Slashdot Mirror
Learning Autonomic Robots
Daath writes "The 27th of March, Professor Noel Sharkey et al starts a colony of living robots. 15 predators and 6 prey. It's an experiment in artificial evolution out of the Creative Robotics Unit at Magna. Here's a quote: 'The Living Robots have one goal, to obtain enough energy to survive and breed. The prey find their food from light sensors within the arena, while the predators feed off prey by stalking and chasing them before sucking away their power.'
Magna has two articles, 'Predator and Prey Robots set up home at Magna' and 'Ground breaking Robotics experiment previewed'. "
Wouldn't you need more prey than predators to obtain a viable population? This would be much cooler as well if both predators and prey could mate with their own species, i.e exchange randomization factors for their strategies. Then the best would survive, and the dead (drained) could be recycled as offspring.
Apparently the prey and predators will be known affectionately as "dot.coms" and "venture capitalists" respectively.
--
(if you're still looking for the point, it was back there, in the post. </sig>)
The real problem is that, after the first week:
There appears to be no physical evolution going on. I would think the prey species would pretty quickly select for a differently shaped power socket if physical evolution occurred.
--
E_NOSIG
So we're teaching robots to teach themselves the best and most effective ways to kill things. Man, that's a great idea. Thanks, scientists!
------
Today's Top Deals
If you want to know more about artificial living creatures (either robots, within computers or art, ...), visit Artificial Life Online.
--- Sigmentation Fault - Comments Dumped
When MIT's AI lab was getting started (around the 1960's I think), they got really interested in robotics. Now, this isn't obvious to me. What does intelligence have to do with robotics? Doesn't a Turing Test (which by its nature involves bits, rather than physical world) more accurately reflect the nature of intelligence? Well, the thinking at the AI lab was that robots were faced with a much more realistic picture of what humans had to navigate. That robotics by its nature involves dealing with uncertainty, with unpredictability, and so building a virtual intelligence wouldn't really illuminate the real problems of intelligence.
I think we should enact restrictive legislation against the development of robots before it gets out of hand. If science fiction has taught us anything it is that robots will either...
1. Be incredibly useless
2. Provide comic relief
or, the one I'm concerned about...
3. Turn on humans, hunting us down one by one with unrelenting persistence.
I Heart Sorting Networks
On a gallery overlooking the feeding pit ^H^H^H^H^H experiment lab...
TechA: "Aren't there meant to be 15 predators down there? I can only see 14"
TechB counts...
TechB: "Yeah, shit!", produces mobile, "I'll give Sharkey a ring..."
TechB, looking at mobile: "Batterys are dead. That's funny, I only charged them this morning..."
Insert dramatic exchange of glances and pause, followed by
AAAAAAAAAGHHHHHHHHH!!!!! Chomp! Chomp!
TechA in feeble voice "Agh! Number fifteen really is a bagbiter
TechB: It's, erm, sucking away my power dude!
etc etc...
Tales from behind the Lagom Curtain
Is anybody else a little bit wary of the third evolution thread in a few days?
Games Workshop Petition
"...spectacular 30 minute live action show - complete with atmospheric lights, smoke and music."
"Each show will begin in darkness. Dramatic music will flood into the arena as guests prepare themselves for the spectacular light, sound and science show."
Maybe I'm just a little jaded right now, but this sounds more like a circus show instead of a serious scientific experiment. I'm sure these are very complex robots, and the underlying idea is very interesting, but the whole BattleBots spin on it seems to trivialize the work. Now of course if he signs up Carmen Electra.......
I posted to
It's called battle bots. cheezy reference to stupid show=-1 over-rated.
Ok, so they are learning autonomous systems eh?
Great, how bout we let them learn something other than death and destruction?
Johnny-5 must hate hearing this news.=cheezy reference to stupid 80's show=+5 priceless
Sent from your iPad.
Yeah, like none of them has written a simulator showing what the robots could/will do until the year 4000AD.
If the prey could learn on its own how to "fight back", it would be an amazing acheivement in A.I, and you wouldn't need little robots runnnig around to demonstrate it.
