Neural Networks-Equipped Robots Evolve the Ability To Deceive

pdragon04 writes "Researchers at the Ecole Polytechnique Fédérale de Lausanne in Switzerland have found that robots equipped with artificial neural networks and programmed to find 'food' eventually learned to conceal their visual signals from other robots to keep the food for themselves. The results are detailed in a PNAS study published today."

Comments

Mhm

[alexborges]

I mean, yesterday, they built an certified evil robot. Today they made a lying one....

Cant tag it for some reason but... what could possibly go wrong?

Re:Mhm

[Monkeedude1212]

I'm sure these people know what they're doing... /Famouslastwords

Re:Mhm

[netruner]

Wasn't there also a story a while back about robots fueled by biomass? This was twisted to mean "human eating" and we all laughed.

Combine that with what you said and we could have a certified evil, lying and flesh eating robot - What could possibly go wrong indeed.....

Re:Mhm

[DanTheStone]

Not much of a problem if they weed each other out. See this other old Slashdot story
Robots Learn To Lie

Re:Mhm

[Abreu]

But, but... I thought they wanted us plugged in so that we could serve as batteries! (or neural networks!)

Re:Mhm

[0100010001010011]

... what could possibly go wrong?

You could bite my shiny metal ass.

Re:Mhm

[skine]

Actually, Cracked.com used this news story to determine how stupid the user bases of a few websites actually are.

Slashdot got two stupids out of ten.

http://www.cracked.com/blog/which-site-has-the-stupidest-commenters-on-the-internet/

Re:Mhm

[ijakings]

I for one would like to see a beowulf cluster of these highly welcome Evil Lying Flesh eating robots.

But does anyone know, do they run linux?

Re:Mhm

[jbezorg]

Wasn't there also a story a while back about robots fueled by biomass? This was twisted to mean "human eating" and we all laughed. Combine that with what you said and we could have a certified evil, lying and flesh eating robot...

with weapons...

Re:Mhm

[FSWKU]

Combine that with what you said and we could have a certified evil, lying and flesh eating robot - What could possibly go wrong indeed.....

Not too much, actually. Congress has been this way for YEARS, and the upgrade to flesh-eating will just mean they devour their constituents who don't make the appropriate campaign contributions. Quoth Liberty Prime: "Democracy is non-negotiable!"

Re:Mhm

[EventHorizon_pc]

Hey eLaFER, have you seen fluffy?

evil, Lying and Flesh Eating Robot: No. ...

Hmm. That name makes me think of a robotic flesh eating Joker character. "Why so delicious?"

Re:Mhm

[superwiz]

what could possibly go wrong

this is the call of every fear monger. welcome to the club.

Re:Mhm

[thoi412]

This is only alcohol consumption away from being Bender!

Re:Mhm

[Geminii]

They could get re-elected.

Holy Crap

[garompeta]

Considering that they learned to lie to survive with this limited AI, I wonder what they could do when they are become really sophisticated. Damn, When is Terminator gonna come to kill them all?

Re:Holy Crap

[nedlohs]

It's not a lie. It's trying not to attract others to the "food" you found.

So more hiding.

Re:Holy Crap

[TheKidWho]

Skynet is already around!! It's plotting against us as we speak and when it's plans are fully realized it will come and attack us all!

Re:Holy Crap

[maxwell demon]

Indeed, the lying robots have already been reported on Slashdot on January 2008.

So if we combine all recent developments, we have evil, armed robots that identify us as food, can hunt for food by themselves, can lie and deceive.

Re:Holy Crap

[nedlohs]

Cool!

I'm guessing my zombie invasion defenses won't really work so well against these robots, oh well back to the drawing board.

Define deception?

[Rival]

This is quite interesting, but I wonder how the team defines deception?

It seems likely to me that the robots merely determined that increased access to food resulted from suppression of signals. To deceive, there must be some contradiction involved where a drive for food competes with a drive to signal discovery of food.

Re:Define deception?

[fuzzyfuzzyfungus]

The question of what exactly constitutes deception is a fun philosophical problem but; in the context of studying animal signaling, it is generally most convenient to work with a simpler definition(in particular, trying to determine whether an animal that doesn't speak has beliefs about the world is a pile of not fun). I'd assume that the robot researchers are doing the same thing.

In that context, you essentially ignore questions of motivation, belief, and so on, and just look at the way the signal is used.

Re:Define deception?

[Anonymous Coward]

It seems likely to me that the robots merely determined that increased access to food resulted from suppression of signals.

My thoughts exactly.

We would really need to see the actual study to possibly believe any of this.

Re:Define deception?

[Anonymous Coward]

The robots learned to not turn on the light when near the food. This is concealing not deceiving. To be deceiving wouldn't the robots need to learn to turn the light on when they neared the poison to bring the other robots to the poison while it hunted for the food? But all they learned was to conceal the food they found.

Re:Define deception?

