A comparison of localised and global niching methods

Abstract

Niching methods are a useful extension of evolutionary computation that allow evolutionary algorithms to be applied in multimodal problem domains. Current niching methods use either one of two methods to promote the formation of species within a population. Genetics-based methods, such as fitness sharing or clearing, work directly on the search space of the problem. Alternatively, spatiallystructured evolutionary algorithms are used to place individuals onto a landscape and restrict mating to within isolated demes of population members. This isolation promotes the driving of geographically distant individuals towards separate parts of the search space. This paper introduces the concept of localised niching (LC). LC takes the traditionally global operations used in genetics-based niching and applies them locally in a spatially-structured population. Testing on two well known and difficult benchmark problems indicates that LC not only has the potential to significantly outperform the traditional global niching methods, but is also more resistant to some of the known limitations of genetics-based species formation.