652 thoughts on “Evolving Wind Turbine Blades”

  1. CharlieM: Because with a real 3 dimensional, 3 blade wind turbine there will be many more variables which will affect the outcome.

    It is certainly true that the simulation here is simple compared to the real world. Do you believe it would not be possible to use selection to improve the propeller blades if more parameters were added to the simulation to make it more realistic?

  2. phoodoo: This has a hard target, that is my whole point. It has ONE target, which has the most efficient rating in a wind tunnel. That’s the only target.

    By analogy, reproductive success is the “target” of evolution in the real world. It doesn’t, therefore, make any substantive difference with respect to the simulation as a representation of a real-world process, to declare it has a “target”.

  3. phoodoo: Will the weasel program work if we give it TWO targets? How about seven? How about 100?

    Why don’t you try it and find out? How do you propose giving it multiple targets would even work?

    What would a WEASEL program with “multiple targets” even look like? What the fuck does that even mean?

    Dude, do you do thought?

  4. phoodoo: Great Adapa, four methods to get at ONE target. Why only one target? How about four methods to get at six targets! The designs will get even worse won’t they? (Clue-the correct answer is yes).

    PROVE IT.

  5. Rumraket,

    Ha. What he is trying to tell you is that the GA is a search for something specific, and that evolution is not a search for anything, it is blind. Thus the two are completely different entities.

    Of course if you give a computer something to look for it can find it. It can find the longest piece of string, it you let it test 10 pieces. What it can’t do is find the BEST piece if you don’t tell it what best is.

    I recommend that you try to understand that distinction before you go on another meaningless tirade.

  6. Rumraket: What would a WEASEL program with “multiple targets” even look like?

    Sort of what evolution would look like if you let it choose the strongest and the dumbest, and the tallest, and the one with most hair. A mess.

    But you can try it if you don’t believe me. Give a computer the target to spell “Me thinks its a weasel.” and also “Rumraket is a silly drunk” and also spell out the first chapter of the DaVinci Code, as well as a target to write out the first five lines of Sun Tzu’s Art of war in Mandarin. Don’t give any of the targets priority, let it keep parts that match any of those, occasionally discard some parts, occasionally flip the letters backwards, wipe out some parts of a phrase with virtual typhoons, and others with mass starvation, give it viruses that destroy some of the letters, and see if it can randomly stumble upon spelling out any of those phrases in their entirety.

    One thing you might learn. The means targets you allow the computer to have, the more your side loses. Variation is not your friend.

  7. phoodoo: What he is trying to tell you is that the GA is a search for something specific

    No, it is simply iteratively producing more and more efficient wind turbines. It isn’t searching for some specific turbine. Nowhere in the algorithm is it stated that it must “find the most efficient one”. That might eventually be the result if it was allowed to run some amazingly long period of time, but merely because that might result from the process doesn’t mean it is valid to say the algorithm is somehow actively searching for that specific best possible solution. It is at best a metaphor for what happens.

  8. phoodoo: Sort of what evolution would look like if you let it choose the strongest and the dumbest, and the tallest, and the one with most hair.A mess.

    But you can try it if you don’t believe me.Give a computer the target to spell “Me thinks its a weasel.” and also “Rumraket is a silly drunk”and also spell out the first chapter of the DaVinci Code, as well as a target to write out the first five lines of Sun Tzu’s Art of war in Mandarin.Don’t give any of the targets priority, let it keep parts that match any of those, occasionally discard some parts, occasionally flip the letters backwards, wipe out some parts of a phrase with virtual typhoons, and others with mass starvation, give it viruses that destroy some of the letters, and see if it can randomly stumble upon spelling out any of those phrases in their entirety.

    One thing you might learn.The means targets you allow the computer to have, the more your side loses. Variation is not your friend.

    LOL. If you give it multiple complete sentences to produce, it will simply produce all of them. If you mix and match words from each individual sentence into a larger nonsense sentence, it will reproduce the mixed and matched one. How the fuck is this supposed to be an argument against evolution?

    There is no search for the “dumbest and smallest” in evolution, there is just differential reproduction due to the physiological capacities of the organism. If there are multiple peaks, the population will simply either move to one peak and stay there or split into two.

    Holy fucking shit your argument is dumb.
    Again, look at the fucking video: https://www.youtube.com/watch?v=4pdiAneMMhU

  11. Phoodoo keeps getting it wrong. Evolution doesn’t target individual characteristics like “tall”, “smart”, etc… it’s overall fitness as a combination of all those factors what determines reproductive success. Funny because that’s exactly why some body plans succeed and why evolution can build those complex, coordinated systems with parts that work well together, because together they increase fitness, but in isolation are worthless

  13. phoodoo,

    Will the weasel program work if we give it TWO targets? How about seven? How about 100?

    Yes.

