General scheme for analyzing running times of parallel evolutionary algorithms

Abstract

We present new methods for the running time analysis of parallel evolutionary algorithms with spatially structured populations. These methods are applied to estimate the speed-up gained by parallelization in pseudo-Boolean optimization. The possible speed-up increases with the density of the topology. Surprisingly, even sparse topologies like ring graphs lead to a significant speed-up for many functions while not increasing the total number of function evaluations. We also give practical hints towards choosing the minimum number of processors that gives an optimal speed-up. © 2010 Springer-Verlag.