A robot Hitler?

June 2, 2009 6:33 AM

Electronic Evolution: Research Show Robots Forming Human-like Societies

Within fifty generations of this electronic evolution, co-operative societies of robots had formed - helping each other to find food and avoid poison. Even more amazing is the emergence of cheats and martyrs. Transistorized traitors emerged which wrongly identified poison zone as food, luring their trusting brethren to their doom before scooting off to silently charge in a food zone - presumably while using a mechanical claw to twirl a silicon carving of a handlebar moustache.

posted by supercres (38 comments total)

I've got the entire history of speculative fiction behind me to tell you this is a bad idea.
posted by The Whelk at 6:36 AM on June 2, 2009 [2 favorites]

Of course, for evolution to take place, there must be some sort of slight (?) intergenerational randomization taking place, like DNA mutation in humans. The article is unfortunately a bit short on details.

How long before they try to incorporate sexual reproduction, I wonder...
posted by supercres at 6:37 AM on June 2, 2009

This seems like a interesting project, but the article is pretty short and only goes into vague details, spending most of the time completely anthropomorphizing the robots.
posted by Midnight Rambler at 6:38 AM on June 2, 2009 [1 favorite]

And they have a plan.
posted by ricochet biscuit at 6:40 AM on June 2, 2009 [8 favorites]

supercres: "How long before they try to incorporate sexual reproduction, I wonder..."

I'd weld it.
posted by Joe Beese at 6:45 AM on June 2, 2009 [6 favorites]

EPFL Laboratory of Intelligent Systems; a paper [pdf] describing what I presume is this experiment.
posted by you at 6:47 AM on June 2, 2009 [1 favorite]

Their programming was initially random, so the first generation staggered around the place like bunch of concussed puppies

That sounds awesome. LOLCONCUSSEDPUPPIES
posted by shothotbot at 6:54 AM on June 2, 2009

Sounds like they did what they were programmed to do, and nothing more.
But then I don't read sci-fi.
posted by rocket88 at 6:55 AM on June 2, 2009

Hm, not actually a real paper (Microsoft Word alert). I can see no reason for using physical robots for this other than cheap PR. Also, I sincerely hope "human-like" is the journalist's invention. And I'm pretty sure game theory problems have been explored using genetic algorithms before.

Still, looks like a fun lab.
posted by you at 6:57 AM on June 2, 2009 [1 favorite]

Their programming was initially random, so the first generation staggered around the place like bunch of concussed puppies.

Hey guys, don't give PETA any ideas about the life hidden within these robots. Remember Sea Kittens? We might get "electronic puppies," which shouldn't be used to experiment upon. I can see a PR disaster stemming from "setting a group of harmless 'electronic puppies' loose in a room with food and poison, just to see which of Čapek's children would survive, with no thought to their suffering."
posted by filthy light thief at 7:01 AM on June 2, 2009

The Whelk is right. In fifty years' time, our descendants will curse us for sowing the seeds that eventually became the fifty-foot robots that will rule the earth then, hogging all of the food and forcing the tattered remnants of humanity to subsist on poison alone.
posted by No-sword at 7:03 AM on June 2, 2009

Robots Evolve from Concussed Puppies to Brainiacs!
"These cute little guys need to take an Ethics 101 class next time," says bemused professor, observing the behavior of his creations, who have been given nicknames like "St. Francis" and "Little Pol Pot."

(From the Weekly Galaxy News)
posted by kozad at 7:08 AM on June 2, 2009

The article describes an open-ended awesome experiment where you make a bunch of tiny, generalized robots and see what happens. The actual paper is limited to a few stereotyped communications about food vs poison to study just the altruism factor.

I never understand why ALL scientists keep taking baby steps like this. Why aren't there people out there just putting a bunch of stuff together and seeing what happens? If there's something of particular interest, it can always be isolated and studied alone.

For instance, make a bunch of tiny, mobile robots. They each have, say, 10 RGB LEDs. Put out the food and poison, but also let them "prey" on each other somehow (like laser tag, the underlying hardware would have to support one robot "killing" another with some signal they couldn't ignore) and so forth. And other features I'm too short-sighted and lazy to think of and type.
posted by DU at 7:10 AM on June 2, 2009

This article explained.
posted by Krrrlson at 7:15 AM on June 2, 2009 [4 favorites]

Yeah interesting but the link you gave is to a summary of this story, which has been online since January 2008. Not the freshest thing.
posted by modernnomad at 7:15 AM on June 2, 2009

oops posted to soon. Meant to add that also the 2nd article I linked to is a summary of this original piece in Current Biology, from March 2007.
posted by modernnomad at 7:17 AM on June 2, 2009

The alpha robot will emerge as a 1000 ft tall titanium-coated, cold-fusion driven spider with the heart of Kim Jong Il, the brains of Stephen Hawking, the bravery of a thousand Spartans and the never-give-up determination of a high-school cheerleading squad. It will spit sulphuric acid out of its spiked, diamond-tipped proboscis as it runs at 500mph over the barren landscape, picking off lesser beings with its nuclear-warhead-tipped legs. It will march towards complete intergalactic domination - manufacturing clones of itself and firing them off to distant suns to crush aliens so that, in 2 billion years, something doesn't do the same to it.

