PAWN

Piece value Analysis With Neural networks

TLDR

Incorporating intermediate position representations derived using a CNN autoencoder as additional inputs to MLP-based chess piece relative value prediction systems significantly increases their accuracy.

TLDR^2

All ♝s on g7 are created equal, but some ♝s on g7 are more equal than others.

Datasets and Models

We open-source Dataset MC-large and TF along with the best MLP and MLP+CNN model configurations trained for both datasets.

Both datasets also include many other helpful and unused columns such as:

• FEN + evaluation pairs for anyone wanting to optimize position evaluations for chess engines
• Misc. metadata like side-to-move, material strings, and opening code/name for data analysis/visualization

Dependencies

Machine Learning (+400% stock price) Stuff

pip install torch torchvision torchaudio scikit-learn

torchvision and torchaudio are likely redundant (I was experimenting with graphs/GNNs for piece value prediction but gave up halfway)

Chess Stuff

pip install python-chess stockfish

General Utility

pip install numpy pandas psutil pyarrow tqdm

Please reference shell scripts included in each folder for additional information about how to run each step!

How to Run This Code

PGN → Piece Value Dataset

1. Gather a selection of chess games/positions in a single file (using some database like the Lichess Open Source Games Database or ChessBase).
2. Download Stockfish (or some other version of it/another chess engine).
3. Run python -u pgn_to_piecevals.py

Piece Value Predictor Training

1. Run python -u pval_train_val_split.py
2. Run python -u train_all_models.py

Miscellaneous Files

• ./pval_stats — contains a helper file used with slurm output files to calculate the number of timeouts per worker (when evaluating positions for calculating piece values).
• ./pvp_example — contains a Jupyter Notebook used to generate figures and piece values used in the main PAWN paper for Figures 1, 3, and 5.
• ./sample_run — contains input/output files from a sample run completed on the Sol supercomputer using 32 games (sample_games.pgn) from GM Garry "Chess" Kasparov's 1985 simul against 32 chess computers (in which he scored 32-0), sourced from: chessprogramming.org/Kasparov_Simul_vs_32_Micros_Hamburg_1985

Citation

@misc{tang2026pawnpiecevalueanalysis,
      title={PAWN: Piece Value Analysis with Neural Networks}, 
      author={Ethan Tang and Hasan Davulcu and Jia Zou and Zhongju Zhang},
      year={2026},
      eprint={2604.15585},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2604.15585}, 
}