Uncommon Descent Serving The Intelligent Design Community

Dave Thomas says, “Cordova’s algorithm is remarkable”

Salvador Cordova
August 17, 2006
Intelligent Design
Share
Facebook
Twitter
LinkedIn
Flipboard
Print
Email


Dave Thomas is in a bit of a tizzy over my humble offering: Tautologies and Theatrics (part 2): Dave Thomas’ Panda Food. He responds at Pandas Thumb with: Calling ID’s Bluff, Calling ID’s Bluff. I thought I’d alert the readers at UD what horrible things I’m accused of, that I might be some sort of vile scoundrel. 🙂

[Dave writes:]

Imagine my surprise, then, when I found Salvador Cordova at Uncommon Descent spewing blatant falsehoods about this work. I was shocked – shocked, I say – to catch the UD Software Engineers in a lie. And quite a lie it is – with the help of mathematicians like Carl Gauss, I’m going to lift the veil from the obfuscations of IDers, and prove it’s a Lie, much as you would prove a mathematical theorem.
….
which the brilliant Gauss found useful

Thomas then correctly identified the formula I implemented via Genetic Algorithm:

[Dave writes:]

The Software Engineering Team of Uncommon Descent has been caught lying – Q.E.D.

Where did I ever claim there wasn’t smoke and mirrors involved in the gimmickery here? Fer cryin’ out loud, my post was talking about mathematical theatrics, and I presented that program as an example of gimmickery! I even alerted the reader with these words before presenting my program, “The following [are] computational theatrics”. Sheesh!

[Dave writes:]

As an exercise in Smoke and Mirrors, Cordova’s algorithm is remarkable

Well dog gone, he actually says something nice about my work. It’s REMARKABLE! 🙂

I fully take pride in the smoke and mirrors I used, I never pretended otherwise. In contrast, Thomas refuses to admit he’s also using smoke and mirrors in his GA. He pretends somehow his steiner-solving program should persuade us that mindless undesigned forces can hit specified targets.

Well, did he have some Chimpanzee create the fitness functions in his software for him? Without intelligent design on his part, his fitness functions will fail to guide the system to the intended target. He has thus snuck the answer his GA, much the same way I snuck the answer in my GA. At least I alert the readers of where the trickery is, but Thomas would rather have his faithful congregation at Pandas Thumb believe that mindless evolution can truly work magic.

As Haeckel said:

Evolution is henceforth the magic word by which we shall solve all the riddles that surround us.
Ernst Haeckel1

Instead of “abracadabra” the Pandas say, “evolution”. Ramen.