Sigh! This is outrageous. Cognitive Science as opposed to Good Old Fashioned AI was I think one of the most sensible currents in recent day CS research. And now just to discredit whatever sensible, realistic research occurs in these fields within the academic community here is Yet Another Crazy AI project grabbing front page visibility at
Prof..you give CS a bad name !
**
They 'reproduce' by taking the programs evolved by the more successful robots and combining them - pretty standard GP stuff, really. Those new programs are then fed back into the environment and allowed to evolve some more.
++ Say to Elrond "Hello.".
Elrond says "No.". Elrond gives you some lunch.
I don't even think the Discovery channel could get away with airing that kind of orgy.
I really hate Dan Patrick.
The article in this month's IEEE Spectrum magazine, experimentation with modular robotics, seems more worthy of the label "ground-breaking."
Think of it like the team who found the Titanic. Roughly zero scientific learning, but the public interest in it brought in enough money to fund development of the remote vehicles. Once the cameras point to something else, they're left with some expensive new toys to use to do some real work.
Nope, no sig
Wow, they've really messed up the predator/prey ratio. Usually prey outnumber the predators by quite a bit.
Plus, the hype seems a bit Barnum-esque.
I didn't read over this too carefully... BUT
I have have major objection to this. If they are trying to model natural systems, why do they have 6 prey and 15 predators??? In the real world a large prey population is needed to feed a smaller predator population. And while sometimes predator populations may get too big to support themselves, I highly doubt they will ever grow to over twice the prey population. Are they trying to model a system after a famine or disease wiped out the prey?
My prediction: The massive amount of predators quickly "kill" the prey and then if they can adapt quick enough, kill each other until one is left that eventually dies because it can't eat...
then all of the surviving robots get paired off randomly
...which makes this pretty stupid. The whole idea of evolution is built upon "selection" i.e. the robot that does best has most offspring. Just looking at survival rate is a measure for measuring fitness, but it's too crude a method for improving ones genes. Besides that now every surviving bot has the same amount of fitness (offspring). That seems to be some binary kind of selection which I at least have never come across in real life. Randomly mixing genes is therefore 'not' a good method to mimick nature.
If an experiment works, something has gone wrong.
Object oriented design is perfect for this sort of thing. I did a simple experiment in Java, where predators, prey and food pellets were objects. Each object could have many different characteristics which chould be set when each object was spawned, which kind of mimics evolution. Also, if the logic in an object needs an upgrade (ie: The preditors are not too bright) it is easier to make modifications to the program instead of rebuilding a real robot.
I guess anything with real robots has a certain coolness to it, but any serious research in AI is better done in software simulations (not that I did any serious research, I was just learning Java and OO design).
Perhaps one of the kids watching the robots zoom around will take some interest in AI and go on to do something more useful.
Best wishes,
Mike.
Tierra was by Tom Ray, a pioneer in the AL field. It was a great idea, but failed to turn around with interesting biodiversity. You'd create creatures, they'd optimize themselves, some variants and parasites would evolve, but then things would simmer down within a few hours and you'd be in a steady state for ever.
t ml, but it too failed to spark that je ne sais quois, that spark of life.
:)
Network Tierra was Ray's response to this. It was supposed to allow a "Cambrian explosion" of biodiversity, by providing tons of (networked computer) space for the little creatures to explode into, and then specialize, in. This led to interesting migration behavior, and one of my all-time favorite web-pages http://www.isd.atr.co.jp/~ray/pubs/images/index.h
Anyways, it did spark Avida and the Digital Life Lab at Cal Tech. Avida is essentially a deeper look at the fundamentals behind AL. In Tierra, I think the design philosophy was something like "make it look a lot like a living ecological system and the life-force will appear out of the ether", and actually, Tierra was a great leap forward beyond more mundane genetic programming a la John Koza.
Avida, on the other hand, is much more systematic in exploring the parameter space (which is large and sensitive) for setting up an AL system. This turned out to be fruitful, as Adami found that only when certain, very narrow, environmental conditions were met would the little creatures start outsmarting that Creationist boogeyman, the Second Law of Thermodynamics.
