Types of selection

[From Bill Powers (941006.0720 MDT)]

Gary Cziko (941005.1944 GMT)--

I'm glad your book is making progress again. What a bummer, that one
referee could hold it up like that. I don't wish you luck, I wish for
the MIT editorial board to act like scientists. I can't imagine holding
up a scientific publication just because someone disagrees with its
content. That's not refereeing, it's gatekeeping. The time to argue with
what you say (as I may well do!) is AFTER you have been heard from.

1. Does it make any sense to think of the normal functioning of a
properly working control system as a type of blind-variation-and-
selection process?


I suppose the thinking here is that the output function of a control
system makes no pre-calculation of how much output to put out. It just
gives whatever its amplication and energy source allows. It only "hits
the brakes" when it gets feedback. Without the feedback, the output
function is totally stupid, totally blind. The stability and success
of the control system is thus due to after-the-fact selection, not a
priori calculation. In spite of my "radical selectionism," I hesistate
characterizing control systems in this way, although reorganization is
certainly seen as a variation-and-selection process in PCT. Comments?

We really have four cases to consider. Variations can be random or
systematic; selection can be natural or goal-directed.

1. Random variation, natural selection

This is the usual meaning of "blind variation and selection." The
variations occur at random, not as a function of prior events. Selection
occurs out of the natural consequences of variation interacting with
environmental events, but does not affect the variation process in any

2. Systematic variation, goal-directed selection

This is normal negative feedback control. Here systematic variations are
driven by a regular selection process based on the outcomes of the
variations, with (perception of) the outcomes being compared with
desired outcomes and the difference driving the variation.

3. Systematic variation, natural selection

This is open-loop or command-driven behavior. Variations are generated
by systematic algorithms, but selection results only from interactions
between the variations and the environment. The variations do not depend
on the selection process.

4. Random variation, goal-directed selection

This is the E. coli or reorganization case. The outcome of a random
variation is sensed and compared with a goal. The difference determines
the interval between random variations, but does not affect the outcome
of any one variation.

The critical difference here is between 1 and 4: random variation
coupled with either natural selection or goal-directed selection.

Imagine the E. coli setup played under two conditions: the player
operating as usual, or the player blindfolded. In either case, survival
depends on getting the moving point to a target in the center within 5
minutes (arbitrary rule).

The blindfolded player and the sighted one can produce only random
variations in direction of the marker. But the sighted player can
refrain from hitting the key when the marker is moving toward the target
and hit it again sooner if the marker is moving away from the target.
The blindfolded player, who has no idea of the direction of motion, can
only press the key blindly -- without seeing the consequences. This
removes the relationship between the consequences and the timing of the
random variations: there is no systematic selection process that can
influence when a random variation will be triggered. Only the natural
geometry of the situation and the natural laws governing random walks
select for survival.

In a large population of "players" of this game, some players under each
condition would survive; even the blindfolded players would sometimes
hit the target. But the sighted players would very rapidly predominate.



2. How does physiological adaptation work? Is all or some of it also
based on a type of variation and selection process, or do muscles just
immediately "know" to get stronger after they have been exposed to

The latter, I think. There would be no advantage of a variation-and-
selection process over a systematic process here. Muscle cells could
simply proliferate if their average tension got too large, and atrophy
if it got smaller. There must be very few cases -- I can't think of any
at all -- in which this would not be the proper action. I find I'm
pretty vague about how muscles add bulk: do the muscle cells themselves
go through mitosis, or are more of them somehow generated from precursor
cells? At any rate, since we can think of a purely systematic process
that would work under just about all conditions, I don't see any
advantage in a process that works in the right direction only some of
the time.

Isn't that the basic question? The vary-and-select process is best where
there is no general rule that will remain applicable, and where any
general rule would frequently prove disastrous if followed under the
wrong conditions. Obviously, the vary-and-select process creates, among
other things, systematic control systems that do not rely on random
variation. As long as these systematic systems keep critical variables
where they belong, there is no need for further vary-and-select
processes. Or put another way, the systematic control processes prevent
application of selection pressures that might end up modifying those

Bill P.