Comments
trrll asked: So where does that information come from? Is it in the few bits of the “shorter is better” fitness criterion? Or the program as a whole?
The information comes from the choices of the the intelligent designer(s) including the computer system designers. Whether they act through a surrogate like a computer does not negate the fact that the the source of the information (the reduction of uncertainty) regresses to an intelligent designer. And even if he does not know what the answer will be before hand, it does not negate that the information in the output channel still proceeds from his choices. The output of the program is merely an alternative representation of the information he put in. The "bits" in the software are not to be confused with the "bits" in the solution space. For example, I could write the following that will "generate" zillions of bits of CSI in an output channel until you shut the program off. The following will generate a long string of 1's:
main(){ while(1) printf("1"); }
The output (from a storage standpoint) has many more bits of information than the source program that generated it. But the information still regresses to the choice of the and engineering of the system. A ZIP file with X number of byte to store it may decompress to a much larger file of Y bytes. A program can in principle be miniscule to the output it decompresses to (the above program was such an example). CSI metrics for information apply a different technique for measuring information rather than rote storage requirements or even algorithmic complexity. This thread is not appropriate for that discussion. Maybe Part 3 of Tautologies and Theatrics, ok? The question whether undesigned nature can create such genetic algorithms in real life that cna originate novel CSI. The displacement theorem describes the likelihood as more remote than random chance on average. Thus GA's do not demonstrate that undesigned nature can evolve complex designs from scratch. No one has yet been able to demonstrate that a GA can popup out of nowhere on its own. NO ONE! Salvador PS (A long string of 1's can be generated by a stuck key on a computer keyboard. Such a circumstance might inadvertently generate a long string of 1's. But these are the pitfalls of computer arguements due to some of the artificialities computers induce into a simulation.)scordova
August 19, 2006
August
08
Aug
19
19
2006
08:51 AM
8
08
51
AM
PDT
"Do you genuinely not recognize the difference between a solution that requires that the final answer be known and one that only requires identification of a constraint that the answer must satisfy? Or are you just trying to confuse the issue because you are not confident of your ability to rebut the actual points raised by Thomas’s simulation?" trrll, for my part, I understand what you're saying, I just disagree with the conclusion. The simulation is over simplified and does not represent lifes diversity in this unique biosphere. A fitness function or genetic algorithm is dependent upon the a narrow viewpoint(contraints) for optimization and leads often to deadends. This is one of the actual problems with genetic algorithms. Without conditional feedback loops for multiple external stimuli and a reactionary pre-coded feedback loop, solutions can be suboptimal without even knowing it. A simple fitness function selected to favor certain reproductive success based upon one narrow contraint across large search spaces is only one function simulated, but it does not extrapolate to the complex interactions of higher level organisms or new morphological changes. Recognition of certain laws and the catastrophic consequences if they are modified only attributes to the precarious situation of which we find ourselves in a finely tuned plan. Thermodynamics does not only apply to this planet, but the universe as we know it. The contraints as you put it here on earth is actually what allows life to exist. We are living within multiple levels of hierarchical contraints which limit exposure to full planetary and life extinction. Otherwise, life can pop up anywhere, in any form, any planet. I think a more simple reduction is simulating eye color combinations; however, popping eyes on the back of ones head that is functional is quite another. There is a breakpoint in morphology, but this should not be confused with thermodynamics on our planet which actually allows life to exist. The fact is you're defining living contraints of the program, thermo being just one of the external considerations in your example. Your cost/efficiency ratios are but one of many conditions that must be optimized, not including interactions, immune systems, repair systems, catalyst(enzymes). But cost/efficient ratios can be overcome by larger energy input/output. So the contraint of such vascular systems is only limited like fish in a fish bowl or those released into a pond, a lake or ocean. These are size and energy considerations yes, but this is not related to morphology, only survival of each species. Famine can kill an elephant or a caterpillar. Optimization of neural, vascular pathways within each does not prove one can evolve into the other. Morphology to me is programmatic - designed. Optimization and variation based upon external stimuli within bounded contraints is but one part of the overall genetic program. The Steiner Tree problem mimics but one component and does not solve Fluid Dynamics contraints. Finally, speaking of chess, re: enzymes as catalyst and time contraints: http://news.biocompare.com/newsstory.asp?id=10433 hattip: linked to by Jonathan Sarfati, former New Zealand Chess Champion; http://www.creationontheweb.com/content/view/3547 I'd appreciate any links which may dispute the articles findings at Biocompare posted in 2003 on a study by Dr. Richard Wolfenden who claims it to be an enigma as to enzymes arising within the currently accepted universal timeline. He states, "As to the uncatalyzed phosphate monoester reaction of 1 trillion years, "This number puts us way beyond the known universe in terms of slowness," he said. "(The enzyme reaction) is 21 orders of magnitude faster than the uncatalyzed case. And the largest we knew about previously was 18. We've approached scales than nobody can grasp."" So, not only are we looking at thermo considerations, fluid, etc., but also catalyst responsible for all life forms which speed reactionary survival mechanisms to milliseconds.Michaels7
August 19, 2006
August
08
Aug
19
19
2006
08:23 AM
8
08
23
AM
PDT
Sal, et.al., LEt's face it, the PT crowd's little choo choo has gone way round the bend and off the tracks! Their language is more and more strident. Their use of the "IDers are lying" phrase is getting, well, tiresome. I keep saying this, but I'll repeat again and again until they get it: for a group that claims to hold logic and scientific reasoning so dear, it is constantly amazing how quickly they abandon it when their own logic, reason, claims and idea are challenged. IT reminds me of the title of Solzhenitsyn's "We Never Make Mistakes". That they actually believe that's the case is laughable and sad at the same time. Keep after them Sal. I'll take your logic and reason over their ad hominems, straw men and violations of the law of non-contradiction any day! Let us know when one of them says something reasonable!DonaldM
August 19, 2006
August
08
Aug
19
19
2006
07:47 AM
7
07
47
AM
PDT
