A curious couple of days for me at Uncommon Descent!
I’ve always wondered why people who are skeptical about Darwinian evolution aren’t persuaded by the power of evolutionary algorithms to find creative solutions.
There seems to me to be a deep misunderstanding of a) the nature of search space (and its structure) and b) the role a Designer plays in a GA.
The oddest objection I find is the objection that the Designer designs the fitness function. If the ID of ID is the fitness function, then ID is the environment! Because that’s the analog surely.
Which is nicely Earth Motherish, I guess.
But almost equally odd is the apparent objection that because the Darwinian algorithm is programmed into the computer, that makes it designed.
A bit bemused.
But this notion of the structure of the search space seems to me to be important. I’m trying to think of a search space that wouldn’t be easily traversed by a Darwinian algorithm. I guess one that consisted entirely of irreducibly complex features, where nothing worked at all unless it was complete.
OK, that helps. But in that case the whole needle-in-a-haystack argument depends on most things being IC. Or at least crucial things being IC. Including, of course, self-replication itself. Or at least Darwinian-capable self-replication.
While mutation would be random, presumably you would make reference to a library of words and phrases for selection. (Otherwise how would it recognize misspelled words?) A computer could do thousands of such comparisons each second. Alternatively, you could use an audience, but judging millions of possible phrases could take centuries, and rather bore the critics. Is that how you envision making such an evolutionary algorithm? Or did you have some other notion?
I meant a naturally occurring landscape 🙂
What do you mean “new solutions”?
Actual experiments reveal that a significant percentage of random sequences code for functional proteins and can replace proteins in actual living organisms. I posted a link to this back on Mark’s blog, in response to gpuccio. That was a day or two before he stopped posting.
Once you have a minimally effective sequence, variation and selection can operate in two ways. the sequence can change in ways that make it more useful, or it can branch. Branching is what we hypothesize is responsible for the common descent of diverse species.
This functionality from random sequences speaks directly to the issue of how rich and possibly connected the “search space” is. GAs demonstrate convincingly that if variations can be connected, they will be.
But it is useless to do what Douglass Axe did. He did not try to illuminate the evolutionary history of proteins. He demonstrated a rather narrow case of Dollo’s law, which states it is improbable to evolve backwards to a branching point and then forward to the alternate branch.
In fact, any change in any direction, specified in advance, is improbable. The problem with specification as a concept is that it is retrospective.
The problem with thought experiments, of course, is that they are generally more difficult to implement in practice than in theory. Yes, a real “language GA” would likely require some vast library of all possible (meaningful) sentences up to the maximum length under consideration (this takes into account the words). A vast crowd of bored humans would fit the bill, certainly. The fitness function would allow a good deal of wiggle-room in locating fits; with misspellings, repeated words etc. allowed up to some reasonable limit that retains meaning.
I think such a set-up would let you play a bit with the problem; how far you could evolve a given phrase, any limitations etc. Sufficient runs would let you answer such questions as “is the space substantially connected, given the allowed range of mutations”.
An interesting example of a mutated sentence, the mutations caused by defective blog software.
“All Sci.â€Ânce so f€År! ”
It’s a joke. But reading these archived blogs is possible, even when they aren’t jokes.
And the game of telephone illustrates that an iterative selection function can turn a garbled message into a different message that is still syntactically correct.
So, we could use a dictionary of words and phrases, or perhaps an algorithm to decide whether a phrase makes some sort of sense. How much sense would it have to make? How flexible?
But if you are only worried about connections within the space, then we don’t have to use random mutation. We can actually inspect the space ourselves to see what point mutation and recombination can do. Like this:
o, or, ore, one, wore, word, words, wordy, ward, war, tar, wars, ware.
war, wan, man, may, mean, many, bean, bear, beer, bee
be.
be-ware.
I actually wrote one a while back.
The way I set it up was that any letter combinations that were pronouncable (from a look up table) were advantageous, any combinations that were real words ditto, as well as combinations of word types (article-noun; adjective noun; article adjective; noun-verb; adjective-verb).
