joe@uw-evolution.UUCP (Joe Felsenstein) (12/18/88)
[this one already sent also to sci.bio newsgroup on Usenet] I note the strong claims made for the magicality of the "genetic algorithm", e.g.: > Genetic algorithms derive largely >from this genetic search theory (GST) which falls within the field >of mathematical genetics.) ... > >The observed efficiency of genetic algorithms at locating ever-better >local optima on a wide variety of complex performance surfaces, >without becoming trapped on a local optimum anywhere, ... >Dan Offutt offutt@caen.engin.umich.edu I recently acted as reviewer on a paper in the optimization-by-simulated- evolution literature, and was disturbed to find in it assertions that biologists have proven that this strategy is incredibly powerful. Now we are told that computer scientists have proofs of powerfulness that we biologists should be reassured by. I suspect neither is true. Genetic systems optimize pretty well, but only when there is a rather uncomplex relationship between genotype and phenotype. In their most sophisticated form (Sewall Wright's "Shifting Balance Theory") they can only cope with moderate interaction among genetic elements. Real genetic systems mostly do not interact enormously strongly (your keratin genes which affect hair and toenails are unlikely to affect the oxygen-carrying properties of your blood, and so on). If they did not have this partial decoupling of effects, they would not evolve very well: every mutant would reduce the organism to a disorganized mess, and no progress would be made. Presumably organisms whose genes interacted too strongly just never evolved well and died out. Does anyone have real proofs that simulated evolution is much more powerful as a mathematical optimization method than biologists have concluded evolution is as a real evolutionary process? I note Offutt's citations of work by Holland, but are there others? What do they really show? I would expect SA to work pretty well, but only if the genotype-to-phenotype mapping is not too tightly interactive. It should not be the magical all-purpose optimization method everyone is looking for. Pandas can't fly. Note -- this is NOT intended as an argument against creationists. --- Joe Felsenstein, Dept. of Genetics SK-50, Univ. of Washington, Seattle WA 98195 BITNET: FELSENST@UWALOCKE INTERNET: joe@evolution.ms.washington.edu or: uw-evolution!joe@entropy.ms.washington.edu UUCP: ... uw-beaver!uw-entropy!uw-evolution!joe