Mixed-initiative systems present the challenge of finding an effective level of interaction between humans and computers. Machine learning presents a promising approach to this problem in the form of systems that automatically adapt their behavior to accommodate different users. This talk presents an empirical study of learning user models in an adaptive assistant for crisis scheduling. I will begin by describing HAZMAT, our synthetic problem domain for hazardous materials incidents, and then INCA, our computational assistant for crisis response. The results of a baseline study show some benefit from learning but leave room for improvement. I will also discuss three subsequent experiments that investigate the effects of problem reformulation on performance. The results reveal that problem reformulation leads to significantly better accuracy without sacrificing the usefulness of the learned behavior. The study also raises several interesting issues in adaptive assistance for scheduling.
This talk describes joint work with Wayne Iba, Pat Langley, and
Stephanie Sage at ISLE.
Date: Weds., June 3; Time: 4:15-5:30PM; Place: Cordura 100
Return to seminar schedule.