The Optimal Zone Breadth for Parallel and Distributed MOEA/D with Virtual Overlap Zone and Exclusively Evaluated Mating
Xinyuan Zhu, Yuji Sato, Mikiko Sato · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2024
In this paper, we propose an index of the optimal overlapping zone breadth for parallel and distributed MOEA/D with a virtual overlapping zone and exclusively evaluated mating. The proposed index aims to address the sparse areas that occur when the T-neighborhood is divided for parallelization computation. Increasing the overlapping zone breadth reduces the sparse areas on the Pareto solutions near the zone boundary and enhances the Hypervolume (HV) value but increases the computation time. To determine the optimal index of optimal overlapping zone breadth, we investigated the HV value and processing time when the number of censored generations was fixed and the T neighborhood size and overlapping zone breadth were changed. We also conducted an experiment to examine the HV value per unit time by varying the degree of parallelism and the breadth of overlap as parameters. Experiments on a constrained knapsack problem show that setting the overlapping zone breadth close to half of the T-neighborhood size not only reduces sparse areas of the Pareto-optimal front but also enhances the HV value.