Conclusion

The description of the planner given in this chapter has been of an informal nature, and particularly, there has been no formal design specification from which to derive an implementation for the planner. This is mitigated by the choice of a high-level programming language, Prolog, for the implementation of the planner. For example, the Prolog program specifies the minimax and evaluation functions using recursive rules that are as useful for a design specification as they are as an implementation. There is a guide to the implementation given in appendix A.

Some functional requirements were set out at the beginning of this chapter. One was that the planner should be based on current theories of dialogue planning, and in basing the plan recognition and focussed forward planning of the system the theories of Carberry [10] and Pollack [53], this has been achieved. Another requirement was that the system should be easy to use. This ease of use stems from the ease with which the starting mental state of the system can be given by the designer. He needs only to give the following: a utility function to evaluate plan trees, and initial beliefs about plan rules. Once this is done, the choice of strategy, management of the dialogue, and initialisation and maintenance of the user model through belief revision is automatic. Perhaps the most important functional requirement was that the system offer some utility gain over a system that cannot use a user model at all, and over a system that uses the traditional logical, rather than a probabilistic belief model. There must be good reason to depart from the well accepted logical model, and so this final requirement is the main topic of the next chapter, in which two examples are used to establish that such a utility gain can be obtained. These examples will also illustrate the operation of the algorithms that have been described in this chapter.