Decomposition-based evolutionary sampling for expensive large-scale multi-objective optimization

Huixiang Zhen, Wenyin Gong, Zhenshou Song, Ling Wang, Xiangyun Hu

Abstract: Surrogate-assisted evolutionary algorithms have demonstrated significant promise for solving expensive largescale multi-objective problems by reducing the number of costly evaluations. Existing decomposition-based frameworks offer advantages in terms of their low modeling costs and errors, yet current approaches still suffer from inefficient subproblem search. To address this issue, we propose a novel decompositionbased evolutionary sampling algorithm (DES), which effectively integrates the decomposition framework with two complementary surrogate-assisted evolutionary sampling strategies, thereby balancing exploration and exploitation in high-dimensional search spaces. DES decomposes the objective space into uniformly distributed reference vectors and generates candidate solutions through combining two kinds of surrogate-guided evolutionary sampling strategies, leading to enhanced convergence behavior and diversity maintenance. Component analyses and parameter study experiments confirm the effectiveness of each module. Comprehensive comparison experiments on four benchmark suites and the problem of estimating the ratio error of voltage transformers demonstrate that DES achieves markedly better convergence accuracy and diversity preservation than six state-of-the-art algorithms, while retaining high efficiency.

External IDs:doi:10.36227/techrxiv.176105870.07117173/v1