摘要
使用BP算法训练多层网络的速度很慢而且事先难于确定隐节点和隐层的适当数目。本文提出一个有效的算法,先构造决策树,然后将构造的决策树转换为神经网。文中使用一个全局准则函数控制决策树的增长。它较好地匹配了树的复杂性和训练样本量及错分率界。实验结果表明,本文的算法比用BP算法训练多层网络要快,而其分类精度不低于用BP算法训练的多层神经网。
Training a multilayer neural net by BP algorithm is slow and it is difficult to choose the number of hidden units and layers in advance. This paper proposes an efficient algorithm for constructing decision tree and then mapping it to neural net. A global criterion function is used to control the growth of the decision tree. It well matches the tree complexity to the training data and the misclassi-fication rate bound. It is showed that the proposed algorithm is much faster than BP algorithm and achieves accuracy as good as the nets trained using BP algorithm.
出处
《计算机应用与软件》
CSCD
1994年第5期15-19,53,共6页
Computer Applications and Software
基金
国家自然科学基金
关键词
神经网络
分类器
决策树
算法
Neural network, BP algorithm, decision tree.