Single-file MARL implementations with sample, customizable, discrete grid world environments. Adapted from natashamjacques/multigrid.
NOTE: This project is in progress. Source code is not stable and unintended behavior might emerge while running. Please notify us of issues so that we can fix them
(Help us by adding some!)
- IPPO (Independent Proximal Policy Optimization)
The metacontroller (housed at multiagent_metacontroller.py) is the middle man between the agents and the MARL environment. While we tried to keep its implementation as generic as possible, it might need to be updates as new reinforcement learning methods are added to agent.py
To run, clone this repo and type the following on the command line:
python main.py <arguments>
Available command line arguments are env_name, mode, with_expert, debug, seed, keep_training, visualize, video_dir, load_checkpoint_from, and wandb_project
A Sphinx documentation of the source code has not been created yet. We will update the README with a link to the readthedocs when completed.
The user interface provides a way to train and visualize agents without any code. To run, clone this repo and type the following on the command line:
python gui.py
Feel free to add new implementations or fix any issues by filing a pull request.