Genetic algorithms, Mona Lisa and JavaScript + Canvas(blog.nihilogic.dk)
blog.nihilogic.dk
Genetic algorithms, Mona Lisa and JavaScript + Canvas
http://blog.nihilogic.dk/2009/01/genetic-mona-lisa.html
21 comments
Wow - 2009. Excellent post.
I recently entered a competition (February 2013) and used this technique without having seen this blog. I also used webworkers and SVG.
Source: https://github.com/binarymax/randriaan Demo: http://binarymax.com/randriaan.html
Sorry, FF or Chrome only (because of how I load the webworkers from the same page)
@tlarkworthy - I had a version that did 12 images for each generation and chose the best - It was a better result but it really throttled the CPU and took down the browser on my weaker machine, so went with the single thread.
I recently entered a competition (February 2013) and used this technique without having seen this blog. I also used webworkers and SVG.
Source: https://github.com/binarymax/randriaan Demo: http://binarymax.com/randriaan.html
Sorry, FF or Chrome only (because of how I load the webworkers from the same page)
@tlarkworthy - I had a version that did 12 images for each generation and chose the best - It was a better result but it really throttled the CPU and took down the browser on my weaker machine, so went with the single thread.
btw, the original article about using Genetic Algorithms to derive Mona Lisa's image also generated quite a bit of interesting discussions here, 4 years ago:
https://news.ycombinator.com/item?id=389727
https://news.ycombinator.com/item?id=389727
I also did the same for that competition (while having seen the blog): http://github.com/aerique/vbwga
My "artistic" spin was I wanted to reproduce the Victory Boogie Woogie using only circles (see vbw-example.pdf and vbw-example-2.pdf in previous link).
My "artistic" spin was I wanted to reproduce the Victory Boogie Woogie using only circles (see vbw-example.pdf and vbw-example-2.pdf in previous link).
How was the event? Still sorry I couldn't make it. Looks like it was a lot of fun.
It was good fun especially the drinks afterwards were a good occassion to talk with the other developers and organizers.
The jury was heavily biased towards algorithms that created VBWs from nothing so we didn't stand a chance :-)
The jury was heavily biased towards algorithms that created VBWs from nothing so we didn't stand a chance :-)
It would work better if you did a local optimization after every crossover and ditched random mutation. I think they call it a memetic algorithm (http://en.wikipedia.org/wiki/Memetic_algorithm).
So take your every triangle vertex (x,y) and differentiate against the fitness (using finite difference), then gradient accent on that parameterization. That will let the triangles fit the edges effectively (I am thinking Mondrian here primarily)
So take your every triangle vertex (x,y) and differentiate against the fitness (using finite difference), then gradient accent on that parameterization. That will let the triangles fit the edges effectively (I am thinking Mondrian here primarily)
It's super mario bros on the background. You can use your keys to control him.
Not "Bros.", "Land"... (This game was a formative experience for me, so please bear with my nitpickery... :)
http://en.wikipedia.org/wiki/Super_Mario_Land
http://en.wikipedia.org/wiki/Super_Mario_Land
I noticed! Very cool.
That's very interesting
It seems the default mutation parameters are a bit low, if you increase it it may lead to better results
I think the main problem with this is that it's using a fixed number of polygons, and apparently is mutating their vertice positions.
If it could be changed for a range of polygons (with a range of vertices, let's say from 3-6) maybe it can converge faster (and better)
But it is nonetheless a very nice experiment. GA/GP is one of those things that really can work with "untractable" questions.
It seems the default mutation parameters are a bit low, if you increase it it may lead to better results
I think the main problem with this is that it's using a fixed number of polygons, and apparently is mutating their vertice positions.
If it could be changed for a range of polygons (with a range of vertices, let's say from 3-6) maybe it can converge faster (and better)
But it is nonetheless a very nice experiment. GA/GP is one of those things that really can work with "untractable" questions.
This is really awesome. It took me a while but I think I found the perfect settings for "Mondrian" (but it really does take a long time)
http://i.imgur.com/OhgYn6X.png
Different images work best with different polygon sizes, different cut-offs etc. and it's really sensitive. If I change the parent cutoff for example, a simple 0.01 change seems to ruin Mondrian from forming, even after waiting several thousand mutations.
Different images work best with different polygon sizes, different cut-offs etc. and it's really sensitive. If I change the parent cutoff for example, a simple 0.01 change seems to ruin Mondrian from forming, even after waiting several thousand mutations.
Note how much faster it runs today: 1966 mutations took 120 minutes back then, today it performed ~2100 mutations in about 4 minutes.
Indeed, I ran the script for 12 minutes and managed 37511 mutations which came up with this:
http://imgur.com/phA7Ocu
DNA here: http://pastebin.com/PEGwKvG4
http://imgur.com/phA7Ocu
DNA here: http://pastebin.com/PEGwKvG4
While not a "learning" technique directly, this reminds me of this amazing work: http://research.microsoft.com/en-us/um/people/kopf/pixelart/
Next step: run a genetic meta-algorithm to optimize the initial parameters... judge fitness based on, say, result_quality / runtime...
(Or fastest-to-achieve-reasonable-result...)
(Or fastest-to-achieve-reasonable-result...)
I'm not trolling, but this appears to work pretty poorly. Does anyone know any better approaches to approximating images like this?
i played Mario there . lol.. btw ..excellent post .
http://alteredqualia.com/visualization/evolve/
[1] comments: https://news.ycombinator.com/item?id=392036