13 January 2007

Revisiting the Weasel Experiment

Richard Dawkins' interesting computer simulation of natural selection with random variation demonstrates a startling ability that, "at first exposure, is every bit as counterintuitive as quantum mechanics." [Scott Maxwell]. It has been criticized by intelligent design folks on grounds that: 1. the answer is known ahead of time so of course it gets the right answer, and 2. unrealistic mutation probabilities are used in the demonstration, and 3. random variation will cause more damage than improvement, resulting in catastrophic results. Objection 1. is that of someone who simply will not trust the experimenter, but can be addressed by defining a "fitness function" that turns the goal into an objective optimization problem. Objection 2 can easily be adjusted. Objection 3. reflects some subtleties that are too often glossed over by evolutionists, including Dawkins. This post is to show that the Dawkins demo has been put online, and that updates have been made to the demo that allow it to be much more closely analogous to real natural selection. In addition I'll point to some other software that gets even more realistic, evolving 2D graphics in a method that observes a distinction between genotype (algorithmic data) and phenotype (translation to graphical products).

First, here is the original Dawkins weasel experiment, except that you can substitute whatever string of data you would like as the target and then watch the process converge on the answer, typically in less than 100 "generations." It's too bad one could not continuously update the target string and watch the selection process demonstrate adaptation, as it dynamically tracks the changing conditions. Ah well, that will come in time.

But here is a better version of the weasel program, especially in the sense that you can download the source code and play with it yourself, or just read it to see what is going on. This one reveals a subtle point that is the source of creationist misunderstandings of this program. The figure shows that quite a large population of "critters" is generated in the process of selecting for the target output string. This is essential because many "bad" mutations occur that are deselected. This isn't stated clearly in Dawkins' original description or the instance described above. It is something that the ID folks believe to make a solution impossible, but clearly it does not do that, though it does require more deselection than is obvious from reading Dawkins or the previous online program.

Next is the much more realistic case where we have a specified population of random strings that are being randomly altered at a specified rate. Moreover, pairs of the strings are sexually mated to produce offspring strings in each generation, as random mutation proceeds. Several selection options can be exercised, as well as several options for the sexual merger. The mutation rate and population rate can be varied as much as one likes, slowing the process to a crawl if desired. So much for objection 2. Again, it would be nice if the target string could be varied as the process proceeds, so we could watch adaptation to new guidance being reflected in the results.

Finally, we have a more complex but visually appealing piece of software that implements a more realistic simulation in which a genotype is set up consisting of parameters controlling diverse graphics primitives and techniques. Each genotype translates to a phenotype as a 2D graphic panel. The user can then exert intelligent selection by either asexually or sexually breeding new generations of graphics, manipulating the graphical products toward whatever s/he may choose. Warning: this gets considerably more involved, though this program insulates one from the details so that everything is accomplished by point and click selection and no actual graphical engineering is required.

So there we have it in gradually more complex forms; a compelling demonstration of the power of selection with random variation, as a way of exploring multi-parameter design spaces and searching for optimal solutions of whatever problem is chosen, whether it be the matching of a particular string of characters, or the creation of a desired graphical image. So much for objection 3.

Does anyone know of simulation programs like these designed to illustrate the criticisms of natural selection that have been made by the intelligent design community? I would like to see what kinds of demonstration can be made of the failure of natural selection to function as advertised. If so, please point to them in the comments and I will compare them with the above cases.

Further Reading:
More than you ever wanted to know can be found in the Creationism Asserted and Creationism Rebutted links in the sidebar. However for a really definitive look at the matters discussed here, TalkOrigins can't be beat.


Anonymous said...

The problem with 'experiments' of this type is that you haven't thought the process through. Darwinian evolution by it's very definition, cannot have a target. So, put a blank field in your experiment as a target and see how long it takes you to get to weasel.

It'll NEVER happen.

Tom Moore said...

This doesn't really address your point, but the NS program converges just fine on a string of 30 spaces. ;=)

Of course you mean that there can be no target in nature, but I think you really mean that no one knows the target in advance.

The target doesn't need to be known in advance, as long as the selections are made. The party making the selections knows the target, and we can infer it approximately from where we are after 5 B years of selection. You can see all that in the experiments.

Now the question is whether an experiment can be constructed that realizes and demonstrates the ID objections to natural selection. If the claims are true, this should be possible.