Parameters calibration for parallel differential evolution based on islands

Abstract

We are studying di erent alternatives to obtain a version of the Di erential Evolution (DE) algorithm that improves the solutions quality properties. One of the parallel alternatives, named Island Model, follows a Master/Worker scheme. With this model, multiple instances of DE are executed in parallel on various computing nodes or islands, each of them considering a di erent population of individuals. Each worker makes the search process, and communicates with each other to exchange information with certain frequency. This model signi cantly promote the exploration of a larger search space, which leads to good solutions quality. The aim of this paper is to analyse the behaviour of this model, when setting each island with di erent input parameters. We apply some input con guration tests for the islands, in order to analyse the impact in the solutions quality and the execution time, taking into account the crossover probability and mutation factor, and the crossing type. These parameters are crucial to guide the search towards certain areas of the search space.