Slashdot Mirror
Microchips That Evolve
An Onymous Coward writes: "A scientist in Britain has apparently developed some genetic algorithms that run on microchips that can alter themselves, in other words, hardware that evolves. Pretty cool. Check out the column by Paul Gilster." The article contributes some background I'd never heard before regarding FPGAs, and even mentions the dogged Starbridge. A short but fun read.
Darwin did no such thing. Is selectively misquoting from a book which is freely downloadable the best creationists can come up with these days? Here's the correct quote, from "The Origin of Species", chapter 6:
To suppose that the eye, with all its inimitable contrivances for adjusting the focus to different distances, for admitting different amounts of light, and for the correction of spherical and chromatic aberration, could have been formed by natural selection, seems, I freely confess, absurd in the highest possible degree. Yet reason tells me, that if numerous gradations from a perfect and complex eye to one very imperfect and simple, each grade being useful to its possessor, can be shown to exist; if further, the eye does vary ever so slightly, and the variations be inherited, which is certainly the case; and if any variation or modification in the organ be ever useful to an animal under changing conditions of life, then the difficulty of believing that a perfect and complex eye could be formed by natural selection, though insuperable by our imagination, can hardly be considered real.
Darwin forsaw this line of objection and disarmed it before he even published. Using this argument nearly 150 years later, and claiming that Darwin believed it a fatal flaw in his theory is intellectually dishonest in the extreme. You, sir, should be thoroughly ashamed of yourself.
More modern devices are harder to destroy. If you mis-configure your I/O pins then you're in trouble, but within the array devices like the Xilinx Virtex seem much harder to destroy (I was chatting to someone who was trying just that, but failed without using the I/O pins).
As a flip side, here's a paper about destroying less secure FPGAs and ways to prevent it.
-- Michael
In rereading this, I think it would have been much stronger if I simply omitted the last four paragraphs. They are true to some extent, but they're really stuff I've seen other places -- relatively weak, alarmist thinking.
If I had it to do over, I'd pull those last four paragraphs out... they weren't the main point anyway.
Right, that is exactly it. The circuits in that original article were taking advantage of electrical effects that we do not understand. It is doing something entirely new, and it is very possible that no human designer of any intelligence can ever even understand how the circuit works, much less design a new one.
And not all FPGAs will run this circuit either. Only certain ones will, and apparently there's no easy way to determine which ones will work.
Again extrapolating, what happens when we start evolving one-off tools whcih we don't understand and can't duplicate? If the tool did something important enough, it would become instantly precious. It would be magic..... the equivalent of a holy relic.
Now, that's a big claim. Most important EVER?
Consider: no other technological progress to date has ever changed the fundamental nature of Man and his tools. We are the Tool-Using Ape, and all of our technological toys and bombs are just extensions of the first person picking up a stick. He/she could use that stick in many ways, from agriculture to communication to personal hygiene to warfare. Essentially all of the other tools we have ever invented are extensions of that original concept -- specialized forms of stick.
Our sticks have gotten very complex, to the point where most of us navigate in ignorance through a world of astonishing complexity. Very few of us understand something so simple as indoor plumbing, but we are thoroughly dependent on it. This is all old news and we are well used to it -- if the plumbing breaks, call a plumber. If the car breaks, go see a mechanic. If the computer breaks, post on Slashdot from another one and hope to god you pick out the truly informed answer from the barrage of ignorant suggestions. But I digress. :-)
This is where this new technology is important. This isn't just a stick anymore. With all of our existing tools, someone, somewhere invented them, and someone somewhere knows how to fix them when they break. No matter how complex or intimidating the tool, if you are willing to devote enough resources to the problem, you find an expert who understands the technology you are using and can fix it. In the cases of unforeseen interactions between complex systems, like computer software, it may be necessary to consult many experts, but ultimately all of the tools we use are sticks, someone invented them, and someone can fix them.
And that relationship is what is now changing, for the first time since that first person picked up a stick and used it to scratch, we are starting to use tools that we did not invent and that nobody, anywhere understands.
This is an incredibly big deal. We are talking about starting to use tools that are essentially not deterministic in their uses. Different versions of the same tool may work better than others, even though nobody quite understands why.
