Searching for a search

Dembski seems to be back online again, with a couple of articles at ENV, one in response to a challenge by Joe Felsenstein for which we have a separate thread, and one billed as a “For Dummies” summary of his latest thinking, which I attempted to precis here. He is anxious to ensure that any critic of his theory is up to date with it, suggesting that he considers that his newest thinking is not rebutted by counter-arguments to his older work. He cites two papers (here and here) he has had published, co-authored with Robert Marks, and summarises the new approach thus:

So, what is the difference between the earlier work on conservation of information and the later? The earlier work on conservation of information focused on particular events that matched particular patterns (specifications) and that could be assigned probabilities below certain cutoffs. Conservation of information in this sense was logically equivalent to the design detection apparatus that I had first laid out in my book The Design Inference (Cambridge, 1998).

In the newer approach to conservation of information, the focus is not on drawing design inferences but on understanding search in general and how information facilitates successful search. The focus is therefore not so much on individual probabilities as on probability distributions and how they change as searches incorporate information. My universal probability bound of 1 in 10^150 (a perennial sticking point for Shallit and Felsenstein) therefore becomes irrelevant in the new form of conservation of information whereas in the earlier it was essential because there a certain probability threshold had to be attained before conservation of information could be said to apply. The new form is more powerful and conceptually elegant. Rather than lead to a design inference, it shows that accounting for the information required for successful search leads to a regress that only intensifies as one backtracks. It therefore suggests an ultimate source of information, which it can reasonably be argued is a designer. I explain all this in a nontechnical way in an article I posted at ENV a few months back titled “Conservation of Information Made Simple” (go here).

 

As far as I can see from his For Dummies version, as well as from his two published articles, he has reformulated his argument for ID thus:

Patterns that are unlikely to be found by a random search may be found by an informed search, but in that case, the information represented by the low probability of finding such a pattern by random search is now transferred to the low probability of finding the informed search strategy.  Therefore, while a given search strategy may well be able to find a pattern unlikely to be found by a random search, the kind of search strategy that can find it itself commensurably improbable i.e. unlikely to be found by random search.

Therefore, even if we can explain organisms by the existence of a fitness landscape with many smooth ramps to high fitness heights, we have are left with the even greater problem of explaining how such a fitness landscape came into being from random processes, and must infer Design.

I’d be grateful if a Dembski advocate could check that I have this right, remotely if you like, but better still, come here and correct me in person!

But if I’m right, and Dembski has changed his argument from saying that organisms must be designed because they cannot be found by blind search to saying that they can be found by evolution, but evolution itself cannot be found by blind search, then I ask those who are currently persuaded by this argument to consider the critique below.

First of all, I think Dembski has managed to mislead himself by boxing himself into the “search” metaphor, without clarifying who, or what, is supposed to be doing the searching.  When I am struggling to understand an argument, and unclear as to whether the laws lies in my own understanding or in the argument being made, I like to translate the argument into E prime, and see whether, firstly, still makes sense, and secondly, leaves out some crucial information (information that has been “smuggled out” of the argument, as it were :)).  In A Search for A Search, Marks and Dembski write:

A search’s difficulty can be measured by its endogenous information defined as

I=−log2p

where p is the probability of a success from a random query. When there is knowledge about the target location or search space structure,

Translated into E-prime (avoiding the verb to be and the passive voice), this becomes:

We can measure the difficulty of a search by its endogenous information, which we define as

I=−log2p

where we represent the probability that the searcher will find the target using a random query as p. When the searcher knows the location of the target, or a way to find the location of the target, the probability of finding it will increase, and we define this increase in probability as the active information [possessed by the searcher].

See what I did there? E-prime forces the writer to specify the hidden doer of each action, and, in this case, reveals that the “active information” is that possessed by the searcher at the start of the search. But who is the searcher?  And how does the information transfer take place?

Now, the searcher doesn’t need to be an Intelligent (i.e intentional) Agent.  It could be a mechanical algorithm, or physical system, that results in something that is special (if not formally specified) in some way (a cool pattern, a functioning organism, a novel feature), and also something that is unlikely to just turn up (“blind search”).  So in keeping with the mathiness of the paper, well denote the Searcher as S, and the special result (which could be one of many possible results that we’d consider Special) as R (I want to avoid T for Target, which I think is another siren that may lure us to the rocks instead of the deepwater channel).

