[From Bill Powers (920914.0900)]
Penni Sibun (920913.2000) --
That was a delightful essay on the Turing test for sex. I think you
have a great sense of humor to go with a sharp mind.
I would add another requirement to the Turing test, although it
doesn't alter the basic criterion of telling the machine from the real
person. The Turing test is basically a test for humanity at the level
of symbolic interaction. One of the kinds of interesting tests that an
AIer is not likely to think of applying is a test for control. A human
being, in a conversation, has (may have) a point he or she is trying
to make. In that case, when one person fails to understand, fails to
agree, shows that the point has been missed, and so on, the other
person will try different methods of correcting the perceived error --
use alternative arguments, paraphrases, illustrations, and so on. So I
would look for the thing at the other teletype to do this sort of
error-correcting communication. If it says "I am a mechanical
engineer," and I say "who made you?" I would look for it to correct my
mistaken understanding of "mechanical" instead of telling me who its
parents are.
With communication limited to a teletype, we can only test the other
entity at the higher levels of organization. At the strictly rule-
driven level, I doubt whether we could distinguish a program from a
person if the program were complex enough. But what would happen when
we try to see the principles or system concepts that the other
"person" is working from? Is it possible, for example, to write a
program that follows principles of humility, arrogance, caution, and
so on? I don't doubt that a program could make use of vocabularies
constructed from the speech of people judged to exhibit these
principles, but could a program in itself generate any new principle?
And if it could, would its principles be indistinguishable from the
ones human beings adopt and control for? Could a computer program
think up something like the Turing test? Could a computer program be
devised to carry out the Turing test with another entity?
Would you be surprised to learn that I am really an 80486/33 with 4
Megs of memory?
···
----------------------------------------------------------------------
-
Greg Williams (920914) --
It may not seem so, but I think we're approaching some sort of
understanding. You're forcing me, at least, to clarify my thinking.
I have still not managed to get across the point I'm making about
abilities that lie (autonomously) inside an organism. Actually I've
had the same failure in getting this point across to S-R types. The
question I'm trying to answer isn't "what effects on the behavior of
another person can you (empirically) demonstrate?" but "on what
internal properties and processes in the other person are you
depending in order to have those effects?"
This question arises when a radical behaviorist talks about the
effects of reinforcement. It can be shown experimentally that making
reinforcement contingent on behavior according to a schedule of
reinforcement results (after a little trial and error and rejecting
some potential test-organisms) in predictable patterns of behavior in
almost all test organisms. So the behaviorist says, "See? I have
complete control of this animal's behavior. It is obviously the
contingency of reinforcement that is causing it to behave as it does,
and that contingency is under my control."
The question I would want to ask in return is "But doesn't this
apparent control depend on properties and processes inside the
organism over which you have no control?" I would really want to ask,
"What if the organism doesn't want the reinforcer you offer?" But that
wouldn't go across big with a behaviorist.
Even under the S-R postulates, the experimenter is depending on many
autonomous functions inside the organism, starting with sensing and
recognizing the reinforcer. The internal connections from the stimulus
to the muscles then have to be established. Then the connections have
to be adjusted so that when the muscles act, they have an effect on
the environment (forget about disturbances) that produces a pattern
that either an apparatus or the experimenter recognizes as the one
that merits a reinforcement. If these things inside the organism
weren't present, and didn't somehow become organized in just the right
way, the contingency would have no effect at all, or at least it would
have an effect quite different from the expected one. And there is
NOTHING the experimenter can do to provide the organism with any of
these internal elements, should any of them prove to be missing. How
do you control an animal that has no sensory organs? The only way is
to pick it up and run it across the floor like a toy car. Of course
such an organism couldn't live, but this is a thought experiment.
In a hierarchy of control that accretes through reorganization, this
situation is compounded in complexity. The external agent can produce
expected behaviors from this hierarchy, but only if they don't violate
the conditions that the reorganizing system must, by its nature,
maintain.
On the surface, the external agent is dealing only with a single
control system, the one whose actions the agent is trying to control.
But as soon as this control has any effect that significantly disturbs
the control hierarchy at the same or a higher level, either direct
resistance will develop or the reference signals in the original
system will begin to shift. So in order to retain control in general,
the external agent has to give up control of that action and pick a
higher-level aspect of behavior to control.
It's inevitable that the general desire to control another must lead
to applying disturbances at higher and higher levels, which of course
becomes more and more difficult. And in the background, at any level
of interaction, the reorganizing system sits watching for any
consequences of these interactions, known to the hierarchy or not,
that have adverse effects on the fundamental variables that signify
the viability of the organism. As soon as any such violation takes
place, the reorganizing system will come into action. Ideally it will
act to prevent exactly the manipulation that is causing the problem,
but more probably it will begin trying new organizations at random.
Either way, the external agent must start over because now its map is
obsolete.
