Use of very small training data subsets in parallel distributed genetic fuzzy rule selection

Abstract

Genetic fuzzy rule selection has been successfully used to design accurate and interpretable fuzzy classifiers from numerical data. Its computation time, however, becomes very long when it is applied to large data sets. To drastically decrease the computation time of genetic fuzzy rule selection without severely degrading the accuracy of obtained fuzzy classifiers, we proposed its parallel distributed implementation on parallel computers with multiple CPU cores. The main feature of our implementation is to divide not only a population but also training data into multiple sub-groups. In this paper, we try to further decrease the computation time by dividing the training data into very small data subsets. While the number of data subsets is the same as the number of CPU cores in our former study, we divide the training data into much more data subsets than CPU cores. Through computational experiments, we examine the effects of using very small data subsets with frequent rotation on the computation time and the accuracy of obtained fuzzy classifiers.