Authors: A. R. AlMomani, Jie Sun, and Erik Bollt
Introduces entropic regression as a fundamentally new approach to nonlinear system identification that is inherently robust to outliers. By replacing least-squares fitting with an information-theoretic objective based on mutual information, the method selects the correct sparse model structure even when data is contaminated with heavy-tailed noise.
The key insight: mutual information measures statistical dependence rather than prediction error, making it insensitive to the magnitude of individual residuals, the exact property needed to resist outlier corruption.
Causal Inference & Discovery (Pillar 1), Explainable & Interpretable ML / AI (Pillar 2)