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
I'd like to ask about the loss of a2c algorithm. There are three part of loss in tianshou's code:
loss = a_loss + self._w_vf * vf_loss - self._w_ent * ent_loss
I'd like to ask about the meaning of the third part of the loss. I try use this to run in CartPole-v0, it converge after 1355 steps,then I try use only the first and second part of loss to run in CartPole-v0, it converge after 770 steps. It seems that only use the first and second part of loss is enough?
The text was updated successfully, but these errors were encountered:
The third loss is about the policy entropy loss. In other repo such as openai baseline, their A2C/A3C implementation has this term, so I add the entropy loss here.
Of course you can set the weight of entropy loss to zero :)
I'd like to ask about the loss of a2c algorithm. There are three part of loss in tianshou's code:
loss = a_loss + self._w_vf * vf_loss - self._w_ent * ent_loss
I'd like to ask about the meaning of the third part of the loss. I try use this to run in CartPole-v0, it converge after 1355 steps,then I try use only the first and second part of loss to run in CartPole-v0, it converge after 770 steps. It seems that only use the first and second part of loss is enough?
The text was updated successfully, but these errors were encountered: