Every Niching Method has its Niche : Fitness Sharing and Implicit Sharing

Various extensions to the Genetic Algorithm (GA) attempt to nd all or most optima in a search space containing several optima. Many of these emulate natural speciation. For co-evolutionary learning to succeed in a range of management and control problems, such as learning game strategies, such methods must nd all or most optima. However, suitable comparison studies are rare. We compare two similar GA specia-tion methods, tness sharing and implicit sharing. Using a realistic letter classiication problem, we nd they have advantages under diierent circumstances. Implicit sharing covers optima more comprehensively, when the population is large enough for a species to form at each optimum. With a population not large enough to do this, tness sharing can nd the optima with larger basins of attraction, and ignore the peaks with narrow bases, while implicit sharing is more easily distracted. This indicates that for a speciated GA trying to nd as many near-global optima as possible, implicit sharing works well only if the population is large enough. This requires prior knowledge of how many peaks exist. In a co-evolutionary GA, individuals are evaluated by how they perform against individuals in the same GA population, or perhaps another GA running in parallel. This is a promising way to learn game strategies 1] 3] 17] 21]. Without speciation, a GA will converge to only one high-tness solution, due to genetic drift. So a co-evolutionary GA will converge and overspecialise to only one strategy for the game being learned. This generalises poorly, and is na vely vulnerable to novel strategies not in the converged population 3]. To overcome this problem, speciation can greatly delay genetic drift 12], and prevent the convergence and overspecialisation that can otherwise happen in co-evolutionary learning 3]. Speciation with co-evolution can also nd a repertoire of high-quality strategies for a game, giving improved generalisation 4]. However , this approach relies heavily on the speciated GA to nd all (or most) good strategies | as speciation nds more high-quality strategies, generalisation ability improves 4, page 92]. So knowing what GA speciation method covers more peaks is important for co-evolutionary learning, not only for games but for other management and control tasks.