Aug 28 (Thurs) @ 10:00am: "Data efficiency and task-oriented exploration for data-driven methods in control, games, and decision making," Sean Anderson, ECE PhD Defense

Research Areas: Communications & Signal Processing, Control Systems
Research Keywords: Control, Game Theory, Data-driven Methods
Location: Engineering Science Building (ESB), Rm 2001 |  https://ucsb.zoom.us/j/9200743303

Abstract

Data-driven control and decision-making systems depend on high-quality data in order to perform well, yet in many applications collecting such data is expensive or impractical. Examples include vehicles operating on Mars, car racing with limited trial runs before a race, or board games where the state space is prohibitively large to explore completely. This defense presents recently developed methods for efficiently gathering data that directly improves task performance. The first part of the talk briefly reviews novel algorithms for control-oriented experiment design/identification. The second part will consider two-player zero-sum games solved via variations of Q-learning, including a novel algorithm that finds minimax solutions with only partial state space exploration—reducing computational demands while retaining strong guarantees. Finally, this defense addresses online decision-making in non-stationary environments, motivated by settings where the reward function depends on an adaptive adversary or environmental factors such as seasonal effects. The proposed algorithm leverages contextual information, includes regret bounds for two different types of context, and demonstrates strong empirical performance.

Bio

Sean Anderson is a Ph.D. candidate in the Department of Electrical and Computer Engineering at the University of California, Santa Barbara, advised by Professor João Hespanha. He received a B.S. from the University of California, Berkeley in Energy Engineering in 2018. He spent time in industry at New Sun Road, P.B.C. as a data scientist and controls engineer before starting his Ph.D. in 2020.

Hosted By: Professor João Hespanha

Submitted By: Sean Anderson <seananderson@ucsb.edu>