Re: Haldane's Dilemma -- talk.oriigns rehash

Bill Hamilton (hamilton@predator.cs.gmr.com)
Wed, 18 Jun 1997 13:28:42 -0400

Wesley R. Elsberry quoted John Queen:
>JQ>pim---
>JQ> You still dont get my point. The logic can go both
>JQ>ways...its random. Can a very ordered set of information can
>JQ>become more ordered every generation(what would of had to
>JQ>happened for us to get here) through random mutations? I get
>JQ>your reasoning that bad mutations would not propogate, but the
>JQ>chances become even more 'out of this world' each generation
>JQ>that some random mutation will actually improve upon what was
>JQ>previously improved upon.
>JQ> This type of lottery type improvement would had to have
>JQ>happened for millions of generations.
>
>Hmmm. I think that I'd like to see your mathematical writeup
>of this. If you are correct, then genetic algorithms can only
>very rarely converge on good solutions. Since they do converge
>quite regularly on good solutions, I suspect that you have a
>problem somewhere in your assumptions or logic.

It looks to me as though John is not compeletly accounting for selection.
True, he is aware that harmful mutations don't propagate, but perhaps he is
not taking account of the fact that beneficial mutations do propagate. The
action of selection at every generation guarantees that if a better
variation is produced, it will have a reproductive advantage. All that is
needed is for the random and nonrandom processes which generate the
variations to produce a sufficiently rich variety of characteristics.
Wesley's point about genetic algorithms is a good one. They essentially do
what evolution does, and they are sufficiently robust that they have become
valuable in solving a variety of problems in engineering design,
scheduling, resource allocation and other fields. Another point that needs
to be made is that selective advantage is not obtained solely through
mutations. In sexual reproduction offspring inherit some genetic material
from each parent. In this way genetic material gets combined in new ways
at reproduction, increasing the variety of characteristics in offspring.
I've used genetic algorithms in a few design problems and have found that
the recombination of genetic traits at reproduction seems more useful than
mutation. Mutation can help you get away from a local extremum, but too
high a mutation rate can actually interfere with convergence to the optimum.

Bill Hamilton
--------------------------------------------------------------------------
William E. Hamilton, Jr, Ph.D. | Staff Research Engineer
Chassis and Vehicle Systems | General Motors R&D Center | Warren, MI
William_E._Hamilton@notes.gmr.com
810 986 1474 (voice) | 810 986 3003 (FAX) | whamilto@mich.com (home email)