The Apprentice Learner architecture: Closing the loop between learning theory and educational data

While Educational Data Mining research has traditionally emphasized the practical aspects of learner modeling, such as predictive modeling, estimating students knowledge, and informing adaptive instruction, in the current study, we argue that Educational Data Mining can also be used to test and improve our fundamental theories of human learning. Using the Apprentice Learner architecture, a computational theory of learning capable of simulating human behavior in interactive learning environments, we generate two models that embody alternative theories of human learning: (1) that humans perfectly recall previous training during learning and (2) that humans only recall a limited window of experience. We evaluate which of these models is better supported by data from two fractions tutoring systems. In general, we find that the model with a complete memory better fits the data than a model recalling only the previous training experience (data-drive theory development). Additionally, we demonstrate that both models are able to predict student performances, as well as, reproduce the main effects of an experimental paradigm without being trained on student data (theory-driven prediction). These results demonstrate how the Apprentice Learner architecture can be used to close the loop between learning theory and educational data.