FEA and Darwinian Computer Simulations

_{Gil Dodgen

June 11, 2011

Intelligent Design}

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

In my work as a software engineer in aerospace R&D I use what is arguably the most sophisticated, universally applicable, finite-element analysis program ever devised by the most brilliant people in field, refined and tested for 35 years since its inception in the mid-1970s for the development of variable-yield nuclear weapons at Lawrence Livermore National Laboratory. It is called LS-DYNA (LS for Livermore Software, and DYNA for the evaluation of dynamic, nonlinear, transient systems).

A finite element is an attempt to descretize on a macro level what occurs at a molecular level in a physical system, so that a result is amenable to a practical computational solution. The learning curve for the use of this sophisticated technology is extremely steep, and the most important thing one learns is that empirical verification of the simulation results is absolutely required to validate the predictions of any FEA model.

In an LS-DYNA simulation, all the laws of physics and the mathematics that describe them are precisely known. In addition, all of the material properties associated with the physical objects are precisely quantified with empirical verification (density, modulus of elasticity, and much more).

The FEA solver computes a physical result by solving millions of differential equations with a minimal integration time step based on the time required for a disturbance traveling at the speed of sound to traverse the smallest finite element with the greatest mass density.

Even with all of this, and countless man-years of experience by sophisticated and experienced users (LS-DYNA has been used for many years in the auto industry for simulating car crashes) empirical verification is always required, by actually crashing a car to validate the FEA results.

In light of all this, consider the typical Darwinian computer simulation and the trust that could be put in one.

Darwinian computer simulations are simply a pathetic joke as they relate to biological reality. This should be obvious to anyone with experience in the field of legitimate computer simulation.

Comments

Gil - can you respond to my question at 48? It's only just been released from moderation, so you might not have seen it earlier. Of course, this comment might not appear for another 4 days either. :-(Heinrich_{June 15, 2011
June
06
Jun
15
15
2011
11:33 PM
11
11
33
PM
PDT}

What gets designed by the GA is an optimally fit phenotype.
The GA itself doesn't design anything.
But those genotypes can be randomly generated, so that the corresponding phenotypes are all over the shop in terms of fitness.
So I could randomly generate some strings and you could use those as the starting population in one of your GA's and it would work fine? We both know that's not true, so what on earth do you mean?
You have to have, effectively, a starting population of genotypes, and some kind of way of relating the “genotype” to a “phenotype” (essentially, what the genotype does).
You have to have, effectively, a starting population of designed genotypes. Am I missing something, or didn't you already admit that? And if you're going to toss in phenotypes, those too would have to be designed. But seriously, phenotypes in a GA? What's an example of a phenotype in one of your GA's?
Whereas in typical GA applications evolution works directly on a population of candidate solutions, in nature there is a separation between genotypes and phenotypes (candidate solutions). - An Introduction to Genetic Algorithms
In fact, there is very little indeed about a GA that is not designed. So sure, if you want to use a GA as analogous to living things and evolution go right ahead. You're actually making a stronger case for design. Since you appear to have missed the relevant material, let me expand:
In genetic algorithms, the term chromosome typically refers to a candidate solution to a problem, often encoded as a bit string. The "genes" are either single bits or short blocks of adjacent bits that encode a particular element of the candidate solution. - An Introduction to Genetic Algorithms
The encoding is designed. It's a critical part of the picture in a GA. You left it out of your things to keep in mind about GAs. Surely you aware of how important it is to get the "chromosome" right in a GA. You can't use just any old chromosome. And surely you aren't going to tell us that the chromosome doesn't have to be designed with the solution in mind. Well, evolution, or so we are told, is not like that.Mung_{June 15, 2011
June
06
Jun
15
15
2011
05:47 PM
5
05
47
PM
PDT}

Well, it doesn't "magically" work, obviously. You have to have, effectively, a starting population of genotypes, and some kind of way of relating the "genotype" to a "phenotype" (essentially, what the genotype does). But those genotypes can be randomly generated, so that the corresponding phenotypes are all over the shop in terms of fitness. What gets designed by the GA is an optimally fit phenotype.Elizabeth Liddle_{June 15, 2011
June
06
Jun
15
15
2011
03:50 PM
3
03
50
PM
PDT}

Elizabeth Liddle:
Mung: the initial population in a GA is usually designed, although the individuals may be extremely simple.
I thought that the initial population was generated randomly. Do you do it differently in your GA's? But it should have been clear that I am talking about the "genome" or "chromosome," the entity that is used to represent an "individual" in the "population." That representation is designed. You cannot just take any old binary string and have it magically work. You have to chose an appropriate chromosome.Mung_{June 15, 2011
June
06
Jun
15
15
2011
03:31 PM
3
03
31
PM
PDT}

Mung:
In a GA, one has to encode the potential solutions into a genome. That is an extremely important part of the process, and it too is designed.
In what sense are the "potential solutions encoded into" the genome of a GA? (I mean, in some GAs they may be, but not the ones I am familiar with, unless we are using the term in different senses, which is possible, given our track record :))Elizabeth Liddle_{June 15, 2011
June
06
Jun
15
15
2011
02:16 PM
2
02
16
PM
PDT}

Mung: the initial population in a GA is usually designed, although the indivuals may be extremely simple. The final population has been "designed", from that simple prototype, by the Darwinian process. You could conclude from this that the original living things must therefore have been designed, but you could also conclude that their final diversity could be accounted to by the Darwinian process. And Darwin's theory does not attempt to address how the original simplest self-replicator came into existence. He specifically excludes this. I don't think we need infer a Designer for that part of the process, but you may want to. But then it wouldn't be an argument against Darwinian evolution, but an argument against natural abiogenesis.Elizabeth Liddle_{June 15, 2011
June
06
Jun
15
15
2011
02:14 PM
2
02
14
PM
PDT}

