Explicity Using Default Knowledge in Concept Learning: An Extended Description Logics Plus Strict and Default Rules

Abstract

This work concerns the use of default knowledge in concept learning from positive and negative examples. Two connectives are added to a description logics, C-CLASSIC, previously defined for concept learning. The new connectives (δ and ∈) allow to express the idea that some properties of a given concept definition are default properties, and that some properties that should belong to the concept definition actually do not (these are excepted properties). When performing concept learning both hypotheses and examples are expressed in this new description logics but prior to learning, a saturation process using default and non default rules has to be applied to the examples in order to add default and excepted properties to their definition. As in the original C-CLASSIC, disjunctive learning is performed using a standard greedy set covering algorithm whose generalization operator is the Least Common Subsumer operator of C-CLASSIC ^δ∈. We exemplify concept learning using default knowledge in this framework and show that explicitly expressing default knowledge may result in simpler concept definitions.