It’s always amusing to wander over to the Discovery Institute’s blogs, and see what kind of nonsense they’re spouting today. So, today, as I’m feeling like steamed crap, I took a wander over. And what did I find? High grade, low-content rubbish from my old buddy, Casey Luskin. Luskin is, supposedly, a lawyer. He’s not a scientist or a mathematician by any stretch of the imagination. There’s nothing wrong with that in the abstract; the amount of time we have to learn during our lives is finite, and no one can possible know everything. For example, I don’t know diddly-crap about law, American or otherwise; my knowledge of western history is mediocre at best; I don’t really speak any language other than english. I know some physics, but my understanding of anything beyond the basics is very limited. Even when it comes to the topic of this blog, math, I’m at best an enthusiastic amateur.
The problem with Casey, and people like him, is that they’re ignorant of a topic where they believe that they’re experts. Growing up, I was taught to call that kind of behavior not just
ignorant, but pig-ignorant. It’s a foolish kind of arrogance, where you believe that you know as much as people who’ve spent years studying something, even though you’ve never even read an elementary textbook. It’s like the dozens of people who’ve emailed my “disproofs” of Cantor’s theorem, when they don’t actually know what “cardinality” actually means.
In this instance, Casey is annoyed because a group of people at NASA used evolutionary algorithms to create a better antenna.
The fascinating thing about the antenna story is that
no one had any idea of just what a “better antenna” would look like. In fact, they wound up with something that looks like a paper clip bent into triangles. Let me repeat the key thing here: a bunch of
engineers wanted a better antenna. They had no idea what that
better antenna would look like. But by throwing it into an evolutionary
algorithm, they produced an antenna better than anything designed by a human being.
That’s pretty damned impressive, and pretty difficult evidence to
confront for anyone who wants to claim that random mutation plus selection can’t produce anything new. Of course, that wouldn’t stop someone like Casey: he, in his masterful brilliance, knows more about this than even the people who did the experiment!
So what’s Casey’s problem with this?
The presumption of evolutionary biologists, of course, is that these “brilliant designs” evolved by natural selection preserving random, but beneficial mutations. Engineers operating under such presumptions have thus tried to mimic not only the “brilliant designs,” but also the evolutionary processes that allegedly produced the designs. Biologic’s article notes that one success story of such methods was the case of NASA engineers who used evolutionary computing to produce a better antenna.
Did they use truly Darwinian “evolutionary computing?” The article goes on to discuss how design parameters were smuggled into the simulation, such that it really wasn’t a truly unguided Darwinian evolutionary scenario.
So what exactly can unguided Darwinian evolutionary computing actually produce? Probably not very much, but this is a research question that Biologic is attempting to tackle. As their research page says, they are exploring “fundamental laws governing the origin of information” by “building and testing computational models that mimic the role of genetic information in specifying functions by means of structure-forming sequences.”
In essence, he claims that the antenna is really designed, because the engineers “smuggled” information into the system, meaning that it’s not truly an “unguided Darwinian evolution scenario”.
First and most important, no one ever claimed that
evolutionary techniques like this are perfect simulations of biological Darwinian evolution. Casey is, as usual, battling against a
straw man. If you were to point out to the engineers involved that this simulation wasn’t a true simulation of biological evolution, their
response would be something along the lines of “Yeah, so?”
Second, nothing was smuggled in to the system. As Casey himself points out, they’re very open about the fact that they provided a lot of data. From their article:
How impressively it works, though, depends on what you were
expecting. You can’t fault the NASA engineers for choosing the automated
evolutionary approach when you consider the alternative–a pair of
needle-nose pliers, half a ton of paper clips, and a whole lot of wrist
strain. But if you really saw evolutionary computing as a high-speed
version of the process that produced all the jaw-dropping designs of
biology, well… you ought to be more than a little disappointed.Equally sobering is the likelihood that this striking
disparity–between the stunning things attributed to evolution and the
modest things we get by harnessing it–will persist.Two major limitations to evolutionary processes seem to assure this.
First, it turns out that if you want these processes to go anywhere, you
really do need to master the design principles specific to your
objective. You’d better believe the NASA team did their homework for the
task they were tackling–they knew what materials to use, they knew the
range of dimensions to explore, they knew what kind of geometric space
to explore, and they knew how to model the performance of any prototype
within those specifications. So the software they used was intelligently
pre-configured for this particular design task and no other.
As they said – to design an antenna using evolutionary techniques, you need to start with an understanding of exactly what you’re looking for, what range of space to try to cover with your mutations,
how to perform the mutations, how to evaluate different results
during the selection phase, and so on.
That doesn’t distract from the amazing outcome. A system based
on replication, mutation, and selection produced a better design for the antenna than the best designed by intelligent human engineers. And none of the engineers could have predicted that outcome.
Casey, as usual, is trying to play the intelligent design game of saying that evolution can’t produce information. From the standpoint
of information theory, as I’ve pointed out time and again, that’s just pure rubbish. Random processes, by definition produce huge quantities of information. In fact, the interesting thing about the
kinds of systems that make up living things isn’t how much information they encode, but how little.
A DNA molecule is an amazing thing. It’s a stable system for
encoding a quantity of information that is essential for the function
of life as we understand it. But from an information-theory standpoint,
it’s really amazingly sparse. It’s a double-helix, where each half of the helix contains exactly the same information as the other half. It’s got a substantial backbone, which is copied over and over and over down the chain. It’s really highly compressible.
But back to Casey. The nub of his argument is that there’s nothing
interesting about the production of this new antenna, because the “information” needed to produce it was “smuggled” in to the simulation. But if that’s really the case, then the question is, why did the engineers bother with an evolutionary process? Why didn’t they just
use their information to figure out what the optimal antenna geometry was? Engineers are known for being very down to earth, practical, results-oriented people. If they could produce the optimal solution
by themselves, they would. But the fact of the matter is, they couldn’t. They didn’t have the information that they needed to figure out
what the optimal antenna could look like.
So no matter what kind of stupid arguments Casey wants to make about information, the final fact remains: that the shape of the optimal antenna was not included in the input the the evolutionary simulation, but the simulation produced something superior to the best efforts of the intelligent, skilled engineers.
I don’t get the thing about information “smuggled in” at all: those parameters are merely mimicking the biological/environmental context of the evolving organism (or, more properly, organ), which needs to survive and thrive in a particular environment with particular parameters. If it violates those parameters, it dies: no point in wasting computer time on that.
http://www.mazepath.com/uncleal/reality.png
The greatest obstacle to understanding reality is not ignorance but the illusion of knowledge. Reality is not a peer vote.
