Data-Driven Methods for Learning and Control of Dynamical Systems

QC 20261010

Abstract

Engineering systems are often analyzed and operated using models that areincomplete, uncertain, or, in some cases, unavailable. In control engineering,the analysis of important system properties commonly relies on approximateparametric models, typically obtained through a prior system identificationstep. This thesis develops data-driven methods for characterizing behaviorallysignificant perturbations in dynamical systems and machine learning models using experimentalobservations rather than complete parametric descriptions.

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

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

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

Link to DiVA