An Ensemble Surrogate-Based Coevolutionary Algorithm for Solving Large-Scale Expensive Optimization Problems.

Area of Science:

• Computational Intelligence
• Optimization Theory
• Machine Learning

Background:

• Surrogate-assisted evolutionary algorithms (SAEAs) excel at expensive optimization problems (EOPs).
• Existing SAEAs often struggle with high-dimensional problems, termed large-scale EOPs (LSEOPs).
• A gap exists in addressing LSEOPs effectively with current SAEAs.

Purpose of the Study:

• To propose a novel ensemble surrogate-based coevolutionary optimizer for LSEOPs.
• To enhance the approximation accuracy of surrogate models for LSEOPs.
• To improve the efficiency and performance of SAEAs on large-scale optimization tasks.

Main Methods:

• Feature selection is employed to train local surrogate models on low-dimensional data subsets.
• A selective ensemble surrogate model is constructed for better LSEOP approximation.
• A coevolutionary optimizer with two populations is designed to solve the LSEOP and a simplified auxiliary problem.
• Information sharing between populations facilitates coevolution and leverages search experience.
• An infill selection criterion is utilized to iteratively update and refine the ensemble surrogate model.

Main Results:

• The proposed algorithm demonstrates superior performance on various benchmark LSEOPs.
• Experimental results show significant advantages over nine state-of-the-art SAEAs.
• The ensemble surrogate model effectively approximates the target LSEOP.

Conclusions:

• The developed ensemble surrogate-based coevolutionary optimizer is highly effective for LSEOPs.
• The coevolutionary strategy enhances the search process by utilizing a simplified auxiliary problem.
• The approach offers a promising direction for solving computationally expensive, high-dimensional optimization problems.