Effects of Periodical Rotation of Training Data Subsets on Evolution Process in Parallel Distributed Genetic Fuzzy Rule Selection

Abstract

In our parallel distributed genetic fuzzy rule selection, a training data set and a population are divided into subgroups. The computation time can be faster by assigning a pair of a training data subset and a sub-population to each of multiple processor cores. To avoid the overfitting of the sub-population to the corresponding training data subsets, the rotation of training data subsets is applied periodically. In this paper, we examine the effects of the rotation operation on the evolution process of the sub-populations and the generalization ability by monitoring the change in the fitness of each sub-population.