Learning Systems of First-Order Rules

Abstract

Similarity-based inductive learning has mainly been concerned with learning concepts from examples and counter-examples. When applied to knowledge base construction, these techniques suffer from a basic limitation: a knowledge base is not a simple juxtaposition of such concepts, but rather a coherent system of rules interacting with one another. These techniques are incapable of learning such systems of rules because they simply ignore the context in which the rules are induced. This paper introduces a new algorithm for learning such rule systems; it allows the representation of these in a powerful subset of first-order predicate logic, thus going beyond the capabilities of systems that learn decision trees or production rules. The approach is kept formal and independent of any application.