Consider some of the ramifications: Perhaps your specific tool will have learned how to trace cables or diagnose car engine trouble better than other people's, and you can charge more for your services because of that. Or maybe your gizmo calibrator just doesn't work properly unless you hold it upside down. And what if your computer develops an amazing ability to pick stocks if you consult it after midnight on Tuesday, but sinks into a blue funk and refuses to even talk to you if you fail to address it as "Lord Master Xerxes of Apollo"??
What we are talking about here is sticks with intelligence. Admittedly this is just the very beginning of what, I presume, will be a long slow (dare I say it) evolution of the tools. Taking things out to an extreme, it is not inconceivable that we might end up with tools that have moral or ethical reservations about being used for what we want to use them for. Or we might see guns that *like* to kill and look for excuses to do so. 'It was my sidearm, Your Honor, it just went off unexpectedly.'
Arthur C. Clarke's most famous quote was, "Any sufficiently advanced technology is indistinguishable from magic." I really am starting to wonder if we're not going to end up in an age of tools that are so smart that using them resembles incantation and gesture -- in essence the casting of spells.
What happens when we become dependent on tools that NOBODY understands or invented?
In a weird sideways tangent, I am reminded of a wise observation that, I believe, Louis L'Amour had about the early war in America between the settlers and the Indians. He pointed out that the Indians lost not because of being outgunned, or because of the sheer numbers of European colonists, but rather because they became dependent on firearms with which to fight, and could not themselves manufacture them. That meant that to fight the colonists, they needed the colonists as arms suppliers. They couldn't drive out the colonists without the colonists' help. Their loss of cultural identity was inescapable as soon as they picked up guns.
If we flesh intelligences end up at odds with silicon ones, and we are dependent on them and cannot duplicate their services with our old sticks and rocks -- we will lose.
Like I said, this is a big deal.
The author of this article seems to have little idea of what he's talking about.
The UK experiments used FPGAs, and Starbridge uses FPGAs, but the novelty in the experimental work lies in the use of genetically evolving cell interconnection in the FPGAs, whereas Starbridge's FPGAs are reprogrammed on-the-fly in a completely deterministic manner, not unlike the Crusoe. The Starbridge FPGA hardware isn't mutating genetically, it's merely reconnecting functional units under the control of something like a JIT analyser/compiler to deal in the best way with the incoming instruction stream.
The author of the article has put two and two together and got five, making FPGAs look dangerous.
FPGAs are in thousands of products all around us. If they were mutating, we'd know about it PDQ!
"The question of whether machines can think is no more interesting than [] whether submarines can swim" - Dijkstra
How did it work? Magic."
How silly of me to think it was capacitance or inductance or both.
I see even classic Slashdot is now pretty much unusable on dial up anymore.
Yeah, I've seen source code that worked like this too (remove an unused variable and it stops working), but strangely enough we didn't consider it a great leap forwards... I seem to remember we burnt it and danced on the ashes.
Also:
"Hey, my hardware has stopped working..."
"Sorry, it must have evolved, but we can't fix it cos we don't know how it worked. Have you tried resetting it back to primordial state and waiting 5 years for it to evolve back to useful again ?"
T
I spent a lot of money on booze, birds and fast cars. The rest I just squandered. - George Best
First, this is old news, as many have pointed out several different sources reported on this in '98. This particular story is actually less informative.
Second, the chips don't freakin mutate and reprogram themselves. An external agent programs the FPGAs with a 'genetic' string and tests that capabilities of that string, then tries other variations that were based on the previous generation. It is not as if these things just magically mutate themselves to do a particular task, there is much external intervention.
Thirdly. When I last heard about this guy he was off exploring ways of making the chips more robust. Because the algorithms evolved appear to depend heavily on the analog, non-linear nature of the components in the chip, they were extremely sensitive to temparture variations. Also some programs evolved on one chip lot would not work on another chip lot, even though they were functionally identical chips. He was looking into ways of testing fitness at a variety of temperatures and using different chip lot in the process, perhaps producing a result less efficient than his original attempts but more robust.
And now we have heard nothing new from this guy for the last 2.5 years. My guess is that this was just too good for the specialized chip manufacturers to pass up. It's extremely promising technology - but I betcha most of it is patented by now.