It produced a few funnies 🙂
I’ve seen someone at AtBC mutate a sentence into a different sentence with each step grammatically correct and all the words spelled correctly.
Relaxing the rules to allow misspelled but recognizable words,and relaxing the grammar to allow incorrect but meaningful syntax would make the problem trivial.
Except for writing the oracle. Decades after Chomsky, we don’t have a universal parser of human language. And individuals differ in their ability to see meaning in garbled messages.
I once thought about building a web site in which would display a population of random strings and asking visitors to vote on the one most like a string of intelligible language. Those receiving the most votes would populate the next generation, with mutations.
No one would have knowledge of how other visitors are voting or why.
Sort of like female choice in Darwinian selection.
I would place a small bet that the strings would quickly converge on meaningful statements.
We were working on an algorithm that used a primitive grammar parser, along with selection for meter and other poetic features. Here’s a few interesting results:
“a rainy land”
“for our bad pet”
“any fast zany cab”
“pain of a rat, pray!”
“a beast may warn”
“what hot red breast”
“our fat years best go”
“any old good best do”
“a man prays nor eats”
“a fast arson might sear”
“a cold God soon berates”
“what hot and fat beast bears far”
“what hot and fat breast bears far”
The project was never completed because we couldn’t get the other side to agree to clear criteria for success.
Those are a lot better than mine!
Although I did get one hilarious output from another program that started from an existing sentence:
The quick brown fox jumps over the lazy dogma.
Could you explain why you think the bolded part is true?
Thanks 🙂
Yes, small words evolve quite nicely, but once you increase the length of the sentence it becomes much more difficult to maintain a coherent thought or instruction. If the difficulty for the protein evolution case increases only linearly (say) with length, then the evolution of proteins is not so difficult. If the increase is exponential, then that’s a problem.
Great, so you accept the idea, in principle, of a library of words and phrases. Do you think we can approach the Universal Probability Bound by such a method?
Could you please provide the link? I’d like to see that.
Well, we have to have some means of determining fitness, so a library is a good starting point, but you could have perfect spelling and grammar and the phrase will still be nonsense, so how do you judge meaning? The UPB question is way beyond me at this stage… I’m just trying to come up with a GA which might be a good model for real evolution at this stage. It’s really hard to define what you should call “fitness” in this context. Is any meaning good enough?
Well, I guess I’m part of the “other” side, and I’m not sure what that criteria would be, either, to be a reasonable representation of “reality”. But still, whatever the criteria we could find out the general properties for that case, then vary the demands to see how the model responds.
I suppose a novel protein would only undergo positive selection if it were able to accomplish some new function, or improve an existing one. Perhaps to modify an existing system would be simpler in principle than creating an entirely new system. I don’t think it is sufficient just for the sentence to have meaning, any more than it would be sufficient that any random protein has its own active regions, but how much meaning? What kind of meaning?
But here is a suggestion, which I hope would be an attempt to model gene duplication: Start with a single, coherent sentence that contains one idea, then duplicate it, and evolve the duplicated sentence until it successfully modifies, or adds to the meaning of the first. Is that fair and reasonable?
Dr. Sean Pitman.
At this point, we are examining your wordscape analogy, not a model of biological evolution. But let’s be sure of this point:
In other words, there are no connections—whether through random or intelligent mutation and recombination—from short words and phrases to more complex structures, such pathways don’t exist. Are you sure?
Moderation queue, please.
SCheesman:
There are some problems with those claims.
1. Evolutionary computation hasn’t been proffered here by anyone as “proof of macro-evolution”.
2a. Evolutionary computation needs nothing more than rankings of candidates. Rankings are not continuous.
2b. My example of approximate solutions to TSP problems found by evolutionary computation also is at odds with the assertion that only “continuous variables with locally differentiable surfaces” can be approached. As for coded systems, the Avida artificial life system shows that those are amenable to evolutionary computation as well.
