Fast Downward

Association

Multiple research institutions.

Description

We use Fast Downward, a planning system developed in the automated planning community, as the solver. For each problem we make an automatic translation from ASP to PDDL, the Planning Domain Definition Language used in the automated planning competitions, after modeling the benchmark as a planning domain. We have restricted ourselves to benchmarks that could be easily seen as planning problems. Apart from the modeling process for each benchmark we have used different configurations of Fast Downward.

Participants

• Malte Helmert - Albert-Ludwigs-Universität Freiburg, Institut für Informatik
• Vidal Alcázar - Universidad Carlos III de Madrid
• Erez Karpas - Technion
• Silvia Richter - NICTA
• Holger H. Hoos - University of British Columbia
• Chris Fawcett - University of British Columbia

System Settings

• Sokoban Decision: --heuristic "h=ff(cost_type=2)" --search "eager_greedy(h, preferred=(h),bound=$BOUND)" ; uses the FF heuristic with plan length as a metric, the search algorithm is greedy best-first search with eager evaluation, preferred operators and a bound on the maximum solution length determined by the maximum number of steps extracted from the problem description.

• Sokoban Optimization: --heuristic "h=ff(cost_type=2)" --search "eager_greedy(h, preferred=(h),bound=$BOUND)" ; uses the lm-cut heuristic with a* as search algorithm and a bound on the maximum solution length determined by the maximum number of steps extracted from the problem description.

• Knight Tour: --heuristic "h1=ff()" --heuristic "h2=goalcount()" --search "eager_greedy(h1,h2,preferred=(h1))"; uses a dual queue sorted by two different heuristics, a goal counting heuristic and the ff heuristic, the search algorithm is greedy best-first search with eager evaluation and preferred operators from the ff heuristic.

• Hanoi Tower: --search "astar(selmax(lmcut(),hmax(),alpha=4,training_set=10),bound=$bound,pathmax=true)"; uses the selective-max heuristic with the landmark-cut heuristic and hmax, the search algorithm is a* with pathmax and a bound on the maximum solution length determined by the maximum number of steps extracted from the problem description.

• Airport Pickup: --heuristic "h=cg()" --search "eager_greedy(h, preferred=(h))"; uses the causal graph heuristic, the search algorithm is greedy best-first search with eager evaluation and preferred operators.

• Solitaire: "hff=ff()" --search "eager_greedy(hff, preferred=(hff))"; uses the FF heuristic, the search algorithm is greedy best-first search with eager evaluation and preferred operators.

• Hydraulik Leaking: --search "astar(lmcut())"; uses the lm-cut heuristic with a* as search algorithm.

• Hydraulic Planning: --search "astar(lmcount(lm_merged(lm_rhw(),lm_hm(m=1)),admissible=true),mpd=true)"; uses the landmark counting heuristic with cost partitioning and a* with multi-path dependence.