Would the antenna benefit from a surface coating that was the geometric mean between the antenna’s impedence and the impedence of free space?
Drilling a hundred 5-micron diameter channels down the long axis of a solid foot-long Plexiglas rod cannot be done, but 7 microns is easy. If a thing problem is really a stuff problem you need a chemist not an engineer. If it really is a thing problem the best solution is not by the book, it rewrites the book. In the whole of human history across the entire planet not one deity has volunteered Novocain. It is a telling omission.
When “Intelligent Design” proponents talk about “intelligence” and “information,” they aren’t really talking about what scientists and mathematicians mean by these terms. They mean something more along the lines of elan vital, a magical force that can only be transmitted by those that already have it. So it doesn’t matter that the engineers did not know how to make an antenna that good–by their “laying on of hands” in designing the simulation, they have smuggled in the magical essence of intelligence. Conveniently, this means that there is no possible experiment that anybody can do to prove that mutation-selection algorithms are capable of discovery and invention, because any experiment requires some form of human intervention, and the ID guys will always insist that the experiment/simulation was contaminated by the “intelligent information” from the experimenters. And since the ID definition of “information” (or “specified complexity,” if you will) is so vague that there is no way to actually quantify it, there is no way to disprove this claim.
It’s so frustrating. My dad and brother are chemists and I went to school on physics scholarship, but eventually wound up an actor. Day to day, I’m surrounded by people, typically those without a hard science background, who stand in opposition to ideas and theories based almost entirely on two objections: 1) grasping the complexity of the theory is a lot of work and so they assume trickery or 2) it does not fit with their current belief system and is therefore dismissed out of hand.
I worked briefly with someone who was an evolution disbeliever. She constantly argued that, given enough time, she was certain she could disprove evolution. She just didn’t have the time. This was someone trained as an actor (not to dismiss that training — I have the same training — but I’ll be the first to admit it’s hardly the training you’d want as a foundation of good scientific exploration), who thought she could out-do any evolutionary biologist who ever lived, if she just had more free time. No amount of pointing out the illogic of her statement helped. She was convinced.
Small correction: you have ‘who’ve emailed my “disproofs”‘, which I believe should be ‘who’ve emailed me “disproofs”‘.
Casey, to me, seems an idiot because he argues against evolution without a better explanation. And it’s obvious to me that ht does that only because of his religious indoctrination.
Re Robert(#5):
I don’t agree that there’s anything wrong with arguing against some idea without having a suitable replacement for it.
Some things are just wrong. I can look at, say, traditional chinese medicine, and make a very strong argument that many of its “cures” are nonsensical – for example, the idea that long yuen can cure glaucoma, because the long yuen fruit looks like an eye. I can easily demonstrate that the idea that the fruit cures the problem doesn’t work. That doesn’t mean that I have any idea of what causes glaucoma, or that I have a better treatment for it.
Similarly, if you could show that evolution is impossible, I don’t think you would need to show that you had a complete replacement for it. If, for example, evolution really did violate the second law of thermodynamics, that would be a sufficient reason to reject it – even without an alternate explanation.
The problem with Casey is that his arguments are based on pure ignorance. He doesn’t understand what he’s talking about; but he’s been told various things which he’s chosen to
believe without actually understanding them. Then he builds up elaborate arguments based on the “facts” that he’s been fed on. Even if the arguments made sense, they’re built on “facts” that aren’t really true. But Casey isn’t capable of judging the truth or falsehood of the basic facts that he uses to build his arguments. He’s been fed facts by people he trusts, and he accepts them because of that trust, and because they confirm his strongly held personal beliefs. He’s never actually taken any time to understand the basic science that underlies his arguments – because he doesn’t care to do it. He believes that he’s already in possession of the truth. There’s no need for him to study, or learn, or experiment to confirm his beliefs. He just knows that they’re true. And everything else that he encounters is filtered through that arrogant, pig-ignorant lens: if it matches his beliefs, it must be true; if it doesn’t match his beliefs, it must be false. No investigation required.
Re #1:
The “information smuggling” rubbish goes back to some work that Dembski did with (I think) Robert Marks. They argued that, following on Dembski’s “No Free Lunch”, that true evolutionary systems were no better than random walks through a fitness landscape. Since it’s easy to demonstrate that there are evolutionary software systems that do perform significantly better than random walks, they needed to come up with some explanation of why those systems were able to do what Dembski had supposedly proven to be impossible.
The excuse they came up with was “information smuggling”. That is, some information about the structure of the fitness landscape was sneakily added into the system – and the result of that was that the evolutionary systems were really following a pre-ordained path determined by the smuggled information.
So, according to them, the evolutionary system didn’t actually do anything, except follow the instructions sneakily added to the system in the form of smuggled information.
It’s a really trashy argument, which is demonstrably wrong. The demonstration is when an evolutionary system produces a result that surprises the people who ran it.
There’s a famous example, which I don’t have a handy reference for at the moment; I learned about it in a speech given by Bill Joy.
A group at Sun Research was looking at using evolutionary systems for circuit design and layout. One of their results
consisted of two electrically isolated components – components that couldn’t influence each other. And yet, they did. It turned out that the system was exploiting a flaw in the manufacturing process – the insulation between circuit traces wasn’t as complete as they thought. So they got a result based on a physical property of their manufacturing process which they were not even aware of.
That’s not smuggled information. You can’t smuggle information that you don’t know. But the evolutionary
experiment exploited it – a randomly generated mutation
put two circuit traces very close together, so that they
influenced each other, and when the circuit containing that
feature was evaluated for correctness, it got the right results. They only discovered the manufacturing process by trying to understand why the circuit generated by their system worked!
Mark wrote
Yup, Dembski’s so-called “Law of Conservation of Complex Specified Information.” He’s also called it the “displacement” problem in criticizing the work with Avida — the “information” is smuggled in via the selective environment. Well, Duh! What evolution by natural selection does is create mutual information between selective environment and population!
Mark also wrote
That same sort of phenomenon turned up in Adrian Thompson’s (I think) work with evolving oscillators using field programmable gate arrays. The evolved circuits took advantage of stuff like parasitic capacitance to generate high frequency oscillating signals.
