|
Author
|
Topic: Darel R. Finley: Three Issues With "No Free Lunch"
|
Moderator
Administrator
Member # 1
|
posted 10. March 2002 18:43
Three Issues With "No Free Lunch"
by Darel R. Finley
dfinley@mdanderson.org
ABSTRACT—Dembski's book is undoubtedly an important addition to the growing set of books in the Intelligent Design movement, and I wholeheartedly recommend it to anyone interested in that movement. However, when I finished reading the book I was left with the uneasy feeling that a few important points were being overlooked; hence, this paper. I present my points in a non-technical form, based primarily on an analogy to finding a specified square on a large grid. However, I have little doubt that these issues are valid, and that applying them to biology is ultimately unavoidable.
To read the entire paper, please click here
[ 05 May 2002, 14:59: Message edited by: Moderator ]
IP: Logged
|
|
William A. Dembski
Member
Member # 7
|
posted 21. March 2002 15:24
I want to thank Darel for his thoughtful engagement of my work. Let me briefly address two of his concerns: (1) the question of real vs. apparent specified complexity and (2) the question of CSI begetting CSI.
Ad (1): For something to exhibit specified complexity it must exhibit a detachable pattern that maps onto an event of small probability (the "complexity" in "specified complexity" is a measure of probability). Now, the determination of small probability is made with respect to all relevant probability distributions that might account for the phenomenon in questions. In practice this means limiting oneself to probability distributions induced by all known naturalistic mechanisms operating in known ways. Evolutionary naturalists often cite this as a defect of specified complexity claiming that it merely provides a sophisticated cloak for ignorance.
Two comments on this objection: (i) the great strength of Darwinian and other naturalistic accounts of evolution was precisely to show that known naturalistic mechanisms operating in known ways could produce all of biological complexity, so at the very least specified complexity is showing that the problems claimed to be solved by naturalistic means have not been solved. (ii) The argument from ignorance objection can be raised for any design inference that utilizes specified complexity, including those where humans are implicated in constructing artifacts. There may be unknown naturalistic mechanisms that lead to identical essays being written by independent agents even though now we routinely refer such coincidences to design (i.e., plagiarism).
To come back to Darels' concern, however, about real vs. apparent specified complexity. Specified complexity, by being defined relative to the relevant probability distributions that might account for the occurence of an event, or alternatively by being defined relative to the known naturalistic mechanisms operating in known ways, might always be defeated by showing that some relevant probability distribution was omitted. That's always a possibility (though as with the plagiarism example and with many other cases, we don't take it seriously). As William James put it, there are live possibilities and then again there are bare possibilities. There are many design inferences which, to question or doubt, requires invoking a bare possibility. Such bare possibilities, if realized, would defeat specified complexity. But how would they defeat specified complexity? Not by rendering the concept incoherent but by dissolving it (or as I put it in NFL, but rendering the specified complexity only apparent).
In fact, that is how Darwinists, complexity theorists, and anyone intent on defeating specified complexity usually does it, namely, by showing that the probability or complexity that was thought to be so extreme really wasn't all that extreme after all (cf. Dawkins's _Climbing Mount Improbable_). Those who want to defeat specified complexity therefore try to show that it isn't real -- that the notion dissolves once we have a better understanding of the underlying causal mechanisms that render the object in question reasonably probable. By contrast, the design theorist argues that the specified complexity is real: that any attempt to palliate the complexities/improbabilities is destined to fail. This can in some cases be argued conclusively, as when the geometry of some biological structure allows complete freedom in possible arrangements of parts (cf. the sequencing of nucleotide bases). Michael Polanyi made such an argument in the 1960s.
Ad (2): As for the concern that CSI might always beget CSI and that we might never need to invoke an irreducible intelligent agent (just some preprogrammed bundle of CSI), the problem here is that backtracking CSI always makes the problem worse (cf. Law of Conservation of Information in ch. 3 of NFL). This law allows for one of the strongest anti-regress results that I know. The quantity of CSI to be accounted for, apart from invoking an irreducible intelligent agent (as opposed to a derived intelligence that merely shuffles around preexisting CIS -- I deal with derived intelligences briefly in the preface to NFL), always gets intensified when one tries to backtrack CSI under the operation of naturalistic mechanisms. Explaining CSI in terms of CSI therefore always comes to an end, especially in a finite universe of finite duration. CSI is not self-explanatory.
