Real-world multi-objective optimization problems are usually with noise. Existing intelligent optimization techniques seem to be difficult when direcfly applied to these optimization problems belonging to multi-objective stochastic optimization (MSO).This is because the optimized quality is greatiy influenced by noisy envirormients and imprecise models so that it is ahnost impossible to obtain true solutions. Even if two popular multi-objective evolutionary algorithms, SPEA2 and NSGA2 respectively proposed by Zitzler, Deb, and their colleagues, have been reported in the literature, they are originally designed for static multi-objective optimization. Thus, if directiy applied to MSO problems without any modification, they do not perform well. So, MSO should be specially investigated, in which vital problems are individual ranking and noisy suppression. For the latter one, some sampling methods in single-objective stochastic optimization (SSO) have been displayed in the literature. They can be categorized into two broader types: static sampling with the same fixed or predefined sampling number for each individual, and adaptive sampling including hypothesis test-based threshold selection, sample-allocation and changing duration time for each generation, and so on. The first scheme is both simple and convenient, but its performance is worse; the second one is a challenging research topic but difficult. Thanks to complexity of MSO, these methods cannot be directly adopted; therefore, new sampling techniques are desired.
展开▼
