Automatic Selection of Macro-Events for Heuristic-Search Temporal Planning

Authors: Alessandro La Farciola, Alessandro Valentini, Andrea Micheli

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we experimentally demonstrate that the proposed approach yields a substantial performance improvement for a state-of-the-art temporal planner.
Researcher Affiliation Academia Fondazione Bruno Kessler, Trento, Italy EMAIL, EMAIL, EMAIL
Pseudocode Yes Algorithm 1 Prototypical Heuristic-Search Algorithm for Temporal Planning (with Macro-Events)
Open Source Code Yes Benchmarks, code and all the scatter plots are available at https://github.com/fbk-pso/step-rl.
Open Datasets Yes We considered the two benchmark domains in (Micheli and Valentini 2021) (enlarging the problem instances to make the problems more challenging).
Dataset Splits Yes For each domain, we perform a 4-fold splitting of the instances using the training part to construct the plan-macroevents and the testing part to evaluate our planner equipped with the selected set of macros against the baseline TAMER planner without macros.
Hardware Specification Yes We run our experiments on a server with 4 AMD EPYC 7413 processors and 528GB of RAM
Software Dependencies No The paper mentions 'Python' and the 'Unified Planning library' but does not provide specific version numbers for these software components. It also mentions TAMER without a version.
Experiment Setup Yes We chose as maximum length of candidates macro-events ℓmax = 5. The size of the extracted CME is 23619 for Kitting and 61378 for MAJSP. The average (wrt the 4-folds) time of execution for macros selection is 5s (Kitting) and 13s (MAJSP) in case FA-, 54s (Kitting) and 355s (MAJSP) in case PA-; in cases with intermediate nodes we allocated 30 minutes for the anytime algorithm. We run our experiments on a server with 4 AMD EPYC 7413 processors and 528GB of RAM, we allocated 4 cores for each planning run with a timeout of 600s and a memory limit of 40GB (the maximum memory used in the experiments was 6.4GB).