|
Author
|
Topic: Nature Refutes ID?: The Evolutionary Origin of Complex Features
|
Roger R
Member
Member # 200
|
posted 15. May 2003 21:11
charlie d. writes,
quote:
Genome length (RR's objection): Clearly program length in the AVIDA simulation makes a big difference in terms of computation cycles, and access to computation cycles is how fitness is rewarded in AVIDA - therefore, a compensation for the loss in efficiency of longer programs was necessary (I guess the alternative would have been to simply give an even bigger reward to more effective programs, but the result would have been the same). In nature, however, very little if any disadvantage is generally associated with increased genome complexity (up to a certain limit at least - but that's pretty high) thus no "leveling of the playing field" for genome complexity is necessary in real-world evolution.
I'm not sure on what basis you think such a compensation was "necessary". I'll admit that it will get them to the goal quicker, but there has never been a question in my mind, or that of most IDers, that such goals are attainable in a GA given the proper tuning by an intelligent agent. Also, why is the alternative to give even a bigger reward to the more effective programs, rather than just dropping a reward for larger genomes? And I am also skeptical of your conclusion that rewarding non-advantageous traits in addition to advantageous ones, produces the same outcome as rewarding only advantageous ones at a higher level. I'll have to chew over the implications of that one.
As far as the cost of larger genomes, I think you are missing the point that with these digital organisms, the genome is essentially the entire organism, so I think you are ignoring the costs in the real world of adding significantly more organism w/o any advantageous traits.
As for your assertion that the authors indeed did claim that the resulting function was IC, I have to disagree. You cited this statement:
quote:
Although the complex feature first appeared as the immediate result of only one or two mutations, its function invariably depended on many instructions that had previously evolved to perform other functions, such that their removal would eliminate the new feature.
I take this to be saying that, for example, the AND operation evolved, and that it was necessary to the final EQU function, and that removing the AND removes the EQU function. That is true, but that doesn't imply that EQU is IC, because the AND operation is selected for itself, and so are the remaining OR and NOT functions. Again, how does this map to the biological world? Some here want to analogize EQU to bacterial motility. I think a more appropriate analogy is logical operation for bacterial motility, and in that case, the knock out leaves other "complex" (though less complex than EQU) logical operations still functioning. Hence, not IC.
As far as your reasoning for why IC is not mentioned in the Nature article, I find both rather puzzling. First, if as you and others contend, this is evidence for Darwinian methods of resolving IC, then wouldn't the appropriate venue be a claim in a peer reviewed journal? Yet the claim doesn't appear. How then can the important peer review process evaluate such claims? Even if they aren't familiar with the nomenclature, an attempt to explain it would seem to be in order if the study is as significant as you claim.
Your second reason seems no more legitimate than the first: b) that given its definitional ambiguities and inherent contradictions (as seen in this thread), IC is currently of very limited, if any, applicability in biology.
No, how am I supposed to view this in light of this later statement: I agree that this discussion will go nowhere until somebody who disagrees with the paper's conclusions takes on the task of doing an objective, systematic analysis of the EQU functions; it's actually not that complicated. However, I do understand the reluctance.
How am I supposed to reach a conclusion about an issue so filled with "definitional ambiguities" no matter how much time I spend studying the EQU functions? Indeed, how could the study even hope to settle the matter, as you and RBH claim it does? Goalpost moving seems to be an equal opportunity pastime.
Not to mention, the fact is the paper doesn't make any claims vis a vis IC, so I really don't have a major beef with the paper, but with some of the conclusions being drawn from folks here about. And those differences can only be handled via a dialogue between the disputing parties.
IP: Logged
|
|
charlie d.
Member
Member # 159
|
posted 15. May 2003 21:57
quote:
I'm not sure on what basis you think such a compensation was "necessary". I'll admit that it will get them to the goal quicker, but there has never been a question in my mind, or that of most IDers, that such goals are attainable in a GA given the proper tuning by an intelligent agent.
No, it's not necessary because it gets to the goal quicker, but because for the programs, unlike living organisms, legthening the genome is a major drawback. quote:
And I am also skeptical of your conclusion that rewarding non-advantageous traits in addition to advantageous ones, produces the same outcome as rewarding only advantageous ones at a higher level. I'll have to chew over the implications of that one.
