Title: Learning to control: from one linear system, to many, to nonlinear ones
Farnaz Adib Yaghmaie
Abstract: How far can one go starting from linear quadratic control? In this talk I begin with our work on model-free reinforcement learning for LQ problems. Then I follow two directions developed in my group. The first moves from one linear system to many. Using domain randomization, we design a single controller for a family of uncertain plants, through optimization, policy iteration, and learning from offline data. The second moves from linear to nonlinear systems. With Koopman operators, nonlinear dynamics become linear in a lifted space, and we study approximation errors and control under parametric uncertainty. I close by showing how the two directions meet: an uncertain nonlinear system lifts to a family of linear models, bringing us back to LQ.
Time: Fri 2026-11-06 11.00 - 12.00
Location: Seminar room 3418
Bio: Farnaz Adib Yaghmaie is an Assistant Professor and Docent in the Division of Automatic Control at Linköping University, where she leads the learn2decide group. She received her PhD from Nanyang Technological University, Singapore, in 2017. Her current work covers reinforcement learning, learning from demonstration, Koopman methods for nonlinear systems, and foundation models combined with reinforcement learning.