Design And Optimization Of Stochastic Dynamic Multi-Objective Benchmark Problems

In the past few decades, the evolutionary algorithms(EA) have widely used in solving the multi-objective optimization problems(MOP). Generally, the MOPs have many conflicting objectives, and the optimization methods must balance those objectives and achieve a set of trade-off solutions for the decision makers. The multi-objective evolutionary algorithm(MOEA) can effectively deal with this kind of problems. Besides, MOEAs have proved to generate a set of solutions with well-convergence and well-distribution in a single run.With the further exploration of MOEA, a special kind of MOP is designed which parameters and objectives are changing over time. The relative researchers have proposed a variety of dynamic MOEAs, and successfully applied into the real-world problems. However, recently, the developing of this domain has been limited because of lacking of dynamic multi-objective test problems. A well-designed dynamic problems are demanded to be not only comprehensively checking the performances of tested algorithms, but also sufficiently reflecting the features of practical situations.In the roughly reviewing of existing test suit, the paper concluded the weakness of them, and proposed some design principles. According to those guidelines, a stochastic benchmark suit is constructed. Many important problem features are introduced into those problems that contain deceptive, multi-modal and bias. Beside,many complicated geometries, such as mixed and disconnected, are also combined with some problems. The experimental analysis of those characteristics confirmed that they are significantly improved the problem's ability of performance testing.As a new reaction method, the center matching strategy(CMS) is proposed to optimize those stochastic problems. The CMS applied the global information reflected by history individuals to predict the population. Once the environmental change is detected, the selected centers are added to the current population. The comparative experiments demonstrated that the CMS has a sufficient capability of handling stochastic variation. Meanwhile, the CMS shown good advantages of well-converging and well-distribution, and other important performances.