Subjects: Evolutionary computation. Information technology–Mathematics.
… the authors establish that their mathematical analysis of search applies to models of evolution.
I have all sorts of fancy stuff to say about the new book by Marks, Dembski, and Ewert. But I wonder whether I should say anything fancy at all. There is a ginormous flaw in evolutionary informatics, quite easy to see when it’s pointed out to you. The authors develop mathematical analysis of apples, and then apply it to oranges. You need not know what apples and oranges are to see that the authors have got some explaining to do. When applying the analysis to an orange, they must identify their assumptions about apples, and show that the assumptions hold also for the orange. Otherwise the results are meaningless.
The authors have proved that there is “conservation of information” in search for a solution to a problem. I have simplified, generalized, and trivialized their results. I have also explained that their measure of “information” is actually a measure of performance. But I see now that the technical points really do not matter. What matters is that the authors have never identified, let alone justified, the assumptions of the math in their studies of evolutionary models.a They have measured “information” in models, and made a big deal of it because “information” is conserved in search for a solution to a problem. What does search for a solution to a problem have to do with modeling of evolution? Search me. In the absence of a demonstration that their “conservation of information” math applies to a model of evolution, their measurement of “information” means nothing. It especially does not mean that the evolutionary process in the model is intelligently designed by the modeler.1
I was going to post an explanation of why the analysis of search does not apply to modeling of evolution. But I realized that it would give the impression that the burden is on me to show that the authors have misapplied the analysis.2 As soon as I raise objections, the “Charles Ingram of active information” will try to turn the issue into what I have said. The issue is what he and his coauthors have never bothered to say, from 2009 to the present. As I indicated above, they must start by stating the assumptions of the math. Then they must establish that the assumptions hold for a particular model that they address. Every one of you recognizes this as a correct description of how mathematical analysis works. I suspect that the authors recognize that they cannot deliver. In the book, they work hard at fostering the misconception that an evolutionary model is essentially the same as an evolutionary search. As I explained in a sidebar to the Evo-Info series, the two are definitely not the same. Most readers will swallow the false conflation, however, and consequently will be incapable of conceiving that analysis of an evolutionary model as search needs justification.
The premise of evolutionary informatics is that evolution requires information. Until the authors demonstrate that the “conservation of information” results for search apply to models of evolution, Introduction to Evolutionary Informatics will be worthless.
1 Joe Felsenstein came up with a striking demonstration that design is not required for “information.” In his GUC Bug model (presented in a post coauthored by me), genotypes are randomly associated with fitnesses. There obviously is no design in the fitness landscape, and yet we measured a substantial quantity of “information” in the model. The “Charles Ingram of active information” twice feigned a response, first ignoring our model entirely, and then silently changing both our model and his measure of active information.
2 Actually, I have already explained why the “conservation of information” math does not apply to models of evolution, including Joe’s GUC Bug. I recently wrote a much shorter and much sweeter explanation, to be posted in my own sweet time.
a ETA: Marks et al. measure the “information” of models developed by others. Basically, they claim to show that evolutionary processes succeed in solving problems only because the modelers supply the processes with information. In Chapter 1, freely available online, they write, “Our work was initially motivated by attempts of others to describe Darwinian evolution by computer simulation or mathematical models. The authors of these papers purport that their work relates to biological evolution. We show repeatedly that the proposed models all require inclusion of significant knowledge about the problem being solved. If a goal of a model is specified in advance, that’s not Darwinian evolution: it’s intelligent design. So ironically, these models of evolution purported to demonstrate Darwinian evolution necessitate an intelligent designer. The programmer’s contribution to success, dubbed active information, is measured in bits.” If you wonder Success at what? then you are on the right track.
The Series
Evo-Info review: Do not buy the book until…
Evo-Info 1: Engineering analysis construed as metaphysics
Evo-Info 2: Teaser for algorithmic specified complexity
Evo-Info sidebar: Conservation of performance in search
Evo-Info 3: Evolution is not search
Evo-Info 4: Non-conservation of algorithmic specified complexity
Evo-Info 4 addendum
I mean, if I crack open a book on evolutionary computation, evolutionary algorithms, genetic algorithms, natural computing, etc., these authors are all in agreement about the nature of the algorithms employed.
Now if Joe has some special knowledge about “models of evolutionary processes” which do not employ these algorithms then they would obviously not be included.
Perhaps Joe can provide a specific example.
But to the larger point. In discussions here at TSZ it is obviously the former sort that take center stage as alleged “proofs” of evolution. They are, rather, proofs of intelligent design.
