Learning models and synthetizing controllers for nonlinear systems directly from data (Prof. dr. ir. Roland Toth, Eindhoven University of Technology)
Systems & Control Seminar
Abstract
In this talk, first an overview of some recent results on learning models of nonlinear systems is given. We show that a Nonlinear (NL) stochastic system in an innovation form can be separated to a deterministic process part and a nonlinear stochastic noise model like in the classical identification setting in the Linear Time-Invariant (LTI) case. Then, we provide a unifying model parametrization of these parts where dynamics from LTI, through Linear Parameter-Varying (LPV) behavior till full NL aspects can be captured by the used Artificial Neural Network (ANN) State-Space (SS) models both in the discrete and continuous-time cases. Next we show how to achieve training of the models with an identification method that (i) benefits from both single and multi-shooting formulation of the cost, (ii) co-estimates a state encoder in the multiple-shooting case, which can be seen as a state- observer, (iii) supported by (group)-Lasso and l2 regularization schemes for model structure learning and sophisticated initialization, (iv) benefits from a JAX-based optimization backend where ADAM and L-FBGS are combined for a lighting fast model training, and last but not least, there is proven stochastic consistency guarantee to obtain an exact model of the system under mild conditions as data goes to infinity. Then we show, how to incorporate existing base-line models and first-principle knowledge into the model learning process in terms of model augmentation using a flexible Linear-Fractional Representation (LFR) form, where the learning process can discover what is the best suiting configuration of the existing model w.r.t. learning component. Furthermore, we provide parametrization of the resulting model that can ensure well-posedness and overall stability of the model estimate in a constraint-free sense. Last, but not least, we revisit our results for behavioral direct data-driven control for LPV systems and give an overview of the existing schemes and how they can provide equilibrium-free stability and dissipativity guarantees when applied to nonlinear systems.
Biographical information
Roland Toth received his Ph.D. degree with Cum Laude distinction at the Delft University of Technology (TUDelft) in 2008. He was a post-doctoral researcher at TUDelft in 2009 and at the Berkeley Center for Control and Identification, University of California in 2010. He held a position at TUDelft in 2011-12, then he joined to the Control Systems (CS) Group at the Eindhoven University of Technology (TU/e). Currently, he is a Full Professor at the CS Group, TU/e and a Senior Researcher at the Systems and Control Laboratory, HUN-REN Institute for Computer Science and Control (SZTAKI) in Budapest, Hungary. He is Senior Editor of the IEEE Transactions on Control Systems Technology and Associate Editor of Automatica. His research interests are in identification and control of linear parameter-varying (LPV) and nonlinear systems, developing data-driven and machine learning methods with performance and stability guarantees for modeling and control, model predictive control and behavioral system theory. On the application side, his research focuses on advancing reliability and performance of precision mechatronics and autonomous robots/vehicles with nonlinear, LPV and learning-based motion control. He has received the TUDelft Young Researcher Fellowship Award in 2010, the VENI award of The Netherlands Organization for Scientific Research in 2011, the Starting Grant of the European Research Council in 2016 and the DCRG Fellowship of Mathworks in 2022 and he is a Fellow of IEEE.
Termin
21. Mai 202610:30 - 11:30
