Boke is an open source GTP engine for 9x9 Go (Baduk/Weiqi) written in pure Python. A version written in C is in development here.
We use a simplified implementation of the Monte Carlo Tree Search variant pioneered by AlphaGo1. For more details, see our report.
To install bokego run
git clone https://github.com/meiji163/BokeGo && cd BokeGo
python setup.py install
You can then use boke with a GUI that supports GTP engines, such as GoGui or Sabaki.
For example, to use Boke with Sabaki go to Manage Engines > Add and enter the absolute path to boke.py.
Specify the weights or time limit with optional flags
usage: boke.py [-h] [-t SEC] [-r R] [-p PATH] [-v PATH] [--simulate]
BokeGo v0.3 GTP Engine
optional arguments:
-h, --help show this help message and exit
-t SEC time limit in seconds for each move
-r R number of rollouts per move
-p PATH path to policy weights
-v PATH path to value weights
--simulate enable simulations to end of game (slow)
If you have a CUDA-enabled GPU, you can move the neural network computations to the GPU by setting
USE_CUDA = True
in boke.py.
Note this requires a PyTorch build with CUDA
GnuGo (B) vs. Boke (W)