The kind of confusion (to put it kindly) Luskin (and Dembski and Marks) display is caused by treating biological evolution as a search process. While search is sometimes a useful metaphor for biological evolution, the technical search literature is of little use in analyzing bio evolution, and can be actively misleading. For one example, biological “search” begins with a population already at a ‘solution’ — a reproducing population is already on a viable node/peak in the space. Second, biological evolution samples a constrained area of the search space near that viable node/peak, where ‘new’ viable solutions are more likely to be found (the space bio evolution operates in is characterized by local non-zero autocorrelations so the fitness of nearby nodes isn’t random with respect to the fitness of the current node). Third, treating biological evolution as search is taken (usually implicitly) to mean that what exists now was some sort of target, and ignores the fact that there are many viable ‘targets’, and which particular target from among many gets landed on is a function of the long string of contingent events in the evolutionary history of a population. Gould got that one right.
Well, I’m a lawyer and that puts you on a par with Luskin. That’s hardly surprising either. Luskin’s spent his entire time since law school acting as a PR flack and Gofer General at the Discovery Institute. You don’t learn much law in law school … that’s why they call it practicing law.
As to “smuggling in information” in the form of parameters, nature does that too … in the form of heredity. Evolution occurs in the context of what came before. Zebras can’t grow machine guns out their asses not because it wouldn’t be a good way to handle lions but because they can’t get there from where their ancestors started. (Example courtesy of Elliot Sober.)
Once, I tried attacking an ICFP contest problem using a quasi-evolutionary algorithm. The problem was to produce a sequence of instructions for driving a race car around a track. In my initial development, I used an elliptical track that had the start and finish line in the same spot.
The thing was, I’d accidentally coded a tiny piece of the physics of the race car universe incorrectly, such that applying the brake on a stopped car would cause the car to drift backwards. It took almost no time for a population to converge on the solution of “stomp on the brake twice, then floor it and cross the finish line”.
There’s nothing like an evolutionary algorithm for smacking some humility into the programmer of the universe the evolution is happening inside.
Bah, that’s just because the dude’s got no clue how hard it is to design an antenna. I did my master’s thesis on an antenna design for a specific application in different materials. It was fairly simple, though, compared to some of the things out there, and it still took a year with a high-powered and expensive simulation program to do it. Just because you may understand the parameters and constraints of a problem doesn’t mean that finding the optimal solution is trivial.
I think the pigs will be terribly offended!
A group at Sun Research was looking at using evolutionary systems for circuit design and layout. One of their results consisted of two electrically isolated components – components that couldn’t influence each other. And yet, they did. It turned out that the system was exploiting a flaw in the manufacturing process – the insulation between circuit traces wasn’t as complete as they thought. So they got a result based on a physical property of their manufacturing process which they were not even aware of.
I’d be really curious to see a more detailed version of this story. It doesn’t sound quite right: presumably the evolutionary algorithm was running a circuit simulator rather than an actual fabrication plant, for reasons of feasibility. Then the simulator shouldn’t have known any physical properties that weren’t programmed into it. My guess is that instead of an unexpected physical effect, they ran into a bug in their simulator. (This isn’t unheard of: a friend of mine’s thesis was nearly derailed by a bug in SPICE.) In other words, they saw the two isolated components and thought it was a bug in their optimization algorithm, but instead it was correctly optimized to take advantage of an obscure bug in the simulator.
Here’s the link to Adrian Thompson’s paper:
http://www.cogs.susx.ac.uk/users/adrianth/ices96/paper.html
“presumably the evolutionary algorithm was running a circuit simulator rather than an actual fabrication plant”
I think we are talking here about programmable logic arrays. in effect, evolving some firmware. So the “fitness” of a configuration is not simulated, but evalueated by actually running it in the real world ™.
Anonymous (#13), if you read the paper I linked to above, you’ll see that the evolving circuit was evaluated running on the hardware, not in simulation. In fact, simulation would never have produced the same result. From the paper:
“Even though this is a digital FPGA, and we are evolving a recurrent network of logic gates, the gates are not being used to `do’ logic. Logic gates are in fact high-gain arrangements of a few transistors, so that the transistors are usually saturated — corresponding to logic 0 and 1. Evolution does not `know’ that this was the intention of the designers of the FPGA, so just uses whatever behaviour these high-gain groups of transistors happen to exhibit when connected in arbitrary ways… This is not a digital system, but a continuous-time, continuous valued dynamical system made from a recurrent arrangement of high-gain groups of transistors…”
A bit below this is where he finds that clamping some gates that are NOT connected to his circuit actually prevent it from working. It’s really awesome.
The paper on the evolutionary FPGA experiments is here – it’s one of my favorite papers ever đŸ™‚
Indeed. The IDers don’t seem smart enough to realize that what they’re saying is that evolution is uniquely impossible to study. That is a phenomenally stupid claim, that’s the caliber of person we’re dealing with here.
Wherein I try to imagine what a response to MCC would be if I were the kind of rare thinker Casey is:
http://www.antievolution.org/cgi-bin/ikonboard/ikonboard.cgi?act=ST;f=14;t=5735;st=2850#entry127911
you just can’t argue with nutjobs. if arguments persuaded them – they wouldn’t be nutjobs.
so nothing is to be gained by debating them…
but i haven’t a clue what the solution might be.
Back before “evolutionary computing” become the latest buzzword, this kind of search strategy was called “simulated annealing”.
And it only really works at all in highly constrained design environments (like this one), and even then needs a fair amount of hand-holding to get good results.
It really has little or nothing to do with the validity of evolutionary theory.
The problems with evolution have more do with
(1) the “survival” fitness function should never have produced higher lifeforms because by this definition life should have stabilised at an extremely simple level.
(2) the “fitness landscape” is so complex and full of peaks and valleys that the probability of a lifeform jumping from its current point to a point with greater fitness is pretty much zero.
Evolutionary computing isn’t but a sexy name given to an optimization algorithm. What it has in common with biological evolution are these couple of new ideas in optimization like crossover/mutation of solution candidates and population/generations of solution candidates.
What often makes genetic algorithms favorable is that they put very few constraints on the search space and target function. And they tend to be faster than other such methods that have loose requirements on those (like simulated annealing).
Still the greatest job involved is to create a model that allows evolutionary computing to come up with “creative” results, just like in any other optimization task.
@Noel: justify your claims 1 and 2. You will find this difficult as they’re both wrong.
For (1): Most organisms on earth are bacteria; this has been true for billions of years and it’s still true now. But an ecosystem consisting only of bacteria is an ecosystem with lots of empty niches for multicellular creatures to inhabit. Given that multicellular creatures manifestly do exist and reproduce, it’s tough to argue that they don’t have a positive value of a survival fitness function… you can tell by the way they’re surviving!
