[From Bill Powers (930617.0730 MDT)]
Martin Taylor (930616.1716) --
One CANNOT make a consistent view of information at a place in
a control system by taking a viewpoint outside the control
system. It HAS to be applied to the control system FROM ITS
OWN POINT OF VIEW.
I know that you've said this, Martin, but the moment you start
talking about how a control system perceives, you begin giving
that system a very elaborate set of capabilities, or so it seems
to me. As you explain how the probabilities are calculated by a
perceptual function, I feel that you're describing operations
that are pretty unlikely to be going on in the perceptual
function itself. Not only do the probability calculations have to
be applied TO the proposed control system from its own point of
view, they have to be applied BY the proposed control system,
without the aid of a human intepreter. The model has to work by
itself while you're out having lunch. If your computer program
contains any calculations of probabilities or information, that
computer program is part of the model, and amounts to a claim
that the control system itself is performing those same
computations. In my models the computations in the program are
intended to represent processes in the perceptual function
itself, not computations ABOUT the perceptual function. I don't
use any auxiliary computations or meta-computations. Everything
the program does (that is part of the model of the behaving
system itself rather than its environment or the user interface)
represents something the system itself does, or is proposed to
do.
We need to talk about this in detail until it's settled.
Otherwise we'll be stuck here forever.
···
-----------------------------------
Insertion of ECSs:
As a reason for not letting outputs jump levels, you say
On the other hand, output signals cannot usefully jump levels,
even though they may not be prohibited from doing so. This
restriction occurs because a level-jumping output would be
affecting the same variables (probably the reference signals of
low-level ECSs) that would be affected by intervening ECSs, and
control by means of those intervening ECSs would be more
effective than affecting the low-level references directly.
The reason I have given is a bit more forceful. If a higher level
system's output skips a level instead of working through the
reference signal of the skipped level, it will be adding a signal
to the same place where the output of the skipped level enters.
This will simply create a disturbance in the skipped control
systems, and the skipped level will, as usual, oppose it. The
skipped level will oppose the output from the higher system just
as effectively as it would oppose any disturbance arising from
outside the system. So the higher system would find it impossible
to affect its own perceptions: its output would be cancelled
without accomplishing anything.
I'll be interested to see how your ECS-inserting model works. As
you're intending to make an actual working model, I won't throw
any roadblocks in the way.
---------------------------------------------------
RE: following citations from Nevin:
1. Each cell in an ECS can control itself using intracellular
control mechanisms such as ion exchange across membranes.2. A cell cannot control another cell with these mechanisms.
3. Yet cooperating cells can together constitute a control
mechanism of a higher order.4. This higher order of control is invisible to and does not
itself affect the constituent cells.5. However, chronic error in the higher-order control system
is perceived by the cells in the form of environmental factors
that are distressing to the cells.
This relates to what I was saying to Hans Blom yesterday. The
implication of what Bruce says is that the higher order of
control is implemented using the same physical cells that
constitute the set of lower order control systems. When I first
read this, I happily incorporated it into my own model by
assuming it meant that the higher level control systems used the
existing lower level ones as an output function, which is how my
hierarchical model works. Now I realize that this is not what
Bruce was saying (from the context of your citation). Bruce is
treating the higher level as if it were simply an emergent
property of the existing set of control systems, "invisible"
because the very same cells are involved and only their relations
to each other express the new level.
This is a concept of hierarchy quite different from mine. In my
hierarchical concept, a new level is composed of physically
distinct cells, specialized to perform new functions and capable
of sending outputs to the old set of cells (through their
membranes) that adjust their internal reference signals.
Thus: the liver cells control the sensed concentration of
glycogen in their immediate vicinity. As there are many cells,
they do interact with one another, and there are probably
phenomena of mass action and interaction involved. But liver
cells contain no mechanisms for adjusting the desired level of
glycogen. That is done by physically different cells that do not
control glycogen, but other variables that depend on the level of
glycogen. Those cells are far from the liver, in the hypothalamus
and perhaps the pituitary. The outputs of the cells in the
hypothalamus or pituitary reach liver cells and raise or lower
their reference levels for sensed glycogen.
The liver cells are not "distressed" by experiencing different
levels of glycogen unless those different levels occur without a
change in the reference signal. The liver cells are just as happy
to maintain a higher level as a lower level, if the received
reference signals so specify. It is not the business of liver
cells to judge where the goal for glycogen concentration should
be set.
The addition of a higher level of control does not affect
existing control organizations at all. It simply makes their
reference signals variable.
