question on freezing target nework

[hashbangCoder]

Hi @yenchenlin1994 , love your implementation!
I went through your code and I can't seem to find where you've frozen the target network?
Unless Im missing something in my excess-caffeine induced brain fade,you continue to update the target every batch?
Wouldn't that hurt your convergence rate badly?

[yenchenlin]

Hello,
Yeah you are right.
Actually I got a reimplemented version.
Will submit soon!
On Wed, Apr 20, 2016 at 17:46 Code-Deep-Blue notifications@github.com
wrote:

Hi @yenchenlin1994 https://github.com/yenchenlin1994 , love your
implementation!
I went through your code and I can't seem to find where you've frozen the
target network?
Unless Im missing something in my excess-caffeine induced brain fade,you
continue to update the target every batch?
Wouldn't that hurt your convergence rate badly?

—
You are receiving this because you were mentioned.
Reply to this email directly or view it on GitHub
#15

[hashbangCoder]

Hi again,
i'm trying to reproduce the results on keras and have trained for 400,000 steps and the bird is unable to cross the first pipe consistently. My loss is low though ( 0.2) and Q-values are in the range of [0,8]. How long did it take for you before it actually started working i.e. cross the first pipe consistently?

[yenchenlin]

I can't remember the exactly number of iterations, but it's no more than ~1000,000 steps

[xiahouzuoxin]

Still cannot find freezing target network in current version's code. It's really no effect?

[zsy372901]

@hashbangCoder
I meet the same question that the silly bird keeps top of the screen.....Did you fix it?

[weijinsong]

I also couldn't find freezing target network code. But thanks for your code. It's helpful for me.

[initial-h]

I write a version base on this repo with freezing target network.FlappyBird_DQN_with_target_network

[patrick-llgc]

Here is another repo with target network. https://github.com/patrick-12sigma/DRL_FlappyBird

I made target network an option. You can turn it on and off and experiment to see how much it affects the convergence of training.

I refactored the network into a class, and added some logging functionalities to track the training process. I also borrowed the human play function from @initial-h. Thanks!