration, one can combine the proposed method with existing work by choosing these goals in a more principled way, i.e. a procedure that is not only goal-oriented, but also information seeking or uncertainty aware, to perform even better exploration. Furthermore, combining the idea of this paper with methods from multitask learning and meta-learning is a promising path to create general-purpose agents that can continuously and efficiently acquire skill. Lastly, there are a variety of robot tasks whose state representation would be difficult to capture with sensors, such as manipulating deformable objects or handling scenes with variable number of objects. It is interesting to see whether the RIG can be scaled up to solve these tasks. [10] A new paper was published last week that built on the framework of goal conditioned Reinforcement Learning to extract state representations based on the actions required to reach them, which is abbreviated ARC for actionable representation for control.<br />
<br />
=Critique=<br />
1. This paper is novel because it uses visual data and trains in an unsupervised fashion. The algorithm has no access to a ground truth state or to a pre-defined reward function. It can perform well in a real-world environment with no explicit programming.<br />
<br />
2. From the videos, one major concern is that the output of robotic arm's position is not stable during training and test time. It is likely that the encoder reduces the image features too much so that the images in the latent space are too blury to be used goal images. It would be better if this can be investigated in future. It would be better, if a method is investigated with multiple data sources, and the agent is trained to choose that source which has more complete information. <br />
<br />
3. The algorithm seems to perform better when there is only one object in the images. For example, in Visual Multi-Object Pusher experiment, the relative positions of two pucks do not correspond well with the relative positions of two pucks in goal images. The same situation is also observed in Variable-object experiment. We may guess that the more information contain in a image, the less likely the robot will perform well. This limits the applicability of the current algorithm to solving real-world problems.<br />
<br />
4. The instability mentioned in #2 is even more apparent in the multi-object scenario, and appears to result from the model attempting to optimize on the position of both objects at the same time. Reducing the problem to a sequence of single-object targets may reduce the amount of time the robots spends moving between the multiple objects in the scene (which it currently does quite frequently). <br />
<br />
=References=<br />
1. Lerrel Pinto, Marcin Andrychowicz, Peter Welinder, Wojciech Zaremba, and Pieter Abbeel. Asymmetric<br />
Actor Critic for Image-Based Robot Learning. arXiv preprint arXiv:1710.06542, 2017.<br />
<br />
2. Pulkit Agrawal, Ashvin Nair, Pieter Abbeel, Jitendra Malik, and Sergey Levine. Learning to Poke by<br />
Poking: Experiential Learning of Intuitive Physics. In Advances in Neural Information Processing Systems<br />
(NIPS), 2016.<br />
<br />
3. Deepak Pathak, Parsa Mahmoudieh, Guanghao Luo, Pulkit Agrawal, Dian Chen, Yide Shentu, Evan<br />
Shelhamer, Jitendra Malik, Alexei A Efros, and Trevor Darrell. Zero-Shot Visual Imitation. In International<br />
Conference on Learning Representations (ICLR), 2018.<br />
<br />
4. Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David<br />
Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. In International<br />
Conference on Learning Representations (ICLR), 2016.<br />
<br />
5. Irina Higgins, Arka Pal, Andrei A Rusu, Loic Matthey, Christopher P Burgess, Alexander Pritzel, Matthew<br />
Botvinick, Charles Blundell, and Alexander Lerchner. Darla: Improving zero-shot transfer in reinforcement<br />
learning. International Conference on Machine Learning (ICML), 2017.<br />
<br />
6. Aravind Srinivas, Allan Jabri, Pieter Abbeel, Sergey Levine, and Chelsea Finn. Universal Planning<br />
Networks. In International Conference on Machine Learning (ICML), 2018.<br />
<br />
7. Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, and Sergey<br />
Levine. Time-contrastive networks: Self-supervised learning from video. arXiv preprint arXiv:1704.06888,<br />
2017.<br />
<br />
8. Alex Lee, Sergey Levine, and Pieter Abbeel. Learning Visual Servoing with Deep Features and Fitted<br />
Q-Iteration. In International Conference on Learning Representations (ICLR), 2017.<br />
<br />
9. Online source: https://bair.berkeley.edu/blog/2018/09/06/rig/<br />
<br />
10. https://arxiv.org/pdf/1811.07819.pdf</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=File:WF_Sec_11Nov25_01.png&diff=41353File:WF Sec 11Nov25 01.png2018-11-26T03:11:58Z<p>Wfisher: </p>
<hr />
<div></div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Unsupervised_Neural_Machine_Translation&diff=41351Unsupervised Neural Machine Translation2018-11-26T03:01:54Z<p>Wfisher: Misc Edits (T,E)</p>
<hr />
<div>This paper was published in ICLR 2018, authored by Mikel Artetxe, Gorka Labaka, Eneko Agirre, and Kyunghyun Cho. Open source implementation of this paper is available [https://github.com/artetxem/undreamt here]<br />
<br />
= Introduction =<br />
The paper presents an unsupervised Neural Machine Translation(NMT) method that uses monolingual corpora (single language texts) only. This contrasts with the usual supervised NMT approach which relies on parallel corpora (aligned text) from the source and target languages being available for training. This problem is important because parallel pairing for a majority of languages, e.g. for German-Russian, do not exist.<br />
<br />
Other authors have recently tried to address this problem using semi-supervised approaches (small set of parallel corpora). However, these methods still require a strong cross-lingual signal. The proposed method eliminates the need for cross-lingual information all together and relies solely on monolingual data. The proposed method builds upon the work done recently on unsupervised cross-lingual embeddings by Artetxe et al., 2017 and Zhang et al., 2017.<br />
<br />
The general approach of the methodology is to:<br />
<br />
# Use monolingual corpora in the source and target languages to learn single language word embeddings for both languages separately.<br />
# Align the 2 sets of word embeddings into a single cross lingual (language independent) embedding.<br />
Then iteratively perform:<br />
# Train an encoder-decoder model to reconstruct noisy versions of sentences in both source and target languages separately. The model uses a single encoder and different decoders for each language. The encoder uses cross lingual word embedding.<br />
# Tune the decoder in each language by back-translating between the source and target language.<br />
<br />
= Background =<br />
<br />
===Word Embedding Alignment===<br />
<br />
The paper uses word2vec [Mikolov, 2013] to convert each monolingual corpora to vector embeddings. These embeddings have been shown to contain the contextual and syntactic features independent of language, and so, in theory, there could exist a linear map that maps the embeddings from language L1 to language L2. <br />
<br />
Figure 1 shows an example of aligning the word embeddings in English and French.<br />
<br />
[[File:Figure1_lwali.png|frame|400px|center|Figure 1: the word embeddings in English and French (a & b), and (c) shows the aligned word embeddings after some linear transformation.[Gouws,2016]]]<br />
<br />
Most cross-lingual word embedding methods use bilingual signals in the form of parallel corpora. Usually, the embedding mapping methods train the embeddings in different languages using monolingual corpora, then use a linear transformation to map them into a shared space based on a bilingual dictionary.<br />
<br />
The paper uses the methodology proposed by [Artetxe, 2017] to do cross-lingual embedding aligning in an unsupervised manner and without parallel data. Without going into the details, the general approach of this paper is starting from a seed dictionary of numeral pairings (e.g. 1-1, 2-2, etc.), to iteratively learn the mapping between 2 language embeddings, while concurrently improving the dictionary with the learned mapping at each iteration. This is in contrast to earlier work which used dictionaries of a few thousand words.<br />
<br />
===Other related work and inspirations===<br />
====Statistical Decipherment for Machine Translation====<br />
There has been significant work in statistical deciphering techniques (decipherment is the discovery of the meaning of texts written in ancient or obscure languages or scripts) to develop a machine translation model from monolingual data (Ravi & Knight, 2011; Dou & Knight, 2012). These techniques treat the source language as ciphertext (known as encrypted or encoded information because it contains a form of the original plain text that is unreadable by a human or computer without the proper cipher to decrypt it) and model the generation process of the ciphertext as a two-stage process including the generation of the original English sequence and the probabilistic replacement of the words in it. This approach is able to take the advantage of the incorporation of syntactic knowledge of the languages. The use of word embeddings has also shown improvements in statistical decipherment.<br />
<br />
====Low-Resource Neural Machine Translation====<br />
There are also proposals that use techniques other than direct parallel corpora to do neural machine translation(NMT). Some use a third intermediate language that is well connected to the source and target languages independently. For example, if we want to translate German into Russian, we can use English as an intermediate language(German-English and then English-Russian) since there are plenty of resources to connect English and other languages. Johnson et al. (2017) show that a multilingual extension of a standard NMT architecture performs reasonably well for language pairs when no parallel data for the source and target data was used during training.<br />
<br />
Other works use monolingual data in combination with scarce parallel corpora. A simple but effective technique is back-translation [Sennrich et al, 2016]. First, a synthetic parallel corpus in the target language is created. Translated sentence and back translated to the source language and compared with the original sentence.<br />
<br />
The most important contribution to the problem of training an NMT model with monolingual data was from [He, 2016], which trains two agents to translate in opposite directions (e.g. French → English and English → French) and teach each other through reinforcement learning. However, this approach still required a large parallel corpus for a warm start (about 1.2 million sentences), while this paper does not use parallel data.<br />
<br />
= Methodology =<br />
<br />
The corpora data is first preprocessed in a standard way to tokenize and case the words. The authors also experiment with an alternative way of tokenizing words by using Byte-Pair Encoding (BPE) [Sennrich, 2016]. BPE has been shown to improve embeddings of rare-words. The vocabulary was limited to the most frequent 50,000 tokens (BPE tokens or words).<br />
<br />
The tokens are then converted to word embeddings using word2vec with 300 dimensions and then aligned between languages using the method proposed by [Artetxe, 2017]. The alignment method proposed by [Artetxe, 2017] is also used as a baseline to evaluate this model as discussed later in Results.<br />
<br />
The translation model uses a standard encoder-decoder model with attention. The encoder is a 2-layer bidirectional RNN, and the decoder is a 2 layer RNN. All RNNs use GRU cells with 600 hidden units. The encoder is shared by the source and target language, while the decoder is different for each language.<br />
<br />
Although the architecture uses standard models, the proposed system differs from the standard NMT through 3 aspects:<br />
<br />
#Dual structure: NMT usually are built for one direction translations English<math>\rightarrow</math>French or French<math>\rightarrow</math>English, whereas the proposed model trains both directions at the same time translating English<math>\leftrightarrow</math>French.<br />
#Shared encoder: one encoder is shared for both source and target languages in order to produce a representation in the latent space independent of language, and each decoder learns to transform the representation back to its corresponding language. <br />
#Fixed embeddings in the encoder: Most NMT systems initialize the embeddings and update them during training, whereas the proposed system trains the embeddings in the beginning and keeps these fixed throughout training, so the encoder receives language-independent representations of the words. This requires existing unsupervised methods to create embeddings using monolingual corpora as discussed in the background.<br />
<br />
[[File:Figure2_lwali.png|600px|center]]<br />
<br />
The translation model iteratively improves the encoder and decoder by performing 2 tasks: Denoising, and Back-translation.<br />
<br />
===Denoising===<br />
<br />
Random noise is added to the input sentences in order to allow the model to learn some structure of languages. Without noise, the model would simply learn to copy the input word by word. Noise also allows the shared encoder to compose the embeddings of both languages in a language-independent fashion, and then be decoded by the language dependent decoder.<br />
<br />
Denoising works by reconstructing a noisy version of a sentence back into the original sentence in the same language. In mathematical form, if <math>x</math> is a sentence in language L1:<br />
<br />
# Construct <math>C(x)</math>, noisy version of <math>x</math>,<br />
# Input <math>C(x)</math> into the current iteration of the shared encoder and use decoder for L1 to get reconstructed <math>\hat{x}</math>.<br />
<br />
The training objective is to minimize the cross entropy loss between <math>{x}</math> and <math>\hat{x}</math>.<br />
<br />
In other words, the whole system is optimized to take an input sentence in a given language, encode it using the shared encoder, and reconstruct the original sentence using the decoder of that language.<br />
<br />
The proposed noise function is to perform <math>N/2</math> random swaps of words that are contiguous, where <math>N</math> is the number of words in the sentence. This noise model also helps reduce reliance of the model on the order of words in a sentence which may be different in the source and target languages. The system will also need to correctly learn the of a language to decode the sentence into the correct order.<br />
<br />
===Back-Translation===<br />
<br />
With only denoising, the system doesn't have a goal to improve the actual translation. Back-translation works by using the decoder of the target language to create a translation, then encoding this translation and decoding again using the source decoder to reconstruct a the original sentence. In mathematical form, if <math>C(x)</math> is a noisy version of sentence <math>x</math> in language L1:<br />
<br />
# Input <math>C(x)</math> into the current iteration of shared encoder and the decoder in L2 to construct translation <math>y</math> in L2,<br />
# Construct <math>C(y)</math>, noisy version of translation <math>y</math>,<br />
# Input <math>C(y)</math> into the current iteration of shared encoder and the decoder in L1 to reconstruct <math>\hat{x}</math> in L1.<br />
<br />
The training objective is to minimize the cross entropy loss between <math>{x}</math> and <math>\hat{x}</math>.<br />
<br />
Contrary to standard back-translation that uses an independent model to back-translate the entire corpus at one time, the system uses mini-batches and the dual architecture to generate pseudo-translations and then train the model with the translation, improving the model iteratively as the training progresses.<br />
<br />
===Training===<br />
<br />
Training is done by alternating these 2 objectives from mini-batch to mini-batch. Each iteration would perform one mini-batch of denoising for L1, another one for L2, one mini-batch of back-translation from L1 to L2, and another one from L2 to L1. The procedure is repeated until convergence. <br />
During decoding, greedy decoding was used at training time for back-translation, but actual inference at test time was done using beam-search with a beam size of 12.<br />
<br />
Optimizer choice and other hyperparameters can be found in the paper.<br />
<br />
=Experiments and Results=<br />
<br />
The model is evaluated using the Bilingual Evaluation Understudy (BLEU) Score, which is typically used to evaluate the quality of the translation, using a reference (ground-truth) translation.<br />
<br />
The paper trains translation model under 3 different settings to compare the performance (Table 1). All training and testing data used was from a standard NMT dataset, WMT'14.<br />
<br />
[[File:Table1_lwali.png|600px|center]]<br />
<br />
The results show that backtranslation is essential for the proposed system to work properly. The denoising technique alone is below the baseline while big improvements appear when introducing backtranslation.<br />
<br />
===Unsupervised===<br />
<br />
The model only has access to monolingual corpora, using the News Crawl corpus with articles from 2007 to 2013. The baseline for unsupervised is the method proposed by [Artetxe, 2017], which was the unsupervised word vector alignment method discussed in the Background section.<br />
<br />
The paper adds each component piece-wise when doing an evaluation to test the impact each piece has on the final score. As shown in Table 1, Unsupervised results compared to the baseline of word-by-word results are strong, with improvement between 40% to 140%. Results also show that back-translation is essential. Denoising doesn't show a big improvement however it is required for back-translation, because otherwise, back-translation would translate nonsensical sentences. The addition of backtranslation, however, does show large improvement on all tested cases.<br />
<br />
For the BPE experiment, results show it helps in some language pairs but detract in some other language pairs. This is because while BPE helped to translate some rare words, it increased the error rates in other words. It also did not perform well when translating named entities which occur infrequently.<br />
<br />
===Semi-supervised===<br />
<br />
Since there is often some small parallel data but not enough to train a Neural Machine Translation system, the authors test a semi-supervised setting with the same monolingual data from the unsupervised settings together with either 10,000 or 100,000 random sentence pairs from the News Commentary parallel corpus. The supervision is included to improve the model during the back-translation stage to directly predict sentences that are in the parallel corpus.<br />
<br />
Table 1 shows that the model can greatly benefit from the addition of a small parallel corpus to the monolingual corpora. It is surprising that semi-supervised in row 6 outperforms supervised in row 7, one possible explanation is that both the semi-supervised training set and the test set belong to the news domain, whereas the supervised training set is all domains of corpora.<br />
<br />
===Supervised===<br />
<br />
This setting provides an upper bound to the unsupervised proposed system. The data used was the combination of all parallel corpora provided at WMT 2014, which includes Europarl, Common Crawl and News Commentary for both language pairs plus the UN and the Gigaword corpus for French- English. Moreover, the authors use the same subsets of News Commentary alone to run the separate experiments in order to compare with the semi-supervised scenario.<br />
<br />
The Comparable NMT was trained using the same proposed model except it does not use monolingual corpora, and consequently, it was trained without denoising and back-translation. The proposed model under a supervised setting does much worse than the state of the NMT in row 10, which suggests that adding the additional constraints to enable unsupervised learning also limits the potential performance. To improve these results, the authors also suggest to use larger models, longer training times, and incorporating several well-known NMT techniques.<br />
<br />
===Qualitative Analysis===<br />
<br />
[[File:Table2_lwali.png|600px|center]]<br />
<br />
Table 2 shows 4 examples of French to English translations, which shows that the high-quality translations are produces by the proposed system, and this system adequately models non-trivial translation relations. Example 1 and 2 show that the model is able to not only go beyond a literal word-by-word substitution but also model structural differences in the languages (ex.e, it correctly translates "l’aeroport international de Los Angeles" as "Los Angeles International Airport", and it is capable of producing high-quality translations of long and more complex sentences. However, in Example 3 and 4, the system failed to translate the months and numbers correctly and having difficulty with comprehending odd sentence structures, which means that the proposed system has limitations. Specially, the authors points that the proposed model has difficulties to preserve some concrete details from source sentences.<br />
<br />
=Conclusions and Future Work=<br />
<br />
The paper presented an unsupervised model to perform translations with monolingual corpora by using an attention-based encoder-decoder system and training using denoise and back-translation.<br />
<br />
Although experimental results show that the proposed model is effective as an unsupervised approach, there is significant room for improvement when using the model in a supervised way, suggesting the model is limited by the architectural modifications. Some ideas for future improvement include:<br />
*Instead of using fixed cross-lingual word embeddings at the beginning which forces the encoder to learn a common representation for both languages, progressively update the weight of the embeddings as training progresses.<br />
*Decouple the shared encoder into 2 independent encoders at some point during training<br />
*Progressively reduce the noise level<br />
*Incorporate character level information into the model, which might help address some of the adequacy issues observed in our manual analysis<br />
*Use other noise/denoising techniques, and analyze their effect in relation to the typological divergences of different language pairs.<br />
<br />
= Critique =<br />
<br />
While the idea is interesting and the results are impressive for an unsupervised approach, much of the model had actually already been proposed by other papers that are referenced. The paper doesn't add a lot of new ideas but only builds on existing techniques and combines them in a different way to achieve good experimental results. The paper is not a significant algorithmic contribution. <br />
<br />
The results showed that the proposed system performed far worse than the state of the art when used in a supervised setting, which is concerning and shows that the techniques used creates a limitation and a ceiling for performance.<br />
<br />
Additionally, there was no rigorous hyperparameter exploration/optimization for the model. As a result, it is difficult to conclude whether the performance limit observed in the constrained supervised model is the absolute limit, or whether this could be overcome in both supervised/unsupervised models with the right constraints to achieve more competitive results. <br />
<br />
The best results shown are between two very closely related languages(English and French), and does much worse for English - German, even though English and German are also closely related (but less so than English and French) which suggests that the model may not be successful at translating between distant language pairs. More testing would be interesting to see.<br />
<br />
The results comparison could have shown how the semi-supervised version of the model scores compared to other semi-supervised approaches as touched on in the other works section.<br />
<br />
Their qualitative analysis just checks whether their proposed unsupervised NMT generates sensible translation. It is limited and it needs further detailed analysis regarding the characteristics and properties of translation which is generated by unsupervised NMT.<br />
<br />
* (As pointed out by an anonymous reviewer [https://openreview.net/forum?id=Sy2ogebAW])Future work is vague: “we would like to detect and mitigate the specific causes…” “We also think that a better handling of rare words…” That’s great, but how will you do these things? Do you have specific reasons to think this, or ideas on how to approach them? Otherwise, this is just hand-waving.<br />
<br />
= References =<br />
#'''[Mikolov, 2013]''' Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. "Distributed representations of words and phrases and their compositionality."<br />
#'''[Artetxe, 2017]''' Mikel Artetxe, Gorka Labaka, Eneko Agirre, "Learning bilingual word embeddings with (almost) no bilingual data".<br />
#'''[Gouws,2016]''' Stephan Gouws, Yoshua Bengio, Greg Corrado, "BilBOWA: Fast Bilingual Distributed Representations without Word Alignments."<br />
#'''[He, 2016]''' Di He, Yingce Xia, Tao Qin, Liwei Wang, Nenghai Yu, Tieyan Liu, and Wei-Ying Ma. "Dual learning for machine translation."<br />
#'''[Sennrich,2016]''' Rico Sennrich and Barry Haddow and Alexandra Birch, "Neural Machine Translation of Rare Words with Subword Units."<br />
#'''[Ravi & Knight, 2011]''' Sujith Ravi and Kevin Knight, "Deciphering foreign language."<br />
#'''[Dou & Knight, 2012]''' Qing Dou and Kevin Knight, "Large scale decipherment for out-of-domain machine translation."<br />
#'''[Johnson et al. 2017]''' Melvin Johnson,et al, "Google’s multilingual neural machine translation system: Enabling zero-shot translation."<br />
#'''[Zhang et al. 2017]''' Meng Zhang, Yang Liu, Huanbo Luan, and Maosong Sun. "Adversarial training for unsupervised bilingual lexicon induction"</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=conditional_neural_process&diff=40759conditional neural process2018-11-22T04:14:56Z<p>Wfisher: Add some related work (T)</p>
<hr />
<div>== Introduction ==<br />
<br />
To train a model effectively, deep neural networks typically require large datasets. To mitigate this data efficiency problem, learning in two phases is one approach: the first phase learns the statistics of a generic domain without committing to a specific learning task; the second phase learns a function for a specific task, but does so using only a small number of data points by exploiting the domain-wide statistics already learned. Taking a probabilistic stance and specifying a distribution over functions (stochastic processes) is another approach -- Gaussian Processes being a commonly used example of this. Such Bayesian methods can be computationally expensive, however. <br />
<br />
The authors of the paper propose a family of models that represent solutions to the supervised problem, and an end-to-end training approach to learning them that combines neural networks with features reminiscent of Gaussian Processes. They call this family of models Conditional Neural Processes (CNPs). CNPs can be trained on very few data points to make accurate predictions, while they also have the capacity to scale to complex functions and large datasets. <br />
<br />
== Model ==<br />
Consider a data set <math display="inline"> \{x_i, y_i\} </math> with evaluations <math display="inline">y_i = f(x_i) </math> for some unknown function <math display="inline">f</math>. Assume <math display="inline">g</math> is an approximating function of f. The aim is you minimize the loss between <math display="inline">f</math> and <math display="inline">g</math> on the entire space <math display="inline">X</math>. In practice, the routine is evaluated on a finite set of observations.<br />
<br />
<br />
Let training set be <math display="inline"> O = \{x_i, y_i\}_{i = 0} ^{n-1}</math>, and test set be <math display="inline"> T = \{x_i, y_i\}_{i = n} ^ {n + m - 1}</math>.<br />
<br />
P be a probability distribution over functions <math display="inline"> F : X \to Y</math>, formally known as a stochastic process. Thus, P defines a joint distribution over the random variables <math display="inline"> {f(x_i)}_{i = 0} ^{n + m - 1}</math>. Therefore, for <math display="inline"> P(f(x)|O, T)</math>, our task is to predict the output values <math display="inline">f(x_i)</math> for <math display="inline"> x_i \in T</math>, given <math display="inline"> O</math>. <br />
<br />
A common assumption made on P is that all function evaluations of <math display="inline"> f </math> is Gaussian distributed. The random functions class is called Gaussian Processes (GPs). This framework of stochastic process allows a model to be data efficient, however, it's hard to get appropriate priors and stochastic processes are expensive in computation, which includes GPs.<br />
<br />
[[File:001.jpg|300px|center]]<br />
<br />
== Conditional Neural Process ==<br />
<br />
Conditional Neural Process models directly parametrize conditional stochastic processes without imposing consistency with respect to some prior process. CNP parametrize distributions over <math display="inline">f(T)</math> given a distributed representation of <math display="inline">O</math> of fixed dimensionality. Thus, the mathematical guarantees associated with stochastic processes is traded off for functional flexibility and scalability.<br />
<br />
CNP is a conditional stochastic process <math display="inline">Q_\theta</math> defines distributions over <math display="inline">f(x_i)</math> for <math display="inline">x_i \in T</math>. For stochastic processs, we assume <math display="inline">Q_{\theta}</math> is invariant to permutations, and in this work, we generally enforce permutation invariance with respect to <math display="inline">T</math> be assuming a factored structure. That is, <math display="inline">Q_\theta(f(T) | O, T) = \prod _{x \in T} Q_\theta(f(x) | O, x)</math><br />
<br />
In detail, we use the following archiecture<br />
<br />
<math display="inline">r_i = h_\theta(x_i, y_i)</math> for any <math display="inline">(x_i, y_i) \in O</math>, where <math display="inline">h_\theta : X \times Y \to \mathbb{R} ^ d</math><br />
<br />
<math display="inline">r = r_i * r_2 * ... * r_n</math>, where <math display="inline">*</math> is a commutative operation that takes elements in <math display="inline">\mathbb{R}^d</math> and maps them into a single element of <math display="inline">\mathbb{R} ^ d</math><br />
<br />
<math display="inline">\Phi_i = g_\theta</math> for any <math display="inline">x_i \in T</math>, where <math display="inline">g_\theta : X \times \mathbb{R} ^ d \to \mathbb{R} ^ e</math> and <math display="inline">\Phi_i</math> are parameters for <math display="inline">Q_\theta</math><br />
<br />
Note that this architecture ensures permutation invariance and <math display="inline">O(n + m)</math> scaling for conditional prediction. Also, <math display="inline">r = r_i * r_2 * ... * r_n</math> can be computed in <math display="inline">O(n)</math>, this architecture supports streaming observation with minimal overhead.<br />
<br />
We train <math display="inline">Q_\theta</math> by asking it to predict <math display="inline">O</math> conditioned on a randomly<br />
chosen subset of <math display="inline">O</math>. This gives the model a signal of the uncertainty over the space X inherent in the distribution<br />
P given a set of observations. Thus, the targets it scores <math display="inline">Q_\theta</math> on include both the observed<br />
and unobserved values. In practice, we take Monte Carlo<br />
estimates of the gradient of this loss by sampling <math display="inline">f</math> and <math display="inline">N</math>. <br />
<br />
This approach shifts the burden of imposing prior knowledge from an analytic prior to empirical data. This has the advantage of liberating a practitioner from having to specify an analytic form for the prior, which is ultimately<br />
intended to summarize their empirical experience. Still, we emphasize that the <math display="inline">Q_\theta</math> are not necessarily a consistent set of conditionals for all observation sets, and the training routine does not guarantee that.<br />
<br />
In summary,<br />
<br />
1. A CNP is a conditional distribution over functions<br />
trained to model the empirical conditional distributions<br />
of functions <math display="inline">f \sim P</math>.<br />
<br />
2. A CNP is permutation invariant in <math display="inline">O</math> and <math display="inline">T</math>.<br />
<br />
3. A CNP is scalable, achieving a running time complexity<br />
of <math display="inline">O(n + m)</math> for making <math display="inline">m</math> predictions with <math display="inline">n</math><br />
observations.<br />
<br />
== Related Work ==<br />
<br />
===Gaussian Process Framework===<br />
<br />
Several papers attempt to address various issues with GPs. These include:<br />
* Using sparse GPs to aid in scaling (Snelson & Ghahramani, 2006)<br />
* Using Deep GPs to achieve more expressivity (Damianou & Lawrence,<br />
2013; Salimbeni & Deisenroth, 2017)<br />
* Using neural networks to learn more expressive kernels (Wilson et al., 2016)<br />
<br />
===Meta Learning===<br />
<br />
Meta-Learning attempts to allow neural networks to learn more generalizable functions, as opposed to only approximating one function. This can be done by learning deep generative models which can do few-shot estimations of data. This can be implemented with attention mechanisms or additional memory.<br />
<br />
== Experimental Result I: Function Regression ==<br />
<br />
Classical 1D regression task that used as a common baseline for GP is our first example. <br />
They generated two different datasets that consisted of functions<br />
generated from a GP with an exponential kernel. In the first dataset they used a kernel with fixed parameters, and in the second dataset the function switched at some random point. on the real line between two functions each sampled with<br />
different kernel parameters. At every training step they sampled a curve from the GP, select<br />
a subset of n points as observations, and a subset of t points as target points. Using the model, the observed points are encoded using a three layer MLP encoder h with a 128 dimensional output representation. The representations are aggregated into a single representation<br />
<math display="inline">r = \frac{1}{n} \sum r_i</math><br />
, which is concatenated to <math display="inline">x_t</math> and passed to a decoder g consisting of a five layer<br />
MLP. The function outputs a Gaussian mean and variance for the target outputs. The model is trained to maximize the log-likelihood of the target points using the Adam optimizer. <br />
<br />
Two examples of the regression results obtained for each<br />
of the datasets are shown in the following figure.<br />
<br />
[[File:007.jpg|300px|center]]<br />
<br />
They compared the model to the predictions generated by a GP with the correct<br />
hyperparameters, which constitutes an upper bound on our<br />
performance. Although the prediction generated by the GP<br />
is smoother than the CNP's prediction both for the mean<br />
and variance, the model is able to learn to regress from a few<br />
context points for both the fixed kernels and switching kernels.<br />
As the number of context points grows, the accuracy<br />
of the model improves and the approximated uncertainty<br />
of the model decreases. Crucially, we see the model learns<br />
to estimate its own uncertainty given the observations very<br />
accurately. Nonetheless it provides a good approximation<br />
that increases in accuracy as the number of context points<br />
increases.<br />
Furthermore the model achieves similarly good performance<br />
on the switching kernel task. This type of regression task<br />
is not trivial for GPs whereas in our case we only have to<br />
change the dataset used for training<br />
<br />
== Experimental Result II: Image Completion for Digits ==<br />
<br />
[[File:002.jpg|600px|center]]<br />
<br />
They also tested CNP on the MNIST dataset and use the test<br />
set to evaluate its performance. As shown in the above figure the<br />
model learns to make good predictions of the underlying<br />
digit even for a small number of context points. Crucially,<br />
when conditioned only on one non-informative context point the model’s prediction corresponds<br />
to the average over all MNIST digits. As the number<br />
of context points increases the predictions become more<br />
similar to the underlying ground truth. This demonstrates<br />
the model’s capacity to extract dataset specific prior knowledge.<br />
It is worth mentioning that even with a complete set<br />
of observations the model does not achieve pixel-perfect<br />
reconstruction, as we have a bottleneck at the representation<br />
level.<br />
Since this implementation of CNP returns factored outputs,<br />
the best prediction it can produce given limited context<br />
information is to average over all possible predictions that<br />
agree with the context. An alternative to this is to add<br />
latent variables in the model such that they can be sampled<br />
conditioned on the context to produce predictions with high<br />
probability in the data distribution. <br />
<br />
<br />
An important aspect of the model is its ability to estimate<br />
the uncertainty of the prediction. As shown in the bottom<br />
row of the above figure, as they added more observations, the variance<br />
shifts from being almost uniformly spread over the digit<br />
positions to being localized around areas that are specific<br />
to the underlying digit, specifically its edges. Being able to<br />
model the uncertainty given some context can be helpful for<br />
many tasks. One example is active exploration, where the<br />
model has a choice over where to observe.<br />
They tested this by<br />
comparing the predictions of CNP when the observations<br />
are chosen according to uncertainty, versus random pixels. This method is a very simple way of doing active<br />
exploration, but it already produces better prediction results<br />
than selecting the conditioning points at random.<br />
<br />
== Experimental Result III: Image Completion for Faces ==<br />
<br />
<br />
[[File:003.jpg|400px|center]]<br />
<br />
<br />
They also applied CNP to CelebA, a dataset of images of<br />
celebrity faces, and reported performance obtained on the<br />
test set.<br />
<br />
As shown in the above figure our model is able to capture<br />
the complex shapes and colours of this dataset with predictions<br />
conditioned on less than 10% of the pixels being<br />
already close to ground truth. As before, given few context<br />
points the model averages over all possible faces, but as<br />
the number of context pairs increases the predictions capture<br />
image-specific details like face orientation and facial<br />
expression. Furthermore, as the number of context points<br />
increases the variance is shifted towards the edges in the<br />
image.<br />
<br />
[[File:004.jpg|400px|center]]<br />
<br />
An important aspect of CNPs demonstrated in the above figure is<br />
its flexibility not only in the number of observations and<br />
targets it receives but also with regards to their input values.<br />
It is interesting to compare this property to GPs on one hand,<br />
and to trained generative models (van den Oord et al., 2016;<br />
Gregor et al., 2015) on the other hand.<br />
The first type of flexibility can be seen when conditioning on<br />
subsets that the model has not encountered during training.<br />
Consider conditioning the model on one half of the image,<br />
fox example. This forces the model to not only predict pixel<br />
values according to some stationary smoothness property of<br />
the images, but also according to global spatial properties,<br />
e.g. symmetry and the relative location of different parts of<br />
faces. As seen in the first row of the figure, CNPs are able to<br />
capture those properties. A GP with a stationary kernel cannot<br />
capture this, and in the absence of observations would<br />
revert to its mean (the mean itself can be non-stationary but<br />
usually this would not be enough to capture the interesting<br />
properties).<br />
<br />
In addition, the model is flexible with regards to the target<br />
input values. This means, e.g., we can query the model<br />
at resolutions it has not seen during training. We take a<br />
model that has only been trained using pixel coordinates of<br />
a specific resolution, and predict at test time subpixel values<br />
for targets between the original coordinates. As shown in<br />
Figure 5, with one forward pass we can query the model at<br />
different resolutions. While GPs also exhibit this type of<br />
flexibility, it is not the case for trained generative models,<br />
which can only predict values for the pixel coordinates on<br />
which they were trained. In this sense, CNPs capture the best<br />
of both worlds – it is flexible in regards to the conditioning<br />
and prediction task, and has the capacity to extract domain<br />
knowledge from a training set.<br />
<br />
[[File:010.jpg|400px|center]]<br />
<br />
<br />
They compared CNPs quantitatively to two related models:<br />
kNNs and GPs. As shown in the above table CNPs outperform<br />
the latter when number of context points is small (empirically<br />
when half of the image or less is provided as context).<br />
When the majority of the image is given as context exact<br />
methods like GPs and kNN will perform better. From the table<br />
we can also see that the order in which the context points<br />
are provided is less important for CNPs, since providing the<br />
context points in order from top to bottom still results in<br />
good performance. Both insights point to the fact that CNPs<br />
learn a data-specific ‘prior’ that will generate good samples<br />
even when the number of context points is very small.<br />
<br />
== Experimental Result IV: Classification ==<br />
Finally, they applied the model to one-shot classification using the Omniglot dataset. This dataset consists of 1,623 classes<br />
of characters from 50 different alphabets. Each class has<br />
only 20 examples and as such this dataset is particularly<br />
suitable for few-shot learning algorithms. They used 1,200 randomly selected classes as<br />
their training set and the remainder as our testing data set.<br />
This includes cropping<br />
the image from 32 × 32 to 28 × 28, applying small random<br />
translations and rotations to the inputs, and also increasing<br />
the number of classes by rotating every character by 90<br />
degrees and defining that to be a new class. They generated<br />
the labels for an N-way classification task by choosing N<br />
random classes at each training step and arbitrarily assigning<br />
the labels 0, ..., N − 1 to each.<br />
<br />
<br />
[[File:008.jpg|400px|center]]<br />
<br />
Given that the input points are images, they modified the architecture<br />
of the encoder h to include convolution layers as<br />
mentioned in section 2. In addition they only aggregated over<br />
inputs of the same class by using the information provided<br />
by the input label. The aggregated class-specific representations<br />
are then concatenated to form the final representation.<br />
Given that both the size of the class-specific representations<br />
and the number of classes are constant, the size of the final<br />
representation is still constant and thus the O(n + m)<br />
runtime still holds.<br />
The results of the classification are summarized in the following table<br />
CNPs achieve higher accuracy than models that are significantly<br />
more complex (like MANN). While CNPs do not<br />
beat state of the art for one-shot classification our accuracy<br />
values are comparable. Crucially, they reached those values<br />
using a significantly simpler architecture (three convolutional<br />
layers for the encoder and a three-layer MLP for the<br />
decoder) and with a lower runtime of O(n + m) at test time<br />
as opposed to O(nm)<br />
<br />
== Conclusion ==<br />
<br />
In this paper they had introduced Conditional Neural Processes,<br />
a model that is both flexible at test time and has the<br />
capacity to extract prior knowledge from training data.<br />
<br />
We had demonstrated its ability to perform a variety of tasks<br />
including regression, classification and image completion.<br />
We compared CNPs to Gaussian Processes on one hand, and<br />
deep learning methods on the other, and also discussed the<br />
relation to meta-learning and few-shot learning.<br />
It is important to note that the specific CNP implementations<br />
described here are just simple proofs-of-concept and can<br />
be substantially extended, e.g. by including more elaborate<br />
architectures in line with modern deep learning advances.<br />
To summarize, this work can be seen as a step towards learning<br />
high-level abstractions, one of the grand challenges of<br />
contemporary machine learning. Functions learned by most<br />
Conditional Neural Processes<br />
conventional deep learning models are tied to a specific, constrained<br />
statistical context at any stage of training. A trained<br />
CNP is more general, in that it encapsulates the high-level<br />
statistics of a family of functions. As such it constitutes a<br />
high-level abstraction that can be reused for multiple tasks.<br />
In future work they are going to explore how far these models can<br />
help in tackling the many key machine learning problems<br />
that seem to hinge on abstraction, such as transfer learning,<br />
meta-learning, and data efficiency.<br />
<br />
== Other Sources ==<br />
# Code for this model and a simpler explanation can be found at [https://github.com/deepmind/conditional-neural-process]<br />
# A newer version of the model is described in this paper [https://arxiv.org/pdf/1807.01622.pdf]<br />
# A good blog post on neural processes [https://kasparmartens.rbind.io/post/np/]<br />
<br />
== Reference ==<br />
Bartunov, S. and Vetrov, D. P. Fast adaptation in generative<br />
models with generative matching networks. arXiv<br />
preprint arXiv:1612.02192, 2016.<br />
<br />
Blundell, C., Cornebise, J., Kavukcuoglu, K., and Wierstra,<br />
D. Weight uncertainty in neural networks. arXiv preprint<br />
arXiv:1505.05424, 2015.<br />
<br />
Bornschein, J., Mnih, A., Zoran, D., and J. Rezende, D.<br />
Variational memory addressing in generative models. In<br />
Advances in Neural Information Processing Systems, pp.<br />
3923–3932, 2017.<br />
<br />
Damianou, A. and Lawrence, N. Deep gaussian processes.<br />
In Artificial Intelligence and Statistics, pp. 207–215,<br />
2013.<br />
<br />
Devlin, J., Bunel, R. R., Singh, R., Hausknecht, M., and<br />
Kohli, P. Neural program meta-induction. In Advances in<br />
Neural Information Processing Systems, pp. 2077–2085,<br />
2017.<br />
<br />
Edwards, H. and Storkey, A. Towards a neural statistician.<br />
2016.<br />
<br />
Finn, C., Abbeel, P., and Levine, S. Model-agnostic metalearning<br />
for fast adaptation of deep networks. arXiv<br />
preprint arXiv:1703.03400, 2017.<br />
<br />
Gal, Y. and Ghahramani, Z. Dropout as a bayesian approximation:<br />
Representing model uncertainty in deep learning.<br />
In international conference on machine learning, pp.<br />
1050–1059, 2016.<br />
<br />
Garnelo, M., Arulkumaran, K., and Shanahan, M. Towards<br />
deep symbolic reinforcement learning. arXiv preprint<br />
arXiv:1609.05518, 2016.<br />
<br />
Gregor, K., Danihelka, I., Graves, A., Rezende, D. J., and<br />
Wierstra, D. Draw: A recurrent neural network for image<br />
generation. arXiv preprint arXiv:1502.04623, 2015.<br />
<br />
Hewitt, L., Gane, A., Jaakkola, T., and Tenenbaum, J. B. The<br />
variational homoencoder: Learning to infer high-capacity<br />
generative models from few examples. 2018.<br />
<br />
J. Rezende, D., Danihelka, I., Gregor, K., Wierstra, D.,<br />
et al. One-shot generalization in deep generative models.<br />
In International Conference on Machine Learning, pp.<br />
1521–1529, 2016.<br />
<br />
Kingma, D. P. and Ba, J. Adam: A method for stochastic<br />
optimization. arXiv preprint arXiv:1412.6980, 2014.<br />
<br />
Kingma, D. P. and Welling, M. Auto-encoding variational<br />
bayes. arXiv preprint arXiv:1312.6114, 2013.<br />
<br />
Koch, G., Zemel, R., and Salakhutdinov, R. Siamese neural<br />
networks for one-shot image recognition. In ICML Deep<br />
Learning Workshop, volume 2, 2015.<br />
<br />
Lake, B. M., Salakhutdinov, R., and Tenenbaum, J. B.<br />
Human-level concept learning through probabilistic program<br />
induction. Science, 350(6266):1332–1338, 2015.<br />
<br />
Lake, B. M., Ullman, T. D., Tenenbaum, J. B., and Gershman,<br />
S. J. Building machines that learn and think like<br />
people. Behavioral and Brain Sciences, 40, 2017.<br />
<br />
LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. Gradientbased<br />
learning applied to document recognition. Proceedings<br />
of the IEEE, 86(11):2278–2324, 1998.<br />
<br />
Liu, Z., Luo, P., Wang, X., and Tang, X. Deep learning face<br />
attributes in the wild. In Proceedings of International<br />
Conference on Computer Vision (ICCV), December 2015.<br />
<br />
Louizos, C. and Welling, M. Multiplicative normalizing<br />
flows for variational bayesian neural networks. arXiv<br />
preprint arXiv:1703.01961, 2017.<br />
<br />
Louizos, C., Ullrich, K., and Welling, M. Bayesian compression<br />
for deep learning. In Advances in Neural Information<br />
Processing Systems, pp. 3290–3300, 2017.<br />
<br />
Rasmussen, C. E. and Williams, C. K. Gaussian processes<br />
in machine learning. In Advanced lectures on machine<br />
learning, pp. 63–71. Springer, 2004.<br />
<br />
Reed, S., Chen, Y., Paine, T., Oord, A. v. d., Eslami, S.,<br />
J. Rezende, D., Vinyals, O., and de Freitas, N. Few-shot<br />
autoregressive density estimation: Towards learning to<br />
learn distributions. 2017.<br />
<br />
Rezende, D. J., Mohamed, S., and Wierstra, D. Stochastic<br />
backpropagation and approximate inference in deep generative<br />
models. arXiv preprint arXiv:1401.4082, 2014.<br />
<br />
Salimbeni, H. and Deisenroth, M. Doubly stochastic variational<br />
inference for deep gaussian processes. In Advances<br />
in Neural Information Processing Systems, pp.<br />
4591–4602, 2017.<br />
<br />
Santoro, A., Bartunov, S., Botvinick, M., Wierstra, D., and<br />
Lillicrap, T. One-shot learning with memory-augmented<br />
neural networks. arXiv preprint arXiv:1605.06065, 2016.<br />
<br />
Snell, J., Swersky, K., and Zemel, R. Prototypical networks<br />
for few-shot learning. In Advances in Neural Information<br />
Processing Systems, pp. 4080–4090, 2017.<br />
<br />
Snelson, E. and Ghahramani, Z. Sparse gaussian processes<br />
using pseudo-inputs. In Advances in neural information<br />
processing systems, pp. 1257–1264, 2006.<br />
<br />
van den Oord, A., Kalchbrenner, N., Espeholt, L., Vinyals,<br />
O., Graves, A., et al. Conditional image generation with<br />
pixelcnn decoders. In Advances in Neural Information<br />
Processing Systems, pp. 4790–4798, 2016.<br />
<br />
Vinyals, O., Blundell, C., Lillicrap, T., Wierstra, D., et al.<br />
Matching networks for one shot learning. In Advances in<br />
Neural Information Processing Systems, pp. 3630–3638,<br />
2016.<br />
<br />
Wang, J. X., Kurth-Nelson, Z., Tirumala, D., Soyer, H.,<br />
Leibo, J. Z., Munos, R., Blundell, C., Kumaran, D., and<br />
Botvinick, M. Learning to reinforcement learn. arXiv<br />
preprint arXiv:1611.05763, 2016.<br />
<br />
Wilson, A. G., Hu, Z., Salakhutdinov, R., and Xing, E. P.<br />
Deep kernel learning. In Artificial Intelligence and Statistics,<br />
pp. 370–378, 2016.<br />
<br />
Damianou, A. and Lawrence, N. Deep gaussian processes.<br />
In Artificial Intelligence and Statistics, pp. 207–215,<br />
2013.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Mapping_Images_to_Scene_Graphs_with_Permutation-Invariant_Structured_Prediction&diff=40011Mapping Images to Scene Graphs with Permutation-Invariant Structured Prediction2018-11-19T15:59:14Z<p>Wfisher: edit (E)</p>
<hr />
<div>The paper ''Mapping Images to Scene Graphs with Permutation-Invariant Structured Prediction'' was written by Roei Herzig* from Tel Aviv University, Moshiko Raboh* from Tel Aviv University, Gal Chechik from Google Brain, Bar-Ilan University, Jonathan Berant from Tel Aviv University, and Amir Globerson from Tel Aviv University. This paper is part of the NIPS 2018 conference to be hosted in December 2018 at Montréal, Canada. This paper summary is based on version 3 of the pre-print (as of May 2018) obtained from [https://arxiv.org/pdf/1802.05451v3.pdf arXiv] <br />
<br />
(*) Equal contribution<br />
<br />
=Motivation=<br />
In the field of artificial intelligence, a major goal is to enable machines to understand complex images, such as the underlying relationships between objects that exist in each scene. Although there are models today that capture both complex labels and interactions between labels, there is a disconnect for what guidelines should be used when leveraging deep learning. This paper introduces a design principle for such models that stem from the concept of permutation invariance and proves state of the art performance on models that follow this principle.<br />
<br />
The primary contributions that this paper makes include:<br />
# Deriving sufficient and necessary conditions for respecting graph-permutation invariance in deep structured prediction architectures<br />
# Empirically proving the benefit of graph-permutation invariance<br />
# Developing a state-of-the-art model for scene graph predictions over a large set of complex visual scenes<br />
<br />
=Introduction=<br />
In order for a machine to interpret complex visual scenes, it must recognize and understand both objects and relationships between the objects in the scene. A '''scene graph''' is a representation of the set of objects and relations that exist in the scene, where objects are represented as nodes and relations are represented as edges connecting the different nodes. Hence, the prediction of the scene graph is analogous to inferring the joint set of objects and relations of a visual scene.<br />
<br />
[[File:scene_graph_example.png|600px|center]]<br />
<br />
Given that objects in scenes are interdependent on each other, joint prediction of the objects and relations is necessary. The field of structured prediction, which involves the general problem of inferring multiple inter-dependent labels, is of interest for this problem.<br />
<br />
In structured prediction models, a score function <math>s(x, y)</math> is defined to evaluate the compatibility between label <math>y</math> and input <math>x</math>. For instance, when interpreting the scene of an image, <math>x</math> refers to the image itself, and <math>y</math> refers to a complex label, which contains both the objects and the relations between objects. As with most other inference methods, the goal is to find the label <math>y*</math> such that <math>s(x,y)</math> is maximized. However, the major concern is that the space for possible label assignments grows exponentially with respect to input size. For example, although an image may seem very simple, the corpus containing possible labels for objects may be very large, rendering it difficult to optimize the scoring function. <br />
<br />
The paper presents an alternative approach, for which input <math>x</math> is mapped to structured output <math>y</math> using a "black box" neural network, omitting the definition of a score function. The main concern for this approach is the determination of the network architecture.<br />
<br />
=Structured prediction=<br />
This paper further considers structured predictions using score-based methods. For structured predictions that follow a score-based approach, a score function <math>s(x, y)</math> is used to measure how compatible label <math>y</math> is for input <math>x</math>. To optimize the score function, previous works have decomposed <math>s(x,y) = \sum_i f_i(x,y)</math> in order to facilitate efficient optimization for each local score function, <math>\max_y f_i(x,y)</math>.<br />
<br />
In the area of structured predictions, the most commonly-used score functions include the singleton score function <math>f_i(y_i, x)</math> and pairwise score function <math>f_{ij} (y_i, y_j, x)</math>. Previous works explored a two-stage architectures (learn local scores independently of the structured prediction goal), and end-to-end architectures (to include the inference algorithm within the computation graph). <br />
<br />
==Advantages of using score-based methods==<br />
# Allow for intuitive specification of local dependencies between labels, and how they map to global dependencies<br />
# Linear score functions offer natural convex surrogates<br />
# Inference in large label space is sometimes possible via exact algorithms or empirically accurate approximations<br />
<br />
The concern for modelling score functions using deep networks is that learning may no longer be convex. Hence, the paper presents properties for how deep networks can be used for structured predictions by considering architectures that do not require explicit maximization of a score function.<br />
<br />
=Background, Notations, and Definitions=<br />
We denote <math>y</math> as a structured label where <math>y = [y_1, \dots, y_n]</math><br />
<br />
'''Score functions:''' for score-based methods, the score is defined as either the sum of a set of singleton scores <math>f_i = f_i(y_i, x)</math> or the sum of pairwise scores <math>f_{ij} = f_{ij}(y_i, y_j, x)</math>.<br />
<br />
Let <math>s(x,y)</math> be the score of a score-based method. Then:<br />
<br />
<div align="center"><br />
<math>s(x,y) = \begin{cases}<br />
\sum_i f_i ~ \text{if we have a set of singleton scores}\\<br />
\sum_{ij} f_{ij} ~ \text{if we have a set of pairwise scores } \\<br />
\end{cases}</math><br />
</div><br />
<br />
'''Inference algorithm:''' an inference algorithm takes input set of local scores (either <math>f_i</math> or <math>f_{ij}</math>) and outputs an assignment of labels <math>y_1, \dots, y_n</math> that maximizes score function <math>s(x,y)</math><br />
<br />
'''Graph labeling function:''' a graph labeling function <math>\mathcal{F} : (V,E) \rightarrow Y</math> is a function that takes input of: an ordered set of node features <math>V = [z_1, \dots, z_n]</math> and an ordered set of edge features <math>E = [z_{1,2},\dots,z_{i,j},\dots,z_{n,n-1}]</math> to output set of node labels <math>\mathbf{y} = [y_1, \dots, y_n]</math>. For instance, <math>z_i</math> can be set equal to <math>f_i</math> and <math>z_{ij}</math> can be set equal to <math>f_{ij}</math>.<br />
<br />
For convenience, the joint set of nodes and edges will be denoted as <math>\mathbf{z}</math> to be a size <math>n^2</math> vector (<math>n</math> nodes and <math>n(n-1)</math> edges).<br />
<br />
'''Permutation:''' Let <math>z</math> be a set of node and edge features. Given a permutation <math>\sigma</math> of <math>\{1,\dots,n\}</math>, let <math>\sigma(z)</math> be a new set of node and edge features given by [<math>\sigma(z)]_i = z_{\sigma(i)}</math> and <math>[\sigma(z)]_{i,j} = z_{\sigma(i), \sigma(j)}</math><br />
<br />
'''One-hot representation:''' <math>\mathbf{1}[j]</math> be a one-hot vector with 1 in the <math>j^{th}</math> coordinate<br />
<br />
=Permutation-Invariant Structured prediction=<br />
<br />
With permutation-invariant structured prediction, we would expect the algorithm to produce the same result given the same score function. For instance, consider the case where we have label space for 3 variables <math>y_1, y_2, y_3</math> with input <math>\mathbf{z} = (f_1, f_2, f_3, f_{12}, f_{13}, f_{23})</math> that outputs label <math>\mathbf{y} = (y_1^*, y_2^*, y_3^*)</math>. Then if the algorithm is run on a permuted version input <math>z' = (f_2, f_1, f_3, f_{21}, f_{23}, f_{13})</math>, we would expect <math>\mathbf{y} = (y_2^*, y_1^*, y_3^*)</math> given the same score function.<br />
<br />
'''Graph permutation invariance (GPI):''' a graph labeling function <math>\mathcal{F}</math> is graph-permutation invariant, if for all permutations <math>\sigma</math> of <math>\{1, \dots, n\}</math> and for all nodes <math>z</math>, <math>\mathcal{F}(\sigma(\mathbf{z})) = \sigma(\mathcal{F}(\mathbf{z}))</math><br />
<br />
The paper presents a theorem on the necessary and sufficient conditions for a function <math>\mathcal{F}</math> to be graph permutation invariant. Intuitively, because <math>\mathcal{F}</math> is a function that takes an ordered set <math>z</math> as input, the output on <math>\mathbf{z}</math> could very well be different from <math>\sigma(\mathbf{z})</math>, which means <math>\mathcal{F}</math> needs to have some sort of symmetry in order to sustain <math>[\mathcal{F}(\sigma(\mathbf{z}))]]_k = [\mathcal{F}(\mathbf{z})]_{\sigma(k)}</math>.<br />
<br />
[[File:graph_permutation_invariance.jpg|400px|center]]<br />
<br />
==Theorem 1==<br />
Let <math>\mathcal{F}</math> be a graph labeling function. Then <math>\mathcal{F}</math> is graph-permutation invariant if and only if there exist functions <math>\alpha, \rho, \phi</math> such that for all <math>k=1, .., n</math>:<br />
\begin{align}<br />
[\mathcal{F}(\mathbf{z})]_k = \rho(\mathbf{z}_k, \sum_{i=1}^n \alpha(\mathbf{z}_i, \sum_{i\neq j} \phi(\mathbf{z}_i, \mathbf{z}_{i,j}, \mathbf{z}_j)))<br />
\end{align}<br />
where <math>\phi: \mathbb{R}^{2d+e} \rightarrow \mathbb{R}^L, \alpha: \mathbb{R}^{d + L} \rightarrow \mathbb{R}^{W}, p: \mathbb{R}^{W+d} \rightarrow \mathbb{R}</math>.<br />
<br />
Notice that for the dimensions of inputs and outputs, <math>d</math> refers to the number of singleton features in <math>z</math> and <math>e</math> refers to the number of edges. <br />
<br />
[[File:GPI_architecture.jpg|thumb|A schematic representation of the GPI architecture. Singleton features <math>z_i</math> are omitted for simplicity. First, the features <math>z_{i,j}</math> are processed element-wise by <math>\phi</math>. Next, they are summed to create a vector <math>s_i</math>, which is concatenated with <math>z_i</math>. Third, a representation of the entire graph is created by applying <math>\alpha\ n</math> times and summing the created vector. The graph representation is then finally processed by <math>\rho</math> together with <math>z_k</math>.|600px|center]]<br />
<br />
==Proof Sketch for Theorem 1==<br />
The proof of this theorem can be found in the paper. A proof sketch below is provided:<br />
<br />
'''For the forward direction''' (function that follows the form set out in equation (1) is GPI):<br />
# Using definition of permutation <math>\sigma</math>, and rewriting <math>[F(z)]_{\sigma(k)}</math> in the form from equation (1)<br />
# Second argument of <math>\rho</math> is invariant under <math>\sigma</math>, since it takes the sum of all indices <math>i</math> and all other indices <math>j \neq i </math>.<br />
<br />
'''For the backward direction''' (any black-box GPI function can be expressed in the form of equation 1):<br />
