Marks and Dembski Acknowledge Withdrawing “Unacknowledged Costs” Paper
Robert Marks and William Dembski co-authored a paper that appeared on their “Evolutionary Informatics Laboratory” website, titled
“Unacknowledged Information Costs in Evolutionary Computing: A Case Study on the Evolution of Nucleotide Binding Sites”. After bragging about how their work was going to overturn the findings of all those folks working on evolutionary computation in the real world, and bragging that they hadn’t had criticism of their latest offerings, people did have a look and made some criticisms. Well, when the criticisms arrived and pointed out that the program that supposedly countered everything people knew about evolutionary computation was, in fact, dependent for its conclusions upon the fact that it improperly initialized some of its variables, the paper was de-linked from the main website. There was no notice given that the criticism specifically discussed as missing by the authors had played a role in the de-linking.
Now, about a month later, there is an acknowledgment that the paper was withdrawn.
Thanks to those who pointed to a bug in our software. This paper has been withdrawn.
For revised analysis, see HERE.
Well, Bob and Bill, while I appreciate the thanks, at this point I’d be astonished if this were actually credited such that the IDC cheerleaders could figure out who exactly deserves that thanks.
And one would expect that the “revised analysis” either would utilize a correct methodology to support the original claims, or it would retract the strong claims made and assert weaker claims, but one would be wrong. Instead, the “revised analysis” continues to claim that “ev”‘s performance is due to “perceptron structure and error measure”, but offers nothing except the fact that it outperforms blind search on the problem of interest as support. It’s a great big instance of begging the question, without even the poor support of a bogus script to spew phony numbers.
I wonder if Tom English thinks Marks is still a guy worth working with in spite of the association with Dembski.
ZOMG! My particle swarm heuristic outperformed blind search on an NP-hard problem! Can I now dispute everything we know about honey bees?
Dembski’s cyncism is breathtaking. Seriously. How totally stupid does he think his followers really are?
No. On second thought, don’t answer….
IIRC, English’s involvement is a commentary on Baylor U. policy, not a positive endorsement of Marks and Dembski. I don’t think Dembski’s reputation plays much of a role in his decision.
I’m way beyond fashionably late to this party. But I want to thank Wes for getting it right in #4.
In fairness to Bob Marks, I should say that I still like his notion of “evolutionary informatics,” no matter that it is ID-friendly by design. Evolutionary processes somehow, in some sense, imbue members of the population with information. We need appropriate definitions of information and rigorous accounting in terms of those definitions.
In criticism of Bob Marks, I have to say that he undercut his denotative definition of “evolutionary informatics” with his ostentative definition. Apparently evolutionary informatics entails the rhetoric of a thermodynamics-challenged mathematician and a minimax tweaker.
It’s hard to believe now that the EvoInfo lab stands any more chance of making it at Baylor than a snowman did in the Branch Davidian compound. As a Waco native, I know when to hold up my hands and walk out.
My opinion is that “active information” as deployed by Marks and Dembski is nothing more than either invented or misleadingly co-opted jargon for Dembski’s breathtakingly bad idea from way back that effective evolutionary computation must be premised upon “smuggling in” information.
Take the analysis of “ev” that Marks and Dembski have in their “unacknowledged costs” essay. They claim that the “perceptron structure and error measure” account for all but a few bits of information in solutions, even though their basis for analysis was demonstrated to be completely “garbage in garbage out”, lacking proper initialization of data structures, mismatch in row versus column layouts of data structures, and misunderstanding of “ev” to boot. Their “revision” in the new version of the essay, now available online, is to remove references to the busted and bogus script that generated their numbers. The bogus numbers remain, now apparently unassailable (in Marks and Dembski’s view) because they don’t bother to tell you how they came up with them.
The underlying perceptron pattern-matching does yield a different probability distribution for matching given alterations in the parameters than an assumption of a uniform distribution. However, once one gets into estimating a real probability for a match on randomly-generated “ev”, one finds that p_s cannot be significantly greater than about 4e-23, about twenty orders of magnitude smaller than the p_s offered (now, without justification or basis!) by Marks and Dembski. That is much greater than the odds of hitting on an exact bit match for the number of bits in an “ev” string, but it is still a significant challenge for solution via evolutionary computation. Marks and Dembski have taken the route of pretending that “active information” explains all but a pitiful residue of what appears in the solution of “ev”. They do not justify their work, because they cannot. It is simply talking up an old, bad idea Dembski had, dressing it up in jargon and numbers that no longer even can be credited as a derivation of something. Certainly evolutionary computation results need to be cast in terms of the actual probability distribution of solution states, but I haven’t seen any indication that the literature even has a problem with respect to that issue. Marks and Dembski simply sneer it up in their conclusions without documenting that there is even an issue for them to sneer at.
Assuming that Marks and Dembski are able to hawk this incredible dog to some unsuspecting journal, I look forward to writing up a response laying out not just the technical errors within it, but also the history of their persistent rejection of critical commentary and advice, and their failure to acknowledge those who tried to set them straight pre-publication. One might be tempted to think that they aren’t finding the peer-review process for it as much of a cakewalk as Dembski earlier stated it would be.
