PAWN: Piece Value Analysis with Neural Networks

2026-03-01 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Gaming & Interactive Media, Data Science & Analytics · Depth: Expert, extended

Summary

The PAWN (Piece Value Analysis with Neural Networks) project introduces an enhanced MLP-based piece value predictor for chess, significantly improving accuracy by incorporating full board state context via a CNN-based autoencoder. Using a dataset of over 12 million piece-value pairs from Grandmaster-level games, with ground-truth labels generated by Stockfish 17, the MLP+CNN architecture reduces validation mean absolute error (MAE) by 16% compared to context-independent MLP systems. The best-performing model achieves an MAE of 65.45 centipawns (approximately 0.65 pawns) on the Dataset TF, which comprises Classical games from 2023. This improvement highlights that encoding the entire problem state provides useful inductive bias for predicting individual component contributions, addressing the challenge of a piece's value depending on its spatial relationships with all other pieces on the board.

Key takeaway

For AI Scientists and Machine Learning Engineers developing predictive systems for complex, interdependent domains, integrating a vector representation of the entire problem state as context is crucial. Your models will achieve substantially higher accuracy and better generalization by capturing global relationships, as demonstrated by the 16% MAE reduction in chess piece value prediction. Consider exploring graph neural networks or transformer-based architectures for richer board representations and improved performance in future work.

Key insights

Full board state context, via CNN autoencoders, significantly enhances chess piece value prediction accuracy.

Principles

• Individual component value depends on global context.
• Ablation-based definition quantifies piece contribution.

Method

A CNN autoencoder generates latent board representations, which are then concatenated with piece-specific features and fed into an MLP for relative piece value prediction, trained with Huber loss and Z-score normalization.

In practice

• Use CNN autoencoders for contextual feature extraction.
• Employ Huber loss for robust outlier handling.
• Consider game-level data splits to reduce overfitting.

Topics

• Chess Piece Value Prediction
• Convolutional Neural Networks
• Autoencoders
• Multi-Layer Perceptrons
• Stockfish 17

Code references

• ethanjtang/PAWN
• asdfjkl/nnue
• ynasu87/nnue
• official-stockfish/WDL_model

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• PAWN: Piece Value Analysis with Neural Networks — Encoding full problem state via latent representations significantly improves individual component value prediction.
• [P] Vibecoded on a home PC: building a ~2700 Elo browser-playable neural chess engine with a Karpathy-inspired AI-assisted research loop — An AI-assisted research loop can rapidly develop high-performance neural chess engines on consumer hardware.
• Chess-World-Model: A 10M-Game Benchmark for Exact State Tracking from Chess Move Sequences — Large-scale, out-of-distribution benchmarks reveal state-tracking failures in models that scale alone conceals.
• The Math That Kills Trillion-Parameter AI Models — Knowledge graphs can serve as implicit reward models, enabling AI to learn logical deduction and generalize reasoning.
• Trained transformer-based chess models to play like humans (including thinking time) [P] — Transformer models can emulate human chess play and thinking times using large-scale Lichess data.
• TableNet A Large-Scale Table Dataset with LLM-Powered Autonomous — An LLM-powered multi-agent system autonomously generates diverse, high-quality table datasets for robust Table Structure Recognition.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.