Replies

The problem can be readily seen in the application of genetic algorithms to computer science. They are almost always used for tuning algorithms, not creating them. And, when they are creating algorithms, they are usually very simplistic (see what sorts of things Avida has been making, for instance).

I'm sympathetic to some of the author's overall argument. But this argument is not very strong. He criticizes Genetic Algorithms because they are used for "tuning." But that's what Genetic Algorithms do--they are just optimizers. It's kind of like criticizing a car for only transporting people.

The author should instead look at the Genetic Programming and the NEAT algorithms. Try Banzhaf, Genetic Programming: An Introduction and search for NEAT on the Internet. Both algorithms evolve structure AND tune the structure and do so quite successfully.

He should also look at the Q-beta replicase experiments in biology (probably from the 60's). There is an extensive description of the Q-Beta-Replicase experiments in Banzhaf, cited above.

In all of these situations, there's a lot more going on than just tuning the parameters.

I also don't know where the author gets the claim that the fitness slope is usually downward with natural selection. At least in the computer simulation world, which the author relies on when he refers to Genetic Algorithms, that is manifestly false. Thousands of peer reviewed publications say otherwise.

I believe that God created the universe and Man. But I also believe that God created a universe in which evolution is a force within that universe. So, as I said, I am sympathetic to the author's argument. But I know a little something about evolutionary computation and I can tell you the author's argument in that regard is a thin reed indeed.

Can you give a more specific reference to the NEAT algorithms?

"I also don't know where the author gets the claim that the fitness slope is usually downward with natural selection."

I think you misunderstood what I was trying to say (I wasn't very clear). The fitness slope is downward if you take an arbitrary system, mutate it, and select it. Most cases of natural selection on computers is done where the selection of parameters for mutation is non-arbitrary, and that's why it works - the parameters were distinctly designed for that.

In biology, natural selection can work because mutations are not random, but instead are directed by cellular processes to be in locations that are biologically meaningful. And so, you can get improvements because they are designed to restrict mutation/selection to specific areas of the genome.

If instead, you randomly mutated the genome, and then did selection, the large-scale result of what you were doing would be to degrade the organism over time, even with natural selection acting (the term for this is genetic load).

Is evolution a force within the universe? Sure. But it is largely a guided force - that is, it follows certain preconceived patterns which are implemented by specific mechanisms in the cell. For an example of one of these mechanisms, see Barbara Wright's A Biochemical Mechanism for Nonrandom Mutations and Evolution (this was the lead-off paper for what has been a very successful research program over the last decade - just look for "Wright BE[Author]" in pubmed).

Anyway, please send me a more specific reference on NEAT, as I can't find what you are referring to.