Boke Go

Boke is an open source Go engine for a 9x9 board. It uses the PUCT (Prediction + Upper Confidence Bound for Trees) variant of Monte Carlo Tree Search. The prediction is provided by a policy network trained on human games and selfplay games.

The current version of Boke won 10/10 games against GnuGo 3.8 with 100 rollouts per move.

GnuGo (B) vs. Boke (W)

Todo

• Add pass/resign
• Improve/parallelize rollouts
• Port to C++

Play Boke

The requirements to test the current python version are PyTorch and NumPy (compatible versions listed in requirements.txt).

pip install -r requirements.txt
cd BokeGo/policy_net_py
python3 bokePlay.py --help
usage: bokePlay.py [-h] [-p PATH] [-c {W,B}] [-r ROLLOUTS] [--mode {gui,gtp}]

Play against Boke

optional arguments:
  -h, --help        show this help message and exit
  -p PATH           path to model
  -c {W,B}          Boke's color
  -r ROLLOUTS       number of rollouts per move
  --mode {gui,gtp}  Graphical or GTP mode

Warning: rollouts are currently single-threaded and very slow.