Faster Stackelberg Planning via Symbolic Search and Information Sharing
Authors: Álvaro Torralba, Patrick Speicher, Robert Künnemann, Marcel Steinmetz, Jörg Hoffmann11998-12006
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our empirical evaluation shows that SLS outperforms previous approaches for Stackelberg planning. The improvement is consistent, and is particularly pronounced in tasks with large leader action spaces and in net-benefit Stackelberg planning where follower subtasks are inherently harder to solve. |
| Researcher Affiliation | Academia | 1 Aalborg University, Denmark 2 CISPA Helmholtz Center for Information Security, Germany 3 Saarland University, Saarland Informatics Campus, Germany alto@cs.aau.dk, {patrick.speicher,robert.kuennemann}@cispa.saarland, {steinmetz,hoffmann}@cs.uni-saarland.de |
| Pseudocode | Yes | Algorithm 1: Symbolic Leader Search (SLS) and Algorithm 2: Regress Plan |
| Open Source Code | Yes | Our source code, benchmarks, and results are publicly available (Torralba et al. 2021). https://doi.org/10.5281/zenodo.4320574. |
| Open Datasets | Yes | Our source code, benchmarks, and results are publicly available (Torralba et al. 2021). https://doi.org/10.5281/zenodo.4320574. It contains instances of classical planning IPC domains, but extended with n leader actions that disable preconditions of some follower actions. |
| Dataset Splits | No | No explicit training, validation, or test dataset split percentages or counts are provided. The paper discusses using instances from benchmark domains for evaluation. |
| Hardware Specification | Yes | We ran our experiments with the Downward Lab toolkit (Seipp et al. 2017) on an Intel Xeon CPU E5-2650 v3, 2.30 GHz. |
| Software Dependencies | No | We implemented our new SLS algorithm on top of the Stackelberg framework by Speicher et al. (2018a), built upon the Fast Downward planning system (Helmert 2006). We ran our experiments with the Downward Lab toolkit (Seipp et al. 2017)... (No specific version numbers are provided for Fast Downward or Downward Lab toolkit). |
| Experiment Setup | Yes | For each task, we set a 30 minute time limit and a 4 GB memory limit, ignoring the translate and preprocessing phase which is equal for all configurations. ... To aggressively look for a solution below the cost bound B, we use a greedy best-first search with the FF heuristic (Hoffmann and Nebel 2001) and a time limit of 1s... |