[From Bill Powers (921004.0800)]
Tom Bourbon (921002.1050) --
Your treatise on good control vs good modeling makes a critically
important point, which is that the control model doesn't fit behavior
well just because it's a control-system model. It has adjustable
parameters that can make it fit both excellent control behavior and
very poor control behavior. With different values of the parameters,
it would fit even expert control behavior very poorly.
The full significance of the numbers you posted isn't self-evident in
the correlations:
A
c-t mod-pers
.987 .984
.987 .984
.463 .880
.404 .954
.430 .930
.986 .989
The "c-t" column indicates, via correlations, control error by the
real person. In three of the cases, the person's actual "cursor
position" (value of the variable A) departed by a very large amount
from the pattern it should have followed for perfect control
("target"). With a correlation of only 0.404, explaining the observed
behavior simply by saying "it's a [perfect] control system" would have
accounted for only 16% of the variance of variable A even using a
measured reference level. But for that particular entry, explaining
the behavior by saying "It's a control system with a [particular
measured] value of integration factor and a [particular measured]
reference level" accounts for 91% of the variance of variable A. Even
for the entry in which the model did the most poorly and the subject
did the best (among those worst three), the percent variance accounted
for rises from 21.4% to 77.4% when the parameters of the model are
adjusted to fit the behavior of the person, rather than for perfect
control.
One lesson to be learned from this is that it means little to explain
behavior by saying "it's a control system." That explanation could
account for 90 percent of the variance of observed behavior (or
better), or for 16 percent of the variance (or worse). It's not just
the architecture of a control system that enables it to predict
behavior; it's the quantitative parameters. When we say that a
control-system model explains behavior, we mean a control-system model
that has the right parameters, not just the bare boxes with arrows
between them. The qualitative model applies only when control is
expert, so there is little difference between the actual control
processes and perfect ones. And when control is expert, you hardly
need the quantitative model to prove that control is going on.
···
-----------------------------------------------------------------
Greg Williams (920928) --
An aside: What's going on in HYPNOSIS in PCT terms???
Another possibility is "suggested" by the fact that hypnosis almost
always involves verbal communication. You're getting sleepy, your arm
is getting heavy, and so on. Induction processes seem aimed at drawing
attention to lower-level perceptions while a voice drones on and on.
If you can get the locus of awareness moved into the lower control
systems, this leaves the higher ones running on automatic. Then words
(and meanings) circulating through the logical and higher levels would
be indistinguishable from self-generated logical processes or
programs, and might begin to be used as if they were the person's own
intentions at the higher levels, particularly the category through
program levels. Disturbing principles or system concepts would lead to
resistance, just as they say. And of course awareness is still present
-- you feel as "in charge" as ever, except that what you happen to
feel like doing is partly under external direction. Oh, blah, blah,
blah.
-----------------------------------------------------------------
Gary Cziko (921002.0200!) --
I like that excerpt from your book very much. It suggests that there
is some conscious participation in and direction of reorganization --
at least direction of WHERE it is applied, if not direction of the
detailed process itself.
As I was reading your examples, I was recalling my own first lessons
in swimming. I really wanted to know how, but the main memory that
persists is the terror involved in making myself go into the deeper
water in order to be taught. I did not, after all, know how to swim at
that point. The biggest problem was in keeping the fear (and
excitement, pretty much the same thing) within bounds so I didn't just
head for shallow water and give up, or panic and start thrashing
around.
I think, as a theorist, that it was the fear/excitement that drove the
reorganization, but clearly what I learned was swimming, not high-
jumping. My existing goal structure, which included wanting to learn
to swim like other people were doing, had to be involved, and the fear
certainly resulted from my putting myself into a situation I felt as
dangerous and wanted to leave (an action that could only have resulted
from operation of my existing control systems and a more or less
deliberate induction of conflict). Also, my mother and father were
giving me verbal instructions and demonstrations, so clearly what I
was learning was something presented to my senses in the present-time
world (as opposed to a genetically-driven process). They not only
taught, but reassured me of my safety because they were there to get
me out of trouble, so this lessened the conflict.
