Refining Complete Hypotheses in ILP
Most ILP systems employ the covering algorithm whereby hypotheses are constructed iteratively clause by clause. Typically the covering algorithm is greedy in the sense that each iteration adds the best clause according to some local evaluation criterion. Some typical problems of the covering algorithm are: unnecessarily long hypotheses, difficulties in handling recursion, difficulties in learning multiple predicates. This paper investigates a non-covering approach to ILP, implemented as a Prolog program called HYPER, whose goals were: use intensional background knowledge, handle recursion well, and enable multi-predicate learning. Experimental results in this paper may appear surprising in the view of the very high combinatorial complexity of the search space associated with the non-covering approach.
KeywordsLearning Problem Covering Algorithm Prolog Program Infinite Loop Multiple Predicate
Unable to display preview. Download preview PDF.
- 1.Bratko, I., Grobelnik, M. (1993) Inductive learning applied to program construction and verification. Proc. ILP Workshop 93, Bled, Slovenia.Google Scholar
- 2.Grobelnik, M. (1992) Markus-an optimised model inference system. Proc. ECAI Workshop on Logical Approaches to Machine Learning, Vienna.Google Scholar
- 4.Pompe, U. (1998) Constraint Inductive Logic Programming. Ph.D. Thesis, University of Ljubljana, Faculty of Computer and Info. Sc. (In Slovenian).Google Scholar
- 5.Quinlan, J.R. (1990) Learning logical definitions from relations. Machine Learning, 5 (1990), 239–266.Google Scholar