|
Author
|
Topic: John R. Bracht: Inventions, Algorithms, and Biological Design
|
Moderator
Administrator
Member # 1
|
posted 26. November 2001 14:04
Inventions, Algorithms, and Biological Design
by John R. Bracht
jbracht@prodigy.net
ABSTRACT—This paper outlines a new model for understanding biological invention, based upon an extensive study of human inventiveness originating in Russia shortly after the Second World War. This science, known as the Theory of Inventive Problem Solving and referred to by the Russian acronym TRIZ, relies upon the study of human patents to reveal general principles of invention. TRIZ recognizes two distinct mechanisms of invention, operating upon two distinct types of problems. The first, trial and error, is a remarkably accurate description of the neo-Darwinian mechanism, natural selection. Furthermore, the trial and error mechanism has been found to be severely limited in the sorts of problems it can solve; these limitations are also found to apply to the Darwinian mechanism. The second mechanism, lacking an explicit TRIZ name but referred to here as the intentional mechanism, is the source of true inventiveness. The science of evolutionary programming gives insight into precisely why the intentional mechanism is required; certain fundamental parameters must be given before the Darwinian mechanism can even operate and these parameters are themselves out of reach of the Darwinian mechanism. Certain key events in the history of life require alterations of this sort of fundamental parameter, and it is precisely these events that the neo-Darwinian model fails to explain. Consequently, it is the inescapable conclusion that there is a second mechanism, an intentional mechanism, which operates in nature and is responsible for the changes that are not accountable via the Darwinian mechanism.
To read the entire paper, please click here
Paper Versions:
-27 November 2001
[ 05 May 2002, 15:55: Message edited by: Moderator ]
IP: Logged
|
|
Elend
Member
Member # 326
|
posted 12. July 2002 10:52
Good attempt in pointing out the inadequacies of current GA for modeling evolution. In fact GA used as search methods are very restricted. I have the following questions though:
1) It is claimed that GA can only search for solutions in a n-dimensional hyperspace, where the number n is fixed. This is true for the classic GA where the variables of a problem (the ones which the user sees as variables) are making up the chromosomes. Note that the number of variables is known, fixed and indeed programmed by the user. Yet a more general GA can be imagined where the number of variables or in other words the length of the chromosomes vary. Mistakes like copying the same part of a chromosome twice is a valid mutation. The general GA is no longer limited to the n-dimensional hypervolume.
In your graph example, one could start with one point chromosomes and allow mutations that increase and decrease the number of points instead of starting directly with 5 points.
In the icebreaker example, the variables of the systems do not have to be only the engine power for example, but also the shape of the ship, the materials, etc.
2) The fixed/tuned parameters of a GA, such as mutation rate, mutation points, crossover points, population size, etc. are indeed important finding solutions QUICKLY. With the limited computation power and memory resources, the speed converging to optimum of such algorithms is imperative. Yet, if these are set by chance (and population size might very likely increase exponentially) the algorithm will still find near-optimal solutions, granted it will take a very long time to do so.
3) The implicit assumption in your paper is that evolution claims organisms to be optimal solutions, and the only solutions for their own environment, and GA must obey into finding these optimal solutions too. But evolution makes no such claims - current organisms are not perfect or unique (the only possible) solutions. While we use GA to search for near-optimal solutions, stopping the GA early we do get better designs than the initial ones.
4) The fittness function is designed in GA because we have goals, we need a specific solution to a specific problem. Natural selection in evolution on the other hand may indeed be just a natural process, responding to physical laws - round rocks roll down a slope because of the gravity, but there is no intelligence behind finding round rocks at the bottom of a slope.
Finally, (classic) GA were inspired by the evolutionary theory, as a search method. They were not ment to model evolution, only certain aspects from it. More general GAs may not suffer from the drawbacks you accurately point out.
IP: Logged
|
|
Frances
Member
Member # 169
|
posted 12. July 2002 13:35
I have been reading Bracht's paper for the last day or so and interestingly enough many of my comments seem to overlap with Elend's.
While it is tempting to extend the basic GA algorithms to evolution, one has to realize that the original mutation, selection algorithms indeed were 'limited' in their solution space.
The question of course is, is evolution limited in such a manner? And I argue that it isn't for the following reasons:
1. All 'innovations' in nature are still encodable in the DNA, thus in theory any mechanism which could search the DNA space could run across 'innovations' and this is important to realize that there is not a simple 1 to 1 mapping from genotype to phenotype. In fact, one example mentioned, the Cambrian explosion, might be understandable through the hox genes, slight variations in developmental timing can lead to significant morphological changes. Thus even if the genome were limited to dimension N, there is no such inherent limitation to the dimension of the morphology.
2. The genome itself is not limited to dimension N either. For instance, gene duplication can give rise to higher dimensions. Thus the idea that RM&NS cannot explore beyond the original dimension of the genome seems to be faulty.
3. Bracht is right that RM&NS is restricted in the sense that selection can only explore space that is explored by mutation and mutation can only explore space that has some advantage or is at least neutral. This is why in evolution we often see how existing structures are reused for a different purpose.
4. Bracht objects to the fitness functions but does not explain why fitness functions in such algorithms may be different from what nature provides.
More later since I have run across some interesting example how a temporally variable fitness function may be the 'driving force' behind innovation.
Here is the link
The conclusion is quite interesting since it shows using simple model that for evolutionary versality to increase, it must be possible for the lengths of the genome to increase. See my argument 2). I argue that this relatively simple model shows that there need not be a bound to Darwinian evolution perse.
