改进的GMDH型神经网络及其在混沌预测中的应用

本文提出基于改进自组织方法的GMDH（Group Method of Data Handling）型神经网络并将它应用于混沌顶测。一般的GMDH型神经网络的自组织功能是通过给定一个准则阈值来确定或直接给定数值来实现，但GMDH型神经网络的自组织准则的阈值难以合适确定，由此提出了一种简单的自组织方法来实现真正意义上的自组织功能。这种用改进了自组织方法所构成的GMDH型神经网络可以应用于混沌时间序列预测。通过仿真实验，证明其预测效果明显比基本的GMDH型神经网络好，即改进GMDH型神经网络优于基本的GMDH型神经网络。

The Improved GMDH-type Neural Network and Its Application to Forecasting Chaotic Time Series

An improved GMDH-type neural network and its application to predicting chaotic time series are proposed. The architecture of conventional GMDH-type neural network can be optimized by means of heuristic self-organization method. In such process, the threshold of the error criterion has to be chosen by means of human experience. It is difficult because of without any rule to be followed. Improved self-organization method is proposed to automatically optimize the structure of both the neurons and GMDH network without manual intervention. The application of proposed GMDH-type neural network to the prediction of chaotic time series demonstrates that this new network has better performance than the conventional one.