One way to think of the hierarchy of control is as a series of layers
of defense erected by the reorganizing system. When a systematic
control system is acquired that has, as a side-effect, the ability to
prevent changes in the environment that cause intrinsic error, the
organism has created a defense, over a limited range, against external
events that previously caused intrinsic error. Unless some marked
change in the properties of the environment occurs, those events are
now controlled by the behavioral hierarchy and prevented from causing
intrinsic error again. As the hierarchy builds up, the number of
events of a kind still capable of creating intrinsic error must
decrease; the effects of the events must become more subtle and
indirect. Control behavior becomes more and more sophisticated as the
levels are added, to the point where in an adult there are few things
that can upset the milieu interieur. Either the adult organism
directly cancels such effects or it knows how to avoid exposing itself
to them.
I define the reorganizing system in a particular way. Its input
function is genetically designed to monitor some set of critical
variables (whose nature is still undetermined). For each such variable
there is a genetically-given reference level. The reorganizing system
contains a comparator that detects the sum of the absolute values of
all these errors, resulting in an intrinsic error signal. The
intrinsic error signal drives the processes of change that
unsystematically alter the organization of the hierarchy. Obviously
this is just the outer envelope of a system that may have much more
detail in it that I can provide. The effects of reorganization may
extend to biochemical systems, too, such as the immune system. I don't
know.
The main point is that if there is a reorganizing system of the kind I
propose, then it is autonomous with respect to the events of a single
lifetime. Its criteria for acting are given in its set of intrinsic
reference levels, which are fixed for life and independent of
particular experiences and particular organizations of the behavioral
hierarchy.
It is this reorganizing system that sets the limits on what an
external agent can get the organism to do or not do. The entire
hierarchy is constructed around the unchangeable requirement that
intrinsic error shall be held near zero. What constitutes intrinsic
error is defined by DNA, so it represents the judgment of the entire
species, acquired over geological time, as to what is good and bad for
any individual of this species. This judgment is unaffected by the
events of a single lifetime.
I have postulated, as you know, that awareness is somehow connected
with the reorganizing system. Apparently, awareness can adopt the
point of view of any level of perception in the hierarchy, so that the
conscious world consists of the perceptual signals in that level and
is shaped by the perceptual interpretations that give rise to those
signals. This is the basis of the Method of Levels, the HPCT design
for psychotherapy.
Apparently, also, awareness can direct the locus of reorganization --
that is, reorganization is concentrated on the systems with which
awareness is most closely identified at a given time, at least in the
neural-behavioral hierarchy.
This is my theory of "free will," which defines both "free" and
(conscious) "will." I do not know the nature of awareness, nor does
anyone else. But anyone who or anything that lacks it clearly gives
the impression of "nobody home." This is a working hypothesis, of
course, nothing more. But it does seem to have noticeable explanatory
power.
The point in bringing this up is that nobody can have any freer a will
than this -- the ability to change point of view, and to direct
reorganization to a place in one's own system, with the same apparent
randomness that we attribute to reorganization itself.
If this is true, then what we call manipulation of one person by
another is simply the interaction of two systems, each attempting to
reorganize itself to suit its own internal definition of what is good
or bad for it. We do NOT have the situation where one all-knowing
omnipotent system of unlimited intellect is manipulating a dimly
conscious and defenseless system of some inferior kind -- that is the
situation only in the relation between a human being and an
emphatically lower order of life. It is not the situation that holds
between normal adult human beings. The manipulator's "will to
manipulate" is not different in kind or effect from the manipulee's
"will to defend."
The term "manipulation" is unfortunate, because it implies power and
will on the side of the manipulator, and weakness and automaticity on
the part of the manipulated: exploiter and exploited, controller and
controlled, agent and effect. This implies that the manipulator has
some abilities or powers that are lacking in the manipulee, an
awareness that is denied to the manipulee, a natural advantage over
the manipulee. This is simply not the case in the HPCT concept of
behavior. All human beings have the same basic abilities, save for
those that are crippled in some striking way. Whatever conceptions of
social life we may have developed, in truth we all have the same
equipment, which is designed to allow us to control what happens to
ourselves -- and especially what happens to our ability to continue
living.
Manipulation is just a one-sided way of looking at interaction.
----------------------------------------------------------------------
Eric Harnden (920914.0900) --
Thanks much for trying out that two-system control thing. I am
relieved that it turned out as I predicted. I think this settles the
question as to whether Stella can be used for significant PCT
modeling.
Can you tell me a bit about Stella's ability to create time plots of
the variables in a simulation? Is there any way for Stella to accept
real-time input from, say, a mouse or a joystick? Would be be possible
to provide Stella with results from an experimental tracking run if
the data were presented as a list of successive values? Does Stella
have any provision for user-defined functions (which could be used to
do input-output even if Stella doesn't have that ability)?
Finally, how Big a Mac is required to run Stella? If one of the Little
Macs could run it, I would consider getting a used Mac just to be able
to communicate through Stella with Mac users.
I hope you're willing to be the Stella guru for a lot of PCTers who
have Macs. You would be helping a whole lot of people beside yourself
to get a real grip on PCT.
I should point out that the Forrester/System-dynamics concepts of
"levels" and "flows" are somewhat narrow concepts. Some rather
unnatural translations would have to be made, for instance, to
describe how a position-control system works. But that's just a matter
of words. I think Stella should be able to do the job.
----------------------------------------------------------------------
--
Best to all,
Bill P.