Beware the Parameters: Estimation of Distribution Algorithms Applied to Circles in a Square Packing

摘要

Simple continuous estimation of distribution algorithms are applied to a benchmark real-world set of problems: packing circles in a square. Although the algorithms tested are very simple and contain minimal parameters, it is found that performance varies surprisingly with parameter settings, specifically the population size. Furthermore, the population size that produced the best performance is an order of magnitude larger that the values typically used in the literature. The best results in the study improve on previous results with EDAs on this benchmark, but the main conclusion of the paper is that algorithm parameter settings need to be carefully considered when applying metaheuristic algorithms to different problems and when evaluating and comparing algorithm performance.