A student model for an intelligent tutoring system helping novices learn object-oriented design

Learning object-oriented design is a challenging task for many beginners. Intelligent Tutoring Systems (ITSs) can aid learners with complex problem-solving. Generally, an ITS contains three main components: an expert evaluator that observes a student's work and identifies errors in his/her solution; a student model that diagnoses the gap in the student's knowledge; and a pedagogical advisor that provides feedback to the student. Existing student models have several common problems: (1) they only consider rule-based behaviors or concepts as students' learning goals, while students oftentimes are confused about the relationship among concepts such as prerequisite, transition, similarity and distinction; (2) they do not represent layered student knowledge, taking into account domain, reasoning, monitoring and reflective knowledge; (3) they often use Bayesian networks requiring exponential time, and hence cannot provide consistent support for real time communicative tutoring; and (4) they rarely simulate students' knowledge history. This dissertation develops a student model applying Atomic Dynamic Bayesian Networks (ADBNs), which consists two connected Atomic Bayesian Networks (ABNs). It advances the state of the art of student models by: (1) representing concepts and important relationships, such as prerequisites and distinction; (2) tracking a history of student learning; and (3) integrating a student model from both open- and close-ended work. This student model does all these things in real time, so that the ITS can be responsive to students as they work on an assigned problem. We evaluated the ABN- and ADBN-based student models with 240 simulated students and 71 human subjects. The evaluation investigates the student models' behavior for different types of students and different slip and guess values. Holding slip and guess to equal and small values, ADBNs are able to produce accurate diagnostic rates for student knowledge states over students' learning history. The results also show that student models with ADBNs perform better than student models with ABNs only. We evaluated the student model which integrates the diagnosed students' learning state from closed- and open-ended exercises with the 71 human subjects. The results show that integrating diagnoses from closed- and open-ended exercises is an effective way to increase the accuracy of student models. In addition, we compared the accuracy of non-advanced-numerical student models with the student models using ADBNs. The results show that student models using ADBNs perform much better than the non-advanced-numerical student models.