Merge-and-Shrink Task Reformulation for Classical Planning
Authors: Álvaro Torralba, Silvan Sievers
IJCAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 6 Experiments We implemented the M&S reformulation framework in Fast Downward (FD) [Helmert, 2006b], using its existing M&S framework [Sievers, 2018] and extending it with weak bisimulation as well as pruning transformations that remove dead labels and irrelevant TSs and labels. [...] 6.1 Search Space Reduction To assess the impact of our task reformulations on the reachable state space, we run uniform-cost search and evaluate the number of expansions until the last f-layer. Fig. 2 compares the FDR representation against a-ls and d-ls with bisimulation (top) and weak bisimulation (bottom) shrinking. |
| Researcher Affiliation | Academia | 1Saarland University, Saarland Informatics Campus, Germany 2University of Basel, Switzerland |
| Pseudocode | No | The paper describes algorithms but does not provide pseudocode or a clearly labeled algorithm block. |
| Open Source Code | Yes | Implementation: https://doi.org/10.5281/zenodo.3232878, dataset with benchmarks: https://doi.org/10.5281/zenodo.3232844. |
| Open Datasets | Yes | We use all STRIPS benchmarks from the optimal/satisficing tracks of all IPCs, two sets consisting of 1827/1816 tasks across 48 unique domains.1 Implementation: https://doi.org/10.5281/zenodo.3232878, dataset with benchmarks: https://doi.org/10.5281/zenodo.3232844. |
| Dataset Splits | No | The paper evaluates on 'STRIPS benchmarks from the optimal/satisficing tracks of all IPCs' which are pre-defined planning tasks, not a single dataset split into train/validation/test sets for model training. Therefore, explicit dataset splits are not provided. |
| Hardware Specification | No | The paper does not provide specific details on the hardware used to run the experiments. |
| Software Dependencies | No | The paper mentions 'Fast Downward (FD) [Helmert, 2006b]' and 'its existing M&S framework [Sievers, 2018]', but does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | We impose a time limit of 900s on the reformulation process. For the overall planning, we use a limit of 3.5 Gi B and 1800s. We consider DFP (d-ls) and sb MIASM (m-ls, called dyn-MIASM originally) [Sievers et al., 2014], with a size limit of 1000 on the resulting product. |