Dynamic distributed genetic algorithms

Abstract

Distributed populations in genetic algorithms can make the search more smart, in that local minima may be skipped. However, when the global population is divided into small sub-populations, the ability of these sub-populations to evolve is set back because of their relatively small sizes. In this paper, a new method to manage the distributed populations in evolution is introduced. A supervising subroutine observes all the sub-populations during evolution. The sizes of these sub-populations are dynamically changed according to their performance. Better sub-populations get more quotas of the total number of individuals, thus get more possibility to produce even better ones. This algorithm is illustrated with an example. Different policies of managing the sub-populations are compared and discussed. The main conclusion is that dynamical rearrangement of the global population can make the process of evolution faster and more stable.