Possible?
How would you approach that with a neural network sort of thing. I don't know enough about them unfortunately but it's been something I have slowly been trying to learn about. There is at least lots of training data available.
AI level builder?
Moderator: Moporators
Re: AI level builder?
if you look around on sla's site you can find like 2 level generators i think (one by ics, maybe another by domovoy? not sure)
so yeah this idea has definitely been explored to some extent already
so yeah this idea has definitely been explored to some extent already
Re: AI level builder?
All level generators of all times:
https://pulu.org/en/60-history-of-elast ... ncy-levels
Also a little related topic: viewtopic.php?f=1&t=9875
https://pulu.org/en/60-history-of-elast ... ncy-levels
Also a little related topic: viewtopic.php?f=1&t=9875
[00:21:58] (@cyre) ismo taas juonu liikaa omenamehuu
Re: AI level builder?
Hmm interesting topic indeed.Ismo wrote: ↑19 Sep 2021, 13:54 All level generators of all times:
https://pulu.org/en/60-history-of-elast ... ncy-levels
Also a little related topic: viewtopic.php?f=1&t=9875
However, it gave me this idea:
Could we build an AI to predict whether some topologically valid combination of polygons is a real level vs. some kind of total gibberish. Then use that to train another AI on how to build levels that feel like they are real levels?
Re: AI level builder?
There are many ways to do this, and I have been experimenting a bit with two approaches. I haven't been able to make anything sensible yet though.
There are AI generators that generates specific text and images (human faces, cats, lyrics): https://thisxdoesnotexist.com/
From this Elma could make levels of different concepts such as barry p pipes, ramone gentle levs, or internals. Then you can approach this in at least two ways: either using thumbnail images (which means you also would have to make an AI creating levels from thumbnails), or you can do it directly from level data (probably a better approach, but could be more difficult if you do it script kiddie style as I have done).
I'm not sure there is enough data for each approach though. I mean if you use a ready made model such as "Elma Lev AI" viewtopic.php?f=1&t=9875 it uses a ready made google model that is used to recognize cats and dogs from images. Since you already have that starter model you don't need as much data. But it only works on images.
Another approach you can use is to turn level data into an image. For example if the first pixel is the x value of the starting object, the second pixel is the y value, and then you have some pixels representing flowers and apples, and then polygons with some kind of polygon separation data. This approach would create ranges that could differ in resolution. It could work quite well in Elma levels, as most levels (except pipes and really difficult levels) might forgivable with small alterations.
If we as a community would classify levels (pipes, easy levels, internal style, hoyl style etc) we could have more specific data for ai creation.
I'm sure some day in the near future AI level generators will come (for other games), and this kind of thing would be quite easy to make for someone who actually knows AI science.
There are AI generators that generates specific text and images (human faces, cats, lyrics): https://thisxdoesnotexist.com/
From this Elma could make levels of different concepts such as barry p pipes, ramone gentle levs, or internals. Then you can approach this in at least two ways: either using thumbnail images (which means you also would have to make an AI creating levels from thumbnails), or you can do it directly from level data (probably a better approach, but could be more difficult if you do it script kiddie style as I have done).
I'm not sure there is enough data for each approach though. I mean if you use a ready made model such as "Elma Lev AI" viewtopic.php?f=1&t=9875 it uses a ready made google model that is used to recognize cats and dogs from images. Since you already have that starter model you don't need as much data. But it only works on images.
Another approach you can use is to turn level data into an image. For example if the first pixel is the x value of the starting object, the second pixel is the y value, and then you have some pixels representing flowers and apples, and then polygons with some kind of polygon separation data. This approach would create ranges that could differ in resolution. It could work quite well in Elma levels, as most levels (except pipes and really difficult levels) might forgivable with small alterations.
If we as a community would classify levels (pipes, easy levels, internal style, hoyl style etc) we could have more specific data for ai creation.
I'm sure some day in the near future AI level generators will come (for other games), and this kind of thing would be quite easy to make for someone who actually knows AI science.
-
"leader status in the Elma against-the-system underground" - Abula
-
ElmaAutoGravityApples 2024
IncrElastoMania - Elma Simulation - Browser Game 2020
Elma Imager - Command Line Tool 2020
"leader status in the Elma against-the-system underground" - Abula
-
ElmaAutoGravityApples 2024
IncrElastoMania - Elma Simulation - Browser Game 2020
Elma Imager - Command Line Tool 2020
Re: AI level builder?
Maybe it will even be possible to describe the level to be created, check out this AI: https://www.youtube.com/watch?v=81rBzfbFLiE
-
"leader status in the Elma against-the-system underground" - Abula
-
ElmaAutoGravityApples 2024
IncrElastoMania - Elma Simulation - Browser Game 2020
Elma Imager - Command Line Tool 2020
"leader status in the Elma against-the-system underground" - Abula
-
ElmaAutoGravityApples 2024
IncrElastoMania - Elma Simulation - Browser Game 2020
Elma Imager - Command Line Tool 2020