Motivation

While a theory of dialogue planning has developed, the use of deeply nested belief models has not filtered through to the development of dialogue planning systems. Instead, most systems employ strategies that do not change with the user's beliefs. McTear [49] reports that the popular architectures used today - finite state and frame-based, do not employ a model of the user. However, agent-based architectures, in which the system models its user as an autonomous agent with a belief state and plan rules, have not been shown to have any definite advantage over the simpler architectures, in terms of the quality of the dialogue [41], [3], [69]. The need for a system that produces measurably better dialogues than the architectures in use today is the primary motivation for this thesis.

A secondary motivation stems from the need for a system that can be easily understood by, and easily programmed by dialogue system designers. They should only be required to know something about the specification of dialogue plan rules, rather than about the use of the user model, or the mechanisms that choose dialogue strategies, or the mechanisms that understand the user's plan. These mechanisms turn out to be quite complex, and so their hiding is important. Today, the VoiceXML language [46] provides such a facility for state-based dialogue systems, whereby the designer need only go so far as to specify states and their transitions. The VoiceXML interpreter then uses this specification to control the dialogue.

A tertiary motivation is the need for a freely available design and implementation of an agent-based planner, that can serve as the basis for further research and development of dialogue management systems.