Data-Driven Methods for Learning and Control of Dynamical Systems

QC 20261010

Abstract

Engineering systems are often analyzed and operated using models that are incomplete, uncertain, or, in some cases, unavailable. In control engineering,vthe analysis of important system properties commonly relies on approximate parametric models, typically obtained through a prior system identification step. This thesis develops data-driven methods for characterizing behaviorally significant perturbations in dynamical systems and machine learning models using experimental observations rather than complete parametric descriptions.

The first part of the thesis concerns robustness analysis under model uncertainty. Given an unknown dynamical system, the objective is to estimatecontrol-relevant quantities directly from input-output data. In particular, we develop methods based on sequential input design to characterize structured perturbations that may destabilize a system and to quantify the mismatch betweena nominal model and the true plant from a closed-loop perspective. This information makes it possible to assess whether stability and performance conclusions obtained from the nominal model remain valid for the actual system.

In the second part of the thesis, we study the complementary problem of identifying minimal perturbations capable of producing a desired change in the behavior of a black-box model. Inspired by explainable artificial intelligence,we develop an adaptive algorithm for identifying feasible and actionable modifications using a limited number of model evaluations. We further apply this method to controller fine-tuning, where the objective is to identify small changes in controller parameters that lead to satisfactory closed-loop behavior while respecting operational constraints.

Together, these contributions show how task-relevant information about system sensitivity can be extracted directly from data, without reconstructing acomplete model of the underlying system.

Link to DiVA