[From Dag Forssell (951022 2100-2)] Part 2 of 2
Thread on geese from CSG archives, continued.
[From Bill Powers (910615.1000)]
On Geese:
Here's another slant on the "open loop" problem. There is lots of
open-loop behavior. For example, when you turn the wheel of a car,
the result is to alter the wind noise, isn't it? So
wind-noise-altering behavior is open loop (wind noise doesn't
affect steering). When you open a can of soup, you create a
particular pattern of serrated indentations around the rim of the
lid. Serration-creating behavior is open loop.
Almost every act that an organism produces generates behavior in an
open-loop manner -- if you define behavior as any effect of motor
activity that you happen to notice. This is what happens when you
take the observer-centered view of behavior. The naive observer has
no way of knowing in advance which effects of another organism's
motor activities are important and which are not. It's an
undeniable fact that when you hear a person in the next room typing
messages on a terminal, that person is creating clacking sounds in
rapid succession. So knowing nothing about what the person at the
terminal intends to be doing, you can only say that the person is
producing clacking behavior. You pick the effect that strikes you
as "salient." From a theoretical point of view, that is equivalent
to picking effects at random.
The chances of stumbling across a true controlled variable in this
way are minuscule. Most of the effects you notice will be
side-effects, of no interest or importance to the system doing the
behaving. The observer simply notices what his or her own
perceptions are concerned with. That is why those who study E.
coli's mode of locomotion can marvel at its ability to "navigate
through space." Of course E. coli does no such thing: all it can do
is control concentrations of various substances at its chemosensory
inputs (or their rates of change). A human being, however, sees
that E. Coli is "swimming up a gradient," a concept that involves
spatial concepts that are not part of E. coli's world. E. coli
simply cannot control spatial variables, even though we can see it
affecting them. E. coli doesn't even know that it's moving.
When Konrad Lorenz sees a goose arching its neck and using its bill
to move an egg, he is seeing all sorts of things that probably mean
nothing to the goose. The arching of the neck creates a graceful
pattern in human perception, because we are looking sideways at the
goose's neck with our eyes, a view that the goose can't possibly
have. The human observer sees the neck and bill cooperating to
bring the egg "closer to the nest." Maybe that concept is part of
the goose's world, and maybe not. The only way to find out what the
goose is controlling is to test hypotheses -- and to be prepared to
enter a perceptual world that is completely different from ours.
The goose is not controlling the curvature of its neck as seen from
outside the goose. It may be varying it, but it isn't controlling
it.
From the standpoint of traditional science, point of view and
perceptual interpretation simply don't arise. The scientist, after
all, looks at the world the way it actually is. If the goose's neck
fits a curve that is expressible as a cubic function of several
variables, then that is a measure of "behavior." The goose's
nervous system must be hooked up to compute that cubic family of
curves. The idea that this curve is a completely accidental and
meaningless side-effect of the goose's actual control processes
would never occur to a scientist who automatically assumes the
existence of a single objective universe -- the one he or she
experiences, of course.
Most of the "problems" that people throw up against control
theorists are of this same nature. They concern effects of motor
activities that look interesting to a particular observer, but have
not been shown to have any significance to the behaving system. And
whenever open-loop behaviors are pointed out (often just to try to
find a hole in control theory), somehow the question never comes up
as to how open-loop behavior could ever have become organized. If
an effect of motor behavior never reflects back through some
external closed loop to inform the system of the effects of that
behavior, how could it be that an organism can produce that same
effect over and over by using motor actions that never repeat
themselves? Even if you consider control theory to be only a
vaguely-possible interpretation of behavior, how can you put up
against it an explanation that depends on something even less
possible?
In Gary Cziko's quotes from Lorenz, did you notice how Lorenz
describes the outputs that produce the visible effects without
apparent concern for just how those outputs instead of others come
to be produced? To speak of movements that bring the egg closer to
the nest is like speaking of steering-wheel movements that bring
the car closer to the center of the lane. If you focus on outputs
alone, you skip right over the central question of behavior, which
is how just the right outputs happen to be produced. Control theory
faces this question and answers it. All the other approaches dodge
it, or beg it.
Tom Bourbon (910615) --
Don't feel too sorry for your students. They may face difficulties
later in life through having come to understand control theory, but
it is up to them to choose between continuing with that
understanding and doing what they can to advance it, or seeking a
more comfortable life by giving in to the majority opinion. I have
seen plenty of evidence that those who understand control theory
elect to continue with it, considering the rewards of understanding
greater than the material rewards available elsewhere in life. Why
shield them from the problems that all control theorists have
faced? They can handle them.
I hope that Wade Harman can come to the meeting and show us his
experiment. At the very least he deserves to experience the
reaction he will get from people who understand what he has done.
···
Subject: Geese; Harman's experiment
Subject: Geese, Control
[From Rick Marken (910615)]
Bill Powers (910615.0100) beat me to the punch once again.