# Construct <math>\phi, \alpha</math> such that second argument of <math>\rho</math> contains all information about graph features of <math>z</math>, including edges that the features originate from<br />
# Assume each <math>z_k</math> uniquely identifies the node and <math>\mathcal{F}</math> is a function only of pairwise features <math>z_{i,j}</math><br />
# Construct <math>H</math> be a perfect hash function with <math>L</math> buckets, and <math>\phi</math> which maps '''pairwise features''' to a vector of size <math>L</math><br />
# <math>*</math>Construct <math>\phi(z_i, z_{i,j}, z_j) = \mathbf{1}[H(z_j)] z_{i,j}</math>, which intuitively means that <math>\phi</math> stores <math>z_{i,j}</math> in the unique bucket for node <math>j</math><br />
# Construct function <math>\alpha</math> to output a matrix <math>\mathbb{R}^{L \times L}</math> that maps each pairwise feature into unique positions (<math>\alpha(z_i, s_i) = \mathbf{1}[H(z_i)]s_i^T</math>)<br />
# Construct matrix <math>M = \sum_i \alpha(z_i,s_i)</math> by discarding rows/columns in <math>M</math> that do not correspond to original nodes (which reduces dimension to <math>n\times n</math>; set <math>\rho</math> to have same outcome as <math>\mathcal{F}</math>, and set the output of <math>\mathcal{F}</math> on <math>M</math> to be the labels <math>\mathbf{y} = y_1, \dots, y_n</math><br />
<br />
<math>*</math>The paper presents the proof for the edge features <math>z_{ij}</math> being scalar (<math>e = 1</math>) for simplicity, which can be extended easily to vectors with additional indexing.<br />
<br />
Although the results discussed previously apply to complete graphs (edges apply to all feature pairs), it can be easily extended to incomplete graphs. However, in place of permutation-invariance, it is now an automorphism-invariance.<br />
<br />
==Implications and Applications of Theorem 1==<br />
===Key Implications of Theorem 1===<br />
# Architecture "collects" information from the different edges of the graph, and does so in an invariant fashion using <math>\alpha</math> and <math>\phi</math><br />
# Architecture is parallelizable, since all <math>\phi</math> functions can be applied simultaneously<br />
<br />
===Some applications of Theorem 1===<br />
# '''Attention:''' the concept of attention can be implemented in the GPI characterization, with slight alterations to the functions <math>\alpha</math> and <math>\phi</math>. The complete details can be found in the supplementary materials of the paper.<br />
# '''RNN:''' recurrent architectures can maintain GPI property, since all GPI function <math>\mathcal{F}</math> are closed under composition. The output of one step after running <math>\mathcal{F}</math> will act as input for the next step, but maintain the GPI property throughout.<br />
<br />
=Related Work=<br />
# '''Architectural invariance:''' suggested recently in a 2017 paper called Deep Sets by Zaheer et al., which considers the case of invariance that is more restrictive.<br />
# '''Deep structured prediction:''' previous work applied deep learning to structured prediction, for instance, semantic segmentation. Some algorithms include message passing algorithms, gradient descent for maximizing score functions, greedy decoding (inference of labels based on time of previous labels). Apart from those algorithms, deep learning has been applied to other graph-based problems such as the Travelling Salesman Problem (Bello et al., 2016; Gilmer et al., 2017; Khalil et al., 2017). However, none of the previous work specifically address the notion of invariance in the general architecture, but rather focus on message passing architectures that can be generalized by this paper.<br />
# '''Scene graph prediction:''' scene graph extraction allows for reasoning, question answering, and image retrieval (Johnson et al., 2015; Lu et al., 2016; Raposo et al., 2017). Some other works in this area include object detection, action recognition, and even detection of human-object interactions (Liao et al., 2016; Plummer et al., 2017). Additional work has been done with the use of message passing algorithms (Xu et al., 2017), word embeddings (Lu et al., 2016), and end-to-end prediction directly from pixels (Newell & Deng, 2017). A notable mention is NeuralMotif (Zellers et al., 2017), which the authors describe as the current state-of-the-art model for scene graph predictions on Visual Genome dataset.<br />
# '''Burst Image Deblurring Using Permutation Invariant Convolutional Neural Networks:''' similar ideas were applied, where Permutation Invariant CNN, are used to restore sharp and noise-free images from bursts of photographs affected by hand tremor and noise. This presented good quality images with lots of details for challenging datasets.<br />
<br />
=Experimental Results=<br />
==Synthetic Graph Labeling==<br />
The authors created a synthetic problem to study GPI. This involved using an input graph <math>G = (V,E)</math> where each node <math>i</math> belongs to the set <math>\Gamma(i) \in \{1, \dots, K\}</math> where <math>K</math> is the number of samples. The task is to compute for each node, the number of neighbours that belong to the same set (i.e. finding the label of the node <math>i</math> if <math>y_i = \sum_{j \in N(i)} \mathbf{1}[\Gamma(i) = \Gamma(j)]</math>) . Then, random graphs (each with 10 nodes) were generated by sampling edges, and the set <math>\Gamma(i) \in \{1, \dots, K\}</math>for each node independently and uniformly.<br />
The node features of the graph <math>z_i \in \{0,1\}^K</math> are one-hot vectors of <math>\Gamma(i)</math>, and each pairwise edge feature <math>z_{ij} \in \{0, 1\}</math> denote whether the edge <math>ij</math> is in the edge set <math>E</math>. <br />
3 architectures were studied in this paper:<br />
# '''GPI-architecture for graph prediction''' (without attention and RNN)<br />
# '''LSTM''': replacing <math>\sum \phi(\cdot)</math> and <math>\sum \alpha(\cdot)</math> in the form of Theorem 1 using two LSTMs with state size 200, reading their input in random order<br />
# '''Fully connected feed-forward network''': with 2 hidden layers, each layer containing 1,000 nodes; the input is a concatenation of all nodes and pairwise features, and the output is all node predictions<br />
<br />
The results show that the GPI architecture requires far fewer samples to converge to the correct solution.<br />
[[File:GPI_synthetic_example.jpg|450px|center]]<br />
<br />
==Scene-Graph Classification==<br />
Applying the concept of GPI to Scene-Graph classification (SGP) is the main task of this paper. The input to this problem is an image, along with a set of annotated bounding boxes for the entities in the image. The goal is to correctly label each entity within the bounding boxes and the relationship between every pair of entities, resulting in a coherent scene graph.<br />
<br />
The authors describe two different types of variables to predict. The first type is entity variables <math>[y_1, \dots, y_n]</math> for all bounding boxes, where each <math>y_i</math> can take one of L values and refers to objects such as "dog" or "man". The second type is relation variables <math>[y_{n+1}, \cdots, y_{n^2}]</math>, where each <math>y_i</math> represents the relation (e.g. "on", "below") between a pair of bounding boxes (entities).<br />
<br />
The scene graph and contain two types of edges:<br />
# '''Entity-entity edge''': connecting two entities <math>y_i</math> and <math>y_j</math> for <math>1 \leq i \neq j \leq n</math><br />
# '''Entity-relation edges''': connecting every relation variable <math>y_k</math> for <math>k > n</math> to two entities<br />
<br />
The feature set <math>\mathbf{z}</math> is based on the baseline model from Zellers et al. (2017). For entity variables <math>y_i</math>, the vector <math>\mathbf{z}_i \in \mathbb{R}^L</math> models the probability of the entity appearing in <math>y_i</math>. <math>\mathbf{z}_i</math> is augmented by the coordinates of the bounding box. Similarly for relation variables <math>y_j</math>, the vector <math>\mathbf{z}_j \in \mathbb{R}^R</math>, models the probability of the relations between the two entities in <math>j</math>. For entity-entity pairwise features <math>\mathbf{z}_{i,j}</math>, there is a similar representation of the probabilities for the pair. The SGP outputs probability distributions over all entities and relations, which will then be used as input recurrently to maintain GPI. Finally, word embeddings are used and concatenated for the most probable entity-relation labels.<br />
<br />
'''Components of the GPI architecture''' (ent for entity, rel for relation)<br />
# <math>\phi_{ent}</math>: network that integrates two entity variables <math>y_i</math> and <math>y_j</math>, with input <math>z_i, z_j, z_{i,j}</math> and output vector of <math>\mathbb{R}^{n_1}</math> <br />
# <math>\alpha_{ent}</math>: network with inputs from <math>\phi_{ent}</math> for all neighbours of an entity, and uses attention mechanism to output vector <math>\mathbb{R}^{n_2}</math> <br />
# <math>\rho_{ent}</math>: network with inputs from the various <math>\mathbb{R}^{n_2}</math> vectors, and outputs <math>L</math> logits to predict entity value<br />
# <math>\rho_{rel}</math>: network with inputs <math>\alpha_{ent}</math> of two entities and <math>z_{i,j}</math>, and output into <math>R</math> logits<br />
<br />
==Set-up and Results==<br />
'''Dataset''': based on Visual Genome (VG) by (Krishna et al., 2017), which contains a total of 108,077 images annotated with bounding boxes, entities, and relations. An average of 12 entities and 7 relations exist per image. For a fair comparison with previous works, data from (Xu et al., 2017) for train and test splits were used. The authors used the same 150 entities and 50 relations as in (Xu et al., 2017; Newell & Deng, 2017; Zellers et al., 2017). Hyperparameters were tuned using a 70K/5K/32K split for training, validation, and testing respectively.<br />
<br />
'''Training''': all networks were trained using the Adam optimizer, with a batch size of 20. The loss function was the sum of cross-entropy losses over all of entities and relations. Penalties for misclassified entities were 4 times stronger than that of relations. Penalties for misclassified negative relations were 10 times weaker than that of positive relations.<br />
<br />
'''Evaluation''': there are three major tasks when inferring from the scene graph. The authors focus on the following:<br />
# '''SGCIs''': given ground-truth entity bounding boxes, predict all entity and relations categories<br />
# '''PredCIs''': given annotated bounding boxes with entity labels, predict all relations<br />
<br />
The evaluation metric Recall@K (shortened to R@K) is drawn from (Lu et al., 2016). This metric is the fraction of correct ground-truth triplets that appear within the <math>K</math> most confident triplets predicted by the model. Graph-constrained protocol requires the top-<math>K</math> triplets to assign one consistent class per entity and relation. The unconstrained protocol does not enforce such constraint.<br />
<br />
'''Models and baselines''': The authors compared variants of the GPI approach against four baselines, state-of-the-art models on completing scene graph sub-tasks. To maintain consistency, all models used the same training/testing data split, in addition to the preprocessing as per (Xu et al., 2017).<br />
<br />
'''Baselines from existing state-of-the-art models'''<br />
# (Lu et al., 2016): use of word embeddings to fine-tune the likelihood of predicted relations<br />
# (Xu et al., 2017): message passing algorithm between entities and relations to iteratively improve feature map for prediction<br />
# (Newell & Deng, 2017): Pixel2Graph, uses associative embeddings to produce a full graph from image<br />
# (Zellers et al., 2017): NeuralMotif method, encodes global context to capture higher-order motif in scene graphs; Baseline outputs entities and relations distributions without using global context<br />
<br />
'''GPI models'''<br />
# '''GPI with no attention mechanism''': simply following Theorem 1's functional form, with summation over features<br />
# '''GPI NeighborAttention''': same GPI model, but considers attention over neighbours features<br />
# '''GPI Linguistic''': similar to NeighborAttention model, but concatenates word embedding vectors<br />
<br />
'''Key Results''': The GPI Linguistic approach outperforms all baseline for SGCIs, and has similar performance to the state of the art NeuralMotifs method. The authors argue that PredCI is an easier task with less structure, yielding high performance for the existing state of the art models.<br />
<br />
[[File:GPI_table_results.png|700px|center]]<br />
<br />
=Conclusion=<br />
<br />
A deep learning approach was presented in this paper to structured prediction, which constrains the architecture to be invariant to structurally identical inputs. This approach relies on pairwise features which are capable of describing inter-label correlations and inherits the intuitive aspect of score-based approaches. The output produced is invariant to equivalent representation of the pairwise terms. <br />
<br />
As future work, the axiomatic approach can be extended. Additionally, exploring algorithms that discover symmetries for deep structured prediction when invariant structure is unknown and should be discovered from data is also an interesting extension of this work. <br />
<br />
<br />
=Critique=<br />
The paper's contribution comes from the novelty of the permutation invariance as a design guideline for structured prediction. Although not explicitly considered in many of the previous works, the idea of invariance in architecture has already been considered in Deep Sets by (Zaheer et al., 2017). This paper characterizes relaxes the condition on the invariance as compared to that of previous works. In the evaluation of the benefit of GPI models, the paper used a synthetic problem to illustrate the fact that far fewer samples are required for the GPI model to converge to 100\% accuracy. However, when comparing the true task of scene graph prediction against the state-of-the-art baselines, the GPI variants had only marginal higher Recall@K scores. The true benefit of this paper's discovery is the avoidance of maximizing a score function (leading computationally difficult problem), and instead directly producing output invariant to how we represent the pairwise terms.<br />
<br />
=References=<br />
Roei Herzig, Moshiko Raboh, Gal Chechik, Jonathan Berant, Amir Globerson, Mapping Images to Scene Graphs with Permutation-Invariant Structured Prediction, 2018.<br />
<br />
Additional resources from Moshiko Raboh's [https://github.com/shikorab/SceneGraph GitHub]</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Mapping_Images_to_Scene_Graphs_with_Permutation-Invariant_Structured_Prediction&diff=40010Mapping Images to Scene Graphs with Permutation-Invariant Structured Prediction2018-11-19T15:57:16Z<p>Wfisher: Remove 'contrast'</p>
<hr />
<div>The paper ''Mapping Images to Scene Graphs with Permutation-Invariant Structured Prediction'' was written by Roei Herzig* from Tel Aviv University, Moshiko Raboh* from Tel Aviv University, Gal Chechik from Google Brain, Bar-Ilan University, Jonathan Berant from Tel Aviv University, and Amir Globerson from Tel Aviv University. This paper is part of the NIPS 2018 conference to be hosted in December 2018 at Montréal, Canada. This paper summary is based on version 3 of the pre-print (as at May 2018) obtained from [https://arxiv.org/pdf/1802.05451v3.pdf arXiv] <br />
<br />
(*) Equal contribution<br />
<br />
=Motivation=<br />
In the field of artificial intelligence, a major goal is to enable machines to understand complex images, such as the underlying relationships between objects that exist in each scene. Although there are models today that capture both complex labels and interactions between labels, there is a disconnect for what guidelines should be used when leveraging deep learning. This paper introduces a design principle for such models that stem from the concept of permutation invariance and proves state of the art performance on models that follow this principle.<br />
<br />
The primary contributions that this paper makes include:<br />
# Deriving sufficient and necessary conditions for respecting graph-permutation invariance in deep structured prediction architectures<br />
# Empirically proving the benefit of graph-permutation invariance<br />
# Developing a state-of-the-art model for scene graph predictions over a large set of complex visual scenes<br />
<br />
=Introduction=<br />
In order for a machine to interpret complex visual scenes, it must recognize and understand both objects and relationships between the objects in the scene. A '''scene graph''' is a representation of the set of objects and relations that exist in the scene, where objects are represented as nodes and relations are represented as edges connecting the different nodes. Hence, the prediction of the scene graph is analogous to inferring the joint set of objects and relations of a visual scene.<br />
<br />
[[File:scene_graph_example.png|600px|center]]<br />
<br />
Given that objects in scenes are interdependent on each other, joint prediction of the objects and relations is necessary. The field of structured prediction, which involves the general problem of inferring multiple inter-dependent labels, is of interest for this problem.<br />
<br />
In structured prediction models, a score function <math>s(x, y)</math> is defined to evaluate the compatibility between label <math>y</math> and input <math>x</math>. For instance, when interpreting the scene of an image, <math>x</math> refers to the image itself, and <math>y</math> refers to a complex label, which contains both the objects and the relations between objects. As with most other inference methods, the goal is to find the label <math>y*</math> such that <math>s(x,y)</math> is maximized. However, the major concern is that the space for possible label assignments grows exponentially with respect to input size. For example, although an image may seem very simple, the corpus containing possible labels for objects may be very large, rendering it difficult to optimize the scoring function. <br />
<br />
The paper presents an alternative approach, for which input <math>x</math> is mapped to structured output <math>y</math> using a "black box" neural network, omitting the definition of a score function. The main concern for this approach is the determination of the network architecture.<br />
<br />
=Structured prediction=<br />
This paper further considers structured predictions using score-based methods. For structured predictions that follow a score-based approach, a score function <math>s(x, y)</math> is used to measure how compatible label <math>y</math> is for input <math>x</math>. To optimize the score function, previous works have decomposed <math>s(x,y) = \sum_i f_i(x,y)</math> in order to facilitate efficient optimization for each local score function, <math>\max_y f_i(x,y)</math>.<br />
<br />
In the area of structured predictions, the most commonly-used score functions include the singleton score function <math>f_i(y_i, x)</math> and pairwise score function <math>f_{ij} (y_i, y_j, x)</math>. Previous works explored a two-stage architectures (learn local scores independently of the structured prediction goal), and end-to-end architectures (to include the inference algorithm within the computation graph). <br />
<br />
==Advantages of using score-based methods==<br />
# Allow for intuitive specification of local dependencies between labels, and how they map to global dependencies<br />
# Linear score functions offer natural convex surrogates<br />
# Inference in large label space is sometimes possible via exact algorithms or empirically accurate approximations<br />
<br />
The concern for modelling score functions using deep networks is that learning may no longer be convex. Hence, the paper presents properties for how deep networks can be used for structured predictions by considering architectures that do not require explicit maximization of a score function.<br />
<br />
=Background, Notations, and Definitions=<br />
We denote <math>y</math> as a structured label where <math>y = [y_1, \dots, y_n]</math><br />
<br />
'''Score functions:''' for score-based methods, the score is defined as either the sum of a set of singleton scores <math>f_i = f_i(y_i, x)</math> or the sum of pairwise scores <math>f_{ij} = f_{ij}(y_i, y_j, x)</math>.<br />
<br />
Let <math>s(x,y)</math> be the score of a score-based method. Then:<br />
<br />
<div align="center"><br />
<math>s(x,y) = \begin{cases}<br />
\sum_i f_i ~ \text{if we have a set of singleton scores}\\<br />
\sum_{ij} f_{ij} ~ \text{if we have a set of pairwise scores } \\<br />
\end{cases}</math><br />
</div><br />
<br />
'''Inference algorithm:''' an inference algorithm takes input set of local scores (either <math>f_i</math> or <math>f_{ij}</math>) and outputs an assignment of labels <math>y_1, \dots, y_n</math> that maximizes score function <math>s(x,y)</math><br />
<br />
'''Graph labeling function:''' a graph labeling function <math>\mathcal{F} : (V,E) \rightarrow Y</math> is a function that takes input of: an ordered set of node features <math>V = [z_1, \dots, z_n]</math> and an ordered set of edge features <math>E = [z_{1,2},\dots,z_{i,j},\dots,z_{n,n-1}]</math> to output set of node labels <math>\mathbf{y} = [y_1, \dots, y_n]</math>. For instance, <math>z_i</math> can be set equal to <math>f_i</math> and <math>z_{ij}</math> can be set equal to <math>f_{ij}</math>.<br />
<br />
For convenience, the joint set of nodes and edges will be denoted as <math>\mathbf{z}</math> to be a size <math>n^2</math> vector (<math>n</math> nodes and <math>n(n-1)</math> edges).<br />
<br />
'''Permutation:''' Let <math>z</math> be a set of node and edge features. Given a permutation <math>\sigma</math> of <math>\{1,\dots,n\}</math>, let <math>\sigma(z)</math> be a new set of node and edge features given by [<math>\sigma(z)]_i = z_{\sigma(i)}</math> and <math>[\sigma(z)]_{i,j} = z_{\sigma(i), \sigma(j)}</math><br />
<br />
'''One-hot representation:''' <math>\mathbf{1}[j]</math> be a one-hot vector with 1 in the <math>j^{th}</math> coordinate<br />
<br />
=Permutation-Invariant Structured prediction=<br />
<br />
With permutation-invariant structured prediction, we would expect the algorithm to produce the same result given the same score function. For instance, consider the case where we have label space for 3 variables <math>y_1, y_2, y_3</math> with input <math>\mathbf{z} = (f_1, f_2, f_3, f_{12}, f_{13}, f_{23})</math> that outputs label <math>\mathbf{y} = (y_1^*, y_2^*, y_3^*)</math>. Then if the algorithm is run on a permuted version input <math>z' = (f_2, f_1, f_3, f_{21}, f_{23}, f_{13})</math>, we would expect <math>\mathbf{y} = (y_2^*, y_1^*, y_3^*)</math> given the same score function.<br />
<br />
'''Graph permutation invariance (GPI):''' a graph labeling function <math>\mathcal{F}</math> is graph-permutation invariant, if for all permutations <math>\sigma</math> of <math>\{1, \dots, n\}</math> and for all nodes <math>z</math>, <math>\mathcal{F}(\sigma(\mathbf{z})) = \sigma(\mathcal{F}(\mathbf{z}))</math><br />
<br />
The paper presents a theorem on the necessary and sufficient conditions for a function <math>\mathcal{F}</math> to be graph permutation invariant. Intuitively, because <math>\mathcal{F}</math> is a function that takes an ordered set <math>z</math> as input, the output on <math>\mathbf{z}</math> could very well be different from <math>\sigma(\mathbf{z})</math>, which means <math>\mathcal{F}</math> needs to have some sort of symmetry in order to sustain <math>[\mathcal{F}(\sigma(\mathbf{z}))]]_k = [\mathcal{F}(\mathbf{z})]_{\sigma(k)}</math>.<br />
<br />
[[File:graph_permutation_invariance.jpg|400px|center]]<br />
<br />
==Theorem 1==<br />
Let <math>\mathcal{F}</math> be a graph labeling function. Then <math>\mathcal{F}</math> is graph-permutation invariant if and only if there exist functions <math>\alpha, \rho, \phi</math> such that for all <math>k=1, .., n</math>:<br />
\begin{align}<br />
[\mathcal{F}(\mathbf{z})]_k = \rho(\mathbf{z}_k, \sum_{i=1}^n \alpha(\mathbf{z}_i, \sum_{i\neq j} \phi(\mathbf{z}_i, \mathbf{z}_{i,j}, \mathbf{z}_j)))<br />
\end{align}<br />
where <math>\phi: \mathbb{R}^{2d+e} \rightarrow \mathbb{R}^L, \alpha: \mathbb{R}^{d + L} \rightarrow \mathbb{R}^{W}, p: \mathbb{R}^{W+d} \rightarrow \mathbb{R}</math>.<br />
<br />
Notice that for the dimensions of inputs and outputs, <math>d</math> refers to the number of singleton features in <math>z</math> and <math>e</math> refers to the number of edges. <br />
<br />
[[File:GPI_architecture.jpg|thumb|A schematic representation of the GPI architecture. Singleton features <math>z_i</math> are omitted for simplicity. First, the features <math>z_{i,j}</math> are processed element-wise by <math>\phi</math>. Next, they are summed to create a vector <math>s_i</math>, which is concatenated with <math>z_i</math>. Third, a representation of the entire graph is created by applying <math>\alpha\ n</math> times and summing the created vector. The graph representation is then finally processed by <math>\rho</math> together with <math>z_k</math>.|600px|center]]<br />
<br />
==Proof Sketch for Theorem 1==<br />
The proof of this theorem can be found in the paper. A proof sketch below is provided:<br />
<br />
'''For the forward direction''' (function that follows the form set out in equation (1) is GPI):<br />
# Using definition of permutation <math>\sigma</math>, and rewriting <math>[F(z)]_{\sigma(k)}</math> in the form from equation (1)<br />
# Second argument of <math>\rho</math> is invariant under <math>\sigma</math>, since it takes the sum of all indices <math>i</math> and all other indices <math>j \neq i </math>.<br />
<br />
'''For the backward direction''' (any black-box GPI function can be expressed in the form of equation 1):<br />
# Construct <math>\phi, \alpha</math> such that second argument of <math>\rho</math> contains all information about graph features of <math>z</math>, including edges that the features originate from<br />
# Assume each <math>z_k</math> uniquely identifies the node and <math>\mathcal{F}</math> is a function only of pairwise features <math>z_{i,j}</math><br />
# Construct <math>H</math> be a perfect hash function with <math>L</math> buckets, and <math>\phi</math> which maps '''pairwise features''' to a vector of size <math>L</math><br />
# <math>*</math>Construct <math>\phi(z_i, z_{i,j}, z_j) = \mathbf{1}[H(z_j)] z_{i,j}</math>, which intuitively means that <math>\phi</math> stores <math>z_{i,j}</math> in the unique bucket for node <math>j</math><br />
# Construct function <math>\alpha</math> to output a matrix <math>\mathbb{R}^{L \times L}</math> that maps each pairwise feature into unique positions (<math>\alpha(z_i, s_i) = \mathbf{1}[H(z_i)]s_i^T</math>)<br />
# Construct matrix <math>M = \sum_i \alpha(z_i,s_i)</math> by discarding rows/columns in <math>M</math> that do not correspond to original nodes (which reduces dimension to <math>n\times n</math>; set <math>\rho</math> to have same outcome as <math>\mathcal{F}</math>, and set the output of <math>\mathcal{F}</math> on <math>M</math> to be the labels <math>\mathbf{y} = y_1, \dots, y_n</math><br />
<br />
<math>*</math>The paper presents the proof for the edge features <math>z_{ij}</math> being scalar (<math>e = 1</math>) for simplicity, which can be extended easily to vectors with additional indexing.<br />
<br />
Although the results discussed previously apply to complete graphs (edges apply to all feature pairs), it can be easily extended to incomplete graphs. However, in place of permutation-invariance, it is now an automorphism-invariance.<br />
<br />
==Implications and Applications of Theorem 1==<br />
===Key Implications of Theorem 1===<br />
# Architecture "collects" information from the different edges of the graph, and does so in an invariant fashion using <math>\alpha</math> and <math>\phi</math><br />
# Architecture is parallelizable, since all <math>\phi</math> functions can be applied simultaneously<br />
<br />
===Some applications of Theorem 1===<br />
# '''Attention:''' the concept of attention can be implemented in the GPI characterization, with slight alterations to the functions <math>\alpha</math> and <math>\phi</math>. The complete details can be found in the supplementary materials of the paper.<br />
# '''RNN:''' recurrent architectures can maintain GPI property, since all GPI function <math>\mathcal{F}</math> are closed under composition. The output of one step after running <math>\mathcal{F}</math> will act as input for the next step, but maintain the GPI property throughout.<br />
<br />
=Related Work=<br />
# '''Architectural invariance:''' suggested recently in a 2017 paper called Deep Sets by Zaheer et al., which considers the case of invariance that is more restrictive.<br />
# '''Deep structured prediction:''' previous work applied deep learning to structured prediction, for instance, semantic segmentation. Some algorithms include message passing algorithms, gradient descent for maximizing score functions, greedy decoding (inference of labels based on time of previous labels). Apart from those algorithms, deep learning has been applied to other graph-based problems such as the Travelling Salesman Problem (Bello et al., 2016; Gilmer et al., 2017; Khalil et al., 2017). However, none of the previous work specifically address the notion of invariance in the general architecture, but rather focus on message passing architectures that can be generalized by this paper.<br />
# '''Scene graph prediction:''' scene graph extraction allows for reasoning, question answering, and image retrieval (Johnson et al., 2015; Lu et al., 2016; Raposo et al., 2017). Some other works in this area include object detection, action recognition, and even detection of human-object interactions (Liao et al., 2016; Plummer et al., 2017). Additional work has been done with the use of message passing algorithms (Xu et al., 2017), word embeddings (Lu et al., 2016), and end-to-end prediction directly from pixels (Newell & Deng, 2017). A notable mention is NeuralMotif (Zellers et al., 2017), which the authors describe as the current state-of-the-art model for scene graph predictions on Visual Genome dataset.<br />
# '''Burst Image Deblurring Using Permutation Invariant Convolutional Neural Networks:''' similar ideas were applied, where Permutation Invariant CNN, are used to restore sharp and noise-free images from bursts of photographs affected by hand tremor and noise. This presented good quality images with lots of details for challenging datasets.<br />
<br />
=Experimental Results=<br />
==Synthetic Graph Labeling==<br />
The authors created a synthetic problem to study GPI. This involved using an input graph <math>G = (V,E)</math> where each node <math>i</math> belongs to the set <math>\Gamma(i) \in \{1, \dots, K\}</math> where <math>K</math> is the number of samples. The task is to compute for each node, the number of neighbours that belong to the same set (i.e. finding the label of the node <math>i</math> if <math>y_i = \sum_{j \in N(i)} \mathbf{1}[\Gamma(i) = \Gamma(j)]</math>) . Then, random graphs (each with 10 nodes) were generated by sampling edges, and the set <math>\Gamma(i) \in \{1, \dots, K\}</math>for each node independently and uniformly.<br />
The node features of the graph <math>z_i \in \{0,1\}^K</math> are one-hot vectors of <math>\Gamma(i)</math>, and each pairwise edge feature <math>z_{ij} \in \{0, 1\}</math> denote whether the edge <math>ij</math> is in the edge set <math>E</math>. <br />
3 architectures were studied in this paper:<br />
# '''GPI-architecture for graph prediction''' (without attention and RNN)<br />
# '''LSTM''': replacing <math>\sum \phi(\cdot)</math> and <math>\sum \alpha(\cdot)</math> in the form of Theorem 1 using two LSTMs with state size 200, reading their input in random order<br />
# '''Fully connected feed-forward network''': with 2 hidden layers, each layer containing 1,000 nodes; the input is a concatenation of all nodes and pairwise features, and the output is all node predictions<br />
<br />
The results show that the GPI architecture requires far fewer samples to converge to the correct solution.<br />
[[File:GPI_synthetic_example.jpg|450px|center]]<br />
<br />
==Scene-Graph Classification==<br />
Applying the concept of GPI to Scene-Graph classification (SGP) is the main task of this paper. The input to this problem is an image, along with a set of annotated bounding boxes for the entities in the image. The goal is to correctly label each entity within the bounding boxes and the relationship between every pair of entities, resulting in a coherent scene graph.<br />
<br />
The authors describe two different types of variables to predict. The first type is entity variables <math>[y_1, \dots, y_n]</math> for all bounding boxes, where each <math>y_i</math> can take one of L values and refers to objects such as "dog" or "man". The second type is relation variables <math>[y_{n+1}, \cdots, y_{n^2}]</math>, where each <math>y_i</math> represents the relation (e.g. "on", "below") between a pair of bounding boxes (entities).<br />
<br />
The scene graph and contain two types of edges:<br />
# '''Entity-entity edge''': connecting two entities <math>y_i</math> and <math>y_j</math> for <math>1 \leq i \neq j \leq n</math><br />
# '''Entity-relation edges''': connecting every relation variable <math>y_k</math> for <math>k > n</math> to two entities<br />
<br />
The feature set <math>\mathbf{z}</math> is based on the baseline model from Zellers et al. (2017). For entity variables <math>y_i</math>, the vector <math>\mathbf{z}_i \in \mathbb{R}^L</math> models the probability of the entity appearing in <math>y_i</math>. <math>\mathbf{z}_i</math> is augmented by the coordinates of the bounding box. Similarly for relation variables <math>y_j</math>, the vector <math>\mathbf{z}_j \in \mathbb{R}^R</math>, models the probability of the relations between the two entities in <math>j</math>. For entity-entity pairwise features <math>\mathbf{z}_{i,j}</math>, there is a similar representation of the probabilities for the pair. The SGP outputs probability distributions over all entities and relations, which will then be used as input recurrently to maintain GPI. Finally, word embeddings are used and concatenated for the most probable entity-relation labels.<br />
<br />
'''Components of the GPI architecture''' (ent for entity, rel for relation)<br />
# <math>\phi_{ent}</math>: network that integrates two entity variables <math>y_i</math> and <math>y_j</math>, with input <math>z_i, z_j, z_{i,j}</math> and output vector of <math>\mathbb{R}^{n_1}</math> <br />
# <math>\alpha_{ent}</math>: network with inputs from <math>\phi_{ent}</math> for all neighbours of an entity, and uses attention mechanism to output vector <math>\mathbb{R}^{n_2}</math> <br />
# <math>\rho_{ent}</math>: network with inputs from the various <math>\mathbb{R}^{n_2}</math> vectors, and outputs <math>L</math> logits to predict entity value<br />
# <math>\rho_{rel}</math>: network with inputs <math>\alpha_{ent}</math> of two entities and <math>z_{i,j}</math>, and output into <math>R</math> logits<br />
<br />
==Set-up and Results==<br />
'''Dataset''': based on Visual Genome (VG) by (Krishna et al., 2017), which contains a total of 108,077 images annotated with bounding boxes, entities, and relations. An average of 12 entities and 7 relations exist per image. For a fair comparison with previous works, data from (Xu et al., 2017) for train and test splits were used. The authors used the same 150 entities and 50 relations as in (Xu et al., 2017; Newell & Deng, 2017; Zellers et al., 2017). Hyperparameters were tuned using a 70K/5K/32K split for training, validation, and testing respectively.<br />
<br />
'''Training''': all networks were trained using the Adam optimizer, with a batch size of 20. The loss function was the sum of cross-entropy losses over all of entities and relations. Penalties for misclassified entities were 4 times stronger than that of relations. Penalties for misclassified negative relations were 10 times weaker than that of positive relations.<br />
<br />
'''Evaluation''': there are three major tasks when inferring from the scene graph. The authors focus on the following:<br />
# '''SGCIs''': given ground-truth entity bounding boxes, predict all entity and relations categories<br />
# '''PredCIs''': given annotated bounding boxes with entity labels, predict all relations<br />
<br />
The evaluation metric Recall@K (shortened to R@K) is drawn from (Lu et al., 2016). This metric is the fraction of correct ground-truth triplets that appear within the <math>K</math> most confident triplets predicted by the model. Graph-constrained protocol requires the top-<math>K</math> triplets to assign one consistent class per entity and relation. The unconstrained protocol does not enforce such constraint.<br />
<br />
'''Models and baselines''': The authors compared variants of the GPI approach against four baselines, state-of-the-art models on completing scene graph sub-tasks. To maintain consistency, all models used the same training/testing data split, in addition to the preprocessing as per (Xu et al., 2017).<br />
<br />
'''Baselines from existing state-of-the-art models'''<br />
# (Lu et al., 2016): use of word embeddings to fine-tune the likelihood of predicted relations<br />
# (Xu et al., 2017): message passing algorithm between entities and relations to iteratively improve feature map for prediction<br />
# (Newell & Deng, 2017): Pixel2Graph, uses associative embeddings to produce a full graph from image<br />
# (Zellers et al., 2017): NeuralMotif method, encodes global context to capture higher-order motif in scene graphs; Baseline outputs entities and relations distributions without using global context<br />
<br />
'''GPI models'''<br />
# '''GPI with no attention mechanism''': simply following Theorem 1's functional form, with summation over features<br />
# '''GPI NeighborAttention''': same GPI model, but considers attention over neighbours features<br />
# '''GPI Linguistic''': similar to NeighborAttention model, but concatenates word embedding vectors<br />
<br />
'''Key Results''': The GPI Linguistic approach outperforms all baseline for SGCIs, and has similar performance to the state of the art NeuralMotifs method. The authors argue that PredCI is an easier task with less structure, yielding high performance for the existing state of the art models.<br />
<br />
[[File:GPI_table_results.png|700px|center]]<br />
<br />
=Conclusion=<br />
<br />
A deep learning approach was presented in this paper to structured prediction, which constrains the architecture to be invariant to structurally identical inputs. This approach relies on pairwise features which are capable of describing inter-label correlations and inherits the intuitive aspect of score-based approaches. The output produced is invariant to equivalent representation of the pairwise terms. <br />
<br />
As future work, the axiomatic approach can be extended. Additionally, exploring algorithms that discover symmetries for deep structured prediction when invariant structure is unknown and should be discovered from data is also an interesting extension of this work. <br />
<br />
<br />
=Critique=<br />
The paper's contribution comes from the novelty of the permutation invariance as a design guideline for structured prediction. Although not explicitly considered in many of the previous works, the idea of invariance in architecture has already been considered in Deep Sets by (Zaheer et al., 2017). This paper characterizes relaxes the condition on the invariance as compared to that of previous works. In the evaluation of the benefit of GPI models, the paper used a synthetic problem to illustrate the fact that far fewer samples are required for the GPI model to converge to 100\% accuracy. However, when comparing the true task of scene graph prediction against the state-of-the-art baselines, the GPI variants had only marginal higher Recall@K scores. The true benefit of this paper's discovery is the avoidance of maximizing a score function (leading computationally difficult problem), and instead directly producing output invariant to how we represent the pairwise terms.<br />
<br />
=References=<br />
Roei Herzig, Moshiko Raboh, Gal Chechik, Jonathan Berant, Amir Globerson, Mapping Images to Scene Graphs with Permutation-Invariant Structured Prediction, 2018.<br />
<br />
Additional resources from Moshiko Raboh's [https://github.com/shikorab/SceneGraph GitHub]</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Mapping_Images_to_Scene_Graphs_with_Permutation-Invariant_Structured_Prediction&diff=40009Mapping Images to Scene Graphs with Permutation-Invariant Structured Prediction2018-11-19T15:52:12Z<p>Wfisher: lower case for consistency (E)</p>
<hr />
<div>The paper ''Mapping Images to Scene Graphs with Permutation-Invariant Structured Prediction'' was written by Roei Herzig* from Tel Aviv University, Moshiko Raboh* from Tel Aviv University, Gal Chechik from Google Brain, Bar-Ilan University, Jonathan Berant from Tel Aviv University, and Amir Globerson from Tel Aviv University. This paper is part of the NIPS 2018 conference to be hosted in December 2018 at Montréal, Canada. This paper summary is based on version 3 of the pre-print (as at May 2018) obtained from [https://arxiv.org/pdf/1802.05451v3.pdf arXiv] <br />
<br />
(*) Equal contribution<br />
<br />
=Motivation=<br />
In the field of artificial intelligence, a major goal is to enable machines to understand complex images, such as the underlying relationships between objects that exist in each scene. Although there are models today that capture both complex labels and interactions between labels, there is a disconnect for what guidelines should be used when leveraging deep learning. This paper introduces a design principle for such models that stem from the concept of permutation invariance and proves state of the art performance on models that follow this principle.<br />
<br />
The primary contributions that this paper makes include:<br />
# Deriving sufficient and necessary conditions for respecting graph-permutation invariance in deep structured prediction architectures<br />
# Empirically proving the benefit of graph-permutation invariance<br />
# Developing a state-of-the-art model for scene graph predictions over a large set of complex visual scenes<br />
<br />
=Introduction=<br />
In order for a machine to interpret complex visual scenes, it must recognize and understand both objects and relationships between the objects in the scene. A '''scene graph''' is a representation of the set of objects and relations that exist in the scene, where objects are represented as nodes and relations are represented as edges connecting the different nodes. Hence, the prediction of the scene graph is analogous to inferring the joint set of objects and relations of a visual scene.<br />
<br />
[[File:scene_graph_example.png|600px|center]]<br />
<br />
Given that objects in scenes are interdependent on each other, joint prediction of the objects and relations is necessary. The field of structured prediction, which involves the general problem of inferring multiple inter-dependent labels, is of interest for this problem.<br />
<br />
In structured prediction models, a score function <math>s(x, y)</math> is defined to evaluate the compatibility between label <math>y</math> and input <math>x</math>. For instance, when interpreting the scene of an image, <math>x</math> refers to the image itself, and <math>y</math> refers to a complex label, which contains both the objects and the relations between objects. As with most other inference methods, the goal is to find the label <math>y*</math> such that <math>s(x,y)</math> is maximized. However, the major concern is that the space for possible label assignments grows exponentially with respect to input size. For example, although an image may seem very simple, the corpus containing possible labels for objects may be very large, rendering it difficult to optimize the scoring function. <br />
<br />
The paper presents an alternative approach, for which input <math>x</math> is mapped to structured output <math>y</math> using a "black box" neural network, omitting the definition of a score function. The main concern for this approach is the determination of the network architecture.<br />
<br />
=Structured prediction=<br />
This paper further considers structured predictions using score-based methods. For structured predictions that follow a score-based approach, a score function <math>s(x, y)</math> is used to measure how compatible label <math>y</math> is for input <math>x</math>. To optimize the score function, previous works have decomposed <math>s(x,y) = \sum_i f_i(x,y)</math> in order to facilitate efficient optimization for each local score function, <math>\max_y f_i(x,y)</math>.<br />
<br />
In the area of structured predictions, the most commonly-used score functions include the singleton score function <math>f_i(y_i, x)</math> and pairwise score function <math>f_{ij} (y_i, y_j, x)</math>. Previous works explored a two-stage architectures (learn local scores independently of the structured prediction goal), and end-to-end architectures (to include the inference algorithm within the computation graph). <br />
<br />
==Advantages of using score-based methods==<br />
# Allow for intuitive specification of local dependencies between labels, and how they map to global dependencies<br />
# Linear score functions offer natural convex surrogates<br />
# Inference in large label space is sometimes possible via exact algorithms or empirically accurate approximations<br />
<br />
The concern for modelling score functions using deep networks is that learning may no longer be convex. Hence, the paper presents properties for how deep networks can be used for structured predictions by considering architectures that do not require explicit maximization of a score function.<br />
<br />
=Background, Notations, and Definitions=<br />
We denote <math>y</math> as a structured label where <math>y = [y_1, \dots, y_n]</math><br />
<br />
'''Score functions:''' for score-based methods, the score is defined as either the sum of a set of singleton scores <math>f_i = f_i(y_i, x)</math> or the sum of pairwise scores <math>f_{ij} = f_{ij}(y_i, y_j, x)</math>.<br />
<br />
Let <math>s(x,y)</math> be the score of a score-based method. Then:<br />
<br />
<div align="center"><br />
<math>s(x,y) = \begin{cases}<br />
\sum_i f_i ~ \text{if we have a set of singleton scores}\\<br />
\sum_{ij} f_{ij} ~ \text{if we have a set of pairwise scores } \\<br />
\end{cases}</math><br />
</div><br />
<br />
'''Inference algorithm:''' an inference algorithm takes input set of local scores (either <math>f_i</math> or <math>f_{ij}</math>) and outputs an assignment of labels <math>y_1, \dots, y_n</math> that maximizes score function <math>s(x,y)</math><br />
<br />
'''Graph labeling function:''' a graph labeling function <math>\mathcal{F} : (V,E) \rightarrow Y</math> is a function that takes input of: an ordered set of node features <math>V = [z_1, \dots, z_n]</math> and an ordered set of edge features <math>E = [z_{1,2},\dots,z_{i,j},\dots,z_{n,n-1}]</math> to output set of node labels <math>\mathbf{y} = [y_1, \dots, y_n]</math>. For instance, <math>z_i</math> can be set equal to <math>f_i</math> and <math>z_{ij}</math> can be set equal to <math>f_{ij}</math>.<br />
<br />
For convenience, the joint set of nodes and edges will be denoted as <math>\mathbf{z}</math> to be a size <math>n^2</math> vector (<math>n</math> nodes and <math>n(n-1)</math> edges).<br />
<br />
'''Permutation:''' Let <math>z</math> be a set of node and edge features. Given a permutation <math>\sigma</math> of <math>\{1,\dots,n\}</math>, let <math>\sigma(z)</math> be a new set of node and edge features given by [<math>\sigma(z)]_i = z_{\sigma(i)}</math> and <math>[\sigma(z)]_{i,j} = z_{\sigma(i), \sigma(j)}</math><br />
<br />
'''One-hot representation:''' <math>\mathbf{1}[j]</math> be a one-hot vector with 1 in the <math>j^{th}</math> coordinate<br />
<br />
=Permutation-Invariant Structured prediction=<br />
<br />
With permutation-invariant structured prediction, we would expect the algorithm to produce the same result given the same score function. For instance, consider the case where we have label space for 3 variables <math>y_1, y_2, y_3</math> with input <math>\mathbf{z} = (f_1, f_2, f_3, f_{12}, f_{13}, f_{23})</math> that outputs label <math>\mathbf{y} = (y_1^*, y_2^*, y_3^*)</math>. Then if the algorithm is run on a permuted version input <math>z' = (f_2, f_1, f_3, f_{21}, f_{23}, f_{13})</math>, we would expect <math>\mathbf{y} = (y_2^*, y_1^*, y_3^*)</math> given the same score function.<br />
<br />
'''Graph permutation invariance (GPI):''' a graph labeling function <math>\mathcal{F}</math> is graph-permutation invariant, if for all permutations <math>\sigma</math> of <math>\{1, \dots, n\}</math> and for all nodes <math>z</math>, <math>\mathcal{F}(\sigma(\mathbf{z})) = \sigma(\mathcal{F}(\mathbf{z}))</math><br />
<br />
The paper presents a theorem on the necessary and sufficient conditions for a function <math>\mathcal{F}</math> to be graph permutation invariant. Intuitively, because <math>\mathcal{F}</math> is a function that takes an ordered set <math>z</math> as input, the output on <math>\mathbf{z}</math> could very well be different from <math>\sigma(\mathbf{z})</math>, which means <math>\mathcal{F}</math> needs to have some sort of symmetry in order to sustain <math>[\mathcal{F}(\sigma(\mathbf{z}))]]_k = [\mathcal{F}(\mathbf{z})]_{\sigma(k)}</math>.<br />
<br />
[[File:graph_permutation_invariance.jpg|400px|center]]<br />
<br />
==Theorem 1==<br />
Let <math>\mathcal{F}</math> be a graph labeling function. Then <math>\mathcal{F}</math> is graph-permutation invariant if and only if there exist functions <math>\alpha, \rho, \phi</math> such that for all <math>k=1, .., n</math>:<br />
\begin{align}<br />
[\mathcal{F}(\mathbf{z})]_k = \rho(\mathbf{z}_k, \sum_{i=1}^n \alpha(\mathbf{z}_i, \sum_{i\neq j} \phi(\mathbf{z}_i, \mathbf{z}_{i,j}, \mathbf{z}_j)))<br />
\end{align}<br />
where <math>\phi: \mathbb{R}^{2d+e} \rightarrow \mathbb{R}^L, \alpha: \mathbb{R}^{d + L} \rightarrow \mathbb{R}^{W}, p: \mathbb{R}^{W+d} \rightarrow \mathbb{R}</math>.<br />
<br />
Notice that for the dimensions of inputs and outputs, <math>d</math> refers to the number of singleton features in <math>z</math> and <math>e</math> refers to the number of edges. <br />
<br />
[[File:GPI_architecture.jpg|thumb|A schematic representation of the GPI architecture. Singleton features <math>z_i</math> are omitted for simplicity. First, the features <math>z_{i,j}</math> are processed element-wise by <math>\phi</math>. Next, they are summed to create a vector <math>s_i</math>, which is concatenated with <math>z_i</math>. Third, a representation of the entire graph is created by applying <math>\alpha\ n</math> times and summing the created vector. The graph representation is then finally processed by <math>\rho</math> together with <math>z_k</math>.|600px|center]]<br />
<br />
==Proof Sketch for Theorem 1==<br />
The proof of this theorem can be found in the paper. A proof sketch below is provided:<br />
<br />
'''For the forward direction''' (function that follows the form set out in equation (1) is GPI):<br />
# Using definition of permutation <math>\sigma</math>, and rewriting <math>[F(z)]_{\sigma(k)}</math> in the form from equation (1)<br />
# Second argument of <math>\rho</math> is invariant under <math>\sigma</math>, since it takes the sum of all indices <math>i</math> and all other indices <math>j \neq i </math>.<br />
<br />
'''For the backward direction''' (any black-box GPI function can be expressed in the form of equation 1):<br />
# Construct <math>\phi, \alpha</math> such that second argument of <math>\rho</math> contains all information about graph features of <math>z</math>, including edges that the features originate from<br />
# Assume each <math>z_k</math> uniquely identifies the node and <math>\mathcal{F}</math> is a function only of pairwise features <math>z_{i,j}</math><br />
# Construct <math>H</math> be a perfect hash function with <math>L</math> buckets, and <math>\phi</math> which maps '''pairwise features''' to a vector of size <math>L</math><br />
# <math>*</math>Construct <math>\phi(z_i, z_{i,j}, z_j) = \mathbf{1}[H(z_j)] z_{i,j}</math>, which intuitively means that <math>\phi</math> stores <math>z_{i,j}</math> in the unique bucket for node <math>j</math><br />
# Construct function <math>\alpha</math> to output a matrix <math>\mathbb{R}^{L \times L}</math> that maps each pairwise feature into unique positions (<math>\alpha(z_i, s_i) = \mathbf{1}[H(z_i)]s_i^T</math>)<br />
# Construct matrix <math>M = \sum_i \alpha(z_i,s_i)</math> by discarding rows/columns in <math>M</math> that do not correspond to original nodes (which reduces dimension to <math>n\times n</math>; set <math>\rho</math> to have same outcome as <math>\mathcal{F}</math>, and set the output of <math>\mathcal{F}</math> on <math>M</math> to be the labels <math>\mathbf{y} = y_1, \dots, y_n</math><br />
<br />
<math>*</math>The paper presents the proof for the edge features <math>z_{ij}</math> being scalar (<math>e = 1</math>) for simplicity, which can be extended easily to vectors with additional indexing.<br />
<br />
Although the results discussed previously apply to complete graphs (edges apply to all feature pairs), it can be easily extended to incomplete graphs. However, in place of permutation-invariance, it is now an automorphism-invariance.<br />
<br />
==Implications and Applications of Theorem 1==<br />
===Key Implications of Theorem 1===<br />
# Architecture "collects" information from the different edges of the graph, and does so in an invariant fashion using <math>\alpha</math> and <math>\phi</math><br />
# Architecture is parallelizable, since all <math>\phi</math> functions can be applied simultaneously<br />
<br />
===Some applications of Theorem 1===<br />
# '''Attention:''' the concept of attention can be implemented in the GPI characterization, with slight alterations to the functions <math>\alpha</math> and <math>\phi</math>. The complete details can be found in the supplementary materials of the paper.<br />
# '''RNN:''' recurrent architectures can maintain GPI property, since all GPI function <math>\mathcal{F}</math> are closed under composition. The output of one step after running <math>\mathcal{F}</math> will act as input for the next step, but maintain the GPI property throughout.<br />
<br />
=Related Work=<br />
# '''Architectural invariance:''' suggested recently in a 2017 paper called Deep Sets by Zaheer et al., which considers the case of invariance that is more restrictive.<br />
# '''Deep structured prediction:''' previous work applied deep learning to structured prediction, for instance, semantic segmentation. Some algorithms include message passing algorithms, gradient descent for maximizing score functions, greedy decoding (inference of labels based on time of previous labels). Apart from those algorithms, deep learning has been applied to other graph-based problems such as the Travelling Salesman Problem (Bello et al., 2016; Gilmer et al., 2017; Khalil et al., 2017). However, none of the previous work specifically address the notion of invariance in the general architecture, but rather focus on message passing architectures that can be generalized by this paper.<br />
# '''Scene graph prediction:''' scene graph extraction allows for reasoning, question answering, and image retrieval (Johnson et al., 2015; Lu et al., 2016; Raposo et al., 2017). Some other works in this area include object detection, action recognition, and even detection of human-object interactions (Liao et al., 2016; Plummer et al., 2017). Additional work has been done with the use of message passing algorithms (Xu et al., 2017), word embeddings (Lu et al., 2016), and end-to-end prediction directly from pixels (Newell & Deng, 2017). A notable mention is NeuralMotif (Zellers et al., 2017), which the authors describe as the current state-of-the-art model for scene graph predictions on Visual Genome dataset.<br />
# '''Burst Image Deblurring Using Permutation Invariant Convolutional Neural Networks:''' similar ideas were applied, where Permutation Invariant CNN, are used to restore sharp and noise-free images from bursts of photographs affected by hand tremor and noise. This presented good quality images with lots of details for challenging datasets.<br />
<br />
=Experimental Results=<br />
==Synthetic Graph Labeling==<br />
The authors created a synthetic problem to study GPI. This involved using an input graph <math>G = (V,E)</math> where each node <math>i</math> belongs to the set <math>\Gamma(i) \in \{1, \dots, K\}</math> where <math>K</math> is the number of samples. The task is to compute for each node, the number of neighbours that belong to the same set (i.e. finding the label of the node <math>i</math> if <math>y_i = \sum_{j \in N(i)} \mathbf{1}[\Gamma(i) = \Gamma(j)]</math>) . Then, random graphs (each with 10 nodes) were generated by sampling edges, and the set <math>\Gamma(i) \in \{1, \dots, K\}</math>for each node independently and uniformly.<br />
The node features of the graph <math>z_i \in \{0,1\}^K</math> are one-hot vectors of <math>\Gamma(i)</math>, and each pairwise edge feature <math>z_{ij} \in \{0, 1\}</math> denote whether the edge <math>ij</math> is in the edge set <math>E</math>. <br />
3 architectures were studied in this paper:<br />
# '''GPI-architecture for graph prediction''' (without attention and RNN)<br />
# '''LSTM''': replacing <math>\sum \phi(\cdot)</math> and <math>\sum \alpha(\cdot)</math> in the form of Theorem 1 using two LSTMs with state size 200, reading their input in random order<br />
# '''Fully connected feed-forward network''': with 2 hidden layers, each layer containing 1,000 nodes; the input is a concatenation of all nodes and pairwise features, and the output is all node predictions<br />
<br />
The results show that the GPI architecture requires far fewer samples to converge to the correct solution.<br />
[[File:GPI_synthetic_example.jpg|450px|center]]<br />
<br />
==Scene-Graph Classification==<br />
Applying the concept of GPI to Scene-Graph classification (SGP) is the main task of this paper. The input to this problem is an image, along with a set of annotated bounding boxes for the entities in the image. The goal is to correctly label each entity within the bounding boxes and the relationship between every pair of entities, resulting in a coherent scene graph.<br />
<br />
The authors describe two different types of variables to predict. The first type is entity variables <math>[y_1, \dots, y_n]</math> for all bounding boxes, where each <math>y_i</math> can take one of L values and refers to objects such as "dog" or "man". The second type is relation variables <math>[y_{n+1}, \cdots, y_{n^2}]</math>, where each <math>y_i</math> represents the relation (e.g. "on", "below") between a pair of bounding boxes (entities).<br />
<br />
The scene graph and contain two types of edges:<br />
# '''Entity-entity edge''': connecting two entities <math>y_i</math> and <math>y_j</math> for <math>1 \leq i \neq j \leq n</math><br />
# '''Entity-relation edges''': connecting every relation variable <math>y_k</math> for <math>k > n</math> to two entities<br />
<br />
The feature set <math>\mathbf{z}</math> is based on the baseline model from Zellers et al. (2017). For entity variables <math>y_i</math>, the vector <math>\mathbf{z}_i \in \mathbb{R}^L</math> models the probability of the entity appearing in <math>y_i</math>. <math>\mathbf{z}_i</math> is augmented by the coordinates of the bounding box. Similarly for relation variables <math>y_j</math>, the vector <math>\mathbf{z}_j \in \mathbb{R}^R</math>, models the probability of the relations between the two entities in <math>j</math>. For entity-entity pairwise features <math>\mathbf{z}_{i,j}</math>, there is a similar representation of the probabilities for the pair. The SGP outputs probability distributions over all entities and relations, which will then be used as input recurrently to maintain GPI. Finally, word embeddings are used and concatenated for the most probable entity-relation labels.<br />
<br />
'''Components of the GPI architecture''' (ent for entity, rel for relation)<br />
# <math>\phi_{ent}</math>: network that integrates two entity variables <math>y_i</math> and <math>y_j</math>, with input <math>z_i, z_j, z_{i,j}</math> and output vector of <math>\mathbb{R}^{n_1}</math> <br />
# <math>\alpha_{ent}</math>: network with inputs from <math>\phi_{ent}</math> for all neighbours of an entity, and uses attention mechanism to output vector <math>\mathbb{R}^{n_2}</math> <br />
# <math>\rho_{ent}</math>: network with inputs from the various <math>\mathbb{R}^{n_2}</math> vectors, and outputs <math>L</math> logits to predict entity value<br />
# <math>\rho_{rel}</math>: network with inputs <math>\alpha_{ent}</math> of two entities and <math>z_{i,j}</math>, and output into <math>R</math> logits<br />
<br />
==Set-up and Results==<br />
'''Dataset''': based on Visual Genome (VG) by (Krishna et al., 2017), which contains a total of 108,077 images annotated with bounding boxes, entities, and relations. An average of 12 entities and 7 relations exist per image. For a fair comparison with previous works, data from (Xu et al., 2017) for train and test splits were used. The authors used the same 150 entities and 50 relations as in (Xu et al., 2017; Newell & Deng, 2017; Zellers et al., 2017). Hyperparameters were tuned using a 70K/5K/32K split for training, validation, and testing respectively.<br />
<br />
'''Training''': all networks were trained using the Adam optimizer, with a batch size of 20. The loss function was the sum of cross-entropy losses over all of entities and relations. Penalties for misclassified entities were 4 times stronger than that of relations. In contrast, penalties for misclassified negative relations were 10 times weaker than that of positive relations.<br />
<br />
'''Evaluation''': there are three major tasks when inferring from the scene graph. The authors focus on the following:<br />
# '''SGCIs''': given ground-truth entity bounding boxes, predict all entity and relations categories<br />
# '''PredCIs''': given annotated bounding boxes with entity labels, predict all relations<br />
<br />
The evaluation metric Recall@K (shortened to R@K) is drawn from (Lu et al., 2016). This metric is the fraction of correct ground-truth triplets that appear within the <math>K</math> most confident triplets predicted by the model. Graph-constrained protocol requires the top-<math>K</math> triplets to assign one consistent class per entity and relation. The unconstrained protocol does not enforce such constraint.<br />
<br />
'''Models and baselines''': The authors compared variants of the GPI approach against four baselines, state-of-the-art models on completing scene graph sub-tasks. To maintain consistency, all models used the same training/testing data split, in addition to the preprocessing as per (Xu et al., 2017).<br />
<br />
'''Baselines from existing state-of-the-art models'''<br />
# (Lu et al., 2016): use of word embeddings to fine-tune the likelihood of predicted relations<br />
# (Xu et al., 2017): message passing algorithm between entities and relations to iteratively improve feature map for prediction<br />
# (Newell & Deng, 2017): Pixel2Graph, uses associative embeddings to produce a full graph from image<br />
# (Zellers et al., 2017): NeuralMotif method, encodes global context to capture higher-order motif in scene graphs; Baseline outputs entities and relations distributions without using global context<br />
<br />
'''GPI models'''<br />
