改进的RBF网络及其参数优化方法

An Improved RBF Neural Network and Parameter Optimization Method

该文提出了一个改进的RBF网络及其参数优化方法。将典型的三层RBF网络改为一个两层RBF和一个单层感知器的串联网络。参数优化方法自动确定核函数个数,并根据核函数输出误差用BP算法修正核函数中心和宽度。根据样本分布的不规则性,引入了子类的概念,使每个类由若干子类覆盖,每个类生成一个单独的网络。实验表明,这种方法能得到较优的网络结构及其参数,并且提高了RBF网络中BP算法的收敛速度。

This paper promotes an improved RBF network and the method of its parameters optimization.The general three -layer RBF network is divided into two parts:a two -layer RBF and a single -layer perceptron(SLP).In the parameters optimization method the number of basis functions is decided automatically,and the centers and width are trained by BP algorithm based on the output of basis function.As the samples usually distributes irregularly,authors use the conception called subclass,let every class be represented by several subclasses and generate an improved RBF network.Using the iris set and letter recognition set,the method gets a quite good neural network structure and parameters,and the converge of BP algorithm is speeded up.