Research On Evolutionary Multiobjective Optimization Algorithms

With the rapid development of human society, the problems the people have to deal with in routine work and daily life are more and more complex. Generally, when constructing the mathematical model of the problem, it may not be sufficient to describe the essential character of the problem using only single objective function. Thus, there are important academic interests and valuable applications in investigating the effective methods to solving the multiobjective optimization problems (MOPs). In this dissertation, the main issue involving MOPs are discussed, and the multiobjective optimization algorithms have been proposed based on evolutionary algorithms for complex problems. The main contents are as follows:First, a brief introduction to basic definitions of MOPs is given with some classic multiobjective optimization algorithms. Then, the current researches on multiobjective evolutionary algorithms (MOEAs) are reviewed. Moreover, the objectives and several key issues in designing MOEAs are proposed, and various popular test functions and quality measures in evaluating MOEAs are also presented.Second, a new steady-state MOEA based on relaxed Pareto dominance relation (ε-Pareto dominance relation), denoted as EDMOEA, is presented. A new archive updating strategy combining Pareto dominance andε-Pareto dominance is used in the EDMOEA. Moreover, the strong elitism strategy with steady-state evolution scheme is also proposed. Experimental results illustrate that our algorithm is able to produce significantly better results in ZDT test problems with less time cost, and achieve better convergence and diversity in DTLZ test problems.Third, the replacement and mating selection strategies in a steady-state cooperative coevolutionary MOEA are discussed. Based on a framework of a steady-state MOEA evolved with two populations, some new replacement strategies and mating selection methods are investigated on the archive population and main population respectively. Through combining these new replacement strategies and mating selection methods together, six variants of the MOEA are explored on two kinds of scalable test problems ( Ptr ue-connected and Ptr ue-disconnected), which illustrates that the high elitism is more efficient in the steady-state MOEAs with two concurrent evolved populations. Moreover, the boundary points in mating selection could be used to effectively expand the finally obtained Pareto front. Fourth, this thesis attempts to investigate that the essential reason that the Pareto-based baseline algorithms, such as NSGA-II and SPEA2, deteriorate significantly with increasing number of objectives. Through introducing some special individuals those have positive offsets from the extreme vertexes of the true Pareto front, it is illustrated that dominance resistant solutions and selection strategies based on density information are essential handicaps for the sustained population evolution. At last, four representative approaches those are originally proposed to address many-objective problems are compared based on a new evolution framework.Finally, the main research contents are summarized at the end of the thesis with an expectation for further study and research.