# '''GPI with no attention mechanism''': simply following Theorem 1's functional form, with summation over features<br />
# '''GPI NeighborAttention''': same GPI model, but considers attention over neighbours features<br />
# '''GPI Linguistic''': similar to NeighborAttention model, but concatenates word embedding vectors<br />
<br />
'''Key Results''': The GPI Linguistic approach outperforms all baseline for SGCIs, and has similar performance to the state of the art NeuralMotifs method. The authors argue that PredCI is an easier task with less structure, yielding high performance for the existing state of the art models.<br />
<br />
[[File:GPI_table_results.png|700px|center]]<br />
<br />
<br />
=Conclusion=<br />
<br />
A deep learning approach was presented in this paper to structured prediction, which constrains the architecture to be invariant to structurally identical inputs. This approach relies on pairwise features which are capable of describing inter-label correlations and inherits the intuitive aspect of score-based approaches. The output produced is invariant to equivalent representation of the pairwise terms. <br />
<br />
As future work, the axiomatic approach can be extended. Additionally, exploring algorithms that discover symmetries for deep structured prediction when invariant structure is unknown and should be discovered from data is also an interesting extension of this work. <br />
<br />
<br />
=Critique=<br />
The paper's contribution comes from the novelty of the permutation invariance as a design guideline for structured prediction. Although not explicitly considered in many of the previous works, the idea of invariance in architecture has already been considered in Deep Sets by (Zaheer et al., 2017). This paper characterizes relaxes the condition on the invariance as compared to that of previous works. In the evaluation of the benefit of GPI models, the paper used a synthetic problem to illustrate the fact that far fewer samples are required for the GPI model to converge to 100\% accuracy. However, when comparing the true task of scene graph prediction against the state-of-the-art baselines, the GPI variants had only marginal higher Recall@K scores. The true benefit of this paper's discovery is the avoidance of maximizing a score function (leading computationally difficult problem), and instead directly producing output invariant to how we represent the pairwise terms.<br />
<br />
=References=<br />
Roei Herzig, Moshiko Raboh, Gal Chechik, Jonathan Berant, Amir Globerson, Mapping Images to Scene Graphs with Permutation-Invariant Structured Prediction, 2018.<br />
<br />
Additional resources from Moshiko Raboh's [https://github.com/shikorab/SceneGraph GitHub]</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Mapping_Images_to_Scene_Graphs_with_Permutation-Invariant_Structured_Prediction&diff=40008Mapping Images to Scene Graphs with Permutation-Invariant Structured Prediction2018-11-19T15:47:14Z<p>Wfisher: Edits (E)</p>
<hr />
<div>The paper ''Mapping Images to Scene Graphs with Permutation-Invariant Structured Prediction'' was written by Roei Herzig* from Tel Aviv University, Moshiko Raboh* from Tel Aviv University, Gal Chechik from Google Brain, Bar-Ilan University, Jonathan Berant from Tel Aviv University, and Amir Globerson from Tel Aviv University. This paper is part of the NIPS 2018 conference to be hosted in December 2018 at Montréal, Canada. This paper summary is based on version 3 of the pre-print (as at May 2018) obtained from [https://arxiv.org/pdf/1802.05451v3.pdf arXiv] <br />
<br />
(*) Equal contribution<br />
<br />
=Motivation=<br />
In the field of artificial intelligence, a major goal is to enable machines to understand complex images, such as the underlying relationships between objects that exist in each scene. Although there are models today that capture both complex labels and interactions between labels, there is a disconnect for what guidelines should be used when leveraging deep learning. This paper introduces a design principle for such models that stem from the concept of permutation invariance and proves state of the art performance on models that follow this principle.<br />
<br />
The primary contributions that this paper makes include:<br />
# Deriving sufficient and necessary conditions for respecting graph-permutation invariance in deep structured prediction architectures<br />
# Empirically proving the benefit of graph-permutation invariance<br />
# Developing a state-of-the-art model for scene graph predictions over a large set of complex visual scenes<br />
<br />
=Introduction=<br />
In order for a machine to interpret complex visual scenes, it must recognize and understand both objects and relationships between the objects in the scene. A '''scene graph''' is a representation of the set of objects and relations that exist in the scene, where objects are represented as nodes and relations are represented as edges connecting the different nodes. Hence, the prediction of the scene graph is analogous to inferring the joint set of objects and relations of a visual scene.<br />
<br />
[[File:scene_graph_example.png|600px|center]]<br />
<br />
Given that objects in scenes are interdependent on each other, joint prediction of the objects and relations is necessary. The field of structured prediction, which involves the general problem of inferring multiple inter-dependent labels, is of interest for this problem.<br />
<br />
In structured prediction models, a score function <math>s(x, y)</math> is defined to evaluate the compatibility between label <math>y</math> and input <math>x</math>. For instance, when interpreting the scene of an image, <math>x</math> refers to the image itself, and <math>y</math> refers to a complex label, which contains both the objects and the relations between objects. As with most other inference methods, the goal is to find the label <math>y*</math> such that <math>s(x,y)</math> is maximized. However, the major concern is that the space for possible label assignments grows exponentially with respect to input size. For example, although an image may seem very simple, the corpus containing possible labels for objects may be very large, rendering it difficult to optimize the scoring function. <br />
<br />
The paper presents an alternative approach, for which input <math>x</math> is mapped to structured output <math>y</math> using a "black box" neural network, omitting the definition of a score function. The main concern for this approach is the determination of the network architecture.<br />
<br />
=Structured prediction=<br />
This paper further considers structured predictions using score-based methods. For structured predictions that follow a score-based approach, a score function <math>s(x, y)</math> is used to measure how compatible label <math>y</math> is for input <math>x</math>. To optimize the score function, previous works have decomposed <math>s(x,y) = \sum_i f_i(x,y)</math> in order to facilitate efficient optimization for each local score function, <math>\max_y f_i(x,y)</math>.<br />
<br />
In the area of structured predictions, the most commonly-used score functions include the singleton score function <math>f_i(y_i, x)</math> and pairwise score function <math>f_{ij} (y_i, y_j, x)</math>. Previous works explored a two-stage architectures (learn local scores independently of the structured prediction goal), and end-to-end architectures (to include the inference algorithm within the computation graph). <br />
<br />
==Advantages of using score-based methods==<br />
# Allow for intuitive specification of local dependencies between labels, and how they map to global dependencies<br />
# Linear score functions offer natural convex surrogates<br />
# Inference in large label space is sometimes possible via exact algorithms or empirically accurate approximations<br />
<br />
The concern for modelling score functions using deep networks is that learning may no longer be convex. Hence, the paper presents properties for how deep networks can be used for structured predictions by considering architectures that do not require explicit maximization of a score function.<br />
<br />
=Background, Notations, and Definitions=<br />
We denote <math>y</math> as a structured label where <math>y = [y_1, \dots, y_n]</math><br />
<br />
'''Score functions:''' for score-based methods, the score is defined as either the sum of a set of singleton scores <math>f_i = f_i(y_i, x)</math> or the sum of pairwise scores <math>f_{ij} = f_{ij}(y_i, y_j, x)</math>.<br />
<br />
Let <math>s(x,y)</math> be the score of a score-based method. Then:<br />
<br />
<div align="center"><br />
<math>s(x,y) = \begin{cases}<br />
\sum_i f_i ~ \text{if we have a set of singleton scores}\\<br />
\sum_{ij} f_{ij} ~ \text{if we have a set of pairwise scores } \\<br />
\end{cases}</math><br />
</div><br />
<br />
'''Inference algorithm:''' an inference algorithm takes input set of local scores (either <math>f_i</math> or <math>f_{ij}</math>) and outputs an assignment of labels <math>y_1, \dots, y_n</math> that maximizes score function <math>s(x,y)</math><br />
<br />
'''Graph labeling function:''' a graph labeling function <math>\mathcal{F} : (V,E) \rightarrow Y</math> is a function that takes input of: an ordered set of node features <math>V = [z_1, \dots, z_n]</math> and an ordered set of edge features <math>E = [z_{1,2},\dots,z_{i,j},\dots,z_{n,n-1}]</math> to output set of node labels <math>\mathbf{y} = [y_1, \dots, y_n]</math>. For instance, <math>z_i</math> can be set equal to <math>f_i</math> and <math>z_{ij}</math> can be set equal to <math>f_{ij}</math>.<br />
<br />
For convenience, the joint set of nodes and edges will be denoted as <math>\mathbf{z}</math> to be a size <math>n^2</math> vector (<math>n</math> nodes and <math>n(n-1)</math> edges).<br />
<br />
'''Permutation:''' Let <math>z</math> be a set of node and edge features. Given a permutation <math>\sigma</math> of <math>\{1,\dots,n\}</math>, let <math>\sigma(z)</math> be a new set of node and edge features given by [<math>\sigma(z)]_i = z_{\sigma(i)}</math> and <math>[\sigma(z)]_{i,j} = z_{\sigma(i), \sigma(j)}</math><br />
<br />
'''One-hot representation:''' <math>\mathbf{1}[j]</math> be a one-hot vector with 1 in the <math>j^{th}</math> coordinate<br />
<br />
=Permutation-Invariant Structured prediction=<br />
<br />
With permutation-invariant structured prediction, we would expect the algorithm to produce the same result given the same score function. For instance, consider the case where we have label space for 3 variables <math>y_1, y_2, y_3</math> with input <math>\mathbf{z} = (f_1, f_2, f_3, f_{12}, f_{13}, f_{23})</math> that outputs label <math>\mathbf{y} = (y_1^*, y_2^*, y_3^*)</math>. Then if the algorithm is run on a permuted version input <math>z' = (f_2, f_1, f_3, f_{21}, f_{23}, f_{13})</math>, we would expect <math>\mathbf{y} = (y_2^*, y_1^*, y_3^*)</math> given the same score function.<br />
<br />
'''Graph permutation invariance (GPI):''' a graph labeling function <math>\mathcal{F}</math> is graph-permutation invariant, if for all permutations <math>\sigma</math> of <math>\{1, \dots, n\}</math> and for all nodes <math>z</math>, <math>\mathcal{F}(\sigma(\mathbf{z})) = \sigma(\mathcal{F}(\mathbf{z}))</math><br />
<br />
The paper presents a theorem on the necessary and sufficient conditions for a function <math>\mathcal{F}</math> to be graph permutation invariant. Intuitively, because <math>\mathcal{F}</math> is a function that takes an ordered set <math>z</math> as input, the output on <math>\mathbf{z}</math> could very well be different from <math>\sigma(\mathbf{z})</math>, which means <math>\mathcal{F}</math> needs to have some sort of symmetry in order to sustain <math>[\mathcal{F}(\sigma(\mathbf{z}))]]_k = [\mathcal{F}(\mathbf{z})]_{\sigma(k)}</math>.<br />
<br />
[[File:graph_permutation_invariance.jpg|400px|center]]<br />
<br />
==Theorem 1==<br />
Let <math>\mathcal{F}</math> be a graph labeling function. Then <math>\mathcal{F}</math> is graph-permutation invariant if and only if there exist functions <math>\alpha, \rho, \phi</math> such that for all <math>k=1, .., n</math>:<br />
\begin{align}<br />
[\mathcal{F}(\mathbf{z})]_k = \rho(\mathbf{z}_k, \sum_{i=1}^n \alpha(\mathbf{z}_i, \sum_{i\neq j} \phi(\mathbf{z}_i, \mathbf{z}_{i,j}, \mathbf{z}_j)))<br />
\end{align}<br />
where <math>\phi: \mathbb{R}^{2d+e} \rightarrow \mathbb{R}^L, \alpha: \mathbb{R}^{d + L} \rightarrow \mathbb{R}^{W}, p: \mathbb{R}^{W+d} \rightarrow \mathbb{R}</math>.<br />
<br />
Notice that for the dimensions of inputs and outputs, <math>d</math> refers to the number of singleton features in <math>z</math> and <math>e</math> refers to the number of edges. <br />
<br />
[[File:GPI_architecture.jpg|thumb|A schematic representation of the GPI architecture. Singleton features <math>z_i</math> are omitted for simplicity. First, the features <math>z_{i,j}</math> are processed element-wise by <math>\phi</math>. Next, they are summed to create a vector <math>s_i</math>, which is concatenated with <math>z_i</math>. Third, a representation of the entire graph is created by applying <math>\alpha\ n</math> times and summing the created vector. The graph representation is then finally processed by <math>\rho</math> together with <math>z_k</math>.|600px|center]]<br />
<br />
==Proof Sketch for Theorem 1==<br />
The proof of this theorem can be found in the paper. A proof sketch below is provided:<br />
<br />
'''For the forward direction''' (function that follows the form set out in equation (1) is GPI):<br />
# Using definition of permutation <math>\sigma</math>, and rewriting <math>[F(z)]_{\sigma(k)}</math> in the form from equation (1)<br />
# Second argument of <math>\rho</math> is invariant under <math>\sigma</math>, since it takes the sum of all indices <math>i</math> and all other indices <math>j \neq i </math>.<br />
<br />
'''For the backward direction''' (any black-box GPI function can be expressed in the form of equation 1):<br />
# Construct <math>\phi, \alpha</math> such that second argument of <math>\rho</math> contains all information about graph features of <math>z</math>, including edges that the features originate from<br />
# Assume each <math>z_k</math> uniquely identifies the node and <math>\mathcal{F}</math> is a function only of pairwise features <math>z_{i,j}</math><br />
# Construct <math>H</math> be a perfect hash function with <math>L</math> buckets, and <math>\phi</math> which maps '''pairwise features''' to a vector of size <math>L</math><br />
# <math>*</math>Construct <math>\phi(z_i, z_{i,j}, z_j) = \mathbf{1}[H(z_j)] z_{i,j}</math>, which intuitively means that <math>\phi</math> stores <math>z_{i,j}</math> in the unique bucket for node <math>j</math><br />
# Construct function <math>\alpha</math> to output a matrix <math>\mathbb{R}^{L \times L}</math> that maps each pairwise feature into unique positions (<math>\alpha(z_i, s_i) = \mathbf{1}[H(z_i)]s_i^T</math>)<br />
# Construct matrix <math>M = \sum_i \alpha(z_i,s_i)</math> by discarding rows/columns in <math>M</math> that do not correspond to original nodes (which reduces dimension to <math>n\times n</math>; set <math>\rho</math> to have same outcome as <math>\mathcal{F}</math>, and set the output of <math>\mathcal{F}</math> on <math>M</math> to be the labels <math>\mathbf{y} = y_1, \dots, y_n</math><br />
<br />
<math>*</math>The paper presents the proof for the edge features <math>z_{ij}</math> being scalar (<math>e = 1</math>) for simplicity, which can be extended easily to vectors with additional indexing.<br />
<br />
Although the results discussed previously apply to complete graphs (edges apply to all feature pairs), it can be easily extended to incomplete graphs. However, in place of permutation-invariance, it is now an automorphism-invariance.<br />
<br />
==Implications and Applications of Theorem 1==<br />
===Key Implications of Theorem 1===<br />
# Architecture "collects" information from the different edges of the graph, and does so in an invariant fashion using <math>\alpha</math> and <math>\phi</math><br />
# Architecture is parallelizable, since all <math>\phi</math> functions can be applied simultaneously<br />
<br />
===Some applications of Theorem 1===<br />
# '''Attention:''' the concept of attention can be implemented in the GPI characterization, with slight alterations to the functions <math>\alpha</math> and <math>\phi</math>. The complete details can be found in the supplementary materials of the paper.<br />
# '''RNN:''' recurrent architectures can maintain GPI property, since all GPI function <math>\mathcal{F}</math> are closed under composition. The output of one step after running <math>\mathcal{F}</math> will act as input for the next step, but maintain the GPI property throughout.<br />
<br />
=Related Work=<br />
# '''Architectural invariance:''' suggested recently in a 2017 paper called Deep Sets by Zaheer et al., which considers the case of invariance that is more restrictive.<br />
# '''Deep structured prediction:''' previous work applied deep learning to structured prediction, for instance, semantic segmentation. Some algorithms include message passing algorithms, gradient descent for maximizing score functions, greedy decoding (inference of labels based on time of previous labels). Apart from those algorithms, deep learning has been applied to other graph-based problems such as the Travelling Salesman Problem (Bello et al., 2016; Gilmer et al., 2017; Khalil et al., 2017). However, none of the previous work specifically address the notion of invariance in the general architecture, but rather focus on message passing architectures that can be generalized by this paper.<br />
# '''Scene graph prediction:''' scene graph extraction allows for reasoning, question answering, and image retrieval (Johnson et al., 2015; Lu et al., 2016; Raposo et al., 2017). Some other works in this area include object detection, action recognition, and even detection of human-object interactions (Liao et al., 2016; Plummer et al., 2017). Additional work has been done with the use of message passing algorithms (Xu et al., 2017), word embeddings (Lu et al., 2016), and end-to-end prediction directly from pixels (Newell & Deng, 2017). A notable mention is NeuralMotif (Zellers et al., 2017), which the authors describe as the current state-of-the-art model for scene graph predictions on Visual Genome dataset.<br />
# '''Burst Image Deblurring Using Permutation Invariant Convolutional Neural Networks:''' Similar ideas were applied, where Permutation Invariant CNN, are used to restore sharp and noise-free images from bursts of photographs affected by hand tremor and noise. This presented good quality images with lots of details for challenging datasets.<br />
<br />
=Experimental Results=<br />
==Synthetic Graph Labeling==<br />
The authors created a synthetic problem to study GPI. This involved using an input graph <math>G = (V,E)</math> where each node <math>i</math> belongs to the set <math>\Gamma(i) \in \{1, \dots, K\}</math> where <math>K</math> is the number of samples. The task is to compute for each node, the number of neighbours that belong to the same set (i.e. finding the label of the node <math>i</math> if <math>y_i = \sum_{j \in N(i)} \mathbf{1}[\Gamma(i) = \Gamma(j)]</math>) . Then, random graphs (each with 10 nodes) were generated by sampling edges, and the set <math>\Gamma(i) \in \{1, \dots, K\}</math>for each node independently and uniformly.<br />
The node features of the graph <math>z_i \in \{0,1\}^K</math> are one-hot vectors of <math>\Gamma(i)</math>, and each pairwise edge feature <math>z_{ij} \in \{0, 1\}</math> denote whether the edge <math>ij</math> is in the edge set <math>E</math>. <br />
3 architectures were studied in this paper:<br />
# '''GPI-architecture for graph prediction''' (without attention and RNN)<br />
# '''LSTM''': replacing <math>\sum \phi(\cdot)</math> and <math>\sum \alpha(\cdot)</math> in the form of Theorem 1 using two LSTMs with state size 200, reading their input in random order<br />
# '''Fully connected feed-forward network''': with 2 hidden layers, each layer containing 1,000 nodes; the input is a concatenation of all nodes and pairwise features, and the output is all node predictions<br />
<br />
The results show that the GPI architecture requires far fewer samples to converge to the correct solution.<br />
[[File:GPI_synthetic_example.jpg|450px|center]]<br />
<br />
==Scene-Graph Classification==<br />
Applying the concept of GPI to Scene-Graph classification (SGP) is the main task of this paper. The input to this problem is an image, along with a set of annotated bounding boxes for the entities in the image. The goal is to correctly label each entity within the bounding boxes and the relationship between every pair of entities, resulting in a coherent scene graph.<br />
<br />
The authors describe two different types of variables to predict. The first type is entity variables <math>[y_1, \dots, y_n]</math> for all bounding boxes, where each <math>y_i</math> can take one of L values and refers to objects such as "dog" or "man". The second type is relation variables <math>[y_{n+1}, \cdots, y_{n^2}]</math>, where each <math>y_i</math> represents the relation (e.g. "on", "below") between a pair of bounding boxes (entities).<br />
<br />
The scene graph and contain two types of edges:<br />
# '''Entity-entity edge''': connecting two entities <math>y_i</math> and <math>y_j</math> for <math>1 \leq i \neq j \leq n</math><br />
# '''Entity-relation edges''': connecting every relation variable <math>y_k</math> for <math>k > n</math> to two entities<br />
<br />
The feature set <math>\mathbf{z}</math> is based on the baseline model from Zellers et al. (2017). For entity variables <math>y_i</math>, the vector <math>\mathbf{z}_i \in \mathbb{R}^L</math> models the probability of the entity appearing in <math>y_i</math>. <math>\mathbf{z}_i</math> is augmented by the coordinates of the bounding box. Similarly for relation variables <math>y_j</math>, the vector <math>\mathbf{z}_j \in \mathbb{R}^R</math>, models the probability of the relations between the two entities in <math>j</math>. For entity-entity pairwise features <math>\mathbf{z}_{i,j}</math>, there is a similar representation of the probabilities for the pair. The SGP outputs probability distributions over all entities and relations, which will then be used as input recurrently to maintain GPI. Finally, word embeddings are used and concatenated for the most probable entity-relation labels.<br />
<br />
'''Components of the GPI architecture''' (ent for entity, rel for relation)<br />
# <math>\phi_{ent}</math>: network that integrates two entity variables <math>y_i</math> and <math>y_j</math>, with input <math>z_i, z_j, z_{i,j}</math> and output vector of <math>\mathbb{R}^{n_1}</math> <br />
# <math>\alpha_{ent}</math>: network with inputs from <math>\phi_{ent}</math> for all neighbours of an entity, and uses attention mechanism to output vector <math>\mathbb{R}^{n_2}</math> <br />
# <math>\rho_{ent}</math>: network with inputs from the various <math>\mathbb{R}^{n_2}</math> vectors, and outputs <math>L</math> logits to predict entity value<br />
# <math>\rho_{rel}</math>: network with inputs <math>\alpha_{ent}</math> of two entities and <math>z_{i,j}</math>, and output into <math>R</math> logits<br />
<br />
==Set-up and Results==<br />
'''Dataset''': based on Visual Genome (VG) by (Krishna et al., 2017), which contains a total of 108,077 images annotated with bounding boxes, entities, and relations. An average of 12 entities and 7 relations exist per image. For a fair comparison with previous works, data from (Xu et al., 2017) for train and test splits were used. The authors used the same 150 entities and 50 relations as in (Xu et al., 2017; Newell & Deng, 2017; Zellers et al., 2017). Hyperparameters were tuned using a 70K/5K/32K split for training, validation, and testing respectively.<br />
<br />
'''Training''': all networks were trained using the Adam optimizer, with a batch size of 20. The loss function was the sum of cross-entropy losses over all of entities and relations. Penalties for misclassified entities were 4 times stronger than that of relations. In contrast, penalties for misclassified negative relations were 10 times weaker than that of positive relations.<br />
<br />
'''Evaluation''': there are three major tasks when inferring from the scene graph. The authors focus on the following:<br />
# '''SGCIs''': given ground-truth entity bounding boxes, predict all entity and relations categories<br />
# '''PredCIs''': given annotated bounding boxes with entity labels, predict all relations<br />
<br />
The evaluation metric Recall@K (shortened to R@K) is drawn from (Lu et al., 2016). This metric is the fraction of correct ground-truth triplets that appear within the <math>K</math> most confident triplets predicted by the model. Graph-constrained protocol requires the top-<math>K</math> triplets to assign one consistent class per entity and relation. The unconstrained protocol does not enforce such constraint.<br />
<br />
'''Models and baselines''': The authors compared variants of the GPI approach against four baselines, state-of-the-art models on completing scene graph sub-tasks. To maintain consistency, all models used the same training/testing data split, in addition to the preprocessing as per (Xu et al., 2017).<br />
<br />
'''Baselines from existing state-of-the-art models'''<br />
# (Lu et al., 2016): use of word embeddings to fine-tune the likelihood of predicted relations<br />
# (Xu et al., 2017): message passing algorithm between entities and relations to iteratively improve feature map for prediction<br />
# (Newell & Deng, 2017): Pixel2Graph, uses associative embeddings to produce a full graph from image<br />
# (Zellers et al., 2017): NeuralMotif method, encodes global context to capture higher-order motif in scene graphs; Baseline outputs entities and relations distributions without using global context<br />
<br />
'''GPI models'''<br />
# '''GPI with no attention mechanism''': simply following Theorem 1's functional form, with summation over features<br />
# '''GPI NeighborAttention''': same GPI model, but considers attention over neighbours features<br />
# '''GPI Linguistic''': similar to NeighborAttention model, but concatenates word embedding vectors<br />
<br />
'''Key Results''': The GPI Linguistic approach outperforms all baseline for SGCIs, and has similar performance to the state of the art NeuralMotifs method. The authors argue that PredCI is an easier task with less structure, yielding high performance for the existing state of the art models.<br />
<br />
[[File:GPI_table_results.png|700px|center]]<br />
<br />
<br />
=Conclusion=<br />
<br />
A deep learning approach was presented in this paper to structured prediction, which constrains the architecture to be invariant to structurally identical inputs. This approach relies on pairwise features which are capable of describing inter-label correlations and inherits the intuitive aspect of score-based approaches. The output produced is invariant to equivalent representation of the pairwise terms. <br />
<br />
As future work, the axiomatic approach can be extended. Additionally, exploring algorithms that discover symmetries for deep structured prediction when invariant structure is unknown and should be discovered from data is also an interesting extension of this work. <br />
<br />
<br />
=Critique=<br />
The paper's contribution comes from the novelty of the permutation invariance as a design guideline for structured prediction. Although not explicitly considered in many of the previous works, the idea of invariance in architecture has already been considered in Deep Sets by (Zaheer et al., 2017). This paper characterizes relaxes the condition on the invariance as compared to that of previous works. In the evaluation of the benefit of GPI models, the paper used a synthetic problem to illustrate the fact that far fewer samples are required for the GPI model to converge to 100\% accuracy. However, when comparing the true task of scene graph prediction against the state-of-the-art baselines, the GPI variants had only marginal higher Recall@K scores. The true benefit of this paper's discovery is the avoidance of maximizing a score function (leading computationally difficult problem), and instead directly producing output invariant to how we represent the pairwise terms.<br />
<br />
=References=<br />
Roei Herzig, Moshiko Raboh, Gal Chechik, Jonathan Berant, Amir Globerson, Mapping Images to Scene Graphs with Permutation-Invariant Structured Prediction, 2018.<br />
<br />
Additional resources from Moshiko Raboh's [https://github.com/shikorab/SceneGraph GitHub]</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Learning_to_Navigate_in_Cities_Without_a_Map&diff=40007Learning to Navigate in Cities Without a Map2018-11-19T15:28:50Z<p>Wfisher: edit (E)</p>
<hr />
<div>Paper: <br />
Learning to Navigate in Cities Without a Map[https://arxiv.org/pdf/1804.00168.pdf]<br />
A video of the paper is available here[https://sites.google.com/view/streetlearn].<br />
<br />
== Introduction ==<br />
Navigation is an attractive topic in many research disciplines and technology related domains such as neuroscience and robotics. The majority of algorithms are based on the following steps.<br />
<br />
1. Building an explicit map<br />
<br />
2. Planning and acting using that map. <br />
<br />
In this article, based on this fact that human can learn to navigate through cities without using any special tool such as maps or GPS, authors propose new methods to show that a neural network agent can do the same thing by using visual observations. To do so, an interactive environment using Google StreetView Images and a dual pathway agent architecture is designed. As shown in figure 1, some parts of the environment are built using Google StreetView images of New York City (Times Square, Central Park) and London (St. Paul’s Cathedral). The green cone represents the agent’s location and orientation. Although learning to navigate using visual aids is shown to be successful in some domains such as games and simulated environments using deep reinforcement learning (RL), it suffers from data inefficiency and sensitivity to changes in the environment. Thus, it is unclear whether this method could be used for large-scale navigation. That’s why it became the subject of investigation in this paper.<br />
[[File:figure1-soroush.png|600px|thumb|center|Figure 1. Our environment is built of real-world places from StreetView. The figure shows diverse views and corresponding local maps in New York City (Times Square, Central Park) and London (St. Paul’s Cathedral). The green cone represents the agent’s location and orientation.]]<br />
<br />
==Contribution==<br />
This paper has made the following contributions:<br />
<br />
1. Designing a dual pathway agent architecture. This agent can navigate through a real city. The agent is trained with end-to-end reinforcement learning.<br />
<br />
2. Using Goal-dependent learning. This means that the policy and value functions must adapt themselves to a sequence of goals that are provided as input.<br />
<br />
3. Leveraging a recurrent neural architecture. Using that, not only could navigation through a city be possible, but also the model is scalable for navigation in new cities.<br />
<br />
4. Using a new environment which is built on top of Google StreetView images. This provides real-world images for agent’s observation. Using this environment, the agent can navigate from an arbitrary starting point to a goal and then to another goal etc. Also, London, Paris, and New York City are chosen for navigation.<br />
<br />
==Related Work==<br />
<br />
1. Localization from real-world imagery. For example, (Weyand et al., 2016), a CNN was able to achieve excellent results on geolocation task. This paper provides novel work by not including supervised training with ground-truth labels, and by including planning as a goal.<br />
<br />
2. Deep RL methods for navigation. For instance, (Mirowski et al., 2016; Jaderberg et al., 2016) used self-supervised auxiliary tasks to produce visual navigation in several created mazes. This paper makes use of real-world data, in contrast to many related papers in this area.<br />
<br />
3. Deep RL for path planning and mapping. For example, (Zhang et al., 2017) created an agent that represented a global map.<br />
<br />
==Environment==<br />
Google StreetView consists of both high-resolution 360-degree imagery and graph connectivity. Also, it provides a public API. These features make it a valuable resource. In this work, large areas of New York, Paris, and London that contain between 7,000 and 65,500 nodes<br />
(and between 7,200 and 128,600 edges, respectively), have a mean node spacing of 10m and cover a range of up to<br />
5km chosen (Figure 2), without simplifying the underlying connections. This means that there are many areas 'congested' with nodes, occlusions, available footpaths, etc. The agent only sees RGB images that are visible in StreetView images (Figure 1) and is not aware of the underlying graph.<br />
<br />
[[File:figure2-soroush.png|700px|thumb|center|Figure 2. Map of the 5 environments in New York City; our experiments focus on the NYU area as well as on transfer learning from the other areas to Wall Street (see Section 5.3). In the zoomed in area, each green dot corresponds to a unique panorama, the goal is marked in blue, and landmark locations are marked with red pins.]]<br />
<br />
==Agent Interface and the Courier Task==<br />
In RL environment, we need to define observations and actions in addition to tasks. The inputs to the agent are the image <math>x_t</math> and the goal <math>g_t</math>. Also, a first-person view of the 3D environment is simulated by cropping <math>x_t</math> to a 60-degree square RGB image that is scaled to 84*84 pixels. Furthermore, the action space consists of 5 movements: “slow” rotate left or right (±22:5), “fast” rotate left or right (±67.5), or move forward (implemented as a ''noop'' in the case where this is not a viable action).<br />
<br />
There are lots of ways to specify the goal to the agent. In this paper, the current goal is chosen to be represented in terms of its proximity to a set L of fixed landmarks <math> L={(Lat_k, Long_k)}</math> which are specified using Latitude and Longitude coordinate system. For distance to the <math> k_{th}</math> landmark <math>{(d_{(t,k)}^g})_k</math> the goal vector contains <math> g_{(t,i)}=\tfrac{exp(-αd_{(t,i)}^g)}{∑_k exp(-αd_{(t,k)}^g)} </math>for <math>i_{th}</math> landmark with <math>α=0.002</math> (Figure 3).<br />
<br />
[[File:figure3-soroush.PNG|400px|thumb|center|Figure 3. We illustrate the goal description by showing a goal and a set of 5 landmarks that are nearby, plus 4 that are more distant. The code <math>g_i</math> is a vector with a softmax-normalised distance to each landmark.]]<br />
<br />
This form of representation has several advantages: <br />
<br />
1. It could easily be extended to new environments.<br />
<br />
2. It is intuitive. Even humans and animals use landmarks to be able to move from one place to another.<br />
<br />
3. It does not rely on arbitrary map coordinates, and provides an absolute (as opposed to relative) goal.<br />
<br />
In this work, 644 landmarks for New York, Paris, and London are manually defined. Furthermore, in each episode, which consists of 1000 steps, the agent starts from a random place with random orientation. when an agent gets within 100 meters of goal, the next goal is randomly chosen. Finally, the reward is proportional to the shortest path between agent and goal when the goal is first assigned (providing more reward for longer journeys).<br />
<br />
==Methods==<br />
<br />
===Goal-dependent Actor-Critic Reinforcement Learning===<br />
In this paper, the learning problem is based on Markov Decision Process, with state space S, action space A, environment Ɛ, and a set of possible goals G. The reward function depends on the current goal and state: <math>R: S×G×A → R</math>. maximize the expected sum of<br />
discounted rewards starting from state <math>s_0</math> with discount Ƴ. Also the expected return from <math>s_t</math> depends on the goals that are sampled. So, policy and value functions are as follows.<br />
<br />
\begin{align}<br />
g_t:π(α|s,g)=Pr(α_t=α|s_t=s, g_t=g)<br />
\end{align}<br />
<br />
\begin{align}<br />
V^π(s,g)=E[R_t]=E[Σ_{k=0}^∞Ƴ^kr_{t+k}|s_t=s, g_t=g]<br />
\end{align}<br />
<br />
Also, an architecture with multiple pathways is designed to support two types of learning that is required for this problem. First, an agent needs an internal representation which is general and gives an understanding of a scene. Second, the agent needs to remember features that are available in a specific place.<br />
<br />
===Architectures===<br />
<br />
[[File:figure4-soroush.png|400px|thumb|center|Figure 4. Comparison of architectures. Left: GoalNav is a convolutional encoder plus policy LSTM with goal description input. Middle: CityNav is a single-city navigation architecture with a separate goal LSTM and optional auxiliary heading (θ). Right: MultiCityNav is a multi-city architecture with individual goal LSTM pathways for each city.]]<br />
<br />
The agent takes image pixels as input. Then, These pixels are passed through a convolutional network. The output of the Convolution network is fed to a Long Short-Term Memory (LSTM) as well as the past reward <math>r_{t-1}</math><br />
and previous action <math>α_{t-1}</math>.<br />
<br />
Three different architectures are described below.<br />
<br />
The '''GoalNav''' architecture (Fig. 4a) which consists of a convolutional architecture and policy LSTM. Goal description <math>g_t</math>, previous action, and reward are the inputs of this LSTM.<br />
<br />
The '''CityNav''' architecture (Fig. 4b) consists of the previous architecture alongside an additional LSTM, called the goal LSTM. Inputs of this LSTM are visual features and the goal description. The CityNav agent also adds an auxiliary heading (θ) prediction task which is defined as an angle between the north direction and the agent’s pose. This auxiliary task can speed up learning and provides relevant information. <br />
<br />
The '''MultiCityNav''' architecture (Fig. 4c) is an extension of City-Nav for learning in different cities. This is done using the parallel connection of goal LSTMs for encapsulating locale-specific features, for each city. Moreover, the convolutional architecture and the policy LSTM become general after training on a number of cities. So, new goal LSTMs are required to be trained in new cities.<br />
<br />
===Curriculum Learning===<br />
In curriculum learning, the model is trained using simple examples in first steps. As soon as the model learns those examples, more complex and difficult examples would be fed to the model. In this paper, this approach is used to teach agent to navigate to further destinations. This courier task suffers from a common problem of RL tasks which is sparse rewarding very sparse rewards. To overcome this problem, a natural curriculum scheme is defined, in which sampling each new goal would be within 500m of the agent’s position. Then, the maximum range increases gradually to cover the full range(3.5km in the smaller New York areas, or 5km for central London or Downtown Manhattan)<br />
<br />
==Results==<br />
In this section, the performance of the proposed architectures on the courier task is shown.<br />
<br />
[[File:figure5-2.png|600px|thumb|center|Figure 5. Average per-episode goal rewards (y-axis) are plotted vs. learning steps (x-axis) for the courier task in the NYU (New York City) environment (top), and in central London (bottom). We compare the GoalNav agent, the CityNav agent, and the CityNav agent without skip connection on the NYU environment, and the CityNav agent in London. We also compare the Oracle performance and a Heuristic agent, described below. The London agents were trained with a 2-phase curriculum– we indicate the end of phase 1 (500m only) and the end of phase 2 (500m to 5000m). Results on the Rive Gauche part of Paris (trained in the same way<br />
as in London) are comparable and the agent achieved mean goal reward 426.]]<br />
<br />
It is first shown that the CityNav agent, trained with curriculum learning, succeeds in learning the courier task in New York, London and Paris. Figure 5 compares the following agents:<br />
<br />
1. Goal Navigation agent.<br />
<br />
2. City Navigation Agent.<br />
<br />
3. A City Navigation agent without the skip connection from the vision layers to the policy LSTM. This is needed to regularise the interface between the goal LSTM and the policy LSTM in multi-city transfer scenario.<br />
<br />
Also, a lower bound (Heuristic) and an upper bound(Oracle) on the performance is considered. As it is said in the paper: "Heuristic is a random walk on the street graph, where the agent turns in a random direction if it cannot move forward; if at an intersection it will turn with a probability <math>P=0.95</math>. Oracle uses the full graph to compute the optimal path using breadth-first search.". As it is clear in Figure 5, CityNav architecture with the previously mentioned architecture attains a higher performance and is more stable than the simpler GoalNav agent.<br />
<br />
The trajectories of the trained agent over two 1000 step episodes and the value function of the agent during navigation to a destination is shown in Figure 6.<br />
<br />
[[File:figure6-soroush.png|400px|thumb|center|Figure 6. Trained CityNav agent’s performance in two environments: Central London (left panes), and NYU (right panes). Top: examples of the agent’s trajectory during one 1000-step episode, showing successful consecutive goal acquisitions. The arrows show the direction of travel of the agent. Bottom: We visualize the value function of the agent during 100 trajectories with random starting points and the same goal (respectively St Paul’s Cathedral and Washington Square). Thicker and warmer color lines correspond to higher value functions.]]<br />
<br />
Figure 7 shows that navigation policy is learned by agent successfully in St Paul’s Cathedral in London and Washington Square in New York.<br />
[[File:figure7-soroush.png|400px|thumb|center|Figure 7. Number of steps required for the CityNav agent to reach<br />
a goal (Washington Square in New York or St Paul’s Cathedral in<br />
London) from 100 start locations vs. the straight-line distance to<br />
the goal in meters. One agent step corresponds to a forward movement<br />
of about 10m or a left/right turn by 22.5 or 67.5 degrees.]]<br />
<br />
A critical test for this article is to transfer model to new cities by learning a new set of landmarks, but without re-learning visual representation, behaviors, etc. Therefore, the MultiCityNav agent is trained on a number of cities besides freezing both the policy LSTM and the convolutional encoder. Then a new locale-specific goal LSTM is trained. The performance is compared using three different training regimes, illustrated in Fig. 9: Training on only the target city (single training); training on multiple cities, including the target city, together (joint training); and joint training on all but the target city, followed by training on the target city with the rest of the architecture frozen (pre-train and transfer). Figure 10 shows that transferring to other cities is possible. Also, training the model on more cities would increase its effectiveness. According to the paper: "Remarkably, the agent that is pre-trained on 4 regions and then transferred to Wall Street achieves comparable performance to an agent trained jointly on all the regions, and only slightly worse than single-city training on Wall Street alone". Training the model in a single city using skip connection is useful. However, it is not useful in multi-city transferring.<br />
[[File:figure9-soroush.png|400px|thumb|center|Figure 9. Illustration of training regimes: (a) training on a single city (equivalent to CityNav); (b) joint training over multiple cities with a dedicated per-city pathway and shared convolutional net and policy LSTM; (c) joint pre-training on a number of cities followed by training on a target city with convolutional net and policy LSTM frozen (only the target city pathway is optimized).]]<br />
[[File:figure10-soroush.png|400px|thumb|center|Figure 10. Joint multi-city training and transfer learning performance of variants of the MultiCityNav agent evaluated only on the target city (Wall Street). We compare single-city training on the target environment alone vs. joint training on multiple cities (3, 4, or 5-way joint training including Wall Street), vs. pre-training on multiple cities and then transferring to Wall Street while freezing the entire agent except for the new pathway (see Fig. 10). One variant has skip connections between the convolutional encoder and the policy LSTM, the other does not (no-skip).]]<br />
<br />
Giving early rewards before agent reaches the goal or adding random rewards (coins) to encourage exploration is investigated in this article. Figure 11a suggests that coins by themselves are ineffective as our task does not benefit from wide explorations. Also, as it is clear from Figure 11b, reducing the density of the landmarks does not seem to reduce the performance. Based on the results, authors chose to start sampling the goal within a radius of 500m from the agent’s location, and then progressively extend it to the maximum distance an agent could travel within the environment. In addition, to asses the importance of the goal-conditioned agents, a Goal-less CityNav agent is trained by removing inputs gt. The poor performance of this agent is clear in Figure 11b. Furthermore, reducing the density of the landmarks by the ratio of 50%, 25%, and 12:5% does not reduce the performance that much. Finally, some alternative for goal representation is investigated:<br />
<br />
a) Latitude and longitude scalar coordinates normalized to be between 0 and 1.<br />
<br />
b) Binned representation. <br />
<br />
The latitude and longitude scalar goal representations perform the best. However, since the all landmarks representation performs well while remaining independent of the coordinate system, we use this representation as the canonical one.<br />
<br />
[[File:figure11-soroush.PNG|300px|thumb|center|Figure 11. Top: Learning curves of the CityNav agent on NYU, comparing reward shaping with different radii of early rewards (ER) vs. ER with random coins vs. curriculum learning with ER 200m and no coins (ER 200m, Curr.). Bottom: Learning curves for CityNav agents with different goal representations: landmark-based, as well as latitude and longitude classification-based and regression-based.]]<br />
<br />
==Conclusion==<br />
In this paper, a deep reinforcement learning approach that enables navigation in cities is presented. Furthermore, a new courier task and a multi-city neural network agent architecture that is able to be transferred to new cities is discussed.<br />
<br />
==Critique==<br />
1. It is not clear that how this model is applicable in the real world. A real-world navigation problem needs to detect objects, people, and cars. However, it is not clear whether they are modeling them or not. From what I understood, they did not care about the collision, which is against their claim that it is a real-world problem.<br />
<br />
2. This paper is only using static google street view images as its primary source of data. But the authors must at least complement this with other dynamic data like traffic and road blockage information for a realistic model of navigation in the world.<br />
<br />
3. The 'Transfer in Multi-City Experiments' results could strengthened significantly from cross-validation (only Wall Street, which covers the smallest area of the four regions, is used as the test case). Additionally, the results do not show true 'multi-city' transfer learning, since all regions are within New York City. It is stated in the paper that not having to re-learn visual representations when transferring between cities is one of the outcomes, but the tests do not actually check for this. There are likely significant differences in the features that would be learned in NYC vs. Waterloo, for example, and this type of transfer has not been evaluated.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Learning_to_Navigate_in_Cities_Without_a_Map&diff=40006Learning to Navigate in Cities Without a Map2018-11-19T15:20:45Z<p>Wfisher: Add more info and advantages (T)</p>
<hr />
<div>Paper: <br />
Learning to Navigate in Cities Without a Map[https://arxiv.org/pdf/1804.00168.pdf]<br />
A video of the paper is available here[https://sites.google.com/view/streetlearn].<br />
<br />
== Introduction ==<br />
Navigation is an attractive topic in many research disciplines and technology related domains such as neuroscience and robotics. The majority of algorithms are based on the following steps.<br />
<br />
1. Building an explicit map<br />
<br />
2. Planning and acting using that map. <br />
<br />
In this article, based on this fact that human can learn to navigate through cities without using any special tool such as maps or GPS, authors propose new methods to show that a neural network agent can do the same thing by using visual observations. To do so, an interactive environment using Google StreetView Images and a dual pathway agent architecture is designed. As shown in figure 1, some parts of the environment are built using Google StreetView images of New York City (Times Square, Central Park) and London (St. Paul’s Cathedral). The green cone represents the agent’s location and orientation. Although learning to navigate using visual aids is shown to be successful in some domains such as games and simulated environments using deep reinforcement learning (RL), it suffers from data inefficiency and sensitivity to changes in the environment. Thus, it is unclear whether this method could be used for large-scale navigation. That’s why it became the subject of investigation in this paper.<br />
[[File:figure1-soroush.png|600px|thumb|center|Figure 1. Our environment is built of real-world places from StreetView. The figure shows diverse views and corresponding local maps in New York City (Times Square, Central Park) and London (St. Paul’s Cathedral). The green cone represents the agent’s location and orientation.]]<br />
<br />
==Contribution==<br />
This paper has made the following contributions:<br />
<br />
1. Designing a dual pathway agent architecture. This agent can navigate through a real city. The agent is trained with end-to-end reinforcement learning.<br />
<br />
2. Using Goal-dependent learning. This means that the policy and value functions must adapt themselves to a sequence of goals that are provided as input.<br />
<br />
3. Leveraging a recurrent neural architecture. Using that, not only could navigation through a city be possible, but also the model is scalable for navigation in new cities.<br />
<br />
4. Using a new environment which is built on top of Google StreetView images. This provides real-world images for agent’s observation. Using this environment, the agent can navigate from an arbitrary starting point to a goal and then to another goal etc. Also, London, Paris, and New York City are chosen for navigation.<br />
<br />
==Related Work==<br />
<br />
1. Localization from real-world imagery. For example, (Weyand et al., 2016), a CNN was able to achieve excellent results on geolocation task. This paper provides novel work by not including supervised training with ground-truth labels, and by including planning as a goal.<br />
<br />
2. Deep RL methods for navigation. For instance, (Mirowski et al., 2016; Jaderberg et al., 2016) used self-supervised auxiliary tasks to produce visual navigation in several created mazes. This paper makes use of real-world data, in contrast to many related papers in this area.<br />
<br />
3. Deep RL for path planning and mapping. For example, (Zhang et al., 2017) created an agent that represented a global map.<br />
<br />
==Environment==<br />
Google StreetView consists of both high-resolution 360-degree imagery and graph connectivity. Also, it provides a public API. These features make it a valuable resource. In this work, large areas of New York, Paris, and London that contain between 7,000 and 65,500 nodes<br />
(and between 7,200 and 128,600 edges, respectively), have a mean node spacing of 10m and cover a range of up to<br />
5km chosen (Figure 2), without simplifying the underlying connections. This means that there are many areas 'congested' with nodes, occlusions, available footpaths, etc. The agent only sees RGB images that are visible in StreetView images (Figure 1) and is not aware of the underlying graph.<br />
<br />
[[File:figure2-soroush.png|700px|thumb|center|Figure 2. Map of the 5 environments in New York City; our experiments focus on the NYU area as well as on transfer learning from the other areas to Wall Street (see Section 5.3). In the zoomed in area, each green dot corresponds to a unique panorama, the goal is marked in blue, and landmark locations are marked with red pins.]]<br />
<br />
==Agent Interface and the Courier Task==<br />
In RL environment, we need to define observations and actions in addition to tasks. The inputs to the agent are the image <math>x_t</math> and the goal <math>g_t</math>. Also, a first-person view of the 3D environment is simulated by cropping <math>x_t</math> to a 60-degree square RGB image that is scaled to 84*84 pixels. Furthermore, the action space consists of 5 movements: “slow” rotate left or right (±22:5), “fast” rotate left or right (±67.5), or move forward (implemented as a ''noop'' in the case where this is not a viable action).<br />
<br />
There are lots of ways to specify the goal to the agent. In this paper, the current goal is chosen to be represented in terms of its proximity to a set L of fixed landmarks <math> L={(Lat_k, Long_k)}</math> which are specified using Latitude and Longitude coordinate system. For distance to the <math> k_{th}</math> landmark <math>{(d_{(t,k)}^g})_k</math> the goal vector contains <math> g_{(t,i)}=\tfrac{exp(-αd_{(t,i)}^g)}{∑_k exp(-αd_{(t,k)}^g)} </math>for <math>i_{th}</math> landmark with <math>α=0.002</math> (Figure 3).<br />
<br />
[[File:figure3-soroush.PNG|400px|thumb|center|Figure 3. We illustrate the goal description by showing a goal and a set of 5 landmarks that are nearby, plus 4 that are more distant. The code <math>g_i</math> is a vector with a softmax-normalised distance to each landmark.]]<br />
<br />
This form of representation has several advantages: <br />
<br />
1. It could easily be extended to new environments.<br />
<br />
2. It is intuitive. Even humans and animals use landmarks to be able to move from one place to another.<br />
<br />
3. It does not rely on arbitrary map coordinates, and provides an absolute (as opposed to relative) goal.<br />
<br />
In this work, 644 landmarks for New York, Paris, and London are manually defined. Furthermore, in each episode, which consists of 1000 steps, the agent starts from a random place with random orientation. when an agent gets within 100 meters of goal, the next goal is randomly chosen. Finally, the reward is proportional to the shortest path between agent and goal when the goal is first assigned (providing more reward for longer journeys).<br />
<br />
==Methods==<br />
<br />
===Goal-dependent Actor-Critic Reinforcement Learning===<br />
In this paper, the learning problem is based on Markov Decision Process, with state space S, action space A, environment Ɛ, and a set of possible goals G. The reward function depends on the current goal and state: <math>R: S×G×A → R</math>. maximize the expected sum of<br />
discounted rewards starting from state <math>s_0</math> with discount Ƴ. Also the expected return from <math>s_t</math> depends on the goals that are sampled. So, policy and value functions are as follows.<br />
<br />
\begin{align}<br />
g_t:π(α|s,g)=Pr(α_t=α|s_t=s, g_t=g)<br />
\end{align}<br />
<br />
\begin{align}<br />
V^π(s,g)=E[R_t]=E[Σ_{k=0}^∞Ƴ^kr_{t+k}|s_t=s, g_t=g]<br />
\end{align}<br />
<br />
Also, an architecture with multiple pathways is designed to support two types of learning that is required for this problem. First, an agent needs an internal representation which is general and gives an understanding of a scene. Second, the agent needs to remember features that are available in a specific place.<br />
<br />
===Architectures===<br />
<br />
[[File:figure4-soroush.png|400px|thumb|center|Figure 4. Comparison of architectures. Left: GoalNav is a convolutional encoder plus policy LSTM with goal description input. Middle: CityNav is a single-city navigation architecture with a separate goal LSTM and optional auxiliary heading (θ). Right: MultiCityNav is a multi-city architecture with individual goal LSTM pathways for each city.]]<br />
<br />
The agent takes image pixels as input. Then, These pixels are passed through a convolutional network. The output of the Convolution network is fed to a Long Short-Term Memory (LSTM) as well as the past reward <math>r_{t-1}</math><br />
and previous action <math>α_{t-1}</math>.<br />
<br />
Three different architectures are described below.<br />
<br />
The '''GoalNav''' architecture (Fig. 4a) which consists of a convolutional architecture and policy LSTM. Goal description <math>g_t</math>, previous action, and reward are the inputs of this LSTM.<br />
<br />
The '''CityNav''' architecture (Fig. 4b) consists of the previous architecture alongside an additional LSTM, called the goal LSTM. Inputs of this LSTM are visual features and the goal description. The CityNav agent also adds an auxiliary heading (θ) prediction task which is defined as an angle between the north direction and the agent’s pose. This auxiliary task can speed up learning and provides relevant information. <br />
<br />
The '''MultiCityNav''' architecture (Fig. 4c) is an extension of City-Nav for learning in different cities. This is done using the parallel connection of goal LSTMs for encapsulating locale-specific features, for each city. Moreover, the convolutional architecture and the policy LSTM become general after training on a number of cities. So, new goal LSTMs are required to be trained in new cities.<br />
<br />
===Curriculum Learning===<br />
In curriculum learning, the model is trained using simple examples in first steps. As soon as the model learns those examples, more complex and difficult examples would be fed to the model. In this paper, this approach is used to teach agent to navigate to further destinations. This courier task suffers from a common problem of RL tasks which is sparse rewarding very sparse rewards. To overcome this problem, a natural curriculum scheme is defined, in which sampling each new goal would be within 500m of the agent’s position. Then, the maximum range increases gradually to cover the full range(3.5km in the smaller New York areas, or 5km for central London or Downtown Manhattan)<br />
<br />
==Results==<br />
In this section, the performance of the proposed architectures on the courier task is shown.<br />
<br />
[[File:figure5-2.png|600px|thumb|center|Figure 5. Average per-episode goal rewards (y-axis) are plotted vs. learning steps (x-axis) for the courier task in the NYU (New York City) environment (top), and in central London (bottom). We compare the GoalNav agent, the CityNav agent, and the CityNav agent without skip connection on the NYU environment, and the CityNav agent in London. We also compare the Oracle performance and a Heuristic agent, described below. The London agents were trained with a 2-phase curriculum– we indicate the end of phase 1 (500m only) and the end of phase 2 (500m to 5000m). Results on the Rive Gauche part of Paris (trained in the same way<br />
as in London) are comparable and the agent achieved mean goal reward 426.]]<br />
<br />
It is first shown that the CityNav agent, trained with curriculum learning, succeeds in learning the courier task in New York, London and Paris. Figure 5 compares the following agents. <br />
<br />
1. Goal Navigation agent.<br />
<br />
2. City Navigation Agent.<br />
<br />
3. A City Navigation agent without the skip connection from the vision layers to the policy LSTM. This is needed to regularise the interface between the goal LSTM and the policy LSTM in multi-city transfer scenario.<br />
<br />
Also, a lower bound(Heuristic) and an upper bound(Oracle) on the performance is considered. As it is said in the paper: "Heuristic is a random walk on the street graph, where the agent turns in a random direction if it cannot move forward; if at an intersection it will turn with a probability <math>P=0.95</math>. Oracle uses the full graph to compute the optimal path using breadth-first search.". As it is clear in Figure 5, CityNav architecture with the previously mentioned architecture attains a higher performance and is more stable than the simpler GoalNav agent.<br />
<br />
The trajectories of the trained agent over two 1000 step episodes and the value function of the agent during navigation to a destination is shown in Figure 6.<br />
<br />
[[File:figure6-soroush.png|400px|thumb|center|Figure 6. Trained CityNav agent’s performance in two environments: Central London (left panes), and NYU (right panes). Top: examples of the agent’s trajectory during one 1000-step episode, showing successful consecutive goal acquisitions. The arrows show the direction of travel of the agent. Bottom: We visualize the value function of the agent during 100 trajectories with random starting points and the same goal (respectively St Paul’s Cathedral and Washington Square). Thicker and warmer color lines correspond to higher value functions.]]<br />
<br />
Figure 7 shows that navigation policy is learned by agent successfully in St Paul’s Cathedral in London and Washington Square in New York.<br />
[[File:figure7-soroush.png|400px|thumb|center|Figure 7. Number of steps required for the CityNav agent to reach<br />
a goal (Washington Square in New York or St Paul’s Cathedral in<br />
London) from 100 start locations vs. the straight-line distance to<br />
the goal in meters. One agent step corresponds to a forward movement<br />
of about 10m or a left/right turn by 22.5 or 67.5 degrees.]]<br />
<br />
A critical test for this article is to transfer model to new cities by learning a new set of landmarks, but without re-learning visual representation, behaviors, etc. Therefore, the MultiCityNav agent is trained on a number of cities besides freezing both the policy LSTM and the convolutional encoder. Then a new locale-specific goal LSTM is trained. The performance is compared using three different training regimes, illustrated in Fig. 9: Training on only the target city (single training); training on multiple cities, including the target city, together (joint training); and joint training on all but the target city, followed by training on the target city with the rest of the architecture frozen (pre-train and transfer). Figure 10 shows that transferring to other cities is possible. Also, training the model on more cities would increase its effectiveness. According to the paper: "Remarkably, the agent that is pre-trained on 4 regions and then transferred to Wall Street achieves comparable performance to an agent trained jointly on all the regions, and only slightly worse than single-city training on Wall Street alone". Training the model in a single city using skip connection is useful. However, it is not useful in multi-city transferring.<br />
[[File:figure9-soroush.png|400px|thumb|center|Figure 9. Illustration of training regimes: (a) training on a single city (equivalent to CityNav); (b) joint training over multiple cities with a dedicated per-city pathway and shared convolutional net and policy LSTM; (c) joint pre-training on a number of cities followed by training on a target city with convolutional net and policy LSTM frozen (only the target city pathway is optimized).]]<br />
