UCI compatible deep neural network chess engine.
Hydra contains two main components:
- Parallel UCT search.
- NN position evaluator.
The project is built on:
- The libchess library for board representation and move generation.
- Facebook's PyTorch library for model backend.
- The AlphaZero paper: Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm
- Arman Maesumi's paper on Playing Chess with Limited Look Ahead
- And various other great internet resources.
The weights found in the repository were trained with 21 million depth 12 Stockfish evaluated positions provided by Mr Maesumi.
First install:
- CUDA v10.2
- cuDNN v8.0.2
- libtorch v1.6.0
Then run:
git clone https://github.com/ABigPickle/HydraChess.git
cd HydraChess
mkdir build
cd build
cmake -DCMAKE_PREFIX_PATH=/path/to/libtorch ..
cmake --build .
Just run with:
./HydraChess -train </path/to/dataset.csv>
Which will train the network with the parameters defined the config.hpp. The dataset should be formatted as a CSV (comma seperated values) file like so:
fen,evaluation
fen,evaluation
...
fen,evaluation
Where each line represents a training example. The fen string should be a game state in FEN format and the evaluation should be a single decimal value from [-1, 1] where 1 is winning for the current side to move and vice versa.
Reinforcement learning is currently not implemented. If you want to take a crack at it, please submit a pull request!
After 25 epochs of training the model on the Stockfish evaluation dataset, the estimated play strength is about 1600 elo.
(Intel i7-7700k, Nvidia GTX 1060 6gb, 2 threads, 16000 iterations per-thread)
Engine (black) against Stockfish level 4 (1600-1700 elo):
Engine (black) against Stockfish level 3 (1300-1400 elo):
