How does describing the moth's `fold up & drop like a stone' routine as
a feedback system amount to anything more than a decision to regard S-R
systems as degenerate cases of ECS's?
Avery.Andrews@anu.edu.au
[From Rick Marken (920318)]
Avery Andrews (920317) asks:
How does describing the moth's `fold up & drop like a stone' routine as
a feedback system amount to anything more than a decision to regard S-R
systems as degenerate cases of ECS's?
It's not just a matter of describing it as a feedback system -- it IS
a feedback system. The best answer to your question is Bill Powers'
article "Quantitative analysis of purposive systems" article in Psych
Review (1978). The main point relevant to your question is that the
moth's "sound control system" is just one of many cases where behavioral
scientists have made the mistake of applying an SR analysis to a feedback
control system. This mistake is even made in my field where psychologists
doing tracking tasks -- and who know control theory -- describe these
tasks in SR terms (for example, the distance from cursor to target is
seen as the stimulus that causes "control action" responses). In the case
of the moth, the sound emitted by the bat is probably seen as a stimulus
for the drop response. Of course, since dropping will occur when anything
produces the proper sound, the ethologists probably think of the sensed
sound as the stimulus for the drop. Since it is likely that sensed
sound is the controlled variable (with a fixed reference of 0) then
the ethologists analysis of the moth is like the psychologists analysis
of tracking (although, with the moth, the target is not "outside" as
part of the stimulus as it is in tracking -- but since the moth's reference
level for sound is probably 0, a sound level meter, which measures sound
pressure relative to a reference 0 sound level- gives a measure of the
"stimulus" which is like discrepency from the target in tracking).
So what is the problem with an SR analysis of the moth's control of sound
intensity? First, I should say that, with a reference fixed at 0 a
control model "looks like" an SR model -- even though it is not an SR model.
There really is no way to model the moth's behavior as an SR model -- because
there is also an R-S connection -- the sensed level of sound influences
the dropping (S-R) but the dropping also influences the sensed level of
sound (R-S). So there is a loop and any model of this process must take
into account the dynamic constraints that make the control loop work -- that's
one of the things in Bill's article cited above; he shows that
a sequential approach (implied by S-R type models) to the moth's behavior
(sound causes drop which then causes lower sound which then causes
no drop, etc) will only work when their is very low gain in the loop; other-
wise, the system become unstable.
So there really is no SR model of control (and the moth is controlling
something). I have read articles about SR models of movement in simulated
bugs -- the bug senses "light" and has outputs that move it toward
the light. The people who build these bugs think of the bug as an SR
device -- sensed light (S) causes response output (R). These bugs
work, however, because the appropriate dynamics have been built in
"accidentally" -- the programmers were not trying to stabilize the
control system. For example, the effect of light on output is to
produce increments in output proportional to the input. This incremental
approach makes the "SR" device into a proportional control system with
integrated output and a fixed reference level (which is implicit in the
equations that transform input to output).
Most SR analyses of behavior (like those of the moth) are usually done
in one's head -- it looks like SR. When working models are actually
built (and they work) they are actually control systems with fixed
reference levels (usually implicit in the S-R equations and, therefore,
at the 0 point of the range of the input variable). These models are
called SR models (the Braitenberg (sp?) "Vehicles" are another example that
I just thought of) but they are NOT.
So, what's the problem with looking at behavior like the moth's in SR
terms? Well, besides the fact that it's wrong and won't lead to the
correct detailed model of the moth's behavior in a real environment, there
are these explicit problems:
1) you won't realize that the reference for the controlled variable
(probably the intensity of sound in a particular frequency region) can
be changed (maybe) by the moth in order to accomplish some other goal
(like, get around a tree -- I dunno). Anyway, SR analysis of control
has this HUGE flaw -- it doesn't notice that variables are controlled; and
possibly at varying reference levels.
2) you don't see the bat as just one of many disturbances to the controlled
variable (sound intensity). Thus, you will observe all kinds of puzzling
variability in the moth's dropping behavior that will likely be attributed
to "error variance" or "random stimuli flying around" or whatever -- and
you will miss the fact that the controlled variable is kept precisely at
its reference level BECAUSE OF THESE APPARENTLY RANDOM VARIATIONS).
Obviously, I think you asked a very important question Avery. I hope
this post helps a bit.
Regards
Rick
···
**************************************************************
Richard S. Marken USMail: 10459 Holman Ave
The Aerospace Corporation Los Angeles, CA 90024
Internet:marken@aerospace.aero.org
(310) 336-6214 (day)
(310) 474-0313 (evening)
Regrettably, I don't have time to respond carefully to everything that has
been written about plummeting moths, since I have to get ready to go
off to Palo Alto in a week, but here's how I see things so far.
S-R psychology is a dead-end because it misrepresents what's going on
in at least three interrelated ways:
a) the nervous system is seen as a discrete transducer over an
utterly ill-defined class of event-types (stimuli and responses)
rather than a continuous transducer over vectors (I'm assuming that
the fact that its actually pulse-trains is irrelevant for the
time-scales relevant to behavior).
b) the functional importance of the R -> S link is ignored
c) it isn't noticed that much of what is going on is the
stabilization of the values of complex functions of inputs rather
than the production of transient effects under given circumstances.
And, just as Gary, Bill et. al. say, even though S->R psychology is
officially dead, it is still seriously alive as an influence in AI,
philosophy of mind, etc. (I guess I differ from Gary about Phil of
Mind in that I think there's plenty stuff to be salvaged from there).
(a) is an erroneous description of the psychological mechanism
(hopelessly vague except for the bit about discreteness, which is
false), while (b) and (c) are basically mistakes in ecology. By
which I mean: suppose someone were to ask why this neural circuit is
hooked this way rather than some other way? You can't (in most
cases) give a sensible answer without including the R -> S relation
in the story, and explain about how negative feedback loops work.
Now, considering my version of the moth (the one that crashes into
the leaf litter), I continue to insist that this is an S-R setup
rather than control, because, w.r.t. (a) in this particular case,
there really is a discrete stimulus producing a discrete response
[but N.B.: the stimulus is not a discrete *event* but a discrete
*condition*, e.g. something that becomes true and stays that way for
a while, and similarly for the response--I'm guessing that this
possibility is catered for in the standard S-R conception]. (b) even
though there is an R->S connection, it does *not*, in this particular
case, contribute anything to the selective advantage conferred by
this piece of gadgetry (except in the peculiar negative sense that if
the action that caused plummeting also caused the sonar signal to
become inaudible, without masking the moth from the bat, the setup
wouldn't to its job). (c) this particular system does not in fact
stabilize any perceptual function in the moth.
None of this means that S-R analysis is a good way to look at
behavior in general, in fact the reverse, as can be seen by
considering all the very specific conditions that have to be
satisfied for this particular arrangement to be useful. It works
because the moths happen to be in an environment whereby
a very simple maneuver will take make them invisible to bat sonar,
and not too vulnerable to other dangers (it will fail and thus not
evolve in environments where the ground is normally swarming with
ants hungry for hunkering moths, and equally fail in swamps).
It seems to me that there can be no harm for CT in recognizing this
kind of thing as an S-R setup, and going on for a bit about all the
rather special circumstances that have to exist in order for it to
work: it's suitable only for some uses by critters who can't afford
much in the way of brainpower, and rely on mass reproduction rather than
individual survivability. Marginality is not the same thing as
nonexistence, and looking at the properties of real cases where
S-R setups can work ought to make it even clearer why, most of the time,
they don't.
Avery.Andrews@anu.edu.aucl