For (2): The higher the dimension of the landscape (and the survival landscape must have a dimension for each independent possible mutation; that’s astronomical!), the less likely it is for a point to be a local maximum or minimum with respect to all possible variables. It’s thus overwhelmingly likely that near to any point there are other points which are as good or better.
Do you have any more ignorance to demonstrate, or are you done?
“For example, I don’t know diddly-crap about law…”
That said, I’d be willing to bet you know more about law than Luskin does about evolution!
Noel is right, at least about the simulated annealing – annealing, genetic algorithms, and other algorithms based on nature have been used for a long time to approximate solutions to NP-hard problems (such as optimizing circuits, schedules, etc.) This isn’t really anything new – though it is interesting đŸ™‚
Noel is wrong about simulated annealing because he claims that genetic algorithms used to be called simulated annealing. Which is a bit like saying that cars used to be called bicycles.
@Noel:
“Back before “evolutionary computing” become the latest buzzword, this kind of search strategy was called “simulated annealing”.”
Nope. Simulated annealing means taking a system and adding randomness to its parameters, then you heat it up (numerically) and cool it down to find minima or maxima in a Monte-Carlo-type of algorithm.
Contrary to evolutionary algorithms, you usually have no selection since you deal with one system at a time (of course you can run several copies with different random numbers). You also have no such thing as crossing-over type of mutations which are quite common in genetic algorithms to combine good solutions of different individuals. Of course you can use combinations of simulated annealing and genetic algorithms, but GAs add a new dimension (selection i.e., removal of unfit individuals in each time step, combinations of individuals genome).
You are right that simulated annealing needs a lot of fine-tuning for the heat cycle – but this is exactly what evolutionary algorithms get rid of. They also need some tuning, for example of mutation rates, but these have not to be as fine-tuned as parameters in simulated annealing and this is also mainly a matter of efficiency – different from nature we usually don’t have the resources to run billions or trillions of system copies to get our optimum. If we had, we could use very loose system specifications.
(If we had enough resources, we could also get rid of lots of the fine-tuning in simulated annealing, but that’s not the issue here…)
God himself could come down and say, “I caused the big bang which initiated the universe, but everything else flowed from evolution.” However, I still don’t think Casey Luskin would believe him.
Daniel Martin,
That was a nice story and I loved the quote, “There’s nothing like an evolutionary algorithm for smacking some humility into the programmer of the universe…”
The story of the “Tower of Babel” makes me wonder if the original “programmer of the universe” ran into the same problem. God says in Genesis 11:6, “If as one people speaking the same language they have begun to do this, then nothing they plan to do will be impossible for them.” Then he confused our languages and spread us through out the world. It almost appears like God is scared of us.
That last post was mine. I forgot to put in my name, and it seems cowardly to make a post like that and not include your name.
Re Noel (#21):
I’m going to ignore the simulated annealing thing; others have dealt with your misunderstanding adequately.
But you comment about fitness landscapes really pushes my buttons; it’s something that I’ve written about and talked about numerous times. For example, it’s the heart of my
main critique in my review of Behe’s latest book.
First, fitness landscapes are a piss-poor way of describing evolution. There’s a lot of reasons for that, but the most important one (in my opinion) is that the fitness landscape model assumes that there’s a fixed, static fitness landscape. That’s just inescapably wrong. In biological evolution, the fitness landspace changes in response to the changes in the organisms that are traversing the landscape.
The only way of escaping the problem with the shifting landscape is to add dimensions – lots of dimensions.
That’s where the second major problem comes in. We think of landscapes as something like a terrain map – that is, a two-dimensional surface of a three dimensional shape. In a landscape like that, there are hills, which form local maxima and valleys which form local minima, and it’s hard to see how things can escape from those.
But we’re not talking about a 3-dimensional landscape. We’re talking about a landscape with thousands of dimensions. For a maximum to be a trap, it needs to be a maximum in every dimension. If there’s even one dimension out of those thousands where a point isn’t a maximum, then it’s not a trap at all.
Finally, the fitness landscape arguments assume that changes can only have very small effects – that, basically, a small change can only result in a small motion in the landscape. That isn’t true. There are plenty of cases where a small genetic change produces a large developmental change. And that, in turn, means that a maximum is only a trap if there’s no higher point within a maximum jumping distance.
Finally, your point about how the fitness landscape should have “stabilized” is just nonsensical. That would only be true if the fitness landscape was stable, and had exactly one maximum reachable from the first organism. The moment you allow two maxima, you wind up with competition, which changes the landscape, which pushes continual development. There’s just no reason to believe that the landscape is so limited or so static. If you want to make an argument of that form, you need to present some kind of real evidence that that’s the real structure of the fitness landscape of biological evolution. But you’re not making that argument – you’re just asserting it, even though it flies in the face of observed facts.
>>
Back before “evolutionary computing” become the latest buzzword, this kind of search strategy was called “simulated annealing”.
>>
Wrong.
Simulated Annealing is from the 1980ies, computer simulation of genetic processes from the 1950ies and 1960ies. Using evolution as an actual means of problem solving is known at least since 1971 (Evolutionsstrategie – Optimierung technischer Systeme nach Prinzipien der biologischen Evolution, Ingo Rechenberg’s PhD thesis). Get your facts right, please.
I am a bit amused when people say that some meta-heuristic search method is actually the same as one other. They are, in practice, quite different animals.
I have yet to see a search method other than Genetic Programming that finds innovative work-arounds for the mistakes the programmer has done, as has happened to me once – the best solutions contained a tricky little sub-routine for multiplication because I had encoded this operation wrongly in the GPs language.
Saying that SAs are actually the same as Evolutionary Computing is about as relevant as saying that finite Turing machines are computationally as powerful as a quad-core Pentium PC in theory. Which one would you rather use to solve problems with?
I personally used Genetic Algorithms (including Genetic Programming) for a lot of research projects. Algorithms are difficult to program and understand but they work when done properly. There will always be people that “Do not believe” and that “I know everything and this is wrong”.
There is nothing wrong with “smuggling of information” into the system. Remember Dawkins’ WEASEL-program? Randomly mutated strings of letters were compared to a “target phrase” and those most similar reproduced faster. As a result the program ended with randomly mutated string equal to the target phrase. Yes, the target phrase was “smuggled into the system”, but that’s exactly what is going on in nature when evolving camouflage! Patterns in environment of an organism are the “target”, organisms mutate, and the most similar to the environment (target pattern) reproduce better, because they are better hidden from predators. So the information about the result is smuggled in the environment and it has nothing to do with intelligent design nor it contradicts darwinian evolution.
