search

jscriptcoder / Upside-Down-Reinforcement-Learning

Landing a Spaceship using Upside-Down Reinforcement Learning (a.k.a ⅂ꓤ)
11 stars 4 forks source link

Results robustness #1

Open TrentBrick opened 4 years ago

TrentBrick commented 4 years ago

Thanks for sharing your code and reproduction!

I am finding that everything runs great when I use the seed=0 you provided but not with other random seeds. Were you also finding your results to not be robust?

Thanks. Trenton

jabowery commented 4 years ago

seed = 0 failed to converge for me Capture Capture1

jscriptcoder commented 4 years ago

Hi Treton,

Good observation, and happy that someone even tried to reproduce it. No, unfortunately I didn't try different seeds since my main goal was to understand and successfully implement the paper testing it on a different environment. I honestly didn't play much with hyperparameters and as you can see I didn't run the algorithm with other seeds for robustness. So there is plenty of room for improvement.

I'd be super happy to see more robust implementations.

Thanks a lot, Fran

On Tue, Jul 14, 2020 at 1:40 AM Trenton Bricken notifications@github.com wrote:

Thanks for sharing your code and reproduction!

I am finding that everything runs great when I use the seed=0 you provided but not with other random seeds. Were you also finding your results to not be robust?

Thanks. Trenton

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/jscriptcoder/Upside-Down-Reinforcement-Learning/issues/1, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAEV2XHAHIMPMQ34ZZZ5ZDTR3OLNZANCNFSM4OY67Y3A .