Seminars

View all Seminars  |  Download ICal for this event

On Robustness and Sample-Efficiency for Data-Driven Control

Series: Bangalore Theory Seminars

Speaker: Vinay Kanakeri, North Carolina State University

Date/Time: Aug 18 11:00:00

Location: CSA Auditorium, (Room No. 104, Ground Floor)

Abstract:
The growing reliance on data to control unknown dynamical systems raises two central questions: how do we ensure robustness to poor-quality data, and how do we use the available data efficiently? This talk addresses both questions for linear time-invariant (LTI) systems.

On robustness, existing finite-sample guarantees for linear system identification hinge on the process noise being Gaussian or sub-Gaussian. However, such assumptions can fail in practice due to sensor glitches, environmental disturbances, or adversarial interference. We instead assume only that the noise has a finite fourth moment, allowing for heavy tails. We introduce Robust-SysID, an algorithm that partitions trajectories into buckets, computes a least-squares estimate within each bucket, and boosts these weakly concentrated estimators via the geometric median. We show that Robust-SysID nearly recovers the sample-complexity bounds achievable under Gaussian noise, with the gap governed by the kurtosis of the noise distribution, a natural measure of tail heaviness. We extend these guarantees to settings with adversarially corrupted trajectories and to partially observed systems.

On sample-efficiency, we study personalized and collaborative policy optimization for a collection of clustered LTI systems with quadratic costs, where the dynamics, task objectives, and cluster memberships are all unknown and only zeroth-order cost feedback is available. Since collaborating across dissimilar systems can destabilize the resulting policies, we design a sequential-elimination clustering algorithm that lets agents safely identify and collaborate within their cluster, yielding faster convergence and lower per-agent sample complexity without incurring bias from system heterogeneity.

Together, these results take early steps toward a statistical learning theory for data-driven control under realistic, non-ideal assumptions on the data-generating process.


Microsoft Teams link

Link


Hosts: KVN Sreenivas, Sreeramji K S, Ritabrata Barat, Venkata Sai Nikhil Srivatsava Ayyadevara