Fitness landscape analysis of differential evolution algorithms

Abstract

Fitness landscape analysis in evolutionary algorithms is commonly done on problems represented as bit strings with Hamming distance based random walks on the landscape. In this study, we aim to do a preliminary fitness landscape analysis of the differential evolution algorithm, which works on continuous search spaces. To the authors' best knowledge, no such fitness landscape analysis has been conducted in literature on continuous problems where search is performed through differential evolution. To achieve this aim, we first propose a suitable neighborhood definition through which a vector-based random walk on the landscape is possible. Then we use this neighborhood definition to conduct a fitness distance correlation and a correlation length analysis on a series of benchmark functions.