Permisse Planning: a Mac to Linking Intern

Because complex real-world domains defy perfect formalization, real-world planners must be able to cope with incorrect domain knowledge. This paper offers a theoretical fhmework fmpemissiveplanning, a machine leamingmethodforimprovingthereal-world behaviorofplanners. Permissive planning aims to acquire techniques that tolerate the inevitable mismatch between the planner’s internal beliefs and the external world. Unlike the reactive approach tothis mismatch, permissive planning embraces projection. The method is both problem-iudependent and domain-independent. Unlike classical planning, permissive planning does not exclude real-world performance from the formal definition of planning. Introduction An important facet of AI planning is projection, the process by which a system anticipates attributes of a future world state from knowledge of an initial state and the intervening actions. A planner’s projection ability is often flawed. A classical planner can prove goal achievement only to be thwarted by reality. The reactive approach; which has received much attention, avoids these problems by reducing reliance on projecticm or disallowing it-altogether. For ah its stimulating effect on the field, however,+ reactivity is only one path around projection problems. It is important to continue searching for and researching alternatives. In this paper we advance one such alternative termedpemissiveplanning. In some ways it is the dual of the reactive approach, relying heavily on a goal projection ability enhanced by machine learning. From a broader perspective, permissive planning embodies an approach to inference which integrates empirical observations into a traditioual apriori domain axiomatization. The research reported in this paper was carried out at the University of Illiuois and was supported by the Office of Naval Research under grant NOOO1491-J-1563. The authors also wish to thank Renee Baillargeon, Pat Hayes, Jon Gratch and the anonymous reviewers for helpful comments. 508 DeJong Scott Bennett [email protected] Systems Research and Applications Corporation 2000 15th St. North Arlington VA 22201