|
Author
|
Topic: I.G.D. Strachan: An Evaluation of "Ev"
|
Moderator
Administrator
Member # 1
|
posted 28. June 2003 09:25
An Evaluation of "Ev"
by I.G.D. Strachan
Abstract- In this paper we assess the validity of the evolutionary simulation described in the paper “Evolution of Biological Information” (Schneider, 2000), which, it is claimed, demonstrates the evolution of new biological information from scratch with no external intervention, as a set of characteristic patterns developed in nucleotide binding sites. The further claim is made that the amount of information that evolves, a quantity designated Rsequence is approximately equal to the amount that is needed to locate the binding sites, given the number of them on the genome. This quantity is designated Rfrequency We find both these claims to be flawed, firstly that information arises from scratch, when in fact all the simulation is demonstrating is a form of “supervised learning” of a fixed target by a simple form of neural network. Secondly, we demonstrate that for the neural network classifier used in the simulation, that Rsequence & Rfrequency except for the unrealistic exceptional case where all the information is confined to one axis (one base in the binding site). We also show that the result showing Rsequence ≈ Rfrequency reported in the Schneider paper is a coincidence, resulting from an incorrect evaluation of the correction for small sample sizes, which fails to take into account the fact that the standard formula for uncertainty (or entropy) is a limiting case that utilizes Stirling's approximation for ln N!, which is not valid for small values of N. Analytical results are backed up by simulations, including an extensive set of runs using Schneider's Pascal program ev.p.
Reference:
Schneider, T. D. (2000). Evolution of biological information. Nucleic Acids Res., 28:2794- 2799. http://www.lecb.ncifcrf.gov/~toms/paper/ev/.
To read the entire paper, click here.
[ 28. June 2003, 09:28: Message edited by: Moderator ]
IP: Logged
|
|
Moderator
Administrator
Member # 1
|
posted 28. June 2003 14:35
Topic is now open for discussion.
[ 28. June 2003, 23:16: Message edited by: Moderator ]
IP: Logged
|
|
yersinia
Member
Member # 324
|
posted 29. June 2003 15:26
This paper is interesting for the recalculation of information measures from Scheider's ev simulation -- I am not qualified to judge the merits, but it appears that because the nearby binding sites are not independent (because of the small size of the genome) the actual information per binding site was ~3 bits rather than ~4 bits, a difference not detected by Schneider because various simplifications happened to cancel out. This by itself, however, amounts to a minor correction although over half of the paper (15 pages) is devoted to demonstrating it.
The key claim, however, is that ev boils down to being just another Dawkins "METHINKS" simulation and thus proves nothing about "generating information from scratch" because all of the information was included at the beginning. Unfortunately, this key part of the argument takes up only a few pages, and really only one page discusses the biological analogy. Strachan's basic argument is that the binding site locations, although randomly generated at the beginning of each simulation, still constitute the pre-specified target.
The obvious answer to this is that in biology, the location of binding sites is pre-determined by the location of genes. The expression of genes is controlled, in part, by the binding of activator and repressor proteins to DNA regions near to the promoter sequences (to which RNA polymerase binds in order to begin transcription of a gene's DNA to RNA). This crudely describes the basic situation in prokaryotes, in eukaryotes regulation is much more complex; an introduction is online here:
http://www.indstate.edu/thcme/mwking/gene-regulation.html
Here is a graphic:
The ev simulation is (IMO) analogous to a situation in which a gene is duplicated without some of its regulatory sequences. There are various processes that can sometimes retain such genes, e.g. the gene might be selected for because it increases dosage of the protein product. If it is advantageous, however, to have the gene copy expressed in a different situation (or perhaps the gene mutates to have a new function that is more advantageous, and then different regulation is advantageous), then there will be selection for binding between activators/repressors and the regulatory region.
Strachan attempts to address this objection, and the section is rather short so I will quote it and comment:
quote:
2.3 Binding sites forming next to pre-existing functional proteins
It might be argued that the pre-specifcation of the binding site locations is "as in nature", or that the exons in the genome are already present, thereby defining the regulatory region. In the paper, Schneider notes that:
quote:
An advantage of the ev model over previous evolutionary models, such as biomorphs, [Dawkins, 1986], Avida, [Lenski et al., 1999], and Tierra, [Ray, 1994], is that it starts with a completely random genome, and no further intervention is required. Given that gene duplication is common and that transcription and translation are part of the housekeeping functions of all cells, the program simulates the process of evolution of new binding sites from scratch.
It is clear that if the process in nature started with a completely random genome, that no binding sites could be evolved using the process described in the ev program because the locations of the binding sites would be unspecified. However, we might argue that in fact the binding site locations are fixed because they are the positions immediately before valid coding regions to define some functional protein. There are two problems here, which mean that it cannot be said that the ev simulation represents this case (of a pre-existing functional protein). The first objection is that the program makes no attempt to preserve any information in the exon regions.
