Methinks it is sort-of like two weasels

[Weasel results]
Part of a run of a Dawkins's Weasel simulation, showing it approaching the target phrase

Eric Holloway has written an article here at the Discovery Institute's site Mind Matters, in which he presents a variant on Richard Dawkins's famous Weasel simulation, one in which there is not just one target phrase, but two. You may recall that the original Weasel program had a target phrase, which was "METHINKS IT IS LIKE A WEASEL". Counting blanks, this is 28 characters long. We start with a random string of 28 characters. Each generation the single string of 28 characters produces offspring and they mutate, each character having a chance of changing to one of the 26 other possibilities (the alphabet is 26 characters plus a blank). If there are 10 offspring in each generation, some may match the target phrase in more positions than other do. In the original Weasel, natural selection occurs by having the the offspring closest to the target string replace the previous parent string.

In Dawkins's original computer simulation, the target phrase was found in a few hundred to a few thousand steps, depending on the number of offspring and the mutation rate. We can compare this to what would occur without natural selection, by having a random offspring string replace the parent string. In that case, the string just changes at random by mutation, and one can show that the number of generations that would be needed to come across the target phrase would be about 1040. That's an awful lot more than a few thousand.

Holloway's article claims to show that by using not one, but two target phrases, it can be seen that natural selection is in most cases ineffective. A commenter, "Roy", at Josh Swamidass's Peaceful Science forum took issue here with Holloway. Roy's argument there is fairly convincing, but there is more to be said.

Let me explain ...

The Weasel

Richard Dawkins introduced his "Weasel" computer simulation in 1976 in his book *The Blind Watchmaker*. It was a dramatic example of how natural selection could achieve a goal far faster than purely undirected random change. Many advocates of creationism are insistent on characterizing natural evolutionary processes as "random". This in hopes of convincing their audience that evolutionary process are hopelessly bad at achieving adaptations. Dawkin's example demonstrates why natural selection is not making "random" change. The example is an enormously effective teaching example.

The result has been that many creationists are obsessed with showing that Dawkins's argument does not work. By contrast, evolutionary biologists are only mildly interested, as they have had many empirical and theoretical examples of natural selection working to achieve adaptations. Theoretical population genetics dates back over 120 years, and the mathematics of natural selection solidified 100 years ago: they have not been waiting to hear from Richard Dawkins. For a discussion of how a Weasel simulation can be made to use the standard Wright-Fisher model of theoretical population genetics, see a post by me at The Skeptical Zone in 2016. But that carries us away from the present issues.

Previous criticisms

Here are some of the most common creationist objections:

(1) "But unlike evolution, Weasel programs have an explicit target". Yes, Richard Dawkins said so himself at the time. The purpose of his simulation was as a teaching example to show readers how natural selection could be strongly nonrandom.

(2) "But Dawkins never provided his source code." He described the simulation clearly, and it is very simple. Many people have written Weasel programs in various computer languages, and all have shown results very similar to his. There is no mystery that would be solved by inspecting the original code.

(3) "But the program is intelligently designed!" So what? We can intelligently design computer simulation programs of many natural phenomena (motion of planets, soil erosion, change of weather, etc.) and that does not show that those natural phenomena must be intelligently designed.

(4) "The Weasel shows 'locking' (or 'a partitioned search'), so that a position does not change once it matches the target. That is why it works." Consideration of examples published by Dawkins shows that they do not show "locking". But when Weasels have "locking" added, they are only slightly better than ones without, and both show adaptation that is far faster than random wandering. So this is a minor issue.

It is rather remarkable how obsessed many creationists are with the Weasel. Consider that many of them declare that change within species is irrelevant to their arguments, because they are only talking about the origin of major changes of body-plan. They claim to not be making a general argument that natural selection is ineffective. Actually, they used to make such arguments, but they have sworn off doing that. But the minute someone claims to have an argument that shows that natural selection is generally ineffective, they "fall off the wagon" and once again succumb to this obsession.

Holloway's criticism

Eric Holloway's modified Weasel has two target strings. In his trials, the second string was "IT LOOKS LIKE A WEASEL TO ME". There is still one parent string, which starts out being random, and it undergoes the same mutational process as before. But this time natural selection acts by checking each offspring against both targets. The score is how many positions match at least one of the targets. Thus if a string has 12 positions that match only the first target, another 3 that match only the second target, and one position that matches both targets, we count 12+3+1 = 16 matches. As before, the offspring that has the highest score replaces the previous string in each generation becoming the new parent..

The result is typically a string that matches one target or the other in every position, but is not identical to either target. An example would be "IE LINKS LTKIS IEA EL EASME". How many outcomes have a score of 28? There are 26 positions in which there are two possible matches, and two positions in which there is only one possible match, positions 7 and 8. The number of strings achieving a perfect score is then 226, which is 67,108,864. Oops! Roy has pointed out that I made the same mistake that Holloway did, and which Roy corrected. There are three positions that match between the two targets, positions 7, 8, and 9, which means that the number of strings achieving a perfect score is 225, which is 33,554,432.

Aha, cries Holloway. The strings he calls targets are two strings out of this large number, and we reach all these possible perfect-scoring strings with equal probability, so the chance that natural selection and mutation succeed in finding one of the two target strings in this case is very small. All he has done, he says, is add an additional target, but natural selection does not know to go to one or the other targets, but instead mistakenly "splits the difference":

It is like natural selection is driving the evolution bus down the road and encounters a fork with either a left or right turn. Unable to choose either alternative, natural selection decides to split the difference and plow straight ahead, leaving evolution a smoking wreck.

"Roy" replies to Holloway

At Josh Swamidass's "Peaceful Science" forum, a frequent commenter whose handle is "Roy" presented a rebuttal of Holloway which started a discussion thread on "Misunderstanding weasel programming". Roy demonstrates some misunderstandings on Holloway's part. His main criticism is

... natural selection does not only select one character at a time. It selects the combination of characters in an individual organism. If a mouse is born with two mutations that give it respectively larger ears and darker eyes, it’s not possible for natural selection to retain one but not the other - that would require only half the mouse to have a litter of pinkies.
...
What happened is that @EricMH implemented the wrong selection criterion. He’s scoring single characters, when he should actually be scoring the entire string.
Roy is correct, as far as Roy's argument goes. I want to discuss the matter a little more generally.

Holloway and different cases

In natural examples, natural selection could work either like Eric's program or like Roy's program. These are two legitimate possibilities (and there are others, too). But what about the target strings? In natural evolutionary processes there are no targets. Different genotypes simply have different fitnesses, and that is the result of how the genotype affects the phenotype, and how that phenotype interacts with the environment. The closest thing to a target is the set of genotypes of high fitness.

So if you use Holloway's program, the targets that it is seeking are not the two sentences. They are the set of all 226 strings that achieve the highest fitness. On the other hand if you use Roy's program, the strings that achieve highest fitness are only the two strings that Holloway describes.

Yes, there is a behavior of "splitting the difference", but it is Holloway, not a simulated evolutionary process, that is doing the splitting. Holloway is running a program that would have 226 targets, but declaring that there are only two target strings. Holloway would be right about that if he were running, not his own program, but Roy's. But he isn't, he's splitting the difference and instead running his program, which has far more than two strings of high fitness.

Both case are possible, but you shouldn't go around arbitrarily declaring strings as being the targets, when evolution is acting as if there are far more or far fewer strings of high fitness. Your argument might be left "a smoking wreck".