Exploiting the Ordering of Observed Problem-Solving Steps for Knowledge Base Refinement: An Apprenticeship Approach

Machine Learning

Volume

Proceedings of the AAAI Conference on Artificial Intelligence, 12

Control Learning

Apprenticeship is a powerful method of learning among humans whereby a student refines his knowledge simply by observing and analyzing the problem-solving steps taken by an expert. This paper focuses on knowledge base (KB) refinement for classification problems and examines how the ordering of the problem-solving steps taken by an observed expert can be used to yield leverage in KB refinement. Questions examined include: What added information can be extracted from attribute ordering? How can this added information be utilized to identify and repair KB shortcomings? What assumptions must be made about the observed expert, and how important of a role do these assumptions play? The principles explored have been implemented in the SKIPPER apprentice, and empirical results are given for the audiology domain.