Uncommon Descent Contest Question 10: Provide the Code for Dawkins’ WEASEL Program

_{Denyse O'Leary

August 26, 2009

Darwinism}

Share: Facebook; Twitter; LinkedIn; Flipboard; Print; Email

Special invitation for Richard Dawkins – but any civil person is entitled to enter.

There’s been some discussion here and elsewhere whether the the recent IEEE article by Dembski and Marks correctly characterizes Richard Dawkins’ famous METHINKS IT IS LIKE A WEASEL program.

Does the program ratchet correct letters or does it let them vary? One is a partitioned or stair-step search, the other a more realistic evolutionary search. From The Blind Watchmaker, where Dawkins describes the program, its performance suggests that it could be either of these options (though he doesn’t say).

On the other hand, from a (video-run of the program , go to 6:15), it seems to be the latter.

It’s easy enough to settle this question: Make the code for the program public. Perhaps Richard Dawkins himself or his friends at RichardDawkins.net can finally provide this code (apparently a program written in BASIC).

The prize is a copy of either Stephen Meyer’s new Signature in the Cell or Richard Dawkins’ soon-to-be-out The Greatest Show on Earth.

Should the winner choose the latter, I will ask Dawkins’s publicist to mail the copy. Given that at his site, he calls himself “the most formidable intellect in public discourse,” I would assume that if he signs the copy, it will be worth millions.

But wait. Let’s see that code first.

Comments

kf, everybody here has understood the latching issue. The main point is that we all agree that there is no proof or evidence that explicit latching is involved. This was the original question of this thread and it has been solved a long time ago. The programm DOES let the characters vary, the subsequent filtering eliminates bad mutations most of the time. Your letter counting ("as a matter of fact that in 200+ cases of letters that if reversions from correct were frequent") is of course wrong because you should a) count only correct letters and b) take into acount that you only see every 10th generation so that wrong letters have a good chance to be restored anyway. How often do we have to go through this? The discussion has moved on. It has now been shown that this is by far not the only misrepresentation in the D+M paper. In fact they use a completely different algorithm while referencing Weasel. This is most easily demonstrated by checking Eq 22 which is incorrect for Weasel. Even Joseph has acknowledged this. And finally, you continue to evade my questions which in turn continues to amuse me. I will repeat them later for your convenience!Indium_{August 31, 2009
August
08
Aug
31
31
2009
04:26 AM
4
04
26
AM
PDT}

--kf sorry, I couldn't spot where in your variably emphasized and enumerated post you answered the question:
Which algorithm from Atoms suite corresponds to Weasel and which one corresponds tpo the Partitioned Search discussed by D+M?
And for your subtlety:
when mutation occurs, the same letter is a candidate, so 1 in 27 times, a mutation will go back to being itself.
Why? That's just a question of implementation. Doing it this way, a possibility of 4% of change for a letter would translate into a de-facto-mutation rate of 3.85% - but you can also achieve a mutation rate of 4% if you chose uniformly only from the other 26 symbols. For practical purposes, it's not much of a difference...DiEb_{August 31, 2009
August
08
Aug
31
31
2009
03:51 AM
3
03
51
AM
PDT}

kf@222: can you explain why you're reproducing that? I think almost everyone here has written an implementation of the weasel algorithm and has observed the fact that with many progeny to choose from in each generation, a letter that is already correct is not going to change in the next generation (unless p(mutation) is extremely high). So what are you revealing by listing a result that's already very well known? Is it that you're trying to prove that RD's original program would have behaved just as he claimed in the book without having to fix letters in place once they'd matched, and therefore we don't need to see the source code? Thanks anyway, but I think we know that already.nephmon_{August 31, 2009
August
08
Aug
31
31
2009
02:55 AM
2
02
55
AM
PDT}

I've often wonder if the fact that the weasel program is performing directed a search at "a distant goal" is as troublesome as even RD states. If you think of the phrase "methinks it is a weasel" as representing the environment in which the mutated, generated phrases are "living", then what the program is doing is optimizing the progeny of the random seed phrase for fitness in that environment. Of course, there are many simplifications involved here: the "environment" happens to also be a specific value of the "genome" that's evolving; there is no analogue of a phenotype resulting from simple genotype that's being modeled (and on which selection would operate in reality); and the "environment" never varies. (This last point is easily addressed by also mutating the target phrase across generations, so that the program will now be chasing a moving target, which is closer to the idea of organisms adapting to slow-changing environments. Of course in this case, it's unlikely the simulation would halt, just like evolution.) There are many other enhancements that could be made to the program to better model real-world evolution, and I'm sure many such simulations exist. But I still think that as a very simple-to-understand illustration of the vast difference between random mutation and the selection of random mutations according to some selective pressure, it holds up pretty well.nephmon_{August 31, 2009
August
08
Aug
31
31
2009
02:45 AM
2
02
45
AM
PDT}

PPS: On a subtlety: when mutation occurs, the same letter is a candidate, so 1 in 27 times, a mutation will go back to being itself. As an effect of that, 1 in 27 times on average, when a correct letter "mutates," it goes back to being itself. That should not be overlooked in analysis.kairosfocus_{August 31, 2009
August
08
Aug
31
31
2009
01:59 AM
1
01
59
AM
PDT}

