From: Iain Strachan (email@example.com)
Date: Sun Aug 20 2000 - 17:52:44 EDT
On Mon, 19 Aug 2002 12:40:23 Dale K Stalnaker wrote:
>The following article was forwarded to me today by a friend. Some of you
>on the ASA list would find this interesting. Scientists have used an
>evolutionary strategy to make a robot learn how to fly by flapping its wings.
>Perhaps another gap just closed in the "God of the Gaps" perspective.
Having worked in the past professionally on genetic algorithms, and
assessing just what they can and can't do, I'd have to say that I'm
very skeptical that this experiment shows anything particularly
interesting. There are a lot of claims about what evolutionary
algorithms do around at the time, and a lot (in my opinion) of hype
attached to them, particularly from blurb-writers. I think the
following observations are relevant here.
(1) The robot evolution was driven by a distant idealized target.
This criticism was precisely the criticism that Richard Dawkins made
of his own famous "METHINKS IT IS LIKE A WEASEL" example in "The
Blind Watchmaker". The problem with the experiment was that the
designers of it decided that this thing was going to fly, and
therefore designed their fitness function to measure explicitly the
capability of flying (or half getting off the ground). But it's
clear that real evolution cannot operate like that; it is a truly
Blind Watchmaker that has no idea that it would be a good idea to
fly. Dawkins likewise points out that his "Weasel" model has a
pre-specified target, and hence was "misleading in important ways".
At least Dawkins is honest enough to admit that his illustration is
not evolution; merely a tutorial example to show the ability of
cumulative selection to make the very improbable happen.
(2) I note that the robot was pre-equipped with wings and motors
etc, so it had the _capability_ of flight right from the start; it
only became necessary to discover the correct sequence of
instructions (up down, rotate etc). One presumes that the genetic
"string" in the program consisted of a sequence of these
instructions, which could each change (via a mutation), thus at each
point making a small change to the path. But it's hopelessly
simplistic to imagine that this is how the flight might be
"programmed" into a real creature. The genome doesn't contain a
sequence of move instructions, but a whole complex of encoded enzymes.
(3) In any case, given the capability to fly, walk, ride a bicycle,
or whatever, living beings don't evolve from
non-flying/walking/bicycle-riding beings over millions of years to
ones that have these abilities; instead they discover how to do it
during their lifetimes by a process of trial and error. Consider the
problem of learning to ride a bicycle; it is achieved by the tuning
of various neural feedback systems over a period of time, and the
tuned parameters are somehow stored in the connections in the brain.
No mutations or genetic processes are involved. I've no doubt you
could program a genetic algorithm to ride a bicycle, but it doesn't
show any resemblance (in my opinion) to real evolution.
When an evolutionary algorithm evolves the bicycle/robot from nothing
and the means to ride it/fly it, then might be the time to be
Join 18 million Eudora users by signing up for a free Eudora Web-Mail
account at http://www.eudoramail.com
This archive was generated by hypermail 2.1.4 : Tue Aug 20 2002 - 14:02:15 EDT