3. The characterization of evolutionary change appeals to facts not in evidence.
Douglas Axe hasn’t come anywhere near establishing the “islands of functionality” claim. (Neither has anyone else, for that matter.) Figure out how many different functional proteins are known (call it N), how many functions Axe actually tested for (call it K), and I think that the following inequality holds readily:
K < 10 << N
(It's been a while since I read Axe's papers, but IIRC his K=1. )
Until K gets to be some appreciable fraction of N, the "islands of functionality" claim is simply unsupported. And other research doesn’t make it look promising for the “islands of functionality” conjecture.
PS: If you must drag an honorific into the discussion, the one that applies to me is “Dr.”. “Mr. Elsberry” refers to no one involved in this discussion.
Earlier you suggested:
Which is not the same request.
And, by the way, this is a sentence that contains a point mutation (“speach” in lieu of “speech”) – yet conveyed the meaning you intended perfectly well. One mutation further and it conveys the coherent, but quite different (and whimsical) notion that survival be related to the flavor of the text (recognizable as English peach).
Hello Reciprocating
Yes, I may develop my ideas as I go along. I hope you don’t take that as a fault. And yes, I have been quite happy all along to allow minor “mutations” in spelling, as long as the meaning remains. You illustrate both “neutral” and “non-neutral” examples with “speech”.
SCheesman,
I hope you begin developing your ideas (as in responding to your responders) rather than just repeating them in slightly different words.
The landscape issue is rather important. Perhaps you think this issue is ignored by the mainstream. But you would be wrong to think so.
If you click on my screen name you can see a direct modelling of the landscape of words in several languages. The connectedness varies from language to language.
As near as possible, I’ve got every (based on Scrabble dictionaries) word in several languages. Selection is not based on a target of any particular word, but on nearness to any word.
To do your simulation, one would have to have a dictionary of every possible sentence. Rather unlikely. Alternatively, one would have to have a program that could determine the likelihood that a human could make sense out of a syntactically imperfect string. Also not likely.
But chemistry does have a perfect unabridged dictionary. Proteins just fold, and we don’t know how to simulate that efficiently. So we have one natural example of an oracle that we can’t model. We know protein folding can be solved quickly; we just can’t match the quickness.
As others have pointed out, changes in living things have effects in hundreds or thousands of dimensions. We can effectively model a few. But nature is the perfect oracle. Selection, like protein folding, happens across all dimensions without computation.
Science simplifies in order to quantify and illustrate principles. One can, as Newton did, understand planetary orbits from quantifying the behavior of falling objects on earth.
Similarly the propagation of alleles in a population can be modeled, even though we can’t simulate all the dimensions of life.
Seems to me that my example meets the requirements of your first suggested model: with just two mutations (one a random error) I generated recognizable speech and a coherent idea (wouldn’t the second of these requirements entail the first, BTW?).
(English Peach is a rather nice meme. I believe it refers to Lizzie.)
Honestly, I don’t see the value of this exercise. What is the value in honing landscapes that GAs cannot traverse, absent independent reasons to conclude that those landscapes model those across which biological organisms are arrayed with some fidelity? Seems to me that your entire thesis still turns on that independent evidence.
I think it is rather easy to imagine or invent or model landscapes that are not analogous to biological landscapes.
Regardless of Dembski’s “laws” GAs can traverse many landscapes, including many useful landscapes. The travelling salesman GA has wide use in real corporations.
GAs work. Philosophical arguments are irrelevant.
The only way to challenge RM_NS is to demonstrate in biology itself that the landscape it too sparse. Douglas Axe get a point for recognizing the problem faced by ID, but he will never follow up. because the balance of research does not support his thesis.
SCheesman: Well, I guess I’m part of the “other” side,
Dr. Sean Pitman.
SCheesman: and I’m not sure what that criteria would be, either, to be a reasonable representation of “reality”.
At this point, we are examining your wordscape analogy, not a model of biological evolution. But let’s be sure of this point:
Zachriel: But if you are only worried about connections within the space, then we don’t have to use random mutation. We can actually inspect the space ourselves to see what point mutation and recombination can do. Like this:
o, or, ore, one, wore, word, words, wordy, ward, war, tar, wars, ware.
war, wan, man, may, mean, many, bean, bear, beer, bee
be.
be-ware.
In other words, there are no connections—whether through random or intelligent mutation and recombination—from short words and phrases to more complex structures, such pathways don’t exist. Are you sure?
