[From Bill Powers (931103.1340 MST)]
Hans Blom (931103) --
An interesting and thorough discussion of feedforward. You make a
number of points with which nobody can argue:
1. Feedforward is the only means of acting possible when feedback
is interrupted. When action is still necessary, the normal
feedback organization can't work and must be supplanted by
bridging the gap with actions estimated to have the required
effects.
2. Feedforward can work only to the extent that the behavior of
the environment continues to follow from actions in a predictable
way for some appreciable length of time. Only during that time
can feedforward continue to have the proper effect.
3. The actions required for feedforward can be based on
monitoring variables known to have an effect on the outcome,
without those variables being indicators of the outcome. The
actions must be calibrated and consistently produced in a
specific relation to the indicator variables. In effect, I add,
the result is to base the actions on direct or indirect
perceptions of the disturbances that can affect the outcome.
There are some points you made with which I can argue or to which
I can offer alternative interpretations.
4. Under HPCT, it is not necessary that a perceptual signal be
available to consciousness or awareness in order for it to play a
part in feedback control. In general, we are aware of the world
only in terms of a few perceptual signals existing primarily at
some high level of organization. This field of awareness can
shift among and within levels, but it is always only a small part
of the whole set of controlled perceptual signals. When you are
walking, you are unlikely to be aware of the fluctuations in
muscular effort that are responsible for moving and placing your
legs, yet control of walking would be impossible if there were
not control systems continually monitoring and controlling
kinesthetic feedback signals by varying muscle tensions. While
you're furiously typing a message, you are probably not directly
aware of the relative positions of your fingertips and the
keyboard, yet those relationships must be under active feedback
control for typing to take place, and so must the kinesthetic
position and force feedback control systems be intact and
working. So lack of consciousness of feedback signals is not an
argument in favor of feedforward.
5. A purely feedforward system can't work, precisely because the
"tuning" of which you speak is necessary. Prior to setting the
feedforward system in action on its own, you (or some higher-
level system) must monitor the critical outcome and adjust the
feedforward response until it consistently produces an outcome
close to the one desired. And after some period of operation it
is necessary to recalibrate the feedforward system because both
it and its environment will suffer secular drifts in sensitivity
to inputs. So every successful feedforward system must initially
and periodically be part of some larger feedback loop controlling
in the normal PCT way. The control action in this case is not
control by adjustment of lower-level reference signals, but
control by adjustment of lower-level parameters.
6. The range of situations under which feedforward control can
perform usefully is limited. It will not work at all at the
kinesthetic levels, because muscles fatigue too rapidly and there
are too many immediate static and dynamic interactions among low-
level systems. Feedforward at higher levels relies on accurate
feedback control of muscle forces and limb positions; open-loop
control at the lower levels would not be nearly precise enough to
produce output forces and positions with the accuracy needed to
project the effects of actions ahead in time by more than a
fraction of a second.
7. The kinds of variables amenable to feedforward control are
those that naturally change slowly, over seconds or longer, even
in the presence of normal disturbances. Male drivers of cars, I
read somewhere, look away from the road ahead for at most two
seconds at a time, and this interval shortens when there is even
moderate traffic. The implication is that this is not really
feedforward, but sampled feedback control, with the output being
held constant between samples. In fact many apparent examples of
feedforward control should more properly be considered as sampled
feedback control, with the intervals between samples depending on
the susceptibility of the environment to disturbances.
8. Your remarkable home heating system, which works strictly by
measuring the outside temperature, depends for its success on the
absence of certain kinds of disturbances. For example, your house
must be built so that there is no differential effect on inside
temperature on cloudy versus sunny days. It must not have any
large windows where the sun can shine in during some parts of the
day. You must not do much cooking, or light a fire in your
fireplace now and then, or have parties at which large numbers of
people might be present at the same time (at about 100 watts of
heat output each), or open windows to clear the air. You must
have vestibules at every door to the outside, so that in the
winter it makes no difference whether the house is always closed,
or people are going in and out frequently. The insulation of the
house must be remarkably good, so that heat losses are unaffected
by outside wind velocity, or by insulating and reflective
blankets of snow on the roof. In short, disturbances other than
changes in outside air temperature must be singularly ineffective
at altering the inside temperature. Or, of course, you might have
decided that as long as the inside temperature remains within 15
c of some average value, the control is "good."
