基于决策树的神经网络规则抽取方法

• 作 者: 侯广坤 张劲峰
• 机 构: 中山大学计算机科学系!广东广州510275,中山大学计算机科学系!广东广州510275
• 出 处: 中山大学学报(自然科学版)
• 发表日期: 2000年04期
• 关键词: 基于决策树的神经网络 规则抽取

将从神经网络中抽取一个可理解的模型视为一个归纳学习任务 ,其中 ,目标概念就是神经网络表达的功能 ,所生成的可理解模型是一个能很好近似神经网络的决策树 .在这个过程中 ,应用了决策树归纳学习的优化原则 ,使得生成的决策树能最简洁、准确地描述神经网络学到的知识 .实验证明 ,生成的决策树可以很好地近似神经网络 ,且比用传统方法生成的决策树具有更好的分类精度 ,同时NNtoDT算法也保持了具有较好的通用性和可扩充性的特性 .

Taking a comprehensible model from neural network as a inductive learning task, the target concept is the function of neural network expression, and the comprehensible model generated is a decision tree which can approximate the neural network well.In this process , The optimization principle of decision tree inductive learning is applied to make the generated decision tree can describe the knowledge learned by neural network in the most concise and accurate way.Experiments show that the generated decision tree can approximate the neural network well, The decision tree generated by the method has better classification accuracy, and the NNtoDT algorithm also retains the characteristics of good universality and scalability.