bounded-prescience

This repo contains code for the paper Shielding Atari Games with Bounded Prescience (https://arxiv.org/abs/2101.08153) by Mirco Giacobbe, Hosein Hasanbeig, Hjalmar Wijk and Daniel Kroening. In addition to replicating the experiments in the paper and the bounded prescience shield, the repo allows the creation of property-labelled versions of the standard Atari gym environments which can expose safety information for other agents and policies.

Installation

The prescience package needs to be installed through pip:

git clone https://github.com/HjalmarWijk/bounded-prescience.git
cd bounded-prescience
pip3 install .

or alternatively:

pip3 install git+https://github.com/HjalmarWijk/bounded-prescience.git

Properties

* See manual for details on this game rule.

To test properties use the test_property.py script with flags --env [Game name] --prop [Property name]

This simulates a random agent and logs violations. To evaluate properties with human play, use --human flag (requires pygame).

Verification

To check properties for the 9 pre-trained agents evaluated in the paper under a variety of settings see the script check_noops.py To run the ChainerRL agents you first need to download them by running download_models.py (the Atari Zoo agents download dynamically). The scripts chainer_no_shield.sh and atari_zoo_no_shield.sh run all the agents for all propertiesm and write results as a csv in the results folder. Note that Atari Zoo agents need Tensorflow 1 (and AtariZoo) installed, while ChainerRL agents needs Tensorflow 2 and ChainerRL.

Shielding

To check properties using prescience shielding use the --lookahead flag for check_noops.py The shield scripts run this check for all algorithms, properties and shield depths up to 5.

Reference

Please use this bibtex entry if you want to cite this repository in your publication:

@misc{bounded_prescience_repo,
  author = {Mirco Giacobbe, Mohammadhosein Hasanbeig, Daniel Kroening, and Hjalmar Wijk},
  title = {Shielding Atari Games with Bounded Prescience Code Repository},
  year = {2020},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/HjalmarWijk/bounded-prescience}},
}

License

This project is licensed under the terms of the BSD-3-Clause