Asymmetric Sharing

Where natural evolution has lead to a wide variety of coexisting species, evolutionary algorithms tend to end up with nearly identical individuals. Techniques known as niching aim to avoid this. However, they often depend on a threshold parameter, known as the sharing radius or niche radius. It serves as a similarity threshold. This is a critical, problem-dependent parameter. How to set it is known as the niche radius problem. Inspired by the self-adaptation of mutation parameters in evolution strategies, this paper proposes a variant of the niching technique fitness sharing, asymmetric sharing. It is an attempt at self-adaptation of the niche radius. Test runs on a variety of multimodal functions indicate asymmetric sharing has an improved ability of finding and maintaining multiple local optima as compared to standard fitness sharing, without the need of an external parameter. The algorithms presented are built and tested within the framework of evolution strategies, but could potentially be transported to a broad class of evolutionary algorithms.