Workshop on data-driven control and analysis of dynamical systems

As indicated in the 2024-2028 GDR MACS roadmap, data-driven methods are on the rise in the control community, thanks in particular to recent advances based on the use of AI (reinforcement learning, representation of a system by an “informative” trajectory thanks to the Willems lemma and the methods derived from it). In automatic control, data-driven approaches are particularly interesting when the construction of a model is complex, and their development could enable to broaden the range of applications for control. Data-driven analysis methods have also been developed to analyse the resulting closed loops.

This topic is of growing interest to the international community, and the objective of this workshop is to gather the GDR MACS members developing such data-driven approaches. This workshop is part of a GDR MACS funded project. We are counting on this workshop to get a picture of the research and teaching activities on this topic at the french national level, and to initiate a national dynamic on these subjects.

• Cédric Join (CRAN, Université de Lorraine)
  Commande sans modèle et à modèle partiel
• Michel Fliess (LIX, École polytechnique & LJLL, Sorbonne Université)
  Toward Model-Free Predictive Control: Possible Connections with AI.
  Corresponding papers : Join, C., Delaleau, E. & Fliess, M. (2025) Model-Free Predictive Control: Introductory Algebraic Calculations, and a Comparison with HEOL and ANNs, Link
• Jean-Philippe Condomines(ENAC)
  Restricted Model-Free Control
• Antoine Girard (L2S, CNRS)
  Safe learning-based nonlinear model predictive control using data-driven set-valued models
  Corresponding papers :
  • Makdesi, A., Girard, A., & Fribourg, L. (2024). Data-Driven Models of Monotone Systems. IEEE Transactions on Automatic Control, 69(8), 5294-5309. Link
  • Makdesi, A., Girard, A., & Fribourg, L. (2023). Safe learning-based model predictive control using the compatible models approach. European Journal of Control, 74, 100849. Link
• Mayank Jha (CRAN, CNRS)
  Safe Reinforcement Learning with provable guarantees: Recent Advancements

This talk presents recent advances in Safe Reinforcement Learning (Safe RL) methodologies applicable to both discrete-time and continuous-time systems with provable guarantees. The talk will introduce reinforcement learning (Adaptive Dynamic Programming) and motivate Safe RL. Then, Safe RL approaches developed for discrete-time systems will be presented where safety is enforced by augumenting Control Barrier Functions (CBFs) into the reward structure, thereby ensuring the forward invariance of predefined safe sets. The presentation then turns to the Safe Exploration problem, highlighting how Input-to-State Stability (ISS) properties can be exploited to maintain safety during the exploration phase. In this context, the concept of Input-to-State Safety (ISSf) is introduced, offering a novel framework that promotes rich, risk-aware exploration while preserving formal safety guarantees, thereby enabling more informative exploration and improved tracking performance. The talk further covers recent developments addressing input saturation constraints, strategies for boundary-focused exploration, and model-free methods for system dynamics learning, contributing to a more robust and generalizable Safe RL framework.
Bio: Dr. Mayank Shekhar Jha obtained his PhD in 2015 at Ecole Centrale de Lille in France and has previously held post-doctoral research position at the Institut National des Sciences Appliquées de Toulouse (INSA Toulouse) France and Research Associate position at Rolls Royce Technology Centre at University of Sheffield, United Kingdom in 2017. Dr. Jha has has authored around 30 publications in prestigious international conferences and journals, leads a Work package (WP) in recently accepted project funded by National Agency for Research (ANR) in France titled “Self-Organizing, Smart and Safe heterogeneous Robots Fleet by collective emergence for a mission (SOS)”, has been Co-PI of three and PI of one industrially funded scientific projects with French National Space Agency (CNES) as well as Co-PI of one project with Dassault Aviation in last 5 years. Dr. Jha is an external collaborator and visiting researcher at NASA Ames Research Centre. Dr. Jha serves on editorial board of Scientific Reports, Nature and Associate Editor of Aerospace Science and Technology, Elsevier. Dr. Jha’s current research interests include Safe Reinforcement Learning, Deep Learning based prognostics and Adiabatic Quantum Computing.
• Oumayma Khattabi (PhD student at L2S under the supervision of Sorin Olaru and Matteo Tacchi)
  Data-driven stability analysis
• Riccardo Bonalli (L2S, CNRS)
  Safely Learning Controlled Stochastic Dynamics
• Stanislav Aranovskiy (IETR, CentraleSupelec)
  Dataset Management in Data-Enabled Predictive Control
• Pauline Kergus (LAPLACE, CNRS)
  Data-driven controller design in the Loewner framework
• Charles Poussot-Vassal (ONERA)
  The Loewner Framework, the Kolmogorov Superposition Theorem and the Curse of Dimensionality
  We extend the Loewner framework to multivariate (dynamical) functions (and systems) with an arbitrary number "n" of variables (or parameters). The presentation go through the following contributions: (i) first, we show how from a n-dimensional data tensor we can construct an approximating n-variable rational function; (ii) second, we detail a methodology to both reduce the computational burden and memory needs to achieve (i) with exact arithmetic, by replacing the null-space construction of a (large-scale) n-dimensional Loewner matrix, with a series of null-space construction of (limited size) 1-dimensional Loewner matrices; (iii) third, as a consequence of (ii), we show that the proposed approach achieves variables decoupling and provides a scalable numerical approach to the Kolmogorov Superposition Theorem applied to rational functions, bridging the gap between approximation theory and the neural network community. A collection of numerical examples and method comparison illustrates the effectiveness and scalability of the proposed method and its ability to tame the Curse of Dimensionality.
  Corresponding paper : On the Loewner framework, the Kolmogorov superposition theorem, and the curse of dimensionality, Athanasios C. Antoulas, Ion Victor Gosea, Charles Poussot-Vassal, preprint, 2025. Link
  
