Farama-Foundation / stable-retro

Retro games for Reinforcement Learning
https://stable-retro.farama.org/
MIT License
176 stars 36 forks source link

Filtered Actions Documentation #3

Open onaclov2000 opened 1 year ago

onaclov2000 commented 1 year ago

Issue summary

It would be nice to add documentation on filtered actions, with some examples. I have ran into issues with it when integrating a new game, as I wasn't sure what I need to prevent.

There are some items in the docs https://retro.readthedocs.io/en/latest/search.html?q=filtered&check_keywords=yes&area=default

But the referenced "discretizer.py" file specifically doesn't really give me an intuitive understanding of the format and meaning of the data.

I've seen similar formatting in the data.json (if I'm recalling) but I don't have a clear understanding of what it is doing.

In my particular game the only "constraint" I can think of is that you can't press left and right at the same time, you can't press up/down at the same time, and maybe it's not useful to press start/select at the same time. (Of course U/D and L/R are constraints because of the physical configuration of the controller, not as a means of reducing the action space for the RL agent, that's a whole different story, but this documentation would be helpful for that too)

I'll continue to dive into how it works and help document, but if someone already has insights, I'd love to see any documentation.

System information

Not relevant

pseudo-rnd-thoughts commented 1 year ago

Hey, this documentation is part of gym-retro not stable-retro. We will look to update the documentation to make this more clear in the future.

onaclov2000 commented 1 year ago

No problem, I had assumed it was gym-retro docs, but either way, I would love to have a bit more info on how to use/interact with certain things, and don't mind helping document as I discover things.

pseudo-rnd-thoughts commented 1 year ago

Thanks, we recommend joining the Farama discord, a number of the changes are discussed there