IP: Logged
|
|
Darel R. Finley
Member
Member # 100
|
posted 14. May 2002 22:21
William Debmski, et al — Thanks very much for publishing my paper; I'm quite honored! A few quick comments on your post:
Apparent Specified Complexity
I think I see a problem with the "nucleotide sequence freedom" argument. In what sense is the placement "free?" Would it be valid to say that the biologically undesirable, strong bonding of a huge, random, blob-like, 10,000-atom molecule to the strand of DNA is another valid possibility in the scope of DNA freedom? If not, why not? Perhaps because it is biologically damaging and does not permit the DNA strand ultimately to reproduce and to serve its function in the cell and organism. But that is also true of many nucleotide sequences — in fact, it is true of the vast majority of nucleotide sequences. How is this fact interpreted by the real-vs-apparent-specified-complexity framework?
CSI Backtracking
I see your point, but I would add that (1) CSI might not need not grow when going back in time — it might be able to stay the same, and (2) I wasn't necessarily suggesting that CSI is ultimately self-explanatory — but rather that the explanation for it may lie at a level of existence to which our science has no access, in the same category as "who designed the designer?". What I was trying to suggest is that human intelligence (if not a supernatural designers' intelligence) may consist entirely of pre-programmed CSI in the human genome, acquired CSI from our observation of the universe around us, and the appropriate neural algorithms to manipulate that CSI. In this scenario, music appreciation is entirely pre-coded by the author of human DNA, but knowledge of say, North American mountain ranges is acquired.
The specific relevance of the above scenario to NFL is that our general experience of CSI coming from intelligence may be simply the observation of humans expressing CSI contained in the human brain — some of it pre-programmed at the inception of the human race, and some of it acquired from the environment. So, we don't have a basis to confidently assert that intelligence is the source of CSI; for all we know, intelligence may be a label we put on the effects of CSI. (Whether this is true of a supernatural intelligence is, at present, beyond our capacity to study.)
IP: Logged
|
|
rossum
Member
Member # 463
|
posted 04. October 2002 23:36
[Moderator's Note: This post was moved from an outside thread into the thread on Finley's paper ]
Since I cannot post in the Archive, I am putting this into Brainstorms. It is a comment on Darel R Finley's article 'Three Issues with "No Free Lunch"', posted in the Archive on 14 May 2002.
1 Introduction
These comments cover the second of Finley's three issues - the persistence of the displacement problem in a non-evolutionary biology. Finley explains the problem thus:
quote:
"To explain this issue, let me begin with an example of the displacement problem in action. Suppose we have a very large chessboard, 2250 by 2250 squares on each side. Thus, the whole board has 2500 squares (more than 10150). This means that all the probabilistic resources our universe has to offer cannot render reasonably probable the selection of a particular (independently specified) square at random, as detailed in section 1.5 of No Free Lunch."
I will examine Finley's conclusions and propose that he has not taken his analysis of the displacement problem far enough. By taking the analysis further the original source of the CSI can be found.
2 Algorithms
After discussing different computer algorithms Finley concludes that any target seeking algorithm needs to know what the target is if it is to be able to find it. As Finley puts it:
quote:
"The formal way to describe it - keeping with the framework laid out by Dembski - is that to contingently select Function 2 from the larger phase space of all possible fitness functions over our super-chessboard, is to select the full CSI of the "Washington" phrase, in advance of any evolutionary algorithm finding it with the aid of the function. A more practical way to describe it is simply to note that any computer implementation of Function 2 would have to include the full 500 bits of the "Washington" phrase in its compiled (launchable) code, whereas a computer implementation of Function 1 would not."
Despite the seeming obviousness of this conclusion I disagree with it. It is not necessary to have the target, the "Washington" phrase, compiled into the code of either the algorithm or the fitness function. For a practical algorithm, we would not hard code the target phrase, or the equivalent coordinates, into the algorithm. For practical use we would read in the coordinates required from some input device. This would allow the algorithm to be general rather than tied to one particular target square.
Taking Finley's first example algorithm, GO-TO, we could modify it to make GO-TO2:
code:
read the coordinates of the specified square
select the desired square
The specification is not preloaded and the algorithm will work for any specified square that gives any desired 100 character sequence. Since the algorithm will work for any specified square it cannot have any specific square pre-programmed into it. The coordinates of the specified square are explicitly read into the algorithm after it has started running.
Similar modifications are possible to the next two algorithms Finley gives, RASTER and RANDOM, to make them general, rather than tied to the particular "Washington" phrase.
Given that the algorithms are generalized so that they will work for any specified square then they cannot have any particular specified square preloaded. If one particular square was somehow preloaded then the algorithms would only work for that one square.
This means that for the generalized algorithms the Complex Specified Information (CSI) is not actually loaded into the algorithm, but it is read in from the input device while the algorithm is running. Therefore the source of the CSI is whatever is read from the input device. The CSI is not inherent in the algorithms. I shall return later to the question of where the input device gets it CSI from.
