[From Bruce Abbott (961003.1325 EST)]
My thanks to all who have participated in the KR (knowledge of results)
discussion. My reason for posing the problem has a serious purpose. Many
people have a disorder of the airways called asthma (myself included), and
for some it is chronic and serious. During an asthma attack, the muscles of
the bronchial tubes constrict and mucus is secreted in the tubes, producing
wheezing and great difficulty in getting air into and out of the lungs.
Typically these episodes, called "asthma attacks," develop rapidly, persist
for perhaps half an hour or so, and then gradually fade, returning the
airway to the pre-attack state. Most attacks are not life-threatening, but
they can be. Sufficient closure of the airways can lead to suffocation.
Attacks may be triggered by a variety agents, including allergens (inhaled
or ingested), cold air entering the bronchial tubes, and even exercise; a
person's sensitivity to these agents can be increased through
irritation/inflammation of the bronchial tubes via airborn particulates and
infections (e.g., bronchitus). For the asthmatic the problem is to gain
control over the state of the airway, to keep it as close as possible to a
reference level of free breathing. Accomplishing this goal requires taking
appropriate steps to abort or diminish an attack when one occurs, to avoid
those disturbances that trigger an attack, and in some cases, to reduce the
ability of disturbances to act as such.
One difficulty in establishing such control is that changes in the
controlled variable (state of the airway) may be difficult to perceive in
the early stage of an attack when action to correct the problem should
begin. Thus, it has been proposed that asthma sufferers be given training
to improve their ability to perceive the relevant changes that indicate the
beginning phase of an attack. In one study conducted to test this idea,
participants (both asthmatics and nonasthmatics) were tested for their
ability to detect a restriction that was sometimes inserted into a tube
through which the participants breathed, in similuation of the restriction
produced by bronchial constriction. During the "training" phase, some
participants were asked to say whether they thought the restriction was
present or absent and then were given "feedback" as to the actual condition.
There was some improvement among these participants in their ability to
tell, as compared to others who were not given "feedback."
The object, of course, was to develop a training program through which
asthmatics could learn to recognize the early changes in lung-state that
precede an asthma attack, so they could take action abort or minimize the
attack before it got worse. To put it in PCT terms, the object was to
improve the perceptual input function so that the controlled environmental
variable (state of bronchial constriction) could be more effectively
controlled. This got me to wondering how to describe the control system
during the training phase (when "KR" was being provided), and how to account
for any resulting improvement in control.
The discussion of KR that followed my question ranged beyond the experiment
for which my question was posed; even so, I have found the proposals and
arguments helpful. We've had some excellent analyses of different KR
arrangements by Bill Powers and Rick Marken, showing how KR might function
as controlled variable or as disturbance, depending on conditions. So far
only Shannon Williams and I have offered proposals to address how the
"perceptual filter" might be tuned. Shannon's proposal used a set of neural
nets whose parameters were optimized to discriminate the presence/absence of
the restriction and to "read" the output of this first net, which I take to
mean another set of parameter adjustments performed to optimize the
relationship between the state of the first net's output (input to the
second net) and the "judgment" of the second net as to presence/absence of
restriction. I presume Shannon has in mind something like back-propagation,
using the KR as the "training" state against which the net output is
compared when determining what weighting adjustments to make. I suggested
that the ability to discriminate the two states depended on the "sorting" of
perceptual-input states into two classes, and am currently imagining some
sort of averaging process that would allow the average differences between
the two states to stand out, something like the way cortical "evoked
potentials" are pulled out of the cortical "noise" background in the
analysis of EEG (electroencephalograph) output.
Regards,
Bruce