NetNews Usenet Archive 1992 #18/NN_1992_18.iso/spool/bit/listserv/csgl/712

Text File | 1992-08-17 | 7.0 KB | 136 lines

Comments: Gated by [e-mail]
Path: sparky!uunet!paladin.american.edu!auvm!AERO.ORG!MARKEN
X-Delivery-Notice:  SMTP MAIL FROM does not correspond to sender.
Posted-Date: Mon, 17 Aug 92 11:08:33 PDT
Message-ID: <[e-mail]>
Newsgroups: bit.listserv.csg-l
Date:         Mon, 17 Aug 1992 11:08:33 PDT
Sender:       "Control Systems Group Network (CSGnet)" <[e-mail]>
From:         [e-mail]
Subject:      Generating Output
Lines: 123

[From Rick Marken (920817.1100)]

Gary Cziko (920816.1715) says:

> Greg was telling me that the arm demo does generate
>output.  It may not calculate trajectories, but it nonetheless generates a
>behavior based on the error signal of the difference between its reference
>level and its perception.

>When I show Demo2 to someone with computer savvy, he or she can say that
>what the model does is a reiteration of
>perception-behavior-perception-behavior, etc. a la TOTE.

>but by using digital computers don't the demos show in fact that
>control can be accomplished by generating output based on input (sounds
>pretty S-R to me)?

I think it is fair to think of components of the loop in s-r terms. External
variables get transformed into perceptions so p = f(i) -- a functional
relationship between a stimulus and a response; perceptions get transformed
into errors, errors (as you note) get transformed into outputs and outputs
get transformed into inputs, completing the circle. As you note, the computer
computes these s-r components of the loop sequentially; but it must mimic the
dynamic aspect of these functions (the fact, for example, that perception does
not instantaeously become p when the input is i) and it does this by only
computing a fraction of the computed "response" on each iteration of the
sequence and integrating each computed fraction over iterations. Because
things work in this nice, dynamic way, it means that the response to a
particular input is never exactly the same -- it depends on what went
before and what else is currently affecting the variables. So its not really
correct to say that the control loop works (ie. controls) by generating
precisely computed outputs, although all the components of the loop do
produce outputs in response to inputs -- at least they "start" to produce
those outputs -- so you can call them sr components. For example, suppose
that the s-r relationship between error and output is o = 10*e. So when
e = .1, o should be 1; but what really happens is o just starts to change
towards 1. So o might be .2 when e = .1. At some later time, as the control
system works to get the perception under control, e (which is always changing)
might be .1 again. At that point, o might already be .9 so when o starts
"responding" to the error, o ends up as 1 -- what it is "supposed to be".
The point is that the dynamics of the loop turn the apparatly exact functional
mapping of error in output (or of perception into output) into a continuously
changing relationship.

But it's probably not worth trying to teach your listener differential
equations in order to get the point across that control systems do not
work by computing functionally exact outputs (at each point in the loop).
The fact of the matter is that when you connect s-r components together
in a dynamically stable negative feedback loop, what that loop does is
keep its perceptual signal equal to its reference signal (if there is
an explicit reference signal) or to zero if there is no explicit reference.
This is the fact that was missed by those who proposed the TOTE loop as a
model of purposeful behavior. A properly functioning TOTE loop is a control
loop; and what it controls is its perceptual variable. The implications of
that simple observation are enormous, but only if you notice that that's
how TOTE loops work (incidentally, TOTE is probably a poor name since control
loops don't EXIT -- unless they are explicitly shut off by another system;
so the TOTE loop, when it works (ie. is a control loop) is really a TOT loop).

If there were a simple way to convince your computer savvy friend (or anyone
else) that this is the way so- called "perception-behavior-perception loops
work (they control perception -- not output) then Bill Powers could have saved
30 or so years and psychology would now be a real science.

John Van Loon (920817) says:

>    To me it would seem that drivers do infact see the disturbances.

A good way to see that this is not the case is to look at my demo described
in chapter 3 (first paper) in my Mind Readings book. Another way to
see this (mentally) is to remember that the visual image through the
windshield of the car is ALWAYS simultaneously the result of disturbances
AND driver outputs. Even the effect of a sudden, transient disturbance
depends to some extent on what the driver (and other disturbances) were
doing at the time it occurred. When control is good, responses (outputs) are
always the exact opposite of the net effect of disturbances on perception;
there is nothing in the perception that could be used to infer the net
effect of the disturbance. That is such an amazing fact that even people who
understand control theory can hardly believe it (right Gary).

> 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

In fact, this illusion is so compelling that probably fewer than .001% of
people currently working in psychology know that it is an illusion. Actions
control perception, not vice versa. I know it's hard to believe -- but it's
true (and demonstarted over and over again, in many different ways); do
read that paper in "Mind Readings"; do the experiment yourself; try to
find something in the sensory input that is related to the output that
affects that input. You might be startled too -- and see why those of us who
understand PCT are so excited about it.

>My
>problem is with the driver not being able to "see" the disturbances and not
>controlling the position of the car on the road.  Could you clarify the
>point a little further it would help me believe in your methods and theories

One way to do this is to just look at the equations and believe that they are
true (which they are). For example, the closed loop equations say that

o = -kd

Output,o,depends on disturbances,d, (environmental events) -- perception
is not part of this relationship (when the loop gain is high enough).

The other way is to set up the demos and run them yourself. Do experiments
yourself. Try, as I said, to find the perceptual variable that guides
responses in a tracking task -- when control is good. You will find that
perception does not guide output in a control loop. Output, however,
guides perception. In fact, behavior is the control of perception (is there
an echo in here?).

Best regards

Rick

     **************************************************************

Richard S. Marken                   USMail: 10459 Holman Ave
The Aerospace Corporation                   Los Angeles, CA 90024
E-mail: [e-mail]
[phone] (day)
[phone] (evening)