基于径向基函数的混合神经网络模型研究

Research on Mixture Neural Network Model Based on Radial Basis Function

随着系统复杂程度的增加,构造一个径向基函数神经网络(RBFNN)所需样本及训练时间都急剧增加,得到的复杂网络往往不能完全揭示问题的层次和结构。采用“分而治之”的思想,提出了一种基于RBF的混合网络模型,通过最短距离均匀聚类方法划分样本空间,构造合适的子样本集和子网络模型对网络进行训练,与采用正交最小二乘法的单独RBF网络在结构、训练时间、泛化能力上做了对比。结果表明其时间复杂度有显著降低,网络的泛化能力与精度比全局RBFNN有明显提高。整个网络模型具有良好的扩展性和应用前景,适合于大样本神经网络的建模和训练问题。

With the increasing complexity of the system, the sample and training time of constructing a radial basis function neural network （RBFNN） will increase rapidly. The obtained network model always can＇t completely post the hierarchy of the problem. Aimed at above case, the thought of ＂divide and conquer＂ was introduced, and a mixture of experts network model based on the RBF network was proposed. The model divided sample space through least distance even cluster, constructed appropriate subset and sub-network to train it, and contrasted to the single RBF network by using OLS algorithm in structure, training time and generalization. The results showed that the model decreased time complexity obviously and increased generalization and accuracy of the network compared to the global RBFNN. The network model had the good extension and the application prospect, suitd to modeling for great sample.