SCheesman,
I added a reply that appears further below. I hadn’t taken in the import of the “Quote” and “Reply” links before.
One other reason probability calculations by ID advocates do not impress is that they apparently do not realize how important conditional probability is. In a GA, the probability of getting a particular string of 1’s and 0’s strongly depends on whether a similar string exists in the previous generation.
SCheesman,
OK, not random. But not pre-selected or designed for function.
My point remains that Axe has not considered or explored the possibility the functional sequences are not particularly sparse.
http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0015364
You don’t make progress by rigging an experiment to give negative results. There are more ways to get negative results than positive results.
Demonstrated that billion year old proteins are isolated says nothing about their origin, possibly as minimally functional proteins.
My apologies, Dr. Elsberry.
I was not aware that I had implied such. Of course any computational experiment can only provide evidence that is illustratory.
My point about the differentiable surface was directed at GAs such as the one provided by Elizabeth Liddle for solving antennae problems. In other places in this thread I have noted that GAs have various other strategies for navigating the solution space. These, however are generally not nearly as efficient, and frequently reduce to little more than blind searches or programmed sweeps of the local parameter space, hence the need for “rankings”.
Well, we could reference papers back and forth. Here’s Axe’s latest reply.
But I’m really not competent myself to comment knowledgably on the current state of such research, so I will grant the possibility, at least in principle, that the search space of functional folded proteins could have been traversed by relatively simple (and mathematically probable) mutational steps in sufficient time to produce those we see today. My sense is that, as the length of proteins increases, the degree of connectedness decreases, and quite rapidly, too.
Yes, I had the same problem… see my reply at your original post above.
Petrushka. The paper you referenced notes the following:
Note the “design”. What results do you think they would obtain if the library was composed of a completely random assortment of 102-residue sequences. I’ll tell you… there wouldn’t be a paper.
In other words, you can get somewhere faster and more easily if you are close to it. Well, that had occured to me.
I am heartened to hear so.
Well, when I can. I have, as I expect most of you do, a day job!
SCheesman,
Nice try, but the point of the paper is not that the experimenters had a million years or so to wait for the right sequences to appear, but that functional sequence space is not pointy. The fact is that a bunch of sequences that were not known to have a function actually have a function and can replace genes in living things.
Sequences that have no “letters” in common with naturally occurring sequences.
The issue here is not OOL, but whether functional space is or is not pointy and whether function can arise from many starting points not explored by life.
Don’t you find this approach more informative than asking whether a squirrel can evolve into a cockroach, or whatever it is that Douglas Axe asked?
It is, after all, the origin of protein families that is interesting, whether there are many or few possible starting points.
SCheesman,
But the issue of random libraries is not being ignored:
http://molbio.mgh.harvard.edu/szostakweb/publications/Szostak_pdfs/Keefe_Szostak_Nature_01.pdf
Whether there is one point or many points is not the problem; it is the total “cross-sectional area” of the points in relation to the full parameter space. If I have needles spaced every inch, its still not a bed I’d want to lie on. We already covered this ground when I allowed that matching sentences could contain synonyms and misspellings as long as the original meaning gets through.
SCheesman,
Selection is based entirely on this fact, so probability calculations without this are worthless. Suppose one wishes to compute the likelihood of Lenski getting all each variant in his E. Coli experiment. The natural starting point would be Lenski’s starting points, not a random genome. There is complete genome information available – how would any of the ID proposed methods (EF, CSI, Specified complexity) and ancillary claims about the structure of the search space be tested against this reality? There are two issues 1) if it is doable, 2) if ID advocates have the ability to do it if it is doable.
Yes, Wesley Elsberry referenced essentially the same 2001 paper earlier. A total of 4 proteins were found that would bind to ATP out of 6×10^12. And that’s just binding, to say nothing of any actual function.
Here’s a followup…
What fraction do you think might actually not be dangerous? Sounds more needle-like to me.
Well, Petrushka, I’m sure you’ve got a come-back, but I think this thread is pretty-well played out.
