ChessMimic: Per-Rating Transformer Models for Human Move, Clock, and Outcome Prediction in Online Blitz Chess
Researchers have developed ChessMimic, a system employing three specialized transformer models to predict human chess moves, thinking times, and game outcomes in online blitz games. Each model is trained for specific Elo rating bands to improve accuracy across different skill levels. The move prediction model surpasses existing benchmarks like Maia-2, while the outcome model achieves a 0.78 AUC, outperforming simpler logistic regressions. Although the thinking time prediction model is not state-of-the-art, it provides a usable signal for analysis. AI
IMPACT Introduces specialized models for predicting human chess performance, potentially aiding AI training and analysis.