[From Bill Powers (931118.0630 MST)]
Bruce Nevin (931117.1545) --
The problem attributed metaphorically to the fish is very
similar to the impossibility of recognizing a given perception
as such (or perceptions of a given level) unless you first go
up a level. Or, conversely, perceiving the world through nth-
level glasses: if you're a fish, all the world is watery.
The way I've always heard the aphorism, it was intended to show
the impossibility of perceiving anything that pervades one's
environment, like language customs. No "unlesses." Your second
suggestion means almost the opposite: that you see the
environment as containing the uibiquitous thing rather than
seeing it as an interpretation of yours. In any case, I agree
with your two interpretations, but I think you've modified the
aphorism as "customarily" used.
Since you understand the intended point -- and I believe you
when you say that you do -- perhaps you can suggest a metaphor
(and aphorism) free of the defect that troubles you.
The one I've always preferred, when an an aphoristic mood, is
that of standing before a window and watching a baseball game
being contested outside the house. If the game is engrossing, you
may forget that you're looking through a window at it.
Another obvious possibility is watching a show on television. One
of the greatest Star Trek: New Gen episodes involved the holodeck
character Moriarty, who accidently achieved self-awareness,
setting up a simulated holodeck that fooled Worf, an engineer
whose name I forget, and Data into believing that Moriarty was
able to leave the holodeck and survive. In that simulated
Enterprise, Data and a simulated Geordi worked out a way to
transport Moriarty's lady-love off the holodeck and into reality,
but when they tried it the transporter log showed that nothing
had happened. The crew on the real holodeck then created a
simulation within the simulation to convince Moriarty that his
lady friend had actually been successfully transported. All this
was revealed only near the end, when Data said "computer, end
program," and they were back in the original simulation, and then
said "computer, end program" again, restoring them to the real
Enterprise. At the end (why give it all away), the last character
to appear looked uncomfortable, and muttered "computer, end
program." When all remained as it was, the character looked
satisfied and walked off the scene. The program (i.e., that
episode) then ended. Wow.
All in all, the window metaphor is shorter.
Bill Cunningham (931117.1700) --
It has struck me that the WHY-HOW mismatch is at the heart of
much of PCT's problems with the outside world. PCT is really a
HOW theory. Sure it opposes other HOW theories, but it must
really irritate the hell out of the WHY askers. That may be a
simplistic assessment, but that's the way it comes across.
And that's the source of PCT's weaknesses.
Only in some people's opinions. Those are the people who like to
see general principles behind specific processes; they aren't
satisfied with knowing HOW a system works, they want to put that
HOW in the context of a higher-level perception of principles.
When they succeed, it seems to them that they have found the WHY
of the system: why it HAS to be organized that way.
This gives the status of incontestible _a priori_ truth to the
principles, and makes it seem that the principles somehow have a
guiding influence on the lower-level processes, as if they were
purposive and forbad any other lower-level processes from
existing. They are purposive, but the purposes are those of the
person who wants those principles to apply. What is forgotten in
all this is that the principles were derived from observations of
the processes; only those principles that are consistent with the
observed processes are perceived. So the principles are dependent
on perceptions of processes, just as any higher-level perception
is a function of lower-level perceptions. If the processes
happened to work differently, different principles would be
perceived: a different WHY would be seen.
Hierarchical relationships between perceptions are confusing
because they are circular. The higher is derived from the lower,
yet once a higher perception is selected as a target, it serves
to guide the way the lower perceptions appear. Each seems to
explain the other. I see a ball BECAUSE there are certain
sensations of shading, curvature, and edges; yet it is BECAUSE I
want to see a ball that those shadings, curvatures, and edges
come into being in my experience.
A principle like conservation of energy (to pick a neutral
example) is not the explanation of why a pendulum works as it
does. It is a consequence of the way the pendulum works; it is a
more generalized way of perceiving the workings of a pendulum. It
is, of course, a very useful principle. But it has no causal
properties beyond those given to it by a purposive system.
