Slashdot Mirror
Music By Natural Selection
maccallr writes "The DarwinTunes experiment needs you! Using an evolutionary algorithm and the ears of you the general public, we've been evolving a four bar loop that started out as pretty dismal primordial auditory soup and now after >27k ratings and 200 generations is sounding pretty good. Given that the only ingredients are sine waves, we're impressed. We got some coverage in the New Scientist CultureLab blog but now things have gone quiet and we'd really appreciate some Slashdotter idle time. We recently upped the maximum 'genome size' and we think that the music is already benefiting from the change."
"Given that the only ingredients are sine waves, we're impressed."
This is different from all other sounds, including regular music, how?
No reply yet and the website can't even load.. now I understand why we don't RTFA!
All signals can be represented with a set of sine waves. That's what makes Fourier transforms so useful.
What would be really impressive is if they had music that can't be represented as a set of sine waves.
Did I just get Rick-rolled?!?
In fact, isn't this playing God?
"Paragraph 7" by Cornelius Cardew, among other works, explored similar ground decades earlier. (Citation to my friend's book on Brian Eno's Another Green World.)
The site has paid ads, one of which apparently has been taken over by the XPAntiVirus people. If you visit the site, it will install malware, unless you are using Firefox and Linux.
A long time ago when I was learning lisp, I worked through an interesting book by Heinrich Taube called Notes from the Metalevel. A very enlightening and interesting work for people interested in both music theory and computer science.
My work here is dung.
...minimalistic electronic music.
IANA biologist, but nothing in the theory of natural selection precludes "intelligent" selection, as far as I know. There simply needs to be some sort of fitness function. Intelligence, in some form or another, factors into this all the time. In this case, fitness is determined by whether a bunch of people like it or not. That's really no different than a plant appearing attractive to a bee, or a beetle tasting nasty to a lizard.
JJ
I remember reading papers on this during my AI classes in the mid 90's. I don't see how this is impressive nearly 15 years later.
Here's the first link I found on G.P. Music from '98 which actually had the computer rate some of the music.
http://graphics.stanford.edu/~bjohanso/papers/gp98/johanson98gpmusic.pdf
If you look at his references, people were doing this in the '80's.
No, I didn't RTFA. I didn't even read the article I linked in this post, so don't get upset if they aren't completely related.
What keeps people from herding it toward an existing copyrighted tune? Even composers accidentally do this all the time.
Table-ized A.I.
>...but now things have gone quiet and we'd really appreciate some Slashdotter idle time.
Your wish is our Slashdotting! That's a name-brand CPU cooling solution you're running, right? Gooood.
Paleotechnologist and connoisseur of pretty shiny things.
That "27K ratings" sort of changes things, don't you think? Certainly, a "rating" sounds more rigorous than traditional, Darwinian, "useful mutations live longer/reproduce more".
"He who would learn astronomy, and other recondite arts, let him go elsewhere. " -- John Calvin, commenting on Genesis 1
It depends to what level of detail you want to look at, but there are some other waveforms that at least have their own names. I mean, to some extent these can still technically be represented by sine waves, but generally seem to have their own characteristics.
I've written a few genetic algorithm/programming things for "music" over the years. However, not being a musician, I approached it only from an algorithmic perspective. The last of these, called "grammidity" can attempt to evolve sequences of midi events based on a kind of grammar that evolves (loosely based on the ideas behind L-systems). I had it online for a couple of years, but it never evolved much of anything interesting. The source code (java) is on sourceforge and includes ways to evolve "plants" and a fuzzer that generates html and which worked quite nicely to break browsers a couple of years back.
Not really. In this experiment, the music is evolving in an environment where "fitness" equates with "what people say they like". The people voting can't make changes to the music, they only get to say how much they like it (on a 5-point scale).
It's not even selective breeding, where a breeder has a trait that they are looking to improve and forces mating among individuals which exhibit that trait.