One aspect of this learning process that's of interest is that it
involved imitation of movements, using control systems already well-
developed but not adjusted to the underwater environment. I was
consciously trying to move my arms and legs in the demonstrated way,
but DISCOVERING the sensory consequences of acting in just that way
under water. I was taught pretty much as you outlined -- learning to
tread water first, and stay afloat. But nobody could tell me how the
resistance of the water was going to affect the way my limbs moved, or
the way that the resulting forces kept me up. I could see how my
mother moved her arms slowly back and forth, sideways, but only when I
tried it did I realize how much force had to be used to create those
slow movements. I did a lot of experimenting, discovering that moving
a flat hand sideways would push me up if the hand was tilted
correctly. I didn't consciously make any changes in my own
organization; I just kept noticing effects and trying to repeat them
(or avoid them, cough, choke). But clearly, the changes that were
going on were all aimed at gaining control over the new perceptions
that were being generated by these new ways of using familiar actions
in a new environment -- repeating some perceptions and trying to make
sure others didn't occur again.
I think we have to separate the process of reorganization itself,
which is random but reasonably efficient, from the processes that
create the need for reorganization and those that direct it to the
appropriate systems. The hierarchy, as Greg Williams has been
suggesting, does get into the act, in the sense that it can actually
create a situation that induces critical error, and can thus (learn
to) cause reorganization independently of any environmental pressures
or accidents -- which, of course, continue to work through their own
effects on critical variables. Mary, reading your chapter excerpt,
pointed out the article in Science News about people who seek risk on
purpose, like rock climbers. They put themselves in more and more
risky situations, but by doing so acquire greater and greater skill.
It must be confidence in their own capacities for reorganization that
makes this attractive -- they know that they will quickly learn the
required skills, so that the fear goes away. I suppose that they
relabel their emotions, so that the sensations of bodily state that
others might call fear are categorized in more positive ways, and not
avoided.
This conscious creation of situations that turn reorganization on
doesn't produce any control of reorganization itself; it merely
creates errors that turn the process on, and confines it to the areas
of the brain pertinent to learning a particular kind of thing. I don't
think we're ever conscious of reorganization itself; only of a
gradually increasing ability to make certain perceptions occur when we
want them to, and of a heightened state that reflects the critical
errors that drive reorganization. The human system, which begins with
a very low degree of organization in comparison with most animals, can
reorganize to the extent that it can even learn to create the
conditions that produce reorganization. It still can't direct that
process in detail, but it can assure that the process continues until
the conscious goal is met.
Anyhow, your descriptions of learning to swim make it clearer than
ever that even if my model of reorganization is basically OK, it is
part of a larger system and doesn't tell the whole story.
Skinner, in his invention of "shaping," anticipated "scaffolding." He
found that animals learn quickest when only small changes have to be
made at any stage. This has important theoretical implications, I'm
sure, although I couldn't say what they are right now. Somehow the
space in which an organism gets from organization A to organization B
must be continuous between A and B -- learning couldn't work in a
universe where there was no coherent path from one mode of behavior to
another, with intermediate steps also making some kind of sense.
--------------------------------------------------------------------
Greg Williams (921003) --
I guess what I object to in your proposals that we can "guide"
reorganization in other people is the assumption that the guider can
arbitrarily set this as a goal and then just carry it out. The whole
discussion of "influence," from your end, seems to be setting up the
influencer as an independent agent, with the influencee simply
reacting as the influencer wants. This puts all the volition and
agency into the influencer and ignores that of the influencee. It
really seems to me to be an attempt to get around the interactive
nature of control and go back to the old picture of the organism's
behavior as a passive consequence of external events.
When we influence the actions of others in systematic ways, we succeed
(when we do) only BECAUSE they are independent controlling agents, not
DESPITE that fact. You can't make a person reorganize without inducing
critical error into that person, or creating a situation that does it
for you. The immediate reaction of the other person will be to find
SOME way to counteract this most invasive act, or to negate or escape
from the situation that produces the error, by the most direct and
immediate means possible including setting the situation back to the
way it was. The other person isn't going to just wait around passively
while you, to suit your own goals, start causing pain or suffering.
You can set problems all you like, but there's no constraint on the
other person to solve them according to the rules you had in mind, or
to seek the particular solution you consider the "right" one, or even
to be interested in the problem in the first place. When you use this
method, you're in constant danger.
I objected that you can predict the outcome of reorganization only
when you arrange the environment so that only one action or learned
method of control will in fact correct the critical error. You took
this to mean that there IS a way of guiding reorganization -- just do
what I described. But I offered this as an objection because it is so
unlikely that you could ever do this with a human adult, a peer,
without some great source of power behind you. It implies that you
know all possible ways of solving the problem to the other's
satisfaction; it implies that you already have far more control of the
local environment than the other person has, so you can freely
rearrange the environment without any effective objection from the
other person. It implies that you have sole control over what the
other person needs.