[[File:figure10-soroush.png|400px|thumb|center|Figure 10. Joint multi-city training and transfer learning performance of variants of the MultiCityNav agent evaluated only on the target city (Wall Street). We compare single-city training on the target environment alone vs. joint training on multiple cities (3, 4, or 5-way joint training including Wall Street), vs. pre-training on multiple cities and then transferring to Wall Street while freezing the entire agent except for the new pathway (see Fig. 10). One variant has skip connections between the convolutional encoder and the policy LSTM, the other does not (no-skip).]]<br />
<br />
Giving early rewards before agent reaches the goal or adding random rewards (coins) to encourage exploration is investigated in this article. Figure 11a suggests that coins by themselves are ineffective as our task does not benefit from wide explorations. Also, as it is clear from Figure 11b, reducing the density of the landmarks does not seem to reduce the performance. Based on the results, authors chose to start sampling the goal within a radius of 500m from the agent’s location, and then progressively extend it to the maximum distance an agent could travel within the environment. In addition, to asses the importance of the goal-conditioned agents, a Goal-less CityNav agent is trained by removing inputs gt. The poor performance of this agent is clear in Figure 11b. Furthermore, reducing the density of the landmarks by the ratio of 50%, 25%, and 12:5% does not reduce the performance that much. Finally, some alternative for goal representation is investigated:<br />
<br />
a) Latitude and longitude scalar coordinates normalized to be between 0 and 1.<br />
<br />
b) Binned representation. <br />
<br />
The latitude and longitude scalar goal representations perform the best. However, since the all landmarks representation performs well while remaining independent of the coordinate system, we use this representation as the canonical one.<br />
<br />
[[File:figure11-soroush.PNG|300px|thumb|center|Figure 11. Top: Learning curves of the CityNav agent on NYU, comparing reward shaping with different radii of early rewards (ER) vs. ER with random coins vs. curriculum learning with ER 200m and no coins (ER 200m, Curr.). Bottom: Learning curves for CityNav agents with different goal representations: landmark-based, as well as latitude and longitude classification-based and regression-based.]]<br />
<br />
==Conclusion==<br />
In this paper, a deep reinforcement learning approach that enables navigation in cities is presented. Furthermore, a new courier task and a multi-city neural network agent architecture that is able to be transferred to new cities is discussed.<br />
<br />
==Critique==<br />
1. It is not clear that how this model is applicable in the real world. A real-world navigation problem needs to detect objects, people, and cars. However, it is not clear whether they are modeling them or not. From what I understood, they did not care about the collision, which is against their claim that it is a real-world problem.<br />
<br />
2. This paper is only using static google street view images as its primary source of data. But the authors must at least complement this with other dynamic data like traffic and road blockage information for a realistic model of navigation in the world.<br />
<br />
3. The 'Transfer in Multi-City Experiments' results could strengthened significantly from cross-validation (only Wall Street, which covers the smallest area of the four regions, is used as the test case). Additionally, the results do not show true 'multi-city' transfer learning, since all regions are within New York City. It is stated in the paper that not having to re-learn visual representations when transferring between cities is one of the outcomes, but the tests do not actually check for this. There are likely significant differences in the features that would be learned in NYC vs. Waterloo, for example, and this type of transfer has not been evaluated.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Learning_to_Navigate_in_Cities_Without_a_Map&diff=40005Learning to Navigate in Cities Without a Map2018-11-19T15:15:53Z<p>Wfisher: Some description of graph</p>
<hr />
<div>Paper: <br />
Learning to Navigate in Cities Without a Map[https://arxiv.org/pdf/1804.00168.pdf]<br />
A video of the paper is available here[https://sites.google.com/view/streetlearn].<br />
<br />
== Introduction ==<br />
Navigation is an attractive topic in many research disciplines and technology related domains such as neuroscience and robotics. The majority of algorithms are based on the following steps.<br />
<br />
1. Building an explicit map<br />
<br />
2. Planning and acting using that map. <br />
<br />
In this article, based on this fact that human can learn to navigate through cities without using any special tool such as maps or GPS, authors propose new methods to show that a neural network agent can do the same thing by using visual observations. To do so, an interactive environment using Google StreetView Images and a dual pathway agent architecture is designed. As shown in figure 1, some parts of the environment are built using Google StreetView images of New York City (Times Square, Central Park) and London (St. Paul’s Cathedral). The green cone represents the agent’s location and orientation. Although learning to navigate using visual aids is shown to be successful in some domains such as games and simulated environments using deep reinforcement learning (RL), it suffers from data inefficiency and sensitivity to changes in the environment. Thus, it is unclear whether this method could be used for large-scale navigation. That’s why it became the subject of investigation in this paper.<br />
[[File:figure1-soroush.png|600px|thumb|center|Figure 1. Our environment is built of real-world places from StreetView. The figure shows diverse views and corresponding local maps in New York City (Times Square, Central Park) and London (St. Paul’s Cathedral). The green cone represents the agent’s location and orientation.]]<br />
<br />
==Contribution==<br />
This paper has made the following contributions:<br />
<br />
1. Designing a dual pathway agent architecture. This agent can navigate through a real city. The agent is trained with end-to-end reinforcement learning.<br />
<br />
2. Using Goal-dependent learning. This means that the policy and value functions must adapt themselves to a sequence of goals that are provided as input.<br />
<br />
3. Leveraging a recurrent neural architecture. Using that, not only could navigation through a city be possible, but also the model is scalable for navigation in new cities.<br />
<br />
4. Using a new environment which is built on top of Google StreetView images. This provides real-world images for agent’s observation. Using this environment, the agent can navigate from an arbitrary starting point to a goal and then to another goal etc. Also, London, Paris, and New York City are chosen for navigation.<br />
<br />
==Related Work==<br />
<br />
1. Localization from real-world imagery. For example, (Weyand et al., 2016), a CNN was able to achieve excellent results on geolocation task. This paper provides novel work by not including supervised training with ground-truth labels, and by including planning as a goal.<br />
<br />
2. Deep RL methods for navigation. For instance, (Mirowski et al., 2016; Jaderberg et al., 2016) used self-supervised auxiliary tasks to produce visual navigation in several created mazes. This paper makes use of real-world data, in contrast to many related papers in this area.<br />
<br />
3. Deep RL for path planning and mapping. For example, (Zhang et al., 2017) created an agent that represented a global map.<br />
<br />
==Environment==<br />
Google StreetView consists of both high-resolution 360-degree imagery and graph connectivity. Also, it provides a public API. These features make it a valuable resource. In this work, large areas of New York, Paris, and London that contain between 7,000 and 65,500 nodes<br />
(and between 7,200 and 128,600 edges, respectively), have a mean node spacing of 10m and cover a range of up to<br />
5km chosen (Figure 2), without simplifying the underlying connections. This means that there are many areas 'congested' with nodes, occlusions, available footpaths, etc. The agent only sees RGB images that are visible in StreetView images (Figure 1) and is not aware of the underlying graph.<br />
<br />
[[File:figure2-soroush.png|700px|thumb|center|Figure 2. Map of the 5 environments in New York City; our experiments focus on the NYU area as well as on transfer learning from the other areas to Wall Street (see Section 5.3). In the zoomed in area, each green dot corresponds to a unique panorama, the goal is marked in blue, and landmark locations are marked with red pins.]]<br />
<br />
==Agent Interface and the Courier Task==<br />
In RL environment, we need to define observations and actions in addition to tasks. The inputs to the agent are the image <math>x_t</math> and the goal <math>g_t</math>. Also, a first-person view of the 3D environment is simulated by cropping <math>x_t</math> to a 60-degree square RGB image that is scaled to 84*84 pixels. Furthermore, the action space consists of 5 movements: “slow” rotate left or right (±22:5), “fast” rotate left or right (±67.5), or move forward.<br />
<br />
There are lots of ways to specify the goal to the agent. In this paper, the current goal is chosen to be represented in terms of its proximity to a set L of fixed landmarks <math> L={(Lat_k, Long_k)}</math> which is specified using Latitude and Longitude coordinate system. For distance to the <math> k_{th}</math> landmark <math>{(d_{(t,k)}^g})_k</math> the goal vector contains <math> g_{(t,i)}=\tfrac{exp(-αd_{(t,i)}^g)}{∑_k exp(-αd_{(t,k)}^g)} </math>for <math>i_{th}</math> landmark with <math>α=0.002</math> (Figure 3).<br />
<br />
[[File:figure3-soroush.PNG|400px|thumb|center|Figure 3. We illustrate the goal description by showing a goal and a set of 5 landmarks that are nearby, plus 4 that are more distant. The code <math>g_i</math> is a vector with a softmax-normalised distance to each landmark.]]<br />
<br />
This form of representation has 2 advantages: <br />
<br />
1. It could easily be extended to new environments.<br />
<br />
2. It is intuitive. Even humans and animals use landmarks to be able to move from one place to another.<br />
<br />
In this work 644 landmarks for New York, Paris, and London is manually defined. Furthermore, in each episode, which consists of 1000 steps, the agent starts from a random place with random orientation. when an agent gets within 100 meters of goal, the next goal is randomly chosen. Finally, the goal is proportional to the shortest path between agent and goal.<br />
<br />
==Methods==<br />
<br />
===Goal-dependent Actor-Critic Reinforcement Learning===<br />
In this paper, the learning problem is based on Markov Decision Process, with state space S, action space A, environment Ɛ, and a set of possible goals G. The reward function depends on the current goal and state: <math>R: S×G×A → R</math>. maximize the expected sum of<br />
discounted rewards starting from state <math>s_0</math> with discount Ƴ. Also the expected return from <math>s_t</math> depends on the goals that are sampled. So, policy and value functions are as follows.<br />
<br />
\begin{align}<br />
g_t:π(α|s,g)=Pr(α_t=α|s_t=s, g_t=g)<br />
\end{align}<br />
<br />
\begin{align}<br />
V^π(s,g)=E[R_t]=E[Σ_{k=0}^∞Ƴ^kr_{t+k}|s_t=s, g_t=g]<br />
\end{align}<br />
<br />
Also, an architecture with multiple pathways is designed to support two types of learning that is required for this problem. First, an agent needs an internal representation which is general and gives an understanding of a scene. Second, the agent needs to remember features that are available in a specific place.<br />
<br />
===Architectures===<br />
<br />
[[File:figure4-soroush.png|400px|thumb|center|Figure 4. Comparison of architectures. Left: GoalNav is a convolutional encoder plus policy LSTM with goal description input. Middle: CityNav is a single-city navigation architecture with a separate goal LSTM and optional auxiliary heading (θ). Right: MultiCityNav is a multi-city architecture with individual goal LSTM pathways for each city.]]<br />
<br />
The agent takes image pixels as input. Then, These pixels are passed through a convolutional network. The output of the Convolution network is fed to a Long Short-Term Memory (LSTM) as well as the past reward <math>r_{t-1}</math><br />
and previous action <math>α_{t-1}</math>.<br />
<br />
Three different architectures are described below.<br />
<br />
The '''GoalNav''' architecture (Fig. 4a) which consists of a convolutional architecture and policy LSTM. Goal description <math>g_t</math>, previous action, and reward are the inputs of this LSTM.<br />
<br />
The '''CityNav''' architecture (Fig. 4b) consists of the previous architecture alongside an additional LSTM, called the goal LSTM. Inputs of this LSTM are visual features and the goal description. The CityNav agent also adds an auxiliary heading (θ) prediction task which is defined as an angle between the north direction and the agent’s pose. This auxiliary task can speed up learning and provides relevant information. <br />
<br />
The '''MultiCityNav''' architecture (Fig. 4c) is an extension of City-Nav for learning in different cities. This is done using the parallel connection of goal LSTMs for encapsulating locale-specific features, for each city. Moreover, the convolutional architecture and the policy LSTM become general after training on a number of cities. So, new goal LSTMs are required to be trained in new cities.<br />
<br />
===Curriculum Learning===<br />
In curriculum learning, the model is trained using simple examples in first steps. As soon as the model learns those examples, more complex and difficult examples would be fed to the model. In this paper, this approach is used to teach agent to navigate to further destinations. This courier task suffers from a common problem of RL tasks which is sparse rewarding very sparse rewards. To overcome this problem, a natural curriculum scheme is defined, in which sampling each new goal would be within 500m of the agent’s position. Then, the maximum range increases gradually to cover the full range(3.5km in the smaller New York areas, or 5km for central London or Downtown Manhattan)<br />
<br />
==Results==<br />
In this section, the performance of the proposed architectures on the courier task is shown.<br />
<br />
[[File:figure5-2.png|600px|thumb|center|Figure 5. Average per-episode goal rewards (y-axis) are plotted vs. learning steps (x-axis) for the courier task in the NYU (New York City) environment (top), and in central London (bottom). We compare the GoalNav agent, the CityNav agent, and the CityNav agent without skip connection on the NYU environment, and the CityNav agent in London. We also compare the Oracle performance and a Heuristic agent, described below. The London agents were trained with a 2-phase curriculum– we indicate the end of phase 1 (500m only) and the end of phase 2 (500m to 5000m). Results on the Rive Gauche part of Paris (trained in the same way<br />
as in London) are comparable and the agent achieved mean goal reward 426.]]<br />
<br />
It is first shown that the CityNav agent, trained with curriculum learning, succeeds in learning the courier task in New York, London and Paris. Figure 5 compares the following agents. <br />
<br />
1. Goal Navigation agent.<br />
<br />
2. City Navigation Agent.<br />
<br />
3. A City Navigation agent without the skip connection from the vision layers to the policy LSTM. This is needed to regularise the interface between the goal LSTM and the policy LSTM in multi-city transfer scenario.<br />
<br />
Also, a lower bound(Heuristic) and an upper bound(Oracle) on the performance is considered. As it is said in the paper: "Heuristic is a random walk on the street graph, where the agent turns in a random direction if it cannot move forward; if at an intersection it will turn with a probability <math>P=0.95</math>. Oracle uses the full graph to compute the optimal path using breadth-first search.". As it is clear in Figure 5, CityNav architecture with the previously mentioned architecture attains a higher performance and is more stable than the simpler GoalNav agent.<br />
<br />
The trajectories of the trained agent over two 1000 step episodes and the value function of the agent during navigation to a destination is shown in Figure 6.<br />
<br />
[[File:figure6-soroush.png|400px|thumb|center|Figure 6. Trained CityNav agent’s performance in two environments: Central London (left panes), and NYU (right panes). Top: examples of the agent’s trajectory during one 1000-step episode, showing successful consecutive goal acquisitions. The arrows show the direction of travel of the agent. Bottom: We visualize the value function of the agent during 100 trajectories with random starting points and the same goal (respectively St Paul’s Cathedral and Washington Square). Thicker and warmer color lines correspond to higher value functions.]]<br />
<br />
Figure 7 shows that navigation policy is learned by agent successfully in St Paul’s Cathedral in London and Washington Square in New York.<br />
[[File:figure7-soroush.png|400px|thumb|center|Figure 7. Number of steps required for the CityNav agent to reach<br />
a goal (Washington Square in New York or St Paul’s Cathedral in<br />
London) from 100 start locations vs. the straight-line distance to<br />
the goal in meters. One agent step corresponds to a forward movement<br />
of about 10m or a left/right turn by 22.5 or 67.5 degrees.]]<br />
<br />
A critical test for this article is to transfer model to new cities by learning a new set of landmarks, but without re-learning visual representation, behaviors, etc. Therefore, the MultiCityNav agent is trained on a number of cities besides freezing both the policy LSTM and the convolutional encoder. Then a new locale-specific goal LSTM is trained. The performance is compared using three different training regimes, illustrated in Fig. 9: Training on only the target city (single training); training on multiple cities, including the target city, together (joint training); and joint training on all but the target city, followed by training on the target city with the rest of the architecture frozen (pre-train and transfer). Figure 10 shows that transferring to other cities is possible. Also, training the model on more cities would increase its effectiveness. According to the paper: "Remarkably, the agent that is pre-trained on 4 regions and then transferred to Wall Street achieves comparable performance to an agent trained jointly on all the regions, and only slightly worse than single-city training on Wall Street alone". Training the model in a single city using skip connection is useful. However, it is not useful in multi-city transferring.<br />
[[File:figure9-soroush.png|400px|thumb|center|Figure 9. Illustration of training regimes: (a) training on a single city (equivalent to CityNav); (b) joint training over multiple cities with a dedicated per-city pathway and shared convolutional net and policy LSTM; (c) joint pre-training on a number of cities followed by training on a target city with convolutional net and policy LSTM frozen (only the target city pathway is optimized).]]<br />
[[File:figure10-soroush.png|400px|thumb|center|Figure 10. Joint multi-city training and transfer learning performance of variants of the MultiCityNav agent evaluated only on the target city (Wall Street). We compare single-city training on the target environment alone vs. joint training on multiple cities (3, 4, or 5-way joint training including Wall Street), vs. pre-training on multiple cities and then transferring to Wall Street while freezing the entire agent except for the new pathway (see Fig. 10). One variant has skip connections between the convolutional encoder and the policy LSTM, the other does not (no-skip).]]<br />
<br />
Giving early rewards before agent reaches the goal or adding random rewards (coins) to encourage exploration is investigated in this article. Figure 11a suggests that coins by themselves are ineffective as our task does not benefit from wide explorations. Also, as it is clear from Figure 11b, reducing the density of the landmarks does not seem to reduce the performance. Based on the results, authors chose to start sampling the goal within a radius of 500m from the agent’s location, and then progressively extend it to the maximum distance an agent could travel within the environment. In addition, to asses the importance of the goal-conditioned agents, a Goal-less CityNav agent is trained by removing inputs gt. The poor performance of this agent is clear in Figure 11b. Furthermore, reducing the density of the landmarks by the ratio of 50%, 25%, and 12:5% does not reduce the performance that much. Finally, some alternative for goal representation is investigated:<br />
<br />
a) Latitude and longitude scalar coordinates normalized to be between 0 and 1.<br />
<br />
b) Binned representation. <br />
<br />
The latitude and longitude scalar goal representations perform the best. However, since the all landmarks representation performs well while remaining independent of the coordinate system, we use this representation as the canonical one.<br />
<br />
[[File:figure11-soroush.PNG|300px|thumb|center|Figure 11. Top: Learning curves of the CityNav agent on NYU, comparing reward shaping with different radii of early rewards (ER) vs. ER with random coins vs. curriculum learning with ER 200m and no coins (ER 200m, Curr.). Bottom: Learning curves for CityNav agents with different goal representations: landmark-based, as well as latitude and longitude classification-based and regression-based.]]<br />
<br />
==Conclusion==<br />
In this paper, a deep reinforcement learning approach that enables navigation in cities is presented. Furthermore, a new courier task and a multi-city neural network agent architecture that is able to be transferred to new cities is discussed.<br />
<br />
==Critique==<br />
1. It is not clear that how this model is applicable in the real world. A real-world navigation problem needs to detect objects, people, and cars. However, it is not clear whether they are modeling them or not. From what I understood, they did not care about the collision, which is against their claim that it is a real-world problem.<br />
<br />
2. This paper is only using static google street view images as its primary source of data. But the authors must at least complement this with other dynamic data like traffic and road blockage information for a realistic model of navigation in the world.<br />
<br />
3. The 'Transfer in Multi-City Experiments' results could strengthened significantly from cross-validation (only Wall Street, which covers the smallest area of the four regions, is used as the test case). Additionally, the results do not show true 'multi-city' transfer learning, since all regions are within New York City. It is stated in the paper that not having to re-learn visual representations when transferring between cities is one of the outcomes, but the tests do not actually check for this. There are likely significant differences in the features that would be learned in NYC vs. Waterloo, for example, and this type of transfer has not been evaluated.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Learning_to_Navigate_in_Cities_Without_a_Map&diff=40002Learning to Navigate in Cities Without a Map2018-11-19T15:12:29Z<p>Wfisher: Add some description of related work (T)</p>
<hr />
<div>Paper: <br />
Learning to Navigate in Cities Without a Map[https://arxiv.org/pdf/1804.00168.pdf]<br />
A video of the paper is available here[https://sites.google.com/view/streetlearn].<br />
<br />
== Introduction ==<br />
Navigation is an attractive topic in many research disciplines and technology related domains such as neuroscience and robotics. The majority of algorithms are based on the following steps.<br />
<br />
1. Building an explicit map<br />
<br />
2. Planning and acting using that map. <br />
<br />
In this article, based on this fact that human can learn to navigate through cities without using any special tool such as maps or GPS, authors propose new methods to show that a neural network agent can do the same thing by using visual observations. To do so, an interactive environment using Google StreetView Images and a dual pathway agent architecture is designed. As shown in figure 1, some parts of the environment are built using Google StreetView images of New York City (Times Square, Central Park) and London (St. Paul’s Cathedral). The green cone represents the agent’s location and orientation. Although learning to navigate using visual aids is shown to be successful in some domains such as games and simulated environments using deep reinforcement learning (RL), it suffers from data inefficiency and sensitivity to changes in the environment. Thus, it is unclear whether this method could be used for large-scale navigation. That’s why it became the subject of investigation in this paper.<br />
[[File:figure1-soroush.png|600px|thumb|center|Figure 1. Our environment is built of real-world places from StreetView. The figure shows diverse views and corresponding local maps in New York City (Times Square, Central Park) and London (St. Paul’s Cathedral). The green cone represents the agent’s location and orientation.]]<br />
<br />
==Contribution==<br />
This paper has made the following contributions:<br />
<br />
1. Designing a dual pathway agent architecture. This agent can navigate through a real city. The agent is trained with end-to-end reinforcement learning.<br />
<br />
2. Using Goal-dependent learning. This means that the policy and value functions must adapt themselves to a sequence of goals that are provided as input.<br />
<br />
3. Leveraging a recurrent neural architecture. Using that, not only could navigation through a city be possible, but also the model is scalable for navigation in new cities.<br />
<br />
4. Using a new environment which is built on top of Google StreetView images. This provides real-world images for agent’s observation. Using this environment, the agent can navigate from an arbitrary starting point to a goal and then to another goal etc. Also, London, Paris, and New York City are chosen for navigation.<br />
<br />
==Related Work==<br />
<br />
1. Localization from real-world imagery. For example, (Weyand et al., 2016), a CNN was able to achieve excellent results on geolocation task. This paper provides novel work by not including supervised training with ground-truth labels, and by including planning as a goal.<br />
<br />
2. Deep RL methods for navigation. For instance, (Mirowski et al., 2016; Jaderberg et al., 2016) used self-supervised auxiliary tasks to produce visual navigation in several created mazes. This paper makes use of real-world data, in contrast to many related papers in this area.<br />
<br />
3. Deep RL for path planning and mapping. For example, (Zhang et al., 2017) created an agent that represented a global map.<br />
<br />
==Environment==<br />
Google StreetView consists of both high-resolution 360-degree imagery and graph connectivity. Also, it provides a public API. These features make it a valuable resource. In this work, large areas of New York, Paris, and London that contain between 7,000 and 65,500 nodes<br />
(and between 7,200 and 128,600 edges, respectively), have a mean node spacing of 10m and cover a range of up to<br />
5km chosen (Figure 2), without simplifying the underlying connections. Also, the agent only sees RGB images that are visible in StreetView images (Figure 1) and is not aware of the underlying graph.<br />
<br />
[[File:figure2-soroush.png|700px|thumb|center|Figure 2. Map of the 5 environments in New York City; our experiments focus on the NYU area as well as on transfer learning from the other areas to Wall Street (see Section 5.3). In the zoomed in area, each green dot corresponds to a unique panorama, the goal is marked in blue, and landmark locations are marked with red pins.]]<br />
<br />
==Agent Interface and the Courier Task==<br />
In RL environment, we need to define observations and actions in addition to tasks. The inputs to the agent are the image <math>x_t</math> and the goal <math>g_t</math>. Also, a first-person view of the 3D environment is simulated by cropping <math>x_t</math> to a 60-degree square RGB image that is scaled to 84*84 pixels. Furthermore, the action space consists of 5 movements: “slow” rotate left or right (±22:5), “fast” rotate left or right (±67.5), or move forward.<br />
<br />
There are lots of ways to specify the goal to the agent. In this paper, the current goal is chosen to be represented in terms of its proximity to a set L of fixed landmarks <math> L={(Lat_k, Long_k)}</math> which is specified using Latitude and Longitude coordinate system. For distance to the <math> k_{th}</math> landmark <math>{(d_{(t,k)}^g})_k</math> the goal vector contains <math> g_{(t,i)}=\tfrac{exp(-αd_{(t,i)}^g)}{∑_k exp(-αd_{(t,k)}^g)} </math>for <math>i_{th}</math> landmark with <math>α=0.002</math> (Figure 3).<br />
<br />
[[File:figure3-soroush.PNG|400px|thumb|center|Figure 3. We illustrate the goal description by showing a goal and a set of 5 landmarks that are nearby, plus 4 that are more distant. The code <math>g_i</math> is a vector with a softmax-normalised distance to each landmark.]]<br />
<br />
This form of representation has 2 advantages: <br />
<br />
1. It could easily be extended to new environments.<br />
<br />
2. It is intuitive. Even humans and animals use landmarks to be able to move from one place to another.<br />
<br />
In this work 644 landmarks for New York, Paris, and London is manually defined. Furthermore, in each episode, which consists of 1000 steps, the agent starts from a random place with random orientation. when an agent gets within 100 meters of goal, the next goal is randomly chosen. Finally, the goal is proportional to the shortest path between agent and goal.<br />
<br />
==Methods==<br />
<br />
===Goal-dependent Actor-Critic Reinforcement Learning===<br />
In this paper, the learning problem is based on Markov Decision Process, with state space S, action space A, environment Ɛ, and a set of possible goals G. The reward function depends on the current goal and state: <math>R: S×G×A → R</math>. maximize the expected sum of<br />
discounted rewards starting from state <math>s_0</math> with discount Ƴ. Also the expected return from <math>s_t</math> depends on the goals that are sampled. So, policy and value functions are as follows.<br />
<br />
\begin{align}<br />
g_t:π(α|s,g)=Pr(α_t=α|s_t=s, g_t=g)<br />
\end{align}<br />
<br />
\begin{align}<br />
V^π(s,g)=E[R_t]=E[Σ_{k=0}^∞Ƴ^kr_{t+k}|s_t=s, g_t=g]<br />
\end{align}<br />
<br />
Also, an architecture with multiple pathways is designed to support two types of learning that is required for this problem. First, an agent needs an internal representation which is general and gives an understanding of a scene. Second, the agent needs to remember features that are available in a specific place.<br />
<br />
===Architectures===<br />
<br />
[[File:figure4-soroush.png|400px|thumb|center|Figure 4. Comparison of architectures. Left: GoalNav is a convolutional encoder plus policy LSTM with goal description input. Middle: CityNav is a single-city navigation architecture with a separate goal LSTM and optional auxiliary heading (θ). Right: MultiCityNav is a multi-city architecture with individual goal LSTM pathways for each city.]]<br />
<br />
The agent takes image pixels as input. Then, These pixels are passed through a convolutional network. The output of the Convolution network is fed to a Long Short-Term Memory (LSTM) as well as the past reward <math>r_{t-1}</math><br />
and previous action <math>α_{t-1}</math>.<br />
<br />
Three different architectures are described below.<br />
<br />
The '''GoalNav''' architecture (Fig. 4a) which consists of a convolutional architecture and policy LSTM. Goal description <math>g_t</math>, previous action, and reward are the inputs of this LSTM.<br />
<br />
The '''CityNav''' architecture (Fig. 4b) consists of the previous architecture alongside an additional LSTM, called the goal LSTM. Inputs of this LSTM are visual features and the goal description. The CityNav agent also adds an auxiliary heading (θ) prediction task which is defined as an angle between the north direction and the agent’s pose. This auxiliary task can speed up learning and provides relevant information. <br />
<br />
The '''MultiCityNav''' architecture (Fig. 4c) is an extension of City-Nav for learning in different cities. This is done using the parallel connection of goal LSTMs for encapsulating locale-specific features, for each city. Moreover, the convolutional architecture and the policy LSTM become general after training on a number of cities. So, new goal LSTMs are required to be trained in new cities.<br />
<br />
===Curriculum Learning===<br />
In curriculum learning, the model is trained using simple examples in first steps. As soon as the model learns those examples, more complex and difficult examples would be fed to the model. In this paper, this approach is used to teach agent to navigate to further destinations. This courier task suffers from a common problem of RL tasks which is sparse rewarding very sparse rewards. To overcome this problem, a natural curriculum scheme is defined, in which sampling each new goal would be within 500m of the agent’s position. Then, the maximum range increases gradually to cover the full range(3.5km in the smaller New York areas, or 5km for central London or Downtown Manhattan)<br />
<br />
==Results==<br />
In this section, the performance of the proposed architectures on the courier task is shown.<br />
<br />
[[File:figure5-2.png|600px|thumb|center|Figure 5. Average per-episode goal rewards (y-axis) are plotted vs. learning steps (x-axis) for the courier task in the NYU (New York City) environment (top), and in central London (bottom). We compare the GoalNav agent, the CityNav agent, and the CityNav agent without skip connection on the NYU environment, and the CityNav agent in London. We also compare the Oracle performance and a Heuristic agent, described below. The London agents were trained with a 2-phase curriculum– we indicate the end of phase 1 (500m only) and the end of phase 2 (500m to 5000m). Results on the Rive Gauche part of Paris (trained in the same way<br />
as in London) are comparable and the agent achieved mean goal reward 426.]]<br />
<br />
It is first shown that the CityNav agent, trained with curriculum learning, succeeds in learning the courier task in New York, London and Paris. Figure 5 compares the following agents. <br />
<br />
1. Goal Navigation agent.<br />
<br />
2. City Navigation Agent.<br />
<br />
3. A City Navigation agent without the skip connection from the vision layers to the policy LSTM. This is needed to regularise the interface between the goal LSTM and the policy LSTM in multi-city transfer scenario.<br />
<br />
Also, a lower bound(Heuristic) and an upper bound(Oracle) on the performance is considered. As it is said in the paper: "Heuristic is a random walk on the street graph, where the agent turns in a random direction if it cannot move forward; if at an intersection it will turn with a probability <math>P=0.95</math>. Oracle uses the full graph to compute the optimal path using breadth-first search.". As it is clear in Figure 5, CityNav architecture with the previously mentioned architecture attains a higher performance and is more stable than the simpler GoalNav agent.<br />
<br />
The trajectories of the trained agent over two 1000 step episodes and the value function of the agent during navigation to a destination is shown in Figure 6.<br />
<br />
[[File:figure6-soroush.png|400px|thumb|center|Figure 6. Trained CityNav agent’s performance in two environments: Central London (left panes), and NYU (right panes). Top: examples of the agent’s trajectory during one 1000-step episode, showing successful consecutive goal acquisitions. The arrows show the direction of travel of the agent. Bottom: We visualize the value function of the agent during 100 trajectories with random starting points and the same goal (respectively St Paul’s Cathedral and Washington Square). Thicker and warmer color lines correspond to higher value functions.]]<br />
<br />
Figure 7 shows that navigation policy is learned by agent successfully in St Paul’s Cathedral in London and Washington Square in New York.<br />
[[File:figure7-soroush.png|400px|thumb|center|Figure 7. Number of steps required for the CityNav agent to reach<br />
a goal (Washington Square in New York or St Paul’s Cathedral in<br />
London) from 100 start locations vs. the straight-line distance to<br />
the goal in meters. One agent step corresponds to a forward movement<br />
of about 10m or a left/right turn by 22.5 or 67.5 degrees.]]<br />
<br />
A critical test for this article is to transfer model to new cities by learning a new set of landmarks, but without re-learning visual representation, behaviors, etc. Therefore, the MultiCityNav agent is trained on a number of cities besides freezing both the policy LSTM and the convolutional encoder. Then a new locale-specific goal LSTM is trained. The performance is compared using three different training regimes, illustrated in Fig. 9: Training on only the target city (single training); training on multiple cities, including the target city, together (joint training); and joint training on all but the target city, followed by training on the target city with the rest of the architecture frozen (pre-train and transfer). Figure 10 shows that transferring to other cities is possible. Also, training the model on more cities would increase its effectiveness. According to the paper: "Remarkably, the agent that is pre-trained on 4 regions and then transferred to Wall Street achieves comparable performance to an agent trained jointly on all the regions, and only slightly worse than single-city training on Wall Street alone". Training the model in a single city using skip connection is useful. However, it is not useful in multi-city transferring.<br />
[[File:figure9-soroush.png|400px|thumb|center|Figure 9. Illustration of training regimes: (a) training on a single city (equivalent to CityNav); (b) joint training over multiple cities with a dedicated per-city pathway and shared convolutional net and policy LSTM; (c) joint pre-training on a number of cities followed by training on a target city with convolutional net and policy LSTM frozen (only the target city pathway is optimized).]]<br />
[[File:figure10-soroush.png|400px|thumb|center|Figure 10. Joint multi-city training and transfer learning performance of variants of the MultiCityNav agent evaluated only on the target city (Wall Street). We compare single-city training on the target environment alone vs. joint training on multiple cities (3, 4, or 5-way joint training including Wall Street), vs. pre-training on multiple cities and then transferring to Wall Street while freezing the entire agent except for the new pathway (see Fig. 10). One variant has skip connections between the convolutional encoder and the policy LSTM, the other does not (no-skip).]]<br />
<br />
Giving early rewards before agent reaches the goal or adding random rewards (coins) to encourage exploration is investigated in this article. Figure 11a suggests that coins by themselves are ineffective as our task does not benefit from wide explorations. Also, as it is clear from Figure 11b, reducing the density of the landmarks does not seem to reduce the performance. Based on the results, authors chose to start sampling the goal within a radius of 500m from the agent’s location, and then progressively extend it to the maximum distance an agent could travel within the environment. In addition, to asses the importance of the goal-conditioned agents, a Goal-less CityNav agent is trained by removing inputs gt. The poor performance of this agent is clear in Figure 11b. Furthermore, reducing the density of the landmarks by the ratio of 50%, 25%, and 12:5% does not reduce the performance that much. Finally, some alternative for goal representation is investigated:<br />
<br />
a) Latitude and longitude scalar coordinates normalized to be between 0 and 1.<br />
<br />
b) Binned representation. <br />
<br />
The latitude and longitude scalar goal representations perform the best. However, since the all landmarks representation performs well while remaining independent of the coordinate system, we use this representation as the canonical one.<br />
<br />
[[File:figure11-soroush.PNG|300px|thumb|center|Figure 11. Top: Learning curves of the CityNav agent on NYU, comparing reward shaping with different radii of early rewards (ER) vs. ER with random coins vs. curriculum learning with ER 200m and no coins (ER 200m, Curr.). Bottom: Learning curves for CityNav agents with different goal representations: landmark-based, as well as latitude and longitude classification-based and regression-based.]]<br />
<br />
==Conclusion==<br />
In this paper, a deep reinforcement learning approach that enables navigation in cities is presented. Furthermore, a new courier task and a multi-city neural network agent architecture that is able to be transferred to new cities is discussed.<br />
<br />
==Critique==<br />
1. It is not clear that how this model is applicable in the real world. A real-world navigation problem needs to detect objects, people, and cars. However, it is not clear whether they are modeling them or not. From what I understood, they did not care about the collision, which is against their claim that it is a real-world problem.<br />
<br />
2. This paper is only using static google street view images as its primary source of data. But the authors must at least complement this with other dynamic data like traffic and road blockage information for a realistic model of navigation in the world.<br />
<br />
3. The 'Transfer in Multi-City Experiments' results could strengthened significantly from cross-validation (only Wall Street, which covers the smallest area of the four regions, is used as the test case). Additionally, the results do not show true 'multi-city' transfer learning, since all regions are within New York City. It is stated in the paper that not having to re-learn visual representations when transferring between cities is one of the outcomes, but the tests do not actually check for this. There are likely significant differences in the features that would be learned in NYC vs. Waterloo, for example, and this type of transfer has not been evaluated.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Learning_to_Navigate_in_Cities_Without_a_Map&diff=40001Learning to Navigate in Cities Without a Map2018-11-19T15:09:05Z<p>Wfisher: Small additions (T)</p>
<hr />
<div>Paper: <br />
Learning to Navigate in Cities Without a Map[https://arxiv.org/pdf/1804.00168.pdf]<br />
A video of the paper is available here[https://sites.google.com/view/streetlearn].<br />
<br />
== Introduction ==<br />
Navigation is an attractive topic in many research disciplines and technology related domains such as neuroscience and robotics. The majority of algorithms are based on the following steps.<br />
<br />
1. Building an explicit map<br />
<br />
2. Planning and acting using that map. <br />
<br />
In this article, based on this fact that human can learn to navigate through cities without using any special tool such as maps or GPS, authors propose new methods to show that a neural network agent can do the same thing by using visual observations. To do so, an interactive environment using Google StreetView Images and a dual pathway agent architecture is designed. As shown in figure 1, some parts of the environment are built using Google StreetView images of New York City (Times Square, Central Park) and London (St. Paul’s Cathedral). The green cone represents the agent’s location and orientation. Although learning to navigate using visual aids is shown to be successful in some domains such as games and simulated environments using deep reinforcement learning (RL), it suffers from data inefficiency and sensitivity to changes in the environment. Thus, it is unclear whether this method could be used for large-scale navigation. That’s why it became the subject of investigation in this paper.<br />
[[File:figure1-soroush.png|600px|thumb|center|Figure 1. Our environment is built of real-world places from StreetView. The figure shows diverse views and corresponding local maps in New York City (Times Square, Central Park) and London (St. Paul’s Cathedral). The green cone represents the agent’s location and orientation.]]<br />
<br />
==Contribution==<br />
This paper has made the following contributions:<br />
<br />
1. Designing a dual pathway agent architecture. This agent can navigate through a real city. The agent is trained with end-to-end reinforcement learning.<br />
<br />
2. Using Goal-dependent learning. This means that the policy and value functions must adapt themselves to a sequence of goals that are provided as input.<br />
<br />
3. Leveraging a recurrent neural architecture. Using that, not only could navigation through a city be possible, but also the model is scalable for navigation in new cities.<br />
<br />
4. Using a new environment which is built on top of Google StreetView images. This provides real-world images for agent’s observation. Using this environment, the agent can navigate from an arbitrary starting point to a goal and then to another goal etc. Also, London, Paris, and New York City are chosen for navigation.<br />
<br />
==Related Work==<br />
<br />
1. Localization from real-world imagery. For example, (Weyand et al., 2016), a CNN was able to achieve excellent results on geolocation task.<br />
<br />
2. Deep RL methods for navigation. For instance, (Mirowski et al., 2016; Jaderberg et al., 2016) used self-supervised auxiliary tasks to produce visual navigation in several created mazes<br />
<br />
3. Deep RL for path planning and mapping. For example, (Zhang et al., 2017) created an agent that represented a global map.<br />
<br />
==Environment==<br />
Google StreetView consists of both high-resolution 360-degree imagery and graph connectivity. Also, it provides a public API. These features make it a valuable resource. In this work, large areas of New York, Paris, and London that contain between 7,000 and 65,500 nodes<br />
(and between 7,200 and 128,600 edges, respectively), have a mean node spacing of 10m and cover a range of up to<br />
5km chosen (Figure 2), without simplifying the underlying connections. Also, the agent only sees RGB images that are visible in StreetView images (Figure 1) and is not aware of the underlying graph.<br />
<br />
[[File:figure2-soroush.png|700px|thumb|center|Figure 2. Map of the 5 environments in New York City; our experiments focus on the NYU area as well as on transfer learning from the other areas to Wall Street (see Section 5.3). In the zoomed in area, each green dot corresponds to a unique panorama, the goal is marked in blue, and landmark locations are marked with red pins.]]<br />
<br />
==Agent Interface and the Courier Task==<br />
In RL environment, we need to define observations and actions in addition to tasks. The inputs to the agent are the image <math>x_t</math> and the goal <math>g_t</math>. Also, a first-person view of the 3D environment is simulated by cropping <math>x_t</math> to a 60-degree square RGB image that is scaled to 84*84 pixels. Furthermore, the action space consists of 5 movements: “slow” rotate left or right (±22:5), “fast” rotate left or right (±67.5), or move forward.<br />
<br />
There are lots of ways to specify the goal to the agent. In this paper, the current goal is chosen to be represented in terms of its proximity to a set L of fixed landmarks <math> L={(Lat_k, Long_k)}</math> which is specified using Latitude and Longitude coordinate system. For distance to the <math> k_{th}</math> landmark <math>{(d_{(t,k)}^g})_k</math> the goal vector contains <math> g_{(t,i)}=\tfrac{exp(-αd_{(t,i)}^g)}{∑_k exp(-αd_{(t,k)}^g)} </math>for <math>i_{th}</math> landmark with <math>α=0.002</math> (Figure 3).<br />
<br />
[[File:figure3-soroush.PNG|400px|thumb|center|Figure 3. We illustrate the goal description by showing a goal and a set of 5 landmarks that are nearby, plus 4 that are more distant. The code <math>g_i</math> is a vector with a softmax-normalised distance to each landmark.]]<br />
<br />
This form of representation has 2 advantages: <br />
<br />
1. It could easily be extended to new environments.<br />
<br />
2. It is intuitive. Even humans and animals use landmarks to be able to move from one place to another.<br />
<br />
In this work 644 landmarks for New York, Paris, and London is manually defined. Furthermore, in each episode, which consists of 1000 steps, the agent starts from a random place with random orientation. when an agent gets within 100 meters of goal, the next goal is randomly chosen. Finally, the goal is proportional to the shortest path between agent and goal.<br />
<br />
==Methods==<br />
<br />
===Goal-dependent Actor-Critic Reinforcement Learning===<br />
In this paper, the learning problem is based on Markov Decision Process, with state space S, action space A, environment Ɛ, and a set of possible goals G. The reward function depends on the current goal and state: <math>R: S×G×A → R</math>. maximize the expected sum of<br />
discounted rewards starting from state <math>s_0</math> with discount Ƴ. Also the expected return from <math>s_t</math> depends on the goals that are sampled. So, policy and value functions are as follows.<br />
<br />
\begin{align}<br />
g_t:π(α|s,g)=Pr(α_t=α|s_t=s, g_t=g)<br />
\end{align}<br />
<br />
\begin{align}<br />
V^π(s,g)=E[R_t]=E[Σ_{k=0}^∞Ƴ^kr_{t+k}|s_t=s, g_t=g]<br />
\end{align}<br />
<br />
Also, an architecture with multiple pathways is designed to support two types of learning that is required for this problem. First, an agent needs an internal representation which is general and gives an understanding of a scene. Second, the agent needs to remember features that are available in a specific place.<br />
<br />
===Architectures===<br />
<br />
[[File:figure4-soroush.png|400px|thumb|center|Figure 4. Comparison of architectures. Left: GoalNav is a convolutional encoder plus policy LSTM with goal description input. Middle: CityNav is a single-city navigation architecture with a separate goal LSTM and optional auxiliary heading (θ). Right: MultiCityNav is a multi-city architecture with individual goal LSTM pathways for each city.]]<br />
<br />
The agent takes image pixels as input. Then, These pixels are passed through a convolutional network. The output of the Convolution network is fed to a Long Short-Term Memory (LSTM) as well as the past reward <math>r_{t-1}</math><br />
and previous action <math>α_{t-1}</math>.<br />
<br />
Three different architectures are described below.<br />
<br />
The '''GoalNav''' architecture (Fig. 4a) which consists of a convolutional architecture and policy LSTM. Goal description <math>g_t</math>, previous action, and reward are the inputs of this LSTM.<br />
<br />
The '''CityNav''' architecture (Fig. 4b) consists of the previous architecture alongside an additional LSTM, called the goal LSTM. Inputs of this LSTM are visual features and the goal description. The CityNav agent also adds an auxiliary heading (θ) prediction task which is defined as an angle between the north direction and the agent’s pose. This auxiliary task can speed up learning and provides relevant information. <br />
<br />
The '''MultiCityNav''' architecture (Fig. 4c) is an extension of City-Nav for learning in different cities. This is done using the parallel connection of goal LSTMs for encapsulating locale-specific features, for each city. Moreover, the convolutional architecture and the policy LSTM become general after training on a number of cities. So, new goal LSTMs are required to be trained in new cities.<br />
<br />
===Curriculum Learning===<br />
In curriculum learning, the model is trained using simple examples in first steps. As soon as the model learns those examples, more complex and difficult examples would be fed to the model. In this paper, this approach is used to teach agent to navigate to further destinations. This courier task suffers from a common problem of RL tasks which is sparse rewarding very sparse rewards. To overcome this problem, a natural curriculum scheme is defined, in which sampling each new goal would be within 500m of the agent’s position. Then, the maximum range increases gradually to cover the full range(3.5km in the smaller New York areas, or 5km for central London or Downtown Manhattan)<br />
<br />
==Results==<br />
In this section, the performance of the proposed architectures on the courier task is shown.<br />
<br />
[[File:figure5-2.png|600px|thumb|center|Figure 5. Average per-episode goal rewards (y-axis) are plotted vs. learning steps (x-axis) for the courier task in the NYU (New York City) environment (top), and in central London (bottom). We compare the GoalNav agent, the CityNav agent, and the CityNav agent without skip connection on the NYU environment, and the CityNav agent in London. We also compare the Oracle performance and a Heuristic agent, described below. The London agents were trained with a 2-phase curriculum– we indicate the end of phase 1 (500m only) and the end of phase 2 (500m to 5000m). Results on the Rive Gauche part of Paris (trained in the same way<br />
as in London) are comparable and the agent achieved mean goal reward 426.]]<br />
<br />
It is first shown that the CityNav agent, trained with curriculum learning, succeeds in learning the courier task in New York, London and Paris. Figure 5 compares the following agents. <br />
<br />
1. Goal Navigation agent.<br />
<br />
2. City Navigation Agent.<br />
<br />
3. A City Navigation agent without the skip connection from the vision layers to the policy LSTM. This is needed to regularise the interface between the goal LSTM and the policy LSTM in multi-city transfer scenario.<br />
<br />
Also, a lower bound(Heuristic) and an upper bound(Oracle) on the performance is considered. As it is said in the paper: "Heuristic is a random walk on the street graph, where the agent turns in a random direction if it cannot move forward; if at an intersection it will turn with a probability <math>P=0.95</math>. Oracle uses the full graph to compute the optimal path using breadth-first search.". As it is clear in Figure 5, CityNav architecture with the previously mentioned architecture attains a higher performance and is more stable than the simpler GoalNav agent.<br />
<br />
The trajectories of the trained agent over two 1000 step episodes and the value function of the agent during navigation to a destination is shown in Figure 6.<br />
<br />
[[File:figure6-soroush.png|400px|thumb|center|Figure 6. Trained CityNav agent’s performance in two environments: Central London (left panes), and NYU (right panes). Top: examples of the agent’s trajectory during one 1000-step episode, showing successful consecutive goal acquisitions. The arrows show the direction of travel of the agent. Bottom: We visualize the value function of the agent during 100 trajectories with random starting points and the same goal (respectively St Paul’s Cathedral and Washington Square). Thicker and warmer color lines correspond to higher value functions.]]<br />
<br />
Figure 7 shows that navigation policy is learned by agent successfully in St Paul’s Cathedral in London and Washington Square in New York.<br />
[[File:figure7-soroush.png|400px|thumb|center|Figure 7. Number of steps required for the CityNav agent to reach<br />
a goal (Washington Square in New York or St Paul’s Cathedral in<br />
London) from 100 start locations vs. the straight-line distance to<br />
the goal in meters. One agent step corresponds to a forward movement<br />
of about 10m or a left/right turn by 22.5 or 67.5 degrees.]]<br />
<br />
A critical test for this article is to transfer model to new cities by learning a new set of landmarks, but without re-learning visual representation, behaviors, etc. Therefore, the MultiCityNav agent is trained on a number of cities besides freezing both the policy LSTM and the convolutional encoder. Then a new locale-specific goal LSTM is trained. The performance is compared using three different training regimes, illustrated in Fig. 9: Training on only the target city (single training); training on multiple cities, including the target city, together (joint training); and joint training on all but the target city, followed by training on the target city with the rest of the architecture frozen (pre-train and transfer). Figure 10 shows that transferring to other cities is possible. Also, training the model on more cities would increase its effectiveness. According to the paper: "Remarkably, the agent that is pre-trained on 4 regions and then transferred to Wall Street achieves comparable performance to an agent trained jointly on all the regions, and only slightly worse than single-city training on Wall Street alone". Training the model in a single city using skip connection is useful. However, it is not useful in multi-city transferring.<br />
[[File:figure9-soroush.png|400px|thumb|center|Figure 9. Illustration of training regimes: (a) training on a single city (equivalent to CityNav); (b) joint training over multiple cities with a dedicated per-city pathway and shared convolutional net and policy LSTM; (c) joint pre-training on a number of cities followed by training on a target city with convolutional net and policy LSTM frozen (only the target city pathway is optimized).]]<br />
[[File:figure10-soroush.png|400px|thumb|center|Figure 10. Joint multi-city training and transfer learning performance of variants of the MultiCityNav agent evaluated only on the target city (Wall Street). We compare single-city training on the target environment alone vs. joint training on multiple cities (3, 4, or 5-way joint training including Wall Street), vs. pre-training on multiple cities and then transferring to Wall Street while freezing the entire agent except for the new pathway (see Fig. 10). One variant has skip connections between the convolutional encoder and the policy LSTM, the other does not (no-skip).]]<br />
<br />
Giving early rewards before agent reaches the goal or adding random rewards (coins) to encourage exploration is investigated in this article. Figure 11a suggests that coins by themselves are ineffective as our task does not benefit from wide explorations. Also, as it is clear from Figure 11b, reducing the density of the landmarks does not seem to reduce the performance. Based on the results, authors chose to start sampling the goal within a radius of 500m from the agent’s location, and then progressively extend it to the maximum distance an agent could travel within the environment. In addition, to asses the importance of the goal-conditioned agents, a Goal-less CityNav agent is trained by removing inputs gt. The poor performance of this agent is clear in Figure 11b. Furthermore, reducing the density of the landmarks by the ratio of 50%, 25%, and 12:5% does not reduce the performance that much. Finally, some alternative for goal representation is investigated:<br />
<br />
a) Latitude and longitude scalar coordinates normalized to be between 0 and 1.<br />
<br />
b) Binned representation. <br />
<br />
The latitude and longitude scalar goal representations perform the best. However, since the all landmarks representation performs well while remaining independent of the coordinate system, we use this representation as the canonical one.<br />
<br />
[[File:figure11-soroush.PNG|300px|thumb|center|Figure 11. Top: Learning curves of the CityNav agent on NYU, comparing reward shaping with different radii of early rewards (ER) vs. ER with random coins vs. curriculum learning with ER 200m and no coins (ER 200m, Curr.). Bottom: Learning curves for CityNav agents with different goal representations: landmark-based, as well as latitude and longitude classification-based and regression-based.]]<br />
<br />
==Conclusion==<br />
In this paper, a deep reinforcement learning approach that enables navigation in cities is presented. Furthermore, a new courier task and a multi-city neural network agent architecture that is able to be transferred to new cities is discussed.<br />
<br />
==Critique==<br />
1. It is not clear that how this model is applicable in the real world. A real-world navigation problem needs to detect objects, people, and cars. However, it is not clear whether they are modeling them or not. From what I understood, they did not care about the collision, which is against their claim that it is a real-world problem.<br />
<br />
2. This paper is only using static google street view images as its primary source of data. But the authors must at least complement this with other dynamic data like traffic and road blockage information for a realistic model of navigation in the world.<br />
<br />
3. The 'Transfer in Multi-City Experiments' results could strengthened significantly from cross-validation (only Wall Street, which covers the smallest area of the four regions, is used as the test case). Additionally, the results do not show true 'multi-city' transfer learning, since all regions are within New York City. It is stated in the paper that not having to re-learn visual representations when transferring between cities is one of the outcomes, but the tests do not actually check for this. There are likely significant differences in the features that would be learned in NYC vs. Waterloo, for example, and this type of transfer has not been evaluated.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Obfuscated_Gradients_Give_a_False_Sense_of_Security_Circumventing_Defenses_to_Adversarial_Examples&diff=39888Obfuscated Gradients Give a False Sense of Security Circumventing Defenses to Adversarial Examples2018-11-19T02:40:03Z<p>Wfisher: Small edits (E)</p>
<hr />
<div>= Introduction =<br />
Over the past few years, neural network models have been the source of major breakthroughs in a variety of computer vision problems. However, these networks have been shown to be susceptible to adversarial attacks. In these attacks, small humanly-imperceptible changes are made to images (that are originally correctly classified) which causes these models to misclassify with high confidence. These attacks pose a major threat that needs to be addressed before these systems can be deployed on a large scale, especially in safety-critical scenarios. <br />
<br />
The seriousness of this threat has generated major interest in both the design and defense against them. In this paper, the authors identify a common technique employed by several recently proposed defenses and design a set of attacks that can be used to overcome them. The use of this technique, masking gradients, is so prevalent, that 7 out of the 9 defenses proposed in the ICLR 2018 conference employed them. The authors were able to circumvent the proposed defenses and successfully brought down the accuracy of their models to below 10%.<br />
<br />
= Methodology =<br />
<br />
The paper assumes a lot of familiarity with adversarial attack literature. The section below briefly explains some key concepts.<br />
<br />
== Background ==<br />
<br />
==== Adversarial Images Mathematically ====<br />
Given an image <math>x</math> and a classifier <math>f(x)</math>, an adversarial image <math>x'</math> satisfies two properties:<br />
