This is meant to serve as a clone of the AlphaZero architecture, discussed here. The architecture is reproduced in both keras and pytorch.
There are still a number of bugs to resolve before results can be generated.
- numpy
- keras
- pytorch
- python-chess (when using the chess game)
The main.py file contains an example of setting up a training session for the AlphaZero architecture. One can select the backend (pytorch or keras) by changing the nnet.keras line to nnet.pytorch. You can play against the trained model by running play.py.
Three games are available, being either TicTacToe, Othello, or Chess. TicTacToe and Othello board sizes can be changed by utilising their parameter setting, e.g. TicTacToe(5) produces a 5x5 TicTacToe board, and Othello(10) would produce a 10x10 Othello board.