样条神经网络的权值直接确定法.pdf

31 11 Vol.31 No. 11 2009 11 Systems Engineering and Electronics Nov. 2009 : 1001506X( 2009) 1126 504 1 1, 2 1 1 1 张雨浓 , 杨逸文 , 肖秀春 , 邹阿金 , 李 巍 ( 1. 中山大学信息科学与技术学院, 广东 广州510275; 2. 中山大学软件学院, 广东 广州510275) : 根据样条逼近理论和神经网络原理构造了 种样条神经网络模型, 以 组样条基函数作为隐神经元 的激励函数依据误差回传( BP) 思想推导出该网络模型的权值修正迭代公式, 利用该公式迭代训练可得到该网 络的最优权值而对于构造的具有特定网络结构的样条神经网络, 依据伪逆思想提出了 种直接计算权值的方 法, 从而避免冗长的迭代训练过程仿真结果表明该权值直接确定法不仅能 步确定权值从而获得更快的运算 速度, 而且能达到更高的计算精度 : 样条函数; 神经网络; 权值直接确定; 伪逆 : TP 1 3 : A Weights direct determination of a spline neural network 1 1, 2 1 1 1 ZHANG Yunong , YANG Yiw en , XIAO Xiuchun , ZOU Ajin , LI W ei ( 1. S chool of Inf ormation S cience and Technology , Sun YatSen Univ . , Guangz hou5102 5, China; 2 . School of S of tware, S un YatS en Univ . , Guangzhou5102 5, China) Abstract: Based on spline approximation theory and neural networks, a spline neural network is constructed, w here the hiddenlayer neurons are activated by a group of spline functions. Based on the standard error backpropagation ( BP) method, the neuralweight s updating formula is derived. More importantly, a pseudoinverse based method is then proposed, which could directly determine the network weights without iter ative training. Computer simulation results show that the onestep weightsdetermination method could be more effective than the standard BP iterativetraining method. Keywords: spline function; neural network; onestep weigthsdetermination; pseudoinverse 0 ) , BP , ; , , [ 14] BP(