[ 13 July 2002, 16:59: Message edited by: Frances ]
IP: Logged
|
|
John Bracht
Member
Member # 5
|
posted 13. July 2002 22:31
Hi Elend,
Thanks for the thoughtful reply. You actually bring up the most common objection to the ideas presented in my paper: "what about gene duplication?" I've dealt with it at length before, and you may be interested in reading what I wrote then. Here's the link:
http://www.arn.org/cgi-bin/ubb/ultimatebb.cgi?ubb=get_topic;f=1;t=001388
(see especially my replies to Erik; you'll have to scan down about half the page to see my exchange with him.)
You are right that we can program GA's such that the number of genes are themselves allowed to vary. Effectively, then, you've introduced a higher-level variable that simply induces another dimension on the multidimensional hypervolume in which the GA operates. For each "setting" of the variable X which tells us how many genes there are, there will be a corresponding f(X), or Y, that delimits the possibilities that GA may explore. You have vastly increased the number of possibilities (or dimensions on the hypervolume in which the GA operates)--but you have not set your GA free from having constraints. In fact, there is no way to get around the fact that all GA's operate within a pre-defined hypervolume of possible outcomes--that the encoding of the problem determines, ahead of time, all possibilities that the algorithm may ever explore.
To see this clearly, (and see why I argue that gene duplication doesn't allow you to invent things), think about a concrete example, the Steiner algorithm in my paper. As you pointed out, one could program it to allow multiple points beyond just five (in fact, the original program had the capability to put in extra "floating points" in random places; as far as I know there was no limitation on the number of floating points). What have we gained? The algorithm can now explore a 6-point, 7-point, 8-point, etc., Steiner solution. But it's still limited to points and lines, and optimizing networks. It cannot do anything new like draw concentric circles around each point to generate pretty pictures, or interference patterns. It cannot solve differential equations. There are still limits on the sorts of solution the algorithm may explore even though we have allowed for limitless expansion in the number of variables. That's why in my paper I not only emphasize the variables that evolution can work on, but also the crucial incorporation of those variables into an overall whole. Another way to think of it is this: every variable comes with some rules attached that determine what that variable "means." It often has some rules about what other variables must work together to produce a given effect. In essence, the hypervolume is defined by a combination of the variables, in concert with the way they are tied together to produce a given structure.
Let's look at another good example, again from the paper: the Biomorphs program. Dawkins himself took a 9-gene program and made a 15-gene program from it. Did he do it just by gene duplication? No. He had to alter the previously existing structure so the new genes would have new functions. They have to "make sense" to the pre-existing network, and that requires the old network be re-wired to integrate the new variables. What happens if we just duplicate genes? We get duplicates of a function we already have--we don't get a whole new functionality.
Part of the problem for a GA is that oftentimes many genes or variables are tied together in the production of any given structure. This is certainly true in biology, and Dawkin's angel wings example demonstrates nicely why gene duplication alone doesn't give inventive change. In order to sprout angel wings, there are a host of new genes needed, and they have to do new things. Genes for new bone, muscle, tendon, blood vessels, etc., must be incorporated into the developmental program of the embryo to have an effect from the very beginning. Such a major change will involve altering a good many other developmental pathways such that this new one is properly integrated and regulated (controlled).
In short, inventive change (re-engineering a hypervolume) comes from not only new variables (or genes) but new interactions with the previously existing system.
You also raise the interesting question: can we just "telescope out" enough in our parameters such that the hypervolume incorporates all possible inventive solutions (eg, so that the inventive icebreaker hull is within the hypervolume of icebreaker possibilities)? The problem here is that natural selection cannot cross over certain boundaries: the technical contradictions. This is the key problem: natural selection can only operate upon what already exists. Inventive solutions do not yet exist. Therefore, natural selection cannot operate upon them. In other words, natural selection can select for and preserve an icebreaker with bigger engines (engines already exist and it's conceptually simple to enlarge them). However, a brand-new hull design simply doesn't exist yet, and so selection cannot drive the inventive change. The fitness function "ends" at the place where the old hull design reaches the contradiction that one needs bigger engines (for more power) and smaller engines (to reduce weight and support systems). At this point the Darwinian mechanism is stuck; it needs to step outside the boundaries on the hypervolume prescribed by the contradictions inherent in inventive problems. And it simply cannot do so.
You bring up a few issues with how quickly and optimally evolutionary algorithms operate. I actually don't make any claims about the efficiency or optimality of GA's, and these issues don't matter at all to the argument I'm making, which deals instead with the basic characteristics of GA's and the Darwinian mechanism.
Your point number 4 seems to be basically a statement of front-loading: the information in in the environment, and is extracted and used by organisms as they evolve. The problem with this argument is that biological complexity is not (as you imply) pre-encoded into natural laws. There is no law-like drive to generate a bacterial flagellum, for example, or any other biological structure. In fact, evolutionists often point out that their mechanism is without plan or purpose. It just preserves whatever pops up and happens to be useful. The problem here is that the mechanism has to rely on chance alone to do the innovating; once the innovation is done, selection (acting in a law-like way) can preserve it. But it's definitely not a mechanism that generates biological complexity in a law-like way comparable to a rock rolling down a hill. In fact, it's not really a mechanism at all--it's a reliance upon pure chance to do the design work, and selection to preserve that design work. The design inference gains traction precisely where the probabilities associated with that chance-driven design work are so low that it would never be expected to happen in the history of the universe (until its heat death).
