Abstract

This report studies player-specific behaviour modeling for elite chess players, extending human-centric chess AI beyond population-level and amateur-player settings. It focuses on a cohort of fourteen elite twentieth-century players, including world champions and top-tier candidates.

Building on the Maia-2 architecture, the project explores parameter-efficient approaches for individual adaptation, including player-specific embedding fine-tuning and a Mixture of Experts framework with Low-Rank Adaptation. Lightweight Monte Carlo Tree Search and Descent Search, constrained by Nucleus Pruning, are also integrated to improve tactical coherence while preserving human alignment.

The work also develops evaluation methods that move beyond top-1 move prediction accuracy. It introduces distributional evaluation on shared board positions and a stylistic manifold framework using AutoEncoders and UMAP to analyse player-specific chess behaviour in a way that better accounts for strategic variance and stochasticity in human play.

Reference

Sogliuzzo, L. (2026). Modeling Player-Specific Behaviors in Elite Chess. Research project report, ENSTA Paris. Hosted at UCLouvain, ICTEAM / INGI.

BibTeX

@techreport{sogliuzzo2026modelingPlayerSpecific,
  author      = {Sogliuzzo, Loris},
  title       = {Modeling Player-Specific Behaviors in Elite Chess},
  institution = {ENSTA Paris},
  year        = {2026},
  type        = {Research project report},
  note        = {Hosted at UCLouvain, ICTEAM / INGI},
  url         = {https://piette.info/eric/master/2026%20-%20Modeling%20Player-Specific%20Behaviors%20in%20Elite.pdf}
}