Noel Grandin wrote that there are two “problems” with evolution:
With reference to #1, biological evolution is not about survival, it’s about differential reproduction. Also, see “dynamics,” especially with reference to physical variables that compose selective environments as well as the biological selective environment — other species. Evolution by natural selection is an algorithm that increases mutual information between selective environments and the genomes of populations. As differentiation of species occurs in response to variation in the physical selective environment, the biological environment (other species/critters) increases in complexity. As a consequence, the information transmitted from the conjunction of the inorganic and biological selective environments to populations’ genomes increases in complexity, and natural selection automatically generates increased complexity in the array of population genomes comprising the biosphere.
With reference to the second, see Fitness Landscapes and the Origin of Species. Uninformed intuitions about the topography of high-dimensioned fitness spaces are unreliable. And Mark’s comment is on point, as is the review he links to.
geez you guys are a bunch of nerds. I love it.
Zebras can’t grow machine guns out their asses not because it wouldn’t be a good way to handle lions but because they can’t get there from where their ancestors started. (Example courtesy of Elliot Sober.)
That was truly wonderful.
Whatever gives the ID folks the idea that evolution doesn’t provide information? Getting killed (or in the case of sexual selection, laid) or not is pretty damned informative feedback.
I agree completely with Mark Chu-Carroll here, even though I’m a published Physicist who has won court cases from what I learned in 15+ years as a paralegal.
“Growing up, I was taught to call that kind of behavior not just ignorant, but pig-ignorant.”
The Intelligent Design frauds don’t even know how to put lipstick on the pig.
The basic difference between this use of a GA and natural evolution: this GA has a fitness function given from outside, while in natural evolution the fitness is implicit, it’s the result of interactions with the environment and other organisms.
It’s hilarious that IDers would complain about information being “smuggled in” to the algorithm, when you can view the process of evolution as the transfer of information from the environment into genomes.
Wow, those Thompson FPGA papers are neat. Now I’m interested in trying out such an approach.
Smuggling. Shipping routes. genetic algorithm. Okay, try this one:
Operational Optimal Ship Routing Using a Hybrid Parallel Genetic Algorithm
Authors: O. T. Kosmas, Z. A. Anastassi, D. S. Vlachos, T. E. Simos
Subjects: Mathematical Physics (math-ph)
Optimization of ship routing depends on several parameters, like ship and cargo characteristics, environmental factors, topography, international navigation rules, crew comfort etc. The complex nature of the problem leads to oversimplifications in analytical techniques, while stochastic methods like simulated annealing can be both time consuming and sensitive to local minima. In this work, a hybrid parallel genetic algorithm – estimation of distribution algorithm is developed in the island model, to operationally calculate the optimal ship routing. The technique, which is applicable not only to clusters but to grids as well, is very fast and has been applied to very difficult environments, like the Greek seas with thousands of islands and extreme micro-climate conditions.
Mark, from what you say about “their article”, it sounds as if you think the article you linked to there is the NASA guys’. It isn’t. It’s written by some people at the Biologic Institute, which is basically a bunch of ID-pushers, and it is indeed ID propaganda, albeit less blatant than Luskin’s. So the fact that it’s “very open about the fact that they provided a lot of data” tells you that, shock horror, ID propagandists are “very open” about the alleged fact that to get useful results out of evolutionary computation you need to put a lot of information in; it doesn’t tell you anything about how “open” the NASA people were. (I expect they *were* “very open”, but you need different evidence.)
I think that by quoting that paragraph the way you do and commenting on it the way you do, you give much more credit to the ID people’s ideas than those ideas deserve.
If information is being “smuggled in”, then presumably Luskin can point to where. Otherwise, he should keep his trap shut.
@42: apparently we smuggle information in when we tell it that the goal is to produce a solution which is a good antenna. Luskin thinks that an evolutionary algorithm should produce a good antenna without our telling it to.
No, seriously.
@43: So we get back to the good old Evolution is a random process and a hurricane blowing through a junkyard can’t create a 747. Haven’t they been told time and time again that Evolution requires selection, it takes a three year old to then determine that if selection is required there has to be some criteria for how that selection is to be performed.
Aghhhh, it’s so frustrating,
Head->Desk Head->Desk Head-Desk.
Luskin thinks that an evolutionary algorithm should produce a good antenna without our telling it to.
Or more accurately “without conditions that select for a good antenna”. Did Luskin expect their setup to produce a toaster?
I found this website because I remembered reading about this at least a year ago (the evolutionary antenna not the Casey guy). What brought me back is just how much potential evolutionary algorithms could have in solving problems. With how powerful supercomputers are, evolution simulations could do soo much.
Also, I’m one of the few minority of Christians who completely agrees with evolution. I came to my own conclusions and then about a year later the “elders” at my church took the same stance on evolution. I can’t see the creation of the Earth and people any more beautiful then through natural evolution.
@43: that’s not really a minority, I think – Catholics have no problems with Evolution, nor do Anglicans and (I think) Orthodoxes and Lutherans, plus probably many more Reformed Churches I do not know about.
AFAIK, the whole °Evolution vs Religion” thing is almost exclusively the domain of some of the more… extreme U.S. Churches: as usual, the problem is just that a few noisy morons are more visible than the reasonable majority – a form of aposematism, I suspect đŸ™‚
As for evolutionary programming: this result is extremely cool, but is such an approach scalable when the space of parameters becomes very big? My (almost completely uneducated) impression is that the computational cost could become huge very quickly…
LOL, thats just too funny dude! Way too funny
http://www.anonymity.at.tc
I am very disappointed by this article, like many before. Applying the term “evolutionary” to the NASA team’s triumph is dishonest and discrediting to them. That team put an immense effort and a lot of brainpower into the process of designing their new antenna, so calling it “evolutionary” and essentially accidental is insulting.
If anything, their efforts are a beautiful case of intelligent design. It was not random, it was incremental. It was not evolutionary in any sense, it was guided, managed, directed design. It would never have happened by accident, nor could probably any other team on Earth have reproduced it.
It is an amazing accomplishment and they deserve full credit for it. Attaching the term “evolution” just cheapens it.
It’s been mentioned before but I want to reiterate the point that what NASA did really isn’t that far off from “Darwinian evolution”.
“to design an antenna using evolutionary techniques, you need to start with an understanding of exactly what you’re looking for, what range of space to try to cover with your mutations, how to perform the mutations, how to evaluate different results during the selection phase, and so on.”
All these things are represented in nature, it’s just that instead of looking for specific features like “eyes that see with the best detail” the only thing measured is the state of it’s existence. Evaluating results is easy, because the state of existence of a failure will be 0, or no existence. So instead of needing to set up a routine to remove these failures from the environment, they are automatically removed.