I hope this helps; thanks much for your feedback.
Sincerely,
John Bracht
[ 13 July 2002, 22:33: Message edited by: John Bracht ]
IP: Logged
|
|
John Bracht
Member
Member # 5
|
posted 13. July 2002 22:48
Frances,
Thanks for the comments. Some of your concerns are addressed above (about gene duplication) but I did want to address your argument about a lack of 1-to-1 correspondence between genes and morphology (genotype and phenotype). This is truly a novel argument and a good attempt at getting out of the problems (for Darwinian evolution) posed by the analysis given in my paper. Here's why it doesn't work: while there may not be a strict 1-to-1 correspondence (the environment has some influence on development etc.,), it's pretty close. There is no doubt that there is a basic correspondence between genotype and phenotype. In fact, all of molecular biology and genetics is based upon this fact. For a given phenotype, there is an underlying genotype that more or less produces that phenotype. We have only to look at identical twins to see this correspondence. It's not perfect, but it's darn good. And if this correspondence were not there, then many modern techniques in genetics and molecular biology (like mutagenesis screens) would be impossible, because there would be little or no correspondence between genes and phenotype; thus, manipulating the genes in an effort to alter morphology would be pointless and futile. Since they do, in reality, correspond,the hypervolume which constrains variation of one also constrains variation in the other (this should be obvious since the phenotype is actually produced via a developmental process encoded into the genotype).
It might also be worthwile to point out that while you are correct that any innovation is in-principle encodable in DNA (this is like saying that any invention is capable of being produced by manipulating matter), it does not follow that a trial-and-error process can generate any concievable innovation. That is the whole point of my paper, and the TRIZ theory of inventiveness in general. The fact is, inventive changes (in biology or human engineering) require re-designing the hypervolume such that the inventive solution is now reachable. In practice, there is no way for a trial-and-error process to achieve this re-engineering of the hypervolume, for the simple reason that they operate within that hypervolume and cannot step outside it.
Sincerely,
John Bracht
P.S. I'm now in grad school, doing a summer research rotation before classes start. So I'm very busy and that's why any further responses may be delayed (it's also why I've been absent from this board for awhile). I'll certainly do my best to be involved as time permits.
[ 13 July 2002, 22:53: Message edited by: John Bracht ]
IP: Logged
|
|
Frances
Member
Member # 169
|
posted 14. July 2002 00:24
John,
Thanks for your response. Your argument about gene duplication needs some clarification, in your response you refer to the Steiner problem to suggest that natural selection cannot cross over certain boundaries. But you have failed to show this, in fact we know that mere variations on DNA can result in innovations. So all that is in principle needed is variation that achieves such a variation. An interesting example for instance is given by the reptile mammal transition were a jaw bone takes on the role of a new structure or the "At the water's edge" by Karl Zimmer in which he suggests how hox genes may have transformed a fin into an appendix with digits.
One may of course argue that this is not innovation since it reuses an existing system but if that's the argument then our observation that evolution seems to have reused a lot of structures suggests that the theory of evolution is sufficient in explaning these 'innovations'. You seem to suggest that evolution cannot think up totally new solutions, and even if you are right, and I argue that you have not shown that this is the case, evolution can come up with 'novel' solutions to deal with new challenges in nature.
John then argues that there is no 'law like drive to generate a flagellum'. This of course is correct in a limited sense that there is no law which results with 100% certainty in the flagellum but there is a law-like mechanism which uses 'random fluctuations' and selection to explore new solutions. As has been shown in limited simulations, such approaches can be quite succesful.
John's argument that it 'relies on pure chance' to do the design work is correct but John does not provide us with a reason why chance would not come up with novel solutions. Once they arise selection can select for them if they provide the organism with a reproductive advantage. As shown by Schneider and Adami, random fluctuations and selection are sufficient in explaining the increase of Shannon information in the genome for instance.
Design inference tries to use our limited knowledge of these occurrences to suggest that there is design but does not seem to make an independent attempt to show that the probability for design is indeed larger than that for non-design.
John then continues to argue that twins are similar but that does not deal with my argument about one-to-one or one-to-many mapping. Since the two genomes are identical and the developmental environment is very similar we should not be surprised to find similarities in twins. Of course we should also not be surprised to find differences in fraternal twins for instance. In fact hox genes for instance show that small delays in timing in the development could lead to relatively large morphological differences. Of course for a given phenotype there is one genotype that generates but this does not mean that small changes in genotype cannot have signficant impact on the genotype. That's what I mean by 1 to many.
You are correct though when you stated that since innovation is in principle encodable in DNA that RM&NS can generate this but that's a whole different issue. The fact that RM&NS in principle cannot be excluded from generating innovation means that one needs to do more than merely showing that certain limited GA's cannot generate innovation beyond a certain limit.
By defining inventive chancges as re-designing the hyper volume John tries to give design a prefered status but he forgets that even if RM&NS cannot redesign the hypervolume in an inventive manner, it can do it through simple processes which could at least in principle generate the same results.
Gene duplication is but one argument that shows that RM&NS can step beyond its given hypervolume. Your paper argues for this in an indirect way when discussing the new Dawkin's biomorph program with additional genes which now generate a 15 dimensional hypervolume versus the original 9. So if the mere addition of genes can in principle increase the dimension of the hypervolume then your argument seems to be self defeating.
If your argument is that selection cannot drive evolution then you have indeed captured the general evolutionary thought but this hardly shows that innovation cannot arise and thus seems more like a red herring to me.
