RBF神经网络的混合结构优化算法  

Hybrid structure optimization algorithm of radial basis function neural network

作　　者：赵志刚[1] 缪凯[1] 吕慧显[2] 

机构地区：[1]青岛大学信息工程学院,青岛266071 [2]青岛大学自动化工程学院,青岛266071

出　　处：《仪器仪表学报》2007年第4期650-656,共7页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(60602040)资助项目

摘　　要：本文改进了递归正交最小二乘(ROLS)算法的停止条件,并用改进的ROLS算法优选RBF神经网络中隐单元的个数;借用细菌群体趋药性(bacterial colony chemotaxis,BCC)算法的思想来确定RBF神经网络隐层神经元的控制参数;把改进的ROLS算法和BCC算法相结合(ROLS-BCC算法),来全结构优化RBF神经网络,使RBF网络不但可以得到合适的结构,同时也可以得到合适的控制参数。实验结果表明,本文算法的优化效率高,优化后RBFNN的结构小,并且提高了RBFNN的泛化能力。In this paper, the stop condition for recursion orthogonal least square （ROLS） algorithm is improved, and the optimal number of hidden neurons in RBFNN is chosen using this improved ROLS algorithm. The thought of bacterial colony chemotaxis （BCC） algorithm is applied to determine the control parameters of the hidden neurons in RBF neural network. The improved ROLS algorithm is combined with the BCC algorithm, i.e. ROLS-BCC algorithm, so as to optimize the overall structure of RBFNN （ including obtaining the appropriate structure and the appropriate control parameters）. The experimental results show that the ROLS-BCC algorithm has a good optimization performance. Furthermore, the RBFNN optimized by ROLS-BCC algorithm has a smaller structure, and the generalization ability of RBFNN is improved.

关 键 词：RBF神经网络 细菌群体趋药性算法 ROLS算法 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]