Wednesday, June 26, 2013

Black-box “optimization” is merely sampling

[Edit: The replacement is here at last. There have been many visits to this page since I retracted the post. I apologize to those who have been waiting.]

The main points of this post were correct, but the math contained some errors. I am working on a replacement.

Dembski’s perennial misconception of fitness

DiEb has begun a response to the latest morph of the creationist model of “search” (Dembski, Ewert, and Marks, “A General Theory of Information Cost Incurred by Successful Search” [pdf]). Here, slightly modified, is a rather general comment I made.

In the conventional “no free lunch” analytic framework, the objective (cost, fitness) function is a component of the problem. Dembski, Ewert, and Marks turn the objective function into an “oracle” that is part of the problem-solver itself. This model is inappropriate to most, if not all, of the evolutionary computations they purport to have analyzed.

Back in the 1990’s, Dembski committed himself to the misconception that Richard Dawkins’ Weasel program uses the fitness function in order to “hit the target.” Various people have tried, with no apparent success, to explain to him that one of the offspring in each generation survives because it is the most fit. The so-called target is nothing but the fittest individual.

To put it simply, the fitness function comes first. The “target” is defined in terms of the fitness function. Dembski gets this backwards. He believes that the target comes first, and that the fitness function is defined in terms of the target.

Dembski and Marks carry this to extreme in “Life's Conservation Law.” They claim that biological targets exist implicitly in nature, and that if Darwinian evolution “hits” them, then fitness functions necessarily have guided evolution. A remarkable aspect of this claim is that they treat fitness functions, which are abstractions appearing in mathematical models of evolution, as though they really exist.

The “search for a search” is another abstraction that they reify. A probability measure on the sample space is a mathematical abstraction. They merely assert that a search practitioner, in selecting a search, searches the uncountably infinite set of probability measures. To that I say, “Give me a physical description of the process.”

Thursday, June 6, 2013

Open access to Biological Information: New Perspectives

I previously raised an eyebrow at an editor of Springer’s “Intelligent Systems Reference Library,” in which the creationist volume Biological Information: New Perspectives (eds. Robert J. Marks II, Michael J. Behe, William A. Dembski, Bruce L. Gordon, and John C. Sanford) was scheduled to appear. The proceedings of the secret scientific symposium of scientists and “scientists”…

In the spring of 2011 a diverse group of scientists gathered at Cornell University with an eye on the major new principles that might be required to unravel the problem of biological information. These scientists included experts in information theory, computer science, numerical simulation, thermodynamics, evolutionary theory, whole organism biology, developmental biology, molecular biology, genetics, physics, biophysics, mathematics, and linguistics. Original scientific research was presented and discussed at this symposium, which was then written up, and constitute most of the twenty-four peer-edited papers in this volume.
… (did I mention science?) that took place at, but not under the auspices of, Cornell University have migrated to World Scientific. You can read the volume online, free of charge.

The big surprise is that “Section Four: Biological Information and Self-Organizational Complexity Theory” comprises two dissenting papers, one by Stuart Kauffman (whose views on many things are similar to my own), and the other by Bruce H. Weber. Although editor Gordon is none too clear on the matter in his introduction to the section, it appears that Kauffman and Weber actually contributed to a previous secret meeting, the proceedings of which were never published.

Their involvement in this project traces back to a 2007 conference I organized in Boston under the auspices of the Discovery Institute’s Center for Science and Culture. The conference commemorated the famous 1967 Wistar Symposium on “Mathematical Challenges to the Neo-Darwinian Interpretation of Evolution.” [...] The general perception among the participants in the Boston symposium, as with the participants in the Cornell University conference giving rise to this compendium, is that the mathematical and biological challenges posed to the modern evolutionary synthesis (neo-Darwinism) have not been resolved, but actually have grown more acute as our knowledge of molecular biology, cell biology, developmental biology, and genetics has exploded.
Gee, that sounds like “these guys are on our side.” But here’s the second half of Weber’s abstract:
Presently, however, there is ferment in the Darwinian Research Tradition as new knowledge from molecular and developmental biology, together with the deployment of complex systems dynamics, suggests that an expanded and extended evolutionary synthesis is possible, one that could be particularly robust in explaining the emergence of evolutionary novelties and even of life itself. Critics of Darwinism need to address such theoretical advances and not just respond to earlier versions of the research tradition.
So Gordon contradicts Weber while trying to paint him as an ally. He makes a fine point of the inadequacy of the “modern evolutionary synthesis (neo-Darwinism),” which is hardly where Darwinian evolutionary theory stands today. Kauffman highlights in his abstract the essential reason that the information measures of Dembski and Marks go nowhere in biology.
Biological evolution rests on both quantum random and classical non-random natural selection and whole-part interactions that render the sample space of adjacent biological possibilities unknowable.
I’ve heard him put it more simply: We don’t know the phase space. This means that it is impossible to assign probabilities to evolutionary trajectories. And taking logarithms of probabilities is how Dembski and Marks get information.

I wrote “scientists and ‘scientists’” above because only two of the five editors are scientists, and because engineers, computer scientists, and mathematicians have contributed heavily.

Unsurprisingly, about half of the “new perspectives” are variations on old themes of why evolution doesn’t work. John C. Sanford, a young-earth creationist who believes that genomes have been going to hell in a handbasket since the Fall of Man, authored seven of the papers and one of the section introductions. Dembski, Marks, MontaƱez and Ewert continue to bash evolutionary computation, including artificial life.

Jonathan Wells shocks us by reporting, “Not Junk After All: Non-Protein-Coding DNA Carries Extensive Biological Information.” Other papers show that the genetic code is fine-tuned, and furthermore that DNA sequences and computer code look much alike, with appropriate visualization. I’m sure there are other sensations to be found on closer inspection of the volume.