As what was being simulated was a stretch of regulatory DNA, this is not surprising. There were no exon regions, the exon would be downstream of the entire regulatory region. As far as I can tell we are talking about one protein that binds at a number of sites at once. Starting with a more-or-less random segment of regulatory DNA (let's not confuse a full organismal genome with the short digital "genome") seems reasonable as binding is not an all-or-nothing thing in real life.
quote:
Indeed, these regions are required to evolve as well as the binding site sequences, in order to avoid false recognitions. To illustrate this situation, we performed a modified ev type simulation, where coding regions were specified, and at the outset, information was introduced into the sequences. In order to do this, we considered each coding region to be divided into "codons" of three nucleotides, but set it up so that one of 8 of the 64 possible values (initially randomly chosen), were selected for each simulated codon. Thus, the initial information content is log2(64=8) = 3 bits of information per codon. We constructed a genome of 605 nucleotides, where the coding regions consisted of 24 nucleotides, and the binding sites of 6 nucleotides. The rest of the genome consisted of the recognizer gene. During the simulation, we monitored the information content of the binding sites, and of the coding regions (ignoring for simplicity the small sample error correction used by Schneider; which does not affect the general pattern of the results). The results are shown in Figure 4. Not surprisingly (because the evaluation function of the simulation does not allow for it), the information content starts at 3 bits per codon, and rapidly decays, reaching its equilibrium level before any significant gain in information has arisen in the binding sites.
All that this shows is that if there is no selection for maintaining exon function then they will not be maintained. It would be interesting to repeat this experiment with an actual realistic exon (say with ~100 codons) attached to the end of the regulatory sequence, with only synonymous mutations in the exon not selected against. Would this really cause any significant difficulty in the evolution of binding sites in the regulatory region? I doubt it.
The point seems to be that because the simulation didn't include coding regions, one can't argue that the location of the regulatory region is specified by the coding region. But surely it would be a trivial modification of the program to add conserved coding regions (or even take an actual DNA sequence as the starting point).
quote:
The second reason why ev does not represent this situation has to do with the information calculations. In order for the Shannon information content at the binding sites to be equal to Rfrequency, there must in fact be a uniform random distribution of the input points given to the perceptron, which is a hyperplane that divides the input (hyper)space into two regions, one with 1/16th of the total volume. But the problem here is that if the distribution is uniform, then when the genome is considered as a whole, there will in fact be no information at all, because a uniform distribution of points will give the highest value of entropy possible. If there is pre-existing information in the genome, which defines where the binding sites are (just the location of the intron-ends), then we should not expect a uniform distribution of points, and hence the information content would not be expected to be equal to Rfrequency.
This seems to be a lead in to Strachan's later recalculation of Rfrequency, which ended up at ~3 bits per binding site rather than 4. It seems to me that if the main qualitative question is "can mutation and selection generate information" then differences in quantitative measures are not particularly important.
quote:
The conclusion is inescapable. The reason the information arose at the simulated binding sites is because it was put in right at the beginning. There was not just a "completely random genome" at the beginning. There was a random genome plus an externally supplied specifcation of where the sites were supposed to be.
But the "externally supplied specification" was randomly generated as well! Is Strachan telling us now telling us that you can get information from a random number generator?
What distresses me about this paper is that, despite Strachan's obvious expertise in information theory, there seems to be relatively little consideration of the differing definitions of "information" employed in the discussion. In a Shannon sense, information is basically the reduction of uncertainty, but Shannon information is premised on there being a pre-existing "message" or "signal" that defines what is actually signal and what is "noise".
We see this confusion in Strachan's discussion of METHINKS and EV (the above quotes and preceding section), where these various things get called "information":
The METHINKS and FORTY-TWO strings (a grammatical sentence)
Randomly generated sequences of codons
Randomly generated locations
In biology, though, it seems that the only thing that can be considered "signal" is function. Binding of molecules to each other is such a function -- in fact, it is one of the most general functions in biology. Many lines of evidence point to the notion that evolutionary processes can produce proteins with new and/or improved binding functions -- some forms of antibiotic resistance, wherein the antibiotic molecule is attacked by an anti-antibiotic protein are one example -- and Schneider's simulation is just an approximation of this process in silico.
Unless Strachan wants to argue that antibiotic molecules are the equivalent of the "METHINKS" target which "pre-specify" the anti-antibiotic sequence; or similarly, argue that "high temperature" is the target which "pre-specifies" the adaptations of microbes to high temperature, then it seems to me that he's going to have to admit that natural processes can create new information and do so regularly.
IP: Logged
|
|
yersinia
Member
Member # 324
|
posted 29. June 2003 15:33
Here's a puzzle for folks:
Consider a simulation which begins with two random letter sequences:
LHBKDFPOEIWHSDSDKJFHSDKJEG...
WEKJHHSDJFNWEOIUHDSJEKNWER...
Let's say we mutate each sequence and retain only those mutations that make the sequences closer to each other.
1) Does anyone doubt that these two sequences will rapidly converge to the same sequence (intermediate between the two originals).
2) Has "information" been generated in this process?
(before you answer, consider the coevolution of binding between two proteins, where the specific shape of the binding region is more-or-less irrelevant as long as the shapes are complimentary).
IP: Logged
|
|
Pim van Meurs
Member
Member # 541
|
posted 29. June 2003 17:13
As I stated clearly in my deleted posting (it would be nice if deleted postings were sent back to the submitter for resubmission since significant effort went into its formulation), my comments were based on a quick browsing of the paper and the abstract.
More later.
My comment was with respect to Iain's suggestion that Ev is similar to Dawkins in that it looks at a fixed target.
" all the simulation is demonstrating is a form of “supervised learning” of a fixed target by a simple form of neural network."
