A multi-source surrogate-assisted evolutionary algorithm for expensive optimization: An optinformatics perspective

Abstract

Surrogate-Assisted Evolutionary Algorithms (SAEAs) have emerged as a powerful paradigm for solving Expensive Optimization Problems (EOPs). However, existing SAEAs often struggle to fully exploit the evolutionary knowledge embedded in the limited historical data during the optimization process, which impedes their ability to locate high-quality solutions for complex EOPs. To address this challenge, this work develops a knowledge-driven framework for expensive optimization from an optinformatics perspective, aiming to extract implicit yet useful knowledge and adaptively utilize it to drive the evolutionary process. Accordingly, we propose a Multi-source Surrogate-Assisted Evolutionary Algorithm (MSAEA) that systematically synthesizes three types of evolutionary knowledge, namely performance, spatial, and temporal knowledge, extracted respectively from the objective space, decision space, and evolutionary sequence. MSAEA adaptively exploits this knowledge to drive three tailored optimization components: a two-stage global surrogate-guided search, an adaptive focusing search, and an adaptive diversity search. Extensive experiments on the classical benchmark problems, the CEC 2017 test suite, and real-world applications demonstrate that MSAEA significantly outperforms state-of-the-art SAEAs, achieving superior solution quality and robustness.