An Improved RBF Network On-Line Learning Algorithm
Zhang Xiao Ming, Ning Guang Liang · 2009
RBF neural network off-line learning algorithm can only be trained by a set of samples at the same time, not by individual samples one by one, therefore its adaptiveness is poor. This paper presents an improved RBF network on-line learning algorithm based on resource allocation network. The network can be trained by individual samples respectively. At first a structure of RBF neural network with no hidden layer is created, then this network is trained using individual samples, according to current errors between output and input, the numbers and locations of the hidden layer nodes are dynamically added or deleted to optimize the structure of network. The strategy of increasing or decreasing nodes on-line is based on the recursive least squares algorithm. For the samples producing larger output response, the weights of neurons are retained otherwise deleted to improve RBF neural network convergence speed and real time. Nonlinear curve fitting tests including natural index, multiplication, trigonometric functions and gas content forecast are carried out by VC++ simulation software to verify the validity of this algorithm. Simulation tests show that the improved on-line learning algorithm for RBF network has higher forecasting accuracy, well generalization ability, fewer hidden nodes. It can be realized in embedded systems and has a good value in many application fields such as gas content forecast.