摘要
首先提出一种双曲函数型神经网络 HFL ANN,设计出一类基于 HFL ANN网络的层次双曲型函数网络HHFL ANN,给出了 HHFL AN N的网络学习算法 ,使其在用于非线性的拟合中体现了较强的优越性 ,对于任意的Volterra级数使用 HHFL ANN网络来逼近是完全可行的 ,该算法较 GMDH算法和 SOP算法 ,具有快速简单的特性 ,它优于 GMDH算法 ,有规律地选取部分多项式 ;优于 SOP算法 ,在构造 SOP网络不需要太多的中间隐层 ,从而加快了学习过程 ,提高了网络的逼近性能 。
In this paper, a new hyperboloid function link artificial neural network(HFLANN) is presented, and a kind of hierarchical HFLANN is designed, and a learning algorithm is presented. The algorithm shows stronger superiority on nonliner fitting, which can approximate a given Volterra series. The algorithm has better precision compared with GMDH algorithm and SOP algorithm. The main advantage of fitting is always based on hyperboloid function transformation. The algorithm is superior to the GMDH algorithm in randomly choosing partial ploynomials and the algorithm is superior to the SOP algorithm in that designing SOP networks needs not too many hidden layers, thus accelerating the learning process and improving the approximate quality of neural network. HHFLANN is very suitable for application domains with hierarchical structures.
出处
《计算机研究与发展》
EI
CSCD
北大核心
2001年第5期587-590,共4页
Journal of Computer Research and Development
