Christian Magele, Michael Jaindl, Alice Köstinger, Werner Renhart, Bogdan Cranganu‐Cretu and Jasmin Smajic
The purpose of this paper is to extend a (μ/ρ, λ) evolution strategy to perform remarkably more globally and to detect as many solutions as possible close to the Pareto optimal…
Abstract
Purpose
The purpose of this paper is to extend a (μ/ρ, λ) evolution strategy to perform remarkably more globally and to detect as many solutions as possible close to the Pareto optimal front.
Design/methodology/approach
A C‐link cluster algorithm is used to group the parameter configurations of the current population into more or less independent clusters. Following this procedure, recombination (a classical operator of evolutionary strategies) is modified. Recombination within a cluster is performed with a higher probability than recombination of individuals coming from detached clusters.
Findings
It is shown that this new method ends up virtually always in the global solution of a multi‐modal test function. When applied to a real‐world application, several solutions very close to the front of Pareto optimal solutions are detected.
Research limitations/implications
Stochastic optimization strategies need a very large number of function calls to exhibit their ability to reach very good local if not the global solution. Therefore, the application of such methods is still limited to problems where the forward solutions can be obtained with a reasonable computational effort.
Originality/value
The main improvement is the usage of approximate number of isolated clusters to dynamically update the size of the population in order to save computation time, to find the global solution with a higher probability and to use more than one objective function to cover a larger part of the Pareto optimal front.