BIO-Complexity Publishes Article Answering Critics Who Promote Tom Schneider’s “ev” Simulation
Over the years William Dembski’s critics have accused him of allegedly not doing research. A few years back, Wesley Elsberry and Jeff Shallit published a response to Dembski which charged that “intelligent design advocates have produced many popular books, but essentially no scientific research.” Given how much peer-reviewed research Dembski himself has published in the field of evolutionary computation, these criticisms are not credible and hardly worth mentioning. On the other hand, some of Dembski’s harshest critics, such as Jeffrey Shallit, are smart guys that have published extensively in mathematics journals but have not cracked into the literature relevant to this field of evolutionary computation. Is it appropriate for Shallit to posture himself as a prestigious academic critic of Dembski when he has not published in the relevant scientific literature?
The same could be said for Kenneth Miller, a biologist, who in his latest book, Only a Theory, attacks Dembski’s arguments about the law of conservation of information. Miller cites the work of Tom Schneider, a researcher who claims to have simulated the production of new biological information via a program called “Ev.” Miller says that the results of ev “are striking” (p. 76), apparently so striking in fact that Miller touts Schneider’s program on eight pages in the book. Miller claims that Schneider did not “rig” the program “at all” stating:
Where’s the new information coming from? Perhaps the investigator is sneaking it into the system at the start? No chance of that, since the starting sequences are completely randomized. Maybe there’s hidden information in the program itself? Not likely. Schneider has made the source code of his program open for inspection, and there isn’t a hint of such nonsense. Did Schneider rig the parameters of the program to get the result he wanted? Not at all. In fact, changing the program in just about any way still results in an increase in measurable information, so long as we keep those three elements–selection, replication, and mutation–intact. Where the information “comes from” is, in fact, from the selective process itself.
(Ken Miller, Only a Theory, p. 78 (Viking, 2008).)
In contrast, in Signature in the Cell, Stephen Meyer wrote the following about ev:
The target sequence involves foresight. So too does the program’s fitness function, which makes selections based upon proximity to future function. Thus, it again simulates a goal-directed foresight that natural selection does not possess. It makes use of information about a functional state (nucleotide binding sites) in a way that natural selection cannot.
(Stephen C. Meyer, Signature in the Cell, p. 284 (HarperOne, 2009).)
Who is right: Miller or Meyer? A new peer reviewed paper published in Bio-Complexity, “A Vivisection of the ev Computer Organism: Identifying Sources of Active Information,” answers that question. Dembski and his coauthors have shown that, contrary to Miller’s claim, the ev program is in fact rigged to produce a particular outcome, and that Stephen Meyer’s description of ev is precisely correct.
According to the paper’s authors, ev “exploit[s] one or more sources of knowledge to make the search successful” and this knowledge “predisposes the search towards its target.” They explain that Ken Miller’s credulity towards ev was unwarranted:
The success of ev is largely due to active information introduced by the Hamming oracle and from the perceptron structure. It is not due to the evolutionary algorithm used to perform the search.
Indeed, other algorithms are shown to mine active information more efficiently from the knowledge sources provided by ev.
Schneider claims that ev demonstrates that naturally occurring genetic systems gain information by evolutionary processes and that “information gain can occur by punctuated equilibrium”. Our results show that, contrary to these claims, ev does not demonstrate “that biological information…can rapidly appear in genetic control systems subjected to replication, mutation, and selection”. We show this by demonstrating that there are at least five sources of active information in ev.
1. The perceptron structure. The perceptron structure is predisposed to generating strings of ones sprinkled by zeros or strings of zeros sprinkled by ones. Since the binding site target is mostly zeros with a few ones, there is a greater predisposition to generate the target than if it were, for example, a set of ones and zeros produced by the flipping of a fair coin.
2. The Hamming Oracle. When some offspring are correctly announced as more fit than others, external knowledge is being applied to the search and active information is introduced. As with the child’s game, we are being told with respect to the solution whether we are getting “colder” or “warmer”.
3. Repeated Queries. Two queries contain more information than one. Repeated queries can contribute active information.
4. Optimization by Mutation. This process discards mutations with low fitness and propagates those with high fitness. When the mutation rate is small, this process resembles a simple Markov birth process that converges to the target.
5. Degree of Mutation. As seen in Figure 3, the degree of mutation for ev must be tuned to a band of workable values.
(George Montañez, Winston Ewert, William A. Dembski, and Robert J. Marks II, “A Vivisection of the ev Computer Organism: Identifying Sources of Active Information,” Bio-Complexity, Vol. 2010(3) (internal citations removed).)