The case of the EVOLUTION algorithm is different. Here the CSI is supposed to be encoded in the fitness function J. Again it is easy to modify the fitness function to read the target coordinates from an external device.
code:
float function J2 (int x, int y);
begin
int a, b;
read specified coordinates (M, N) into (a, b);
return (1 / (1 + (x - a)^2 + (y - b)^2));
end J2.
J2 is based on Finley's F2, but in this case M and N are supplied externally and read into a pair of internal variables, a and b. Being variables, a and b cannot start with CSI since they can take any values. It is only after their values have been set by the read statement that CSI is present.
Here again, the fitness function J2 would work for any specified square (M, N) so it cannot have any particular square preloaded into it. Again the source of the CSI is the input device from which the specified coordinates are read.
In passing it is worth noting that the J2 version of the fitness function will still work if the target, the specified square, is changed over time. Since it reads the specified coordinates each time it is called it will react to any change of target. This is in contract to the version where the target is hard coded into the fitness function, which will not react at all to any change in the target while it is running.
Evolution can react to changing targets within the DNA phase space, so it would seem that the separation of the algorithm from the target information gives a better model than where the target is hard-wired into the algorithm. For example, the earth's atmosphere changed from reducing to oxidizing with the release of free oxygen by photosynthesizing bacteria. With this change, parts of the DNA phase space became non-viable while other parts became newly viable. It is just as well that evolution can react to such changes because humanity occupies one of those newly viable areas.
3 The Mysterious Device
If CSI is not inherent in the algorithms then where does it come from? It is explicitly read in from an external device. Where is this device getting its CSI? As a mere mechanical device surely it cannot generate CSI itself.
First we need to step back and look at the definition of Complex Specified Information. In his opening paragraph, quoted above, Finley talks about "the selection of a particular (independently specified) square at random". This independent specification is essential if we are to have CSI. Without the independent specification the information is not specified; at most it can be merely Complex Information.
Implicit in this definition is that the independent specification is accessible. Does a given arrangement of rocks exhibit CSI? Perhaps the arrangement exactly matches a constellation in the night sky of a particular planet in the Andromeda galaxy. However without access to that planet we are unable to say that the arrangement of rocks is specified. Hence it is valid to assume that a specification is accessible and hence we can be sure that CSI is indeed CSI.
Given that there is an independent specification, and that it is accessible, then it is possible for the specification to be read by the general algorithms we were discussing. The source of CSI is the specification of the CSI. This exists independently of the CSI itself and can be accessed by the general algorithms and fitness functions. This seems reasonable, since without independent specification CSI cannot exist. Therefore we should not be surprised if the source of the CSI turns out to be the very specification on which it depends.
Even without rewriting the algorithms it is possible to ask how the coder of the algorithms got the information necessary to preload the target into the algorithms. There are plenty of computer programs that write computer programs so it is not possible to assume that any given algorithm will actually be written by a human being. Whoever or whatever is constructing the algorithm must have got the target information from somewhere. That source can only be the independent specification.
Taking the example of Dawkins' METHINKS IT IS LIKE A WEASEL program, we can see that for all versions of the program, whoever wrote them, the original source of the target specification was Richard Dawkins himself. Whether the phrase is hard coded into the program or is entered while the program is running, that particular phrase has to be typed at some point. It was Richard Dawkins' original selection of that particular phrase, which both created the CSI and provided the target information that must be passed to the program.
4 Conclusions
1 Finley is correct to say that any target seeking algorithm needs to know what the target is if it is to work correctly. As Finley points out the GO-TO algorithm cannot know where to jump to if it does not know the target coordinates. Similarly the RASTER or RANDOM algorithms will not know when to stop if they do not know what the target is.
2 However it is not necessary for any algorithm to have the target "in its compiled (launchable) code." By making the algorithms more general and by explicitly loading the target information from some external device the source of the target information can be detached from the algorithm.
3 This separation of a static algorithm from a possibly changing target allows some algorithms, such as EVOLUTION, to react to a target that changes with time.
4 If there is Complex Specified Information then it must be specified, i.e. there must exist an independent specification of the information. As a corollary this independent specification must also be accessible.
5 The separation of the algorithm and the target information allows the original source of the CSI implicit in the selection of the target to be seen more clearly. The source is the independent specification itself. An example of this is the numerous versions of the METHINKS IT IS LIKE A WEASEL program, which all depend on the original specification of that particular phrase by Richard Dawkins.
IP: Logged
|
|
|
Printer-friendly view of this topic