Re #46 for an object to be designed (as opposed to appearing to be designed) it must not only satisfy the designer's purpose but a) should be the result of some activity by the designer (a heavy shower suits my purpose for watering the garden but I didn't design the shower) b) achieve the designer's purpose in the fashion that the designer planned (if I lay out a hose to water the garden and the garden actually gets watered because of a leak in the hose you can't say I designed the solution) Dave Thomas had no idea how the resulting patterns were going to achieve short paths so he did not design (b). They do however give the appearance of design because they look like the kind of pattern someone might have thought up if told to produce a pattern with a short path.Mark Frank
August 19, 2006
August
08
Aug
19
19
2006
06:09 AM
6
06
09
AM
PDT
Dave T.: "That’s like dismissing as nitpicking the objection that an alleged perpetual motion machine only needs a little bit of input energy to keep it working." The phrase "a little bit pregnant" comes to mind, too.j
August 19, 2006
August
08
Aug
19
19
2006
04:16 AM
4
04
16
AM
PDT
Me (30): "Salvatore" Uggh. Sorry, Salvador -- I was pressed for time. I really do know better. I've only read your name here about, what, maybe 2000 times? ofro (31): "It seems to me that you are implying that it is not possible to write a program that behaves in the manner postulated by blind/dumb/purposeless Darwinian evolution. In other words, one cannot write a code that behaves like this hypothetical, very mechanistic process because as soon as I write it, I have put in a non-Darwinian goal?" No. Darwinian evolution programs exist, just not any that design anything. Google "Darwinbots", for example. Mark Frank (35): "The power of a programme like David Thomas’s is not to simulate all aspects of evolution. It is simply to show that small mutations plus selection applied repeatedly can generate innovative solutions that give the illusion of design." It's not an "illusion of design" when something was intended. It's actual design. See definitions 1 and 2 (the original definitions) of the noun at www.m-w.com/dictionary/design :
1 a : a particular purpose held in view by an individual or group {he has ambitious designs for his son} b : deliberate purposive planning {more by accident than design} 2 : a mental project or scheme in which means to an end are laid down
Use of the word to designate a "the arrangement of elements or details in a product or work of art" is recent. Dave T.: "That’s like dismissing as nitpicking the objection that an alleged perpetual motion machine only needs a little bit of input energy to keep it working."j
August 19, 2006
August
08
Aug
19
19
2006
04:14 AM
4
04
14
AM
PDT
Re #38 (and others). I am not doing well at explaining my point. I will try a different approach. 1. Salvador and others write as though there were two things: a problem to be solved and a selection algorithm for solving it. (Then of course you can object that the algorithm has been intelligently designed to solve the problem). But they are not two separate things. Evolution is not trying to solve a separate problem from reproducing - reproducing/getting selected *is* the problem. It would be nice to simulate the complex, subtle and everchanging ways that organisms get selected in reality - but that is asking too much of a simulation. So we substitute a different selection algorithm. To that extent the programme is a very partial simulation of evolution. 2. It was perhaps unfortunate that Dave Thomas put his programme in the context of finding a Steiner solution. That is just a by-product of his programme. Imagine he had never mentioned the Steiner solution. The programme still works. 3. It is a trivial result that if you repeatedly mutate, select and inherit for attribute A then you will end up with a population that has more and more of attribute A. I don't think anyone on the list would challenge that or find it interesting. What is interesting about programmes like Dave's is that the "solutions" generated are novel (ie.e. not predicted by the writer of the programme) and give the appearance of being designed. 4. Of course intelligence is required to write a computer programme including a selection algorithm. This is because it is an artifical simulation of reality. In the same way you need intelligence to write a climate simulation programme - that doesn't mean the climate was intelligently designed. 5. I think this is the real issue: Only some selection algorithms lead to novel solutions that appear to be designed. So a selection algorithm on the lines of "accept if closer to the sum of the first 1000 digits, reject if further away" is unlikely to lead to a novel solution. After all the solution is only going to be a number. So some selection algorithms lead to novel solutions which appear to be designed; others don't. So the real question is does the ineffable selection process of natural selection fall into the first category or the second? That's not a question that will be solved by computer programmes. All they can do is show that at least some selection algorithms are able to generate novelty.Mark Frank
August 19, 2006
August
08
Aug
19
19
2006
12:29 AM
12
12
29
AM
PDT
However, such problems are rare in which a GA can “design” a solution. In contrast, a GA in and of itself would be a very poor way of designing chess strategies. Chess software does not employ GA solutions as much at employs search heuristics, and brute force searches, and good guesses (the technical term is “static evaluation”). Thus, contrary to Haeckels claim that evolution is the word that can solve all our problems, in the world of engineering, that is not true, and its not even a common solution!
It would almost certainly be possible to design a genetic algorithm to solve chess problems. One could start with a set of genes regulating the connectivity of a neural network, train each network over the equivalent of a decade or so of human chess competition, and then allow it to "reproduce" based upon its chess ranking. Of course, the resultant program might well run too slowly to be of much use, considering the requirement to simulate the firing of perhaps millions of neurons. Chess, after all, is a relatively simple game, amenable to brute force look-ahead strategies. Similarly, your trivial math problem is hardly worth the effort of a genetic algorithm. But when you get to really difficult problems—NP complete problems like the Traveling Salesman—then the power of an evolutionary approach becomes more evident. I am curious about your perspective regarding the information generated by the simulation. After all, one could add in a random number to generate random arrays of nodes, and the program could go on indefinitely generating designs for Steiner networks, outputting a huge amount of information. And it is hardly trivial information—even intelligent human beings have difficulty designing efficient networks. So where does that information come from? Is it in the few bits of the "shorter is better" fitness criterion? Or the program as a whole? Even that doesn't come close to the number of bits output by the program. Do you deny that the solution to a network problem (and presumably, all other NP Complete problems) constitutes real complex specified information? Of course, there is a sense in which the solution to any problem can be said to be implicit in the problem definition. But in this case the design of every viable organism can reasonably be said to be implicit in the laws of physics and chemistry, in which case a hypothetical intelligent designer of life would no more be adding information than an intelligent traveling salesman who works out an efficient route.trrll
August 18, 2006
August
08
Aug
18
18
2006
04:52 PM
4
04
52
PM
PDT
