Skip to content

jendrikseipp/scorpion

Repository files navigation

Scorpion

Scorpion is a classical planning system that extends Fast Downward. The main extensions are:

  • novel state-of-the-art algorithms for optimal classical planning
  • additional search algorithms
  • several new plugin options and utilities

See below for a detailed list of extensions. We regularly port the latest changes from Fast Downward to Scorpion and also integrate some features from Scorpion back into Fast Downward.

Citing Scorpion:
Jendrik Seipp, Thomas Keller and Malte Helmert.
Saturated Cost Partitioning for Optimal Classical Planning.
Journal of Artificial Intelligence Research 67, pp. 129-167. 2020.

Instructions

Apptainer image

To simplify the installation process, we provide an executable Apptainer container (formerly known as Singularity). It accepts the same arguments as the fast-downward.py script (see below).

# Download the image,
apptainer pull scorpion.sif oras://ghcr.io/jendrikseipp/scorpion:latest

# or build it yourself.
apptainer build scorpion.sif Apptainer

# Then run the recommended configuration (for solving STRIPS tasks optimally).
./scorpion.sif --transform-task preprocess-h2 --alias scorpion [DOMAIN_FILE] PROBLEM_FILE

Manual compilation

Install the dependencies (the table below lists which versions are tested):

sudo apt install cmake g++ git make python3

For plugins based on linear programming (e.g., ocp(), pho()) you need to add an LP solver. Then compile the planner with

./build.py

and see the available options with

./fast-downward.py --help  # driver
./fast-downward.py --search -- --help  # search component

For more details (including build instructions for macOS and Windows), see the documentation about compiling and running the planner. The plugin documentation shows which plugins are available (heuristics, search algorithms, etc.) and how to use them.

Recommended configurations

In case you want to solve tasks quickly and do not require optimality, we recommend using the first iteration of LAMA with an added type-based open list:

./fast-downward.py \
  --transform-task preprocess-h2 \
  [DOMAIN_FILE] PROBLEM_FILE \
  --search "let(hlm, landmark_sum(lm_reasonable_orders_hps(lm_rhw()), transform=adapt_costs(one)),
    let(hff, ff(transform=adapt_costs(one)),
    lazy(alt([single(hff), single(hff, pref_only=true), single(hlm), single(hlm, pref_only=true),
    type_based([hff, g()])], boost=1000), preferred=[hff, hlm], cost_type=one)))"

For solving STRIPS tasks optimally, we recommend using the --alias scorpion shortcut

./fast-downward.py --transform-task preprocess-h2 --alias scorpion PROBLEM_FILE

which is equivalent to

./fast-downward.py \
  --transform-task preprocess-h2 \
  [DOMAIN_FILE] PROBLEM_FILE \
  --search "astar(scp_online([
      projections(sys_scp(max_time=100, max_time_per_restart=10)),
      cartesian()],
      saturator=perimstar, max_time=1000, interval=10K, orders=greedy_orders()),
      pruning=limited_pruning(pruning=atom_centric_stubborn_sets(), min_required_pruning_ratio=0.2))"

The preprocess-h2 call prunes irrelevant operators in a preprocessing step. The search configuration uses partial order reduction and maximizes over diverse, subset-saturated cost partitioning heuristics computed online during the search. The underlying abstractions are Sys-SCP pattern databases and Cartesian abstractions.

(In Downward Lab you can use add_algorithm(name="scorpion", repo="path/to/repo", rev="scorpion", component_options=[], driver_options=["--transform-task", "preprocess-h2", "--alias", "scorpion"] to run the recommended Scorpion configuration.)

For solving tasks with conditional effects optimally, we recommend using

./fast-downward.py \
  --transform-task preprocess-h2 \
  [DOMAIN_FILE] PROBLEM_FILE \
  --search "astar(scp_online([projections(sys_scp(
        max_time=100, max_time_per_restart=10, max_pdb_size=2M, max_collection_size=20M,
        pattern_type=interesting_non_negative, create_complete_transition_system=true),
      create_complete_transition_system=true)],
    saturator=perimstar, max_time=100, max_size=1M, interval=10K, orders=greedy_orders()))"

IPC versions

If you prefer to run the Scorpion versions from the IPC 2018 or 2023 (which are based on an older Fast Downward version and use different abstractions), we recommend using the Apptainer images from the Scorpion 2018 or Scorpion 2023 repos.

