chess_foehn

This is me trying my way with machine learning using my hobby, chess.

Idea

Instead of using images of boards or hard-coded engines like Stockfish, I’m feeding in raw PGN and FEN files (yep, those game logs with all the move information). The idea is simple:

• Train a model that understands chess moves directly from PGNs
• It should never make illegal moves (coz rules are built-in, Python-Chess helps)
• Try some pretraining tricks and forcing the model to guess them
• Seeing if a hybrid of transformer or Reinforcement learning works
• Later: see if it can play fast games (like blitz/bullet) where intuition matters more than brute-force search

This isn’t about beating Stockfish-like engines (they are far too powerful).
It’s more about exploring:

• Can ML learn patterns & style from millions of human games better than traditional chess engines without any methodical bias?
• Can it make creative, human-like moves?
• Can it move more like a human?, in contrast to the chess engines that make very illogical but accurate moves

Plans

• Explore the Searchless chess architecture
• Identify the bridges to RL
• What is the difference to AlphaZero?
• Build a simple interface to play against the model

Acknowledgments

This project is continuously getting inspired by the work of Ruoss et al. and their paper:

Ruoss, A., Delétang, G., Medapati, S., Grau-Moya, J., Wenliang, L. K., Catt, E., ... & Genewein, T. (2024). Amortized Planning with Large-Scale Transformers: A Case Study on Chess. 38th Conference on Neural Information Processing Systems (NeurIPS 2024). arXiv preprint arXiv:2402.04494.