Algorithms like Dave Thomas’s do an excellent job of optimizing for a particular physical variable. In the case of the Steiner solution, the algorithm optimizes for shortest length or smallest area. And there is no doubt about the results: A computer running the algorithm can find an optimal solution faster than a human being. But notice that the choice of what physical variable the algorithm optimizes for is not decided by the algorithm but built into it by the programmer. With respect to the algorithm, the “most fit” solution is always the one that minimizes length or area. Now “fitness” for Darwinism means “survival”. But survivability does not map to a single or even a small set of physical variables. It maps to a virtually infinite set of physical variables. For one organism, increasing length might increase survival, for another decreasing length might increase survival, for another survival might have nothing to do with length at all. In fact, it may involve an entirely novel physical characteristic, which is the point of evolution in a creative sense in the first place.
The choice of what biological variables to optimize for is not something that is decided by evolution—it is something that is imposed by nature. For example, the Steiner network problem models a biological problem that organisms must solve, the problem of "designing" efficient networks. Networks are used heavily in biology—neural networks, vascular networks. They need to connect crucial targets within the body, but thermodynamics imposes an energetic cost per unit length. An organism that grows excessively long vascular and neural pathways will be at a disadvantage compared to an otherwise identical organism that grows its pathways more efficiently, because it will require more nutrition and be at greater risk when food is scarce. There is no choice open to evolution of what physical variable to optimize for, because thermodynamics is inflexible; whatever else the organism might do to improve its resource utilization, mutations that result in shorter, more efficient networks will always enhance survivla. So the programmer of the simulation is effectively playing the role of the laws of thermodynamics. Allowing the fitness criterion to mutate, as some have suggested, is clearly unrealistic, because the laws of thermodynamics are fixed. It may well be that evolution won't work in a universe in which the laws of nature mutate randomly from moment to moment, but that is not the question that the simulation is designed to test.trrll
August 18, 2006
August
08
Aug
18
18
2006
04:34 PM
4
04
34
PM
PDT
Tom, Thank you again for sharing your expertise. If I may offer a couple anecdotes and the invite you to make some more comments for the benefit of the readers, especially those without a PhD in Computer Science like yourself. :-) I was very fascinated with Artificial Intelligence (AI) at first, and then, its very ambitious hopes seemed to fall short of expectations. At first, it was quite amazing to see these "AI" programs play chess, play checkers, and so forth, but at the end of the day these were not really thinking or creative machine, at least not in the way we conceive of what a thinking being is.... In fact, the definition of AI is now somewhat nebulous.... All this to say, it seems to me that systems that somewhat mimic an intelligent designer in its activities still require a great deal of front loaded intelligent design to give the system all its marvelous abilities. This of course has bearing on the issue of how much intelligent front loading nature would require for nature to be able to create life from scratch and confer the complexities we see today. If hypothetically life came about through a process of selection and mutation, how rare would such mutations and events have to be? If functional designs we see in the biological world as are easy to come by as fog in London, then we should not at all be amazed. However if such events are rare, then one would have to wonder: 1. whether "intelligently designed selection" is a more appropriate metaphor versus "natural selection" 2. whether intelligently designed prescribed evolution is a better metaphor for some kinds of evoltuion 3. some intial amounts of special creation of the first life (even Darwin believed in limited special creation) 4. some combination of the above The selection that I see in GAs fits into category #1. That is, Genetic Algorithms do their thing because an intelligent agency is designing the selection that is used. The selection in such an algorithms is intelligently designed to achieve a goal. Self-extracting zip files correspond to #2, where data de-compresses (de-represses using Davison terminology). #3 would correspond to the existence of the computer systems in the first place. A self-extracting ZIP file of a GA would be analogous to #4. The point is, it seems to me there is not a lot of room for thinking such events would be common place, and because of their rarity, intelligence seems at least a plausible candidate for their ultimate origin. That’s kind of where all this debate falls. Along the lines of a successful Genetic Algorithm’s (GA) rarity, let me offer some thoughts as I think it speaks of the improbability of a successful evolutionary pathway for life. Here are the requirements for a human GA to succeed: 1. The problem has to be solvable 2. The problem is amenable to be solved by an evolutionary algorithm 3. The problem can be tractably analyzed such that a trustworthy selection strategy appropriate to the problem can be formulated 4. An intelligent agency is available to put the evolutionary system together such that evolution can happened toward a desired goal The first example that comes to mind for all the requirement being met is a GA that solves the Travelling Salesman Problem
Given a number of cities and the costs of travelling from any city to any other city, what is the cheapest round-trip route that visits each city exactly once and then returns to the starting city?
However, such problems are rare in which a GA can “design” a solution. In contrast, a GA in and of itself would be a very poor way of designing chess strategies. Chess software does not employ GA solutions as much at employs search heuristics, and brute force searches, and good guesses (the technical term is "static evaluation"). Thus, contrary to Haeckels claim that evolution is the word that can solve all our problems, in the world of engineering, that is not true, and its not even a common solution! For example, I provided 5 programs that pumped out the final string of "500500", but the most obtuse-looking by far was the genetic algorithm. The problem was simply not amenable to a GA. I had to actually concoct an extremely circuitous way for a GA to compute the answer. But beyond the rarity of GA’s an problems amenable to GA’s is the fact that every GA we have seen come into existence from scratch came about through intelligence or pre-existing life. Let us hypothetically assume an evolutionary route via selection was taken. Would that route require design? My initial response is, “yes”.scordova
August 18, 2006
August
08
Aug
18
18
2006
03:37 PM
3
03
37
PM
PDT
BC, I wrote my first FORTAN program in 1981 on punch cards, fed into an IBM370. One single comma out of place, or a missing punch card could spell disaster for the programs Compilation or infinite series of loops or serious wrong answers and simulation failures. So your decades of experience holds no sway over me. I built along with my buddy at the time a Texas Instruments computer and have programmed in Assembler, PL/I, COBOL, Basic and other 4th generation languages. I ran Differential Equations on my poor little VIC20 to score extra credit in third year Calculus. I worked with a leading edge software company for 7 years who revolutionized the document industry supplying both IBM and Xerox with key software components that included Image Compression algorithms, before starting my own consulting firm for another 7 years working with fortune 500 companies in securing cost-savings of millions of dollars in the elimination of outdated legacy document management systems that included intelligent, knowledge based systems that built on-demand technology for real-time solutions. What took weeks, we reduced to seconds and allowed for interactive response. We did this across broad-based input and output systems, including translation with centralized or network processing. Now, all bravado aside and with all due respect to your software engineering skills, the problem is not in understanding what the program is doing, but extrapolating from it grand illusions. One little snippet of code does not provide evidence for a materialist solution. We are only just unraveling the mystery of life. I once believed in simple evolutionary steps before I became more interested and read with intense interest about Genetics, cell structure, regulatory systems, the repair mechanisms, signal processing and the host of other complex sub routines being performed in the most simplest of life forms. What I find most extraordinary is the claim that code on a computer simulating the most basic of instructions can be evidence of full blown evolution as pushed by NDEs. You are talking to one of the converted, a former evolutionary believer. And much like someone who quits smoking cigarettes, the awful smell of a once bad habit reminds one all to well why the habit was dropped. Why it may "look cool" at one point in youthful experimentation among peers, the cost far outweighs any temporary benefit(peer-acceptance), and as society awakens to new information, youth matures, hindsight shines light upon the folly of past mistakes and uninformed or bad input. While I in no way equate myself on the level of Dr. Sanford, or for that matter any scientist or PhD. here, it was in reading more information, not less which transformed my opinion and influenced me to cross-check unanswered questions on evolution. The more I read, the more I find RM&NS cannot account for the diversity of life as we see it or its complexity on the micro levels and genetic code. The task at hand is enormous and systems biologist all agree that much of the knowledge base and coordination must be systematically re-programmed for better understanding to unlock the code of life. I'm not the one saying this, but leaders in the fields. The simulation process is still directed and not by simple materialist processes, but by complex design that includes conditional processing logic. Anytime you allow variables into a system upon which that system must conditionally react to multiple inputs and levels, then pre-programmed responses are found which include the inherit capability to conserve itself. Simple replication and reproductive differentials does not merit extrapolation for all we see in cellular design and genetic code. What we see in life is Code Conservation wins and mutations End. The simulation is to simplistic and does not account for accumulated intelligence or myriad external conditions. At the same time we see how precarious life forms are, that adaptation is limited, bounded and mutations do not suffice for lifes morphology, but more often lead to extinction. These are plain, simple, straight forward observations. All laboratory experiments to this day, tell this tale of woe for random mutations limits and destructive force. If evolutionary lab results produced any observational, repeated outcome to the contrary it would be trumpeted thru out the entire world. Our way forward in unlocking lifes code is thru the Design Paradigm. Simulations and mathematical concepts will apply, but not without overall design. Again, we are confusing random mutations with to much power. Confusing the designed purpose of bacteria's promiscuous role and that of complex multi-cellular organisms.Michaels7
August 18, 2006
August
08
Aug
18
18
2006
03:15 PM
3
03
15
PM
PDT
taciturnus wrote:
Now “fitness” for Darwinism means “survival”. But survivability does not map to a single or even a small set of physical variables. It maps to a virtually infinite set of physical variables...So a Darwinian algorithm needs to do more than optimize for a prespecified physical variable. It needs to figure out what variable to optimize for.
Dave, Darwinian evolution doesn't need to "figure out" what variable to optimize for. Mutations occur across the genome. Whichever ones promote survival and reproduction are retained, regardless of what "physical variables" they affect. At no point is it necessary for evolution to single out a particular variable for optimization. For a genetic algorithm, the difference is that we know the precise problem we are trying to solve. The colors of the lines in a Steiner network are obviously irrelevant to its optimality, so the programmer doesn't bother to mutate color, focusing instead on the relevant variables.sophophile
August 18, 2006
August
08
Aug
18
18
2006
03:08 PM
3
03
08
PM
PDT
Ofro, Relax, you were not my "target" for calling Salvador a liar, though certainly I see how you can make such an "assumption" since my initial aim signaled an address at first to you. My "assumption" was that all knew it was Thomas I was "targeting" in relation to the programs limits and his false accusation against Salvador. Thus I stated the following... "What was uncalled for is calling Salvador a liar. Anyone reading his initial premise understood exactly what he intended in his reponse and it was not misleading in the least. I think Dave’s intentions while I’m sure well intentioned at first lead the lay person to believe design in the simulation..." The first line should read, "What was uncalled for is Dave Thomas calling Salvador a liar". Again, when we go on "assumptions", then information loss(in this case, one or two symbolic names) leads to false-positives and even more "assumptions" which lead to chaos if intelligence does not intervene to provide corrective measures and address the specific misunderstanding between sender and receiver. A result of the random mutational loss of one or two keywords is corrected not thru a materialistic blind mechanism, but thru appropriate intelligent keywords being restored, symbols shared by both sender and receiver to resume a more directed and informed pathway of discussion away from miscontrued or uniformed "assumptions". In this case for you and possibly others; symbols as names, though not hard-coded or hard wired as the original genetic datastream does make a difference. Blind processes built upon materialistic assumptions can never lead to this obvious correction in our misunderstood communications. The mechanisms in place to ascertain your feedback, retrieve, coordinate appropriate reponse, and send new symbolic data expands enormous programmatic and conditional responses which can vary dependent upon a willful recognition and anticipation of feedback from you the original sender or others. Our very misunderstanding, interaction, correction and recognition demonstrate materialistic processes alone woefully inadequate for the reactions to information found within and outside of us. We are simply at the very beginnings of the understanding of lifes animated dance. And what a wondeful choreographed dance it is.Michaels7
August 18, 2006
August
08
Aug
18
18
2006
02:44 PM
2
02
44
PM
PDT
