Cost-Partitioned Merge-and-Shrink Heuristics for Optimal Classical Planning

Authors: Silvan Sievers, Florian Pommerening, Thomas Keller, Malte Helmert

IJCAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that SCP significantly improves M&S on standard planning benchmarks. We implemented our techniques on top of the M&S implementation of Fast Downward [Helmert, 2006], version 19.12. To evaluate them, we use the following state-of-the-art M&S configuration... Experiments were run on Intel Xeon Silver 4114 CPUs...
Researcher Affiliation Academia University of Basel, Switzerland {silvan.sievers,florian.pommerening,tho.keller,malte.helmert}@unibas.ch
Pseudocode Yes Algorithm 1 Merge-and-shrink algorithm extended to compute SCP heuristics.
Open Source Code Yes The code, benchmarks, and experimental data are published online [Sievers et al., 2020a]. Sievers et al., 2020a] Silvan Sievers, Florian Pommerening, Thomas Keller, and Malte Helmert. Code, benchmarks and experiment data for the IJCAI 2020 paper Cost-Partitioned Merge-and-Shrink Heuristics for Optimal Classical Planning . https://doi.org/10.5281/zenodo. 3775871, 2020.
Open Datasets Yes As benchmarks, we use the tasks of all optimal tracks of all International Planning Competitions, a set consisting of 1827 tasks across 65 domains. The code, benchmarks, and experimental data are published online [Sievers et al., 2020a].
Dataset Splits No The paper uses a set of planning benchmark tasks but does not describe any specific training, validation, or test dataset splits as is common in machine learning models.
Hardware Specification Yes Experiments were run on Intel Xeon Silver 4114 CPUs
Software Dependencies Yes We implemented our techniques on top of the M&S implementation of Fast Downward [Helmert, 2006], version 19.12.
Experiment Setup Yes To evaluate them, we use the following state-of-the-art M&S configuration: shrinking is based on bisimulation and uses a size limit of 50000 on transition systems (allowing exact shrinking also when shrinking is not necessary to respect the size limit); the merge strategy is SCC-DFP [Sievers et al., 2016]; we use exact label reduction; and the main loop is limited to 900s. Each planner run is limited to 1800s and 3.5 Gi B.