Desperately Seeking Subscripts: Towards Automated Model Parameterization

This paper addresses the laborious task of specifying parameters within a given model of student learning. For example, should the model treat the probability of forgetting a skill as a theory-determined constant? As a single empirical parameter to fit to data? As a separate parameter for each student, or for each skill? We propose a generic framework to represent and mechanize this decision process as a heuristic search through a space of alternative parameterizations. Even partial automation of this search could ease researchers’ burden of developing models by hand. To test the framework’s generality, we apply it to two modeling formalisms – a dynamic Bayes net and learning decomposition – and compare how well they model the growth of children’s oral reading fluency.