I was able to do some rating for a while, and I think the results are fairly cool, but it may not produce anything very interesting for a couple reasons.
The first is that there isn't strong enough evolutionary pressure. There are too many people rating with very different opinions of what sounds good. I think it would be much more interesting to create different channels. Classical, jazz, ambient, electronica, whatever. It's still a very broad definition but not so much that our ratings aren't just noise.
Secondly, the algorithms used to generate the music are really important. I couldn't find any information on it, but the way the notes are put together seems fairly random. I think it's important to stick to what we do know sounds good... to an extent. For example, the gene could contain information on which way to move the current note, rather than the specific note. That way you could limit it to 2 or 3 steps and lay it over a scale or mode. The willy nillyness of it will guarantee that we pick 'safe' consonant sounding harmonies. 5ths and 4ths with beep boop melodies.
Very interesting though, I can't wait to see what happens with this.
I tweaked some Apache config with help from the hosting provider, removed some unnecessary audio content from the front page, and it seems to be responding better now...
And it's sounding sweet!
No problems loading here. After a couple of minutes, I'm tired of it. Time to kill it. No cowbells. No drums. No strings. No piano. Nada. Just strange synthesizer noises from decades ago - as has already been pointed out. A couple of loops almost sound good, but mostly just boring repetition.
"Windows is like the faint smell of piss in a subway: it's there, and there's nothing you can do about it." - Charlie Br
This is an academic site and there are no paid ads. It hasn't been compromised either, as far as I can tell.
Set up a Facebook group to fight against the continued xmas dominance of Simon Cowell's evolved algorithms.
Head to Evolectronica when the slashdot dust has settled. I'm planning to give it a make-over and some banging new evo-tunes.
I guess this isn't really user generated content.
I see their individual loops are covered by the creative commons license for non-commercial use with attribution, but I'm search of new On Hold music, so I'm hoping we can come up with some sort of solution. It's a problem when you have zero budget though. heh
I'm looking forward to future generations when they start to do good transitions between different loops, that will be interesting.
Gonzo Granzeau
"Nothing the god of biomechanics wouldn't let you into heaven for.." -Roy Batty
Comment removed based on user account deletion
No cowbells. Bruce Dickinson, is that you?
I've abandoned my search for truth; now I'm just looking for some useful delusions.
Although the loops sound repetitive and annoying, if you listen to some of the earlier loops, you can tell they've made a lot of progress. I'm hoping something resembling a human voice turns up but I don't think this is possible unless they increase the amount of complexity allowed. They already had to increase the allowed complexity once, I wouldn't be surprised if this becomes the primary limiting factor.
Having been both an EE and a CS major I had an idea exactly like this. Alas I didn't do anything about it. Hopefully this project will go somewhere and end up creating really complex music.
So far the tunes sound meh...
I wonder if voice can be generated too?
I remember reading papers on this during my AI classes in the mid 90's.
And I was writing software to use genetic algorithms to generate and/or harmonize melodies in 95/96. And I'm still impressed.
For one thing, even if the general idea isn't totally new, most of the world never gets off their butt and actually does *anything*.
The other thing is that even when the general idea is fairly straightforward, getting an implementation that not only works but sounds pleasant can be non-trivial. It's often not as simple as throwing a machine learning technique or grammar or what have you at the problem -- you'll either end up arbitrarily constraining the output (which is fine -- that can be an art) or you'll have to work at some level of sophistication. Either way, getting to the point where you have listenable output can be something worthy of respect.
Most music itself recycles ideas that've been around since Montverdi, but some of it's still pretty impressive, and hey, even the stuff that sucks... well, the people writing it are at least not frittering away their lives watching TV.
Tweet, tweet.
Found myself with some time to kill, so I had a go at this. Here are my thoughts:
1. The original loop (linked to in the summary) has a recognizable beat, even if many of the accompanying tones sound dreadful together. I'll put it this way: generation 0 sounded way better than a lot of the stuff I've seen try to pass for "electronic music" on YouTube. The original loop already had a fair amount of complexity to start with. I'd be more impressed if they began with a loop that had several sine waves with completely random attributes and then evolved that into something resembling music, and then evolved that into something resembling good music.
