Performance Indicator-Based Adaptive Model Selection for Offline Data-Driven Multiobjective Evolutionary...

Area of Science:

• Optimization
• Computational Science
• Machine Learning

Background:

• Real-world multiobjective optimization problems (MOPs) often rely on pre-existing data, termed offline data-driven MOPs.
• Surrogate models in evolutionary algorithms (EAs) can approximate objective functions but suffer from accumulated approximation errors, potentially misranking solutions.

Purpose of the Study:

• To propose a novel surrogate-assisted indicator-based EA for effectively solving offline data-driven MOPs.
• To enhance the robustness of EAs against surrogate model approximation errors in data-limited optimization scenarios.

Main Methods:

• An indicator-based EA serves as the baseline optimizer, leveraging its inherent robustness to surrogate model inaccuracies.
• Kriging models and radial basis function networks (RBFNs) are utilized as surrogate models.
• An adaptive model selection mechanism dynamically chooses between Kriging and RBFNs based on a maximum acceptable approximation error threshold.

Main Results:

• The proposed algorithm demonstrates effectiveness on benchmark MOPs with up to ten objectives.
• Performance was validated on problems with a significant number of decision variables (up to 20 and 30).
• The adaptive mechanism successfully mitigates the misleading effects of approximation errors by selecting appropriate surrogate models.

Conclusions:

• The developed surrogate-assisted indicator-based EA is a promising approach for offline data-driven MOPs.
• Adaptive surrogate model selection is crucial for maintaining search accuracy in data-limited optimization.
• The algorithm shows strong performance and scalability for complex, high-dimensional optimization tasks.