The nature of niching: genetic algorithms and the evolution of optimal, cooperative populations

Genetic algorithms (GAs) with have been analyzed and successfully applied to problems in search and optimization, while GAs using various types of have been incorporated into classifiers, immune system models, artificial ecologies, artificial economies, etc. Both types of sharing are based on the same observation of nature: dividing a finite resource among competing organisms limits the size of populations dependent on that resource. If multiple resources are involved, each resource can be considered a niche, and each subpopulation exploiting a niche can be considered a species. By treating population slots as a finite resource, we obtain the traditional, fixed-size population simple GA. By further treating the figure of merit (i.e., the function to be optimized) as a limited resource, we obtain fitness If our problem domain manifests explicit resources, such as rewards for correct classification of examples, we can enforce resource Both fitness and resource sharing have been studied separately, and usually under the assumption of non-overlapping niches, or sharing. Little effort has been made to understand the more complicated situation of niche overlap. Yet the resolution of niche overlap is critical to successful, useful niching. Non-overlapping niches can be covered by non-competing species, while heavily overlapped niches induce competition between species in which only the best will survive to represent the overlapping resources. Somewhere in between the two extremes of overlap must lie a critical boundary between cooperation (all species survive) and competition (one dominant species survives). We seek this boundary for the simplest case of niche overlap: two niches. We first define the general mechanism of sharing, and investigate where sharing-induced-niching fits into the larger frameworks of context-dependent function optimization, natural ecologies, and models of cooperation and competition. We then map fitness sharing to resource sharing, and show that they are identical when there is no niche overlap (i.e., perfect sharing). We go on to analyze the three cases: perfect sharing, fitness sharing (with overlap) and resource sharing (with overlap). We find that even under severe selective pressure with high degrees of niche overlap, a stable equilibrium population is quickly found and indefinitely maintained, a population consisting of diverse species covering the best niches. We find that maintenance of the sharing equilibrium degrades gradually as niche overlap and fitness discrepancies increase. We create a control map for the two-niche case, plotting the boundary between surviving niche pairs (cooperation) and untenable niche pairs (competition). Controlling this boundary will allow us to use sharing to evolve the types of cooperation and competition appropriate to the problem at hand. We extend the above analysis by looking at some aspects of the gen