That's not what I said. Non-advantageous traits (I imagine you mean genome length) are not rewarded. Simply, a compensation mechanism is introduced so that genome length is neutral in the simulation, as is in nature. quote:
As far as the cost of larger genomes, I think you are missing the point that with these digital organisms, the genome is essentially the entire organism, so I think you are ignoring the costs in the real world of adding significantly more organism w/o any advantageous traits.
Actually, in the simulation the phenotype is the logical operations the programs can carry out. The selective pressure is applied by the replication reward programs get if they perform certain logical operations. But even according to your analogy, in biology there are advantages and disadvantages to increased body mass - for instance, it takes more energy to grow and move around, but it also protects against predators, decreases heat loss and in some cases increases mating chances. Organisms would not evolve larger sizes if it weren't advantageous, all things considered. In AVIDA, however, genome length is a net drawback. quote:
As for your assertion that the authors indeed did claim that the resulting function was IC, I have to disagree. You cited this statement:
quote:
--------------------------------------------------Although the complex feature first appeared as the immediate result of only one or two mutations, its function invariably depended on many instructions that had previously evolved to perform other functions, such that their removal would eliminate the new feature.
-------------------------------------------------
I take this to be saying that, for example, the AND operation evolved, and that it was necessary to the final EQU function, and that removing the AND removes the EQU function. That is true, but that doesn't imply that EQU is IC, because the AND operation is selected for itself, and so are the remaining OR and NOT functions. Again, how does this map to the biological world? [quote]And in the flagellum, its subcomponents may have evolved as secretory systems, attachment devices, pores, channels, etc. That is precisely the point. [quote]Some here want to analogize EQU to bacterial motility. I think a more appropriate analogy is logical operation for bacterial motility, and in that case, the knock out leaves other "complex" (though less complex than EQU) logical operations still functioning. Hence, not IC.
So, you are one of the proponents of the "IC is IC if it cannot have evolved" definition. Fine, but it's a tautology to then claim that IC is special because it can't evolve. quote:
As far as your reasoning for why IC is not mentioned in the Nature article, I find both rather puzzling. First, if as you and others contend, this is evidence for Darwinian methods of resolving IC, then wouldn't the appropriate venue be a claim in a peer reviewed journal? Yet the claim doesn't appear. How then can the important peer review process evaluate such claims? Even if they aren't familiar with the nomenclature, an attempt to explain it would seem to be in order if the study is as significant as you claim.
They did. The problem of complexity predates Behe's IC definiton(s) by a long long time, Darwin himself discussed it. That's what the authors are addressing. IC is a meaningless term to >99% of Nature readers. quote:
Your second reason seems no more legitimate than the first: b) that given its definitional ambiguities and inherent contradictions (as seen in this thread), IC is currently of very limited, if any, applicability in biology.
No, how am I supposed to view this in light of this later statement: I agree that this discussion will go nowhere until somebody who disagrees with the paper's conclusions takes on the task of doing an objective, systematic analysis of the EQU functions; it's actually not that complicated. However, I do understand the reluctance.
How am I supposed to reach a conclusion about an issue so filled with "definitional ambiguities" no matter how much time I spend studying the EQU functions? Indeed, how could the study even hope to settle the matter, as you and RBH claim it does? Goalpost moving seems to be an equal opportunity pastime.
Uh? Just look around how many definitons of IC have appeared in this thread (including yours). All I am saying is, choose a definition and do the work, then we can talk. quote:
Not to mention, the fact is the paper doesn't make any claims vis a vis IC, so I really don't have a major beef with the paper, but with some of the conclusions being drawn from folks here about. And those differences can only be handled via a dialogue between the disputing parties.
Well, for the paper not having made any claim about IC at all, y'all ID advocates are sure putting a lot of effort trying to argue that it actually didn't!
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 15. May 2003 22:03
Roger wrote quote:
As far as your reasoning for why IC is not mentioned in the Nature article, I find both rather puzzling. First, if as you and others contend, this is evidence for Darwinian methods of resolving IC, then wouldn't the appropriate venue be a claim in a peer reviewed journal? Yet the claim doesn't appear. How then can the important peer review process evaluate such claims? Even if they aren't familiar with the nomenclature, an attempt to explain it would seem to be in order if the study is as significant as you claim.
