Neural network evolutionary settings optimization

[vicnaum]

Hello,

I was rendering one image, and it was doing more than 300 passes, but DOF noise still wasn't going off. I guessed - what if a renderer could see that there's DOF noise, and increase DOF samples to clean it?

So I had an idea - what if a renderer could optimize it's settings via evolutionary methods?

For example - it's rendering the passes...

On one pass it sets default settings.

On the other pass it varies the params slightly (samples, thresholds, etc). For example - sets one Light samples a bit higher.
Then it monitors - how fast is noise leaving.

If raising samples on this particular light helps to improve noise levels signficantly - it remains them a bit higher. If only the time increases, and the noise isn't much better - then it lowers the samples again.

And so on, and so on - on each pass it sets slightly different settings - and monitors the evolution of best settings for current scene.

I've seen some similar method here:
http://www.boxcar2d.com/

It uses evolutionary methods to develop the best passable car.

So, what do you think - is it practical to apply these methods to render?

[Ondra]

this is a thesis topic of one of my colleagues ;). But the approach is slightly different, you can go much deeper than just adjusting the few parameters for users, and there are more suitable methods than evolution algorithms. Hopefully I'll implement at least some basic adaptivity soon.

[vicnaum]

Wow, great! Will be waiting for it.

[Polymax]

And will one button in render setting - "make beautifully" :)

[ecximer]

And it will not make the "basic adaptive" side problems that had already begun in VRay quite nauseated?

[Polymax]

And it will not make the "basic adaptive" side problems that had already begun in VRay quite nauseated?

+1
I do not want that!