PPS: TH, you can see an implicitly latched run of Atom's adjustable Weasel hit home in 31 gens here. I reproduce, FYI, as it seems hitting on a link is not popular: ___________________ RUN B, 500 pop/gen, 4% per letter mut rate: 1. MEL LSI YHXMAJLMDGMVKTSKGW 2. MEL LSI YHXIAJLMDNMVKTSKGW 3. MEL LSI YHXISJLMDNMJKTSKGW 4. MEL LSI YHXISJLMDN JKTSKGW 5. MEL LNI YHXISJLDDN JKTSKGW 6. MEL LNI YHXISJLDDN JKTEKGW 7. MEL LNB BHXISJLDDN JKTEKGE 8. MEL LNB BHXISJLIDN JKTEKGE 9. MEL LNB BHXISJLIDN JKTEKSE 10. MEL LNB BHXISJLIDN JKTEKSEL 11. MEL LNK BHXISJLIDN JKTEKSEL 12. MEL LNK BHXIS LIDN JKTEKSEL 13. MET LNKV BHXIS LIDN JKTEKSEL 14. MET LNKV BHXIS LIDN AKTEKSEL 15. MET LNKV BHXIS LIDE AKFEKSEL 16. MET LNKV BHXIS LIKE AKFEKSEL 17. MET LNKS BHXIS LIKE AKFEKSEL 18. MET LNKS BH IS LIKE AKFEKSEL 19. MET LNKS BH IS LIKE AKFEKSEL 20. MET LNKS BH IS LIKE AKWEKSEL 21. MET INKS BH IS LIKE AKWEKSEL 22. MET INKS BH IS LIKE AKWEKSEL 23. MET INKS BH IS LIKE AKWEKSEL 24. MET INKS IH IS LIKE AKWEKSEL 25. MET INKS IH IS LIKE A WEKSEL 26. MET INKS IH IS LIKE A WEASEL 27. MET INKS IH IS LIKE A WEASEL 28. METHINKS IH IS LIKE A WEASEL 29. METHINKS IH IS LIKE A WEASEL 30. METHINKS IH IS LIKE A WEASEL 31. METHINKS IT IS LIKE A WEASEL _________________ And this one hit home in 21 gens: _______________ Case D: 999/gen [maxed out], 8%: ________________ 1. MMCJXLTPPCNATTMLKDXOBDKMBJQX 2. MMCJXL PPCNATT LKDXOBDKMAJQX 3. MMCJXL PPCNATT LKDXOB KMAJUX 4. MECJXL PPCLATT LIDXOB KMAJUX 5. MECJXL PPWPOVS LIDXOB WMAJUX 6. MECLXL PPWPOVS LIDXOY WMAJUL 7. MECLXL SPWPOVS LIDXOY WMAJUL 8. MECLXL SJWPOIS LITXOY WMAJUL 9. MECLXL SJWP IS LIZXOY WTASUL 10. MECLXL S WP IS LIZAOY WTASUL 11. MECLXL S IP IS LIZAOY WTASEL 12. MECLXL S IT IS LIZAOY WTASEL 13. MECLXL S IT IS LIKNOY WTASEL 14. MECLXL S IT IS LIKEOY WTASEL 15. MECHXL S IT IS LIKE Y WUASEL 16. METHXZ S IT IS LIKE Y WUASEL 17. METHXZ S IT IS LIKE A WUASEL 18. METHKN S IT IS LIKE A WUASEL 19. METHKN S IT IS LIKE A WEASEL 20. METHIN S IT IS LIKE A WEASEL 21. METHINKS IT IS LIKE A WEASEL __________________ Here we see speeding up of run to the target.kairosfocus_{August 31, 2009
August
08
Aug
31
31
2009
01:49 AM
1
01
49
AM
PDT}