That ought to be front page news here at TSZ, given all those who have denied that it is so.
Joe:
Mung:
Mung,
Who are these people? I’d be interested in seeing their denials and in hearing how they would respond to your charge.
I don’t see the source code. If you need help posting the source code to a site like GitHub let me know.
🙂
It’s miraculous. I have no objection.
I don’t believe you.
You claim that evolutionary/genetic algorithms are search algorithms and that no one here at TSZ has ever said otherwise?
Mung,
I’m not making any claims. I’m asking a question.
Read it again:
Mmmmm. Vital stuff. I anticipate that someone will have agreed that evolutionary algorithms can be used as searches but denied that’s what they all are. Much as a rock can be used as a doorstop but that’s not what they all are. They Will Pay.
Allan:
Remember, it’s Mung we’re dealing with here.
Joe was even more specific:
So according to Mung, there are people here who deny that such algorithms involve the search for solutions. I suppose it’s possible, but I don’t recall anyone making such an argument, so I would like to see Mung identify “all those who have denied that it is so.”
Note Joe says “as employed”. They are the same algos (not different sorts), just used differently.
Now if those algos have no foresight, no teleology, at best they’d support some non-teleological version of theistic evolution, definitely not any kind of Intelligent Design
But of course even that is not granted. Just because engineers use models of natural processes to produce “designs” doesn’t mean the natural processes were designed. I also would argue those are not intelligent designs, more like evolutionary designs
I’m not sure what Mung’s point is anyway. He’s on the record as being a full supporter of the reality based community’s understanding of evolution. There’s some telic component at some point to him, but he does not dispute anything in the field.
So I do wonder what phrases like:
actually mean. Is it that Mung is a full on the record supporter of something that only has alleged “proofs”? That seems odd.
Trolling troll is trolling. That’s all.
Recall that I was making the distinction between “models of evolution” as used by people like theoretical population geneticists, and “evolutionary algorithms” used by engineers to find solutions to engineering problems. I linked to my theoretical population genetics text (there are number of others, though most not free online).
The text has 11 chapters, every one of which has many sets of equatiions defining models of change of gene frequencies and genotype frequencies.
Chapters 1 through 4 deal with the effects of random mating, natural selection, mutation, and migration. Chapter 5 deals with inbreeding. Chapter 6 with random genetic drift. Chapter 7 with the combination of random genetic drift with natural selection, mutation, and migration. Later chapters with multiple-locus models.
I don’t give source code there because it is not necessary. For example, from Chapter 2, equation II-31 shows the gene frequency in the next generation
as a function of the current gene frequency 
in a case with an infinite population with one locus that has two alleles, when the three possible genotypes have relative fitnesses 
, 
, and 
,
The book is full of models of different evolutionary processes, and of these processes in combination. If you prefer models of finite populations, chapter VII, section VII.9, page 324 of the book (which is page 342 of its PDF) explains the Wright-Fisher model for one locus with two alleles in a finite population of size
. Again, equations that define the stochastic process, but no source code.
But OK, is there source code out there? Well, yes. For use in teaching about evolutionary processes, I have a program, PopG, which simulates that very case, the Wright-Fisher model with two alleles. You will find it here. It is in Java, and as explained in its documentation, the source code is available in the files you download.
Recall that the larger issue is that Marks, Dembski, and Ewert have theorems that show that “evolutionary searches” typically do very badly. They have argued that models used by evolutionary biologists are included in their “evolutionary searches”. Does that show that such models that have multiple genotypes, with natural selection, mutation, migration, and random genetic drift are ineffective at showing improvement of fitness? No, as Tom English and I have shown. Because MDE also include in all “evolutionary searches” lots of models that do horribly because they don’t have fitnesses, or because they actually reward genotypes with the poorest fitnesses. And their theorems describe average performance over all the “evolutionary searches”, not the ones used by evolutionary biologists to model evolutionary processes.
Their Weasel analysis did not make its way into Chapter 6, “Analysis of Some Biologically Motivated Evolutionary Models.” (The outrageous title is not just a slip.) Their analysis of a mathematical model — no code — from Chaitin’s metabiology did make the cut.
Excellent. That’s precisely what I dispute. As I indicated earlier in the thread, Marks et al. address a kind of model that specifies an evolutionary process along with an event that tends to occur in the process. The math that they gather under the “conservation of information” rubric assumes that the event is determined independently of the evolutionary process. If you look at the theorems in their technical papers, you’ll see easily enough that they start by identifying a fixed “target” T — the set of all solutions to a problem. That makes perfectly good sense. The choice of a solution-generating process generally depends on the problem, but the problem does not depend on the choice of solution-generating process.
