Saturday, September 27, 2014

Conservation of Information in Evolutionary Search - Talk by William Dembski - part 3

For an introduction to this post, take a look here. There is some interaction with the audience (15'30" - 18'00") which I wasn't able to understand fully. Any help is appreciated!

Part 3: 12' 45" - 31' 25"

Topics: What is an evolutionary search?

William Dembski: Now let's add this next term evolutionary. What does evolutionary - when we put it in front of search - add to the discussion? I think it changes one key aspect here. Whereas we were looking at some query feedback, now this query feedback takes the form of fitness: how good is it? Query feedback can be quite general. Maybe the query feedback is nothing, when we examine it. Or maybe the query feedback may just say "I'm in the target" or "I'm not in the target". That would be very simple. Fitness is going to give some sort of range of values that ideally identify how close am I to the target.

William Dembski: There are examples of evolutionary search. There is the Dawkins' weasel example from his book "The Blind Watchmaker", that is the one I'm going to focus on here. Then there are various - what I would regard as - embellishments of that, because I don't think that there is anything fundamentally new about them. There is MSU's Avida program, Tom Ray's Tierra, Schneider's ev. What is at the heart of these programs that these are computer programs which mimic - try to mimic - Darwinian evolutionary processes. What are they supposed to show? That is interesting. Look at the history of this field of evolutionary computing and there is a reason why people wanted to do evolution in the computer. That is because the computer would allow evolution to be done in real time, because we cannot really see it in real time in the wild.

William Dembski: Nils Barricelli in 1962: "The Darwinian idea that evolution takes place by random hereditary changes and selection has from the beginning been handicapped by the fact that no proper test has been found to decide whether such evolution was possible and how it would develop under controlled conditions".

William Dembski: J. L. Crosby says substantially the same thing in '67.

William Dembski: Heinz Pagels in a popular book in 1989 wrote "The only way to see evolution in action is to make computer models because in real time these changes take aeons, and experiment is impossible".
Now, there is Richard Lenski at Michigan who - I think - has run 30 - 40,000 generations of E. coli, which probably corresponds to a million or so years of primate evolution. But I'd say that he has not seen a whole lot of changes, at the end of the game E. coli is still E. coli. So if you want to see some massive saltations, I think what Heinz Pagels says does still apply.

member of the audience: Can I ask you something?

William Dembski: Yes.

member of the audience: The Times had a very interesting article, very recently, exactly about this point. It was about a book that was written by Peter and Rosemary Grant. They looked at finches. And the claim is that they actually did exert evolution in forty years of time. They were basically looking at the evolution of finches in the Galapagos Islands. So, can you speak to it?

William Dembski: Finch beak variation, yes, in this case it was [???], they saw some. There were some changes which Richard Lenski saw in E. coli, but I think what is supposed to make evolution interesting is not how finches' beaks vary, but how you get beaks in the first place, how you get birds in the first place. That is the sort of evolution that I think these people who are talking about evolution in silico are thinking about: that we can really speed it up, so that we can see some of these big, impressive evolutionary changes.

member of the audience: So small evolutionary changes don't bother you.

William Dembski: It's not a question of bothering me. They are there. I mean, the evidence for them is clear. I think there is even evidence for large-scale evolutionary changes. The question is: what is driving them? For Darwinians, it is natural selection. For Non-Darwinian, those mechanisms seem to be insufficient.

William Dembski: You are standing up [???]

member of the audience: [???] About two plants growing together and there cells fusing. [???] You get new species. And that's how we make new species in real-time. So, evolution can occur. There is a 1954 [???] Scientific American [???] cataclysmic variation [???]

William Dembski: That happens only in plants. I don't know of any case like that in animals.

Leo Kadanoff: Okay, you made your point. [???] Go ahead. I'd build my argument for example [???] two plants [???].

William Dembski: You might argue better by using other plants. Let's look at this example. I don't know how many of you have read the book "The Blind Watchmaker". This is an example that [???] worked countlessly, even in literature trying to justify the power of Darwinian processes to create information. Underscore that word "create", because that what it is about: is it creating or is it shuffling about already existing information.

William Dembski: Let's look at this example in vantage of search - I had these seven key components. What is that example. What you are trying to do, you take a random string of 28 letters and spaces - that's the reference class, that's the search space: letter and spaces. So there are $27^{28}$ possibilities. Start out with a random sequence - that is the initialization. Your target is "METHINKS_IT_IS_LIKE_A_WEASEL", this is a line from Shakespeare's Hamlet. You have a fitness that is going to measure how many letters correspond in a given sequence to the target sequence, so, that is basically a Hamming measure. You are going to have an update rule which is going to say "take an existing sequence and then - one possibility would be - generate 50 offspring by some sort of random mutation process and then take the one that is closest, and that becomes the next one", so that becomes the update rule. Stop criterion is "you stop when you hit the target sequence". And then the query limit is going to be whatever your computational resources allow. The thing is, with this setup, you are going to evolve to this final target sequence very, very quickly.

William Dembski: I'm just trying to give you a sense that all the components are there in this example. The fitness function in this case is a unimodal fitness where basically you are counting the distance letter by letter from the target sequence. For instance, here we have a score of 27, because you have a "J" where there should be a space. So, when the "J" disappears, then we are actually there. That's the example. I will talk about it a bit more.

William Dembski: I'm throwing that in as a type of digression. There is a kind of lunatic vitality to this example. I keep seeing it in places, and people keep challenging me on the Internet because I come back to this example as though this somehow misses something fundamental or that it is to simplified. But in fact this example just keeps getting reworked. Most recently - I thank a member of the audience for pointing that out to me - Michael Yarus in his 2010 book [???]. The target phrase for him is NOTHING IN BIOLOGY MAKES SENSE EXCEPT IN THE LIGHT OF EVOLUTION. There is a popular book by Jeffrey Satinover, in the "The Quantum Brain" MONKEYS WROTE SHAKESPEARE. Bern-Olaf Küppers in the 1990s, his target phrase was EVOLUTION THEORY. This type of example, where you are evolving symbol strings to some target, keeps getting used in the evolutionary literature to justify biological evolution. That is where we want to go with this. The question is: evolutionary search as I've described it to you, this is widely done, in some ways it is part of computational intelligence, in the sense of evolutionary computing, genetic algorithms, even falls under operation research as some kind of optimization procedure. How does this compare to real life evolution? Now, there are people who think that actually the computational case does provide justification for real life.

William Dembski: Robert Pennock for instance, who worked on this AVIDA program, he says: "I do scientific research on experimental evolution and evolutionary design using evolving computer organisms, including work showing how evolutionary mechanism can produce the kinds of complex features creationists say is impossible... My colleagues and I have demonstrated experimentally that a Darwinian mechanism can discover irreducibly complex systems." I think he is overstating his case, there are some details his leaves behind. The thing to get from this is that he is using what is happening in computational evolutionary searches to justify biological evolution.

William Dembski: Ken Miller in his 2008 book "Only a Theory" - he is a biologist at Brown University - says what is needed to drive biological evolution (that is the question he poses): "Just three things: selection, replication, and mutation... Where the information 'comes from' is , in fact, from the selective process itself." I would say that this is actually the received view, that the Darwinian mechanism is able to produce all these nifty things that you see, that all this biological information can be handed over to Darwinian mechanisms, and there we go. I want to address this from the vantage of what I call the "Conservation of Information", but before I do this, I want to create some doubts for you that this can be the whole story. Not by invoking anything like "Conservation of Information", but by actually going back to somebody at the time of Darwin who was looking at the logic of induction, and raised a method of induction, that actually - I think - undercuts this kind of Darwinian mechanism to produce, to create biological information.

William Dembski: This is Mill's method of difference. He formulated this in his "System of Logic" in 1843. It run to eight editions, the last edition was 1882, so he is a contemporary of Darwin. Mill's method of difference shows that the Darwinian mechanism by itself cannot generate biological information. How does that work?

William Dembski: The method of difference says: "To explain a difference in effects, one must identify a difference in causes." What does that mean?

William Dembski: Common causes cannot explain differences in effects. Imagine, here is a difference in effect: Slowed reflexes versus ordinary reflexes.

William Dembski: Watching television, combing hair, o-oh, consuming alcohol. Alcohol is the difference maker. One person consumed it, the other person didn't. You have people watching television or not watching television, that is not making any difference. The difference maker which accounts for the slowed reflexes versus the ordinary reflexes is consuming the alcohol. Now let's look at the Darwinian mechanism.

