deepQN_tensorflow icon indicating copy to clipboard operation
deepQN_tensorflow copied to clipboard

Published 20 hours ago • verified publisher icon mrkulk

deepQN doesn't learn

Open shendiaomo opened this issue 8 years ago • 5 comments

I've tried this program on a Mac, the agent never got total score more than 10..

Is there some well-trained model as a demo?

shendiaomo avatar Mar 10 '16 04:03 shendiaomo

Yeah It never get score higher than 10~20 because I removed fixed target network which is essential to prevent oscilation and reach high score. I'm working on new version and it's almost done. I'll upload that soon!

gliese581gg avatar Mar 12 '16 03:03 gliese581gg

Great news!

shendiaomo avatar Mar 14 '16 03:03 shendiaomo

It seems in the paper of 2015 there was no target network in the algorithm, and this code is based on this version.

In the paper of 2016, they say that a target network is necessary. This is what you called the new version, isn't is?

But now I'm very curious how the deepmind people made the first version work...

shendiaomo avatar Mar 14 '16 10:03 shendiaomo

target network is necessary for learning.

mrkulk avatar Mar 21 '16 02:03 mrkulk

If someone is interested, I have a working implementation in tensorflow that can reach similar score (in the only game I've tested) to DeepMind's paper.

ppwwyyxx avatar Jun 06 '16 07:06 ppwwyyxx