Implementing RNNs into RL algorithms

[dtransposed]

Hey!

First of all thank you for this library!

I would like to take your actors and critics and implement RNN-enhanced TD3 algorithm as described here: https://arxiv.org/pdf/1710.06537.pdf.

I have investigated the source code and it seems that you are not supporting recurrent feature in your implementation of rl algorithms. Have you considered it? Is there anything you can recommend which may possibly help me with seamless transition to memory-based approach?

Cheers!

Edit: The "history len" parameter in the code, would this be something I am looking for?

[dtransposed]

bump

[Scitator]

Hi, sorry for a long answer (look like there are some issues with github notifications)
We are working on Catalyst.RL - second edition on top of new Core concept of the Catalyst, https://github.com/catalyst-team/catalyst. If you are interested in such initiative, please join our slack https://join.slack.com/t/catalyst-team-core/shared_invite/zt-d9miirnn-z86oKDzFMKlMG4fgFdZafw.

[dtransposed]

Hi @Scitator. Thanks for the reply. So I understand this repository is deprecated and all the upcoming questions should be addressed in the general Catalyst repo?

[Scitator]

Yup :)
This repo is good really good example for distributed RL organisation, but we want to create something more user-friendly... but still powerful. Current Catalyst.Core works really nice for our RL PoCs :)