Large-Scale Multiobjective Evolutionary Algorithm Guided by Low-Dimensional Surrogates of Scalarization Functions.

Area of Science:

• Computational Intelligence
• Optimization Algorithms
• Evolutionary Computation

Background:

• Surrogate-assisted multiobjective evolutionary algorithms efficiently solve computationally intensive problems.
• Existing algorithms struggle with more than 200 decision variables due to unreliable surrogate models.
• Large-scale expensive multiobjective optimization remains a significant challenge.

Purpose of the Study:

• To develop a novel large-scale multiobjective evolutionary algorithm.
• To address the performance deterioration of existing methods with increasing decision variables.
• To enable efficient optimization of problems with up to 1000 decision variables.

Main Methods:

• Proposed a large-scale multiobjective evolutionary algorithm (LDS-AF).
• Utilized low-dimensional surrogate models of scalarization functions.
• Reduced decision space dimensionality using principal component analysis.
• Employed a two-stage modeling and convergence control strategy.

Main Results:

• LDS-AF demonstrated superior performance on large-scale multiobjective optimization problems.
• The algorithm successfully handled problems with up to 1000 decision variables.
• Achieved promising results with only 500 real function evaluations.
• Maintained a balance between convergence and diversity, avoiding premature local optima.

Conclusions:

• LDS-AF overcomes the limitations of existing algorithms for large-scale problems.
• The proposed method offers an effective approach for computationally expensive multiobjective optimization.
• LDS-AF shows significant potential for solving complex optimization tasks with high dimensionality.