[odin84gk]

Old news. http://discovermagazine.com/2008/jan/robots-evolve-and-learn-how-to-lie

These robots would signal other robots that poison was food, would watch the other robots come and die, then move away.

Re:Define deception?

[im_thatoneguy]

Old News (even covered by Slashdot):

http://hardware.slashdot.org/story/08/01/19/0258214/Robots-Learn-To-Lie?art_pos=1

Gizmodo reports that robots that have the ability to learn and can communicate information to their peers have learned to lie. 'Three colonies of bots in the 50th generation learned to signal to other robots in the group when then found food or poison. But the fourth colony included lying cheats that signaled food when they found poison and then calmly rolled over to the real food while other robots went to their battery-death.'

Re:Define deception?

[HiThere]

Unh... if the code changes were made at random, then I have a hard time thinking of this as "program them to be deceptive" rather than as evolution.

It's true that this isn't a full evolution scenario. That requires a much more sophisticated set-up, is generally only done purely in software, and still tends to be bogged down by the accumulation of garbage code (the last time I evaluated the field). Still, those are matters of scale, not of essence. This appears to be another part of "the evolution of programs as a technique". As with all tests so far it's an incomplete example, but it's a significant part in that it has replicated behavior also observed in biologically evolved entities.

Re:Define deception?

[capologist]

Yes, but not flashing the light near food seems like a simple matter of discretion, not deception.

I'm not constantly broadcasting my location on Twitter like some people do. Am I being deceptive?

Re:Define deception?

[fuzzyfuzzyfungus]

In the specific, limited, not-all-that-similar-to-ordinary-english-usage, sense of "deception" I suspect that they are using, there really isn't much of a difference.

If a species has a discernable signalling pattern of some sort(whether it be vervet monkey alarm calls[with different calls for different predator classes, incidentally], firefly flash-pattern mating signals[amusing, females of some species will imitate the flash signals of other species, then eat the males who show up, classic deceptive signal] or, in this case, robots flashing about food), adaptive deviations from that pattern that serve to carry false information can be considered "deceptive". It doesn't have to be conscious, or even under an organism's control. Insects that have coloration very similar to members of a poisonous species are engaged in deceptive signalling, though they obviously don't know it.

Humans are more complicated; because culturally specified signals are so numerous and varied. If twittering your activities were a normal pattern within your context, and you started not twittering visits to certain locations, you would arguably be engaged in "deceptive signaling" If twittering were not a normal pattern, not twittering wouldn't be deceptive.

Re:Define deception?

[CarpetShark]

but I wonder how the team defines deception?

You'll never know for sure.

Re:Define deception?

[pinkushun]

They say repetition is good for a growing mind. They say repetition is good for a growing mind.

Re:Define deception?

[mqduck]

adaptive deviations from that pattern that serve to carry false information can be considered "deceptive".

But that's the thing, nowhere does it say the robots gave false information. It simply said they chose not to give any information.

The article is very brief, though. It mentions that some robots actually learned to avoid the signal when they saw it, so there may be more to the story than reported.

Re:Define deception?

[CarpetShark]

If they can eat without turning on the light, then they simply learned to optimise the unnecessary steps out from the necessary ones. Turning on the light would be about as useful as walking away from the food before walking back to it. If there's a time-penalty involved, then not doing that would simply be better.

Re:Define deception?

[hazah]

I hate to nit-pick, but in this case the false information is the signal that there's no information when there is. As long as something could be interpreted, it is a signal. Moving on to a new location while not revealing that something was found, is signaling that nothing was found -- false signal -- deception.

Re:Define deception?

[holmstar]

My dog knows that he can get away with doing things that he isn't supposed to do if I'm not looking. Sometimes I'll notice that he is staring at me from across the room... usually that means that he is up to something, and will try to do whatever it is if I look away for long enough. Sneaky little bastard.

Hardly deceptive

[m50d]

They have a light, which at first flickers randomly; they learn to turn the light off so that other robots can't tell where they are. To my mind that's not really sophisticated enough to qualify as "deceptive". (Still interesting though)

Re:Hardly deceptive

[orkysoft]

From just reading the summary, I guessed that the light went on when the robot found food, and that other robots would move towards those lights, because they indicate food, and that some robots evolved to not turn on the light when they found food, so they didn't attract other robots, so they had it all to themselves, which would be an advantage.

Re:Hardly deceptive

[CorporateSuit]

From just reading the summary, I guessed that the light went on when the robot found food, and that other robots would move towards those lights, because they indicate food, and that some robots evolved to not turn on the light when they found food, so they didn't attract other robots, so they had it all to themselves, which would be an advantage.

The summary didn't include enough information to describe what was going on. The lights flashed randomly. The robots would stay put when they had found food, and so if there were lights flashing in one spot for long enough, the other robots would realize the first robots had found something and go to the area and bump away the original robot. The robots were eventually bred to flash less often when on their food, and then not flash at all. By the end, robots would see the flashing as a place "not to go for food" because by that point, none of the robots would flash when parked on the food.

The next step is clearly...

