Learning of mixed-initiative human-computer interaction models

Mixed-initiative interaction facilitates a collaboration between intelligent agents and their users that takes advantage of their complementary capabilities by supporting each of them in taking the initiative to perform the tasks for which they are most qualified, at the appropriate time.;This thesis presents a learning-based approach to the development of mixed-initiative interaction models that govern the interaction between an end-user and a multi-agent system consisting of a collection of knowledge-based assistants specialized in helping the user perform different tasks. In general, developing such interaction models is a software engineering task of programming complex interfaces. Our approach transforms this software engineering task into a knowledge engineering one of representing the interaction models into a knowledge base. Moreover, the knowledge engineer does not need to manually define the reasoning rules that govern the user-agent interactions. Instead, the knowledge engineer teaches the agent how to interact with the user based on specific interaction scenarios from which the agent learns general interaction rules. This learning ability allows the agent to also adapt to the changing needs and preferences of the user.;At the basis of our approach is a task analysis methodology that results in the learning of executable interaction models by the mixed-initiative assistants. It includes general methods and guidelines for translating conceptual interaction plans into interaction models executable by a state-based interaction engine, the conceptualization of the interactions into an interaction ontology, and the adaptation and application of methods for learning general problem solving rules to the learning of general interaction patterns.;The task analysis methodology is supported by a general architecture for the mixed-initiative interaction of a multi-agent system. We have developed two assistants in this architecture, the Assumption Assistant and the Modeling Assistant, each with its own interaction model. The Assumption Assistant helps its user to solve problems in application domains with incomplete or uncertain information, by facilitating the definition of hypothetical solutions to the unsolved sub-problems. The Modeling Assistant helps a user to extend the partial reasoning tree generated by an agent by suggesting plausible ways to reduce unsolved problems.;The mixed-initiative interaction framework developed and its associated methods have been implemented as an extension of the Disciple learning agent shell. This shell allows subject matter expert to teach an agent how to solve problems in an application domain. Our mixed-initiative methods allow a knowledge engineer (and the expert) to teach the agent how to more efficiently interact during the problem solving process.