|
Author
|
Topic: ID Predictions made by an Engineer
|
RBH
Member
Member # 380
|
posted 22. January 2003 11:27
In reply to me, Irving wrote quote:
Thanks RBH for the reply, but I believe your referrences again miss the point, L-systems and generative representations are all about evolving the layout, or footprint design not the integration issue.
and to Francis he wrote quote:
If you go back a few posts to the post containing the circuit board picture, and maybe a few before it, I've attempted to illustrate that scalable architectures are only possible with scalable components. The rule idea, as well as the computer model RBH has suggested deal with the arrangement of components, or how they scale. Not the inherent attribute of scalability designed into the component in the first place.
As I read the Hornby papers and references therein, and some other material on similar research, I'm not sure Irving's point holds. From the Hornby page: quote:
Our approach is based on the use of a generative representation as the method to encode individuals in the population. Unlike a direct representation of a design, a generative representation is an algorithm for creating a design.
The emphasis in that research is precisely on the kind of representation of individual components/entities. If those components/entities are described to the evolutionary algorithm as generative representations, they are more likely to evolve to hierachically modular structures (scaled architecture). The "layout" emerges from an evolutionary process if the substrate - the components - are described as generative representations rather than one-to-one gene-phene mappings.
Several issues strike me in this specific connection. There is the issue of level of analysis, the level at which components are amenable to the kind of generative representation in the L-system and hierararchically modular work. Digging down (in a reductionistic sense), at some level the structure of matter seems to be appropriate for sclaed architectures - crystal shapes in the large reflect micro-properties of the components of the crystals. So the component substrate for hierarchically modular structures is in some sense present at the most basic levels. I'm not arguing that is the source of the property necessary for, say, biological evolution to display the same macro-structure, merely that it's not foreign to nature.
Second, it's not necessary that the "inherent attribute of scalability [be] designed into the components in the first place." If the generative representation work has any relevance, it is to show that it's not the structure of the component that counts; it's the nature of the representation presented to the evolutionary algorithm that's important in whether hierarchical modularity or scaled architectures appear at a macro-level.
This topic deserves more time and effort than I can give it now, but I'll try to get back to it soon. I'm not at all sure what "inherent attribute of scalability" means in this context. I can think of properties that would render scalability more or less difficult, and representations that are more or less likely to lead an evolutionary algorithm to produce outcomes describable as scaled architectures, but I can't get my head around an "inherent attribute."
RBH
Added in late edit: What I think I'm suggesting is that "scalability" is not an "inherent attribute" of components; it is an attribute of the kind of representation of components that is accessible to the search algorithm.
[ 22. January 2003, 12:03: Message edited by: RBH ]
IP: Logged
|
|
Frances
Member
Member # 169
|
posted 22. January 2003 12:45
irving you ask
quote:
An aside question...it seems in order to get a discussion going (a debate it seems) every information, complexity, and design issue has to be placed in opposition to evolution. Why is that?
Perhaps it may be helpful to remind you of some of your arguments
quote:
It is my conjecture that the scalable properties of this image indicate beyond reasonable disagreement--intelligent design. I would further
consider (though may be persuaded otherwise) that this image is impossible to arrive at via random mutation and natural selection (RM & NS).
It seems that you are arguing here that scalable properties are 'beyond reasonable disgreement' evidence of intelligent design and could not be arrived at through evolutionary processes.
As RBH and others have argued, there may be good reason for disagreement with your premise.
An interesting thesis which addresses the evolvability of components
An introduction to Lindenmayer systems
quote:
The following pictures were created by the author, using a Genetic Algorithm with genotypes inspired by L-systems. The fitness function employed was based on current evolutionary hypotheses concerning the factors that have had the greatest effect on plant evolution.
 
A L-System parser has been developed which allows mutations to change the form
There are also good lecture notes on L-Systems which show examples of evolution of L-Systems.
[ 22. January 2003, 13:20: Message edited by: Frances ]
IP: Logged
|
|
Irving
Member
Member # 535
|
posted 22. January 2003 15:15
Just a quick response for RBH and Frances, I do have a day job and don't have a lot of time at present. I hope to further engage this in the future.
For RBH,
quote:
Second, it's not necessary that the "inherent attribute of scalability [be] designed into the components in the first place." If the generative representation work has any relevance, it is to show that it's not the structure of the component that counts; it's the nature of the representation presented to the evolutionary algorithm that's important in whether hierarchical modularity or scaled architectures appear at a macro-level.
