文章基本信息

标题：On the Extinction of Evolutionary Algorithm Subpopulations on Multimodal Landscapes
本地全文：下载
作者：Lutz Schönemann ; Michael Emmerich ; Mike Preuss 等
期刊名称：Informatica
印刷版ISSN：1514-8327
电子版ISSN：1854-3871
出版年度：2004
卷号：28
期号：4
页码：345-351
出版社：The Slovene Society Informatika, Ljubljana
摘要：Population based evolutionary algorithms (EA) are frequently used to optimize on multimodal functions. A common assumption is that during search several subpopulations might coexist in different attraction regions of the search space. Practical experience and takeover--time considerations suggest that this is not true in general. We therefore analyze the stability of subpopulations within a simplified EA on a two-attractor model, focusing on two extreme cases: (1) Function values of both local minima are exactly the same and (2) function values on the first attractor are always better than on the second. Realistic scenarios for bimodal optimization are assumed to be located in between these two extremes, such that upper and lower bounds for extinction times can be estimated, e.g. by Markov chain analysis and empirical studies. The obtained results provide new insights into the effect of μ+,λ selection on the stability of subpopulations and the effect of genetic drift. Moreover, the effect of idealized niching on the same scenarios is investigated, leading to an immense increase of the EA's ability to perform concurrent search. Our model and the findings based thereupon do not depend on the number of problem dimensions.
关键词：Evolutionary algorithms; multimodal fitness landscape; niching techniques; random genetic drift