[From Bill Powers (950831.1530 MDT)]
Rick Marken (950831.1330) --
Where did you get the idea that Hans' system nearly killed a patient?
From Hans' description of the event on the net.
Correction: From your interpretation of Hans' description of the event
on the net. The "patient" was a pig.
The model in Hans' model-based controller is a model of the
function that connects the outputs of a controller to the results
produced by that controller; it is also a model of the predictable
effects of other variables (distrubances) on the results produced
by the controller. The kind of model based control you are talking
about is control of imagined (modelled) perception; Hans' model is
not a model-based controller in this sense.
In Hans' posted model, the world-model is affected by the output signal
of the controller which also affects the real system in the environment.
The output of the world-model is an imaginary representation of the
output of the real system in the environment (i.e., the controlled
variable), providing an imaginary perceptual signal x which is compared
against the real perceptual signal y as the basis for modifying the
model.
Suppose that in my head I have a model of another person. If someone
tells this model "You ought to be ashamed of yourself," the model will
compute that it should feel contrition and will decide never to do such
a shameful thing again. Thus if I say those words, my imagined
perception of the other person will cease to do the thing for which it
should be ashamed, and my perception of its behavior will be the one I
want.
At the same time, the same words reach the ears of the real person in
the environment. This person replies "screw you" and redoubles the
unwanted behavior. The real perception does not match my imaginary
perception, so I must revise my model of the other person (if I'm
looking at all at the actual behavior). After many revisions, my
perception of the world-model begins to respond to my words in the same
way that my perception of the real person does and the model no longer
requires revisions. Of course I also begin to use different words,
because there is a control loop involving the model.
That is how Hans' model works. It has many merits, particularly for
higher levels of behavior. I don't accept it as a model of all of
behavior, nor do I think it can handle disturbances in a realistic way.
But it does address a kind of behavior that exists and it can do some
things that a real-time negative feedback control system without a model
can't do. Some day, when someone is set up to do real experiments with
people, Hans' model can be tested and improved. Right now it's an
interesting guess.
As I recall, the model-based blood pressure controller is
essentially an open loop system that administers a drug on the
basis of an algorithm that continuously converts measures of blood
pressure into drug dosage; if the system controls anything it is an
output (drug dosage); not an input (measured blood pressure).
From the paper:
Due to instabilities in the arterial pressure signal caused by all
kinds of artifacts, the signal can be invalid for periods (much)
longer than 5 seconds. Feedback control is then impossible and
feedforward (open loop) control is, for safety reasons, allowed for
only a short time.
Actually, the controller measures mean arterial pressure which is
derived from a fluctuating measure of pressure affected by the heartbeat
and by many other disturbances that can render momentary readings
invalid. One main algorithm has to do with deciding when the perceptual
signal can be trusted as a measure of the variable the doctor wants to
control, the actual mean arterial pressure. The situation is difficult;
it is like having random disturbances between the controlled variable
and the signal that represents it, or is supposed to represent it.
The perceptual signal that results, which could be called the RMAP or
reliable mean arterial pressure, is compared with an adjustable set
point, and the error signal is converted through another complex
algorithm into a rate of infusion of the blood-pressure-reducing drug.
This algorithm takes into account such things as safe limits on rates on
infusion and changes in response characteristics during unvoidable large
transients. Otherwise, the basic loop is an ordinary negative feedback
control system.
All in all, it is a very difficult control problem. In some ways it's an
impossible control problem because the variable the surgeon wants
controlled is not and can't be the same as the perceptual signal that
represents it. Also, the organism is continually trying to maintain a
_different_ blood pressure, so the artificial system is in continuous
conflict with the real one. A lot of outside intelligence is needed to
determine when and in what way the pressure reading should be
controlled. The human user is an essential part of the system, because
there are many conditions under which it will simply stop working ("air
in line; occlusion; low fluid; malfunction"). Despite all these
problems, however, the system could work for long periods before,
during, and after infusions, without human intervention.
I'm not sure what this system has to do with human control systems, but
I have to admit that it's a hell of a control system. If it were this
hard for human control systems to manage their own controlled variables,
I doubt that there would be any human beings.
Mind you, I'm not saying I would have come up with the same design that
Hans did if handed the same problem. Or that I wouldn't have. This is a
one-of-a-kind control problem, and like Hans I would accept gratefully
any solution that worked.
···
----------------------------------------------------------------------
Best,
Bill P.