Evaluation of population partitioning schemes in Bayesian classifier EDAs: Estimation of distribution algoithms

Several algorithms within the field of Evolutionary Computation have been proposed that effectively turn optimisation problems into supervised learning tasks. Typically such hybrid algorithms partition their populations into three subsets, high performing, low performing and mediocre, where the subset containing mediocre candidates is discarded from the phase of model construction. In this paper we will empirically compare this traditional partitioning scheme against two alternative schemes on a range of difficult problems from the literature. The experiments will show that at small population sizes, using the whole population is often a better approach than the traditional partitioning scheme, but partitioning around the midpoint and ignoring candidates at the extremes, is often even better.