A diverse population is the primary source of evolutionary algorithms' robustness, countering the tendency for local optima to trap search. A converged population provides no advantage, behaving like a single individual. The exploratory behavior of evolutionary algorithms can be controlled by algorithm parameters - population size, mutation rates and step sizes, and selection pressure -but the effects are limited. Thus, diversity-promoting mechanisms, inspired by natural evolutionary systems, have been heavily studied. The most widely used methods rely either on resource competition promoting evolutionary niches (fitness-sharing methods), or on restricted interaction (island models) in which subpopulations evolve independently, except for a trickle of individuals migrating between islands. Island models are particularly amenable to parallel implementation, so are increasingly important as parallel systems dominate hardware development.
Was this article helpful?