As I pointed out Schneider is quite clear that the recognizer and the binding sites co-evolve which make any claims to similarity with Dawkins' Weasel program quite tentative.
Additionally Iain seems to claim that the information is front loaded through definition of the location of the binding sites but if that were the case then would the information not evolve whether or not selection were present? In fact Schneider has shown how selection is essential for information to evolve.
More later but these two claims caught my eye and deserve some additional comments.
I have another question for Iain: He observes that in order to calculate Rseq, for small sequences one has to use the full formula for entropy. My question is: Is this relevant to the finding? Would Rseq and Rfreq not be equally affected by this correction?
===============
Yersinia, thanks for your comments. Some of your feedback mirrors my initial feelings about the paper namely that information did seem to increase in the genome although Iain argued not necessarily as much as predicted (I have my questions about this observation though).
Iain seems to suggest however that the information was 'pre-loaded' in the random selection of the locations of the binding sites. As Yersinia has argued, similar selections happen in nature so it may be that selection in nature may after all be the relevant factor in increasing the information in the genome. And in fact that is what we observe: remove the selection and no information increase happens.
As Yersinia I would be interested in a more thorough exploration of the claim that "the Dawkins simulation is simply a special case of the Schneider simulation" and an exploration of the relevance of this. After all a time-invariant solution is merely a special case of the time-varying solution but that does not mean that the arguments against a steady case solution necessarily apply to the more general solution. In fact the main argument against Dawkins' Weasel seem to be that the target is pre-selected. But in Schneider's case the target is not pre-selected but co-evolves.
From the paper
quote:
Note that f (.) is a many-to-one mapping, allowing multiple outcomes, whereas the Dawkins simulation is a one-to-one mapping, with a fixed outcome. However, this does not alter the fact that a fixed target has to be specified in order to make the simulation work.
This seems to be self contradictory. A fixed target versus allowing multiple outcomes. In fact the distribution of f(.) in case of Schneider seems to be Gaussian distributed around Rseq. It makes this somewhat hard to argue that there is a fixed target.
[ 29. June 2003, 21:29: Message edited by: Pim van Meurs ]
IP: Logged
|
|
Moderator
Administrator
Member # 1
|
posted 29. June 2003 20:20
Pim, check your private messages. Your comments were sent back to you upon deletion.
Also, papers submitted to ISCID typically get discussed in Brainstorms. This is our policy, as we feel that articles submitted to and accepted by our society will be of interest to Brainstorms participants.
[ 29. June 2003, 20:24: Message edited by: Moderator ]
IP: Logged
|
|
|
|
Erik
Member
Member # 160
|
posted 04. July 2003 19:37
Micah Sparacio, may I suggest that you move your announcement of your GA (and with this reply) to a new thread? I think Strachan's argument is flawed, but it represents the most sophisticated analysis by an ID advocate to date, so I think it deserves its own thread. (<--- I had not read far enough into Strachan's paper to see the appendix, and apparently missed the brief mention of the simulation in the first half of the main text. I apologize for this mistake.)
As for your GA, your list of hypothetical claims and your responses, here are a few points:
1. You need to define "target" more precisely. I can think of several interpretations that should not be conflated. For instance, let f(g) (aka "the fitness function") be the expected number of offspring that the genotype g will generate (possibly relative to the population's mean fitness). Let u(g) denote the impressiveness of the genotype g according to the GA programmer (or someone else external to the simulated world, e.g. you). Let g0 be a particularly interesting genome at the end of the GA run. Here's a few guesses at what you may mean by "target":
(i) The set of genotypes having at least as high fitness as g0, i.e. the target is {g | f(g) >= f(g0)}.
(ii) The set of genotypes having at least as high fitness as some predetermined cut-off f0, i.e. the target is {g | f(g) >= f0}.
(iii) The set of genotypes having at least as high impressiveness as g0, i.e. the target is {g | u(g) >= f(g0)}.
(iv) The set of genotypes having at least as high impressiveness as some predetermined cut-off u0, i.e. the target is {g | u(g) >= u0}.
(v) The region of genotype space that the GA programmer wanted the GA to reach.
Let me know if any of these guesses is correct or what you actually mean by "target".
2. Concerning irreducible complexity, you need to make clear what the IC Definition of the Week is. For instance, there was a week when you insisted that IC systems must be composed of mechanical parts. With this definition the hypothetical claim is trivially refuted: No mechanical parts have been identified, so the evolved system has not been shown to be IC. The fact that you didn't mention this leads me to believe that you prefer a different IC definition this week (perhaps "a genome (sub)sequence is IC if there couldn't have been any selection pressure to evolve it"?).
Erik
PS. I think a simpler, but equivalent for the purposes of this discussion, example would be the evolution of a genome that is a palindrome, given a fitness function that is some increasing function of the Hamming distance between the first and the reversed second part.
It would be potentially interesting to discuss a toy model of coevolution where there are two species and the reproductive successes of genotypes depends on the Hamming distance to genotypes in the other species (analogy: one species could, e.g., be a toxic plant and the other species could be a plant-eater, whose resistance to the toxic coevolves with the plant's toxic). In such a model, the fitness of a genotype will depend on the ecology (i.e. the genotypes of the other species). DS.
------------------
Added in edit: The italic parenthesis above.