In fact there is even some suggestions that selection can drive evolution, that is through evolution of evolvability.
It seems to me that John is restating the idea that evolution is not forward looking and that design is which means that it can take leaps that evolution cannot take. But that is not sufficient evidence that evolution could not find solutions to these problems, perhaps not as optimal as a designer would but that seems what we are seeing in the world around us.
The questions about GA's being able to extend their hypervolume are interesting but irrelevant since it seems that gene mutation itself is sufficient at least in principle to do this. All that can be shown is that evolution may be more constrained than a designer but it hardly follows that evolution is prohibited from finding solutions similar or even better than those found by a designer.
Some interesting links
Limb evolution
Fish to amphibian transition
Missing links and the origin of biochemical complexity
Provides evidence of a gene duplication, recruiting and amplification a new ability evolved.
A classification of possible routes of Darwinian evolution. By Richard H. Thornhill and David W. Ussery
I fali to see what the relevance of TRIZ is with respect to evolution. All it shows as far as I can tell is that intelligence may have fewer restrictions on designing innovations than nature. Is this surprising? Is the problematic for evolution? I do not think so.
It might present us with a way to detect design but that would be not dissimilar from Behe's ICness argument and thus have similar problems.
Actually some of the "Laws" seem to be not too dissimilar from Behe IC.
Static Law (can we say IC?) quote:
Every technical system includes four basic parts. The system will not work if one of these parts is missing, or does not perform well. The generic names of these parts are the ENGINE, the TRANSMISSION, the LIMBS, and the CONTROLS. The engine is the source of energy.
Too me TRIZ sounds a bit as 'hype' with their claims of 'laws'
Check out Triz website for more information.
Are these "laws" really what they claim to be? Or is the term "law" used here inappropriately?
Also after reading about TRIZ it does not seem that the claim is that trial-and-error cannot generate innovation, just that it is more cumbersome and requires more resources and time.
An interesting paper on the importance of non-coding areas for the convergence of solutions can be found here.
An interesting project on ontogeny
quote:
In (Bongard & Pfeifer, 2001) we demonstrated that the developmental encoding is compact, and how such an encoding leads to repeated, higher-order phenotypic structures in the evolved agents. In a forthcoming publication we will show in more detail how evolution shapes the underlying genetic regulatory networks of the virtual agents. Finally because the encoding scheme is highly evolvable, it will be shown that it can be used to evolve agents which exhibit increasingly non-trivial behaviour, which an outside observer may classify as intelligent behaviour.
In fact Bongard et al also report in a presentation that the combination of evolution of morphology and the controlling genetics outperform agents with fixed morphology. In fact in variable morphology cases, the combination of variations in control and morphology allow searches to move along extradimensional bypasses.
This seems quite relevant to NFL since it shows how the combined evolution can smooth the fitness landscape.
More on one-to-many relationships
Pleiotropy: e ability of a single gene to have multiple phenotypic effects.
[ 14 July 2002, 15:32: Message edited by: Frances ]
IP: Logged
|
|
Elend
Member
Member # 326
|
posted 14. July 2002 15:32
Thank you both for the links.
John,
I agree that evolution does not "invent" designs as we do. It needs intermediate designs to arrive to better ones. Yet, this seems to be exactly the case in nature. I would be very interested in species that exhibit unique designs, that are completely novel and independent on other designs - maybe you can provide such a reference for me. (is the flagellum one of these?)
You argue that increasing the size of the chromosomes does not change their interpretation, and the functionality (resulting from interpretation) remains unchanged. As for example increasing the number of points in the graph program would not make it draw circles. I partially agree - yet if the interpretator is general enough this is not a problem. I suggest that a Universal Turing Machine (UTM) would do the trick. It employs only four very simple operations (read/write symbol and shift tape left/right) and a tape. A UTM is equivalent with any computer. Using the tape as chromosomes I believe it would be possible to evolve programs that exhibit new functionality. Such as like going from points and lines to circles if this is required for some reason.
Yet, there are indeed problems that the UTM can not solve: such as the Halting Problem.
(can't give a url, my browser behaves strangely with cut&paste, but I'm confident you already know about the UTM)
[ 14 July 2002, 15:39: Message edited by: Elend ]
IP: Logged
|
|
Frances
Member
Member # 169
|
posted 14. July 2002 15:49
Behe made the attempt to show that IC systems could not have evolved naturally but in fact it was shown that while direct evolutionary pathways may not have been directly relevant, indirect pathways do in principle exist.
Are there any systems which show a unique design which trades off several constraints? One of the claims of TRIZ is that contradictions allow for innovative solutions. I would argue that similar constraints are also seen in evolution, that is contradictions lead to innovative solutions. For instance "sexual reproduction, symbiosis" seem to be examples in which two contradicting constraints lead to innovative approaches. Even TRIZ recognizes that trial-and-error could lead to innovations but the problem is that inventors do not have the time to do the necessary trials. TRIZ addresses issues such as "how can the time required to invent be reduced".
Perhaps TRIZ's claims of laws should be renamed to heuristics?
Altschuller defined 40 inventive principles. These principles include segmentation, extraction, nesting, dynamicity, feedback, copying. TRIZ encourages to look for analogous solutions and adapt to the situation.
[ 14 July 2002, 15:55: Message edited by: Frances ]
IP: Logged
|
|
John Bracht
Member
Member # 5
|
posted 16. July 2002 01:37
Frances,
You bring up some good points. You are probing some key issues surrounding my main thesis, which is good. I still have some disagreements with your assertions, and I'll try to spell them out as clearly as possible.
