An Inductive Approach to Modeling Affective Reasoning in Interactive Synthetic Agents

Show full item record

Title: An Inductive Approach to Modeling Affective Reasoning in Interactive Synthetic Agents
Author: McQuiggan, Scott William
Advisors: Dr. R. Michael Young, Committee Member
Dr. Munindar P. Singh, Committee Member
Dr. James C. Lester, Committee Chair
Abstract: Recent years have witnessed significant progress on synthetic agents. With a broad range of applications in education, training, and entertainment, foundational work on synthetic agents has yielded expressive models of embodied cognition and behavior that support rich interaction. Complementing advances in cognition and behavior, affective reasoning has begun to play a central role in synthetic agents. A key challenge posed by affective reasoning in synthetic agents is devising empirically informed models of affect that enable synthetic agents to accurately respond in social situations. This thesis presents an inductive affective modeling paradigm for learning models of affect by observing human-human social interactions. First, in training sessions, one trainer directs a synthetic agent to perform a sequence of tasks while another trainer manipulates a synthetic agent's affective states to produce appropriate behaviors. These include spoken language, gestural behaviors, and posture. Second, the model generator tracks observable situational attributes pertaining to locational, intentional, and temporal information to induce a model of affect. Finally, at runtime, a synthetic agent applies the model by tracking precisely the same observable attributes and using them to drive situation-appropriate behaviors. The inductive affective reasoning framework has been implemented in a model generator that induces models of empathy for synthetic agents. A 31-subject experiment indicates that a data-driven approach can generate models of empathy that are both efficient and accurate. In the experiment, naive Bayes affective classifiers and decision tree affective classifiers were learned to model situational assessment (when to perform an affective behavior) and interpretation (which affective behavior to select). Results suggest that inductively generated models satisfy the real-time performance requirements of interactive environments and can provide the basis for empirically informed affective reasoning in synthetic agents.
Date: 2005-11-22
Degree: MS
Discipline: Computer Science

Files in this item

Files Size Format View
etd.pdf 1.739Mb PDF View/Open

This item appears in the following Collection(s)

Show full item record