前馈神经网络权值学习综合算法  

Integrated algorithm for weight learning in feed-forward neural network

作　　者：李娟[1] 徐晋[2] 付灵丽[1] 

机构地区：[1]河北工业大学计算机科学与软件学院,天津300130 [2]上海交通大学管理学院,上海200030

出　　处：《兰州理工大学学报》2004年第4期84-87,共4页Journal of Lanzhou University of Technology

基　　金：国家自然科学基金(10271025)

摘　　要：目前基于高斯牛顿法及其衍生算法的前馈神经网络虽然可以达到局部二阶收敛速度,但只对小残量或零残量问题有效,对大残量问题则收敛很慢甚至不收敛.为了实时解决神经网络学习过程中可能遇到的小残量问题和大残量问题,引入NL2SOL优化算法与GaussNewton法相结合,并引入熵误差函数,构建基于GaussNewton NL2SOL法的前馈神经网络.仿真实例表明,该神经网络较好地解决了残量问题,具有良好的收敛性和稳定性.　Although a local second-order convergence can be obtained by means of the feed-forward neural network with the algorithms of Gauss-Newton and its derivation,yet these algorithm are effective only for the problems with small or zeroth residuum;for large residual problems,the convergence is very slow and even non-convergent.In order to real-timely deal with the problems with small and large residuum,which may be encountered in the learning process of neural network,a new feed-forward neural network based on the integrated algorithm of Gauss-Newton and NL2SOL is developed by means of integrating the optimization algorithm and Gauss-Newton one as well as introducing an entropy-error function.It is shown by simulation that this kind of neural network can effectively deal with the problems with residuum,having better convergence and stability.

关 键 词：前馈神经网络 GaussNewton法 NL2SOL法 残量问题 收敛性 稳定性 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]