基于AIC准则的RBF神经网络在GPS高程拟合中的应用  

RBF neural network application in GPS height fitting based on AIC criterion

作　　者：任超[1] 吴伟[1,2] 黄征凯[1] 焦元元[1] 

机构地区：[1]桂林理工大学测绘地理信息学院/广西空间信息与测绘重点实验室,广西桂林541004 [2]中国水电顾问集团贵阳勘测设计研究院,贵阳550004

出　　处：《测绘科学》2013年第2期77-79,共3页Science of Surveying and Mapping

基　　金：国家自然科学基金项目(41071294);广西自然科学基金资助项目(桂科自0640178);广西科学基金资助项目(桂科基0991023)

摘　　要：本文采用AIC准则优化RBF神经网络参数的方法进行GPS高程拟合,在建立网络模型过程中,对不同的聚类半径由最近邻聚类法求出不同类别的聚类数目及相应的聚类中心和初始扩展常数,通过对不同类别分别进行调整扩展常数的网络训练,求出其最小AIC值,再根据AIC准则确定结构最优的RBF网络模型。实验结果表明:这种方法为确定最优RBF网络模型的隐节点数目及相应参数提供了途径;拟合精度较高,在较平坦测区可以替代三等水准测量。In this paper, it used the AIC criterion to optimize the parameters of RBF neural network for the GPS height fitting. In the process of establishing the model, different types of cluster numbers, relevant cluster centers and initial extended constants to dif- ferent cluster radiuses, were all made out by the cluster method of closest neighbor. Different types were net-trained to adjust extending constants respectively, in order to obtain the minimum AIC values, then the optimal network modal based on AIC was established. The experimental results showed that this way could offer a way to solve the key issue of ensuring concealed nodes numbers and relating pa- rameters, and it could replace the third level measurement in a flat area with high accuracy.

关 键 词：赤池信息量准则(AIC) RBF神经网络 GPS高程拟合 拟合精度 

分 类 号：P228.41[天文地球—大地测量学与测量工程]