[ 05. July 2003, 12:40: Message edited by: Erik ]
IP: Logged
|
|
Micah Sparacio
Member
Member # 6
|
posted 05. July 2003 08:40
Erik,
May I suggest that you read through Strachan's paper to learn whose GA I posted.
Hint: it was not my "list of hypothetical claims and [] responses"
[ 05. July 2003, 08:43: Message edited by: Micah Sparacio ]
IP: Logged
|
|
Erik
Member
Member # 160
|
posted 05. July 2003 13:06
Micah Sparacio (and I.G.D. Strachan), sorry about not noticing that the new Java simulation was an implementation of something mentioned in the article.
Whether or not the my comments should be addressed to Sparacio or Strachan, they still remain. No IC definition of the week has been chosen and the meaning and significance of targets (or lack thereof) is unclear. For instance, since the point of simulations like Schneider's is to investigate the transformation of a particular initial state into a particular final state, I'm not sure how the concept of a target is useful for the discussion. Furthermore, it is unclear if it is logically possible to make a reasonable computer simulation that doesn't have a target (does a computer simulation of the weather have a target?). I suggest that Strachan's concerns would be better formulated and discussed in terms of the (il)legitimacy in using a fitness function that is highly correlated with some feature we find impressive, and the biological analogues of such choices of fitness functions.
Erik
IP: Logged
|
|
Paul A. Nelson
Member
Member # 26
|
posted 05. July 2003 13:27
Erik wrote:
quote:
I suggest that Strachan's concerns would be better formulated and discussed in terms of the (il)legitimacy in using a fitness function that is highly correlated with some feature we find impressive, and the biological analogues of such choices of fitness functions.
I'm making my way through Strachan's paper, and agree that this is probably the central issue. [Incidentally, the biological analogues of several aspects of the Lenski et al. simulation also provide an issue worth discussing, a thread I hope to start later this month.]
Since you're currently thinking about all this, Erik, could you say what you take the biological analogue of (for instance) sitelocations to be? (See page 7 of Strachan's article.)
Thanks.
[ 05. July 2003, 13:32: Message edited by: Paul A. Nelson ]
IP: Logged
|
|
Pim van Meurs
Member
Member # 541
|
posted 05. July 2003 15:06
Micah, thank you for posting the links to the simulator. I have some questions about the relevance of this simulation to the work of Schneider since in Schneider the binding sites are not a 'fixed target' nor is the recognizer. In fact the two co-evolve rather than being pre-specified.
Thus when it is claimed that
quote:
The difference is that here, we are quite explicit in stating up front that there is a fixed target string (which in the simulation is the phrase "THE ANSWER IS FORTY TWO"), whereas in both the Schneider and Lenski et al simulations, the existence of the fixed target has to be deduced from the text (and examining the source code of the Schneider simulation).
I believe that the author is confused about Schneider and Lenski since there is no attempt to pre-specify the target as is done in this simplified example which is merely an extension of Dawkin's weasel but does not capture the essence of either Schneider or Lenski.
This can easily be established by recognizing that repeated runs will give different 'answers' unlike the simulation presented here.
On the other hand, if I understand Ev correctly:
quote:
The ev model can also be used to succinctly address two other creationist arguments. First, the recognizer gene and its binding sites co-evolve, so they become dependent on each other and destructive mutations in either immediately lead to elimination of the organism
Similarly in the paper it is stated that
quote:
Mutations are applied to all positions in the genome, so the binding sites and the weight matrix co-evolve.
This seems to be significantly different from a Weasel or Weasel-like approach.
[ 05. July 2003, 16:34: Message edited by: Pim van Meurs ]
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 05. July 2003 19:12
Introduction
The author of the description of the Web VETE program claims in effect that VETE is a model of two recent evolutionary simulations, Tom Schneider's ev program and the Avida simulation used in Lenski, et al., discussed in the Literature Review thread. That is a strong claim, and it is false. The Web VETE simulation implements a program provided by Strachan in his critique of Schneider's ev, but the author of the VETE description generalizes it to a model (and critique) of the Lenski, et al. Avida simulation as well. Here I will discuss only its applicability (or really, its inapplicability) to the Avida simulation, since I'm much more familiar with Avida than with ev.
The sole property of the Lenski, et al., simulation that VETE models is the presence of an experimenter-defined fitness function. And that is all. In no other relevant respect does the VETE simulation model the Avida simulation.
The author of the VETE Web description (hereinafter "Author") asserts that it has been claimed that the Avida simulation has no fixed target. quote:
There have been some more recent attempts at evolutionary simulation, notably the Schneider 'EV' simulation, and also the Lenski et al 'AVIDA' Nature paper. In both of these simulations, there is, it is claimed, no fixed target to aim at; indeed, there are many possible different outcomes and hence, it appears, they overcome Dawkins' objection to his own model that Evolution has no long-term goal.
However, in both these cases, all the authors have succeeded in doing it to push the Dawkins objection one step further back. Both algorithms in fact still aim to reproduce a distant ideal target (in Dawkins' words), but with a subtle difference. The target is not the genome that arises at the end of the simulation, but the result of applying a processing algorithm to the final genome.
No references are given, so one doesn't know the source of the "it is claimed" assertion about no fixed targets to aim at. It is the case that the Lenski, et al., simulation had no fixed target in the sense of having no pre-specified genome as the desired product of the evolutionary runs. However, it did have a fitness function that allowed the evolutionary operators (mutations of various kinds) to induce fitness landscapes characterized by varying topography rather than a flat surface. Whether that is some sort of fatal flaw in the Avida simulation will be discussed below. I will note here only that the so-called "subtle difference" is in fact a very important difference.
