Feedback, Unsegregated information

[Richard_Marken]

[From Rick Marken (960725.1300)]

Me:

If you are controlling for hearing "you're getting warmer" then these
phrases are states of a controlled perception

Bruce Abbott (960725.1225 EST) --

It's a bit more complicated than that...The lower-level system provides a
nice example of e-coli style control: Hearing "colder" leads to a change in
direction; hearing "warmer" leads to maintaining the current heading.

It's even more complicated that that. Among other things, the subject must
learn the relationship between the words ("warmer" and "colder") and his
location in the world; this learning is usually accomplished by telling the
subject that "warmer" means "closer to the target" and "colder" means
"farther". The subject also has to learn the relationship between his actions
and the words; that changing his location in the world influences which
word is said and that blinking and scratching and such don't.

Once all this is understood, the subject can start varying his location in
the world in order to control for hearing "warmer" (if he wants to find the
target). The variations in location are probably not random, like E. coli. A
canny subject can remember whether the results of previous actions produced
the perceeption of "warmer" or "colder" and bias the selection of actions
based on previous results.

Jeff Vancouver (960725.13:00) --

Bottom line, I think some control theorists need to study both kinds of
feedback. I am.

I don't understand. "Feedback" refers to only one thing in a control loop;
the effect of a variable on itself. This effect can be positive, negative or
zero. Only negative feedback results in control. People who don't know a
control system from a banana peel use the word "feedback" to refer to
disturbance variables (variables that are independent of the control system's
output, like unsolicited comments about a presentation) and controlled
variables (like the words "warmer" and "colder" in the "warm/cold" game).

What are the two kinds of feedback that control theorists need to study? It
seems to me that we already are studying all variables relevant to the
behavior of a control system: disturbances, outputs, controlled quantities,
controlled perceptions, reference signals. What are we missing?

Me:

Are you saying that I get information about a.1*o + b.1*d1+b.2*d2+...b.n* dn
but that I don't segregate the information about b.1*d1+b.2*d2+ ...b.n* dn
out?

Martin Taylor (960725 1400) --

Yes.

OK. Now let's simplify. In our compensatory tracking experiments we typically
set things up so that a.1 = 1, b.1 = 1 and b.2 ...b.n = 0 so

p = o + d

That means that the disturbing influence on p (b.1*d) is equivalent to the
disturbing variable (d): they are both d. So you are saying that in such a
tracking experiment I get information about the disturbing influence, d, but
I don't segregate that information from informationh about the other
influences on p, in this case only o. Is that correct?

the very idea of the information being segregated verges on the absurd.

I agree.

The demonstration that it _could be_ segregated and used independently to
generate an approximation of the disturbing influence waveform was not
intended to show that this segregation is actually done in a control system.

Ok. So you demonstrated that information about d _could be_ segregated from
information about o. I take it that this was demonstated by your solution of
the control system equations for d given p. And now you say that this
demonstration was _not_ intended to show that the segregation of information
about d from that about o is actually done by the control system. Am I
following?

If so, then could you explain the use to which the control system puts the
unsegregated information about o and d (that is, the scalar value p)? Is
there any difference between saying "the system uses p as the basis for
output" and saying "the system uses the unsegregated information about o
and d as the basis for output"? If not, wouldn't it be simpler to just drop
the part about the "unsegregated information" and just talk about control of
perception.

Best

Rick

[Bruce_Gregory9]

[From Bruce Gregory (960725.1610 EDT)]

Rick Marken (960725.1300)]

"Feedback" refers to only one thing in a control loop;
the effect of a variable on itself. This effect can be positive, negative or
zero. Only negative feedback results in control. People who don't know a
control system from a banana peel use the word "feedback" to refer to
disturbance variables (variables that are independent of the control system's
output, like unsolicited comments about a presentation) and controlled
variables (like the words "warmer" and "colder" in the "warm/cold" game).

Stepping on a banana peel is an example of positive feedback,
no?

Bruce

[_Martin_Taylor]

[Martin Taylor 960726 10:30]

Rick Marken (960725.1300)

And now you say that this
demonstration was _not_ intended to show that the segregation of information
about d from that about o is actually done by the control system. Am I
following?

Your comments suggest to me that you are following, but remain puzzled as
to what it's all about. I think that everything you have said up to this
point is correct.

If so, then could you explain the use to which the control system puts the
unsegregated information about o and d (that is, the scalar value p)?

I'll continue to try, but it needs more than what I had written in
my the text I started to develop for the Web page on the topic
(Martin Taylor 960703 10:30).

Remember that the intention is to deal with an analytic approach that
is analogous to a Laplace Transform analysis. So your question is akin
to asking "Could you explain the use to which the control system puts
the Laplace Transform of the Perceptual Input Function." It doesn't,
as such, but changing the Laplace Transform of the PIF of a control
system changes the behaviour of the control system. Likewise, changing
the information rates between disturbance and perceptual signal (for
example) changes the behaviour of the control system.

If not, wouldn't it be simpler to just drop
the part about the "unsegregated information" and just talk about control of
perception.

Yes, if all you want to do is to say that there _is_ some degree of control.
No, if you want to say more than that about the behaviour of the control
system.

What comes next in the discussion is probably an extension of the analogy
between filters, amplifiers, and observers that was begun at the end of
the text I posted. Thereafter, we follow the information relations backwards
around the loop, from disturbance to output to error to reference and
perception to CEV to output...

The end product is a kind of statement about the degree to which control
can protect the next level from being affected by the environment and allow
it to be affected by its own output (that provides the reference signals
for the lower level control). In linear systems there are other, easier
ways to get the same answers, but informational analysis does not know
anything about linearity or nonlinearity, and can be used with any system.

Anyway, except when I get exasperated enough to try again to explain the
informational approach, I _do_ just talk about "control of perception", as
you would like. You'll notice that I seldom talk about Laplace Transform
analysis of control systems, either. It's when one wants an informal
analysis of what might improve control, or where problems might lie, that
informational considerations are most potent, because many of the results
are of the form "condition A <= condition B", rather than of the form
f1(x1,...) = f2(y1...).

In our compensatory tracking experiments we typically
set things up so that a.1 = 1, b.1 = 1 and b.2 ...b.n = 0 so

p = o + d

That's so, but remember that you really ought to write p(t) = o(t) + d(t).
More properly, both environmental feedback functions should be written as
extended over time, but that (speaking loosely) could be included in o and d.
Writing "p = o + d" leads you into trouble when you complete the loop
equations in the same vein. You wind up with p = f (p,...), whereas you
should wind up with p(t) = f(p(t-tau),...).

I'm away again from tomorrow until around Aug 16, and then back for about
10 days. So there won't be much more, if anything, on this from me until
then. And I've only just caught up with the e-mail backlog from my last
absence, with several messages awaiting replies from that time:-(

Martin