[From Bill Powers (930903.1330 MDT)]
Martin Taylor (930903.1130) --
Very instructive comments on "coarse coding." You're right on all
counts: PCT is an "as-if" model. There's another phenomenon which
must be very closely related: what Mark Olsen called "place
coding" at one point last year. A "place" in the visual field is
mapped into a place in a neural net. If one place represents the
target position and another the fingertip position, we don't
actually have a magnitude coding of x,y, and z differences as in
the Little Man. Something has to be able to compare the locations
in the map where a signal represents either fingertip or target,
and derive an error signal by some means other than simple
subtraction. The motor output has to be aimed to move the
fingertip's place to the target's place, in the neural net.
It could be that there are perceptual systems higher than the
neural map that are deriving positions as magnitudes, and that
the mapping simply preserves the spatial relationships until the
level is reached where the conversion to x,y,z magnitudes can
take place. I don't know of any actual explicit data saying that
the error is derived within the map. But there's a strong
suggestion of a way of designing a control system that entails
space-like operations within a neural map where contiguity,
angles, and separations are preserved. I've tried now and then to
imagine how it might work, without much progress.
···
---------------------------------
(Parenthetically, this suggests another component of the Test:
the tester must succeed in changing the value of a lower-level
CEV, and determine whether the proposed CEV that has it as an
element is kept under control. Simply observing a proposed CEV
at its own level is not enough.)
Good point. However, it isn't a CEV at the lower level that has
to be changed by a disturbance, but a UCEV -- an uncontrolled
complex environmental variable. If you arbitrarily change an
input that is under control, you'll be met with instant
opposition; you can produce a known arbitrary change, then, only
by overwhelming the opposing control system. The effectiveness of
the control system will depend on the stiffness of your method of
applying the disturbance. If you insist on creating a known
amount of change in the controlled variable, the control system
will lose control while the disturbance is being applied (unless
it can produce more output than you can).
What you have to do is to alter some variable that _influences_
but does not _determine_ the state of the proposed controlled
variable. This is why we define the disturbing quantity
separately from the controlled variable.
An example: if a person is controlling the distance between a
finger and a target, we insert a known disturbance by moving the
target by a known amount -- NOT by moving the finger by a known
amount. If we want to disturb the finger directly, we use some
other variable loosely linked to the finger -- we apply, in
effect, a _position-independent force_ that can be resisted by
alterations in the muscle outputs without necessarily altering
the finger position. This leaves the control system free to
operate normally while the disturbance is being applied.
------------------------------
I'd change Bill's second quoted sentence to "If you can't tell
when and by how much a theory is wrong, you can't tell
when it will be nearly enough right to be useful, either."
That's better than my original. It's so easy to speak of
quantative matters qualitatively!
--------------------------------
What Michael seems to be doing is to try to model the
observable behaviour of an organization AS IF it were a
hierarchic control system, without needing to assert that it
works by having trans-personal ECSs that receive percepts from
and provide references to ECSs within individual persons.
I have no objection to modeling the observable behavior of an
organization as a system. That's the only way to do it. You
characterize each discriminable component of the system in terms
of input-output properties (even if the components are control
systems), and describe the ways in which outputs interact
directly or through the environment, and the effects the outputs
have on inputs of other components. From this you get the system
equations, which you then try to solve, or simulate, to deduce
the behavior of the aggregate.
The observable input-output behavior of the human components of
an organization is the relationship between disturbances and
actions that oppose them, at a number of levels. Complicating
this description of input-output behavior is the variability of
reference signals at each level, which alters the balance-point
between actions and disturbances. The components themselves show
varying input-output relationships, depending on the level at
which you characterize them and on the action of higher levels
that may not be taken into account.
The outcome of such an analysis is simply the behavior of the
whole system that emerges from the interactions of the components
(and the effects of independent perturbations). There is no _a
priori_ reason for this emergent interaction to have the
characteristics of a control system. Even if the individual
active components are control systems, from an external viewpoint
that fact is evident only as actions that arise in response to
perturbations, which is equivalent to an input-output transfer
function. The controlled perceptions are not evident without some
added formal experimental procedures which would not be part of
the operation of the system under investigation. So as far as the
analysis of the organization is concerned, the situation is
simply that of a bunch of black boxes with inputs and outputs,
interconnected in observable ways. Because there are no natural
constraints on the kinds of interconnections that might exist,
there is no reason why the aggregate system should behave like
any particular kind of organization.
The only way we could justify analyzing an organization as if it
were, _a priori_, a control system would be to show some reason
that organizations must converge naturally to the control-system
kind of organization. This justification might be developed,
although I don't see how right now.
--------------------------------------------------------------
Best to all,
Bill P.