Introduction

In this chapter, some open questions arising from the work presented in this thesis are discussed. These are questions that are worth exploring in the future, but are not thus far adequately addressed. Instead, sketches will be given of ways in which they may be addressed will be provided that may form the basis of future work. Two directions of importance are related to the evaluation of the planner. A comparison should be made with reinforcement learning techniques in order that measurements are made of the efficiency gain of the planner over a reinforcement learning system, given the same amount of training material. The planner must also be shown to efficient in a human setting, where bounded rationality threatens the expectations that the planner has about the outcome of the dialogue, in assuming an ideal counterpart.

Section 6.2 describes experiments in comparing the planner with a planner based on reinforcement learning. It also discusses the possible limitations of the planner with respect to dialogues with human beings, the shortcomings of the simulation approach to evaluation, and how the planner might be further developed once experimentation with human subjects is used. Section 6.3 discusses ways in which the planner's design can be developed to encompass a more general planning model, or to improve its efficiency. Section 6.4 looks at some example dialogues problems that were not covered in the thesis, but which could be tackled by the planner. Section 6.5 looks at issues in negotiation dialogues. Section 6.6 shows how to integrate the planner with a statistical speech recognition system.