If the modeler does not say that the event in the model is determined independently of the evolutionary process in the model, then the model is not one in which an evolutionary process generates a solution to a given problem, and the analysis of Marks et al. does not apply. I cannot imagine why a modeler would say such a thing. And no one can put the words into the modeler’s mouth. The model is what the modeler says it is. When in doubt as to the intent of the modeler, what do you do? Ask the modeler, of course.
What Tom emphasizes:
(I do so love stories of how a large community of scientists succumbs to irrationality obvious to anyone not blinded by sin.)
Marks, Dembski, and Ewert indeed can apply their math only when a goal (target, problem) is specified in advance. Never, in any of their publications, have they acknowledged the if I’ve just highlighted. They’ve merely asserted that the event in the model is specified in advance, calling it a target. Modelers in fact jointly specify evolutionary processes and events that tend to occur in them. The event is not “specified in advance” of the evolutionary process in which it tends to occur. But I am not the one claiming to analyze the models. It falls to Marks et al. to show that the assumptions of their math are warranted: “If a goal of a model is specified in advance…” The fact that they’ve never bothered is itself a major indictment of evolutionary informatics.
Please don’t lapse into an argument about an argument about… an argument. I think we’re communicating pretty well now (apart from the fact that I’m moving in slow motion with responses).
Asserted, not argued. The issue is whether the event of interest to the modeler is determined independently of the evolutionary process specified by the modeler. In the GUC Bug, the event is defined in terms of the evolutionary process.
Their response to ev is outrageous, but relatively difficult to explain. Schneider wasn’t out to “prove” that binding sites and a binding site recognizer coevolve rapidly. His interest was in providing an abstract model of entropy reduction (information gain) at binding sites. He began with observations of nature. He didn’t start with an event determined independently of the evolutionary process.
Particularly if there is a space of genotypes that each have a fitness, and one argues that the “target” is some subset of them, the ones having high fitnesses, and calls the evolutionary process a “search”. Kind of hard for the target to be chosen independently of the process in those cases.
It is an endless argument, one that I am embarrassed to contribute to further … but hey. If one has an evolutionary algorithm that starts with a single genotype and allows it to run with mutation only, no fitness criteria, one sees a particular kind of evolutionary behaviour. If one allows this to run and run, and observes the shifting pattern … what is one searching for?
So it’s the presence of selection – the subsetting of all genotypes – that makes it a ‘search’ in some eyes, not the being-an-evolutionary process? But even then, you may not be remotely interested in actual genotypes, which inevitably form a subset of the whole when fitnesses are not equal. The moment the fitness landscape starts to ripple, you’re ‘searching’ for the heights.
What are the heights? How do you decide?
It is a matter of fact that individuals of different genotypes tend to leave different numbers of offspring. It would be quite an extraordinary circumstance if the number of offspring left by an organism did not depend on its heritable traits. The latter would require some special explanation. The former does not.
But Allan is talking about evolutionary algorithms. So if in an evolutionary algorithm fitness is defined by how many, then this is of course the fitness function, thus the whole “give a computer a goal, let it search for that goal” aspect of a computer program can never be overcome. It can never do what it is claimed evolution does-which is give it no goal and see if it still makes meaningful, intelligent things.
A search algorithm is a search algorithm is a search algorithm.
It’s an algorithm. By definition, it’s teleological.
Compared to what? Are they alone in thinking that? How do we measure the performance of an evolutionary search?
Is that in dispute? I didn’t think that was in dispute.
Sorry, I took keiths off ignore. Fixed that.
You’re right to bring that up. I started to address his use of the term earlier, but got bogged down, and gave up. Sorry.
It seems to me that what Allan means by evolutionary algorithm is an algorithm for simulating evolution.
I want to address the rest of your comment, but don’t have it in me at the moment.
He’s pretending that he’s talking about evolutionary algorithms. How does he measure performance?
Here’s Wikipedia on Evolutionary Algorithm:
They are the mirror images of the holes.
Now if we just had a way to measure performance.
From the Preface:
From the OP:
Search me too!
Tom English,
Well, I mean any algorithm that operates upon a string population with analogues of birth and death, whether problem solving – for the purpose of ‘finding’ certain kinds of genotype – or analytical, investigating the behaviour of such processes given various conditions. If I’ve misused a technical term, allow me to say “Waaaaaah!”! I didn’t capitalise it, does that get me off the hook?
phoodoo,
You don’t need to decide in my conceptual illustration. A fitness landscape is either completely flat, or has regions of relative height and depth. I don’t see any alternative.
