Chess bot

This is a bot that learns chess in two steps:

Supervised learning

The first step focuses on learning from a Stockfish engine how to evaluate a chess board. To do this, we downloaded a chess database (KingBase 2018 & KingBase 2019) containing games of GM players that we evaluated with Stockfish. Then, we designed a neural network similar to Giraffe (see docs: giraffe) which has been trained to score similarly as the Stockfish engine

Reinforcement learning (ongoing)

The second step of this project (still ongoing) is to manage to learn further how to play chess by applying TD-leaf(lambda) algorithm (see docs: giraffe). The idea of this algorithm is to minimize the error between Giraffe's board evaluation at a time t and its own evaluation of the board n steps further in the future. By doing this, we hope that the network keeps temporal consistency on the evaluation of the board.

How to use:

Warning: All the scripts use all your CPU cores by default, to change this simply set the N_PROC variable of each script to the number of cores you want to actually use

Build Stockfish database

A Stockfish database is expected in the path data/csv. If you want to create it you can run data_preparation script. Just make sure to set the max_game parameter (the number of games you want to learn from) to the value you desire. Note: The data_preparation script expects pgn file in the path data/pgn. These files can be downloaded from KingBase 2018 & KingBase 2019

Learn from Stockfish

Use stockfish_training.py script. This will learn from the Stockfish database you have built and store the model in the model directory

Self-learning (ongoing)

Use td_leaf_learning.py script. This is currently in progress

Play actual game

You can visualize a game between two neural networks in the minimax.py script

Expected folder structure

├── board_encoding.py
├── data
│   ├── csv
│   │   ├── KingBase2018-A00-A39.csv
│   │   ├── ...
│   │   ├── test.csv
│   │   └── train.csv
│   └── pgn
│   ├── KingBase2018-A00-A39.pgn
│   ├── ...
│   └── KingBase2019-A00-A39.pgn
├── data_preparation.py
├── docs
│   ├── giraffe.pdf
│   └── td_leaf.pdf
├── GiraffeNet.py
├── minimax.py
├── model
│   ├── giraffe_net_td_07.pt
│   ├── stockfish_net_4.pt
│   ├── stockfish_net_5.pt
│   └── stockfish_net.pt
├── README.md
├── stockfish_testing.py
├── stockfish_training.py
└── td_leaf_training.py