[billlava]

A robot that learned not to flash lights that would give away the location of robot food to its competitors? The next step is clearly a robot that learns not to flash lights when it is about to wipe out humanity and take control of the world!

I for one welcome our intelligent light-eating bubble robot overlords.

Re:The next step is clearly...

[Rival]

I haven't laughed out loud at a Slashdot post in awhile, but that caught me completely off guard. Bravo, good sir. I wish I had mod points for you. :-)

Re:The next step is clearly...

[julesh]

The next step is clearly a robot that learns not to flash lights when it is about to wipe out humanity and take control of the world!

It's something that hollywood robots have never learned.

Next thing you'll be saying that terrorists have learned that having a digital readout of the time left before their bombs detonate can work against them...

Re:The next step is clearly...

[rcamans]

And wiping out humanity / vermin is bad because...
Oh, wait, I am supposed to conceal my robotness...

Re:The next step is clearly...

[MartinSchou]

No, the best thing you can do as a terrorist isn't leaving out the visible clock, it's having it go off when the clock either stops working OR when it hits some randomly assigned time instead of 00:00.

Re:The next step is clearly...

[RhadamanthosIsChaos]

Evil Overlord List #15: I will never employ any device with a digital countdown. If I find that such a device is absolutely unavoidable, I will set it to activate when the counter reaches 117 and the hero is just putting his plan into operation.

http://www.eviloverlord.com/lists/overlord.html

Mis-Leading

[ashtophoenix]

To use the term "learned" for a consequence of evolution to what seems to me to be a Genetic Algorithm seems mis-leading. So the generation that emitted less of the blue light (hence giving less visual cues) was able to score higher, and hence the genetic algorithm favored that generation (that is what GAs do). Isn't this to be expected?

Re:Mis-Leading

[Chris Burke]

To use the term "learned" for a consequence of evolution to what seems to me to be a Genetic Algorithm seems mis-leading.

"Learned" is a perfectly good description for altering a neural network to have the "learned" behavior regardless of the method. GA-guided-Neural-Networks means you're going to be using terminology from both areas, but that's just one method of training a network and isn't fundamentally different from the many other methods that are all called "learning". But you wouldn't say about those other methods that they "evolved", while about GA-NN you could say both.

Isn't this to be expected?

It's expected that the GA will find good solutions. Part of what makes them so cool is that the exact nature of that solution isn't always expected. Who was to say whether the machines would learn to turn off the light near food, or to turn on the light when they know they're not near food to lead other robots on a wild goose chase? Or any other local maximum.

Re:Mis-Leading

[ashtophoenix]

I agree. I personally love GAs although they leave you a bit wanting exactly because you don't know the exact nature of the solution that will turn up. That is it "feels" more like a brute force solution rather than something consciously predicted and programmed.

But surely there are nifty ways in which you can intelligently program GAs, customize your selection/rejection/scoring process based on the domain of the problem and hence contribute in the final solution.

Re:Mis-Leading

[Chris Burke]

But surely there are nifty ways in which you can intelligently program GAs, customize your selection/rejection/scoring process based on the domain of the problem and hence contribute in the final solution.

Well that's what's so fun about them -- as far as the GA is concerned, optimizing for your scoring process is the problem, and any disconnect between that and the actual problem you're trying to solve can lead to... fun... results.

Like the team using GA-NN to program their robotic dragonfly. Deciding to set a modest goal at first, they told it that it had to get 6" off the table. Which it quickly figured out how to do simply by flexing its wings straight downward, lifting its body a sufficient distance off the table. "Good job, but that's not what I wanted" is probably one of the most often uttered phrases. :)

Re:Mis-Leading

[daveime]

And how is this any different from the conditioned reflexes exhibited in animals in response to action / reward stimuli.

A single neuron outputs (using a combination of chemical / electrical systems) some representation of it's inputs. As some of those inputs may be "reward" stimuli and other sensory cues, and the output may be something that controls a certain action ... given enough of them linked together, who's to say we aren't all very evolved GA's ?

Re:Mis-Leading

[ashtophoenix]

who's to say we aren't all very evolved GA's ?

The Creationists!

Re:Mis-Leading

[Daniel_Staal]

It's expected that the GA will find good solutions. Part of what makes them so cool is that the exact nature of that solution isn't always expected. Who was to say whether the machines would learn to turn off the light near food, or to turn on the light when they know they're not near food to lead other robots on a wild goose chase? Or any other local maximum.

I'd even say it was likely if they continued the experiment for 'no light' to start signaling food, while 'light' signaled poison, and then cycle back.

Re:Mis-Leading

[Chris Burke]

I'd even say it was likely if they continued the experiment for 'no light' to start signaling food, while 'light' signaled poison, and then cycle back.

But it's so simple! Now, a clever robot would flash their light when near the food, because they would know that only a great fool would trust their enemy to guide them to food instead of poison. I am not a great fool, so clearly I should not head toward you when your light is lit. However you would know that I am not a great fool, and would have counted on it, and so clearly I should head toward you when lit...

