PhD Thesis Defense

Zoom: https://caltech.zoom.us/j/82674474782

Increasing adoption of learning-based controllers in autonomous systems presents significant challenges to ensuring reliability, that traditional model-based approaches often enjoy. These controllers are particularly vulnerable to distribution shifts, mismatches between training and deployment data, which can be catastrophic in safety-critical applications. While classical robust control offers protection, it is often overly conservative, discarding valuable statistical information in favor of worst-case guarantees.

To address these limitations, my talk explores two complementary approaches that integrate statistical modeling into control design. First, I present distributionally robust control as a principled method for guarding against temporally correlated distribution shifts. By modeling uncertainty with a Wasserstein ball around a nominal distribution, I derive a tractable formulation that balances robustness and performance in a data-informed manner. However, these methods face scalability issues with longer time horizons. In the second part of the talk, I introduce the nonrational control framework, which enables the synthesis of scalable and stabilizing controllers for a broad class of infinite-horizon problems, including distributionally robust control. This framework allows for the design of finite-dimensional controllers even when the optimal solution is infinite-dimensional (nonrational). I demonstrate its generality through several examples, including mixed H2/H∞ control.