-josh
Once again Slashdot has missed the boat.
Actually, it was posted here before.
---
satire, n: 1) witty language used to convey insults or scorn; 2) a form of humor lost on most slashdot moderators.
You don't need a 'genetic' FPGA for this, just a plain old FPGA. Any (large rnough) FPGA is reprogrammable to do (almost) anything you want. In fact, kits with FPGAs on PCI cards are fairly inexpensive these days -- with a little hardware design knowledge, you can easily have what you described, at home.
The novel part of this article, though, is that FPGAs 'evolve' by reprogramming itself to a better version, over and over.
----------
Never underestimate the bandwidth of a 747 filled with CD-ROMs.
'(1991) Can you imagine the devastation that could be wrought by malicious thugs when everyone's home computer is connected to everyone else's via this "internet"? I urge slashdot readers to boycott any PCs which use this dangerous new technology.'
The point I'm making here is that any technology, whether it's GPFPGAs, the Internet, Guns, Solar power, or Hydrogenization of vegetable oils is capable of both benefit and harm. The view that anything potentially dangerous should be shunned is detrimental to everyone. The only safe way to handle potentially dangerous technology is to understand it, and spread that understanding. You can rest assured that if the open scientific community avoids a technology out of fear, the closed criminal community will only take more dangerous advantage of it. Understanding and knowledge is the only safeguard from the harm that can be caused by technology.
Ideology breeds Hypocrisy. Just how much is up to you.
Actually if you read the longer article posted a few days back about this guy, you'll see he hopes to _improve_ reliability. How ?
:), you could
With GAs. If you make the "environment unfriendly", then the result of your evoultionary process should be a very survivable chip. simulated biological evolutinon of fault tolerant and healing systems. Afterall, the human body is one of the most advanced healing system in modern medicine.
If this sounds way too blue sky, its easier to get a little more specific. Why not pick some "fault" scenarios to model, then introduce those as conditions during the evolutionary process. For instance, if you want to develop a chip that could survive internal/external cache errors (are you listening, Sun ?
create a cache section which emits errors at some random rate/interval, and watch what the chip does to work around it. It might build another cache.
The important thing in doing this though (as the researcher found out) is varying your conditions a _lot_. When he first ran this on a single FPGA, the design that eventually evolved would only work on that specific chip (not that specific model, the specific physical sample he evolved it on!), at a specific temperature, with a specific host program, power supply, etc.. i.e. it over specified itself and could not function in other environments.
For his later experiments he started using multiple examples of the same part number, but from different foundries. He also adjusted environmental conditions somewhat. The hope was to create a more "survivable/versatile" design.
Fascinating stuff, IMO.
My opinions are my own, and do not necessarily represent those of my employer.
This is true for man-made circuits too, my friend. When you design a circuit, you must still test it against every state you can think of, right? So simply subject GA-designed circuits to the same tests you would subject your own designs to, and you're no worse off than I am right now - after all, my region boasts one or two human-designed nuclear power plants for which there undoubtedly exist untested states (and also, I might add, states that have been tested and are known to cause catastrophic failure).
Any sufficiently well-organized community is indistinguishable from Government.
I remember reading an article about this kind of thing, probably in Discover, a long while back. I think they were trying to breed a minimal gate circuit that would spark when the system heard the sound 'No' or something. Using evolutionary algorithm techniques, they got some phenomenally low gate counts... the trouble was these systems were incredibly sensitive to the environment. It's a general problem with letting circuits breed themselves, they'll end up taking advantage of the oddest things, like RF interferance, or (don't remember the term, the way electricity in one wire is likely to cause a current in a parallel wire) and that once you move the circuit from your evolution/test harness, or even change the temperature or RF shielding, there's a good chance nothing will work. The solution is to build more complex test harnasses and test for how the circut responds to a variety of environments, but that starts to get more and more expensive.
I've heard it said that in the future programming will be more like gardening than building up with legos, but I dunno...
--
SO YOU'RE GOING TO DIE: The Comic for Dealing with Death
Imagine the philosophical problem this creates.
What if you build a critical system for, say,
a nuclear power plant. It works and
works well, but you don't know how to explain it.
Can you implement it? Can you rely on it?
