Implementation of an alphaZero like algorithm to play games from the openSpiel library such as breakthrough or connect-four. To train a new network, run train.py
Overview of important files: train.py: Main script to run to train an AlphaZero PV-network, also contains the trainer class that trains a network. Current configuration reproduces data for Figure 7.
network.py: ResNet implementation for board games.
examplegenerator.py: File deals with parallelization of games on multiple threads and GPU's.
alphazerobot.py: OpenSpiel bot implementation of an AlphaZero agent. Selects actions based on MCTS with some value and policy evaluator (a Neural Network)
game_utils.py: Contains the functions to play games. Note that here the difference between value targets is made.
mcts.py: Implementation of MCTS guided by a policy-value evaluator.
Here is a broad overview of the code structure.