I repeat: The "organisms" in a GA are designed. Lizzie left this out. The obvious conclusion is that living organisms are also designed.Mung_{June 15, 2011
June
06
Jun
15
15
2011
01:55 PM
1
01
55
PM
PDT}

GAs are "general purpose" search methods... In genetic algorithms, the term chromosome typically refers to a candidate solution to a problem, often encoded as a bit string. The idea of searching among a collection of candidate solutions for a desired solution is so common in computer science that it has been given it's own name: searching in a "search space." Each chromosome can be thought of as a point in the search space of candidate solutions. The GA most often requires a fitness function that assigns a score (fitness) to to each chromosome in the current population. The fitness of a chromosome depends on how well that chromosome solves the problem at hand. ...candidate solutions to a problem are encoded as abstract chromosomes encoded as string of symbols, with fitness functions defined on the resulting space of strings. A genetic algorithm is a method for searching such fitness landscapes for highly fit strings. - Melanie Mitchell, An Introduction to Genetic Algorithms
Candidate solutions, not solutions. Encoded (by the designer). A point in the search space (designed). How well it solves a [target] problem (designed). Searching for = teleological. Target. Solution. Designed. Not like Darwinian evolution.Mung_{June 14, 2011
June
06
Jun
14
14
2011
03:10 PM
3
03
10
PM
PDT}

The first issue to be faced is how to represent the individuals (organisms) that make up an evolving population. A fairly general technique is to describe an individual as a fixed length vector of L features that are chosen presumably because of their (potential) relevance to estimating an individuals fitness. - Kenneth A. De Jong, Evolutionary Computation
1. It's the population that evolves, not the solution. 2. There needs to be a mapping (also designed) of the individuals to the designed fitness function.Mung_{June 14, 2011
June
06
Jun
14
14
2011
03:02 PM
3
03
02
PM
PDT}

Elizabeth Liddle:
Not sure which aspect you mean, unless it’s the computer I run it on.
See my post at @54 In a GA, one has to encode the potential solutions into a genome. That is an extremely important part of the process, and it too is designed. Chalk up one more for intelligent design.Mung_{June 14, 2011
June
06
Jun
14
14
2011
02:57 PM
2
02
57
PM
PDT}

Elizabeth Liddle @55: Thank you for the trouble of your response. I think that part of the misunderstanding (and I believe it is a misunderstanding) arises because someone who generally accepts evolutionary theory and also publishes a simulation is assumed to be asserting that their simulation proves evolutionary theory. Arguably, it is assumed that the simulation in fact abides by the tenets of evolutionary theory, the core tenets of which are random mutation and natural selection. Any simulation that "short-cuts" these two tenets, while possibly of some practical benefit in analyzing data to classify it or automate the generation of probability distributions, can not honestly be labeled an "evolutionary" simulation, certainly not without a multitude of "it depends" caveats. In reality, what the simulation does is test a specific hypothesis. That hypothesis may be supported, and the conclusion may in turn support evolutionary theory, but the idea that such simulations are the core of the evidence for evolutionary theory is, I think, simply wrong. Arguably, your simulations seem more like tests of varieties of Mendelian inheritance and searching for patterns in existing genomes (somewhat like cladistic analysis). You seem to be testing the probabilities of some pre-determined mutation becoming fixed in a population. You seem to be using a random number only to trigger whether the mutation manifests, but not the actual character of the mutation itself (the kind of mutation, its size, where on the genome it occurs, etc., are not random but was predetermined by you for your particular simulation purpose). What you are not simulating is a random mutation arising in some genotype, being manifested in a phenotype, and increasing the reproductive and survival fitness of parents and offspring, and becoming fixed in the population. You are not simulating random mutation and natural selection giving rise to a few beneficial but mostly deleterious de-novo traits (what "evolution" is assumed to be). Arguably, your simulations seem more atuned to genetic engineering research in which selected traits (not random traits) are being probabalistically analyzed for success. While I may have mischaracterized what you've actually simulated, I won't play "twenty questions" to ferret out the essential details. Obviously it all "depends". That is a given. But did your simulations actually depend on random mutation and natural selection? Seems not. Were they actually lab and field tested? Depending on the kind of classifying you intended, yes. But depending on trying to understand some facet of evolutionary theory, again it seems not, you implied you could develop such a simulation but not that you actually did. It also seems you don't use a single simulator which implements a consistent algorithm, but rather a variety of simulators each with customized algorithms, depending again on what you intend to learn. In and of itself that is not unreasonable. But I wonder how much mathematical and theoretical consistency all the simulators share in common. If they all shared the same computational routines, then perhaps there is a lot of consistency. But if they each use individually 'tweaked' or customized routines, then they collectively don't simulate a single consistent "evolutionary theory" but rather simulate several different varieties (or modes?) of a theory, which necessitates further caveats. There is no “core” evidence – there is simply IMO an abundance of circumstantial evidence for so many aspects of the theory that much of it we can take for (virtually) granted. There was likewise an abundance of circumstantial evidence for so many aspects of the (now defunct) theories of a geocentric solar system, spontaneous generation, time is constant, etc., and yet as greater attention was paid to the inconsistent details in the circumstantial evidence, as the hidden mechanisms at work were understood, those quaint theories were proven false. If there were few or no inconsistencies in evolutionary theory, it might be reasonable to take it (virtually) for granted. But that isn't the case. While theories are revised to reflect new observations, the core tenets of those theories are regardless repeatedly tested against reality and almost never taken for granted. Contrast 'taking evolutionary theory virtually for granted' with the standard model of physics and the "law" of gravity: as successful as they are, they are not taken for granted because there remain many nagging inconsisencies. Especially in view of how "difficult life science is" as you note, all the more reason to not "take it (virtually) for granted". While there is merit in simulations intended to explore a narrow set of data, the error of generalization (which you rightly cite) is committed by those who implicity or explicitly characterize such simulations as consistent with evolution, when in fact no such consistency is in evidence, certainly no consistency on evolution's core tenets of random mutation and natural selection having been closely modeled in the simulation.Charles_{June 14, 2011
June
06
Jun
14
14
2011
05:16 AM
5
05
16
AM
PDT}

