Abstract: The present work is a theoretical and experimental study in the evolutionary computation domain.The first part is an introduction to the artificial evolution with a synthesis of the principal approaches. The second part is theoretical study devoted to handling constraints in evolutionary computation. It presents an extensive review of previous constraint handling methods in the literature and their limitations. Two solutions are then proposed.The first aims to improve genetic operator exploration capacity for constrained optimization problems. It propose the logarithmic mutation operator conceived to explore both locally and globally the search space. The second solution introduce the original Adaptive Segregational ConstraintHandling Evolutionary Algorithm, which the main idea is to maintain population diversity. In order to achieve this goal, three main ingredients are used: An original adaptive penalty method that uses global information of the population to adjust the penalty coefficients; a constraint-driven recombination, where in some cases feasible individuals can only mate with infeasible individuals; a segregational selection that distinguishes between feasible and infeasible individuals to enhance the chances of survival of the feasible ones. Moreover, a niching method with an adaptive radius is added to ASCHEA in order to handle multimodal functions. Finally, to complete the ASCHEA system, a new equality constraint handling strategy is introduced, that reduce progressively the feasible domain of the explored solution in order to get is as close as possible to the reel domain (with null measure) at the end of the evolution.The third part is a case study tackling a real-word problem. The goal is to design the 2-dimensional profile of an optical lens (phase plate) in order to control focal-plane irradiance of some laser beam. The aim is to design the phase plate such that a small circular target on the focal plane is uniformally illuminated without energy loss. ; Cette thèse est une étude théorique et ...
No Comments.