-----------------------
If the world's responses to outputs were linear, a one-level
system of orthogonal ECSs would be perfectly adequate.
Adequate for what? For controlling the rate at which a variable
changes back and forth between two values? For adjusting the
placement of flowers in a bowl so they have a pleasing look? For
raising the garage door before rather than after you back the car
out?
You've said before that you think a perceptron using weighted
sums of its inputs is adequate for modeling all of perception. I
don't think you're considering all the kinds of perceptions there
are, and how one kind must be controlled in order to control
others because of the way the external world is organized. I know
that there's a degrees-of-freedom argument in the background, but
there's also a practicality argument: it's just not practical to
accomplish all the different higher-level perceptions by means of
a single level of transformation, even though mathematically the
difference isn't visible. Nonlinearity isn't the only factor. And
the mathematics is limited.
---------------------------------------------------------------
Hans Blom (930617) --
I appreciate your clarification; it introduces a nontrivial
consideration.
The CROWD program contains ample illustrations of the case you
cite.
Suppose we have set up the program so that person A seeks a
specific distance from person B, person B from C, and C from D,
while D simply seeks the position of a distant target location.
The result is that a little procession forms, and when the leader
finally arrives at the destination the followers stop in a line
trailing behind the leader. This single-file procession is a
vivid example of an emergent social phenomenon that is not part
of the intentions of any participant, just as much so as the arcs
and rings.
It is easy to set up this situation so that about midway during
the march across the screen, this processsion encounters a loose
group of people traveling on a path crossing at an angle through
the procession. A series of disturbances occurs, with the first
line of marchers dodging about and deviating far from the
original simple lineal order. After the encounter, the single-
file march reforms and continues to the goal.
So this does indeed look like a control system that resists
disturbances of a social variable, restoring a social
relationship to a specific form after a disturbance. But does it
pass the Test for the Controlled Variable?
It's easy to forget that resistance to disturbance is only one
part of the Test. There are three other parts. One essential part
is to make sure that the apparent resistance of the controlled
variable is not due entirely to a natural tendency of the
controlled variable to resist disturbance. Another part is to
interrupt the ability of the supposed control system to perceive
the controlled variable, and show that the resistance to
disturbance becomes unsystematic -- that control is lost. And the
final part is to remove the control system entirely (when
possible) and show what the effect of the disturbance on the
controlled variable is when the system is not opposing it (there
might be some other system acting that we haven't noticed). The
effect should be much larger than when the proposed control
system is present.
The "control" of the lineal progression passes the first part of
the test, but fails all the others. We can't even find a control
system that is sensing the state of the proposed controlled
variable, the linearity of the formation, so clearly we can't
interrupt its perception of that variable. We can't find the
variable output by which this control system affects the shape of
the formation, so we can't do that part of the test, either. In
fact, all we can find is the set of control systems which are
each controlling a variable that has nothing to do with the
lineal formation. We must conclude that the lineal formation is
simply a property of the collection of control systems with their
reference signals set as they are, and is not the controlled
variable of any superordinate control system. The social behavior
we observe is completely explainable through knowing the
properties of the individual control systems. No postulate of
social control is required.
There's a much simpler example that makes the same point. Suppose
that we find a marble in the bottom of a bowl. By applying a
known force to the marble, we can made it deviate from the
resting position. As the deviation occurs, a counterforce
develops; the marble comes to rest partway up the side of the
bowl where the restoring force is equal to the applied force. Is
this an example of a control system?
Again, strictly on the basis of opposition to disturbance, yes.
But again, none of the other parts of the Test will be passed. We
can find nothing sensing the position of the marble. We can find
no system that varies its output force in opposition to our
disturbance of the marble; the cause of the restoring force,
gravity, doesn't vary at all. In fact, the restoring force is
produced directly by a physical effect of the disturbance on the
marble. No control system is required to explain the
observations.
We would require a different explanation if we found that the
marble sought a position partway up the side of the bowl, and
resisted forces tending to move it from that position. We would
have to find the force that is holding it up against gravity, and
that force might be coming from a control system. Also, it might
not: there might be a magnet off to one side, and the marble
might be made of magnetic material. One simply has to explore all
the possibilities. Control is not the only possible explanation.
We accept the control postulate only when all other simpler
explanations have been ruled out.
This is why the Test has four parts, not just one. If all parts
of the Test are passed, we can be sure that we have found a
control system that works like the model used in PCT.
--------------------------------------------------------------
Best to All,
Bill P.