The paper itself did not mention IC specifically, most likely for the reasons charlie mentioned. However, that doesn't prevent anyone else - those of us here on this board who read it, for example - from looking at the paper, its methodology, and results, and concluding that the data address the question of the evolvability of IC structures/processes, given Definition #1 above, the "present traits" definition.
I also think you've got a somewhat different notion of the "important peer review process" than is the actual case. Peer reviewers don't "evaluate such claims." They are not in general content filters. In the reviewing of journal submissions I've done, and that has been done by folks I know, the review is intended to try to ensure that (a) the research addresses an interesting or worthwhile question in the domain of interest; (2) the paper displays sufficient awareness of past work so as to be put into a context; (3) the research methodology is appropriate to the question addressed; and (d) the analyses are handled appropriately from a technical point of view. Typically a reviewer wouldn't ding a paper on the author's interpretation of his results unless they're completely off the wall. There's some consideration of writing clarity (though not much!), and no explicit consideration of whether the data and interpretations agree with the reviewer's opinions, prejudices, or predispositions. I've recommended papers for publication where I disagreed with the author's interpretations but regarded it as the author's business how he interpreted the data: It's his research, after all. The reviewer does not pass on the larger meaning of the article. Peer review is intended to exercise some level of quality control on the technical side, but not so much on the content side. Once the paper is published, the content and interpretation and meaning and implications get reviewed by readers, much as is happening here.
RBH
IP: Logged
|
|
Argon
Member
Member # 276
|
posted 16. May 2003 10:12
One should note that Richard Lenski is doing parallel work in microbial continuous cultures. He has been running 12 cultures for about 11 years and archiving samples to track changes. This is generating data that will take grad students and post-doc years to analyze (It's tough to tease apart the genomes of thousands of strains of E.coli and track the changes in functional interactions over time!). There are at least a couple other labs performing related experiments in other model systems. I believe the utility of continuous cultures in evaluating ID proposals was discussed on a previous Brainstorm thread.
Lenski has also done work investigating the cost of antibiotic resistance and the compensatory mechanisms acquired by bacteria to relieve such costs. I suspect these interacting mutations (resistance followed by suppression of growth inhibition) may also be evaluated in terms of their "IC-ness". There are a couple fairly recent reviews listed in his publications list.
Richard's home page is here.
IP: Logged
|
|
YZ2
Member
Member # 91
|
posted 16. May 2003 13:04
I have to admit I have never quite read Behe's book, so I may not know some of the concepts as well as many of you here. Here is my question:
Isn't that in the theory of ID,
IC implies ID (to avoid false positive)
but nothing is said about the other cases, i.e.
non-IC may or may not imply ID
ID may or may not imply IC.
Now, in terms of evolution history and IC, the relationship is even more murky, ie
IC is the current state of property, whether it is reducible or irreducible. It may or may not be related to whether it has evolved in the past or not. Something that evolves in the past, can now become IC. The two questions are quite different.
Now if EQU is IC, the theory only says it is ID.
[ 16. May 2003, 16:15: Message edited by: YZ2 ]
IP: Logged
|
|
Argon
Member
Member # 276
|
posted 16. May 2003 17:51
YZ2 writes:
"Isn't that in the theory of ID, IC implies ID (to avoid false positive)"
That is the relationship some ID theorists would like to establish. But even in his book, Behe mentions that IC systems could potentially arise via indirect routes. Therefore the IC-to-ID link is not absolute.
[...]
"IC is the current state of property, whether it is reducible or irreducible."
IC is a state. By definition, an IC system is irreducible. A system is said to be irreducible with respect to a particular function if one cannot delete or easily mutate a component without losing that particular function. However, the same system may not by IC with respect to a different function. Thus, IC-ness is context dependent. In determining whether a system is IC, one must clearly define: a) The exact components, b) The exact function under consideration, and c) The local context/biochemical environment specific to the function and the system. If system was "reducible" in the past then it was not an IC system at that time. If the same system subsequently acquired a change that made it "irreducible" that system would then be called IC.
[...]
"Now if EQU is IC, the theory only says it is ID."
In theory, theory and practice are the same, but in practice, they're different.
