Probilistic Neural Network For Pattern Recognitions

Probabilistic neural network (PNN) is a classification network, which is based on Bayesian decision theory and probability function estimation theory. Because the structure is simple , PNN is an effective model for many classification problems and many pattern recognitions.The main factors of probabilistic neural network including the hidden neuron size, hidden central vector and the smoothing parameter, to influence the PNN classification , are analyzed; The XOR problem is implemented by using PNN. A new supervised learning algorithm for the PNN is developed : the learning vector quantization is employed to group training samples and the Genetic algorithms (GA's) is used for training the network's smoothing parameters and hidden central vector for determining hidden neurons. Simulations results show that, the advantage of our method in the classification accuracy is over other unsupervised learning algorithms for PNN.The convergence of the PNN decision rule for the Bayesian decision rule is proved with probability. An explicit formulation of PNN generalization ability is proposed. And the accuracy rate of a classification network is explained as the maximum likelihood estimation of the generalization ability. The upper bound of the PNN's generalization ability is less than the accuracy rate computed by Bayesian decision rule.