|
Author
|
Topic: I.G.D. Strachan: An Evaluation of "Ev"
|
Erik
Member
Member # 160
|
posted 16. July 2003 17:17
RBH,
0. I use the term "model" in the same sense as statisticians do (well, maybe I mean it a little more generally), i.e. as a relation between observable quantities. By a "simulation" I mean a sequence of states, generated by a computer program using some rule for updating the state from time step to the next (I'm not sure I'd like to commit myself to this sense for all of science, but it is my definition for evolutionary biology). In a simulation, we can observe not only the initial and final states, but also all the intermediate states, and compare the fit to empirical observations of the real world. A simulation can therefore be considered as a model of all the states (not just initial and final).
Dawkins's WEASEL, Strachan's VETE, Schneider's Ev and Adami's AVIDA can all be considered simulations of the genotype frequencies of evolving populations. If we find a biological population of genotypes, whose reproductive success we know (or are willing to assume) can be described by, say, Strachan's fitness function and genotype space, then we could use Strachan's VETE to simulate the time-evolution of genotype frequencies. Adami's AVIDA is, in addition, also a simulation of a lower-level reproduction process, whereas the three former GAs leave that part to our imagination. In AVIDA, the fitness function emerges as a higher-order result of simulated reproduction processes and other interactions. In WEASEL, VETE, and Ev the fitness function is given a priori, and then a genotype frequency dynamics consistent with the description given by the fitness function is enforced. Whether a GA simulates both the reproduction process and the genotype frequency dynamics, or just the latter, has nothing to with "goal-directedness" or "locality".
The calculation of Dawkins's fitness function does not necessarily involve an explicit comparison to a distant genotype. For instance, if the genotype space consists of all binary sequences of length 100, then the fitness function can be the sum of all the individual bits in the genotype. This is of course mathematically equivalent to the interpretation of the fitness function as 100 minus the Hamming distance to the sequence 11...1. All fitness functions can be written as sums of distances to fixed points.
1. You're supposed answer "Yes, up to isomorphism!". Isomorphic models make exactly the same predictions and therefore it is at best a task for metaphysicians, not scientists, to argue that one isomorphic model is better than another.
2. If gravity is not a good topic for examples, then why not pick any scientific equation of your choice, give the veridical expression for the equation, and explain why this expression is the veridical one? Pick the ideal gas law or any equation in this excellent survey of mathematical models in evolutionary biology or whatever is the most convenient example.
3. My point was not that you should feel psychological discomfort. My point was that learning that physicists typically do not agree with your interpretation of mathematical models is a good reason to reevaluate the reasons that led you to the opposite conclusion.
Erik
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 16. July 2003 19:53
Erik,
Now that we've got the semantic issues sorted out, as far as I can see we do not disagree. Let me try these propositions:
1. In your sense of "model" I agree that there is no more or less veridical model of gravitation; all equations formally equivalent to F = GMm/r^2 are on equal footing as statistician-type "models." Therefore your question in point #2 is moot: there are no choices to be made. I don't interpret physicists' mathematical models as frameworks for simulations, and thus there is nothing with which to disagree. In this sense of "model" I do not disagree with physicists.
2. Simulations, on the other hand, are better or worse, more or less veridical (up to isomorphism, of course!). I used the term "theory-guided" above because substantive theory is what is supposed to identify the relevant aspects that a candidate simulation should map. To the extent that a simulation appropriately maps the variables, entities, processes, and relationships that theory identifies as relevant, the simulation can serve as a device for testing theories even in contexts where we cannot write the equations of 'motion' of the system. That use of simulations is the object of the game for the various evolutionary simulations under discussion.
3. The several evolutionary simulations differ significantly in how they represent the various processes identified by evolutionary theory, and thus it's appropriate to consider whether one or another is more veridical, whether one or another more faithfully corresponds to the variables, etc., identified by theory as relevant. Thus showing that two simulations are mathematically equivalent does not imply that they are equally veridical simulations. That "All fitness functions can be written as sums of distances to fixed points" does not imply that the various different operations by which a simulation might evaluate the fitness of a genotype are all equivalent in the eyes of theory.
4. Evaluating fitness as the count of the individual bits (1s) in a genotype (on the assumption that there is a theory-based reason for more 1s being reproductively advantageous) is not the same evaluation operation (in a simulation) as calculating 100 minus the Hamming distance to the sequence 11 ... 1. In the first case the simulation is using only local information, the number of bits in the genotype together with a counting operation and comparisons of the counts for existing genotypes to determine relative fitness in the population. In the second case the simulation is using information about some distant state (11...1) -- the simulation has and uses information about the distance to the remote state 11...1.