Yes, the proposal is that IC systems are unlikely to arise by natural means. Thus if EQU is IC (which it is), and it was not intelligently designed (which it wasn't), then perhaps IC isn't quite a gold-standard indicator of design. Now, there will be attempts to revise the definition of "IC-ness" in light of this (as demonstated in this thread). To avoid future confusion (too late, I suppose), I suggest that any new criteria not be named "Irreducible Complexity" or even IC-2. Better to use a name that properly describes the particular set of criteria (e.g. "UFB" for "Unbridgeable Functional Gap" or "RMSNM" for "Requires Many Sequential Neutral Mutations", or "RTFM" for "Requires Too Fortunate a Mutation"). Otherwise we'll get into a confusion situation like calling every color "red". Imagine the conversation in that situation:
>>> "Boy the sea was really red today."
>>> "What? Red like the sun or red like the grass?"
>>> "No, I mean red like the dirt."
[ 16. May 2003, 17:53: Message edited by: Argon ]
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 17. May 2003 21:11
I've finally gotten around to looking at the Brig Klyce reference Micah supplied a few days ago. It's disappointing. There are two non-trivial errors in the first paragraph's straight description of the paper, one of them a real whopper, so it isn't really worth reading. (Hint: For the whopper, read the Avida instruction set the experiment used.)
There's another error of some significance in the second paragraph of Klyce's discussion. Like others who have criticized the paper, he says quote:
To enable "EQU" to evolve, the team employed a system that rewarded specific intermediate steps on the route to it. When intermediate steps were not rewarded, the "EQU" function did not evolve a single time in 50 runs. Rewarding intermediate steps assures that the fitness landscape has gradual slopes leading to the prescribed summit.
The locution "rewarded specific intermediate steps on the route to it" seems to imply that the experimenters figured out 'the' evolutionary route to EQU beforehand and then supplied the necessary intermediates to ensure a smooth path to it. Nothing like it. In fact, 23 different routes to EQU were taken in the 50 fully enabled runs, with an additional 124 routes to EQU showing up in lineages in a 360-run control condition with partially disabled fitness functions. Though the authors don't mention how many of those 124 were unique, they had to be mostly different from the 23 in the fully-enabled condition since the control runs in which the 124 appeared did not reward one or a pair of the simpler logic operations.
Klyce (and not a few other people) clearly missed the description of that last control condition in the paper, which showed that eliminating any one or any two of the simpler logic functions from the fitness function did not eliminate the evolution of lineages that produced genotypes that performed EQU. As I suggested elsewhere, the subpopulations evolving in Avida deke and juke around, ducking here, backing up there, and generally don't follow a "gradual slope." The notion that EAs are simple-minded monotonic hill-climbers is simply false. "Incremental evolution" does not mean "monotonic hill-climbing."
Finally, Klyce remarked that quote:
Furthermore, in real life the summits are not prescribed.
Well D'oh! It's an experiment, Klyce! It was intended to ascertain whether certain kinds of systems can evolve. So yeah, the experimenters defined the summit, to see if (under the conditions specified) the algorithm can find that kind of summit. Turns out it can.
RBH
[ 17. May 2003, 21:21: Message edited by: RBH ]
IP: Logged
|
|
Micah Sparacio
Member
Member # 6
|
posted 19. May 2003 09:22
Hey RBH,
I'm more interested in what happens if you remove a nand primitive or even a pop, swap, etc.
The point here is that the Avida instruction set is perfectly suited for moving data around in a static architecture (three registers, 1 stack, output, instructions) and modifying the data via nand. These are the building blocks that are significant. These are the building blocks that ensure that data in an Avida program can be modified towards an EQU or other logic function.
Removing the selective advantage of a some of the logic function intermediates does not do much for me. You still need to select for some of a specific set of intermediates and all of the intermediates are intimately related as logic functions to the EQU function.
I'm surprised that you're so avidly rejecting the notion of a relatively smooth path. Does anyone question the fact that if there are readily available, small steps between intermediates, leading up to a target that the function has a reasonable chance of finding the target? What does this answer? What does it do for us? An above 50% success rate indicates to me that the path is smooth...in fact, this should be exactly what jumps out as being a problem for the experiment.
