Seminars are held Wednesdays, at 3:30 pm, in Seaver Science Library, Room 150 (SSL 150), unless otherwise noted. Refreshments are served at 3:00 pm. Call (213) 740-8762 for further information.
Robustness Guarantees with Learning in the Loop
Postdoctoral Scholar in EECS
Unversity of California at Berkeley
Motivated by the push to incorporate learning enabled components into safety-critical systems such as autonomous ground and aerial vehicles, we seek to understand what types of robustness guarantees can be provided when the output of a machine learning (ML) algorithm is used within a robust control (RC) algorithm. ML algorithms typically view the world as stochastic, and provide probabilistic guarantees on estimation or prediction accuracy, whereas RC algorithms view the world as adversarial, and provide worst-case deterministic guarantees on system stability and performance. How then, if at all, can methods from these two fields be leveraged together? In this talk we propose the “Coarse-ID Control” framework as a way of bridging the gap between ML and RC, and apply it to a classical problem from the control literature, the Linear Quadratic Regulator (LQR), with the added twist that now the system dynamics are unknown. We provide, to the best of our knowledge, the first end-to-end, finite data robustness and performance guarantees for learning and control in an LQR problem that do not require restrictive or unrealistic assumptions. A key technical tool used in deriving this result is our recently developed System Level Approach (SLA) to Controller Synthesis. The SLA provides a transparent connection between system structure, constraints, and uncertainty and their effects on controller synthesis, implementation, and performance — we exploit these properties to combine results from contemporary high-dimensional statistics and robust controller synthesis in a way that is amenable to non-asymptotic analysis. We then show how the solution to the “Learning-LQR” problem can be incorporated into an adaptive polynomial-time algorithm with non-asymptotic convergence rate guarantees. We conclude with an overview of current experimental and simulation based validation of the Coarse-ID Control framework, and of our ongoing efforts to extend our theoretical guarantees to broader classes of systems and control algorithms.
Nikolai Matni is a postdoctoral scholar in EECS at UC Berkeley working with Benjamin Recht. Prior to that, he was a postdoctoral scholar in Computing and Mathematical Sciences at the California Institute of Technology. He received the B.A.Sc. and M.A.Sc. in Electrical Engineering from the University of British Columbia, and the Ph.D. in Control and Dynamical Systems from the California Institute of Technology in June 2016 under the advisement of John C. Doyle. His research interests broadly encompass the use of learning, layering, dynamics, control, and optimization in the design and analysis of safety-critical data-driven systems. He was awarded the IEEE Conference on Decision and Control 2013 Best Student Paper Award, the IEEE American Control Conference 2017 Best Student Paper Award (as co-advisor), and was an Everhart Lecture Series speaker at Caltech.
Wednesday, August 22, 2018
Seaver Science Library, Room 150 (SSL 150)
Refreshments will be served at 3:15 pm.