Concurrent Planning and Execution Using Dispatch-Dependent Values
Authors: Andrew Coles, Erez Karpas, Eyal Shimony, Shahaf Shperberg, Wheeler Ruml
IJCAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | An experimental evaluation on problems with time pressure shows that the new method significantly outperforms the previous state-of-the-art. An empirical study shows that this new approach achieves state-of-the-art performance in concurrent planning and execution. |
| Researcher Affiliation | Academia | Andrew Coles1 , Erez Karpas2 , Eyal Shimony3 , Shahaf Shperberg3 , Wheeler Ruml4 1King s College London, UK 2Technion, Israel 3Ben-Gurion University, Israel 4University of New Hampshire, USA EMAIL, EMAIL EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes the approach conceptually and formally defines value functions, but it does not present any structured pseudocode or algorithm blocks. |
| Open Source Code | No | Source code freely available from the authors upon request. |
| Open Datasets | Yes | Similarly to previous work on concurrent planning and execution [Coles et al., 2024], we evaluate our techniques on the PDDL encoding of the Robocup Logistics League (RCLL) [Niemueller et al., 2015]. |
| Dataset Splits | No | The paper mentions using "100 RCLL scenarios" and creating "3 versions for each one, with 1, 2, and 3 robots," which describes the experimental instances. However, it does not provide specific training/test/validation dataset splits in the context of machine learning models, as this is a planning paper. |
| Hardware Specification | No | We used values for EPS ranging from 10 expansions per second (a very slow CPU) to 1000 expansions per second (a very fast CPU). |
| Software Dependencies | No | we have implemented it within the planner OPTIC [Coles et al., 2012]. |
| Experiment Setup | Yes | For our evaluation, for the f(i) open list (WA*), we set the W value to 5 the default used in OPTIC. Action dispatch reasoning is performed according to Coles et al. [2024], but without the dispatch reasoning constraining search to expanding a promising subtree, as the divergence between the best states according to f(i) and pf (i) is an alternative to this. The only remaining parameter for the dispatch reasoning is the dispatch threshold, which we will explore within this evaluation. For dispatch threshold, we considered three possible values (0.025, 0.1, 0.25) for both disp and dual. |