Differential evolution induced many objective optimization

Abstract

We propose a novel approach to solve the many objective optimization (MaOO) problem using a ranking policy, instead of the Pareto ranking, supposing that a solution is unlikely to perform well for all objectives in a MaOO problem. A solution is thus evolved with respect to a specific objective only, which it may proficiently optimize. First, all objectives of the MaOO problem are individually optimized by evolutionary algorithms in parallel. The second step is concerned with judiciously selecting and filtering the quality solutions obtained by individual optimization of all objectives in parallel. A unique ranking policy is proposed to grade the members of the union set of quality solutions based on their extent of optimization of individual objectives. The evolutionary algorithm used for parallel optimization of all objectives in a MaOO here has been realized with differential evolution (DE). The mutation strategy of DE is also amended here with an aim to allow controlled communication between population members, concerned with parallel optimization of different objectives of a MaOO problem. Experiments undertaken with DTLZ and WFG test suits reveal that the proposed algorithm outperforms the state-of-art techniques with respect to inverted generational distance and hypervolume metrics.