An efficient estimation of distribution algorithm with rank-one modification and population reduction

Area of Science:

• Artificial Intelligence
• Optimization Algorithms
• Computational Intelligence

Background:

• Estimation of Distribution Algorithms (EDAs) are model-based metaheuristics utilizing statistical learning.
• Gaussian EDAs (GEDAs) often suffer from premature convergence, limiting their effectiveness.
• Improving GEDA performance requires addressing model estimation and population size tuning.

Purpose of the Study:

• To enhance the performance of Gaussian EDAs (GEDAs) by refining their model estimation and population management.
• To introduce a novel EDA variant, EDA-R1M-PR, integrating improved strategies.
• To evaluate the effectiveness of the proposed strategies in improving global optimization and search efficiency.

Main Methods:

• Developed a new model estimation method for GEDAs, incorporating a rank-one modification (R1M) of the covariance matrix.
• Introduced a population reduction (PR) strategy to dynamically adjust population size during evolution.
• Combined R1M and PR strategies to create the EDA-R1M-PR algorithm.

Main Results:

• Theoretical analysis confirmed the R1M strategy effectively adjusts search scope and direction.
• The PR strategy balances exploration and exploitation, optimizing resource utilization.
• Experimental results demonstrated that EDA-R1M-PR significantly outperforms state-of-the-art evolutionary algorithms on benchmark functions.

Conclusions:

• The R1M and PR strategies substantially enhance the global optimization ability of GEDAs.
• The proposed EDA-R1M-PR variant offers improved search efficiency and superior performance compared to existing algorithms.
• This research provides a promising approach for advancing model-based metaheuristic optimization.