Regarding MESA, you should be aware that we didn't suppose that all GA's are smooth monotonic hills. We just wanted to see what happens on a landscape with an ideal fitness function and environment when multiple coordinated mutations were required for selective advantage.
Finally, let me tell you that it has been a secret desire of mine for sometime to see some form of A-life or genetic algorithm do something really cool (I was especially fascinated with Tom Ray's Tierra and his vision for a Tierra network). What excites me is the dynamic potential of such things. Unfortunately, the system we are discussing in this thread is limited to finding sets of 1 and 2 input logic functions operating bitwise on 32 bit strings. In fact, it is limited in its ability to even discover some 2 input logic functions:
http://scitec.uwichill.edu.bb/cmp/online/P10F/logic%20circuits.htm
Not to mention a plethora of 3 input logic functions:
code:
logic 3a + 0 # A and B and C
logic 3b + 0 # A and B and ~C
logic 3c + 0 # A and ~B and ~C
logic 3d + 0 # ~A and ~B and ~C
logic 3e + 0 # A and (B xor C)
logic 3f + 0 # A & (B | C)
logic 3g + 0 # A + B + C = 2
logic 3h + 0 # A + B + C >= 2
logic 3i + 0 # A & ~(B xor C)
logic 3j + 0 # A xor (B & C)
logic 3k + 0 # A | (B & C)
logic 3l + 0 # A & (B | ~C)
logic 3m + 0 # (A & ~B) | (~A & B & C)
logic 3n + 0 # (A & ~B) | (B & C)
logic 3o + 0 # A & (B nand C)
logic 3p + 0 # A xor (B & C)
logic 3q + 0 # A | (B & C)
logic 3r + 0 # (A xor B) & ~C
logic 3s + 0 # ~A & ( B xor C) ) | (A & B & C)
logic 3t + 0 # (A & ~B) | (~A & B & ~C)
logic 3u + 0 # A & (B | ~C) | (~A & ~B & C)
logic 3v + 0 # (A xor B) | (A & C)
logic 3w + 0 # (A nor B) nor C
logic 3x + 0 # (~A & (B | C)) | (B & C)
logic 3y + 0 # (~A & B) | (~A & C) | (B & ~C)
logic 3z + 0 # A | (B & ~C)
logic 3aa + 0 # (A & ~B) | (A & ~C) | (~A & B)
logic 3ab + 0 # A + B + C = 1
logic 3ac + 0 # A xor B xor C
logic 3ad + 0 # (A & ~C) | (B & ~C) | (~A & ~B & C)
logic 3ae + 0 # ???
logic 3af + 0 # ???
logic 3ag + 0 # A | (B xor C)
logic 3ah + 0 # ~( (A & B & C) | (~A & ~B & ~C) )
logic 3ai + 0 # A or B or C
logic 3aj + 0 # (A & B & C) | (~A & ~B & ~C)
logic 3ak + 0 # A nor (B xor C)
logic 3al + 0 # ???
logic 3am + 0 # ???
logic 3an + 0 # (C & (A | B)) | (~A & ~B & ~C)
logic 3ao + 0 # A xor ~(B xor C)
logic 3ap + 0 # A + B + C != 1
logic 3aq + 0 # (~A & ~B) | (A & B & C)
logic 3ar + 0 # (A xor B) nor (A & C)
logic 3as + 0 # ~A & (B | ~C)
logic 3at + 0 # (A & ~B) | (A & C) | (~B & ~C)
logic 3au + 0 # ???
logic 3av + 0 # ???
logic 3aw + 0 # ???
logic 3ax + 0 # A | (B nor C)
logic 3ay + 0 # (~A & (~B | C)) | (A & (B xor C))
logic 3az + 0 # A | ~(B xor C)
logic 3ba + 0 # A nor (B & C)
logic 3bb + 0 # A equ (B & C)
logic 3bc + 0 # (A & ~B) nor (B & C)
logic 3bd + 0 # (A & ~B) nor (~A & B & C)
logic 3be + 0 # A nor (B & C)
logic 3bf + 0 # A equ (B & C)
logic 3bg + 0 # A nand B nand C
logic 3bh + 0 # ~A | (B & ~C)
logic 3bi + 0 # A nand ~(B xor C)
logic 3bj + 0 # ~A | B | C
logic 3bk + 0 # A + B + C <= 1
logic 3bl + 0 # A + B + C != 2
logic 3bm + 0 # A nand (B | C)
logic 3bn + 0 # A nand (B xor C)
logic 3bo + 0 # ~A | ~B | C
logic 3bp + 0 # ~A | ~B | ~C
[ 19. May 2003, 09:34: Message edited by: Micah Sparacio ]
IP: Logged
|
|
Pim van Meurs
Member
Member # 541
|
posted 19. May 2003 15:25
Micah: I'm surprised that you're so avidly rejecting the notion of a relatively smooth path. Does anyone question the fact that if there are readily available, small steps between intermediates, leading up to a target that the function has a reasonable chance of finding the target?
