Robustness of classifier-in-the-loop control systems: A hybrid-systems approach

This paper studies continuous-state dynamical systems that use a classifier to determine the control input. Since the classifier output belongs to a finite set, the feedback control is a piece-wise constant function of the state. We therefore model the closed-loop system as a switched system. The decision surfaces of the classifier in the feature space dictate determine the switching surfaces in the state space. Therefore, the classifier affects the stability of the closed-loop system. Any analysis of the nominal closed-loop system may not be valid in states or environments that the training data for the classifier do not represent. We propose techniques to determine when the stability of the nominal system can be extended to unseen states and/or environments.