基于决策树的神经网络规则抽取方法

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 .

A Rule Extraction Method Based on Decision Tree

An NNtoDT algorithm is presented. NNtoDT views the problem of extracting a comprehensible hypothesis from a trained neural network as inductive learning task. In this task, the target concept is the function represented by the neural network, and the hypothesis produced by the learning algorithm is a decision tree that approximates the network. In the process of constructing tree, three optimization principles are adopted to concisely and accurately describe the knowledge that the networks have learned. The experiments demonstrate that NNtoDT is able to extract decision tree that closely approximates the hypothesis learned by the networks, provides superior predictive accuracy to tree learned directly by conventional algorithm such as C4 5, and possesses the characteristics of generality and scalability.