    I bet a 10 year old could see the problem, if he was given a five minute explanation. But you geniuses….

    You (and fifthmonarchyman I’ve noticed) place more faith in the assumed capacities of children than of university professors. Because the former see things like you do. That is not half so devastating to your targets as you intend.

  14. Frankie,

    .Searches require sight and mind. Yet evolution is supposed to be blind and mindless.

    Whoops

    OK, it’s not a search then, if that’s an essential feature. Please yourself. God pops into and out of existence depending on whether I say ‘yes’ or ‘no’ to the question. What fun.

    It’s a search. Hi, God!
    It’s not a search. Hey! Where’d ‘e go?

    Have a banana.

  16. phoodoo,

    One thing you might learn. The means [more?] targets you allow the computer to have, the more your side loses. Variation is not your friend.

    And you just kinda … know this?

    The method was modelled directly on nature. Variation and all. Amazing that it works, considering it doesn’t work in nature. Fucking amazing.

  17. I think that part of phoodoo’s confusion (probably a small part, but that’s a whole other discussion…) comes from thinking that evolution, and the turbine GA, have a “target”. This “search” analogy leads to the misconception that there is a specific structure that the algorithm is trying to find. Under this misconception, if the algorithm is given multiple conflicting “targets”, then it won’t find anything of any use whatsoever. This is wrong. There are not “targets” to be found, but rather properties to be optimized.
    The algorithm can have multiple metrics that are used to assess utility. They may be conflicting(short and fast), but so long as they are not mutually exclusive(short and tall), then the algorithm can optimize for many of them simultaneously.
    The delicious irony here is that phoodoo’s caricature

    The traits that make you faster hinder your ability to make you fatter. Which does the computer choose for the next round? Ok, choose the fatter. Next round choose the fatter one that also has good fur.

    Turns out to be a rather good way of optimizing for multiple traits. If you have multiple, conflicting attributes that you want to optimize for, combining them into a single invariant measure of fitness can lead your beasties to get stuck at a local optimum that is an artefact of the weights you gave to the different traits. Phoodoo’s “one round you select for fat, the next round you select for fast” is actually a good way of avoiding getting stuck in local optima. You are, in effect, jiggling the fitness surface, thereby allowing greater exploration. Quite unexpected solutions appear.
    Well, it worked for me, anyway.

  19. sez phoodoo:

    The entire problem is when you DON’T have a target, you know like evolution for example! The claim that it could also work with multiple variables is pure and utter bullshit. It couldn’t work with TWO variables, if the two variables were contradictory.

    Groovy. Now all you need to do is demonstrate that “it” cannot work with TWO variables, when the two variables are not contradictory.

  20. I’ve lost track (been busy). One of the most interesting aspects of this genetic algorithm is that the information that ends up in its genome is not encoded in the algorithm. So objections that say that “well this may succeed but the information was already there, encoded in the target phrase” do not apply here.

    Among all the arguments about whether this “is a search” or not, or whether it is realistic enough (realistic enough for what?) I don’t see anyone addressing this issue and showing that the information is in the algorithm, hidden somewhere, before it ends up in the genome.

  21. Joe Felsenstein:
    I’ve lost track (been busy).One of the most interesting aspects of this genetic algorithm is that the information that ends up in its genome is not encoded in the algorithm.So objections that say that “well this may succeed but the information was already there, encoded in the target phrase” do not apply here.

    My go-to reply to Creationists who assert that the solutions generated by GAs are somehow based on information “smuggled into” the GA by their creator(s), is to point out that there are known instances of GAs coming up with solutions which no human being understands how they work. How, exactly, does a designer-of-GAs “smuggle” that sort of “information” into a GA?

  22. cubist: My go-to reply to Creationists who assert that the solutions generated by GAs are somehow based on information “smuggled into” the GA by their creator(s), is to point out that there are known instances of GAs coming up with solutions which no human being understands how they work. How, exactly, does a designer-of-GAs “smuggle” that sort of “information” into a GA?

    Exactly. A fine example is this:

    An evolved circuit, intrinsic in silicon, entwined with physics

    where a GA was used to evolve a circuit of logic gates in a FPGA, and the result ended up using the digital logic in combination with analogic signals produced by the very physics of that particular chip, and nobody has a clue how that works. There are clusters of gates that are not interconnected but still interact somehow. Simply awesome

  23. Joe Felsenstein,

    I’ve lost track (been busy). One of the most interesting aspects of this genetic algorithm is that the information that ends up in its genome is not encoded in the algorithm. So objections that say that “well this may succeed but the information was already there, encoded in the target phrase” do not apply here.

    Among alll the arguments about whether this “is a search” or not, or whether it is realistic enough (realistic enough for what?) I don’t see anyone addressing this issue and showing that the information is in the algorithm, hidden somewhere, before it ends up in the genome.

