Late parallelization and feedback approaches for distributed computation of evolutionary multi-objective optimization algorithms

Abstract

Distributing of the multi-objective evolutionary algorithms into various computational devices in a parallel fashion is a method for speeding up the execution time of the algorithms. When the processors are increased in number, the gain from parallelization decreases. For this reason, the aim of the parallelization method is not only to decrease the overall algorithm execution time, but also to obtain a higher gain from the use of parallel processors. Therefore, in this study two new parallelization approaches are proposed and discussed to increase the benefit from the computational devices. The proposed schemes are named as late parallelization and feedback approaches. These proposed approaches are based on four-layered structure, and each layer (i) has the $$2^i$$ number of computational units where population at each unit is divided into two equal subpopulations at ordered layers. Consequently, the sum of the members at every layer equals to each other. Of these layer-based cognate schemes, in late parallelization, the layers are executed sequentially beginning with the first layer. Until the last layer is executed, a delay time is defined such that layers are executed more than one except the last layer. However, at feedback approach, the order of the layers is different and the last layer is connected with the front layer. So, the population is not only divided into subpopulations but also they are gathered with the in-osculation of the subpopulations. Finally, the performances of these approaches are evaluated on multi-objective test problems with respect to the statistical properties of the number of function evaluations.