With this I disagree, and it appears that Hornby also disagrees.
quote:
The main criticism of the evolutionary design approach, however, is that it is doubtful whether it will reach the high complexities necessary for practical engineering. Here we claim that for automatic design systems to scale in complexity the designs they produce must be made of re-used modules [1].
[1] Hornby, Gregory. S., Lipson, Hod and Pollack, Jordan. B. (2001). Evolution of Generative Design Systems for Modular Physical Robots.
IEEE International Conference on Robotics and Automation.
Neither the L-systems nor the model developed by Hornby address the integration issue, in fact Hornby scopes his research by saying that it only applies when using scalable components. His paper does not discuss how scalable components may be developed. I have another issue with the paper and the use of a high order language and the use of an operating system stack, but that is for a future topic on properly coding EA's(?) as it were...
For Frances,
quote:
It seems that you are arguing here that scalable properties are 'beyond reasonable disgreement' evidence of intelligent design and could not be arrived at through evolutionary processes.
You are correct in that I've been trying to position my concept in just such a way...so that I'd get responses from people in this forum. I also maybe mispoke...I'm not trying to imply that the concept has been developed enough that it currently "is beyond reasonable disagreement," I offered it as my "conjecture" that the concept (if developed fully) might offer something beyond resonable disagreement. For any confusion I am very sorry.
Thank You,
Added in late edit:
Again for RBH,
quote:
Second, it's not necessary that the "inherent attribute of scalability [be] designed into the components in the first place." If the generative representation work has any relevance, it is to show that it's not the structure of the component that counts; it's the nature of the representation presented to the evolutionary algorithm that's important in whether hierarchical modularity or scaled architectures appear at a macro-level.
What I believe this is saying, is that if the nature of the component is scalability, then the EA is more likely to develop a scalable architecture. If the nature is not scalable, then the EA is more likely to produce a hierarchical modularity.
quote:
Added in late edit: What I think I'm suggesting is that "scalability" is not an "inherent attribute" of components; it is an attribute of the kind of representation of components that is accessible to the search algorithm.
Looks like we need to spend some time working on defintions. To me one of the representations of a component is it's scalability. Scalability is not an inherent attribute of all components, it is an inherent attribute of components in a scalable architecture.
quote:
I can think of properties that would render scalability more or less difficult, and representations that are more or less likely to lead an evolutionary algorithm to produce outcomes describable as scaled architectures, but I can't get my head around an "inherent attribute."
I'm sorry we seem to be having difficulty with this. Perhaps if I knew your background a little, I might be able to cast the issue into something familar. Perhaps I may have developed a little tunnel vision in my field, and am unable to represent my ideas properly. If there are any other design engineers (or even integration engineers) lurking about who understand what I'm talking about, please jump in if you have a better way of conveying this...
When designing an integration project from scratch, you must first develop an attribute list for each of the components to be integrated. When reverse engineering a scalable architecture, the attibute of scalablibility is considered to be inherent in the scaled components...
[ 22. January 2003, 20:00: Message edited by: Irving ]
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 22. January 2003 20:11
This is starting to get confusing, and I suspect Irving and I are talking past each other a bit. As I see it there are several issues involved in the question of the evolvability of scalable architectures and the evolution of what Hornby calls hierarchical modularity.
First is the component level. It's at this level that you assert that there must be an "inherent scalability," a attribute or attributes of the components themselves that allow them to form the kind of complex structures we're talking about.
The second issue is the kind of representation of the problem and components in the evolutionary processes, the evolutionary algorithm. It is at this level that Hornby's work is applicable. His work is all about representations. For example the abstract of his thesis says quote:
First, properties of representations are identifed as: combination, control-flow, and abstraction. Using these properties, representations are classifed as non-generative, or generative. Whereas non-generative representations use elements of encoded artifacts at most once in translation from encoding to actual artifact, generative representations have the ability to reuse parts of the data structure for encoding artifacts through control-flow (using iteration) and/or abstraction (using labeled procedures). Unlike non-generative representations, which do not scale with design complexity because they cannot capture design dependencies in their structure, it is argued that evolution with generative representations can better scale with design complexity because of their ability to hierarchically create assemblies of modules for reuse, thereby enabling better search of large design spaces. Second, GENRE, an evolutionary design system using a generative representation, is described. Using this system, a non-generative and a generative representation are compared on four classes of designs: three-dimensional static structures constructed from voxels; neural networks; actuated robots controlled by oscillator networks; and neural network controlled robots. Results from evolving designs in these substrates show that the evolutionary design system is capable of ?nding solutions of higher fitness with the generative representation than with the non-generative representation. This improved performance is shown to be a result of the generative representation's ability to capture intrinsic properties of the search space and its ability to reuse parts of the encoding in constructing designs.
