Gmk0

Learn playing gomoku from zero.

Result of training

Training is available. Test training has started. See improve.md.

• After playing 150'000 games, Gmk0 is as strong as an ordinary alphabeta algorithm program.
• But it's still far away from top level AI (3-27 Yixin16).

Python version

Run pysrc/Main.py to selfplay or play with human, or run pysrc/Trainer.py to train a nn from selfplay data.

require:

• python3.6
• numpy-1.13
• tensorflow-1.4.0 or tensorflow-gpu-1.4.0

cpp version

cpp version is faster when selfplay. compile and run Gmk0.exe --mode selfplay --network "weight.txt" --playout 400 to generate selfplay data

The neural network need some library, see cppsrc/NN/README.md