MARL - Agent Abstraction

This repository contains the source code used for our experiments on agent abstraction in multi-agent reinforcement learning. It includes, but is not limited to, the code necessary for reproducing the plots from our ICLR 2022 workshop paper, Summarizing Societies: Agent Abstraction in Multi-Agent Reinforcement Learning.

Mila Cluster First Time Setup:

If you are in the interactive mode:

source /etc/profile

Then:

module load python/3.7 pytorch/1.8.1
cd ~; git clone https://github.com/amimem/marl.git
cd marl; python -m venv ./env/marl_env
source ./env/marl_env/bin/activate
pip install --no-index -r requirements.txt

Once the setup is done, for future use just:

module load python/3.7 pytorch/1.8.1
source ~/marl/.env/marl_env/bin/activate

And run the script that you want:

python ray/playground.py

Compute Canada Setup:

The code does not work on CC due to dependency and package build issues.

Script Help:

For now ray includes all the scripts; main experiments are in playgorund.py, once the reliablity and usability of data is established, we can move the setting to separate files (DQN.py, PPO.py for example).

magent_wrappers.py includes the custom wrappers we wrote for RLLib + PettingZoo. Only the AEC (non-parallel) wrapper is tested.