The third issue is what "scalable architecture" and "hierarchical modularity" mean, and are they different? Your last posting suggests they are. I'm not so sure.
The fourth issue, and one I don't think we're addresing yet, is whether the representation issue is relevant to biological evolution. As Janitor has repeatedly argued (sometimes to deafening silence ![[Smile]](smile.gif) ), it clearly is, and embodies a mare's nest of issues to be sorted out.
What I take from the various papers on the effect of the kind of representation on the evolution of complex structures is that the representation does indeed make a large difference, with representations that are non-generative being less likely to produce outcomes displaying scalable architecture or hierarchical modularity, while generative representations raise the likelihood that those sorts of structures will evolve. I've scanned several of the papers at the site referenced above, and scanned Hornby's thesis, but I'm going to have to read them a lot more carefully to fully understand them.
The question remains, though, of what kind of components are necessary and sufficient for those kinds of representations to operate. Is there an "inherent scalability" attribute of components (as distinguished from their representations) that is necessary? What are the properties of components that can be said to contribute to "inherent scalability"? That's not at all clear to me.
RBH
P.S. Having written this offline and then logged on, I see that Irving has discussed some of the same confusions in different words. I'll read that, think about it a bit, and perhaps generate another response. I too have a day job that often runs into the late night and my time is unpredictably available. The last paragraph of Irving's post, however, raises the question of whether the attribute of scalability is associated with the components or with the fact that they composed a structure characterized by a scalable architecture. If the latter, then there's a problem associated with independently defining those attributes.
[ 22. January 2003, 20:13: Message edited by: RBH ]
IP: Logged
|
|
Irving
Member
Member # 535
|
posted 22. January 2003 21:17
We may be talking past each other, but maybe we can make some progress as well? Let me try and address your four points.
quote:
First is the component level. It's at this level that you assert that there must be an "inherent scalability," a attribute or attributes of the components themselves that allow them to form the kind of complex structures we're talking about.
Close, but not exactly. Complex structures is too broad a term. I'm referring to scalable architectures in particular, not just complex structures in general.
quote:
The second issue is the kind of representation of the problem and components in the evolutionary processes, the evolutionary algorithm. It is at this level that Hornby's work is applicable. His work is all about representations. For example the abstract of his thesis says...
Yes, it is applicable as far as it goes. I had to surf around a bit and download a PDF file that describes a bit of the design of his GENRE program to see what he was doing. He hints at it a little in his abstract by talking about "labled procedures," you have to understand just what GENRE is trying to accomplish and recognize that when coding in a high-level language you need to separate what the EA is doing from what the compiler does for you...the use of global versus local variables in procedure calls, etc...). If your not careful, the compiler will perform a certain level of integration for you (as a function of the design of the program language itself, i.e. pushing and popping off a stack). This integration is a result of the compilation itself, and not a result of the execution of the compiled code.
Fortunately it appears that Hornby recognizes this and therefore makes the statement that the entire enterprise is contingent upon re-use components provided, and not that the program creates re-useable components in and of itself. (At least that's my disection of his program design...could be convinced otherwise if his source-code is available).
quote:
The third issue is what "scalable architecture" and "hierarchical modularity" mean, and are they different? Your last posting suggests they are. I'm not so sure.
If in "hierarchical modularity" all the modules are exact duplicates of each other, then it would be the same as a scalable architecture. However my reading of the term "hiearchical modularity" places no requirement that the modules be duplicates.
quote:
The fourth issue, and one I don't think we're addresing yet, is whether the representation issue is relevant to biological evolution. As Janitor has repeatedly argued (sometimes to deafening silence [Smile] ), it clearly is, and embodies a mare's nest of issues to be sorted out.
Exactly! I had hoped to avoid biological relevance for a while, but as I've stated, unless you cast your concepts in opposition to evolution, you don't seem to get much discussion. When Kyle7 started this thread, I looked at the topic and thought about how a design engineer would approach ID irrespective of biological implications. I've become convinced that there is a small population of people who will argue that RM & NS is capable of developing anything! Anything at all from mousetraps to pocketwatches...etc.
