Research On RBF Hybrid Structure Neural Network Classifier Based On Adaptive Kernel Learning

Classification is one of the most common decision-making tasks in human activities.Nowadays,the neural network algorithms and models used for classification have been widely used in various fields,such as business,science,industry and medicine.Although a variety of neural network algorithms and models have been greatly developed,because of the difference of algorithm and model,and the diversity and complexity in different nonlinear problems,the network performance of traditional neural network classifiers is often limited.A typical convolution neural network in deep learning,adding multiple different types of hidden layers on the network structure,can realize the classification processing of complex problems,where the general method is to first map the input samples through convolution kernel,and the output is processed in a series of processes and cascaded with multilayer perceptrons(MLP),and then the back propagation(BP)algorithm is used to update the weights of all layers.However,the method of RBF cascading radial basis function(RBF)with BP network has not been well studied at present.For the problem of supervised classification,this paper is based on constructing optimized neural network classifiers and matching learning algorithms as the goal,to explore the selection and optimization of RBF network structure and kernel parameters,the design of RBF hybrid structure neural network classifier and the classification method of kernel global partition,which better adapts the requirement of different nonlinear classification problem.In the process of research,the following research results have been obtained:1.A RBF-BP hybrid structure neural network with pre-kernel is proposed.The RBF network structure and BP network structure are cascaded to adjust,where the output of the original RBF network hidden layer is cascaded with the hidden layer of BP network.For the improved network classifier,the RBF network is used to realize the localized kernel mapping of the original samples,and the BP network is used for nonlinear classification.In this way,the local nonlinear mapping ability of the RBF network and the global nonlinear classification ability of the BP network can be combined together.The experimental results show that the proposed network structure can improve the classification performance of the single RBF network and the BP network,respectively,and reduce the dependence on the node parameter selection of RBF network and BP network hidden,respectively.2.A RBF-BP hybrid structure neural network classifier based on kernel number adaptive learning is proposed.The spatial distribution information in each categeory of the training sample set is fully utilized.By introducing the method of potential function clustering to measure the density of different regions,the corresponding RBF hidden nodes can be established to complete the coverage of different regions,which can incrementally generate RBF hidden nodes.The experimental results show that this method can automatically estimate the number and the center of hidden nodes in RBF network in the RBF-BP hybrid structure neural network with pre-kernel,and has better classification performance.3.A RBF-BP hybrid structure neural network classifier based on heterogeneous repulsive force kernel optimization is proposed.Based on the method of potential function clustering to generate initial RBF network hidden nodes,the neighbor information of the region covered by each RBF hidden node is taken into account,and the center-oriented heterogeneous sample repulsive force model is designed to achieve the optimization of the center.In this way,the width,the number of RBF network,and the optimization learning algorithm of the whole network can be constructed.The quantitative condition of the separability enhancement when the original samples go through the kernel structure-adaptive RBF network is theoretically analyzed and proved.The experimental results show that the proposed method has good classification ability,especially when the dimension of the sample space is relatively low and the number of training samples is sufficient,the proposed method has unique advantages.4.A RBF-BP hybrid structure neural network classifier based on kernel global learning and division is proposed.The proposed method utilizes the RBF kernels as the whole training goal,and the advantages of the RBF kernel global learning and division is expounded.The internal sample generation and optimization screening mechanism of RBF kernels are designed to achieve the approximation effect of the RBF kernel global learning and division.On the basis of generating appropriate size of the sample set,the existing network classifier is used for training and classification.The experimental results show that our proposed method can effectively improve the sparsity problem of the sample space caused by the small size or too high dimension of training samples,thereby improving the robustness and generalization ability of network classifiers.5.A RBF-ELM hybrid neural network classifier with adaptive kernel structure is proposed.This method is a generalization of the kernel adaptive RBF hybrid structure neural network classifier.The heterogeneous repulsive force optimized RBF network and the ELM network are cascaded to adjust,where the heterogeneous repulsive force optimized RBF network is used for nonlinear localization kernel mapping of sample space in different regions,and the ELM network is used for nonlinear classification in the kernel space.In this way,the complementary network model can be constructed.The learning algorithm of RBF-ELM hybrid neural network classifier with adaptive kernel structure is given.The superiority of the heterogeneous repulsive force optimized RBF network is clarified,and the experiments on multiple data sets show that the proposed method can significantly improve the classification performance of ELM networks.