Font Size: a A A

Research On Probabilistic Neural Network Algorithm In Pattern Classification

Posted on:2019-09-25Degree:MasterType:Thesis
Country:ChinaCandidate:X YanFull Text:PDF
GTID:2428330548958933Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
We can identify a wide variety of things in our lives and categorize things of different attributes,this behavior can be expressed in a professional term--pattern recognition(pattern classification).Human beings hope that computers can also implement this intelligent behavior.Based on the development of pattern recognition,machine learning and deep learning,artificial intelligence has become one of the hottest topics and fields at present.Bayesian decision theory is the basic theoretical method in statistical decision-making model,which can solve many classification problems in pattern recognition.Given the known prior probability and the conditional probability density,the posterior probability is obtained by the Bayesian formula to make the decision of the minimum error rate.Artificial neural network(ANN)is a hot model in the field of artificial intelligence.This model imitates the behavioral characteristics of animal neural network and carries out distributed parallel information processing.The Probability neural network model(PNN)was proposed by D.F.Specht in 1990,which applys Bayesian decision theory to the ANN model.PNN is simple,fast,accurate and robust,so it is widely used in pattern classification.PNN has high requirements on the representativeness and quantity of training samples,the performance of small-scale data can not be guaranteed,and a large number of training samples will increase the computational cost.On the other hand,although a single global variable makes the training of the neural network easier,it will also mask the partial properties and heterogeneity of the pattern.This paper puts forward several improvements on the basis of PNN model: add a feature layer to improve the dimension;adopt heterogeneous PNN(Mixed Gaussian model is used in the model layer,the input of each model is a Gaussian model with a separate parameter.);take the maximum value as output in the summation layer.Finally,the paper verifies the performance of the improved PNN on the datasets: wine,Iris,and breast cancer,which are from the UCI database.The experiment consists of two steps: first,we use all of the data for training and testing;then the data is divided into training data and test data in a certain proportion and then used for training and testing.The experimental results show that the improved PNN classifies the training data very well,andthe classification accuracy on the three data sets can reach 100%,but the performance on the test data is general,which shows that the improved PNN exists over-fitting problem.
Keywords/Search Tags:Pattern classification, Pattern recognition, Artificial neural network, Probability neural network, Bayesian decision theory, Parzen window
PDF Full Text Request
Related items