One point you make is that TRIZ doesn't preclude trial and error processes from reaching inventive solutions; it just says that trial and error is prohibitively inefficient when dealing with inventive problems. There are some places where the literature suggests this, but I think that the real innovation occurred via an intuitive leap to the new, inventive solution, not by the trial and error process (I think this makes more sense in terms of the actual claims made by TRIZ authors). In fact, the motivation behind TRIZ was to give inventors an alternative to trial and error; Altschuller always talks about overcoming psychological inertia (the tendency to think in terms of pre-existing ideas and structures) to allow the invention of something really NEW. In other words, he was always trying to get people to quit thinking in terms of trial and error. Trial and error may take us to a point from which it is easier to make that mental leap (hence, you could say it may sometimes help one to invent things) but it is fundamentally a different (inventive) process whereby the inventor leaves the realm of pre-existing things and comes up with something genuinely new. As I've argued before, the technical contradictions constitute absolute barriers which the process of trial and error may not pass. Certainly, an inventor may make use of his/her own capacity for inventive problem-solving and get past his/her psychological inertia. Does that show that trial and error processes solved the problem? No, only that humans are capable of getting past the limitation imposed by trial and error: the requirement that all candidate solutions be variants of pre-existing forms. Darwinian evolution, on the other hand, is not free in the same way. It always tests variants of pre-existing forms, and so it is held captive to what I term "morphological inertia." It simply cannot think "outside the box"; the human can.
This brings us to the key issue. Darwinian theory presupposes that all biological forms can be reached by creating variants of pre-existing forms. In other words, that there is no real innovation in biology, just a lot of routine problem-solving. In a Darwinian view, inventiveness is not a part of the biological realm. Because if there are indeed inventive problems that have been solved in biology, TRIZ theory tells us that trial and error (the Darwinian mechanism) cannot have effected that change. The reason is as I have described before, that inventive change absolutely requires creating something absolutely novel, something that cannot be generated by merely altering something that existed before.
As I've stated before, an inventive change in biology doesn't just involve adding new genes (though that is important) but also in integrating those genes into the existing network of genes (the regulatory network) in such a way that they mean something new; produce a new structure or feature. The intuition pump in this case came from Dawkins who talked about the difficulty of sprouting angel wings from a mammal's back. The process of embryonic development defines a hypervolume of possibilities which can be explored by genetic mutation. The ability to sprout angel wings is not in that hypervolume. Why? Because sprouting angel wings would require re-working the entire process of embryonic development, adding new genes and enmeshing them within the framework of other genes such that wings were formed. That re-working of the "rules" of gene expression and embryonic development, along with adding extra genes, is what I'm referring to as an inventive change in biology. I've never seen an example where that sort of change was observed to be the result of Darwinian processes. The reason, I argue, is that the process of embryonic development is a set of basic parameters which define the hypervolume in which the organism resides. Mutations to the genome serve to move about within that hypervolume (by their effect, "filtered" through development), but they do nothing to alter the fixed paramters of embryonic development in which the organism resides. (Notice I'm not saying that no mutations ever affect embryonic development; surely some mutations cause it to go badly awry or to terminate early, etc. But these do not alter the hypervolume that the organism exists in; they only mess up what is already established.)
Gene regulation is another factor in establishing a hypervolume of possibilities. Does a particular gene mutation kill the organism or create a favorable mutant? It depends on how it impacts embryonic development and also the expression of other genes. It is the logical interrelationship of genes that determines what changes are viable and hence selectable. Perhaps we can envision the process of embryological development as the interactions of genes through time: a program of certain interactions that give rise to a certain type of organism.
So the question becomes: is there anything in biology that I can point to as a definite inventive change? For concreteness I want to present a an example case to illustrate precisely what I'm talking about: the transition from single-celled organisms to multicelled metazoans.
Why do I contend that this was an inventive change? Because here a whole lot of genuinely new stuff was added. I'll give a brief (by no means exhaustive) list of what comes immediately to my mind:
-- cell-cell adhesion molecules
-- cell-cell signalling pathways
-- developmental program
-- tissue differentiation:
- muscle
-nerve
-gut
-reproductive tract
-circulatory system
-repiratory system
-etc
The amount of genetic material needed to put together even a simple metazoan is about an order of magnitude greater than the amount needed by a bacterium. Furthermore, these genes have no function until they are in the metazoan and properly integrated with other genes. It does the bacterium no good to have genes for multicellular adhesion molecules or signalling pathways. The developmental program simply doesn't make sense for an organism that reproduces by budding asexually. Having the genes for a complex mouthpart or an arm or leg do you no good if you are a single cell. I like how Christian Schwabe put it:
quote:
The phenotypes of the Quaternary (our period) are so different from the Cambrian creatures, it is argued by Darwinists, that only a long time of adaptation, fine-tuning so to speak, by natural selection would have converted them into all subsequent species. One must object to that proposition on the following grounds: As the Precambrian stem cells burst onto the animal scene they brought along from the single-cellular (or colony) stage their appendices, sensory and reproductive organs, their feeding machinery, and armor for protection. But since these features, save reproduction, are without meaning in single-cellular life from which the animals just emerged, their appendages could not have been adaptations to any need for defense that multi-cellular life forms developed purportedly to deal with each other. What kind of evolutionary pressure would have prepared single-cellular organisms for what lay one short step ahead of them once they transgressed the line to multi-cellularity?
