Learning to Predict User Operations for Adaptive Scheduling

Abstract

Mixed-initiative systems present the challenge of find-ing an effective level of interaction between humans and computers. Machine learning presents a promis-ing approach to this problem in the form of systems that automatically adapt their behavior to accommo-date different users. In this paper, we present an em-pirical study of learning user models in an adaptive assistant for crisis scheduling. We describe the prob-lem domain and the scheduling assistant, then present an initial formulation of the adaptive assistant’s learn-ing task and the results of a baseline study. After this, we report the results of three subsequent experiments that investigate the effects of problem reformulation and representation augmentation. The results suggest that problem reformulation leads to significantly bet-ter accuracy without sacrificing the usefulness of the learned behavior. The studies also raise several inter-esting issues in adaptive assistance for scheduling.