ASP encoding to ground planning tasks

Encodes the relaxed-reachability of a PDDL planning task as a logic program. The canonical model of this program is equivalent to all reachable atoms (and actions). This can be used to find all possible ground atoms that are relaxed-reachable in the planning task. In fact, this is how Fast Downward grounds the input tasks. Similarly, this is how Powerlifted extracts delete-relaxation heurisitics without grounding the problem.

You can either do the traditional grounding in one pass or do a two-step grounding, which is slower but can ground larger tasks. For details, see Corrêa et al. (ICAPS 2023).

Installation

We recommend using a Python virtual environment. You need to use Python >= 3.7 to execute the scripts. Once inside the virtual environment, you can run

$ python setup.py install

or

$ pip install -e .

while in the root directory of the repository. You also need to have the grounder you want to use in your PATH (see below).

Basic Usage

There are two main scripts in the repo: generate-asp-model.py and count-ground-actions.py. The first one simply encodes the task as a Datalog program (with some options) and then grounds it using some off-the-shelf grounder; the second one counts or grounds (depending on the parameters) the actions of the task given a set of relaxed-reachable atoms. Usually, to run count-ground-actions.py, you need to have the output model from generate-asp-model.py stored.

To run the program generate-asp-model, execute

$ ./generate-asp-model.py -i /path/to/instance.pddl [-m MODEL-OUTPUT] [-t THEORY-OUTPUT]

where /path/to/instance.pddl is the path to a PDDL instance. It is necessary that there is a PDDL domain file in the same directory as instance.pddl, though. The script will infer the domain file automatically or you can pass it with parameter -d.

The program will generate the Datalog encoding corresponding to the PDDL task and ground it using some selected grounder (default: gringo). The Datalog file will be saved in MODEL-OUTPUT (default: output.theory); the canonical model (together with any other output from the grounder) will be saved in THEORY-OUTPUT (default: output.model).

There are some extra options one can use:

• --ground-actions: Use the encoding from Helmert (AIJ 2009) where action predicates are listed explcitly.
• --grounder: Select grounder to be used to ground the Datalog program. Current options are clingo, gringo, idlv, and newground. You must have the selected grounder on the PATH. (Default: gringo)
• --lpopt-preprocessor: Uses the lpopt preprocessor to rewrite the Datalog program. This option expects an environment variable called LPOPT_BIN_PATH to point to the binary file of lpopt.
• --fd-split: Uses Fast Downward preprocessor to split rules of the Datalog program. This uses the method by Helmert (AIJ 2009).
• --htd-split: (This option is not fully functional yet.) Splits the rules based on the hypertree decompositions of the rule bodies. It expects BalancedGo to be on the PATH.

To run the count-ground-actions program for upper-bounding the number of action atoms in the grounding and for computing alternative groundings, you must set the following environment variables:

• LPCNT_AUX_PATH pointing to /path/to/repo/build/bdist.linux-x86_64.
• GRINGO_BIN_PATH pointing to gringo.
• CLINGO_BIN_PATH pointing to clingo.
• LPCNT_BIN_PATH pointing to the desired counting/solving script.

Then, you can execute

$ ./ count-ground-actions.py -m MODEL-OUTPUT -t THEORY-WITH-ACTIONS [--choices] [--output] [--extendedOutput] [--counter-path SCRIPT]

where MODEL-OUTPUT is the path to the MODEL-OUTPUT obtained with a call to the program above, and THEORY-WITH-ACTIONS is the path to the theory file containing action predicates obtained by the same call above (this file is automatically generated with the name output-with-actions.theory by the generate-asp-model.py script).

The key to this program is the counting/solving script used. The ones used in the experiments (Corrêa et al. ICAPS 2023) are:

• lpcnt: Simple counting script to count number of ground actions. Outputs with + Sign indicate a lower bound.
• lpgrnd: Grounding via solving (see Corrêa et al. ICAPS 2023). But note that this script does not output the model by default. This is on purpose because some of the instances in the tested benchmark produce files larger than 10 GiB.
• lpgrnd_io: Same as lpgrnd but outputting all ground actions. Be careful to not produce very large files.

There are other counting scripts in the codebase (lpcnt_nopp, lpcnt_omni, lpcnt_opt). They are different configurations that might work better for specific instances. However, they were not used in the aforementioned experiments.

There are some extra options that one can optionally turn on:

• --choices: Use ASP choice rules, which is an alternative encoding that directly uses ASP choices, instead of SAT-like rules for guessing / deciding whether an action is contained in the grounding.
• --output: Compute the alternative grounding encoding and write it to stdout.
• --extendedOutput: Compute an alternative (extended) grounding encoding and write it to stdout.
• --counter-path: Set the used counting solver environment variable, giving the execution path within the path specified by the environment variable VAR. This option expects that SCRIPT points to a valid script for counting/solving. (default: LPCNT_BIN_PATH)