Moonfish is a chess engine I developed in Python a few years ago to understand how engines work under the hood. The code favors simplicity and readability over performance optimization.
GitHub: https://github.com/luccabb/moonfish (MIT License)
The engine implements:
- Negamax
- Layer-based Parallelization: Distributes work at specific search depths (L1P, L2P algorithms)
- Lazy SMP
- Move Ordering: MVV-LVA (Most Valuable Victim - Least Valuable Attacker)
- Null Move Pruning
- PeSTO Evaluation Function with Tapered Evaluation
- UCI protocol
- Integrates with lichess bot platform
- Web API
- Uses Cerebellum as opening book
- Endgame tablebases support
- Distributed via PyPI, you can access the engine from your custom python code, check the README
- Bratko-Kopec test suite
- Custom test suite to ensure basic functionality. Not sure how much ELO it tests for, but if these tests are passing it your custom engine search implementation is likely not super off. If it does fail then your search algorithm _likely_ has a problem
- You can control how the engine behaves via CLI arguments, `moonfish --help` to check all options.
On Performance:
- ~2000 Elo when tested against lichess stockfish bots.
- it beats stockfish lvl 5 ~2000 Elo.
- mostly loses to stockfish lvl 6 ~2300 Elo.
- When testing online on lichess against other engines it performs at ~1700 Elo
- The above is when running on a Macbook M1 Pro, this will vary based on hardware and parameters passed to the engine.
- No time control implemented—deeper searches take proportionally longer
For a list of resources and inspirations that helped shape Moonfish, check out the references in the repository.