Surrogate-guided multi-objective optimization (SGMOO) using an efficient online sampling strategy

In this paper, we present a new multi-objective global optimization algorithm SGMOO for computationally expensive black-box problems, where Radial basis functions are used to build dynamically updated surrogate models for each objective. Moreover, an efficient online sampling strategy that includes three infilling criteria “Multi-objective-based exploitation on RBF, Single-objective-based exploitation on RBF, and Evolutionary-computation-based exploration” is presented to capture promising samples in each cycle. In the first criterion, a distance-based data mining strategy is proposed to pick out the valuable samples from the predicted Pareto solution set, speeding up the convergence to the true Pareto frontier. In the second criterion, single-objective surrogate-based sampling approach is used to enhance the local infilling performance at the bounds of Pareto frontier. Furthermore, the dynamically updated expensive sample set is regarded as a population to generate offspring by non-dominated sorting, and a novel prescreening operator considering hypervolume and space infilling performance is presented to select elite individuals in the third infilling criterion. With the help of the cooperation of the three infilling criteria, SGMOO builds a reasonable balance between global exploration and local exploitation. Compared with 4 well-known multi-objective algorithms, SGMOO has more stable and impressive performance on 25 benchmark cases and the shape optimization design of a blended-wing-body underwater glider.