An Evolutionary Algorithm Assisted by an Ensemble of Pareto-Optimal Surrogate Models

Area of Science:

• Computational intelligence
• Machine learning
• Optimization

Background:

• Surrogate models enhance surrogate-assisted evolutionary algorithms (SAEAs) by improving prediction quality and robustness.
• Effective ensemble surrogate models require careful design of fitness landscape approximation smoothness, an aspect often overlooked.
• Existing adaptive/ensemble SAEAs primarily focus on prediction accuracy, neglecting other crucial model properties.

Purpose of the Study:

• To propose an adaptive ensemble SAEA that automatically constructs ensemble models by optimizing parameter settings.
• To address the under-exploration of explicit tuning for surrogate model smoothness in ensemble approaches.
• To develop a robust ensemble modeling strategy that balances approximation error and model complexity.

Main Methods:

• An adaptive ensemble SAEA is proposed, optimizing radial basis function network (RBFN) parameters.
• A bi-objective minimization approach is used to balance approximation error and model complexity.
• A novel infill criterion is designed to leverage surrogate models with varying smoothness for solution prescreening.

Main Results:

• The proposed algorithm reduces over/under-fitting by optimizing RBFN structures.
• Ensemble models with varying degrees of smoothness are robustly constructed.
• Experimental results show statistical superiority over state-of-the-art SAEAs on benchmark and real-world problems.

Conclusions:

• The adaptive ensemble SAEA offers improved performance in expensive optimization settings.
• Optimizing surrogate model smoothness alongside accuracy leads to more robust and effective ensembles.
• The developed algorithm provides a statistically superior alternative to existing SAEAs.