Parallel Genetic Algorithms with Distributed Panmictic Populations

Genetic algorithms (GAs) are commonly parallelized using multiple communicating populations or by keeping one population and dividing the task of evaluating the tness among several processors. This paper examines an algorithm where the population is physically distributed, but behaves like a single panmictic unit. This is a desirable property because much more is known about single-population GAs than about multi-population algorithms. The paper analyzes the algorithm and shows how to nd the optimal number of processors that minimizes the execution time. The results show that|despite frequent interprocessor communications|the algorithm may eeectively integrate a large number of processors. The results may also be applicable to a bounding case of multi-population GAs that is very similar to the single distributed population.