[Question] Result better than baseline

[Seraphli]

I run some tests on dopamine. When I run the game 'Qbert', the result seems better than the result presented in the baseline. I use here config located in dopamine/agents/rainbow/configs/rainbow.gin. I only change the environment name to Qbert.

[mgbellemare]

That does seem a little better. Our baseline results are reporting 1) training scores, 2) with sticky actions. Are you using 2)?

[Seraphli]

I use here config located in dopamine/agents/rainbow/configs/rainbow.gin. So the configuration seems with sticky actions.

[mgbellemare]

Strange. Maybe the code is getting better as it ages? :)

Is the x axis on your tensorboard million agent steps, or million frames (x4 steps)? The numbers would match the former.

[Seraphli]

I didn't change any training code, nor tensorboard summary code. So I think it matches the baseline. BTW, I run the training for Pong before and the score didn't reach 15 as the baseline showed. I didn't know the reason, so I run rainbow for the game Qbert.

[satyakesav]

@Seraphli, Were you using a GPU to run the code and how long did it take to get the above results?

[Seraphli]

@satyakesav 1d13h for first 100 iterations on Qbert

[satyakesav]

@Seraphli, Ahh. I see. What GPU did you train it on?

[Seraphli]

@satyakesav 1080ti

[GoingMyWay]

@satyakesav 1d13h for first 100 iterations on Qbert

That is really fast, normally for DQN, one iteration takes 40 minutes on V100.