View article

One function of a student model in tutoring systems is to select future tasks that will best meet student needs. If the inference procedure that updates the model is inaccurate, the system may select non-optimal tasks for enhancing students’ learning. Poor selection may arise when the model assumes multiple knowledge components are required for a single correct student behavior. When the student makes an error, a deliberately simple model update procedure uniformly reduces the probability of all components even though just one may be to blame. Until now, we have had no evidence that this simple approach has any bad consequences for students. We present such evidence. We observed problem selection thrashing in analysis of log data from a tutor designed to adaptively fade (or reintroduce) instructional scaffolding based on student performance. We describe a conjunctive knowledge tracing approach, based on techniques from Bayesian networks and psychometrics, and show how it may alleviate thrashing. By applying this approach to the log data, we show that a third (441 of 1370) of the problems students were assigned may have been unnecessary.