# <math>D(x,x') < \epsilon </math><br />
# <math>c(x') \neq c^*(x) </math><br />
<br />
Where <math>D</math> is some distance metric, <math>\epsilon </math> is a small constant, <math>c(x')</math> is the output ''class'' predicted by the model, and <math>c^*(x)</math> is the true class for input x. In words, the adversarial image is a small distance from the original image, but the classifier classifies it incorrectly.<br />
<br />
==== Adversarial Attacks Terminology ====<br />
#Adversarial attacks can be either '''black''' or '''white-box'''. In black box attacks, the attacker has access to the network output only, while white-box attackers have full access to the network, including its gradients, architecture and weights. This makes white-box attackers much more powerful. Given access to gradients, white-box attacks use back propagation to modify inputs (as opposed to the weights) with respect to the loss function.<br />
#In '''untargeted''' attacks, the objective is to ''maximize'' the loss of the true class, <math>x'=x \mathbf{+} \lambda(sign(\nabla_xL(x,c^*(x))))</math>. While in '''targeted''' attacks, the objective is to ''minimize loss for a target class'' <math>c^t(x)</math> that is different from the true class, <math>x'=x \mathbf{-} \epsilon(sign(\nabla_xL(x,c^t(x))))</math>. Here, <math>\nabla_xL()</math> is the gradient of the loss function with respect to the input, <math>\lambda</math> is a small gradient step and <math>sign()</math> is the sign of the gradient.<br />
# An attacker may be allowed to use a single step of back-propagation ('''single step''') or multiple ('''iterative''') steps. Iterative attackers can generate more powerful adversarial images. Typically, to bound iterative attackers a distance measure is used.<br />
<br />
In this paper the authors focus on the more difficult attacks; white-box iterative targeted and untargeted attacks.<br />
<br />
== Obfuscated Gradients ==<br />
As gradients are used in the generation of white-box adversarial images, many defense strategies have focused on methods that mask gradients. If gradients are masked, they cannot be followed to generate adversarial images. The authors argue against this general approach by showing that it can be easily circumvented. To emphasize their point, they looked at white-box defenses proposed in ICLR 2018. Three types of gradient masking techniques were found:<br />
<br />
# '''Shattered gradients''': Non-differentiable operations are introduced into the model, causing a gradient to be nonexistent or incorrect.<br />
# '''Stochastic gradients''': A stochastic process is added into the model at test time, causing the gradients to become randomized.<br />
# '''Vanishing Gradients ''': Very deep neural networks or those with recurrent connections are used. Because of the vanishing or exploding gradient problem common in these deep networks, effective gradients at the input are small and not very useful.<br />
<br />
== The Attacks ==<br />
To circumvent these gradient masking techniques, the authors propose:<br />
# '''Backward Pass Differentiable Approximation (BPDA)''': For defenses that introduce non-differentiable components, the authors replace it with an approximate function that is differentiable on the backward pass. In a white-box setting, the attacker has full access to any added non-linear transformation and can find its approximation. <br />
# '''Expectation over Transformation [Athalye, 2017]''': For defenses that add some form of test time randomness, the authors propose to use expectation over transformation technique in the backward pass. Rather than moving along the gradient every step, several gradients are sampled and the step is taken in the average direction. This can help with any stochastic misdirection from individual gradients. The technique is similar to using mini-batch gradient descent but applied in the construction of adversarial images.<br />
# '''Re-parameterize the exploration space''': For very deep networks that rely on vanishing or exploding gradients, the authors propose to re-parameterize and search over the range where the gradient does not explode/vanish.<br />
<br />
= Main Results =<br />
[[File:Summary_Table.png|600px|center]]<br />
<br />
The table above summarizes the results of their attacks. Attacks are mounted on the same dataset each defense targeted. If multiple datasets were used, attacks were performed on the largest one. Two different distance metrics (<math>\ell_{\infty}</math> and <math>\ell_{2}</math>) were used in the construction of adversarial images. Distance metrics specify how much an adversarial image can vary from an original image. For <math>\ell_{\infty}</math> adversarial images, each pixel is allowed to vary by a maximum amount. For example, <math>\ell_{\infty}=0.031</math> specifies that each pixel can vary by <math>256*0.031=8</math> from its original value. <math>\ell_{2}</math> distances specify the magnitude of the total distortion allowed over all pixels. For MNIST and CIFAR-10, untargeted adversarial images were constructed using the entire test set, while for Imagenet, 1000 test images were randomly selected and used to generate targeted adversarial images. <br />
<br />
Standard models were used in evaluating the accuracy of defense strategies under the attacks,<br />
# MNIST: 5-layer Convolutional Neural Network (99.3% top-1 accuracy)<br />
# CIFAR-10: Wide-Resnet (95.0% top-1 accuracy)<br />
# Imagenet: InceptionV3 (78.0% top-1 accuracy)<br />
<br />
The last column shows the accuracies each defense method achieved over the adversarial test set. Except for [Madry, 2018], all defense methods could only achieve an accuracy of <10%. Furthermore, the accuracy of most methods was 0%. The results of [Samangoui,2018] (double asterisk), show that their approach was not as successful. The authors claim that is is a result of implementation imperfections but theoretically the defense can be circumvented using their proposed method.<br />
<br />
==== The defense that worked - Adversarial Training [Madary, 2018] ====<br />
<br />
As a defense mechanism, [Madry, 2018] proposes training the neural networks with adversarial images. Although this approach is previously known [Szegedy, 2013] in their formulation, the problem is setup in a more systematic way using a min-max formulation:<br />
\begin{align}<br />
\theta^* = \arg \underset{\theta} \min \mathop{\mathbb{E_x}} \bigg{[} \underset{\delta \in [-\epsilon,\epsilon]}\max L(x+\delta,y;\theta)\bigg{]} <br />
\end{align}<br />
<br />
where <math>\theta</math> is the parameter of the model, <math>\theta^*</math> is the optimal set of parameters and <math>\delta</math> is a small perturbation to the input image <math>x</math> and is bounded by <math>[-\epsilon,\epsilon]</math>. <br />
<br />
Training proceeds in the following way. For each clean input image, a distorted version of the image is found by maximizing the inner maximization problem for a fixed number of iterations. Gradient steps are constrained to fall within the allowed range (projected gradient descent). Next, the classification problem is solved by minimizing the outer minimization problem.<br />
<br />
This approach was shown to provide resilience to all types of adversarial attacks.<br />
<br />
==== How to check for Obfuscated Gradients ====<br />
For future defense proposals, it is recommended to avoid using masked gradients. To assist with this, the authors propose a set of conditions that can help identify if defense is relying on masked gradients:<br />
# If weaker one-step attacks are performing better than iterative attacks.<br />
# Black-box attacks can find stronger adversarial images compared with white-box attacks.<br />
# Unbounded iterative attacks do not reach 100% success.<br />
# If random brute force attempts are better than gradient based methods at finding adversarial images.<br />
<br />
==== Recommendations for future defense methods to encourage reproducibility ====<br />
<br />
= Detailed Results =<br />
<br />
== Non-obfuscated Gradients ==<br />
<br />
==== Cascade Adversarial Training, [Na, 2018] ====<br />
'''Defense''': Since to the method of [Madry, 2018], the authors of [Na, 2018] propose a new training method. The main difference is that instead of using iterative methods to generate adversarial examples at each mini-batch, a separate model is first trained and used to generate adversarial images. These adversarial images are used to augment the train set of another model.<br />
<br />
'''Attack''': The authors found that this technique does not use obfuscated gradients. They were not able to reduce the performance of this method. However, they point out that the claimed accuracy is much lower (%15) compared with [Madry, 2018] under the same perturbation setting.<br />
<br />
== Gradient Shattering ==<br />
<br />
==== Thermometer Coding, [Buckman, 2018] ====<br />
'''Defense''': Inspired by the observation that neural networks learn linear boundaries between classes [Goodfellow, 2014] , [Buckman, 2018] sought to break this linearity by explicitly adding a highly non-linear transform at the input of their model. The non-linear transformation they chose was quantizing inputs to binary vectors. The quantization performed was termed thermometer encoding,<br />
<br />
Given an image, for each pixel value <math>x_{i,j,c}</math>, if an <math>l</math> dimensional thermometer code, the <math>kth</math> bit is given by:<br />
\begin{align}<br />
\tau(x_{i,j,c})_k = \bigg{\{}\begin{array}{ll}<br />
1 \space if \thinspace x_{i,j,c} >k/l \\<br />
0 \space otherwise \\<br />
\end{array}<br />
\end{align}<br />
Here it is assumed <math>x_{i,j,c} \in [0, 1] </math> and <math>i, j, c</math> are the row, column and channel index of the pixel respectively. This encoding is essentially like one-hot encoding, except all the points (not just one) greater than the target value are set to 1. This quantization technique preserves pairwise ordering between pixels.<br />
<br />
On CIFAR-10, the model gave 50% accuracy against <math>\ell_\infty</math> adversarial images with <math>\epsilon=0.031</math> attacks.<br />
<br />
'''Attack''': The authors attack this model using there BPDA approach. Given the non-linear transformation performed in the forward pass, <math>\tau(x)</math>, they develop a differentiable counterpart,<br />
\begin{align}<br />
\hat{\tau}(x_{i,j,c})_k = \min ( \max (x_{i,j,c} - \frac{k}{l}), 1 )<br />
\end{align}<br />
and use it in place of <math>\tau(x)</math> on the backward pass. With their modifications they were able to bring the accuracy of the model down to 0%.<br />
<br />
==== Input Transformation, [Guo, 2018] ====<br />
'''Defense''':[Gou, 2018] investigated the effect of including different input transformation on the robustness to adversarial images. In particular, they found two techniques provided the greatest resistance: total variance minimization and image quilting. Total variance minimization is a technique that removes high frequency noise while preserving legitimate edges (good high frequency components). In image quilting, a large database of image patches from clean images is collected. At test time, input patches, that contain a lot of noise, are replaced with similar but clean patches from the data base.<br />
<br />
Both techniques, removed perturbations from adversarial images which provides some robustness to adversarial attacks. Moreover, both approaches are non-differentiable which makes constructing white-box adversarial images difficult. Moreover, the techniques also include test time randomness as the modifications made are input dependent. The best model achieved 60% accuracy on adversarial images with <math>l_{2}=0.05</math> perturbations.<br />
<br />
'''Attack''': The authors used the BPDA attack where the input transformations were replaced by an identity function. They were able to bring the accuracy of the model down to 0% under the same type of adversarial attacks.<br />
<br />
==== Local Intrinsic Dimensionality, [Ma, 2018] ====<br />
'''Defense''' Local intrinsic dimensionality (LID) is a distance-based metric that measures the similarity between points in a high dimensional space. Given a set of points, let the distance between sample <math>x</math> and its <math>ith</math> neighbor be <math>r_i(x)</math>, then the LID under the choose distance metric is given by,<br />
<br />
\begin{align}<br />
LID(x) = - \bigg{(} \frac{1}{k}\sum^k_{i=1}log \frac{r_i(x)}{r_k(x)} \bigg{)}^{-1}<br />
\end{align}<br />
where k is the number of nearest neighbors considered, <math>r_k(x)</math> is the maximum distance to any of the neighbors in the set k. <br />
<br />
First, <math>L_2</math> distances for all training and adversarial images. Next, the LID scores for each train and adversarial images were calculated. It was found that LID scores for adversarial images were significantly larger than those of clean images. Base on these results, the a separate classifier was created that can be used to detect adversarial inputs. [Ma, 2018] claim that this is not a defense method, but a method to study the properties of adversarial images.<br />
<br />
'''Attack''': Instead of attacking this method, the authors show that this method is not able to detect, and is therefore venerable to, attacks of the [Carlini and Wagner, 2017a] variety.<br />
<br />
== Stochastic Gradients ==<br />
<br />
==== Stochastic Activation Pruning, [Dhillon, 2018] ====<br />
'''Defense''': [Dhillon, 2018] use test time randomness in their model to guard against adversarial attacks. Within a layer, the activities of component nodes are randomly dropped with a probability proportional to its absolute value. The rest of the activation are scaled up to preserve accuracies. This is akin to test time drop-out. This technique was found to drop accuracy slightly on clean images, but improved performance on adversarial images.<br />
<br />
'''Attack''': The authors used the expectation over transformation attack to get useful gradients out of the model. With their attack they were able to reduce the accuracy of this method down to 0% on CIFAR-10.<br />
<br />
==== Mitigation Through Randomization, [Xie, 2018] ====<br />
'''Defense''': [Xie, 2018] Add a randomization layer to their model to help defend against adversarial attacks. For an input image of size [299,299], first the image is randomly re-scaled to <math>r \in [299,331]</math>. Next the image is zero-padded to fix the dimension of the modified input. This modified input is then fed into a regular classifier. The authors claim that is strategy can provide an accuracy of 32.8% against ensemble attack patterns (fixed distortions, but many of them which are picked randomly). Because of the introduced randomness, the authors claim the model builds some robustness to other types of attacks as well.<br />
<br />
'''Attack''': The EOT method was used to build adversarial images to attack this model. With their attack, the authors were able to bring the accuracy of this model down to 0% using <math>L_{\infty}(\epsilon=0.031)</math> perturbations.<br />
<br />
== Vanishing and Exploding Gradients ==<br />
<br />
==== Pixel Defend, [Song, 2018] ====<br />
'''Defense''': [Song, 2018] argues that adversarial images lie in low probability regions of the data manifold. Therefore, one way to handle adversarial attacks is to project them back in the high probability regions before feeding them into a classifier. They chose to do this by using a generative model (pixelCNN) in a denoising capacity. A PixelCNN model directly estimates the conditional probability of generating an image pixel by pixel [Van den Oord, 2016],<br />
<br />
\begin{align}<br />
p(\mathbf{x}= \prod_{i=1}^{n^2} p(x_i|x_0,x_1 ....x_{i-1}))<br />
\end{align}<br />
<br />
The reason for choosing this model is the long iterative process of generation. In the backward pass, following the gradient all the way to the input would not be possible because of the vanishing/exploding gradient<br />
problem of deep networks. The proposed model was able to obtain an accuracy of 46% on CIFAR-10 images with <math>l_{\infty} (\epsilon=0.031) </math> perturbations.<br />
<br />
'''Attack''': The model was attacked using the BPDA technique where back-propagating though the pixelCNN was replaced with an identity function. With this apporach, the authors were able to bring down the accuracy to 9% under the same kind of perturbations.<br />
<br />
==== Defense-GAN, [Samangouei, 2018] ====<br />
<br />
= Conclusion =<br />
In this paper, it was found that gradient masking is a common technique used by many defense proposals that claim to be robust against a very difficult class of adversarial attacks: white-box, iterative attacks. However, the authors found that they can be easily circumvented. Three attack methods are presented that were able to defeat 7 out of the 8 defense proposal accepted in the 2018 ICLR conference for these types of attacks.<br />
<br />
Some future work that can come out of this paper includes avoiding relying on obfuscated gradients for perceived robustness and use the evaluation approach to detect when the attach occurs. Early categorization of attacks using some supervised techniques can also help in critical evaluation of incoming data.<br />
<br />
= Critique =<br />
# The third attack method, reparameterization of the input distortion search space was presented very briefly and at a very high level. Moreover, the one defense proposal they chose to use it against, [Samangouei, 2018] prove to be resilient against the attack. The authors had to resort to one of their other methods to circumvent the defense.<br />
# The BPDA and reparameterization attacks require intrinsic knowledge of the networks. This information is not likely to be available to external users of a network. Most likely, the use-case for these attacks will be in-house to develop more robust networks. This also means that it is still possible to guard against adversarial attack using gradient masking techniques, provided the details of the network are kept secret. <br />
## A notable exception to this case could be applications that are built using open-source (or even published) models that are paired with model-agnostic defense mechanisms. For example, A ResNet-50 using the model-agnostic 'input transformations' technique by [Guo, 2018] may be used in many different image classification tasks, but could still be successfully attacked using BPDA. <br />
# The BPDA algorithm requires replacing a non-linear part of the model with a differentiable approximation. Since different networks are likely to use different transformations, this technique is not plug-and-play. For each network, the attack needs to be manually constructed.<br />
<br />
<br />
= Other Sources =<br />
# Their re-implementation of each of the defenses and implementations of the attacks are available [https://github.com/anishathalye/obfuscated-gradients here].<br />
<br />
= References =<br />
#'''[Madry, 2018]''' Madry, A., Makelov, A., Schmidt, L., Tsipras, D. and Vladu, A., 2017. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083.<br />
#'''[Buckman, 2018]''' Buckman, J., Roy, A., Raffel, C. and Goodfellow, I., 2018. Thermometer encoding: One hot way to resist adversarial examples.<br />
#'''[Guo, 2018]''' Guo, C., Rana, M., Cisse, M. and van der Maaten, L., 2017. Countering adversarial images using input transformations. arXiv preprint arXiv:1711.00117.<br />
#'''[Xie, 2018]''' Xie, C., Wang, J., Zhang, Z., Ren, Z. and Yuille, A., 2017. Mitigating adversarial effects through randomization. arXiv preprint arXiv:1711.01991.<br />
#'''[song, 2018]''' Song, Y., Kim, T., Nowozin, S., Ermon, S. and Kushman, N., 2017. Pixeldefend: Leveraging generative models to understand and defend against adversarial examples. arXiv preprint arXiv:1710.10766.<br />
#'''[Szegedy, 2013]''' Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I. and Fergus, R., 2013. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199.<br />
#'''[Samangouei, 2018]''' Samangouei, P., Kabkab, M. and Chellappa, R., 2018. Defense-GAN: Protecting classifiers against adversarial attacks using generative models. arXiv preprint arXiv:1805.06605.<br />
#'''[van den Oord, 2016]''' van den Oord, A., Kalchbrenner, N., Espeholt, L., Vinyals, O. and Graves, A., 2016. Conditional image generation with pixelcnn decoders. In Advances in Neural Information Processing Systems (pp. 4790-4798).<br />
#'''[Athalye, 2017]''' Athalye, A. and Sutskever, I., 2017. Synthesizing robust adversarial examples. arXiv preprint arXiv:1707.07397.<br />
#'''[Ma, 2018]''' Ma, Xingjun, Bo Li, Yisen Wang, Sarah M. Erfani, Sudanthi Wijewickrema, Michael E. Houle, Grant Schoenebeck, Dawn Song, and James Bailey. "Characterizing adversarial subspaces using local intrinsic dimensionality." arXiv preprint arXiv:1801.02613 (2018).<br />
# '''[Na, 2018]''' Na, T., Ko, J.H. and Mukhopadhyay, S., 2017. Cascade Adversarial Machine Learning Regularized with a Unified Embedding. arXiv preprint arXiv:1708.02582.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Robot_Learning_in_Homes:_Improving_Generalization_and_Reducing_Dataset_Bias&diff=39878Robot Learning in Homes: Improving Generalization and Reducing Dataset Bias2018-11-19T01:52:29Z<p>Wfisher: Misc additions (T,E)</p>
<hr />
<div>==Introduction==<br />
<br />
<br />
Using data-driven approaches in robotics has increased in the last decade. Instead of using hand-designed models, these data-driven approaches work on large-scale datasets and learn appropriate policies that map from high-dimensional observations to actions. Since collecting data using an actual robot in real-time is very expensive, most of the data-driven approaches in robotics use simulators in order to collect simulated data. The concern which arises here is whether these approaches are able to be robust enough to domain shift and to be used for real-world data. It is an undeniable fact that there is a wide reality gap between simulators and the real world.<br />
<br />
On the other hand, the declining costs of hardware to expand collecting data for a variety of tasks push the robotics community to collect real-world physical data. This approach has been quite successful at tasks such as grasping, pushing, poking and imitation learning. However, the major problem is that the performance of these learning models is not good enough and tends to plateau fast. Furthermore, robotic action data did not lead to similar gains in other areas such as computer vision and natural language processing. As the paper claimed, the solution for all of these obstacles is using “real data”. Current robotic datasets lack diversity of environment. Learning-based approaches need to move out of simulators in the labs and go to real environments such as real homes so that they can learn from real datasets. <br />
<br />
Like every other process, collecting real data and working with it has several challenges. First, there is a need for cheap and compact robots to collect data in homes but current industrial robots (i.e. Sawyer and Baxter) are too expensive. Secondly, cheap robots are not accurate enough to collect reliable data. Also, collecting data in homes cannot have a supervisor at all times. These challenges in addition to some other external factors can have a result in having noisy data. In this paper, a first systematic effort has been presented for collecting a dataset inside people's homes which has the following parts: <br />
<br />
-A cheap robot which is appropriate for use in homes<br />
<br />
-Collecting training data in 6 different homes and testing data in 3 homes<br />
<br />
-An approach for modeling the noise in the labeled data<br />
<br />
[[File:aa1.PNG|600px|thumb|center|]]<br />
<br />
==Overview==<br />
<br />
<br />
This paper emphasizes the importance of diversifying the data for robotic learning in order to have a greater generalization, by focusing on the task of grasping. A diverse dataset also allows for removing biases in the data. By considering these facts, the paper argues that even for simple tasks like grasping, datasets which are collected in labs suffer from strong biases such as simple backgrounds and same environment dynamics. Hence, the learning approaches cannot generalize the models and work well on real datasets.<br />
<br />
As a future possibility, there would be a need for having a low-cost robot to collect large-scale data inside a huge number of homes. For this reason, they introduced a customized mobile manipulator. They used a Dobot Magician which is a robotic arm mounted on a Kobuki which is a low-cost mobile base. The resulting robot arm has five degrees of freedom (DOF) (x, y, z, roll, pitch). The gripper is a two-fingered electric gripper with a 0.3kg payload. They also add an Intel R200 RGBD camera to their robot which is at a height of 1m above the ground. An Intel Core i5 processor is also used as an onboard laptop to perform all the processing. The whole system can run for 1.5 hours with a single charge.<br />
<br />
As there is always a trade-off, when we gain a low-cost robot, we are actually losing accuracy for controlling it. So, the low-cost robot which is built from cheaper components than the expensive setups such as Baxter and Sawyer suffers from higher calibration errors and execution errors. This means that the dataset collected with this approach is diverse and huge but it has noisy labels. To illustrate, consider when the robot wants to grasp at location <math> {(x, y)}</math>. Since there is a noise in the execution, the robot may perform this action in the location <math> {(x + \delta_{x}, y+ \delta_{y})}</math> which would assign the success or failure label of this action to a wrong place. Therefore, to solve the problem, they used an approach to learn from noisy data. They modeled noise as a latent variable and used two networks, one for predicting the noise and one for predicting the action to execute.<br />
<br />
==Learning on low-cost robot data==<br />
<br />
This paper uses patch grasping framework in its proposed architecture. Also, as mentioned before, there is a high tendency for noisy labels in the datasets which are collected by inaccurate and cheap robots. The cause of the noise in the labels could be due to the hardware execution error, inaccurate kinematics, camera calibration, proprioception, wear, and tear, etc. Here are more explanations about different parts of the architecture: <br />
<br />
<br />
===Grasping Formulation===<br />
<br />
Planar grasping is the object of interest in this architecture. It means that all the objects are grasped at the same height and vertical to the ground (ie: a fixed end-effector pitch). The final goal is to find <math>{(x, y, \theta)}</math> given an observation <math> {I}</math> of the object, where <math> {x}</math> and <math> {y}</math> are the translational degrees of freedom and <math> {\theta}</math> is the rotational degrees of freedom (roll). For the purpose of comparison, they used a model which does not predict the <math>{(x, y, \theta)}</math> directly from the image <math> {I}</math>, but samples several smaller patches <math> {I_{P}}</math> at different locations <math>{(x, y)}</math>. Thus, the angle of grasp <math> {\theta}</math> is predicted from these patches. Also, in order to have multimodal predictions, discrete steps of the angle <math> {\theta}</math>, <math> {\theta_{D}}</math> is used. <br />
<br />
Hence, each datapoint consists of an image <math> {I}</math>, the executed grasp <math>{(x, y, \theta)}</math> and the grasp success/failure label g. Then, the image <math> {I}</math> and the angle <math> {\theta}</math> are converted to image patch <math> {I_{P}}</math> and angle <math> {\theta_{D}}</math>. Then, to minimize the classification error, a binary cross entropy loss is used which minimizes the error between the predicted and ground truth label <math> g </math>. A convolutional neural network with weight initialization from pre-training on Imagenet is used for this formulation.<br />
<br />
===Modeling noise as latent variable===<br />
<br />
In order to tackle the problem of inaccurate position control, they found a structure in the noise which is dependent on the robot and the design. They modeled this structure as a latent variable which is shown is figure 2: <br />
<br />
<br />
[[File:aa2.PNG|600px|thumb|center|]]<br />
<br />
<br />
<br />
The grasp success probability for image patch <math> {I_{P}}</math> at angle <math> {\theta_{D}}</math> is represented as <math> {P(g|I_{P},\theta_{D};R )}</math> where <math> {R}</math> represents environment variables that can add noise to the system.<br />
<br />
The conditional probability of grasping for this model is computed by:<br />
<br />
<br />
\[ { P(g|I_{P},\theta_{D}, R ) = ∑_{( \hat{I}_{P} ϵ P)} P(g│z=\hat{I}_{P},\theta_{D},R ). P(z=\hat{I}_{P} |(\theta_{D},I_{P} ,R ) } \]<br />
<br />
<br />
<br />
<br />
Here, <math> {z}</math> models the latent variable of the actual patch executed, and <math> {\hat{I}_{P}}</math> belongs to a set of possible neighboring patches <math> {P}</math>. <math> {P(z=\hat{I}_{P} |(\theta_{D},I_{P} ,R )}</math> shows the noise which can be caused by <math> {R}</math> variables and is implemented as the Noise Modelling Network (NMN). <math> {P(g│z=\hat{I}_{P},\theta_{D},R )}</math> shows the grasp prediction probability given the true patch and is implemented as the Grasp Prediction Network (GPN). The overall Robust-Grasp model is computed by marginalizing GPN and NMN.<br />
<br />
===Learning the latent noise model===<br />
<br />
<br />
They assume that <math> {z}</math> is conditionally independent of the local patch-specific variables <math> {(I_{P}, \theta_{D})}</math>. To estimate the latent variable <math> {z}</math>, they used direct optimization to learn both NMN and GPN with noisy labels. The entire image of the scene and the environment information are the inputs of the NMN, as well as robot ID and raw-pixel grasp location.. The output of the NMN is the probability distribution of the actual patches where the grasps are executed. Finally, a binary cross entropy loss is applied to the marginalized output of these two networks and the true grasp label g.<br />
<br />
===Training details===<br />
<br />
<br />
They implemented their model in PyTorch using a pretrained ResNet-18 model. They concatenated 512 dimensional ResNet feature with a 1-hot vector of robot ID and the raw pixel location of the grasp for their NMN. Also, the inputs of the GPN are the original noisy patch plus 8 other equidistant patches from the original one.<br />
Their training process starts with training only GPN over 5 epochs of the data. Then, the NMN and the marginalization operator are added to the model. So, they train NMN and GPN simultaneously for the other 25 epochs.<br />
<br />
==Results==<br />
<br />
<br />
In the results part of the paper, they show that collecting dataset in homes is essential for generalizing learning from unseen environments. They also show that modeling the noise in their Low-Cost Arm (LCA) can improve grasping performance.<br />
They collected data in parallel using multiple robots in 6 different homes, as shown in Figure 3. They used an object detector (tiny-YOLO) as the input data were unstructured due to LCA limited memory and computational capabilities. With an object location detected, class information was discarded, and a grasp was attempted. The grasp location in 3D was computed using point cloud data. They scattered different objects in homes within 2m area to prevent collision of the robot with obstacles and let the robot move randomly and grasp objects. Finally, they collected a dataset with 28K grasp results.<br />
<br />
[[File:aa3.PNG|600px|thumb|center|]]<br />
<br />
<br />
To evaluate their approach in a more quantitative way, they used three test settings:<br />
<br />
- The first one is a binary classification or held-out data. The test set is collected by performing random grasps on objects. They measure the performance of binary classification by predicting the success or failure of grasping, given a location and the angle. Using binary classification allows for testing a lot of models without running them on real robots. They collected two held-out datasets using LCA in lab and homes and the dataset for Baxter robot.<br />
<br />
- The second one is Real Low-Cost Arm(Real-LCA). Here, they evaluate their model by running it in three unseen homes. They put 20 new objects in these three homes in different orientations. Since the objects and the environments are completely new, this tests could measure the generalization of the model.<br />
<br />
- The third one is Real Sawyer(Real-Sawyer). They evaluate the performance of their model by running the model on the Sawyer robot which is more accurate than the LCA. They tested their model in the lab environment to show that training models with the datasets collected from homes can improve the performance of models even in lab environments.<br />
<br />
They used baselines for both their data which is collected in homes and their model which is Robust-Grasp. They used two datasets for the baseline. The dataset collected by (Lab-Baxter) and the dataset collected by their LCA in the lab(Lab-LCA).<br />
They compared their model with the noise independent patch grasping model (Patch-Grasp). They also compared their data and model with DexNet-3.0 (DexNet) for a strong real-world grasping baseline.<br />
<br />
<br />
<br />
===Experiment 1: Performance on held-out data===<br />
<br />
<br />
Table 1 shows that the models trained on lab data cannot generalize to the Home-LCA environment. However, the model trained on Home-LCA hasa good performance on both lab data and home environment.<br />
<br />
[[File:aa4.PNG|600px|thumb|center|]]<br />
<br />
<br />
<br />
===Experiment 2: Performance on Real LCA Robot===<br />
<br />
<br />
In table 2, the performance of the Home-LCA is compared against a pre-trained DexNet and the model trained on the Lab-Baxter. Training on the Home-LCA dataset performs 43.7% better than training on the Lab-Baxter dataset and 33% better than DexNet. The low performance of DexNet can be described by the possible noise in the depth images that are caused by the natural light. DexNet, which requires high quality depth sensing, cannot perform well. By using cheap commodity RGBD cameras in LCA, the noise in the depth images is not a matter of concern, as the model has no expectation of high quality.<br />
<br />
[[File:aa5.PNG|600px|thumb|center|]]<br />
<br />
===Performance on Real Sawyer===<br />
<br />
<br />
To compare the performance of the Robust-Grasp model against the Patch-Grasp model, they used Lab-Baxter which is an accurate robot. The Sawyer robot is used for testing to ensure that the testing robot is different from both training robots. As shown in Table 3, the Robust-Grasp model trained on Home-LCA outperforms the Patch-Grasp model and achieves 77.5% accuracy. Furthermore, the visualizations of predicted noise corrections in Figure 4 shows that the corrections depend on both the pixel locations of the noisy grasp and the robot.<br />
<br />
<br />
[[File:aa6.PNG|600px|thumb|center|]]<br />
<br />
[[File:aa7.PNG|600px|thumb|center|]]<br />
<br />
==Related work==<br />
<br />
<br />
Over the last few years, the interest of scaling up robot learning with large scale datasets has been increased. Hence, many papers were published in this area. A hand annotated grasping dataset, a self-supervised grasping dataset, and grasping using reinforcement learning are some examples of using large scale datasets for grasping. There were also many papers that worked on other robotic tasks like material recognition or pushing objects. However, none of these papers worked on real data in real environments like homes. They just used high-cost hardware and lab data.<br />
<br />
<br />
Furthermore, since grasping is one of the basic problems of robotic, there were some efforts to improve grasping. Classic approaches focused on physics-based issues of grasping and required 3D models of the objects. However, recent works focused on data-driven approaches which learn from visual observations to grasp objects. The point here is that they usually require high quality depth as input which seems to be a barrier for practical use of robots in real environments.<br />
<br />
<br />
Most labs use industrial robots or standard collaborative hardware for their experiments. Therefore, there is few research that used low cost robots. One of the examples is learning using a cheap inaccurate robot for stack multiple blocks, although it is not clear whether learning approaches are used in it alongside mapping and planning.<br />
<br />
<br />
Learning from noisy inputs is another challenge specifically in computer vision. A controversial question which is often raised in this area is whether learning from noise can improve the performance. Some works show it could have bad effects on the performance; however, some other works find it valuable when the noise is dependent of the environment. In this paper, they used a model that can exploit the noise and learn a better grasping model.<br />
<br />
==Conclusion==<br />
<br />
All in all, the paper presents an approach for collecting large-scale robot data in real home environments. They implemented their approach by using a mobile manipulator which is a lot cheaper than the existing industrial robots. They collected a dataset of 28K grasps in six different homes. In order to solve the problem of noisy labels which were caused by their inaccurate robots, they presented a framework to factor out the noise in the data. They tested their model by physically grasping 20 new objects in three new homes and in the lab. The model trained with home dataset showed 43.7% improvement over the models trained with lab data. Their results also showed that their model can improve the grasping performance even in lab environments. They also demonstrated that their architecture for modeling the noise improved the performance by about 10%.<br />
<br />
==Critiques==<br />
<br />
This paper is not a significant algorithmic contribution. They are just combining a large number of data engineering techniques for the robot learning problem. The authors claim that they have obtained 43.7% more accuracy than baseline models, but it does not seem to be a fair comparison as the data collection happened in simulated settings in the lab for other methods, whereas the authors use the home dataset. The authors must have also discussed safety issues when training robots in real environments as against simulated environments like labs. The authors are encouraging other researchers to look outside the labs, but are not discussing the critical safety issues in this approach.<br />
<br />
<br />
The paper argues that the dataset collected by the LCA is noisy, since the robot is cheap and inaccurate. It further asserts that in order to handle the noise in the dataset, they can model the noise as a latent variable and their model can improve the performance of grasping. Although learning from noisy data and achieving a good performance is valuable, it is better that they test their noise modeling network for other robots as well. Since their noise modelling network takes robot information as an input, it would be a good idea to generalize it by testing it using different inaccurate robots to ensure that it would perform well.<br />
<br />
==References==<br />
<br />
#Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. "Domain randomization for transferring deep neural networks from simulation to the real world." 2017. URL https://arxiv.org/abs/1703.06907.<br />
#Xue Bin Peng, Marcin Andrychowicz, Wojciech Zaremba, and Pieter Abbeel. "Sim-to-real transfer of robotic control with dynamics randomization." arXiv preprint arXiv:1710.06537,2017.<br />
#Lerrel Pinto, Marcin Andrychowicz, Peter Welinder, Wojciech Zaremba, and Pieter Abbeel. "Asymmetric actor critic for image-based robot learning." Robotics Science and Systems, 2018.<br />
#Lerrel Pinto and Abhinav Gupta. "Supersizing self-supervision: Learning to grasp from 50k tries and 700 robot hours." CoRR, abs/1509.06825, 2015. URL http://arxiv.org/abs/1509. 06825.<br />
#Adithyavairavan Murali, Lerrel Pinto, Dhiraj Gandhi, and Abhinav Gupta. "CASSL: Curriculum accelerated self-supervised learning." International Conference on Robotics and Automation, 2018.<br />
# Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. "End-to-end training of deep visuomotor policies." The Journal of Machine Learning Research, 17(1):1334–1373, 2016.<br />
#Sergey Levine, Peter Pastor, Alex Krizhevsky, and Deirdre Quillen. "Learning hand-eye coordination for robotic grasping with deep learning and large scale data collection." CoRR, abs/1603.02199, 2016. URL http://arxiv.org/abs/1603.02199.<br />
#Pulkit Agarwal, Ashwin Nair, Pieter Abbeel, Jitendra Malik, and Sergey Levine. "Learning to poke by poking: Experiential learning of intuitive physics." 2016. URL http://arxiv.org/ abs/1606.07419<br />
#Chelsea Finn, Ian Goodfellow, and Sergey Levine. "Unsupervised learning for physical interaction through video prediction." In Advances in neural information processing systems, 2016.<br />
#Ashvin Nair, Dian Chen, Pulkit Agrawal, Phillip Isola, Pieter Abbeel, Jitendra Malik, and Sergey Levine. "Combining self-supervised learning and imitation for vision-based rope manipulation." International Conference on Robotics and Automation, 2017.<br />
#Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. "Revisiting unreasonable effectiveness of data in deep learning era." ICCV, 2017.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Robot_Learning_in_Homes:_Improving_Generalization_and_Reducing_Dataset_Bias&diff=39875Robot Learning in Homes: Improving Generalization and Reducing Dataset Bias2018-11-19T01:41:27Z<p>Wfisher: Add misc info (T)</p>
<hr />
<div>==Introduction==<br />
<br />
<br />
Using data-driven approaches in robotics has increased in the last decade. Instead of using hand-designed models, these data-driven approaches work on large-scale datasets and learn appropriate policies that map from high-dimensional observations to actions. Since collecting data using an actual robot in real-time is very expensive, most of the data-driven approaches in robotics use simulators in order to collect simulated data. The concern which arises here is whether these approaches are able to be robust enough to domain shift and to be used for real-world data. It is an undeniable fact that there is a wide reality gap between simulators and the real world.<br />
<br />
On the other hand, the declining costs of hardware to expand collecting data for a variety of tasks push the robotics community to collect real-world physical data. This approach has been quite successful at tasks such as grasping, pushing, poking and imitation learning. However, the major problem is that the performance of these learning models is not good enough and tends to plateau fast. Furthermore, robotic action data did not lead to similar gains in other areas such as computer vision and natural language processing. As the paper claimed, the solution for all of these obstacles is using “real data”. Current robotic datasets lack diversity of environment. Learning-based approaches need to move out of simulators in the labs and go to real environments such as real homes so that they can learn from real datasets. <br />
<br />
Like every other process, collecting real data and working with it has several challenges. First, there is a need for cheap and compact robots to collect data in homes but current industrial robots (i.e. Sawyer and Baxter) are too expensive. Secondly, cheap robots are not accurate enough to collect reliable data. Also, collecting data in homes cannot have a supervisor at all times. These challenges in addition to some other external factors can have a result in having noisy data. In this paper, a first systematic effort has been presented for collecting a dataset inside people's homes which has the following parts: <br />
<br />
-A cheap robot which is appropriate for use in homes<br />
<br />
-Collecting training data in 6 different homes and testing data in 3 homes<br />
<br />
-An approach for modeling the noise in the labeled data<br />
<br />
[[File:aa1.PNG|600px|thumb|center|]]<br />
<br />
==Overview==<br />
<br />
<br />
This paper emphasizes the importance of diversifying the data for robotic learning in order to have a greater generalization, by focusing on the task of grasping. A diverse dataset also allows for removing biases in the data. By considering these facts, the paper argues that even for simple tasks like grasping, datasets which are collected in labs suffer from strong biases such as simple backgrounds and same environment dynamics. Hence, the learning approaches cannot generalize the models and work well on real datasets.<br />
<br />
As a future possibility, there would be a need for having a low-cost robot to collect large-scale data inside a huge number of homes. For this reason, they introduced a customized mobile manipulator. They used a Dobot Magician which is a robotic arm mounted on a Kobuki which is a low-cost mobile base. The resulting robot arm has five degrees of freedom (DOF) (x, y, z, roll, pitch). The gripper is a two-fingered electric gripper with a 0.3kg payload. They also add an Intel R200 RGBD camera to their robot which is at a height of 1m above the ground. An Intel Core i5 processor is also used as an onboard laptop to perform all the processing. The whole system can run for 1.5 hours with a single charge.<br />
<br />
As there is always a trade-off, when we gain a low-cost robot, we are actually losing accuracy for controlling it. So, the low-cost robot which is built from cheaper components than the expensive setups such as Baxter and Sawyer suffers from higher calibration errors and execution errors. This means that the dataset collected with this approach is diverse and huge but it has noisy labels. To illustrate, consider when the robot wants to grasp at location <math> {(x, y)}</math>. Since there is a noise in the execution, the robot may perform this action in the location <math> {(x + \delta_{x}, y+ \delta_{y})}</math> which would assign the success or failure label of this action to a wrong place. Therefore, to solve the problem, they used an approach to learn from noisy data. They modeled noise as a latent variable and used two networks, one for predicting the noise and one for predicting the action to execute.<br />
<br />
==Learning on low-cost robot data==<br />
<br />
This paper uses patch grasping framework in its proposed architecture. Also, as mentioned before, there is a high tendency for noisy labels in the datasets which are collected by inaccurate and cheap robots. The cause of the noise in the labels could be due to the hardware execution error, inaccurate kinematics, camera calibration, proprioception, wear, and tear, etc. Here are more explanations about different parts of the architecture: <br />
<br />
<br />
===Grasping Formulation===<br />
<br />
Planar grasping is the object of interest in this architecture. It means that all the objects are grasped at the same height and vertical to the ground (ie: a fixed end-effector pitch). The final goal is to find <math>{(x, y, \theta)}</math> given an observation <math> {I}</math> of the object, where <math> {x}</math> and <math> {y}</math> are the translational degrees of freedom and <math> {\theta}</math> is the rotational degrees of freedom (roll). For the purpose of comparison, they used a model which does not predict the <math>{(x, y, \theta)}</math> directly from the image <math> {I}</math>, but samples several smaller patches <math> {I_{P}}</math> at different locations <math>{(x, y)}</math>. Thus, the angle of grasp <math> {\theta}</math> is predicted from these patches. Also, in order to have multimodal predictions, discrete steps of the angle <math> {\theta}</math>, <math> {\theta_{D}}</math> is used. <br />
<br />
Hence, each datapoint consists of an image <math> {I}</math>, the executed grasp <math>{(x, y, \theta)}</math> and the grasp success/failure label g. Then, the image <math> {I}</math> and the angle <math> {\theta}</math> are converted to image patch <math> {I_{P}}</math> and angle <math> {\theta_{D}}</math>. Then, to minimize the classification error, a binary cross entropy loss is used which minimizes the error between the predicted and ground truth label <math> g </math>. A convolutional neural network with weight initialization from pre-training on Imagenet is used for this formulation.<br />
<br />
===Modeling noise as latent variable===<br />
<br />
In order to tackle the problem of inaccurate position control, they found a structure in the noise which is dependent on the robot and the design. They modeled this structure as a latent variable which is shown is figure 2: <br />
<br />
<br />
[[File:aa2.PNG|600px|thumb|center|]]<br />
<br />
<br />
<br />
The grasp success probability for image patch <math> {I_{P}}</math> at angle <math> {\theta_{D}}</math> is represented as <math> {P(g|I_{P},\theta_{D};R )}</math> where <math> {R}</math> represents environment variables that can add noise to the system.<br />
<br />
The conditional probability of grasping for this model is computed by:<br />
<br />
<br />
\[ { P(g|I_{P},\theta_{D}, R ) = ∑_{( \hat{I}_{P} ϵ P)} P(g│z=\hat{I}_{P},\theta_{D},R ). P(z=\hat{I}_{P} |(\theta_{D},I_{P} ,R ) } \]<br />
<br />
<br />
<br />
<br />
Here, <math> {z}</math> models the latent variable of the actual patch executed, and <math> {\hat{I}_{P}}</math> belongs to a set of possible neighboring patches <math> {P}</math>. <math> {P(z=\hat{I}_{P} |(\theta_{D},I_{P} ,R )}</math> shows the noise which can be caused by <math> {R}</math> variables and is implemented as the Noise Modelling Network (NMN). <math> {P(g│z=\hat{I}_{P},\theta_{D},R )}</math> shows the grasp prediction probability given the true patch and is implemented as the Grasp Prediction Network (GPN). The overall Robust-Grasp model is computed by marginalizing GPN and NMN.<br />
<br />
===Learning the latent noise model===<br />
<br />
<br />
They assume that <math> {z}</math> is conditionally independent of the local patch-specific variables <math> {(I_{P}, \theta_{D})}</math>. To estimate the latent variable <math> {z}</math>, they used direct optimization to learn both NMN and GPN with noisy labels. The entire image of the scene and the environment information are the inputs of the NMN, as well as robot ID and raw-pixel grasp location.. The output of the NMN is the probability distribution of the actual patches where the grasps are executed. Finally, a binary cross entropy loss is applied to the marginalized output of these two networks and the true grasp label g.<br />
<br />
===Training details===<br />
<br />
<br />
They implemented their model in PyTorch using a pretrained ResNet-18 model. They concatenated 512 dimensional ResNet feature with a 1-hot vector of robot ID and the raw pixel location of the grasp for their NMN. Also, the inputs of the GPN are the original noisy patch plus 8 other equidistant patches from the original one.<br />
Their training process starts with training only GPN over 5 epochs of the data. Then, the NMN and the marginalization operator are added to the model. So, they train NMN and GPN simultaneously for the other 25 epochs.<br />
<br />
==Results==<br />
<br />
<br />
In the results part of the paper, they show that collecting dataset in homes is essential for generalizing learning from unseen environments. They also show that modeling the noise in their Low-Cost Arm (LCA) can improve grasping performance.<br />
They collected data in parallel using multiple robots in 6 different homes, as shown in Figure 3. They used an object detector (tiny-YOLO) as the input data were unstructured due to LCA limited memory and computational capabilities. They scattered different objects in homes within 2m area to prevent collision of the robot with obstacles and let the robot move randomly and grasp objects. Finally, they collected a dataset with 28K grasp results.<br />
<br />
[[File:aa3.PNG|600px|thumb|center|]]<br />
<br />
<br />
To evaluate their approach in a more quantitative way, they used three test settings:<br />
<br />
- The first one is a binary classification or held-out data. The test set is collected by performing random grasps on objects. They measure the performance of binary classification by predicting the success or failure of grasping, given a location and the angle. Using binary classification allows for testing a lot of models without running them on real robots. They collected two held-out datasets using LCA in lab and homes and the dataset for Baxter robot.<br />
<br />
- The second one is Real Low-Cost Arm(Real-LCA). Here, they evaluate their model by running it in three unseen homes. They put 20 new objects in these three homes in different orientations. Since the objects and the environments are completely new, this tests could measure the generalization of the model.<br />
<br />
- The third one is Real Sawyer(Real-Sawyer). They evaluate the performance of their model by running the model on the Sawyer robot which is more accurate than the LCA. They tested their model in the lab environment to show that training models with the datasets collected from homes can improve the performance of models even in lab environments.<br />
<br />
They used baselines for both their data which is collected in homes and their model which is Robust-Grasp. They used two datasets for the baseline. The dataset collected by (Lab-Baxter) and the dataset collected by their LCA in the lab(Lab-LCA).<br />
They compared their model with the noise independent patch grasping model (Patch-Grasp). They also compared their data and model with DexNet-3.0 (DexNet) for a strong real-world grasping baseline.<br />
<br />
<br />
<br />
===Experiment 1: Performance on held-out data===<br />
<br />
<br />
Table 1 shows that the models trained on lab data cannot generalize to the Home-LCA environment. However, the model trained on Home-LCA have a good performance on both lab data and home environment.<br />
<br />
[[File:aa4.PNG|600px|thumb|center|]]<br />
<br />
<br />
<br />
===Experiment 2: Performance on Real LCA Robot===<br />
<br />
<br />
In table 2, the performance of the Home-LCA is compared against a pre-trained DexNet and the model trained on the Lab-Baxter. Training on the Home-LCA dataset performs 43.7% better than training on the Lab-Baxter dataset and 33% better than DexNet. The low performance of DexNet can be described by the possible noise in the depth images that are caused by the natural light. So that’s why DexNet which requires high quality depth sensing cannot perform well. By using cheap commodity RGBD cameras in LCA, the noise in the depth images is not a matter of concern.<br />
<br />
[[File:aa5.PNG|600px|thumb|center|]]<br />
<br />
===Performance on Real Sawyer===<br />
<br />
<br />
To compare the performance of the Robust-Grasp model against the Patch-Grasp model, they used Lab-Baxter which is an accurate robot. Sawyer robot is used for testing to ensure that the testing robot is different from both training robots. As shown in Table 3, the Robust-Grasp model trained on Home-LCA outperforms the Patch-Grasp model and achieves 77.5% accuracy. Furthermore, the visualizations of predicted noise corrections in Figure 4 shows that the corrections depend on both the pixel locations of the noisy grasp and the robot.<br />
<br />
<br />
[[File:aa6.PNG|600px|thumb|center|]]<br />
<br />
[[File:aa7.PNG|600px|thumb|center|]]<br />
<br />
<br />
<br />
==Related work==<br />
<br />
<br />
Over the last few years, the interest of scaling up robot learning with large scale datasets has been increased. Hence, many papers were published in this area. A hand annotated grasping dataset, a self-supervised grasping dataset, and grasping using reinforcement learning are some examples of using large scale datasets for grasping. There were also many papers that worked on other robotic tasks like material recognition or pushing objects. However, none of these papers worked on real data in real environments like homes. They just used high-cost hardware and lab data.<br />
<br />
<br />
Furthermore, since grasping is one of the basic problems of robotic, there were some efforts to improve grasping. Classic approaches focused on physics-based issues of grasping and required 3D models of the objects. However, recent works focused on data-driven approaches which learn from visual observations to grasp objects. The point here is that they usually require high quality depth as input which seems to be a barrier for practical use of robots in real environments.<br />
<br />
<br />
Most labs use industrial robots or standard collaborative hardware for their experiments. Therefore, there is few research that used low cost robots. One of the examples is learning using a cheap inaccurate robot for stack multiple blocks, although it is not clear whether learning approaches are used in it alongside mapping and planning.<br />
<br />
<br />
Learning from noisy inputs is another challenge specifically in computer vision. A controversial question which is often raised in this area is whether learning from noise can improve the performance. Some works show it could have bad effects on the performance; however, some other works find it valuable when the noise is dependent of the environment. In this paper, they used a model that can exploit the noise and learn a better grasping model.<br />
<br />
==Conclusion==<br />
<br />
All in all, the paper presents an approach for collecting large-scale robot data in real home environments. They implemented their approach by using a mobile manipulator which is a lot cheaper than the existing industrial robots. They collected a dataset of 28K grasps in six different homes. In order to solve the problem of noisy labels which were caused by their inaccurate robots, they presented a framework to factor out the noise in the data. They tested their model by physically grasping 20 new objects in three new homes and in the lab. The model trained with home dataset showed 43.7% improvement over the models trained with lab data. Their results also showed that their model can improve the grasping performance even in lab environments. They also demonstrated that their architecture for modeling the noise improved the performance by about 10%.<br />
<br />
==Critiques==<br />
<br />
This paper is not a significant algorithmic contribution. They are just combining a large number of data engineering techniques for the robot learning problem. The authors claim that they have obtained 43.7% more accuracy than baseline models, but it does not seem to be a fair comparison as the data collection happened in simulated settings in the lab for other methods, whereas the authors use the home dataset. The authors must have also discussed safety issues when training robots in real environments as against simulated environments like labs. The authors are encouraging other researchers to look outside the labs, but are not discussing the critical safety issues in this approach.<br />
<br />
<br />
The paper argues that the dataset collected by the LCA is noisy, since the robot is cheap and inaccurate. It further asserts that in order to handle the noise in the dataset, they can model the noise as a latent variable and their model can improve the performance of grasping. Although learning from noisy data and achieving a good performance is valuable, it is better that they test their noise modeling network for other robots as well. Since their noise modelling network takes robot information as an input, it would be a good idea to generalize it by testing it using different inaccurate robots to ensure that it would perform well.<br />
<br />
==References==<br />
<br />
#Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. "Domain randomization for transferring deep neural networks from simulation to the real world." 2017. URL https://arxiv.org/abs/1703.06907.<br />
#Xue Bin Peng, Marcin Andrychowicz, Wojciech Zaremba, and Pieter Abbeel. "Sim-to-real transfer of robotic control with dynamics randomization." arXiv preprint arXiv:1710.06537,2017.<br />
#Lerrel Pinto, Marcin Andrychowicz, Peter Welinder, Wojciech Zaremba, and Pieter Abbeel. "Asymmetric actor critic for image-based robot learning." Robotics Science and Systems, 2018.<br />
#Lerrel Pinto and Abhinav Gupta. "Supersizing self-supervision: Learning to grasp from 50k tries and 700 robot hours." CoRR, abs/1509.06825, 2015. URL http://arxiv.org/abs/1509. 06825.<br />
#Adithyavairavan Murali, Lerrel Pinto, Dhiraj Gandhi, and Abhinav Gupta. "CASSL: Curriculum accelerated self-supervised learning." International Conference on Robotics and Automation, 2018.<br />
# Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. "End-to-end training of deep visuomotor policies." The Journal of Machine Learning Research, 17(1):1334–1373, 2016.<br />
#Sergey Levine, Peter Pastor, Alex Krizhevsky, and Deirdre Quillen. "Learning hand-eye coordination for robotic grasping with deep learning and large scale data collection." CoRR, abs/1603.02199, 2016. URL http://arxiv.org/abs/1603.02199.<br />
#Pulkit Agarwal, Ashwin Nair, Pieter Abbeel, Jitendra Malik, and Sergey Levine. "Learning to poke by poking: Experiential learning of intuitive physics." 2016. URL http://arxiv.org/ abs/1606.07419<br />
#Chelsea Finn, Ian Goodfellow, and Sergey Levine. "Unsupervised learning for physical interaction through video prediction." In Advances in neural information processing systems, 2016.<br />
#Ashvin Nair, Dian Chen, Pulkit Agrawal, Phillip Isola, Pieter Abbeel, Jitendra Malik, and Sergey Levine. "Combining self-supervised learning and imitation for vision-based rope manipulation." International Conference on Robotics and Automation, 2017.<br />
#Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. "Revisiting unreasonable effectiveness of data in deep learning era." ICCV, 2017.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Robot_Learning_in_Homes:_Improving_Generalization_and_Reducing_Dataset_Bias&diff=39874Robot Learning in Homes: Improving Generalization and Reducing Dataset Bias2018-11-19T01:34:17Z<p>Wfisher: Add info on robot (T)</p>
<hr />
<div>==Introduction==<br />
<br />
<br />
Using data-driven approaches in robotics has increased in the last decade. Instead of using hand-designed models, these data-driven approaches work on large-scale datasets and learn appropriate policies that map from high-dimensional observations to actions. Since collecting data using an actual robot in real-time is very expensive, most of the data-driven approaches in robotics use simulators in order to collect simulated data. The concern which arises here is whether these approaches are able to be robust enough to domain shift and to be used for real-world data. It is an undeniable fact that there is a wide reality gap between simulators and the real world.<br />
<br />
On the other hand, the declining costs of hardware to expand collecting data for a variety of tasks push the robotics community to collect real-world physical data. This approach has been quite successful at tasks such as grasping, pushing, poking and imitation learning. However, the major problem is that the performance of these learning models is not good enough and tends to plateau fast. Furthermore, robotic action data did not lead to similar gains in other areas such as computer vision and natural language processing. As the paper claimed, the solution for all of these obstacles is using “real data”. Current robotic datasets lack diversity of environment. Learning-based approaches need to move out of simulators in the labs and go to real environments such as real homes so that they can learn from real datasets. <br />
