ReadMe.md

AlphaZero reproduction

• GO code is available in goban.py. It respects the rules provided here.
• Vanilla MCTS player is available in mctsPlayer.py.
• AlphaZero player is available in mctsPlayer.py, the model can be found in alphazeroModel.py. Missing computational power to train it on a 19x19 goban. The player is not optimized at all and maybe still bugged.

Requirements

joblib==0.14.1
scipy==1.4.1
matplotlib==3.1.2
numpy==1.16.2
tensorflow==2.1.0

Usage

It is possible to simulate game using play.py.

    goban = Goban()
    player1, player2 = AlphaZeroPlayer(
        goban._WHITE, goban.get_board_size(), "saves/test1/best/model"), MCTSPlayer(goban._BLACK, goban.get_board_size())
    play(player1, player2, goban)

To train the AlphaZero player, simply launch from the root of the project:

python trainMcts.py --name model_name