Background
In the Dawkins WEASEL program with which Author begins his exposition, there was an output string (the 'WEASEL' sentence) that was the declared target of selection (not "evolution"!). As Dawkins explicitly noted in his original description, that program was not intended to be a model of evolution. It was explicitly intended only to demonstrate the difference between cumulative selection and single step selection. Dawkins himself made that very clear in what Author, following Strachan, strangely calls Dawkins' "criticism" of his own program. That wasn't a criticism; it was an effort to make clear just what was and was not being demonstrated by the WEASEL illustration. That his effort to make that clear wasn't successful is evidenced by the various creationist criticisms of it on precisely the grounds that Dawkins described, that it wasn't a demonstration of evolution!
Author then describes the Lenski, et al., work in a section strangely titled "Recent attempts to improve on the WEASEL model." The implication is that the Lenski, et al., work is somehow descended from the WEASEL program. Nothing like it: Avida is descended from a line of research on evolutionary simulations in which Dawkins' WEASEL program (either in particular or as a generic representative) plays no significant role whatsoever. The line of descent of Avida can be traced back to Holland's work on classifier systems in the mid-1970s through the genetic programming work of Koza in the early 1990s and Ray's Tierra work also in the 1990s. From the point of view of the development of evolutionary algorithms, Dawkins' WEASEL demonstration of cumulative selection is a tiny side path of that line of research.
The Lenski, et al., simulation used Avida, an artificial life research platform. Let me first deal with some inaccuracies, misconceptions, and plain errors in Author's description of the Avida simulation.
First, Author says quote:
In the Lenski et al simulation, a string of letters (1 of 26) is evolved, and then each of the 26 possible codes is treated as a machine instruction for a "virtual machine", running a language called AVIDA. The program is required to output certain logic functions (the most complex being EQU; a test for the equality of two 32-bit inputs). Graded rewards are given for the complexity of logic function computed, with EQU having the highest reward because it is the most complex.
That is mistaken in two respects. First, the programs (plural; 3,200 individual assembly language programs evolved in parallel in the Avida simulation) were not required to output anything at all. Each evolution run in Avida starts with a population of 3,200 identical replicators, 3,200 assembly language programs that could replicate themselves but could do nothing else. Lineages of replicators in Avida compete with each other for 'living space' on a matrix. More efficient replicators - those that could replicate themselves faster than their brethren - succeed in producing more copies of themselves than slower replicators and thus their lineages gain living space. Nothing else is "required" of them. It is the case that the selective environment differentially rewarded - provided extra reproductive resources - for mapping some input patterns into output patterns, but that was not "required" of the evolving programs. That resource was available to be exploited if the replicators could evolve to do so, but it was not required of them.
Second, the programs - Avida replicators - do not "output" logic functions. They can gain replication resources by performing logic functions by mapping inputs into outputs. By appropriately operating on environmental stimuli - bit strings - and producing appropriate bit strings as outputs in response, the Avida replicators could acquire additional replication resources and could therefore replicate faster than their cousins that could not perform those operations. That is, evolving on a non-uniform fitness landscape, the Avida replicators gained in reproductive fitness by performing appropriate mappings of environmental inputs into outputs.
The experimenters, of course, provided the external fitness function that (in conjunction with the evolutionary operators - various sorts of mutations) induced the fitness landscapes on which the population of replicators evolved. The basic question of the Lenski, et al., paper was, 'Given a complex behavior (a complicated input to output mapping) that is "fit" (gains reproductive resources) in a selective environment, can ordinary evolutionary processes start from scratch with a population of replicators that can do nothing but replicate, and produce lineages of replicators that can perform that behavior?'
Continuing the quotation above, Author says quote:
This [Avida] simulation still corresponds to having a fixed target; in order to determine survival likelihood of a digital creature, each program is tested with a given set of 32 bit inputs, and then the 32 bit output is compared with a fixed set of required outputs (i.e. the EQU function had to be pre-calculated externally to the simulation, and it was then used for a target). In fact they had to do a lot more than that, in offering intermediate rewards to logical functions that were less complex to compute.
Again, there are two misconceptions in that description. First, a given digital creature's "survival likelihood" was unaffected by whether it could or could not perform a logic function. Its reproductive fitness was affected, but not its survival. In Avida, lineages evolve as some kinds of replicators become faster reproducers than others, but no digital creature lives or dies because it cannot perform a logic function. Individual "survival" based on fitness is not a component of the Avida simulation.
Second, it is obvious bordering on trivial that the selective environment embodied in the Avida simulation evaluates the replicators' performances of the various logic functions in the fitness function. That is what selective environments do! They 'evaluate' the behaviors of organisms, differentially rewarding and punishing them in reproductive advantage terms. The simulation accurately represents what selective environments do. And that fitness landscape embodied the experimental question: 'Can the population of replicators evolve to perform the input-output mapping corresponding to performing EQU?' So in some sense Author is correct: the experimenters built into the simulation shell the 'knowledge' of correct performance of EQU and the other logic functions. However, that 'knowledge' was built into the topography of the selective environment in order to study an experimental question as part of the experimental design. (See below for more on "built into the environment.")
