BP算法各种改进算法的研究及应用

The Study of Improving Algorithms of BP Network

在简述ＢＰ网结构、算法及优缺点的基础上，详细研究了ＢＰ算法的改进方法及措施，即“步长自适应”、“训练集归一化”、“导数提升”、“各层变尺度”、“训练集重组”及“隐含层节点数的确定”。本文以实例在ＢＰ网的训练时间和预报精度两方面，将上述各改进的算法与经典算法做出对比，对比结果表明，合适地选取改进算法将大大加速ＢＰ网的训练时间并有效地提高ＢＰ网的预报精度。

Considers the structure, algorithms, advantages and disadvatages of BP network briefly,and then deals with the methods of improving Ba algorithms, such as,step adapation, rearrangement of training set, normalization of training set,and determination of hidden layer modes' This paper compares the above algorithms with classic BP algorithms in training time and predication exactitude with examples. The results show that they are greatly improved with proper method of improving algorithms.