Eleven domain families were used for IPC-2008:
- Cyber security (BAMS)
- Model train
- Peg solitaire
We plan to eventually provide detailed information for each of these domains. In the meantime, please send an email to the competition organizers (see home page) if you have questions about the IPC-2008 benchmark suite which are not answered by these pages.
Domain archive and directory structure
Archive with all IPC-2008 domains: ipc2008.tar.bz2 (5 MB; expands to 666 MB)
Archive with all IPC-2008 domains except cyber security: ipc2008-no-cybersec.tar.bz2 (401 KB; expands to 19 MB)
The problem instances are grouped by track, domain and domain formulation (see below) following the general pattern ipc2008/<TRACK>/<DOMAIN>-<FORMULATION>/p<NUMBER>.pddl. Problem instances within each domain are numbered from 1 to 30. For each instance, a separate domain file is provided. In most domain formulations, all 30 domain files are identical.
ipc2008/seq-opt/openstacks-strips/p17.pddl: sequential optimization track, openstacks domain, STRIPS formulation, problem instance #17
ipc2008/seq-opt/openstacks-strips/p17-domain.pddl: the corresponding domain file
ipc2008/seq-opt/openstacks-adl/p17.pddl: an equivalent formulation of the same problem instance using ADL features
ipc2008/seq-sat/openstacks-strips/p17.pddl: an openstacks instance from the sequential satisficing track; much larger than instance #17 of the sequential optimization track
Domains and domain families
In cases where a domain family was used in more than one track, different domains were used as appropriate for that track. For example, three different peg solitaire domains were used:
sequential peg solitaire: Uses action costs to define the number of jump sequences as the optimization criterion.
temporal peg solitaire: Parallel jumps are allowed if they don't interfere. The objective is to minimize makespan.
net benefit peg solitaire: Unsolvable instances are used, and the objective is to clear away as many pegs as possible.
The sequential satisficing and sequential optimization tracks used the same domains apart from the Cyber Security domain which was limited to the satisficing track due to the large problem sizes. However, the satisficing tracks typically used larger (or otherwise more difficult) problem instances than the optimization track. The exceptions to this were the PARC printer, peg solitaire and Scanalyzer-3D domains, where both tracks used the same instance set.
If you refer to these benchmark suites in your papers, make sure to mention whether you mean the sequential or optimization track instances.
As in previous IPCs, we sometimes include different formulations of the same domain, for example a more natural formulation that requires ADL features and a compiled STRIPS version. Whenever a domain has different alternative formulations, these model exactly the same problem set and have exactly the same optimal solution quality, so results between different formulations are comparable. For example, the net benefit openstacks domain has three formulations:
netben/openstacks-stripsneg: Requires support for negative literals in preconditions.
netben/openstacks-adl: Requires support for ADL features.
netben/openstacks-numericadl: Requires support for numeric fluents and ADL features.
All these formulations additionally require support for action costs and soft goals. (These are general requirements of the net benefit track.) However, see below for versions of these domains with soft goals compiled away.
Bug: After the competition, we noticed that the numeric formulations of the temporal openstacks domain are not equivalent to the non-numeric formulations. Until this bug is fixed, please refrain from using the numeric formulations of temporal openstacks. This bug does not affect the competition outcome because the only planner which was capable of dealing with this domain formulation was separated from the planners above and below in the ranking by a wide enough score margin.
Compiled net benefit track instances
Emil Keyder and Hector Geffner have shown that soft goals do not add expressive power to the classical planning model with action costs, and a procedure for compiling planning instances with soft goals into equivalent instances with only action costs was described in Soft Goals Can Be Compiled Away. The tar file linked below contains versions of all instances from the net benefit track with soft goals compiled into action costs. Note that this formulation of the domains was not used in the competition. Please email questions, comments and bug reports concerning these instances to emil.keyder at upf.edu.
Note that the domain file may vary in these formulations from instance to instance, even if the files are identical for the corresponding original domain.