Code, Data and Models for the AAAI 2025 paper "Learning More Expressive General Policies for Classical Planning Domains"

Code

The file 'code.zip' contains the source code used for training and testing. To train new or use existing models, we recommend compiling the source code using the included Apptainer recipes. See README.md for more information on how to use them. To train a model with the same configuration and hyperparameters we used in the paper, we suggest opening the '*.hparams' file in a text editor: it is just a json file and contains the arguments used to initialize the model. In general, each hyperparameter corresponds to an argument to the training program. For example:

• "pair_embeddings": true must be enabled to train the more expressive models, and
• "composition_depth": 0 is the t parameter in the paper.

Some arguments override others, e.g. "all_compositions": true is used for the baseline R-GNN_2, to use all possible compositions rather than those based on t.

Data

The file 'data.zip' contains all training and test instances used in the paper.

Models

The file 'models.zip' contains all the models that are used in the paper.