基于多分支RNN快速学习算法的混沌时间序列预测

Chaotic time series of fast learning algorithm of multi branch prediction based on RNN

针对传统递归神经网络中出现的网络结构与计算复杂性,提出了使用多分支递归神经网络学习算法,并将其应用到混沌时间序列预测领域。首先缩减了部分冗余的分支,只保留了节点与自身之间以及节点与代表以后时刻的节点之间的分支;然后使用规则导数代替惯用的一般偏导数,有助于同时反映权值对目标函数的直接影响和间接影响;最后使学习率根据学习情况进行动态调整,有助于加快学习算法的收敛速度。仿真实验表明,当参数的选取合理时,多分支递归神经网络能够达到较高的性能。

In view of the network structure and the computational complexity of the traditional recurrent neural network, this paper proposed the use of multi branch of recurrent neural network learning algorithm, and its application to the prediction of chaotic time series field. First it cut the branch part of redundant, branch only retained between node and node and between itself and the later time points. Secondly, it replaced general partial derivative by using the rules of derivative, helped to reflect both the direct effects of weight on the objective function and the indirect influence. Finally it made the learning rate according to the study of dynamic adjustment, helped to speed up the convergence of learning algorithm. Simulation results show that, when the reasonable selection of parameters, the performance of muhi branch recursive neural network can achieve higher.