Improving Many-Objective Optimization Performance by Sequencing Evolutionary Algorithms

Evolutionary multiobjective optimization (EMO) has been successfully applied to various real-world scenarios with usually two or three contradicting optimization goals. However, several studies have pointed out a great deterioration of computational performance when handling more than three objectives. In order to improve the scalability of multiobjective evolutionary algorithms (MOEAs) onto higher-dimensional objective spaces, techniques using e.g. scalarizing functions and preference-or indicator-based guidance have been proposed. Most of those proposals require a-priori information or a decision maker during optimization, which increases the complexity of the algorithms. In this paper, we propose a divide and conquer method for many-objective optimization. First, we partition a problem into lower-dimensional subproblems for which standard algorithms are known to perform very well. Our key improvement is the sequential usage of MOEAs, utilizing the results of one suboptimization as initial population for another MOEA. This technique allows modular optimization phases and can be applied to common evolutionary algorithms. We test our enhanced method on the hard to solve multiobjective Quadratic Assignment Problem (mQAP), using a variety of established MOEAs.