Gomoku implemention of reinforcement learning by AlphaZero methods using mxnet cpp-package.
- On windows, compiled Mxnet shared libray is needed, edit CMakeLists.txt to include and link correct directories.
- On linux, if you don't bother to compile Mxnet yourself, you can use python
pip install mxnet
to download mxnet, then edit build.sh, add library path to LD_LIBRARY_PATH. - In any way, it is indispensible to include correct header.
- Pay attention to your c++ compiler version, it must fully support c++11.
Enter gomoku <command>
to see subcommand help in detail.
These are common Gomoku commands used in various situations:
config Print global configure
train Train model from scatch or parameter file
play Play with trained model
benchmark Benchmark between two mcts deep players
The model supplied has 8x8 board size, 64 filters, 3 residual blocks,
trained on 1cpu for about 1.5 days, 9336 backward updates to the network.
Above shows a game played between human(first hand, represented by x
) and AI(represented by o
)
using pretrained model with command gomoku play 0 9336
.