Beliefs about plan rules

In the current implementation of the planner, there is only one game tree, which is constructed using beliefs about plan rules from only level 1 and level 2 of the belief set. This single tree is then evaluated using a minimax-like algorithm. In future, there should be a different tree from the perspective of each agent, because each has different beliefs about the applicable plan rules. The agent at level 1 would construct a tree using the level 1 and level 2 beliefs, whereas it would expect the agent at level 2 to be using levels 2 and 3, and so on. By a series of recursive calls the evaluator can call the planner to construct each game tree in context, evaluate it, and use the resulting choice of the agent to form the best play.

For many dialogue problems, agents do not change their beliefs about plan rules, and so for the purposes of this thesis, the current implementation is adequate. Negotiation rules would never change. However, in a negotiation over a domain plan, it may happen that one of the agents, a novice, has little expertise, whereas the other, an expert, has lots, and the point of the dialogue would be for the expert to convey the rules to the novice as efficiently as possible. The planner could be used for example, to plan out a tell act where the expert tells the novice how to make a pavlova, with the domain plan appended at the end of the instructional dialogue, in much the same way as the negotiation dialogues were planned in Chapter 5. There would be a chance node in the domain plan to represent uncertainty about whether the novice knows the plan rule. The value of information of the instructional dialogue could then be calculated. The expert might also find out what the novice does and does not know by observing his domain plans and using belief revision on the plan rule beliefs, thereby triggering hints as it executes the plan. In another scenario, a negotiation may involve experts from different domains, who in their negotiation must explain why their proposals are reasonable. In both of these scenarios, the ability to generate different game trees from different perspectives is essential.