Avaliação de técnicas de regularização aplicadas as redes neurais nebulosas

Abstract

This paper proposes a new training algorithm for fuzzy neural networks that is able to generate parsimonious models with some degree of interpretability. In some cases, as in fuzzy neural networks learning can become a very slow and complex task. In this work learning is performed based on concepts of extreme learning machines to estimate parameters and a feature selection technique based on regularization and resampling called bootstrap lasso, to define the network topology. The use of regularization in the inner layers of the model enables it to be more precise in its answers, and concise set of fuzzy rules can be extracted from the resulting topology allowing the interpretability of the results. Numerical results are presented for pattern classification problems using real bases of large and small dimensions. The results are compared to other classifiers reference in the literature. Statistical analysis of the results suggests that the proposed algorithm has a similar accuracy to regularized extreme machine learning models, but with an interpretable topology.