一种双隐层RBF神经网络的算法研究

针对RBF神经网络的结构和学习算法的缺点,提出了一种双隐层RBF神经网络(DRBF)模型,并结合网络结构的动态更新策略对网络结构进行实时更新,以梯度下降法对网络参数进行修正,即确保了网络结构的最简化,提高了网络的逼近精度和泛化能力,同时也加快了网络的训练速度.将本算法和传统RBF神经网络算法应用于非线性逼近和电信企业客户流失分类进行性能比较,实验仿真结果证明了本算法的有效性和高效性.

The Study of Two Hidden Layer of RBF Neural Network Algorithm

To the disadvantage of the RBF neural network structure and learning algorithms, a two-hidden layer RBF neural network （DRBF） model was proposed. It combined with the dynamic update policies of the network structure for real-time updates to the network structure, and utilized gradient descent to correct parameters of network. This method not only ensure the network structure is most simplified, improved the approximation accuracy and the generalization ability of the network, but also accelerating the speed of network training. The algorithm is applied to non-linear approximation and the loss of classification of the communication enterprise customers, and compared with traditional RBF neural network algorithm, simulation results proved the algorithm is more effecient.