So instead of looking at biology and trying to see something that had yet to be explained by evolution, I decided to approach the issue from the opposite side. Is there any possible design that can't be produced by RM & NS? Step outside of people's subjective (and sometimes religious) passions and see if there is any conceivable limit to the power of RM & NS (biology or not). If we can define what RM & NS can and can't do, then by analyzing the qualities of the things RM & NS can't do, we can formalize certain aspects of ID theory--which I believe can apply to things other than biology.
Hence I don't know it's relevance to biology, but it might.
Perhaps Janitor may be able to advance the discussion here?
Irving
[ 22. January 2003, 21:21: Message edited by: Irving ]
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 22. January 2003 22:51
Irving wrote quote:
Fortunately it appears that Hornby recognizes this and therefore makes the statement that the entire enterprise is contingent upon re-use components provided, and not that the program creates re-useable components in and of itself. (At least that's my disection of his program design...could be convinced otherwise if his source-code is available).
Some of the source code is in an Appendix in his thesis.
Added in edit: It includes some of the source code for the generative-representation compiler, but he is not real big on internal documentation!
I'm not sure Janitor will clarify anything, and he may set the discussion back, but I've hollered at him that we're taking his name in vain over here.
RBH
[ 22. January 2003, 23:02: Message edited by: RBH ]
IP: Logged
|
|
Janitor@MIT
Member
Member # 125
|
posted 23. January 2003 12:43
“…what ID predictions could be made based upon engineering concepts? I'd propose that an element of engineering not sucseptible to selection pressures would be future expansion and possibly scalability.”—Irving
We are used to viewing evolution as designing for immediate utility, necessity, contingency, and the detection of a number of design “-ilities” (such as “scalability,” “reachability,” etc.) in life must certainly be in some sense illusory or paradoxical, because even though we may see or anticipate the obvious “selective advantage” of incorporating such features into the design, they are in a real sense designing for the future. (I believe Darwin himself made a statement about evolution, as he conceived of it, making no provision for the future, lacking "foresight," and always only adapting or fitting organisms for the here and now.)
But the expanded vocabulary of modern evolutionary biology, specifically the terms and concepts (global and less than precisely defined) of “adaptability” and “evolvability,” needs to be more closely examined for what exactly such terms are conveying. Terms like these may be unfamiliar to an engineer, but he is, no doubt, well acquainted with the ideas that they convey: provisioning for the possible, designing for the foreseeable!
Assuming a rather doctrinaire position wrt to biological theory (and I don’t know that anyone does this anymore, except in public debate), I’m not sure that ideas about “adaptability” and “evolvability” are easily reconcilable.
“Though life may exhibit scalability through random chance, it is unlikely that more than a few distinct features, randomly developed, would exhibit scalability. Is scalability an attribute that can be selected?”—Irving
Now your professional experience many be helpful here. Just over the last several years the discovery that genomes are organized into scale-invariant networks has become a hot topic. When I first did a little reading on the subject my first strong suspicion was that the feature was an artifact of sampling and that diverse organisms cannot all have the same global architecture. But my second impression was: What a marvelous adaptation for the future evolution of the system! The feature permits local optimization while maintaining a persistently, globally optimized state, facilitates intragenomic communication and control, is a highly scalable architecture, and is the solution to the problem of reachability in an exploded space.
That evolution may have blindly latched onto a solution of this sort is not so implausible (although there are real difficulties: “Uniformity in an environment susceptible to randomness is an indicator of design.”--Irving). But it’s interesting to consider why it is plausible—because if I were a particularly ingenious designer I would have done it no other way!
The “representation problem” is quite broad and touches upon issues from artificial intelligence to coding and programming theory to the accuracy of computerized models and simulations of biological phenomenon. Rather than clarify anything I’ll cast a veil over everything: Much of design, and certainly Irving and RBH are more experienced with it than I am, is the “negotiation” of the protocols of the system. Often I get the impression that people, unawares, think of design as magical, and certainly it is arcane and largely inaccessible. (Engineers appear to be largely indifferent to the “popularizing” mania that grips many natural scientists.) But every engineer knows that we don’t conjure material objects into existence according to our whims, we negotiate both with Mother Nature and the systems themselves. We are in a sense mediators of design and not the creators of it.
But certainly this is a strange idea. The way we realize together some functioning organization of matter is by arguing, arbitrating, and negotiating with it. In this process we are informing it, and by informing it we are conforming it according to a (our) logic of design. What ties together these exceedingly complicated heterogeneous agglomerations of matter and makes them work is the logic of design. The principle of design involved is so fundamental that we don’t even think about: abstraction. The material realization of abstract concepts, compositions, imperatives, declarations, and logics, which are in any case difficult (as no doubt the biologists are discovering) to “reverse engineer.” What ties it altogether is an idea(s) that we make real.
I’m really agreeing with RBH, that these design features are not inherent to their material instantiations or realizations but are in some sense supplied “from outside.”
So, if I haven’t made this hopelessly obtuse, a very fundamental question can be asked of biological theory. And I think we are doing that aren’t we?
(That’ll teach ya to “take my name in vain.” LOL)
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 23. January 2003 13:54
This is intended to be a process post, and therefore very brief, rather than a substantive post. Janitor wrote quote:
So, if I haven't made this hopelessly obtuse, a very fundamental question can be asked of biological theory. And I think we are doing that aren't we?
I vote that before we explicitly plunge into biology, we first get to some (more-or-less clear) notion of what we think about things like scalability, hierarchical modularity, and what is required/assumed by way of component properties, representations, and algorithms that (may or may not) produce them. After that we can begin to sneak up on biology, perhaps by enlisting a biologist or two who are willing to advise, and see what sort of model/analogy/metaphor the EA design stuff might be (or not be) for biological evolution. In other words, let's shoot first for understanding, and then let the blood flow! ![[Wink]](wink.gif)
RBH
Added in edit: Another thought is that this is by rights Kyle7's thread. Since he's our host, so to speak, maybe we should start another thread just on this stuff? Kyle7? Whatcha think about being hijacked?
RBH
[ 23. January 2003, 13:56: Message edited by: RBH ]
IP: Logged
|
|
Irving
Member
Member # 535
|
posted 23. January 2003 14:54
quote:
I’m really agreeing with RBH, that these design features are not inherent to their material instantiations or realizations but are in some sense supplied “from outside.”
Okay, I'll budge a little...the "outside" supplied design becomes inherent in the material instantiation when that instantiation is then used in further applications. When elements are combined into compounds (H2O), the compounds have inherent physical properties (freezing point, boiling point). These inherent properties are critical in the appropriateness (or even possible) use of the compound in certain design applications.
quote:
I vote that before we explicitly plunge into biology, we first get to some (more-or-less clear) notion of what we think about things like scalability, hierarchical modularity, and what is required/assumed by way of component properties, representations, and algorithms that (may or may not) produce them. After that we can begin to sneak up on biology, perhaps by enlisting a biologist or two who are willing to advise, and see what sort of model/analogy/metaphor the EA design stuff might be (or not be) for biological evolution. In other words, let's shoot first for understanding, and then let the blood flow!
Agreed! Caution that we don't get too broad that we fail to achieve depth.
quote:
Added in edit: Another thought is that this is by rights Kyle7's thread. Since he's our host, so to speak, maybe we should start another thread just on this stuff? Kyle7? Whatcha think about being hijacked?
Another good point, if I had started this thread I would have titled it, "Are the any limits to what RM & NS can design?" Though Kyle7's original topic "ID Predicitions made by an Engineer" is appropriate to the spirit of the approach.
Irving
IP: Logged
|
|
|
|
Janitor@MIT
Member
Member # 125
|
posted 24. January 2003 00:58
OK, sorry I’m not very helpful. But with Kyle7’s indulgence, probably I’m not understanding where exactly the difficulty is. Where does scaling enter the equation: at the level of the elements of the system, or at some higher level? Seems to me that the answer is Yes. LOL Scalability must certainly be an attribute (either intrinsic or designed) of the (some, many, all) elements of the system (Aren’t I now contradicting myself?), and the (scalable) coupling of the elements of the systems with each other and throughout the levels of the hierarchy at the interfaces, and the capacity to take full advantage of any rescaling, must be defined by the protocols, assuming simply that there must be either given or inherent some system of “legalities” (another “-ility”) about how objects can combined/recombined, merged, sorted, etc..