Differences between Scorpion and Fast Downward

Diff between the latest merged version of Fast Downward and Scorpion: https://github.com/jendrikseipp/scorpion/compare/main...scorpion

  • Scorpion comes with the h²-preprocessor by Vidal Alcázar and Álvaro Torralba that prunes irrelevant operators. Pass --transform-task preprocess-h2 to use it.
  • The --transform-task command allows you to run arbitrary preprocessing commands that transform the SAS+ output from the translator before passing it to the search.
  • Scorpion uses incremental search for Cartesian abstraction refinement.
  • Scorpion uses a phmap::flat_hash_set to check for duplicate states, which often drastically reduces the peak memory usage, compared to Fast Downward's IntHashSet.
  • If ccache is installed (recommended), Scorpion uses it to cache compilation files.

New translator options

  • Use --dump-predicates and --dump-static-atoms to write files with information that's useful for learning domain control knowledge.

New plugin options

  • {cegar/cartesian}(..., pick_flawed_abstract_state={batch_min_h, ...}): find all current flaws, then iteratively repair the flaw that's closest to the goal (paper, default=batch_min_h).

  • {cegar/cartesian}(..., pick_split={max_cover, ...}, tiebreak_split={max_refined, ...}): smarter strategies for splitting a flawed abstract state (paper, default=max_cover and max_refined for tiebreaking).

  • {cegar,cartesian}(..., dot_graph_verbosity={silent, write_to_console, write_to_file}): write intermediate abstractions as Graphviz dot files to stdout or to files (default=silent).

  • hillclimbing(..., max_generated_patterns=200): limit the number of patterns generated by hill climbing.

  • systematic(..., pattern_type=interesting_general): compute interesting patterns for general cost partitioning.

New cost partitioning algorithms for abstraction heuristics

We use Cartesian abstractions in the example configurations below ([cartesian()]). You can also use pattern database heuristics, e.g., [projections(systematic(2))], or mix abstractions, e.g., [projections(systematic(3)), cartesian()]. Some of the algorithms below are also part of vanilla Fast Downward, but are only implemented for PDB heuristics.

  • Optimal cost partitioning: ocp([cartesian()])
  • Canonical heuristic: canonical_heuristic([cartesian()])
  • Uniform cost partitioning: ucp([cartesian()], opportunistic=false)
  • Opportunistic uniform cost partitioning: ucp([cartesian()], ..., opportunistic=true)
  • Greedy zero-one cost partitioning: gzocp([cartesian()], ...)
  • Saturated cost partitioning: scp([cartesian()], ...) (offline), scp_online([cartesian()], ...) (online)
  • (Saturated) post-hoc optimization: pho([cartesian()], ..., saturated={false,true}) (offline), operatorcounting([pho_abstraction_constraints([cartesian()], saturated={false,true})]) (online)

You can also compute the maximum over abstraction heuristics:

  • maximize([cartesian()])

The plugin documentation shows all options for cost partitioning heuristics.

New pattern collection generators

  • Systematic patterns with size limits: sys_scp(max_pattern_size=X, max_pdb_size=Y, max_collection_size=Z, ..., saturate=false)
  • Sys-SCP patterns: sys_scp(...)

New cost partitioning algorithms for landmark heuristics

Example using A* search and saturated cost partitioning over BJOLP landmarks:

--evaluator
  "lmc=landmark_cost_partitioning(lm_merged([lm_rhw(), lm_hm(m=1)]),
  cost_partitioning=suboptimal, greedy=true,
  reuse_costs=true, scoring_function=max_heuristic_per_stolen_costs)"
--search
  "astar(lmc, lazy_evaluator=lmc)"

Different cost partitioning algorithms for landmark heuristics:

  • Optimal cost partitioning (part of vanilla Fast Downward): landmark_cost_partitioning(..., cost_partitioning=optimal)
  • Canonical heuristic: landmark_cost_partitioning(..., cost_partitioning=canonical)
  • Post-hoc optimization: landmark_cost_partitioning(..., cost_partitioning=pho)
  • Uniform cost partitioning: landmark_cost_partitioning(..., cost_partitioning=suboptimal, greedy=false, reuse_costs=false)
  • Opportunistic uniform cost partitioning (part of vanilla Fast Downward): landmark_cost_partitioning(..., cost_partitioning=suboptimal, greedy=false, reuse_costs=true, scoring_function=min_stolen_costs)
  • Greedy zero-one cost partitioning: landmark_cost_partitioning(..., cost_partitioning=suboptimal, greedy=true, reuse_costs=false, scoring_function=max_heuristic)
  • Saturated cost partitioning: landmark_cost_partitioning(..., cost_partitioning=suboptimal, greedy=true, reuse_costs=true, scoring_function=max_heuristic_per_stolen_costs)

New search engines

  • Breadth-first search (without overhead of the more general eager() search): brfs()
  • Depth-first search: dfs()
  • Exhaustive search (useful for dumping the reachable state space of small input tasks): dump_reachable_search_space()
  • IDA* search: idastar(cegar(cache_estimates=false))
  • Iterative width search: iw(width=2)

Fast Downward

Fast Downward is a domain-independent classical planning system.

