fffasttime/gmk0

Learning 15x15 gomoku from zero!

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

gmk0's Issues

NN forward function implement in cpp

Hi! I just finished the work in https://github.com/syys96/Gmk0/tree/master/NN. Because I did my work in master branch, it's hard to pr directly. I hope it'll work and bug report will be welcome.

Bad mcts playout

Simlpe nn policy is stronger than mcts with playout.
threre should be bugs in mcts search.

what is gmkdata.txt?

can't train..

nn error

tensorflow.python.framework.errors_impl.FailedPreconditionError: Attempting to use uninitialized value bn11/batch_normalization/moving_mean

运行Main.py出现的，tensorflow1.6

NN forward throw exception

@syys96
Now cppsrc/Estimate.cpp calls NNforward for evaluation. But the result is not as expected.

Exception occored sometimes in this line.

Gmk0/cppsrc/NN/Network.cpp

Line 745 in 4d0aa8a

throw std::runtime_error("OpenCL self-check mismatch.");

At the same time, test shows the output seems randomly generated.
https://github.com/fffasttime/Gmk0/blob/master/cppsrc/Gmk0.cpp#L21

forget to remove invalid position

Gmk0/cppsrc/Evaluation.cpp

Line 63 in 94d6066

sum_policy += ret.first[i];