A critical problem in benchmarking and analysis of evolutionary computation methods

journal article in Web of Science

en

Benchmarking constitutes a cornerstone in the analysis and development of computational methods. Especially in the field of evolutionary computation, where theoretical analysis of the algorithms is almost impossible, benchmarking is at the center of attention. In this text, we show that some of the frequently used benchmark functions that have their respective optima in the center of the feasible set pose a critical problem for the analysis of evolutionary computation methods. We carry out the analysis of seven recent methods, published in respected journals, which contain a center-bias operator that lets them find these optima with ease. This makes their comparison with other methods (that do not have a center-bias) meaningless on such types of problems. We perform a computational comparison of these methods with two of the oldest methods in evolutionary computation on shifted problems and on more advanced benchmark problems. The results show a serious problem, as only one of the seven methods performed consistently better than the pair of old methods, three performed on par, two performed very badly, and the worst one performed barely better than a random search. We also give several suggestions that could help to resolve the presented issues.

Benchmarking constitutes a cornerstone in the analysis and development of computational methods. Especially in the field of evolutionary computation, where theoretical analysis of the algorithms is almost impossible, benchmarking is at the center of attention. In this text, we show that some of the frequently used benchmark functions that have their respective optima in the center of the feasible set pose a critical problem for the analysis of evolutionary computation methods. We carry out the analysis of seven recent methods, published in respected journals, which contain a center-bias operator that lets them find these optima with ease. This makes their comparison with other methods (that do not have a center-bias) meaningless on such types of problems. We perform a computational comparison of these methods with two of the oldest methods in evolutionary computation on shifted problems and on more advanced benchmark problems. The results show a serious problem, as only one of the seven methods performed consistently better than the pair of old methods, three performed on par, two performed very badly, and the worst one performed barely better than a random search. We also give several suggestions that could help to resolve the presented issues.

Evolutionary computation; Metaheuristics; Benchmarking; Zero-bias

12.12.2022

2522-5839

4

1238–1245

8