Mrs O'Leary (and onlookers): Let us first draw attention to the issue for this thread: absent someone out there having a copy that is currently unknown, no copy of Weasel c 1986 code will be forthcoming. It therefore remains to correct a few mis-impressions and to underscore the real central issue and achievement of the recent IEEE paper (and its stable-mates). For, the clutch of papers introduce an extension to the achievements of design theory: active information as an explanatory factor for the superior performance of intelligent search over blind, random search in large configuration spaces. And, we may again underscore that Mr Dawkins, in 1986, inadvertently pointed the way to that conclusion when he developed Weasel as a targetted, proximity increment rewarding search for a preloaded target: ____________ >> [Weasel] . . . begins by choosing a random sequence of 28 letters ... it duplicates it repeatedly, but with a certain chance of random error – 'mutation' – in the copying. The computer examines the mutant nonsense phrases, the 'progeny' of the original phrase, and chooses the one which, however slightly, most resembles the target phrase, METHINKS IT IS LIKE A WEASEL . . . . What matters is the difference between the time taken by cumulative selection, and the time which the same computer, working flat out at the same rate, would take to reach the target phrase if it were forced to use the other procedure of single-step selection: about a million million million million million years. This is more than a million million million times as long as the universe has so far existed . . . . Although the monkey/Shakespeare model is useful for explaining the distinction between single-step selection and cumulative selection, it is misleading in important ways. One of these is that, in each generation of selective 'breeding', the mutant 'progeny' phrases were judged according to the criterion of resemblance to a distant ideal target, the phrase METHINKS IT IS LIKE A WEASEL. Life isn't like that. Evolution has no long-term goal . . . In real life, the criterion for selection is always short-term, either simple survival or, more generally, reproductive success. [TBW, Ch 3, as cited by Wikipedia, various emphases added] >> ______________ 1 --> Now, clearly, Dawkins accounts for the advantages of his "cumulative selection" that rewards mere increments in proximity to "a distant ideal target" by "nonsense phrases" over the "single step selection" -- i.e. random mutation that has to achieve significant function before it can hill-climb -- he would dismiss by the active information contained in the target, the comparison to target and reward of hotter as opposed to colder NON-FUNCTIONAL phrases. 2 --> So, the core point of the current series of papers was inadvertently acknowledged from 1986. By Mr Dawkins. 3 --> Q: So, why has there been a debate over a plainly secondary (and often simply contentious and distractive) issue -- latching and/or ratcheting -- since December last? __________ ANS: i] Of course, the answer has long been accessible one click away, from App 7, the always linked. ii] And, yes, there is a bit of Darwinist revisionism of history at work here, projectively shifting the locus of "obsession" to those who have pointed to the reason why explicit and implicit latching are valid interpretations of the description and showcased printoffs from Weasel c 1986. iii] In the main, this is because the issue was raised again and again and again by Darwinist objectors to those who pointed to the evident ratcheting action published by Dawkins in 1986; objectors who to often insist on mischaracterisations and resist the most reasonable and simple corrections. iv] Presumably, they raised it in the main because it is a point where they think they can discredit those who they object to, while focussing on what they want to highlight rather than the more embarrassing -- for them -- details of Mr Dawkins' statements as cited above. v] Unfortunately for them, they thereby show for the discerning reader several of the standard Darwinist rhetorical tactics of misdirection, mischaracterisation (of both arguments and persons) and dismissal. vi] For instance, they are unwilling to acknowledge that the plain fact is that in the published, showcased -- presumably "good" but representative -- print runs of 1986, the excerpted runs of champions show as a matter of fact that in 200+ cases of letters that if reversions from correct were frequent we would see a few reversions, we see NONE. (Cf. on the law of large numbers and observability of low-probability phenomena here in App 7. [Do we not find it interesting how adverse comments are routinely being advanced by the Darwinists without seeking to fairly represent those being criticised.]) vii] Equally, they do not wish to acknowledge that in his gushing comments on the prowess of Weasel's cumulative selection in BW, CRD used a term that normally means: "Increasing or enlarging by successive addition," which when multiplied by the highlighted runs strongly supports the conclusion that at minimum, "good" runs of weasel produced lines of generational champions that routinely behaved like the published excerpts. viii] So, CRD has the dilemma of either using misleading examples/excerpts, or using misleading words. (And that in a context of acknowledging that the is being in key parts misleading -- Weasel as described is intelligently designed, targetted search, not a good stand-in for chance variation and natural selection based on differential FUNCTIONALITY.) ix] Similarly, there has been a campaign to discredit the concept of ratcheting or of latching. In fact these are quite simple: based on what we see in the 1986 runs of champions, when Weasel is working "right" it will hold on to correct letters, and it will add more until the target is achieved, generation by generation. As in what seems to be clearly going on in the 1986 runs -- at least if we are unprepared to swallow the "don't believe yer lyin' eyes" rhetoric. x] So far, we have highlighted the pattern observable/ supported by the showcased runs c 1986 and the commentary on them. The question is where did these come from? xi] One obvious possibility is explicit latching, and this is a legitimate interpretation of the information c. 1986. However, on further commentary and claims c 2000 and currently from Dawkins and proxies, the other possibility seems to be more likely: IMPLICIT LATCHING; as is discussed in the App 7 and in earlier UD threads. (The mechanism that Darwinists seem to want to either dismiss as unreal or accept as trivial.) xii] But in fact, since CRD has in effect accepted that Weasel type algorithms as implemented in recent years are more or less typical of what his program did, we can rely on the demonstrated and long since publicly posted fact of implicit latching shown by Atom's adjustable weasel. (I'll bet some onlookers would not realise that such is demonstrated fact from the objections above and elsewhere!) xiii] Thus, we can see that indeed under certain conditions where the population size per generation, the mutation rate per letter and the filter work together right, the sequence of generation champions on closest yet approach to target will show latching and ratcheting. (How desperate are some objectors to dismiss the observed fact as non-existent or poorly defined!) xiv] How this likely happens: under not too hard to achieve conditions (enough pop and a reasonably low mut rate)the mutant population form a seed will almost certainly contain at least one unchanged member -- cf calculation and explanation in the App 7 on this. xv] That means that practically, progress to date on proximity will be preserved, given the stipulation on the filter. And, if in addition, single step advances are favoured otherwise because the mut rate is fairly low [so that 1 mutation per child phrase is typical], then mutations that affectt he correct letters will tend to be retreats, and will be eliminated by the filter. xvi] Most mutations of non-correct letters will make little or no difference to the proximity too (or if they hit a correct letter they will most likely revert it to incorrect for that member of the pop). But when a newly correct letter arises and the presently correct ones are undisturbed for a member of a generation, we have a one-letter distance increment in proximity. if such a member is present, it will be picked by the filter in the normal case. [Obserfve how int eh Weasel 1986 runs as showcased, we saw 40+ and 60+ gens to target for "good" runs. That is no-change won out about half the time, and single step increments in proximity dominated the rest.] xvii] Of course, where mutation rate or population size is high enough, we will see occasional cases where double mutations pop up which may give a two-letter advance, or may substitute one correct letter for another: the latter marking a LOSS of latching behaviour. xviii] On the other end, if the pop size per gen is too small, not enough members of population will be present to have a predictable pattern within the generation, so reversions may win, and latching vanishes again. (Hence the discussion on tuning or matching of pop size, mut rate and filter. The filter, of course, can also affect outcomes, as is mentioned in the App 7. As Apollos inadvertently demonstrated some months back, this can sometimes have a dramatic effect.) xix] If reversions are rare, we see near-latching or as I have termed it quasi-latching with an occasionally slipping ratcheting action. xx] Indeed, given that the above are probabilistic phenomena, do enough runs and you will see reversions in cases where latching "often" turns up. For a probabilistic barrier is not an absolute barrier. xxi] That brings up the 1987 videotaped runs. Most likely, they showcase a detuned run, which does not latch implicitly, and the onscreen reversions may reflect that. (It is also possible that the video is highlighting the members of the population, which are not the same as the run of generational champions published in the 1986 printoffs. The two are not mutually exclusive.) xxii] Beyond that, we may see behaviour that does not in any wise resemble latching, i.e cases with a rather high mutation rate will find it hard to close the deal, and maybe in some cases will never hit home to the target. _________ 4 --> Now, none of the above is really new, having been put up several times -- including in this thread. Not to mention, it is presented in the always linked app 7. Why then is it apparently so hard for Darwinists to see what should be obvious? 5 --> Partly, because it does not fit with what they expect to see. (It is hard to see what is really there when that does not fit with what you expect to see.) 6 --> For instance, just above we see TA @ 219: The point I would repeat is that latching and ratcheting are either undefined, arbitrarily defined, or so broadly defined (pick one), as to be useless. 7 --> Compare that ill-informed, counter-factual statement with the facts as again outlined. How could a reasonably informed person not see that here is a sufficiently precise definition, with publicly accessible demonstrations [starting with the showcased Weasel printooff excerpts of 1986]? And counter-examples that show contrasting cases?
ANS: because the actual facts cut across the expectations. [Starting with "don't believe yer lyin' eyes" on what Weasel 1986 credibly did.]
8 --> In that context, it is easy to see as well why a focus on an imaginary flaw can distract from the actual outright acknowledgments in BW of:
[1] use of artificial selection, [2] resulting material dis-analogy of this computer simulation to chance variation and natural selection, [3] the further resulting fundamentally "misleading" nature of the Weasel icon of evolution, [4] the actual reason for improvement in performance over blind search: active information in the form of a preset, embedded target and a hotter-colder beacon that draws in non-functional "nonsense phrases" until they hit home in a few dozen generations if the program is set up right. [5] the distraction from and dismissal of the need to de novo originate complex functional information through chance variation before hill-climbing/ warmer-colder algorithms etc can be legitimately employed.
9 --> And this last cluster extends directly to the more modern and more sophisticated Genetic Algorithms. GA's may reasonably mimic certain aspects of micro-evolution, but hey do not credibly account for the origin of complex information based functionality in the absence of intelligent action and the associated injection of active information. 10 --> And, that is a fatal defect. GEM of TKI PS: BillB of course -- as is sadly usual -- manges to excerpt out of context and mischaracterise. The context for my remark on halting was that because of halting, Weasel type algors will lock up as they hit home. that means that when we talk about probabilities of getting home AT gen x, we are talking about just that -- we would not get to to gen X if there had been an earlier hitting of the target.kairosfocus_{August 31, 2009
August
08
Aug
31
31
2009
01:41 AM
1
01
41
AM
PDT}

