Abstract: Learning algorithms are increasingly deployed for control and decision-making in autonomous systems. Despite their significant computational benefits, ensuring the safety and reliability of such learning-enabled systems is challenging due to the high dimensionality of the learning models and the inherent nonlinearity of system dynamics. In this talk, we leverage tools and techniques from control theory to develop theoretical and algorithmic methods for certifying the safety of learning-enabled systems. Our approach studies safety through a reachability lens, leveraging mixed monotone theory to decouple the cooperative and competitive effects of control inputs on the reachable sets of these systems. This decoupling enables the integration of state-of-the-art machine learning verification algorithms into our reachability framework, resulting in fast and computationally efficient methods for the safety verification of learning-based systems. 

Bio: Saber Jafarpour is a Research Assistant Professor in the Department of Electrical, Computer, and Energy Engineering at the University of Colorado Boulder. Before that, he was a postdoctoral research fellow at the Georgia Institute of Technology and a postdoctoral research fellow at the University of California Santa Barbara. He completed his Ph.D. in the Department of Mathematics and Statistics at Queen’s University, Canada, in 2016. His research interests include safe control and learning in autonomous systems with applications to robotic systems and multi-agent cyber-physical networks