Effects of parallel distributed implementation on the search performance of Pittsburgh-style genetics-based machine learning algorithms

Abstract

Pittsburgh-style genetics-based machine learning (GBML) algorithms have strong search ability for obtaining rule-based classifiers. However, when we apply them to data mining from large data, we need huge computation time for fitness evaluation. In our previous studies, we have proposed parallel distributed implementation of fuzzy GBML for fuzzy classifier design from large data. The basic idea of our parallel distributed implementation is to divide not only a population but also a training data set into N sub-populations and N training data subsets, respectively. A pair of a sub-population and a training data subset is assigned to each of N CPU cores in a workstation or a cluster. This dual division strategy achieved a quadratic speedup (i.e., N2 times faster than the use of a single CPU core) while maintaining the generalization ability on test data. In this paper, we apply our parallel distributed implementation to GAssist which is a non-fuzzy Pittsburgh-style GBML algorithm. We examine the effects of the number of divisions on the search ability comparing with the parallel distributed fuzzy GBML.