基于AIC准则优化的径向神经网络微地形曲面重构

Surface Reconstruction for Micro-landform Based on RBF Neural Network Optimized by AIC Criterion

采用RBF网络模型进行复杂微地形曲面重构,建立了适应于曲面重构的RBF网络模型。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量,再根据赤池信息量准则确定最优结构的RBF神经网络模型,从而进行复杂微地形的曲面重构。实验结果表明:该方法能较好地反映原始地形;这种基于AIC准则将样本输入信息与样本输出信息同时考虑,进行RBF网络结构优化的方法,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。

Surface reconstruction for complex micro-landform is proposed using RBF network model in this paper and RBF network fit for curved surface reconstruction is accordingly established. During the course of structuring network model, different types of cluster numbers, relevant cluster centers and initial extended constants to different cluster radiuses, are all made out by the cluster method of closest neighbor. Different types are net-trained to adjust extending constants respectively, and obtain the minimum AIC values. Then based on Akaike information criterion, the RBFNN model is ensured with the optimized structure. So complex (micro-)landform curved surface can be reconstructed. The results show the original tiny terrain are preferably reproduced using this method. Based on AIC criterion, and taking into account simultaneously both input and output of the sample information, this method, which optimizes RBF net structure, offers a way to solve the key issue of ensuring concealed nodes' numbers and relating parameters of the optimized RBF network model.