Attributes docstrings

[TimotheeMathieu]

For now there is practically no attribute documentation except in Agent and AgentWithSimplePolicy. It would be nice to have this for every agent. and every environment also.

[KohlerHECTOR]

PR: #349 : Done with DQN-PPO-REINFORCE.
To do torch: A2C-MDQN.
To do non-torch: all of them.

[KohlerHECTOR]

To do torch: A2C-MDQN.
To do non-torch: all of them.

[riccardodv]

So far, we have put all attributes. These are often just a copy and paste of the parameters, leading to some redundancy. I understand it could be helpful to have both things, but do you think this is the best way to do it? The docstring will be quite long in this way.

[TimotheeMathieu]

Maybe it would be sufficient to document the attributes that are not arguments but created during the fit. A bit like in scikit-learn ;) . in the case of scikit-learn, the attributes that could be reused by the user are denoted with a "" at the end (e.g. regressor.coef ) they say "an attribute that have a meaningful value after fit() was called". Maybe we could take the same convention for rlberry agents ?