ethanjtang/PAWN-piece-value-predictors

PAWN: Piece Value Analysis with Neural Networks

Overview

Best-performing MLP and MLP+CNN piece value prediction models from PAWN: Piece Value Analysis with Neural Networks.

We define piece value as the difference in Stockfish evaluation between the original position and the position with that piece removed.

Models

MLP (MC-Large) — Best MLP model trained on Dataset MC-Large (6,925 Magnus Carlsen games, 11.7M piece value entries).
MLP (TF) — Best MLP model trained on Dataset TF (7,656 GM-level Classical games, 12.3M piece value entries).
MLP+CNN (MC-Large) — Best MLP+CNN model trained on Dataset MC-Large.
MLP+CNN (TF) — Best MLP+CNN model trained on Dataset TF.

Datasets/Usage

Training Data — HF: ethanjtang/PAWN-datasets
Training Loop — GitHub: ethanjtang/PAWN/sample_run
Model Inference — GitHub: ethanjtang/PAWN/PAWN_demonstration.ipynb

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}, 
}