DanielTakeshi
commented
7 years ago Actually, no, the game input is the same, it's in the range 0 to 255 for each pixel. Then the q-network will scale it down before getting it as input, but my human net will simply take the original image as it is, as it was trained to do so. Thus, there's no need to do input re-scaling, though I could try and do that to see if I get better performance.
A continuation of this previous issue.
The code in spragnur divides the state by the input scale. So I think I have to do the same thing, i.e. train the human-guided net on input which was scaled by 255.