(from here.) (emphasis added.)
What good are the multitude various genes specifically adapted to join cells together into an integrated unit, allow them to communicate, allow the system to exchange oxygen, circulate nutrients, or send nerve impulses before there were cells to join together, a system which can do the oxygen exchanging, nutrient gathering, or nerve-impulse encoding/decoding? In the bacteria, all these functions were integrated into one cell. In the multicelled organism, suddenly all these functions had to be divided between multiple different cells, and all this differentiation (which had no conceptual precursor in a bacteria) was controlled by a developmental pathway (again, meaningless in the single-celled progenitor). There is no half-way point. Either a system is multicelled (and needs all these different genetic characteristics) or it is single-celled (and many of these characteristics would be meaningless or useless--or worse, evolutionarily disadvantageous baggage). Even reproduction, which is needed whether one is single celled or multicelled, becomes vastly more complex with a dedicated germline that becomes either egg or sperm, a separate reproductive system that manages the task of producing germ cells and joining them together. So not only must all this biological complexity arise with a single, enormous leap, but it must happen twice to give rise to male and female (or, another way to put it, the developmental system must contain two pathways: one for male and one for female). This is an irreducibly complex system where both male and female are needed before the species can procreate.
Now, what's going on here? The rules of the system, the basic paramters, are changing. The bacteria had very few, simple rules. Suddenly, with the advent of multicellularity, a whole new set of genes and regulatory/developmental pathways is required. And there is no way for the transition to happen gradually; there's no selection pressure for the advanced functions until they are required (all of them!) in fully functional form. These functions/rules are part of what define the system, not what can change within the system. In a real sense, these rules determine how the system is "allowed" to change. Thus, they exist at a level more fundamental than mere genetic mutation.
Precisely because the new organism is so novel, natural selection cannot drive its evolution; it cannot select for a future goal. It can only modify what already exists; it cannot bring into existence a completely new structure or system. Yet that is what we have in the case of multicellularity.
Just as a side note, I want to deal with the inevitable question about co-optation that comes up: what if all the genes needed for multicellularity were coopted from genes providing some other function? The problem with this argument is that it has to posit that all the new genetic material was pre-existing in the ancestral cell (remember, about an order of magnitude increase), and had to be performing some other cellular task (otherwise the genes would become nonfunctional pseudogenes and virtually impossible to resurrect). Now, suddenly the organism becomes multicellular and all these genes have to switch to their novel functions simultaneously to provide the needed functions. Not only must they switch in a coordinated fashion (remember this is all driven by chance alone, supposedly), but they must not disrupt whatever their original function in the cell was. Keep in mind that they could not be performing the "multicellular" functions before the advent of multicellularity, so there must have been a change of function. And they must not disrupt any vital cell functions with their new function. Once you begin really thinking about the cooptation scenario, you realize that it's not really a viable option--it's little more than wishful speculation.
You might be interested in knowing precisely what contradiction was overcome in coming to the innovative multicellular organism. I'm not entirely sure we can always reverse-engineer the technical contradiction or problem that was solved in generating an invention, but I can venture a guess. I'll assume that the inventive drive is to make "bigger" organisms (physically bigger), that somehow there was a selective advantage to becoming larger. Perhaps bigger organisms can outcompete their smaller relatives somehow. Perhaps they can eat them. But as organisms become larger (still single-celled), they will reach a technical contradiction: they cannot diffuse oxygen in, and waste products out, fast enough to keep up with metabolic requirements; the problem is that the surface area to volume ratio falls below a critical point. Somehow, the organism needs to get big (that's the selective drive), and yet stay small (for gas diffusion). There's nothing in the pre-existing hypervolume of possibilities that allows the organism to overcome this limitation. Much like the leap to a new hull design in the icebreaker example, we can imagine an inventive leap to the multicellular organism, in which the "big" organism is constructed of many small, semi-autonomous cells. That way, the organism is both big (for size and competition advantage) but small (for gas exchange). Here's another example: sexual reproduction is a beautiful solution to the technical contradiction that the gene pool needs to be interconnected (to promote genetic variation) but also discrete (since each organism is separate and autonomous). Sex, with recombination and gene mixing, allows for the gene pool to be both interconnected and discrete at the same time. So the inventive solutions overcome technical contradictions by satisfying both sides of the equation at the same time--doing the seeming impossible by resolving a logical contradiction. They promote a positive-sum game in which all sides get what they want, but they have to step outside the limitations of what currently exists to reach the possibilities that lie beyond. Trial and error, and the Darwinian mechanism, lack the resources to overcome these contradictions because they cannot make the intuitive leap to a distant solution and instead founder when fitness (or overall function) begins to decline.
John Bracht
[ 16 July 2002, 01:49: Message edited by: John Bracht ]
IP: Logged
|
|
John Bracht
Member
Member # 5
|
posted 16. July 2002 01:44
Elend,
Your point about the Universal Touring Machine (UTM) was interesting but I think you missed a key flaw: a UTM doesn't just go around solving any particular problem it comes across. It has to be specifically set up (encoded) with the parameters for the problem to be solved. And the setting up or encoding of the UTM is never done by the UTM or by random chance; it's done by an intelligent agent who wishes to solve the problem. It's this setting up of parameters beforehand (parameters that allow problems to be solved) that I argue cannot be done without an inventive, intelligent source. So no, the UTM couldn't go from the Steiner problem to generating interference patterns without intelligent input--intelligence is required for re-setting the encoding of the problem into the machine.
In fact, much of my paper is dedicated to dealing with algorithms, and the computers these algorithms run on are basically Turing machines. Turing machines run algorithms. And it's the encoding of the fixed parameters that makes all the difference with algorithms.
John Bracht
[ 16 July 2002, 01:46: Message edited by: John Bracht ]
IP: Logged
|
|
Frances
Member
Member # 169
|
posted 16. July 2002 02:24
John Bracht hypothesizes about the origins of multicellularity. I would like to point John to the interesting experiment of the Chlorella vulgaris. In an accident, a predator was released. What was found was that while the predator managed to almost kill off the whole population, a few had managed to survive by aggregating as multi-cells (8 cells in total).
This number seemed to be a trade-off between protection and metabolism requirements. The cells all were similar, no specialization occured (yet).
Was it the 'invention' of gene regulation which helped the differentiation?
There is an interesting paper on this which looks at cooption and conflict in evolution of multicellularity.
What is interesting is that indeed the solution found had to deal with two opposing or conflicting goals. So all the opportunity for innovation to happen here. Could evolution have played a role here? I see no reason to reject evolution as I believe it has been shown that evolution can explore search space beyond the original dimensions, either through gene duplication or through a integration of the gene regulation with the morphological changes.
The Volvox theory is not the only one
John seems to be fixated on the idea that jumps to leap to a distant solution are the only truely innovative ways. But I would argue that 1) John has failed to show that RM&NS cannot bypass certain parts of the fitness landscape to reach solutions (as I have shown there are ways for RM&NS to by-pass or smooth the fitness functions) and 2) that John has failed to show that such leaps cannot be reached over a longer time via a more trial-and-error route.
Perhaps evolution was 'front loaded' in the sense that the simple original organism(s) had to build up a repertoire of capabilities which would help them become so succesful. Gene regulation, co-option, replication etc were basic requirements. It seems that by combining these through perhaps non-linear feedbacks, evolution could finally bypass many of the obstacles. Certainly the Hox genes for instance seem to suggest that something like this may have happened. Do we know all the details? Nope. Does this mean that evolution could not possibly 'innovate'? Certainly not to the extent of innovations as found in nature. What we do see is that evolution is more constrained which is why so much of life forms nested hierarchies and many innovations seem to reuse bits and pieces of other systems.
Perhaps not that 'innovative' in the sense required for patents and surely not as profitable but over the long term perhaps quite effective nevertheless.
To conclude: While John may have pointed out some problems with some simple (simplistic) models of evolution, more advanced versions of algorithms suggest that 'intelligent behavior' may have arisen quite naturally and that by combining control and morphology, despite an increase in search space, better solutions could be found by bypassing or smoothing the fitness functions.
Evolution of evolvability may have been a key ingredient for this to happen.
IP: Logged
|
|
Elend
Member
Member # 326
|
posted 16. July 2002 10:31
John,
In regard to the UTM being set up. I just want to point out that there is a difference between a normal Turing Machine (solving a specific problem) and the Universal Turing Machine. You are right in saying that the states of the Turing Machine have to be set up before the machine starts processing the tape. Yet, for the UTM this is done at run-time, through the tape that feeds the machine. The tape contains both the program (what the machine should do) and the data. For example using a UTM together with the binary encoding of all integers covers all possible Turing Machines (or all computers with their programs). All the "intelligence" has to be on the tape. If one regards the tape as chromosomes in a GA (with variable length) I see no reason why a GA can not come up with "novel" algorithms or inventions, granted it might take a considerable amount of time.
I hope I explained myself better.
IP: Logged
|
|
Janitor@MIT
Member
Member # 125
|
posted 16. July 2002 11:26
My perspective, albeit sympathetic, is quite different from John Bracht’s (?).
The “evolution of evolvability,” raised explicitly by Frances, but implicit in everything said so far, is a key question for me. It relates to Darwin’s ambivalence over “pre-adaptation.”
Darwin’s intent was not merely to subvert a natural theological perspective (“special creationism”), but also to distinguish his theory of evolution from predecessors which (with one exception) relied on such “providential,” “teleological,” or “law-like” mechanisms or directionality (all falling under some “orthogenetic” banner). Nonetheless, Darwin’s peers pretty much rejected (or debated vigorously) his perspective for ~60 years, while opting for some form of “orthogenesis.”
Throughout the middle decades of the last century this idea was definitely a minority opinion (as the Neo-Darwinian theorists re-established a sort of “Darwinian fundamentalism,” they were even more “Darwinian” than Darwin). Even so “mainstream” theorists have always played around the edges of the idea (orthogenesis) with considerations of “canalization” of development, epigenetic inheritance, constrained optimization, pleiotropic-epistatic interactions, etc.
Usually these considerations have taken the form of more-or-less well-defined constraints or limits on the process, without explicit consideration that in a dynamic process model (a true process model) a constraint always imposes an intrinsic “directionality” upon the process. The impression is that even though such “constraints” would appear to be a possible avenue for solving some particularly stubborn problems, such as panchronism, convergence, parallelism, etc., the theorists are uncomfortable 1) with the orthogenetic “flavor,” not being quite consistent with Darwin’s intent and 2) the idea that constraints are not “real,” i.e., being imbued with Darwin’s sense of evolution at play in the world of limitless possibilities, potentialities, opportunities, etc. (Darwin was possessed of the “Compulsive Gambler’s Syndrome,” seeing oppressive “improbabilities” as positive “possibilities.” Admirably optimistic, but not very realistic.)
