Interpretable machine learning analysis of nonlinear error amplification under time pressure and positional ambiguity...

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Game Theory

Background:

• Sequential decision-making systems like chess are susceptible to cognitive constraints.
• Time pressure and positional ambiguity are key factors affecting human performance.
• The interactive and nonlinear effects of these constraints require quantification.

Purpose of the Study:

• To quantify the interactive and nonlinear relationship between time pressure and positional ambiguity on blunder probability in chess.
• To develop a quantitative framework for context-sensitive error modeling in human-AI systems.

Main Methods:

• Analysis of 39,922 ply-level chess positions from elite players on Lichess.
• Utilized Stockfish 14.1 engine evaluation for positional ambiguity assessment.
• Applied cluster-robust logistic regression and histogram-based gradient boosting (HGB) models.
• Employed permutation importance and SHAP values for explainability.

Main Results:

• Blunder probability increases nonlinearly with combined low time and high positional ambiguity.
• The Amplification Index (AMPIND) quantifies an approximate 5.1% error multiplier under extreme conditions.
• HGB model demonstrated high discriminative performance (AUC [Formula: see text]), with ambiguity and time pressure as key predictors.

Conclusions:

• Human errors in chess are concentrated under specific combinations of time pressure and positional ambiguity.
• Findings support a quantitative framework for context-sensitive error modeling.
• Results can inform the development of adaptive decision support systems in human-centered AI.