Feature: Remote control RL processes

[damajor]

In order to train on high-end hardware (high core density CPUs or multiple professional GPUs) would it be possible to dissociate game execution/control/metrics gathering from the compute part ?

[AechPro]

I'm not sure I understand the question. If you're asking if the game and RLGym can be run on a compute cluster, the answer is that it probably can't be without a lot of effort. Rocket League really wants a GPU and Windows.

[damajor]

Well having a windows agent to manage RL itself and communicate with linux compute cluster for training. Inference can be done on windows.

[AechPro]

This doesn't sound like an RLGym question. RLGym is just an interface to rocket league, it does not train anything. If you run an agent on Windows and use it to interact with RLGym you can do whatever you want with that data afterwards.

[damajor]

Humm well nevermind I need to get my hands on it before arguing :)