I came across DeepChess and decided to implement it to learn TensorFlow and to experiment with Deep Learning methods.
To play:
Install python-chess, and then from the main directory, run: python game.py
To train: This model was trained with:
- CUDA 7.5
- Tensorflow 0.10.0
Run python train.py to train the model on the data available in the folder 'pGames'
Some older network checkpoints can be found in the folder 'net'.
To mine a different dataset:
Run python get_data.py, but be sure to change the file name in the source code.
Some notes: The basic idea of the paper is that we can get a deep network to play chess by teaching it an evaluation function that takes in 2 positions and outputs the better one. The network can then be used in a modified Alpha-Beta pruning algorithm, where instead of comparing between two positions' evaluations (as numbers), we compare between the positions themselves.
The network consists of two main components, namely "Pos2Vec", and a fully connected MLP. The "Pos2Vec" component is a Deep Belief Network that consists of 4 stacked autoencoders, that are trained layer by layer, unsupervised. Two identical "Pos2Vec" components lay side by side and feed into a 4 layer MLP. The whole network is trained on 1 million pairs of positions, wherein the pre-training serves as the inital weights of the "Pos2Vec" components. The network trains on the CCRL dataset.