Machine learning methods for control, identification, and estimation

Over the last decades, the landscape of control theory and system identification has changed significantly in response to the new challenges arising from the industry. This is not surprising: classical model-based techniques are not suitable to handle real-world applications for which it is often too expensive to derive even an approximate model using first principles. Data-driven approaches represent a solution to such an issue. Thanks to the ever-increasing availability of a large quantity of data, they have quickly become central topics within the control theory community. This thesis collects some results regarding using machine learning approaches to answer some open questions in control theory by formulating novel techniques and lessening some undesirable aspects of existing methods. We first present a system identification approach based on deep learning to learn state-space models for nonlinear systems. We then propose a data-driven virtual sensor synthesis approach, inspired by the Multiple Model Adaptive Estimation framework, for reconstructing normally unmeasurable quantities such as scheduling parameters in parameter-varying systems. Three data-driven control approaches, two of which are based on the well-known Virtual Reference Feedback Tuning framework, are finally presented to synthesize constrained controllers for unknown nonlinear dynamical systems from the data without identifying first a model of the plant. Tuning guidelines for the proposed methods are also provided.