Well, first off the concept of height and depth in a computer program is completely artificial and meaningless. There is no actual shape to any landscape, its just however one wants the pixels to fall on their computer screen. Do you think an X, Y graph is the only alternative?
Thus, what is the “landscape?” There are some of A, some of B and some of C… Maybe there are more of A than of B. Is that good? Or bad? Maybe A is like a cancer cell, so less is better. Maybe C is in between A and B? Is that good? If we tell the computer to look for the group that is not the most and not the least, then maybe C is good.
Again, meaningless until you decide what you are searching for.
phoodoo,
It is artificial but not meaningless. It is one way of representing the concept of a fitness differential between genotypes (which can be zero, hence a ‘flat’ landscape). There are others.
You don’t need to be searching for anything to have a fitness landscape. Fitness landscape is another way of representing the different genotypes in the space of all possible genotypes in terms of their propensity to be copied. You don’t have to be searching for those genotypes, nor to make any particular decision about them, for this to be true.
In a simulation of evolution, you aren’t searching for anything.
What does that even mean? What discussion in particular, and in what way is some program hailed as “proof” of evolution? Citations please.
If evolution is a search, Mung, presumably the designer has decided what is being searched for? How does it do that? Is the target in the DNA? Can we look at it? Is the target outside of the organism? Where?
How does Intelligent Design Evolution work Mung?
Then, presumably, you can demonstrate proof of intelligent design via biology, as above. Please do so.
I am not sure if you are intentionally or unintentionally struggling with the conceptual problem inherent in every evolutionary computer algorithm. It seems it has been pointed out to you so many times, that perhaps you are just being facetious.
First the program needs to know what it is attempting to do. if it is making different genotypes, than it can just make any random genotypes and no one genotype can ever be considered any better or worse than another, so there is no way to arrange them into any landscape. BUT, AS SOON AS you decide to tell the computer to arrange the groups into a counting, or census of individuals, you have made a choice, a search. You have searched for which ones are more plentiful and which are less.
But what is the “fitness”? There is no such thing as fitness in this example, until you decide what you are calling fit. If you say more equals fit, than its an arbitrary determination that the programmer, not the computer has generated. The programmer says, “More is good” show me what is more. But the programmer could just as easily say “Less is fit.”
So then, the MOST fit genotypes are those which don’t exist at all. And the least fit are those which are most numerous. If eventually one genotype takes over and is nearly the entire population, then it has taken over because it is extremely unfit.
You have shown nothing, other than you can call anything you want fit or unfit, and you haven’t done anything remotely close to simulating evolutionary theory.
No, it doesn’t. Programs don’t know anything for any common sensical definition of the word ‘know’.
Look, I get what you’re saying. Somebody has to design into the program some idea of fitness. The program has to have some way of “making” some things more fit than others.
Here’s the issue though: No, it doesn’t actually have to do that.
It is possible (and it has been done) to make a computer program that simulates organisms and an environment. These organisms make copies of themselves, and they mutate when they do. And those copies then either are better or worse at copying themselves. At no point has anyone made any code that tells the program which copies are better or worse at copying themselves. Those that ARE better at copying themselves, eventually make the worse ones extinct.
That computer program is called avida. There isn’t any line (or collection of lines) of code in Avida that says that some particular trait that emerges from the mutational mechanism in the copying process, is an advantageous (aka high-fitness) trait.
So no, while it is true that for SOME programs a sort of fitness landscape has been “defined” by the programmers, this is by no means true of ALL programs in which the concept of fitness can be used.
phoodoo,
This is a fundamental error. A fair starting point for understanding the concept of a fitness landscape is that very situation. If no genotype can be considered better or worse, the landscape is flat, not nonexistent.
The concept of a fitness landscape can be useful. But hey, it is not central, and not vital that you understand it. I’m not going to expend too many words getting you to do so. My original comment about ‘ripples’ was directed to someone who understands the concept, not you.
Perhaps incorrect points don’t stop being incorrect with repetition?
What is “better” at copying, and why is it better?
The answer to THAT the computer being told what is fit and what isn’t.
If you just tell a computer, start copying things, but make mistakes, then some mistakes will occur more than others. It is only when you say that the fact that there are more of one mistake than another and we want to call them better, that the programmer has decided what is fit. Why aren’t you calling the mistakes that are less numerous -“fit”?
