This is a Chess Bot that I made with deep learning that can play at a 1650 elo (top 4%) level. The architecture is my Dynamic Policy Net architecture, which computes a scalar score for each next legal board position and takes a softmax over those scores to get a probability distribution over all the possible moves it can make.
This bot depends on the following packages
numpy
matplotlib
python-chess
h5py
pytorch
pyjet
The first 4 packages can be installed through pip like so:
pip3 install PACKAGE_NAME
Installation for pytorch is included in their website. PyJet is my own custom front-end I wrote for PyTorch so I could use a Keras-like API for PyTorch without sacrificing its flexibility. Installation instructions are included in the PyJet Repo.
To get GPU speedup, you'll want to install CUDA and CudNN. If you're working off of an AWS EC2 machine or Google Cloud Compute Machine, you can use my 3-step installation.
You'll want to download pgn files of games to train the network on. I used games played by humans with >2000 elo from the FICS Games Database. Once you've downloaded all the games you want, run the following command:
python3 read_games.py PATH_TO_H5_OUT PATH_TO_EACH_PGN_FILE...
Once you have the dataset ready, you'll want to create a folder DeepJetChess/models to save models in. You'll also need to modify the FN_IN global in the train.py file to whatever you saved your compiled H5 dataset to. Once you've made the modifications you can run
python3 train.py
Once you've initialized your model with the supervised learning on human games, you'll need to modify line 187 in reinforce.py to load whatever the name of your saved supervised learning model is from the training step. Then you can just run
python3 reinforce.py
Each successive checkpoint will be saved as PrevNeti.state where i is the checkpoint number.
To test the model's elo, you can change the MODEL_NAME global (line 66) in elo_test.py to whatever model you want to test. Then run
python3 elo_test.py
and the script will print the moves corresponding to this online elo test where you'll have to manually enter the moves.