Improving Relaxed-Plan-Based Heuristics

Relaxed-plan-based (RPB) heuristics were first proposed by Hoffmann and Nebel for their FF system and are still used by current top-performing planners. Their main characteristic is that they are computed by computing a so-called relaxed plan, which is a plan for a relaxed version of the problem that ignores negative effects of actions. However, still in some domains that humans consider simple, they provide bad guidance. Arguably, the reason is that disregarding deletes oversimplifies those domains. Consequently, relaxed plans ignore key parts of the domain’s structure. This paper describes preliminary work that attempts to identify how it is possible to compute better relaxed plans that will better respect the structure of the original (un-relaxed) problem. To that end, we propose two techniques for extracting improved relaxed plans. The first (domain-independent) technique identifies missing actions that would have to be performed if the relaxed plan was to be executed in the real (un-relaxed) world. The second (domain-dependent) technique uses domain knowledge, in the form of simple state constraints, to attempt to extract a relaxed plan that respects key information of the domain. We prove that the first technique can significantly improve the performance and quality of solutions obtained with a vanilla RPB heuristic and enforced hill climbing on a family of simple blocks-world problems. We experimentally show that both techniques improve search efficiency in example domains.