Elizabeth Liddle:
I think it is really important to distinguish three things in a GA: The solution to the problem. This is not designed by the GA designer, but evolves.
That makes no sense to me. What do you mean the solution to the problem evolves?
OK, let me give a very simple (and not terribly typical) example. I have two sets of structural brain images. Each brain image consists of a 3D matrix of "voxels". I want to know what patterns reliably distinguish the Images A from Images B. So I set a template binary logistic regression equation, in which a linear combination of some parameter times the value of some voxel gives a probability that the brain in question belongs to one group or the other. I have no idea at the beginning what the difference between the brains is, if any. And I start by generating a population of "equations", each of which randomly selects any number of voxels, from any part of the brain, multiplies each by a randomly drawn parameter, and outputs a probability, for each brain, that it belongs to one group or another. And then I look at the accuracy of each equation (well, it's automated of course). Most of course will get 50% right and 50% wrong. But some fortuitously, will get slightly more than 50% right. So I "breed" from those (I usually use "sexual reproduction", so the "offspring" are random combinations of two "parents", but I also "mutate" the parameters, the number of voxels selected, and which voxels selected"). And very quickly I find myself with a population of equations that classify the brain images. rather well. I then test my classifier on a completely different set of brain images. Often this goes wrong, of course (my "population" of equations has "evolved" to survive well in a the very specific "environment" of the initial dataset, but can't cope in a different environment), so we go back to the drawing board. But if all goes well, we find we have a classifier that not only can tell, with considerable reliability, which images belong to set A and which to set B, but which tells us what the patterns are that distinguish the two. That is the sense (in this example) in which the solution evolves. We start off with a random equation that performs no better than chances, and we end up with an equation that can not only distinguish between categories of brain images, but can tell us what patterns distinguish them. In other words I end up with information, in a very real sense, that I did not have at the beginning, and that I did not put into the algorithm. The solution is what you have at the end of a run. (Or not. Perhaps no solution is found.) Either way, if a solution to the problem is located or not, it makes no sense to speak of a solution that evolves. Well, it does to me. Half way through I might have a classfier that is better than chance, but still produces false negatives and/or false positives. By the end, I hope for near-perfect classification. "Evolving" seems to me exactly what the solution does.
And really, you left out the most important aspect of a GA. Also designed, I might add.
Not sure which aspect you mean, unless it's the computer I run it on. If so, sure.Elizabeth Liddle_{June 14, 2011
June
06
Jun
14
14
2011
01:57 AM
1
01
57
AM
PDT}

DrBot @53:
If you stop the process part way through you still have a solution (or rather a population of candidate solutions) – it just won’t usually be as good a solution as if you leave it running for longer.
Precisely my point. You don't have a solution. You have a population of "candidate solutions" which may or may not actually constitute a solution to the problem that is trying to be solved by the GA. You don't have a "solution" until you choose one of the candidates, and even then it may not actually be a solution.
That is why it is entirely correct to say that the solutions(s) evolve.
Only by mangling the English language beyond recognition. The potential solutions "evolve." Or are you claiming that all the potential solutions "evolve" to converge on the same solution of you just let the GA run long enough?
The effectiveness of the solution is gradually increased over time using a process of replication, modification and selection (usually called evolution)
Do GA's always solve the problem posed? If not, then you're arguing that the effectiveness of non-solutions is gradually increased. In what world? Look, we're not morons here. Give us some credit.Mung_{June 14, 2011
June
06
Jun
14
14
2011
01:53 AM
1
01
53
AM
PDT}