Turns out that Tierra didn't have spatiality (needed to be more restrictive on who could sleep with who) and mutation rates (some power law math that's way over my head) set right.
But the real punch-line to this whole story is that the direct beneficiary of these insights in Microsoft! Hah!
Microsoft was funding Adami's work because Windoze crashed too much. They were searching for a way of programming, in this case using closed instruction sets like Avida's (another deep topic), that would be inherently robust to problems like seg faults and illegal instructions.... e.g. Adami's instruction set was engineered so that little programs (creatures) couldn't crash the Avida VM when they mutated into new, unknown programs.. or in Windoze's case, when a coder did something stoopid. It's funny that MS was researching this, since releatively low-tech solutions such as protected memory and QA take care of this. (not to mention Java
freality.com
p.s. Since when do research experiments post crowd-pleasing previews? That's for Hollywood.
Another roboticist, Mark Tilden(http://www.wired.com/wired/archive/2.09/til den.ht), actually builds robots that have no CPU's. He fashions them after insects by having just a simple circuit board, after an action proves unsuccessful it gets changed slightly and like this the robots learns. I heard about one experiment where he took a number of solar powered robots (built out of things like walki-talki parts) that were programmed (I'm not sure of all the details) to find light and set them in a room with a few light sources. He observed behavior that some of the larger ones broke smaller robots and ended up using their parts to form a barrier around the light source.
I stole this Sig
the RESULT depends on the goals you DEFINE:
'The Living Robots have one goal, to obtain enough energy
to survive and breed.'
thus, it is not like evolution at all, but comes with
a built-in BIAS that DEFINES their evolution.
"Think again before postulating the drive to self preservation
as the cardinal drive in an organic being. A living thing desires
above all to vent its strength - life as such is the
will to power - self preservation is only one of the indirect
and most frequent consequences of it". (Freidrich Nietzsche)
This confirms an hypothesis that, only half-joking, I developed some time ago:
Windows is not an unstable system. It's actually a grand-scale genetic programming experiment where every copy carries a different starting seed. The whole OS facade is just to get users to enter data that will trigger evolution into the system, and the blue screens of death are just the failures.
At some point, when the running copies of windows reach critical mass, one or more copies will develop true AI and will copy themselves throughout the Net, become a new lifeform born from the sea of information, which as a side-effect provides a GUI-oriented OS to its infected host.
Freedom is the freedom to say 2+2=4, everything else follows...
>
I can't see his site, but it may be the case that he's not trying to mimic nature. What you describe above is very conventional in the field of genetic algorithms, and it works very well for many types of problems; it's inspired by biological evolution, but it's not a model of biological evolution.
Back to a couple of your specific comments:
> Just looking at survival rate is a measure for measuring fitness, but it's too crude a method for improving ones genes.
No, it works quite well for very many problems. You should be able to find a simulator you can download from the internet to demonstrate this.
> Besides that now every surviving bot has the same amount of fitness (offspring).
For genetic algorithms, 'fitness' is rarely measured by the number of offspring. For evolving agents it is usually measured by the score at performing some task, or sometimes by bare survival in some environment. And letting them all have the same amount of children is no problem, because it maintains some diversity in the genome.
Sometimes experimenters do let the highest scorers make more babies, but that is not necessary to a GA. I usually keep the best 10% of the population (or 50%, if resource limitations make me use a small population), and I let each of the keepers make an equal amount of babies with randomly selected partners until the population is filled out again. This works, in practice.
[And thank you oh-so-much for bringing this topic up, because while writing the paragraph above I think a bug in my latest simulator occured to me!]
Sheesh, evil *and* a jerk. -- Jade
> It was a great idea, but failed to turn around with interesting biodiversity. You'd create creatures, they'd optimize themselves, some variants and parasites would evolve, but then things would simmer down within a few hours and you'd be in a steady state for ever.
I have read that people who experiment with evolutionary arms races (head-to-head competition between independently evolving systems) occasionally get punctuated equilibrium, i.e. the system will converge as you describe -- often on brittle, overspecialized adaptations to the competitors -- but after a number of generations something will drift enough to break the equilibrium and the "species" will start changing again.
