摘要
提出一种新型的广义RBF神经网络模型,将径向基输出权值改为权函数,采用高次函数取代线性加权。给出网络学习方法,并通过仿真分析研究隐单元宽度、权函数幂次等参数的选取对网络逼近精度以及训练时间的影响。结果表明,和传统的RBF神经网络相比,该网络具有良好的逼近能力和较快的计算速度,在系统辨识和控制中具有广阔的应用前景。
A new type of general RBF neural network model, which replaces weight of outer layer with weight function, i.e. replaces linear weight with high order function, is proposed. Network training method is brought forward. Parameters selection such as hide layer width and power of weight function, which have effect on approximation precision and training time of the network, are investigated through simulation. The results indicate the general RBF network has better approximation ability and faster calculation speed than traditional RBF neural network , which promotes a good prospect in the field of system identification and control.
出处
《计算技术与自动化》
2007年第1期9-13,共5页
Computing Technology and Automation
关键词
RBF神经网络
训练方法
函数逼近
RBF neural network
training method: function approximation