[From Bill Powers (960224.0200 MST)]
Chris Cherpas (960223.1327 PT) --
Q: How can references _ever_ match inputs?
Here's why I ask: PCT makes the claim that specific responses
can't possibly be what is learned since you're always in different
orientations/positions when you act, so the same set of muscle
tensions won't work for a changing environment. But what about
the "stimulus" side? Don't we see, hear, feel, etc., everything
from all sorts of orientations, viewpoints, angles, etc.?
Therefore, how can references possibly work under such variable
conditions? How can any inputs match a reference if the world
is so variable?
When we perceive objects in the environment, we see them from many
positions, orientations, and distances. Yet we see the "same" objects.
The visual system's input functions seem able to remove the point of
view of the observer and represent objects in relation to other objects
(of which the observer is just one possible object). These relationships
are perceptually represented so they become independent of the point of
view of the observer. A stable perceptual world is generated which is
almost invariant with respect to the observer's point of view.
An example is aligning a screwdriver with the axis of a screw. When the
observer is driving screws, the angle of the screw relative to the line
of sight can vary a lot from one screw to the next, and the line of
sight can vary, yet it is possible to manipulate the screwdriver shaft
so it stays oriented along the axis of the screw while the observer
turns it.
In a program called "14deg.c" I developed some simple kinesthetic
control systems for a jointed arm and hand that can work in this way. It
is possible to combine joint-angle signals to create higher-level
perceptual signals which represent positions of a hand in an x-y-z or
pitch-roll-yaw space, and also which represent angles of rotation of the
hand about any axis (I demonstrated only the "screwdriver" effect, where
the hand holding a tool can be rotated around an axis at an angle to the
forearm). This provides the necessary degrees of freedom in the action-
control systems to work with visual systems that represent the world in
an externalized frame of reference.
Handling the full visual control systems that use these kinethetic
control systems is still beyond me, but I can see how a person with
better analytical skills could solve the whole problem. Robot designers
have solved many parts of the problem, but have not yet been able to
develop the kind of visual perception model that would be needed (at the
top levels, industrial robots operate mostly open-loop, depending on
mechanical precision and precalculation of movements to achieve accurate
positioning). One reason that PCT models are still pretty simple is that
the problem of changing viewpoints that you mention is a very tough one.
The solution to the problem is there -- but getting from here to there
is mathematically difficult, at least for me. That's one reason I keep
hoping that we will run across some control engineers who are both
skillful enough to solve the problem and interested enough in the
general PCT concept to want to solve it.
All that said, there is another side to the question you ask. It's not
just changing points of view that require control actions to vary in
order to compensate for them. There are also independent forces in the
environment which, even when there is only one simple point of view
involved (as when we're looking at a computer display), can affect the
perception that is under control.
Steering a car is a favorite PCT example, because there is only a single
point of view to consider, and one simple action that affects the way
the road looks relative to the car in the windshield. Yet in this
example, there are many forces other than those due to turning the
steering wheel that affect the path of the car. There are crosswinds,
bumps, tilts, and many other factors that tend to make the car drift
left or right. They can't be anticipated, and there are no sensory cues
that provide measures of their influence -- not accurate enough measures
on which to base an adequate compensation. This is an excellent example
because the steering accuracy is (from the open-loop perspective)
impossibly high. After a 100-mile trip, the car will still be within a
foot of the middle of its lane. Accounting for this accuracy without
using a feedback model is impossible: we're talking about an accuracy of
one part in half a million -- and steering errors are _cumulative_.
Every driver has an intuitive understanding of how disturbances affect
steering. Driving down a straight road, you can be exerting no steering
effort at all, or a substantial twisting effort either to the left or
the right -- with the car in exactly the same position on the road. In a
stiff crosswind, you can steer the car around a gentle curve to the left
-- and realize that you're exerting a perceptible steering effort to the
right. You have no direct way to sense the crosswind, yet you
automatically produce the right amount of steering effort in whatever
direction is required to keep the _perceived result_ the same.
This is the effect to which I refer when I point out that in general you
can't just learn a fixed response to a stimulus and expect the
consequence to be the same. Only in very special circumstances is this
possible. Those circumstances, unfortunately, are exactly those we find
in the well-regulated environment of a typical laboratory experiment. If
an action always has the same consequence (or average consequence, if
random variations are deliberately introduced), then it will seem that
producing the same consequence requires the same action to be produced
every time. From this mistaken observation it is an easy step to
concluding that what is learned is "the" action that will produce the
given consequence.
If we apply disturbances directly to the consequence, however, and if
that consequence happens to be under control by the organism, we will
find that the action _changes_ in just the right way to compensate for
the effects of the disturbance. We now have essentially the same (net)
consequence being produced by a different amount or even direction of
action. Now it becomes clear that what is learned can't be a specific
action: what is learned has to be the internal part of a whole negative
feedback control loop, a loop that is able to _vary_ the action in any
way needed to keep the consequence the same.
This point is hard to get across to EABers (and others). One reason is
that in discussing behavior, hardly anybody actually talks about
actions. Most behaviors are defined in terms of consequences of actions.
If those consequences are controlled, and it is the thesis of PCT that
all regular consequences ARE controlled, then the consequences will
repeat quite reliably, and the fact that disturbances are present will
simply not be noticed. Why should we notice something that has no
apparent effects? But more important, it will not be noticed that the
actions producing the behavior are varying right along with the
disturbances, cancelling their effects. One _assumes_ that the more
detailed actions involved in a behavior must be repeating if the
behavior is repeating. And this is the assumption that makes S-R theory
or any other theory that assumes a regular causal chain false.
It's easy to miss this point, because one can admit that there are
variations in behavior without seeing that they are systematically
related to disturbances. The basic fact that PCT brings out is that the
one variable that _ought_ to vary when behavior varies (the consequence)
_does not_ vary in the way it should -- it varies much less than it
would in a lineal causal chain. In EAB in particular, this odd effect is
effectively masked by the custom of using variable schedules, in which a
very large random component is introduced in the feedback function. The
resulting closed-loop behavior becomes so noisy that the basic
relationships can't be seen without the kind of extended statistical
analysis that nobody every bothers with. And anyway, even with fixed
schedules the basic effect can't be seen because independent
disturbances of the consequence are almost never used. When there are no
disturbances, the same action DOES produce the same effect.
I hope this doesn't stop you from thinking in circles. That's the right
way to think about control systems. And I'm very glad to see you
struggling with these important basic questions. That's how you're
really going to learn PCT.
···
-----------------------------------------------------------------------
Best,
Bill P.