So what would it actually mean for S to be in possession of Active Information that would make R more likely?  Well, let’s take some concrete examples.  Let’s say S is  Lizzie looking for her car keys, and the Search Space is her house.  If Lizzie pats every square inch of surface in her house with her eyes closed, and with no clue on which surface she might have left her keys, her probability of finding them on any given pat will be 1 divided by the number of square inches of space in her house. But if she knows she left them on the kitchen table (i.e. knows the location, or a subset of possible locations that must contain the location), or knows that if she thinks back and remembers what she did when she last came in from the car (knows how to acquire knowledge of the location), hee probability of finding them will go up considerably, in other words, 1 divided by the number of places she has to pat will be quite small.

But let’s say R is not an object, like keys, but some kind of physical pattern or configuration with some rare property, for example, a run of 500 coin tosses in which the product of the runs-of-heads is large (i.e. one of a rare and specified subset of all possible runs of 500 coin tosses), as in my thread here. We can compute (as we do in that thread) just how large I is for patterns of coin tosses with a given magnitude of product-of -runs-of-heads by computing how rare those patterns are when generated by real tosses of a fair coin, and can regard R as any patterns with an I value over some threshold.  So what would we have to do to make R more likely?  Dembski and Marks quite reasonably, say that anything we do to make R more likely will itself be something less likely than a simple coin-tosser (and coin-tossings are fairly common, therefore fairly likely). Well, we could get a human being to sit down and work out a few runs that had high I values, and manually place them on the table.  In which case, presumably Dembski and Marks would argue, the human “searcher”, S, would now be in possession of Information commensurate with, or greater than, the Information I .  Which I am happy to accept (whether such an agent is rarer than a coin-tosser, I don’t know, but probably, and in any case in this scenario we are positing something – an intelligent human being, which may itself by much less probable than say some simple physical process that by which coin-like objects regularly fall off cliffs).

But let’s now say that instead of  Lizzie using her intelligence to work out a good run, and then lay it down manually, Lizzie wants to set up a system that will, all by itself, with high probablity, result in – find – a run of coin-tosses with high I.  To do this, she decides (as I did) to write an evolutionary algorithm, in which the starting population consists of a population of runs of 500 coin-tosses generated by a quasi-coin-tossing method (each successive coin toss independent of the previous one, with .5 probability of each being heads), i.e runs that are the result of “blind search”, but on each iteration, the members of the population of runs “reproduce”, with random mutations, and those runs with the lowest product-of-runs-of-heads are culled, leaving the higher performers in the game for the next round.

Clearly this is an informed search.  Lizzie has constructed a “fitness landscape” in which runs that have more of the desired feature (high product-of-runs-of-heads) are “fitter” (more likely to breed) than ones with less of it.  So we can picture this “fitness landscape” as a histogram, in which there are a great many short bars, representing runs with smallish, products-of-runs-of-heads; a few very short bars, representing runs with extremely small products-of-runs-of-heads, and a range of taller bars, with the tallest bar representing the run with the maximum possible product-of-runs-of-heads.  However, that is not all she has to do.  So far, the fitness landscape has no specified structure. The bars are all jumbled up, with high ones next to low ones, next to medium sized ones.

This is what the fitness “landscape” would look like if the randomly mutated offspring of each of successful run had a product-of-runs-of-heads that was unlikely to resemble that of the parent run. The fitness landscape is “rugged”, and R will remain improbable

Note that in this example, we have both a genotype – which is the run of coin-tosses itself – and a phenotype – which is the product-of-runs-of-heads.  The fitness criteria only applies to the phenotype, and it turns out that in this system that quite similar-looking genotypes can have very different products-of-runs-of-heads, resulting in a very rugged fitness landscape.

This means that original Searcher, Lizzie, the Intelligent Designer, needs not only to Design a fitness function (a system in which the closer a phenotype is to R, the more likely it is to reproduce), i.e. the fitness histogram, but also something that will arrange the bars of the fitness histogram in such a way that the population of runs-of-coin-tosses is can “move” from the lower bars to the higher – make it into a smoother, less rugged, “landscape”. To do this, she must ensure that the ways in which offspring can differ genetically from their parents includes ways in which they can inherit not just the phenotype but the genotype.

And it turns out that point-mutations are not very good at doing this.  So, being an Intelligent Designer, Lizzie thinks again, and adds some a different kind of mutations – she includes adding an extra coin-toss to random positions in the run, and then trimming the end to keep it the same length.  Now, it turns out, offspring are much more likely to resemble their parents phenotypically as well as genotypically, and the fitness landscape histogram has arranged itself so that similar height bars tend to be adjacent to each other, and the “landscape” is quite smooth.