Mark, I wouldn't expect Dave Thomas's algorithm to model all aspects of evolution. But when the selling point of Darwinian evolution is that it involves no intelligent intervention at all, it's not mere nitpicking to point out that the algorithm has been deliberately designed to optimize the one physical feature that has been specified a priori as equivalent to survival. That's like dismissing as nitpicking the objection that an alleged perpetual motion machine only needs a little bit of input energy to keep it working. Cheers, Dave T.taciturnus
August 18, 2006
August
08
Aug
18
18
2006
02:39 PM
2
02
39
PM
PDT
steveh, I do apologize for not preserving what you wrote last night. I have opened a thread where in the future, if I make editorial decision, I can deposit yours and others work so you and others can have access for it. Sorry about what happened earlier, and I hope you'll take this as a sign that I hope to hear from you again. See: [off topic experiment] cutting room floor (version 1) Salvadorscordova
August 18, 2006
August
08
Aug
18
18
2006
01:03 PM
1
01
03
PM
PDT
Salvador, "What my position regarding fitness function adequacy is the following tautology: supplying a fitness function that will solve the problem is supplying a fitness function that will solve the problem, and supplying a fitness function that will not solve the problem is supplying a fitness function that will not solve the problem." Then may I conclude that the "supplying" is irrelevant? The fitness function is what it is, regardless of how it originated. If there is complex specified information in the function -- the function itself, not the implementation of the function -- then it is there to be detected and measured. If I paint a picture of the Cliffs of Dover, there is CSI in the painting, but presumably not in the cliffs. If I write code in a programming language to describe a function that sums the Euclidean lengths of line segments, the CSI in the code is not the CSI in the function. The GA gains information from the function, not the code implementing the function. There are infinitely many correct implementations of the fitness function, and the GA behaves identically with all of them. You make it sound as though the problem is extrinsic to the fitness function. The problem for the GA is to maximize the fitness function itself. Dave perhaps muddied the waters by introducing the notion of a Steiner solution too early. The problem is not to find a Steiner solution. The problem is to find a network that connects the fixed nodes with minimal sum of link lengths. The (un)fitness function could not be much more straightforward: If the network connects all fixed nodes to one another, return the sum of link lengths. Otherwise, return a high "length" value. The GA usually finds networks of high fitness, but rarely finds an optimal network (i.e., Steiner solution). "Because not all fitness functions will solve the problem, in fact the overwhelming majority of possible function will guide the algorithm away from the solution." I have proved that almost all fitness functions are algorithmically random, or very nearly so. A highly random fitness function does not guide toward or away from anything. Intuition might suggest that such a function is hard to optimize, but the opposite is actually true. See "Optimization Is Easy and Learning Is Hard in the Typical Function," http://members.cox.net/tom.english/cec2000.pdfTom English
August 18, 2006
August
08
Aug
18
18
2006
11:54 AM
11
11
54
AM
PDT
Re 33. Dave T. Your proposed test is interesting. But you need to be truly analogous to evolution. The Darwinian approach is to generate objects (bit strings presumably) at random, subject them to the unknown algorithm, take those that survive, and slightly mutate some of the survivors, repeat indefinitely. What is the realistic alternative? As you say the real conditions for survival are ever-changing and unpredictable. So it is unreasonable to suppose the designer knows what those conditions are. So the designer should design an object that will do well under an unknown algorithm. I wonder which approach has the best chance? Actually all the above is only marginally important(but kind of fun to make the comparison). The power of a programme like David Thomas's is not to simulate all aspects of evolution. It is simply to show that small mutations plus selection applied repeatedly can generate innovative solutions that give the illusion of design. I think it illustrates that point rather nicely.Mark Frank
August 18, 2006
August
08
Aug
18
18
2006
11:14 AM
11
11
14
AM
PDT
"but I just wanted to make it clear I don’t think supplying a fitness function is tantamount to telling the GA how to solve the problem”. For this to be the case, it has to be the appropriate fitness function." I think the appropriate fitness function is given to us directly by the problem in this case. I apologise if the following comes across as patronising, it's not intended to be so, I'm just trying to explain the steps in my reasoning as clearly as I can: "Find the shortest network which connects all of a set of given points and any number of additional variable points" could IMO, be reworded as "Find the fittest network connecting a series of points, where "fit" is defined as follows: solution A is fitter than solution B if the total network length of A is shorter than that of B and all the points in A are connected". Have I added any new information in that formulation? ( I don't believe I have) Ok so let's say you, the human, find what you believe to be an optimum network using intelligence and I also provide a solution by some means, and we both have a list of line segments defined by endpoints - how will we decide who has won? Will we need to know the optimal solution in order to work out if your solution is better than mine? I suggest we will not. Also, we will have to agree on the meaning of some simple terms such as "length" and be agreed on how length can be calculated from end coordinates etc. but in agreeing these terms will we being be giving away the shape of the optimal solution? Is it cheating to build these basic definitions into the fitness function? Is Dave's fitness function adding information that we haven't agreed to in taking up the challenge, or different from a "function" we would use to determine if your solution is better than mine? Let's also do the same and compete to see who can find the best approximation to the sum of numbers from 1 to 1000. Could you rephrase the problem in a non-circular way? Can you a write a function which will judge if your solution is better than mine that doesn't have to implicitly or explicitly know the correct answer? Steve p.s. I'm not going to try and reconstruct my discarded post. My motivation is flagging.steveh
August 18, 2006
August
08
Aug
18
18
2006
11:02 AM
11
11
02
AM
PDT
