Machine learning methods for control, identification, and estimation

Over the last decades, the landscape of control theory andsystem identification has changed significantly in responseto the new challenges arising from the industry. This is notsurprising: classical model-based techniques are not suitableto handle real-world applications for which it is often tooexpensive to derive even an approximate model using firstprinciples. Data-driven approaches represent a solution tosuch an issue. Thanks to the ever-increasing availability of alarge quantity of data, they have quickly become central topicswithin the control theory community.This thesis collects some results regarding using machinelearning approaches to answer some open questions in controltheory by formulating novel techniques and lesseningsome undesirable aspects of existing methods. We firstpresent a system identification approach based on deeplearning to learn state-space models for nonlinear systems.We then propose a data-driven virtual sensor synthesis approach,inspired by the Multiple Model Adaptive Estimationframework, for reconstructing normally unmeasurable quantitiessuch as scheduling parameters in parameter-varyingsystems. Three data-driven control approaches, two of whichare based on the well-known Virtual Reference FeedbackTuning framework, are finally presented to synthesize constrainedcontrollers for unknown nonlinear dynamical systemsfrom the data without identifying first a model of theplant. Tuning guidelines for the proposed methods are alsoprovided.