hill-a / stable-baselines

A fork of OpenAI Baselines, implementations of reinforcement learning algorithms
http://stable-baselines.readthedocs.io/
MIT License
4.16k stars 725 forks source link

[feature request] Remove erroneous episode from replay buffer #1197

Open WreckItTim opened 1 month ago

WreckItTim commented 1 month ago

Hello, I am using SB3 with a third party simulation environment that is a bit unstable. If running for a few hours, it will crash at some point. I made a crash handler that will automatically reboot the simulation on crash, however it requires resetting the environment to the beginning of the episode because I can not access the full state of the simulation on a step-by-step basis. I would like to remove the impartial episode, that resulted from the crash, from the replay buffer. I can't seem to find a way to handle this natively with SB3, so would like to submit a feature request for this. Perhaps something in the state returned at step? Or maybe this already exists and I just couldn't find it. Or maybe it would be better to have a general method for removing specific episodes from the buffer?

Thanks!