Search

Purpose The purpose of this paper is to propose and analyze the use of local surrogate models to improve differential evolution’s (DE) overall performance in computationally expensive problems.Design/methodology/approach DE is a popular metaheuristic to solve optimization problems with several variants available in the literature. Here, the offspring are generated by means of different variants, and only the best one, according to the surrogate model, is evaluated by the simulator. The problem of weight minimization of truss structures is used to assess DE’s performance when different metamodels are used. The surrogate-assisted DE techniques proposed here are also compared to common DE variants. Six different structural optimization problems are studied involving continuous as well as discrete sizing design variables.Findings The use of a local, similarity-based, surrogate model improves the relative performance of DE for most test-problems, specially when using r-nearest neighbors with r = 0.001 and a DE parameter F = 0.7.Research limitations/implications The proposed methods have no limitations and can be applied to solve constrained optimization problems in general, and structural ones in particular.Practical/implications The proposed techniques can be used to solve real-world problems in engineering. Also, the performance of the proposals is examined using structural engineering problems.Originality/value The main contributions of this work are to introduce and to evaluate additional local surrogate models; to evaluate the effect of the value of DE’s parameter F (which scales the differences between components of candidate solutions) upon each surrogate model; and to perform a more complete set of experiments covering continuous as well as discrete design variables. [ABSTRACT FROM AUTHOR]

Differential Evolution is a simple and efficient stochastic population-based heuristics for global optimization over continuous spaces. As with other nature inspired techniques, there is no provision for constraint handling in its original formulation, and a few possibilities have been proposed in the literature. In this paper an adaptive penalty technique (APM), which has been shown to be quite effective within genetic algorithms, is adopted for constraint handling within differential evolution. The technique, which requires no extra parameters, is based on feedback obtained from the current status of the population of candidate solutions, and automatically defines, for each constraint, its corresponding penalty coefficient. Equality as well as inequality constraints can be dealt with. In this paper we additionally introduce a mechanism for dynamically selecting the mutation operator, according to its performance, among several variants commonly used in the literature. In order to assess the applicability and performance of the proposed procedure, several test-problems from the structural and mechanical engineering optimization literature are considered. [ABSTRACT FROM AUTHOR]