Evolving Fuzzy-Probabilistic Neural Network and Its Online Learning

Abstract

In paper rapid learning algorithm for evolving fuzzy-probabilistic neural-network (EFPNN) in which architecture and characteristics of activation functions are configured is proposed. The proposed approach is based on methods of lazy learning, the concept "Neurons in data points", selflearning based on "Winner takes all", classic supervised learning and evolving systems, allowing to change the network architecture simultaneously with its parameter tuning. The peculiarity of EFPNN is ability of classification task solving in case of fuzziness - overlapping classes. In doing so, in classification process probability of membership of every image to particular class and fuzzy membership levels of this data are estimated. Proposed EFPNN is designed to solve fuzzy classification tasks in general problem of Data Stream Mining.