However, there are still deep valleys between peaks, and populations tend to get “stuck” on these local maxima – they find themselves on a high-ish histogram bar, but the only route to a yet higher bar is across a valley.  In other words, a given genotype may be quite fit, but the only way its descendents can ultimately be fitter is if some of them are less fit, and with the culling of the unfit being fairly ruthless, this is a low probability event.  In practice, in this example, this is because runs with lots of three-head and four-head segments are quite fit, but to convert a genotype with lots of threes and fours into much fitter one with mostly fours, by point mutation or insertion only, too often involves first breaking up some of the fours into a one and a three, which lowers the fitness.

So she thinks yet again, and now she includes snipping out pairs of segments and swapping their positions; and duplicating segments, where a segment is repeated, over-writing another segment, and deleting a segment entirely and replacing it with random heads or tails.

And lo and behold, this new system system tends to produce R much more readily (with higher probability), and not only that, the very highest possible peak is reliably achieved. This set of mutational methods has resulted n a fitness landscape in which there are at least some sets of bars in the histogram that form a series of steadily ascending steps, from the low bars to the very tallest bar of all.  The peaks are high, the landscape is smooth, and the valleys are shallow.

So the take home message for me was: my successful fitness landscape, consisted of three Designed elements –

  1. the fitness criterion, by which fitness is defined, which is the same as defining R;
  2. the relationship of genotype to phenotype, which ensures that fit parents tend to have fit offspring, and makes the landscape smooth,
  3. the variance generation mechanisms are such that the valleys are shallow.

Now, Dembski and Marks would say, presumably, that in my final set up, with several variance-generating mechanism, and which reliably produced R, i.e. making R a high probability result, itself contains at least as much information as that represented by R patterns when R could only be generated by old-fashioned coin-tossing runs.

And we know that that Active Information came from Lizzie, an Intelligent Designer.  I was the original Searcher, possessing Information as to how to find R, and I transferred that Information into my fitness landscape, which in turn became the Searcher, and which reliabley led to R.

But is an Intelligent Designer the only possible source of such information?

Let’s imagine that some future OOL researcher, let’s call her Tokstad, discovers a chemical reaction, involving molecules known to be around in early earth, and conditions also likely to be present in early earth, that results in a double chain polymer of some sort, that tends to split into two single chains under certain cyclical temperature conditions, whereupon those single chain atracts with monomers in the soup to become double chains again, but now with two identical double chain polymers where before there were one.  And let’s say, moreover (as we have some clues here), that this soup also contains lipids that form into vesicles that tend to expand, become unstable, then divide, and which, moreover, it being soupy and all, enclose some of these self-replicating polymers.  Let’s further suppose that the polymers don’t replicate with absolute fidelity – bits get added, chopped off, shorter chains sometimes join up to form longer chains, etc, and finally, let’s suppose that certain properties of some of these varied chains (length, constituent monomers) affect osmotic pressure differences between the vesicle and the soup, and/or the permeability of the vesicle to monomers in the ambient soup, affecting the vesicle’s chances of dividing into two, and of its enclosed polymers self replicating successfully.

It’s quite a big suppose, and possibly impossible, but not beyond the bounds of chemically plausible science fiction.

But here is the point:  IF such a system emerged from a primordial soup, it would be a system in which:

  1. There was a fitness function (some polymer containing vesicle variants replicate more successfully than others).
  2. There is a link between genotype and phenotype (similar polymers have similar effects on the properties of the vesicle)
  3. There are several different ways in which genetic variance can arise (duplicating, adding, deleting, replacing)

In other words, we would have, potentially, a system in which is located, according to Dembski, Active Information in the form of a smooth fitness landcape, shallow valleys, and high fitness peaks representing vesicles with high I (unlikely to emerge spontaneously were the chemistry to be something that did not provide these parameters).

So we have a Informed Searcher and an R, but no Lizzie, just the fitness-landscape itself, spontaneously arising from primordial chemistry.  So how do we measure how much Information that Search contains?  Well, that depends on how improbable the conditions that generated the components of the fitness landscape: the polymers, the vesicles, cyclical temperature changes, the chemical properties of the atoms that make up the molecules themselves, actually are.  In other words: in how many, of all possible worlds might such conditions exist?

AND WE HAVE NO WAY TO CALCULATE THIS.

We do not know whether they are the result of an extraordinarily fluke (or Intelligent Design)  by which, out of all possible universes, one in which this could happen was the one that eventuated, or whether this is the only possible universe, or whether an infinite number of universes eventuated, of which only those that have properties that give rise to polymer chemistsry result in intelligent life capable of asking how intelligent life itself originated.