    This is why I asked phoodoo to point to the target. He responded . . . typically.

  24. cubist,

    Groovy. Now all you need to do is demonstrate that “it” cannot work with TWO variables, when the two variables are not contradictory.

    Even contradictory variables could work, given some analogue of reproductive isolation (even a probabilistic genetic distance should allow partition of the separate ‘searches’.).

  25. If the objective (a forlorn one) were to convince a Creationist that there was no funny business going on, one could get the process to ‘pick’ its own fitness function at runtime. Or get the Creationist to pick it.

  26. Allan Miller:
    If the objective (a forlorn one) were to convince a Creationist that there was no funny business going on, one could get the process to ‘pick’ its own fitness function at runtime. Or get the Creationist to pick it.

    The advantage of this simulation is that even the fitness function is not coded into the algorithm, or existing in a file somewhere. The simulated physics evaluates the genotype/phenotype, and all that exists in the algorithm is some simple function turning performance into fitness. No one can successfully argue that the detailed information about the desired genotype and phenotype are coded into the algorithm.

    There are other simulations that work similarly. I have cited here in comments on another thread some nice work by Karl Sims, an implementation of that called breve, and a fun simulation called Boxcar2d. All of these do not build a target into the algorithm but use simulated physics.

  28. DNA_Jock,

    If that would work so well, why do you think the guy who made the program didn’t do it that way?

    He just didn’t think of it huh?

    Imagine how much better his turbines would be if the computer also selected for those that had the most number of blades, and also those that had the most musical rhythm to them to keep birds away, and also was easiest to clean and also the most resembling an Exxon logo. Do you think you would get the most wind efficient shape? Would you get almost the most wind efficient?

    You would get a frickin missmash of poor design, that is what you would get. Perhaps you think the human eye is a missmash of poor design.

  30. Adapa,

    Adapa,

    Let me ask you question.

    When you are slapping around your baby Capuchin, does he sometimes smear feces in your face?

  31. Phoodoo, its okay you don’t understand evolution. You’re not doing science nor driving opinion. Find comfort in your church.

  33. Rumraket: It is certainly true that the simulation here is simple compared to the real world. Do you believe it would not be possible to use selection to improve the propeller blades if more parameters were added to the simulation to make it more realistic?

    Correct me if I’m wrong but the aim of the simulation is to produce the 2 dimensional shape which will return the highest positive number when a calculation is made of the relationship to points on its perimeter with a constant set of data (the simulated laminar airflow). The very first one that is picked out to breed from is the one among a group which returns the highest positive number and those with the lowest numbers are eliminated. It is searching for and keeping a population with a higher average number than its predecessor.

    If he were to replace the uni-directional stream of simulated wind forces with a more realistic multidirectional torbulent stream of forces I do not think that it would home in on the target as the current simulation does. As it stands the population average is tending towards the highest possible number. With a variable wind force the “better” shapes added to the next generation may turn out to be poorer performers the next time round.

    The constant wind direction of the original simulation guarantees that the more positive numbers are kept, a turbulent flow, in my opinion, would not.

  34. CharlieM: The constant wind direction of the original simulation guarantees that the more positive numbers are kept, a turbulent flow, in my opinion, would not.

    You’d simply run a monte carlo analysis with wind direction pulled from known distributions.

  35. phoodoo: If that would work so well, why do you think the guy who made the program didn’t do it that way?

    He only evinced interest in optimizing efficiency.

    He just didn’t think of it huh?

    Imagine how much better his turbines would be if the computer also selected for those that had the most number of blades, and also those that had the most musical rhythm to them to keep birds away, and also was easiest to clean and also the most resembling an Exxon logo. Do you think you would get the most wind efficient shape? Would you get almost the most wind efficient?

    Let’s suppose that, in addition to optimizing for efficiency, he also optimized for Exxon-logoness; it’s a fairly safe bet that the resulting co-optimized turbines would have near-neighbors who were slightly more wind efficient and less Exxon-like. However, it could simultaneously be the case that the co-optimized turbines would be more wind efficient than the ones found by optimizing for efficiency alone. Particularly if the algorithm varies the weighting of the different metrics during the run. I don’t expect you will be able to understand this.

    You would get a frickin missmash of poor design, that is what you would get.

    I absolutely adore the way you assert without any foundation, yet with religious certainty. Reminds me of your “one black M&M in a bag of 1000 will always go extinct” debacle.

    Start with one, see where it goes. It will die every time. I can tell you that even without a little computer program.

    Comedy gold.

    Perhaps you think the human eye is a missmash of poor design.

    Well there is that blind spot problem…to me, it looks like a highly optimized frozen accident.

  36. Richardthughes,

    Well, varying the wind speed during the iterations would be interesting and useful, and have the added benefit of proving phoodoo wrong…
    However, given the vertical axis, the wind direction is immaterial.
    Modeling gustiness and wind shear would complicate the physical simulation quite a bit, I expect.
    First, get the viscosity right, though.