I claim that in order to establish the conditions under which this
kind of influence (and many other kinds) could be even apparently
effective, you would have had to establish a background of power over
the other person and over the environment, power denied to the other
person. In many of your examples of means of influence, I believe you
are assuming (perhaps without knowing it) surrounding conditions that
amount to establishing a far greater degree of coercive control over
the other person than the degree of influence indicated by the simple
example. This is particularly true when you speak of influencing or
guiding another person's reorganizations. But it is true even of such
innocuous means of influence as rubber-banding. How are you going to
keep the other from seeing what you're doing? Or, after seeing it,
from objecting to it or deliberately changing the goal simply to
frustrate your attempt at influence?
When we talk about influencing children in the classroom, we tend to
overlook the fact that they are in that classroom, and are trying to
do passably well in their courses, not out of their own choice but
because they have been physically forced to attend school, stay in
their seats, be quiet, listen, study, take tests, and keep trying
until they pass the course. The children are not offered any choices
at this level of description. It does not matter whether they want to
be there, whether they want to learn the course material, whether they
want to take the test. They MUST be there, and if they want to avoid
pain and disgrace, they MUST pass the course. Throughout the years of
mandatory education, the children are in prison. If they try to escape
they will be seized, punished, and returned to class. The elementary
and high-school system is fundamentally based on coercion and control
by people who have the power to withhold rewards, restrict freedom,
and punish. This philosophy naturally carries over into higher
education. The situation is quite analogous to that of a rat in a
Skinner box, where the experimenter can put the rat into the box
regardless of where the rat wants to be, and give or withhold reward
while restricting the means of getting it to performing just the
action that the experimenter wants to see.
This is not "influence," but flat-out control. It is based on brute
physical power capable of crushing any resistance to it.
The same sort of glossing-over of the background of brute power that
we see in mandatory education also exists in the world of business.
Even in enlightened business organizations where people are allowed to
choose their own hours, their own tasks, their own ways of executing
those tasks, and even some of the orders that management gives them,
the underlying principle is that of control by overwhelming physical
force. The very idea that people are "allowed" to do these things
implies a surrounding wall of coercion. A person who does not want to
perform any task at all will simply be left to starve. Any attempt to
avoid starvation in a way other than achieving the company's
objectives will be punished by removal of the paycheck; any attempt to
get the paycheck anyway will be punished by the Law. No matter how
varied and enriched the environment within the cage, it is still a
cage in which the person is kept by brute force.
When, in the background, you have the power to control a person's
physical well-being or even existence, the role of mild "influences"
becomes an illusion. The company can offer "incentives" in the form of
"bonuses" for good work, but the clear implication is that bad work
will be punished by loss of the bonus; you can't have the bonus anyway
just because you desperately need it, or because your salary has been
cut to finance the incentive program. Mild, friendly, helpful
influences succeed mostly because failure to be influenced in the
direction obviously wanted will result in loss of the job. The
objective of these influences, keeping the worker or student in
conformity to the larger objectives of the owners of the business or
the administrators of the school, is backed up by the credible threat
of force -- credible because it is used whenever necessary, which is
often.
You propose as one effective means of controlling another:
3. A arranges B's environment so as to trigger learning
/reorganization in B's control system resulting in actions which A
wants to perceive.
The natural reaction of B is to prevent A from arranging B's
environment in that way, perhaps by sinking a knife into A or by
driving a car past A's house and letting loose with an Uzi. This
method of control dooms itself to failure. That's why I don't treat it
as a serious proposal: it can't work, except temporarily.
There is another reason I don't treat it seriously: that is because
reorganization will NOT result in the actions which A wants to
perceive, but in whatever actions B decides may work best for B. You
assume that A can arrange the environment so that only ONE action can
succeed. But you overlook the action of changing the environment back
to its original state, you assume that it's POSSIBLE for A to make
this arrangement without any interference from B, and you assume that
A can anticipate all possible outcomes of reorganization (such as
suicide). I think these assumptions are unwarranted, in general.
------------------------------------------------------------------
I encourge you not to give up the PCT goal of 0.99+ correlations.