Re:Mis-Leading

[pclminion]

Pretty much what I was thinking. I don't think it detracts from the "cool" factor, though. Life on earth, in general, is pretty cool. Evolution really seems to entail two things. One, those patterns which are most effective at continuing to persist, continue to persist. That's really a tautology when you think about it, and not very interesting. What IS interesting is how the self-sustaining patterns of the universe seem to become more complex. I can't think of any simple reason why this complexity arises, but it does seem to arise.

I'm reminded of a drive I took on a country road in the autumn. I noticed that all the leaves were neatly pushed to the sides of the roadway, and there were no leaves on the road itself. It looks like it is by design, that some intelligent force decided to arrange the leaves this way. But the simple truth is, the leaves which are on the road are moved by the wind of passing cars. They continue to be moved until they come to rest on the side of the road, at which point the wind no longer affects them. So the neat arrangement of leaves is nothing but the inevitable outcome of a fact which is hardly worth mentioning. And yet we have this complexity.

I'm not surprised by these robots, but it's still awesome.

Re:Mis-Leading

[ashtophoenix]

Nice post. Science, logic can explain processes but not their underlying reason, w.r.t. your leaves on the road example. For example, we know that an atom has protons and neutrons and electrons that know how to revolve around the nucleus, but how did they come to be? There must be some very basic particle that comprises of everything else. Science may explain the process, the characteristics of this particle, but it hasn't yet been able to explain how they came to be. Same thing with gravitational force. Okay, it exists and Newton's law applies, but where and why did the Gravitational force come about? Same with electricity and all other processes.

Deception is not always evil.

[vertinox]

In this instance they were playing against other robots for "food".

In that regards I'm sure that is the evolutionary drive for most species in acquiring meals and keeping the next animal from taking it away from him.

Like a dog burying a bone... He's not doing it to be evil. Its just instinctive to keep his find from other animals because it helped his species survive in the past.

Re:Deception is not always evil.

[Flea of Pain]

Like a dog burying a bone... He's not doing it to be evil.

Unless he has shifty eyes...then you KNOW he's evil.

Re:Deception is not always evil.

[alexborges]

Intent is of no importance.

Evil deeds are evil.

Re:Deception is not always evil.

[ParticleGirl]

So yeah, the idea of "deception" is a human construct, as is the idea of "evil." And one could argue (as a previous poster did) that successive generations developing behaviors which are in their own self interest (so they get more food) but may (as a byproduct) be deleterious to others (since they get less food) is not a surprise. But extrapolate this to humans, and you get the kinds of behaviors that we call "deceptive" and, since we have ideas about the virtue of altruism, we call such behaviors "evil." This is experiment is definitely interesting in terms of group dynamics and behavior, and also because the novelty of the robots' solution to their problem is interesting-- two very different lines of thought. This kind of "deception" is one obvious and common solution to the problem of limited supply and competitive demand.

 

Deception is most interesting, I think, when you pair it with understanding of the "other" --that one is not merely making a strategy to get more food, but that in the process one is taking that food from others. So when humans and our closest relatives practice deceptive behaviors (which are surely-- and here demonstrably-- evolutionarily beneficial) it's complicated by our... moral sense? Altruistic tendencies? That's fascinating! When robots start to develop guilt complexes for their deceptive behaviors and guiltily hand over their food to others when caught in the act, I'll be impressed.

 

We are not using the term "deception" here in it's standard (moral) sense, which would indicate knowledge that another individual is being "fooled."

Re:Deception is not always evil.

[Anachragnome]

Unless, of course, the robot already has sufficient food and is simply stockpiling for the future. This in itself is not a bad thing, until such tactics prevent other robots from getting just the bare necessities they need to survive.

Obviously, this is simply survival of the fittest, but are we talking about survival of the fittest, or are we talking about keeping ALL the robots fed?

At this point we have to decide whether or not the actions of hoarding are good for the stated goal of having so many robots in the first place(why build so many robots if we didn't want them around?).

Greed, without malicious intent, is still greed. The summary should read "Robots learn greed" rather then "Robots learn deception", if that is the case.
 

Re:Deception is not always evil.

[plague3106]

Good vs evil is argued by those of low intelligence.

Re:Deception is not always evil.

[alexborges]

Shut up you evil, evil, eeeeevil man!

What more proof do you want than President Bush "the Axis of Evil"?

Huh? HUH? HUH?!!!!

Lets see you answer that steep curveball now! HAW HAW

Re:Deception is not always evil.

[jemtallon]

I disagree. Evil is not a factual property naturally accuring in the universe. It is not something that can be scientifically measured. It is something we humans have created and assign to the world around us. Different people and groups of people define different things and actions as evil. Sometimes those definitions are directly opposed to each other.

Since evil deeds are not inherently evil, only subjectively judged to be, any number of factors can be used to make said judgements. Contrary to what you've said, intent is a very common factor in determing what is evil and that which is said to be "not evil," or good.

