Fuzzy evolutionary probabilistic neural networks

Abstract

One of the most frequently used models for classification tasks is the Probabilistic Neural Network. Several improvements of the Probabilistic Neural Network have been proposed such as the Evolutionary Probabilistic Neural Network that employs the Particle Swarm Optimization stochastic algorithm for the proper selection of its spread (smoothing) parameters and the prior probabilities. To further improve its performance, a fuzzy class membership function has been incorporated for the weighting of its pattern layer neurons. For each neuron of the pattern layer, a fuzzy class membership weight is computed and it is multiplied to its output in order to magnify or decrease the neuron's signal when applicable. Moreover, a novel scheme for multi-class problems is proposed since the fuzzy membership function can be incorporated only in binary classification tasks. The proposed model is entitled Fuzzy Evolutionary Probabilistic Neural Network and is applied to several real-world benchmark problem with promising results. © 2008 Springer-Verlag Berlin Heidelberg.