A Dynamic Breathing Surrogate-Assisted Particle Swarm Optimization Algorithm for Expensive Multi-Objective Optimization Problems

When solving expensive multi-objective optimization problems, traditional evolutionary algorithms rely on a large number of fitness evaluations, which leads to insufficient exploration of the solution space when the evaluation budget is limited. Meanwhile, purely machine learning methods typically require abundant high-quality data and are prone to modeling bias or getting stuck in local optima when training data is scarce or the objective function is a complex black-box. Therefore, in recent years, the combination of machine learning and evolutionary computation has become the primary approach to addressing expensive optimization problems. However, the integration of surrogate models into evolutionary algorithms often lacks depth, focusing primarily on superficial enhancements rather than comprehensive synergy based on a framework. To address these issues, this paper proposes a dynamic breathing surrogate-assisted particle swarm optimization algorithm (DB-SAPSO). Within the proposed unified framework, the core innovation of DB-SAPSO lies in its dynamic breathing mechanism, which iteratively expands and efficiently contracts the population and conducts expensive real evaluations on only a small number of high-potential solutions. This achieves better performance compared to existing algorithms, validating the potential of integrating machine learning and evolutionary computation within a unified framework.