[From Bill Powers (950104.1445 MST)]
Martin Taylor (950104.1010) --
I'll wait for my previous message to get through, but just one comment:
it seems unnatural for the mature system to have outputs from the
intrinsic variables providing the reference signals for the
"principles" level or perceptual control--the top level of the
"normal" hierarchy.
Just to emphasize: the version of the reorganization system that I have
proposed does not send reference signals to the hierarchy at all. Its
outputs are strictly organization-altering outputs. Just trying to make
sure we're clear about differences between our ideas of reorganization.
···
----------------------------------------------------------------------
Peter Burke (950104) --
Obviously, we need to measure the disturbances to the system as
well as the perceptions, reference signals and outputs. Without
measures of all the important factors, no methodology will yield an
accurate picture.
Hidden disturbances don't have to stay hidden from the analyst. Once we
see that variations in behavior are maintaining some controlled variable
in a reasonably constant state, we can ask why those variations are
necessary. That will lead to a search for other physical variables that
are influencing the controlled variable, and when we find them we will
know what the disturbances are.
Most motor behavior is subject to at least one pronounced disturbance
which is routinely opposed even though its cause can't be sensed:
gravity. When we recognize that this disturbance is present, we realize
that even simple motor actions have to be control processes -- reaching
out to pick up a glass of water requires the generation of upward forces
that keep the arm, and then arm plus glass, from accelerating downward
at 1 G.
Disturbances are always present, but they're disguised by the very
process of control. When you open a drawer to get out some clean socks,
you just open it. Your control systems change the position of the drawer
without requiring any attention to the fact that some drawers resist
more than others, and if pulled open too rapidly would tend to go on
sliding out if you didn't apply a counterforce to decelerate them to a
stop. The net effect of all these little disturbance-countering
adjustments in our actions is that the disturbances have no noticeable
effects -- so we tend not to notice them. We just "do" things.
To understand how behavior really works, it's very important that we
recognize the impossibility of just "doing" things. Whatever outcomes we
are producing we are producing by performing specific acts, and if we
look carefully at those acts we will see that they are never the same
twice in a row even if the outcome repeats exactly. When we notice those
little (and often not so little) variations, we are led to the
realization that other influences must be acting at the same time, and
that is how we are led to recognizing the disturbances that are always
present. If you look for them you can find them. They're really there.
All this is without proposing any model of the insides of the behaving
system. You can see all this from outside it. We can't really "measure"
perceptual signals and reference signals and output signals. Those are
parts of an explanatory model which we _hope_ is similar in construction
to what goes on inside an organism, but which normally we can't measure
directly (so few of us are neurosurgeons).
We have to start by looking closely at visible behavior and its relation
to the physical environment. That's where we find the direct evidence of
control: stabilized variables, and the actions that are stabilizing
them, and finally the disturbances that are making those actions
necessary. Those observations establish that control is going on. Once
we've established that, we can ask what kind of internal organization
the behaving system has to have in order to generate this phenomenon.
That's where we end up with PCT.
I wonder if most disturbances are in fact hidden?
Yes, most of them are. The ones we can see are not perceivable
accurately enough to account for the precision with which we oppose
their effects. A driver of a car can see trees bending in the wind, and
dust and debris blowing across the road, but those indications aren't
within five orders of magnitude precise enough to allow keeping the car
on the road, in its own lane, for 200 miles (or even 200 yards).
The fact that we often have some sensory indication that disturbances
are present has been given too much importance. Control requires
_quantitative_ opposition to disturbances. Qualitatively noticing that a
disturbance is present is not nearly enough to account for the
quantitative precision of control. It's more like "Wow, look at the
stuff blowing across the road! No wonder I've got the wheel cranked over
so far." Our first sensory indication that disturbances are acting is
very often noticing the efforts we are producing to oppose them -- even
before we have tracked down the cause of the disturbance. A driver can
think the car is in crosswind when the real problem is a flat tire.
----------------------------------------------------------------------
Bruce Abbott (950104.1215 EST) --
Bill: Keep in mind that I agree that the properties of control
systems are best analyzed in light of the control system model.
What I am saying is that there are many questions about human and
animal behavior that can be researched quite effectively using
other methods. I see the control systems analysis as a subset of
"IV-DV" methods, not as an alternative to them.
What I'm trying to get at is that situations traditionally, and
apparently appropriately, analyzed in IV-DV terms can easily turn out to
be unrecognized control situations. I'm trying to find a demonstration
where the natural thing to do would be to handle it in IV-DV terms, but
where in fact this would be a mistake. I'm not asking you to drop back
into the role of a traditionalist here, but to put on that frame of mind
to judge what would be the traditional approach to the situation
(without implying that you would favor it).
The problem here is not with statistical analysis, but with the initial
assessment of a behavioral phenomenon that seems to suggest the
appropriateness of an IV-DV analysis. Control processes can easily be
misinterpreted as cause-effect processes; in fact, without control
theory, what else could happen?
I would really like to have some demonstrations that look convincingly
like a situation where IV-DV analysis would be the obvious choice, but
which are actually control processes. The point is to show people how
mistakes of this kind can be made, so the next time they look at an
obvious candidate for traditional of analysis, they might pause and ask
whether there aren't some signs that control may be going on.
Perhaps a pursuit-tracking demo would work better than simply
stabilizing a variable. Martin Taylor has already suggested that it
seems that the target-cursor separation is the IV and the handle
position or velocity is the DV, and such explanations have actually been
offered. In the pursuit tracking case, the target movements are
independent of handle movements, so there might be more for the
traditionalist to get hold of.
I've run the program; what is it you wish the "traditional
analysis" to analyze? Knowing nothing about control systems,
negative feedback, and the folly of trying to analyze such systems
in unidirectional cause- effect terms, and having run the program,
I would treat mouse movement as the _independent_ variable and the
observed changes in screen display as the _dependent_ variable.
OK, that's a start. Actually this might all be simpler if I just make
the display into three cursors next to three stationary targets, to get
away from all the complications. I'll do that -- it will also simplify
the data storage. I'll post the revision shortly. This will create a
very simple situation: three stimuli and one response. You can do the
analysis (if you're willing) with assumed causality running either way
-- it will make no difference.
I need to move the mouse to various positions and then repeatedly
sample and record the screen variables at each position.
The experimental run takes care of that: the subject picks one of the
cursors and holds it next to its target position during the run. We want
the subject to actually be controlling, of course, because the point is
how control behavior can be misidentified as IV-DV behavior.
-----------------------------
Bill, your comments about the misuse of statistical analysis in
psychology are well taken, but you have made some errors in your
analysis of the probabilities and in their interpretation. You
might want to ask Rick, an expert in such matters, to explain the
problems to you. When he's done that, I'll respond.
Actually your book has a lot more pages in it than Rick's does, so I'll
be happy to hear your corrections of my errors first. As I'm sure you
can tell, I'm not well-versed in this statistical stuff. What I was
trying to get at probably needs a much more careful analysis that I knew
how to give it.
-----------------------------------------------------------------------
Best to all,
Bill P.