Using landscape topology to compare continuous metaheuristics: a framework and case study on EDAs and ridge structure

Area of Science:

• Computational intelligence
• Optimization algorithms
• Machine learning

Background:

• Continuous metaheuristic optimization algorithms are widely used but their behavior on different problem landscapes is not fully understood.
• Explicit dependency modeling in Estimation of Distribution Algorithms (EDAs) is a key area for improving performance, but its effectiveness varies.
• A robust methodology is needed to systematically evaluate algorithm performance across diverse problem instances.

Purpose of the Study:

• To investigate the impact of explicit dependency modeling in continuous Estimation of Distribution Algorithms (EDAs).
• To analyze the behavior of metaheuristic optimization algorithms on parameterized, two-dimensional landscapes with linear ridge structures.
• To develop and apply a comparative experimental methodology for understanding algorithm performance relative to landscape features.

Main Methods:

• Utilized a previously proposed randomized landscape generator to create two-dimensional landscapes with parameterized, linear ridge structures.
• Employed a comparative experimental methodology, performing pairwise comparisons of algorithm instances.
• Applied heat maps for visualizing and comparing algorithm performance across numerous landscape instances and trials; conducted a meta-search in the landscape parameter space.

Main Results:

• Identified specific landscape features where explicit dependency modeling in EDAs is beneficial, detrimental, or has a neutral impact on mean performance.
• Demonstrated that algorithm performance is highly dependent on the interplay between algorithm characteristics and landscape structure.
• Discovered new insights into the relationship between dependency modeling in EDAs and problem landscape characteristics, extending prior intuition.

Conclusions:

• The developed landscape generator and comparative methodology provide a generalizable framework for studying continuous optimization algorithms.
• Explicit dependency modeling in EDAs has a nuanced effect on performance, contingent on specific landscape features.
• The findings offer a basis for designing more effective optimization algorithms tailored to specific problem structures.