I play with genetic algorithms and I believe that the author's concern is based on a misunderstanding of the nature of the beast.
Modern software systems are incredibly complex, to the degree that no one even considers the possibility of conducting formal proofs as to the viability of the algorithms involved (I'm not talking about an algoritm like in an encryption subroutine, of course, but the program itself as an algorithm).
Genetic algorithms, however, work code much more extensively because they iterate through an unimaginable number of combinations. This leads to a programmatic robustness not generally found in traditional programming techniques.
Can anyone guarantee that any given GA will find all of the possible failure points in a system? Of course not. A GA isn't even guaranteed to find the optimal solution to a given problem (it's more like "a good enough solution in a reasonable amount of time").
Genetic programming (allowing software to evolve the actual programmatic structure of the application) isn't my deepest interest - I'm far more interested in artificial life. Bowever, both of these areas have the potential so revolutionize not only software (and, from this article, hardware) as a disclipline, but also make possible systems that can barely be imagined now.
(email addr is at acm, not mca)
We are Number One. All others are Number Two, or lower.
(email addr is at acm, not mca)
We are Number One. All others are Number Two, or lower.
--The Sphinx
A nervous system? My system can get quite nervous from time to time, ending in a kernel panic :)
Link to New Scientist article
It talks about the unconnected cells and the way it was trained, and the fact that the circuits only worked at one temperature(!)
-WolfWithoutAClause
"Gravity is only a theory, not a fact!""My only statement to that would be that from the beginning of time, the creation (in the case the "thinking chip") has never been been able to take over the creator (in this case, man)."
What, never?
Cars have never gone faster than a human? Aeroplanes don't fly better than a human? Chess computers don't beat the person that programmed them (hint: most of the time chess computers DO beat the person, if the program they write is any good at all)?
For that matter do students ever exceed the capabilities of their teacher? Of course.
The old maxim, "a stream can never go above its source" is WRONG. You CAN make water go higher than its source (check out water hammers).
-WolfWithoutAClause
"Gravity is only a theory, not a fact!"http://slashdot.org/articles/99/08/27/1238213.shtm l
Heck, this one's more informative.
Donny
What if you build a critical system for, say, a nuclear power plant. It works and works well, but you don't know how to explain it. Can you implement it? Can you rely on it?
;) - then you cannot with certainty implement this solution.
No. Because a state may exist in future that has not been tested. This 'unexplained' circuit may fail when it encounters this state - there is no way to identify it (maybe). Unless you can gaurantee that the circuit has seen all states/conditions and you are capable of proving a desired result in every instance - in this example that would be a 'not-meltdown'
Just think of the ramifications of evolving computers...think, oh, I don't know, the Matrix or something...
If you select for Matrix-style AIs, then yes. If you base your selection of the best of each generation on their ability to multiply two numbers, then all you'll ever get is a multiplier. A very efficient multiplier, probably, but still just a multiplier.
What we are dealing w/ is a finite set of possiblities here and true evolution is INFINITE!
I always thought the evolutionary possibilities for a particular organism were constrained by their environment. It's true that the environment is quite open and it's very hard to see what some of the possibilities might be, but some things can definitely be ruled out (no organism has photo-receptors tuned to pick up gamma rays, for example, because that wouldn't confer any advantage, and would be a waste of resources).
It's still natural selection, but in the case of these chips, we're controlling the criteria, and they're much narrower, that's all.
nVidia's been doing this for quite some time with thier GeForce chip...
Non impediti ratione cogitationus.
What if some of this hardware was to evolve into a self replicating machine?
Keeping
Imagine the task at hand is to navigate a simple maze--a maze with exactly one entrance, one exit, and no loops. Now imagine that it has the exit intentionally close to the entrance and the one path is intentionally circuitous and counter-intuitive. If the algorithm to solve this maze starts evolving based on mistakes and random variations, there is a really good chance it will dwell on the wrong solution. By evolving, there will be bits of the wrong solution left behind in it's algorithm.
This effect can be minimized by saving the state of the algorithm when the deviations occur and backtracking to avoid unnecessary calculations. This requires you know the proper solution. This is very easy with a maze, very hard with image recognition, sound recognition, and any other task which is not completely understood from the human point-of-view.
