Parallel heterogeneous genetic algorithms for continuous optimization

Abstract

In this paper we address the physical parallelization of a very efficient genetic algorithm (GA) known as gradual distributed real-coded GA (GD-RCGA). This search model naturally provides a set of eight subpopulations residing in a cube topology having two faces for promoting exploration and exploitation. The resulting technique has been shown to yield very accurate results in continuous optimization by using crossover operators tuned to explore and exploit the solutions inside each subpopulation. Here, we encompass the actual parallelization of the technique, and get deeper into the importance of running a synchronous versus an asynchronous version of the basic GD-RCGA model. We also present the evaluation of the parallel execution of GD-RCGA over two local area networks, a Fast-Ethernet network and a Myrinet network. Our results indicate that the GD-RCGA model maintains a very high level of accuracy for continuous optimization when run in parallel, and we also demonstrate the relative advantages of each algorithm version over the two networks. Finally, we show that the async parallelization scales better than the sync one, what suggests future research lines for WAN execution and new models of search based on the original two-faced cube.