Learning Complex Teamwork Tasks Using a Given Sub-task Decomposition

This repository contains code for the AAMAS 2024 paper "Learning Complex Teamwork Tasks Using a Given Sub-task Decomposition" by Fosong et. al.

Requirements

To install dependencies, run

pip install -e .

QMIX experiments were performed using EPyMARL. You may wish to use this fork to run VMAS efficiently via epymarl.

Instructions

To run code, try

python scripts/simple_env_runner.py -cn CONFIG_NAME

The codebase uses hydra for config management. Config files can be found in the configs directory.

Loading pre-trained agents from source tasks is done via an 'agent zoo' interface. An 'agent zoo' is a directory containing config files of agents, as well as saved models and experience buffers. Each agent in the zoo has a name, by which it can be referred. See configs/cooking_medoe.yaml for an example of a config file which loads zoo agents. We provide the agent zoos used in this paper in our data upload.

Results and Data Sharing

Experimental data, agent zoo directories, and plotting scripts can be found via the University of Edinburgh's DataShare service.

Citation

@inproceedings{fosongLearningComplexTeamwork2024,
  title = {Learning Complex Teamwork Tasks Using a Given Sub-task Decomposition},
  booktitle = {Proceedings of the 23rd International Conference on Autonomous Agents and Multiagent Systems},
  author = {Fosong, Elliot and Rahman, Arrasy and Carlucho, Ignacio and Albrecht, Stefano V.},
  year = {2024},
  address = {Auckland, New Zealand},
}