Evolutionary AI

[Mray3460]

Just over a week ago, I read an article here on the Escapist about Hatsune Miku, a.k.a. the computer generated singer/pop star. What sparked my interest/imagination, however, was the number of commentators who mentioned that a ?real? person had written the lyrics and melody. I quickly recalled a scene in a documentary on chaos theory (?The Secret Life of Chaos?) that I?d watched several months earlier. In the scene (Linked below, watch to 6:36 from the set start point), programmers created 100 randomized, virtual ?brains? to operate a virtual body, brains that performed better at causing the body to walk, were ?bred together,? to produce new brains, upon which, the process was repeated. This eventually produced brains capable of walking, and maintaining balance, perfectly. Soon the principle was applied to full bodies, which encountered, and learned to deal with, complex problems such as uneven ground, being hit, or falling, as well as being able to get back up, or even fight each other.

Where these two linked together, in my mind at least, was this: What if we created new, virtual ?brains,? that would produce a series of notes or scores (Initially, at random) either with complete originality, or with some standard input. Then, large groups of people would listen to the melodies produced by the brains, and rank them by which ones ?sounded best.? Then, the most consistently high rated brains would be bred, producing new brains. Repeat the process ad infinitum, to eventually produce computer programs that can compose melodies (Lyrics would probably be a bit harder, but it could still be conceivable).

Further food for thought and discussion: What if we applied a similar process to a group of chat-bots, having the brains ranked by which are most convincingly human, eventually producing programs indistinguishable from human beings (I.E. able to pass a Turing Test).

Citations:

Segment from The Secret Life of Chaos (Watch to 6:36):
http://www.youtube.com/watch?v=-1x-7ZLKhjw#t=02m18s

NaturalMotion (The company that originally produced the brains. They apparently use the tech to produce character movements for video games):
http://www.naturalmotion.com/

Original Escapist Article:
http://www.escapistmagazine.com/news/view/105250-Japanese-Pop-Star-Draws-Crowds-Despite-Being-a-Hologram

First part of The Secret Life of Chaos, in case anyone is interested:
http://www.youtube.com/watch?v=HACkykFlIus

Addendum: If this has been done before, please link to the original thread.

 

[Lilani]

I think it's a good idea, and territory worth exploring. Vocaloid created the first virtual diva, perhaps their next great innovation will be the first virtual songwriter.

 

[DiMono]

I look forward to serving our robot overlords.

Seriously though, that's pretty cool.

 

[Lyx]

The only problem with that approach, is that it isn't really "intelligence" (then again, most "AI" has little to do with intelligence). Your idea does work evolutionary, and yes it may (with enough iterations) lead to "programs" that manage a given challenge well - but the problem is: It didn't learn it, because it does not include the ability to learn. That's because efficient learning requires understanding (see where i'm going here?).

I case you didn't manage to follow my chain of arguments: Via your evolution idea, "adaption" is outsourced to reproduction - the "learning process" doesn't really happen INSIDE that "AI" but instead outside of it. Why this may be a problem:

1. It will mean that the AI-instance which you end up with, may handle the CURRENT scenario, but will be unable to adapt to changing circumstances.

2. For the selection process to work well, you need some method to.... well, SELECT which combos "survive". More often than not, you'll not be able to leave that up to "survival of the fittest". Let's take an FPS AI vs Player deathmatch with respawning. Problem: Getting killed is a binary criterion - either it gets killed by the player, or it doesn't. Now, how is the selection process supposed to improve, when it cannot notice that a combination DOES work better against the player, yet not good enough yet for survival?

Sure, you could just take "survival time" as a criterion - but this time-measurement is no longer plain "survival of the fittest" in the original sense. You the programmer are arbitrarily making up criteria, and then intervene in the selection process according to how an AI scored in misc criteria. This difference is what i meant with #2.

P.S.:

3. The above two issues also imply another shortcoming: Such an "AI" does not invent any new methods - rather, it just modifies EXISTING parameters and methods (ones defined by the programmer) and then during reproduction considers what worked well.

 

[Lullabye]

They can do this, but I'm still stuck killing retarded NPCs in games. *sigh*