An Exploratory Study On Dominance Resistant Solutions In Many Objectives Optimization

Liu Liu,Minqiang Li,Lin Dan

2009-01-01

Abstract:In spite of many approaches have been proposed to improve the performance of evolutionary multiobjective algorithms on many objectives optimization, little work was to explore the essential reason that the algorithms, such as the NSGA-II and SPEA2, deteriorates distinctly with increased number of objectives. One of the popular explanations is that the high proportion of nondominated solutions in the population breaks down the search of the algorithms. However, this paper attempts to explain that some dominance resistant solutions (DRSs) (except the extreme individuals which are located on coordinate axis greater than the extreme points of the true Pareto front), which are hard to be dominated and far away from the true Pareto front, constitute the essential handicap for the population evolution. It is observed this kind of solution is generated at the beginning of the evolution, and further delays the convergence of the algorithms due to diversity promoting methods. We make an analytical explanation of the four representative approaches that were originally proposed to address many objectives problems. At last, under a new framework of MOEA evolved with only nondominated solutions, our experimental results verify their performance on DTLZ1 and DTLZ2 problems with 4, 5, 6 objectives respectively.