Adaptive Learning of Probabilistic Neural Network in Situation of Overlapping Classes in Classification Task

0
346

Abstract

The adaptive probabilistic neural networks for classification task in situation of overlapping classes is proposed. This network is designed to solve data classification task when data are fed sequentially in the online mode, and forming classes are mutually overlapped – the fuzzy case. The distinct feature of the network is that the learning process of the pattern layer uses the sliding window. This allows us to keep the constant number of neurons in this layer. Another point of the learning process is the tuning ability of activation functions’ widespread parameters in online mode. The described advantage allows us to improve the classification quality. Last but not least is the ability to compute both the probability and membership levels of each observation to each of forming classes. The proposed adaptive probabilistic neural network with fuzzy interference is simple in numerical implementation and has high learning speed. The results of experiments confirmed the correctness of approach under consideration.