Planning with Multiple Action-Cost Estimates

Abstract

AI Planning require computing the costs of ground actions. While often assumed to be negligible, the run-time of this computation can become a major component in the overall planning run-time. To address this, we introduce planning with multiple action cost estimates, a generalization of classical planning, where action cost can be incrementally determined using multiple estimation procedures, which trade computational effort for increasingly tightening bounds on the exact cost. We then present ACE, a generalized A*, to solve such problems. We provide theoretical guarantees, and extensive experiments that show considerable run-time savings compared to alternatives.