Then, one day, a rat will take it down.
posted by jimmythefish at 7:18 AM on June 2, 2009 [2 favorites]

I'd like to see the scientists modify the experiment so that the robots are rewarded if they independently create a form of religion and worship. Imagine the look on religious zealot's faces who propose that God passed down religion to man.
posted by digsrus at 7:23 AM on June 2, 2009 [1 favorite]

I never understand why ALL scientists keep taking baby steps like this. Why aren't there people out there just putting a bunch of stuff together and seeing what happens? If there's something of particular interest, it can always be isolated and studied alone.

Because in 99.99999% of cases what happens is: nothing. Swarm robotics in particular is a field of "emergent behaviour" -- you take many simple agents, like ants, with simple programming (do A when you find food, do B when you have no food) and simple abilities (move, sense food, pick up food, drop pheromones), and their collective behaviour somehow turns out to be very complex. The relationship between the simple programs and the emergent behaviour is very poorly understood except for a few examples borrowed from nature, and the design space is infinitely large.
posted by Krrrlson at 7:30 AM on June 2, 2009 [3 favorites]

10 Find Jewbot
20 Kill Jewbot
30 Goto 10
posted by klangklangston at 7:35 AM on June 2, 2009

All I know is that the robots will have an easier time determining which ones to trust when they finally add Autobot and Decepticon insignias.
posted by yeloson at 7:35 AM on June 2, 2009 [2 favorites]

I can see no reason for using physical robots for this other than cheap PR.

Your chances of getting your paper in a robotics conference or journal increase significantly if you have demonstrated your work on physical prototypes.
posted by Krrrlson at 7:36 AM on June 2, 2009 [1 favorite]

I can see no reason for using physical robots for this other than cheap PR.

Rodney Brooks is of the opinion, and I tend to agree, that avoiding real-world issues such as noise and physical limitations of materials, etc can be unhelpful. A virtual robot can give an inflated impression of how well your algorithm works in the real world. In fact, you can see this even in non-robotics contexts. Simulated sample data is never as valuable as real data, the more the better.
posted by DU at 7:55 AM on June 2, 2009

Swarm robotics in particular is a field of "emergent behaviour"

Right, which is why it's interesting if you let the emergent behavior evolve. Put a bunch of stuff together, add in evolution, and stand back. If something evolves, it'll be interesting. (If the same thing evolves every time, that's even more interesting.) If nothing does evolve, that's also interesting. Why would evolution "fail" in some contexts? What's the minimum set of hardware/software/behaviors that evolution needs to get a toehold, etc?
posted by DU at 7:59 AM on June 2, 2009

MetaFilter: Put a bunch of stuff together, add in evolution, and stand back
posted by Joe Beese at 8:08 AM on June 2, 2009 [1 favorite]

Reminds me of how the Boppers came about in Software
posted by brundlefly at 8:38 AM on June 2, 2009

DU:

If 99.99% of the time nothing happens, then when nothing happens it's not interesting. If I wanted to set up a variety of artificial life programs with food and predator/prey relationships, that's pretty easy to do. I remember buying the Rudy Rucker's Artificial Life Lab and playing around with the software therein. It lead to some interesting looking patterns, but unless you were setting up an art installation, that's about it.
posted by zabuni at 9:10 AM on June 2, 2009

As long as the robots cannot yet navigate around "hot lava" by stepping only on objects raised from the floor there is still time for me to send my 3-year-old niece to destroy them.
posted by snofoam at 9:56 AM on June 2, 2009 [1 favorite]

"Why would evolution "fail" in some contexts? What's the minimum set of hardware/software/behaviors that evolution needs to get a toehold, etc?"

Now that is an awesome question. But I wonder if it's just a matter of time. On Earth life appeared, what, a billion years after the planet formed? So there must have been hundreds of millions of years of waiting after the rocks cooled, when conditions were right but the chemicals never quite combined in the proper way. Maybe they formed replicating molecules numerous times but, like the scientist's algorithms, they fizzled out, until finally one line sucessfully replicated and kicked off life as we know it today. Maybe evolution fails most of the time, but given enough years (and the proper conditions) it will always succeed, eventually.
posted by Kevin Street at 10:19 AM on June 2, 2009 [1 favorite]

Incidentally, this experiment strikes me as more of a demonstration of how memes evolve, rather than physical organisms, since the robots don't change, but the programming inside of them is selected for efficiency with each "generation." The robots may not actually be thinking like we do, but they've still managed to come up with concepts like deception and cooperation.
posted by Kevin Street at 10:25 AM on June 2, 2009

Your chances of getting your paper in a robotics conference or journal increase significantly if you have demonstrated your work on physical prototypes.

