Welcome to the FF Domain Collection

This page was created by Joerg Hoffmann, to form a starting point for people who want to perform large-scale empirical studies in planning. The FF domain collection provides (randomized, where possible) generators for 20 STRIPS and ADL planning benchmark domains, including the examples used in both competitions. Below, we give, for each domain, information about the origin, made adaptions (if any), the parameters of the generator, and the randomization strategy. Downloads are available for the domain files and the C source code of the generators. Click here to download the whole package. The generators should be self-explanatory in terms of how to build them (makefiles are included), and how to run them. In the less obvious cases, we have also included a README file. Contact us with questions.

Note: We do not make any claims whatsoever about the validity of the problem ranges generated, nor about the adequacy of the kind of problems that are generated within a domain. We have made an effort to imitate closely the examples known from published problem suites, and we have generally chosen the most obvious first-guess randomization strategy. Some of the generators, like the one for Assembly, are quite a hack. Nevertheless, we believe that the generators form an invaluable tool for experimentation --- they have definetely done so in our own experiments. Please contact us with any comments or suggestions.

Acknowledgments: We thank all persons involved in the development of any of the domains below. We have mentioned all names of persons we knew were involved, and apologize if someone was left out. Please contact us with any additional information on this. Thanks also go to Malte Helmert for discussions on the validity of problem ranges, and to Ulrich Scholz for pointing out bugs.

Assembly

• Origin: Drew McDermott. Used in the AIPS-1998 competition.
• Adaptions: None.
• Description: Typed ADL domain using complex pre- and effect-conditions, as well as conditional effects. Assemble one goal item by repeatedly putting together several items that are part of the same meta-item. The part-of relation defines a tree, where the root is the goal item and the leafs are the base items that do not need to be assembled. Any item can need a resource (like an ofen) during the process of its being assembled. It may be that a ``transient'' part has to be assembled somewhere and removed afterwards, that parts of the same meta-item need to be assembled in a specified order (obeying assemble order constraints) and that some other part needs to be assembled before the transient part is removed again (obeying remove order constraints).
• Parameters:
  • -d depth of part-of tree
  • -m maximal number of sons to any node in the tree
  • -h probability that a node has any sons
  • -n number of distinct resources
  • -r probability that a non-base node requires a resource
  • -t probability that an item is transient part for any higher-in-tree item
  • -a probability that a pair of items has an assemble order constraint
  • -o probability that an item has a remove ordering constraint with a transient part
• Generation: Create a tree of depth -d. Nodes have sons with probability -h. If they do have sons, then a random number between 1 and -m, biased to be lower the deeper the node is in the tree. With probability -r, the node needs a random resource. An item is assigned a transient part relation to any item in a higher tree level with probability -t, parts or transient parts of the same item have an assemble ordering constraint with probability -a, and for a transient part relation (A, B) the parts of B are given a remove ordering constraint to A with probability -o. Cycles in the assemble and remove order constraints are avoided by arbitrarily ordering the respective items, and allowing constraints only between pairs A and B with A < B.
• Downloads: domain.pddl, GENERATOR.tar.gz