Meta-Dialogue Behaviours: Improving the Efficiency of Human-Machine Dialogue -- A Computational Model of Variable Initiative and Negotiation in Collaborative Problem-Solving

This is a study of communication in collaborative problem-solving. We provide a model of collaboration that integrates {\bf natural language dialogue} and {\bf multi-agent planning}. Agents are assumed to be autonomous, but they only possess limited knowledge and capabilities. Although the agents may share the same common goal, they do not have direct access to their collaborators'' knowledge and plans. Agents must use natural language dialogue to request resources from other agents, provide assistance to others, coordinate problem-solving and negotiate conflicts. Kartram and Wilkins argue that the most important issues for evaluating single-agent planners are {\em soundness}, {\em completeness}, {\em optimality}, {\em efficiency} and {\em search control}. We believe these same criteria should be used in evaluating multi-agent problem-solvers. In this thesis, we identify the domain conditions and dialogue behaviors necessary for soundness and completeness within our model of collaboration. We provide an account of efficiency and search control within the model. We show that the dialogue mechanisms of {\bf variable initiative} and {\bf negotiation} vastly improve the efficiency and search control of the collaborative problem-solving between two agents. We demonstrate that the benefit of variable initiative and negotiation mechanisms cannot be realized without proper plan recognition. We show that a particular class of {\bf summarizing statements} can enhance plan recognition and thereby greatly improve the efficacy of these mechanisms. We examine and verify our model of collaboration using both analytical and experimental means. We analyze the collaborative model by presenting an explicit algorithm that each collaborating agent uses during problem-solving. We call this algorithm the {\bf Collaborative Algorithm}. Using this explicit formalism, we can determine the properties necessary for soundness, completeness and efficiency of the algorithm. The Collaborative Algorithm has been implemented, and computer--computer problem-solving experiments have been conducted in several domains. These experiments validate the conclusions of the formal analysis. Furthermore, these computer--computer problem-solving sessions provide a technique for exploring behaviors of the Collaborative Algorithm that cannot be formally analyzed.