Charles, thanks for this:
Ok, help me understand what you’re specifically talking about, and then I’ll translate my argument to your specifics.
Well, what I'm doing depends on what I'm trying to do! So the answers to your questions vary. But I'll try to give some typical answers:
In your program: 1) how do you model an individual’s genome
Again, it depends what I am trying to do. If it's really simple, I simply model it as a repertoire of possible outputs with a frequency distribtuion. Or it could be as an linear equation, in which the parameters are randomly adjusted; or it could be a potentially non-linear equation in which the terms themselves can be added or removed. Those approaches are useful for classifiers. If (for fun, mainly) I'm trying to model some aspect of evolution (as opposed to learning, or trying to solve a problem) then the genome may consist of a randomly adjustable "code" where different code sequences result in different "phenotypic" behaviour.
2) how many mutations per generation
That's usually an adjustable parameter. Sometimes it's interesting to see what the threshold is at which "evolution" breaks down, or does not occur. Again, it depends on the purpose of the program. However, I often use recombination as part of my "breeding" algorithm (as in a sexually reproducing population). With a large enough genome, sometims very few "external" mutations are required, because plenty of variants are produced by combining sequences from one parent genome with sequences from the other. In which case the cut points would also usually be random.
3) how much of the genome is mutated with each generation
Again, that depends on the purpose of the simulation.
4) how does a random number determine a specific mutation
In various ways. What I usually do is to generate random numbers between 0 and 1 and have a threshold (another adjustable parameter) below which a component of the genome is mutated; if it is to be mutated, again, it depends on what I am trying to do. It could be that I replace an element of the genome with another element, drawn from a probability distribution either determined a priori, or, in some cases, that pdf itself could be part of the simulation. Or, in the simplest models, I just add random "noise" to the behavioural repertoire.
5) what is the size of each mutation
Again, it depends. Sometimes it will be a change in an individual element of the genome. However, with recombination, the mutations can be quite large, in that the "child" allele may be a different size from the parent, and may differ from it in various ways, including numbers of repeats, and actual sequences.
6) how is a mutation’s benefit or cost determined
Depends on the purpose of the program. For learning models, I will build in "feedback" ("reward" or "penalty"). For classifiers, similarly, the better the classification (the more correct assignments) the better chance that genome has of replication. But it's possible (and I've so far only played around with this) to dispense with a formal fitness function, and have the ability to reproduce itself determine how likely it is to reproduce! That would be closer to a "real life" fitness function. However, normally, I want my critters to solve my problems, not theirs :)
7) how large is the starting population
Adjustable, and either remains startic or itself can grow or shrink. Depends on the purpose of the program. For practical problem solving it's a tradeoff between time and memory.
8) how is an individual’s breeding fitness scored
Well, of course, for problem solving, I'm not specifically interested in the breeding fitness, but whether it has figured out my problem! And in any case, fitness scores are relative to something (sometimes the ancestral population, sometimes the previous generation). So I haven't actually scored breeding fitness, although I would in the case of my good-replicators-replicate-model. But that's WIP :)
9) what is the criteria for a mutation to be adopted into the population
No criteria. Whether a mutation propagates is simply a function of its success; no additional criteria is required. It's output, in other words, not input.
10) how is an individual’s normal life-span modelled
Again, it depends on the program. Often by a cull of the least good performers, but again, it's output, not input. Good performers will outlive poor ones.
How then do you lab test or field test your simulation results?
Well, again, it depends on what the model is for. For classifiers, you "evolve" or "train" your classifier on a "training set" of data, in which the correct categorisation is fed back to the model, then you let it loose on a "test" set, in which you know the correct categorisation, and you see how well it does. If it does well, then you do that again a few times. Once you have a consistent classifier, the most interesting output is usually the method it uses, because that tells you what pattern it has found most reliably distinguishes the categories. But you may also want to use it, as, say, a diagnostic, where the categories are unknown even to you. For a behavioral model, you compare the behaviour of the model with the behaviour you are trying to model; so a good model will have similar learning curves, and make similar error types to the people whose behaviour you are trying to understand. For a model in which you are trying to understand some facet of evolutionary theory (by what steps an "irreducibly complex" function evolves, for instance) you would examine in detail the lineage of the IC function. But again, it depends what specific question you are trying to answer. And I suspect this is where the bone of contention lies. Gil is obviously a smart guy, and so, obviously was his father. But it is important, when evaluating the scientific integrity of other projects to understand what the specific purpose of any given study is. I know of no study designed to "prove Darwinism" or "model the evolution of species X". Most simulations are designed to solve a specific problem or test a specific hypothesis. And these are as rigorous as any in science, but, as with all hypothesis testing, you have to be very careful about the generalisability of your conclusions. I think that part of the misunderstanding (and I believe it is a misunderstanding) arises because someone who generally accepts evolutionary theory and also publishes a simulation is assumed to be asserting that their simulation proves evolutionary theory. In reality, what the simulation does is test a specific hypothesis. That hypothesis may be supported, and the conclusion may in turn support evolutionary theory, but the idea that such simulations are the core of the evidence for evolutionary theory is, I think, simply wrong. There is no "core" evidence - there is simply IMO an abundance of circumstantial evidence for so many aspects of the theory that much of it we can take for (virtually) granted. Other parts are unknown, and require further investigation. Sometimes these investigations deliver surprises (exciting surprises!) and we know that in many respects Darwin was wrong. But then all scientist are :) That's why it's fun.Elizabeth Liddle_{June 14, 2011
June
06
Jun
14
14
2011
01:34 AM
1
01
34
AM
PDT}

Elizabeth Liddle @45:
I’m not clear what aspect of that you think is not Darwinian?
I'm betting he thinks NONE of it is Darwinian. Charles @46:
1) how do you model an individual’s genome
BINGO! Elizabeth failed to mention that in her "important to distinguish three things in a GA" post. That's quite an oversight, really. I have to assume it was intentional, because she really should know better. Right Elizabeth? You're telling us the important things to distinguish about GA's, and you forget a little thing like encoding the genomes? How do you decide what the genome is going to look like? I mean, is there some cookie cutter genome that all GA's use?Mung_{June 14, 2011
June
06
Jun
14
14
2011
01:34 AM
1
01
34
AM
PDT}

