一种提高神经网络泛化能力的新方法  

A New Approach to Improve the Generalization Ability of Neural Network

作　　者：冯乃勤[1] 邱玉辉[1] 王芳[1] 

机构地区：[1]西南师范大学计算机与信息科学学院

出　　处：《计算机科学》2006年第2期201-204,共4页Computer Science

基　　金：河南省自然科学资助项目(0511012500);教育部科技司资助项目"基于禁忌搜索的模糊神经控制研究"。

摘　　要：提出了一种改进神经网络泛化能力的新方法——“缩放法”。这种方法通过对输入向量的缩放处理,来缩小或模糊化训练样本和新的模式之间的差别,从而使神经网络的泛化能力得以提高。文中提出的新算法——α算法,可以找到合适的缩放因子,进而得到泛化能力更强的新网络。一些实验例证了“缩放法”和α算法的有效性,并从理论上对其进行了分析和讨论。实验和分析表明,这种方法简单可靠,对许多神经网络和模式分类问题效果明显。A new approach to improve the generalization ability of neural network is presented based on an angle of fuzzy theory. This approach is put into effect through shrinking or magnifying the input vector, thereby reducing the difference between training set and testing set. It is called ＂Shrinklng-Magnifying Approach＂（SMA）. At the same time in this paper, a new algorithm--α-algorithm is presented in order to find out the appropriate shrinking factor a and obtain better generalization ability of neural network. Quite a few simulation experiments serve to study about SMA andaalgorithm. The results of experiments and analyses show that the new approach is not only simpler and easier, but also is applicable to many neural networks and many classification problems.

关 键 词：神经网络 泛化 误识率 模糊理论 

分 类 号：TP274[自动化与计算机技术—检测技术与自动化装置] TP183[自动化与计算机技术—控制科学与工程]