基于RBF神经网络的高斯混合近似算法

Gaussian mixture approximation algorithm based on radius basis function neural network

在分析RBF神经网络基本结构的基础上,提出一种基于RBF神经网络求解非高斯概率密度近似为高斯概率密度和的方法。该方法通过选取高斯函数作为神经网络的径向基函数,提取训练好的网络参数,运用这些参数构建混合成分的函数模型。理论分析与仿真证明,与传统采用EM近似算法相比,该算法具有求解跟初值的选取无关、能避免发散、收敛快的特点。

A algorithm based on radius basis function （RBF） neural network is presented, in which any nonlinear function can be approximated as a limited Gauss function mixture, on the basis of analysing the structure of RBF neural network. The Gauss function is selected as a radius basis function in the proposed algrithom, and the network parameters to have been trained are drawn and are used to build a mixture function. The results of theoretical analysis and simulation verify that the proposed algorithm is independent of initial values and is convergent rapidly compared with the traditional EM （expectation maximum） algorithm.