Explaining Soft-Goal Conflicts through Constraint Relaxations
Authors: Rebecca Eifler, Jeremy Frank, Jörg Hoffmann
IJCAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on 4 resource-centric domains and 3 domains with time windows show that both algorithms perform significantly better than the baseline in practice, and that they are complementary to each other in terms of finding explanations on resource- and time-centric domains. We implemented both algorithms in the Fast Downward planning system [Helmert, 2006], extending the code base of Eifler et al. [2020b] and using hmax as a base heuristic for the pruning function1. The experiments were run on Intel E52660 machines running at 2.20 GHz, with a time (memory) limit of 2h (4GB) per benchmark instance. |
| Researcher Affiliation | Academia | 1Saarland University, Saarland Informatics Campus, Germany 2NASA Ames Research Center, Mountain View, CA, USA 3German Research Center for Artificial Intelligence (DFKI), Saarbr ucken, Germany |
| Pseudocode | Yes | Algorithm 1 Internal Constraint Reuse (ICR) and Algorithm 2 Search Space Reuse (SSR) |
| Open Source Code | Yes | 1The source code and the benchmark are available at: https:// github.com/XPP-explainable-planning |
| Open Datasets | Yes | Our benchmark consists of 4 resource-constraint domains (Blocksworld, No Mystery, Rovers R, TPP) and 3 domains (Parent s Afternoon, Rovers T, Satellite) with time constraints. The former part builds on the resource constraint benchmark by Eifler et al. [2020b]. |
| Dataset Splits | No | The paper uses benchmark instances and evaluates coverage, but does not specify training, validation, or test dataset splits in the conventional sense (e.g., percentages or counts). |
| Hardware Specification | Yes | The experiments were run on Intel E52660 machines running at 2.20 GHz, with a time (memory) limit of 2h (4GB) per benchmark instance. |
| Software Dependencies | No | The paper mentions implementation in the 'Fast Downward planning system' but does not provide specific version numbers for this system or any other software dependencies. |
| Experiment Setup | Yes | In each instance there are two individual resources R. For each resource ρ R we generated one benchmark instance, scaling initρ between 0 and two times the initial value in the original instance. The experiments were run on Intel E52660 machines running at 2.20 GHz, with a time (memory) limit of 2h (4GB) per benchmark instance. |