ON ASKING THE RIGHT QUESTIONS

Abstract

Concept learning systems have a great deal to gain by showing more initiative; in particular, by actively posing test examples rather than passively waiting for more examples to appear. This paper briefly reviews different approaches to selecting examples, and goes on to explore the ramifications of one in detail. Current implementations are overly simplistic because they assume a hierarchically-structured network of concepts. It is shown how they break down in more general partially-ordered domains, and a new method which copes with this situation is described and illustrated. Finally, the techniques are related to the version-space approach to provide them with a well-understood theoretical underpinning.