Well, here's one that I wrote up. Non-latching, highly configurable, and doesn't require the start string to be 28 characters (it can be any length). You can furthermore choose penalty weighting, etc. Should that work, I'd go with Dawkin's book. In any case, the program is really quite simple, and Panda's Thumb had a nice breakdown of the convergence times of latching and non-latching programs (though all with 28 character starting strings). Direct program link here.tigerhawkvok_{August 30, 2009
August
08
Aug
30
30
2009
11:40 PM
11
11
40
PM
PDT}

BillB writes> I think KF has even tried to imply that a blind random search has a latching mechanism here: AND KF writes> One of the peculiarities of Weasel algors, is that they halt when they hit home; of course halting being a key property of algorithms. I think BillB is correct: This seems to be saying that any algorithm that halts is latching, and that Weasel is peculiar for doing so? Therefore even a blind random walk that happens to find the target - and halts - is latching. I state as a given that halting is a universally accepted property of good coding practice. However, since good practice is not requirement, any search in an infinite loop might be considered non-latching. If halting is latching, I could be forced to admit this meets my request (in #146) for an example of a non-latching search algorithm. It would be a really stupid non-useful example though. Usefulness aside, by this definition I can now claim that all algorithms that halt, or might halt, exhibit latching. The very concept of latching and ratcheting is now even less meaningful than before. @KF: You wrote a lot, and I am willing to give you the benefit of the doubt and allow the quote above is perhaps taken out of context. I don't think it matters. The point I would repeat is that latching and ratcheting are either undefined, arbitrarily defined, or so broadly defined (pick one), as to be useless.Tomato Addict_{August 30, 2009
August
08
Aug
30
30
2009
01:11 PM
1
01
11
PM
PDT}

Cabal: in this case, "implicitly" simple means "having the appearance of". If you look at the output of an unlatched weasel program using a suitable offspring population size and mutation probability, you will see that once a letter has matched the target, it will never be corrupted again in the next generation, giving the appearance that it's latched. Of course, amongst the 200 (say) candidate offspring, many of them WILL have a corruption of already correct positions, but these ones never (or with a vanishingly small p, anyway) get selected to go forward to the next generation. If you make the offspring population small enough (or the chance of mutation high enough), then you WILL see letters flipping between correct and incorrect, and indeed it will never converge on the target string. I'm sure someone will say "Aha! So you have to TUNE the values to make it work!" to which I would reply, evolution is quite capable of tuning such meta properties of the process itself.nephmon_{August 30, 2009
August
08
Aug
30
30
2009
12:28 PM
12
12
28
PM
PDT}

kairosfocus:
In short, more word twisting and strawmanising.
No, it's a simple statement of uncontroversial fact. The EIL page presents math (the same math as the IEEE paper) that purportedly apply's to Dawkins' algorithm. The simple fact is that it does not apply to Dawkins' algorithm. This is made abundantly clear by the fact that the math does not take into account population size or mutation rate. But if you still don't believe it, run the numbers yourself, as I did above. There is no controversy here -- the numbers don't lie.R0b_{August 30, 2009
August
08
Aug
30
30
2009
10:59 AM
10
10
59
AM
PDT}

