B-spline recurrent neural network and its nonlinear modelling

Abstract

A new recurrent neural network based on B-spline function approximation is presented. The network can be easily trained and its training converges more quickly. Moreover, in order to overcome the slowness of training caused by constant learning rate, an adaptive weight updating algorithm used for training the recurrent network is proposed and an optimal estimate of learning rate is given. Examples are given to show that the optimal learning rate can be used to train the B-spline recurrent neural network, and this recurrent neural network can be applied to modelling of a nonlinear dynamic system.

提出一種新的基于基本樣條逼近的循環神經網絡，該網絡易于訓練且收斂速度快。此外為克服定長學習步長訓練速度慢的問題，提出一種用于該網絡訓練的自適應權值更新算法，給出了學習步長的最優估計。該最優學習步長的選擇可用于基本樣條循環神經網絡的訓練以及對非線性系統的建模。