A switched parameter differential evolution with optional blending crossover for scalable numerical optimization

다국어 초록 (Multilingual Abstract)

<P>Differential Evolution (DE) is currently one of the most competitive Evolutionary Algorithms (EAs) for optimization problems involving continuous parameters. This article presents three very simple modifications to the basic DE scheme such that its performance can be improved and made scalable for optimizing functions having a real-valued moderate-to-high number of variables (dimensions) while focusing on preservation of the simplicity offered by its algorithmic framework. Contrary to the common policies of coupling a complicated scheme for adaptation of the control parameters or introducing additional local search algorithms, we present here a (population member and generation specific) control parameter choosing strategy which uniformly and randomly switches the values of the mutational scale factor and crossover rate between two extremities of their feasible ranges. Furthermore, each population member is mutated either by using the DE/rand/1 scheme or a proposed version of the DE/current-to-best scheme. The mutation scheme for a population member is chosen based on its performance in the current generation. Hence if a mutation strategy successfully replaces a target vector by the corresponding trial vector, then it is reused by the population member of the same index in the following generation, else a switch of mutation method is executed. In the crossover phase, each member undergoes either the common binomial crossover or the BLX-alpha-beta crossover modified for applying in DE, with equal probabilities. Our experiments using the benchmark optimization functions proposed for the IEEE Congress on Evolutionary Computation (CEC) 2013 competition on real parameter optimization and CEC 2010 competition on large-scale global optimization demonstrate that the basic DE optimizer when coupled with these elementary alterations and/or schemes can indeed provide a very competitive result against some of the most prominent state-of-the-art algorithms. (C) 2017 Elsevier B.V. All rights reserved.</P>