9. An adaptive compensatory system can't learn to control
"better" unless its outer loop is a feedback control system. The
judgment of how good control is depends on comparing a perception
of the controlled outcome with a reference-perception.
10. In your electrifying Autobahn example, the highest-level
control system was the one that perceived the ineffectiveness of
the normal steering control system and turned it off, instituting
a different set of control processes that stood at least some
chance of preventing an accident. The control of this alternative
feedforward system was very poor: it depended entirely on the
unforseeable circumstance that a patch of dry road would appear
near the edge of the road, allowing the normal control system to
be used again. If that patch of dry road had not appeared, the
feedforward system would have failed because it behaved in the
way appropriate to the average circumstances when it had been
used in the past, and was incapable of adjusting itself according
to its current effects. Feedforward systems can be tuned only to
the average conditions; they rely on differences from that
average being small enough. You are lucky to be alive.
11. It is a common conceit among certain doctors that they can
see what is wrong with a patient in the first 30 seconds of
visual observation. I advise staying far away from such doctors.
You say
It has been estimated that many family doctors or general
practitioners (now extinct in the US?) accurately recognize
about 80 percent of patient's problems within 30 seconds of
meeting the patient.
Family doctors and general practitioners judge the correctness of
their diagnoses mainly on whether the prescribed treatment
results in a cure. Since it has also been estimated that 80% of
patients will recover from their illness without any intervention
at all, the doctors' self-confidence is, to say the least, based
on doubtful evidence.
I knew a doctor of the sort you mention. He was knowm to his less
flambouyant colleagues mainly for the inaccuracy of his
diagnoses, resulting primarily from his insistence that his first
impression was nearly always right. He also refused to consider
subsequent evidence against him (including death of the patient
from some undiagnosed cause). Such outcomes were never his fault.
So he would claim that feedforward worked just fine.
While, as you say, apparent cases of feedforward abound, a great
many of them are based on superstition and after-the-fact
"prediction." The kind of control achieved by feedforward is
necessarily very poor except under specially controlled
circumstances, most of which involve reliable artificial devices
and actuators rather than sloppy human nerve and muscle. This
does not prevent people from trying to control in a feed-forward
fashion: for example, consider all the "planning" think-tanks
that try to select policies that will have desired effects a year
or so into the future. There are strong motivations for people
who use such methods to claim a far better track record than they
actually demonstrate, and to blame inadequate data for the
failures of prediction rather than the simple fact that
feedforward control is helpless in the face of unpredicted
disturbances.
Far more common are cases of negative feedback control in the
manner of PCT. It is easy to overlook them, because we are used
to thinking of our actions as "just happening." The idea that we
have to continually _vary_ our detailed actions in order to
repeat the _same_ outcome is hard to grasp until you look very
carefully at behavior. It is so easy for us to reproduce outcomes
that we don't even realize how differently we are acting almost
every time we "do same thing." We take successive gulps from a
stein of beer, repeating the "same act" of lifting the stein
until it is empty, never realizing that each time we lifted it,
the dynamic and sustaining forces were less by the weight of the
beer absent from the stein and present in our bellies. The
evidence for PCT is subtle, but it is everywhere. Once you
realize what you have to look for, it is incontrovertible.
Many cases that appear on the surface to be feedforward or open-
loop behaviors turn out to require negative feedback control when
you consider them more carefully. This, in fact, is why S-R
psychology exists at all. A careful examination of all behavior
will rapidly deplete the world of apparent cases of open-loop
behavior, and swell the number of obvious cases of feedback
control.
I can agree with almost everything you say about the nature and
characteristics of feedforward control. I just can't agree that
it accounts for very much of human behavior.
ยทยทยท
-------------------------------------------------------------
Best,
Bill P.