Keeping in mind that I’m not a biologist (and only play one on the Internet): In the physical implementation biology is quite different, using exclusively “smart materials” (and especially carbon) and is therefore distinctly scalable at the elemental level in its hierarchy. Some molecules are wonderfully “scalable” in that by simply adding to them we can extend the length ad infinitum. But scalability is, as Irving has indicated, a specified functionality and not merely additivity. But the complexity of typical biopolymers, the functional demands made upon them, and their integration into complex networks of controlled interactions, also limits their intrinsic reusability/scalability, just as complexifying any system constrains/deconstrains.
It’s plain that reusability/scalability is a fundamental organizational principle in biological evolution. (The HGP reports indicated that it is the most salient feature in the difference between yeast, worm, fly, mouse (?), man, reflecting the thinking, if not exactly in those terms, that has gone before.) But two opposed dynamics are at work here (as always, as in design, as in biological evolution). Reusability/scalability has its limits. Gunther Wagner and Lee Altenberg (“Complex adaptations and the evolution of evolvability,” Evolution 50: pp. 967-976 (1996). Treating the subject at the next (third?) highest level in the hierarchy.) have theorized that continued evolution requires a relaxation in the level of system integration (pleiotropy). (I’m citing from my bibliography, and not looking at the paper, but I believe they highlight the difficulties of affecting such “relaxations” via the usual evolutionary operations.) While Ancel and Fontana (“Evolutionary Lock-in and the Origin of Modularity in RNA Structure”) have shown, using an RNA prototype, that at the molecular-genetic level of the hierarchy (the second level?) selection imposes modularity by reducing “plasticity” (as defined by a thermodynamic stability index). Notice at this level modular design emerges as a functional constraint on plasticity = variability, which results in modularity = stability. (And see a PDF online by Slotine, “Modularity, evolution, and the binding problem: a view from stability theory,” for a “robotics” argument for the selection of stability as a key to the binding problem, i.e., the integration problem.)
So you can see a “sort of” design-evolution prescription working itself out here: We begin with elements with some intrinsic/designed variability/reusability/scalability. Imposing some functionality upon these elements reduces variability to modularity. Combining modules/submodules adds back variability, and presumably the previous steps are repeated. Further combining (affecting a rescaling) modules into integrated functional units successively over and across higher and higher levels of abstraction (polygeny/pleiotropy) results in a heterarchical organization. Presumably all this organization is coordinated in time-space by a still higher level of abstraction, the communications/control protocols—the OS. Voila! A genome! LOL
Now, Irving has suggested, tentatively and innocently (but not naively) that scalability is problematic on conventional terms. But it may be that scalability is a definitive and universal systems level design principle for biology. If it’s a problem, it’s a big problem.
E. Ravasz, et al, “Hierarchical Organization of Modularity in Metabolic Networks,” Science 297:30 AUGUST 2002 and Dassow, G. von & E. Munro, “Modularity in Animal Development and Evolution: Elements of a Conceptual Framework for EvoDevo,” J. Exp. Zool. (MolDevEvol) 285:307–325 (1999), also provide some interesting insights. (Notice in the footnotes of the von Dassow article the amusing tussle between the authors and a referee over certain terms and concepts.)
[I believe everything I’ve cited here is online, sorry no links though.]
Thanks, Kyle7.
(And thanks for the link, Frances.)
IP: Logged
|
|
Irving
Member
Member # 535
|
posted 24. January 2003 08:34
Thanks Janitor, Frances, and RBH for helping out with this...
quote:
Now, Irving has suggested, tentatively and innocently (but not naively) that scalability is problematic on conventional terms. But it may be that scalability is a definitive and universal systems level design principle for biology. If it’s a problem, it’s a big problem.
It may be a universal principle for biology; however, before combining various scalable architectures with various biological scenarios, I'd like to see if people on all sides of this issue can come to agreement on an upper-limit to what RM & NS can design. Granted biological components and materials offer complex and not fully understood material properties, yet at a certain level of abstraction it should be possible to define a limit to RM & NS as a result of certain fundemental principles of RM & NS.
From Janitor:
quote:
(I believe Darwin himself made a statement about evolution, as he conceived of it, making no provision for the future, lacking "foresight," and always only adapting or fitting organisms for the here and now.)
If it is possible to define what RM & NS is in principle, then it may be possible to define what is outside the capability of RM & NS in principle.
What has led me to scalable architectures is that scalable architectures are not just modular architectures, but replication architectures that work (not just models) at almost infinite replication AND work at differing levels of scale. In the last image I offered, notice that the architecture is more than just a simple L-system, the component design at the "leaf" level is the same design that works at the "branch" level, as well as the same design that works at the over-all architecure level. Compare this image to the excellent L-system derived examples offered by Frances above. (There may be more complex rule-sets that leverage off of L-systems that may result in the image...rule-sets that may be of specified complexity LOL).
