基于决策树算法确定一个最小神经网络  

Determination of minimal artificial neural network based on decision tree algorithm

作　　者：刘栋[1] 冯勇[1] 赵向辉[1] 

机构地区：[1]中国科学院成都计算机应用研究所,成都610041

出　　处：《计算机应用》2010年第A01期83-85,共3页journal of Computer Applications

基　　金：国家自然科学基金资助项目(10771205);中国科学院知识创新重要方向项目(KJCX2-YW-S02)

摘　　要：神经网络隐层神经元的个数对于网络的性能有着重要的影响,通常情况下,对于一个特定问题来说,没有一个确定的方法来决定隐含层到底应该有多少个神经元,一般采用试探的方法通过多次实验来达到理想效果。在分类问题中,决策树和神经网络的结构有着一定的关联性,通过把决策树映射到神经网络结构中来确定隐层神经元的个数的方法能够有效地设计神经网络的结构,从而提高训练的效率并达到良好的分类效果。实验结果表明,该方法能够得到一个有着良好识别率的最小神经网络。方法简单有效,直观且易于操作。It is shown that multilayer networks can be used to approximate almost any function, if neurons are enough. In neural networks, people want to use the simplest network that can adequately represent the training set. Here is a common rule： Do not use a bigger network when a smaller network will work. However, no definite approach can be used to determinate the number of hidden layer neurons. In general, people determine the number through carrying out experiments many times. This paper presented a simple method for determining the number of hidden layer neurons in a neural network. This method integrated two techniques： decision tree based on maximum information gain of entropy and neural network. The depth of decision tree was used to determine the number of hidden layer neurons. This approach for determining the number of hidden layer neurons can produce a minimum scale network with good recognition rate. Two different experiments were done to show its validity.

关 键 词：决策树 人工神经网络 熵 分类 隐层神经元 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]