Scaling genetically guided fuzzy clustering

Describes improvements to previous work on the use of genetic algorithms and evolutionary strategies to generate fuzzy partitions of unlabeled data. It was found that genetically guided clustering could be used in some domains to produce fuzzy partitions for which the objective function does not get trapped in local extrema. Gray code representation, two-point crossover tournament selection, and variable crossover and mutation rates combine for improved performance in terms of the final partition and the required population size. Also, methods to allow this approach to scale to problems such as magnetic resonance imaging (22000 pixels in 3D) are detailed. Initially, using properly chosen subsamples of the full data allows the genetically guided approach to move to good regions in the search space quickly. Examples are given to show how local extrema of an objective function (and thereby poor data partitions) can be avoided in domains with a large number of patterns.