NetNews Usenet Archive 1992 #19/NN_1992_19.iso/spool/bit/listserv/csgl/883

Text File | 1992-09-02 | 5.4 KB | 103 lines

Comments: Gated by [e-mail]
Path: sparky!uunet!paladin.american.edu!auvm!BEN.DCIEM.DND.CA!MMT
Message-ID: <[e-mail]>
Newsgroups: bit.listserv.csg-l
Date:         Wed, 2 Sep 1992 00:53:58 EDT
Sender:       "Control Systems Group Network (CSGnet)" <[e-mail]>
From:         Martin Taylor <[e-mail]>
Subject: Re: Delayed; misc subjects
Lines: 92

[Martin Taylor 920902 0030]
(Bill Powers 920901.1000)

>
>But doesn't this say that there "is" a CEV with its own definition Out There in
> the environment, and that the percept simply represents it?
>
>I'm trying to make the percept the central concept, because that's what the
> control system knows and what it controls. The CEV is found (by somebody who
> knows more than we do) by working backward through the perceptual function and
> finding what combinations of external variables would be sufficient to produce
> that percept via that input function. In general there are many combinations
 of
> external variables that will lead to the same percept -- that's the other side
> of the concept "invariant."

More precision of language is clearly required.  I'm almost sure we have the
same ideas, but not quite sure.  There is a problem distinguishing the
"something" that might be "out there" and the percept of that something that
is the actual object of control.  I introduced the concept of the CEV some
time ago with a vague notion that it was useful to separate the two concepts,
and that the C*E*V (emphasizing the E) might be different from the percept
to which it gives rise.  (This is really difficult to put into precise words,
partly because the idea is not precisely formed).  Take an example--optical
illusion.  Suppose you are cutting the birthday cake so that "one cut, the
other choose" is fair.  But knowing the nature of illusion, you cut so that
the bigger-looking piece is actually smaller.  The other satisfies (apparently)
the reference for maximizing cake, but actually chooses the smaller piece.
Perceptually it was bigger, and only knowledge of a different kind allowed
you, the cutter, to be satisfied that you actually got the bigger piece.
I don't know whether that's a good example.  Maybe it gets the idea across.

The base problem is that no ECS can know anything about what is "really" out
there, so from one point of view, the idea of an external CEV is nonsense.
But there are different ways of determining percepts that "should" be the same.
Weighing the cake pieces "should" give the same result as looking at them,
so far as telling which is bigger.  But the two percepts that "should" be the
same are not.  We (external to either) infer that there is something there
that is being misperceived in one way or another, even though from the point
of view of either ECS, there is no problem.  You are quite right in calling
it a third-party issue.
>
>
>I do want to cavil about control existing when one person happens to control
> something the way somebody else wants it controlled. I said this to Greg
> Williams but it's worth repeating. Control requires a closed loop. If one
> person is controlling a variable, it's highly unlikely that another can
 control
> it at the same time without generating conflict.

What about four people carrying a bed by the four corners? At one level, they
are all controlling the location of the bed in coordination with each other
(Tom Bourbon demonstration, if I remember correctly).  At another, they are
controlling it to satisfy a reference that an employer should be happy, and
the employer has provided input that each of the four adopts as a reference
for the location of the bed.
>
> Exploratory behaviour, then, is NOT >the control of perception, but the
> discovery of how to control >perception.
>
>Tsk, tsk, Martin. Are you saying that the goal is to learn how to control, but
> without any perception of whether you've learned to control? You forgot to go
> up one more level.

I didn't forget.  Maybe I wrote too much, and it got lost in the shuffle, but
I explicitly limited the non-control aspect of exploration to one level,
and more properly one ECS.  An ECS above may well be in control, using, say,
the gain of that ECS as something to be controlled (as you mentioned a
month or so ago as a possibility).  And certainly ECSs below would be in
control during the exploration.  But the point of the exploration is that
the explorer cannot guess what percept will occur, and should not attempt
to develop any particular percept.  That would be "looking for", not "seeing
what," which is the kind of exploration I'm considering.  The explorer is,
I still think, finding out what is likely to happen to its percept under
different contextual circumstances when it generates an output.

>
>Exploratory _behavior_ -- meaning observable actions -- is just like any other
> kind of behavior. If you see someone sharpening a pencil, you can't tell
> whether he's sharpening it because he doesn't know what will happen and wants
> to find out, or because he does know what will happen.

Yes, that's the permanent "Third Party Problem."  You can't know, ever, though
you may often make guesses that YOU think are pretty good. He may be
sharpening the pencil just to get you wondering about that very question,
with no concern either for getting the pencil sharp or for finding out what
the sharpener does.

==========
It's still fun, but Paris beckons and I've got to get more sleep.  Bye.

Martin