一种基于傅里叶变换的RBF神经网络函数逼近方法

本文提出了一种基于傅里叶变换的RBF神经网络函数逼近方法。基于聚类算法的RBF网络中心与宽度确定方法侧重于考察信号在时空的分布规律。与之相比，本文通过分析信号所含谐波分量的幅度和相位随频率分布的情况，用前有限个频率的正弦波分量的频谱特征构造RBF网络，并采用单调指数法合并隐层节点，最后用增加微调节点的方法提高网络的局部逼近精度。一个应用实例表明，本文方法具有良好的函数逼近能力。

A RBF Neural Network Based on Fourier Transformation for Function Approximation

In this paper we present a new model of the RBF neural network based on Fourier transformation to deal with function approximation problems. Compared with the traditional clustering algorithms,which lay emphasis upon the sample distribution rules in the time space, the proposed algorithm focuses particularly on the distribution law of amplitude and the phase of the partial sine waves in the spectrum domain. Firstly, we construct an initial model with the spectrum characteristics of the first several sine waves to approximate the complicated signal. Secondly,we adopt the monotonous index to prune the hidden neurons of the RBF neural network Finally, we use incremental tiny regulative neurons to improve the local approximation precision. A test result shows that the presented method is quite effective for solving function approximation problems.