Powerlifted is a domain-independent classical planner that uses only lifted representations.
(See References for more details.)
The powerlifted.py
script solves a PDDL task provided as input. It also builds
the planner if the --build
parameter is passed. To run a single search, you
can use the following algorithms:
$ ./powerlifted.py [-d DOMAIN] -i INSTANCE -s SEARCH -e EVALUATOR -g GENERATOR [--state STATE REPR.] [ADDITIONAL OPTIONS] [--build]
The options for each parameter are described below. If you do not pass any value for SEARCH
, EVALUATOR
, and GENERATOR
, the planner will use the best (known) configuration for satisficing planning (i.e., no optimality guaranteed). (See next section for more details.)
It is also possible to perform multiple search algorithms on the same task iteratively. See the section "Multiple Search Algorithms" below.
You can either use the build.py
script to build the planner first, or pass the --build
flag to build the planner prior to the search execution.
Currently, the best configuration for satisficing planning (with respect to total coverage) is the following:
$ ./powerlifted.py [-d DOMAIN] -i INSTANCE -s alt-bfws1 -e ff -g yannakakis [ADDITIONAL OPTIONS] [--build]
These are also the default values for -s
, -e
, and -g
. To maximize
coverage, we also recommend adding --unit-cost
(see below) to the ADDITIONAL OPTIONS
.
alt-bfws1
andalt-bfws2
: [R_x, h] with w=1 and w=2, respectively. The choice of h is given theEVALUATOR
option. (See Corrêa and Seipp 2022.)astar
: A* Searchbfs
: Breadth-First Search (This option was previously callednaive
. You can still usenaive
with thepowerlifted.py
script but the planner will internally use the new keywordbfs
.)bfws1
andbfws2
: Best-First Width Search with w=1 and w=2, respectively.bfws1-rx
andbfws2-rx
: BFWS(R_x) with w=1 and w=2, respectively. (See Corrêa and Seipp 2022.)dq-bfws1-rx
anddq-bfws2-rx
: DQ(R_x) with w=1 and w=2, respectively. (See Corrêa and Seipp 2022.)gbfs
: Greedy Best-First Searchiw1
andiw2
: Iterated Width Search (with w=1 and w=2, respectively)lazy
: Lazy Best-First Searchlazy-po
: Lazy Best-First Search with Boosted Dual-Queuelazy-prune
: Lazy Best-First Search with pruning of states generated by non-preferred operators
add
: The additive heuristicblind
: No Heuristicff
: The FF heuristicgoalcount
: The goal-count/STRIPS heuristichmax
: The hmax heuristicrff
: The rule-based FF heuristic
join
: Join program using the predicate order given in the PDDL filerandom_join
: Randomly ordered join programordered_join
: Join program ordered by the arity of the predicatesfull_reducer
: Generate successor for acyclic schemas using the full reducer method; for cyclic schemas it uses a partial reducer and a join program.yannakakis
: Same as above but replaces the final join of the full reducer method by the Yannakakis' project-join program.
sparse
: Use the sparse state representation where a state is only represented by the facts that are true in this state.extensional
: Use the extensional representation where a state is a bitset where the ith-bit is true if the fact associated to it is true in this state. This representation requires the grounding of facts (but not of actions) which, right now, is performed in the search component. Warning: this setting does not support allEVALUATOR
options.
[--novelty-early-stop]
: Flag if the novelty evaluation of a state should stop as soon as the return value is defined. (See Corrêa and Seipp 2022.)[--only-effects-novelty-check]
: Flag if the novelty evaluation of a state should only consider atoms in the applied action effect. Warning: for state-of-the-art performance, you must use this option when running BFWS-based search engines. (See Corrêa and Seipp 2022.)[--plan-file]
: Name of the file used to output the plan(s) found. If you are using sequential iterations, the plan found at the x-th iteration will have.x
appended to its name.[--preprocess-task]
: Preprocess domain and problem PDDL files so it is translated into a fragment supported by Powerlifted. You need to have CPDDL installed, and an environment variableCPDDL_BIN
pointing to thecpddl/bin
directory after it is locally compiled. Note: This is still not fully functional. CPDDL still leaves negated static precondition on the PDDL file, which Powerlifted does not support. Also, it produces two new intermediate files, which now have hardcoded names.[--seed RANDOM SEED]
: Random seed for the random number generator.[--translator-output-file TRANSLATOR_FILE]
: Output of the intermediate representation to be parsed by the search component will be saved intoTRANSLATOR_FILE
. (Default:output.lifted
)[--unit-cost]
: Use unit cost (i.e., all costs are equal to 1) instead of the costs specified in the domain file.[--validate]
: Runs VAL after a plan is found to validate it. This requires VAL to be added asvalidate
to thePATH
.
You can use the flag --iteration
to specify one single search iteration for
the planner. You can pass as many --iteration
arguments as you wish, and each
argument will execute a different search.
The syntax to specify a search iteration is the following:
$ ./powerlifted.py -i INSTANCE --iteration S,E,G
where S
is a search algorithm, E
is an evaluator, and G
a successor generator. For example, to execute Greedy Best-First Search with FF followed by a Lazy Best-First Search with the additive heuristic (and both using the Yannakakis' algorithm for successor generation), you should run
$ ./powerlifted.py -i INSTANCE --iteration gbfs,ff,yannakakis --iteration lazy,add,yannakakis
The plan founds are then numbered based on its iterations. If the first iteration finds a plan, it will be called plan.1
; the second will be called `plan.2; etc.
Unfortunately, the planner has the limitation that additional options are set for all the iterations.
You can also build a Singularity image to run the planner. This might be useful
in the case where you are not able to compile the planner locally, for
example. To do so, first remove the builds/
directory, in case you have any
builds already in your system. Then, you can run the following command to create
the planner image:
sudo singularity build powerlifted.sif Singularity
Be aware that this might take a while. Once the image powerlifted.sif
is
created, you can run it with the same parameters as the powerlifted.py
script. The only exception is that, by default, VAL is not installed in the
container, so it is not possible to use the --validate
flag with the
Singularity image. However, you can run VAL with the sas_plan
file created by
the planner after the execution. The following command is a usage example on
how to run the planner with the Singularity image:
./powerlifted.sif -i /path/to/instance.pddl -s lazy-po -e add -g yannakakis --datalog-file model.lp --translator-output-file output.lifted
- A C++17-compliant compiler
- CMake 3.9+
- Python 3.5+
- Boost
- Axioms: not supported
- Conditional effects: not supported
- Negated preconditions: only inequality
- Quantifiers: not supported
- Corrêa, A. B.; Pommerening, F.; Helmert, M.; and Francès, G. 2020. Lifted Successor Generation using Query Optimization Techniques. In Proc. ICAPS 2020, pp. 80-89. [pdf]
- Corrêa, A. B.; Francès, G.; Pommerening, F.; and Helmert, M. 2021. Delete-Relaxation Heuristics for Lifted Classical Planning. In Proc. ICAPS 2021, pp. 94-102. [pdf]
- Corrêa, A. B.; Pommerening, F.; Helmert, M.; and Francès, G. 2022. The FF Heuristic for Lifted Classical Planning. In Proc. AAAI 2022. [pdf]
- Corrêa, A. B.; and Seipp, J. 2022. Best-First Width Search for Lifted Classical Planning. In Proc. ICAPS 2022. [pdf]