Viewing Landmarks as Temporally Extended Goals

Landmarks are facts that must hold true at some point in all solutions to a planning problem. The exploitation of landmarks and associated ordering constraints has been shown to be very effective in computing high-quality plans, quickly. Indeed the success of LAMA, winner of the 2008 International Planning Competition sequential satisficing track, can be attributed in some measure to its ability to generate reliable landmark and ordering information and to exploit them via a pseudo-heuristic. In this paper we view landmarks and their associated orderings as temporally extended goals (TEGs) and adapt an existing heuristic for TEGs to provide guidance to a planner in two ways: by estimating distance to achieving landmarks and goals while respecting orderings, and by identifying dead-ends in the search tree that can be pruned by the planner. We integrate our heuristic as an extension to LAMA. Experiments show that our planner outperforms LAMA in several domains, while exploring significantly less of the search space in over 50% of the experiments run.