Deep Reinforcement Learning in Continuous Action Spaces a Case Study in the Game of Simulated Curling
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 [Online Source], published in ICML 2018
- 1 Introduction and Motivation
- 1.1 Curling
- 1.2 Related Work
- 1.3 General Background of Algorithms
- 2 Methods
- 2.1 Network Design
- 2.2 Continuous Action Search
- 2.3 Supervised Learning
- 2.4 Self-Play Reinforcement Learning
- 2.5 Long-Term Strategy Learning
- 3 Experimental Procedure and Results
- 4 Commentary
- 5 Future Work
- 6 References
Introduction and Motivation
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.
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.
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.
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.
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.
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).
Curling offers many possible high-level actions, which are directed by a team member to the throwing member. An example set of these includes:
- Draw: Throw a rock to a target location
- Freeze: Draw a rock up against another rock
- Takeout: Knock another rock out of the house. Can be combined with different ricochet directions
- Guard: Place a rock in front of another, to block other rocks (ex: takeouts)
Challenges for AI
Curling offers many challenges for curling based on its physics and rules. This sections lists a few concerns.
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.
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.
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.
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.
- End: A round of the game
- House: The end of the sheet of ice, which contains
- Hammer: The team that throws the last rock of an end 'has the hammer'
- 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.
- Back Line: think line drawn just behind the house. Rocks that pass this line are removed from play.
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.
The use of both policy and value networks are reflected in this paper's work.
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.
The unification of networks, and self-play are also reflected in this paper.
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.
Monte Carlo Tree Search
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.
Curling Physics and Simulation
Several references in the paper refer to the study and simulation of curling physics.
General Background of Algorithms
Policy and Value Functions
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.
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.
Monte Carlo Tree Search
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).
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.
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.
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.
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.
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.
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.
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 [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.
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.
A typical kernel is the Gaussian kernel, shown below. The formula for calculating estimated value is shown below as well (sources: Lee et al.).
In this case, the combination of the two act to weigh scores of samples closest to x more strongly.
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.
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:
the input to this network is the following:
- Location of stones
- Order to tee (the center of the sheet)
- A 32x32 grid of representation of the ice sheet, representing which stones are present in each grid cell.
The authors do not describe how the stone-based information is added to the 32x32 grid as input to the network.
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.
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.
Continuous Action Search
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.
Actions in the continuous space are generate using a MCTS algorithm, with the following steps:
From a given state, the list of already-visited actions is denoted as At. 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):
The UCB formula is then used to select an action to expand.
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
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.
Standard backpropogration is used, updating both the values and number of visits stored in the path of parent nodes.
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.
=== Policy Network
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.
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, (st,at), a state 'd' steps ahead was generated, st+d. This process dealt with uncertainty by considering all actions in this rollout to have no uncertainty, and allowing uncertainty in the last action, at+d-1. The value network is used to predict the value for this state, and the value is used for learning the value at st.
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:
- Algorithm: stochastic gradient descent
- Batch size: 256
- Momentum: 0.9
- L2 regularization: 0.0001
- Training time: ~100 epochs
- Learning rate: initialised at 0.01, reduced twice
A multi-task loss function was used. This takes the summation of the cross-entropy losses of each prediction:
Self-Play Reinforcement Learning
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.
At a game state st:
1) the algorithm outputs a prediction zt. This is en estimate of game score probabilities. It is based on similar past actions, and computed using kernel regression.
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
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.
It is not clear how the improved network is used, as MCTS seems to be the driving process at this point.
Long-Term Strategy Learning
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.
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.
Experimental Procedure and Results
- 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)