But not only do you have a way to determine success, you also have parameters that guide the process, such as terrain, scarcity of resources, proximity to other organisms, et al. In nature, the above parameters are used, so when an organism that requires humidity to function develops in a desert, it is swiftly eliminated from existence. In an evolutionary algorithm like one used by NASA, you have things like antenna position and material specified (making these up here, not clear on the exact parameters used), and when something no longer meets these criterion they are eliminated.
To say that adding these parameters in unfairly guides the simulation shows a complete lack of understanding in the subject at hand.
@Rob Williams
I’m sure the NASA team did put a lot of effort and brainpower into the new antenna design, but only in the same way a teacher puts a lot of effort into moulding student minds. After the training is complete, the teacher doesn’t take credit for the designs of students.
@Voideka
Surely it fuels a creationist’s doubts to say “when an organism that requires humidity to function develops in a desert, it is swiftly eliminated.”
Better to stress that desert dwellers with an increased tolerance for dryness have a better chance of survival, and therefore a greater chance to pass on their genes. I don’t see how anyone could argue with a logical conclusion like that.
But then, I too might be disappointed …
Re #47: “is such an approach scalable when the space of parameters becomes very big?”
Yes.
That’s part of the point of the Genetic Algorithm.
It is inherently parallel in its operation. Making the “chromosome” twice as long (i.e. as a string with twice as many parameters) does not cut the evolution rate in half, or by a factor of 2^2 = 4, or anything of the kind.
Remember — you, personally, have some 30,000 genes. That’ shows scaling up rather nicely.
When, as part of my PhD dissertation research at University of Massachusetts at Amherst, in the CS department where I’d earned my M.S. in Artificial Intelligence and Cybernetics (1975) I beta tested John Holland’s version of the Evolutionary Algorithm (1975-1977) and became the first person to artifically evolve software (string of characters in the language APL) I was interested in Holland’s description of a higher level of parallism.
He pointed out that if the “chromosome” was simplified to be a string of N bits (i.e. N genes, each in two different alleles) then there was more going on than the 2^N possible bit strings being sampled by the database of strings in the population undergoing simulated evolution.
One is also implicitly dealing with pattersn, which he called “schema.” At each gene, one has this set of possibilities defining a schema: {0, 1, don’t care}.
Hence there are 3^N schema being sampled by the database of strings in the population undergoing simulated evolution.
The major theorem that John Holland proved was this. The first-order evolution is evidenced by the mean value of each instantiation of a chromosome (of the fitness function applied to the string) increases with rate of increase proportional to the ratio of that string’s fitness normalized by mean or total fitness of the population.
But, at the same time, if we estimate the fitness of a schema as the mean fitness of all strings present in the population which are instantiations of that schema, the mean value of each schema (of the fitness function applied to the strings instantiating that schema) increases with rate of increase proportional to the ratio of that schema’s fitness normalized by mean or total fitness of the schema instantiated in the population.
That’s scaling up, with a vengeance!
Holland also proved that his Genetic Algorithm was exponentially faster than certain naive earlier optimization algorithms.
Re #49:
Why is it insulting to the NASA engineers to accurately characterize their methods?
The people involved in the project set up a system based on mutation and selection. How can you possibly claim that such a system is anything but an evolutionary approach?
And since the final result was unexpected, and better than what any of the engineers was able to design without the use of the evolutionary approach, it makes no sense to argue that the evolutionary system didn’t do something “creative”.
Speaking as an engineer myself, the best criteria for judging an engineer is to look at how they do their work. Given a problem, a good engineer analyzes the problem, and selects the best approach to solving that problem. Sometimes the best approach is very hands-off.
In this case, a team of engineers recognized that they didn’t know how to analytically find the best point in a configuration space. So they used evolutionary techniques – reproduction with mutation and selection – to “search” the state space.
Accurately describing that as using an evolutionary system is in no way diminishing the accomplishment of the engineers. The fact that they found a technique for finding a superior solution is brilliant; accurately describing how they did it is actually high praise. I’d be damned proud to be one of the engineers who used an evolutionary technique to solve a difficult problem.
Putting aside the particular words used in Luskin’s article, is it fair to ask how much the fitness landscape* was restricted before running the GA? Would it be accurate to say that a given GA is less powerful if you have to give it lots of “help”, such as complex fitness analysis (in essence, giving more granularity and perhaps more continuity to the landscape), tight restrictions on parameters, etc.? Surely, the goal is for humans to be able to let computers do most of the work, without laborious discovery of tweaks required to get simulation time down to a reasonable amount? Is it not fair to say that there are many ways to specify fitness and restrict parameters, and the more you do, the more you are starting out with a more designed system? I’m sure the antenna algorithm could have been vastly simplified, at the cost of requiring an inordinate amount of computing power.
RBH’s #8 comment, critiquing the analogy of biological evolution as a “search process” is interesting. I’m not sure the point is particularly interesting, because it appears that “optimization process” survives all of his criticisms. If it valid to use the term power as an aggregate way to measure the “performance-to-parameter” ratio of GAs [for a given problem], then I would think that a fundamental understanding of how powerful we can make GAs could be very enlightening. On the one hand, we could find that with the right kinds of mutations and other, [relatively] problem-agnostic configuration, GAs can do a whole lot without much laborious characterization of the problem, like the NASA folks had to do. On the other hand, we might find that there are just theoretical limits on what GAs can do, and just perhaps, that might reveal specifics about biological evolution and how it might cheat. **
Now, am I completely off-base here? Maybe I’m being much too nice to all the ID and creationist folks, but this is one direction I could see them wanting to go, without knowing enough to spell it out like I’ve attempted above. If that’s not where they want to go, then damnit, I’m interested in the above. That, or I want to know why all the questions I’ve asked are bad and not worth answering.
* Mark, I know you said in #30 that you don’t really like this term, but your “most important [reason]”, that of assuming a “fixed, static fitness landscape”, is actually valid for this situation. In fact, this is actually fairly common with GAs. I’m not sure if this FPGA experiment were discussed above, but there was on experiment where the goal was to recognize different frequencies. It worked, but only at the lab’s standard temperature, and perhaps with ambient EM fields relatively unchanged. Hopefully we can move away from the static landscape assumption, but I see nothing to contradict it as a good assumption for many simulations, for now.
** I put that in there just for those who had so much fun with Luskin’s use of the word “smuggling”. If you are about to lambast me for using the word “cheat”, you need to go outside and take a deep breath of fresh air.
