论文部分内容阅读
提出一种新的基于基本样条逼近的循环神经网络,该网络易于训练且收敛速度快。此外为克服定长学习步长训练速度慢的问题,提出一种用于该网络训练的自适应权值更新算法,给出了学习步长的最优估计。该最优学习步长的选择可用于基本样条循环神经网络的训练以及对非线性系统的建模。
A new cyclic neural network based on basic spline approximation is proposed, which is easy to train and converges fast. In addition, in order to overcome the slow training speed of fixed-length learning step, an adaptive weight updating algorithm for the network training is proposed, and an optimal estimation of the learning step length is given. The choice of the optimal learning step can be used to train the basic spline cyclic neural network and to model the nonlinear system.