一种改进PSO优化RBF神经网络的新方法

为了克服神经网络模型结构和参数难以设置的缺点,提出了一种改进粒子群优化的径向基函数(RBF)神经网络的新方法.首先将最近邻聚类用于RBF神经网络隐层中心向量的确定,同时对引入适应度值择优选取的原则对基本粒子群算法进行改进,采用改进粒子群(IMPSO)算法对最近邻聚类的聚类半径进行优化,合理的确定了RBF神经网络的隐层结构.将改进PSO优化的RBF神经网络应用于非线性函数逼近和混沌时间序列预测,经实验仿真验证.与基本粒子群(PSO)算法,收缩因子粒子群(CFA PSO)算法优化的RBF神经网络相比较,其在识别精度和收敛速度上都有了显著的提高.

A Novel Radial Basis Function Neural Network Method Based on Improved Particle Swarm Optimization

To prevent the problem that structure and parameters of neural network are hard to be tuned, a novel radial basis function (RBF)neural network method based on improvod particle swarm optimization(IMPSO) is pro-posed. In the proposed method, firefly, Nearest neighbor cluster algorithm is utilized to RBF neural network, second-ly, an improved particle swarm optimization(PSO) which synthesizes the existing models of PSO, Half of the poor fit-ness particles are updated by the other half to good fitness particles and the particles focus on the optimum space. Cluster distance factor is searched by the improved PSO for radial basis function (RBF) neural network ,and units in RBF layer is determined. The new training algorithm is used to approximate polynominal function and predict chaotic time series, compared with PSO, and CFA PSO, the algorithm improves the velocity of convergence, and has much higher accuracy.