Information in the organism about the disturbance

[From Bill Powers (940421.1815 MDT)]

Wayne Hershberger (940421.1153) --

     I'm not sure what you mean by "a control system which
     transmits information." What does a control system
     transmit information from, and to?

From environment (i.e., disturbance) to organism/mechanism
(i.e.,output). The more closely the output mirrors the
disturbance the more fully the organism/mechanism is informed
about the state of the environment.

"The organism" perhaps (only perhaps), but not the specific control
system doing the controlling. Information contained in the output
tells the organism nothing unless some effect of the output is
sensed. As you said in _Volitional Action_, it is possible for some
system other than the controlling system in question to sense the
output effort, and from that deduce that there must be some
corresponding disturbance at work. That must be done by some system
other than the control system in question. The control system in
question is controlling a visual image, not deducing anything.

Even that method is limited, however. From examining the output of
another control system, the most that can be said is that there is a
disturbance having, or tending to have, an effect on the controlled
variable of that system. The magnitude of the effect can't be
estimated unless there is information about the physical link
between the control system's output and the controlled variable. The
nature of that physical feedback link must be known in order to
estimate the magnitude of the effect of a given output on the
controlled variable, which is needed in order to estimate the
magnitude of the mirrored effect of the (unknown) disturbing
variable or variables.

All this can be illustrated using our favorite example, the driver
of a car. A driver keeps the visual field in a certain condition by
applying output forces to the steering wheel. That visual-motor
control system senses visual variables, controlling them by varying
motor outputs. But another system in the driver can, at the same
time, sense the applied force through kinesthetic sensors, or see a
different aspect of the visual field, the angle at which the
steering wheel is turned. From the sensed effort or the angle of the
wheel, the driver can know whether the effects of environmental
disturbances are relatively large or relatively small, or whether
they are averaging some non-zero amount in one direction.

However, to deduce the actual magnitude of environmental effects,
the driver would have to be able to translate the sensed effort or
the angle of the wheel into an equivalent sideward force on the car
due to the angling of the front wheels (part of the feedback
function). This would require knowlege of the mechanical advantage
built into the steering linkage, and the physics of a rolling tire
being maintained at an angle to the direction of travel. Such
information is not likely to be readily available. So while the
driver can know that there is a disturbance acting, and perhaps
sense its momentary direction, there is no way (even with perfect
control going on) to extract from sensory information the nature of
the mechanical linkages and frictional effects where the rubber
meets the road.

Even if that information were magically made available, the driver
would still know only the net sideward force on the car being
exerted by the angled front tires, as deduced from sensed effort and
wheel angle and the equations transforming them into the corrective
force being generated by the tires. There would be no indication of
the physical processes in the environment that were adding up to a
corresponding amount of sideward force in the other direction on the
car. The same steering effort would be required to counteract the
effects of a crosswind blowing from 30 degrees to the right at 20
mph and a crosswind blowing from 150 degrees to the right at some
lesser velocity (not proportional to the sine, because aerodynamic
forces are nonlinear with velocity). Likewise, if multiple
disturbing variables were at work, sensing the applied effort and
translating it into tire force could not distinguish among the
multiple causes of lateral forces: flat tires, tilts in the road,
wind, etc. Each of them individually and all of them together would
simply result in single a net sideward force on the car.

I have done my best to find a way in which your answer could be seen
as partially right. But we were talking about a single control
system, and in that context your answer is entirely wrong.

The output of a control system consists of physical effects that
propagate outward into the environment, outside the control system.
It is irrelevant to the control system whether those outputs carry
information or not: even if they do, the information is travelling
away from the control system, and can't be used by the control
system.

The only way information can get INTO the control system is via the
perceptual signal. And here we are back to the apparent
misconception against which Rick argued so vehemently that he ended
up receiving enough flak to drive him off the net for a while, and I
don't blame him.

The better the control is, the more closely the controlled variable
and perceptual signal are maintained in a match with the reference
signal, and the less effect disturbances can have on the perception.

The perceptual signal contains information about what we see as the
controlled variable. On that I am sure we can agree. So what we
really have to ask is whether the _controlled variable_ contains
information about the environment, about physical events in the
environment capable of causing changes in the controlled variable.
If the controlled variable contains such information about the
environment, then the perceptual signal will contain most of that
information, because in general the perceptual signal can be assumed
to be an accurate representation of the controlled variable (I'm not
quibbling about epistemology here). On the other hand, if the
controlled variable does NOT contain information about the causes of
its own changes, then the perceptual signal can't contain any
information of that kind either.