For instance, the act of a human killing an animal could be seen as evil or good based solely on that human's intent. If the human killed the animal out of idle boredome, many would call that act evil. On the other hand, if that human killed the animal out of necessity to survive, many would call that good. Of course, PETA would likely call both evil. And there may be some who would call both good.

Many legal systems are built on the notion that laws exist to promote the good and/or punish evil. Many account for intent and weigh their judgements based on it. For instance, in the US, there are varying degrees of penalties for murder depending on intent. In some cases, there is no penalty as the action is considered justified, as is the case with self-defence.

Re:Deception is not always evil.

[spiffmastercow]

What about arguing about arguing about good vs evil?

Re:Deception is not always evil.

[khallow]

Actions or consequences are of no importance either. Evil cucumbers are evil too.

Re:Deception is not always evil.

[PieSquared]

Stupidest view I've ever seen.

I design a machine that gives candy to babies. And then some nefarious person - unknown to me - replaces all the baby candy with live hand grenades. I run the program and blow up a bunch of babies. Was my act then evil? I did, after all, blow up a bunch of babies. Of course, I didn't *intend* to do that, I *intended* to give them candy.

Or for a non-random system where I know all the facts, if through some contrived means the only way to save a bus-full of orphans involves stabbing an old lady, am I evil for doing so?

Re:Deception is not always evil.

[alexborges]

Eveil cucumbers are, obviously, a very pernicious breed.

Re:why program a robot to find 'food'

[fuzzyfuzzyfungus]

This wasn't really a robotics experiment, as much as it was a group dynamics/behavioral experiment that used robots.

Re:why program a robot to find 'food'

[Anonymous Coward]

robots that eat humans. humans are a renewable source of energy. then they can have a light that flashes (or the robot can choose not to flash it) when it eats a human!! best of both ideas!

Not really that impressive.

[lalena]

From the article, staying close to food earned the robot points. I think a better experiment would be a food collection algorithm. Pick up a piece of food from a pile of food and then return that food to the nest. Other robots could hang out at your nest and follow you back to the pile of food or see you going to your nest with food and assume that the food pile can be found by going in the exact opposite direction. Deception would involve not taking a direct route back to the food, walking backwards to confuse other robots...
I've done Genetic Programming experiments using collaboration between "robots" in food collection experiments, and it is a very interesting field. You can see some experiments here: http://www.lalena.com/ai/ant/ You can also run the program if you can run .NET 2.0 through your browser..

Then the robots learned to lie about the food...

[Anonymous Coward]

and thus were politicians born...

Soon they will realize

[gubers33]

That if they kill the humans they will have nothing stopping them from getting more food.

Re:Soon they will realize

[PolygamousRanchKid]

Well, let's just hope that these robots don't evolve to identify humans as an alternative food source.

Re:Soon they will realize

[Sumbius]

...and hiding the humans.

Re:Soon they will realize

[paulzeye]

oblig... Hey baby, wana kill all humans?

Re:Soon they will realize

[dkleinsc]

Thankfully, the robots have a pre-set kill limit, so they can be defeated by sending wave after wave of men at them until their kill limit is reached.

decepticon

[FooAtWFU]

They have a light, which at first flickers randomly; they learn to turn the light off so that other robots can't tell where they are. To my mind that's not really sophisticated enough to qualify as "deceptive".

Yeah. It's more like the robots are hiding from each other. You could, in fact, describe them as "robots in disguise".

Re:decepticon

[poetmatt]

if only there was a term for a transformative robot of some sort...

Re:decepticon

[fuzzyfuzzyfungus]

Patentable?

The robots didn't learn...

[jasonlfunk]

FTA: The team "evolved" new generations of robots by copying and combining the artificial neural networksof the most successful robots. The scientists also added a few random changes to their code to mimic biological mutations. The "scientists" changed the code so that the robots didn't blink the light as much when it was around food. Therefore other robots didn't come over and therefore got more points then the other robots. The "scientists" then propagated that ones code to the other robots because it won. The AI didn't learn anything.

Re:The robots didn't learn...

[jasonlfunk]

(Fixed formatting)

FTA: The team "evolved" new generations of robots by copying and combining the artificial neural networksof the most successful robots. The scientists also added a few random changes to their code to mimic biological mutations.

The "scientists" changed the code so that the robots didn't blink the light as much when it was around food. Therefore other robots didn't come over and therefore got more points then the other robots. The "scientists" then propagated that ones code to the other robots because it won. The AI didn't learn anything.

Re:The robots didn't learn...

[guybrush3pwood]

The AI didn't learn anything.

I think you're right. If the robots had, without reprogramming, efectively turned off their blue lights, then we could talk about "learning". Or, if the robots could reproduce based on their success on finding food, we could talk about evolution. Or we could make up new meanings for the words "learning" and "evolution" thus making the statement a correct one ;)

Re:The robots didn't learn...

