Directional Conformal Uncertainty Quantification from Learned Model Discrepancy
TL;DR
A conformal prediction framework for quantifying the error of physics-based predictors used in control, where simple models are preferred for synthesis, certification, and real-time use, and is extended to the multivariate case via a Minkowski-gauge score yielding a jointly calibrated disturbance set.
Abstract
We propose a conformal prediction framework for quantifying the error of physics-based predictors used in control, where simple models are preferred for synthesis, certification, and real-time use. Because these models are selected for compatibility with the intended application rather than for maximal predictive accuracy, their error combines process noise with a state-dependent discrepancy. A data-driven discrepancy estimate defines an asymmetric nonconformity score: errors consistent with the learned discrepancy are penalized less than equally large in the opposite direction. The sets remain in the nominal model's error coordinates and are physics-consistent, i.e., they contain a ball at the origin. The construction is agnostic to the discrepancy model (kernel, neural-network, or other), preserves finite-sample marginal validity under exchangeability, and provably narrows the interval over a characterizable state-input region. We further show that, for RKHS models, the power function provides a local confidence measure for adaptive score design and we extend the construction to the multivariate case via a Minkowski-gauge score yielding a jointly calibrated disturbance set.