Toward Rational Deployment of Multiple Heuristics in A

David Tolpin, Tal Beja, Solomon Eyal Shimony, Ariel Felner, Erez Karpas

Abstract: The obvious way to use several admissible heuristics in A* is to take their maximum. In this paper we aim to reduce the time spent on computing heuristics. We discuss Lazy A*, a variant of A* where heuristics are evaluated lazily: only when they are essential to a decision to be made in the A* search process. We present a new rational meta-reasoning based scheme, Rational Lazy A*, which decides whether to compute the more expensive heuristics at all, based on amyopic value of information estimate. Both methods are examined theoretically. Empirical evaluation on several domains supports the theoretical results, and shows that Lazy A* and Rational Lazy A* are state-of-the-art heuristic combination methods.