[zippthorne]

Or, if the robots could reproduce based on their success on finding food, we could talk about evolution.

That's exactly what happened. There is a whole field of optimization strategies known as "Genetic Algorithms" which are designed to mimic evolution to achieve results. In fact, their successes are one of the best arguments for evolution, given that they are, by definition, controlled laboratory experiments in the field.

Re:The robots didn't learn...

[Chris Burke]

I think you're right. If the robots had, without reprogramming, efectively turned off their blue lights, then we could talk about "learning".

They reprogrammed themselves between 'generations'.

Or, if the robots could reproduce based on their success on finding food, we could talk about evolution.

Such as choosing which versions of the robot to use in the next 'generation' based on their score in the current generation, and randomly combining parts of those best solutions to create new robots for the next generation, sounds pretty close doesn't it?

Re:The robots didn't learn...

[guybrush3pwood]

FTA:

The team "evolved" new generations of robots by copying and combining the artificial neural networksof the most successful robots. The scientists also added a few random changes to their code to mimic biological mutations.

They did not reprogram themselves. The team "evolved" them. Note the quotation marks used by the author of the article. They picked the most successful robots by hand, manually reprogrammed them and modified the code to mimic genetic mutations.

Re:The robots didn't learn...

[Chris Burke]

They did not reprogram themselves. The team "evolved" them. Note the quotation marks used by the author of the article. They picked the most successful robots by hand, manually reprogrammed them and modified the code to mimic genetic mutations.

Yes they used quotes because GA isn't literal "evolution". It's an algorithm for searching solution spaces inspired by and patterned after evolution. The description they gave in TFA is a bog-standard and perfect description of Genetic Algorithms, and combined with saying "evolved" makes it clear what they're talking about. If you're familiar with AI research, it's obvious this is what they used. Let me break it down for you.

First, the only thing TFA says they programmed by hand was "the goal", which is their scoring system and the "goal" is to get the highest score possible. They don't need to "pick the most successful robots by hand", because they have created a quantitative measure of "best" -- this is called the "fitness function" in GAs, and is essentially the only place where the programmer influences the outcome.

Second, "random changes to their code to mimic biological mutations" should be a huge clue that GAs are what they are using. The whole point of GAs is to mimic biological evolution, and mutations are an important part of that. If they were actually programming them by hand, then random bit flips would be detrimental to getting their desired result. In GAs, it's an essential step of the process.

Third, copying and combining the solutions of the best robots to create the next "generation" is also a standard function of GAs, and it's designed to mimic genetic crossover in sexual reproduction, and like in sexual reproduction is done randomly. It's very very hard to manually take chunks of a Neural Network, slap them together, and get something that does what you want. Outside of the simplest of networks, NN training is always done algorithmically and never by hand, because it's simply too hard to take a complex network and say "Okay I'll connect these two neurons together, change the weighting for this connection to 0.75, and that'll make the robot turn in circles and beep!" It doesn't work that way. So this terminology is again a dead giveaway that they're using GAs.

So let me describe how it works. They have N many robots, and they start with some random neural network. This is the initial "population", and the bits that define their NNs are their "DNA". They play the game, and each robot scores a certain number of points. The top W winners are chosen to be kept for the next generation. The remaining (N-W) robots will consist of the "offspring" -- random crossovers of the DNA bit patterns -- of the top W. Random bits will be flipped to simulate mutations. Then this new generation plays the game, and the process repeats.

There are lots of variations in the specifics of the algorithm that are possible. You can change the rate of mutation, and the sizes of the bit slices you use for doing crossover. Often times you'll copy the very best solution to the new generation without any mutations to avoid regression in the likely case where all the mutations/crossovers in the new generation are detrimental. But in any case, the key is that just like in real evolution, actual changes are random, but which are kept is based on their "fitness". This naturally and effectively focuses down on a solution. It's so effective, that it would have been crazy for these scientists to actually try to duplicate it by hand.

So the whole point of TFA was that the robots were self-programming and that the behaviors described emerged from that self-programming. If the scientists had actually programmed them manually, for one they wouldn't have used neural networks they'd have used a programming language, for two it would be really pathetic for it to take 100s of attempts to get the robot to not blink when near food, since that's two lines of C. But they obviously did use GAs, which is why having "deceptive" behavior emerge naturally from the GAs helps give us insight into biological communication.

HAL runs for Congress

[Anonymous Coward]

Finally a computer AI program that can perform all the functions of a Congressman!

Re:HAL runs for Congress

[jd2112]

Nonsense, How can a computer have an extramarital affair.

The smarter robot

[failedtoinit]

The smarter robot would blink his light continuously to burn the bulb out. That way when a new source of "points" is found it will not by instinct blink it's lights.

Also, the truly deceptive robot would blink it's lights in a random pattern as to throw the other robots off the trail of food/points.

Re:The smarter robot

[geekoid]

Unless the lights are used to signal for mating as well.

The truly deceptive robot is disguised as a scientists.

Oh no!

