文章基本信息

标题：Cumulative Scoring-Based Induction of Default Theories
本地全文：下载
作者：Farhad Shakerin ; Gopal Gupta
期刊名称：OASIcs : OpenAccess Series in Informatics
电子版ISSN：2190-6807
出版年度：2018
卷号：64
页码：1-15
DOI：10.4230/OASIcs.ICLP.2018.2
出版社：Schloss Dagstuhl -- Leibniz-Zentrum fuer Informatik
摘要：Significant research has been conducted in recent years to extend Inductive Logic Programming (ILP) methods to induce a more expressive class of logic programs such as answer set programs. The methods proposed perform an exhaustive search for the correct hypothesis. Thus, they are sound but not scalable to real-life datasets. Lack of scalability and inability to deal with noisy data in real-life datasets restricts their applicability. In contrast, top-down ILP algorithms such as FOIL, can easily guide the search using heuristics and tolerate noise. They also scale up very well, due to the greedy nature of search for best hypothesis. However, in some cases despite having ample positive and negative examples, heuristics fail to direct the search in the correct direction. In this paper, we introduce the FOLD 2.0 algorithm - an enhanced version of our recently developed algorithm called FOLD. Our original FOLD algorithm automates the inductive learning of default theories. The enhancements presented here preserve the greedy nature of hypothesis search during clause specialization. These enhancements also avoid being stuck in local optima - a major pitfall of FOIL-like algorithms. Experiments that we report in this paper, suggest a significant improvement in terms of accuracy and expressiveness of the class of induced hypotheses. To the best of our knowledge, our FOLD 2.0 algorithm is the first heuristic based, scalable, and noise-resilient ILP system to induce answer set programs.
关键词：Inductive Logic Programming; Negation As Failure; Answer Set Programming; Default reasoning; Machine learning