CrazyAra is an open-source UCI chess engine that leverages deep learning and Monte Carlo Tree Search (MCTS) to play chess variants. It targets chess enthusiasts and researchers interested in AI game playing, offering a powerful engine trained on human data and reinforcement learning.

How It Works

The engine combines a neural network for move evaluation and policy prediction with an MCTS algorithm, inspired by AlphaZero. It features both a Python-based initial version and a more performant C++ implementation. The C++ version utilizes libraries like MXNet and TensorRT for efficient deep learning inference, and Multi Variant Stockfish for board representation and move generation.

Quick Start & Requirements

• Installation: Pre-built binaries are available for Linux and Windows with NVIDIA GPUs (CUDA 11.3, cuDNN 8.2.1, TensorRT-8.0.1) or Intel CPUs (MXNet 1.8.0, Intel oneAPI MKL 2021.2.0). Mac support uses MXNet and MKL.
• Dependencies: Requires a UCI-compatible GUI. GPU acceleration is recommended for optimal performance.
• Resources: Model files need to be downloaded and placed in a specified directory.
• Documentation: Wiki provides installation and usage details.

Highlighted Details

• Supports multiple chess variants including Crazyhouse, Chess960, King of the Hill, and Three-Check.
• Trained on human games from lichess.org and via reinforcement learning.
• Published research papers and a master's thesis detail its methodologies.
• Includes custom artwork generated with a Stable Diffusion model.

Maintenance & Community

The project is actively developed, with multiple academic publications and theses associated with it. Links to related projects and research are provided.

Licensing & Compatibility

Licensed under GNU General Public License v3.0 (GPL v3). This is a copyleft license, requiring derivative works to also be open-sourced under GPL v3.

Limitations & Caveats

The engine's performance is heavily dependent on the quality and size of the trained neural network models. While binaries are provided, compiling from source may require specific versions of dependencies like CUDA and MXNet.