Martin Taylor (931117.1745) --
Why don't those who use this old adage ask scoffingly if
people perceive the existence of air?
<I thought of doing just that. Don't we perceive cold, wind,
smells, and the like, rather than "air," just as the fish might
be expected to perceive the proximity of other fish,
turbulence, shearing currents and such stuff?
This is a matter of the level of perception. How do you perceive
a keyboard in front of you, if not through perceptions of colors,
edges, shadings, movements, and so forth? The higher perception
is a function of the lower. Air is a perception derived from a
collection of experiences at lower levels. If we were fish, water
would be a perception derived from collection of experiences at
I'll probably shut up about it soon, unless something does
come up. How about Bernard Gibbs's J-curves as a possibility?
If EVERY ONE of Gibb's subjects showed the reversal for the 2nd
and 4th (or whatever) lights, I might try to model this
phenomenon. If not, how can I construct a model to be the right
explanation when it would be flatly incorrect a good part of the
RE: idea of recalibration
In its favour, however, is a point made last year by Jeff
Hunter and a couple of days ago by Tom Bourbon, that one can
estimate a segregated effect by applying reference steps that
can be correlated with effects on the CEV. No matter what the
disturbance, the output->perception relational function can be
estimated by this procedure in an analytic way, if the
environment function is stable enough.
What do you mean, "one can estimate?" It's the behaving system
that must do the estimation, and to do it it must perceive
certain things and apply certain processes to the perceptions.
All those things and processes have to be put into the model and
it must be shown that they produce the effect we observe, every
time. I'm not doubting that this can be done (after all, Tom
reported its being done). But to model it is considerably beyond
our current capabilities. There are lots of things WE can do when
looking at a DIFFERENT system. The question is, how can a system
do these things for itself?
Recalibration has its own problems: what's the standard?
... there is no standard; what has to be sensed is a
relationship between output steps and perception changes.
Once that relationship has been sensed, what is the _right_
relationship? There must be a standard; otherwise the perceived
relationships is just whatever it is, and there is no reason to
change it toward any other relationship.
If it turns out that the accepted reorganization approach (an
e-coli convergence to the appropriate function) does not work,
then there is a problem. If it does work, a continuous
updating of the model would fall out naturally.
Not out of reorganization alone. There must be something sensed,
and there must be a defined reference condition, in order for the
e. coli method to work.
We are still miscommunicating about getting to bed in the dark.
I believe that the way you say it is done is essentially
correct. But I have always thought that "control of
perception" was taken to imply that the perception was of a CEV
in the real world. If one cannot determine the state of that
CEV through the senses, one cannot control it.
I have thought this was your conception since our early
discussions of the "mirror world." The CEV exists in the real
world; if we perceive it correctly (mirror it correctly in the
internal world), we can control it. Otherwise the CEV will remain
uncontrolled, or will be incorrectly controlled in some respect.
Other arguments have seemed more important, but now I guess we
must have it out.
The organism can know of CEVs only as perceptions. That goes for
you, too. You have a concept of a CEV, but that concept is in
your head. It is part of your perceptions. You have no way of
knowing what actually corresponds to your perceptions. Nobody
does. All we have are the perceptions. The rest is hypothesis. It
is the external world that is the mirror of the perceived world.
The imagined mirror.
All that a control system needs to know about the environment is
that if it acts in certain ways, perceptions change in certain
ways. It does not need to know how those actions work to affect
the perceptions that change; the mechanisms linking action to
perception exist outside the organism, outside of you and me,
where we have no direct access to them. Perceptions can be
controlled without any knowledge of their external counterparts,
A CEV is the inverse perceptual function of a perceptual signal.
Normally, input functions have multiple inverses, so many
different conditions of the environment suffice to create a given
magnitude of a given perceptual signal. There is, therefore,
nothing external corresponding to "the" CEV. "The" CEV is a
ghost, a projection of internal perceptions onto an external
universe which is not experienced.
