摘要
根据多项式理论,构造了一种以Laguerre正交多项式作为隐层神经元激励函数的前向神经网络模型.根据标准BP算法,导出了权值修正的迭代公式(包括标量形式和矩阵形式).区别于这种需要迭代训练获得最优权值的方法,针对该网络模型,进一步提出了一种基于伪逆的直接计算权值的方法.该权值直接确定法避免了以往的权值反复迭代训练的冗长过程.仿真结果显示其具有比传统的BP迭代法更快的计算速度,并且能够能达到更高的工作精度.
Based on network is constructed. tivated by Laguerre orth polynomial curve-fitting theory, a Laguerre orthogonal basis feed-forward neural The model adopts a three-layer structure, where the hidden-layer neurons are acogonal polynomial functions. In order to obtain optimal weights, weights -updating formula is derived firstly by standard BP training method. A pseudo-inverse based method is finally proposed, which determines the network weights without lengthy iterative training. Simulation results show that the weights-directly-determined method is more efficient and effective than conventional BP iterativetraining algorithms.
出处
《暨南大学学报（自然科学与医学版）》
CAS
CSCD
北大核心
2008年第3期249-253,共5页
Journal of Jinan University(Natural Science & Medicine Edition)
基金
国家自然科学基金(60643004)
中山大学科研启动费和后备重点课题
关键词
Laguerre正交多项式
激励函数
前向神经网络
BP迭代法
权值直接确定法
Laguerre orthogonal polynomials
activation function
feed-forward neural network
iterative-training algorithm
weights-direct-determination method
