Robust Control, Feedback and Learning

Professor Michael G. Safonov, University of Southern California, Electrical Engineering - Systems

A sound understanding of robust control, feedback and learning requires a theoretical framework that admits an open-eyed consideration of the implications of evolving measurement data. To address this issue, a data-driven behavioral formulation of the feedback control problem has been developed based on the paradigm of controller unfalsification. The theory lays a firmer foundation for feedback control, and it offers a means to directly identify controllers that are consistent with performance objectives and past experimental data --- possibly even before controllers are ever inserted in the feedback loop. The theory provides a precise quantitative characterization of the information-theoretic interplay between (1) data, (2) admissible controllers and (3) performance goals. Interestingly, neither plant models nor uncertainty models are essential to the application of the theory. Examples illustrate the use of unfalsified control methods for the design adaptive control switching procedures that allow one to quickly and reliably discover high-precision stabilizing controllers in real-time, directly from evolving closed-loop experimental data.