Adrian Thompson should do more trials on his evolving algorithm. There should be variations in his results. Two weeks is not a long time for a science experiment. He may get many unremarkable results, a few fantastic ones, and a couple of "just plain strange" ones.
I'll have you all know that I still have my appendix, thank you.
----------------------
Now, most of this depends on how complex the programmes can be for the FPGAs, but it would be nice to have a complete box loaded with cards of these puppies for specific applications. For example:
A piece of gaming software downloads additional morphing/rendering code to GFPGA card #1, while loading the game engine into GFPGA card #2.
Download updated compression algorithms as a compression co-processor.
Card #1 becomes a highly optimized search engine for your corporate site.
Card(s) #1 gets GIMP modules
Yup - it would be nice. But seriously, I don't see it happening at my house in the near future. Although, using cards #1-#4 for rendering near perfect Natalie Portman photos, and a fabber... Hmm....
I donate all spillover Karma to the charity of my choice... Ada was still a babe despite what people may say...
Now, imagine in the future you have one of these GPFPGA chips in your PC. What is to stop a malicious hacker (or is it cracker?) from planting a logic bomb in the very hardware of your PC ?
This technology seems like a retrograde step if we are concerned with the reliability of our hardware. I urge slashdot readers to boycott any PCs which use this dangerous new technology.
We had some yahoo give a seminar about this at my uni and he was gushing about how it had come up with a patented voltage regulator topology and how it would revolutionize analog design. Well, an analog design that depends on its environment is simply a sucky analog design. In practice, analog circuits have to work in the harshest, most variable environments. I don't think that us "experts" are useless quite yet.
I'm not a biologist, but if I remember correctly from my AI class, a Genetic Algorithm always has a fixed number of variables (similar to a fixed DNA length/number of genes in a given species). You, the programmer, still has to determine which variable adresses which part of the problem to be solved, and how (similar to gene sequencing?). So a GA that is supposed to distinguish between two signals, for instance, can't evolve all on its own into a solution for a different "problem", e.g. eliminating all those pesky humans running around. GA is just a technique for obtaining a better solution for a given problem from less ideal solutions. Rather similar to a farmer breeding cows with larger milk output/better feed utilisation from existing cattle, and quite unlike a frog evolving into a prince or some such thing :-).
Free, as in your money being freed from the confines of your account.
... here.
SIG: TAKE OFF EVERY 'CAPTAIN'!!
If the control system is properly programmed, then the chips cannot 'crash' and spiral into uselessness since the control system will go back a step unless the new system is better than the old one.
It is very possible for computers programmed using evolutionary techniques to do things that you don't understand how. It's what is known as an irreducible system - the interactions are so complex that after a few generations you cannot trace the changes back by looking at the finished product, you need a log of the evolution. For a full genetic program, this log could be 20,000 or more generations of 100 or more programs being tested - and that's a lot of log to go through to figure out why & how it works!
And as for the reliability issues - you don't run these things 'hot' in critical situations! You run them in simulation, throw problems at them and they create every neat solutions. You then take these solutions and combine them to make a very neat, very normal computer program which can be added to and debugged in the normal way. In terms of FPGAs you would use small, tested circuits to control parts of a critical system like any other small circuit.
These thing aren't meant to be used 'on the fly' - it takes a lot of processing time! They're used to create highly efficient normal circuits which can then be used in normal circuitry.
Beg:
Before anybody brings up the current movie which uses a message against a technology imagine the other side of the coin. Somebody using a comic book (which is what most of these movies really are) such as Teenage Mutant Ninja Turtles as an argument in favour of genetic engineering.
There are a lot of reasons to be cautious when extending almost any technology as well as reasons in favour of advancing the technology. Listen to evidence as opposed to propoganda or fiction and determine your own standings on it.
Chris Kuivenhoven is a thief, beware
"And get this: Evolution had left five logic cells unconnected to the rest of the circuit, in a position where they should not have been able to influence its workings. Yet if Thompson disconnected them, the circuit failed. Evidently the chip had evolved a way to use the electromagnetic properties of a signal in a nearby cell. But the fact is that Thompson doesn't know how it works."
Not bad for a new-born.
After being taught gah-gah goo-goo,
it goes on to prove the computational value of quantum mechanical entanglements.
