[Question] User needs to reset gSDE noise when using learned model? #1767
Description
❓ Question
Many thanks for the great library!
I have been trying out gSDE lately, which seems to be working well for my problem, but I have found that when I simulate the learned model using the model.predict() approach described in the examples in the documentation (e.g., here), it gives deterministic behaviour (in the sense that the behaviour in each simulated episode will look the samel) even if I set deterministic=False in the call to predict(). After some digging, I think I understand that this is due to the sde_sample_freq setting not being made use of in predict(), which sort of makes sense because that function doesn't have access to the environment.
So my question is just: Am I correct in understanding that when running models learned with gSDE, if the user wants the same non-deterministic behaviour as at the end of learning, the user needs to keep track of n_sde_freq themselves and do the model.policy.reset_noise(env.num_envs) themselves at appropriate intervals? If so, it's possibly something to mention in the documentation? (Happy to have a go at contributing with such edit(s) if appropriate.)
Checklist
- I have checked that there is no similar issue in the repo
- I have read the documentation
- If code there is, it is minimal and working
- If code there is, it is formatted using the markdown code blocks for both code and stack traces.