Controlled evolution

[From Gary Cziko 940810.0425 GMT}

responding to Bill Powers (940808.1110 MDT)

The difference between goal-directed evolution and blind evolution, as I
see it, is that if the result of a mutation is a decrease in fitness, in
the goal-directed case there is another mutation right away in the
relatively unfit population of offspring, whereas in blind evolution the
next mutation occurs no sooner than if the result had been beneficial.
In the relatively more fit population, the next mutation is postponed in
the directed case, while in the blind case the next mutation takes place
at the same time as usual (on the average). So the goal-directed case
will end up producing fit offspring in the species much more often (per
unit time) than in the blind case, and will allow reproduction of the
fit offspring for a longer time than in the blind case.

Thanks for the clarification. The contrast between "blind" and
"goal-directed" is what threw me since in BOTH cases mutations (and sexual
recombinations) are blind as to their effect. And people now use the term
"directed mutations" to mean that organisms can somehow "know" what
mutations to produce when placed under stress (e.g., Cairns).

All this has to be worked out in simulation, taking both the normal
evolutionary factors and the addition of this systematic variation in
mutation (or crossover) rate into account. Do you know any people using
GA who might be persuaded to try this? It shouldn't be hard -- the ECOLI
algorithm is pretty simple.

I'm not sure what the advantages would be for genetic algorithm (GA) users
who use GAs to find a solution that will work in an unchanging environment
and then just use the solution. I can see, however, how it might be useful
for use in changing environments.

Responding to Bill Powers (940809.1107 MDT)]

A great deal is said about the environment "selecting" organisms. I can
see nothing in the environment, other than organisms, capable of
"selecting" anything. To speak of natural selection as something the
environment does is to use a metaphor as a way of subtly asserting what
does not exist. It seems much more plausible to me to say that species
contain selection mechanisms which include the production of random
variations as a means of creating more possibilities among which to
select. Without those capabilities, organisms could not evolve any
faster than a rock does. Evolution is a continuing experiment conducted
by a species, not by the environment. That metaphor, I think, will stand
up to close scrutiny.

I recall that the co-discoverer of natural selection, Charles Russell
Wallace, had much the same objection to Darwin's use of "natural selection"
and recommended instead "survival of the fittest" to get away from the idea
that something external to the organisms (God, Mother Nature) was doing the
selecting.

Here is a nice quote on selection by Henry Plotkin (which he calls
"generate-test-regenerate" or "g-t-r" from his 1994 book _Darwin Machines
and the Nature of Knowledge_ (p. 139). He discusses two characteristics of
g-t-r"

"One is that it takes the logical form of induction, generalizing into the
future what worked in the past. That is, the successful variants are fed
back into the gene pool where they will be available for sampling by future
organisms. This is the conservative, pragmatic part of the heuristic. The
other is the generation of novel variants by chance processes. This is the
radical, inventive component of the heuristic. It is nature's way of
injecting new variants into the system in order, possibly, to make up for
the deficiencies that may occur if what worked in the past no longer does
so because the world has changed. When John Odling-Smee and I first wrote
about this in the 1970s, we noted: 'In effect the g-t-r
[generate-test-regenerate] heuristic "gambles" that the future will be the
same as the past. At the same time it hedges its bets with aleatoric
(chance) jumps, just in case it is not.'"

So what Bill is saying is that it is good to become more radical when the
old, conservative solutions are no longer working very well. Maybe that's
why tonight for the first time ever in the U.S. Senate they are discussing
legislation for a national health plan!--Gary

[From Bill Powers (940808.1110 MDT)]

Gary Cziko (940808.0544 GMT) --

You're still alive. Splendid.

Bill, I appreciate your comments about the artificiality of artificial
selection that selects for movement toward a goal. But I'm having a
hard time (as you can see) following your other arguments. Perhaps you
could clarify a bit for the man on the net fascinated by both evolution
and perceptual control?--Gary

My argument is quantitative but unfortunately I don't have the numbers.
The difference between goal-directed evolution and blind evolution, as I
see it, is that if the result of a mutation is a decrease in fitness, in
the goal-directed case there is another mutation right away in the
relatively unfit population of offspring, whereas in blind evolution the
next mutation occurs no sooner than if the result had been beneficial.
In the relatively more fit population, the next mutation is postponed in
the directed case, while in the blind case the next mutation takes place
at the same time as usual (on the average). So the goal-directed case
will end up producing fit offspring in the species much more often (per
unit time) than in the blind case, and will allow reproduction of the
fit offspring for a longer time than in the blind case.

How much difference this makes depends on how much range there is in the
rapidity of mutations. With low baseline mutation rates, the main effect
would come from the relatively unfit population left over after the
previous mutation; the mutation rate would go up, increasing the chances
for favorable results per unit time, although not per mutation.

When this gets confusing, just go back to the ECOLI demo. Pretend that
you're monitoring some aspect of the species that would correlate
causally with fitness. When this aspect changes, there would be a
corresponding effect on fitness. You're trying to keep this aspect
changing toward some goal-state, represented by the goal shown on the
screen. The only way you can act is by hitting the space bar and causing
a random change in direction. Yet clearly you can get the "aspect"
(distance) to the goal state (zero) very handily.

On the other hand, if you just close your eyes and hit the space bar at
approximately regular intervals (or, really, any intervals you like),
the chances of maintaining an increase in this indicator of fitness
become far, far lower. If there were a population of you-s doing this
for a population of organisms, far fewer organisms would reach the goal
in a given length of time when you have your eyes shut.

Perhaps the key concept here is that of a biochemical variable which is
causally correlated with fitness. If the organism should contain some
such indicator, then relative fitness could be compared with some
reference fitness (in terms of the state of the indicator) without
requiring death of the organism, or total failure to reproduce, to
measure fitness. This removes a fatal objection to control of the
results of mutation, which is that no organism individually could
benefit from the knowledge gained from its failure to reproduce.

Actually, isn't the criterion of fitness usually stated in too extreme a
manner? All that is really required for differential survival is that
one variant reproduce _faster_ than another, and that there be
competition for resources. The losers in this game thus can stick around
for a long time, many generations, before finally being wiped out.
During that time, the effects of the competition would surely be felt
biochemically -- through relative starvation, for example. If the
biochemical changes were used as the basis for varying the mutation

rate, the losers could start churning out variants faster, and thus
produce more competitive offspring more often.

All this has to be worked out in simulation, taking both the normal
evolutionary factors and the addition of this systematic variation in
mutation (or crossover) rate into account. Do you know any people using
GA who might be persuaded to try this? It shouldn't be hard -- the ECOLI
algorithm is pretty simple.

···

-----------------------------------------------------------------------
Best,

Bill P.