Synchronous parallel GA with nearly zero sequential computations

Abstract

Generally speaking, genetic algorithms (GAs) can provide global optimal solutions with higher probability. However, compared with conventional approaches, GAs are usually more time consuming. To solve this problem, different parallel GAs (PGAs) have been proposed in the literature. A direct method is to employ many slave processors for individual evaluation, and a master processor to synchronize the whole evolution process. This method is good in the sense that it does not change any properties (say, convergence) of GAs. However, according to Amdahl's (1967) law, this method cannot speedup the evolution process very much, because the percentage of sequential computation is relatively large. We make two proposals: one is to keep all genetic materials inside the slave processors, and the other is to ask the slaves to perform crossover and mutation as well. The former reduces data transmission time, and the latter reduces sequential computing time directly. With these proposals, we have succeeded to achieve almost m times speedup when we use m slaves. This is verified through experiments.