Conclusion

This chapter has described work that is both related to and that directly forms the foundation for models of dialogue planning, and for the dialogue planner that will be developed in the remainder of this thesis. The main areas of emphasis were planning and plan recognition, user modelling, and dialogue management systems. In all of the theories of dialogue planning, and in most of the planning and user modelling systems described in this chapter, there has been an emphasis on planning as something that produces a set of possible plans that can achieve a goal, without comparing the utility of the different alternatives. Plans are treated as merely valid or invalid, and beliefs are modelled using logical rather than statistical models. The only exception is the systems based on reinforcement learning, where different strategies are compared according to a quantitative metric. However, while reinforcement learning can be used for many of the same problems as a dialogue planner, its approach is one of brute force, using only raw statistics to decide between strategies. On the other hand, dialogue planners seek to explain the dialogue by inferring the mental state of the speaker, inferring from the data not just that one strategy is better than another, but determining the belief state that explains the choice of strategy as well. While such a planner, which combines the best of both worlds, is desirable, there is currently no satisfactory planning model or implemented system. In the next chapter, a dialogue planning model, and a design for such a system is developed. By using game theory, and particularly by following Harsanyi's model of the Bayesian game, the planner retains the existing theory of dialogue planning, but uses the alternative plans that are generated according to this theory as the alternatives in a game. As a result, a planner is obtained that can choose efficient dialogue plans.