Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Study of the Proper NNUE Dataset

Daniel Tan, Neftali Watkinson Medina

TL;DR

The paper tackles the problem of poorly understood NNUE dataset construction by proposing a replicable pipeline to generate and filter quiet positions. It combines data from 44,000 grandmaster Xiangqi games with 30,000 synthetic self-play games, then uses margins $M_1$ and $M_2$ to prune tactically volatile positions, producing robust training data. A compact NNUE model with $3240$ inputs ($1620$ per side) and a $256$-unit hidden layer demonstrates substantial strength gains, achieving around $+100$ Elo and a $65\%$ win rate against a handcrafted baseline, with gains preserved when paired with a more complex evaluator. The work argues for a general, replicable dataset-generation approach that can extend to other chess variants and evaluation functions, potentially guiding future NNUE training pipelines.

Abstract

NNUE (Efficiently Updatable Neural Networks) has revolutionized chess engine development, with nearly all top engines adopting NNUE models to maintain competitive performance. A key challenge in NNUE training is the creation of high-quality datasets, particularly in complex domains like chess, where tactical and strategic evaluations are essential. However, methods for constructing effective datasets remain poorly understood and under-documented. In this paper, we propose an algorithm for generating and filtering datasets composed of "quiet" positions that are stable and free from tactical volatility. Our approach provides a clear methodology for dataset creation, which can be replicated and generalized across various evaluation functions. Testing demonstrates significant improvements in engine performance, confirming the effectiveness of our method.

Study of the Proper NNUE Dataset

TL;DR

The paper tackles the problem of poorly understood NNUE dataset construction by proposing a replicable pipeline to generate and filter quiet positions. It combines data from 44,000 grandmaster Xiangqi games with 30,000 synthetic self-play games, then uses margins and to prune tactically volatile positions, producing robust training data. A compact NNUE model with inputs ( per side) and a -unit hidden layer demonstrates substantial strength gains, achieving around Elo and a win rate against a handcrafted baseline, with gains preserved when paired with a more complex evaluator. The work argues for a general, replicable dataset-generation approach that can extend to other chess variants and evaluation functions, potentially guiding future NNUE training pipelines.

Abstract

NNUE (Efficiently Updatable Neural Networks) has revolutionized chess engine development, with nearly all top engines adopting NNUE models to maintain competitive performance. A key challenge in NNUE training is the creation of high-quality datasets, particularly in complex domains like chess, where tactical and strategic evaluations are essential. However, methods for constructing effective datasets remain poorly understood and under-documented. In this paper, we propose an algorithm for generating and filtering datasets composed of "quiet" positions that are stable and free from tactical volatility. Our approach provides a clear methodology for dataset creation, which can be replicated and generalized across various evaluation functions. Testing demonstrates significant improvements in engine performance, confirming the effectiveness of our method.

Paper Structure

This paper contains 14 sections, 2 equations, 4 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background
  3. Definitions
  4. Negamax
  5. Quiescence Search
  6. Centipawn
  7. Piece Square Tables
  8. Elo
  9. Platform
  10. Methodology
  11. Implementation
  12. Results
  13. Future Work
  14. Conclusion

Figures (4)

  • Figure 1: In this position, the rook checks the king. Because of the check, this is not a quiet position to be included in the dataset.
  • Figure 2: In this position, the knight can capture the rook for free. Because of the free capture of the rook, this is not a quiet position to be included in the dataset.
  • Figure 3: In this position, the knight can fork the king and rook, winning a free rook. Because of the fork, this is not a quiet position to be included in the dataset.
  • Figure 4: This position contains no significant changes in material imbalance, whether through a fork or tactic, only exchanges of relatively equal pieces. This is a quiet position useful to the dataset.