FastChess
FastChess,
a didactic open source chess engine by Thomas Dybdahl Ahle, written in Python, licensed under the GPL v3.0.
FastChess predicts the next move by probing a one-layer neural network softmax model, using the fastText text classification library.
The model takes the board state as input, and outputs a vector of probabilities for each possible move. That simple linear model might further be combined with a Monte-Carlo tree search along with the PUCT selection to improve the quality of play [2].
Training
FastChess' model is trained by feeeding a set of pgn files to a special training procedure, creating the neural network weights in form of a model.bin file, which is later used to play chess [3].
Tuning
FastChess' hyperparameters can be tuned with black box optimization through scikit optimize [4].
See also
Forum Posts
- Re: A question to MCTS + NN experts by Thomas Dybdahl Ahle, CCC, August 04, 2019
- New Tool for Tuning with Skopt by Thomas Dybdahl Ahle, CCC, August 25, 2019 [5]
- Re: AlphaZero by Thomas Dybdahl Ahle, CCC, May 05, 2020
External Links
- GitHub - thomasahle/fastchess: Predicts the best chess move with 27.5% accuracy by a single matrix multiplication
- GitHub - thomasahle/noisy-bayesian-optimization: Bayesian Optimization for very Noisy functions
References
- ↑ fastchess/README.md at master · thomasahle/fastchess · GitHub - Screenshot
- ↑ fastchess/README.md at master · thomasahle/fastchess · GitHub - Teaching FastText to play Chess
- ↑ fastchess/README.md at master · thomasahle/fastchess · GitHub - Train the model
- ↑ New Tool for Tuning with Skopt by Thomas Dybdahl Ahle, CCC, August 25, 2019
- ↑ skopt API documentation