前馈式神经网络的最小二乘学习算法  

Least square learning algorithm for feed-forward neural network

作　　者：谭松波[1] 冯英浚[1] 王雪峰[1] 

机构地区：[1]哈尔滨工业大学数学系,黑龙江哈尔滨150001

出　　处：《哈尔滨工业大学学报》2003年第4期403-404,409,共3页Journal of Harbin Institute of Technology

基　　金：国家自然科学基金资助项目 (699740 13 )

摘　　要：通过对Sigmoid函数求逆 ,把非线性极值问题转化为线性方程组来处理 ,巧妙地避开了梯度 ,从而可以克服BP算法的一些缺点 ,提高了算法的收敛速度 .同时采用最小二乘法来求解方程组 ,进一步提高了收敛速度 .算法的计算过程为每次处理一个节点的所有前一层连接权 ,轮换处理 ,直到收敛到最小点 .The nonlinear extremum question is changed into linear equations by computing the inverse of the Sigmoid function. So the algorithm avoids the gradient method. As a result, some drawbacks of the BP algorithm are overcome and the convergence speed of the algorithm is improved while the least square algorithm is employed to solve the equations. The procedure of the algorithm is that the computation does not stop until the minimum is found by dealing with the whole prior connection weights of one node in turn.

关 键 词：前馈式神经网络 最小二乘学习算法 BP算法 收敛速度 LS算法 SIGMOID函数 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]