A critic might protest that some of these items entail the proper modeling of Darwinian evolution, but the way that ev uses these processes is unlike Darwinian evolution. For example, in (1), we see that the program’s use of a “perceptron” causes the output to be highly biased towards matching the target. It’s a way of cheating to ensure the program reaches its target sequence. Likewise, in (2) and (4), the program can effectively look ahead and march in the right direction towards the target, whereas unguided Darwinian evolution would have no “look ahead” capability. Thus, the active information in the Hamming Oracle allows the program to look ahead towards the target. This is unlike the evolution of real binding sites where there may be no binding capability until multiple mutations are fixed.
It would seem that contra Miller, mutation and selection are not the causes of success in these genetic algorithms. Yes random mutation occurs and yes there is selection. But selection is performed by a fitness function that is encoded by the programmer. And in programs like ev, the programmer intentionally shapes the fitness function to be amenable to stepwise Darwinian evolution. This effectively assumes the truth of the Darwinian evolution. But in the real world of biology, fitness functions might look very different: there might be lonely islands of function in a vast sea of nonfunctional sequences. Indeed, if one were to use a randomized fitness function, the search performs poorly and might not even outperform a blind search.
Thus choosing the right fitness function (from the set of possible fitness functions) requires as much or more information than choosing the right string from the set of possible strings in your search space. The fitness function itself is an information-rich structure. The program starts with this information rich fitness function, and then produces something much less information rich–the target sequence. And as the paper shows, ev does this in a relatively inefficient way: using the same information rich fitness function you can find the target 700 times more efficiently using simple single-agent stochastic hill climbing. Active information is smuggled into the fitness function. In the words of Ken Miller, “there’s hidden information in the program itself.” The program is designed to evolve.
Meyer, Not Miller, was Right
It would appear that Stephen Meyer’s words in Signature in the Cell were correct when he wrote:
[Robert] Marks shows that despite claims to the contrary by their sometimes overly enthusiastic creators, algorithms such as Ev do not produce large amounts of functionally specified information “from scratch.” Marks shows that, instead, such algorithms succeed in generating the information they seek by providing information about the desired outcome (the target) from the outset, or by adding information incrementally during the computer program’s search for the target. … In his critique of Ev as well as other evolutionary algorithms, Marks shows that each of these putatively successful simulations of undirected mutation and selection depends on several sources of active information. The Ev program, for example, uses active information by applying a filter to favor sequences with the general profile of a nucleotide binding site. And it uses active information in each iteration of its evaluation algorithm or fitness function.
(Stephen C. Meyer, Signature in the Cell, pp. 284-285 (HarperOne, 2009).)
Meyer cites work from Robert Marks, which we’ve previously covered here, here, and here. Meyer explains other sources of active information in ev which appear similarly vindicated by the new paper from Montañez, Ewert, Dembski, and Marks:
Before Ev applies its fitness function, it applies a filter to the crop of mutated sequences. The filter favors sequences that have the general profile of a binding site. Like the fitness function, this course filter makes use of information about the functional requirements of binding sites to favor some sequences over others. As such, it imparts information based on knowledge that Thomas Schneider, not natural selection or the environment, has imparted into the Ev simulation. Ev exhibits the genius of its designer.
(Stephen C. Meyer, Signature in the Cell, p. 284 (HarperOne, 2009).)
How Will Dr. Schneider Respond?
With such praise for Dr. Schneider’s genius, one would hope that his replies to criticisms of his work would be tempered. Unfortunately, in the past Dr. Schneider has acquired a reputation for responding to his critics by calling them “creationists,” as if by merely labeling them such he can defeat their arguments. Jeffrey Koperski said the following about Schneider’s common, though fallacious tactic:
In my view, labeling those who doubt the efficacy of genetic mutation and natural selection “creationists” is a rhetorical strategy, what some logic texts call “stereotyping.” Cable television provides ready exemplars for both the creationist stereotype and its cousin, the fundamentalist. Critics try to shape the debate by connecting ID to these templates. If successful, little work needs to be done. The labels tell us who represents the side of rationality over and against the side of ignorance. Having sorted us and them, what they actually say matters little, whoever they happen to be. We must recognize that although this is a common argumentative strategy in talk radio and presidential politics, it is not itself a logical critique. Placing the black hat on one’s opponent is no substitute for an argument.
(Jeffrey Koperski, “Two Bad Ways to Attack Intelligent Design and Two Good Ones,” Zygon, Vol. 43(2):433-449 (June, 2008).)
Whether Dr. Schneider attempts a substantive rebuttal to this paper’s forceful rebuttal to ev or resorts to name-calling against the authors as “creationists” remains to be seen.