摘要
对多层前向神经网络的函数逼近能力进行了研究,讨论了用多层前向BP网络来逼近非线性函数时,输入激励信号的选择和增加隐层层数和每层神经元个数对逼近精度的影响。为了在隐层层数、每层神经元个数有限的情况下,加快网络学习速度,改善逼近效果,本文提出了利用对被逼近函数的先验知识,在隐层前加一函数层的思想,并通过仿真证明了其有效性。
The approximation ability of multilayer feedforward neural networks is studied. The problem of choosing input stimulating signals and the effect of increasing the number of hidden layers and the number of nets in each layer on the approximation precision of nonlinear functions by multilayer feed forward BP networks are discussed. Under constraints of limited hidden lapers and the number of nets in each layer the learning speed of neural networks and the approximation precsision are improved by using the prior knowledge of approxunated functions and adding one layer before the hidden layers Simulation results are given to verify the above statements.
出处
《南京航空航天大学学报》
CAS
CSCD
1994年第S1期191-195,共5页
Journal of Nanjing University of Aeronautics & Astronautics
关键词
神经网络
函数
逼近
非线性函数
函数逼近
neural networks
functions
approximation
nonlinear function
function approximation