2. Note that they're only generating four-bar loops, not an entire song. My guess is that after they consider the experiment finished, they'll take the best ones and paste them into a song. Which won't be difficult, as this music fits into the electronic genre after all, and there is not that much variation amongst the loops that I heard.
3. I'd be curious to know what the constraints on the evolution algorithm are. For example, in the loops that I heard, a mis-matched set of notes was extremely rare so I wonder if they have chords or progressions that the algorithm can select from when generating a new loop.
4. On the front page, there's a video where a professor argues that evolution drives all of human music. I strongly disagree with this on just about every level. Music styles are based on culture, individual expression, and the power of suggestion. Yes, there is a kind of selection at work, but it only occurs within genres and cultures, and very little about it is natural. On top of that, the power of persuasion is immense. People can be told what to like. Today, it's commercially engineered pop music. Two centuries ago, it was religious hymns. Stephen Colbert could screw up this entire experiment just by taking pity on the horrid-sounding loop 76 and giving it his famous "Colbert Bump."
5. The loops really sound like they were made with a Tenori-On, so maybe these "researchers" are just fucking with us about the whole "evolved music" thing.
Not me. I just threw in an overused meme. ;^)
"Windows is like the faint smell of piss in a subway: it's there, and there's nothing you can do about it." - Charlie Br
It is called "artificial selection". Darwin started out by pointing at the example of humans who breed - evolve - animals and plants using artificial selection and then went on to show that a similar process takes place in nature, causing animals and plants to evolve on their own, only that nature does the selection rather than humans, thus the term "natural selection" to be contrasted to "artificial selection".
The phase space of this experiment is too large to explore with the simple rating system. No wonder the "survivors" all sound hyper-sequenced and repetitive, and nothing like Beethoven. What's happening is a bifurcation of the binary number space, because a music sequence is just a binary value occurring on a binary timeline, and each vote of plus or minus is a bifurcation of that space. An "i love it" vote is no less a simple plus, just a plus with extra survival chances.
The problem is that within the regions of binary space that are voted down, there could be much Beethoven, or even all of Beethoven - who could never be mistaken for OMD or The Magnetic Fields (sans lyrics).
The "evolutionary pressure" here is not astute, not deep. The "survivors" might have a glittering sonic quality, enough to get drive-by votes, but the set-up of the experiment could never teach us anything about music aesthetics. I suggest that this algorithm be re-purposed to selecting blinking patterns for Christmas light displays. Same thing, really - but it has nothing to teach us about music.
The first time I clicked the link...
... I got a bogus system scan web page and then it tried to get me to run an EXE file. I tried the link a few minutes later and it seemed okay. I'm perplexed as to what happened. From my browser history, the bad link was...
http://www.newscientist.com/blogs/culturelab/2009/12/amanda-gefter-books-arts.php
h t t p : / / n i s s a n - r e n t . c n / g o . p h p ? i d = 2 0 0 6 - 5 1 & k e y = 0 5 2 2 c 7 0 6 6 & d = 1
I'm using Opera 10.10 (latest) and haven't been anywhere other than major news sites today. Just thought I'd mention it in case anyone else sees the same.
I did a similar thing earlier this year. However, I took a top-down approach: I picked a music style, did a several simple preset musical patterns, and defined a set of rules how to mutate them to something different. There's no system to pick the good mutations though, so it's not really evolutionary, but at least it sounds somewhat musical. The app is already available as online applet: http://www.rinki.net/pekka/acid/ ...
Some people are having problems with the New Scientist link. AardvarkCelery has some info in the post currently below this
http://entertainment.slashdot.org/comments.pl?sid=1485740&cid=30521010
Anyone remember Terry Riley's In C? This reminds me of it; now if they just had several sources going through the fragments at different rates...