A New Differential Evolution Algorithm with Dynamic Population Partition and Local Restart

Abstract

This paper will introduce a new differential evolution (DE) algorithm called DE/cluster. DE/cluster applies a simple hierarchical clustering model to mine the distribution information of the DE population every K generations to make a dynamic partition of the population. One special cluster formed by the single-individual clusters will use a slower convergence mutation strategy to do the global search. The other clusters will use more greedy searching strategy to do the local search. As long as the subpopulations may be trapped by local minima, the "dead" state is defined for a cluster and clusters will be checked in every generation and the "dead" clusters will be restarted in the current searching range. This local restart strategy can make the performance of DE/cluster even be better than DE/rand on some multimodal test functions that are not linearly separable. The DE/cluster algorithm is tested on a test suite with 24 functions and it shows promising performance compared with the current best DE variants.