You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
First of all, congratulations for your paper. I really enjoyed reading it, the idea is quite refreshing and I was happy to see I'm not the only one using Unity for RL research. ;)
I do have a (few) technical question(s) however. I have been wondering for several days now how to deal with message passing when learning in mini-batches.
Do you have parallel environments ? If so, there is an important deal of observations preprocessing, right ? I'm talking about the fact that the observations have to be properly lined in order to feed each parent its child message (in the bottom-up case).
Do you consider the message to be part of the output action ? If not, how do you backpropagate through the message sending head? In the appendix, it is written that only the sensory inputs and the action (torque + link/unlink) are considered.
There are a few details that I do not yet completely graps, but I really enjoyed the paper overall.
Thanks a lot for your time !
The text was updated successfully, but these errors were encountered:
Hey !
First of all, congratulations for your paper. I really enjoyed reading it, the idea is quite refreshing and I was happy to see I'm not the only one using Unity for RL research. ;)
I do have a (few) technical question(s) however. I have been wondering for several days now how to deal with message passing when learning in mini-batches.
There are a few details that I do not yet completely graps, but I really enjoyed the paper overall.
Thanks a lot for your time !
The text was updated successfully, but these errors were encountered: