This is just an effort to help keep Joe’s thread focused and to help keep it from being derailed. People can use it or not. I hope they will.
CharlieM: Can someone explain to me, why is all of this not just a model of directed evolution? Surely it is set up to be directed towards a target?
Allan Miller: It gets towards the target by means of variation and selection of genotypes in the current population. The programmer does not direct it towards the target. Indeed, there would be no point in writing GAs for problem solution if it were simply a matter of specifying a target and directing the program to find it.
Consider a small modification – instead of distance from target, evaluate fitness by adding up the ASCII bits. Those with the greater sum are fitter than those with a lesser. There is no mention of a distant target – although it is clear that the program will converge on a string of all Zs, the program doesn’t know this. It is not drawn by, nor directed towards, that target. It is simply doing generational evaluation of fitter and less fit genotypes in the current population.
Discuss.
CharlieM,
Yes, there is a target in some implementations, but all that happens in any selective arena is that genes are sifted according to their effect on survival. When there is a target, survival depends on approach to that target. When there is no target, survival depends on ‘something else’.
Such as my targetless/targeted ‘ASCII bitsum’ examples, which seem to be being ignored for some reason.
There is no fundamental difference at gene-within-population level. This seems a surprisingly difficult point to get across to your side. Perhaps due to your constant resort to inappropriate analogies, rather than getting your hands grubby with code.
CharlieM,
There’s nothing to stop the dimensionality and interactivity being increased, subject to computing limitations. It’s only not done in typical GAs because exhaustively modelling life isn’t the objective.
The remarkable thing is that GAs were inspired by life. Something that you think doesn’t work ‘in the real world’ was based precisely upon it and finds tons of practical applications. Food for thought?
It would be a simple thing to randomly change a character of the target every once in a while. The algorithm will track the changes, just as changes in the environment are changed in life.
Every single organism alive today is an example of evolutionary success. Evolutionary success was achieved in each single generation leading up to these organisms. Even those organisms that have become extinct are necessary for this evolutionary success. Hands can’t be formed without the death of countless bone cells and extant life could not be as it is without the death of countless lifeforms. There is as much direction in the latter as in the former. The only difference is that we can view the former as an external spectator, the latter we are in the midst of and so do not see the whole process.
CharlieM,
Evidence?
You cannot get your hands grubby with abstractions. If you want to see real models they are in the life around us. The life of a butterfly is a model for the life of a flowering plant and the life of a flowering plant is a model for the life of a butterfly. This is not a pre-conceived idea but a conclusion I have come to.
Study in detail the life cycle, habits, activities and relationships of butterflies and flowering plants, compare them and hold them together in your thoughts. This is an example of Goethe’s gentle empiricism where nature is revealed in its true reality and nothing is forced. I’m not asking anyone to believe this just to try it if they feel inclined. There are close connections in nature that genetics will never show.
CharlieM,
If you want to critique GAs, it helps enormously if you have a crack at writing one and try to understand what is going on.
You seem rather patronisingly to assume that I have no grasp of the complexities of the living world.
And you are dodging the point. GAs are based entirely on what happens within populations of living organisms subject to differential reproduction (NS). They work. How come the thing that inspired them doesn’t work like that at all, yet they succeed?
I didn’t say that you don’t appreciate the complexities of the natural world.
I think I know what is going on with GAs, they succeed in modelling natural selection to various degrees, and that is all very well. But what if natural selection acts not so much like the designer of an aircraft but more like a modification program? The airgraft is put into service and through use various faults and problems become apparent, modifications are embodied to change the aircraft in a way that better fits its purpose.
This is how I see natural selection operating and GAs do a reasonable job of mimicing this, but it doesn’t get to the actual creativity in nature. Weasel is very good at finding a sentence which was already in existance but it will never have anything resembling the creative abilities of Shakespeare.
And neither will you. So what?
I resent that remark. William McGonagall often gets the credit for the following gem, but in reality it was me:
“on yonder hill,
there stood a coo,
it’s no there noo,
it must have shifted.”
I am beginning to share Allan’s & Petrushka’s frustration!
It is becoming very clear to me that some present suffer a fundamental misunderstanding of how mathematical and scientific models are “useful”.
A model is akin to some conceptual “map”. A map is not a miniature recreation of an entire physical landscape. Understanding this would go a long way to explaining George Box’s aphorism “All mathematical models are wrong, however some are useful.”
By definition: models are necessarily simplifications or approximations, that can still provide interesting answers to very specific questions, thereby fulfilling Popper’s criterion of falsifiability.
No let’s reexamine the original question Joe Felsenstein was attempting to address. In Joe’s own words:
The purpose of the program is to show that creationist orators who argue that evolutionary biology explains adaptations by “chance” are misleading their audiences.
In other words, the evolution-doubters aka “creationist orators” conjured a most trivial bordering on outrageously silly mathematical rebuttal to the cogency of the current Evolutionary Theory.
Again in Joe Felsenstein’s own words:
Although Dawkins’s Weasel algorithm is a dramatic success at making clear the difference between pure “chance” and selection, it differs from standard evolutionary models…
As far as I can tell, a trivial mathematical rebuttal was required to slay (once and for all) the mathematically trivial bordering on silly evolution-doubting-creationist argument Joe mentioned.
The premise of Mung’s current thread is premised on blatant ignorachio elenchi that fails to appreciate the initial intent of the weasel algorithm. The weasel algorithm handily slays that initial incarnation of the silly “evolution as purely random/chance is mathematically impossible, dontcha know” shibboleth as originally proposed. Gainsayers are being most disingenuous by disregarding their initial mathematical naïveté while simultaneously and surreptitiously moving the goal-posts.
The fact that the weasel-algorithm can also provide even further insight and understanding than initially intended constitutes extra gravy on the “Wurst”… to paraphrase a German aphorism. (I am thinking of course of the contentious Neutral vs Adaptionist debates raging on other blog sites. Of course, the weasel algorithm cannot in principle answer every conceivable question thrown at it. No model can do that! Ever!!! Any such consideration in no way detracts from weasel-algorithm cogency (in what it initially set out to do).
I already touched on most of this with my earlier post suggesting that we were rehashing philosophical arguments from long ago?
No one is claiming that GAs are not useful. Why even think such a thing, must less level that accusation?
I thought the OP was quite clear. The purpose was two-fold. To help keep Joe’s thread from getting derailed by discussions about Weasel and to address the question about whether it is an example of directed evolution. The original intent of the Weasel program hardly matters to either.
I don’t quite understand these comments. I did address your example. And you already know what the outcome would be:
“There is no mention of a distant target – although it is clear that the program will converge on a string of all Zs, the program doesn’t know this.”
What’s the point of writing code just to tell us something we already know?
It is not drawn by, nor directed towards, that target.
It sure looks and acts like it is!
Perhaps someone else will take you up on that. I did not claim that evolution is guided and I really have no interest in debating with you about it.
I asked first, Allan. It was your claim. If your argument is going to reduce to well, yeah, by any reasonable definition it’s guided/directed, and all the experts agree with Mung, “but that’s just semantics,” then I don’t see any need to discuss it further.
If you have some ideas about how we could test the claim I’d be interested.
Evolution is “guided” by mutation and selection. GAs model that “guidance.” No spooks required.
So?
Does it follow from this that Weasel isn’t a model of directed evolution?
@ Mung… Just to be clear here; would you consider “artificial selection” to be an example of “directed evolution”?
Chuckle.
Mung,
As others have already explained to you, that depends on what you regard as “directed evolution”.
The bottom line is that GAs are not “directed” in any sense that should give comfort to a selection-phobic IDer.
But they are directed and I am glad we can agree on that. As for selection-phobic IDers, I don’t know of any.
CharlieM,
I wish you’d lay off the analogies! Natural selection acts like natural selection. That is, certain genotypes do better or worse in a current environment, because they increase or decrease in a gene pool in a manner correlated with their effect on phenotype.
You have thereby retracted your argument that GAs model directed evolution, as far as I can see. You put the direction somewhere else, creating the bare bones as it were, and selection tunes. But not in a directed sense. Or do you think some ‘director’ is actively killing more of the bacteria that lack antibiotic resistance, say, or more of the finches with a suboptimal beak shape?
Shakespeare was not in the business of finding genotypes fit for a particular environmental constraint.
Mung,
Only under certain not-very-useful definitions of what it means for a GA to be “directed”.
Got a mirror handy?
You’ve been after that wascally Weasel and its demonstwation of cumuwative sewection for a long time.
Mung,
They were addressed to Charlie, but I guess they apply to you too. I have yet to see any of you GA-skeptics – phoodoo, Cordova, Charlie, you – y’know – write a bloody GA!
I really feel it would help. Would you take it well if your business users tell you how code works, despite them having never written any?
This is you ‘addressing my example’:
The rather obvious conclusion is that there’s more than one way to specify a target. And again, you have not shown that in this case we are not dealing with a model of directed evolution.
All that shows is that you have missed the point. Here it is again.
Program A counts bits in an ASCII string and gives higher fitness to higher values. What is the ‘target’ in this? It’s pretty obvious where it will end up, but does that make it a target?
Program B counts bits in the ASCII string, subtracts that from the bitsum in target phrase ZZZZZZZZZZZZZZZZZZZZZZZZZZZZ and gives higher fitness to lower values.
Which of these is directed, by what and why?
I’ll add a program C: it counts bits in the ASCII string, subtracts that from 56, and gives higher fitness to lower values. Many strings will match that; does that make them all targets?
So: which of these is directed, by what and why?
Of course! My intention was to analyse an example where the outcome is fucking obvious, so as not to be sidetracked by unknowns. Was that unreasonable?
Haha. The point is to illustrate something about selection, and about notions of directedness.
Well, that’s evolution for ya! It always looks as if the result was what was desired all along. Hence The Blind Watchmaker, and other books dealing with Paleyesque arguments, and the ease with which products of selection can look like intentional outcomes.
I wouldn’t call mathematical models wrong, I would just call them abstractions and often “useful” abstractions. I would say the misunderstanding comes when the abstractions are taken for reality.
Just look at how Dawkins enthuses about his computer biomorphs in “The Blind Watchmaker:
Does he really think that the little stick figures his tinkering with his computer produced can in any way be compared to real insects?
Simon Conway Morris wrote In “The Crucible of Creation”:
“…species with very different adult forms may reach this final stage via markedly different developmental pathways.”
Maybe the Weasel algorithm models some aspects of evolution more closely than Dawkins would have liked.
Mung,
If you think natural selection is a form of direction, there is indeed no need to discuss further. I said that early on in the thread ***, but you seemed not to agree.
Channelling Joe G … if a program picks the string against which fitness is to be evaluated, I don’t see how the programmer could be accused of directing the program to that target. That was a suggestion I made some time back. It really does not matter which string is chosen, the program will find it by cumulative selection, using much the same algorithm as any other GA outside of the selection part.
[*** eta the relevant quote being “ISTM that all implementations of NS in a GA would be regarded as ‘guided’ in their terms. If that’s the case – if NS and ‘guiding’ are synonyms – the dispute seems terminological.”
I see no reason why the program couldn’t, at random intervals, mutate the target string.
petrushka,
It doesn’t need to do that either. 1 extra iteration of the population initialise routine could provide a random, unknown target. Most random strings are likely to be at approx the same Hamming distance – about 26 – from all other such strings. The chance of it being less is about equal to the chance that a random member of the start population is already that far on its way to Weasel.
So for example you can create your alphabet on the fly and create your target phrase on the fly. But so what.
ALPHABET_SIZE = 27
PHRASE_LENGTH = 28
char_set = (32..126).map {|i| i.chr}.shuffle.slice(0, ALPHABET_SIZE)
target_phrase = PHRASE_LENGTH.times.map {char_set.sample}
You won’t know in advance what your target phrase is, or even the specific characters it will be made of. But so what?
How is that conceptually different from biology?
I mean, how is variation and selelection conceptually different from one implementation to another?
Mung:
Says Mung, who thought this sort of thing was rilly, rilly important a couple of months ago:
Mung,
So no accusations can be made that the programmer is directing the program towards the target. He/She doesn’t even know what it is. There has to be an evaluation routine of some kind. You can’t have consistent survival differentials without it. But many different such routines could be envisaged, and they don’t need to involve a distant goal. That’s why I looked at ASCII bit sums. It’s not the distant goal of all Z’s that draws the program upwards, but the simple fact that in any pool, higher sums will breed more. Z’s happens to be a limit, in this character set, not a target.
Ultimately, a digital-string GA subsets a character space. It may or may not subset the space down to a single slot. When it does, you seem to regard that as a ‘target’. That’s what happens in Weasel. There is a single peak of maximum fitness and smooth slopes leading to it.
But one could change the evaluation routine in a vast number of ways, to subset the space differently, with many more fitness peaks (all targets), and lethal space or regions of detriment that block progress in some directions. That is but a small step, and is really no different from biology.
The only ‘direction’ comes from NS. If that involves a directOR, it must be mimicking the actions of the environment, actively killing (say) more antibiotic-susceptible bacteria than resistant ones, and keeping tabs. But why can’t the environment cause this differential itself?
It was even analyzed mathematically, back in 2000 or 2001. Name of the author eludes me at the moment.
Direction from natural selection is still direction is it not?
If I’m not mistaken this is Dembski and Mark’s point. Any result an algorithm achieves is by definition already present in the inputs to the algorithm.
What selection does is separate the desired result from the noise.
We direct the process by choosing what will be selected.
The only way to mimic Darwinism in software AFAIK is to let the memory constraints of the computer do the selecting with out human input.
Set up a reproducing mechanism and let the code run indefinitely with no artificial constraints whatsoever till it fills the available memory and crashes then look to see if anything interesting rose to the surface.
Just a drive by comment
peace
Mung,
The Wright-Fisher model assumes only that the contributions of the sites to fitness are independent and multiplicative. What enters into the calculation that Joe did is the ratio of the numbers of fit and non-fit alleles at a site. It does not matter at all which alleles contribute to fitness. If the environment changes, and what contributes to fitness changes at a particular site, the evolutionary process, as modeled by Wright-Fisher, simply takes some time to return to equilibrium.
We’ve seen, in the plots I’ve posted on Joe’s thread, that an evolutionary process at equilibrium can be high in average fitness, and yet rarely, if ever, achieve maximum fitness.
As I mentioned in a comment on Joe’s thread, the math works out the same when 2 of 54 possible alleles at a site contribute identically to fitness as when exactly 1 of 27 possible alleles contributes. For 28 sites like that, the so-called “target” contains
length-28 sequences of alleles.
As I said in Keith’s thread, what really matters is that the alleles contribute independently to fitness. I have a very hard time seeing an absence of interaction as design to achieve an end. It is the absence of interaction, not the specification of particular alleles at sites as “good” or “bad,” that makes the math pan out as it does. To put that more concretely, you cannot show me a target in the math Joe presented, and you cannot show me a target in the program I wrote (and rewrote for you).
And if, hypothetically, the program spewed out
OOOWWOOWOOWWOWOOOOWWWOOOOOWW
that would be further proof that it was not reaching any target intended by the programmer.
I hope you weren’t hoping that Mung would implement your idea.
Naah.
P.S.–I should have referred to the specific case of the Wright-Fisher model we’re considering, and not the model in general. And I slipped up again, and referred to the average number of fit alleles in the parent as its fitness. If the number of fit alleles is
, then the fitness is 
Apt metaphor.
fifthmonarchyman,
Depends what you mean by ‘direction’. It is completely trivial if you are using ‘direction’ as a synonym for sample bias. I don’t know why we have so many threads on it if that‘s all that’s at stake.
That cannot be correct. Well, not unless you define the result of an algorithm as ‘that which was already present in the inputs’, which would be a pretty stupid definition. Why run an algorithm?
If a result is ‘desired’. But undesired results will also rise above the noise. All that is required is that there be a correlated differential in replication, not that someone wishes there to be.
We could easily let the process choose its own evaluation criteria.
If there is no mechanism differentiating between genomes, there will be nothing but noise (although there will be evolution, just by drift). So, you are declaring it impossible to model selection. Just like that. That is, you think it impossible to create a model that has a consistent correlation between different genotypes and their survival chances.
It’s a point of view, I suppose.
Presumably it is also impossible to model antibiotic resistance.
Allan has a short memory. Cordova’s remarkable algorithm:
You even made comments afterwards. So what’s your excuse for forgetting?
You should make a retraction.
This is perhaps a version of the
determinismpredestination of Calvinism.I am still not clear on what “directed evolution” is.
Say a population of wolves is preying on a population of deer. We may expect that natural selection will bring about changes in the length of limbs, the reactivity of the nervous system, and the acuity of vision of the deer. Of course since we do not have the detailed genomics of the deer, and do not know what other aspects of fitness confront the deer with a tradeoff, we cannot know exactly what response to expect.
Is such a case “directed evolution”, with the wolves doing the directing?
It’s not. What they say is that information is “inputted in the construction of the search,” or something very close to that. However, fifthmonarchyman has got one thing right. Their notion of active information is tautological.
ETA: There is no input to a so-called search.
I’d suggest a definition for Directed evolution as “directed evolution is like directional selection toward a non-existent conceptual trait.”
Travelling salesman converges on the same conceptual target even if exact mutations are different, i.e. different seeds for the random number generator still result in the same final outcome.
NOTES:
https://en.wikipedia.org/wiki/Directional_selection
And the deer directing evolution of the wolves?
Directed evolution also converges on the same target independent of initial condition — like starting out with a bird or tree or starfish and still getting a human at the end.
The devil is in the word converge. The travelling salesman solutions will converge toward better solutions, but not the same solution.
I’m not sure what your intended meaning of tree to human evolution might be. If you meant it literally, it isn’t true.
I would suppose if the initial conditions of the ancestors were:
“METHINK IT IS LIKE A DONKEY”
or
“METHINK IT IS LIKE A FAWCET”
or
“METHINK IT IS LIKE A CAMERO”
it would still evolve to
“METHINK IT IS LIKE A WEASEL”
That’s directed evolution.