First of all let me point out that the presence or absence of a smooth path has nothing to do with the definition of IC. Secondly it may be interesting to look at the protein and RNA networks. I have argued in the past that the scale free network format of RNA/DNA has some very interesting impact. First of all there are a few very common structures, spread out through sequence space and in close distance from eachother. I am not sure if Avida allows for neutral evolution but the works by Schuster and many others gives us some interesting insight in the closeness of proteins in sequence space. Application of neutral evolution and selective evolution would explain both stasis and instances of increased evolutionary rates. During stasis the mechanism is mainly neutral until a 'novelty' is reached after which selective evolution may take over again.
But the pathway does not seem to be that smooth afterall, nor is the term desired target very accurate. As RBH points out 23 different pathways
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 19. May 2003 18:31
Pim wrote quote:
Micah: I'm surprised that you're so avidly rejecting the notion of a relatively smooth path. Does anyone question the fact that if there are readily available, small steps between intermediates, leading up to a target that the function has a reasonable chance of finding the target?
Actually, it was me who was saying that the path is not "smooth," in the sense of a nice monotic slope up some single-function hill. Rather, it seems clear that the paths (plural) are jagged and winding, with detours downslope sometimes to get to another ridge or slope. And the plain old evolutionary processes can operate on that kind of fitness landscape; the landscape does not have to be carefully and specially "tuned." I'm fighting (perhaps needlessly) the notion that all the algorithm had to do was step from stone to stone, each stone representing "progress" up some nice smooth tuned fitness slope, in order to reach a genotype capable of performing EQU.
I sure hope I get that Beowulf machine for a couple of months. I really want to mess around with this stuff!
RBH
[ 19. May 2003, 18:34: Message edited by: RBH ]
IP: Logged
|
|
Pim van Meurs
Member
Member # 541
|
posted 19. May 2003 18:46
The paper itself, figure 3 gives us some insight in the fitness per generation which according to the paper is the product of the replication efficiency and computational merrit.
quote:
The presence of deleterious mutations along the line of descent is more surprising. Fifteen of the 18 deleterious mutations reduced fitness by ,3% relative to the parent, and might have hitchhiked with beneficial mutations that arose soon after in the same genetic background. However, two mutations reduced fitness by .50%. One was a point mutation that disrupted replication efficiency. Its harmful effect was eliminated by the next mutation in the line of
descent, which occurred at a distant site in the genome. The other very deleterious step was a point mutation, at depth 110, that knocked out NAND, one of the simplest logic functions. Only two individuals had this maladapted genotype, yet their descendants emerged as eventual winners. In fact, in the very next step, this genotype produced the mutation that gave rise to EQU. Was that deleterious mutation extremely lucky to hitchhike with such a beneficial mutation? Or was the deleterious mutation a prerequisite for producing the EQU function within that genome context? To distinguish between these hypotheses, we reversed this one-stepprior mutation in the genotype that first expressed EQU. This reversal eliminated the EQU function. Therefore, a mutation that was highly deleterious when it appeared was highly beneficial in combination with a subsequent mutation. The evolution of a
complex feature, such as EQU, is not always an inexorably upward climb toward a fitness peak, but instead may involve sideways and even backward steps, some of which are important.
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 19. May 2003 23:38
Micah wrote quote:
Unfortunately, the system we are discussing in this thread is limited to finding sets of 1 and 2 input logic functions operating bitwise on 32 bit strings. In fact, it is limited in its ability to even discover some 2 input logic functions:
I'm not sure what you're suggesting here. Clearly the experimenters in designing the study limited the number of one- and two- input logic functions that entered the fitness function. I don't think that was an "in principle" limitation, though. I don't doubt that more could have been entered and subsequently performed by the critters. That wasn't a question they were interested in, it seems.