Yes, I'm using leaf & branch analogies and the image does resemble a fern, but I only used the diagram to spark interest, since my circuit board picture failed...I'd like to bring biology and biologists into this later as RBH has suggested.
quote:
Combining modules/submodules adds back variability, and presumably the previous steps are repeated. Further combining (affecting a rescaling) modules into integrated functional units successively over and across higher and higher levels of abstraction (polygeny/pleiotropy) results in a heterarchical organization. Presumably all this organization is coordinated in time-space by a still higher level of abstraction, the communications/control protocols—the OS. Voila! A genome! LOL
Yes, Voila! LOL, I sort of see this heterarchical organization as a stage two discussion. Before we address the development and integration issues of modular design (with functionally specified modules)would it be simpler to first tackle the issue of whether or not it is probable that RM & NS would produce a modular design in which each module is a duplicate of the others (stage one)...and still works?
RBH suggests
quote:
I vote that before we explicitly plunge into biology, we first get to some (more-or-less clear) notion of what we think about things like scalability, hierarchical modularity, and what is required/assumed by way of component properties, representations, and algorithms that (may or may not) produce them.
RBH & I may have been talking past each other earlier, and I agree that a uniform denifition of terms and concepts needs to be arrived at before significant progress can be made. In grasping about for different analogies or scenarios to describe what I call "the integration problem" with RM & NS regarding scalable components...let me offer a fanciful story...
In 1978 the video arcade game known as "Space Invaders" arrived. Some would say it was a perfect "fit" for the late 1970's environment. It took the video arcade industry by storm. The Space Invader population expanded seemingly overnight. For a time, it dominated the video arcade market.
Now an enterprising video arcade manager wanted to have the biggest and best Space Invader's experience possible. His thought was to combine multiple Space Invader games into a sinlgle functioning game experience. He may have even used L-system modeling to determine optimal placement of the machines in his building.
Putting his plan in action, he ordered his first two machines. He received Space Invader machines serial numbers 35565 & 35567, and placed them right next to each other. However, no matter how many times he played them, they wouldn't play together. They never seemed to sync-up...the controls on the one machine never seemed to impact the game on the other machine in anyway. The frustrated manager then called his Space Invader's vendor and asked him, "What do I have to do to get my machines to work together? Why weren't the machines designed to plug-n-play with each other. Why didn't you anticipate that this might be something I might want to do?"
Okay, it's kind of a rough analogy and misses a lot of issues, but at a simple level it speaks of an important issue with scalable components in a scalable architecture.
Irving
[ 24. January 2003, 08:42: Message edited by: Irving ]
IP: Logged
|
|
Janitor@MIT
Member
Member # 125
|
posted 24. January 2003 11:34
“…I'd like to see if people on all sides of this issue can come to agreement on an upper-limit to what RM & NS can design… at a certain level of abstraction it should be possible to define a limit to RM & NS as a result of certain fundemental principles of RM & NS…”--Irving
I don’t know if this can be done, Irving. The idea of varying and selecting from amongst the variations is so broad that I can’t imagine that there is much in the phenomenological world that couldn’t be “explained” on this principle. Many feel that this is indeed the “explanatory power” of the principle, but the ironic criticism of it, oft repeated, is that a principle that can explain virtually anything and everything really explains virtually nothing. (But, of course, that depends on what you are willing to accept as an “explanation.”)
Accordingly I’ve pursued two tracks: 1) Showing that a principle of selection also explains the actions of any conceivable intelligent agent (per Darwin’s original analogy); 2) It never really suffices as a bare explanation even for those who believe in its “explanatory power”—they invariably fall back on the real explanation and that is that we can always rationalize an engineering design criteria/explanation for why this or that instance was selected. The appeal of the principle of natural selection is certainly seductive: It can explain virtually anything via an intelligent design rationale!
