Planning with Perspectives -- Decomposing Epistemic Planning using Functional STRIPS
Authors: Guang Hu, Tim Miller, Nir Lipovetzky
JAIR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We ran evaluations on well-known epistemic planning benchmarks to compare an existing state-of-the-art planner, and on new scenarios that demonstrate the expressiveness of the PWP approach. The results show that our PWP planner scales significantly better than the state-of-the-art planner that we compared against, and can express problems more succinctly. |
| Researcher Affiliation | Academia | Guang Hu EMAIL Tim Miller EMAIL Nir Lipovetzky EMAIL School of Computing and Information Systems The University of Melbourne, Australia |
| Pseudocode | No | The paper provides concrete code examples in Appendix 7.1 (e.g., C++ functions for 'sees' in different domains), but it does not include structured pseudocode or algorithm blocks in a language-agnostic format within the main text. |
| Open Source Code | Yes | The source code of our implementation along with all experiments can be found at https://github.com/guanghuhappysf128/benchmarks. |
| Open Datasets | Yes | We ran evaluations on well-known epistemic planning benchmarks to compare an existing state-of-the-art planner, and on new scenarios that demonstrate the expressiveness of the PWP approach. We implement three benchmarks problems, Corridor (Kominis & Geffner, 2015), Gossip (Cooper et al., 2016), and Grapevine (Muise et al., 2015a) to compare the computational performance, and model two new domains, Big Brother Logic and Social-media Network, to examine the expressiveness of our model. ... our experiments, we compare these three methods state-based, action-based (full) and action-based (relative) with the baseline of Cooper et al. (2019) s method using their tool for generating their Gossip Generator8. Their generator compiles Gossip problems into classical planning problems... 8. Downloadable from https://github.com/Faustine Maffre/Gossip Problem-PDDL-generator |
| Dataset Splits | No | The paper describes using various planning benchmarks (Corridor, Grapevine, Gossip) and custom domains (BBL, Social-media Network) for evaluation. It varies parameters like the number of agents, depth of queries, and number of goals to test scalability. However, it does not mention specific training, validation, or test dataset splits in the context of typical machine learning experimentation. |
| Hardware Specification | Yes | We ran the problems with both planners on a Linux machine with 8 CPUs (Intel Core i7-7700K CPU @ 4.20GHz 8) and 16 gigabytes of memory. |
| Software Dependencies | No | We use F-STRIPS as the modelling language, and BFWS(R0) (Franc es et al., 2017) as the planner. ... In our implementation, external functions are programmed in C++ for scalability and flexibility. ... in this domain, there was not a significant performance difference with respect to the original planner used by Muise et al. (2015a), the FF planner. ... The planner is available through https://github.com/aig-upf/2017-planning-with-simulators. We used options --driver sbfws --options bfws.rs=none |
| Experiment Setup | No | The paper describes the general approach to encoding problems in F-STRIPS and the use of external functions. It provides goal conditions and example plans for various scenarios. However, it does not explicitly provide hyperparameter values, specific optimizer settings, learning rates, or other detailed system-level training configurations typically found in experimental setups for machine learning or complex simulations. |