Learning Generalized Unsolvability Heuristics for Classical Planning
Authors: Simon Ståhlberg, Guillem Francès, Jendrik Seipp
IJCAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our empirical results show that our approach often captures important classes of unsolvable states with high classification accuracy. Additionally, the logical form of our heuristics makes them easy to interpret and reason about, and can be used to show that the characterizations learned in some domains capture exactly all unsolvable states of the domain. |
| Researcher Affiliation | Academia | Simon St ahlberg1, Guillem Franc es2 and Jendrik Seipp1 1Link oping University, Sweden 2Universitat Pompeu Fabra, Spain |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | All source code, benchmarks and datasets are available online.1 [1https://doi.org/10.5281/zenodo.4740386] |
| Open Datasets | Yes | Table 1 provides an overview of the datasets. For each domain we choose at most 10K states from the state spaces of the training set tasks to train on. All source code, benchmarks and datasets are available online.1 [1https://doi.org/10.5281/zenodo.4740386] |
| Dataset Splits | Yes | We partition them into training and test sets such that both sets have roughly the same size, both contain instances with small and large state spaces, and in both sets each generator parameter has at least two different values. For each domain, we choose the tree that performs best on the training set (F1 score) in a 10-fold crossvalidation that evaluates different combinations of maximum tree depth (1, . . . , 10) and maximum feature complexity (1, . . . , maxf F K(f)). |
| Hardware Specification | No | The paper mentions using 'resources from the Swedish National Infrastructure for Computing (SNIC)' but does not provide specific hardware details such as CPU/GPU models, memory, or other specifications. |
| Software Dependencies | No | The paper mentions using 'Open-WBO Weighted Max-SAT solver [Martins et al., 2014]' and 'Scikit-learn machine learning library [Pedregosa et al., 2011]'. However, it only provides publication years as citations and not specific software version numbers. |
| Experiment Setup | Yes | For all domains, we limit the feature complexity by k=16 and the number of concepts by n=80K in the feature generation step (see Section 3.3). We give each method a maximum of five hours to learn an unsolvability heuristic. |