Symbolic Search for Optimal Total-Order HTN Planning

Authors: Gregor Behnke, David Speck11744-11754

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental An empirical analysis shows that our symbolic approach outperforms the current state of the art for optimal totally-ordered HTN planning. We conclude with an empirical evaluation. Tab. 1 shows the overall coverage of all five planners on the benchmark instances. Figure 4 shows the coverage over time. Figure 5 shows a per-instance runtime comparison between the two best planners overall, auto Sym and SAT(bin).
Researcher Affiliation Academia Gregor Behnke and David Speck University of Freiburg Freiburg im Breisgau, Germany behnkeg, speckd @informatik.uni-freiburg.de
Pseudocode Yes Algorithm 1: Overall BDD-based Algorithm Algorithm 2: Adding state pairs Algorithm 3: Primitive actions and empty methods Algorithm 4: Handling methods with two subtasks
Open Source Code Yes We implemented the presented approach auto Sym and empirically compared it with other optimal (TO)HTN planners.4 auto Sym is based on the PANDA planning framework in its C++ version panda PI (H oller et al. 2021). Since auto Sym operates on a grounded model, we use parser and grounder of panda PI (Behnke et al. 2020).5 4https://github.com/galvusdamor/panda PIengine Symbolic 5https://github.com/panda-planner-dev/panda PIparser and https://github.com/panda-planner-dev/panda PIgrounder
Open Datasets No The paper mentions collecting domains from former evaluations of HTN planners and listing them (e.g., Barman, Blocks-GTOHP, Childsnack, etc.) with the total number of instances. However, it does not provide explicit access information (links, DOIs, citations with author/year) for these datasets to indicate public availability.
Dataset Splits No The paper describes using a set of benchmark instances for evaluation but does not specify any training, validation, or test splits. It focuses on the number of instances solved by each planner.
Hardware Specification Yes Every planner was given 4 GB of RAM and 30 minutes of runtime on a compute cluster with nodes equipped with two Intel Xeon Gold 6242 32-core CPUs, 20 MB cache and 188 GB shared RAM running Ubuntu 18.04 LTS 64 bit.
Software Dependencies Yes auto Sym is based on the PANDA planning framework in its C++ version panda PI (H oller et al. 2021). To represent and manipulate BDDs we use CUDD 3.0.0 (Somenzi 2015).
Experiment Setup Yes Every planner was given 4 GB of RAM and 30 minutes of runtime on a compute cluster with nodes equipped with two Intel Xeon Gold 6242 32-core CPUs, 20 MB cache and 188 GB shared RAM running Ubuntu 18.04 LTS 64 bit. auto Sym uses the default panda PI variable ordering for the BDDs.