Knowledge engineering and large problems

The problems discussed so far have had a relatively small number of plan rules. When a human partner is introduced, the planner will have to use large sets of plan rules that in turn generate larger game trees than have been seen so far. The problems of acquiring these rules and working with complex planning problems have yet to be addressed. There is no procedure for inducing the rules from example dialogues. One approach is by expert hand-crafting of the rules, another is by automatic inference from dialogue examples. It is conventional that a "tell" act has a precondition that the speaker believes what he is telling, but it may be necessary to examine many examples to make this generalisation from a lot of "tell"s. As for learning the plan structures, there needs to be some method of choosing rule sets that can generate a set of observed example dialogues. There has been work in grammar induction, and particularly under the assumption that the language is generated by a context free grammar, that might be applicable to the task of automatically finding the dialogue rules [4].