Does this implementation solve the inconsistent parameter issue from the original TF universe-starter-agent implementation?

ikostrikov / pytorch-a3c

PyTorch implementation of Asynchronous Advantage Actor Critic (A3C) from "Asynchronous Methods for Deep Reinforcement Learning".

MIT License

1.23k stars 280 forks source link

Does this implementation solve the inconsistent parameter issue from the original TF universe-starter-agent implementation? #6

Closed ethancaballero closed 7 years ago

ethancaballero commented 7 years ago

I'm referencing this issue: https://github.com/tensorflow/tensorflow/issues/6360#issuecomment-271486091

ikostrikov commented 7 years ago

I'm not sure. Probably, we should ask authors of PyTorch.

But is it a big issue? I thought that original A3C doesn't use locking so parameters might be inconsistent.

ethancaballero commented 7 years ago

@apaszke

apaszke commented 7 years ago

@ethancaballero please don't tag me all over GitHub. I'm not the only person who can answer this quesition. If you still have doubts ask on the forums.