  37. CharlieM,

    Honestly, I have no idea of how good OpenFOAM is at simulating airflow, or if it can simulate turbulence.
    Of course if there’s a better way to test designs, they will perform better once built, but that’s all irrelevant to whether it converges on some particular designs or not.

    First off, it doesn’t home in on the target. Every run converges into something (into something different every run, so there’s no “target”) because the algorithm doesn’t model random variation in a way that allows it to jump “far” enough to find another hill.

    Having a better airflow simulation may help by smoothing the landscape, or it may make matters worse, who knows

  38. DNA_Jock:
    Richardthughes,

    Well, varying the wind speed during the iterations would be interesting and useful, and have the added benefit of proving phoodoo wrong…
    However, given the vertical axis, the wind direction is immaterial.
    Modeling gustiness and wind shear would complicate the physical simulation quite a bit, I expect.
    First, get the viscosity right, though.

    The wind direction is a very important component if it includes different directions within the one stream,

  39. dazz: Now if a GA was meant to properly model evolution, it should also model mutational meltdown, which means that some branches of small populations would actually work their way downwards into extinction. Of course that would make no sense for engineering applications, but my point is that the closest we get to modeling evolution, the clearer it gets that the algo doesn’t search for anything, it’s just a process that happens to beat the crap out of design in many applications even if it doesn’t have a tiny fraction of the detail, time and resources that evolution has had throughout the history of life in the last 3 billion years or so.

    As near as I can tell, you are the only one here claiming GA’s are meant to model evolution.

  40. dazz: To turn a GA into a search, you need some external criteria to the algo to determine when to stop.

    So a search, in order to be a search, must come to an end, and the GA in the OP is not a search because it has to be manually ended?

    That’s your argument?

  42. Mung: So a search, in order to be a search, must come to an end, and the GA in the OP is not a search because it has to be manually ended?

    That’s your argument?

    Maybe we should clarify what we mean by “search”.
    It can go two ways:

    1. It targets specific solutions, which is wrong
    2. The tautological: it searches for better solutions

    And if you think I’m the only one who thinks GA’s model evolution (to some extent/level of detail) ask others or read again, actually they model Darwinian evolution

  43. Richardthughes: I don’t know, nor care, Phoodoo. I just try and learn about things before I talk about them. You might want to try it, to save future embarrassment.

    Richardthughes was so disappointed when he found that youtube video of the Encabulator, after he had spent all that money on the university course. He’ll never make that mistake again. From now on it’s youtube or ignorance!

  44. A guy designs a program to design wind turbine blades and it does so. How does that model Darwinian evolution and not evolution by design?

  45. Rumraket: I’m fine with calling it a search as long as we understand evolution isn’t consciously or deliberately searching for anything. That we use the word ‘search’ as a metaphor for how the evolutionary process results in adaptations in living organisms. If you’re married to the word search and just can’t let it go, then fine I can live with you using it.

    First, the quote you were responding to wasn’t about evolution, it was about a GA.

    Second, you’re contradicting yourself. In an earlier post you wrote:

    Oh for fucks sake, the measure of “better” for real organisms is reproductive success. It’s not “tall and short” or “furry and scaled” or “walking and burrowing”. It’s reproductive success full stop. Whatever compromise solution is better able to reproduce is the measure of success, regardless of how that is actually physiologically achieved.

    So it’s got nothing to do with adaptations. Full Stop.

    But if it is about adaptations, then you have nothing to connect your model to the adaptation, because “It’s reproductive success full stop.” Once you introduce adaptations you have targets. And a search. And indeed, the whole idea is to explain adaptations (the appearance of design).

    It’s like a shell game. Trying to follow the pea.

  46. Rumraket: By analogy, reproductive success is the “target” of evolution in the real world.

    It has nothing to do with adaptations then. Fine. If it cannot explain adaptations it is no alternative to design. Back to the drawing board it is then!

  47. Rumraket: What would a WEASEL program with “multiple targets” even look like?

    It would look like a WEASEL program with multiple targets. Are you saying it can’t be done?

  48. Mung: It has nothing to do with adaptations then. Fine. If it cannot explain adaptations it is no alternative to design. Back to the drawing board it is then!

    Of course it does, because fitness is defined as reproductive success, and better fitness = adaptation. Will it ever sink? It’s really not that hard

  49. Mung: It has nothing to do with adaptations then

    Don’t you think being better adapted to your environment affords you more chance for reproductive success? If not, why not?

  50. Patrick: Please point to the doll simulation and show where the bad man inserted the target.

    Don’t be silly, it makes it look like you’re not posting in good faith and I would not want to think that you are posting in bad faith.

Leave a Reply