Have trouble reproducing "Appendix B/Figure.9" in the paper
lesleychou opened this issue · 5 comments
Hi,
I am trying to reproduce the "Figure.9: Sampling_frequency", but with my ADR training, the Sampling_frequency distribution does not show a clear trend like the one in the paper.
Figure.9 in the paper:
Sampling_frequency fig from my own ADR training: MES in [8, 20]
My training cmd:
python -m experiments.domainrand.experiment_driver lunar \ --experiment-name=unfreeze-policy --experiment-prefix="ADR" --max-agent-timesteps=4e6 \ --seed=5 --continuous-svpg --randomized-env-id="LunarLanderRandomized-v0" \ --plot-frequency=5
May I ask what is your experiment setting to produce Figure.9?
Thanks for your great work and public code!
Look forward to receiving your comments, appreciate any of your hints!
Second this, my sampling distribution looks much closer to uniform than presented in the paper.
My command:
python -m experiments.domainrand.experiment_driver lunar\ --continuous-svpg --experiment-name=unfreeze-policy\ --agent-name=adrold --experiment-prefix=adrold\ --randomized-env-id=LunarLanderRandomized-v0
The sampling distribution:
Hi Lesley, I apologize for the late comment; when I left research, I unsubscribed from most old repos.
In our paper, we used seeds 101-105; would you mind trying those and showing distributions? In addition, #5 found a bug with our repo, which seeds correctly, but the code pre- merge request is what we used for our work. Hopefully that helps.
I've re-subscribed to the repository, so hopefully I can answer more quickly in the future
@lesleychou can you replicate the ADR results in the paper? Could you let me know your email if possible, thanks.
@harish-kamath could you replicate the ADR sampling results? My plots look exact the same as yours...