Parallelization of Multi-objective Evolutionary Algorithms Using Clustering Algorithms

Abstract

Abstract. While Single-Objective Evolutionary Algorithms (EAs) parallelization schemes are both well established and easy to implement, this is not the case for Multi-Objective Evolutionary Algorithms (MOEAs). Nevertheless, the need for parallelizing MOEAs arises in many real-world applications, where fitness evaluations and the optimization process can be very time consuming. In this paper, we test the 'divide and conquer' approach to parallelize MOEAs, aimed at improving the speed of convergence beyond a parallel island MOEA with migration. We also suggest a clustering based parallelization scheme for MOEAs and compare it to several alternative MOEA parallelization schemes on multiple standard multi-objective test functions.

A Hierarchical Approach in Distributed Evolutionary Algorithms for Multiobjective Optimization

“…Three variants were analyzed: the data were uniformly split in 5, 10 and 20 subsets leading to 5,10 and 20 jobs, respectively. The rankings obtained are: (1,3,5,6,2,4,7,8,9,10) in the case of 5 subsets, (1,6,3,5,2,4,8,7,9,10) in the case of 10 subsets and (1,3,6,5,4,2,7,8,9,10) in the case of 20 subsets. All results are in concordance with the generated data.…”

“…The division of the search space can be made before starting the evolution (apriori division rule) or dynamically during the evolution (dynamic division rule). In the last years different strategies have been proposed, most of them being based on dynamic division rules [4,7]. Dynamic division rules usually involve some operations aiming to periodically reorganize the structure of the search space.…”

“…Dynamic division rules usually involve some operations aiming to periodically reorganize the structure of the search space. In some cases this operations could be costly by themselves, as for instance in [7], where a clustering step involving the elements of all subpopulations is executed, or in [4], where all subpopulations are gathered and their elements are sorted. On the other hand, applying apriori division rules (e.g.…”

“…However, for complex problems characterized by a large number of decision variables and/or objective functions they need large populations and a lot of iterations in order to obtain a good approximation of the Pareto optimal set. In order to solve this problem, different variants for parallelizing and distributing multiobjective evolutionary algorithms (MOEAs) have been proposed in the last years [3,4,7]. Choosing the appropriate variant for a particular problem is a difficult task, thus simultaneously applying different variants and combining their results could be beneficial.…”

1821 Discussion of Clustering and Network Inversion for Multi-Objective Genetic Algorithms

“…In such parallel models, MOGA operations are executed in many computational resources sharing the processes among them. Such parallel studies have been reported by many researchers, most of which have made use of the master-slave or island model [3,4,5,6,7,8]. Especially, it is reported that Deb's parallel model can obtain good Pareto optimal solutions while reducing the calculation time.…”