Google and CSU team up to advance the state of artificial intelligence
August 18, 2014 4:38 AM Subscribe
Evolving QWOP gaits is the first work that samples video from the QWOP game to drive the fitness function of a genetic algorithm, which allows a fully autonomous simulated runner to kind of slowly shuffle forward, effectively achieving human-like levels of performance.
Computer models still aren't as smart as human brains. I got frustrated and just banged on the keys in quick succession after 2 tries and figured out the same technique.
posted by Potomac Avenue at 5:25 AM on August 18, 2014
posted by Potomac Avenue at 5:25 AM on August 18, 2014
This is a great example of an optimizer getting stuck at local maxima. A properly implemented optimization technique should be introducing enough noise to bump them out of the local peak in order to find the global maximum.
That said, given how easy it is to fall over in QWOP, I imagine that the optimization surface is SUPER spiky, which would make finding the global maximum pretty hard in any kind of hill climber.
posted by leotrotsky at 7:04 AM on August 18, 2014 [6 favorites]
That said, given how easy it is to fall over in QWOP, I imagine that the optimization surface is SUPER spiky, which would make finding the global maximum pretty hard in any kind of hill climber.
posted by leotrotsky at 7:04 AM on August 18, 2014 [6 favorites]
This seems to me to be a task that genetic algorithms would be particularly bad at... I'd predict that it would get caught up on a whole bunch of local maxima. It's surprising that it ended up on the same strategy that a bunch of humans use. But I think that indicates that QWOP is so difficult that humans just naturally land on using something like a genetic algorithm themselves instead of doing anything more complicated.
posted by painquale at 7:05 AM on August 18, 2014 [1 favorite]
posted by painquale at 7:05 AM on August 18, 2014 [1 favorite]
jinx
posted by leotrotsky at 7:20 AM on August 18, 2014 [1 favorite]
posted by leotrotsky at 7:20 AM on August 18, 2014 [1 favorite]
I just know I couldn't convince anyone I was human if this were the Turing Test.
posted by straight at 8:40 AM on August 18, 2014 [1 favorite]
posted by straight at 8:40 AM on August 18, 2014 [1 favorite]
what's funny is that we imagine we've probably found just about the best way to walk using our own limbs.
posted by jepler at 8:45 AM on August 18, 2014
posted by jepler at 8:45 AM on August 18, 2014
I do like the thought that we're just trapped in a local maximum and that there's actually a much better way to walk using our limbs. Shuffling around like a QWOP runner or crabwalking like Regan in The Exorcist or something like that.
posted by painquale at 9:04 AM on August 18, 2014 [2 favorites]
posted by painquale at 9:04 AM on August 18, 2014 [2 favorites]
Oh man, this is great.
I do like the thought that we're just trapped in a local maximum and that there's actually a much better way to walk using our limbs.
Once we start talking about human gait, the tricky thing is to figure out what the fitness function is. I mean, we have aggressively explored outside the common local maxima (natural walking & running & crawling movements) in service of both sports and survival/warfare situations; high level competitive track-and-field stuff for example is a great big stew of weird movements that work better than "I dunno, just run fast" or "I dunno, just jump as high/far as you can", and if you look at the kind of body movements used in like jungle recon or other such Tom Clancy shit stuff gets really goofy in service of maintaining balance and quiet while scouting, etc.
Martial arts in general are, while not primarily focused on locomotion, basically a whole giant pile of iterated theories of non-intuitive ways to optimize human movement.
But at the end of the day, is walking as a basic human motion something that justifies optimization beyond a probable local maximum? Speedwalking techniques work or they wouldn't be what they are, but if you're strolling to the store you're probably not going to be in enough of a hurry often enough to switch into that gait, and if you are in that much of a hurry that often you'll more likely use some other tech solution to the problem (bike, car, bus) instead.
Folks with transient or chronic joint/movement impairments are probably the best picture we have of the idea of routing to a different local maximum in walking gaits—the basic concept of a limp is just this, reprioritizing parts of the walking movement to accommodate a sensitive limb or a displaced joint movement, but folks who deal with serious hip/knee/ankle/foot injuries or mobility-impairing diseases end up manifesting a whole variety of specific movement pattern alterations that pretty precisely capture the idea of a common trait/practice suddenly becoming non-viable (in a general model as e.g. an extinctions event somewhere in the modeled ecosystem). Adaptability under pressure really is an impressive thing.
posted by cortex at 9:19 AM on August 18, 2014 [4 favorites]
I do like the thought that we're just trapped in a local maximum and that there's actually a much better way to walk using our limbs.
Once we start talking about human gait, the tricky thing is to figure out what the fitness function is. I mean, we have aggressively explored outside the common local maxima (natural walking & running & crawling movements) in service of both sports and survival/warfare situations; high level competitive track-and-field stuff for example is a great big stew of weird movements that work better than "I dunno, just run fast" or "I dunno, just jump as high/far as you can", and if you look at the kind of body movements used in like jungle recon or other such Tom Clancy shit stuff gets really goofy in service of maintaining balance and quiet while scouting, etc.
Martial arts in general are, while not primarily focused on locomotion, basically a whole giant pile of iterated theories of non-intuitive ways to optimize human movement.
But at the end of the day, is walking as a basic human motion something that justifies optimization beyond a probable local maximum? Speedwalking techniques work or they wouldn't be what they are, but if you're strolling to the store you're probably not going to be in enough of a hurry often enough to switch into that gait, and if you are in that much of a hurry that often you'll more likely use some other tech solution to the problem (bike, car, bus) instead.
Folks with transient or chronic joint/movement impairments are probably the best picture we have of the idea of routing to a different local maximum in walking gaits—the basic concept of a limp is just this, reprioritizing parts of the walking movement to accommodate a sensitive limb or a displaced joint movement, but folks who deal with serious hip/knee/ankle/foot injuries or mobility-impairing diseases end up manifesting a whole variety of specific movement pattern alterations that pretty precisely capture the idea of a common trait/practice suddenly becoming non-viable (in a general model as e.g. an extinctions event somewhere in the modeled ecosystem). Adaptability under pressure really is an impressive thing.
posted by cortex at 9:19 AM on August 18, 2014 [4 favorites]
There are ways our of our current local maximum.
posted by Monday, stony Monday at 9:27 AM on August 18, 2014 [1 favorite]
posted by Monday, stony Monday at 9:27 AM on August 18, 2014 [1 favorite]
« Older occasionally, she wears flannel. | Our Microbiome May Be Looking Out for Itself Newer »
This thread has been archived and is closed to new comments
That is not a compliment when it comes to QWOP.
posted by Lemurrhea at 5:11 AM on August 18, 2014 [4 favorites]