摘要： 针对傅里叶神经网络采用最速下降法导致局部极小、学习速度慢以及泛化能力差的问题，提出一种基于DFP校正拟牛顿法的新学习算法。该算法计算复杂度低，能保证网络具有良好的泛化能力和全局最优性。通过2个数值算例检验该算法，同时和BP神经网络以及另外2种傅里叶神经网络作比较。结果表明，该算法计算复杂度约为最速下降法的5%，为最小二乘学习算法的80%，具有较好的泛化 能力。

关键词: 傅里叶神经网络, BP神经网络, 最速下降法, 最小二乘法, 拟牛顿法, DFP校正拟牛顿法

Abstract: This paper proposes a novel Fourier Neural Network(FNN) based on DFP emendatory Quasi-Newton method in dealing with the problems of local minimum, slow learning rate and poor generalization ability of the FNN based on steepest descent method. The newly FNN has low computational complexity, good generalization ability and global optimization. Two numerical examples are utilized to validate the proposed learning algorithm by comparing with BP neural network and two kinds of FNNs. Numerical example results show that the computational complexity is 5% of the steepest descent method’s and 80% of the least squares method’s, and the new learning algorithm has good generalization capacity.

Key words: Fourier neural network, BP neural network, steepest descent method, least squares method, Quasi-Newton method, DFP emendatory Quasi-Newton method

中图分类号: 

• TP18