<br />
Like every other process, collecting real data and working with it has several challenges. First, there is a need for cheap and compact robots to collect data in homes but current industrial robots (i.e. Sawyer and Baxter) are too expensive. Secondly, cheap robots are not accurate enough to collect reliable data. Also, collecting data in homes cannot have a supervisor at all times. These challenges in addition to some other external factors can have a result in having noisy data. In this paper, a first systematic effort has been presented for collecting a dataset inside people's homes which has the following parts: <br />
<br />
-A cheap robot which is appropriate for use in homes<br />
<br />
-Collecting training data in 6 different homes and testing data in 3 homes<br />
<br />
-An approach for modeling the noise in the labeled data<br />
<br />
[[File:aa1.PNG|600px|thumb|center|]]<br />
<br />
==Overview==<br />
<br />
<br />
This paper emphasizes the importance of diversifying the data for robotic learning in order to have a greater generalization, by focusing on the task of grasping. A diverse dataset also allows for removing biases in the data. By considering these facts, the paper argues that even for simple tasks like grasping, datasets which are collected in labs suffer from strong biases such as simple backgrounds and same environment dynamics. Hence, the learning approaches cannot generalize the models and work well on real datasets.<br />
<br />
As a future possibility, there would be a need for having a low-cost robot to collect large-scale data inside a huge number of homes. For this reason, they introduced a customized mobile manipulator. They used a Dobot Magician which is a robotic arm mounted on a Kobuki which is a low-cost mobile base. The resulting robot arm has five degrees of freedom (DOF) (x, y, z, roll, pitch). The gripper is a two-fingered electric gripper with a 0.3kg payload. They also add an Intel R200 RGBD camera to their robot which is at a height of 1m above the ground. An Intel Core i5 processor is also used as an onboard laptop to perform all the processing. The whole system can run for 1.5 hours with a single charge.<br />
<br />
As there is always a trade-off, when we gain a low-cost robot, we are actually losing accuracy for controlling it. So, the low-cost robot which is built from cheaper components than the expensive setups such as Baxter and Sawyer suffers from higher calibration errors and execution errors. This means that the dataset collected with this approach is diverse and huge but it has noisy labels. To illustrate, consider when the robot wants to grasp at location <math> {(x, y)}</math>. Since there is a noise in the execution, the robot may perform this action in the location <math> {(x + \delta_{x}, y+ \delta_{y})}</math> which would assign the success or failure label of this action to a wrong place. Therefore, to solve the problem, they used an approach to learn from noisy data. They modeled noise as a latent variable and used two networks, one for predicting the noise and one for predicting the action to execute.<br />
<br />
==Learning on low-cost robot data==<br />
<br />
This paper uses patch grasping framework in its proposed architecture. Also, as mentioned before, there is a high tendency for noisy labels in the datasets which are collected by inaccurate and cheap robots. The cause of the noise in the labels could be due to the hardware execution error, inaccurate kinematics, camera calibration, proprioception, wear, and tear, etc. Here are more explanations about different parts of the architecture: <br />
<br />
<br />
===Grasping Formulation===<br />
<br />
Planar grasping is the object of interest in this architecture. It means that all the objects are grasped at the same height and vertical to the ground. The final goal is to find <math>{(x, y, \theta)}</math> given an observation <math> {I}</math> of the object, where <math> {x}</math> and <math> {y}</math> are the translational degrees of freedom and <math> {\theta}</math> is the rotational degrees of freedom. For the purpose of comparison, they used a model which does not predict the <math>{(x, y, \theta)}</math> directly from the image <math> {I}</math>, but samples several smaller patches <math> {I_{P}}</math> at different locations <math>{(x, y)}</math>. Thus, the angle of grasp <math> {\theta}</math> is predicted from these patches. Also, in order to have multimodal predictions, discrete steps of the angle <math> {\theta}</math>, <math> {\theta_{D}}</math> is used. <br />
<br />
Hence, each datapoint consists of an image <math> {I}</math>, the executed grasp <math>{(x, y, \theta)}</math> and the grasp success/failure label g. Then, the image <math> {I}</math> and the angle <math> {\theta}</math> are converted to image patch <math> {I_{P}}</math> and angle <math> {\theta_{D}}</math>. Then, to minimize the classification error, a binary cross entropy loss is used which minimizes the error between the predicted and ground truth label <math> g </math>. A convolutional neural network with weight initialization from pre-training on Imagenet is used for this formulation.<br />
<br />
===Modeling noise as latent variable===<br />
<br />
In order to tackle the problem of inaccurate position control, they found a structure in the noise which is dependent on the robot and the design. They modeled this structure as a latent variable which is shown is figure 2: <br />
<br />
<br />
[[File:aa2.PNG|600px|thumb|center|]]<br />
<br />
<br />
<br />
The grasp success probability for image patch <math> {I_{P}}</math> at angle <math> {\theta_{D}}</math> is represented as <math> {P(g|I_{P},\theta_{D};R )}</math> where <math> {R}</math> represents environment variables that can add noise to the system.<br />
<br />
The conditional probability of grasping for this model is computed by:<br />
<br />
<br />
\[ { P(g|I_{P},\theta_{D}, R ) = ∑_{( \hat{I}_{P} ϵ P)} P(g│z=\hat{I}_{P},\theta_{D},R ). P(z=\hat{I}_{P} |(\theta_{D},I_{P} ,R ) } \]<br />
<br />
<br />
<br />
<br />
Here, <math> {z}</math> models the latent variable of the actual patch executed, and <math> {\hat{I}_{P}}</math> belongs to a set of possible neighboring patches <math> {P}</math>. <math> {P(z=\hat{I}_{P} |(\theta_{D},I_{P} ,R )}</math> shows the noise which can be caused by <math> {R}</math> variables and is implemented as the Noise Modelling Network (NMN). <math> {P(g│z=\hat{I}_{P},\theta_{D},R )}</math> shows the grasp prediction probability given the true patch and is implemented as the Grasp Prediction Network (GPN). The overall Robust-Grasp model is computed by marginalizing GPN and NMN.<br />
<br />
===Learning the latent noise model===<br />
<br />
<br />
They assume that <math> {z}</math> is conditionally independent of the local patch-specific variables <math> {(I_{P}, \theta_{D})}</math>. To estimate the latent variable <math> {z}</math>, they used direct optimization to learn both NMN and GPN with noisy labels. The entire image of the scene and the environment information are the inputs of the NMN. The output of the NMN is the probability distribution of the actual patches where the grasps are executed. Finally, a binary cross entropy loss is applied to the marginalized output of these two networks and the true grasp label g.<br />
<br />
===Training details===<br />
<br />
<br />
They implemented their model in PyTorch using a pretrained ResNet-18 model. They concatenated 512 dimensional ResNet feature with a 1-hot vector of robot ID and the raw pixel location of the grasp for their NMN. Also, the inputs of the GPN are the original noisy patch plus 8 other equidistant patches from the original one.<br />
Their training process starts with training only GPN over 5 epochs of the data. Then, the NMN and the marginalization operator are added to the model. So, they train NMN and GPN simultaneously for the other 25 epochs.<br />
<br />
==Results==<br />
<br />
<br />
In the results part of the paper, they show that collecting dataset in homes is essential for generalizing learning from unseen environments. They also show that modeling the noise in their Low-Cost Arm (LCA) can improve grasping performance.<br />
They collected data in parallel using multiple robots in 6 different homes, as shown in Figure 3. They used an object detector (tiny-YOLO) as the input data were unstructured due to LCA limited memory and computational capabilities. They scattered different objects in homes within 2m area to prevent collision of the robot with obstacles and let the robot move randomly and grasp objects. Finally, they collected a dataset with 28K grasp results.<br />
<br />
[[File:aa3.PNG|600px|thumb|center|]]<br />
<br />
<br />
To evaluate their approach in a more quantitative way, they used three test settings:<br />
<br />
- The first one is a binary classification or held-out data. The test set is collected by performing random grasps on objects. They measure the performance of binary classification by predicting the success or failure of grasping, given a location and the angle. Using binary classification allows for testing a lot of models without running them on real robots. They collected two held-out datasets using LCA in lab and homes and the dataset for Baxter robot.<br />
<br />
- The second one is Real Low-Cost Arm(Real-LCA). Here, they evaluate their model by running it in three unseen homes. They put 20 new objects in these three homes in different orientations. Since the objects and the environments are completely new, this tests could measure the generalization of the model.<br />
<br />
- The third one is Real Sawyer(Real-Sawyer). They evaluate the performance of their model by running the model on the Sawyer robot which is more accurate than the LCA. They tested their model in the lab environment to show that training models with the datasets collected from homes can improve the performance of models even in lab environments.<br />
<br />
They used baselines for both their data which is collected in homes and their model which is Robust-Grasp. They used two datasets for the baseline. The dataset collected by (Lab-Baxter) and the dataset collected by their LCA in the lab(Lab-LCA).<br />
They compared their model with the noise independent patch grasping model (Patch-Grasp). They also compared their data and model with DexNet-3.0 (DexNet) for a strong real-world grasping baseline.<br />
<br />
<br />
<br />
===Experiment 1: Performance on held-out data===<br />
<br />
<br />
Table 1 shows that the models trained on lab data cannot generalize to the Home-LCA environment. However, the model trained on Home-LCA have a good performance on both lab data and home environment.<br />
<br />
[[File:aa4.PNG|600px|thumb|center|]]<br />
<br />
<br />
<br />
===Experiment 2: Performance on Real LCA Robot===<br />
<br />
<br />
In table 2, the performance of the Home-LCA is compared against a pre-trained DexNet and the model trained on the Lab-Baxter. Training on the Home-LCA dataset performs 43.7% better than training on the Lab-Baxter dataset and 33% better than DexNet. The low performance of DexNet can be described by the possible noise in the depth images that are caused by the natural light. So that’s why DexNet which requires high quality depth sensing cannot perform well. By using cheap commodity RGBD cameras in LCA, the noise in the depth images is not a matter of concern.<br />
<br />
[[File:aa5.PNG|600px|thumb|center|]]<br />
<br />
===Performance on Real Sawyer===<br />
<br />
<br />
To compare the performance of the Robust-Grasp model against the Patch-Grasp model, they used Lab-Baxter which is an accurate robot. Sawyer robot is used for testing to ensure that the testing robot is different from both training robots. As shown in Table 3, the Robust-Grasp model trained on Home-LCA outperforms the Patch-Grasp model and achieves 77.5% accuracy. Furthermore, the visualizations of predicted noise corrections in Figure 4 shows that the corrections depend on both the pixel locations of the noisy grasp and the robot.<br />
<br />
<br />
[[File:aa6.PNG|600px|thumb|center|]]<br />
<br />
[[File:aa7.PNG|600px|thumb|center|]]<br />
<br />
<br />
<br />
==Related work==<br />
<br />
<br />
Over the last few years, the interest of scaling up robot learning with large scale datasets has been increased. Hence, many papers were published in this area. A hand annotated grasping dataset, a self-supervised grasping dataset, and grasping using reinforcement learning are some examples of using large scale datasets for grasping. There were also many papers that worked on other robotic tasks like material recognition or pushing objects. However, none of these papers worked on real data in real environments like homes. They just used high-cost hardware and lab data.<br />
<br />
<br />
Furthermore, since grasping is one of the basic problems of robotic, there were some efforts to improve grasping. Classic approaches focused on physics-based issues of grasping and required 3D models of the objects. However, recent works focused on data-driven approaches which learn from visual observations to grasp objects. The point here is that they usually require high quality depth as input which seems to be a barrier for practical use of robots in real environments.<br />
<br />
<br />
Most labs use industrial robots or standard collaborative hardware for their experiments. Therefore, there is few research that used low cost robots. One of the examples is learning using a cheap inaccurate robot for stack multiple blocks, although it is not clear whether learning approaches are used in it alongside mapping and planning.<br />
<br />
<br />
Learning from noisy inputs is another challenge specifically in computer vision. A controversial question which is often raised in this area is whether learning from noise can improve the performance. Some works show it could have bad effects on the performance; however, some other works find it valuable when the noise is dependent of the environment. In this paper, they used a model that can exploit the noise and learn a better grasping model.<br />
<br />
==Conclusion==<br />
<br />
All in all, the paper presents an approach for collecting large-scale robot data in real home environments. They implemented their approach by using a mobile manipulator which is a lot cheaper than the existing industrial robots. They collected a dataset of 28K grasps in six different homes. In order to solve the problem of noisy labels which were caused by their inaccurate robots, they presented a framework to factor out the noise in the data. They tested their model by physically grasping 20 new objects in three new homes and in the lab. The model trained with home dataset showed 43.7% improvement over the models trained with lab data. Their results also showed that their model can improve the grasping performance even in lab environments. They also demonstrated that their architecture for modeling the noise improved the performance by about 10%.<br />
<br />
==Critiques==<br />
<br />
This paper is not a significant algorithmic contribution. They are just combining a large number of data engineering techniques for the robot learning problem. The authors claim that they have obtained 43.7% more accuracy than baseline models, but it does not seem to be a fair comparison as the data collection happened in simulated settings in the lab for other methods, whereas the authors use the home dataset. The authors must have also discussed safety issues when training robots in real environments as against simulated environments like labs. The authors are encouraging other researchers to look outside the labs, but are not discussing the critical safety issues in this approach.<br />
<br />
<br />
The paper argues that the dataset collected by the LCA is noisy, since the robot is cheap and inaccurate. It further asserts that in order to handle the noise in the dataset, they can model the noise as a latent variable and their model can improve the performance of grasping. Although learning from noisy data and achieving a good performance is valuable, it is better that they test their noise modeling network for other robots as well. Since their noise modelling network takes robot information as an input, it would be a good idea to generalize it by testing it using different inaccurate robots to ensure that it would perform well.<br />
<br />
==References==<br />
<br />
#Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. "Domain randomization for transferring deep neural networks from simulation to the real world." 2017. URL https://arxiv.org/abs/1703.06907.<br />
#Xue Bin Peng, Marcin Andrychowicz, Wojciech Zaremba, and Pieter Abbeel. "Sim-to-real transfer of robotic control with dynamics randomization." arXiv preprint arXiv:1710.06537,2017.<br />
#Lerrel Pinto, Marcin Andrychowicz, Peter Welinder, Wojciech Zaremba, and Pieter Abbeel. "Asymmetric actor critic for image-based robot learning." Robotics Science and Systems, 2018.<br />
#Lerrel Pinto and Abhinav Gupta. "Supersizing self-supervision: Learning to grasp from 50k tries and 700 robot hours." CoRR, abs/1509.06825, 2015. URL http://arxiv.org/abs/1509. 06825.<br />
#Adithyavairavan Murali, Lerrel Pinto, Dhiraj Gandhi, and Abhinav Gupta. "CASSL: Curriculum accelerated self-supervised learning." International Conference on Robotics and Automation, 2018.<br />
# Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. "End-to-end training of deep visuomotor policies." The Journal of Machine Learning Research, 17(1):1334–1373, 2016.<br />
#Sergey Levine, Peter Pastor, Alex Krizhevsky, and Deirdre Quillen. "Learning hand-eye coordination for robotic grasping with deep learning and large scale data collection." CoRR, abs/1603.02199, 2016. URL http://arxiv.org/abs/1603.02199.<br />
#Pulkit Agarwal, Ashwin Nair, Pieter Abbeel, Jitendra Malik, and Sergey Levine. "Learning to poke by poking: Experiential learning of intuitive physics." 2016. URL http://arxiv.org/ abs/1606.07419<br />
#Chelsea Finn, Ian Goodfellow, and Sergey Levine. "Unsupervised learning for physical interaction through video prediction." In Advances in neural information processing systems, 2016.<br />
#Ashvin Nair, Dian Chen, Pulkit Agrawal, Phillip Isola, Pieter Abbeel, Jitendra Malik, and Sergey Levine. "Combining self-supervised learning and imitation for vision-based rope manipulation." International Conference on Robotics and Automation, 2017.<br />
#Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. "Revisiting unreasonable effectiveness of data in deep learning era." ICCV, 2017.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Countering_Adversarial_Images_Using_Input_Transformations&diff=39872Countering Adversarial Images Using Input Transformations2018-11-19T01:15:03Z<p>Wfisher: Misc additional info (T)</p>
<hr />
<div>The code for this paper is available here[https://github.com/facebookresearch/adversarial_image_defenses]<br />
<br />
==Motivation ==<br />
As the use of machine intelligence has increased, robustness has become a critical feature to guarantee the reliability of deployed machine-learning systems. However, recent research has shown that existing models are not robust to small, adversarially designed perturbations of the input. Adversarial examples are inputs to Machine Learning models that an attacker has intentionally designed to cause the model to make a mistake. Adversarially perturbed examples have been deployed to attack image classification services (Liu et al., 2016)[11], speech recognition systems (Cisse et al., 2017a)[12], and robot vision (Melis et al., 2017)[13]. The existence of these adversarial examples has motivated proposals for approaches that increase the robustness of learning systems to such examples. In the example below (Goodfellow et. al) [17], a small perturbation is applied to the original image of a panda, changing the prediction to a gibbon.<br />
<br />
[[File:Panda.png|center]]<br />
<br />
==Introduction==<br />
The paper studies strategies that defend against adversarial-example attacks on image-classification systems by transforming the images before feeding them to a Convolutional Network Classifier. <br />
Generally, defenses against adversarial examples fall into two main categories -<br />
<br />
# Model-Specific – They enforce model properties such as smoothness and in-variance via the learning algorithm. <br />
# Model-Agnostic – They try to remove adversarial perturbations from the input. <br />
<br />
Model-specific defense strategies make strong assumptions about expected adversarial attacks. As a result, they violate the Kerchkoffs principle, which states that adversaries can circumvent model-specific defenses by simply changing how an attack is executed. This paper focuses on increasing the effectiveness of Model Agnostic defense strategies. Specifically, they investigate the following image transformations as a means for protecting against adversarial images:<br />
<br />
# Image Cropping and Re-scaling (Graese et al, 2016). <br />
# Bit Depth Reduction (Xu et. al, 2017) <br />
# JPEG Compression (Dziugaite et al, 2016) <br />
# Total Variance Minimization (Rudin et al, 1992) <br />
# Image Quilting (Efros & Freeman, 2001). <br />
<br />
These image transformations have been studied against Adversarial attacks such as the fast gradient sign method (Goodfelow et. al., 2015), its iterative extension (Kurakin et al., 2016a), Deepfool (Moosavi-Dezfooli et al., 2016), and the Carlini & Wagner (2017) <math>L_2</math>attack. <br />
<br />
From their experiments, the strongest defenses are based on Total Variance Minimization and Image Quilting. These defenses are non-differentiable and inherently random which makes it difficult for an adversary to get around them.<br />
<br />
==Previous Work==<br />
Recently, a lot of research has focused on countering adversarial threats. Wang et al [4], proposed a new adversary resistant technique that obstructs attackers from constructing impactful adversarial images. This is done by randomly nullifying features within images. Tramer et al [2], showed the state-of-the-art Ensemble Adversarial Training Method, which augments the training process but not only included adversarial images constructed from their model but also including adversarial images generated from an ensemble of other models. Their method implemented on an Inception V2 classifier finished 1st among 70 submissions of NIPS 2017 competition on Defenses against Adversarial Attacks. Graese, et al. [3], showed how input transformation such as shifting, blurring and noise can render the majority of the adversarial examples as non-adversarial. Xu et al.[5] demonstrated, how feature squeezing methods, such as reducing the color bit depth of each pixel and spatial smoothing, defends against attacks. Dziugaite et al [6], studied the effect of JPG compression on adversarial images.<br />
<br />
==Terminology==<br />
<br />
'''Gray Box Attack''' : Model Architecture and parameters are Public<br />
<br />
'''Black Box Attack''': Adversary does not have access to the model.<br />
<br />
'''Non Targeted Adversarial Attack''': The goal of the attack is to modify a source image in a way such that the image will be classified incorrectly by the network.<br />
<br />
'''Targeted Adversarial Attack''': The goal of the attack is to modify a source image in way such that image will be classified as a ''target'' class by the network.<br />
<br />
'''Defense''': A defense is a strategy that aims make the prediction on an adversarial example h(x') equal to the prediction on the corresponding clean example h(x).<br />
<br />
== Problem Definition ==<br />
The paper discusses non-targeted adversarial attacks for image recognition systems. Given image space X, a source image x ∈ X, and a classifier h(.), a non-targeted adversarial example of x is a perturbed image x', such that h(x) ≠ h(x') and d(x, x') ≤ ρ for some dissimilarity function d(·, ·) and ρ ≥ 0 (Euclidean distance is often used).<br />
<br />
From a set of N clean images <math>[{x_{1}, …, x_{n}}]</math>, an adversarial attack aims to generate <math>[{x^{'}_{1}, …, x^{'}_{n}}]</math> images, such that (<math>x^{'}_{n}</math>) is an adversary of (<math>x_{n}</math>).<br />
<br />
The success rate of an attack is given as: <br />
<br />
[[File:Attack.PNG|200px |]],<br />
<br />
which is the proportions of predictions that were altered by an attack.<br />
<br />
The success rate is generally measured as a function of the magnitude of perturbations performed by the attack. In this paper, L2 perturbations are used and are quantified using the normalized L2-dissimilarity metric:<br />
[[File:diss.png|150px |]]<br />
<br />
A strong adversarial attack has a high rate, while its normalized L2-dissimilarity given by the above equation is less.<br />
<br />
==Adversarial Attacks==<br />
<br />
For the experimental purposes, below 4 attacks have been studied.<br />
<br />
1. '''Fast Gradient Sign Method (FGSM; Goodfellow et al. (2015)) [17]''': Given a source input x, and true label y, and let l be the differentiable loss function used to train the classifier h(.). Then the corresponding adversarial example is given by:<br />
<br />
[[File:FGSM.PNG|200px |]]<br />
<br />
2. '''Iterative FGSM ((I-FGSM; Kurakin et al. (2016b))[14]''': iteratively applies the FGSM update, where M is the number of iterations.It is given as:<br />
<br />
[[File:IFGSM.PNG|300px |]]<br />
<br />
3. '''DeepFool ((Moosavi-Dezfooliet al., 2016) [15]''': projects x onto a linearization of the decision boundary defined by binary classifier h(.) for M iterations. This can be particularly effictive when a network uses ReLU activation functions. It is given as:<br />
<br />
[[File:DeepFool.PNG|400px |]]<br />
<br />
4. '''Carlini-Wagner's L2 attack(CW-L2; Carlini & Wagner (2017))[16]''': propose an optimization-based attack that combines a differentiable surrogate for the model’s classification accuracy with an L2-penalty term which encourages the adversary image to be close to the original image. Let Z(x) be the operation that computes the logit vector (i.e., the output before the softmax layer) for an input x, and Z(x)k be the logit value corresponding to class k. The untargeted variant<br />
of CW-L2 finds a solution to the unconstrained optimization problem. It is given as:<br />
<br />
[[File:Carlini.PNG|500px |]]<br />
<br />
Below figure shows adversarial images and corresponding perturbations at five levels of normalized L2-dissimilarity for all four attacks, mentioned above.<br />
<br />
[[File:Strength.PNG|center|600px |]]<br />
<br />
==Defenses==<br />
Defense is a strategy that aims make the prediction on an adversarial example equal to the prediction on the corresponding clean example. <br />
Five image transformations that alter the structure of these perturbations have been studied:<br />
# Image Cropping and Re-scaling, <br />
# Bit Depth Reduction, <br />
# JPEG Compression, <br />
# Total Variance Minimization, <br />
# Image Quilting.<br />
<br />
'''Image cropping and Rescaling''' has the effect of altering the spatial positioning of the adversarial perturbation. In this study, images are cropped and re-scaled during training time. At test time, the predictions of randomly cropped are averaged.<br />
<br />
'''Bit Depth Reduction( Xu et. al)''' performs a simple type of quantization that can remove small (adversarial) variations in pixel values from an image. Images are reduced to 3 bits in the experiment.<br />
<br />
'''JPEG Compression and Decompression (Dziugaite etal., 2016)''' removes small perturbations by performing simple quantization. The authors use a quality level of 75/100 in their experiments<br />
<br />
'''Total Variance Minimization [9]''' :<br />
This combines pixel dropout with total variance minimization. This approach randomly selects a small set of pixels, and reconstructs the “simplest” image that is consistent with the selected pixels. The reconstructed image does not contain the adversarial perturbations because these perturbations tend to be small and localized.Specifically, we first select a random set of pixels by sampling a Bernoulli random variable <math>X(i; j; k)</math> for each pixel location <math>(i; j; k)</math>;we maintain a pixel when <math>(i; j; k)</math>= 1. Next, we use total variation, minimization to constructs an image z that is similar to the (perturbed) input image x for the selected<br />
set of pixels, whilst also being “simple” in terms of total variation by solving:<br />
<br />
[[File:TV!.png|300px|]] , <br />
<br />
where <math>TV_{p}(z)</math> represents <math>L_{p}</math> total variation of '''z''' :<br />
<br />
[[File:TV2.png|500px|]]<br />
<br />
The total variation (TV) measures the amount of fine-scale variation in the image z, as a result of which TV minimization encourages removal of small (adversarial) perturbations in the image.<br />
<br />
'''Image Quilting(Efros & Freeman, 2001)[8]'''<br />
Image Quilting is a non-parametric technique that synthesizes images by piecing together small patches that are taken from a database of image patches. The algorithm places appropriate patches in the database for a predefined set of grid points and computes minimum graph cuts in all overlapping boundary regions to remove edge artifacts. Image Quilting can be used to remove adversarial perturbations by constructing a patch database that only contains patches from "clean" images ( without adversarial perturbations); the patches used to create the synthesized image are selected by finding the K nearest neighbors ( in pixel space) of the corresponding patch from the adversarial image in the patch database, and picking one of these neighbors uniformly at random. The motivation for this defense is that resulting image only contains pixels that were not modified by the adversary - the database of real patches is unlikely to contain the structures that appear in adversarial images.<br />
<br />
=Experiments=<br />
<br />
Five experiments were performed to test the efficacy of defenses. The first two experiments focus on gray-box and black-box attacks, respectively. Further experiments test other possiblities. The final experiment compares the authors' defences with prior work.<br />
<br />
'''Set up:'''<br />
Experiments are performed on the ImageNet image classification dataset. The dataset comprises 1.2 million training images and 50,000 test images that correspond to one of 1000 classes. The adversarial images are produced by attacking a ResNet-50 model, with different kinds of attacks mentioned in Section5. The strength of an adversary is measured in terms of its normalized L2-dissimilarity. To produce the adversarial images, L2 dissimilarity for each of the attack was set as below:<br />
<br />
- FGSM. Increasing the step size <math>\epsilon</math>, increases the normalized L2-dissimilarity.<br />
<br />
- I-FGSM. We fix M=10, and increase <math>\epsilon</math> to increase the normalized L2-dissimilarity.<br />
<br />
- DeepFool. We fix M=5, and increase <math>\epsilon</math> to increase the normalized L2-dissimilarity.<br />
<br />
- CW-L2. We fix <math>k</math>=0 and <math>\lambda_{f}</math> =10, and multiply the resulting perturbation <br />
<br />
The hyperparameters of the defenses have been fixed in all the experiments. Specifically the pixel dropout probability was set to <math>p</math>=0.5 and regularization parameter of total variation minimizer <math>\lambda_{TV}</math>=0.03.<br />
<br />
Below figure shows the difference between the set up in different experiments below. The network is either trained on a) regular images or b) transformed images. The different settings are marked by 8.1, 8.2 and 8.3 <br />
[[File:models3.png]] <br />
<br />
==GrayBox- Image Transformation at Test Time== <br />
This experiment applies a transformation on adversarial images at test time before feeding them to a ResNet -50 which was trained to classify clean images. Below figure shows the results for five different transformations applied and their corresponding Top-1 accuracy. Few of the interesting observations from the plot are: All of the image transformations partly eliminate the effects of the attack, Crop ensemble gives the best accuracy around 40-60 percent, with an ensemble size of 30. The accuracy of Image Quilting Defense hardly deteriorates as the strength of the adversary increases. However, it does impact accuracy on non-adversarial examples.<br />
<br />
[[File:sFig4.png|center|600px |]]<br />
<br />
==BlackBox - Image Transformation at Training and Test Time==<br />
ResNet-50 model was trained on transformed ImageNet Training images. Before feeding the images to the network for training, standard data augmentation (from He et al) along with bit depth reduction, JPEG Compression, TV Minimization, or Image Quilting were applied on the images. The classification accuracy on the same adversarial images as in the previous case is shown Figure below. (Adversary cannot get this trained model to generate new images - Hence this is assumed as a Black Box setting!). Below figure concludes that training Convolutional Neural Networks on images that are transformed in the same way at test time, dramatically improves the effectiveness of all transformation defenses. Nearly 80 -90 % of the attacks are defended successfully, even when the L2- dissimilarity is high.<br />
<br />
<br />
[[File:sFig5.png|center|600px |]]<br />
<br />
<br />
==Blackbox - Ensembling==<br />
Four networks ResNet-50, ResNet-10, DenseNet-169, and Inception-v4 along with an ensemble of defenses were studied, as shown in Table 1. The adversarial images are produced by attacking a ResNet-50 model. The results in the table conclude that Inception-v4 performs best. This could be due to that network having a higher accuracy even in non-adversarial settings. The best ensemble of defenses achieves an accuracy of about 71% against all the other attacks. The attacks deteriorate the accuracy of the best defenses (a combination of cropping, TVM, image quilting, and model transfer) by at most 6%.<br />
<br />
<br />
[[File:sTab1.png|600px|thumb|center|Table 1. Top-1 classification accuracy of ensemble and model transfer defenses (columns) against four black-box attacks (rows). The four networks we use to classify images are ResNet-50 (RN50), ResNet-101 (RN101), DenseNet-169 (DN169), and Inception-v4 (Iv4). Adversarial images are generated by running attacks against the ResNet-50 model, aiming for an average normalized <math>L_2</math>-dissimilarity of 0.06. Higher is better. The best defense against each attack is typeset in boldface.]]<br />
<br />
==GrayBox - Image Transformation at Training and Test Time ==<br />
In this experiment, the adversary has access to the network and the related parameters (but does not have access to the input transformations applied at test time). From the network trained in-(BlackBox: Image Transformation at Training and Test Time), novel adversarial images were generated by the four attack methods. The results show that Bit-Depth Reduction and JPEG Compression are weak defenses in such a gray box setting. In contrast, image cropping, rescaling, variation minimization, and image quilting are more robust against adversarial images in this setting.<br />
The results for this experiment are shown in below figure. Networks using these defenses classify up to 50 % of images correctly.<br />
<br />
[[File:sFig6.png|center| 600px |]]<br />
<br />
==Comparison With Ensemble Adversarial Training==<br />
The results of the experiment are compared with the state of the art ensemble adversarial training approach proposed by Tramer et al. [2] 2017. Ensemble Training fits the parameters of a Convolutional Neural Network on adversarial examples that were generated to attack an ensemble of pre-trained models. The model release by Tramer et al [2]: an Inception-Resnet-v2, trained on adversarial examples generated by FGSM against Inception-Resnet-v2 and Inception-v3<br />
models. The results of ensemble training and the preprocessing techniques mentioned in this paper are shown in Table 2.<br />
The results show that ensemble adversarial training works better on FGSM attacks (which it uses at training time), but is outperformed by each of the transformation-based defenses all other attacks.<br />
<br />
<br />
<br />
[[File:sTab2.png|600px|thumb|center|Table 2. Top-1 classification accuracy on images perturbed using attacks against ResNet-50 models trained on input-transformed images and an Inception-v4 model trained using ensemble adversarial. Adversarial images are generated by running attacks against the models, aiming for an average normalized <math>L_2</math>-dissimilarity of 0.06. The best defense against each attack is typeset in boldface.]]<br />
<br />
=Discussion/Conclusions=<br />
The paper proposed reasonable approaches to countering adversarial images. The authors evaluated Total Variance Minimization and Image Quilting and compared it with already proposed ideas like Image Cropping- Rescaling, Bit Depth Reduction, JPEG Compression, and Decompression on the challenging ImageNet dataset.<br />
Previous work by (Wang et al) [10], shows that a strong input defense should, be nondifferentiable and randomized. Two of the defenses - namely Total Variation Minimization and Image Quilting, both possess this property. Future work suggests applying the same techniques to other domains such as speech recognition and image segmentation. For example, in speech recognition, total variance minimization can be used to remove perturbations from waveforms and "spectrogram quilting" techniques that reconstruct a spectrogram could be developed. The input transformations can also be studied with ensemble adversarial training by Tramèr et al.[2]<br />
<br />
<br />
=Critiques=<br />
1. The terminology of Black Box, White Box, and Grey Box attack is not exactly given and clear.<br />
<br />
2. White Box attacks could have been considered where the adversary has a full access to the model as well as the pre-processing techniques.<br />
<br />
3. Though the authors did a considerable work in showing the effect of four attacks on ImageNet database, much stronger attacks (Madry et al) [7], could have been evaluated.<br />
<br />
4. Authors claim that the success rate is generally measured as a function of the magnitude of perturbations, performed by the attack using the L2- dissimilarity, but the claim is not supported by any references. None of the previous work has used these metrics.<br />
<br />
=References=<br />
<br />
1. Chuan Guo , Mayank Rana & Moustapha Ciss´e & Laurens van der Maaten , Countering Adversarial Images Using Input Transformations<br />
<br />
2. Florian Tramèr, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, Patrick McDaniel, Ensemble Adversarial Training: Attacks and defenses.<br />
<br />
3. Abigail Graese, Andras Rozsa, and Terrance E. Boult. Assessing threat of adversarial examples of deep neural networks. CoRR, abs/1610.04256, 2016. <br />
<br />
4. Qinglong Wang, Wenbo Guo, Kaixuan Zhang, Alexander G. Ororbia II, Xinyu Xing, C. Lee Giles, and Xue Liu. Adversary resistant deep neural networks with an application to malware detection. CoRR, abs/1610.01239, 2016a.<br />
<br />
5.Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. CoRR, abs/1704.01155, 2017. <br />
<br />
6. Gintare Karolina Dziugaite, Zoubin Ghahramani, and Daniel Roy. A study of the effect of JPG compression on adversarial images. CoRR, abs/1608.00853, 2016.<br />
<br />
7.Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, Adrian Vladu .Towards Deep Learning Models Resistant to Adversarial Attacks, arXiv:1706.06083v3<br />
<br />
8.Alexei Efros and William Freeman. Image quilting for texture synthesis and transfer. In Proc. SIGGRAPH, pp. 341–346, 2001.<br />
<br />
9.Leonid Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms. Physica D, 60:259–268, 1992.<br />
<br />
10.Qinglong Wang, Wenbo Guo, Kaixuan Zhang, Alexander G. Ororbia II, Xinyu Xing, C. Lee Giles, and Xue Liu. Learning adversary-resistant deep neural networks. CoRR, abs/1612.01401, 2016b.<br />
<br />
11. Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. CoRR, abs/1611.02770, 2016.<br />
<br />
12. Moustapha Cisse, Yossi Adi, Natalia Neverova, and Joseph Keshet. Houdini: Fooling deep structured prediction models. CoRR, abs/1707.05373, 2017 <br />
<br />
13. Marco Melis, Ambra Demontis, Battista Biggio, Gavin Brown, Giorgio Fumera, and Fabio Roli. Is deep learning safe for robot vision? adversarial examples against the icub humanoid. CoRR,abs/1708.06939, 2017.<br />
<br />
14. Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. CoRR, abs/1607.02533, 2016b.<br />
<br />
15. Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: A simple and accurate method to fool deep neural networks. In Proc. CVPR, pp. 2574–2582, 2016.<br />
<br />
16. Nicholas Carlini and David A. Wagner. Towards evaluating the robustness of neural networks. In IEEE Symposium on Security and Privacy, pp. 39–57, 2017.<br />
<br />
17. Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In Proc. ICLR, 2015.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39529Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-17T01:12:13Z<p>Wfisher: /* Value Network */</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, [5]) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, [6]) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, [7]) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail, balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or even a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Variable Definitions ==<br />
<br />
The following variables are used often in the paper:<br />
<br />
* <math>s</math>: A state in the game, as described below as the input to the network.<br />
* <math>s_t</math>: The state at a certain time-step of the game. Time-steps refer to full turns in the game<br />
* <math>a_t</math>: The action taken in state <math>s_t</math><br />
* <math>A_t</math>: The actions taken for sibling nodes related to <math>a_t</math> in MCTS<br />
* <math>n_{a_t}</math>: The number of visits to node a in MCTS<br />
* <math>v_{a_t}</math>: The MCTS value estimate of a node<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 1 convolutional layer followed by 9 residual blocks, each block consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is distribution of probabilities of the actions to select the best shot out of 32x32x2 set of actions. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games). These are likely estimated rather than simply taken from the MCTS statistics to help account for the differences in a continuous action space.<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
The actions that are taken in the simulator appear to be drawn from a gaussian centered around <math>a_t</math>. This allows exploration in the continuous action space.<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times.<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 ([8]). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network ===<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, <math>(s_t, a_t)</math>, a state 'd' steps ahead was generated, <math>s_{t+d}</math>. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, <math>z_t</math>, and the value is used for learning the value at ''s<sub>t</sub>''.<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
After initialisation by supervised learning, the algorithm uses self-play to further train itself. During this training, the policy network learns probabilities from the MCTS process, while the value network learns from game outcomes.<br />
<br />
At a game state ''s<sub>t</sub>'':<br />
<br />
1) the algorithm outputs a prediction ''z<sub>t</sub>''. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.<br />
<br />
2) the algorithm outputs a prediction <math>\pi_t</math>, representing a probability distribution of actions. These are proportional to estimated visit counts from MCTS, based on kernel density estimation.<br />
<br />
It is not clear how these predictions are created. It would seem likely that the policy-value network generates these, but the wording of the paper suggests they are generated from MCTS statistics.<br />
<br />
The policy-value network is updated by sampling data <math>(s, \pi, z)</math> from recent history of self-play. The same loss function is used as before.<br />
<br />
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.<br />
<br />
== Long-Term Strategy Learning ==<br />
<br />
Finally, the authors implement a new strategy to augment their algorithm for long-term play. In this context, this refers to playing a game over many ends, where the strategy to win a single end may not be a good strategy to win a full game. For example, scoring one point in an end, while being one point ahead, gives the advantage to the other team in the next round (as they will throw the last stone). The other team could then use the advantage to score two points, taking the lead.<br />
<br />
The authors build a 'winning percentage' table. This table stores the percentage of games won, based on number of ends left, and difference in score (current team - opposing team). This can be computed iteratively, and using the probability distribution estimation of one-end scores.<br />
<br />
== Final Algorithms ==<br />
<br />
The authors make use of the following versions of their algorithm:<br />
<br />
=== KR-DL ===<br />
<br />
''Kernel regression-deep learning'': This algorithm is trained only by supervised learning.<br />
<br />
=== KR-DRL ===<br />
<br />
''Kernel regression-deep reinforcement learning'': This algorithm is trained by supervised learning (ie: initialised as the KR-DL algorithm), and again on self-play. During self-play, each shot is selected after 400 MCTS simulations of k=20 randomly selected actions. Data for self-play was collected over a week on 5 GPUS, and generated 5 million game positions. The policy-value network was continually updated using samples from the latest 1 million game positions.<br />
<br />
=== KR-DRL-MES ===<br />
<br />
''Kernel regression-deep reinforcement learning-multi-ends-strategy'': This algorithm makes use of the winning percentage table generated from self-play.<br />
<br />
= Testing and Results =<br />
<br />
Testing is done with a simulated curling program [9]. This simulator does not deal with changing ice conditions, or sweeping, but does deal with stone trajectories and collisions.<br />
<br />
== Comparison of KR-DL-UCT and DL-UCT ==<br />
<br />
The first test compares an algorithm trained with kernel regression with an algorithm trained without kernel regression, to show the contribution that kernel regression adds to the performance. Both algorithms have networks initialised with the supervised learning, and then trained with two different algorithms for self-play. KR-DL-UCT uses the algorithm described above. The authors do not go into detail on how DL-UCT selects shots, but state that a constant is set to allow exploration.<br />
<br />
As an evaluation, both algorithms play 2000 games against the DL-UCT algorithm, which is frozen after supervised training. 1000 games are played with the algorithm taking the first, and 100 taking the 2nd, shots. The games were two-end games. The figure below shows each algorithm's wining percentage given different amounts of training data. While the DL-UCT outperforms the supervised-training-only-DL-UCT algorithm, the KR-DL-UCT algorithm performs much better.<br />
<br />
[[File:curling_KR_test.png | 400px]]<br />
<br />
== Matches ==<br />
<br />
Finally, to test the performance of their multiple algorithms, the authors run matches between their algorithms and other existing programs. Each algorithm plays 200 matches against each other program, 100 of which are played as the first-playing team, and 100 as the second-playing team. Only 1 program was able to out-perform the KR-DRL algorithm. The authors state that this program, ''JiritsukunGAT'17'' also uses a deep network, and hand-crafted features. However, the KR-DRL-MES algorithm was still able to out-perform this. The figure below shows the Elo ratings of the different programs. Note that the programs in blue are those created by the authors.<br />
<br />
[[File:curling_ratings.png | 400px]]<br />
<br />
= Critique =<br />
<br />
== Strengths ==<br />
<br />
This algorithm out-performs other high-performance algorithms (including past competition champions).<br />
<br />
I think the paper does a decent job of comparing the performance of their algorithm to others. They are able to clearly show the benefits of many of their additions.<br />
<br />
The authors do seem to be able to adapt strategies similar to those used in Go and other games to the continuous action-space domain. In addition, the final strategy needs no hand-crafted features for learning.<br />
<br />
== Weaknesses ==<br />
<br />
I found this paper difficult to follow at times. One problem was that the algorithms were introduced first, and then how they were used was described. So when the paper stated that self-play shots were taken after 400 simulations, it seemed unclear what simulations were being run, and what stage of the algorithm (ex: MCTS simulations, simulations sped up by using the value network, full simulations on the curling simulator). In particular, both the MCTS statistics and policy-value network could be used to estimate both action probabilities and state values, so it is difficult to tell which is used in which case. There was also no clear distinction between discrete-space actions and continuous-space actions.<br />
<br />
While I think the comparing of different algorithms was done well, I believe it still lacked some good detail. There were one-off mentions in the paper which would have been nice to see as results. These include the statement that having a policy-value network in place of two networks lead to better performance.<br />
<br />
At this point, the algorithms used still rely on initialisation by a pre-made program.<br />
<br />
There was little theoretical development or justification done in this paper.<br />
<br />
While curling is an interesting choice for demonstrating the algorithm, the fact that the simulations used did not support many of the key points of curling (ice conditions, sweeping) seems very limiting. Another game, such as pool, would likely have offered some of the same challenges, but offered more high-fidelity simulations/training.<br />
<br />
While the spatial placements of stones was discretized in a grid, the curl of thrown stones was discretized to only +/-1. This seems like it may limit learning high- and low-spin moves. It should be noted that having zero spin is not commonly used, to the best of my knowledge.<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/<br />
# Silver, D., Huang, A., Maddison, C., Guez, A., Sifre, L.,Van Den Driessche, G., Schrittwieser, J., Antonoglou, I.,Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe,D., Nham, J., Kalchbrenner, N.,Sutskever, I., Lillicrap, T.,Leach, M., Kavukcuoglu, K., Graepel, T., and Hassabis,D. Mastering the game of go with deep neural networksand tree search. Nature, pp. 484–489, 2016.<br />
# Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou,I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L.,van den Driessche, G., Graepel, T., and Hassabis, D.Mastering the game of go without human knowledge.Nature, pp. 354–359, 2017.<br />
# Yamamoto, M., Kato, S., and Iizuka, H. Digital curling strategy based on game tree search. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 474–480, 2015.<br />
# Ohto, K. and Tanaka, T. A curling agent based on the montecarlo tree search considering the similarity of the best action among similar states. In Proceedings of Advances in Computer Games, ACG, pp. 151–164, 2017.<br />
# Ito, T. and Kitasei, Y. Proposal and implementation of digital curling. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 469–473, 2015.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39528Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-17T01:10:53Z<p>Wfisher: /* Weaknesses */</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, [5]) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, [6]) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, [7]) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail, balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or even a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Variable Definitions ==<br />
<br />
The following variables are used often in the paper:<br />
<br />
* <math>s</math>: A state in the game, as described below as the input to the network.<br />
* <math>s_t</math>: The state at a certain time-step of the game. Time-steps refer to full turns in the game<br />
* <math>a_t</math>: The action taken in state <math>s_t</math><br />
* <math>A_t</math>: The actions taken for sibling nodes related to <math>a_t</math> in MCTS<br />
* <math>n_{a_t}</math>: The number of visits to node a in MCTS<br />
* <math>v_{a_t}</math>: The MCTS value estimate of a node<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 1 convolutional layer followed by 9 residual blocks, each block consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is distribution of probabilities of the actions to select the best shot out of 32x32x2 set of actions. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games). These are likely estimated rather than simply taken from the MCTS statistics to help account for the differences in a continuous action space.<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
The actions that are taken in the simulator appear to be drawn from a gaussian centered around <math>a_t</math>. This allows exploration in the continuous action space.<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times.<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 ([8]). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network ===<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, <math>(s_t, a_t)</math>, a state 'd' steps ahead was generated, <math>s_{t+d}</math>. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, and the value is used for learning the value at ''s<sub>t</sub>''. The authors do not specify, but this is likely done through back-propogating the estimated value, and then training the value network to learn the updated score.<br />
<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
After initialisation by supervised learning, the algorithm uses self-play to further train itself. During this training, the policy network learns probabilities from the MCTS process, while the value network learns from game outcomes.<br />
<br />
At a game state ''s<sub>t</sub>'':<br />
<br />
1) the algorithm outputs a prediction ''z<sub>t</sub>''. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.<br />
<br />
2) the algorithm outputs a prediction <math>\pi_t</math>, representing a probability distribution of actions. These are proportional to estimated visit counts from MCTS, based on kernel density estimation.<br />
<br />
It is not clear how these predictions are created. It would seem likely that the policy-value network generates these, but the wording of the paper suggests they are generated from MCTS statistics.<br />
<br />
The policy-value network is updated by sampling data <math>(s, \pi, z)</math> from recent history of self-play. The same loss function is used as before.<br />
<br />
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.<br />
<br />
== Long-Term Strategy Learning ==<br />
<br />
Finally, the authors implement a new strategy to augment their algorithm for long-term play. In this context, this refers to playing a game over many ends, where the strategy to win a single end may not be a good strategy to win a full game. For example, scoring one point in an end, while being one point ahead, gives the advantage to the other team in the next round (as they will throw the last stone). The other team could then use the advantage to score two points, taking the lead.<br />
<br />
The authors build a 'winning percentage' table. This table stores the percentage of games won, based on number of ends left, and difference in score (current team - opposing team). This can be computed iteratively, and using the probability distribution estimation of one-end scores.<br />
<br />
== Final Algorithms ==<br />
<br />
The authors make use of the following versions of their algorithm:<br />
<br />
=== KR-DL ===<br />
<br />
''Kernel regression-deep learning'': This algorithm is trained only by supervised learning.<br />
<br />
=== KR-DRL ===<br />
<br />
''Kernel regression-deep reinforcement learning'': This algorithm is trained by supervised learning (ie: initialised as the KR-DL algorithm), and again on self-play. During self-play, each shot is selected after 400 MCTS simulations of k=20 randomly selected actions. Data for self-play was collected over a week on 5 GPUS, and generated 5 million game positions. The policy-value network was continually updated using samples from the latest 1 million game positions.<br />
<br />
=== KR-DRL-MES ===<br />
<br />
''Kernel regression-deep reinforcement learning-multi-ends-strategy'': This algorithm makes use of the winning percentage table generated from self-play.<br />
<br />
= Testing and Results =<br />
<br />
Testing is done with a simulated curling program [9]. This simulator does not deal with changing ice conditions, or sweeping, but does deal with stone trajectories and collisions.<br />
<br />
== Comparison of KR-DL-UCT and DL-UCT ==<br />
<br />
The first test compares an algorithm trained with kernel regression with an algorithm trained without kernel regression, to show the contribution that kernel regression adds to the performance. Both algorithms have networks initialised with the supervised learning, and then trained with two different algorithms for self-play. KR-DL-UCT uses the algorithm described above. The authors do not go into detail on how DL-UCT selects shots, but state that a constant is set to allow exploration.<br />
<br />
As an evaluation, both algorithms play 2000 games against the DL-UCT algorithm, which is frozen after supervised training. 1000 games are played with the algorithm taking the first, and 100 taking the 2nd, shots. The games were two-end games. The figure below shows each algorithm's wining percentage given different amounts of training data. While the DL-UCT outperforms the supervised-training-only-DL-UCT algorithm, the KR-DL-UCT algorithm performs much better.<br />
<br />
[[File:curling_KR_test.png | 400px]]<br />
<br />
== Matches ==<br />
<br />
Finally, to test the performance of their multiple algorithms, the authors run matches between their algorithms and other existing programs. Each algorithm plays 200 matches against each other program, 100 of which are played as the first-playing team, and 100 as the second-playing team. Only 1 program was able to out-perform the KR-DRL algorithm. The authors state that this program, ''JiritsukunGAT'17'' also uses a deep network, and hand-crafted features. However, the KR-DRL-MES algorithm was still able to out-perform this. The figure below shows the Elo ratings of the different programs. Note that the programs in blue are those created by the authors.<br />
<br />
[[File:curling_ratings.png | 400px]]<br />
<br />
= Critique =<br />
<br />
== Strengths ==<br />
<br />
This algorithm out-performs other high-performance algorithms (including past competition champions).<br />
<br />
I think the paper does a decent job of comparing the performance of their algorithm to others. They are able to clearly show the benefits of many of their additions.<br />
<br />
The authors do seem to be able to adapt strategies similar to those used in Go and other games to the continuous action-space domain. In addition, the final strategy needs no hand-crafted features for learning.<br />
<br />
== Weaknesses ==<br />
<br />
I found this paper difficult to follow at times. One problem was that the algorithms were introduced first, and then how they were used was described. So when the paper stated that self-play shots were taken after 400 simulations, it seemed unclear what simulations were being run, and what stage of the algorithm (ex: MCTS simulations, simulations sped up by using the value network, full simulations on the curling simulator). In particular, both the MCTS statistics and policy-value network could be used to estimate both action probabilities and state values, so it is difficult to tell which is used in which case. There was also no clear distinction between discrete-space actions and continuous-space actions.<br />
<br />
While I think the comparing of different algorithms was done well, I believe it still lacked some good detail. There were one-off mentions in the paper which would have been nice to see as results. These include the statement that having a policy-value network in place of two networks lead to better performance.<br />
<br />
At this point, the algorithms used still rely on initialisation by a pre-made program.<br />
<br />
There was little theoretical development or justification done in this paper.<br />
<br />
While curling is an interesting choice for demonstrating the algorithm, the fact that the simulations used did not support many of the key points of curling (ice conditions, sweeping) seems very limiting. Another game, such as pool, would likely have offered some of the same challenges, but offered more high-fidelity simulations/training.<br />
<br />
While the spatial placements of stones was discretized in a grid, the curl of thrown stones was discretized to only +/-1. This seems like it may limit learning high- and low-spin moves. It should be noted that having zero spin is not commonly used, to the best of my knowledge.<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/<br />
# Silver, D., Huang, A., Maddison, C., Guez, A., Sifre, L.,Van Den Driessche, G., Schrittwieser, J., Antonoglou, I.,Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe,D., Nham, J., Kalchbrenner, N.,Sutskever, I., Lillicrap, T.,Leach, M., Kavukcuoglu, K., Graepel, T., and Hassabis,D. Mastering the game of go with deep neural networksand tree search. Nature, pp. 484–489, 2016.<br />
# Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou,I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L.,van den Driessche, G., Graepel, T., and Hassabis, D.Mastering the game of go without human knowledge.Nature, pp. 354–359, 2017.<br />
# Yamamoto, M., Kato, S., and Iizuka, H. Digital curling strategy based on game tree search. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 474–480, 2015.<br />
# Ohto, K. and Tanaka, T. A curling agent based on the montecarlo tree search considering the similarity of the best action among similar states. In Proceedings of Advances in Computer Games, ACG, pp. 151–164, 2017.<br />
# Ito, T. and Kitasei, Y. Proposal and implementation of digital curling. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 469–473, 2015.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39527Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-17T01:10:19Z<p>Wfisher: /* Selection */</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, [5]) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, [6]) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, [7]) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail, balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or even a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Variable Definitions ==<br />
<br />
The following variables are used often in the paper:<br />
<br />
* <math>s</math>: A state in the game, as described below as the input to the network.<br />
* <math>s_t</math>: The state at a certain time-step of the game. Time-steps refer to full turns in the game<br />
* <math>a_t</math>: The action taken in state <math>s_t</math><br />
* <math>A_t</math>: The actions taken for sibling nodes related to <math>a_t</math> in MCTS<br />