Third, the experimenters didn't "have to" provide intermediate rewards. That again was part of the question at issue: Given intermediates, could a population of replicators evolve to perform the input-output mapping corresponding to performing EQU? In fact, Lenski, et al., ran a series of control conditions to determine whether intermediates are really necessary, and if so, whether particular intermediates are necessary. Those control conditions required a much larger number of evolutionary runs than the main experimental condition of 50 runs. As it turned out, having some intermediates was necessary for evolutionary runs to produce replicators that could perform the EQU input-output mapping, six (of the main experiment's eight) intermediates anyway, but no specific intermediate or specific pair of intermediates was necessary.
Author concludes quote:
But in summary, both the Ev and the AVIDA simulations require a fixed target to be specified prior to the evolution, and the target is to be the outcome of applying a computer algorithm to the genome. Both are claimed as being "target free" in that there are many possible outcomes that will work, but both are in fact tied to fixed targets.
This paragraph contains an interesting misconception. "Target free" is in quotation marks, but no reference is given. I am not aware of any claim that there were no "targets" in the Lenski, et al., work in the sense of having a flat fitness landscape. The actual claim (and it is correct) is that no specific kind of replicator was the target of evolution. No particular way of performing the various input-output mappings corresponding to the several logic functions was differentially rewarded with additional reproductive resources. What gained replicators reproductive advantage was performing those mappings somehow or other. But obviously (and trivially) had there been no fitness function there would have been no evolution to perform those mappings. (However, there would have been evolution: One can run Avida with no externally-defined fitness function - equivalent to a flat external fitness landscape - and watch lineages of replicators compete for living space and evolve based on the fact that space in the Avida world is limited and is therefore a scarce resource. Under those circumstances lineages of replicators evolve by 'tuning' their initial replication code to become more efficient.)
So, the Avida simulation has a "target," reproduction rate, and replicators gain reproductive advantage and thereby increase their reproduction rate by performing input-output mappings that are rewarded by the selective environment. If that is what is meant by a "target" in Author's description of the Lenski, et. al, study, then I proclaim that, yes indeed, there was a "target."
But we see that there are a number of misconceptions and plain errors in Author's description of the Lenski, et al., Avida platform and research design, and those misconceptions and errors invalidate the claim that the VETE simulation models the Avida work in any useful way.
The VETE Program
The structure of the VETE simulation is quite different from the Avida platform's structure. In VETE, a genome consists of an "encryption key" (a string of letters symbolizing variable letter shifts in a modified Caesar code) and a "text string" that represents an encrypted string. The two coevolve, in the sense that both are subject to mutation and selection - both are mutated, and selection is on the basis of their joint behavior. The "target" of the coevolving strings is a combination of an encryption key and text string that produces as output "THE ANSWER IS FORTY TWO." (Thank you, Douglas Adams. ![[Smile]](smile.gif) )
Both strings are initially randomized, and both are randomly mutated in the course of a run. Selection (from a population of progeny of unknown size) is on the basis of similarity of the target string ("THE ANSWER IS FORTY TWO") to the result of applying the key to the text string. As I read the description, in each 'round' or 'generation,' a single initial candidate is repeatedly mutated (with point mutations?) some unknown number of times, and the set of "progeny" thereby generated is compared with the target string. The one 'offspring' most similar to the target string then serves as the basis for a subsequent round of mutations and selection. This is very similar to Dawkins' Biomorph program except that in the latter there is no fixed target string. (Added in late edit: The program's structure indicates VETE's roots in applied work on GAs as search algorithms rather than in biological modeling.)
So in the VETE simulation there is no population. Or more correctly, there is a series of temporary populations of unknown size. A new temporary population is generated at each step of the program, each temporary population the result of a round of mutations of a single founder individual. Once a round of mutations produces a temporary population, the population immediately goes extinct except for one individual that enters the next round of mutation as its founder.
Given that each 'evolutionary' run in VETE is initialized with a different randomization of the key and text, the exact key and text strings that are produced by the simulation are virtually certain to be different from run to run. Yet all produce the same output when key is applied to text: "THE ANSWER IS FORTY TWO." So there is a many-one mapping from key/text strings to output string.
Local Targets, Global Targets, and Selective Environments
Once again the claim of Author is that the Lenski, et al., simulation was not "target free," and that is correct in the sense that evolution occurred in an experimenter-defined selective environment in which some behaviors - some input-output mappings - gained additional reproductive resources for the replicators that performed them. It's analogous to a biological critter wandering around in the real world. If it appropriately maps inputs into outputs - on sight of food, approach; on sight of predator, flee; on sight of potential mating partner, approach; on sight of cliff, retreat - it is more likely to reproduce and its lineage is therefore more likely to persist through time. Appropriate input-output mappings provide a replicator with additional reproductive resources relative to its brethren that do not perform the appropriate mappings. Once again, the essence of the Avida simulation is competition among lineages for limited living space. Replicators that gain additional reproductive resources out-replicate their brethren and their lineage prospers.
