Modeling Player-Specific Behaviors in Elite Chess
(2026)
Files
2026 - Modeling Player-Specific Behaviors in Elite.pdf
Open Access - Adobe PDF
- 2.81 MB
Details
- Authors
- Abstract
- (en) The development of chess models has historically prioritized superhuman strength, often producing uninterpretable strategies that diverge from human decision-making. While recent human-centric models, such as Maia-2 and Maia4All, successfully capture behavior at population and individual levels, they primarily target amateur to intermediate players, leaving the modeling of elite worldclass players as a significant challenge. Furthermore, current evaluations heavily rely on top-1 move prediction accuracy, a metric that fails to account for the natural diversity and stochasticity inherent in human play. This research extends individual behavior modeling to a cohort of 14 elite chess players active in the 20th century, encompassing world champions and top-tier candidates, while developing evaluation methodologies that respect strategic variance. Building upon the unified architecture of Maia-2, the proposed behavior generation pipeline employs parameterefficient techniques, specifically player-specific embedding fine-tuning and a Mixture of Experts (MoE) framework with Low-Rank Adaptation (LoRA). To mitigate tactical loop vulnerabilities while preserving human alignment, lightweight Monte Carlo Tree Search (MCTS) and Descent Search, constrained by Nucleus Pruning, are integrated. Finally, to move beyond deterministic metrics like move accuracy and Centipawn error, this work introduces a distributional evaluation pipeline across shared board positions (𝐽𝑆𝐷action) and a stylistic manifold framework (𝐽𝑆𝐷spatial) leveraging AutoEncoders and UMAP dimensionality reduction to compress 2304-dimensional board transitions into a 2D spatial grid.
- Affiliations
Citations
Sogliuzzo, L. (2026). Modeling Player-Specific Behaviors in Elite Chess. Sci-Comm Consulting Ltd. https://hdl.handle.net/2078.5/279530