learning-rate so sensitive to training result?

facebookresearch / torchbeast

A PyTorch Platform for Distributed RL

Apache License 2.0

734 stars 113 forks source link

learning-rate so sensitive to training result? #5

Closed 1576012404 closed 4 years ago

1576012404 commented 4 years ago

train PongNoFrameskip-v4 in example expriment, when i change learning-rate from 0.0004 to 0.0008,i can not train to good result in 30000000 frames. why? when i choose a larger batch-size, such as 8,training speed slow down,and so how to speed up training?

1576012404 commented 4 years ago

why close this issue? i am intersting in the answer of this question.thank you.

heiner commented 4 years ago

Hey 1576012404,

While we're more than happy you chose to use TorchBeast, this issue tracker is for bugs. Yours seems to be better described as miscellaneous questions. I appreciate that RL is hard and hyperparameter dependent. Lots of people are working to improve that situation though so stay tuned and keep reading the literature. :)