[DO NOT CLOSE] Library TODOs and call for contributions - Githubissues

facebookresearch / BenchMARL

A collection of MARL benchmarks based on TorchRL

https://benchmarl.readthedocs.io/

MIT License

191 stars 23 forks source link

[DO NOT CLOSE] Library TODOs and call for contributions #94

Open matteobettini opened 1 month ago

matteobettini commented 1 month ago

Hello people!

In this issue I will list the things I would really like to have in BenchMARL and will tick them off as they are implemented!

It is also a really good place to find something you would like to contribute.

Features

[ ] Parameter sharing between actors and critics (started in #95, still need to implement sharing in PPO)
[ ] support for variable number of agents (dead agents) (#103)
[ ] support for turn-based environments (#76)
[ ] Parallel collection for non-vectorized environments
[x] Improve checkpointing (just keep last checkpoints) (#102)
[ ] Prioritized replay buffers

Models

[ ] RNN models (https://github.com/pytorch/rl/pull/1948) (#103)
[x] DeepSets model (#96)

Algorithms

[ ] MPO and V-MPO
[ ] HARL algorithms (#52)

Environments

[ ] Unity ML-Agents (https://github.com/pytorch/rl/pull/2161)