William Dembski: We have replication, heritability, random variation, natural selection, all these basic components of the Darwinian mechanism. When you run a Darwinian mechanism, if you are a Darwinist, then you would say in a cellular context it is going to produce, we are going to see a lot of interesting evolution. But there are cases - for instance, Sol Spiegelman had an experiment back in the sixties in which he looked at polynucleotide synthesis and found instead of these evolving polynucleotides becoming more and more complex and more interesting, in fact, they tended towards simplicity, where the replicators would replicate as quick as possible. What supposed to make evolution interesting is that we go from monad to man, right? It is not that we go from cave-fish or cave-fishes that have working eyes to cave-fishes with eye-knobs, because in a case of use it or lose it, in this dark environment they have lost it and now they have eye-knobs. That is evolution, but that is not interesting evolution. It is how you these eyes in the first place, how you get the beaks in the first place, how you get the birds.
Cellular automata: You can have cellular automata that follow Darwinian principles and never go anywhere. And artificial life, [???] the same thing. You can have cases of interesting evolution and evolution that goes in a simplifying direction, that goes nowhere, with all these features. If this is the case, if the Darwinian mechanism is common to cases where you have interesting evolution and evolution that is not going anywhere, then something besides the Darwinian mechanism must being involved. That is the logic. It seems to me that this should be uncontroversial.

William Dembski: But Stuart Kauffmann, a complexity theorist who is not Darwinian, and not an Intelligent Design guy like me, has seen this problem. I think he puts it very well in his book "Investigations". He says: "In the absence of any knowledge, or constraint, on the fitness landscape, on average, any search procedure is as good as any other."
This is a no-free-lunch theorem, which actually really upset people. Jon Holland and the evolutionary [???] community back in the nineties - I have a colleague who was there on one of their meetings when this happened.
"But life uses mutation, recombination, and selection. These search procedures seem to be working quite well. Your typical bat or butterfly has managed to get itself evolved and seems a rather impressive entity.... If mutation, recombination, and selection only work well on certain kinds of fitness landscapes, yet most organisms are sexual, and hence use recombination, and all organisms use mutation as a search mechanism, where did these well-wrought fitness landscapes come from, such that evolution manages to produce the fancy stuff around us?... No one knows"
When I pose this to Darwinians, they often say: "Well, it is just the environment. That is where we get the fitness." I will revisit that. I think Kauffmann asked the right question here, it is a question that many people do not even see is a question. Let's go back: there are seven key components of our evolutionary search. Question is: where is the information coming from? We do this in a computational context, this is usually where it is, it is put there in the fitness, it is put in the update rule. My friend Bob Marks had a colleague at Boeing who called himself a himself a "penalty function artist". If you had the right penalty function, the optimization problem was solved. What is a penalty function? That is basically the inverse of a fitness. [???] That is usually where it comes in. Where does the information come in in this METHINKS IT IS LIKE A WEASEL? It came in obviously in setting up the fitness. You have a unimodal fitness function which measures how close you are to this METHINKS IT IS LIKE A WEASEL target phrase. You could have set up a fitness for any other phrase, for gibberish, and it would have evolved there. It was by choosing that fitness that you got it to evolve where it did. By the way, there are about $10^{40}$ ($27^{28}$) sequences of length 28 having 27 possible characters. Any idea how many unimodal Hamming-distance fitness landscapes there are over that space? It is the same: $10^{40}$. For every possible element there you got a unimodal fitness landscape. What he has done there is to say "I evolve this thing to the target sequence", but what he has not told you is "In doing that, I had a fitness landscape which I have carefully adapted". The search for the target phrase became the search for the right unimodal fitness landscape. This is a expression Paul Nelson - a good friend and colleague of mine - gave to me, which I use over and over again: "Filling one hole by digging an other".

A longer excerpt this time, and one which include a few gems, though I apologize for not getting everything which was said. My thoughts:
  • "At the end of the game E. coli is still E. coli." Yes, William Dembski really did say this.
  • The audience seems to have expected to be confronted with a creationist like Ken Ham, but Dembski has not problem with evolution, neither on the small scale nor on the large scale.
  • However, he uses the term "interesting evolution" as a kind of straw-man: things have to get more complex and considerably diverse. Is the creation of a tiny bit of information by a Darwinian process unproblematic for him? I doubt it...
  • I'm not totally convinced by Dembski's application of the method of differences, he seems to ignore the influence of chance altogether: neglecting the influence of chance, two guys playing Russian Roulette should end up both dead and both alive...
  • Flogging the WEASEL takes an awful amount of time. Why does he not talk about a the Traveling Salesman Problem? Because the information "smuggled in" cannot be detected? Who searched for the fitness landscape?
  • BTW, at 30'15'', there is an impressive animation illustrating how big the number 40 is....

