AlphaZero Gomoku

A multi-threaded implementation of AlphaZero

Features

  • Free-style Gomoku
  • Tree/Root Parallelization with Virtual Loss/LibTorch
  • Gomoku and MCTS are written in C++
  • SWIG wrap C++ extension
  • Update 2019.7.10: support Ubuntu and Windows

Args

Edit config.py

Environment

  • Python 3.6+
  • PyGame 1.9+
  • PyTorch 1.0+
  • LibTorch 1.0+
  • MSVC14.0/GCC6.0+
  • CMake 3.8+
  • SWIG 3.0.12+

Run

# Add LibTorch/SWIG to environment variable $PATH

# Compile Python extension
mkdir build
cd build
cmake .. -DCMAKE_PREFIX_PATH=path/to/libtorch -DCMAKE_BUILD_TYPE=Release
cmake --build

# Run
cd ../test
python learner_test.py train # train model
python learner_test.py play  # play with human

Pre-trained models

Trained 2 days on GTX1070

Link: https://pan.baidu.com/s/1c2Otxdl7VWFEXul-FyXaJA Password: e5y4

References

  1. Mastering the Game of Go without Human Knowledge
  2. Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm
  3. Parallel Monte-Carlo Tree Search
  4. An Analysis of Virtual Loss in Parallel MCTS
  5. A Lock-free Multithreaded Monte-Carlo Tree Search Algorithm
  6. github.com/suragnair/alpha-zero-general

GitHub