SymK is a state-of-the-art classical optimal and top-k planner based on symbolic search that extends Fast Downward. The main extensions are described below.
With SymK, it is possible to find a single optimal plan or a set of k different best plans with the lowest cost for a given planning task. In addition, SymK natively supports a variety of PDDL features that are rarely supported by other planners, such as conditional effects, derived predicates with axioms, and state-dependent action costs. See this readme file for more information on running SymK and the various configurations. We appreciate citations when SymK is used in a scientific context (see References for more details).
- Getting Started
- Single Optimal Solution
- Multiple Solutions
- Plan Selection Framework
- Pitfalls and Troubleshooting
- References
- License
Currently we only support Linux systems. The following should install all necessary dependencies.
sudo apt-get -y install cmake g++ make python3 autoconf automakeSymK should compile on MacOS with the GNU C++ compiler and clang with the same instructions described above.
./build.py To simplify the installation process, we alternatively provide an executable Apptainer container (formerly known as Singularity). It accepts the same arguments as SymK (fast-downward.py script; see below).
# Download the image,
apptainer pull symk.sif oras://ghcr.io/speckdavid/symk:latest
# or build it yourself.
apptainer build symk.sif Apptainer
# Then run the desired configuration (for other configurations see below).
./symk.sif domain.pddl problem.pddl --search "sym_bd()"We recommend using the following configuration which uses bidirectional search.
./fast-downward.py domain.pddl problem.pddl --search "sym_bd()"Other configurations are forward or backward search: --search "sym_fw()" or --search "sym_bw()".
If you are interested in more options, you can run ./fast-downward.py --search -- --help sym_bd to view the help for sym_bd.
Note: By default, the planner performs a relevance analysis, eliminating variables and actions that are irrelevant to achieving the goal, thus avoiding uninteresting plans. For details on how to disable this optimization, see Pitfalls and Troubleshooting.
We recommend using the following configuration which uses bidirectional search and
reports the best k plans. Note that you can also specify num_plans=infinity if you want to find all possible plans.
./fast-downward.py domain.pddl problem.pddl --search "symk_bd(plan_selection=top_k(num_plans=**k**,dump_plans=false))"Note that with ./fast-downward.py --plan-file sas_plan dom... you can specify the path and name of the generated plan files (here: sas_plan), and by setting the dump_plans argument to true, all plans found will also be dumped to the console.
If you are interested in more options, you can run ./fast-downward.py --search -- --help symk_bd to view the help for symk_bd.
We recommend using the following configuration which uses bidirectional search and
reports the k plans with quality bound q. Quality 1<=q<=infinity is a multiplier that is multiplied to the cost of the cheapest solution.
For example, q=1 reports only the cheapest plans, where quality=infinity corresponds to the top-k planning.
./fast-downward.py domain.pddl problem.pddl --search "symq_bd(plan_selection=top_k(num_plans=**k**,dump_plans=false),quality=**q**)"It is possible to generate loopless/simple plans, i.e., plans that do not visit any state more than once. In general, the option to consider and generate only simple plans can be combined with any SymK search engine and with different plan selectors by setting the simple parameter to true. See the following two examples and our ICAPS 2022 Paper.
./fast-downward.py domain.pddl problem.pddl --search "symk_bd(simple=true,plan_selection=top_k(num_plans=**k**,dump_plans=false))"./fast-downward.py domain.pddl problem.pddl --search "symq_bd(simple=true,plan_selection=top_k(num_plans=**k**,dump_plans=false),quality=**q**)"It is possible to run SymK also with forward or backward search instead of bidirectional search, e.g., with --search "symk_fw(...)" or --search "symk_bw(...)". Depending on the domain, one of these configurations may be faster than bidirectional search ("--search symk_bd(...)").
It is possible to create plans until a number of plans or simply a single plan is found that meets certain requirements. For this purpose it is possible to write your own plan selector. During the search, plans are created and handed over to a plan selector with an anytime behavior.
An example of a plan selector is the unordered_selector, which considers two plans as equivalent if their action multi-sets are equivalent. In other words, plans with the same multi-set of actions form an equivalence class and only one representative plan is reported for each equivalence class. Note that plan selectors can be combined with the different planning configurations.
We recommend to use the following configurations which use bidirectional search.
./fast-downward.py domain.pddl problem.pddl --search "symk_bd(plan_selection=unordered(num_plans=**k**))"./fast-downward.py domain.pddl problem.pddl --search "symq_bd(plan_selection=unordered(num_plans=**k**),quality=**q**)"Two simple examples of plan selectors are the top_k_selector and the top_k_even_selector. For this purpose it is possible to write your own plan selector. The most important function is add_plan, in which you can specify whether a newly generated plan shall be accepted or rejected. To create your own plan selector, you can copy the .cc and .h files of one of these two selectors and adjust them accordingly. Also add the new file name to DownwardFiles.cmake, similar to the other selection files. Finally, if you want to find a plan with your awesome_selector selector (the name of the selector you specified for the plugin in the .cc file), you can use the following command.
./fast-downward.py domain.pddl problem.pddl --search "symk_bd(plan_selection=awesome_selector(num_plans=1))"Note, that you can also search for the best k plans using your selector.
By default, the planner performs a relevance analysis and removes components such as variables and actions that are irrelevant to achieving the goal. Although such variables and actions can in principle lead to further (simple) plans, they are classified as irrelevant and removed when translating PDDL to SAS+. If you wish to obtain all plans (even the non-relevant ones), please use the following options. Note that this can have a negative impact on the performance of the planner.
./fast-downward.py domain.pddl problem.pddl --translate-options --keep-unimportant-variables --preprocess-options --keep-unimportant-variables --search-options --search "symk_bd(plan_selection=top_k(num_plans=**k**))"Note that several components of SymK have been developed and published separately. We appreciate citations of these sources when used.
- David Speck, Robert Mattmüller, Bernhard Nebel: Symbolic Top-k Planning. AAAI 2020: 9967-9974 [pdf] [bib & more]
- Julian von Tschammer, Robert Mattmüller, David Speck: Loopless Top-K Planning. ICAPS 2022: 380-384 [pdf] [bib & more]
- David Speck, Florian Geißer, Robert Mattmüller, Álvaro Torralba: Symbolic Planning with Axioms. ICAPS 2019: 464-472 [pdf] [bib & more]
- David Speck: Symbolic Search for Optimal Planning with Expressive Extensions. Ph.D. thesis: University of Freiburg (2022) [pdf] [bib & more]
You can find examples of domains with state-dependent action cost here.
We want to acknowledge that SymK is based on:
- Fast Downward (23.06+): http://www.fast-downward.org/ and FD_README.md
- Symbolic Fast Downward: https://people.cs.aau.dk/~alto/software.html
Finally, SymK uses some external software that can be found in the src/search/ext folder.
SymK is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or (at
your option) any later version.
SymK is distributed in the hope that it will be useful, but
WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <https://www.gnu.org/licenses/>.