一种基于信息熵的神经网络规则提取方法

• 作 者: 袁红春 熊范纶
• 机 构: 中国科学技术大学自动化系,中国科学院合肥智能机械研究所,
• 出 处: 模式识别与人工智能
• 发表日期: 2002年02期
• 关键词: 神经网络 信息熵 决策树 规则提取

从训练后的神经网络中提取规则已成为当前研充热点.已有的网络规则提取方法常需网络修剪和再训练过程,因而计算成本较高.本文提出一种基于信息熵的神经网络规则提取方法.它在网络无需重复训练的情况下能够从训练过的神经网络中快速提取规则.其算法主要有四个过程组成:网络训练、决策树构建和相关隐单元识别、相关输入连接的识别及规则产生.文章以异或问题和棉花病害诊断规则提取为例进行实验,结果表明,基于信息熵的神经网络规则提取方法是有效可行的.

Extracting rules from the trained neural network has become the hot spot of research currently.Existing network rule extraction methods usually require network pruning and retraining processes, so the computational cost is higher.This paper proposes a neural network rule extraction based on information entropy Method.It can quickly extract rules from the trained neural network without repeated training.It has four main processes: network training, decision tree construction and related hidden unit identification, identification of related input connections and Rules are taken as examples.Experimental results of XOR problems and diagnosis rules of cotton diseases show that the method of neural network rule extraction based on information entropy is effective and feasible.