DRAMA at the PettingZoo

Dynamically Restricted Action Spaces for
Multi-Agent Reinforcement Learning Frameworks

This repository contains the reference implementation of the DRAMA framework as introduced in Oesterle et al. (2024): DRAMA at the PettingZoo: Dynamically Restricted Action Spaces for Multi-Agent Reinforcement Learning Frameworks. Submitted to HICSS 2024.

Installation

To install the DRAMA library:

$  pip install drama-wrapper

Usage

In analogy to the AEC of PettingZoo

env.reset()
for agent in env.agent_iter():
    observation, reward, termination, truncation, info = env.last()
    action = env.action_space(agent).sample() # this is where you would insert your policy
    env.step(action)

the DRAMA loop can be imported and used as follows:

from drama.restrictors import Restrictor
from drama.wrapper import RestrictionWrapper

env = ...
restrictor = Restrictor(...)
wrapper = RestrictionWrapper(env, restrictor)
policies = {...}

wrapper.reset()
for agent in wrapper.agent_iter():
    observation, reward, termination, truncation, info = wrapper.last()
    action = policies[agent](observation)
    wrapper.step(action)

Please refer to getting-started.ipynb for a first full example.

Documentation

The full documentation of the code can be found here.

Citation

To cite this project in a publication, please use

@misc{oesterle-2023-drama,
    author = {Oesterle, Michael and Grams, Tim},
    title = {DRAMA},
    year = {2023},
    url = {https://github.com/michoest/hicss-2024}
}

or use the CITATION.cff file which is part of the package.