However an input function is organized, it creates a CEV: it
creates a perceptual signal that is seen as an entity in the
environment. Given that perceptual signal and one state of it
selected as a reference, a control system can be constructed that
will control the perceptual signal. The construction entails
choosing the signs of various output effects of the error signal
so that each loop is a negative feedback loop (or so that most of
them are). The choosing, we have shown in simple cases, can
result from a random reorganizing effect based on the error
signal itself, with the implicit goal of making the error signal
zero. No hypothesis about the external world is required for this
to work. If a particular output connection makes the error worse,
remove that connection. That is all that is required.
The resulting control does not imply that anything in the
environment is being brought to just one state. All that is
required is that the outputs bring the environment to one of the
states that is equivalent in terms of creating the same
perceptual signal. The "bellringers" analogy is worth repeating
"Of course if there are coordinates for which there are no ropes,
but which enter into the expression for the energy, then, if the
motion of these coordinates is periodic, there will be "adynamic
vibrations" communicated to the ropes, and by these the men below
will know that there is something peculiar going on above them.
But if they pull the ropes in proper time, they can either quiet
these adynamic vibrations or strengthen them, so that in this
case these coordinates cannot be ignored."
"If the machinery above has more degrees of freedom than there
are ropes, the coordinates which express those degrees of freedom
must be ignored. There is no help for it."
(paragraph order reversed)
Maxwell, J. C.: Thompson and Tait's Natural Philosophy. Nature,
Vol. XX, 1867. From Niven; The scientific papers of James Clerk
Maxwell, pp. 776-785. New York: Dover (1965).
If you were blind ...
Yes. Your control hierarchy would have to skip the on-line
control of the relational perception you<->bed.
No. Your control hierarchy would control that relationship in a
space defined kinesthetically instead of visually. A blind person
can put a cane down and then reach for it and pick it up. The
cane has a perceived location, and so does the reaching hand. The
coordinates of this space are simply not x, y, and radius as
Kinesthetic control is partially sample-and-hold control. Instead
of a continuous update of the visual model of the world, you get
only an intermittent update of a kinesthetic model. The bed is
where you last felt it to be until you can feel it again. The
cane, tapping inquistively around, is updating the model, moving
the chair or the sofa or the lamp a little, altering the position
of the dot that is you in this model. To control relationships
you alter the part of the relationship under your direct control.
All you're saying is that continuous control is more effective
than sampled control; that continuous perception works better
than intermittent perception in a control process. I would agree,
I continue to believe that the output provided to the world by
the relational perception ECS is open-loop when the lights go
out, whether or not the ECS is, by that output, affecting a
controlled imagined perception.
In the going-to-bed case, one perception in the relationship is
still under continuous control: the perception of where you are
on the map. Judging from joint angles you can alter your
perceived position according to number of strides and stride
length, and your perceived direction of movement by adding up
twists at the ankles and hips as you walk. It is not necessary
for BOTH inputs to a relational perceptual function to be real in
order for closed-loop control of a relationship perception to
exist. It is only necessary that they both be real if you want
the controlled variable to be veridical. The controlled variable
is ALWAYS a perception.
If you would stop thinking of relationships as exclusively visual
in nature, this problem would go away.
Bob Clark (931117.1700) --
Nice summary of points that have been made here and there on the
In relation to ballistic types of control, one factor seldom
brought up is that the ball or projectile must be given a
specific vector velocity in order to reproduce a previous
trajectory accurately. This requires lower-level control systems
that control velocity very accurately up to the point of release.
Open-loop activation of muscles according to a stored program
would not be anywhere near accurate enough. Gary Cziko has a neat
demonstration of this lower-level control, in which he can apply
a substantial elastic restraint unexpectedly while a person is
tossing a ball underhanded toward a target. There is only a
slight disturbance of the trajectory, and the person throwing the
ball can feel the extra effort that suddenly appears, opposing
the effect of the restraint.
Chuck Tucker (direct communication)
Please post your demo of feedforward on the net. It's worth a
thousand abstract arguments.
Best to all,