Where does selection come from? In biology, it comes from the environment, where "environment" includes other members of the population of which a replicator is a member (its most formidable competitors); other species that may be prey, predators, parasites, or competitors for the same resources; and the non-biological physical environment. Jointly, those three sets of factors (along with the particular evolutionary operators available to the population) determine the set of fitness landscapes on which a biological critter and the population of which it is a member must operate. Similarly, in Avida the selective environment is the other lineages in a replicator's immediate vicinity and (optionally) an external environment provided by the experimenter. Jointly, those two sets of factors (with the particular evolutionary operators available) determine the set of fitness landscapes on which a digital critter and the lineage of which it is a member must operate.
In evolutionary simulations, selective environments come from experimenters. The simulations are run in order to test hypotheses of interest. In order to test hypotheses, experimenters must design environments appropriate to the research question under study. Thus in the Lenski, et al., research the question was 'Given a complex behavior (a complex input-output mapping), can ordinary evolutionary processes start from scratch with a population of replicators that can do nothing but replicate, and produce replicators that can perform that behavior?' Given that question, they designed 38 selective environments for a study in which the question could be meaningfully asked and answered.
Someone in the Literature Review thread dealing with the Lenski, et al., study wondered what would have happened had they not "rewarded" performance of the mapping appropriate to performing EQU. The answer is trivial: Avida digital critters that perform EQU would not have evolved. But that question signaled a profound misconception. Selection, and a fortiori, a selective environment, is an integral part of the evolutionary process. No selection, no evolution. What a surprise!
Does the fact that experimenters, including Author, set a distant (from the initial conditions) "target" for an evolutionary simulation invalidate the assertion that in biology, evolution proceeds without long-term goals? No. The Avida digital organisms are not 'striving' toward any goal; they are not required to perform any particular input-output mapping in order to replicate. They are merely differentially reproducing at a rate that is a function of the reproductive resources that they have managed to accumulate from behaving in their environment. And when they do things that their selective environment rewards with reproductive advantage, they out-reproduce their brethren. But at every single time slice they are responding to their immediate environment; they know nothing of long-term "targets" or distant goals. In our experimenter-as-God role we know that the selective environment will reward this or that behavior, but the replicators don't. All they 'know' is what they can do in the immediate context.
Evolution is unaware of distant targets: it is local, undirected, and blind. That we know something about the fitness landscape's topography says nothing about what the evolving critters 'know.' In the Avida simulation the evolving replicators know nothing of targets or selective environments or fitness landscapes. They are operating purely locally, reproducing at variable rates. If the fitness function rewards some behaviors and punishes others, then those local rewards and punishments will change the genetic composition of the population over generations, with some lineages prospering and some faltering. (Providing, of course, that there is a population to alter, which is not the case in VETE.) Most biological lineages falter sooner or later: to a first approximation all species are extinct: the vast majority of biological species (and Avida lineages) go extinct.
The illegitimacy of the VETE demonstration as some sort of refutation of the Lenski, et al., study is due to its definition of fitness according to a global criterion (similarity to a distant goal) as opposed to local criteria (local topography of a fitness landscape). That is a comparison that neither biological evolution nor Avida make. The Avida shell evaluates a replicator's performance of the several logic functions, implementing the selective environment, but reproductive resources are not awarded on the basis of an explicit evaluation of the graded similarity of the replicator's output to the distant EQU peak. Reproductive resources are awarded on the basis of what the replicator can currently do, and not on a finely graded judgement of how similar its current behavior is to some distant target.
The last assertion is not strictly true. While replicators were not explicitly graded for similarity to some distant target, the various intermediate logic functions differed in reward values according to their complexity, where "complexity" was defined in terms of the minimum number of nand operations necessary to perform them, from 1 to 4. Since performing EQU requires a minimum of 5 nand operations, in a sense there was a graded evaluation based on a loose index of similarity to a distant goal, where "similarity" means "uses more of the same constituents." In the Literature Review thread Micah asked what the most significant weakness of the Lenski, et al., study was. I regard that as the most significant question one might ask about the Lenski, et al., simulation, and when we get Version 1.6 of Avida running on the Beowulf cluster here I will test that evaluation function. The test question will be 'What kinds of fitness landscapes are conducive to evolving digital organisms capable of performing the EQU input-output mapping? Do those fitness landscapes veridically model those in biology?'
However, the general proposition that different intermediates confer differing reproductive advantage is not in itself implausible. Which particular complex functions can be 'reached' by an evolutionary process may be constrained by those differences among intermediates that are themselves reproductively fit in their immediate circumstances. In particular, fitness landscapes in nature are correlated in their several dimensions, with sloping multi-dimensional hills and ridges. Natural fitness landscapes are not random, and hence there will be peaks and valleys and ridges, and 'distant' targets will sometimes be accessible via intermediates. Different intermediates 'lead to' - allow reaching - different peaks, and hence some constraints due to the availability of intermediates are implied. But the evaluation of fitness is purely local in the sense that similarity to distant target is not in itself sufficient to garner additional reproductive resources. Further, the control runs in the Lenski, et al. paper demonstrated that the 4-nand intermediates are not necessary to evolve lineages that perform 5-nand EQU, so there's at least some mitigation of the criticism deriving from the misconception of a simple-minded graded similarity-to-distant-peak fitness evaluation.
Nevertheless, the existence of some sort of intermediates is clearly necessary to evolve complicated structures and processes, and that's also true of organic evolution. To criticize the Avida simulation on the ground that it was possible to evolve EQU-performing lineages only because reproductively advantageous intermediates exist is tantamount to arguing that intermediates are never available in nature and that natural fitness landscapes are uniformly flat in all their dimensions. As long as fitness is evaluated solely in terms of immediate local reproductive success, the mere existence of distant peaks on an experimental fitness landscape does not in itself invalidate a simulation.
