Distributed parallel genetic algorithms based on multi-agent cooperation

An architecture of a distributed parallel genetic algorithm was developed to improve computing resource allocation in parallel genetic algorithms. The architecture was defined on a network consisted of several computing nodes each of which had several computing units. The algorithm mapped the physical computing nodes to logical computing units using certain parallel model and carried through individual migrating between neighboring units. A system with only two units was used to analyze the performance of the architecture. Four parallel modes, the serial, the simple parallel, the quasi-parallel and the migrating parallel, as well as the individual migrating fractions were introduced. The experiments of searching for the global maximum of the Schaffer function show that the quasi-parallel model is better than the simple parallel mode or the serial model when there is only one node, while the migrating parallel mode is better than all the other modes for certain range of computing power re-mapping fraction. Heuristic arguments were also provided for the understanding of these observations