A Study for Parallelization of Multi-Objective Evolutionary Algorithm Based on Decomposition and Directed Mating

This work proposes a parallel multi-objective evolutionary algorithm based on decomposition for solving constrained multi-objective optimization problems. A representative decomposition-based algorithm, MOEA/D, decomposes multi-objective problems into a number of single-objective sub-problem using weight vectors and a scalarizing function. It keeps only the best solution for each sub-problem and neighbor solutions are used to generate offspring. Therefore, to independently execute solution generation in parallel by using multi-core, at least two solutions have to be included in a core. Hence, maximum parallel number of MOEA/D-based parallel algorithm is the population size over 2. However, in proposed parallel algorithm, it can be the population size since it keeps not only the best feasible solution but also an archive population of useful infeasible solutions for each sub-problem. The experimental results using discrete knapsack problems with 2 objectives and {2, 6, 10} constraints show that the proposed parallel algorithm achieves higher search performance by utilizing infeasible solutions even if the number of parallelization is higher than a parallel decomposition-based algorithm.