moratodpg/imp_marl

IMP-MARL: a Suite of Environments for Large-scale Infrastructure Management Planning via MARL

imp_marl's Issues

Possibility to access trained models

Hi,

I like the implementation of IMP-MARL and that you made the training data of your runs accessible! Would it be possible to give access to the trained models as well? This would alleviate the need to retrain the models and make reproduction of the results very easy.

Yours
Yannik

old branches

Hey Pablo,

I think we can delete the old plot branch.

What about your heuristic branch?

Cheers

Clean RL

Clean RL is not a modular library but focus on a single-file implementation of RL aglos (see their README).

Right now they do not support Gymnasium but they support olg Gym.

One only has to modify their single-file implementations to register imp_marl in the avaialble gym environments.
Kind of:

from imp_marl.imp_wrappers.gymnasium import gym_wrap_sa_struct
gym.register("imp_marl", entry_point=gym_wrap_sa_struct.GymSaStruct)

Where do you want we mention that in our repo?

Shall we include a normalized timestep in the state info for only one or all components?

https://github.com/moratodpg/marl_struct/blob/831de86f5b4ed703f83dae67cc39d31e9849c29a/struct_env/pymarl_ma_struct.py#L136

Let's confirm this and, if necessary, adapt it.