With computers being binary devices, how can 'implicit latching' be implemented? To me, implicit latching have a ring like 'implicitly pregnant'?Cabal_{August 30, 2009
August
08
Aug
30
30
2009
10:56 AM
10
10
56
AM
PDT}

To be honest, Cabal, the answers to these questions will probably no longer be of much interest now that kf has more or less admitted that D+M discuss something that is NOT Weasel. I bet he will still avoid them anyway, otherwise he will have to contradict himself even more. In fact, all questions have now more or less been answered: - the explicit/implicit latching obfuscation has been exposed (about 100 times now) - it is obvious that Dembski and Marks don´t use a proper Weasel and therefore misrepresent Dawkins The only one left on the sinking ship is kf, so who cares except for people interested in the argument regarding design?Indium_{August 30, 2009
August
08
Aug
30
30
2009
10:49 AM
10
10
49
AM
PDT}

Kairosfocus,
In short, the cites from Mr Dawkins raise — or miss — more questions than they answer. [And the ad hominems are unfortunately of a piece with his longstanding contempt for those who differ with him: ignorant, stupid, insane or wicked. Mr Dawkins: FYI, I have what seem to me to be some serious questions, and I am not simply setting out to smear and distort, I would like to have them resolved in a positive fashion. In that light I do not find off- the- cuff prejudice and dismissal civil; but instead to sound far too much like the incivility and regrettably ill-informed commentary that I and others, e.g, start here, find in your The God Delusion.
I suppose your words are heartfelt, but wouldn’t it have been easier and more conducive to the matter at hand, once and for all to offer a couple of words just to answer that simple question:
I ask again: Which algorithm from Atoms suite corresponds to Weasel and which one corresponds tpo the Partitioned Search discussed by D+M?
or
Which of Atoms algorithms is the partitioned search and which one Weasel?
Since you have made a lot of references to Atom’s program:
Nor is this theory, there are actual runs that do that. (HT: Atom and EIL.)

using replication of results as a good cross-check, e.g through Atom’s adjustable Weasel, here.

j] As already noted and linked, these patterns have been demonstrated through actual runs of Atom’s adjustable Weasel, from EIL
I think it would be a good idea to answer the question, or at least tell us why you don’t consider it relevant? I believe a lot of people are interested but I am afraid your performance so far has not convinced many of them.Cabal_{August 30, 2009
August
08
Aug
30
30
2009
10:36 AM
10
10
36
AM
PDT}

nephmon, observing how kf tries to defend some truly obvious errors with more and more distractions/obfuscations provides a certain amount of amusement to some people.Indium_{August 30, 2009
August
08
Aug
30
30
2009
10:35 AM
10
10
35
AM
PDT}

I've just been observing with a mix of amusement and astonishment. It is a vary minor but clear issue, these are two different algorithms, Dembski and Marks made a mistake in their citation, they should issue a correction, Kariosfocus can stop making a mockery of reason (and ID), everyone can move on.Excession_{August 30, 2009
August
08
Aug
30
30
2009
10:26 AM
10
10
26
AM
PDT}

Oops. Sorry, I didn't realize before that kf is actually suffering from a pathology. I'm astonished at the stamina of those who are arguing with him, but somewhat perplexed as to why they bother. Is someone who is so hung up up on the specifics of a program written 20+ years ago, and whose algorithm has been shown time and time again to exhibit just the behavior that was claimed for it, really worth arguing with? The really funny thing about all this is that RD's claims for the program were so tiny, and his acknowledgment of its limitations so clear, you really have to sympathize with someone whom it causes to spill so much virtual ink. Good luck with the debate, all! (Yes, I don't mention the Dembski algorithm because it is, of course, a huge irrelevancy.)nephmon_{August 30, 2009
August
08
Aug
30
30
2009
10:16 AM
10
10
16
AM
PDT}

BillB, as far as I can see only kf continues with this marathon of distractions. I am quite sure that Dr. Dembski will of course correct his misrepresentatiojn soon, it´s easy enough! For a man with his integrity it will only take a few days or 1-2 weeks to prepare a correction. I would bet a (very small) amount of money that he has so far not been made aware of these problems. So, may I ask again: Instead of waiting for Dawkins to show up here, which is very unlikely, why doesn´t Dr. Dembski explain the issue on his own blog? Can somebody ask him for a clarification? On the other hand he closed comments on the thread where he announced the paper. Thanks Clive and O´Leary for opening new ones, btw!Indium_{August 30, 2009
August
08
Aug
30
30
2009
08:32 AM
8
08
32
AM
PDT}