* <math>n_{a_t}</math>: The number of visits to node a in MCTS<br />
* <math>v_{a_t}</math>: The MCTS value estimate of a node<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 1 convolutional layer followed by 9 residual blocks, each block consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is distribution of probabilities of the actions to select the best shot out of 32x32x2 set of actions. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games). These are likely estimated rather than simply taken from the MCTS statistics to help account for the differences in a continuous action space.<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
The actions that are taken in the simulator appear to be drawn from a gaussian centered around <math>a_t</math>. This allows exploration in the continuous action space.<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times.<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 ([8]). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network ===<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, <math>(s_t, a_t)</math>, a state 'd' steps ahead was generated, <math>s_{t+d}</math>. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, and the value is used for learning the value at ''s<sub>t</sub>''. The authors do not specify, but this is likely done through back-propogating the estimated value, and then training the value network to learn the updated score.<br />
<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
After initialisation by supervised learning, the algorithm uses self-play to further train itself. During this training, the policy network learns probabilities from the MCTS process, while the value network learns from game outcomes.<br />
<br />
At a game state ''s<sub>t</sub>'':<br />
<br />
1) the algorithm outputs a prediction ''z<sub>t</sub>''. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.<br />
<br />
2) the algorithm outputs a prediction <math>\pi_t</math>, representing a probability distribution of actions. These are proportional to estimated visit counts from MCTS, based on kernel density estimation.<br />
<br />
It is not clear how these predictions are created. It would seem likely that the policy-value network generates these, but the wording of the paper suggests they are generated from MCTS statistics.<br />
<br />
The policy-value network is updated by sampling data <math>(s, \pi, z)</math> from recent history of self-play. The same loss function is used as before.<br />
<br />
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.<br />
<br />
== Long-Term Strategy Learning ==<br />
<br />
Finally, the authors implement a new strategy to augment their algorithm for long-term play. In this context, this refers to playing a game over many ends, where the strategy to win a single end may not be a good strategy to win a full game. For example, scoring one point in an end, while being one point ahead, gives the advantage to the other team in the next round (as they will throw the last stone). The other team could then use the advantage to score two points, taking the lead.<br />
<br />
The authors build a 'winning percentage' table. This table stores the percentage of games won, based on number of ends left, and difference in score (current team - opposing team). This can be computed iteratively, and using the probability distribution estimation of one-end scores.<br />
<br />
== Final Algorithms ==<br />
<br />
The authors make use of the following versions of their algorithm:<br />
<br />
=== KR-DL ===<br />
<br />
''Kernel regression-deep learning'': This algorithm is trained only by supervised learning.<br />
<br />
=== KR-DRL ===<br />
<br />
''Kernel regression-deep reinforcement learning'': This algorithm is trained by supervised learning (ie: initialised as the KR-DL algorithm), and again on self-play. During self-play, each shot is selected after 400 MCTS simulations of k=20 randomly selected actions. Data for self-play was collected over a week on 5 GPUS, and generated 5 million game positions. The policy-value network was continually updated using samples from the latest 1 million game positions.<br />
<br />
=== KR-DRL-MES ===<br />
<br />
''Kernel regression-deep reinforcement learning-multi-ends-strategy'': This algorithm makes use of the winning percentage table generated from self-play.<br />
<br />
= Testing and Results =<br />
<br />
Testing is done with a simulated curling program [9]. This simulator does not deal with changing ice conditions, or sweeping, but does deal with stone trajectories and collisions.<br />
<br />
== Comparison of KR-DL-UCT and DL-UCT ==<br />
<br />
The first test compares an algorithm trained with kernel regression with an algorithm trained without kernel regression, to show the contribution that kernel regression adds to the performance. Both algorithms have networks initialised with the supervised learning, and then trained with two different algorithms for self-play. KR-DL-UCT uses the algorithm described above. The authors do not go into detail on how DL-UCT selects shots, but state that a constant is set to allow exploration.<br />
<br />
As an evaluation, both algorithms play 2000 games against the DL-UCT algorithm, which is frozen after supervised training. 1000 games are played with the algorithm taking the first, and 100 taking the 2nd, shots. The games were two-end games. The figure below shows each algorithm's wining percentage given different amounts of training data. While the DL-UCT outperforms the supervised-training-only-DL-UCT algorithm, the KR-DL-UCT algorithm performs much better.<br />
<br />
[[File:curling_KR_test.png | 400px]]<br />
<br />
== Matches ==<br />
<br />
Finally, to test the performance of their multiple algorithms, the authors run matches between their algorithms and other existing programs. Each algorithm plays 200 matches against each other program, 100 of which are played as the first-playing team, and 100 as the second-playing team. Only 1 program was able to out-perform the KR-DRL algorithm. The authors state that this program, ''JiritsukunGAT'17'' also uses a deep network, and hand-crafted features. However, the KR-DRL-MES algorithm was still able to out-perform this. The figure below shows the Elo ratings of the different programs. Note that the programs in blue are those created by the authors.<br />
<br />
[[File:curling_ratings.png | 400px]]<br />
<br />
= Critique =<br />
<br />
== Strengths ==<br />
<br />
This algorithm out-performs other high-performance algorithms (including past competition champions).<br />
<br />
I think the paper does a decent job of comparing the performance of their algorithm to others. They are able to clearly show the benefits of many of their additions.<br />
<br />
The authors do seem to be able to adapt strategies similar to those used in Go and other games to the continuous action-space domain. In addition, the final strategy needs no hand-crafted features for learning.<br />
<br />
== Weaknesses ==<br />
<br />
I found this paper difficult to follow at times. One problem was that the algorithms were introduced first, and then how they were used was described. So when the paper stated that self-play shots were taken after 400 simulations, it seemed unclear what simulations were being run, and what stage of the algorithm (ex: MCTS simulations, simulations sped up by using the value network, full simulations on the curling simulator). In particular, both the MCTS statistics and policy-value network could be used to estimate both action probabilities and state values, so it is difficult to tell which is used in which case.<br />
<br />
While I think the comparing of different algorithms was done well, I believe it still lacked some good detail. There were one-off mentions in the paper which would have been nice to see as results. These include the statement that having a policy-value network in place of two networks lead to better performance.<br />
<br />
At this point, the algorithms used still rely on initialisation by a pre-made program.<br />
<br />
There was little theoretical development or justification done in this paper.<br />
<br />
While curling is an interesting choice for demonstrating the algorithm, the fact that the simulations used did not support many of the key points of curling (ice conditions, sweeping) seems very limiting. Another game, such as pool, would likely have offered some of the same challenges, but offered more high-fidelity simulations/training.<br />
<br />
While the spatial placements of stones was discretized in a grid, the curl of thrown stones was discretized to only +/-1. This seems like it may limit learning high- and low-spin moves. It should be noted that having zero spin is not commonly used, to the best of my knowledge.<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/<br />
# Silver, D., Huang, A., Maddison, C., Guez, A., Sifre, L.,Van Den Driessche, G., Schrittwieser, J., Antonoglou, I.,Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe,D., Nham, J., Kalchbrenner, N.,Sutskever, I., Lillicrap, T.,Leach, M., Kavukcuoglu, K., Graepel, T., and Hassabis,D. Mastering the game of go with deep neural networksand tree search. Nature, pp. 484–489, 2016.<br />
# Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou,I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L.,van den Driessche, G., Graepel, T., and Hassabis, D.Mastering the game of go without human knowledge.Nature, pp. 354–359, 2017.<br />
# Yamamoto, M., Kato, S., and Iizuka, H. Digital curling strategy based on game tree search. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 474–480, 2015.<br />
# Ohto, K. and Tanaka, T. A curling agent based on the montecarlo tree search considering the similarity of the best action among similar states. In Proceedings of Advances in Computer Games, ACG, pp. 151–164, 2017.<br />
# Ito, T. and Kitasei, Y. Proposal and implementation of digital curling. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 469–473, 2015.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39526Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-17T01:01:24Z<p>Wfisher: /* Strengths */</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, [5]) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, [6]) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, [7]) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail, balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or even a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Variable Definitions ==<br />
<br />
The following variables are used often in the paper:<br />
<br />
* <math>s</math>: A state in the game, as described below as the input to the network.<br />
* <math>s_t</math>: The state at a certain time-step of the game. Time-steps refer to full turns in the game<br />
* <math>a_t</math>: The action taken in state <math>s_t</math><br />
* <math>A_t</math>: The actions taken for sibling nodes related to <math>a_t</math> in MCTS<br />
* <math>n_{a_t}</math>: The number of visits to node a in MCTS<br />
* <math>v_{a_t}</math>: The MCTS value estimate of a node<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 1 convolutional layer followed by 9 residual blocks, each block consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is distribution of probabilities of the actions to select the best shot out of 32x32x2 set of actions. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games). These are likely estimated rather than simply taken from the MCTS statistics to help account for the differences in a continuous action space.<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times.<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 ([8]). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network ===<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, <math>(s_t, a_t)</math>, a state 'd' steps ahead was generated, <math>s_{t+d}</math>. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, and the value is used for learning the value at ''s<sub>t</sub>''. The authors do not specify, but this is likely done through back-propogating the estimated value, and then training the value network to learn the updated score.<br />
<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
After initialisation by supervised learning, the algorithm uses self-play to further train itself. During this training, the policy network learns probabilities from the MCTS process, while the value network learns from game outcomes.<br />
<br />
At a game state ''s<sub>t</sub>'':<br />
<br />
1) the algorithm outputs a prediction ''z<sub>t</sub>''. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.<br />
<br />
2) the algorithm outputs a prediction <math>\pi_t</math>, representing a probability distribution of actions. These are proportional to estimated visit counts from MCTS, based on kernel density estimation.<br />
<br />
It is not clear how these predictions are created. It would seem likely that the policy-value network generates these, but the wording of the paper suggests they are generated from MCTS statistics.<br />
<br />
The policy-value network is updated by sampling data <math>(s, \pi, z)</math> from recent history of self-play. The same loss function is used as before.<br />
<br />
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.<br />
<br />
== Long-Term Strategy Learning ==<br />
<br />
Finally, the authors implement a new strategy to augment their algorithm for long-term play. In this context, this refers to playing a game over many ends, where the strategy to win a single end may not be a good strategy to win a full game. For example, scoring one point in an end, while being one point ahead, gives the advantage to the other team in the next round (as they will throw the last stone). The other team could then use the advantage to score two points, taking the lead.<br />
<br />
The authors build a 'winning percentage' table. This table stores the percentage of games won, based on number of ends left, and difference in score (current team - opposing team). This can be computed iteratively, and using the probability distribution estimation of one-end scores.<br />
<br />
== Final Algorithms ==<br />
<br />
The authors make use of the following versions of their algorithm:<br />
<br />
=== KR-DL ===<br />
<br />
''Kernel regression-deep learning'': This algorithm is trained only by supervised learning.<br />
<br />
=== KR-DRL ===<br />
<br />
''Kernel regression-deep reinforcement learning'': This algorithm is trained by supervised learning (ie: initialised as the KR-DL algorithm), and again on self-play. During self-play, each shot is selected after 400 MCTS simulations of k=20 randomly selected actions. Data for self-play was collected over a week on 5 GPUS, and generated 5 million game positions. The policy-value network was continually updated using samples from the latest 1 million game positions.<br />
<br />
=== KR-DRL-MES ===<br />
<br />
''Kernel regression-deep reinforcement learning-multi-ends-strategy'': This algorithm makes use of the winning percentage table generated from self-play.<br />
<br />
= Testing and Results =<br />
<br />
Testing is done with a simulated curling program [9]. This simulator does not deal with changing ice conditions, or sweeping, but does deal with stone trajectories and collisions.<br />
<br />
== Comparison of KR-DL-UCT and DL-UCT ==<br />
<br />
The first test compares an algorithm trained with kernel regression with an algorithm trained without kernel regression, to show the contribution that kernel regression adds to the performance. Both algorithms have networks initialised with the supervised learning, and then trained with two different algorithms for self-play. KR-DL-UCT uses the algorithm described above. The authors do not go into detail on how DL-UCT selects shots, but state that a constant is set to allow exploration.<br />
<br />
As an evaluation, both algorithms play 2000 games against the DL-UCT algorithm, which is frozen after supervised training. 1000 games are played with the algorithm taking the first, and 100 taking the 2nd, shots. The games were two-end games. The figure below shows each algorithm's wining percentage given different amounts of training data. While the DL-UCT outperforms the supervised-training-only-DL-UCT algorithm, the KR-DL-UCT algorithm performs much better.<br />
<br />
[[File:curling_KR_test.png | 400px]]<br />
<br />
== Matches ==<br />
<br />
Finally, to test the performance of their multiple algorithms, the authors run matches between their algorithms and other existing programs. Each algorithm plays 200 matches against each other program, 100 of which are played as the first-playing team, and 100 as the second-playing team. Only 1 program was able to out-perform the KR-DRL algorithm. The authors state that this program, ''JiritsukunGAT'17'' also uses a deep network, and hand-crafted features. However, the KR-DRL-MES algorithm was still able to out-perform this. The figure below shows the Elo ratings of the different programs. Note that the programs in blue are those created by the authors.<br />
<br />
[[File:curling_ratings.png | 400px]]<br />
<br />
= Critique =<br />
<br />
== Strengths ==<br />
<br />
This algorithm out-performs other high-performance algorithms (including past competition champions).<br />
<br />
I think the paper does a decent job of comparing the performance of their algorithm to others. They are able to clearly show the benefits of many of their additions.<br />
<br />
The authors do seem to be able to adapt strategies similar to those used in Go and other games to the continuous action-space domain. In addition, the final strategy needs no hand-crafted features for learning.<br />
<br />
== Weaknesses ==<br />
<br />
I found this paper difficult to follow at times. One problem was that the algorithms were introduced first, and then how they were used was described. So when the paper stated that self-play shots were taken after 400 simulations, it seemed unclear what simulations were being run, and what stage of the algorithm (ex: MCTS simulations, simulations sped up by using the value network, full simulations on the curling simulator). In particular, both the MCTS statistics and policy-value network could be used to estimate both action probabilities and state values, so it is difficult to tell which is used in which case.<br />
<br />
While I think the comparing of different algorithms was done well, I believe it still lacked some good detail. There were one-off mentions in the paper which would have been nice to see as results. These include the statement that having a policy-value network in place of two networks lead to better performance.<br />
<br />
At this point, the algorithms used still rely on initialisation by a pre-made program.<br />
<br />
There was little theoretical development or justification done in this paper.<br />
<br />
While curling is an interesting choice for demonstrating the algorithm, the fact that the simulations used did not support many of the key points of curling (ice conditions, sweeping) seems very limiting. Another game, such as pool, would likely have offered some of the same challenges, but offered more high-fidelity simulations/training.<br />
<br />
While the spatial placements of stones was discretized in a grid, the curl of thrown stones was discretized to only +/-1. This seems like it may limit learning high- and low-spin moves. It should be noted that having zero spin is not commonly used, to the best of my knowledge.<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/<br />
# Silver, D., Huang, A., Maddison, C., Guez, A., Sifre, L.,Van Den Driessche, G., Schrittwieser, J., Antonoglou, I.,Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe,D., Nham, J., Kalchbrenner, N.,Sutskever, I., Lillicrap, T.,Leach, M., Kavukcuoglu, K., Graepel, T., and Hassabis,D. Mastering the game of go with deep neural networksand tree search. Nature, pp. 484–489, 2016.<br />
# Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou,I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L.,van den Driessche, G., Graepel, T., and Hassabis, D.Mastering the game of go without human knowledge.Nature, pp. 354–359, 2017.<br />
# Yamamoto, M., Kato, S., and Iizuka, H. Digital curling strategy based on game tree search. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 474–480, 2015.<br />
# Ohto, K. and Tanaka, T. A curling agent based on the montecarlo tree search considering the similarity of the best action among similar states. In Proceedings of Advances in Computer Games, ACG, pp. 151–164, 2017.<br />
# Ito, T. and Kitasei, Y. Proposal and implementation of digital curling. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 469–473, 2015.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39525Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-17T00:57:49Z<p>Wfisher: /* Weaknesses */</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, [5]) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, [6]) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, [7]) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail, balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or even a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Variable Definitions ==<br />
<br />
The following variables are used often in the paper:<br />
<br />
* <math>s</math>: A state in the game, as described below as the input to the network.<br />
* <math>s_t</math>: The state at a certain time-step of the game. Time-steps refer to full turns in the game<br />
* <math>a_t</math>: The action taken in state <math>s_t</math><br />
* <math>A_t</math>: The actions taken for sibling nodes related to <math>a_t</math> in MCTS<br />
* <math>n_{a_t}</math>: The number of visits to node a in MCTS<br />
* <math>v_{a_t}</math>: The MCTS value estimate of a node<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 1 convolutional layer followed by 9 residual blocks, each block consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is distribution of probabilities of the actions to select the best shot out of 32x32x2 set of actions. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games). These are likely estimated rather than simply taken from the MCTS statistics to help account for the differences in a continuous action space.<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times.<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 ([8]). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network ===<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, <math>(s_t, a_t)</math>, a state 'd' steps ahead was generated, <math>s_{t+d}</math>. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, and the value is used for learning the value at ''s<sub>t</sub>''. The authors do not specify, but this is likely done through back-propogating the estimated value, and then training the value network to learn the updated score.<br />
<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
After initialisation by supervised learning, the algorithm uses self-play to further train itself. During this training, the policy network learns probabilities from the MCTS process, while the value network learns from game outcomes.<br />
<br />
At a game state ''s<sub>t</sub>'':<br />
<br />
1) the algorithm outputs a prediction ''z<sub>t</sub>''. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.<br />
<br />
2) the algorithm outputs a prediction <math>\pi_t</math>, representing a probability distribution of actions. These are proportional to estimated visit counts from MCTS, based on kernel density estimation.<br />
<br />
It is not clear how these predictions are created. It would seem likely that the policy-value network generates these, but the wording of the paper suggests they are generated from MCTS statistics.<br />
<br />
The policy-value network is updated by sampling data <math>(s, \pi, z)</math> from recent history of self-play. The same loss function is used as before.<br />
<br />
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.<br />
<br />
== Long-Term Strategy Learning ==<br />
<br />
Finally, the authors implement a new strategy to augment their algorithm for long-term play. In this context, this refers to playing a game over many ends, where the strategy to win a single end may not be a good strategy to win a full game. For example, scoring one point in an end, while being one point ahead, gives the advantage to the other team in the next round (as they will throw the last stone). The other team could then use the advantage to score two points, taking the lead.<br />
<br />
The authors build a 'winning percentage' table. This table stores the percentage of games won, based on number of ends left, and difference in score (current team - opposing team). This can be computed iteratively, and using the probability distribution estimation of one-end scores.<br />
<br />
== Final Algorithms ==<br />
<br />
The authors make use of the following versions of their algorithm:<br />
<br />
=== KR-DL ===<br />
<br />
''Kernel regression-deep learning'': This algorithm is trained only by supervised learning.<br />
<br />
=== KR-DRL ===<br />
<br />
''Kernel regression-deep reinforcement learning'': This algorithm is trained by supervised learning (ie: initialised as the KR-DL algorithm), and again on self-play. During self-play, each shot is selected after 400 MCTS simulations of k=20 randomly selected actions. Data for self-play was collected over a week on 5 GPUS, and generated 5 million game positions. The policy-value network was continually updated using samples from the latest 1 million game positions.<br />
<br />
=== KR-DRL-MES ===<br />
<br />
''Kernel regression-deep reinforcement learning-multi-ends-strategy'': This algorithm makes use of the winning percentage table generated from self-play.<br />
<br />
= Testing and Results =<br />
<br />
Testing is done with a simulated curling program [9]. This simulator does not deal with changing ice conditions, or sweeping, but does deal with stone trajectories and collisions.<br />
<br />
== Comparison of KR-DL-UCT and DL-UCT ==<br />
<br />
The first test compares an algorithm trained with kernel regression with an algorithm trained without kernel regression, to show the contribution that kernel regression adds to the performance. Both algorithms have networks initialised with the supervised learning, and then trained with two different algorithms for self-play. KR-DL-UCT uses the algorithm described above. The authors do not go into detail on how DL-UCT selects shots, but state that a constant is set to allow exploration.<br />
<br />
As an evaluation, both algorithms play 2000 games against the DL-UCT algorithm, which is frozen after supervised training. 1000 games are played with the algorithm taking the first, and 100 taking the 2nd, shots. The games were two-end games. The figure below shows each algorithm's wining percentage given different amounts of training data. While the DL-UCT outperforms the supervised-training-only-DL-UCT algorithm, the KR-DL-UCT algorithm performs much better.<br />
<br />
[[File:curling_KR_test.png | 400px]]<br />
<br />
== Matches ==<br />
<br />
Finally, to test the performance of their multiple algorithms, the authors run matches between their algorithms and other existing programs. Each algorithm plays 200 matches against each other program, 100 of which are played as the first-playing team, and 100 as the second-playing team. Only 1 program was able to out-perform the KR-DRL algorithm. The authors state that this program, ''JiritsukunGAT'17'' also uses a deep network, and hand-crafted features. However, the KR-DRL-MES algorithm was still able to out-perform this. The figure below shows the Elo ratings of the different programs. Note that the programs in blue are those created by the authors.<br />
<br />
[[File:curling_ratings.png | 400px]]<br />
<br />
= Critique =<br />
<br />
== Strengths ==<br />
<br />
I think the paper does a decent job of comparing the performance of their algorithm to others. They are able to clearly show the benefits of many of their additions.<br />
<br />
The authors do seem to be able to adapt strategies similar to those used in Go and other games to the continuous action-space domain.<br />
<br />
== Weaknesses ==<br />
<br />
I found this paper difficult to follow at times. One problem was that the algorithms were introduced first, and then how they were used was described. So when the paper stated that self-play shots were taken after 400 simulations, it seemed unclear what simulations were being run, and what stage of the algorithm (ex: MCTS simulations, simulations sped up by using the value network, full simulations on the curling simulator). In particular, both the MCTS statistics and policy-value network could be used to estimate both action probabilities and state values, so it is difficult to tell which is used in which case.<br />
<br />
While I think the comparing of different algorithms was done well, I believe it still lacked some good detail. There were one-off mentions in the paper which would have been nice to see as results. These include the statement that having a policy-value network in place of two networks lead to better performance.<br />
<br />
At this point, the algorithms used still rely on initialisation by a pre-made program.<br />
<br />
There was little theoretical development or justification done in this paper.<br />
<br />
While curling is an interesting choice for demonstrating the algorithm, the fact that the simulations used did not support many of the key points of curling (ice conditions, sweeping) seems very limiting. Another game, such as pool, would likely have offered some of the same challenges, but offered more high-fidelity simulations/training.<br />
<br />
While the spatial placements of stones was discretized in a grid, the curl of thrown stones was discretized to only +/-1. This seems like it may limit learning high- and low-spin moves. It should be noted that having zero spin is not commonly used, to the best of my knowledge.<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/<br />
# Silver, D., Huang, A., Maddison, C., Guez, A., Sifre, L.,Van Den Driessche, G., Schrittwieser, J., Antonoglou, I.,Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe,D., Nham, J., Kalchbrenner, N.,Sutskever, I., Lillicrap, T.,Leach, M., Kavukcuoglu, K., Graepel, T., and Hassabis,D. Mastering the game of go with deep neural networksand tree search. Nature, pp. 484–489, 2016.<br />
# Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou,I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L.,van den Driessche, G., Graepel, T., and Hassabis, D.Mastering the game of go without human knowledge.Nature, pp. 354–359, 2017.<br />
# Yamamoto, M., Kato, S., and Iizuka, H. Digital curling strategy based on game tree search. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 474–480, 2015.<br />
# Ohto, K. and Tanaka, T. A curling agent based on the montecarlo tree search considering the similarity of the best action among similar states. In Proceedings of Advances in Computer Games, ACG, pp. 151–164, 2017.<br />
# Ito, T. and Kitasei, Y. Proposal and implementation of digital curling. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 469–473, 2015.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39524Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-17T00:55:52Z<p>Wfisher: Add details</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, [5]) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, [6]) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, [7]) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail, balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or even a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Variable Definitions ==<br />
<br />
The following variables are used often in the paper:<br />
<br />
* <math>s</math>: A state in the game, as described below as the input to the network.<br />
* <math>s_t</math>: The state at a certain time-step of the game. Time-steps refer to full turns in the game<br />
* <math>a_t</math>: The action taken in state <math>s_t</math><br />
* <math>A_t</math>: The actions taken for sibling nodes related to <math>a_t</math> in MCTS<br />
* <math>n_{a_t}</math>: The number of visits to node a in MCTS<br />
* <math>v_{a_t}</math>: The MCTS value estimate of a node<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 1 convolutional layer followed by 9 residual blocks, each block consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is distribution of probabilities of the actions to select the best shot out of 32x32x2 set of actions. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games). These are likely estimated rather than simply taken from the MCTS statistics to help account for the differences in a continuous action space.<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times.<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 ([8]). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network ===<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, <math>(s_t, a_t)</math>, a state 'd' steps ahead was generated, <math>s_{t+d}</math>. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, and the value is used for learning the value at ''s<sub>t</sub>''. The authors do not specify, but this is likely done through back-propogating the estimated value, and then training the value network to learn the updated score.<br />
<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
After initialisation by supervised learning, the algorithm uses self-play to further train itself. During this training, the policy network learns probabilities from the MCTS process, while the value network learns from game outcomes.<br />
<br />
At a game state ''s<sub>t</sub>'':<br />
<br />
1) the algorithm outputs a prediction ''z<sub>t</sub>''. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.<br />
<br />
2) the algorithm outputs a prediction <math>\pi_t</math>, representing a probability distribution of actions. These are proportional to estimated visit counts from MCTS, based on kernel density estimation.<br />
<br />
It is not clear how these predictions are created. It would seem likely that the policy-value network generates these, but the wording of the paper suggests they are generated from MCTS statistics.<br />
<br />
The policy-value network is updated by sampling data <math>(s, \pi, z)</math> from recent history of self-play. The same loss function is used as before.<br />
<br />
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.<br />
<br />
== Long-Term Strategy Learning ==<br />
<br />
Finally, the authors implement a new strategy to augment their algorithm for long-term play. In this context, this refers to playing a game over many ends, where the strategy to win a single end may not be a good strategy to win a full game. For example, scoring one point in an end, while being one point ahead, gives the advantage to the other team in the next round (as they will throw the last stone). The other team could then use the advantage to score two points, taking the lead.<br />
<br />
The authors build a 'winning percentage' table. This table stores the percentage of games won, based on number of ends left, and difference in score (current team - opposing team). This can be computed iteratively, and using the probability distribution estimation of one-end scores.<br />
<br />
== Final Algorithms ==<br />
<br />
The authors make use of the following versions of their algorithm:<br />
<br />
=== KR-DL ===<br />
<br />
''Kernel regression-deep learning'': This algorithm is trained only by supervised learning.<br />
<br />
=== KR-DRL ===<br />
<br />
''Kernel regression-deep reinforcement learning'': This algorithm is trained by supervised learning (ie: initialised as the KR-DL algorithm), and again on self-play. During self-play, each shot is selected after 400 MCTS simulations of k=20 randomly selected actions. Data for self-play was collected over a week on 5 GPUS, and generated 5 million game positions. The policy-value network was continually updated using samples from the latest 1 million game positions.<br />
<br />
=== KR-DRL-MES ===<br />
<br />
''Kernel regression-deep reinforcement learning-multi-ends-strategy'': This algorithm makes use of the winning percentage table generated from self-play.<br />
<br />
= Testing and Results =<br />
<br />
Testing is done with a simulated curling program [9]. This simulator does not deal with changing ice conditions, or sweeping, but does deal with stone trajectories and collisions.<br />
<br />
== Comparison of KR-DL-UCT and DL-UCT ==<br />
<br />
The first test compares an algorithm trained with kernel regression with an algorithm trained without kernel regression, to show the contribution that kernel regression adds to the performance. Both algorithms have networks initialised with the supervised learning, and then trained with two different algorithms for self-play. KR-DL-UCT uses the algorithm described above. The authors do not go into detail on how DL-UCT selects shots, but state that a constant is set to allow exploration.<br />
<br />
As an evaluation, both algorithms play 2000 games against the DL-UCT algorithm, which is frozen after supervised training. 1000 games are played with the algorithm taking the first, and 100 taking the 2nd, shots. The games were two-end games. The figure below shows each algorithm's wining percentage given different amounts of training data. While the DL-UCT outperforms the supervised-training-only-DL-UCT algorithm, the KR-DL-UCT algorithm performs much better.<br />
<br />
[[File:curling_KR_test.png | 400px]]<br />
<br />
== Matches ==<br />
<br />
Finally, to test the performance of their multiple algorithms, the authors run matches between their algorithms and other existing programs. Each algorithm plays 200 matches against each other program, 100 of which are played as the first-playing team, and 100 as the second-playing team. Only 1 program was able to out-perform the KR-DRL algorithm. The authors state that this program, ''JiritsukunGAT'17'' also uses a deep network, and hand-crafted features. However, the KR-DRL-MES algorithm was still able to out-perform this. The figure below shows the Elo ratings of the different programs. Note that the programs in blue are those created by the authors.<br />
<br />
[[File:curling_ratings.png | 400px]]<br />
<br />
= Critique =<br />
<br />
== Strengths ==<br />
<br />
I think the paper does a decent job of comparing the performance of their algorithm to others. They are able to clearly show the benefits of many of their additions.<br />
<br />
The authors do seem to be able to adapt strategies similar to those used in Go and other games to the continuous action-space domain.<br />
<br />
== Weaknesses ==<br />
<br />
I found this paper difficult to follow at times. One problem was that the algorithms were introduced first, and then how they were used was described. So when the paper stated that self-play shots were taken after 400 simulations, it seemed unclear what simulations were being run, and what stage of the algorithm (ex: MCTS simulations, simulations sped up by using the value network, full simulations on the curling simulator).<br />
<br />
While I think the comparing of different algorithms was done well, I believe it still lacked some good detail. There were one-off mentions in the paper which would have been nice to see as results. These include the statement that having a policy-value network in place of two networks lead to better performance.<br />
<br />
At this point, the algorithms used still rely on initialisation by a pre-made program.<br />
<br />
There was little theoretical development or justification done in this paper.<br />
<br />
While curling is an interesting choice for demonstrating the algorithm, the fact that the simulations used did not support many of the key points of curling (ice conditions, sweeping) seems very limiting. Another game, such as pool, would likely have offered some of the same challenges, but offered more high-fidelity simulations/training.<br />
<br />
While the spatial placements of stones was discretized in a grid, the curl of thrown stones was discretized to only +/-1. This seems like it may limit learning high- and low-spin moves. It should be noted that having zero spin is not commonly used, to the best of my knowledge.<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/<br />
# Silver, D., Huang, A., Maddison, C., Guez, A., Sifre, L.,Van Den Driessche, G., Schrittwieser, J., Antonoglou, I.,Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe,D., Nham, J., Kalchbrenner, N.,Sutskever, I., Lillicrap, T.,Leach, M., Kavukcuoglu, K., Graepel, T., and Hassabis,D. Mastering the game of go with deep neural networksand tree search. Nature, pp. 484–489, 2016.<br />
# Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou,I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L.,van den Driessche, G., Graepel, T., and Hassabis, D.Mastering the game of go without human knowledge.Nature, pp. 354–359, 2017.<br />
# Yamamoto, M., Kato, S., and Iizuka, H. Digital curling strategy based on game tree search. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 474–480, 2015.<br />
# Ohto, K. and Tanaka, T. A curling agent based on the montecarlo tree search considering the similarity of the best action among similar states. In Proceedings of Advances in Computer Games, ACG, pp. 151–164, 2017.<br />
# Ito, T. and Kitasei, Y. Proposal and implementation of digital curling. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 469–473, 2015.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39523Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-17T00:35:14Z<p>Wfisher: Reference curling program</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, [5]) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, [6]) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, [7]) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail, balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or even a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 1 convolutional layer followed by 9 residual blocks, each block consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is distribution of probabilities of the actions to select the best shot out of 32x32x2 set of actions. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games):<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times.<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 ([8]). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network ===<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, ''(s<sub>t</sub>,a<sub>t</sub>)'', a state 'd' steps ahead was generated, ''s<sub>t+d</sub>''. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, and the value is used for learning the value at ''s<sub>t</sub>''.<br />
<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
After initialisation by supervised learning, the algorithm uses self-play to further train itself. During this training, the policy network learns probabilities from the MCTS process, while the value network learns from game outcomes.<br />
<br />
At a game state ''s<sub>t</sub>'':<br />
<br />
1) the algorithm outputs a prediction ''z<sub>t</sub>''. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.<br />
<br />
2) the algorithm outputs a prediction <math>\pi_t</math>, representing a probability distribution of actions. These are proportional to estimated visit counts from MCTS, based on kernel density estimation<br />
<br />
The policy-value network is updated by sampling data <math>(s, \pi, z)</math> from recent history of self-play. The same loss function is used as before.<br />
<br />
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.<br />
<br />
== Long-Term Strategy Learning ==<br />
<br />
Finally, the authors implement a new strategy to augment their algorithm for long-term play. In this context, this refers to playing a game over many ends, where the strategy to win a single end may not be a good strategy to win a full game. For example, scoring one point in an end, while being one point ahead, gives the advantage to the other team in the next round (as they will throw the last stone). The other team could then use the advantage to score two points, taking the lead.<br />
<br />
The authors build a 'winning percentage' table. This table stores the percentage of games won, based on number of ends left, and difference in score (current team - opposing team). This can be computed iteratively, and using the probability distribution estimation of one-end scores.<br />
<br />
== Final Algorithms ==<br />
<br />
The authors make use of the following versions of their algorithm:<br />
<br />
=== KR-DL ===<br />
<br />
''Kernel regression-deep learning'': This algorithm is trained only by supervised learning.<br />
<br />
=== KR-DRL ===<br />
<br />
''Kernel regression-deep reinforcement learning'': This algorithm is trained by supervised learning (ie: initialised as the KR-DL algorithm), and again on self-play. During self-play, each shot is selected after 400 MCTS simulations of k=20 randomly selected actions. Data for self-play was collected over a week on 5 GPUS, and generated 5 million game positions. The policy-value network was continually updated using samples from the latest 1 million game positions.<br />
<br />
=== KR-DRL-MES ===<br />
<br />
''Kernel regression-deep reinforcement learning-multi-ends-strategy'': This algorithm makes use of the winning percentage table generated from self-play.<br />
<br />
= Testing and Results =<br />
<br />
Testing is done with a simulated curling program [9]. This simulator does not deal with changing ice conditions, or sweeping, but does deal with stone trajectories and collisions.<br />
<br />
== Comparison of KR-DL-UCT and DL-UCT ==<br />
<br />
The first test compares an algorithm trained with kernel regression with an algorithm trained without kernel regression, to show the contribution that kernel regression adds to the performance. Both algorithms have networks initialised with the supervised learning, and then trained with two different algorithms for self-play. KR-DL-UCT uses the algorithm described above. The authors do not go into detail on how DL-UCT selects shots, but state that a constant is set to allow exploration.<br />
<br />
As an evaluation, both algorithms play 2000 games against the DL-UCT algorithm, which is frozen after supervised training. 1000 games are played with the algorithm taking the first, and 100 taking the 2nd, shots. The games were two-end games. The figure below shows each algorithm's wining percentage given different amounts of training data. While the DL-UCT outperforms the supervised-training-only-DL-UCT algorithm, the KR-DL-UCT algorithm performs much better.<br />
<br />
[[File:curling_KR_test.png | 400px]]<br />
<br />
== Matches ==<br />
<br />
Finally, to test the performance of their multiple algorithms, the authors run matches between their algorithms and other existing programs. Each algorithm plays 200 matches against each other program, 100 of which are played as the first-playing team, and 100 as the second-playing team. Only 1 program was able to out-perform the KR-DRL algorithm. The authors state that this program, ''JiritsukunGAT'17'' also uses a deep network, and hand-crafted features. However, the KR-DRL-MES algorithm was still able to out-perform this. The figure below shows the Elo ratings of the different programs. Note that the programs in blue are those created by the authors.<br />
<br />
[[File:curling_ratings.png | 400px]]<br />
<br />
= Critique =<br />
<br />
== Strengths ==<br />
<br />
I think the paper does a decent job of comparing the performance of their algorithm to others. They are able to clearly show the benefits of many of their additions.<br />
<br />
The authors do seem to be able to adapt strategies similar to those used in Go and other games to the continuous action-space domain.<br />
<br />
== Weaknesses ==<br />
<br />
I found this paper difficult to follow at times. One problem was that the algorithms were introduced first, and then how they were used was described. So when the paper stated that self-play shots were taken after 400 simulations, it seemed unclear what simulations were being run, and what stage of the algorithm (ex: MCTS simulations, simulations sped up by using the value network, full simulations on the curling simulator).<br />
<br />
While I think the comparing of different algorithms was done well, I believe it still lacked some good detail. There were one-off mentions in the paper which would have been nice to see as results. These include the statement that having a policy-value network in place of two networks lead to better performance.<br />
<br />
At this point, the algorithms used still rely on initialisation by a pre-made program.<br />
<br />
There was little theoretical development or justification done in this paper.<br />
<br />
While curling is an interesting choice for demonstrating the algorithm, the fact that the simulations used did not support many of the key points of curling (ice conditions, sweeping) seems very limiting. Another game, such as pool, would likely have offered some of the same challenges, but offered more high-fidelity simulations/training.<br />
<br />
While the spatial placements of stones was discretized in a grid, the curl of thrown stones was discretized to only +/-1. This seems like it may limit learning high- and low-spin moves. It should be noted that having zero spin is not commonly used, to the best of my knowledge.<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/<br />
# Silver, D., Huang, A., Maddison, C., Guez, A., Sifre, L.,Van Den Driessche, G., Schrittwieser, J., Antonoglou, I.,Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe,D., Nham, J., Kalchbrenner, N.,Sutskever, I., Lillicrap, T.,Leach, M., Kavukcuoglu, K., Graepel, T., and Hassabis,D. Mastering the game of go with deep neural networksand tree search. Nature, pp. 484–489, 2016.<br />
# Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou,I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L.,van den Driessche, G., Graepel, T., and Hassabis, D.Mastering the game of go without human knowledge.Nature, pp. 354–359, 2017.<br />
# Yamamoto, M., Kato, S., and Iizuka, H. Digital curling strategy based on game tree search. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 474–480, 2015.<br />
# Ohto, K. and Tanaka, T. A curling agent based on the montecarlo tree search considering the similarity of the best action among similar states. In Proceedings of Advances in Computer Games, ACG, pp. 151–164, 2017.<br />
# Ito, T. and Kitasei, Y. Proposal and implementation of digital curling. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 469–473, 2015.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39473Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-16T16:42:27Z<p>Wfisher: /* Weaknesses */</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, [5]) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, [6]) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, [7]) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail, balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or event a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 9 residual blocks, each consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is a 32x32x2 set of action probabilities. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games):<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times<br />
<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 ([8]). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network ===<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, ''(s<sub>t</sub>,a<sub>t</sub>)'', a state 'd' steps ahead was generated, ''s<sub>t+d</sub>''. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, and the value is used for learning the value at ''s<sub>t</sub>''.<br />
<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
After initialisation by supervised learning, the algorithm uses self-play to further train itself. During this training, the policy network learns probabilities from the MCTS process, while the value network learns from game outcomes.<br />
<br />
At a game state ''s<sub>t</sub>'':<br />
<br />
1) the algorithm outputs a prediction ''z<sub>t</sub>''. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.<br />
<br />
2) the algorithm outputs a prediction <math>\pi_t</math>, representing a probability distribution of actions. These are proportional to estimated visit counts from MCTS, based on kernel density estimation<br />
<br />
The policy-value network is updated by sampling data <math>(s, \pi, z)</math> from recent history of self-play. The same loss function is used as before.<br />
<br />
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.<br />
<br />
== Long-Term Strategy Learning ==<br />
<br />
Finally, the authors implement a new strategy to augment their algorithm for long-term play. In this context, this refers to playing a game over many ends, where the strategy to win a single end may not be a good strategy to win a full game. For example, scoring one point in an end, while being one point ahead, gives the advantage to the other team in the next round (as they will throw the last stone). The other team could then use the advantage to score two points, taking the lead.<br />
<br />
The authors build a 'winning percentage' table. This table stores the percentage of games won, based on number of ends left, and difference in score (current team - opposing team). This can be computed iteratively, and using the probability distribution estimation of one-end scores.<br />
<br />
== Final Algorithms ==<br />
<br />
The authors make use of the following versions of their algorithm:<br />
<br />
=== KR-DL ===<br />
<br />
''Kernel regression-deep learning'': This algorithm is trained only by supervised learning.<br />
<br />
=== KR-DRL ===<br />
<br />
''Kernel regression-deep reinforcement learning'': This algorithm is trained by supervised learning (ie: initialised as the KR-DL algorithm), and again on self-play. During self-play, each shot is selected after 400 MCTS simulations of k=20 randomly selected actions. Data for self-play was collected over a week on 5 GPUS, and generated 5 million game positions. The policy-value network was continually updated using samples from the latest 1 million game positions.<br />
<br />
=== KR-DRL-MES ===<br />
<br />
''Kernel regression-deep reinforcement learning-multi-ends-strategy'': This algorithm makes use of the winning percentage table generated from self-play.<br />
<br />
= Testing and Results =<br />
<br />
Testing is done with a simulated curling program (TODO). This simulator does not deal with changing ice conditions, or sweeping, but does deal with stone trajectories and collisions.<br />
<br />
== Comparison of KR-DL-UCT and DL-UCT ==<br />
<br />
The first test compares an algorithm trained with kernel regression with an algorithm trained without kernel regression, to show the contribution that kernel regression adds to the performance. Both algorithms have networks initialised with the supervised learning, and then trained with two different algorithms for self-play. KR-DL-UCT uses the algorithm described above. The authors do not go into detail on how DL-UCT selects shots, but state that a constant is set to allow exploration.<br />
<br />
As an evaluation, both algorithms play 2000 games against the DL-UCT algorithm, which is frozen after supervised training. 1000 games are played with the algorithm taking the first, and 100 taking the 2nd, shots. The games were two-end games. The figure below shows each algorithm's wining percentage given different amounts of training data. While the DL-UCT outperforms the supervised-training-only-DL-UCT algorithm, the KR-DL-UCT algorithm performs much better.<br />
<br />
[[File:curling_KR_test.png | 400px]]<br />
<br />
== Matches ==<br />
<br />
Finally, to test the performance of their multiple algorithms, the authors run matches between their algorithms and other existing programs. Each algorithm plays 200 matches against each other program, 100 of which are played as the first-playing team, and 100 as the second-playing team. Only 1 program was able to out-perform the KR-DRL algorithm. The authors state that this program, ''JiritsukunGAT'17'' also uses a deep network, and hand-crafted features. However, the KR-DRL-MES algorithm was still able to out-perform this. The figure below shows the Elo ratings of the different programs. Note that the programs in blue are those created by the authors.<br />
<br />
[[File:curling_ratings.png | 400px]]<br />
<br />
= Critique =<br />
<br />
== Strengths ==<br />
<br />
I think the paper does a decent job of comparing the performance of their algorithm to others. They are able to clearly show the benefits of many of their additions.<br />
<br />
The authors do seem to be able to adapt strategies similar to those used in Go and other games to the continuous action-space domain.<br />
<br />
== Weaknesses ==<br />
<br />
I found this paper difficult to follow at times. One problem was that the algorithms were introduced first, and then how they were used was described. So when the paper stated that self-play shots were taken after 400 simulations, it seemed unclear what simulations were being run, and what stage of the algorithm (ex: MCTS simulations, simulations sped up by using the value network, full simulations on the curling simulator).<br />
<br />
While I think the comparing of different algorithms was done well, I believe it still lacked some good detail. There were one-off mentions in the paper which would have been nice to see as results. These include the statement that having a policy-value network in place of two networks lead to better performance.<br />
<br />
At this point, the algorithms used still rely on initialisation by a pre-made program.<br />
<br />
There was little theoretical development or justification done in this paper.<br />
<br />
While curling is an interesting choice for demonstrating the algorithm, the fact that the simulations used did not support many of the key points of curling (ice conditions, sweeping) seems very limiting. Another game, such as pool, would likely have offered some of the same challenges, but offered more high-fidelity simulations/training.<br />
<br />
While the spatial placements of stones was discretized in a grid, the curl of thrown stones was discretized to only +/-1. This seems like it may limit learning high- and low-spin moves. It should be noted that having zero spin is not commonly used, to the best of my knowledge.<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/<br />
# Silver, D., Huang, A., Maddison, C., Guez, A., Sifre, L.,Van Den Driessche, G., Schrittwieser, J., Antonoglou, I.,Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe,D., Nham, J., Kalchbrenner, N.,Sutskever, I., Lillicrap, T.,Leach, M., Kavukcuoglu, K., Graepel, T., and Hassabis,D. Mastering the game of go with deep neural networksand tree search. Nature, pp. 484–489, 2016.<br />
# Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou,I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L.,van den Driessche, G., Graepel, T., and Hassabis, D.Mastering the game of go without human knowledge.Nature, pp. 354–359, 2017.<br />
# Yamamoto, M., Kato, S., and Iizuka, H. Digital curling strategy based on game tree search. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 474–480, 2015.<br />
# Ohto, K. and Tanaka, T. A curling agent based on the montecarlo tree search considering the similarity of the best action among similar states. In Proceedings of Advances in Computer Games, ACG, pp. 151–164, 2017.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39472Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-16T16:40:24Z<p>Wfisher: /* Weaknesses */</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, [5]) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, [6]) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, [7]) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail, balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or event a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 9 residual blocks, each consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is a 32x32x2 set of action probabilities. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games):<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times<br />
<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 ([8]). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network ===<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, ''(s<sub>t</sub>,a<sub>t</sub>)'', a state 'd' steps ahead was generated, ''s<sub>t+d</sub>''. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, and the value is used for learning the value at ''s<sub>t</sub>''.<br />
<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
After initialisation by supervised learning, the algorithm uses self-play to further train itself. During this training, the policy network learns probabilities from the MCTS process, while the value network learns from game outcomes.<br />
<br />
At a game state ''s<sub>t</sub>'':<br />
<br />
1) the algorithm outputs a prediction ''z<sub>t</sub>''. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.<br />
<br />
2) the algorithm outputs a prediction <math>\pi_t</math>, representing a probability distribution of actions. These are proportional to estimated visit counts from MCTS, based on kernel density estimation<br />
<br />
The policy-value network is updated by sampling data <math>(s, \pi, z)</math> from recent history of self-play. The same loss function is used as before.<br />
<br />
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.<br />
<br />
== Long-Term Strategy Learning ==<br />
<br />
Finally, the authors implement a new strategy to augment their algorithm for long-term play. In this context, this refers to playing a game over many ends, where the strategy to win a single end may not be a good strategy to win a full game. For example, scoring one point in an end, while being one point ahead, gives the advantage to the other team in the next round (as they will throw the last stone). The other team could then use the advantage to score two points, taking the lead.<br />
<br />
The authors build a 'winning percentage' table. This table stores the percentage of games won, based on number of ends left, and difference in score (current team - opposing team). This can be computed iteratively, and using the probability distribution estimation of one-end scores.<br />
<br />
== Final Algorithms ==<br />
<br />
The authors make use of the following versions of their algorithm:<br />
<br />
=== KR-DL ===<br />
<br />
''Kernel regression-deep learning'': This algorithm is trained only by supervised learning.<br />
<br />
=== KR-DRL ===<br />
<br />
''Kernel regression-deep reinforcement learning'': This algorithm is trained by supervised learning (ie: initialised as the KR-DL algorithm), and again on self-play. During self-play, each shot is selected after 400 MCTS simulations of k=20 randomly selected actions. Data for self-play was collected over a week on 5 GPUS, and generated 5 million game positions. The policy-value network was continually updated using samples from the latest 1 million game positions.<br />
<br />
=== KR-DRL-MES ===<br />
<br />
''Kernel regression-deep reinforcement learning-multi-ends-strategy'': This algorithm makes use of the winning percentage table generated from self-play.<br />
<br />
= Testing and Results =<br />
<br />
Testing is done with a simulated curling program (TODO). This simulator does not deal with changing ice conditions, or sweeping, but does deal with stone trajectories and collisions.<br />
<br />
== Comparison of KR-DL-UCT and DL-UCT ==<br />
<br />
