Sep 24 (Thu) @ 10:00am: "Online Algorithms for Learning and Control under Safety Constraints," Spencer Hutchinson, ECE PhD Defense

Location: Engineering Science Bldg (ESB), Room 2001

Abstract

Learning and control algorithms are integral to modern engineered systems, such as power grids, autonomous vehicles, and robotics. In many settings, they must operate online under uncertain and evolving conditions while achieving performance objectives and satisfying strict operational constraints at all time steps. Accordingly, this dissertation studies online algorithms for learning and control that simultaneously achieve strong performance guarantees and ensure constraint satisfaction. We study this challenge within the frameworks of stochastic bandits, online convex optimization and adaptive control. We first introduce an approach for handling linear constraints in these settings, which we call ROFUL. This algorithm operates by first optimizing over an optimistic set (which overestimates the constraint set) and then scaling this optimizer in to the pessimistic set (which underestimates the constraint set). We show that this approach matches or improves the state-of-the-art regret bounds for linear constraints in stochastic bandits, online convex optimization and adaptive control. Then, for handling convex constraints, we introduce the Polyak feasibility steps algorithm, which applies projections on to the half-space formed by the first-order approximation of the constraint function. In the setting of online convex optimization with convex constraints, this approach enjoys optimal regret bounds and is the first to ensure constraint satisfaction at all time steps under first-order constraint feedback.

Bio

Spencer Hutchinson is a PhD candidate in Electrical and Computer Engineering at UC Santa Barbara, advised by Prof. Mahnoosh Alizadeh. Previously, he received the B.S. in Electrical Engineering from Colorado School of Mines in 2021, and M.S. in Electrical and Computer Engineering from UC Santa Barbara in 2024. His research focuses on designing and analyzing algorithms for machine learning, optimization, and control with a particular interest on problems with constraints.

Hosted By: ECE Professor Mahnoosh Alizadeh

Submitted By: Spencer Hutchinson | Email