An imitation-based framework for large-scale sparse multi-objective optimization

Abstract

Many real-world problems can be modeled as large-scale sparse multi-objective optimization problems (large-scale sparse MOPs). Recent studies on such problems have developed algorithms that significantly outperform conventional multi-objective evolutionary algorithms (MOEAs). However, these state-of-the-art algorithms typically treat decision variables as independent entities and optimize them in isolation. Moreover, they do not fully exploit the information embedded in non-dominated solutions, which hinders further performance gains. This paper introduces an imitation-based framework, dubbed SIF. The framework leverages value combinations of binary variables and distribution characteristics of real variables from non-dominated solutions to generate additional promising offspring solutions. Existing large-scale sparse MOEAs can be integrated with SIF to achieve enhanced performance. Experimental results on the SMOP benchmark problems and three real-world applications validate the effectiveness of the proposed SIF framework.