The first test compares an algorithm trained with kernel regression with an algorithm trained without kernel regression, to show the contribution that kernel regression adds to the performance. Both algorithms have networks initialised with the supervised learning, and then trained with two different algorithms for self-play. KR-DL-UCT uses the algorithm described above. The authors do not go into detail on how DL-UCT selects shots, but state that a constant is set to allow exploration.<br />
<br />
As an evaluation, both algorithms play 2000 games against the DL-UCT algorithm, which is frozen after supervised training. 1000 games are played with the algorithm taking the first, and 100 taking the 2nd, shots. The games were two-end games. The figure below shows each algorithm's wining percentage given different amounts of training data. While the DL-UCT outperforms the supervised-training-only-DL-UCT algorithm, the KR-DL-UCT algorithm performs much better.<br />
<br />
[[File:curling_KR_test.png | 400px]]<br />
<br />
== Matches ==<br />
<br />
Finally, to test the performance of their multiple algorithms, the authors run matches between their algorithms and other existing programs. Each algorithm plays 200 matches against each other program, 100 of which are played as the first-playing team, and 100 as the second-playing team. Only 1 program was able to out-perform the KR-DRL algorithm. The authors state that this program, ''JiritsukunGAT'17'' also uses a deep network, and hand-crafted features. However, the KR-DRL-MES algorithm was still able to out-perform this. The figure below shows the Elo ratings of the different programs. Note that the programs in blue are those created by the authors.<br />
<br />
[[File:curling_ratings.png | 400px]]<br />
<br />
= Critique =<br />
<br />
== Strengths ==<br />
<br />
I think the paper does a decent job of comparing the performance of their algorithm to others. They are able to clearly show the benefits of many of their additions.<br />
<br />
The authors do seem to be able to adapt strategies similar to those used in Go and other games to the continuous action-space domain.<br />
<br />
== Weaknesses ==<br />
<br />
I found this paper difficult to follow at times. One problem was that the algorithms were introduced first, and then how they were used was described. So when the paper stated that self-play shots were taken after 400 simulations, it seemed unclear what simulations were being run, and what stage of the algorithm (ex: MCTS simulations, simulations sped up by using the value network, full simulations on the curling simulator).<br />
<br />
While I think the comparing of different algorithms was done well, I believe it still lacked some good detail. There were one-off mentions in the paper which would have been nice to see as results. These include the statement that having a policy-value network in place of two networks lead to better performance.<br />
<br />
At this point, the algorithms used still rely on initialisation by a pre-made program.<br />
<br />
There was little theoretical development or justification done in this paper.<br />
<br />
While curling is an interesting choice for demonstrating the algorithm, the fact that the simulations used did not support many of the key points of curling (ice conditions, sweeping) seems very limiting. Another game, such as pool, would likely have offered some of the same challenges, but offered more high-fidelity simulations/training.<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/<br />
# Silver, D., Huang, A., Maddison, C., Guez, A., Sifre, L.,Van Den Driessche, G., Schrittwieser, J., Antonoglou, I.,Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe,D., Nham, J., Kalchbrenner, N.,Sutskever, I., Lillicrap, T.,Leach, M., Kavukcuoglu, K., Graepel, T., and Hassabis,D. Mastering the game of go with deep neural networksand tree search. Nature, pp. 484–489, 2016.<br />
# Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou,I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L.,van den Driessche, G., Graepel, T., and Hassabis, D.Mastering the game of go without human knowledge.Nature, pp. 354–359, 2017.<br />
# Yamamoto, M., Kato, S., and Iizuka, H. Digital curling strategy based on game tree search. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 474–480, 2015.<br />
# Ohto, K. and Tanaka, T. A curling agent based on the montecarlo tree search considering the similarity of the best action among similar states. In Proceedings of Advances in Computer Games, ACG, pp. 151–164, 2017.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39471Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-16T16:37:22Z<p>Wfisher: Add a critique</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, [5]) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, [6]) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, [7]) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail, balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or event a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 9 residual blocks, each consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is a 32x32x2 set of action probabilities. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games):<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times<br />
<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 ([8]). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network ===<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, ''(s<sub>t</sub>,a<sub>t</sub>)'', a state 'd' steps ahead was generated, ''s<sub>t+d</sub>''. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, and the value is used for learning the value at ''s<sub>t</sub>''.<br />
<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
After initialisation by supervised learning, the algorithm uses self-play to further train itself. During this training, the policy network learns probabilities from the MCTS process, while the value network learns from game outcomes.<br />
<br />
At a game state ''s<sub>t</sub>'':<br />
<br />
1) the algorithm outputs a prediction ''z<sub>t</sub>''. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.<br />
<br />
2) the algorithm outputs a prediction <math>\pi_t</math>, representing a probability distribution of actions. These are proportional to estimated visit counts from MCTS, based on kernel density estimation<br />
<br />
The policy-value network is updated by sampling data <math>(s, \pi, z)</math> from recent history of self-play. The same loss function is used as before.<br />
<br />
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.<br />
<br />
== Long-Term Strategy Learning ==<br />
<br />
Finally, the authors implement a new strategy to augment their algorithm for long-term play. In this context, this refers to playing a game over many ends, where the strategy to win a single end may not be a good strategy to win a full game. For example, scoring one point in an end, while being one point ahead, gives the advantage to the other team in the next round (as they will throw the last stone). The other team could then use the advantage to score two points, taking the lead.<br />
<br />
The authors build a 'winning percentage' table. This table stores the percentage of games won, based on number of ends left, and difference in score (current team - opposing team). This can be computed iteratively, and using the probability distribution estimation of one-end scores.<br />
<br />
== Final Algorithms ==<br />
<br />
The authors make use of the following versions of their algorithm:<br />
<br />
=== KR-DL ===<br />
<br />
''Kernel regression-deep learning'': This algorithm is trained only by supervised learning.<br />
<br />
=== KR-DRL ===<br />
<br />
''Kernel regression-deep reinforcement learning'': This algorithm is trained by supervised learning (ie: initialised as the KR-DL algorithm), and again on self-play. During self-play, each shot is selected after 400 MCTS simulations of k=20 randomly selected actions. Data for self-play was collected over a week on 5 GPUS, and generated 5 million game positions. The policy-value network was continually updated using samples from the latest 1 million game positions.<br />
<br />
=== KR-DRL-MES ===<br />
<br />
''Kernel regression-deep reinforcement learning-multi-ends-strategy'': This algorithm makes use of the winning percentage table generated from self-play.<br />
<br />
= Testing and Results =<br />
<br />
Testing is done with a simulated curling program (TODO). This simulator does not deal with changing ice conditions, or sweeping, but does deal with stone trajectories and collisions.<br />
<br />
== Comparison of KR-DL-UCT and DL-UCT ==<br />
<br />
The first test compares an algorithm trained with kernel regression with an algorithm trained without kernel regression, to show the contribution that kernel regression adds to the performance. Both algorithms have networks initialised with the supervised learning, and then trained with two different algorithms for self-play. KR-DL-UCT uses the algorithm described above. The authors do not go into detail on how DL-UCT selects shots, but state that a constant is set to allow exploration.<br />
<br />
As an evaluation, both algorithms play 2000 games against the DL-UCT algorithm, which is frozen after supervised training. 1000 games are played with the algorithm taking the first, and 100 taking the 2nd, shots. The games were two-end games. The figure below shows each algorithm's wining percentage given different amounts of training data. While the DL-UCT outperforms the supervised-training-only-DL-UCT algorithm, the KR-DL-UCT algorithm performs much better.<br />
<br />
[[File:curling_KR_test.png | 400px]]<br />
<br />
== Matches ==<br />
<br />
Finally, to test the performance of their multiple algorithms, the authors run matches between their algorithms and other existing programs. Each algorithm plays 200 matches against each other program, 100 of which are played as the first-playing team, and 100 as the second-playing team. Only 1 program was able to out-perform the KR-DRL algorithm. The authors state that this program, ''JiritsukunGAT'17'' also uses a deep network, and hand-crafted features. However, the KR-DRL-MES algorithm was still able to out-perform this. The figure below shows the Elo ratings of the different programs. Note that the programs in blue are those created by the authors.<br />
<br />
[[File:curling_ratings.png | 400px]]<br />
<br />
= Critique =<br />
<br />
== Strengths ==<br />
<br />
I think the paper does a decent job of comparing the performance of their algorithm to others. They are able to clearly show the benefits of many of their additions.<br />
<br />
The authors do seem to be able to adapt strategies similar to those used in Go and other games to the continuous action-space domain.<br />
<br />
== Weaknesses ==<br />
<br />
I found this paper difficult to follow at times. One problem was that the algorithms were introduced first, and then how they were used was described. So when the paper stated that self-play shots were taken after 400 simulations, it seemed unclear what simulations were being run, and what stage of the algorithm (ex: MCTS simulations, simulations sped up by using the value network, full simulations on the curling simulator).<br />
<br />
While I think the comparing of different algorithms was done well, I believe it still lacked some good detail. There were one-off mentions in the paper which would have been nice to see as results. These include the statement that having a policy-value network in place of two networks lead to better performance.<br />
<br />
At this point, the algorithms used still rely on initialisation by a pre-made program.<br />
<br />
There was little theoretical development or justification done in this paper.<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/<br />
# Silver, D., Huang, A., Maddison, C., Guez, A., Sifre, L.,Van Den Driessche, G., Schrittwieser, J., Antonoglou, I.,Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe,D., Nham, J., Kalchbrenner, N.,Sutskever, I., Lillicrap, T.,Leach, M., Kavukcuoglu, K., Graepel, T., and Hassabis,D. Mastering the game of go with deep neural networksand tree search. Nature, pp. 484–489, 2016.<br />
# Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou,I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L.,van den Driessche, G., Graepel, T., and Hassabis, D.Mastering the game of go without human knowledge.Nature, pp. 354–359, 2017.<br />
# Yamamoto, M., Kato, S., and Iizuka, H. Digital curling strategy based on game tree search. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 474–480, 2015.<br />
# Ohto, K. and Tanaka, T. A curling agent based on the montecarlo tree search considering the similarity of the best action among similar states. In Proceedings of Advances in Computer Games, ACG, pp. 151–164, 2017.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39469Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-16T16:30:30Z<p>Wfisher: /* Future Work */</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, [5]) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, [6]) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, [7]) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail, balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or event a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 9 residual blocks, each consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is a 32x32x2 set of action probabilities. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games):<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times<br />
<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 ([8]). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network ===<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, ''(s<sub>t</sub>,a<sub>t</sub>)'', a state 'd' steps ahead was generated, ''s<sub>t+d</sub>''. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, and the value is used for learning the value at ''s<sub>t</sub>''.<br />
<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
After initialisation by supervised learning, the algorithm uses self-play to further train itself. During this training, the policy network learns probabilities from the MCTS process, while the value network learns from game outcomes.<br />
<br />
At a game state ''s<sub>t</sub>'':<br />
<br />
1) the algorithm outputs a prediction ''z<sub>t</sub>''. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.<br />
<br />
2) the algorithm outputs a prediction <math>\pi_t</math>, representing a probability distribution of actions. These are proportional to estimated visit counts from MCTS, based on kernel density estimation<br />
<br />
The policy-value network is updated by sampling data <math>(s, \pi, z)</math> from recent history of self-play. The same loss function is used as before.<br />
<br />
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.<br />
<br />
== Long-Term Strategy Learning ==<br />
<br />
Finally, the authors implement a new strategy to augment their algorithm for long-term play. In this context, this refers to playing a game over many ends, where the strategy to win a single end may not be a good strategy to win a full game. For example, scoring one point in an end, while being one point ahead, gives the advantage to the other team in the next round (as they will throw the last stone). The other team could then use the advantage to score two points, taking the lead.<br />
<br />
The authors build a 'winning percentage' table. This table stores the percentage of games won, based on number of ends left, and difference in score (current team - opposing team). This can be computed iteratively, and using the probability distribution estimation of one-end scores.<br />
<br />
== Final Algorithms ==<br />
<br />
The authors make use of the following versions of their algorithm:<br />
<br />
=== KR-DL ===<br />
<br />
''Kernel regression-deep learning'': This algorithm is trained only by supervised learning.<br />
<br />
=== KR-DRL ===<br />
<br />
''Kernel regression-deep reinforcement learning'': This algorithm is trained by supervised learning (ie: initialised as the KR-DL algorithm), and again on self-play. During self-play, each shot is selected after 400 MCTS simulations of k=20 randomly selected actions. Data for self-play was collected over a week on 5 GPUS, and generated 5 million game positions. The policy-value network was continually updated using samples from the latest 1 million game positions.<br />
<br />
=== KR-DRL-MES ===<br />
<br />
''Kernel regression-deep reinforcement learning-multi-ends-strategy'': This algorithm makes use of the winning percentage table generated from self-play.<br />
<br />
= Testing and Results =<br />
<br />
Testing is done with a simulated curling program (TODO). This simulator does not deal with changing ice conditions, or sweeping, but does deal with stone trajectories and collisions.<br />
<br />
== Comparison of KR-DL-UCT and DL-UCT ==<br />
<br />
The first test compares an algorithm trained with kernel regression with an algorithm trained without kernel regression, to show the contribution that kernel regression adds to the performance. Both algorithms have networks initialised with the supervised learning, and then trained with two different algorithms for self-play. KR-DL-UCT uses the algorithm described above. The authors do not go into detail on how DL-UCT selects shots, but state that a constant is set to allow exploration.<br />
<br />
As an evaluation, both algorithms play 2000 games against the DL-UCT algorithm, which is frozen after supervised training. 1000 games are played with the algorithm taking the first, and 100 taking the 2nd, shots. The games were two-end games. The figure below shows each algorithm's wining percentage given different amounts of training data. While the DL-UCT outperforms the supervised-training-only-DL-UCT algorithm, the KR-DL-UCT algorithm performs much better.<br />
<br />
[[File:curling_KR_test.png | 400px]]<br />
<br />
== Matches ==<br />
<br />
Finally, to test the performance of their multiple algorithms, the authors run matches between their algorithms and other existing programs. Each algorithm plays 200 matches against each other program, 100 of which are played as the first-playing team, and 100 as the second-playing team. Only 1 program was able to out-perform the KR-DRL algorithm. The authors state that this program, ''JiritsukunGAT'17'' also uses a deep network, and hand-crafted features. However, the KR-DRL-MES algorithm was still able to out-perform this. The figure below shows the Elo ratings of the different programs. Note that the programs in blue are those created by the authors.<br />
<br />
[[File:curling_ratings.png | 400px]]<br />
<br />
=Commentary=<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/<br />
# Silver, D., Huang, A., Maddison, C., Guez, A., Sifre, L.,Van Den Driessche, G., Schrittwieser, J., Antonoglou, I.,Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe,D., Nham, J., Kalchbrenner, N.,Sutskever, I., Lillicrap, T.,Leach, M., Kavukcuoglu, K., Graepel, T., and Hassabis,D. Mastering the game of go with deep neural networksand tree search. Nature, pp. 484–489, 2016.<br />
# Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou,I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L.,van den Driessche, G., Graepel, T., and Hassabis, D.Mastering the game of go without human knowledge.Nature, pp. 354–359, 2017.<br />
# Yamamoto, M., Kato, S., and Iizuka, H. Digital curling strategy based on game tree search. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 474–480, 2015.<br />
# Ohto, K. and Tanaka, T. A curling agent based on the montecarlo tree search considering the similarity of the best action among similar states. In Proceedings of Advances in Computer Games, ACG, pp. 151–164, 2017.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39468Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-16T16:30:18Z<p>Wfisher: Describe Results</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, [5]) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, [6]) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, [7]) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail, balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or event a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 9 residual blocks, each consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is a 32x32x2 set of action probabilities. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games):<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times<br />
<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 ([8]). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network ===<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, ''(s<sub>t</sub>,a<sub>t</sub>)'', a state 'd' steps ahead was generated, ''s<sub>t+d</sub>''. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, and the value is used for learning the value at ''s<sub>t</sub>''.<br />
<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
After initialisation by supervised learning, the algorithm uses self-play to further train itself. During this training, the policy network learns probabilities from the MCTS process, while the value network learns from game outcomes.<br />
<br />
At a game state ''s<sub>t</sub>'':<br />
<br />
1) the algorithm outputs a prediction ''z<sub>t</sub>''. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.<br />
<br />
2) the algorithm outputs a prediction <math>\pi_t</math>, representing a probability distribution of actions. These are proportional to estimated visit counts from MCTS, based on kernel density estimation<br />
<br />
The policy-value network is updated by sampling data <math>(s, \pi, z)</math> from recent history of self-play. The same loss function is used as before.<br />
<br />
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.<br />
<br />
== Long-Term Strategy Learning ==<br />
<br />
Finally, the authors implement a new strategy to augment their algorithm for long-term play. In this context, this refers to playing a game over many ends, where the strategy to win a single end may not be a good strategy to win a full game. For example, scoring one point in an end, while being one point ahead, gives the advantage to the other team in the next round (as they will throw the last stone). The other team could then use the advantage to score two points, taking the lead.<br />
<br />
The authors build a 'winning percentage' table. This table stores the percentage of games won, based on number of ends left, and difference in score (current team - opposing team). This can be computed iteratively, and using the probability distribution estimation of one-end scores.<br />
<br />
== Final Algorithms ==<br />
<br />
The authors make use of the following versions of their algorithm:<br />
<br />
=== KR-DL ===<br />
<br />
''Kernel regression-deep learning'': This algorithm is trained only by supervised learning.<br />
<br />
=== KR-DRL ===<br />
<br />
''Kernel regression-deep reinforcement learning'': This algorithm is trained by supervised learning (ie: initialised as the KR-DL algorithm), and again on self-play. During self-play, each shot is selected after 400 MCTS simulations of k=20 randomly selected actions. Data for self-play was collected over a week on 5 GPUS, and generated 5 million game positions. The policy-value network was continually updated using samples from the latest 1 million game positions.<br />
<br />
=== KR-DRL-MES ===<br />
<br />
''Kernel regression-deep reinforcement learning-multi-ends-strategy'': This algorithm makes use of the winning percentage table generated from self-play.<br />
<br />
= Testing and Results =<br />
<br />
Testing is done with a simulated curling program (TODO). This simulator does not deal with changing ice conditions, or sweeping, but does deal with stone trajectories and collisions.<br />
<br />
== Comparison of KR-DL-UCT and DL-UCT ==<br />
<br />
The first test compares an algorithm trained with kernel regression with an algorithm trained without kernel regression, to show the contribution that kernel regression adds to the performance. Both algorithms have networks initialised with the supervised learning, and then trained with two different algorithms for self-play. KR-DL-UCT uses the algorithm described above. The authors do not go into detail on how DL-UCT selects shots, but state that a constant is set to allow exploration.<br />
<br />
As an evaluation, both algorithms play 2000 games against the DL-UCT algorithm, which is frozen after supervised training. 1000 games are played with the algorithm taking the first, and 100 taking the 2nd, shots. The games were two-end games. The figure below shows each algorithm's wining percentage given different amounts of training data. While the DL-UCT outperforms the supervised-training-only-DL-UCT algorithm, the KR-DL-UCT algorithm performs much better.<br />
<br />
[[File:curling_KR_test.png | 400px]]<br />
<br />
== Matches ==<br />
<br />
Finally, to test the performance of their multiple algorithms, the authors run matches between their algorithms and other existing programs. Each algorithm plays 200 matches against each other program, 100 of which are played as the first-playing team, and 100 as the second-playing team. Only 1 program was able to out-perform the KR-DRL algorithm. The authors state that this program, ''JiritsukunGAT'17'' also uses a deep network, and hand-crafted features. However, the KR-DRL-MES algorithm was still able to out-perform this. The figure below shows the Elo ratings of the different programs. Note that the programs in blue are those created by the authors.<br />
<br />
[[File:curling_ratings.png | 400px]]<br />
<br />
=Commentary=<br />
<br />
=Future Work=<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/<br />
# Silver, D., Huang, A., Maddison, C., Guez, A., Sifre, L.,Van Den Driessche, G., Schrittwieser, J., Antonoglou, I.,Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe,D., Nham, J., Kalchbrenner, N.,Sutskever, I., Lillicrap, T.,Leach, M., Kavukcuoglu, K., Graepel, T., and Hassabis,D. Mastering the game of go with deep neural networksand tree search. Nature, pp. 484–489, 2016.<br />
# Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou,I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L.,van den Driessche, G., Graepel, T., and Hassabis, D.Mastering the game of go without human knowledge.Nature, pp. 354–359, 2017.<br />
# Yamamoto, M., Kato, S., and Iizuka, H. Digital curling strategy based on game tree search. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 474–480, 2015.<br />
# Ohto, K. and Tanaka, T. A curling agent based on the montecarlo tree search considering the similarity of the best action among similar states. In Proceedings of Advances in Computer Games, ACG, pp. 151–164, 2017.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=File:curling_ratings.png&diff=39467File:curling ratings.png2018-11-16T16:29:06Z<p>Wfisher: Source: http://proceedings.mlr.press/v80/lee18b/lee18b.pdf</p>
<hr />
<div>Source: http://proceedings.mlr.press/v80/lee18b/lee18b.pdf</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39466Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-16T16:16:16Z<p>Wfisher: Discuss some results</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, [5]) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, [6]) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, [7]) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail, balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or event a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 9 residual blocks, each consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is a 32x32x2 set of action probabilities. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games):<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times<br />
<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 ([8]). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network ===<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, ''(s<sub>t</sub>,a<sub>t</sub>)'', a state 'd' steps ahead was generated, ''s<sub>t+d</sub>''. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, and the value is used for learning the value at ''s<sub>t</sub>''.<br />
<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
After initialisation by supervised learning, the algorithm uses self-play to further train itself. During this training, the policy network learns probabilities from the MCTS process, while the value network learns from game outcomes.<br />
<br />
At a game state ''s<sub>t</sub>'':<br />
<br />
1) the algorithm outputs a prediction ''z<sub>t</sub>''. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.<br />
<br />
2) the algorithm outputs a prediction <math>\pi_t</math>, representing a probability distribution of actions. These are proportional to estimated visit counts from MCTS, based on kernel density estimation<br />
<br />
The policy-value network is updated by sampling data <math>(s, \pi, z)</math> from recent history of self-play. The same loss function is used as before.<br />
<br />
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.<br />
<br />
== Long-Term Strategy Learning ==<br />
<br />
Finally, the authors implement a new strategy to augment their algorithm for long-term play. In this context, this refers to playing a game over many ends, where the strategy to win a single end may not be a good strategy to win a full game. For example, scoring one point in an end, while being one point ahead, gives the advantage to the other team in the next round (as they will throw the last stone). The other team could then use the advantage to score two points, taking the lead.<br />
<br />
The authors build a 'winning percentage' table. This table stores the percentage of games won, based on number of ends left, and difference in score (current team - opposing team). This can be computed iteratively, and using the probability distribution estimation of one-end scores.<br />
<br />
== Final Algorithms ==<br />
<br />
The authors make use of the following versions of their algorithm:<br />
<br />
=== KR-DL ===<br />
<br />
''Kernel regression-deep learning'': This algorithm is trained only by supervised learning.<br />
<br />
=== KR-DRL ===<br />
<br />
''Kernel regression-deep reinforcement learning'': This algorithm is trained by supervised learning (ie: initialised as the KR-DL algorithm), and again on self-play. During self-play, each shot is selected after 400 MCTS simulations of k=20 randomly selected actions. Data for self-play was collected over a week on 5 GPUS, and generated 5 million game positions. The policy-value network was continually updated using samples from the latest 1 million game positions.<br />
<br />
=== KR-DRL-MES ===<br />
<br />
''Kernel regression-deep reinforcement learning-multi-ends-strategy'': This algorithm makes use of the winning percentage table generated from self-play.<br />
<br />
= Testing and Results =<br />
<br />
Testing is done with a simulated curling program (TODO). This simulator does not deal with changing ice conditions, or sweeping, but does deal with stone trajectories and collisions.<br />
<br />
== Comparison of KR-DL-UCT and DL-UCT ==<br />
<br />
The first test compares an algorithm trained with kernel regression with an algorithm trained without kernel regression, to show the contribution that kernel regression adds to the performance. Both algorithms have networks initialised with the supervised learning, and then trained with two different algorithms for self-play. KR-DL-UCT uses the algorithm described above. The authors do not go into detail on how DL-UCT selects shots, but state that a constant is set to allow exploration.<br />
<br />
As an evaluation, both algorithms play 2000 games against the DL-UCT algorithm, which is frozen after supervised training. 1000 games are played with the algorithm taking the first, and 100 taking the 2nd, shots. The games were two-end games. The figure below shows each algorithm's wining percentage given different amounts of training data. While the DL-UCT outperforms the supervised-training-only-DL-UCT algorithm, the KR-DL-UCT algorithm performs much better.<br />
<br />
[[File:curling_KR_test.png | 400px]]<br />
<br />
=Commentary=<br />
<br />
=Future Work=<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/<br />
# Silver, D., Huang, A., Maddison, C., Guez, A., Sifre, L.,Van Den Driessche, G., Schrittwieser, J., Antonoglou, I.,Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe,D., Nham, J., Kalchbrenner, N.,Sutskever, I., Lillicrap, T.,Leach, M., Kavukcuoglu, K., Graepel, T., and Hassabis,D. Mastering the game of go with deep neural networksand tree search. Nature, pp. 484–489, 2016.<br />
# Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou,I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L.,van den Driessche, G., Graepel, T., and Hassabis, D.Mastering the game of go without human knowledge.Nature, pp. 354–359, 2017.<br />
# Yamamoto, M., Kato, S., and Iizuka, H. Digital curling strategy based on game tree search. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 474–480, 2015.<br />
# Ohto, K. and Tanaka, T. A curling agent based on the montecarlo tree search considering the similarity of the best action among similar states. In Proceedings of Advances in Computer Games, ACG, pp. 151–164, 2017.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=File:curling_KR_test.png&diff=39465File:curling KR test.png2018-11-16T16:14:09Z<p>Wfisher: Source: http://proceedings.mlr.press/v80/lee18b/lee18b.pdf</p>
<hr />
<div>Source: http://proceedings.mlr.press/v80/lee18b/lee18b.pdf</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39447Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-16T03:30:20Z<p>Wfisher: Add references</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, [5]) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, [6]) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, [7]) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail, balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or event a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 9 residual blocks, each consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is a 32x32x2 set of action probabilities. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games):<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times<br />
<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 ([8]). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network ===<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, ''(s<sub>t</sub>,a<sub>t</sub>)'', a state 'd' steps ahead was generated, ''s<sub>t+d</sub>''. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, and the value is used for learning the value at ''s<sub>t</sub>''.<br />
<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
After initialisation by supervised learning, the algorithm uses self-play to further train itself. During this training, the policy network learns probabilities from the MCTS process, while the value network learns from game outcomes.<br />
<br />
At a game state ''s<sub>t</sub>'':<br />
<br />
1) the algorithm outputs a prediction ''z<sub>t</sub>''. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.<br />
<br />
2) the algorithm outputs a prediction <math>\pi_t</math>, representing a probability distribution of actions. These are proportional to estimated visit counts from MCTS, based on kernel density estimation<br />
<br />
The policy-value network is updated by sampling data <math>(s, \pi, z)</math> from recent history of self-play. The same loss function is used as before.<br />
<br />
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.<br />
<br />
== Long-Term Strategy Learning ==<br />
<br />
Finally, the authors implement a new strategy to augment their algorithm for long-term play. In this context, this refers to playing a game over many ends, where the strategy to win a single end may not be a good strategy to win a full game. For example, scoring one point in an end, while being one point ahead, gives the advantage to the other team in the next round (as they will throw the last stone). The other team could then use the advantage to score two points, taking the lead.<br />
<br />
The authors build a 'winning percentage' table. This table stores the percentage of games won, based on number of ends left, and difference in score (current team - opposing team). This can be computed iteratively, and using the probability distribution estimation of one-end scores.<br />
<br />
=Experimental Procedure and Results=<br />
<br />
<br />
=Commentary=<br />
<br />
=Future Work=<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/<br />
# Silver, D., Huang, A., Maddison, C., Guez, A., Sifre, L.,Van Den Driessche, G., Schrittwieser, J., Antonoglou, I.,Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe,D., Nham, J., Kalchbrenner, N.,Sutskever, I., Lillicrap, T.,Leach, M., Kavukcuoglu, K., Graepel, T., and Hassabis,D. Mastering the game of go with deep neural networksand tree search. Nature, pp. 484–489, 2016.<br />
# Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou,I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L.,van den Driessche, G., Graepel, T., and Hassabis, D.Mastering the game of go without human knowledge.Nature, pp. 354–359, 2017.<br />
# Yamamoto, M., Kato, S., and Iizuka, H. Digital curling strategy based on game tree search. In Proceedings of the IEEE Conference on Computational Intelligence and Games, CIG, pp. 474–480, 2015.<br />
# Ohto, K. and Tanaka, T. A curling agent based on the montecarlo tree search considering the similarity of the best action among similar states. In Proceedings of Advances in Computer Games, ACG, pp. 151–164, 2017.</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39446Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-16T03:24:14Z<p>Wfisher: /* Supervised Learning */</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, TODO) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, TODO) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, TODO) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail (TODO), balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or event a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 9 residual blocks, each consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is a 32x32x2 set of action probabilities. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games):<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times<br />
<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 (TODO: reference). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network ===<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, ''(s<sub>t</sub>,a<sub>t</sub>)'', a state 'd' steps ahead was generated, ''s<sub>t+d</sub>''. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, and the value is used for learning the value at ''s<sub>t</sub>''.<br />
<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
After initialisation by supervised learning, the algorithm uses self-play to further train itself. During this training, the policy network learns probabilities from the MCTS process, while the value network learns from game outcomes.<br />
<br />
At a game state ''s<sub>t</sub>'':<br />
<br />
1) the algorithm outputs a prediction ''z<sub>t</sub>''. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.<br />
<br />
2) the algorithm outputs a prediction <math>\pi_t</math>, representing a probability distribution of actions. These are proportional to estimated visit counts from MCTS, based on kernel density estimation<br />
<br />
The policy-value network is updated by sampling data <math>(s, \pi, z)</math> from recent history of self-play. The same loss function is used as before.<br />
<br />
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.<br />
<br />
== Long-Term Strategy Learning ==<br />
<br />
Finally, the authors implement a new strategy to augment their algorithm for long-term play. In this context, this refers to playing a game over many ends, where the strategy to win a single end may not be a good strategy to win a full game. For example, scoring one point in an end, while being one point ahead, gives the advantage to the other team in the next round (as they will throw the last stone). The other team could then use the advantage to score two points, taking the lead.<br />
<br />
The authors build a 'winning percentage' table. This table stores the percentage of games won, based on number of ends left, and difference in score (current team - opposing team). This can be computed iteratively, and using the probability distribution estimation of one-end scores.<br />
<br />
=Experimental Procedure and Results=<br />
<br />
<br />
=Commentary=<br />
<br />
=Future Work=<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39445Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-16T03:20:11Z<p>Wfisher: /* Self-Play Reinforcement Learning */</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, TODO) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, TODO) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, TODO) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail (TODO), balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or event a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 9 residual blocks, each consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is a 32x32x2 set of action probabilities. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games):<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times<br />
<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 (TODO: reference). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, ''(s<sub>t</sub>,a<sub>t</sub>)'', a state 'd' steps ahead was generated, ''s<sub>t+d</sub>''. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, and the value is used for learning the value at ''s<sub>t</sub>''.<br />
<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
After initialisation by supervised learning, the algorithm uses self-play to further train itself. During this training, the policy network learns probabilities from the MCTS process, while the value network learns from game outcomes.<br />
<br />
At a game state ''s<sub>t</sub>'':<br />
<br />
1) the algorithm outputs a prediction ''z<sub>t</sub>''. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.<br />
<br />
2) the algorithm outputs a prediction <math>\pi_t</math>, representing a probability distribution of actions. These are proportional to estimated visit counts from MCTS, based on kernel density estimation<br />
<br />
The policy-value network is updated by sampling data <math>(s, \pi, z)</math> from recent history of self-play. The same loss function is used as before.<br />
<br />
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.<br />
<br />
== Long-Term Strategy Learning ==<br />
<br />
Finally, the authors implement a new strategy to augment their algorithm for long-term play. In this context, this refers to playing a game over many ends, where the strategy to win a single end may not be a good strategy to win a full game. For example, scoring one point in an end, while being one point ahead, gives the advantage to the other team in the next round (as they will throw the last stone). The other team could then use the advantage to score two points, taking the lead.<br />
<br />
The authors build a 'winning percentage' table. This table stores the percentage of games won, based on number of ends left, and difference in score (current team - opposing team). This can be computed iteratively, and using the probability distribution estimation of one-end scores.<br />
<br />
=Experimental Procedure and Results=<br />
<br />
<br />
=Commentary=<br />
<br />
=Future Work=<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39444Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-16T03:18:09Z<p>Wfisher: Add self-play sections</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, TODO) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, TODO) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, TODO) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail (TODO), balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or event a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 9 residual blocks, each consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is a 32x32x2 set of action probabilities. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games):<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times<br />
<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 (TODO: reference). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, ''(s<sub>t</sub>,a<sub>t</sub>)'', a state 'd' steps ahead was generated, ''s<sub>t+d</sub>''. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, and the value is used for learning the value at ''s<sub>t</sub>''.<br />
<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
After initialisation by supervised learning, the algorithm uses self-play to further train itself. During this training, the policy network learns probabilities from the MCTS process, while the value network learns from game outcomes.<br />
<br />
At a game state ''s<sub>t</sub>'':<br />
<br />
1) the algorithm outputs a prediction ''z<sub>t</sub>''. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.<br />
<br />
2) the algorithm outputs a prediction ''<math>\pi_t</math>, representing a probability distribution of actions. These are proportional to estimated visit counts from MCTS, based on kernel density estimation<br />
<br />
The policy-value network is updated by sampling data <math>(s, \pi, z)</math> from recent history of self-play. The same loss function is used as before.<br />
<br />
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.<br />
<br />
<br />
== Long-Term Strategy Learning ==<br />
<br />
Finally, the authors implement a new strategy to augment their algorithm for long-term play. In this context, this refers to playing a game over many ends, where the strategy to win a single end may not be a good strategy to win a full game. For example, scoring one point in an end, while being one point ahead, gives the advantage to the other team in the next round (as they will throw the last stone). The other team could then use the advantage to score two points, taking the lead.<br />
<br />
The authors build a 'winning percentage' table. This table stores the percentage of games won, based on number of ends left, and difference in score (current team - opposing team). This can be computed iteratively, and using the probability distribution estimation of one-end scores.<br />
<br />
=Experimental Procedure and Results=<br />
<br />
<br />
=Commentary=<br />
<br />
=Future Work=<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39443Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-16T02:57:38Z<p>Wfisher: Add information on supervised learning</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, TODO) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, TODO) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, TODO) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail (TODO), balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or event a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 9 residual blocks, each consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is a 32x32x2 set of action probabilities. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games):<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times<br />
<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 (TODO: reference). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
=== Policy Network<br />
<br />
The policy network was trained to learn the action taken in each state. Here, the likelihood of the taken action was set to be 1, and the likelihood of other actions to be 0.<br />
<br />
=== Value Network ===<br />
<br />
The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, ''(s<sub>t</sub>,a<sub>t</sub>)'', a state 'd' steps ahead was generated, ''s<sub>t+d</sub>''. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, ''a<sub>t+d-1</sub>''. The value network is used to predict the value for this state, and the value is used for learning the value at ''s<sub>t</sub>''.<br />
<br />
<br />
=== Policy-Value Network ===<br />
<br />
The policy-value network was trained to maximize the similarity of the predicted policy and value, and the actual policy and value from a state. The learning algorithm parameters are:<br />
<br />
* Algorithm: stochastic gradient descent<br />
* Batch size: 256<br />
* Momentum: 0.9<br />
* L2 regularization: 0.0001<br />
* Training time: ~100 epochs<br />
* Learning rate: initialised at 0.01, reduced twice<br />
<br />
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:<br />
<br />
[[File:curling_loss_function.png | 300px]]<br />
<br />
== Self-Play Reinforcement Learning ==<br />
<br />
=Experimental Procedure and Results=<br />
<br />
<br />
=Commentary=<br />
<br />
=Future Work=<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=File:curling_loss_function.png&diff=39442File:curling loss function.png2018-11-16T02:53:09Z<p>Wfisher: Source: http://proceedings.mlr.press/v80/lee18b/lee18b.pdf</p>
<hr />
<div>Source: http://proceedings.mlr.press/v80/lee18b/lee18b.pdf</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39441Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-16T02:42:19Z<p>Wfisher: Start creating supervised learning section</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even more recent years, the use of CNN's has allowed neural networks to out-perform humans in many difficult games, such as Go. However, many of these cases involve a discrete state or action space, and these methods cannot be directly applied to continuous action spaces.<br />
<br />
This paper introduces a method to allow learning with continuous action spaces. A CNN is used to perform learning on a discretion state and action spaces, and then a continuous action search is performed on these discrete results.<br />
<br />
Curling is chosen as a domain to test the network on. Curling was chosen due to its large action space, potential for complicated strategies, and need for precise interactions.<br />
<br />
== Curling ==<br />
<br />
Curling is a sport played by two teams on a long sheet of ice. Roughly, the goal is for each time to slide rocks closer to the target on the other end of the sheet than the other team. The next sections will provide a background on the gameplay, and potential challenges/concerns for learning algorithms. A terminology section follows.<br />
<br />
=== Gameplay ===<br />
<br />
A game of curling is divided into ends. In each end, players from both teams alternate throwing (sliding) eight rocks to the other end of the ice sheet, known as the house. Rocks must land in a certain area in order to stay in play, and must touch or be inside concentric rings (12ft diameter and smaller) in order to score points. At the end of each end, the team with rocks closest to the center of the house scores points.<br />
<br />
When throwing a rock, the curling can spin the rock. This allows the rock to 'curl' its path towards the house, and can allow rocks to travel around other rocks. Teammembers are also able to sweep the ice in front of a moving rock in order to decrease friction, which allows for fine-tuning of distance (though the physics of sweeping are not implemented in the simulation used).<br />
<br />
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:<br />
<br />
* Draw: Throw a rock to a target location<br />
* Freeze: Draw a rock up against another rock<br />
* Takeout: Knock another rock out of the house. Can be combined with different ricochet directions<br />
* Guard: Place a rock in front of another, to block other rocks (ex: takeouts)<br />
<br />
=== Challenges for AI ===<br />
<br />
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.<br />
<br />
The effect of changing actions can be highly nonlinear and discontinuous. This can be seen when considering that a 1-cm deviation in a path can make the difference between a high-speed collision, or lack of collision.<br />
<br />
Curling will require both offensive and defensive strategies. For example, consider the fact that the last team to throw a rock each end only needs to place that rock closer than the opposing team's rocks to score a point, and invalidate any opposing rocks in the house. The opposing team should thus be considering how to prevent this from happening, in addition to scoring points themselves.<br />
<br />
Curling also has a concept known as 'the hammer'. The hammer belongs to the team which throws the last rock each end, providing an advantage, and is given to the team that does not score points each end. It could very well be good strategy to try not to win a single point in an end (if already ahead in points, etc), as this would give the advantage to the opposing team.<br />
<br />
Finally, curling has a rule known as the 'Free Guard Zone'. This applies to the first 4 rocks thrown (2 from each team). If they land short of the house, but still in play, then the rocks are not allowed to be removed (via collisions) until all of the first 4 rocks have been thrown.<br />
<br />
=== Terminology ===<br />
<br />
* End: A round of the game<br />
* House: The end of the sheet of ice, which contains<br />
* Hammer: The team that throws the last rock of an end 'has the hammer'<br />
* Hog Line: thick line that is drawn in front of the house, orthogonal to the length of the ice sheet. Rocks must pass this line to remain in play.<br />
* Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.<br />
<br />
<br />
== Related Work ==<br />
<br />
=== AlphaGo Lee ===<br />
<br />
AlphaGo Lee (Silver et al., 2016, TODO) refers to an algorithm used to play the game Go, which was able to defeat internation champion Lee Sedol. Two neural networks were trained on the moves of human experts, to act as both a policy network and a value network. A Monte Carlo Tree Search algorithm was used for policy improvement.<br />
<br />
The use of both policy and value networks are reflected in this paper's work.<br />
<br />
=== AlphaGo Zero ===<br />
<br />
AlphaGo Zero (Silver et al., 2017, TODO) is an improvement on the AlphaGo Lee algorithm. AlphaGo Zero uses a unified neural network in place of the separate policy and value networks, and is trained on self-play, without the need of expert training.<br />
<br />
The unification of networks, and self-play are also reflected in this paper.<br />
<br />
=== Curling Algorithms ===<br />
<br />
Some past algorithms have been proposed to deal with continuous action spaces. For example, (Yammamoto et al, 2015, TODO) use game tree search methods in a discretized space. The value of an action is taken as the average of nearby values, with respect to some knowledge of execution uncertainty.<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search algorithms have been applied to continuous action spaces. These algorithms, to be discussed in further detail (TODO), balance exploration of different states, with knowledge of paths of execution through past games.<br />
<br />
=== Curling Physics and Simulation ===<br />
<br />
Several references in the paper refer to the study and simulation of curling physics.<br />
<br />
== General Background of Algorithms ==<br />
<br />
=== Policy and Value Functions ===<br />
<br />
A policy function is trained to provide the best action to take, given a current state. Policy iteration is an algorithm used to improve a policy over time. This is done by alternating between policy evaluation and policy improvement.<br />
<br />
A value function is trained to estimate the value of a value of being in a certain state. It is trained based on records of state-action-reward sets.<br />
<br />
<br />
=== Monte Carlo Tree Search ===<br />
<br />
Monte Carlo Tree Search (MCTS) is a search algorithm used for finite-horizon tasks (ex: in curling, only 16 moves, or thrown stones, are taken each end).<br />
<br />
MCTS is a tree search algorithm similar to minimax. However, MCTS is probabilistic, and does not need to explore a full game tree, or event a tree reduced with alpha-beta pruning. This makes it tractable for games such as GO, and curling.<br />
<br />
Nodes of the tree are game states, and branches represent actions. Each node stores statistics on how many times it has been visited by the MCTS, as well as the number of wins encountered by playouts from that position. A node has been considered 'visited' if a full playout has started from that node. A node is considered 'expanded' if all its children have been visited.<br />
<br />
MCTS begins with the '''selection''' phase, which involves traversing known states/actions. This involves expanding the tree by beginning at the root node, and selecting the child/score with the highest 'score'. From each successive node, a path down to a root node is explored in a similar fashion.<br />
<br />
The next phase, '''expansion''', begins when the algorithm reaches a node where not all children have been visited (ie: the node has not been fully expanded). In the expansion phase, children of the node are visited, and '''simulations''' run from their states.<br />
<br />
Once the new child is expanded, '''simulation''' takes place. This refers to a full playout of the game from the point of the current node, and can involve many strategies, such as randomly taken moves, the use of heuristics, etc.<br />
<br />
The final phase is '''update''' or '''back-propagation''' (unrelated to the neural network algorithm). In this phase, the result of the '''simulation''' (ie: win/lose) is update in the statistics of all parent nodes.<br />
<br />
A selection function known as Upper Confidence Bound (UCT) can be used for selecting which node to select. The formula for this equation is shown below [[https://www.baeldung.com/java-monte-carlo-tree-search source]]. Note that the first term essentially acts as an average score of games played from a certain node. The second term, meanwhile, will grow when sibling nodes are expanded. This means that unexplored nodes will gradually increase their UCT score, and be selected in the future.<br />
<br />
[[File:mcts_uct_equation.png | 500px | centered]]<br />
<br />
Sources: 2,3,4<br />
<br />
=== Kernel Regression ===<br />
<br />
Kernel regression is a form of weighted averaging. Given two items of data, '''x''', each of which have a value '''y''' associated with them, the kernel functions outputs a weighting factor. An estimate of the value of a new, unseen point, is then calculated as the weighted average of values of surrounding points.<br />
<br />
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).<br />
<br />
[[File:gaussian_kernel.png | 400 px]]<br />
<br />
[[File:kernel_regression.png | 350 px]]<br />
<br />
In this case, the combination of the two act to weigh scores of samples closest to '''x''' more strongly.<br />
<br />
= Methods =<br />
<br />
== Network Design ==<br />
<br />
The authors design a CNN, called the 'policy-value' network. The network consists of a common network structure, which is then split into 'policy' and 'value' outputs. This network is trained to learn a probability distribution of actions to take, and expected rewards, given an input state.<br />
<br />
=== Shared Structure ===<br />
<br />
The network consists of 9 residual blocks, each consisting of 2 convolutional layers with 32 3x3 filters. The structure of this network is shown below:<br />
<br />
[[File:curling_network_layers.png]]<br />
<br />
<br />
the input to this network is the following:<br />
* Location of stones<br />
* Order to tee (the center of the sheet)<br />
* A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.<br />
<br />
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.<br />
<br />
=== Policy Network ===<br />
<br />
The policy head is created by adding 2 convolutional layers with 2 3x3 filters to the main body of the network. The output of the policy head is a 32x32x2 set of action probabilities. The actions represent target locations in the grid, and spin direction of the stone.<br />
<br />
=== Value Network ===<br />
<br />
The value head is created by adding a convolution layer with 1 3x3 filter, and dense layers of 256 and 17 units, to the shared network. The 17 output units represent a probability of scores in the range of [-8,8], which are the possible scores in each end of a curling game.<br />
<br />
== Continuous Action Search ==<br />
<br />
The policy head of the network only outputs actions from a discretized action space. For real-life interactions, and especially in curling, this will not suffice, as very fine adjustments to actions can make significant differences in outcomes.<br />
<br />
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:<br />
<br />
=== Selection ===<br />
<br />
From a given state, the list of already-visited actions is denoted as A<sub>t</sub>. Scores and the number of visits to each node are estimated using the equations below (the first equation shows the expectation of the end value for one-end games):<br />
<br />
[[File:curling_kernel_equations.png | 500px]]<br />
<br />
The UCB formula is then used to select an action to expand.<br />
<br />
<br />
=== Expansion ===<br />
<br />
The authors use a variant of regular UCT for expansion. In this case, they expand a new node only when existing nodes have been visited a certain number of times<br />
<br />
<br />
=== Simulation ===<br />
<br />
Instead of simulating with a random game playout, the authors use the value network to estimate the likely score associated with a state. This speeds up simulation (assuming the network is well trained), as the game does not actually need to be simulated.<br />
<br />
=== Backpropogation ===<br />
<br />
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.<br />
<br />
<br />
== Supervised Learning ==<br />
<br />
During supervised training, data is gathered from the program AyumuGAT'16 (TODO: reference). This program is also based on both a MCTS algorithm, and a high-performance AI curling program. 400 000 state-action pairs were generated during this training.<br />
<br />
The policy network was trained to learn the action taken in each state. The value network was trained by 'd-depth simulations and bootstrapping of the prediction to handle the high variance in rewards resulting from a sequence of stochastic moves' (quote taken from paper). In this case, ''m'' state-action pairs were sampled from the training data. For each pair, ''(s<sub>t</sub>,a<sub>t</sub>)''<br />
<br />
=Experimental Procedure and Results=<br />
<br />
<br />
=Commentary=<br />
<br />
=Future Work=<br />
<br />
=References=<br />
# Lee, K., Kim, S., Choi, J. & Lee, S. "Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling." Proceedings of the 35th International Conference on Machine Learning, in PMLR 80:2937-2946 (2018)<br />
# https://www.baeldung.com/java-monte-carlo-tree-search<br />
# https://jeffbradberry.com/posts/2015/09/intro-to-monte-carlo-tree-search/<br />
# https://int8.io/monte-carlo-tree-search-beginners-guide/</div>Wfisherhttp://wiki.math.uwaterloo.ca/statwiki/index.php?title=Deep_Reinforcement_Learning_in_Continuous_Action_Spaces_a_Case_Study_in_the_Game_of_Simulated_Curling&diff=39440Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling2018-11-16T02:30:35Z<p>Wfisher: Finish off MCTS sections</p>
<hr />
<div>This page provides a summary and critique of the paper '''Deep Reinforcement Learning in Continuous Action Spaces: a Case Study in the Game of Simulated Curling''' [[http://proceedings.mlr.press/v80/lee18b/lee18b.pdf Online Source]], published in ICML 2018<br />
<br />
= Introduction and Motivation =<br />
<br />
In recent years, Reinforcement Learning methods have been applied to many different games, such as chess and checkers. In even m