So why publish in a robotics journal? Oh right it's because an experiment like this in a simulated world wouldn't even be interesting anymore.

The only 'new' thing here is that they did this using real robots rather then virtual agents. This kind of swarm modeling is pretty old hat.
posted by delmoi at 11:14 AM on June 2, 2009

Incidentally, this experiment strikes me as more of a demonstration of how memes evolve, rather than physical organisms, since the robots don't change, but the programming inside of them is selected for efficiency with each "generation." The robots may not actually be thinking like we do, but they've still managed to come up with concepts like deception and cooperation.

Except memes are spread by communication, whereas what these robots are doing is being spread by 'reproduction'. You wouldn't call an ants behavior a meme that's being spread memetically, rather those things are built into their DNA.
posted by delmoi at 11:18 AM on June 2, 2009

But there is communication going on - mediated through the scientists who are conducting the experiment. Through the scientists, the most successful algorithms can combine and spread into all the robots, just like memes spread in a human population.
posted by Kevin Street at 11:26 AM on June 2, 2009

Why would evolution "fail" in some contexts?

There are probably lots of reasons, but in my (very limited) experience evolution (or a genetic algorithm really) fails when the search space is too large or not smooth enough. That is to say that there are huge numbers of possibilities that are just crap relative to ones that even kind of work, and that if you do stumble on a solution that works, modifying, even very slightly, it is very likely to leave you with something that doesn't work at all, rather than something that works slightly differently.

When I was fooling around with genetic algorithms I had a simulation that was not too far from what we're discussing here. I had an agent that was on a two dimensional grid of cells that could either come across food or poison, to use the same terminology as the article. I wanted to evolve it to look for food. It could see a few squares in front of it, move, and turn. I figured this would be a piece of cake, and then I'd move on to more interesting stuff, like having the agents eat each other or something. No such luck! I ran hundreds of thousands of generations of the algorithm/simulation and discovered pretty quickly that I had given it far too much freedom.

The behaviour of the agent is controlled by a neural network. Initially, I had everything about the network subject to the genetic algorithm. Connection weights, activation functions, number of layers, number of nodes per layer, etc. I tried that for a while, and not only did it not come up with any good behaviour, it was pretty slow to run because some of the neural networks generated by the genotypes were quite large. I decided to scale back the amount of freedom and hard coded a bunch of stuff that was previously selected by the algorithm. I ran it again. The result was still crap. I hard coded every thing except the connection weights between the neurons, which were now part of a very small, simple network. Finally, it started to find some behaviours that worked! The best agents were the ones that went to the edge of the map and ate all the food on the perimeter by running around it in a loop until they died of exhaustion.

Even then, those agents were frequently rendered useless by mutation or crossover, sending the whole thing back to square one, with agents that just spun in place until dying, never even venturing off their starting square to look for food. One random mutation to a single gene is all it takes!

I could have put tons of them in a big area and given them the ability to eat each other, but they probably never would have found each other, and if they did, they probably wouldn't know what to do.

That said, I'm not surprised at all to read that these researchers are taking "baby steps" like this, because even though it might be interesting when something doesn't work, researchers need to publish, and to paraphrase a friend of mine, nobody wants to publish in the journal-of-shit-that-didn't-work.
posted by benign at 11:44 AM on June 2, 2009 [4 favorites]

So why publish in a robotics journal? Oh right it's because an experiment like this in a simulated world wouldn't even be interesting anymore.

I was wondering when the self-proclaimed doctor of everything would weigh in.
posted by Krrrlson at 12:12 PM on June 2, 2009

The behaviour of the agent is controlled by a neural network. Initially, I had everything about the network subject to the genetic algorithm. Connection weights, activation functions, number of layers, number of nodes per layer, etc. I tried that for a while, and not only did it not come up with any good behaviour, it was pretty slow to run because some of the neural networks generated by the genotypes were quite large. I decided to scale back the amount of freedom and hard coded a bunch of stuff that was previously selected by the algorithm. I ran it again. The result was still crap. I hard coded every thing except the connection weights between the neurons, which were now part of a very small, simple network. Finally, it started to find some behaviours that worked! The best agents were the ones that went to the edge of the map and ate all the food on the perimeter by running around it in a loop until they died of exhaustion.

Why do I get the distinct feeling I'm reading the foreward for the 21st century?
posted by The Whelk at 12:14 PM on June 2, 2009

10 Find JewbotKlangKlangston
20 Kill JewbotKlangKlangston
30 Goto 10

fear my elite hacking skills

No, hang on, |=3|> oh fuck it
posted by Sparx at 5:32 PM on June 2, 2009

And how would you program the titular robot?
posted by klangklangston at 8:03 PM on June 2, 2009