a puzzle about the input of strategical timed attack

[GongYanfu]

Do you understand why the input of strategical timed attack is four images? what's the function of these four images?

[GongYanfu]

It seems that the shape of all observations got from env is 4x84x84. is it gray image, not rgb image?

[davide97l]

The reason is that for atari games usually an observation consists of 4 stacked frames so as to encode temporal information. That's not specific of strategical timed attack. All observations are also converted to gray image (84x84) and becomes 4x84x84 after frames stacking.

[GongYanfu]

yeah now I understand. but the input of my model is only one rgb image, I think fro the rgb input image I need do some modify on your code.

[davide97l]

Yes in this case you need to modify the code, but usually is more efficient to train on grayscale images, as an experiment you could try both ways.

[GongYanfu]

ok, thanks for your suggestion. I'll try