Landmark Heuristics for Lifted Classical Planning

Authors: Julia Wichlacz, Daniel Höller, Jörg Hoffmann

IJCAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments show that (while only using LMs is less effective) this combination outperforms the state of the art in lifted planning, making landmark heuristics a valuable tool in this area. We implemented our heuristics in the Power Lifted (PWL) planner [Corrˆea et al., 2020]2. All configurations use greedy best first search (GBFS). We compare against goal counting (h GC), the lifted add heuristic with (h Laddpo) and without (h Ladd) preferred operators [Corrˆea et al., 2021] and the best-performing configurations of Lauer et al. [2021], goal count using unary relaxation heuristic as tiebreaker, with (h GC,ur-d) and without (h GC,ur) disambiguation of static predicates. Table 1 gives the coverage results.
Researcher Affiliation Academia Julia Wichlacz , Daniel H oller and J org Hoffmann Saarland University, Saarland Informatics Campus, Saarbr ucken, Germany {wichlacz, hoeller, hoffmann}@cs.uni-saarland.de
Pseudocode No The paper describes algorithms and methods in textual form and with logical steps but does not include any dedicated pseudocode blocks or sections explicitly labeled as "Algorithm".
Open Source Code Yes The code can be found at: https://github.com/minecraft-saar/powerlifted
Open Datasets No The paper states: "We use the benchmark set used by Lauer et al. [2021] that was introduced to evaluate lifted planners." While it references an established benchmark and paper, it does not provide concrete access information (e.g., URL, DOI) to the specific dataset itself.
Dataset Splits No The paper refers to a "benchmark set" and specific "domains" for evaluation, but does not provide explicit details regarding training, validation, or test dataset splits (e.g., percentages, sample counts, or specific split files).
Hardware Specification Yes The experiments were run on a cluster of machines with Intel Xeon E5-2650 CPUs with a clock speed of 2.30GHz.Timeout and memory limits were set to 30 min and 4GB respectively for all runs.
Software Dependencies No The paper mentions software like the "Power Lifted (PWL) planner" and the "Fast Downward (FD) system" but does not specify version numbers for these or any other software components, libraries, or programming languages used for reproducibility.
Experiment Setup Yes All configurations use greedy best first search (GBFS). Timeout and memory limits were set to 30 min and 4GB respectively for all runs. For our system we included 8 configurations: We combine necessary subgoal landmarks with GBFS to the configurations h LNS and h LNS; where the former uses landmark ordering information and the latter does not. The same is done for the FAM landmarks, denoted h LFAM and h LFAM. Then, we also use the LAMA-style search instead of GBFS, resulting in the configurations h LNS LAMA, h LNS LAMA, h LFAM LAMA, and h LFAM LAMA .