The solution is what you have at the end of a run. (Or not. Perhaps no solution is found.) Either way, if a solution to the problem is located or not, it makes no sense to speak of a solution that evolves.
If you stop the process part way through you still have a solution (or rather a population of candidate solutions) - it just won't usually be as good a solution as if you leave it running for longer. That is why it is entirely correct to say that the solutions(s) evolve. The effectiveness of the solution is gradually increased over time using a process of replication, modification and selection (usually called evolution)DrBot_{June 14, 2011
June
06
Jun
14
14
2011
01:25 AM
1
01
25
AM
PDT}

What Tierra Is Not
Life on Earth is the product of evolution by natural selection operating in the medium of carbon chemistry. However, in theory, the process of evolution is neither limited to occuring on the Earth, nor in carbon chemistry. Just as it may occur on other planets, it may also operate in other media, such as the medium of digital computation. And just as evolution on other planets is not a model of life on Earth, nor is natural evolution in the digital medium. http://life.ou.edu/tierra/whatis.html
Seems that what Tierra can and cannot do is pretty irrelevant.Mung_{June 14, 2011
June
06
Jun
14
14
2011
01:01 AM
1
01
01
AM
PDT}

Elizabeth Liddle:
I think it is really important to distinguish three things in a GA: The solution to the problem. This is not designed by the GA designer, but evolves.
That makes no sense to me. What do you mean the solution to the problem evolves? The solution is what you have at the end of a run. (Or not. Perhaps no solution is found.) Either way, if a solution to the problem is located or not, it makes no sense to speak of a solution that evolves. And really, you left out the most important aspect of a GA. Also designed, I might add.Mung_{June 14, 2011
June
06
Jun
14
14
2011
12:55 AM
12
12
55
AM
PDT}

Elizabeth Liddle @43
Well, what I’m proposing on another thread is just that: a “targetless” model in which things that “breed” better breed more of themselves, thereby concentrating the traits that promote better breeding in the evolving population.
So then is it just coincidence that you chose not to use a GA? Aren't GA's "targetless"?Mung_{June 14, 2011
June
06
Jun
14
14
2011
12:44 AM
12
12
44
AM
PDT}

MathGrrl:
What do you think of Tom Ray’s Tierra? Does it meet your criteria of being “targetless”?
It's teleological.Mung_{June 14, 2011
June
06
Jun
14
14
2011
12:40 AM
12
12
40
AM
PDT}

Integrity is of paramount importance in science, I agree. But I do not think you have made your case that "Darwinism" lacks such. That doesn't mean that there don't exist scientists whose integrity is wanting. Unfortunately there is indeed a lot of sloppy science around. I don't see any evidence that it is particularly rife in evolutionary biology though. (Evolutionary psychology maybe....)Elizabeth Liddle_{June 14, 2011
June
06
Jun
14
14
2011
12:02 AM
12
12
02
AM
PDT}

Gil - can you release my comment that's in moderation. I'd be interested to see how you will respond. I think models have a far wider utility than you are allowing, and that the relationship between models and empirical data can be more subtle than you think. beyond the bluster, I think there are some interesting questions raised about how we can use simulations to tell us about the real world. I actually think it's something ID will have to face as groups like the EIL develop their models.Heinrich_{June 13, 2011
June
06
Jun
13
13
2011
11:44 PM
11
11
44
PM
PDT}

My father, who is 89, is one of the few remaining living scientists who developed the atomic bomb during WWII. He and his colleagues did so with computational tools no more sophisticated than a slide rule. They had to get it right, in a hurry and with no excuses for failure, because, as my father told me when I was growing up in the 1950s, if Hitler got the bomb first he would have the power to rule the world. These are the standards of scientific integrity, discipline, and accountability with which I was groomed during my formative years. Darwinism represents the antithesis of such standards.GilDodgen_{June 13, 2011
June
06
Jun
13
13
2011
09:03 PM
9
09
03
PM
PDT}

Elizabeth Liddle @45: I’m talking about programs (well part of what I’m talking about) where a starting population of individuals is bred, with random variance, where the probability of breeding depends on their score on a fitness criterion. Ok, help me understand what you're specifically talking about, and then I'll translate my argument to your specifics. In your program: 1) how do you model an individual's genome 2) how many mutations per generation 3) how much of the genome is mutated with each generation 4) how does a random number determine a specific mutation 5) what is the size of each mutation 6) how is a mutation's benefit or cost determined 7) how large is the starting population 8) how is an individual's breeding fitness scored 9) what is the criteria for a mutation to be adopted into the population 10) how is an individual's normal life-span modelled How then do you lab test or field test your simulation results?Charles_{June 13, 2011
June
06
Jun
13
13
2011
03:36 PM
3
03
36
PM
PDT}

Hi Charles :) You wrote:
Elizabeth Liddle @39: The glaring deficiency is that *none* of the evolutionary simulation implementations actually employ random number generators when defining a “mutation”. Random mutation and natural selection are *the* core principles of Darwinian evolution and while the algorithms discuss “random mutation”, the actual programmed implementations are *not* random, they’re not even pseudo-random.
Well, I use a random-number generator in mine, always. I guess I could use some true random generator, like the ones you can get from atmospheric noise, but I honestly don't see that that is critical, and the great thing about a random number generator is that you can choose whether to reset the seed or not, which is usefuul if you want to re-run a previous run for some reason. Or am I not understanding you? I'm certainly not aware that other people don't use random-number generators, and all the ones I've looked at do.
Further, “natural selection” is simulated by matching against a pre-determined forward-looking fitness pattern rather than a backward-looking simulation of what has previously survived, or even culling some percentage of mutations based on empirical observations of how few mutations (as a percentage) survive.
Well, that may be true of some,I guess, depending on the purpose of the simulation, but again, in mine, I don't do that. Unless I'm misunderstanding what you mean by "matching". Of course you match the output to your fitness function, but that is merely the equivalent, in nature, of an individual's probability of reproduction being determined by some interaction between the trait it bears and the environment it inhabits. Could you clarify? In the case of the ones I write (either to solve a problem or to simulate learning) I generally select the top 2/3rds or so of each generation (sometimes the top half) for "mating" and breeding. Or sometimes I pick pairs at random and have the winners of each pair "mate". There are lots of ways of doing it.

