Joseph Rollinson

Senior Thesis

Paper Poster Presentation

Abstract

Intelligent tutoring systems provide their students with an adaptive personalized learning experience. To do so, intelligent tutoring systems attempt to capture the state of their students through a student model. Student models have two primary uses: prediction of future student performance and instructional decision making. Since prediction performance is easier to quantify, student models are frequently judged by their predictive power. This has bred student models that are very powerful predictors, but cannot be easily used in decision making. In this work, we leverage these powerful predictors using novel decision algorithms that are compatible with almost any predictive student model. In particular we consider two decision problems: when to stop providing questions to the student and which skill to practice next. Our results suggest that our when-to-stop decision algorithm acts similarly on existing decision algorithms with the added benefit of stopping when students are unable to progress given the current material. Our preliminary work on deciding between skills suggests that logistic regression models, previously only used for prediction, can be used to pick between skills and even learn a skill hierarchy.