Dissipative Imitation Learning for Robust Dynamic Output Feedback

TL;DR

This paper introduces a dissipative imitation learning approach that stabilizes uncertain linear systems using output feedback, without requiring precise plant models, by leveraging QSR-dissipativity properties.

Contribution

It proposes a novel method for robust imitation learning that ensures stability in poorly modeled plants through dissipativity constraints on the controller.

Findings

Successfully stabilizes uncertain systems with parametric variations.

Enforces stability via open-loop QSR-dissipativity properties.

Applicable to systems with linear dynamic output feedback.

Abstract

Robust imitation learning seeks to mimic expert controller behavior while ensuring stability, but current methods require accurate plant models. Here, robust imitation learning is addressed for stabilizing poorly modeled plants with linear dynamic output feedback. Open-loop input-output properties are used to characterize an uncertain plant, and the feedback matrix of the dynamic controller is learned while enforcing stability through the controller's open-loop QSR-dissipativity properties. The imitation learning method is applied to two systems with parametric uncertainty.