Luke:
I don’t think it’s unfair to point out that the landscape in this experiment was a fixed restricted one. But I also don’t think that pointing that out helps the ID case at all.
One of the standard tricks that the ID folks like to pull is using the “no free lunch” theorems to claim that evolutionary processes can’t work. All of the stuff that Casey says about the crafted landscape is an attempt to claim that this isn’t really an evolutionary system; after all, evolutionary systems can’t work because of NFL, therefore there must be some cheating here, and that cheating is the shape of the landscape.
You see,”no free lunch” says that averaged over all fitness landscapes, no search process is better than random walk. IDists try to translate that to the simpler statement that “nothing can do better than a random walk”.
But evolution – any kind of evolution, whether it’s biological evolution or evolutionary algorithms – doesn’t operate over the set of all possible landscapes. It operates on one particular landscape. That’s the “escape clause” from NFL. No evolutionary process is expected to succeed in all possible fitness landscapes. (In fact, “all possible fitness landscapes” necessarily includes both totally random unstructured landscapes, and adversarial landscapes where the landscape has a structure that is specifically structured to mislead the evolutionary search.
Totally random landscapes don’t match observed reality; and
adversarial landscapes only make sense if you assume that the structure of the universe is controlled by an intelligent adversary. But the whole argument is based on trying to show that the universe can’t exist without an omnipotent agent, the argument against evolution can’t assume that the reason evolution doesn’t exist is because there’s an omnipotent agent manipulating it.)
I’m not even sure what averaged over all fitness landscapes means.
Isn’t the issue here whether a GA can optimize in a reasonable time frame? Clearly, the NASA folks had to “help it along” by providing “hints” in the form of constraints. If it weren’t for these hints, optimization would take much too long. Right? If this is the case, then does it not make sense to ask whether our current model of mutation, natural selection, …, has enough “hints” (and/or, a granular enough fitness evaluation algorithm)? There are clearly biological constraints (such as DNA correction functionality), as there were NASA-imposed constraints.
As usual, please let me know if I’ve wandered into nonsense-land.
Luke, Re: the “averaged over all fitness landscapes” thing:
In NFL, you model things like evolution as a search function over a fitness landscape. The performance of a search function on a given landscape can be described (loosely) as a function of time; the performance value at any given time T is the value of the fitness function at the position arrived at by the search after T iterations.
To get the overall performance of a fitness function f, you can choose a couple of options, but one reasonable one is the average performance computed over all times.
So, given a search function and a landscape, you can compute a performance measure. Given a set of multiple landscapes, you can compute the performance of your search function on each landscape separately. Then you can describe the average performance of your search on your set of landscapes.
If the set of landscapes is enumerable, then you can (theoretically) compute the average performance of your search function over all possible landscapes. (Even if the the set of landscapes isn’t enumerable, you can still talk theoretically about the average performance over all landscapes; you just can’t compute it.)
I’ll probably end up converting this to a post with some pictures to clarify it, but if you think of a one-dimensional landscape – that is, a curve f(x), where the fitness at a point x is the value of f(x), then the set of all landscapes includes nice, smooth polynomials, which produce curves with smooth slopes to maxima and minima, and also totally discontinuous curves where there’s no structure at all. (Try generating a curve on a computer where f(x) = currenttime / current_cpu_load * random.)
The set of all landscapes also includes adversarials. To explain what that means, imagine that the fitness function is “closeness to the zero of a specific curve”. For cases like y = x^2 – 2, You can compute that very easily using Newton’s method, which happens to be a nice example of a search function. If you’ve ever taken a calculus class where they taught Newton’s method, you’ll remember that there are curves where Newton’s method
doesn’t converge to the answer – because the structure of the curve produces a varying slope that keeps bouncing you away from the solution every time you get close.
So the average performance over all possible landscapes is a strange idea. And it’s obvious on an intuitive level why there’s no way that a single search function can possibly perform well on all possible landscapes. On some landscapes, it will do really well. On some (the random ones, which make up the majority of all possible landscapes),
it will be effectively the same as a random walk. And on others, it will perform worse than random, because the adversarial landscape is perfectly structured to make the search strategy travel towards the worst possible places.
Ok, this is getting long… I’ll address the next point in a separate comment.
Luke, re: hints.
Sorry, but you’re talking nonsense.
The “hints” that were provided in the NASA experiment were the shape of the landscape, and the set of possible mutations.
The hints weren’t time optimizations. They’re not there because without it would take too long. They’re there because, by definition, evolutionary processes take place in a specific landscape, with a specific definition of fitness, and a particular set of possible mutations.
In the antenna experiment, fitness was antenna performance for a specified range of radio frequencies. Mutations were things like changing the antenna length, adding or removing a bend, and changing the angle of a bend. Without providing those “hints”, what does it mean to do an evolutionary search at all?
In biological evolution, fitness for an organism is something like “number of live children”. The possible mutations are changes in genetic code: point mutations,
reversals, copies, deletions.
In both cases, they’re doing a search of a particular landscape using a particular
fitness function, and the search is based on reproduction with mutation + selection.
Reproduction with mutation + selection is an extremely successful search function for structured landscapes.
When you look at life, you see reproduction and mutation, and a very structured
landscape. The problem with life is that you’re also bumping into issues of scale. The
“fitness landscape” for biological evolution has a nearly incomprehensible number of dimensions. And the “search” runs at a pace which is difficult to comprehend.
To give you an idea of how mind-boggling it is: the surface of the body of the average human being hosts 1012 bacteria. On average, bacteria reproduce every 20 minutes. That means that a back of the envelope calculation shows that on an average day, just on the surface of my body, bacteria reproduction – assuming no deaths – roughly 4×1033 times. Now, obviously, lots of bacteria are dying on my body. But suppose that just one in 10^12 manages to reproduce at every reproduction interval. That’s 1021 reproductions
per day.
And it’s happening on that scale on every living thing on the planet.
Even looking at human beings: there are something around 6 billion humans on earth. We reproduce on average every 20 years. And every single one of us is home to several unique mutations.
Biological evolution is hard to comprehend for precisely that reason: the landscape is too big for us to really comprehend, and the scale of the search is just too big for us to comprehend. But the basic fact remains, that reproduction with mutation + selection
turns out to be an extremely effective and scalable search function for structured landscapes; and our universe is an extremely structured landscape.
#55 is quite interesting: “adversarial landscapes only make sense if you assume that the structure of the universe is controlled by an intelligent adversary.”
Why are there no obvious Satanist Intelligent Design advocates?
