Solutions to Parallel and Distributed Computing Problems: Lessons from Biological Sciences

2.7 SUMMARY

In this chapter we have presented a brief survey of the main parallel models of evolutionary algorithms, analyzed the potential speedup by parallel execution, and described an efficient implementation of the global selection operation.

Furthermore, we reported some major results from the intensive experimental investigation into the influence of the parameters of the island and neighborhood models of evolutionary algorithms performed by U. Kohlmorgen [15]. By appropriate scaling and normalization of the observed evolutionary processes we could devise statements on the behavior of the evolutionary algorithms over a range of different problem types and problem sizes. In particular, our experiments showed a clear positive influence of the number of islands on the results of the optimization.

In summary, unless one is restricted to using very few generations, it should always be advantageous to use the island model of evolutionary algorithms. If a parallel machine or a workstation cluster is available, this will lead to significant speedup, but even on a single workstation the island model should lead to better results. Furthermore, if a massively parallel machine is available or if run time is not the limiting constraint, the neighborhood model with relatively small neighborhoods should be the preferred choice.