【作者】 赵明华；

【导师】 王熙照；

【作者基本信息】 河北大学 ， 计算机应用， 2003， 硕士

【摘要】 基于ID3算法的决策树归纳学习是归纳学习的一个重要分支，可用于知识的自动获取过程。随着归纳学习研究的深入，具有精确描述特征的示例学习已不能适应一个系统中不精确知识自动获取的要求，研究不确定环境中的示例学习已非常必要，进而产生了传统ID3算法的模糊推广——模糊ID3算法。在模糊决策树的产生过程中，用模糊熵选择的扩展属性不能像经典决策树那样将类清晰的分开，而是属性术语所覆盖的例子之间有一定的重叠，因此树的整个产生过程在给定的显著性水平α的基础上进行，参数α的引入能在一定程度上减少这种重叠，从而减少分类的不确定性，提高模糊决策树的分类结果。而它一般由领域专家根据经验或需要直接给出，这种人为的参与过分依赖于专家知识，从而可能使实际分类结果在规则数、准确率方面达不到最优。 本文在Visual C++软件开发平台及模糊ID3算法的基础上，从解析的角度出发，通过分析参数α与模糊熵之间的函数关系式，讨论了随着α的增加，模糊熵函数的变化趋势，进一步分析了参数α对模糊决策树的分类结果在训练准确率、测试准确率、规则数等方面所表现出的敏感性，探讨了得到最优参数α的实验方法。实验证明，利用这一方法得到的最优参数α的值，可以使模糊决策树的分类结果达到最好的效果，从而为人们用模糊决策树进行分类时选取参数以获得最优的分类结果，提供了良好的理论依据。更多还原

【Abstract】 Induction learning of decision tree based on ID3 algorithm is an important branch of inductive learning now, which can be used to automatic acquisition of knowledge. With the deeper research of inductive learning, it can’t meet the automatic acquisition of non-crisp knowledge because of its crisp description. It appears to be very important to research inductive learning in uncertainty condition and therefore the fuzzy extension of traditional ID3-fuzzy ID3 is proposed. In building fuzzy decision tree, each expanded attribute can’t classify the class label clearly like decision tree, but the cases covered with the attribute-values have some overlap. So the entire process of building trees is based on a significant level a, the import of a can reduce such overlap in some degree, decrease the uncertainty of classification and improve classification result. But the value of a is given directly by domain expert based on experience or requirement, which depend on expert’s knowledge excessively, therefore do not gain the best classification result possibly.By analyzing expression between a and fuzzy entropy from the view of analytics, this paper analyses the relationship of between a and fuzzy entropy and the changing trend of fuzzy entropy function with the increase of a, then discusses the sensitivity of the parameter a to classification result such as total nodes, rule number, classification accuracy of fuzzy decision tree, proposes an experimental method of obtaining optimal a , It is proved by experiment that the optimal value a obtained by this method can make the classification result of fuzzy decision tree best, and therefore provides the academic evidence of selecting parameter a in order to gain the best classification result.更多还原