I think it would be interesting to see if given an instruction set lacking the nand primitive, the population could nevertheless evolve a string of instructions that performs nand (obviously nand would have to be defined to have some adaptive value), and whether the population could then use that string in more complex logic functions. That is, how primitive can the primitives be?
I too was excited about Tierra, and Avida is a superset of Tierra - one can (supposedly) run a Tierra simulation within Avida. I really want to get version 1.6 running on that Beowulf cluster.
Tell you what, Micah. If you want to design a study to run in this area using version 1.6 of Avida, testing some interesting hypothesis related to the study under discussion here, and if you can generate the appropriate control files, if I get enough access to the Beowulf cluster I'll run it for you and ship you the output data. (On the assumption that its run time doesn't exceed my remaining lifetime, that is!) I won't even ask for co-author credit! ![[Wink]](wink.gif)
RBH
IP: Logged
|
|
YZ2
Member
Member # 91
|
posted 20. May 2003 09:48
There is actually another issue that the experiment has raised. If the goal is to obtain EQU from NAND, isn't there more direct methods such as testing an incremental search of combining the NAND functions? This is certainly faster than a life-time of mutations and testing.
IP: Logged
|
|
warren_bergerson
Member
Member # 262
|
posted 20. May 2003 11:15
XZ2 raises a good question. It is useful to expand upon the question raised. If the survival of a species depended upon an evolutionary change logically similar to the change from NAND to EQU, if time available for the required change was less than the 15,000 generations required for a mutation-natural selection process, if, as XZ2 suggests there are faster more efficient mathematical methods of producing the change, then isn’t it likely/possible that evolutionary change can and does occur by some process other than mutation and natural selection.
The demonstration presented in the article shows that some types of increases in complexity can be produced by some types of search routines. As was discussed earlier, the study does not demonstrate that the type of search performed in the demonstration is ‘biologically possible or realistic’. The evidence, in fact, suggests that the demonstration used somewhat unrealistic assumptions to make if possible for the identified complexity to evolve by a mutation and natural selection process.
As XZ2 suggests, there are mathematical search routines which could produce the identified ‘evolution of complexity’ much faster and with greater efficiency. Given the known complexity of biological systems, given the speed with which certain types of complexity are known to have evolved (human intelligence would be a reasonable example), and given the demonstrated inefficiency of (non-modified) mutation and natural selection search routines, it would seem reasonable to suggest that evolutionary change is the result of some process other than mutation and natural selection.
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 20. May 2003 11:28
YZ2 wrote quote:
There is actually another issue that the experiment has raised. If the goal is to obtain EQU from NAND, isn't there more direct methods such as testing an incremental search of combining the NAND functions? This is certainly faster than a life-time of mutations and testing.
The goal wasn't to "obtain EQU from NAND." The goal of the experiment was to see if given a pool of primitive instructions, including NAND, regular evolutionary processes (mutations of various sorts plus reproduction plus selection) could evolve an assembly language program that performs EQU when some simpler logic operations are selectively advantageous. They can. And since all 23 of the programs that evolved were irreducibly complex on Definition 1 above (Present State as determined by knockout testing), that implies that irreducibly complex objects are not only evolvable, they're quite easily evolvable. (OF course, if irreducible complexity is defined as "not evolvable," as some in this thread have suggested, then the EQU-performing programs are not IC. But that definition descends into either question-begging or an "IC of the gaps" position.)
John Bracht made the same sort of suggestion as YZ2 earlier in this thread when he remarked that he thought a 'tornado in a junkyard' model could accommodate the results. Unfortunately, the 50-run 'no intermediate selection' control condition shows that notion is untenable. Moreover, there is sufficient information in the paper to calculate the probability of obtaining a program capable of performing EQU from the primitives. John has not yet supplied a calculation to show that his suggestion is plausible, say nothing of calculating the probability that 23 different lineages would each produce a different program capable of performing EQU in 50 evolutionary runs.
RBH
IP: Logged
|
|
|
Printer-friendly view of this topic
