next up previous

4.3.6 Computational Considerations

The procedures controlling the generation of new solutions are so simple that the computational cost of implementing an ES is usually dominated by that associated with the evaluation of the problem functions. It is therefore important that these evaluations should be performed efficiently and is essential if the optimization is to be performed on a serial computer. Advice on how these calculations can be accelerated can be found in the chapter appropriate to the form of the system equations to be solved.

Like GAs, ESs are particularly well-suited to implementation on parallel computers. Evaluation of the objective function and constraints can be done simultaneously for a whole population, as can the production of the new population by mutation and recombination.

If it is possible to parallelize the evaluation of individual problem functions effectively, some thought and, perhaps, experimentation will be needed to determine the level at which multitasking should be performed. This will depend on the number of processors available, the intended population size and the potential speed-ups available. If the number of processors exceeds the population size, multi-level parallelization may be possible.