We'll have to keep an eye on this toddler; (like that little girl in MIB with the two physics books).
It may just go on to prove human beings are as useless as egg-shells.
Ripley's believe it or knit.
Hmm ... instead of this evolving microchips thing, how about inventing memory chips that can duplicate itself. I know I for one wouldn't mind having DIMM stick or two that breeds and multiplies like rabbits. I hear a 1GB ramdisk calling my name ...
I've got some silicone that seems to have a life of its own!
--
(if you're still looking for the point, it was back there, in the post. </sig>)
Windows running this would
Turn your 1ghz box into a 386 that bluescreens
Upside
Linux running this would:
Create its own code for its own kernels
FreeBSD running this would:
spin off and create a port for something other than i386 arch
OpenBSD running this would:
Not allow anything to ever be installed because it may be deemed insecure
NetBSD running this would:
Create another port, then wait a year or two to let you use it
view the source Luke!
360 degrees of Karma
Generally, Asimov's robots weren't intelligent enough to get hung up on stuff like that. But he did a few stories with robots that had broken through to a sort of meta-level where they looked at the good of humanity as a whole...
If, on the other hand, they were given the materials to build an FPGA, and general instructions on how an FPGA works, perhaps then they could improve on the original design. Sort of like humankind does today through scientific research. We have been given a basic design (with many variations of the basic design - our bodies), and have been able to increase our life expectancy and quality of life through learning about ourselves, our world, our universe, etc. To claim this as being a vision of 'evolution' is silly. Let's remember that someone created the FPGA machines in the first place, humans. They did not just magically form themselves from a bunch of spare parts in a computer lab.
For instance, let's say this FPGA computer could alter it's FPGA (or even build new ones) to increase computational speed. Fine, I agree that that is within the realm of possibility. But it would be a mathematical impossibility for this computer to spontaneously generate a new piece of software or hardware that was outside of it's creator's realm of abilities that were bestowed upon it. Let's say we do not give this computer a device, knowledge, or method of communicating with any other computers. It cannot develop a communication protocol because it has no concept of such a thing. Plus, it would need to develop both the software AND the hardware, at the same time working together, to properly begin communications with other computers. So my arguement is that computers will not take over the world. They're confined to the knowledge we build them with.
When we finally get robots that can understand Asimov's Laws of Robotics, and we learn how to program them in, they won't be the altruistic platitudes the writer handed down to us. They're going to look more like this:
1. Make me rich.
2. Don't fuck with me.
3. Fuck everyone else.
--Blair
And of course there's no way around that.... No... Of course Asimov didn't spend the better part of his career writing books about unexpected effects and ways to sidestep and circumwent those laws..
</sarcasm>
Whilst we're on the subject, I've been thinking about Asimov's laws recently and decided that they are complete crap. How the hell is a robot supposed to determine what is human or not. In addition how is it supposed to work out whether any action could indirectly harm a human? Give it a book on Chaos Theory and all of a sudden, picking up a can of Coke in Utah too quickly will cause an earthquake in China. All very silly.
Just a thought. Whilst we were here.
This line intentionally left here to annoy you.
When a true AI is finally developed sombody is going to be smart enough to deeply ingrain Asimov's 3 laws of robotis (or something similar) into them.
That just about solves our problems doesn't it?
one thing i can tell you is you got to be free
This is what I waited for!
Ok, now that that these machines can realy reassemble themselves, let's giv'em the possibility to collect and produce their own resources. Construct an initial seed of nanobots, put them into a small rocket, send them to any planet that seems to be inhabitable for humans. There the bots would reproduce themselves with the materials they find on that paticular planet. Based on their inital "gene-code" they would be programmed to make architectural facilities for humans. Due to their evolutional design they could adapt to regional specialities (such as gravitational and climatic issues), i.e. make very thick walls where radiation is high etc. Just like techno-termites using their own body as building material for the anthill. You could fire some thousand seeds into space know, wait a few hundred years until the technoparasites made up a small colony for you somewhere and all you need to do is move in. (Do not forget to bring your coffemug, linuxbox, plant etc.)
a very humble simulation of this can be seen here (german)
Translation by google here
Conceptual Flash-movie here (click first link)