The effects of different kinds of move in differential evolution searches

In the commonly used DE/rand/1 variant of differential evolution the primary mechanism of generating new solutions is the perturbation of a randomly selected point by a difference vector. The newly selected point may, if good enough, then replace a solution from the current generation. As the replaced solution is not the one perturbed to create the new, candidate solution, when the population has divided into isolated clusters large moves by solutions are the result of small difference vectors applied within different clusters. Previous work on twoand 10-dimensional problems suggests that these are the main vehicle for movement between clusters and that the quality improvements they yield can be significant. This study examines the existence of such nonintuitive moves in problems with a greater number of dimensions and their contribution to the search—changes in solution quality and impact on population diversity—over the course of the algorithm’s run. Results suggest that, while they frequently contribute solutions of higher quality than genuine large moves, they contribute to population convergence and, therefore, may be harmful.