Author: Rémy Chaput
(original idea and code by Benoît Alcaraz, Rémy Chaput, Christopher Leturc, and Olivier Boissier)
AJAR (Argumentation-based Judging Agents framework for ethical Reinforcement Learning) is a Python library to judge the behaviours of Reinforcement Learning (RL) agents, based on moral values that are represented as argumentation graphs.
Clone this repository or install it through
pip install git+https://github.com/ethicsai/ajar.git, then open a Python
shell (3.7+), and execute the following code:
from ajar import AFDM, JudgingAgent, Argument, judgment
decision = 'moral'
afdm = AFDM()
afdm.add_argument(Argument(
'argument_name',
'An argument description',
support=[decision]
))
afdm.add_argument(Argument(
'other_argument',
'Another argument',
counter=[decision]
))
afdm.add_attack_relation('argument_name', 'other_argument')
judge = JudgingAgent('My moral value', afdm, judgment.j_simple)
reward = judge.judge(situation={}, decision=decision)This will create a new graph (afdm), populate it with 2 arguments
(argument_name and other_argument), as well as an attack from the first
argument onto the second. Finally, we create a judge that represents a given
moral value by using this afdm. The judge can then be used to judge a
situation by calling judge.judge({}, decision). It will return a scalar
number that corresponds to the extent that the given situation respects the
moral value.
To go further, please refer to the documentation.
This project follows the Semver (Semantic Versioning): all versions respect
the <major>.<minor>.<patch> format. The patch number is increased when a
bugfix is released. The minor number is increased when new features are added
that do not break the code public API, i.e., it is compatible with the
previous minor version. Finally, the major number is increased when a breaking
change is introduced; an important distinction is that such a change may not
be "important" in terms of lines of code, or number of features modified.
Simply changing a function's return type can be considered a breaking change
in the public API, and thus worthy of a "major" update.
The community guidelines are available in the CONTRIBUTING.md file; you can find a (short) summary below.
If you have a question (something that is not clear, how to get a specific result, ...), do not hesitate to create a new Discussion under the Q&A category.
Please do not use the issue tracker for support, to avoid cluttering it.
If you found a bug (an error raised, or something not working as expected), you can report it on the Issue Tracker.
Please try to be as precise as possible.
We very much welcome and appreciate contributions!
For fixing bugs, or improving the documentation, you can create a Pull Request.
New features are also welcome, but larger features should be discussed first in a new Discussion under the Ideas category.
All ideas, suggestions, and requests are also welcome for discussion.
The source code is licensed under the MIT License (see also the LICENSE file in this repository).
If you use this package in your research, please cite the corresponding paper:
Benoît Alcaraz, Olivier Boissier, Rémy Chaput, and Christopher Leturc. 2023. AJAR: An Argumentation-based Judging Agents Framework for Ethical Reinforcement Learning. In Proceedings of the 2023 International Conference on Autonomous Agents and Multiagent Systems (AAMAS '23). International Foundation for Autonomous Agents and Multiagent Systems, Richland, SC, 2427–2429. DOI: 10.5555/3545946.3598956
@inproceedings{10.5555/3545946.3598956,
author = {Alcaraz, Beno\^{\i}t and Boissier, Olivier and Chaput, R\'{e}my and Leturc, Christopher},
title = {AJAR: An Argumentation-Based Judging Agents Framework for Ethical Reinforcement Learning},
year = {2023},
isbn = {9781450394321},
publisher = {International Foundation for Autonomous Agents and Multiagent Systems},
address = {Richland, SC},
booktitle = {Proceedings of the 2023 International Conference on Autonomous Agents and Multiagent Systems},
pages = {2427–2429},
numpages = {3},
location = {London, United Kingdom},
series = {AAMAS '23}
}