论文部分内容阅读
时变神经网络结构可简单地取为常规神经网络连接形式,但连接权却是时变的.如何确定时变权是应用时变神经网络时的难题.迭代学习方法是一种合理的选择,它不同于将时变连接权展成Taylor级数,通过训练多项式系数的处理方法.而且,后者的处理方式不可避免地存在截断误差.对于有限区间连续时变非线性系统的神经网络建模与辨识,借助于重复运行过程,以迭代学习算法调整权值,进行网络训练.不计逼近误差,提出的学习算法能够使得辨识误差在整个区间上渐近收敛于零.为处理非零但有界的逼近误差,采用带死区的迭代学习算法.逼近误差界值已知时,文中证明带死区修正的迭代学习算法使得辨识误差在整个区间上渐近收敛于由死区界定的邻域内.对于逼近误差界值未知的情形也进行了讨论.
The time-varying neural network structure can simply be taken as a conventional neural network connection, but the connection right is time-varying. How to determine the time-varying weight is a difficult problem when applying the time-varying neural network.Iterative learning method is a reasonable choice, It is different from the method of expanding the time-varying connection weights into Taylor series by training the polynomial coefficients, and the truncation error inevitably exists in the latter processing method.For the neural network modeling of finite interval continuous-time nonlinear systems And recognition, iterative learning algorithm is used to adjust the weights by means of repeated running process, and the network training is carried out.Without the error of approximation, the proposed learning algorithm can make the identification error converge to zero asynchronously over the entire interval.In order to deal with non-zero but boundless , An iterative learning algorithm with dead zone is adopted.When the bound of the error is known, the iterative learning algorithm with dead zone correction is proved to make the identification error converge asymptotically within the neighborhood defined by the dead zone over the entire interval. The case of unknown bounds of the approximation error is also discussed.