That doesn’t make life-sciences junk science – it makes them difficult science!
Science as a discipline is about factual observation and intellectal honesty. No degree of difficulty excuses substituting imagination for fact and claiming something has been simulated. Science is not a video game. Or is it now?
No, of course it isn't. But you left out a crucial component of science, which is hypothesis testing! Simulations are just one of the many kinds of models we test.
When engineers simulate heat transfer from a liquid into a solid, they don’t compute the motion of every molecule based on physics, rather it is a statistical averaging of the effects of molecules treated as a multitude of finite groups in progressive “slices” adjacent to and across the boundaries. Empirically tested “rules of thumb” accurately and realistically “emulate” (yes, quantitatively replace) the net effect of the particular liquid and solid instead of attempting to compute the energy transfer at each molecular collision. Engineers can do this reliably because a) they factually know the composititon of the materials involved, and b) the properties and behavior of these materials are accurately known because they have been measured and recorded specifially for use in exactly these kinds of computations. The handbooks and databases literally fill libraries.
Yes indeed. My husband was a physicist before he became a life-scientist :)
This is indeed difficult for evolutionay biologists because they a) don’t know the properties of what they are “simulating” and b) don’t know what rules to apply because they neither understand the behavior nor have any rules, but inexcusably have even ignored most of the structures in question on the presumption they’re junk and don’t matter.
Well, no. I simply disagree with you here. Or, at best, I don't know what kind of "simulation" you have in mind. What you say isn't mapping on to my experience (as a life scientist), anyway.
If an engineer were to build a flight simulator from a duck decoy hanging by a string with a fan in front of it, and his excuse was that birds appear lighter than air and the only question is how they push forward through a wind, that simulation would never lead to any insight of how airflow over a curved surface produces lift. Yet evolutionary biologists claim to gain insight into evolution by random mutation and natural selection from non-random algorithims and pre-determined fitness patterns, the only question in their (closed) minds being how rapidly do variances spread through a population.
Far more questions that one flow through our very open minds :) But yes, we do gain insight (not that I am an evolutioanry biologist, but what I do is realted) from simulations, which is why we do them!
In the same breath you assert: So I don’t think anyone expects a “Darwinian simulation” to resemble anything like a real-life scenario. No scenario could. … And so we can, using simulations, demonstrate that facets of Darwin’s theory (the core, in fact) work: that if a population breeds with variance, and if that variance results in phenotypic differences in reproductive success within a given environment, the population will evolve and adapt [Charles: ostensibly by random mutation and natural selection, right?]. We can also test this in the field and in the lab, with rigour. … Where is the “rigour” in ignoring the junk DNA? Where is the rigour in simulating random mutation with non-random algorithms? Where is the rigour in simulating natural selection with forward looking pattern matching?
Who is "ignoring" "junk" DNA? Non-coding DNA (non-coding for proteins anyway) is an absolutely vital part of genetic research, especially in my field. And the Grants' work in the Galapagos was nothing if not rigorous and meticulous, as was Endler's work with guppies, and indeed Lenski's work with E.coli. I accept, from your comments, that you do not agree, but I suggest that you may have an inaccurate model of how biological research is done, and what simulations are used for.
How can you excuse using non-random, forward-looking algorithms as ‘demonstrating that core facets of Darwin’s theory in fact work’, and while such ‘simulations do not resemble real-life scenarios’ they can be tested in the field with rigour? Do you not expect to find real-life in the field. Are random mutation and natural selection not core facets of Darwin’s theory? If random mutation and natural selection are not simulated then claims to have genetic or evolutionary algorithms can not be sustained. If you will next argue that real-life Darwinian random mutation and natural selection are just too difficult to simulate, then you really have no basis to assert Darwinian evidenciary standards are high, do you. They are at best, arbitrary, and their simulations are nothing more than “just so” programs.
I don't think I am following you. I'm not talking about "non-random, forward looking algorithms". I'm talking about programs (well part of what I'm talking about) where a starting population of individuals is bred, with random variance, where the probability of breeding depends on their score on a fitness criterion. I'm not clear what aspect of that you think is not Darwinian?Elizabeth Liddle_{June 13, 2011
June
06
Jun
13
13
2011
03:00 PM
3
03
00
PM
PDT}

