例外を有する一般規則を学習する帰納推論システムの研究

在归纳逻辑程序设计（ILP）的大多数以前的工作中，明确的霍恩程序或经典子句程序被认为是学习逻辑程序的形式。然而，对人工智能中知识表示的研究，特别是对非单调推理的研究表明，这种单调程序不足以表示我们的常识知识。为了学习分类学层次结构中的默认规则或概念，需要一种处理非单调推理的学习机制。我们提出了一个学习系统LELP，学习扩展逻辑程序（ELP）。ELP允许两种否定，并且可以表示不完整的知识。LELP可以学习ELP形式的例外的默认规则，给定不完整的正面和负面的例子和背景知识。在LELP中，也可以通过递归调用异常识别算法来学习分层默认值。此外，当一些实例可能被分为积极和消极的时候，不确定规则也可以被学习。本文的主要工作如下：1.证明了LELP的正确性，并分析了其性质; 2.实现了LELP的多种版本，分别采用自顶向下和自底向上两种算法生成一般规则。为了提高系统的运行速度，我们还在Java中重新实现了LELP，并将遗传算法应用到LELP中，自动生成搜索偏差。3.我们考虑了LELP的一个扩展，它不仅可以学习ELP，还可以学习溯因逻辑程序（ALP）。在ALP中发现新的可溯项的新技术也可以应用于非单调推理中学习偏好知识的方法。

In most previous work on Inductive Logic Programming (ILP), definite Horn programs or classical clausal programs are considered in the form of learned logic programs. However, research on knowledge representation in AI, in particular work on nonmonotonic reasoning, has shown that such monotonic programs are not adequate to represent our commonsense knowledge. To learn default rules or concepts in taxonomic hierarchy, a learning mechanism that deals with nonmonotonic reasoning is necessary. We propose a learning system LELP that learns Extended Logic Programs (ELPs). An ELP allows two kinds of negation, and can represent incomplete knowledge. LELP can learn default rules with exceptions in the form of ELPs, given incomplete positive and negative examples and background knowledge. In LELP, hierarchical defaults can also be learned by recursively calling the exception identification algorithm. Moreover, when some instances are possibly classified as both positive and negative, nondeterministic rules can also be learned.In this research, we further developed methodologies to learn ELPs. The contributions of the research can be summarized as follows.1.We proved the correctness of LELP, and analyzed its properties.2.We implemented various versions of LELP, by applying both top-down and bottom-up algorithms for producing general rules. To speed up the system, we also reimplemented LELP in Java, applied Genetic Algorithm to LELP, and automatically generated the search bias.3.We considered an extension of LELP which can learn not only ELPs but also Abductive Logic Programs (ALPs). The new technique to discover new abducibles in ALPs can also be applied to a method to learn preference knowledge in nonmonotonic reasoning.