Symbolic Search for Cost-Optimal Planning with Expressive Model Extensions
Authors: David Speck, Jendrik Seipp, Alvaro Torralba
JAIR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, our empirical evaluations demonstrate that the presented symbolic search algorithms complement and frequently show superior performance compared to other planning approaches from the literature. This holds across various planning domains, for each of the model extensions individually and in combination. Each of these sections contains a theoretical analysis of symbolic search with the extensions and an empirical comparison to other state-of-the-art techniques. |
| Researcher Affiliation | Academia | David Speck EMAIL University of Basel, Switzerland; Jendrik Seipp EMAIL Link oping University, Sweden; Alvaro Torralba EMAIL Aalborg University, Denmark |
| Pseudocode | Yes | Algorithm 1: Symbolic BDD forward search with non-zero operator costs. Algorithm 2: Plan reconstruction for symbolic forward search. Algorithm 3: Symbolic EVMDD forward search with non-zero operator costs. Algorithm 4: Axiom evaluation for explicit states (Helmert, 2008). |
| Open Source Code | Yes | To foster more research on model extensions, we make the benchmarks, the baseline planners and our code available online (Speck, Seipp, & Torralba, 2024). |
| Open Datasets | Yes | To foster more research on model extensions, we make the benchmarks, the baseline planners and our code available online (Speck, Seipp, & Torralba, 2024). |
| Dataset Splits | No | The paper discusses benchmark tasks and domains, and reports the number of solved instances within each domain, but does not provide specific training/test/validation dataset splits or methodologies. |
| Hardware Specification | Yes | For each planner run, we allocate 30 minutes and a memory limit of 8 GiB and use Downward Lab (Seipp, Pommerening, Sievers, & Helmert, 2017) to run our experiments on Intel Xeon Gold 6130 CPUs. |
| Software Dependencies | Yes | For the BDD representation we use the CUDD library (Somenzi, 2015) and for the EVMDD representation we use MEDDLY (Babar & Miner, 2010). Both planners are based on Fast Downward (Helmert, 2006). CUDD: CU decision diagram package release 3.0.0. https://github.com/ivmai/cudd. |
| Experiment Setup | Yes | For each planner run, we allocate 30 minutes and a memory limit of 8 GiB and use Downward Lab (Seipp, Pommerening, Sievers, & Helmert, 2017) to run our experiments on Intel Xeon Gold 6130 CPUs. All our code, the code for other planners, the benchmarks, and the experiment data are available online (Speck et al., 2024). In bidirectional search, we impose initial time and BDD size limits of 1 minute and 10 million nodes per search step in both directions, doubling these limits if both directions exceed the limits in the current step. |