Applying Probabilistic Models for Knowledge Diagnosis and Educational Game Design

• In Event: Promising Scholarship in Education: Dissertation Fellows and Their Research

Abstract

Computer-based learning environments oer the potential for innovative assessments of student knowl-
edge and personalized instruction for learners. However, there are a number of challenges to realizing this
potential. Many psychological models are not specic enough to directly deploy in instructional systems,
and computational challenges can arise when considering the implications of a particular theory of learning.
While learners' interactions with virtual environments encode signicant information about their understand-
ing, existing statistical tools are insucient for interpreting these interactions. This research addresses these
challenges by developing computational models of teaching and learning and combining these models with
machine learning algorithms to work through their implications. This approach results in frameworks for
interpreting learners' behavior and personalizing instruction that can be adapted to a variety of educational
domains, with the frameworks clearly separating components that can be shared across technologies and
components that are customized based on the educational content. Using this approach, my research has
address three major questions: (1) how can we diagnose learners' knowledge from their behavior in games
and virtual laboratories? (2) how can we predict whether a game will be diagnostic of learners' knowledge?
and (3) how can we customize instruction in a computer-based tutor based on a model of learning in a
domain?
The rst part of the dissertation focuses on automatically assessing student knowledge via observed be-
havior in complex interactive environments, such as virtual laboratories and games. These environments
require students to plan out behavior and take multiple actions to achieve their goals. Unlike in traditional
assessments, students' actions in these environments are not independent given their knowledge and each
individual action cannot be classied as correct or incorrect. To address this issue, I developed an inverse
planning framework for inferring learners' knowledge from their actions. The framework is a variation of
inverse reinforcement learning and uses Markov decision processes to model how people choose actions given
their knowledge. Through behavioral experiments, I show that this framework matches learners' stated be-
liefs, with accuracy similar to human observers, and that feedback based on the framework improves learning
eciency. To extend this framework to educational applications outside of the laboratory, I developed an
online algebra tutor that can detect why a student makes errors and direct them to personalized feedback,
including links to Khan Academy videos and targeted problem solving practice.

Author

• Anna Rafferty, University of California - Berkeley