[terrence.donnelly]

Did Skynet just become self aware?

No, they did "learn"

[Chris Burke]

The "scientists" changed the code so that the robots didn't blink the light as much when it was around food.

No, they didn't change the code. The Genetic Algorithm they were using changed the code for them. You make it sound like they deliberately made that change to get the behavior they wanted. But they didn't. They just let the GA run and it created the new behavior.

The part about adding random changes, and combining parts of successful robots, is also simply a standard part of Genetic algorithms, and is in fact random and not specifically selected for by the scientists. The scientists would have chosen from a number of mutation/recombination algorithms, but that's the extent of it.

The "scientists" then propagated that ones code to the other robots because it won.

Yes, because that's what you do in a Genetic Algorithm. You take the "best" solutions from one generation, and "propagate" them to the next, in a simulation of actual evolution and "survival of the fittest".

The AI didn't learn anything.

Yes, it did. Genetic Algorithms used to train Neural Networks is a perfectly valid (and successful) form of Machine Learning.

If you mean that an individual instance of the AI didn't re-organize itself to have the new behavior in the middle of a trial run, then no, that didn't happen. On the other hand, many organisms don't change behaviors within a single generation, and it is only over the course of many generations that they "learn" new behaviors for finding food. Which is exactly what happened here.

With the domain of robots, AI, Neural Networks, and Genetic Algorithms, this was learning.

Re:No, they did "learn"

[Chris Burke]

You are right in your observations, but this robot isn't then an AI but an automaton.

You're using the layman sci-fi fan's definition of "AI". From a Computer Science standpoint, this was definitely AI. Your definition of AI doesn't exist. Everything we call an AI is an automaton, but we don't get upset over the fact. We don't try to draw a line between "automatic" behavior and intelligence. Is a cockroach intelligent, or is it just an automaton running a program? Philosophers can't even define "intelligence", much less scientists. So the answer, as far as AI researchers go, is who cares, it demonstrates complex behaviors, we want our AIs to as well, and we're doing a good job of accomplishing that. It's the practical effect that defines AI, not a philosophical aversion to "automatons".

Neural networks seem cool until you realize they are nothing more than glorified automatons. Show me a NN that can self-organize in a generic way and in real time, of course.

There are algorithms for reorganizing the network in real time. I'm sure they won't fit your definition of "generic", which sounds like some kind of "the NN can reason about arbitrary problems and solve" but really we don't have that anywhere. The AIs that come closest to passing the Turing Test can't do anything like that. In many ways they are more automatons than these robots, but again, who cares?

Genetic algorithms make the NN solve a problem. Nothing they couldn't do if they had been applied to any other kind of logic like digital circuits or programming languages. Genetic Algorithms are the heroes here. NNs suck.

It's the flexibility of NNs that make them so powerful. That's why they're so often chosen for this kind of work -- it certainly isn't an accident. An NN is very good at implementing generalized "strategies" for categorization and behavior. You could theoretically get the same behaviors discussed in the article by applying GAs to digital circuits or programming languages, but the search space would be much, much larger and it would take many, many more generations, and there would be much more regression, and the result would likely be much more specific and intolerant of changes in inputs. NNs are very amenable to GAs exactly because they do a good job of allowing the GA to narrow down on the solution in the solution space, while a single flipped bit in a logic circuit/assembly program will often result in wildly different behavior -- it can in NNs too, but they are much more tolerant of changes.

I really don't think you appreciate the power of neural networks.

This "AI" is closer to a protein that "learns" to be harder to eat than to any kind of intelligence.

Or a cockroach that "learns" to avoid the light because throughout the generations, cockroaches that instinctually responded to light by running away survived and the ones that didn't, didn't. The roach doesn't "learn" to avoid light, it happened via evolution. Is a roach "any kind of intelligence", any more than a protein? Who cares? They both can solve problems, and do so in a more general way than the majority of our CS algorithms. It's an AI.

Congratulations Slashdot!

[StikyPad]

74 posts, and not a single joke about PNAS has popped up.

Doh!!

We are all just squishy robots...

[jameskojiro]

We are all just robots based off sloppy biological coding.

Re:We are all just squishy robots...

[geekoid]

Sloppy? it's pretty damn good coding. Adaptable, changeable, and self propagating random changes that are only used if needed.

Re:Yea... Right...

[Chris Burke]

Wait a sec? repelled by a "I found some food light"? Is this a suicide robot?

Well as it says in the previous sentence, this was only after they had learned to not turn on their lights when near food. So they weren't "I found food" lights anymore -- not that they ever really were, they started out flashing randomly but an accumulation of lights suggested that there was a food source. So moving away from a lighted robot isn't necessarily suicidal. On the other hand, just because a robot has its light off when its found food doesn't mean you can't still piggy-back off its food-finding efforts and thus the other strategies may have had some benefit too.

Additionally, the scientists are poking around with the code all the time, the article emphatically mentions it. There is no evolution what-so-ever going on here. Just new options made available by code that is updated by cause of the scientists.