So what information does the controlled variable contain about the
rest of the environment? If control were perfect -- that is, if
actions changed so as to exactly mirror the effects of external
variables on the controlled variable -- then clearly the controlled
variable could not contain any information about the rest of the
environment, or anything else, because it would never vary no matter
what the rest of the environment did.

Control is never perfect, so the controlled variable will in fact
change slightly as external effects impinge upon it. But again, all
that can be known from examining the controlled variable is the
_effect_ of an external variable. To deduce the state of the
external variable itself, we would have to know _a priori_ first how
many such variables were simultaneously contributing to the state of
the controlled variable, and second the form of the linkage through
which each such variable is acting. Even then we would be looking at
an outcome of many causes, and deducing a unique inverse would
normally be impossible.

Neither the multiplicity of external causes of disturbances nor the
forms of any of the linkages to the controlled variable can be
discerned simply by examining the controlled variable. Since the
perceptual signal directly represents the controlled variable, it
follows that the perceptual signal cannot contain any information
about the number of external disturbing variables, their individual
states, or their individual modes of action on the controlled
variable.

We are left with the simple observation that the perceptual signal
contains information about the state of the controlled variable and
nothing else. While the controlled variable can be perturbed by
external influences, the control system itself acts to keep those
perturbations small -- and even if it didn't, the perturbations
themselves contain no information from which we could identify the
causes of the perturbations, either the physical variables
responsible or the linkages through which they acted.

In focussing on minor perturbations of the controlled variable, we
can forget an overwhelmingly important fact: the state of the
controlled variable reflects with great precision the state of the
system's reference signal, and so, therefore, does the state of the
perceptual signal. As the reference signal changes, the controlled
variable is forced by feedback effects to change in almost exactly
the same manner. The changes due to the reference signal are in
general tens of times larger than any changes due to normal
environmental influences; they can be hundreds of times larger. Or
to put that more properly, changes in a controlled variable due to
disturbances are tens to hundreds of times smaller than those due to
changes in the reference signal. We can easily forget this, because
in our experiments we try to encourage the participant to maintain a
constant reference signal. I have occasionally demonstrated what
happens when the subject is asked to establish a different reference
level for the controlled variable, but these demonstrations seem to
slip past without making any impression.

When we remember the role of the reference signal, we see
immediately that the controlled variable is taken over by an
organism and made almost independent of external forces that
normally would be capable of changing it. If a higher system varies
the reference signal, the controlled variable simply follows it,
with normal external influences having only the most minor effects.
The behavior of the controlled variable becomes the behavior that
the organism wants to experience, and no longer reflects the effects
of the environment (except to a very small degree, of little
consequence in comparison to the effects of the reference signal).

The only problem that a control system design has to handle is how,
given a difference between a controlled variable and the reference
state implied by a reference signal, the system can act to bring
that difference to zero. Clearly, this problem can be solved without
reference to the environment (other than the part that constitutes
the controlled variable and the feedback link connecting the
system's output to it). If the reference signal changes, the
resulting error must be converted into an output that forces the
controlled variable toward the new reference state. When the system
has been designed so that the controlled variable follows the
fastest changes in the reference signal with as little error as
possible, a limit set by the physical properties of the components,
control is as good as it can get. And all this can be achieved
without ever considering even the existence of possible
disturbances.

As it happens, the same design that makes the control system able to
keep the controlled variable close to a varying reference state
quite automatically, and at no added expense, makes the controlled
variable resistant to externally-caused disturbances. The system
does not distinguish between errors caused by changes in the
reference signal and changes caused by external influences. In
either case, it simply acts directly on the controlled variable in a
direction that makes the error smaller. Don't be misled by "optimal
control" advocates; the design differences they speak about for
optimizing control for disturbances vs reference signals are minor
design differences and are not applicable in ordinary situations.

So a control system can be successfully designed without ever taking
into account any information about external disturbances in the
environment. While it makes common sense to assume that in order to
control against disturbances, a control system must be getting some
kind of information about the causes of disturbances, this bit of
common sense is simply wrong. The only information the control
system needs is its own perception of the state of the controlled
variable.

