8 ♜ ♞ ♝ ♚ . ♝ . ♜
7 ♟ . # ♟ # ♟ ♟ ♟
6 . # ♟ # ♟ # . ♞
5 # . # . # . # .
4 . ♟ . # . # ♙ ♙
3 ♙ ♙ ♙ ♛ # ♙ # ♗
2 . # ♕ ♙ ♙ ♔ . #
1 ♖ ♘ ♗ . # . ♘ ♖
a b c d e f g h
A GPU-accelerated Neural Network Chess Engine written in C++ with Tensorflow. The engine uses a Monte Carlo Tree Search (MCTS) reinforement learning model, with the search process pruned by a Convolutional Neural Network. This is the same approach used in the famous paper Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm by Google's DeepMind Team.
This engine will not give Stockfish or Leela Chess Zero a run for its money, but it is a fully-functional neural network chess engine, built with simplicity and compatibility in mind. The engine can be built and run on any system with a C++17-compatible compiler.
This project does not require an installation of the C++ Tensorflow distribution. Instead, it uses the Tensorflow C API, which is much easier to install.
The easiest and most reliable way to get started is to run the model in a Docker Container. First you will need to make sure Docker is installed on your system.
To build the mcts-chess Docker container, run the following command in the project directory:
docker build --no-cache -t mcts-chess:latest .
Building the image may take some time. Once the image builds, it should appear with the tag mcts-chess:latest when you run docker images.
To start the container with a bash shell prompt, run:
docker run -it --rm --gpus all mcts-chess bash
If you do not want to use gpu acceleration, omit the the --gpus all flag. Once the container starts, you should be inside the /mcts-chess directory. To start the program, simply run:
./bin/main
To build the project from source, you will first need to install the Tensorflow C API. If you want to use GPU acceleration (compatible with NVIDIA GPUs), be sure to download the correct Tensorflow binary with GPU support.
Installing the C++ Tensorflow distribution from source is not needed, as this project uses cppflow, a self contained library that calls the C API.
Compiling the source requires GNU Make and GCC to be installed, with at least C++17 compatibility.
To compile, simply run the make command (and hope for the best).
The easiest way to run the notebooks is through the docker container (see above). Simply starting the docker container with forwarding to port 8888 will start the Jupyter notebook server:
docker run -it --rm --gpus all -p 8888:8888 mcts-chess
The notebook server can be accessed in your browser at http://127.0.0.1:8888/?token=... . You will need to need to supply the access token that is displayed when the container starts.
You can omit the --gpus all flag if GPU support is not desired. For troubleshooting, see Tensorflow's Docker Page.
To run the Jupyter notebooks in a regular python environment, you will need to install the dependencies with:
pip3 install --upgrade notebook tensorflow
To start the notebook server in the current directory, run:
jupyter-notebook .
- Create a Docker Image with Dockerfile
- Add a PGN file parser for training on expert games.
- Include a better Pre-Trained Model.
- Add Better Documentation to Jupyter Notebook.
- Include some examples of different nnet architectures (e.g Residual Nets)
- Implement Rick's Method for rollout of states.