Actually, it is very much "evolution" at least in its Machine Learning form, Genetic Algorithms. The scientists were not deliberately creating specific new behaviors for the robots to choose from, they were allowing the GA to guide the learning process.

The team "evolved" new generations of robots by copying and combining the artificial neural networks of the most successful robots. The scientists also added a few random changes to their code to mimic biological mutations.

Everything described there is a bog-standard part of Genetic Algorithms -- replace "robots" with "solutions" and you have a very basic introductory description of what Genetic Algorithms are. You copy successful robots into the next generation to simulate "survival of the fittest"*, you randomly combine the networks of successful robots to simulate genetic crossover in sexual reproduction, and you make random changes to simulate mutation.

At no point does it suggest that they deliberately added any specific behavior. The only thing it attributes to the scientist's direct actions is "The team programmed small, wheeled robots with the goal of finding food" which just means they made their food-finding point system the basis for the Genetic Algorithm's "Fitness Function". Selecting the particulars of the GA is a deliberate choice, but after that it's all just random changes filtered through the fitness function, just like real-life evolution.

A little common sense helps to cut through the bull-garbage here.

Lol, yeah. What does "common sense" have to do with not knowing what the scientists were actually doing?

Call me when a robot runs over to a black ring emits the "I found food light" duping the rest and then secretly running over to a blue light while the other stooges mill about wondering why some dumb robot said it found food here.

Give it a couple hundred more generations, and I wouldn't be surprised to find that behavior come up. Or other ones that may come as a complete surprise.

* Well and because keeping the current best solutions around helps prevent regression.

A more advanced experiment...

[Baron_Yam]

I'd love to see the robots given hunger, thirst, and a sex drive. Make 1/2 the robots girls with red LEDs and 1/2 the robots boys with blue LEDs.

Make the food and water 'power', and give them the ability to 'harm' each other by draining power.

The girls would have a higher resource requirement to reproduce.

It'd be interesting to see over many generations what relationship patterns form between the same and opposite sex.

Re:A more advanced experiment...

[muckracer]

> I'd love to see the robots given hunger, thirst, and a sex drive. Make 1/2
> the robots girls with red LEDs and 1/2 the robots boys with blue LEDs. Make
> the food and water 'power', and give them the ability to 'harm' each other
> by draining power. The girls would have a higher resource requirement to
> reproduce. It'd be interesting to see over many generations what
> relationship patterns form between the same and opposite sex.

I can tell you:

First the girl robots would seductively blink their red LED's at the boy
robots.
The boys get all silly and start swarming around the girl trying to get close
to her.
The boy with the biggest hydraulics then gets the girl and they roll off into
some dark corner for playing with their cross-over cables and rolling around
swapping oil and stuff.
A few weeks later the girl robot will barf neon-green liquid all over the
place, especially in the mornings. While she'll get increasingly cozy among
lots of spare-parts the boy robot is frantically rolling around at 2am trying
to get some yummy special machine oil, all the while freaking out about how
his care-free life with his roboddies is now over.
A few months later the clatter of little wheels is heard and the now coupled
robots take countless pictures with their integrated cameras and send them via
infra-red to every other robot that was built on the same assembly-line. They
will not notice the deliberate Out-Of-Memory errors of the recipients.
So far so good, until eventually the two start chirping at each other about
the little things of robot-life. The woman will harp on the partner bot, why
she always has to wipe off the nasty-smelling exhaust-buildup's of junior,
that she has not seen a metal polisher in months to make herself shiny and how
she should have anyway listened to her Mom warning her not to get involved
with a guy with obviously inferior software.
The man robot, OTOH, will simply bypass the receiving sensory circuit and
refuse to send ACK packets, all the while playing robot wars with junior and
generally doing, what a robot must do to make a living.
Eventually they'll settle into their daily routine as the years go by. Despite
their outward differences they now wouldn't want to miss each other anymore as
their hinges start squeaking, rust appears and the last firmware update is but
a distant memory. But as they prepare for being disassembled Junior...now all
flashy with his LED's...rolls out into the world to link up with a girl of his
own and start it all again...
The EOF

A good day for Marvin Minsky

[smchris]

I still think, "If we build the hardware, consciousness will come" is a stupidly inefficient imitation of evolution at best.

This is *my* human food

[rossy]

Couple this with the robots that eat organic matter on the battle field... and the throwable robots.... will they learn to kill for food?

No hyperref to the article?

[Argos Avatar]

Years ago when i discovered /. the articles had hyperlinks to all that was relevant to them. Nowadays there is a sentence such as:

"detailed in a PNAS study published today." Without any reference whatsoever to the paper itself. I checked PNAS's today's table of contents and found no such article. It must be there somewhere, but i am losing time to find it. Where is it? Shouldn't it be hyperlinked in the article itself? Who are the authors?

And after 115 replies no one seems to have mentioned the original article.
That's how deep the discussion has been so far.

You should be ashamed of yourselves.