Not working?

[pmarcin92]

Did anyone try to learn it so the agent really can play pong? I tried to learn it for over 30h on Tesla K80 and it doesn't look good at all.
I have also once concern about saving and restoring the learned weights. I modified the code to save the session once every 100000 iterations and I restore it like that:

    saver = tf.train.Saver()
    sess.run(tf.initialize_all_variables())
    saver = tf.train.import_meta_graph('pong-dqn-1300000.meta')
    saver.restore(sess, tf.train.latest_checkpoint('./'))

Is it me doing something wrong or there is a bug somewhere in the code preventing it from learn the pong?

[addisonhuddy]

@piorunm I had similar results when running locally. After about 5min pygame gets really slow. After letter it train for another couple of hours, no improvement.

[anthdm]

This is not working at all.

[wh33ler]

The original code has some issues... have a look at a working version here
https://github.com/wh33ler/QNet_Pong

[j0el]

Thanks, this is very helpful. Any chance you could push your model? 700K iterations take a while on my machine.

[wh33ler]

sure why not. A added my current checkpoint of 975k steps and added a USE_MODEL Mode which ignores the training aspect.

[to-bee]

Thanks @wh33ler Very nice improvement!

[john-theo]

Hi, @wh33ler is it normal that at 300k timesteps, the ai player moves almost the same as the first 3k steps? I cloned your repo, make a few irrelevant changes(such as rename variables), and I got stupid results. You disabled issue functionality in your repo so I have to question here, LOL.

[wh33ler]

I am not sure what exactly you mean. It has been a while since I looked at it. But it might take some time until the AI gets it :)

[Traven16]

I think on line 22 you have to set USE_MODEL = False for the net to actually train.