Intelligent tutoring systems: The practical implementation of constraint-based modelling
Thesis DisciplineComputer Science
Degree GrantorUniversity of Canterbury
Degree NameDoctor of Philosophy
An Intelligent Tutoring System (ITS) differs from other educational systems because it uses knowledge to guide the pedagogical process. It attempts to optimise the student's mastery of domain knowledge by controlling the introduction of new problems, concepts and instruction/feedback. Central to this process is the student model, which provides information about what the student knows. The state of the art in student modelling is model tracing, which compares student actions against an "ideal" procedure.Constraint-based modelling is a new domain and student modelling method that describes only pedagogically informative states, rather than following the procedure the student used to arrive at their answer. Ohlsson introduced the idea, which is based on learning from performance errors, but did not provide details of how it should be implemented. Even his definition of constraints is very broad. SQL-Tutor is an existing ITS that uses a constraint-based model. The representation of constraints within this system is as loose as Ohlsson's description. The constraints in SQL-Tutor are LISP code fragments, where domain structural knowledge is incorporated into the constraints via ad hoc functions. In this thesis we present a more specific representation for constraints that obviates the need for complex user-defined functions. Constraints (and their associated taxonomies and domain-specific functions) are specified as pattern matches. This new approach has two advantages: the constraints are simpler to author, and they can be used to generate solutions on demand. We have used the new representation to create algorithms for solving problems and correcting student mistakes, and for generating novel problems to present to the student. We present the details of these algorithms and the results of both laboratory and classroom evaluations. The solution generation algorithm is demonstrated in laboratory testing to be practical, and the problem generation algorithm, together with a new problem selection method, exhibits improved learning performance in the classroom. We also present the design and implementation of an authoring system for constraint-based tutors and demonstrate its efficacy in authoring tutors for two domains. One of these, a tutor for English language skills, was evaluated in an elementary school classroom. This evaluation was a success. The students enjoyed using the tutor, found the interface easy to use, and felt that they had learned a lot. An analysis of their mastery of the constraints suggested that they did indeed learn the underlying principles in the course of the session. The authoring tool enabled us to develop this system quickly using a spelling resource book as the source of both the domain taxonomy from which to produce the problems (i.e. a vocabulary of words to use) and the principles for the constraints. The authoring tool provided all other functions. This evaluation therefore showed that our authoring tool allows the rapid creation of an effective ITS.