Some limits are explored however,e.g., in the Ancel & Fontana paper and it is interesting to set in some historical context. Early, pre-Darwinian theories of selection (Wells, Matthews, and Blyth, all independently) emphasized the opposite of what Darwin emphasized. These all emphasized that selection was, in effect, a reinforcement of system-intrinsic functional limits and constraints and did not make of it a mechanism for boundless, open-ended evolution the way Darwin did. They did not make of it a theory of evolution, as we would understand it (and don’t even appear in the lists of evolutionary theorists, even though they are plainly theories of natural selection). Ancel & Fontana show that imposing a function upon some (simple) system to be evolved results in a rapid and drastic reduction in its “plasticity.” It quickly converges upon a singular state and ceases to evolve any further. At each instance selection systematically eliminates from the pool of possible functional variants (even with replacement) from which the selection can be made until there is no longer any variation left. The “converge and purge” problem. As always there is a design strategy to escape these singularities—by relaxing the constraints (per Wagner & Altenberg, albeit on another level than Ancel & Fontana consider the matter), or by appealing to an “oracle” (such as a randomizer), e.g..
It’s a real question and problem, but I’m not sure that biology is at the stage of analysis where it can set some global limits on the power of natural selection. Nor are they wont to do so for some strange, to my mind, “cultural” reasons. In more public pronouncements this principle of selection is invoked as eliminating “intelligence” as an explanatory category for biological phenomena, and as simultaneously eliminating the “old argument from design.” But it actually does neither. There seems to be a lot of “wishful thinking” and “self-fulfilling prophesying” going on here. If anyone is suspicious that the principle of natural selection doesn’t do all that it is said to do, then I am sympathetic.
(“I'm not sure Janitor will clarify anything, and he may set the discussion back…” Ah, RBH already knows my work well. LOL)
IP: Logged
|
|
Irving
Member
Member # 535
|
posted 24. January 2003 13:35
quote:
“…I'd like to see if people on all sides of this issue can come to agreement on an upper-limit to what RM & NS can design… at a certain level of abstraction it should be possible to define a limit to RM & NS as a result of certain fundemental principles of RM & NS…”--Irving
I don’t know if this can be done, Irving. --Janitor
Great! As long as you don't know that it can't be done.
quote:
The idea of varying and selecting from amongst the variations is so broad that I can’t imagine that there is much in the phenomenological world that couldn’t be “explained” on this principle.
To be sure trees are falling in the forest over in the transcendance thread. Though everytime I watch 2001, I can't help but get the feeling that there was something special to do with that obelisk. At least the monkey's found it curious.
quote:
It’s a real question and problem, but I’m not sure that biology is at the stage of analysis where it can set some global limits on the power of natural selection.
Perhaps then it may be advantageous to look elsewhere than biology for the answer?
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 24. January 2003 15:43
My participation is going to be severely limited for a week due to obligations offline. Let me make just two remarks about Irving's post. He wrote quote:
Yes, Voila! LOL, I sort of see this heterarchical organization as a stage two discussion. Before we address the development and integration issues of modular design (with functionally specified modules)would it be simpler to first tackle the issue of whether or not it is probable that RM & NS would produce a modular design in which each module is a duplicate of the others (stage one)...and still works?
I'm real cautious about "whether or not it is probable" approaches because I never know what the appropriate PDF is and what assumptions inform its selection. I'm much more comfortable with a question like "What are the properties/attributes of elements and what are the aggregation/assembly/combinatorial processes (for lack of a better term for the stuff that puts elements together) that might lead to hierarchically modular (or scalable) architectures, and what kinds of elements and/or processes are unable to produce them?"
Second, "RM & NS" is far from the whole story in EAs. Recombination is definitely non-negligible, and the nature of the representation presented to the EA (as in the L-system stuff, which is a kind of primitive phrase structure grammar allowing embedding phrases) is clearly important. One of the things we've found in building applications using EAs that are used out there in the world to control real processes is that things like the kinds of mutations available (point, substitution, insertion, deletion, duplication), the way(s) in which recombination is allowed to occur (e.g., subsystem-preserving vs. random crossover point), and how selection and replacement works (e.g. generational population replacement vs. individual head-to-head competitions), all affect what comes out the other end of the EA. Even the size of the population and enabling (or not) isolated subpopulations to emerge and occasionally trade genes have an effect that can't be neglected. So I'm wary of "RM & NS" by themselves as the focus of attention.
And as a mostly irrelevant aside, I loved the definition of "heterarchy" I heard a century or so ago in grad school: "a democratic hierarchy." Thanks for using that word, Irving! It's nice to be reminded of those days.
RBH
[ 24. January 2003, 15:49: Message edited by: RBH ]
IP: Logged
|
|
|
Printer-friendly view of this topic