Statistical Learning Theory for Safety-Critical Control: Fundamental Limits and Optimal Algorithms

Learning algorithms are poised to play an increasing role in the control of safety-critical systems. When these systems fail, the consequences can be dire.

Thus, for us to reliably use learning algorithms in safety-critical dynamical systems, a solid theoretical understanding, including failure modes, performance guarantees and fundamental limitations, is necessary.

Moreover, real-world applications go far beyond our understanding, and include sophisticated components such as deep neural networks and perception maps (e.g., cameras/LIDARs).

The aim of this proposal is to provide a principled theoretical understanding of the fundamental limitations of these learning algorithms in such nonlinear and imperfectly observed settings by leveraging tools from control-, information- and statistical learning theory.

A fine-grained theoretical understanding of how learning and control mesh sets the stage for us to successfully integrate machine learning into the control of modern safety-critical systems. We further expect to develop novel, more resilient algorithms appropriate for use in these systems.

The applicant will lead this line of work, which is organized into three thrusts centered around control and learning in the context of 1) nonlinear dynamic phenomena, 2) large-scale systems and 3) perception maps and representation learning.

Toward these goals, the applicant will be supported by leading professors and departments both in the U.S. (UPenn) and in Sweden (Lund).