In fact, the second case simulates a teleological evolutionary system while the first simulates a blind evolutionary system. (That they would produce exactly identical system dynamics on the same fitness landscape is of no little interest. It implies that "teleology" in the context of evolutionary processes is a superfluous notion.) Thus the different but mathematically equivalent ways of calculating a fitness value translate into different simulation operations that map to theory differently. Evolutionary theory asserts that only local information is used in evolution, and thus the first operation (counting genotype bits) is the preferable operation for calculating fitness in a simulation of evolutionary theory; it is more veridical. One of the criticisms of the use of GAs and evolutionary algorithms in general as simulations of evolutionary processes is that they are inherently teleological in how they determine fitness. But that criticism fails if the implementation of the evaluation of fitness in a simulation uses solely local information. (I'll note in passing that in some instances one doesn't know what a peak or reference fitness is so the second case isn't possible to implement in a simulation.)
Again, I don't think we disagree significantly; I think we are merely using words differently.
RBH
[ 17. July 2003, 03:58: Message edited by: RBH ]
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 17. July 2003 02:52
For those interested, I've posted some preliminary remarks on a few calibration and familiarization Avida runs in the Literature Review forum.
RBH
IP: Logged
|
|
roncram_2000
Member
Member # 834
|
posted 17. July 2003 16:57
Micah,
I am not a biologist, but I am interested in Ev and how it relates to Dembski's theory. The Strachan evaluation is informative, although I need to reread it.
One of the points Schneider made on his website is that he was greatly influenced by Stephen Jay Gould and it was after understanding Gould that he wrote Ev.
Schneider claims Ev confirms the evolutionary viewpoint advanced by Gould (a cambrian-like explosion of species). Schneider's statement that he was influenced by Gould prior to writing Ev makes me think he may have programmed in the ability to learn, especially to learn in a way that would confirm Gould.
Now I read your comment:
"The cool thing is that we end up with a different organism in each run."
I do not understand. Why is that a desired result?
IP: Logged
|
|
Micah Sparacio
Member
Member # 6
|
posted 17. July 2003 17:25
multiple solutions (organisms) to the same problem.
IP: Logged
|
|
Micah Sparacio
Member
Member # 6
|
posted 17. July 2003 17:29
Pim indicated that the disanalogy between VETE and Ev was that VETE specified the target whereas Schneider did not.
In both VETE and EV, an algorithm is applied to the genome, that outputs a string of symbols. In VETE there are 27 distinct symbols, and the algorithm is the Vignere decoding. In EV, the algorithm is the perceptron scanning across the genome, and it outputs a string of symbols, of which there are two different kinds ( true and false ).
In both cases, survival is determined by comparing the string to a target string. In the case of VETE, the target sentence is "THE ANSWER IS FORTY TWO" (which could in principle be a randomly chosen string). In Schneider's case it's the array of true and false booleans in sitelocations.
IP: Logged
|
|
Pim van Meurs
Member
Member # 541
|
posted 18. July 2003 13:26
Micah: In both cases, survival is determined by comparing the string to a target string. In the case of VETE, the target sentence is "THE ANSWER IS FORTY TWO" (which could in principle be a randomly chosen string). In Schneider's case it's the array of true and false booleans in sitelocations.
I am not sure what you are trying to say here, there are indeed some superficial similarities but the main difference is that in Vete, the user specifies the goal while in Schneider the target and recognizer coevolve. Thus in Vete the outcome is always "the answer is forty two" while in Schneider the answer is the result of mutation and selection. Thus while in the former case the argument could be made that the information was pre-loaded, such argument becomes harder in case of Schneider.
To conclude: While there may be some superficial similarities between Schneider and Vete, they also differ in a significant area namely the 'target'. In Schneider mutation and selection lead to the increase of information in the genome without the need for the user to pre-load this information.
Roncram:
quote:
Schneider's statement that he was influenced by Gould prior to writing Ev makes me think he may have programmed in the ability to learn, especially to learn in a way that would confirm Gould
I object to such statement for two reasons. 1) you suggest that Schneider programmed Ev to make it have a certain outcome 2) you have access to the source code to see if such an accusation, and yes I consider this a serious assertion in scientific terms, can be supported by the evidence. Schneider provides the methodology and the program for all to inspect.
