Incremental, Approximate Planning

Abstract

This paper shows how using a nonmonotonic logic to describe the effects of actions enables plausible plans to be discovered quickly, and then refined if time permits. Candidate plans are found by allowing them to depend on unproved assumptions. The nonmonotonic logic makes explicit which antecedents of rules have the status of default conditions, and they are the only ones that may be left unproved, so only plausible candidate plans are produced. These are refined incrementally by trying to justify the assumptions on which they depend. The new planning strategy has been implemented, with good experimental results.