Before digging into #6, I’d like to relate an anecdote. Back in the mid-90’s, one of my students was looking at bang-bang control of the cart-pole system, working with system equations in one of David Fogel’s papers. He discovered that the gravitational constant was negated, and that the pole was falling up, instead of down. When I passed the word on to David, he admitted immediately that the constant was wrong, and furthermore determined that the pole had fallen up in various studies prior to his.
David fixed the constant and repeated a large set of experimental runs. Fortunately, his evolutionary system obtained controllers that kept the pole from falling down. He wrote up the revised study, identifying clearly how it differed from the previous one, and presented it at a conference. I know David pretty well, and I believe he’d have published even if the revised study had violated his expectations.
This is certainly the kind of behavior we should expect from all researchers.
I’m coming up short on details for a detailed response. This is not surprising, given that I last read the paper in August. I’ll read the revision and post in the afternoon.
If you need the original for comparison, let me know and I’ll send it to you tomorrow.
I see the definitions of endogenous, exogenous, and active information as an attempt to salvage something legitimate from the notion of specified complexity. The only spinning I see in the basic framework, which is laid out much better in an earlier paper, is in the choice of the term “active information.” Marks and Dembski suggest repeatedly that active information comes from an actor (an agent). The “active” information is merely information gain, as suggested by their notation I_+. Alternatively, it is bias. It’s easy to overlook that the gain is just a log-ratio of probabilities, log(p_s / p).
Although I don’t want to show my hand entirely — I’ve been struggling a number of months to establish a key theorem — I will say that the simple fact that physical search processes address only problem instances that “fit” in the observable universe is an enormous bias. While an agent may introduce bias, the presence of bias does not signal agency.
Dembski’s penchant for defining instance complexity in terms of surprisals has always given me the heebie-jeebies, but I’ve never seen how to raise a formal objection. I’d like to make it clear that what I wrote about conservation of information in my EP96 paper, available at my web site, was in terms of Shannon information (entropy) of distributions. What I’ve had to say about instances was in terms of Kolmogorov complexity.
When I first read the paper, I was surprised by the modes at “all zeros” and “all ones” in the empirical distribution, but placated by the claims about eigenvectors. The other “bumps” in the empirical distribution bothered me, and I suppose in retrospect I should have known something was wrong. Assuming you got the computation right, you’ve located an additional 60-70 bits the ev process yielded. That’s a lot.
As for peer review of the paper, I think there are more superficial aspects that any competent reviewer would raise serious objections to. I’m not going to post free advice on the matter. I hope the paper is actually not under review. Bob Marks does have an outstanding reputation, and he should not risk it on something like this.
I certainly hear you on the “free advice” bit, especially since Dembski has a terrible track record on taking criticism in a scholarly way, and that whole bit about “using” critics with online releases of draft essays. The situation with that paper, though, was going to come out, therefore I felt that I might as well make the criticism that was being offered unmissable.
Yes, Dembski’s notion that populations in their environment obtain information “for free” has always struck me as weird. I’m glad to hear that you are making progress on a formal treatment of this issue.
I’ll reserve judgment on the “spinning” bit; it seems pretty clear to me where Dembski is coming from with this line of argument, and it certainly wasn’t justified when he first started using it.
For all the talk of “free” information in biological evolution, I can’t recall any genetic evidence. This is important, because apparently-free information can and does come from the environment.
What strikes me most about the human genome, viewing it from an informatics perspective, is how little information it contains. A high estimate of the information content is 4.5 billion bits, and with a low estimate that life has existed on earth for 3 billion years, that amounts to gain of at most 1.5 bits per year.
No one claims that a human genotype codes for a baby. The phenotype of a neonate requires many more bits to describe than the genotype contains. This does not signal a miracle of gestation. The “free” information comes from the intrauterine environment. In terms of the “genetic program” metaphor, the program is “linked” with a complex “operating environment.” This is why the result of executing the program may contain much more information than is present in the program itself.
Everyone acknowledges that large differences in phenotype are sometimes attended by small differences in genotype. What line of reasoning leads to the conclusion that the punctuation in punctuated equilibrium is attended by large gains of information in genomes? And even if new species do sometimes differ “suddenly” in genetic information from prior species, what line of reasoning leads to the conclusion that the differences cause, rather than follow, reproductive isolation? Rapid genetic change is improbable in a large population, but not in a small population exploiting a new environmental niche. The first place to look for a source of new information is the new niche, and not supernatural intelligence.
I am perhaps parading my ignorance of biology here. But I will hazard to suggest that evolutionists at the fore of the debate call on IDists to provide detailed genetic evidence of rapid information gain. It seems to me that observations of phenotypes do not suffice. “Oh, can you not produce DNA from the Cambrian explosion, Dr. Dembski? I’m so sorry.”