• Romain Postoyan (CRAN, CNRS)
  A hybrid systems framework for data-based adaptive control of linear time-varying systems
  We consider the problem of stabilization of unknown discrete-time linear time-varying systems using an adaptive data-driven control approach. The controller is defined as a linear state-feedback law whose gain is adapted to the plant changes through a data-based event-triggering rule. To do so, we monitor the evolution of a data-based Lyapunov function along the solution. When this Lyapunov function does not satisfy a designed desirable condition, an episode is triggered to update the controller gain and the corresponding Lyapunov function using the last collected data. The resulting closed-loop dynamics hence exhibits both physical jumps, due to the system dynamics, and episodic jumps, which naturally leads to a hybrid discrete-time system. We leverage the inherent robustness of the controller and provide general conditions under which various stability notions can be established for the system. Two notable cases where these conditions can be established a-priori based on system’s properties and time-variations are treated, and numerical results illustrating the relevance of the approach and comparing it with alternative approaches are discussed. This is a joint work with Andrea Iannelli (Univ. of Stuttgart, Germany).
• Vincent Andrieu (LAGEPP, CNRS)
  A nonlinear KKL framework for theoretical analysis and guarantees of neural network observers
  This presentation addresses the state estimation problem for nonlinear dynamical systems. Modern approaches that use Recurrent Neural Networks (RNNs) for the observer’s internal dynamics and Multi-Layer Perceptrons (MLPs) for the output mapping are popular but often lack theoretical guarantees. We show how these structures can be analyzed through the lens of Kazantzis-Kravaris-Luenberger (KKL) observers. We extend this theory to the case where the RNN implements nonlinear contracting dynamics, leveraging the theory of input-driven contracting systems. We demonstrate the existence of a transformation enabling state reconstruction, analogous to the KKL map, and prove that the structure comprising a Contracting Nonlinear Filter (implementable by an RNN) and an MLP output map constitutes a tunable observer, thereby providing convergence guarantees. Finally, we discuss the potential benefits of using specific nonlinearities to improve the trade-off between convergence speed and noise robustness.
• Moein Sarbandi (PhD student at LS2N under the supervision of Franck Plestan and Mohamed Hamida)
  Data-Based Sliding Mode Control - Application to Floating Offshore Wind Turbines
  This talk presents a data-driven control approach combining online neural networks with sliding mode theory to handle nonlinear systems with unknown dynamics. The control design employs an integral sliding mode control (ISMC) approach enhanced by two neural network (NN) approx- imators. These NNs are adapted online using Lyapunov-based adaptation laws and require no offline pre-training. This ensures real-time adaptability and robustness against disturbances and modeling uncertainties. As a case study, the framework is applied to a floating offshore wind turbine (FOWT) operating in above-rated wind conditions, where aerodynamic and hydrodynamic complexities pose major control challenges. Simulations using the NREL 5-MW turbine model within the OpenFAST environment demonstrate that the proposed method significantly outperforms both standard ISMC and the widely used reference open-source controller (ROSCO)

• The workshop will take place at ENSEEIHT, 2 rue Charles Camichel, 31000 Toulouse, France, in room C002 ("salle des thèses")
• For people not working at ENSEEIHT, a valid ID will be asked to register at the ENSEEIHT entrance.
• Registration is free but mandatory : registration link
• Only coffee breaks are included for now, more information to come regarding lunch options.

Contact