A scaled down version of AlphaGo Zero, playing on a 5x5 board
This projects aims to reimplement the AlphaGo Zero paper to play on a 5x5 board.
AlphaGo Zero uses end-to-end reinforcement learning to learn Go from scratch, without any human player data.
-
GoGame: Go simulation code.
-
MCTS: Implmentation for Monte-carlo tree search.
-
Model: Neural network architecture
-
Selfplay: Module for managing the games between agents
-
Shared: Miscellaneous functions and constants
-
Training: Class for generating gameplay data and performing training and evaluation
Generate games and train the model:
python3 train.pyTest a trained model:
python3 test.py