Performance Analysis Of GRNN And PNN In Classification Algorithm

The Artificial Neural Network is a hot topic and the main direction of many scientists and biologists since the mid-19 th century.It is a highly nonlinear system that mimics biological neurons.It has a lot of advantages such as powerful parallel computing power,self-organizing of network,adaptive parameter and distributed information storage.After half a century of development of neural networks,a variety of achievements have been made.The learning types of the network model can be divided in to supervised learning and unsupervised learning neural network.There are various applications of neural network in optical,such as optical neural network,information optics,quantum optics,integrated optics and adaptive optics.A lot of achievements have been achieved.GRNN(Generalized Regression Neural Network)and PNN(Probabilistic Neural Networks)are both supervised learning neural networks,which are based on the types of neural network of radial basis functions.Since the network training of GRNN and PNN is a single direction and the learning process is one-way,there is no need for multiple iterations.Once the network is formed,it is not necessary to adjust the parameters in the network.Therefore,the learning speed of GRNN and PNN is much faster than that of traditional neural network based on convergence algorithm.GRNN and PNN networks have been widely applied in optical image processing and information optics.There are many fields such as optical character recognition system,optical detection technology,digital watermarking and optical measurement.The pros and cons of GRNN and PNN application and emphasis are different,especially in the practical application of modeling and scientific research.In order to make sure clearly what kind of network model will solve the problem better,the iris data is applied as the test database,and the simulation model is established with MATLAB software simulation platform in this paper.The performance of the two network models is compared in levels of the self-learning ability,the false rejection ability,the robustness,the effect on the prediction accuracy because of the difference between the spread value and the Gaussian amplitude value a.The main work is as follows.(1)The test sample is maintained constant.Every time a sample is added to the training sample.Then the prediction accuracy is showed in the simulation results.And the statistical average of the prediction accuracy is maintained in the process of repeating.Finally,the relationship between the number of samples and the accuracy of predicting is showed in the image.Comparative experiment is taken by changing the noise gradient and increasing the number of simulation repetitions.The analysis shows that the self-learning ability of PNN model is stronger than that of GRNN model.And if the training samples continue to be added,the GRNN model will improve the accuracy rate.(2)One test sample is replaced by one noise at a time while training samples remain stable.Then,the statistical average of the prediction accuracy is obtained by simulation.Finally,the relationship between the number of noise to replace the samples and the corresponding prediction error is plotted.And the noise gradient is added as well as the number of repetitions increased to increase the contrast experiment.The results show that the GRNN model and the PNN model both work well in the false rejection ability.Prediction error are acceptable.However,the false rejection ability of GRNN model is stronger than that of the PNN model.Therefore,we should give priority to the GRNN model in some areas of high precision requirements.(3)Each time a sample is replaced with one noise while the test sample is remained constant.This method is imitating the test of the false rejection ability.Then,the statistical average of the prediction accuracy is obtained by simulation.Finally,the relationship between the number of noise to replace the sample and the corresponding accuracy of predicting is plotted.Comparative experiment is taken by changing the noise gradient,increasing the number of simulation repetitions and setting the relative proportions of the test sample and the training sample to increase the contrast experiment.The robustness of PNN model based on the experimental analysis is stronger than that of GRNN model.However,after the same trend is reduced to about 50% of the training sample error rate,the GRNN model descends much faster than the PNN model.(4)Spread is a variable increasing from 0.1 at a certain rate.Then,the statistical average of the prediction accuracy is obtained by simulation.Finally,the relationship between the spread and the corresponding prediction accuracy is plotted.Observation of image shows that they both keep balance with the same trend at first.But the GRNN model drops at a greater rate after the spread value of 0.3.And the minimum value of the prediction accuracy of the GRNN model is also much smaller than the minimum value of PNN model.It shows that the changing of spread value has a greater influence on the GRNN model than the PNN model.Experimental analysis of imitation spread values show that the value of a Gaussian amplitude value is more appropriate to set at about 0.5.And the PNN model should be prioritized when the amplitude a is same.