icyElephant icon indicating copy to clipboard operation
icyElephant copied to clipboard

a chinese chess engine using conv neural network with zero look-ahead

Icy Elephant

Like Go, Chinese chess is an ancient chess game, After I trained a CNN model to play Go ( https://github.com/bupticybee/icygo ), I wonder if the similar approach can be used to play chinese chess.

So I trained a CNN to do the same thing in chinese chess. I played with it, Althrough the network preduces some interesting result, it is not strong enough, in the 10 games I played with the Neural network, I win all of them by large margin.

Here are the first five move in a CNN self-play:

The data I use can be downloaded here: https://pan.baidu.com/s/1JCmweZUREJxjIMXQlL-o9g

Follow the code in chess_policy_resnet10.ipynb to train the model. After you trained your model, follow the code in play_against_computer.ipynb to play with the model you train.

Have Fun