A Self-Guided Reference Vector Strategy for Many-Objective Optimization.

Area of Science:

• Computational Intelligence
• Optimization Algorithms
• Evolutionary Computation

Background:

• Decomposition-based evolutionary algorithms for many-objective optimization (MaOEA/Ds) commonly use reference vectors (RVs) for search direction and diversity.
• The performance of MaOEA/Ds is sensitive to the alignment between reference vectors and the true Pareto front (PF).

Purpose of the Study:

• To propose a novel self-guided reference vector (SRV) strategy to enhance the performance of MaOEA/Ds.
• To develop an adaptive method for generating reference vectors directly from the population data.

Main Methods:

• A modified k-means clustering method is employed to extract reference vectors from the population.
• An angle-based density measurement strategy is utilized for robust initialization of cluster centroids.
• The proposed SRV strategy is integrated into existing MaOEA/Ds to evaluate its effectiveness.

Main Results:

• The SRV strategy demonstrates superior performance compared to fixed reference vector approaches in MaOEA/Ds.
• The method effectively adapts reference vectors to match both regular and irregular Pareto front shapes.
• Simulation results confirm the strategy's ability to improve search direction and population diversity.

Conclusions:

• The self-guided reference vector strategy offers a significant improvement for decomposition-based many-objective evolutionary algorithms.
• This adaptive approach enhances the robustness and efficiency of optimization processes across diverse problem landscapes.
• The SRV method provides a promising direction for future advancements in many-objective optimization research.