Implement Meta-RL REPTILE algorithm

[zxnga]

Hello !

I would like to propose a meta-RL extension for the repository to be able to use the REPTILE meta-learning algorithm (https://arxiv.org/pdf/1803.02999) along with SB3 algorithms that would act as the inner loop agents.

Motivation:

Currently, there is only few and outdated options for model-free Meta-RL. In particular:

• Many Meta-RL implementations are either research-only or focused on supervised learning experimentations that make them hard to be easily transfered to RL tasks
• MAML was removed from RLlib
• SB3 already has the core inner loop algorithms which simplify the development of some of the meta RL algorithms

Specifically i already implemented an SB3 extension for the REPTILE algorithm for my thesis on a separate repo (https://github.com/zxnga/reptile_sb3) that could also easily be extended to the FOMAML version of the algorithm.

If the idea aligns with the project’s, I’d be happy to open a draft PR as i already have a clean workable solution the reproduces the results of the paper on supervised tasks.

[araffin]

Hello,

similar to DLR-RM/stable-baselines3#1758, Meta RL is out of the scope of SB3.

However, once your implementation is a bit more documented (at least a README explicit what is it and how to run the code), I would be happy to add a link to our documentation (in the project section).