Contrôle Générique de Paramètres pour les Algorithmes Evolutionnaires
Jorge Maturana · HAL (Le Centre pour la Communication Scientifique Directe) · 2009
Parameters of Evolutionary Algorithms (EAs) greatly influence their ability to produce good results. These parameters define structural and behavioral aspects of the EA, ranging from choosing which features will be included (e.g. set of operators, encoding, selection scheme) to how these features will be used (e.g. operators' application rates). Parametrization of EAs have been longly a specialist domain, and even though many previous studies have addressed the parametrization problem, there is a lack of generic approaches that could be applied to a wide range of EAs in a simple way. This thesis deals with the problem of creating a generic controller, that could be included in any EA with a minimum effort. Generality is accomplished by incorporating a learning/ adaptive component, that monitors the state of the search and modify parameter values. The controller focuses in common goals of all EAs, that is to say, to maximize both the diversity of the population and the quality of the individuals, in order to maintain a convenient balance between exploration and exploitation. Several configurations of learning tools and adjustment methods were tried and analyzed, using different EAs solving well known combinatorial problems. Positive results suggest that our goal of building a generic, easy-to-use controller is a feasible approach.