Schneider describes the history
quote:
HISTORICAL NOTE: After getting the idea that Rsequence is close to Rfrequency from data on ribosome binding sites (on 1982 December 7), it became clear that this had to have evolved. Eventually I recognized that Rsequence must evolve toward a more or less set Rfrequency. (See the paper Schneider.ev2000 for a discussion of this.) The challenge was to test this hypothesis. My first thinking was that a simulation of the evolution would take thousands of years on a computer. However, I read the book "Ever Since Darwin, Reflections in Natural History" by Stephen J. Gould and found in there an interesting chapter "Is the Cambrian Explosion a Sigmoid Fraud?" (Gould1977.sigmoid) This described Gould and Eldredge's idea of punctuated equilibrium. I realized that with selection, if an organism has an advantage it can take over a population exponentially. By 1984 I had written up my thesis (on the observation that Rs ~ Rf, Schneider1986) and had a week before my defense in which I had nothing more to do. So I wrote the ev program and it showed that Rs does indeed evolve toward Rf. I don't recall that I even mentioned this during the defense, but it was a nice backup
[ 18. July 2003, 13:31: Message edited by: Pim van Meurs ]
IP: Logged
|
|
roncram_2000
Member
Member # 834
|
posted 19. July 2003 02:50
Pim,
Thank you for posting the quote from Schneider that I referenced earlier. The section that is especially important is this sentence:
"I realized that with selection, if an organism has an advantage it can take over a population exponentially."
I heard somewhere that Dembski had asked someone how Schneider had "snuck in" the ability to learn. As with any computer modeling, I am certain that a number of assumptions were programmed into Ev.
It seems to me the area Dembski should be looking at is the assumptions Schneider used regarding selection. For example, I believe he killed off 50% of organisms with the most mistakes.
How do we know that was the right number to use? What would happen if only 1% were killed off? Would organisms eventually become infertile and the entire class become extinct? Or what would happen if 99% were killed off before reproducing? Is it possible that change would be so slow as to be insignificant?
Which of these numbers would most correspond with observations of biology? Is it even possible to find a number that would correspond?
Since I am not prepared to run these simulations, I can only wish that someone else would.
I readily admit there is much about this whole area I do not understand.
[ 19. July 2003, 02:53: Message edited by: roncram_2000 ]
IP: Logged
|
|
Pim van Meurs
Member
Member # 541
|
posted 19. July 2003 13:45
RonCram: You are certainly right that the program made certain assumptions and these assumptions are there for all to inspect. Without understanding the details it is more helpful to suggest how Dembski may run the model rather than make assertions about the program(mer) of Ev that may or may not be relevant.
And yes the relevant part is 'selection' since as Schneider has shown, without selection nothing much happens. In fact this is quite easily to understand by realizing that selection is what generates information in the genome via the environment.
IP: Logged
|
|
roncram_2000
Member
Member # 834
|
posted 19. July 2003 20:27
Pim,
You wrote:
"And yes the relevant part is 'selection' since as Schneider has shown, without selection nothing much happens. In fact this is quite easily to understand by realizing that selection is what generates information in the genome via the environment."
If Schneider has shown anything, he has shown that selection generates information in certain computer models. This is not the same as showing that it generates information "in the genome via the environment."
Pim, I would like to read your comments on Strachan's three main points:
"1. The information that is supposed to arise “from scratch” without any external intervention, is in fact put into the simulation from the start. In order to be able to locate the binding sites, their locations have to be specified before the simulation can run. What follows during the simulation is simply a form of Supervised Learning, of a simple type of neural network (called a Perceptron). The “Supervision” in a supervised learning procedure alludes to the process where at each round of training, the outputs of the neural network are compared to a specified target and corrections for errors are made accordingly."
"2. The calculation of the information content for the binding sites is done in the paper by computing the information content at each of the six locations in the binding site separately, using all 16 sites as a sample, and then running them. This relies on an important simplifying assumption that the values of the bases at each of the locations are statistically independent. This assumption is acknowledged in [Schneider et al, 1986], where the exact formulae for the information calculations are given, but it appears that in the current paper, the assumption is taken for granted. However, it can be easily demonstrated that this assumption does not apply in general for the simple perceptron recognizer used in the simulation. The only case where the perceptron would conform to this independence assumption in fact corresponds to the case where all the information is concentrated onto one axis (i.e. input variable to the perceptron), and the rest are unspecified; a situation that does not, and could not occur in the simulation."
