Closed slerman12 closed 2 years ago
Hi @slerman12, the standard methodology since the release of DQN is that the agent will observe grayscaled frames, down sampled to 84x84 and stacked in the depth channel with the previous 4 observations.
There's a lot of nuance in the preprocessing stack for the ALE. I would highly suggest using something that's tried and tested. I would recommend looking at:
processors.AtariEnvironmentWrapper
(this is what I'd consider the gold-standard implementation).atari_lib.AtariPreprocessing
atari_wrapper.AtariWrapper
We provide the option to retrieve RGB observations from gym.make
but in the future I plan on building the complete preprocessing pipeline into the ALE itself so there isn't all these different postprocessing implementations.
Let me know if you have any other questions about this, I know it can be confusing, the most common questions I receive are around postprocessing.
I've been using rgb.
gym.make(env, obs_type='rgb')
However, StableBaselines3 uses grayscale.
Which one is preferred/standard?
Thanks.