Multi-agent Continual Coordination via Progressive Task Contextualization

This repository contains implementation for Multi-agent Continual Coordination via Progressive Task Contextualization (MACPro).

Environment Installation

Build the environment by running:

pip install -r requirements.txt

or

conda env create -f environment.yaml

Install the Level Based Foraging (LBF) environment by running:

pip install -e src/envs/lb-foraging

Install the Predator-Prey (PP) environment by running:

pip install -e src/envs/mpe/multi_agent_particle

Install the StarCraft Multi-Agent Challenge (SMAC) environment by running:

pip install -e src/envs/smac

Run an experiment

python src/main.py --config=[Algorithm name] --task-config=[Benchmark name]

The config files act as defaults for an algorithm or benchmark. They are all located in src/config. --config refers to the config files in src/config/algs including MACPro and Finetuning. --task-config refers to the config files in src/config/tasks, including lbf as the LB-Foraging benchmark (https://github.com/semitable/lb-foraging), pp as the Predator-Prey benchmark (https://github.com/openai/multiagent-particle-envs), and marines, sz as the StarCraft Multi-Agent Challenge benchmark (https://github.com/oxwhirl/smac).

All results will be stored in the results folder.

For example, run MACPro on PP benchmark:

python src/main.py --config=qmix_macpro_rnn --task-config=pp

Run Finetuning on Marines benchmark:

python src/main.py --config=qmix_attn --task-config=marines

Publication

If you find this repository useful, please cite our paper:

@article{macpro,
  title   = {Multi-agent Continual Coordination via Progressive Task Contextualization},
  author  = {Lei Yuan and Lihe Li and Ziqian Zhang and Fuxiang Zhang and Cong Guan and Yang Yu},
  journal = {IEEE Transactions on Neural Networks and Learning Systems},
  year    = {2024}
}