JAIR paper source code & dataset & weights & results

Release v5.0.0 contains a large refactoring on the experiment scripts (command line interface) as well as the module organization.

• Bunch of doman-specific launcher code is moved to latplan/main/[domain].py , and they now have argparse-based docstring that describe the usage.
• Various mixins are moved to latplan/mixins/*.py .
• Output activations, loss functions and output visualization/rendering rules are abstracted into one class. For example, an unactivated output + L2 loss is named GaussianOutput, and sigmoid output + binary cross entropy loss is named ProbabilityOutput. The result partially resembles probabilistic programming.

The attached binaries are:

• cylinders-4-dataset-raw.tar.bz2 contains raw images generated by https://github.com/IBM/photorealistic-blocksworld/tree/jair , from which more compact dataset archives are created.
• datasets.tar contains .npz files storing the dataset.
• backup-propositional.tar.bz2 contains problem instances for planning.
• tables.tar.bz2 contains csv files containing various summary statistics of trained models as well as planning results. The files expands into tables/ directory which already exists in the repository. Running make will generate plots, and also a sqlite database file as a byproduct that makes it easy to investigate the results. (use utilities such as sqlitebrowser)
• samples-*.bz2 contains trained weights for AMA4, kltune (prior = bernoulli(0.1) ) configurations. Due to the file size, we included the top-5 ELBO configurations only.

Trained weights

This release contains the final trained weights used for the planning experiments in the IJCAI20 paper.
We included the ones for the proposed approach (Cube-Space AE trained with all losses) which was used for the planning experiments.
Other weights, such as those for the ablation study, is not included for the file size.

Uncompress the archive in the root of the latplan directory. Each subdirectory contains the intermediate files as follows:

• action+ids.csv, binary representation (text file) of the transitions and their corresponding action labels.
• action_add3+ids.csv / action_del3+ids.csv , the binary effect representations with the corresponding action labels.
• actionlearner4-3-actions_both+ids.csv-0-1.00.pddl --- the PDDL files generated from these binary representations.
• aux.json --- hyperparameter that was selected as a result of hyperparameer tuning. (this complements grid-search.log files which have been already available in the repository)
• net0.h5 --- weight files.
• performance.json --- misc performance (accuracy) records.