Otherwise, I believe, you will not be able to decide WHICH PCT
model is the correct one. Fairly low (but higher than typically
found in the behavioral sciences) correlations are convincing (at
least to me) that PCT models are correct, generically, for
tracking. But I next want to know WHICH PCT models are better than
others, and this requires comparing various PCT models'
correlations with what people do UNDER CONDITIONS AFFORDING ENOUGH
SENSITIVITY TO SHOW SIGNIFICANT DIFFERENCES IN CORRELATIONS. Such
conditions, I claim, are "poor" control conditions.
I agree with you, mostly, but I think we should give up on
correlations altogether. They're just not sensitive enough when they
get up into the 90s. It's much simpler to use RMS error over peak-to-
peak value of the predicted variable, which happens to be a standard
measure of noise-to-signal ratio in electronics (or 1/SNR). It's
intuitive, in that you can visualize how the random fluctuations
relate to the range of the variable of interest, and it indicates
approximately the expected percent error of any given value of the
variable. The trouble with the standard statistical measures like
correlations, percent variance accounted for, and so on is that they
tend to distort the situation in favor of making it look better than
it is. You'd think that a correlation of 0.8 would be 80 percent as
good as a correlation of 1.00, wouldn't you? But any physicist who got
an 80% correlation from his measurements of physical quantities
(outside quantum physics) would immediately start dismantling his
apparatus to see what's wrong with it. He'd probably do that with a
correlation of 0.9, too, and even 0.99. You don't start getting good
data until you're in the 0.998s (roughly 5 percent error of
prediction, if I've figured that right).
Most of the difficulty in finding a control-system model for a given
behavior is not so much in finding the right form of the model or the
right definition of the controlled variable, but in setting the
parameters of the model correctly. If you don't have the controlled
variable pretty well defined, you're not going to get much of a match
with the model at all. When you get the right definition, you're going
to get high correlations right away with just about any old parameters
-- 0.8, 0.9. Usually you start with essentially perfect control as the
assumption. Then you back off on the parameters to get the best match
with the actual behavior, which will usually bring you up toward 0.95
or 0.98. Then, if you still want more accuracy, you start playing with
the details of the model -- whether there's slowing in the perceptual
function, how much integral, proportional, or derivative component
goes in the output function. That will get you into the .99s or
0.999s. If you still feel it's worthwhile to push for more accuracy,
you can start in on nonlinearities, thresholds, and transport lags.
This should be the case for almost any continuous type of control.
"Tracking" isn't a particularly special case; it represents control of
all sorts of things, as I show in Demo 1 where you can control half a
dozen continuous or near-continuous variables including size, shape,
pitch, and even a digital number that changes on the screen.
My view is that to build a new science of psychology, the first thing
we have to nail down with simple experiments is the EXISTENCE of
control of all kinds of perceptual variables. For this you don't need
to refine the model into the stratospheric correlations. By the time
you reach 0.999, the chances of this NOT being a control phenomenon
have risen into the billions to one. I'd settle for a million to one.
But not less. Why leave room for argument?
I think that finding exactly the correct control model can't really be
done without neural circuit-tracing. There are just too many ways to
accomplish the perception and control of the same variable in exactly
the same way. There's also the question of what you gain by having a
neurally exact model. There's no reason to believe that the next
person you investigate in such depth will control exactly the same
variable, or if it's the same variable, do it with the same circuitry.
Up to a point there will be close similarities in the way different
people do, say, pursuit tracking tasks. Beyond that point you can't
generalize any more to say this is how "people" do pursuit tracking.
There will be different numbers of muscle fibers involved, different
numbers and placements of sensors, different mechanical advantages,
different lengths of neural pathways, different visual acuities, and
different settings of parameters generating different trajectories of
movement, stabilities, linearities, and limits of control. The only
reason to go past some reasonable level of predictive accuracy would
be to characterize an individual down to the last synapse, something I
wouldn't ever need to do even if I could.
I don't think that we will ever be able to say which control model is
the absolutely correct one -- if only because different people will
accomplish the same control tasks using different control
organizations. It's more important to establish control as a
phenomenon at many levels by experimental demonstration, in the manner
of Galileo, and to establish methods for refining models as far as we
please when such is the point. It took about 300 years after Galileo
for physics to claim that further advances would be in the sixth
decimal place. We aren't even 50 years into this revolution.
--------------------------------------------------------------
I move and second that we propose Bill Powers as a speaker for >MIT's
Revolving Seminar.
I can't think of a single reason for going through such a painful and
futile experience.
---------------------------------------------------------------
Best to all,
Bill P.