Throughout the course of this "open-loop" brouhaha I forgot to
invoke my own favorite mantra -- "control theory is a theory of
CONTROL". If a result produced by an organism is not a controlled
result then the theory of control is obviously irrelevant. Open
loop behavior is simply uncontrolled results of action -- like the
clicks made while I type. They are accidental side effects. That's
what my whole "Mind reading" demo is all about. If, indeed,
something the goose is "doing" is open-loop then control theory
just doesn't apply. There is no reason for alarm about this --
there are many things that the goose does that are unquestionably
better explained by open loop models. The goose's acceleration as
it falls off the goose coup is one example. I think this is also
relevant to Tom's comments about the lack of acceptance of control
theory by colleagues -- something I am indeed quite familiar with.
I forget that one nice approach to dealing with it is to just say
to yourself (and your colleagues) -- "I'm just interested in a
different kind of behavior than you. I am interested in studying
control. And there is plenty of it to study. You go ahead and study
that other kind of behavior -- the kind that is just emitted
output. Have a good time and work hard. Just don't point to the
existence of your kind of behavior as evidence against control
theory. Control theory isn't trying to explain your kind of
behavior. It just tries to understand control. Have a nice research
program. Bye"
Love Rick
Subject: Geese and Monkeys
[from Gary Cziko] Sat, 15 Jun 91 22:24:13 -0500
Note: This note is a bit of a ramble and I think I've answered my
own questions in writing it, so it might not be worth the trouble
reading through if geese, mango-slurping monkeys, ethology and
evolution are not your interests.
Tom Bourbon; Bill Powers; Rick Marken (all 910615):
Thanks so much for your attempts to understand my problems with
Lorenz's goose. I can follow without difficulty your reasoning,
but I still have a problem with using THE TEST with the goose's far
to near egg-rolling behavior, and the computer terminal clacking
behavior now raises another problem for me.
Concerning THE TEST of the controlled variable, let's say that the
goose is controlling for a given amount of push (force) against the
egg in a certain direction (back toward the test). We have already
seen how the lateral movement of the egg seemed to be controlled by
the goose (Lorenz calls this taxis). Now, if I wanted to make a
robot arm that could do the same as the goose, I can't see how I
could use anything but a control system for lateral movement, but
it still seems to me I could use an open-loop system to bring the
egg closer. I would just have it set up to provide a given amount
of force to the egg, no more, no less and let it go. So what can
I do with the goose to show that the far-to-near movement is
controlled? What would be THE TEST that would make it clear to us,
and hopefully to an ethologist as well, that it is not a fixed
motor pattern? Would it just be the observation that the goose's
neck is not arched exactly the same way each time? Hm, perhaps it
would be as simple as putting a weight on the goose's neck to see
if it could still roll the egg back. Hey, I think I've just
answered my own question. Can it be that simple? Just add some
mass to the systems and if it is still successful it is clearly not
open loop?
Now, let's get to the open-loop behavior that accompanies the
closed-loop stuff. Let's say that in eating mangoes, a monkey
makes a certain kind of slurping sound. Of course, he is only
controlling for getting the mango in his mouth. But the slurping
sound is loud and attracts and turns on all the surrounding
females. Of course, he takes advantage of the queue of receptive
females waiting for him to finish the meal and so leaves lots of
progeny including males with an appetite for mangoes and female
with an appetite for mango-eating males. So mango-slurping can
become very important evolutionarily even though it is not a
controlled variable. The monkey is not eating mangoes or slurping
to attract females, but these behaviors in fact turn out to be a
very important behavior. Somehow, this doesn't seem right.
But wait. I bet these male monkeys are too smart not to catch on
and will start to slurp even when there isn't a mango in sight but
females are. And then it will be the males who ARE controlling for
slurping sounds (and successfully so against disturbances, such as
lack of mangoes) who will be the most successful to reproduce. Of
course, what they REALLY are controlling for is attracting females,
with slurping just part of the program. Have I answered my own
question again? Let me know. It seems like a perceptual control
theory perspective on the evolution of controlled behavior could
provide some real insights.
All of a sudden, I feel better about taking on Konrad Lorenz
again!--Gary
Subject: Geese and monkeys
[From Bill Powers (910616.0730)]
Gary Cziko (910615) --
Re: Lorenz's Goose:
It's turning out to be an interesting exercise.
...if I wanted to make a robot arm that could do the same as
the goose, I can't see how I could use anything but a control
system for lateral movement, but it still seems to me I could
use an open-loop system to bring the egg closer. I would just
have it set up to provide a given amount of force to the egg,
no more, no less and let it go.
How do you set up a robot arm to "provide a given amount of force
to the egg, no more, no less?" The force applied to the egg doesn't
depend on the arm position alone, but on the inertia of the egg
combined with resistance from rolling the ovoid (what else?) egg
over bumps and from friction caused by the rolling movement at the
point of contact with the arm (beak). If the arm is not in contact
with the egg, no force is applied to the egg. If there is contact,
the amount of force applied depends on how the egg rolls.