What a simple concept, and boy do evolutionists struggle with it.
Allan appears to be waving the white flag.
phoodoo:
Oh, the irony.
phoodoo,
No, we did that, had a friendly game of soccer in no-man’s land, and now you’re back to your usual modus operandi of dogged intransigence. Yes, I know you think it’s me. Whatevs.
At the moment I’m waiting for the library of the Tierärztliche Universität in Hanover (which for some reason bought a hard-copy of the book) to send it to my local library: it takes a few of days, but I don’t have to buy the book.
I wonder which definition for search is used in the new book: the weird initiator – terminator – inspector – navigator – nominator – discriminator process? IMO this was created mainly to separate the function which is to be optimized from the target: In the usual definition, you succeed if you find the element on which this function has its optimum, with DEM, this element is generally not the target. This makes symmetrical search spaces the norm (for which the no-free-lunch apply), but in reality, asymmetrical spaces seem to be very common and quite interesting…
Those that make more copies than others.
Better in the sense that I used the word here, is only a description of the fact that they out-compete those that are worse at making copies of themselves.
Because of the random mutations.
I’m not the one deciding they’re “better”. The program doesn’t decide, and has not been designed to somehow “make sure” that more copies are “better” than less copies.
Did someone have to somehow design the world such that one bacterium eats less than two bacteria? That seems a bit silly. Even if God only designed a single bacterium, it wouldn’t really make sense to say that God somehow intended or “designed” the fact that, if there were two bacteria, they would eat more than a single one did.
It seems to be almost unavoidable that more copies dominate and overwhelm fewer copies. Nobody has to somehow “make it” so. In fact, it seems to me that if you wanted to avoid this, you’d have to deliberately alter the rules to prevent it.
And what does it mean to say that evolutionary algorithms support ID anyway? What was supposed to be designed? The whole process? The fitness landscape (the environment)? The mutations? And how does that undermine evolution? All one could get from that crap is that evolution works exactly as it’s supposed to but either some aspect of it, or all of it, was somehow “designed”
Rumraket,
My goodness, do you not understand that in a computer algorithm nothing is better or worse, and there is no competition, until you make a competition and decide what wins and what loses. Why does more copies dominate less copies in a computer algorithm? Why does one eat another unless you tell the program to do that? Does the existence of two versus the existence of one in a computer screen, means that two will eat one?
What you have just suggested is so far away from the understanding of the problem, that I am gobsmacked. More copies of something on a computer screen does not in any way whatsoever “out-compete” less copies of something on a computer screen, until you describe a competition and what the definitions of winning are.
Until you define that, all you have are copies and mistakes of copies. Some mistakes more than others, there is no fitness, no direction towards anything, no meaning, no winning or losing. You haven’t simulated anything. What is the relationship to evolution??
If you put 5 red dots, 3 green dots and 6 yellow dots on a computer screen, and then say, Ok, my game is over, which color won-there is no answer to that question, because we don’t know what the meaning of winning is. In golf the objective is to score less, in bowling it is to score more, in swimming it is to go faster, and in figure skater it is to look graceful. If 3 green dots look more graceful than 6 yellow dots, perhaps 5 red dots is the winner
Holy cow.
You are one confused dude, phoodoo.
I have the book. Looking through it, I find that none of those functions are described in the book (though it does cite their several papers that include those terms). Nor are those terms found in the book’s index.
Tom, who has a higher pain threshold than I do, has read all those papers closely and reports that they went through at least three different schemes of that sort. In any case, when they were showing how bad the behavior of an average “evolutionary search” they didn’t actually use any of that. You’d expect that, if they took that scheme seriously, that they would then consider randomizing over all possible terminators, inspectors, etc. to get a randomly chosen “evolutionary search”. But they didn’t do anything of the sort.
Instead, in those papers, they identify an “evolutionary search” with a distribution of outcomes, and talk about all such distributions. They then find (ta-da!) that the result of a randomly chosen “evolutionary search” is just a random element in the space of genotypes, so that the typical “evolutionary search” does no better than random.
As I have noted in this thread before, that includes among “evolutionary searches” all sorts of horrible processes that can look for the worst possible fitness, or just look at random. And as Tom and I showed in last year’s post at Panda’s Thumb, as soon as we make the reasonable requirement that an “evolutionary search” have genotypes that have fitnesses, and thus reward the reproduction of more-fit genotypes, the “evolutionary searches” do much better than random, and thus, in DEM’s terminology, incorporate “active information” without need for a Designer.