Dissertation/ Thesis
Achieving reliable control : robustness and stability in nonlinear systems via DNN-based feedback design ; Obtenir un contrôle fiable : robustesse et stabilité des systèmes non linéaires grâce aux lois des commandes basée sur les DNNs
| العنوان: | Achieving reliable control : robustness and stability in nonlinear systems via DNN-based feedback design ; Obtenir un contrôle fiable : robustesse et stabilité des systèmes non linéaires grâce aux lois des commandes basée sur les DNNs |
|---|---|
| المؤلفون: | Zoboli, Samuele |
| المساهمون: | Laboratoire d'automatique, de génie des procédés et de génie pharmaceutique (LAGEPP), Université Claude Bernard Lyon 1 (UCBL), Université de Lyon-Université de Lyon-École Supérieure de Chimie Physique Électronique de Lyon (CPE)-Centre National de la Recherche Scientifique (CNRS), Université Claude Bernard - Lyon I, Vincent Andrieu, Jilles Steeve Dibangoye, Daniele Astolfi |
| المصدر: | https://theses.hal.science/tel-04631894 ; Automatic Control Engineering. Université Claude Bernard - Lyon I, 2023. English. ⟨NNT : 2023LYO10172⟩. |
| بيانات النشر: | HAL CCSD |
| سنة النشر: | 2023 |
| المجموعة: | HAL Lyon 1 (University Claude Bernard Lyon 1) |
| مصطلحات موضوعية: | Nonlinear systems, Robustness, Stability, Deep neural networks, Contraction, Multi-agent synchronization, Reinforcement learning, Systèmes non linéaires, Robustesse, Stabilité, Réseaux de neurones profonds, Synchronisation multi-agent, Apprentissage par renforcement, [INFO.INFO-AU]Computer Science [cs]/Automatic Control Engineering |
| الوصف: | This thesis focuses on the integration of robustness and stability guarantees in feedback controllers modeled by deep neural networks (DNNs), also known as neural controllers. The main objective is to combine machine learning tools and control theoretic approaches to derive neural controllers with strong theoretical guarantees. In the first part of the manuscript, we investigate how control theory results can be used to enhance modern learning-to-control approaches with stability and robustness guarantees. One key issue is the challenge of understanding whether these control laws, trained in simulated environments, can provide stability guarantees and maintain them in real-world scenarios. Therefore, the first part of this manuscript focuses on discrete-time nonlinear. The first chapter studies robustness to model uncertainties. We establish conditions for the transfer of stability properties based solely on norms of model mismatches. This analysis determines whether the existence and stability of equilibrium points for a nominal system imply the existence and stability of equilibrium points for sufficiently similar systems. In turn, we justify the use of accurate simulators for training neural controllers that are intended for real-world scenarios. Moreover, these results motivate the use of integrators to enhance the robustness of discrete-time controllers (including learned ones), for which formal justification is currently absent in the literature. In the second chapter, we investigate how guaranteed stability properties can be embedded in DNN-based feedback designs. Our objective is to propose an algorithm-agnostic methodology that incorporates local stability guarantees in such controllers, independently from their training process or time. To achieve this, we combine local guaranteed control laws with DNNs. Our approach enables the development of robustly stabilizing neural controllers. In the third and last chapter of the first part of this thesis, we shift our focus from the study of robust stability of ... |
| نوع الوثيقة: | doctoral or postdoctoral thesis |
| اللغة: | English |
| Relation: | NNT: 2023LYO10172; tel-04631894; https://theses.hal.science/tel-04631894; https://theses.hal.science/tel-04631894/document; https://theses.hal.science/tel-04631894/file/TH2023ZOBOLISAMUELE.pdf |
| الاتاحة: | https://theses.hal.science/tel-04631894 https://theses.hal.science/tel-04631894/document https://theses.hal.science/tel-04631894/file/TH2023ZOBOLISAMUELE.pdf |
| Rights: | info:eu-repo/semantics/OpenAccess |
| رقم الانضمام: | edsbas.D472626C |
| قاعدة البيانات: | BASE |
| الوصف غير متاح. |