Elizabeth Liddle @39: The glaring deficiency is that *none* of the evolutionary simulation implementations actually employ random number generators when defining a "mutation". Random mutation and natural selection are *the* core principles of Darwinian evolution and while the algorithms discuss "random mutation", the actual programmed implementations are *not* random, they're not even pseudo-random. Further, "natural selection" is simulated by matching against a pre-determined forward-looking fitness pattern rather than a backward-looking simulation of what has previously survived, or even culling some percentage of mutations based on empirical observations of how few mutations (as a percentage) survive. That doesn’t make life-sciences junk science – it makes them difficult science! Science as a discipline is about factual observation and intellectal honesty. No degree of difficulty excuses substituting imagination for fact and claiming something has been simulated. Science is not a video game. Or is it now? When engineers simulate heat transfer from a liquid into a solid, they don't compute the motion of every molecule based on physics, rather it is a statistical averaging of the effects of molecules treated as a multitude of finite groups in progressive "slices" adjacent to and across the boundaries. Empirically tested "rules of thumb" accurately and realistically "emulate" (yes, quantitatively replace) the net effect of the particular liquid and solid instead of attempting to compute the energy transfer at each molecular collision. Engineers can do this reliably because a) they factually know the composititon of the materials involved, and b) the properties and behavior of these materials are accurately known because they have been measured and recorded specifially for use in exactly these kinds of computations. The handbooks and databases literally fill libraries. This is indeed difficult for evolutionay biologists because they a) don't know the properties of what they are "simulating" and b) don't know what rules to apply because they neither understand the behavior nor have any rules, but inexcusably have even ignored most of the structures in question on the presumption they're junk and don't matter. If an engineer were to build a flight simulator from a duck decoy hanging by a string with a fan in front of it, and his excuse was that birds appear lighter than air and the only question is how they push forward through a wind, that simulation would never lead to any insight of how airflow over a curved surface produces lift. Yet evolutionary biologists claim to gain insight into evolution by random mutation and natural selection from non-random algorithims and pre-determined fitness patterns, the only question in their (closed) minds being how rapidly do variances spread through a population. In the same breath you assert:
So I don’t think anyone expects a “Darwinian simulation” to resemble anything like a real-life scenario. No scenario could. ... And so we can, using simulations, demonstrate that facets of Darwin’s theory (the core, in fact) work: that if a population breeds with variance, and if that variance results in phenotypic differences in reproductive success within a given environment, the population will evolve and adapt [Charles: ostensibly by random mutation and natural selection, right?]. We can also test this in the field and in the lab, with rigour. ...
Where is the "rigour" in ignoring the junk DNA? Where is the rigour in simulating random mutation with non-random algorithms? Where is the rigour in simulating natural selection with forward looking pattern matching? How can you excuse using non-random, forward-looking algorithms as 'demonstrating that core facets of Darwin’s theory in fact work', and while such 'simulations do not resemble real-life scenarios' they can be tested in the field with rigour? Do you not expect to find real-life in the field. Are random mutation and natural selection not core facets of Darwin's theory? If random mutation and natural selection are not simulated then claims to have genetic or evolutionary algorithms can not be sustained. If you will next argue that real-life Darwinian random mutation and natural selection are just too difficult to simulate, then you really have no basis to assert Darwinian evidenciary standards are high, do you. They are at best, arbitrary, and their simulations are nothing more than "just so" programs.Charles_{June 13, 2011
June
06
Jun
13
13
2011
02:14 PM
2
02
14
PM
PDT}

@ NeilBJ #41
Re: Elizabeth Liddle @#39
And so we can, using simulations, demonstrate that facets of Darwin’s theory (the core, in fact) work: that if a population breeds with variance, and if that variance results in phenotypic differences in reproductive success within a given environment, the population will evolve and adapt.
Yes, we can demonstrate that facets of Darwin’s theory work, but only because we overlay our understanding of Darwin’s theory on the program that is written. Demonstrating with a program how we think evolution works is not the same as replicating a real evolutionary sequence.
I agree.
Do you agree that Avida is a demonstration of how evolution works? Does it not have a conscious, that is, intelligent selection algorithm in it? Does evolution have a conscious selection algorithm?
Both Avida and evolutionary process select. Avida does this by means of an algorithm that scores the functionality of each individual (IIRC) according to a table. Those that score highest have the greatest chance of breeding (can't remember how stochastic Avida is). In evolution the "scoring" is intrinsic - the organisms who have whatever functions it takes to maximise their probability of breeding simply do, so no algorithm is needed. But the principle is the same: in the Avida simulation managing to perform some cool high scoring function helps you breed; in the wild, managing to do something cool like be well camouflaged also helps you breed, but instead of being artificially ("intelligent") give a score for it, you just breed well because you don't get eaten!
As I have thought about evolutionary algorithms, I have wondered how a “targetless” algorithm could be written. If the evolutionary variation occurs independently with respect to need, it would seem that it would be impossible to write an honest algorithm.
Well, what I'm proposing on another thread is just that: a "targetless" model in which things that "breed" better breed more of themselves, thereby concentrating the traits that promote better breeding in the evolving population. I won't know in advance what the trick is (as those who use GAs to solve real world problems don't know in advance with the trick is - they just set up the fitness algorithm so that solving it enhances survival).
The variation that occurs is unpredictable, and whether or not the organism survives as a result of that variation or in spite of it is an after-the-fact observation. And this does not even begin to solve the problem of how an algorithm could be written to demonstrate a significant morphological change.
Well, not realistically, obviously, because life is very complicated! But in principle, I don't see why not. Do you know the clock-evolving video?
The programmer seems to have no choice but to overlay a target in the algorithm. The target the programmer implements obviously biases the program in favor of that target. It would seem that a program with a very large number of targets would come incrementally closer to approximating a real evolutionary process, but at what cost?
I think it is really important to distinguish three things in a GA: The fitness function (the criterion, decided by the GA designer, as to whether the output of an individual matches the desired output) The kind of mutations the individuals under go (but not which ones) - again this is decided by the designer. The solution to the problem. This is not designed by the GA designer, but evolves. So in a GA, I would argue, the designer has two main inputs, the fitness function and the search space, and leaves the evolving critters the job of figuring out how to solve the problem. In evolution, the fitness function is simply performed by the environment, and actually constantly changes as the population itself is part of its own environment. The search space - well, that's where, conceivably, a Designer might be inserted (cf Behe) but may be itself constrained by fitness (in other words populations that tend to mutate in a certain way may survive as populations better than populations that mutate to fast, or too slowly, or too radically). The individuals - well the individual in life are the direct analogs of the individuals in the GA.
I am reminded of Dr. David Berlinski’s observation that true Darwinian algorithms do not work and genetic algorithms that do work are not Darwinian.
Well, it's witty, but not entirely true IMO :) Cheers LizzieElizabeth Liddle_{June 13, 2011
June
06
Jun
13
13
2011
11:48 AM
11
11
48
AM
PDT}

