Inner and Outer Boundaries of Literals A Mechanism for Computing Domain Specific Information

A number of works have shown that planning can be speeded up, often very significantly, by utilizing extra domain knowledge [KM81, BK00, BK96, KS98, SK98, DK99, Rei99, NCLMA99, vBC99]. The question that immediately arises is “where does this extra information come from?” We have gained considerable experience with utilizing extra domain information in a planner through implementing numerous planning domains in the TLPLAN system [BK00]. The TLPLAN system is a planning system specifically constructed to utilize extra domain knowledge. In fact, the underlying planning algorithm is very simple—a simple forward chaining engine. Hence almost all of TLPLAN’s performance is due to the extra domain knowledge it utilizes. We have noticed that a number of systematic principles can be applied to uncover extra domain knowledge, and that once one has developed an intuitive feel for these principles it becomes quite easy to identify and write down useful extra domain knowledge. Some of these principles seem to involve proofs by induction, at which people can be quite accomplished but which seem to be difficult to automate. Nevertheless, other principles seem to involve much simpler reasoning. In this paper we present an idea that we are working on that appears to have considerable potential for generating non-inductive domain knowledge. We formalize the idea and show how it can be used in various ways to compute extra domain knowledge.