[From Rick Marken (920925.1000)]
Jeff Hunter (920923 - 1) --
This gives us two common meanings for the word "control"
a) I can make something do what I want it to
b) I can make something do what it doesn't want to
(or I can make it choose the lesser of two evils)
I think the reason Greg is concerned about the discussion
so far is that the PCT term for control seems to follow meaning (a) for
inanimate objects, but suddenly jumps to meaning (b) for people.
I like the idea of having just one technical meaning associated with
the word control. I suggest: acting to keep a perceptual variable equal
to a reference variable. Your meaning (a) captures this technical
meaning when "I" is the controlling system, "something" is the perceptual
variable, "make" is the actions of the control system, and "do" is
keeping the perceptual variable equal to the reference variable. If the
perceptual variable is also controlled by another control system (relative
to a reference variable that is generally uncorrelated, over time, with
the reference variable of the original system) then we have the situation
described by your meaning (b). I didn't mean to use two different meanings
of "control". The fact of the matter, however, is that if you try to control
a perception that is controlled by another control system then you get
CONFLICT. Conflict occurs BECAUSE systems control (in my technical
sense). But when there is conflict, one or both of the systems involved
will not be able to control successfully! Maybe this is one the problems
in this whole discussion of controlling behavior. My technical definition
of control assumes that control is working (perception matches reference;
there is virtually no error). But it is possible to be organized as a
control system with respect to some variable and to be unable to control
that variable; this is what happens when there is conflict. When there is
conflict (which occurs when control systems try to control the same
perceptual variable relative to difference references) there are
control systems involved -- that's the only way that conflict can occur.
But one or both of the systems will not be controlling SUCCESSFULLY (if at
least one of the systems has sufficiently high gain -- so that it would be
able to control the variable successfully when there is no conflict). If one
system has MUCH higher gain and output range than the other then it will
successfully control the variable while the other system experiences massive
error (while producing maximum output). The latter system is definitely
NOT in control -- though it IS controlling FOR a particular perception.
So maybe we need to distinguish between the PROCESS of control and the
RESULT of control. My technical definition focuses on the result;
keeping perception matching reference. This result does not occur
when there is conflict; but the process of control DOES (and must) occur
when there is conflict. The process of control refers to a high gain neg-
ative feedback relationship between action and perception. Nominally, this
process results in perception matching a reference. When there is
conflict, the process may or may not produce this result -- depending
on the strength of the control system.
···
-----
Re: The Turing Test
Jeff Hunter (920924 - 2)]
The interrogator is clearly trying to find which variables
in the sonnet the witness is trying to maintain against disturbance.
Yes. What you described is part of "the test". I found another example
of "test-like" interrogation described in Hofstadter's "Metamagical
Themes". On p 501-502 there is a discussion of how you might tell if
the machine had "emotiuonal responses". It is suggested that, when asked
its reactions to certain pieces of music or literature, the
machine might "steer clear of emotional topics". Hofstadter's
discussion gives the impression that he has some notion that a
living system might be trying to control perceptions -- like "degree
to which I discuss emotional topics". But in fact, when you read the
entire exchange it is clear that Hofstadter has no idea that the
evasion he mentions might be a sign of a controlled variable;
indeed, on p. 502 he clearly misses this point completely. He has one
of his conversationalists say: "This would certainly make me suspicious.
Any consistent pattern of avoiding certain issues would raise serious
doubts in my mind as to whether I was dealing with a thinking being".
A PCTer would immediately see "avoiding" a symptom of control. And this
quote might suggest that Hofstadter might notice this too. But NO. It
turns out that the reason that the speaker would be suspicious of answers
that avoid emotional topics is because "I simply can't believe emotions
and though can be divorced." So these Turing tester's are looking at
the CONTENT of the machines answers, rather than what the answers
suggest about what the machine might be trying to accomplish. Thus, if,
when asked about it's reaction to a certain piece of music the machine
says "I don't know that piece" this is taken as evidence that the
machine has no emotions (because it is trying avoid a discussion of
emotional topics) -- the "avoiding" that MIGHT be happening is considered
important only in terms of it being an indication of the existence of
emotions (or lack thereof) on the part of the respondent.
So I am not getting "frisson" over your example of "the test" being
done as part of the Turing Test. What I am impressed by is how
people can come SO CLOSE to the notion of testing for controlled
variables -- and still miss the point completely. If, in your
sonnet example, there was an explicit hypothesis about the variable
being controlled and efforts to redefine this variable based on
constant questioning aimed at disturbing the hypothesized variable,
THEN I would be "frisson"-ing. But, as it sits, I would say that
the conventional Turing testers are as far from understanding the
notion of controlled variables as are your basic Skinnerian operant
conditioner (who is unwittingly doing a version of the test when s/he
tries to figure out what constitutes a reinforcer).
Best regards
Rick
**************************************************************
Richard S. Marken USMail: 10459 Holman Ave
The Aerospace Corporation Los Angeles, CA 90024
E-mail: marken@aero.org
(310) 336-6214 (day)
(310) 474-0313 (evening)