Towards Transferrable Affective Models for Educational Play

  • Samuel Spaulding Massachusetts Institute of Technology

Abstract

Modern computational agents in adaptive educational systems primarily rely on cognitive (i.e. curricular performance) data, while ignoring important multimodal affect cues which human tutors use to personalize their interactions with students. Students’ affective responses are highly idiosyncratic, noisy, and dependent on interactive context, challenges which defy many standard assumptions of computational player modeling. As a result, recent research efforts to model student affective response have focused on specific, single-task interactions, limiting the amount and variety of affective input from an individual player. For my thesis research, I plan to address these limitations in two ways. First, by developing a new paradigm for modeling student affective data, not as a scalar reward signal, but as a policy label, i.e., feedback on an agent’s recent behavior, and additionally by developing transfer learning methods to apply this policy feedback data across multiple game tasks. Together, these two advances may lead to more data-efficient learning and more flexible and generalizable affective models of players

Published
2020-10-01