seolhokim / distributedrl-pytorch-ray Goto Github PK

@seolhokim This repo is very helpful for me. My question is since there is impala which is central optimizing with data collected to the central buffer, and dppo which sync the communications between learner and actor. How can I run ppo with distributed sampling workers just like the partial combination between impala and dppo. Thanks

@seolhokim Thank you for providing the perfect code. It might help me a lot. Want to know whether this library will continue to be updated? If there is a corresponding tutorial, it would be even better. I look forward to your updates to this library.

think you for this good work,but i have some problem, why i can not train in acrobot-v1

Hello, code is very clear ,but I encountered some problems while running the code.

There are no errors ,but the reward of "CartPole-v0" running with dppo does not rise.
I think there may be bugs in agents/runners/learners/dppo_learner.py.

On line 13（ "gradients_ids = ( [agent.run.remote() for agent in actors]）" ），
the agent.brain's gradient has been calculated,but catually it is None.
Therefore, all brain's weights can not be updated.

Besides, On line 14 ( "while len(gradients_ids):" ) , whether the length of the gradients_ids will be a constant, and then lead to wrong results.

Thank you for your contribution,the code is very good,but when I try to run some complex environment like nasim:Medium-v0 , the trajectory length will reach step limit.I will very appreciate if you can run the environment mentioned above when you have some time,Thank you very much!