(I also note that the ev program apparently did explicitly evaluate fitness according to graded similarity to a distant target, and if that's the case then at least part of the criticism in the Strachan paper may be valid, though I caution again that I am not intimately familiar with the ev simulation.)
Concluding Remarks
Any selective environment that is not absolutely featureless necessarily rewards some behaviors or traits and punishes others, where rewards and punishments are denominated in relative reproductive resources and where fitness evaluations are local to the immediate context. The process is precisely the same whether a fitness landscape is constructed by an experimenter to test a specific research question or is given by nature in the form of other organisms and the physical environment. The process is the same.
Hence one (and only one) of Author's claims about the Avida simulation is correct: quote:
With this idea in mind, we present here a much simpler demonstration that is exactly the same in this respect to both the Schneider and Lenski et al simulations.
"..in this respect" refers to the fact that the experimenters designed the fitness function so that it was not flat. Period. In no other relevant respect is the VETE simulation similar to the Lenski, et al., simulation.
With respect to the Avida simulation, the VETE simulation is aimed at a straw man. No one I am aware of has claimed that the Lenski, et al., simulation was "target free" in the sense of having a uniformly flat fitness landscape. What has been claimed is that no specific "target" genotype was defined as adaptive, and no particular evolutionary route through the space defined by the fitness function was mapped out beforehand. And both of those claims are true.
Author freely and inappropriately mixes two levels of analysis: Experimenter-as-God and population-evolving-locally-on-landscape. That the Experimenter-as-God knows (and perhaps designed) the topography of a fitness landscape does not necessarily tell one anything at all about how a population-evolving-locally-on-landscape behaves on that landscape.
In particular, the claim that the Experimenter-as-God's knowledge that there is a distant fitness peak somehow determines the behavior of a population-evolving-locally-on-landscape is tantamount to the claim that evolutionary processes are capable of traversing the kind of landscape that was modeled, since evolutionary processes are all that operate within the Avida simulation. Lenski, et al.'s question was 'Can populations evolve to perform complex input-output mappings in this kind of selective environment" and the answer was 'Yes.' Experimenters design experimental conditions to test hypotheses, and in testing evolutionary hypotheses in computer simulations, designing the selective environment is part of the experimental design.
What the VETE simulation shows is that a program very loosely modeled on biological evolution can solve a complicated puzzle. It also shows that such a program can evolve (again, very loosely modeling biology) many different solutions to a given problem. But VETE does not model the Avida simulation in critical respects nor does it model biological evolution, and it therefore has little relevance to either.
In Strachan's (page 6) terms, evolution, whether biological or in the Lenski, et al. simulation, is a form of "supervised learning," where 'supervision' is provided by the selective environment created by the dynamic interaction of other organisms and the non-organic physical environment. The selective environment's origin - whether it comes from a researcher designing a simulation or from a combination of the naturally occurring physical and biological environments - is irrelevant to the evolutionary process. Evolution occurs when populations of imperfect replicators with heritable variation inhabit a selective environment that is not constant, on a fitness landscape that is not uniformly flat. By discovering that a differentiated selective environment is necessary in order for evolution to occur, Author has finally caught up with Darwin.
RBH
Edited for a plethora of typos.
Erratum: I said above that the population in the Avida simulation was 3,200. That's wrong. It was 3,600 - the Lenski, et al., simulation used a 60x60 matrix.
[ 06. July 2003, 23:22: Message edited by: RBH ]
IP: Logged
|
|
Pim van Meurs
Member
Member # 541
|
posted 05. July 2003 22:52
Thanks RBH,
Your expertise in artificial intelligence has proven itself once again. I wish I could have expressed my ideas as clearly and to the point as you have. Truely impressive
Peter, I am looking forward to your simulations to determine the validity of your suggestions.
Thanks
Pim
[ 06. July 2003, 00:59: Message edited by: Pim van Meurs ]
IP: Logged
|
|
peter borger
Member
Member # 722
|
posted 06. July 2003 00:20
If I understand it properly Avida's target is an increase of reproduction rate. In colonies of coexisting species the fastest reproducer will always outcompete the slower reproducers. The whole point is that reproductive succes has been linked to increased complexity. The assumption here is thus that complexity is directly associated with reproductive succes. It is however nothing but an assumption and what we observe in biology is exactly the opposite. As mentioned in another thread bacteria that increase their genomes by duplicating fractions of it require more energy and resource (and thus time) to duplicate their genetic content and will be selected against. Unless selective constraint is permanently present to keep such duplications in the genome. That is a law of nature. This law of nature is demonstrated in cocultures of bacteria, in cancer, in organisms that become antibiotic resistant, etcetera.
I wonder what happens if you start with a higly complex cyberorganism (algorithm) that looses all complex information that is not immediately required for reproductive succes and is linked to increasde reproductive succes? (this is the realistic scenario, see examples above). The fastest (less complex) replicator wins through getting rid of embellishment, would be my guess. Evolution or devolution, that's the question.
Selection is always on the level of reproduction. Not on complexity. It's a biological law.
Best wishes,
PB
IP: Logged
|
|
|
Printer-friendly view of this topic