Titre de la thèse : System performance optimization based on identified models: application to projectile design and feed-forward control of Wiener Systems Rapporteurs : - Stéphane Victor, Université de Bordeaux - Jonathan Weber, Université de haute-Alsace Examinateur : - Guillaume Mercère, Université de Poitiers Directeurs et co-encadrante de thèse : - Xavier Bombois, EC Lyon - Marie Albisser, Institut Saint-Louis - Marion Gilson, Université de Lorraine Résumé : In many engineering applications, the objective is to find the optimum performance of a system. To evaluate the performance, it is necessary to measure the output of the system for given inputs. However, in several real-life scenarios, systems are often expensive to evaluate making it difficult to perform the optimization tasks. To address this issue, data-driven models are often identified to estimate the expensive objective functions, associated with the systems, and are employed to approximate their optimum. If poor performances are obtained using these models, they must be improved by re-identifying them with new data. However, since the systems are expensive to evaluate, the data must be chosen carefully. The aim of this thesis is to develop approaches which can be used to improve identified models employed in system performance optimization. These approaches are applied in two different applications. The first one is the aerodynamic design where the goal is to find the optimum dimensions of a projectile based on criteria associated with aerodynamic coefficients. These coefficients are costly to acquire, hence the projectile geometry configurations to evaluate, to find the optimum, must be selected with care. This is usually achieved using approaches such as Bayesian Optimization where a Gaussian Process model is employed to model the static relationship between the projectile configuration and the objective function. In this thesis, a procedure similar to Bayesian Optimization but where Neural Networks are employed as data-driven models instead of Gaussian Processes is developed, to enable scalability for larger datasets. Both approaches are used to solve the aerodynamic design problem, and it is shown that they allow to reduce the costs associated to aerodynamic optimization. The second application concerns control engineering and more particularly the framework of identification for control. The focus is on feed-forward controller design for non-linear systems which can be represented by Wiener structures. More particularly, it is shown how a model of such systems can be used to design the controller. A procedure to iteratively improve the model and re-design the controller is also introduced in the case where the initially designed one does not allow to obtain optimal performances. Overall, the developed approaches provide effective solutions to minimizing system evaluation costs during optimization tasks in diverse engineering fields