Evolutionary Optimization of Expensive Multiobjective Problems With Co-Sub-Pareto Front Gaussian Process Surrogates.

Area of Science:

• Computational intelligence
• Optimization algorithms
• Machine learning

Background:

• Multiobjective optimization problems (MOPs) are common in science and engineering.
• Computationally expensive objectives in MOPs necessitate efficient surrogate-assisted optimization.
• Existing methods struggle to effectively exploit correlations across different parts of the Pareto front.

Purpose of the Study:

• To develop a novel surrogate augmentation strategy for evolutionary multiobjective optimization.
• To address the challenge of computationally expensive objective functions.
• To improve the efficiency and effectiveness of finding Pareto optimal solutions.

Main Methods:

• Decomposition of multiobjective problems into subproblems, each approximating a sub-Pareto front (subPF).
• Application of a multitask Gaussian process (GP) model for joint surrogate learning across subproblems.
• Development of a new utility function criterion and solution management strategy for efficient model building.

Main Results:

• The proposed Gaussian process based co-sub-Pareto front surrogate augmentation strategy effectively exploits correlations between subproblems.
• Joint learning enables knowledge transfer across approximated subPFs, enhancing optimization performance.
• Experimental results demonstrate superior performance compared to state-of-the-art algorithms for expensive MOPs.

Conclusions:

• The proposed method offers a significant advancement in surrogate-assisted evolutionary multiobjective optimization for expensive problems.
• Jointly inferring multiple subproblems and transferring knowledge across them is a key strength.
• The strategy provides a more efficient and effective approach to approximating the Pareto front in complex scenarios.