Um estudo sobre a configuração automática do algoritmo de evolução diferencial

Abstract

The great development in the area of evolutionary algorithms in recent decades has increased the range of applications of these tools and improved its performance in different fronts. In particular, the Differential Evolution (DE) algorithm has proven to be a simple and efficient optimizer in several contexts. Despite its success, its performance is closely related to the choice of variation operators and the parameters which control these operators. To increase the robustness of the method and the ease of use for the average user, the pursuit for methods of self-configuration has been increasing as well. There are several methods in the literature for setting parameters and operators. In order to understand the effects of these approaches on the performance of DE, this paper presents a thorough experimental analysis of the main existing approaches. Furthermore, a new approach is presented for selecting operators based on reinforcement learning. The results show that simple approaches are able to bring significant improvements to the performance of DE.