Monitor and Recover: A Paradigm for Future Research on Distribution Shift in Learning-Enabled Cyber-Physical Systems

TL;DR

The paper tackles distribution shift in learning-enabled cyber-physical systems and critiques inference-time detect-and-abstain strategies as overly conservative for real-time control. It introduces a monitor and recover paradigm comprising robust runtime safety monitoring and inference-time data transformation-based recovery to maintain reliable operation under shift. Two concrete instantiations are presented: a runtime monitor that leverages STL robustness and adaptive conformal prediction (with incremental learning) and SuperStAR, a reinforcement-learning-based data-transform pipeline that recovers the input distribution without retraining. Empirical results from simulated autonomous-vehicle scenarios and image-classification tasks show timely safety alarms under shift and partial recovery of performance, suggesting practical viability and a path toward deployment.

Abstract

With the known vulnerability of neural networks to distribution shift, maintaining reliability in learning-enabled cyber-physical systems poses a salient challenge. In response, many existing methods adopt a detect and abstain methodology, aiming to detect distribution shift at inference time so that the learning-enabled component can abstain from decision-making. This approach, however, has limited use in real-world applications. We instead propose a monitor and recover paradigm as a promising direction for future research. This philosophy emphasizes 1) robust safety monitoring instead of distribution shift detection and 2) distribution shift recovery instead of abstention. We discuss two examples from our recent work.

Figures (1)

• Figure 1: Example recent works following the monitor and recover paradigm. a) An example of robust safety monitoring lin2025safety. A trajectory predictor equipped with incremental learning predicts trajectories of the system states. On this prediction, a conformal region over the STL robustness score is computed using adaptive conformal prediction (ACP). A simple check indicates whether a safety violation is predicted to occur. b) Distribution shift recovery via SuperStAR lin2024dc4l. A sequence of transforms bring the data closer to the training distribution. The transforms are selected by a reinforcement learning agent. The data is represented by state $S_i$ and undergoes transforms $\mathbb{T}_j$ with reward $\mathcal{R}$ defined by the distance $d$ between the validation state $S_v$ and the current state $S_i$.