Seminar on Computational Learning and Adaptation


 
Logic for Learning: Achievements and Challenges

Stephen Muggleton
Department of Computer Science
University of York, UK
stephen@cs.york.ac.uk

In this seminar we will discuss the use of Logic, in the form of first-order predicate calculus, as a representation language for Machine Learning and describe ongoing developments in this area. Both Turing (1950) and McCarthy (1959) viewed a combination of Logic and Machine Learning as central to the development of Artificial Intelligence research. Inductive Logic Programming (ILP) employs a first-order Horn clause representation which provides not only high expressivity but also a natural representation for examples, hypotheses and background knowledge. ILP has had especially successful within prediction problems in Bioinformatics and Natural Language. The notion of "generality" in Machine Learning coincides closely with logical entailment between formulae. The semi-decidability of entailment led Plotkin (1970) to investigate subsumption as a restricted, decidable replacement. Conducting effective search is one of the central difficulties in employing highly expressive representations. Approaches to search include lattice operators (Plotkin), minimal refinement operators (Shapiro), inverse resolution (Muggleton and Buntine) and inverse entailment (Muggleton and Yamamoto). Within Inductive Logic Programming, one of the key challenges presently is the sound incorporation of probabilistic information into both 1) the inductive search and 2) the representation of examples, hypotheses and background knowledge. For 1) Bayesian priors have been used to provide effective preference biases. Stochastic logic programs (Muggleton, 1987) have been used for purpose 2). Initial approaches to inductive construction of stochastic logic programs (Muggleton, 2000) will be described.


Date: Thurs., July 6

Time: 4:15-5:30PM

Place: Gates Building, room 104


Return to the seminar schedule