regular-dev / chess_trainer

Train and play with your own chess bot using nevermind-neu and pleco

Chess trainer

Train and play with your own chess bot using nevermind-neu and pleco. Just follow simple instructions to train your own bot.

Usage

1. Clone repository

2. Run cargo build --release

3. Download any lichess pgn database from https://database.lichess.org/ (.pgn.zst) to chess_trainer/py folder, i suggest to choose not large file, for example "2014 - January" - 100 mb. cd py

wget https://database.lichess.org/standard/lichess_db_standard_rated_2014-01.pgn.zst

unzstd lichess_db_standard_rated_2014-01.pgn.zst - unpack zstd archive

4. Convert pgn file to sqlite3 database with columns - [ fen , stockfish eval ] with python code. For engine evaluation you need to download stockfish engine binary. For ArchLinux it could be installed from AUR https://aur.archlinux.org/packages/stockfish. Otherwise you could download it from official site and provide path to stockfish-binary in pgn_to_db.py.

   python <unpacked_pgn_file>

5. Run training process for both sides(black and white) sequentially from project directory

   cargo run --release train --dataset=py/chess_db_white.db --ocl --out=net_white --epochs=55

   Then we need to train network evaluate positions from black side

   cargo run --release train --dataset=py/chess_db_black.db --ocl --out=net_black --epochs=55

   --ocl flag enables OpenCL computations on GPU --epochs spicifies number of epochs, could be modified.

6. Play with bot using some trained snapshot

   cargo run --release play --state_white=<net_white...state> --state_black=<net_black...state> --ocl --unicode --depth=4

   --unicode flag enables pretty unicode board state displaying

   --depth specifies the depth of move search for alpha-beta algorithm. I suggest to use values from 1 to 4. Big depth values(>4) will make the algorithm take a lot of time to search best move.

GIF