The aim of this study is to build an intelligent authoring environment for Cognitive Tutors in which the author need not manually write a cognitive model. Writing a cognitive model usually requires days of programming and testing even for a well-trained cognitive scientist. To achieve our goal, we have built a machine learning agent -- called a Simulated Student-- that automatically generates a cognitive model from sample solutions demonstrated by the human domain expert (i.e., the author). This paper studies the effectiveness and generality of the Simulated Student. The major findings include (1) that the order of training problems does not affect a quality of the cognitive model at the end of the training session, (2) that ambiguities in the interpretation of demonstrations might hinder machine learning, and (3) that more detailed demonstration can both avoid difficulties with ambiguity and prevent search complexity from growing to impractical levels.

20 pages

*Human-Computer Interaction Institute, Carnegie Mellon University

This page maintained by reports@cs.cmu.edu