Reasoning from Data Rather than Theory

The current framework for constructing intelligent tutoring systems (ITS) is to use psychological/pedagogical theories of learning, and encode this knowledge into the tutor. However, this approach is both expensive and not sufficiently flexible to support reasoning that some system designers would like intelligent tutors to do. Therefore, we propose using machine learning to automatically derive models of student performance. Data are gathered from students using the tutor. These data can come from either the current student or from previous users of the tutor. We have constructed a set of machine learning agents that learn how to predict "high-level" student actions, and use this knowledge to learn how to teach students to fit a particular learning goal. We discuss other related work at using machine learning to construct models within ITS. By combining several different systems, nearly all of an ITS’s decision-making could be performed by machine learner derived models. An open question is the complexity of combining these diverse architectures into a single system ̄