PDBs Go Numeric: Pattern-Database Heuristics for Simple Numeric Planning
Authors: Daniel Gnad, Lee-Or Alon, Eyal Weiss, Alexander Shleyfman
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our numeric PDBs by adapting established pattern generation methods and using the canonical heuristic to combine multiple PDBs admissibly. We compare against state-of-the-art optimal planners for simple numeric tasks. Our empirical evaluation confirms that these solutions work well on common numeric planning benchmarks and bring numeric PDBs in reach of the strongest admissible heuristics in simple numeric planning. |
| Researcher Affiliation | Academia | Daniel Gnad1, Lee-or Alon2, Eyal Weiss2,3, Alexander Shleyfman2 1Link oping University, Link oping, Sweden 2Bar-Ilan University, Ramat Gan, Israel 3Technion Israel Institute of Technology, Haifa, Israel EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1: Numeric PDB Construction Input: IRT Π = V, A, s0, G , pattern P, max. # of states N Output: PDB heuristic h PDB(s) |
| Open Source Code | Yes | Code https://github.com/dgnad/numeric-fast-downward |
| Open Datasets | Yes | Datasets https://doi.org/10.5281/zenodo.14502394 We extend it with existing benchmarks from the literature (Scala, Haslum, and Thi ebaux 2016; Scala et al. 2017, 2020; Shleyfman, Kuroiwa, and Beck 2023) and introduce two benchmark sets ourselves, forestfire and minecraft. |
| Dataset Splits | No | The paper does not explicitly detail training/test/validation dataset splits. It describes how it uses benchmark sets for evaluation and samples states for pattern generation, but not explicit train/test splits for the main experimental evaluation. |
| Hardware Specification | Yes | The experiments were conducted on a cluster of Intel Xeon Gold 6130 CPUs using Downward Lab 8.2 (Seipp et al. 2017), adopting the runtime and memory limits of 30 min and 8Gi B from the recent International Planning Competition (IPC) 2023 (Taitler et al. 2024). |
| Software Dependencies | Yes | Our planner is based on the Numeric Fast Downward framework (NFD) (Aldinger and Nebel 2017; Kuroiwa et al. 2022). We adapted the implementation of pattern-database heuristics from Fast Downward (Helmert 2006) to support numeric variables. All experiment data and code is available online (Gnad et al. 2024), and our code is maintained in a public repository.1 The experiments were conducted on a cluster of Intel Xeon Gold 6130 CPUs using Downward Lab 8.2 (Seipp et al. 2017) |
| Experiment Setup | Yes | We impose a limit on the number of abstract states that are generated during the progression phase for numeric PDBs. This limit has a large impact both on the accuracy of the resulting heuristic and the computational cost of constructing the abstract state space. We empirically determined these limits for each pattern generation method. ... For the greedy PDB, we limit the number of abstract states to 100k. For the systematic pattern generation, we compute all interesting patterns with two variables, which performed best in our evaluation, and limit the number of abstract states by 50k. ... We limit the collection size (sum of individual pattern sizes) of i PDB by 10 million abstract states, the hill-climbing time by 15min, and initially sample 1000 states to estimate improvement, similar to the finite-domain variant of i PDB. We limit the number of abstract states to 10k and use 1 million as an upper bound on the domain-size product of variables allowed in a pattern. ... State sampling is done using NFD s default random seed. |