facebookresearch / e3b Goto Github PK

Hi there!

Firstly, I want to express my gratitude for sharing your code with the community. I've been able to successfully reproduce the E3B and obtain the expected results. However, I've encountered the following error during the environment setup and initialization process:

mylevel.des: line 30, pos 0: Map too large at (25 x 22), max is (76 x 21) at "ENDMAP"
mylevel.des: 1 errors detected for level "mylevel". No output created!

While the training ultimately produces the desired outcomes, I'm curious to understand the source of this error. It appears to be related to either the Minihack or Gym-Minigrid package. Could you provide some insights into the origin of this issue and its potential implications?

Thank you once again for your contribution, and I appreciate any assistance you can offer in resolving this matter.

I have tried using several different hyperparameters, but I have been unable to get my model to converge on some difficult exploration games in Atari(Gravitar and H.E.R.O), even after 5 billion timesteps. Can you suggest some effective hyperparameters for these games or provide an explanation for why you think my model is not converging?

Here are the hyperparameters i've tested : intweight (3e-07,2.0), reward_norm (all,intr,ext). I tried both extremes, it seems like the inverse dynamic loss stays around the same value and doesn't decrease.

Hi!

Where can I find the Multiroom-N10 environments that you mention in the paper? I only see up to N6 here

Also, did you ever consider using the Minihack-MazeWalk envs?

Thanks!