Theoretical foundations for intelligent tutoring systems

This paper considers the case for formalising aspects of intelligent tutoring systems in order to derive more reliable implementations, as opposed to the present use of informal theories to build experimental systems which are then studied empirically. Some recent work in theoretical AI is suggested as a possible source for the elements of a 'theory of ITS'. I n t r o d u c t i o n The engineering of any complex device (such as an ITS) gradually relies less on empirical experimentation and more on mathematical or scientific theory. As yet, there is no significant 'theory of ITS': all of the recent ITS texts (e.g. Wenger, 1987; Mandl and Lesgold, 1988; Polson and Richardson, 1988) are entirely discursive and attempt no kind of formalisation of their content. The aim of this paper is to suggest that it is not premature for ITS research to begin an attempt to complement a short-term emphasis on pragmatic aspects (Kearsley, 1989) by seeking theoretical foundations for its implementations. Most AI researchers regard ITSs as peripheral applications of AI, an understandable opinion in view of the virtual absence of ITS papers from the major AI journals and conferences. But Clancey (1986) has argued that work on ITSs is not a "mere matter of putting well-known AI methods into practice" but is (or should be) "broadening the meaning of AI research". Historically, ITS research began within AI, but AI researchers have retreated from the ITS arena as they have come to appreciate the need for more fundamental work on mental models, language understanding, knowledge representation, etc., leaving others to move into an intrinsically multi-disciplinary field. However, if there is ever to be a formal theory of (aspects of) ITS then it will be derived from elements of AI. Moreover, recent AI research begins to indicate what those elements might be.