A clustering-based surrogate-assisted evolutionary algorithm (CSMOEA) for expensive multi-objective optimization

This paper presents a novel surrogate-assisted evolutionary algorithm, CSMOEA, for multi-objective optimization problems (MOPs) with computationally expensive objectives. Considering most surrogate-assisted evolutionary algorithms (SAEAs) do not make full use of population information and only use population information in either the objective space or the design space independently, to address this limitation, we propose a new strategy for comprehensive utilization of population information of objective and design space. The proposed CSMOEA adopts an adaptive clustering strategy to divide the current population into good and bad groups, and the clustering centers in the design space are obtained, respectively. Then, a bi-level sampling strategy is proposed to select the best samples in both the design and objective space, using distance to the clustering centers and approximated objective values of radial basis functions. The effectiveness of CSMOEA is compared with five state-of-the-art algorithms on 21 widely used benchmark problems, and the results show high efficiency and a good balance between convergence and diversity. Additionally, CSMOEA is applied to the shape optimization of blend-wing-body underwater gliders with 14 decision variables and two objectives, demonstrating its effectiveness in solving real-world engineering problems.