Differential evolution with auto-enhanced population diversity.

Area of Science:

• Computational Intelligence
• Optimization Algorithms
• Evolutionary Computation

Background:

• Differential evolution (DE) algorithms commonly employ parameter adaptation for mutation factor (F) and crossover probability (CR).
• Existing parameter adaptation methods struggle to effectively resolve issues of premature population convergence and stagnation.
• Population diversity is a critical factor influencing the performance of DE algorithms.

Purpose of the Study:

• To propose a novel mechanism, auto-enhanced population diversity (AEPD), for automatically enhancing population diversity in DE.
• To address the limitations of existing methods in preventing premature convergence and stagnation.
• To improve the robustness and performance of DE algorithms.

Main Methods:

• Investigated population adaptation by analyzing population diversity at the dimensional level.
• Developed AEPD to identify and rectify population convergence or stagnation in specific dimensions.
• Integrated AEPD into a standard DE algorithm and evaluated its effectiveness on benchmark functions.

Main Results:

• AEPD significantly enhanced the performance of the base DE algorithm across 25 CEC2005 benchmark functions.
• The AEPD mechanism reduced the sensitivity of the DE algorithm to population size.
• The DE algorithm incorporating AEPD demonstrated superior performance compared to several peer algorithms.

Conclusions:

• AEPD is an effective mechanism for automatically managing population diversity in DE algorithms.
• The proposed method successfully mitigates premature convergence and stagnation, leading to improved optimization performance.
• AEPD offers a robust solution for enhancing DE algorithm efficiency and adaptability across various problems.