Its worth repeating this: D&M's process: -- With each programme loop you check each letter against the target, if it matches you do nothing, if it doesn’t you pick a letter at random to replace it. Repeat. -- Dawkins method: -- With each programme loop copy each letter of the previous winning phrase, with each letter copied having a probability (P) of being randomly replaced. Now add up the number of complete letters to get the fitness score. Perform this G times (where G is the number of generations). Now look through the list of scores and pick the highest, or any one of the highest if there are more than one. Repeat. -- So should we consider these two processes to be the same? They can't be written the same way as code, they don't function the same way and they don't produce the same results. Can they be considered the same for D&M's purposes (according to KF). Well I don't see how - the question about the amount of active info is very pertinent here - but this is also irrelevant to the main topic under debate, one which it is easy to loose sight of what with KF's army of oily burning sixty foot straw men marching across these pages. D&M describe a different process than Dawkins does, the differences are not trivial, particularly to anyone familiar to the topic of search algorithms. Because of this their reference to Dawkins work is incorrect and inappropriate; it should be corrected - and this will have no impact, as far as I can see, on the content of their paper. The surreal and farcical attempts to avoid this quite simple issue begs the question - is this really about search algorithms at all or is it about poking Dawkins with a stick.BillB_{August 30, 2009
August
08
Aug
30
30
2009
08:10 AM
8
08
10
AM
PDT}

KF, I think after reading the comments at Pharyngula, virtually everyone will agree that Dawkins' WEASEL did not use partitioned search. That's really the relevant point here, not whether the algorithm involved quasi or implicit latching, concepts which don't appear in the paper and which D&M don't discuss.yakky d_{August 30, 2009
August
08
Aug
30
30
2009
07:47 AM
7
07
47
AM
PDT}

(Hmm, for some reason it just posted my comment for me before I finished - here is the rest) 1 You code it so that there is a running list of incorrect letters. You work though this list on each iteration of the software and randomise letters on it. When the list is zero the program implicitly halts (or loops for ever doing nothing) 2 With each programme loop you check each letter, if it matches you do nothing, if it doesn't you randomise it. Option 2 is simpler but functionally equivalent as far as I can see EXCEPT that option two will re-converge on a new target if one is set. The latching mechanism in option 1 is permanent but in option 2 it is only permanent (i.e. non reverting) if the target does not move. I think both could be argued as legitimate interpretations? Of course there is still an explicit mechanism required for D&M's compared to Dawkins algorithms. The issue over halting conditions is a rather sad but amusing attempt to stretch the idea of latching behaviour even further - it would apply to any algorithm that can achieve its goal and renders the idea of latching, in the context of D and M's representation of WEASEL, meaningless. When looked at clearly - without KF's will-full attempts at obfuscation, his handy portable goal posts and the bulldozer he is using to try and re-shape the playing field - there is no way you can follow both algorithm descriptions and come up with the same code. Onlookers may also have noted how he laces all his comments with poisonous accusations of dishonesty. An upstanding example of civil discourse for sure!BillB_{August 30, 2009
August
08
Aug
30
30
2009
07:40 AM
7
07
40
AM
PDT}

Tomato Addict:
I have to conclude that the concept of latching and ratcheting is arbitrary, or that it is a property of all search algorithms other than blind random search. Either way, it seems to be a useless concept.
I think KF has even tried to imply that a blind random search has a latching mechanism here:
One of the peculiarities of Weasel algors, is that they halt when they hit home; of course halting being a key property of algorithms. As a result, if they hit home before a generation number G, they do not get to G. Thus, if Q means number of mutants to date, then if size of generation is S, Q = G*S. Immediately, if a run completes [all L letters correct] at gen G, it cannot have been complete before G. So, the paper’s p.1055 discussion would on this reading of Q would be of a ratcheted run that shows a march of champions with latching up to G, when it halts: completion AT G. (In an explicitly latched Weasel this would be automatic, in a version that on being tuned and giving a good run latches — as observed [our first swallowed camel . . . ] — this implies that when metric falls to distance to target = 0, there is a latching action imposed.)
KF is trying to make an art out of semantic mud wrestling. He is claiming that if an algorithm has a halting condition then this is a latching mechanism and because it is assumed that both Dawkins and D&M's algorithms are assumed to halt when they find the target, they both contain a latching mechanism. KF presumably doesn't regard the fact that this latching has no affect on the search as important. Dawkins also doesn't specify a halting condition and one is not actually required, although it would seem daft not to write one in. You can leave WEASEL running after it has found the target and it will continue to produce new generations, and the fittest will sometimes show incorrect letters. This becomes vital if you change the target to a new phrase because WEASEL will start to converge on this new target. Contrast this with Dembski and Mark's algorithm. They appear to specify a halting condition "The search for these letters is over" but it is slightly ambiguous. I realise that you could interpret and code it either way as follows: 1BillB_{August 30, 2009
August
08
Aug
30
30
2009
07:23 AM
7
07
23
AM
PDT}

--kf, I just wanted to add my voice to the choir:
Which algorithm from Atoms suite corresponds to Weasel and which one corresponds tpo the Partitioned Search discussed by D+M?
DiEb_{August 30, 2009
August
08
Aug
30
30
2009
06:47 AM
6
06
47
AM
PDT}

