Ensemble Fuzzy Rule-Based Classifier Design by Parallel Distributed Fuzzy GBML Algorithms

We have already proposed an island model for parallel distributed implementation of fuzzy genetics-based machine learning (GBML) algorithms. As in many other island models, a population of individuals is divided into multiple subpopulations. Each subpopulation is assigned to a different island. The main characteristic feature of our model is that training patterns are also divided into multiple training data subsets. Each subset is assigned to a different island. The assigned subset is used to train the subpopulation in each island. The assignment of the training data subsets is periodically rotated over the islands (e.g., every 100 generations). A migration operation is also periodically used. Our original intention in the use of such an island model was to decrease the computation time of fuzzy GBML algorithms. In this paper, we propose an idea of using our island model for ensemble classifier design. An ensemble classifier is constructed by choosing the best classifier in each island. Since the subpopulation at each island is evolved using a different training data subset, a different classifier may be obtained from each island to construct an ensemble classifier. This suggests a potential ability of our island model as an ensemble classifier design tool. However, the diversity of the obtained classifiers from multiple islands seems to be decreased by frequent training data subset rotation and frequent migration. In this paper, we examine the effects of training data subset rotation and migration on the performance of designed ensemble classifiers through computational experiments.