Interpretable modeling of human decision-making from user interactions in a dynamic stabilization task

This paper introduces the Extended Takagi-Sugeno Fuzzy Model Transformation (ETSFM), a non-parametric, data-driven framework grounded in the TS fuzzy model transformation combined with Close-to-Normal (CNO) and Inverted Relaxed Normal (IRNO) transformations. Without requiring pre-defined membership templates, ETSFM turns interaction data into compact, interpretable fuzzy linguistic rule sets capable of explaining human decision strategies in black-box and gray-box dynamic environments. Applied to an abstract stabilization task as a demonstrative example, the approach isolates three robust patterns in decision-making: reliance on self-stabilization regions (state-space zones where small or no inputs are sufficient to maintain control), transition zones where control effort steeply increases, and blind spots characterized by repeated failures. The systematic asymmetry uncovered in human control manifests as a directional bias: kinematically mirrored states between left and right elicit qualitatively different control decisions. The resulting rule sets are low-complexity yet high-fidelity, capturing the dominant decision logic while remaining auditable. Based on these results, the paper provides two contributions to the literature. Methodologically, we provide a principled workflow for transforming raw data into transparent, linguistically labeled rules without relying on pre-defined fuzzy variables or fuzzy membership templates. In terms of human decision behavior in black-box and gray-box environments, we uncover a systematic asymmetry in human control that endures across a range of dynamic conditions. In a practical sense, the framework can serve as a diagnostic tool for benchmarking human-AI decision systems: it is capable of mapping error-prone states, supporting targeted training and UI refinement, while permitting side-by-side comparison with an optimized controller. The method is transfer-ready to domains where decisions are inferred from interaction data under partial model knowledge, such as driver assistance or process control.