NetNews Usenet Archive 1992 #18/NN_1992_18.iso/spool/bit/listserv/csgl/713

Text File | 1992-08-17 | 9.3 KB | 168 lines

Comments: Gated by [e-mail]
Path: sparky!uunet!paladin.american.edu!auvm!VAXF.COLORADO.EDU!POWERS_W
X-Envelope-to: [e-mail]
X-VMS-To: @CSG
MIME-version: 1.0
Content-transfer-encoding: 7BIT
Message-ID: <[e-mail]>
Newsgroups: bit.listserv.csg-l
Date:         Mon, 17 Aug 1992 12:40:24 -0600
Sender:       "Control Systems Group Network (CSGnet)" <[e-mail]>
From:         "William T. Powers" <[e-mail]>
Subject:      Re: separating action and perception
X-To:         [e-mail]
Lines: 152

[From Bill Powers (920817.1130)]

John van Loon (920817) --

Hello, John, welcome to the net as a real live speaker!

>To me it would seem that drivers do in fact see the disturbances. It is
>true that they do not see each individual one but they do see the net
>effect of them by noting the position of the car in the lane (this
>includes any steering errors also).

"Disturbance" is an ambiguous term; it can refer either to the variable
that is causing the disturbance (wind velocity, bump in the road) or to the
effect that the disturbance has (change in direction of the car's motion).
Just lately I'm experimenting with using "independent variable" to mean the
CAUSES of disturbances, and "disturbance" to mean the effect, if any.

>The action of steering the car so that it stays in the lane seems like >a
negative feedback type of action solely controlled by sight >(excepting for
large disturbances that change the wheel position and >the change is "felt"
and reacted to).

You're right, of course: this is a negative feedback system, which is what
I mean by a control system. The controlled variable, from the point of view
of an external observer, is the objective position of the car on the road.
But when we look at the situation from the standpoint of the driver, it is
a PERCEPTION of the car's position that's being controlled, which is not
necessarily the same as the objective position. When the driver sees a
satisfactory picture in the windshield, the car may not be centered in its
lane. So (and this is the basis thesis of the PCT approach) what the driver
is actually controlling is the perception, not its objective counterpart.

Note, by the way, that the perception doesn't control the action; the
action controls the perception.

The turning of the steering wheel is based on the difference between the
perceptual signal representing the car's position and an internal reference
signal, which is more or less a picture of how the road and hood SHOULD
look as framed in the windshield (in German, the name for reference signal
is Sollwert -- "should-be"). The difference, the error signal, is converted
into steering efforts in a direction corresponding to the sign of the
error. There are some dynamic filters involved to make the whole loop
stable, but that's not necessary for a discussion of the basic idea.

The controlled variable, the car's position, is neither a dependent nor an
independent variable. It's part of a causal loop. It depends on the action
(applying a torque to the steering wheel) at the same time that the action
depends on deviation of the perceived position from its reference state
determined by the driver. There are two independent variables in this
situation. One is the driver's internal reference signal. The other is the
set of all independent variables that can affect the car's position
independently of the driver's actions -- what I have been referring to as
"the disturbance." I'll say "independent variable" when I refer to the
environment, and "reference signal" when I refer to the driver's brain.
They're both really independent variables with respect to the steering
control system.

What drivers don't see are the independent variables in the environment.
They don't see the wind, and even if they see some dust blowing or trees
bending, they can't see enough to calculate the forces being generated on
the car. They can't see soft tires, little tilts of the road, and so on.
All they perceive is where the car is, which they judge relative to where
they want to percieve it. If the car swerves, there is no way of knowing
what caused the swerve; there might be many simultaneously acting
independent variables, which remain inseparable as their effects simply add
together, or just one.

This is actually the real power of a control system. It doesn't need to
know the causes of disturbances of the controlled variable. It simply
monitors the controlled variable itself, compares what it senses with some
reference-state for the perception, and acts to keep the difference as
small as possible.

The point I was making to Penni Sibun concerned the relationship between
the driver's actions (applying torques to the steering wheel) and the
perceptions that are controlled as a result (the observed position of the
car on the road, as the driver sees it). In some of the materials Penni
cited, the statement appears that action and perception are inseparable.
But when there are independent variables in the environment which have just
as much influence on the outcome as the actions do, it is the outcome that
remains the same, while the actions vary to oppose the effects of those
independent variables. So the perception (of the car's position) remains
essentially stable, but the actions (torques applied to the steering wheel)
vary as the independent variables in the environment vary.

As a result, you can be driving down a nice straight road with the steering
wheel in just about any position, depending on how much crosswind there is,
how much the road is tipped, and how well aligned your front end is.
There's no way to tell, just by looking at the car's behavior, what the
steering wheel is doing. So action and perception become decoupled; the
variations in action correlate mostly with the external independent
variables, and hardly at all with the controlled variable. The angle of the
wheel correlates highly with the sideward forces due to the crosswind, but
because of that, shows very little relationship with the direction the car
is going.

The reason that seeing or not seeing independent variables is important is
best understood in relation to conventional interpretations of behavior.
From the standpoint of an external observer, it seems that the independent
variables are causing the behavior -- the crosswind is causing the driver
to turn the steering wheel. If you were from Mars, and didn't know anything
about driving cars, you might conclude that the driver is somehow sensing
the wind, the deviation of the road from level, the state of the tires, and
all the other variables than influence the car's direction, and is
responding to them by turning the wheel by a calibrated amount. You would
see what seems to be a stimulus-response situation.

If you didn't ask too many questions about HOW the driver senses these
things, and how he or she does so with such precision, and how these
stimuli get translated so precisely into just the steering torques that are
needed, you might think you have an adequate explanation of the phenomenon.
It would never occur to you that there is another variable involved, the
path of the car, that is actually under active control. The fact that the
steering forces balance out all the external forces so accurately that the
car stays in its lane for mile after mile would not seem remarkable if you
weren't a physicist -- that's just good luck for the driver. It wouldn't be
at all obvious that what the driver is REALLY sensing is the position of
the car; the driver isn't sensing any of those supposed "stimuli."

This, I maintain, is the story of all conventional approaches to
understanding behavior. The supposed causes of behavior, the independent
variables or "stimuli," are really just influencing perceptions that the
organism has under control. The actions that result from applying these
independent variables are really opposing disturbances of the controlled
variable -- usually very successfully. Organisms are such good control
systems, unfortunately, that controlled variables do little varying that
can be related to external events. The result is that they're easy to
overlook. Statistical analysis is most likely to reject them as
insignificant, even if the experimenter accidentally includes them in the
list of variables to be tested for significance.

The AI and AL approaches (artificial intelligence and artificial life)
uniformly assume that the consequences of acting simply follow regularly
from the actions. As in most conventional sciences of life, they name
actions by their consequences -- "turning left" rather than "applying a
torque to the steering wheel." They would assume, for example, that
steering a car around a curve simply results from commanding that the arms
turn the steering wheel by a specified amount. As every driver knows,
however, it's sometimes both possible and necessary to go around a curve to
the left by turning the steering wheel to the right, if the curve is too
steeply banked for your speed or if a crosswind is blowing the car into the
curve, or both. There's no one "command" that can produce turning left in a
real environment. The system has be be organized to turn the wheel by ANY
AMOUNT and IN ANY DIRECTION that's required at the moment. Only a control
system can behave like that. A command-driven system can't.

I once heard it said that asking Kenneth Orr a question was like trying to
get a drink of water from a fire hose. I'm glad I'm not like that.

Best,

Bill P.