Using Recursive Control Rules in Planning

摘要

We present an approach to learning recursive control rules from universal plans. Such rules represent generalized policies to generate optimal transformation sequences (plans) of arbitrary states of a problem domain into a recursively gen-eralizable goal, such as stacking blocks in a fixed sequence. If a recursive control rule is stored together with a domain specification, all problems with a goal addressed by such a rule can be solved without search by applying the given rule. Thereby, an optimal plan is generated with the time effort for evaluating the rule, which is often only linear.