Code and experiment data for the ICAPS 2024 paper "Versatile Cost Partitioning with Exact Sensitivity Analysis"

Code

The hoeft-et-al-icaps2024-code.zip directory contains our modified version of the Scorpion planner (https://github.com/jendrikseipp/scorpion) which in turn is based on the Fast Downward planning system (https://github.com/aibasel/downward).
All implemented SPhO variants can be found in the folder "src/search/operator_counting/".
The experiments externally depend on the CPLEX LP solver and were conducted with version 22.11.
Detailed instructions for building the planner can be found in the README.md file in the code except for adding the LP solver support. We provide those in the uploaded file "LPBuildInstructions - Fast Downward Homepage.pdf".

The different versions of the SPhO algorithm can be invoked with command-line arguments as follows:

./fast-downward.py PDDL_TASK --search 
"astar(spho($abstractions,strategy=$strategy,group_heuristics={false,true},group_operators={false,true},tiebreak={false,true}))"

where curly brackets indicate choices.

depending on the experiment $abstractions was replaced with:

• [projections(systematic(2,0,interesting_general))] for sys-1-2 patterns
• [projections(systematic(2,1,interesting_general))] for sys-2 patterns
• [projections(systematic(1,0,interesting_general))] for sys-1 patterns

and $strategy with:

• always always compute the lp
• percent_sa for 100% rule based sensitivity analysis
• exact for exact sensitivity analysis

group_heuristics toggles abstraction grouping, group_operators operator grouping and tiebreak the increase_weight tiebreaking strategy.

Dependencies

The experiment scripts were invoked with Python 3.11.2 and lab (https://github.com/aibasel/lab) version 7.3 and reports.zip contains a full requirements.txt. The exact Python version can be installed through Python management systems like conda or pyenv. We used CPLEX version 22.11.

Benchmarks

The file ipc-benchmarks-optimal-strips-1998-2018.zip contains the STRIPS PDDL benchmarks from sequential optimization tracks of IPC 1998-2018 used for the experiments.

Experiment Data:

The remaining zip files contain the raw experiment data (raw-runs), parsed values, and basic reports (report) for the experiments in the paper.
The code directories and benchmark files have been removed to avoid duplication and to save space.