Ofro, I've been following this discussion without contributing, and now I think I understand what everyone says the problem is. Maybe I can explain it in a way that helps. Algorithms like Dave Thomas's do an excellent job of optimizing for a particular physical variable. In the case of the Steiner solution, the algorithm optimizes for shortest length or smallest area. And there is no doubt about the results: A computer running the algorithm can find an optimal solution faster than a human being. But notice that the choice of what physical variable the algorithm optimizes for is not decided by the algorithm but built into it by the programmer. With respect to the algorithm, the "most fit" solution is always the one that minimizes length or area. Now "fitness" for Darwinism means "survival". But survivability does not map to a single or even a small set of physical variables. It maps to a virtually infinite set of physical variables. For one organism, increasing length might increase survival, for another decreasing length might increase survival, for another survival might have nothing to do with length at all. In fact, it may involve an entirely novel physical characteristic, which is the point of evolution in a creative sense in the first place. So a Darwinian algorithm needs to do more than optimize for a prespecified physical variable. It needs to figure out what variable to optimize for. IDer's think Dave Thomas is stealing a base algorithmically by designing the algorithm to optimize a specific physical feature, and then defining fitness as optimization of that feature. It should not be a surprise that a computer will outperform a human designer in these circumstances. Moreover, the same species will need to change the physical characteristics it optimizes as the environment changes. And, since Darwinian evolution is supposed to account for novel structures, the algorithm must somehow optimize for features that are not yet in existence, and only come into existence by the algorithm. If we wished to improve Dave Thomas's contest of algorithm vs human design by making it more realistic, we could frame the contest like this: Design an algorithm that will optimize for "survivability", with survivability meaning the ability to persist in an environment specified only when the algorithm is run. Examples of the specification of survivability in an environment at runtime would be: - whatever calculates pi to the greatest number of digits survives. - whatever adds the first 20 integers in the shortest time survives. - whatever comes up with the optimal Steiner solution for a set of points survives. - etc., etc. The algorithm would be run and the human would do his best, and whoever maximizes the relevant criterion survives. The human could at least make a crack at solving the problem no matter what it is. The problem for the algorithm, of course, is that it must have prior knowledge of what the survivability criterion will be before it is even designed, let alone run. But no such apriori criterion for survivability exists for Darwinism. Therefore, all the artificial Darwinian algorithms suffer the same flaw insofar as they have designed in optimization criteria that are not available in the real world. At least, that is what I think the problem is... Cheers, Dave T.taciturnus
August 18, 2006
August
08
Aug
18
18
2006
09:49 AM
9
09
49
AM
PDT
Surely the fitness function corresponds to natural selection, which is an external force. The only difference here is that the fitness function doesn't change, which I guess would be the equivalent of the organism living in a stable environment. I imagine that many features of organisms could be said to have some kind of optimum value.Chris Hyland
August 18, 2006
August
08
Aug
18
18
2006
09:13 AM
9
09
13
AM
PDT
comment by j: "What needs to be demonstated is Darwinian evolution doing what its supposed (by many) to be able to do. Darwinian evolution is blind/dumb/purposeless. For a program to demonstrate Darwinian evolution, it can’t be given goals, either explicitly or implicitly. The (right) fitness functions need to evolve, too." I am still not clear about the whole issue. It seems to me that you are implying that it is not possible to write a program that behaves in the manner postulated by blind/dumb/purposeless Darwinian evolution. In other words, one cannot write a code that behaves like this hypothetical, very mechanistic process because as soon as I write it, I have put in a non-Darwinian goal? It seems to me that the necessary conclusion would be that this falsifies a Darwinian mechanism a priori. That sounds like testing a null-hypothesis, and something tells me that this is not a valid test.ofro
August 18, 2006
August
08
Aug
18
18
2006
06:45 AM
6
06
45
AM
PDT
Strangelove wrote (4): "How should genetic algorithms look if they are to accurately demonstrate evolution?" It's not a matter of demonstrating "evolution." Demonstrating evolution is easy -- Thomas's and Salvatore's programs do so. Many implementations of algorithms for obtaining mathematical solutions can also be considered to demonstrate "evolution," too. But these are all teleological. What needs to be demonstated is Darwinian evolution doing what its supposed (by many) to be able to do. Darwinian evolution is blind/dumb/purposeless. For a program to demonstrate Darwinian evolution, it can't be given goals, either explicitly or implicitly. The (right) fitness functions need to evolve, too.j
August 18, 2006
August
08
Aug
18
18
2006
05:05 AM
5
05
05
AM
PDT
Tom commented: If merely supplying a fitness function is tantamount to telling the GA how to solve the problem,
I don't believe "supplying a fitness function is tantamount to telling the GA how to solve the problem" is the case. I hope that is not the impression my writings gave, but if so, I should speedily clarify that that is not my position. What my position regarding fitness function adequacy is the following tautology: supplying a fitness function that will solve the problem is supplying a fitness function that will solve the problem, and supplying a fitness function that will not solve the problem is supplying a fitness function that will not solve the problem. Sorry for the redundant redundancy, but I just wanted to make it clear I don't think "supplying a fitness function is tantamount to telling the GA how to solve the problem". For this to be the case, it has to be the appropriate fitness function.
then why are there so many problems GA’s can’t solve well?
Because not all fitness functions will solve the problem, in fact the overwhelming majority of possible function will guide the algorithm away from the solution.scordova
August 18, 2006
August
08
Aug
18
18
2006
12:21 AM
12
12
21
AM
PDT
Tom, I discovered the hard way that WordPress does not like the less than or equal sign. I see you've made that discovery as well. :=) Salvadorscordova
August 17, 2006
August
08
Aug
17
17
2006
11:57 PM
11
11
57
PM
PDT
[continuing a post truncated by the blog software] Bill neglects to mention that I(B) is less than or equal to I(A). Necessity may eliminate in B some or all of the information in the antecedent A. Here A is the input string of bits, and B is the output chromosome, and B is necessary when A is input to the GA. That is, the GA is a deterministic algorithm. It follows by definition that I(A) = N bits. Under the zero-knowledge assumption that all chromosomes are equally likely to be output as the solution, I(B) = K bits. Clearly K is much less than N -- i.e., the number of bits in a chromosome is much less than the number of bits input over the entire GA run by the random number generator. The upshot is that the GA works by selectively eliminating information, not by generating information. Keep in mind that the population not only holds competitive individuals, but serves as the memory of the GA. At the time of reproduction-with-variation, new information is entered into the population (memory). At the time of selection of parents, non-parents are culled from the population (memory). To the degree that a culled individual is distinct from individuals remaining in the population, information is eliminated. The fitness function does not tell the GA how to find an optimal solution. Neither the GA nor the fitness function knows the optimal fitness value. The fitness function gives the GA information on the fitness of individuals currently in the population relative to one another, but not relative to the unknown global optimum. It is important to note that GA's do not perform well for all fitness functions. If they did, there would be a free lunch in optimization for GA's, contradicting a well known theorem of Wolpert and Macready. Stuart Kaufmann formulated the NK landscape as a tool for studying problem hardness. (Different settings of parameters N and K give fitness surfaces with different properties.) There is a large literature addressing what are known as GA-hard and GA-deceptive problems. If merely supplying a fitness function is tantamount to telling the GA how to solve the problem, then why are there so many problems GA's can't solve well?Tom English