4 comments:

  1. What does evolutionary - when we put it in front of search - add to the discussion? I think it changes one key aspect here. Whereas we were looking at some query feedback, now this query feedback takes the form of fitness: how good is it? Query feedback can be quite general. Maybe the query feedback is nothing, when we examine it. Or maybe the query feedback may just say "I'm in the target" or "I'm not in the target". That would be very simple. Fitness is going to give some sort of range of values that ideally identify how close am I to the target.

    I have drawn pictures in my demonstration that what is called "search" or "black-box optimization" in computing is actually sampling. Indeed there is feedback of data. But data-processing does not generate information about the fitness of elements of the sample space yet to be evaluated. Any inference about the fitness of unevaluated elements is necessarily inductive, i.e., an expression of assumptions about the relation of what has been observed to what has yet to be observed. It's highly ironic that Dembski should, in a talk about conservation of information, make the mistake of saying that feedback is informative.

    Now, Bill Dembski, where is your picture of what's going on in nature? Identify the natural entity that generates "queries" and registers "feedback" about their fitness. Identify the natural entity that responds to queries about the fitness of organisms. They do not exist. You've made the rookie error of reifying the model -- and quite a simplistic model, at that. A computer program may implement a system evolving according to Darwinian principles, but that does not mean that you can ascribe components of the implementation to nature. "The map is not the territory."

    I think there is even evidence for large-scale evolutionary changes. The question is: what is driving them? For Darwinians, it is natural selection. For Non-Darwinian, those mechanisms seem to be insufficient.

    This may seem trivial, but it is actually a crucial point: natural selection is not a mechanism. Nothing in nature is doing it. Ernst Mayr referred to natural selection as a phenomenon, and emphasized differential reproduction. Now, Bill Dembski, tell me what extrinsic entity is reproducing organisms. I'm suffering the delusion that organisms reproduce themselves, with differing degrees of success that depend on the nonlinear interaction of a host of factors, some of which we can identify, and most of which we cannot. The factors are different at different times and in different places. The notion of reproductive fitness as a function of just the organism (or some part of it, e.g., the genotype) is a modeling fiction, useful under certain restrictive conditions.

    ReplyDelete
  2. I have never before seen Dembski reduced to such pathetic deflection.

    member of the audience: [???] About two plants growing together and their cells fusing. [???] You get new species. And that's how we make new species in real-time. So, evolution can occur. There is a 1954 [???] Scientific American [???] cataclysmic variation [???]

    William Dembski: That happens only in plants. I don't know of any case like that in animals.

    Leo Kadanoff: Okay, you made your point.


    No, that was point, game, and match. Dembski pretended otherwise, and fell on his face trying to jump the net. Phenomena in plant evolution don't count unless they are known also to occur in animal evolution? Dembski truly plays the Über-IDiot here, managing also to slip in a false dichotomy. What about eukaryotic bacteria engulfing prokaryotic bacteria that evolved, as endosymbionts, into mitochondria and chloroplasts?

    ReplyDelete
  3. In his botched attempt at applying John Stuart Mill's (1843) method of difference, Dembski writes (with abbreviation) that replication + heredity + random variation + natural selection $\neq$ causal difference. This isn't even a good approximation to Darwinism. There is sure to be differential reproduction under the conditions of heredity, variety, and fecundity. However, as I indicated above, there are huge differences in what accounts for differences in reproductive success of different biological types at different times and different places. The notion that you can simply invoke the general term for an enormous range of outcomes, "natural selection," and reduce it logically to a single cause is doubly preposterous.

    ReplyDelete
  4. The audience seems to have expected to be confronted with a creationist like Ken Ham, but Dembski has not problem with evolution, neither on the small scale nor on the large scale.

    He has more problem with evolution than he's saying in this talk. He's made it clear that he has doubts about common descent, and when he was teaching at Southwest Baptist Seminary he was required to adhere to a Young Earth perspective (which he did grudgingly and which school he finally left). He's saying that here is a problem even if he grants the reality of evolution and the effectiveness of natural selection. But just as a hypothetical.

    ReplyDelete