···

---------------------------------------------------------------
Best,

Bill P.

From Tom Bourbon [940422.0903]

[From Bill Powers (940421.1815 MDT)]

Wayne Hershberger (940421.1153) --

Bill:

     I'm not sure what you mean by "a control system which
     transmits information." What does a control system
     transmit information from, and to?

Wayne:

From environment (i.e., disturbance) to organism/mechanism
(i.e.,output). The more closely the output mirrors the
disturbance the more fully the organism/mechanism is informed
about the state of the environment.

Bill:

"The organism" perhaps (only perhaps), but not the specific control
system doing the controlling. Information contained in the output
tells the organism nothing unless some effect of the output is
sensed. As you said in _Volitional Action_, it is possible for some
system other than the controlling system in question to sense the
output effort, and from that deduce that there must be some
corresponding disturbance at work. That must be done by some system
other than the control system in question. The control system in
question is controlling a visual image, not deducing anything.

. . .
Bill continues:

We are left with the simple observation that the perceptual signal
contains information about the state of the controlled variable and
nothing else. While the controlled variable can be perturbed by
external influences, the control system itself acts to keep those
perturbations small -- and even if it didn't, the perturbations
themselves contain no information from which we could identify the
causes of the perturbations, either the physical variables
responsible or the linkages through which they acted.

In focussing on minor perturbations of the controlled variable, we
can forget an overwhelmingly important fact: the state of the
controlled variable reflects with great precision the state of the
system's reference signal, and so, therefore, does the state of the
perceptual signal. As the reference signal changes, the controlled
variable is forced by feedback effects to change in almost exactly
the same manner. The changes due to the reference signal are in
general tens of times larger than any changes due to normal
environmental influences; they can be hundreds of times larger. Or
to put that more properly, changes in a controlled variable due to
disturbances are tens to hundreds of times smaller than those due to
changes in the reference signal. We can easily forget this, because
in our experiments we try to encourage the participant to maintain a
constant reference signal. I have occasionally demonstrated what
happens when the subject is asked to establish a different reference
level for the controlled variable, but these demonstrations seem to
slip past without making any impression.

When we remember the role of the reference signal, we see
immediately that the controlled variable is taken over by an
organism and made almost independent of external forces that
normally would be capable of changing it. If a higher system varies
the reference signal, the controlled variable simply follows it,
with normal external influences having only the most minor effects.
The behavior of the controlled variable becomes the behavior that
the organism wants to experience, and no longer reflects the effects
of the environment (except to a very small degree, of little
consequence in comparison to the effects of the reference signal).

Tom, now:
Excellent points, Bill, concerning the fact that in even modestly
successful control, fluctuations of the controlled perception (and of
associated environmental controlled variables that might be related to the)
presence of disturbances are usually *minuscule* when compared to the
effects of an organism that adopts, or changes, reference signals. And your
assesment of the (non)response of people to your earlier attempts to
describe them has a familiar ring.

I tried to discuss (or portray) similar points in my little handout at CSG
on "a modest proposal to neuro-cognitive-behavioral scientists" and in my
piece in Closed Loop on repetition and mimicry. In both places, I was
demonstrating what can happen when peopole adopt different reference
perceptions of for relationships between a cursor, a target and their own
hand movements. Relationships between the cursor and target change clearly,
unambiguously and even statistically significantly (for those who cherist
statistics) with each change in reference perception, and all the while the
person prevents the disturbance from having any sizeable effect on the
controlled variable and controlled perception. It is *very easy* to draw
a mistaken impression from instances of control during which the reference
perception does not change -- the impression that the disturbance is more
"influential" than it is seen to be when you compare even modest control
with no control at all, or compare control with a refence perception for
seeing cursor aligned with target with control when the reference changes to
seeing the cursor far above (or below) the target or seeing the cursor
moving in the opposite direction but twice as far as the target. The
results of those changes in the reference perception are orders of magnitude
more powerful than the effects of the disturbances present -- including the
changing disturbance pulled from a table of random numbers and applied to
the cursor, and the mother of all constant disturbances, gravity that acts
on the person. The cursor simply goes where the person decides it will go.
And a two-level PCT model performs as does the person. (A second level to
adopt different reference signals according to a program of conditional
relationships -- while t < X, p* := Z; while t (>X and <Y), p* := Q; and
so on.)

Later,

Tom