Multipopulation-Based Differential Evolution for Large-Scale Many-Objective Optimization.

Area of Science:

• Evolutionary Computation
• Optimization Algorithms
• Computational Intelligence

Background:

• Many-objective optimization evolutionary algorithms struggle with scalability in large-scale problems.
• High-dimensional spaces in large-scale many-objective optimization problems (LSMaOPs) can lead to loss of diversity and premature convergence.
• Existing algorithms often exhibit poor convergence or diversity maintenance when tackling LSMaOPs.

Purpose of the Study:

• To propose an efficient and effective multipopulation-based differential evolution algorithm, LSMaODE, for solving LSMaOPs.
• To address the scalability challenges faced by existing algorithms in high-dimensional optimization landscapes.
• To enhance both convergence and diversity maintenance for nondominated solutions in LSMaOPs.

Main Methods:

• A multipopulation strategy divides the population into two groups for distinct optimization approaches.
• Randomized coordinate descent optimizes 10% of individuals for independent decision variable exploitation and diversity maintenance.
• Nondominated guided random interpolation optimizes the remaining 90% for rapid convergence and objective space distribution.

Main Results:

• LSMaODE demonstrates efficient and effective performance on LSMOP test suites, evaluating scalability in decision and objective dimensions.
• Comparative analysis against five state-of-the-art algorithms shows competitive results.
• The proposed algorithm successfully maintains diversity and achieves good convergence in high-dimensional spaces.

Conclusions:

• LSMaODE offers a robust solution for large-scale many-objective optimization problems.
• The multipopulation approach with distinct strategies effectively balances exploration and exploitation.
• LSMaODE presents a promising advancement in evolutionary algorithms for complex optimization tasks.