In human interaction, a prototypical controller (e.g., a prison commandant)
has the power to affect the controlled in arbitrary ways and is relatively
impervious to the actions of the controlled. That arbitrarity is important in
the intuitive idea is illustrated by the fact that, if arbitrarity is reduced,
the controller is perceived as being less potent. Political leaders are
perceived as powerful to the extent that they can get their noses into things;
a basketball player with good "ball control" can do a lot more than just
dribble in a straight line. That imperviousness is important is illustrated
by the fact that we see a person as being more controlling the more he or she
resists (or is able to resist), in cases of conflict, the overtures of the
controlled for mercy.
Now the engineering concept, based on unidirectionality in influence and
independent operation, captures the imperviousness aspect but not the
arbitrarity aspect. This partiality is not in itself a problem -- it is a
_stipulative_ decision made (implicitly) based on both ease of formalization
and usefulness in particular application contexts. But it is important not to
import the notion uncritically into new domains, and it is also necessary to
be clear about intuitions deriving from applications in old ones, because we
will often tend to conflate in the arbitrarity aspect even though it is not
part of the definition. Consider, for instance, the thermostat:
> "The feedback in these systems is the source of all confusion. In many
> engineering systems, the controlling subsystem is easily identified and its
> function tested out in a non-feedback mode. You might have tried this out
> yourself if you ever had installed a thermostat for a heating or cooling
> unit. After installation, one simply rotates the dial or presses some button
> a few times to check out if the ON/OFF function of the thermostat really
> works. So another test of a controller in any system, natural or artificial,
> is that it can also be operated in a non-feedback mode. And this
> non-feedback mode of operation clearly establishes which subsystem controls
> whom; it should clearly show that the controlling subsystem is not dependent
> on the feedback of information from the subservient system, whereas the
> subservient system clearly depends on the inputs from the controlling
> subsystem for its operation."
Here, the application of the independence (unidirectionality, or
"imperviousness") principle actually seems inaccurate. There is a chain of
influence from air temperature to metal length to switch state to heater
function back to air temperature. The installer is simply testing that
portion of the loop he feels is simultaneously most likely to exhibit
unintended performance and amenable to compensatory adjustment if so. If we
could locally adjust the laws of physics determining metal heat expansion then
we might well be worrying about "control" from a different part of the loop.
The only *unidirectionality* here is the influence of the installer on (say)
calibration of the device (see below) -- but there IS a significant
localization of *arbitrarity*: the thermostat portion of the loop has
structure that can cause it to behave in ways more varying (albeit actually
_determined_ by elements outside the immediate system/level being considered)
than the other components. In effect, it has more `options' for behavior and
is thus seen as the controller, while the other systems are more reactive
(less behaviorally arbitrary).
The inaccuracy in application of the engineering notion mentioned above
doesn't hurt things because control flows in a loop structure and whatever
mathematics is typically used still applies. But when loops are more extended
or diffusive, it breaks down. This is where confusion arises in the TV case.
> .... Suppose I set up the TV on an auto remote
> control. Let's say a person, who is not watching the TV, simply switches the
> TV ON and OFF every 10 minutes or some other fixed time interval. ...
First, notice that the counterexample (to the noncontrolling human viewpoint)
of the "10-minute plan" does not escape the recursivity problem. The person
who decides on the plan of turning on and off every 10 minutes was influenced
in this formulation by the nature of the TV (although not its particular
temporally- varying image signal), so the same question of who controls whom
comes up again because imperviousness seems even on both sides. Casting about
based on the thermostat analysis reveals that there are other things BESIDES
the nature of the TV that contribute to the 10-minute plan (e.g., knowledge of
the present discussion!) so that, seemingly, the human is more arbitrary in
determining the aspects of its behavior that affect the other thing than vice
versa. Further inspection reveals the TV is subject to other influences as
well (electricity from the power station, ambient conditions, etc.) that
determine to what extent IT responds to (and hence influences) the human, but
-- and this is the crucial point -- the sources and variety of outside factors
mediated through the human outweigh those mediated through the TV, and so the
human is seen as the controller by arbitrarity.
Thus, arbitrarity DOES seem salient for applying the notion of control in
certain cases -- at least, if we want the application to jibe with our (often
insuppressible) intuitions that so often guide our thinking. It is not
messier or less formalizable than unidirectionality (information-theoretical
entropy could be used, for instance), but it hasn't been that important in the
kinds of systems engineers typically work with, so it hasn't received
attention there. What about in the original case of interest here -- the
brain?
In fact, after all this, the solely imperviousness-based engineering
definition seems most useful for understanding brain function, because in
being faced with such a complex system to analyze, we want to break it up as
much as possible into smaller parts, and the utility of this enterprise is
enhanced when the interactions between the parts are (at least relatively)
one-way. (Then they can be just ignored or set to constants (or
constantly-varying functions) while the intra-part dynamics are worked out.)
In this view we have the hypothalamus acting as a controller because it
connects nonreciprocally to nuclei in the basal ganglia (a subsystem thought
to be involved in action selection), and even though the feedback through the
action-environment-reaction chain is there, its lesser immediacy is
imperviousness enough for the control notion to be practically useful
(although not all agree with this -- see Varela/Thompson/Rosch (1991) for
example). To be consistent, though, blood sugar levels, throat-dryness,
etc. must be seen as controllers of the hypothalamus.
As far as other structures in the brain are concerned, the strategy suggested
is to look for nonreciprocal connections -- the more so (loop-closing through,
e.g., only 3 or 4 stages rather than 1 or 2) the better. In some cases, like
the relation between the cerebral cortex and the brainstem modulatory systems,
though, the arbitrarity dimension or something related to it might start to
seem important, and in these cases it would be worth considering the pros and
cons of employing a formalization of it.
Adrian Robert
Department of Cognitive Science, UC San Diego