The most natural and widely used method to parallelize GA (Genetic Algorithm) is to divide the global population into several sub-populations, and allocate each sub-population on the processors to be evolved in parallel. In this parallelizing scheme, each processor runs the genetic algorithm on its own subpopulation, while periodically exchanging some information with other processors. However, the size of sub-population and the way to exchange the information such as the selection of partner processors, the amount (and/or contents) of exchanged information, and the time and frequency of exchanging, are adjustable parameters when designing a parallel genntic algorithm. This paper aims to classify and investigate the existing parallel genetic algorithms with respect to the above design considerations.
抄録全体を表示