Definição automática de classificadores fuzzy probabilísticos

Abstract

This work presents a new approach for the automatic design of Probabilistic Fuzzy Classifiers (PFCs), which are a special case of Probabilistic Fuzzy Systems. As part of the design process we consider methods for reducing the dimensionality like the principal component analysis and the Fisher discriminant. The clustering methods tested for partitioning the universe of input variables are Gustafson-Kessel and Supervised Fuzzy Clustering, both consolidated in the literature. In addition, we propose a new clustering method called Gustafson-Kessel with Focal Point as part of the automatic design of PFCs. We also tested the capacity of this method to deal with ellipsoidal and non-ellipsoidal clusters. Highly correlated data represent a challenge to fuzzy clustering due to the inversion of the fuzzy covariance matrix. Therefore, a regularization method is necessary for this matrix and a new one is proposed in this work. In the proposed PFCs, the combination of antecedents and consequents provides a rule base in which all consequents are possible, each one associated with a probability measure. In this work, the probability is calculated based on the Bayes Theorem by updating, through the likelihood function, a priori information concerning every consequent in each rule. The main innovation is the calculus of the likelihood functions which is based on the “ideal region” concept, aiming to improve the estimation of the probabilities associated with rules’ consequents. The proposed PFCs are compared with fuzzy-bayesian classifiers and other ones traditional in machine learning over artificial generated data, 30 different benchmarks and also on data directly extracted from real world like the problem of detecting bearings fault in industrial machines. Experiments results show that the proposed PFCs outperform, in terms of accuracy, the fuzzy-bayesian approaches and are competitive with the traditional non-fuzzy classifiers used in the comparison. The results also show that the proposed regularization method is an alternative to the Gustafson-Kessel clustering technique (with or without focal point) when using linearly correlated data.