But that’s not an argument for ID from probability and statistics, it’s an argument for ID from metaphysics.

Alternatively, if Dembsk and Marks are relying on Tokstad not discovering conditions in from which can emerge fitness landscapes in which ever-fitter self-replicators are the result, then they have backed a perfectly falsifiable horse.

But the important point is that observing effective fitness landscapes in the natural world does not, and cannot, tell us that there is an external source of Information that must have been transferred into the natural world. All it tells us is that the world that we observe has structure. There is no way of knowing whether this structure is probable or massively improbable, and therefore no way, by Dembski’s definition, of knowing whether it contains Information.  It seems to me it does, but that’s because I don’t define Information as something possessed by an event with low probability, and therefore don’t attempt to infer an Intelligent Designer from data I don’t, and can’t, possess.

 

153 thoughts on “Searching for a search

  1. Elizabeth said: “You are seeing stall, denial and dismissal tactics where they do not exist, William.

    Or, I’m seeing them where those that employ them do not themselves understand that this is what they are doing.

    1. Anyone that claims that that it biological features do not appear to be designed is denying the trivially obvious. The overwhelming appearance of design – even to materialist evolutionary biologists – is prima facie evidence of design

    2. Anyone that claims that there is no reason to consider that chance and necessity might be categorically insufficient to produce novel, complex biological function is denying the trivially obvious – see #1.

    3. In the face of such obvious prima facie evidence to consider design and doubt the capacity of chance and necessity to produce such features, we are left with an ethical responsibility not to claim such categorical (chance and necessity) forces sufficient, and not to claim ID unnecessary, until, and only if, a metric of some sort can establish that position as reasonably supported.

    4. Anti-ID advocates claim there is no such metric, therefore it is their ethical duty to avoid making claims that ID is unnecessary, otherwise they are engaged in ideologically-driven denial and dismissal tactics.

    5. If I remember correctly, you, Elizabeth, and others, have claimed that there is no reason to consider ID as a potentially necessary causal category and that known evolutionary processes are sufficient, so there is no need to provide a discerning metric to substantiate the claim that non-intelligent causal forces provide adequate explanation. To begin with, describing a physical process does not vet either the existence of the process itself or its product as explicable through chance and necessity. So, as per #1, there is a reason to consider ID as potentially necessary, and as per #2, you (and others) are guilty – either knowingly or unconsciously – of engaging in denial and dismissal tactics against the trivially obvious.

  2. William J. Murray:
    Elizabeth said: “You are seeing stall, denial and dismissal tactics where they do not exist, William.

    Or, I’m seeing them where those that employ them do not themselves understand that this is what they are doing.

    That is always a possibility, William, but to come to that conclusion too quickly, without considering deeply the possibility that our own perception of the other person’s position may be in error, is dangerous, because it can lead to a closed loop in which we cannot hear any argument that potentially invalidates our own position.

    There are two issues now, where you seem to have profoundly misunderstood my position. You seem to think:

    1. I don’t think it is ever possible to infer a designer from a designed product alone
    2. That I think that multiverses may be some sort of solution to a probability problem.

    Neither of these are views I hold. Yet not only do you continue to argue against these positions, you insist that I hold them without apparently knowing I hold them!

    How is communication even possible under these conditions?

    1. Anyone that claims that that it biological features do not appear to be designed is denying the trivially obvious.The overwhelming appearance of design – even to materialist evolutionary biologists – is prima facie evidence of design

    To which I answer: biological organisms share certain features with complex items known to have been designed by humans. Let’s call that set of features F. It is therefore possible that all things that display F have a similar cause. We know that biological organisms, although they display F, cannot all have been designed by humans, because most predate humans. Therefore we posit a non-human intelligence, or intelligence-like process as the origin of F in biological organisms. And we ask ourselves: what non-human intelligence, or intelligence-like processes do we know of that could produce F? One candidate is evolutionary processes, which we know, from human-designed evolutionary processes, can result in F. And we also know that evolutionary processes require that the F bearing item is self-replicating, with variance. And we also note that biological organisms are self-replicating, with variance.

    Therefore we have a candidate cause of F in biological organisms.

    2. Anyone that claims that there is no reason to consider that chance and necessity might be categorically insufficient to produce novel, complex biological function is denying the trivially obvious – see #1.

    Your question presuppose that “chance, necessity and design” are distinguishable causal agents. I think this is fallacious – I think the presupposition involves a category error. I’d say that “Necessity” is simply the Laws of the universe as we observe it, that “Chance” are those events for which we cannot track the probably chain of necessity, and “Design” is a proximal cause, for which the Laws of Nature are a distal cause.

