search

PKU-MARL / HARL

Official implementation of HARL algorithms based on PyTorch.
484 stars 59 forks source link

Questions about learning rate decay in shared_param actors #24

Closed Yuxin916 closed 9 months ago

Yuxin916 commented 9 months ago

Hi, Thank you for your amazing work.

I have a question regarding the OnPolicyBaseRunner. As the init function, if actors are parameter sharing, the self.actor.append(self.actor[0]). It is a list with N identical actor. Then why in run function, only apply lr_decay to the first actor? self.share_param: self.actor[0].lr_decay(episode, episodes).

Thank you and look forward to your reply.

Ivan-Zhong commented 9 months ago

Hello, thank you for acknowledging our work. If actors share parameters, they all correspond to actor[0]. Thus decaying the lr of actor[0] should suffice.

Yuxin916 commented 9 months ago

Oh i see i see. Thank you!