This is a pytorch implementation of Google Deep Mind's AlphaZero algorithm for chess.
Standard python libraries.
The entry point to the chess engine is the python file playchess.py. Good parameters for strong, long-thinking moves would be:
python3 playchess.py --model weights/AlphaZeroNet_20x256.pt --verbose --rollouts 1000 --threads 10 --mode h
The current position is displayed with an ascii chess board. Enter your moves in long algebraic notation. Note that running the engine requires a weights file.
Download the CCRL Dataset, reformat it using reformat.pyand run train.py.
The algorithm is based on this paper. One very important difference between the algorithm used here and the one described in that paper is that this implementation used supervised learning instead of reinforcement learning. Doing reienforcement learning is very computationally intensive. As said in that paper, it took thousands of TPUs to generate the self play games. This program, on the other hand, trains on the CCRL Dataset, which contains 2.5 million top notch chess games. Because each game has around 80 unique positions in it, this yields about 200 million data points for training on.
I have only tested it a few times against the computers on lichess and chess.com. It was able to draw, but had to think for about a minute while its opponent only got a few seconds. Compared to most chess engines which use the Alpha-Beta algorithm and a hand written evaluation function, it evaluates a lot fewer positions every second. Most top programs consider about 45 million chess positions every second, whereas this programs considers about 300, even with GPU and multicore speedups.