DQN : Improve performance of obtain_evaluation_episodes

Open aruhela opened this issue 3 years ago • 1 comments

Hi Garage Developers,

The DQN algo uses 1 worker for running _num_eval_episodes, which seems to be a bottleneck in its performance. Has anyone tried leveraging all the compute cores to make this step faster? Any reason why it hasn't been done before?

File : garage/torch/algos/dqn.py Line : 173-180

Thanks Amit Ruhela

Mar 31 '23 19:03 aruhela

Are there any active developers in the Garage team who would like to comment on my question? Is that too trivial to reply?

Apr 03 '23 02:04 aruhela