Code and Data for the IJCAI 2021 paper "Learning Generalized Unsolvability Heuristics for Classical Planning"

Classifiers

The three classifiers in this paper are in the following files.

• File 'perfect-classifier.tar.gz' contains the perfect classifier.
• File 'feature-generator-and-safe-classifier.tar.gz' contains the safe classifier.
• File 'decision-trees.tar.gz' contains the decision-tree classifier.

The input to each classifier is a feature file, which is described next.

Experiment Data

The data used to produce features in this paper is generated in three steps, in the following order.

• File 'train-test-pddl.tar.gz' contains the PDDL files used to sample states from.
• File 'train-test-states.tar.gz' contains the states samples from the PDDL files. States are sampled by software in file 'state-space-expansion.tar.gz'.
• File 'train-features.tar.gz' contains the features generated from the sampled states. Features are generated by software in  file 'feature-generator-and-safe-classifier.tar.gz'.

Benchmarks

The file 'modified-fast-downward.tar.gz' contains a modified version of Fast Downward planning system (https://fast-downward.org). The benchmarks computed with it is in the file 'baselines.tar.gz'.