Tuning parameters of evolutionary algorithms using ROC analysis

Abstract

Evolutionary algorithms mimic the natural evolution of the species in biological systems and they have been applied to optimization problems with significant success. However, these algorithms have stochastic nature since the search involves probabilistic rules and the setting of several parameters is difficult. Thus, it is crucial to investigate how the parameters influence the performance of the algorithms for distinct classes of optimization problems. Receiver Operating Characteristic (ROC) analysis, over the years, has become a powerful tool to measure diagnostic performance and, in this work, was used to assist the algorithm parameters setting. The study is conducted over a significant number of difficult optimization problems and a single parameter of an evolutionary algorithm. The ROC analysis seems to be helpful to identify parameters values that allow a satisfactory performance of the algorithm for different classes of problems. © 2009 Springer-Verlag Berlin Heidelberg.