Clustering-based multipopulation approaches in MOEA/D for many-objective problems

This work presents a new multipopulation framework for the multiobjective evolutionary algorithm based on decomposition (MOEA/D). In this case, clustering methods are used to reinforce mating restrictions by splitting the MOEA/D evolutionary population into multiple subpopulations of similar individuals for independent evolution. Using subpopulations leads to a natural parallel implementation by assigning each subpopulation to a different processor. The proposed multipopulation MOEA/D (mpMOEA/D) is evaluated using three clustering methods: k-Means, spectral-based clustering, and a method based on the shape of objective vectors. Additionally, a random partitioning approach is tested. Metrics measuring convergence, diversity and computation time are used to compare the results of the mpMOEA/D alternatives and the original MOEA/D using DTLZ and WFG problems with 3, 4, 8 and 10 objectives. Evaluation using the Wilcoxon test and the Friedman rank reveals the importance of using clustering procedures for population division, especially in cases with many objectives. The results show the viability of the clustering-based multipopulation approach in enhancing the performance of evolutionary methods for many-objective problems.