Normalization implementation is actually computing standardization
Opened this issue · 2 comments
Hello! This issue is regarding the "normalization" implemented using RunningMeanStd. Algorithms that use the normalize_input and normalize_value config params use RunningMeanStd internally to keep track of running means and variances and then compute a transformation of
For reference, normalization (in range [0,1]) is implemented using
References
hi @anishhdiwan, Yeah you are right it is an old names. Hard to rename some variables because afraid to break configs.
I've tested normalization vs standardization. Including normalization using p05 and p95. Standardization worked much better for most of the envs. I think I didn't merge it but I have branch somewhere.
Hey @Denys88, thanks for the response. I understand that renaming might be a bit messy. But it's nice to know that standardization vs normalization has been tested in the past. Perhaps a comment in the configs (or readme) might be a simple temporary solution to avoid errors in scientific communication (would be nice to make the distinction clear as rl_games is often part of research codebases)?
In any case, thanks again :)