@kairosfocus, #185 By all means, the behavior you call latching exists. You have twice demonstrated such and ignored the question. It appears the definition of latching/ratcheting in effect (explicit, implicit, or quasi-) is that if the search moves toward the target, then "it must be latching because it looks like latching". I suggested a more formal mathematical definition, and I am unaware of any other definition. I have to conclude that the concept of latching and ratcheting is arbitrary, or that it is a property of all search algorithms other than blind random search. Either way, it seems to be a useless concept.Tomato Addict_{August 30, 2009
August
08
Aug
30
30
2009
06:35 AM
6
06
35
AM
PDT}

kf, you are so fixated on the latching stuff and the imagined differences between 1986 and 1987 it is honestlyn amazing. And of course you still avoid my questions! Hilarious! You are more and more retreating to the strange position that the algorithm that D+M describe is not really Weasel (admitting what we say all the time) but is still relevant even if the math is completely incorrect. Could you elaborate? How much active info does the real Weasel then have? Why don´t D+M discuss Weasel when say reference it and boast about their critic on this website? So, have fun discussing this with Dr. Dembski, who seems to think that he criticized Weasel and not something else.Indium_{August 30, 2009
August
08
Aug
30
30
2009
06:24 AM
6
06
24
AM
PDT}

BTW, these numbers are calculable, you don't have to take my word for them...DiEb_{August 30, 2009
August
08
Aug
30
30
2009
06:21 AM
6
06
21
AM
PDT}

Your links above and quotes are, unfortunately, not helping us resolve the evident differences between the showcased runs of 1986, and the videotape of 1987. The differences can be easily explained by the size of the population used in both examples: Using a mutations probability of 4%, a population of 10 takes 1305 generations on average, a population of 100 takes 78 generations. Obviously, Dawkins wanted to have the program running for the length of his talk, so he chose a smaller population than for the book. I'm sure that Dawkins would answer along these lines - if he was asked.DiEb_{August 30, 2009
August
08
Aug
30
30
2009
06:19 AM
6
06
19
AM
PDT}

BL: Your links above and quotes are, unfortunately, not helping us resolve the evident differences between the showcased runs of 1986, and the videotape of 1987. In short, the cites from Mr Dawkins raise -- or miss -- more questions than they answer. [And the ad hominems are unfortunately of a piece with his longstanding contempt for those who differ with him: ignorant, stupid, insane or wicked. Mr Dawkins: FYI, I have what seem to me to be some serious questions, and I am not simply setting out to smear and distort, I would like to have them resolved in a positive fashion. In that light I do not find off- the- cuff prejudice and dismissal civil; but instead to sound far too much like the incivility and regrettably ill-informed commentary that I and others, e.g, start here, find in your The God Delusion.] Some of the following has been raised several times but it will plainly bear repetition: 1 --> the showcased runs of 1986, with gushing commentary on the wonderful power of CUMULATIVE selection, demonstrate that there is good reason to infer that the showcased runs did ratchet and latch their way to the target: ________________ We may conveniently begin by inspecting the published o/p patterns circa 1986, thusly [being derived from Dawkins, R, The Blind Watchmaker , pp 48 ff, and New Scientist, 34, Sept. 25, 1986; p. 34 HT: Dembski, Truman]: 1 WDL*MNLT*DTJBKWIRZREZLMQCO*P 2? WDLTMNLT*DTJBSWIRZREZLMQCO*P 10 MDLDMNLS*ITJISWHRZREZ*MECS*P 20 MELDINLS*IT*ISWPRKE*Z*WECSEL 30 METHINGS*IT*ISWLIKE*B*WECSEL 40 METHINKS*IT*IS*LIKE*I*WEASEL 43 METHINKS*IT*IS*LIKE*A*WEASEL 1 Y*YVMQKZPFJXWVHGLAWFVCHQXYPY 10 Y*YVMQKSPFTXWSHLIKEFV*HQYSPY 20 YETHINKSPITXISHLIKEFA*WQYSEY 30 METHINKS*IT*ISSLIKE*A*WEFSEY 40 METHINKS*IT*ISBLIKE*A*WEASES 50 METHINKS*IT*ISJLIKE*A*WEASEO 60 METHINKS*IT*IS*LIKE*A*WEASEP 64 METHINKS*IT*IS*LIKE*A*WEASEL [Cf my comments in the always linked app 7] _______________ 2 --> This pattern from 1986 -- 200+ candidates that could revert in the samples if no latching is happening, but do not ever do so [consider on the import of the law of large numbers on this; i.e the simplest explanation that fits the facts is that the sample reflects the run of generational champions here . . . ] -- cannot simply be dismissed by making an offhand comment on events of 1987. For, it is known separately that adjustment of parameters and/or different runs will produce cases that latch and cases that do not. 3 --> In the case of the 1987 run, if the intent by Mr Dawkins is to say that he is showing the run of generational champions [a la 1986] -- and that seems hard to avoid -- then he is not explaining an apparent difference in showcased runs c 1986 and 1987, but distracting from it. (And Mr Dawkins is known to make significant use of misdirection and mischaracterisation in his public discussions, as the linked rebuttal to his The God Delusion documents.) 4 --> If instead, CRD is showing the run of individual members of the population in the video, that these do not latch letters is a built in characteristic of the implicit latching case: it is the run of champions -- the evidence that we are seeking to explain -- that would show latching in the case where: [1] to a sufficiently high likelihood no-change members appear in the generation's pop, [2] single step advances prevail otherwise. On this case, we will see latching of the run of champions as the current state will be locked in by [1] and advances if they happen will appear at [2] preserving the advances to date and occasionally adding to them. (And in some cases one run will latch and another will not on similar parameters. So, we are back at: wha tis being showcased in 1986 and in 1987, and how represenattive of the runs and population is that.) 5 --> Not to mention, how does the behaviour ofthe population fit with the gushing claims from BW that accompany the above runs: ____________ >> It [Weasel] . . . begins by choosing a random sequence of 28 letters … it duplicates it repeatedly, but with a certain chance of random error – ‘mutation’ – in the copying. The computer [indirectly, the programmer!] examines the mutant nonsense phrases, the ‘progeny’ of the original phrase, and chooses the one which, however slightly, most resembles the target phrase, METHINKS IT IS LIKE A WEASEL . . . . What matters is the difference between the time taken by cumulative selection, and the time which the same computer, working flat out at the same rate, would take to reach the target phrase if it were forced to use the other procedure of single-step selection [i.e. "more than a million million million times as long as the universe has so far existed" -- this is where CRD admits that active information has played a key role in the speedup] . . . >> _________________ 6 --> If double changes etc appear with sufficient frequency, we will see another demonstrated characteristic: substitutions where one letter advances, and another retreats -- which then tend to run across decadal generation boundaries. [this would be in either a quasi-latched or an unlatched case.] 7 --> In the showcased 1986 runs, it took 40+ and 60+ generations to hit target, so that no-advance won the generation contests about half the time, i.e. we have separate reason to see that we may well be in the implicit latching regime described. 8 --> And, if the showcased runs were unrepresetnative of he performance of Weasel c 1986, then CRD has some expalining to do on his gushign comments about cumulative selection. 9 --> Lastly, he seems to be suggestig that he weawsels out there more or less capture the patterns of behaviour. In this case, then explict latching and implicit latchigtn are legitimate interpretations, and ont he evidence that he presented he4 has showcased implicitly latched runs c 1986, and either internal members of po per gen in 18987 or a case taht is distinctly dfifferent formt he 1986 runs. 10 --> Or else, if the showcased 1986 runs did not ratchet to target, then his published excerpts were misleading - and that in a context that he admitted was in the broader sense "misleading." GEM of TKI PS: BL, 190: You are missing the point, and are reading INTO the remarks in the paper something that is not drawn out of it. This happens fairly frequently, so we do have to watch out for it. Try reading on a different "scenario," and see if this helps you see where I am coming from: "On p. 1055 M & D presented a didactic example of what partitoning of a search means, not a real world output of a program or illustration of an algorithm's typical output in one step." This, I have pointed out step by step from 161 above on, as a more credible explanation. In short, I see from my angle that on inference to best explanation, M & D are NOT presenting an algorithm or program or output from a run, but they are explaining a concept, then they go on to use the evident ratcheting as given above for Weasel 1986, to show its mathematical consequences. Runs from 1987 are irrelevant to the 1986 result which do show no reversions among the run of champions, in a sufficiently large sample that with the gushing commentary on cumulative selection, one may reasonably conclude that the program runs showcased did latch or ratchet, manifesting partitioning explicitly or -- more likely on balance of evidence and testimony -- implicitly achieved. That is the M & D Analysis is about what searches that show cumulative, ratcheted, latching selection do, and is thought to be relevant to the 1986 Weasel runs as showcased for reasons of the size of sample [as measured in letters] and the lack of reversions therein. Such ratcheting can be achieved by explicit or implicit latching effects, and on balance of evidence the later is most likely. In any case the underlying analysis is relevant and applicable as CRD's commentary in BW reveals -- inadvertently -- that the active information supplied by use of a preloaded target and a proximity hotter-colder metric and filter, makes the program outperform random walks by a wide margin.kairosfocus_{August 30, 2009
August
08
Aug
30
30
2009
06:04 AM
6
06
04
AM
PDT}