It would be possible to set up the robot arm as a position-control
system, and then give it a smoothly varying reference signal
sweeping from "far" to "near." This control system, however, would
not behave as the goose does when a heavy egg is substituted or
when something stops the egg. When the motion is disturbed by a
large opposing force, the control system would simply increase its
output force enough to keep the position matching the changing
reference signal. So the force on the egg would rise if the egg hit
an obstacle or became heavier. This would also happen if the
control system controlled velocity of movement. Anything that
slowed the arm would result in an increase in applied force.
What we need is for the velocity of the arm to depend on the amount
of resistance that is felt. This can be accomplished in several
ways using tactile sensors on the arm. One way would be to have the
reference-velocity be determined by the output of a force-sensing
control system of higher level. With zero sensed force the velocity
would be at some maximum (not very high). When the arm comes in
contact with the egg, starting to compress the force sensor, the
sensed force begins to rise rapidly toward the reference level and
the velocity slows. Velocity ceases to decrease when the egg starts
to roll; the sweep then continues at a slower speed. If we measured
the velocity of the goose's beak during the rolling, we should
observe that it decreases as the egg rises over obstacles and
increases as it rolls down the slope after passing over an
obstacle. If the obstacle is high enough, or if a heavy egg is
substituted, the velocity will slow until the sensed force exceeds
the reference force by some amount, at which point the velocity
will drop to zero: the beak will stop moving.
This isn't the only design that will work, but the point is that
without sensing the resistance of the egg to a push, there is no
way to control the force applied to the egg. There must be a
force-control system. In a real goose we would probably find the
effects of touch occurring right in the spinal reflex loops; in
human beings at least, tactile sensors feed back negatively to
spinal motoneurones. But I'm just trying here to find a sufficient
model, not a realistic one.
If Lorenz had been inclined to quantitative measurements, he might
have used a high-speed camera, or a pair of them, so that the
position and velocity of the beak and the orientation of the egg
could be measured instant by instant. The above hypotheses then
could be tested, and a better picture of the controlled variable
could be obtained. Even better would be to equip the goose's beak
with force transducers so we could get an idea of what the goose is
feeling. It isn't likely that a naturalist would do experiments
this way, though.
Have you tried this with a real egg? I think it always helps to put
yourself in the control system's place if you can. Put an egg on a
blanket (on a table) with some folds in it. Use your extended
forefinger to pull the egg toward you -- with your eyes shut (I
assume the goose can't see the egg under its chin very well). I
think it will be pretty clear that you control the way the egg
feels against your finger as you pull it toward you and that you
sense where the egg is in terms of the kinesthetically-sensed
position of your arm and hand. If you use a raw egg and a nice
clean expensive blanket, you will be quite careful not to apply too
much pressure. You will also understand how the goose uses multiple
tactile sensors to sense the direction in which the egg is rolling
and keep that under control. If the egg rolls ahead of your finger,
you will probably increase the velocity a little until you feel the
egg again. If the egg rolls all the way to the target, or if
someone reaches in and snatches it away, your finger will execute
the whole sweep until it reaches the target position. You're
feeling for the egg, but you don't want to be moving too fast when
you encounter it again. If someone reaches in and stops the egg,
you will push a little harder, but not TOO much harder. And your
arm will stop moving.
Most observations like those of Lorenz need to be done all over by
a control theorist. The necessary observations just weren't made.
------------------------------------
Monkeys Munching Mangos --
There's something in this example and the way you worked it out
that could be interesting in studies of learning.
The implication is that there can be SYSTEMATIC learning of new
control systems. Of course the effect of the side-effect has to
relate to some variable for which the animal already has a
reference level, and perhaps something that is already under
control (but not completely satisfactory control). Slurping on the
mango proves to reduce the error in some other control system. The
next thing is for that other control system to get connected so
that when it experiences error, the error signal is routed to the
slurping control system as well as to systems already being used.
I would predict, therefore, that the monkey would slurp AND do all
the other things it was used to doing to attract females. The
eating of mangos becomes a lower-order control system (one of
several) with respect to the system that attracts females (as well
as being in the hierarchy concerned with getting food). I suppose
that, as you say, the actual mangos might drop out if the slurping
alone attracted them -- but if the mangos attract the girls, maybe
the monkey just imagines the mango part when there aren't any real
ones, not being smart enough to know that nobody else can
experience what it's imagining. On the other hand, slurping alone
might well alert females that somebody in the vicinity has found a
mango, through their own imagination connections. I should think,
though, that hungry males would tend to become a problem. They'll
respond to the prospect of mangos too, although they wouldn't
respond to sexual innuendos from another male (usually). The
slurping strategy might result in more conflict than is acceptable.
It seems like a perceptual control theory perspective on the
evolution of controlled behavior could provide some real
insights.
Naturally, I agree.
---------------------------
End archive