Plan-Based Dialogue Management in a Physics Tutor

Plan-Based Dialogue Management in a Physics Tutor Reva Freedman Learning Research and Development Center University of Pittsburgh Pittsburgh, PA 15260 freedrk+@pitt, edu http://www.pitt, edu/~freedrk Abstract This paper describes an application of APE (the Atlas Planning Engine), an integrated planning and execution system at the heart of the Atlas dialogue management system. APE controls a mixed- initiative dialogue between a human user and a host system, where turns in the 'conversation' may include graphical actions and/or written text. APE has full unification and can handle arbitrarily nested discourse constructs, making it more powerful than dialogue managers based on finite- state machines. We illustrate this work by describing Atlas-Andes, an intelligent tutoring system built using APE with the Andes physics tutor as the host. 1 Introduction The purpose of the Atlas project is to enlarge the scope of student interaction in an intelligent tutoring system (ITS) to include coherent conversational sequences, including both written text and GUI actions. A key component of Atlas is APE, the Atlas Planning Engine, a "just-in- time" planner specialized for easy construction and quick generation of hierarchically organized dialogues. APE is a domain- and task-independent system. Although to date we have used APE as a dialogue manager for intelligent tutoring systems, APE could also be used to manage other types of human-computer conversation, such as an advice- giving system or an interactive help system. Planning is an essential component of a dialogue-based ITS. Although there are many reasons for using natural language in an ITS, as soon as the student gives an unexpected response to a tutor question, the tutor needs to be able to This research was supported by NSF grant number 9720359 to CIRCLE, the Center for Interdisciplinary Research on Constructive Learning Environments at the University of Pittsburgh and Carnegie-Mellon University. plan in order to achieve its goals as well as respond appropriately to the student's statement. Yet classical planning is inappropriate for dialogue generation precisely because it assumes an unchanging world. A more appropriate approach is the "practical reason" approach pioneered by Bratman (1987, 1990). According to Bratman, human beings maintain plans and prefer to follow them, but they are also capable of changing the plans on the fly when needed. Bratman's approach has been introduced into computer science under the name of reactive planning (Georgeff and Ingrand 1989, Wilkins et al. 1995). In this paper we discuss the rationale for the use of reactive planning as well as the use of the hierarchical task network (HTN) style of plan operators. Then we describe APE (the Atlas Planning Engine), a dialogue planner we have implemented to embody the above concepts. We demonstrate the use of APE by showing how we have used it to add a dialogue capability to an existing ITS, the Andes physics tutor. By showing dialogues that Atlas-Andes can generate, we demonstrate the advantages of this architecture over the finite-state machine approach to dialogue management. 2 Integrated planning and execution for dialogue generation 2.1 'Practical reason' and the BDI model For …

Content truncated to 500 words