A Python chess engine using deep neural network evaluation and a quiescent variant of the Predictor + UCB applied to trees algorithm (PUCT).
Estimated ELO ~1600 but aspires to be the strongest Python chess engine!
Please contact me on Discord (@yoman417) or email walterliu417@gmail.com if there are any UCI bugs.
A list of UCI options can be found in uci_options.md in the docs folder.
For technical documentation on the engine, see technical.md in the docs folder.

Usage
In the Colab Notebooks folder, use parakeet_implementation.ipynb to run as a Lichess bot, and parakeet_tuning.ipynb to run games on fastchess. Parakeet uses the UCI protocol.

Building from source
On any terminal: Acquire the source code.

git clone https://github.com/walterliu417/parakeet.git
cd parakeet
python -m venv venv

Activate the virtual Python environment.
Linux:

source venv/bin/activate

Windows:

venv/Scripts/activate.bat

Install required libraries:

pip install numpy chess

On a GPU-enabled environment:

pip install onnxruntime-gpu

Otherwise:

pip install onnxruntime

Run the engine:

python parakeet.py

If you wish to obtain an executable, simply use

pip install pyinstaller
pyinstaller --onefile parakeet.py

Future plans

• Adaptive time management
• Gradient boosting networks (predicting the error between the current network and the target)
• More hyperparameter tuning instead of using random search
• C++ implementation for best performance.

Credits

• AlphaZero and Leela Chess for inspiring this project.
• Engine Programming Discord for all the help the friendly people there gave me.
• SOAP: Improving and Stabilising Shampoo using Adam. https://arxiv.org/abs/2409.11321
• Lichess Evaluation and Puzzle Databases.
• Fastchess, for testing the engine and tuning hyperparameters.
• ChatGPT, for basically giving me a crash course on supervised learning and acting as a platform to bounce ideas off of.