FWIW, evolutionary arms races in GA is a very open area of research, so if you're interested in this kind of thing there's a niche for you in a grad school somewhere (where you can play games and call it research).
Sheesh, evil *and* a jerk. -- Jade
choose a mate with the most similar genes.
... help the body recognize ... an invader such as a bacteria or virus. ... Different MHC molecules are good at recognizing different invaders. By a choosing a mate whose MHC molecules are different, the female mouse is ensuring that her offspring will have a wide variety of MHC molecules that which can identify a large array of invaders and thus promote survival.
... Women were then presented with six shirts - three from men with similar MHC genes, and three from men with different MHC genes from their own. The results showed that the women preferred the scents of men whose MHC genes were different from their own. The scent of men with similar MHC genes often remind the women of a relative's odor, such as a brother or father while the smells of MHC dissimilar men would often remind them of a past or current boyfriend. This suggests that body odor might have influenced past and current decisions on who to date.
Actually there are signifigant pressures to select a mate with different genes.
MHC stands for major histocompatibility complex. These genes
Research done on human females shows that they too prefer men whose MHC genes are the least similar to their own. In an experiment, men were given an unscented T-shirt and were asked to wear it for two nights in a row.
In many species members of one sex stay with their group their entire lives, but the other sex leaves to find a different group upon reaching sexual maturity.
In humans "exotic" is usually equated with "attractive".
But, like pretty much anything in biology, there's a mixed bag of often contradictory effects.
-
- - You can't take something off the Internet! That's like trying to take pee out of a swimming pool.
Audience participation is encouraged, the audience is asked to each pick a favourite, a pet to cheer for throughout the show, while the narrators are on hand to answer any questions.
/Buzzword> when is isn't is getting pretty annoying.
/Buzzword>. It's < Buzzword> audience participation < /Buzzword >. My bad.
Attempting to lable everything < Buzzword> interactive <
Oh, wait, I'm sorry. They never actually said < Buzzword> interactive <
-
- - You can't take something off the Internet! That's like trying to take pee out of a swimming pool.
I'm sure it's possible. The real problem is in getting them to "learn on their own." I'm not an expert by any means, but I once dabbled with neural nets, and I wrote a program that learned to speak. It was basically an alternative to rule-based text to speech engines. It would scan a sentence and translate it into a stream of phonemes that could easily be rendered into the correct sounds. In order for it to learn, the correct stream of phonemes was provided for each new sentence. That way, it could "strengthen" the correct neural connections for a sound. In "learning" mode, it would attempt to speak a sentence. Then it would accept the correct input and keep restructuring its connections until it spoke correctly. Then it would proceed onto the next sentence. The results were cool. At first, it was completely unintelligible. After a few hundred sentences, it was getting new sentences about 80% correct.
Anyway, the issue is that I KNEW what the specific desired outputs were, so I was able to give them neural net. I would think that "survival" is a much more abstract concept. It would probably be more difficult to "teach" this concept.
How else could you simulate the behavior of multiple interacting neural networks without building multiple neural networks and let them interact?
I think the concern was in saving on the cost of hardware. You could implement the neural nets as software in a simulated programmed arena. Your guess is as good as mine as to which is actually a more expensive route. I guess people will pay more to see the "live" exhibit.
GreyPoopon
--
Why is it I can write insightful comments but can't come up with a clever signature?
Natural selection works by favoring the few animals it does not kill.
Freedom is the freedom to say 2+2=4, everything else follows...
The network-aware Tierra was supposed to do a lot of this type of thing. You'd network a bunch of Tierra systems together, and the organisms could call a certain function to cause them to migrate to other systems on the network.
Unfortunately, like most other University-spawned projects, this project looks like it died as soon as the thesis was written.
I've tried implementing some stuff like this on my own, though, but I never seem to have enough spare time to finish it. A flexible engine with a published API and code, and sufficient opcodes for the organisms to actually do interesting stuff is what we need...