One positive development, published as a “Perspective” piece in PNAS several years ago was Kirschner, M. & J. Gerhart, “Evolvability,” PNAS 95: 8430-8427, July 1998. These authors emphasize a “constraint-deconstraint” view, focusing more on how intrinsic/evolved factors “facilitate” or “enable” evolution. (Although the authors don’t mention it, one of my favorite examples is the DNA-code itself, which seems to have some marvelous built-in features that facilitate evolutionary searches. E.g., GA’s are said to be distinguished from random search because “revisiting” is limited or forbidden altogether in its “design,” so that the search is always driven forward. The DNA-code works this way as well. So much for Mayr’s statement that evolution is not a “genetic process.”)
Kirschner & Gerhart offer many positive examples of the kind of “evolvability” they have in mind, but don’t touch upon the “broader issue” for “Darwinian fundamentalism,” pre-adaptation. As I indicated, life forms are at their very core “adapted” to adaptability, “evolved” to evolvability. So that evolutionary theorists while strongly averse (not to mention bizarrely “schizophrenic”) to “teleology” must eventually come to grips with the "teleological" issue of evolvability.
Let me relate to the evolutionary computation models of biological evolution: Some researchers, in noting the very problems John Bracht has highlighted, i.e., the “canonical” form is distinctly limited in its simplicity (is only ever “roughly” analogous to biology), its efficiency (ETA to convergence), suffers from significant drawbacks (stalls in local optima, “converge-and-purge”), and cannot escape its representational limits, its not open and compositional or “self-composing” (the “representation problem” generally), which is what John Bracht is treating—whatever “innovative” capabilities it has are invested in it by the “Oracle” who isn’t only its designer, but its “god of the gaps.”
For these researchers, the solution is to augment the search strategy with heuristics, learning, “self-adapting” operators, “fine-tuning,” “hand-tooling,” etc. In this the researchers whose intent is to somehow “mimic” biological evolution have simply reduplicated research already conducted extensively by engineers in the 1950’s: they are optimizing the optimization process itself. In other words, the problem ceases altogether to be a problem in “philosophical” evolutionism, and has been changed into one of design (!).
To me the dilemma for received theory is plain: 1) Darwin’s model simply cannot be reproduced in silico except in a way that highlights its intrinsic and significant shortcomings and 2) the solution to this problem is to explicitly encode evolvability-adaptability; to “pre-adapt.”
One might well argue that the models, despite the exaggerated claims made for them, are indeed gross oversimplifications. But then we note that their success is not merely a matter of “complexifying” them, but of designing them to evolve-adapt more effectively, following (unwittingly apparently) the precedent and direction the research took ~50 years ago in engineering design theory. (One might take this as an implicit design prediction, made fifty years before the problem was identified!)
So I suppose its one for the evolutionary epistemologists and philosophers. Darwin took a very narrow view of pre-adaptation, believing it to be “incidental,” or “fortuitous,” so to speak, and overall not a significant factor in his theory. (Certainly because he didn’t realize its extent, but also de-emphasized for personal and philosophical reasons. There is also an apparent contradiction, in that his theory did not rely on "innovations," and therefore must always rely on "pre-adaptations, if you see what I mean.)
Life doesn’t merely appear to be “designed,” it appears to be designed to evolve-adapt. Is there any possible way to reconcile this with a “non-teleological” perspective or without involving ourselves in an “ontogenetic paradox”? In this regard Darwin’s theory harbors a hidden paradox: If the sole criterion of evolutionary “success,” its “drive,” is to maximize survivability-reproducibility, and if the Darwinian process is itself a manifestly inefficient way to do this, then any species that “blunders upon” a more efficient evolutionary process (optimizes optimization) would “take over the world”! In the eventuality (inevitability?!) biological evolution ceases to be Darwinian! Am I nuts, or what?! EC simply treats the issue-problem in an ad hoc way, while attempting to “save the phenomenon.” But the only way to “save the phenomenon” is to dump Darwin on his arse!
[Sorry, Mr. Moderator, this turned into a novella. But I think you see how it relates.]
IP: Logged
|
|
Elend
Member
Member # 326
|
posted 16. July 2002 16:29
Janitor@MIT
quote:
One might well argue that the models, despite the exaggerated claims made for them, are indeed gross oversimplifications. But then we note that their success is not merely a matter of “complexifying” them, but of designing them to evolve-adapt more effectively, following (unwittingly apparently) the precedent and direction the research took ~50 years ago in engineering design theory. (One might take this as an implicit design prediction, made fifty years before the problem was identified!)
I don't see the GA models as simplifications but rather as restrictive and focused on the problem. Consequently, what you mean by "complexify-ing" them I see it rather as a generalization. Allowing them to adapt their performance instead of pre-coding it seems more like generalization. Everything allowing more freedom in a model is a generalization, a step closer to reality, even if the actual implementation is more complex. Note also that our implementations of physically accurate models may increase in complexity to be in fact more general, and closer to reality (i.e. newtonian vs. relativistic mechanics).
[ 16 July 2002, 16:47: Message edited by: Elend ]
IP: Logged
|
|
Frances
Member
Member # 169
|
posted 16. July 2002 20:27
Janitor: You raise some interesting points but Darwin's ideas have 'evolved' as well with our increasing understanding of genetics. I find it not surprising that evolvability has evolved as well.
IP: Logged
|
|
|
Printer-friendly view of this topic