Training classifiers for feedback control with safety in mind

Abstract

The sensors of many autonomous systems provide high-dimensional and information-rich measurements. The state of the system is a part of this information, however it is challenging to extract it from such measurements. An autonomous system cannot use traditional feedback control algorithms without knowledge of the state. We propose computational algorithms for the analysis and synthesis of classifier-enabled control architectures. We show how to train classifiers based on criteria that relate to both learning from data and properties of the resulting closed-loop system. The approach to deriving these algorithms involves modeling the resulting closed-loop system as a piecewise affine differential inclusion. The training method is based on the projected gradient descent algorithm. An application of this method to a navigation problem for a mobile robot demonstrates the capabilities of this approach.