Maybe we can resume on something else in another one later. You can have the last word…
http://www.plosone.org/article…..ne.0007385
t_p_hamilton:
Well, I expect any observed variation will fall within the bounds of probability ID proposes.
That’s it for me on this thread… time to give it a rest.
Obviously my reference got out of order above… at least it showed up at the bottom.
I agree that it is quite likely, given that it happened. The question is whether this probability can be estimated by any ID advocate, given complete information. If it can’t be done, then why should we give credence to ID claims of probability for unknown pathways?
I will tell you what the value would be. IF you could actually predict that probability of developing antibiotic resistance for just one drug was below the UPB, and be right, you could probably get the Nobel prize in Medicine. Just think of what a triumph that would be. Too bad experience seems to indicate this in itself is unlikely, and this in turn indicates something about the sparseness of searchable protein space.
Why? Do you consider that to be a big number?
There are more bacteria than that in a shovel full of topsoil. And consider that they have several generations a day. More in warm weather. And thousands of years available between novel proteins. Take a look at how this numbers game played out with Behe at the Dover trial. The full transcript is available.
You are welcome to drop out at any time. It’s usually when actual research on functional space starts showing up that ID proponents drop out.
Or at UD, the evilutionists get banned or called liars.
You really need to chill…
SCheesman:
Wasn’t that what this quote from you meant?
SCheesman:
Is that supposed to be a distinction that makes a difference?
You brought up GAs working on continuous inputs and locally differentiable surfaces as a reason to exclude them from lending any insight into biological evolution, and contrasting your (erroneous) view of what that implied by way of contrast. Now, given that your original assertion was demonstrated to be so much hogwash, you dismiss the assertion — but retain the conclusion that you originally had that supposedly was due to it. Your further commentary on the topic of evolutionary computation is, sadly, yet more hogwash. Rankings are not specialized things that establish “need” in any sense of fragility, as your rhetoric clearly poses them. The fact that evolutionary computation backends can operate on rankings alone demonstrates just how little information is needed to succeed with evolutionary computation, and that no “continuous inputs” or “locally differentiable surfaces” are necessary parts of it, contrary to your prior assertions. Those “other places” in the discussion don’t look to me to reveal any better comprehension of evolutionary computation on your part than in what I’ve quoted. Your claims about efficiency are false and are founded upon misconceptions of process in EC. Why do you bother to assert nonsense about a topic? Why not simply say you don’t understand the argument and thus find it unconvincing?
Oh. Maybe you get your information from what IDC advocates claim about evolutionary computation. If one credulously believed the codswallop they peddle, one could easily become deeply confused about the topic while getting a false sense that one has acquired knowledge instead of codified ignorance.
Axe doesn’t appear to address the K < 10 << N disparity in the linked reply. Am I surprised? No.
Why, are you going to put me on moderation?
And you really need to give some thought to the structure of your argument.. As Elizabeth has argued at UD, we can easily demonstrate that Darwinian processes can find solutions that can be connected by small variations (which in biology include horizontal transfer), so the only valid argument that ID can make is that there are, in fact, gaps that cannot be bridged by known mechanisms of variation.
That’s the same argument confronted by Darwin.
It’s been updated to include DNA, but it’s the same argument that was faced 160 years ago.
And the argument made by Axe is exactly equivalent to asking why squirrels don’t evolve into birds or lizards. He looks at the gaps between modern proteins and argues that it would be unlikely for one to evolve into another.
But of course, that scenario has nothing to do with evolution as understood by biologists.
SCheesman,
You may have missed our response above.
According to your previous comments, there are no connections within phrasespace—whether through random or intelligent mutation and recombination—from short words and phrases to more complex structures, such pathways don’t exist. Are you sure?
I’d really like a source for that Axe quote.
I have to modify what I said about Axe. He does start with a modern sequence, but then he disables it with random mutations and then estimate how many forward mutations are useful.
He is still not addressing the actual history. He’s jumping into a part of the sequence space not traversed by evolution.
The interesting thing is that safecrackers used to use Darwinian methods to find the combination to locks.
The problem with breaking modern encryption isn’t the search space; it’s the lack of cumulative feedback.