The issue of “adversarial landscapes” is quite important in these kinds of contexts:
(1) business competition against companies that have read your Annual Report and are trying to use the marketplace against you, by changing their profit structure or pricing or alliances or whatever. The competitors are quite willing to take a loss for a quarter or a year if they think that this will be forcing you towards your bankruptcy.
(2) Counterterrorism (or counter-intelligence) where the national security or intelligence agency is trying to make the landscape as lethal as possible for terrorists or spies.
(3) In general, arenas (including military) where Disinformation is important (this includes bluffing in Poker).
Stanislaw Ulam was the first to analyze the problem of “20 questions asked of your adversary, who is allowed to lie once, and is presumed not to waste that lie at random, but to disrupt your search for the right answer as much as possible.” If the 20 questions game is to guess a positive integer in the range of 1 to 1,000,000 you can obviously get it in 20 questions because 2^20 is greater than 10^6. If the adversary lies once, it will take you, with perfect strategy on both sides, 23 questions.
The results were worked out by ad hoc means for the 20 questions game to guess a positive integer in the range of 1 to 1,000,000, with the adversary allowed to lie exactly twice.
Then, when ad hoc methods solved the 20 questions game to guess a positive integer in the range of 1 to 1,000,000, with the adversary allowed to lie exactly 3 times. Then someone noticed that the structure was eactly that of the Golay Code. No surprise, in retrospect, that the search problem connects deeply with Error Detecting Error Correcting Codes.
The general problem remains open. The NFL means that no army can win every battle against every foe any better than acting at random. But, in the real world, you use the army that you have, the soldiers that you have, the materiael that you have, the logistical support that you have, and the intelligence that you have RIGHT NOW against the enemy who is actually there to fight you.
Mark CC is 100% correct, yet again.
#55 is quite interesting: “adversarial landscapes only make sense if you assume that the structure of the universe is controlled by an intelligent adversary.”
Why are there no obvious Satanist Intelligent Design advocates?
The issue of “adversarial landscapes” is quite important in these kinds of contexts:
(1) business competition against companies that have read your Annual Report and are trying to use the marketplace against you, by changing their profit structure or pricing or alliances or whatever. The competitors are quite willing to take a loss for a quarter or a year if they think that this will be forcing you towards your bankruptcy.
(2) Counterterrorism (or counter-intelligence) where the national security or intelligence agency is trying to make the landscape as lethal as possible for terrorists or spies.
(3) In general, arenas (including military) where Disinformation is important (this includes bluffing in Poker).
Stanislaw Ulam was the first to analyze the problem of “20 questions asked of your adversary, who is allowed to lie once, and is presumed not to waste that lie at random, but to disrupt your search for the right answer as much as possible.” If the 20 questions game is to guess a positive integer in the range of 1 to 1,000,000 you can obviously get it in 20 questions because 2^20 is greater than 10^6. If the adversary lies once, it will take you, with perfect strategy on both sides, 23 questions.
The results were worked out by ad hoc means for the 20 questions game to guess a positive integer in the range of 1 to 1,000,000, with the adversary allowed to lie exactly twice.
Then, when ad hoc methods solved the 20 questions game to guess a positive integer in the range of 1 to 1,000,000, with the adversary allowed to lie exactly 3 times. Then someone noticed that the structure was eactly that of the Golay Code. No surprise, in retrospect, that the search problem connects deeply with Error Detecting Error Correcting Codes.
The general problem remains open. The NFL means that no army can win every battle against every foe any better than acting at random. But, in the real world, you use the army that you have, the soldiers that you have, the materiael that you have, the logistical support that you have, and the intelligence that you have RIGHT NOW against the enemy who is actually there to fight you.
Mark CC is 100% correct, yet again.
I don’t disagree with your [implied] point that the term averaged over all fitness landscapes is useless and misleading. We need to talk about the type of landscapes we actually have, and perhaps average over those. Evolutionary algorithms are clearly [relatively] poor for problems with easy to compute (and non-GA) algorithms.
When I use the term “hint” in this context, I mean something that restricts the domains of one or more dimensions of the landscape. In other words, hints simply reduce the search space. Describing this as changing the “shape” of the landscape is a bit nebulous — we aren’t affecting the fitness values, just where we look for fitness values. We aren’t altering the function x^2, we’re just looking for x > 0.
I suppose “possible mutations” can be construed as a hint, but unless taken to the extreme, we have a direct analogy between biological mutation and simulated, “mechanical” mutation. At the very basic level, we only need the simplest set of mutations that can compose to produce the solutions. I suppose “the extreme” would be to have so much knowledge about the solution space as to radically reduce the possible mutations so that all the peaks on the fitness landscape are retained. Compare this to the biological situation, where most mutations do nothing and those that do affect the phenotype, usually do something bad. If we have a simulated system where a [relatively] larger portion of mutations made positive contributions, we have a good reason to say that the simulated evolutionary algorithm is quite different from biological systems.
Could you elaborate a bit on “[Hints are] not there because without it would take too long.”? Surely, people would rather be lazier and get the right answer, than laboriously limit the search space? The less time scientists have to spend on each EA, the more EAs they can run. You seem very much opposed to this idea that we can take a basic EA, which has the desired solutions in its search space, and optimize it so that the search space is smaller, but we know that the desired solutions are still in it. (We can’t be entirely sure, but that’s just life.)
My mind isn’t really boggled at your very large numbers; I’m not sure if this is a good thing or a bad thing. I know IDists throw out lots of crazy-large numbers that mean anything, so ignoring those is a good thing. I just do not like assuming that there are “enough” mutations just because the rest of evolutionary theory appears to indicate this is how things happen. Is that really a bad attitude to take?
Switching topics slightly, I think it’s important to note that evolutionary algorithms have to start out somewhat to fully synthetic. Part of modeling is to take the simplest model possible first, and then build upon it. I’ve been in the position of starting to make something cool, only to have people bash me because they didn’t see where I was headed. Then again, I need to be careful not to make claims that are way out of proportion with what I’ve actually shown. Might scientists not be a little guilty of this, misconstruing the level of accomplishment that was achieved and where it is “sure” to head?
Perhaps one point about the NASA experiment might help: they would in no way claim they have the “best possible” answer – they have **an** answer, that performs better than others they derived, and in a way they had not predicted.
They make no claim about being exhaustive, and **don’t care**. No one claims evolution produces the **best** results, just **a** set of results. There may well have been other paths NASA could have searched (curves, loops, floating segments,etc), but didn’t. The whole point of EA is that they build on what they have, rather than leaping to/from utterly random paths.