Perakh, where I got the reference to Wein, is more in the ad hominem business. It's hard to take him at all seriously, with his disavowal of the term "Darwinism" which shows up in the evolutionary literature and Gould's essays (maybe they are "secular creationists"...) and his trafficking of consensus science -- which is noteworthy-- Perakh says that 1 in 142 scientists admit to doubting "the main tenets of modern evolutionary biology" which in context seems to be those elements of "the neo-Darwinian synthesis."
That many?! I thought I was on the ground floor of a scientific revolution. Don't worry, that virus has only infected 1% of the cells of your body. You're fine! No wonder they are desperate. The NCSE has even diversified into climate politics. They've had only 5 decades to get in on that.
I'll read Wolpert's disavowal of Dembski's interpretation of the NFL theorems again, but in writing his rather brief and politic statement, Wolpert succeeded at (a) saying that he was not trying to say what Dembski is saying nor meant his work to be taken as a challenge to the almighty force of natural selection and (b) noting that the "sit down and shut up" tactics of the Darwinist mob is actually a losing strategy at this point.
Maybe Wein is trying to do better than the mob.
But some things stand out to me already:
Wein quotes some researchers saying "averaging over all different fitness functions does not match the situation of black-box optimization in practice." That is hardly surprising. Solving any real problem (and any number of contrived problems) with stochastic/genetic methods generally involves parameterizing the problem such that gradients can be followed and that concerns remain separated enough to avoid -- or the algorithm is altered in order to not get stuck in local optima.
But real world problems often have some really nonlinear interactions between the parameters, complicating nonlinear responses to a single parameter. I could make a fitness function modeling the real world performance, or I could build into the fitness function some heuristics on the relationship between parameters. For parameters a and b, I might decide (this is pure fancy) that f '(sqrt(a ^2 + 2b^2) is an approximation of actual fitness F(a,b) that is likely to avoid getting stuck in local optima. For Wein, this is just making sure that the algorthm is properly designed to get the answer to the correct problem. Tbere is no concern of how this might actually be unnecessary to get a suboptimal answer to the correct problem but necessary to get an optimal answer to the problem.
"The trick in genetic algorithms is to find schemes that do this mapping from a binary bit-string to an engineering design efficiently and elegantly, rather than by brute-force..." In other words, applying engineering knowledge to understand the fitness landscape in advance, properly parameterizing the problem so that the phase space can be searched effectively, and building into the algorithm ways of dealing with obstacles.
"Deciding the best values for these parameters in a given application remains a black art, driven more by blind intuition and communal tradition than by sound engineering principles." Sounds like engineering in general to me. For Wein (quoting Wolpert), for more important than abstract generalizations of hill-climbing algorithms are the particular problems where you "start with the given [fitness function] f, determine certain salient features of it, and then construct a search algorithm, a, specifically tailored to match those features." This is what Schneider calls making the fitness function behave like the environment in natural selection -- "This is exactly how it works!" After all, we know that natural selection did it, the simulation is just demonstrating how.
Where are the solid mathematical foundations here? These guys need to get out the classics on population genetics and read them!
It could be (although I don't know) that the "crooked wire genetic antennas" Wein refers to might be a particular problem that lends itself to hill-climbing (to an apparently counter-intuitive solution) without a lot of teasing from the algorithm-designer. But Wein makes it pretty clear that you don't just rely on the fitness function to tell your algorithm everything it needs to know. You also need to "determine certain salient features of it, and then construct a search algorithm, a, specifically tailored to match those features."
Now, Wein (and this is a subject that deserves further exposition some time) claims that Dembski takes two bites from the cherry by offering that it is possible that the Darwinian fitness landscape-verse could be finely tuned for evolution. This seems unlikely given what we know of it. But are many of the biological design problems solved over the millenia the sorts of problems where you need to "determine certain salient features of it, and then construct a search algorithm, a, specifically tailored to match those features"?
Question: Is the landscape-multiverse well-suited to natural selection? Of course! After all, here we are...
No comments:
Post a Comment