To Oatmeal’s excellent history of the whole affair at 181 , I would only add the following- the search shown in TBW cannot be a partitioned search, for the reasons Blue Lotus points out in 190 – the difference between generation 1 and generation 2. Something that D&M should be aware of. Kf has carefully avoided my post (164). He keeps saying “once latching occurs the D&M analysis applies”. This has been demonstrated to be false. Unfortunately (for him) he has now been backed into the position that Q is the total number of queries.
kf@185: Where I do have what I think on further thought overnight is an error, is that I have taken Q to be number of generations at the first, in that context. Q is — on second thought — number of generations multiplied by size of generations.
I insist that anyone who is having trouble following the math here, just go the http://www.evoinfo.org/Resources/WeaselWare/WeaselGUI.html and run a search for ‘Methinks it is like a weasel’. Then compare the Query Count for the Partitioned Search (about 98) with the Query Count for the “Proximity Reward Search” (which kf states (@163) is the Weasel algorithm) These were my runs (using G=50, MR=4%):
Target: METHINKS IT IS LIKE A WEASEL Search Type: English Phrase Partitioned / Deterministic / Prox Reward 1) 68 25 5850 2) 85 26 4400 3) 94 27 6800 4) 156 27 6100 5) 88 27 7800 6) 78 23 11950 7) 96 26 4800 8) 90 27 9800 9) 100 27 4000 10) 93 26 5600 11) 62 27 8700 12) 127 27 4500 13) 85 27 10150 14) 115 27 5200 15) 53 27 11850 16) 90 27 6350 17) 98 26 13950 18) 83 26 5650 19) 160 25 7700 20) 121 27 5700
So partitioned takes between 60 and 160 queries, but Weasel(50,4%) takes (using kf's definitions throughout) between 4,000 and 13,000. These are "equivalent"? But don’t rely on anyone’s assertions - try it for yourself! Enjoy!DNA_Jock_{August 30, 2009
August
08
Aug
30
30
2009
05:59 AM
5
05
59
AM
PDT}

Sorry again. First sentence should begin with "how", not "you".Indium_{August 30, 2009
August
08
Aug
30
30
2009
05:20 AM
5
05
20
AM
PDT}

Prev 1 … 4 5 6 7 8 … 13 Next

You must be logged in to post a comment.

Leave a Reply