Towards Robustness Evaluation of Models for Cyber-Physical Systems

Abstract:
This paper presents a novel approach to evaluate robustness of models for Cyber-Physical Systems (CPS). We consider a CPS to be a hybrid system and learn models from data using decision tree regressors. We define robustness as the capability of the learned model to maintain its performance under varying unseen conditions. By systematically incorporating a perturbation process and calculating key statistical measures on the differences in model predictions, our method provides a comprehensive evaluation framework when modeling CPS. This approach enhances our understanding of a model’s robustness in the face of uncertainties, ultimately contributing to the development of more robust models for CPS. We demonstrate the application of our method for a simulated temperature control system.