Abstract: The pervasive integration of edge intelligence has provided opportunities for proactive demand management in societal-scale cyber-physical systems (CPS) such as power or transportation networks. Existing approaches often involve soliciting customer preferences or navigating complex negotiation protocols to coordinate aggregate customer behavior before real-time operations, making them challenging to implement in the real world. This has ignited growing interest in data-driven methods that gradually and autonomously learn to manage customer demand, all without the need for direct communication with users. But such data-driven actions may lead to violation of the networks’ physical safety constraints as they are uncertain about the users’ response. Inspired by this application, this talk will explore two recent methods developed in our group which guarantee that data-driven prices (or other control actions) in safety-critical systems are network-safe regardless of model uncertainty, and discuss their accompanying theoretical performance bounds on the growth of regret. The first approach considers a safe bandit optimization formulation, whereas the second approach views this problem from a distributed optimization lens.

Bio:  Mahnoosh Alizadeh is an Associate Professor of Electrical and Computer Engineering at the University of California Santa Barbara. She received her Ph.D. degree in Electrical and Computer Engineering  from the University of California Davis in 2014. From 2014 to 2016, she was a postdoctoral scholar at Stanford University. Her research is focused on the design of network control, learning and optimization algorithms for societal-scale cyber-physical systems. She is a recipient of the National Science Foundation CAREER award and currently serves as an AE for the IEEE Transactions on Control of Network Systems and the IEEE Open Journal of Control Systems.

A few of us will remain to discuss the talk and catch up until ~3:00pm, so please stay if you are able.