改进的径向基函数神经网络预测模型  

An Improved Prediction Model Based on Radial Basis Function Neural Network

作　　者：梁斌梅[1] 韦琳娜[1] 

机构地区：[1]广西大学数学与信息科学学院,广西南宁530004

出　　处：《计算机仿真》2009年第11期191-194,共4页Computer Simulation

基　　金：广西教育厅科研项目资助(2006026)

摘　　要：在提高网络传输性能的研究中,径向基函数神经网络(RBF网络)的基函数个数、中心及宽度的确定一直是难解决的问题,为提高RBF网络泛化能力是当前一个重要的研究问题。分析了传统RBF网络工作原理及不足,提出了改进。采用梯度下降法训练径向基函数中心和宽度,提高网络泛化性能。改进最优停止训练算法,使算法效率提高,且避免过拟合现象,最终使RBF网络获得更优的泛化能力。用改进的RBF网络对iris及wine数据集建立预测模型,进行仿真。结果表明,梯度下降方法训练出更优的基函数参数,改进的最优停止训练方法缩短了训练时间、提高预测精度,网络泛化能力有明显提高。It is difficult to determine the number, the center and the width of basis function for radial basis function neural network（RBF network） ,and at the same time,it is also currently an important research to enhance the generalization ability of RBF network. The paper analyzes the working principle and defects of the traditional RBF network , and puts forward two improving suggestions ： （ 1 ） Use the gradient descent method to train the centers and widths of RBF for enhancing the generalization ability of RBF network;（2）Modify the optimal stopping training method to make it more efficient, to avoid overfitting and bring higher generalization ability to RBF network as a result. The improved RBF network is applied to build a prediction model for the data sets of iris and wine, and the experimental results show that the gradient descent method can better train the basis function parameters, and the modified optimal stopping training method succeeds in shortening the training time, increasing the prediction accuracy, and the generalization ability of network has been improved significantly.

关 键 词：径向基函数神经网络 梯度下降法 最优停止训练法 泛化 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]