August 17, 2006
August
08
Aug
17
17
2006
11:28 PM
11
11
28
PM
PDT
[continuing a post truncated by the blog software] Bill neglects to mention that 0 Tom English
August 17, 2006
August
08
Aug
17
17
2006
11:25 PM
11
11
25
PM
PDT
Salvador, Irrespective of how good a model of natural evolution Dave's genetic algorithm is, it should explainable in terms of design theory. Right? To simplify analysis, let's say that N bits are randomly generated (i.i.d. uniform) and stored in a file prior to each run of the GA. The GA's random number generator is modified to work by reading bits from the file. N is exactly the number of bits it will need in a run of the GA. Now the GA can be seen as single-valued function mapping strings of N bits to strings of K bits, where K is the number of bits in the most fit chromosome of the final population. I hate relying on old work by Bill Dembski, but it seems more appropriate here than more recent work: "Because information presupposes contingency, necessity is by definition incapable of producing information, much less complex specified information. For there to be information there must be a multiplicity of live possibilities, one of which is actualized, and the rest of which are excluded. This is contingency. But if some outcome B is necessary given antecedent conditions A, then the probability of B given A is one, and the information in B given A is zero. If B is necessary given A, Formula (*) reduces to I(A&B) = I(A), which is to say that B contributes no new information to A. It follows that necessity is incapable of generating new information. Observe that what Eigen calls "algorithms" and "natural laws" fall under necessity." http://www.arn.org/docs/dembski/wd_idtheory.htm Bill neglects to mention that 0 Tom English
August 17, 2006
August
08
Aug
17
17
2006
11:23 PM
11
11
23
PM
PDT
from the previous “Panda Food” Thread: ofro wrote: “What I don’t understand is the basic premise of your example, which apparently already has an explicit solution of the problem built into the program.” scordova replied: “I’m afraid that isn’t quite correct because if you go to ga.c, and do a text search for 500500 you won’t find it. The solution was never explicitly stored anywhere.” I would have eventually figured it out myself that the Gauss solution to the problem was part of your program, but your admission makes it easier. I have to say now that your reply was anything but “thruth and nothing but the thruth”, and more like “I have sinned through my own fault, in my thoughts and in my words, in what I have done, and in what I have _failed to do_.” I am convinced now that at that point you wanted to “code-bluff” me (see https://uncommondescent.com/index.php/archives/1449). And I am not pleased.
Ofro, I was not trying to be mean or demeaning. I was only trying to point out the solution can be implicitly stored in a program (analogous to driving directions), it does not have to be explicitly stored (analogous to an explicit street address). This fact makes a lot of GA theatrics possible, where there is nothing explicitly stated, but the answers are lurking and so diffuse that short of running the program, one will not see that the program has the solution implicitly built in. The most basic example was brute.c, where one might not off the top of one's head know the answer, but one would know a proven strategy (simply adding all the numbers) that would succeed, and hence one could program an implicit solution which would eventually reveal an explicit solution. GA will often reveal solutions which we do not know in advance even though the the answer is effectively stored in the definition of the search strategy, the explicit answer often only appears during execution. We may have a thousand random numbers whose sum we don't know in advance. A program however with the appropriate solution seeking strategy can find the answer. The answer is effectively snuck in by matching the right problem-solving strategy to the right problem.scordova
August 17, 2006
August
08
Aug
17
17
2006
09:19 PM
9
09
19
PM
PDT
BC wrote: Nature guides the evolution of organisms by killing off the organisms that have bad mutations and proliferating the ones that have good ones.
That is not completely accurate. For starters, bad and good are teleological conceptions, Nature has no conception of design teleology. This again is Darwinian double speak, not science.
Consider the eyes of cave organisms who live in total darkness. If eyes are expensive to make, selection can wreck their exquisite engineering just as surely as it built it. An optic nerve with little or no eye is most assuredly not the sort of design one expects on an engineer’s blueprint, but we find it in Gammarus minus. Whether or not this kind of evolution is common, it betrays the fundamental error in thinking of selection as trading in the currency of Design. Allen Orr
Lewontin further points out the futility of fitness to define the value of inherent functionality in: SFI Bulletin 2003
It is easy to say that fitness of a type is its “relative probability of survival and reproduction” but turning that phrase into a coherent measure that can do work in evolutionary explanation is not so easy.... The problem is that it is not entirely clear what fitness is. Darwin took the metaphorical sense of fitness literally. The natural properties of different types resulted in their differential “fit” into the environment in which they lived. The better the fit to the environment the more likely they were to survive and the greater their rate of reproduction. This differential rate of reproduction would then result in a change of abundance of the different types. In modern evolutionary theory, however, “fitness” is no longer a characterization of the relation of the organism to the environment that leads to reproductive consequences, but is meant to be a quantitative expression of the differential reproductive schedules themselves. Darwin’s sense of fit has been completely bypassed. ... How, then, are we to assign relative fitnesses of types based solely on their properties of reproduction? But if we cannot do that, what does it mean to say that a type with one set of natural properties is more reproductively fit than another? This problem has led some theorists to equate fitness with outcome. If a type increases in a population then it is, by definition, more fit. But this suffers from two difficulties. First, it does not distinguish random changes in frequencies in finite populations from changes that are a consequence of different biological properties. Finally, it destroys any use of differential fitness as an explanation of change. It simply affirms that types change in frequency. But we already knew that. Richard Lewontin, 2003
scordova
August 17, 2006
August
08
Aug
17
17
2006
09:00 PM
9
09
00
PM
PDT
steveh asked: Could you mail a copy back to me please, so I can post it at ATBC or PT?
I tried to find out if the system stored it anywhere. I did not see it so I won't be able to get you a copy. I'm sorry. I do appreciate your efforts however at posting here.scordova
August 17, 2006
August
08
Aug
17
17
2006
08:45 PM
8
08
45
PM
PDT
1 2 3

Leave a Reply