Copyright 2003-2023 Fast Downward contributors (see below).

For further information:

Scientific experiments

We recommend to use the latest release instead of the tip of the main branch. The Downward Lab Python package helps running Fast Downward experiments. Our separate benchmark repository contains a collection of planning tasks.

Supported software versions

The planner is mainly developed under Linux; and all of its features should work with no restrictions under this platform. The planner should compile and run correctly on macOS, but we cannot guarantee that it works as well as under Linux. The same comment applies for Windows, where additionally some diagnostic features (e.g., reporting peak memory usage when the planner is terminated by a signal) are not supported. Setting time and memory limits and running portfolios is not supported under Windows either.

This version of Fast Downward has been tested with the following software versions:

OS Python C++ compiler CMake
Ubuntu 22.04 3.10 GCC 11, GCC 12, Clang 14 3.22
Ubuntu 20.04 3.8 GCC 10, Clang 12 3.16
macOS 12 3.10 AppleClang 14 3.24
macOS 11 3.8 AppleClang 13 3.24
Windows 10 3.8 Visual Studio Enterprise 2019 (MSVC 19.29) and 2022 (MSVC 19.31) 3.22

We test LP support with CPLEX 22.1.1 and SoPlex 6.0.3+. On Ubuntu we test both CPLEX and SoPlex. On Windows we currently only test CPLEX, and on macOS we do not test LP solvers (yet).

Build instructions

See BUILD.md.

Contributors

The following list includes all people that actively contributed to Fast Downward, i.e., all people that appear in some commits in Fast Downward's history (see below for a history on how Fast Downward emerged) or people that influenced the development of such commits. Currently, this list is sorted by the last year the person has been active, and in case of ties, by the earliest year the person started contributing, and finally by last name.

  • 2003-2023 Malte Helmert
  • 2008-2016, 2018-2023 Gabriele Roeger
  • 2010-2023 Jendrik Seipp
  • 2010-2011, 2013-2023 Silvan Sievers
  • 2012-2023 Florian Pommerening
  • 2013, 2015-2023 Salomé Eriksson
  • 2015, 2021-2023 Thomas Keller
  • 2018-2023 Patrick Ferber
  • 2018-2020, 2023 Augusto B. Corrêa
  • 2021-2023 Clemens Büchner
  • 2022-2023 Remo Christen
  • 2023 Simon Dold
  • 2023 Claudia S. Grundke
  • 2023 Victor Paléologue
  • 2023 Emanuele Tirendi
  • 2021-2022 Dominik Drexler
  • 2016-2020 Cedric Geissmann
  • 2017-2020 Guillem Francès
  • 2020 Rik de Graaff
  • 2015-2019 Manuel Heusner
  • 2017 Daniel Killenberger
  • 2016 Yusra Alkhazraji
  • 2016 Martin Wehrle
  • 2014-2015 Patrick von Reth
  • 2009-2014 Erez Karpas
  • 2014 Robert P. Goldman
  • 2010-2012 Andrew Coles
  • 2010, 2012 Patrik Haslum
  • 2003-2011 Silvia Richter
  • 2009-2011 Emil Keyder
  • 2010-2011 Moritz Gronbach
  • 2010-2011 Manuela Ortlieb
  • 2011 Vidal Alcázar Saiz
  • 2011 Michael Katz
  • 2011 Raz Nissim
  • 2010 Moritz Goebelbecker
  • 2007-2009 Matthias Westphal
  • 2009 Christian Muise

History

The current version of Fast Downward is the merger of three different projects:

  • the original version of Fast Downward developed by Malte Helmert and Silvia Richter
  • LAMA, developed by Silvia Richter and Matthias Westphal based on the original Fast Downward
  • FD-Tech, a modified version of Fast Downward developed by Erez Karpas and Michael Katz based on the original code

In addition to these three main sources, the codebase incorporates code and features from numerous branches of the Fast Downward codebase developed for various research papers. The main contributors to these branches are Malte Helmert, Gabi Röger and Silvia Richter.

License

Fast Downward 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.

Fast Downward 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/>.