NeilBJ,
As I have thought about evolutionary algorithms, I have wondered how a “targetless” algorithm could be written. If the evolutionary variation occurs independently with respect to need, it would seem that it would be impossible to write an honest algorithm.
What do you think of Tom Ray's Tierra? Does it meet your criteria of being "targetless"? What do you mean by "an honest algorithm"
The programmer seems to have no choice but to overlay a target in the algorithm.
I don't think this is done in Tierra, but I'm curious to see if you agree.MathGrrl_{June 13, 2011
June
06
Jun
13
13
2011
09:04 AM
9
09
04
AM
PDT}

Re: Elizabeth Liddle @#39 And so we can, using simulations, demonstrate that facets of Darwin’s theory (the core, in fact) work: that if a population breeds with variance, and if that variance results in phenotypic differences in reproductive success within a given environment, the population will evolve and adapt. Yes, we can demonstrate that facets of Darwin’s theory work, but only because we overlay our understanding of Darwin’s theory on the program that is written. Demonstrating with a program how we think evolution works is not the same as replicating a real evolutionary sequence. Do you agree that Avida is a demonstration of how evolution works? Does it not have a conscious, that is, intelligent selection algorithm in it? Does evolution have a conscious selection algorithm? As I have thought about evolutionary algorithms, I have wondered how a “targetless” algorithm could be written. If the evolutionary variation occurs independently with respect to need, it would seem that it would be impossible to write an honest algorithm. The variation that occurs is unpredictable, and whether or not the organism survives as a result of that variation or in spite of it is an after-the-fact observation. And this does not even begin to solve the problem of how an algorithm could be written to demonstrate a significant morphological change. The programmer seems to have no choice but to overlay a target in the algorithm. The target the programmer implements obviously biases the program in favor of that target. It would seem that a program with a very large number of targets would come incrementally closer to approximating a real evolutionary process, but at what cost? I am reminded of Dr. David Berlinski’s observation that true Darwinian algorithms do not work and genetic algorithms that do work are not Darwinian. No life-science experiment will be as conclusive as a physics or an engineering experiment, because there are far more variables, most of which are unknown, and whose pdfs have to be estimated, or even guessed at. That is all I am trying to say in this post.NeilBJ_{June 13, 2011
June
06
Jun
13
13
2011
08:35 AM
8
08
35
AM
PDT}

Elizabeth; despite your unwarranted fondness of it, neo-Darwinian evolution IS, without a doubt, Pseudo-science; Is evolution pseudoscience? Excerpt:,,, Thus, of the ten characteristics of pseudoscience listed in the Skeptic’s Dictionary, evolution meets nine. Few other?pseudosciences — astrology, astral projection, alien abduction, crystal power, or whatever — would meet so many. http://creation.com/is-evolution-pseudoscience =============== "Certainly, my own research with antibiotics during World War II received no guidance from insights provided by Darwinian evolution. Nor did Alexander Fleming's discovery of bacterial inhibition by penicillin. I recently asked more than 70 eminent researchers if they would have done their work differently if they had thought Darwin's theory was wrong. The responses were all the same: No. Philip S. Skell - Professor at Pennsylvania State University. http://www.discovery.org/a/2816 Podcasts and Article of Dr. Skell http://www.evolutionnews.org/2010/11/giving_thanks_for_dr_philip_sk040981.html Science Owes Nothing To Darwinian Evolution - Jonathan Wells - video http://www.metacafe.com/watch/4028096 ================ And though neo-Darwinian evolution is absolutely horrid as to being a rigorous science, anyone who dares question it is 'EXPELLED" EXPELLED - Starring Ben Stein - Part 1 of 10 - video http://www.youtube.com/watch?v=Fj8xyMsbkO4 Slaughter of Dissidents - Book "If folks liked Ben Stein's movie "Expelled: No Intelligence Allowed," they will be blown away by "Slaughter of the Dissidents." - Russ Miller http://www.amazon.com/Slaughter-Dissidents-Dr-Jerry-Bergman/dp/0981873405 Academic Freedom Under Fire — Again! - October 2010 Excerpt: All Dr. Avital wanted to do was expose students to some of the weaknesses inherent in Darwin’s theory. Surely there’s no harm in that — or so one would think. But, of course, to the Darwinian faithful, such weaknesses apparently do not exist. http://www.evolutionnews.org/2010/10/academic_freedom_under_fire_-_038911.html Journal Apologizes and Pays $10,000 After Censoring Article - Granville Sewell episode - June 2011 http://www.evolutionnews.org/2011/06/journal_apologizes_and_pays_10047121.htmlbornagain77_{June 13, 2011
June
06
Jun
13
13
2011
04:30 AM
4
04
30
AM
PDT}

Prev 1 2 3 4 5 Next

You must be logged in to post a comment.

Leave a Reply