    Clearly your view is different, but given my conceptualisation of Chance, Necessity and Design, it is not “trivially obvious” that we must infer one if we exclude the others. In fact, it makes no sense, because they are not, in that conceptualisation mutually exclusive categories. “Chance” merely means “unpredictable” and “Design” is a system of causality that involves, inter alia, iterative selection and rejection.

    3. In the face of such obvious prima facie evidence to consider design and doubt the capacity of chance and necessity to produce such features,we are left with an ethical responsibility not to claim such categorical (chance and necessity) forces sufficient, and not to claim ID unnecessary, until, and only if, a metric of some sort can establish that position as reasonably supported.

    See above.

    4. Anti-ID advocates claim there is no such metric, therefore it is their ethical duty to avoid making claims that ID is unnecessary, otherwise they are engaged in ideologically-driven denial and dismissal tactics.

    Well, I don’t know whether you consider me “Anti-ID” but I don’t claim there is no such metric.

    5. If I remember correctly, you, Elizabeth, and others, have claimed that there is no reason to consider ID as a potentially necessary causal category and that known evolutionary processes are sufficient, so there is no need to provide a discerning metric to substantiate the claim that non-intelligent causal forces provide adequate explanation.

    I think that “Design” and “Necessity” are not mutually exclusive causes, but represent different levels of explanation, the first being more proximal, the second more distal. I think that “Chance” is yet another category, and not “causal” at all – I don’t think “Chance” causes things. I think some Necessitous processes are intrinsically stochastic, and that we also use the word “chance” to describe causal events for which we have no predictive model. In data analysis, it’s what we put in the error term. However, we can reduce that error term by bringing in more known causal factors; therefore chance itself is not a causal factor, it is the term we employ to model unpredicted causes.

    To begin with, describing a physical process does not vet either the existence of the process itself or its product as explicable through chance and necessity. So, as per #1, there is a reason to consider ID as potentially necessary, and as per #2, you (and others) are guilty – either knowingly or unconsciously – of engaging in denial and dismissal tactics against the trivially obvious.

    I suggest you read my post above carefully. As I’ve said before, you may still disagree with it, and I may still be wrong, but I hope it will help you understand that my position is perfectly coherent, and that I (and others), are not denying and dismissing the trivially obvious but rather have a different, and perfectly coherent conceptualisation of the issues, under which your conclusion is far from obvious.

    William, if, coming back from the pub one night, I am concussed by a robber with a blunt instrument, it is by Design (someone meant to rob me); it is also by Chance (I tossed a coin to decide whether to go to the pub or not); and also Necessity (a blunt instrument descending with force on my head will cause my brain to compress itself against my skull.

    They are all perfectly valid causal explanations. Explanations don’t filter in the way you seem to think (even Dembski has abandoned the EF).

  3. Lizzie:
    Your question presuppose that “chance, necessity and design” are distinguishable causal agents.I think this is fallacious – I think the presupposition involves a category error.I’d say that “Necessity” is simply the Laws of the universe as we observe it, that “Chance” are those events for which we cannot track the probably chain of necessity, and “Design” is a proximal cause, for which the Laws of Nature are a distal cause.

    Clearly your view is different, but given my conceptualisation of Chance, Necessity and Design, it is not “trivially obvious” that we must infer one if we exclude the others.In fact, it makes no sense, because they are not, in that conceptualisation mutually exclusive categories.“Chance” merely means “unpredictable” and “Design” is a system of causality that involves, inter alia, iterative selection and rejection.
    . . .
    I think that “Design” and “Necessity” are not mutually exclusive causes, but represent different levels of explanation, the first being more proximal, the second more distal.I think that “Chance” is yet another category, and not “causal” at all – I don’t think “Chance” causes things.I think some Necessitous processes are intrinsically stochastic, and that we also use the word “chance” to describe causal events for which we have no predictive model. In data analysis, it’s what we put in the error term.However, we can reduce that error term by bringing in more known causal factors; therefore chance itself is not a causal factor, it is the term we employ to model unpredicted causes.
    . . . .

    Very nicely put. This topic is deserving of a top level post, in my opinion. The phrase “chance, necessity, and design” has long struck me as a shibboleth at UD, identifying some shared assumptions that permeate the discussion. Addressing this directly could help expose some of the core differences between IDC proponents and opponents.

    Whether or not that leads to greater understanding and better communication remains to be seen.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.