typical protocol for windows-only games given lack of jax cuda support?

[nullonesix]

what's the typical way for doing RL for windows-only games when jax cuda has no support on windows?

[0xlouis]

I don't think there's a “standard” (maybe I'm wrong), from what I've seen, everyone does their own thing.

At the time of writing, I'm training a DreamerV3 agent (from this repo) on the game “Lethal Company” (Windows).
To do this, I wrote a Windows-side script that injects the keyboard/mouse and reads the screen (and the game's memory to compute the reward).
The script makes all this available on an MQTT broker which allows me to run several games in parallel through the network.
The training script and the gym environment(s) run on a single GNU/Linux system.
Basically, the gym environment interacts with the game(s) via the network.

I'll publish the project on my github soon.

In the past, I've done the same thing with Switch games via a Yuzu emulator (but Yuzu has since been removed by Nintendo).