Reasoning about interaction in mixed-initiative artificial intelligence systems

Abstract

This thesis presents computational models for the design of mixed-initiative artificial intelligence systems that can make rational decisions about interaction with potentially helpful users. Mixed-initiative systems are ones in which either the system or the user can take the initiative to direct the dialogue or the problem solving. These systems have been designed for such diverse applications as robotics, military planning, intelligent tutoring and trip scheduling. One challenge in designing these systems is to specify when the system should take the initiative to interact with the user. The main contribution of the thesis is to provide designers of mixed-initiative systems with a systematic approach for constructing systems that can reason in a principled way about interaction with the user, regardless of the area of application. Our approach is to model the user, the task and the dialogue simultaneously. Specific factors are proposed that must be modeled, and methods are developed for how to combine these factors in order to make rational decisions about interaction, based on whether the perceived benefits of communication exceed the expected costs. Some examples and experiments are described, to demonstrate the value of the models and to justify decisions that were made in determining the role of each factor in the computational models. In particular, we emphasize the value of making decisions about interaction based on a careful evaluation of the needs, preferences and abilities of the user, leading to mixed-initiative systems that are user-specific and therefore result in greater overall user satisfaction.