[From Bill Powers again --]
For those wondering why I found the Pruszinski, Kurtzer, et al. paper so
upsetting, annoying, or whatever (I haven’t yet figured out which
reaction is on top), there are some reasons more defensible than just
that I want to be appreciated.
The basic intellectual reason is that the paper cites as
“influential theories” only the main one against which I have
argued for some 50 years. It’s known now as “optimal control
theory.” That’s why I found it so surprising that Dr. Kurtzer should
be found subscribing to it – he has heard and I had thought long ago
understood and agreed with my arguments, and now appears to have rejected
them. I’d like to know why, and also I’d like to have my objections
recorded and at least discussed in public.
Optimal control theory is basically a strategy for designing control
systems that, for all I know, has virtues in general engineering
applications, but is to my mind an extremely unlikely or even flatly
wrong model of the behavior of any organism. It is based on deriving a
specification for a “control” (what we in PCT would call an
“output”) that will cause some external system to behave in a
particular way – open loop. Even if feedback (as we define it in
PCT) is involved in the form of eventual knowledge of the results of an
action, the feedback is used as a basis for revising the model from which
the output effects of the control are predicted, not as the primary
variable under control as in classical control theory and PCT.
In order to predict the effect of a vector of outputs, the predictor must
take into account (correctly) all the properties of the nervous
system and muscles, the kinematics of the body, the dynamics of the
interactions between the body and the physical world outside, the
behavior of independent variables in the environment that can disturb the
outcome as it is developing, and all the physical and chemical
interactions that occur between the actual outputs of the organism and
the final result that is to be achieved.
The authors of this paper acknowledge that there is great difficulty with
the multiple ways in which an output can have a particular final effect.
They note the ambiguity that arises when an attempt is made to trace
cause and effect from the environment to a perturbation of a joint angle
(shoulder in this case), since as they say there is an infinity of
combinations of causes that would have the same effect. That fact makes
for difficulties in designing an output “control” that will
have just one particular effect. How can the predictor know, given only
an experienced effect, which external cause is operating, so as to be
able to compute an opposing action to cancel the unwanted effect? The
correct answer is, “It can’t.”
Behind the optimal control approach is, I believe, a notion of control
that grew up because engineering is basically a commercial enterprise.
Because control systems are built by engineers to accomplish ends that
their customers want to be accomplished, the focus is on what the
control system produces to satisfy someone else’s desires. Therefore
engineers call the variable being controlled the “output” of
the controller, since it is the ultimate point of the design. The output
of a cruise control in a car is the speed of the car, because that’s what
the driver (not the controller) wants to be controlled. This, despite the
fact that the only output an actual cruise control device can produce by
itself is displacement of a linkage to a throttle. It can’t even sense a
headwind or tailwind or the upslope or downslope of the road or a soft
tire or a dragging brake. Talk about ambiguity!
In optimal control theory, it is said that the system controls its
behavior, planning its actions just so and thereby producing the desired
result. If you think of the desired result simply as a consequence of a
carefully adjusted behavior, it makes sense to think of control as
control of action. Control the action just right, and the desired result
will follow.
Of course that is all perfectly correct, in an exactly reproducible
environment free of unanticipated disturbances, with actions being
carried out by precisely adjustable actuators which never suffer changes
in their calibration, energy supplies, or physical integrity. However, it
is a pipe dream in the world in which organisms must learn to survive
using their own inherited equipment.
In a cruise control, what is actually controlled is not the speed of the
car, but a sensory signal representing the speed. If the sensor changes
its calibration, the cruise control will maintain the same sensed speed,
not the same actual speed. If the sensor is reading high, the actual
speed will be too low. Only a mythical or hypothetical cruise control can
sense the calibration of its own speed sensor. Real ones can’t.
Every practical control system works this way in the real world. What is
controlled is not the actuator output, the visible behavior, but the
sensory input. If a thermostat’s sensor becomes too sensitive to air
temperature, the room will be colder at the same setting of the
adjustment lever. If the gyrocompass starts reading too far west of
north, the airplane’s autopilot will faithfully steer a course too far
east of north. If a putative Mars Lander’s perception of its orbit is
expressed, accidentally, in the wrong units – say English rather than
metric – it will never be (and never was) heard from again.
Living control systems, and most nonliving ones, control their sensory
inputs, not their motor outputs.
The authors seem to have appreciated this because they speak in their
abstract of the “intelligent manipulation of sensory feedback.”
But in the world of optimal control theory, feedback is not from the
system’s output to its own input, but from the sensory input to the
output action. The effect is fed back from the input of the system to the
observer of the system. So now the model depends on converting a sensory
input effect into the right motor output, which makes it even less
believable as a theory of organismic behavior.
The whole problem arises from choosing the wrong point of view from which
to see a control process. Optimal control is described as seen by the
engineer building the system, not as the system itself would see it. The
system itself can know nothing of the external world but what its sensors
tell it; it can control nothing but what its sensors report to be
happening. Behavior of a living control system is the control of
perception, not action.
I could go on about the felicitous properties of hierarchical
multidimensional classical control, but I think the foregoing points are
enough to explain my frustration. I think there is room for discussion
here, if anyone would show up to discuss anything. Is that ever going to
happen? At the age of 85, I can’t wait too long to find out.
Best,
Bill P.