Please help build JaxMARL into the best possible tool for the MARL community.
We actively welcome your contributions!
- If adding an environment or algorithm, check with us that it is the right fit for the repo.
- Fork the repo and create your branch from main.
- Add tests, or show proof that the environment/algorithm works. The exact requirements are listed below.
- Add a README explaining your environment/algorithm.
Environment Requirements
- Unit tests (in
pytestformat) demonstrating correctness. If applicable, show correspondence to existing implementations. If transitions match, write a unit test to demonstrate this (example). - Training results for IPPO and MAPPO over 20 seeds, with configuration files saved to
baselines.
Algorithm Requirements
- Performance results on at least 3 environments (e.g. SMAX, MABrax & Overcooked) with at least 20 seeds per result.
- If applicable, compare performance results to existing implementations to demonstrate correctness.
We use Github's issues to track bugs, just open a new issue! Great Bug Reports tend to have:
- A quick summary and/or background
- Steps to reproduce (Be specific and give example code if you can)
- What you expected would happen
- What actually happens
- Notes (possibly including why you think this might be happening, or stuff you tried that didn't work)
All contributions will fall under the project's original license.
Some improvements we would like to see implemented:
- improved RNN implementations. In the current implementation, the hidden size is dependent on "NUM_STEPS", it should be made independent.
- S5 RNN architecture.