"3. Further to this, we find htat the actual formula for computing R sequence (L), the information content for one location L in the site, is also flawed, relying on a standard formula for statistical entropy, or uncertainty, which itself can be derived only as an asymptotic limit as the number of samples tends to infinity, after appying Stirling'’ approximation for In(N!). It can be demonstrated that this approximation is invalid for the small sample sizes (N = 16) that are used in the simulation."
IP: Logged
|
|
Pim van Meurs
Member
Member # 541
|
posted 19. July 2003 20:58
Roncram, more later in detail but let me point out that Scheider's simulations seem to model mutation and selection in nature quite well. But even granting you that they did not show this in a real genome, they did show, contrary to what Dembski suggests in NFL that complex specified information can increase and that the increase is caused solely through mutation/selection.
In nature we can see similar examples in which mutation and selection can lead to increase in information and this is not that hard to understand in terms of Shannon information and mutual entropy. It's in fact the environment which provides for the information for the genome through linking particular genomes to a higher survival than others.
Specification of binding sites is of no interest since the amount of information at the beginning is close to zero, and the binding sites can be chosen randomly.
2. is irrelevant for the issue of increase in information but affects the amount of increase in information. Again an interesting comment that deserves additional study but it does not seem to affect the results
3. Even if the formula is flawed, information still is shown to increase, just not as much.
2 and 3 are thus in my opinion strawmen arguments when discussing increase of complex specified information but need to be addressed for other reasons.
1. seems irrelevant since if the information were prespecified in the selection of the binding sites how come that the information still varies from zero the expected value only in the presence of mutation AND selection. Secondly, the binding sites are chosen randomly again undermining the need for intelligent designers.
So far I would first like to resolve to what extent Weasel and Vete and Schneider share similarities AND differences. The main difference seems to be that the information is NOT pre-specified unlike Vete and Weasel "The answer is forty two". The fact that the outcome is contingent makes it very different from Weasel and Vete, especially when it comes to claims of 'front loading'.
[ 19. July 2003, 21:10: Message edited by: Pim van Meurs ]
IP: Logged
|
|
Micah Sparacio
Member
Member # 6
|
posted 19. July 2003 22:25
Pim,
Despite your insisting the opposite (and I'm sure you'll continue to do so, so please surprise me), there is a fairly well defined target in Schneider. You shouldn't keep denying this. The fact that the target in VETE is "out in the open for all to see" should not count against it.
Unfortunately, it does count against it in your eyes.
IP: Logged
|
|
roncram_2000
Member
Member # 834
|
posted 20. July 2003 02:03
Pim,
Thank you for your comments. I look forward to your post discussing this in greater detail, esp. point #1.
Micah,
Your view, agreeing with Strachan, that Schneider does specify a target is interesting in that you do not see it as a problem for Schneider.
According to Strachan, Ev makes corrections for errors based on the specified target. Strachan believes that this is putting the information "into the simulation at the start." Why do you disagree?
IP: Logged
|
|
Pim van Meurs
Member
Member # 541
|
posted 20. July 2003 02:28
Micah, you write:
Despite your insisting the opposite (and I'm sure you'll continue to do so, so please surprise me), there is a fairly well defined target in Schneider.
And yet this well "defined target" is neither predefined not predictable. Even the location of the bindings sites is randomized and seem to me to be a red herring since the information does not increase through the selection of the binding sites, indeed it requires mutation and selection for information to increase.
Well defined target? In what sense Micah? I am not insisting on the opposite without any reason, the findings that Ev does not evolve to a pre-specified solution is an essential difference.
[ 20. July 2003, 03:34: Message edited by: Pim van Meurs ]
IP: Logged
|
|
Micah Sparacio
Member
Member # 6
|
posted 20. July 2003 09:21
This will be my last exchange on the topic, because really, there isn't much at issue here (we're talking past each other). Anyway, Pim says, in reference to Schneider, that
"this well "defined target" is neither predefined not predictable"
Two points. First, many of its properties are predefined and predictable, and you've agreed with this in the past. Second, if we focus on what we're supposed to focus on in VETE (the proper analogue) then what we see is that the genome and cipher text are not predefined or predictable (though like in Ev, many of their properties are). They co-evolve.
Pim, not to be blunt, but it is fairly clear that you are mistaking analogues in the simulations. Never in VETE do we end up with "The Answer Is Forty Two." We always end up with a different genome and cipher text. You'll have to deal with this before the discussion can go any further.
IP: Logged
|
|
|
Printer-friendly view of this topic