[From Bill Powers (941025.1625) --
Bruce Abbott (941025.1025 EDT) --
RE: "control"
Rick Marken is arguing, of course, from the PCT definition of control,
which is very narrow and precise. But I think people tend to use the
term control correctly when something about the situation brings out the
difference between controlling and either "influencing" or
"determining."
I think that if presented with the explicit differences in a situation
that would call for using _control_, _influence_, and _determine_, most
people would correctly assign _control_ in the same way we do in PCT,
just from common usage and experience. Looser usages of the term
control, which encompass the meanings of the other two words, simply
reflect not thinking the situation through -- or perhaps come from never
having realized what the difference is.
When we say that A _influences_ B, we mean that a change in A is likely
to result in a change in B, but that B can change for other reasons,
too, so we can't predict the exact effect of A on B, or the final state
of B. When we say that A _determines_ B, we mean to rule out all other
influences; given A, the effect on B is predetermined.
When would we say that A controls B, and not use either "determine" or
"influence"? I think it is when B is subject to influences in addition
to A, yet A varies in such a way as to maintain B in a predetermined
condition. There are many scientists, apparently, who do not realize
that this is possible.
When we say that a driver's actions control the position of a car on a
road, we do not mean that for each action there is one and only one
position on the road that will result. We know that this is not true: a
given action does not _determine_ the car's position. Likewise, we would
not say that the driver's actions only _influence_ the position of the
car, because that implies that the car could be found in any position
relative to the road, depending on other influences that are acting at
the same time. When we say that the driver's actions _control_ the
position of the car, we mean that different actions may be needed at
different times to achieve a given position and that other influences
explain why the actions must differ, but that nevertheless we expect the
car to be held very near a particular position. We expect a consistent
outcome to be maintained by variable actions.
Of course if a person has never realized that it is possible for a
system to create consistent ends by variable means, then there will not
seem to be any need to differentiate control from either influence or
determination. It's very easy for a person to observe consistent
consequences of behavior, and simply assume that the consistent
consequences must be been created by consistent actions. This assumption
gets in the way of seeing the variability of behavior when it exists, so
the word control doesn't need to have any special meaning.
This is generally the case in the behavioral sciences. Behaviors are
named for the consistent ends that are produced, not for the variable
actions that produce them. We talk about "bar-pressing" behavior, naming
the regular consequences of the variable efforts and limb positions
which create the depression of the bar. We talk about "goal-seeking"
behavior, referring to the final effect rather than the variable path.
We talk about "scratching an itch" or "hitting a home run" as if they
were actions by the behaving system instead of consequences of the
actions. Skinner spoke of "emitting responses," as if the consequences
of muscle tensions just squirted out of the ends of the motor nerve
fibers, fully formed in terms of the organized effects they create.
So basically I argue on Rick's side; there is more to consider here than
customary usage. There are critical distinctions that need to be made,
and by confusing control with influence and determination we fail to
make them. I hope to persuade behavioral scientists that control is a
class of phenomenon quite different from any other, and that it should
never be confused with influence or determination. To do that, of
course, we must make sure that behavioral scientists understand the
phenomenon of control.
ยทยทยท
------------------------------------
At present I see absolutely no conflict between this selectionist
principle of reinforcement and B:CP's "reorganization" principle
whereby control systems are modified when they fail to do the job.
I agree, there is no basic difference although the reorganization model
is more explicit about proposing the conditions under which changes of
organization will occur. Under reorganization theory, for example, we
would never attribute the cause of learning to some property of a
reinforcer. We would look for the intrinsic variable that is affected by
the reinforcer, and say that _anything_ that can disturb that intrinsic
variable will become a reinforcer.
The main important difference, however, is that reorganization theory
applies strictly to the process by which behavioral organization
changes, as distinguished from the process by which any given
organization operates in relation to the environment. Skinner made no
distinction, as far as I know, between learning and performance. He
would have applied the concept of reinforcement just as much to the
change of behavior involve in the initial learning of a FR10 schedule as
to the change of behavior involved when the schedule shifts to FR20.
In our control-system models of operant behavior, we assume that
learning is essentially finished; our models can't reorganize themselves
so they become able to press a bar where before they could not do so in
the right relationship to reinforcers. We step in, as modelers, once the
animal has learned that pressing the bar will produce reinforcers. Then
we predict behavior under various schedules under the assumption that
there is _no_ change in organization across schedules -- in other words,
we are looking for one model which will predict behavior rates over the
whole range of schedules without any change in organization or
parameters.
So as matters stand at least for the present, we can't account for the
"conditioning" part of operant conditioning (except in somewhat vague
and general terms). But once we do have a model for a given asymptotic
behavior, we can apply the same model, without change, to a whole range
of schedules, and show that the organism is behaving in the same way
over the whole range: there is no more "conditioning" even when the
schedule changes, and even though the rate of bar-pressing changes
radically. To explain how an animal learns to couple a particular
behavior to a particular reinforcer, we would have to come up with a
model of reorganization that could do this. But to explain how an animal
"adapts" to different schedules, once the basic relationships have been
learned, we need only find a model with specific parameters that will
work for all possible schedules.
What PCT adds is a new definition of what gets selected--responses that
reduce error between reference and perception.
In the light of what I said above, you can see that this isn't quite
right. What's learned is not a specific response to a specific input or
error. It's a _relationship_ between the state of an input or error and
the state of a response variable. What's learned is a control system,
not an activity. The animal learns to press or peck at a rate
proportional to an error signal, whatever the value of the error signal.
It learns to generate an error signal as the difference between a
reference signal and a perceptual signal. It learns to generate a
perceptual signal as a function of certain environmental variables (if
that part's not built in). The learning concerns the connectivity of
these functions, and the parameters that relate outputs to inputs for
each function. The same result of learning applies no matter how the
signals behave. It's not a matter of learning a specific response to a
specific stimulus. What gets selected are proportionality factors, time
constants, and connections that make one variable depend on another.
The _outcome_ of this learning is that reponses reduce error between
reference and perception. But that is not what is learned. What is
learned is a little subsystem that has the properties required to make
this outcome happen.
-------------------------
What is not generally appreciated outside the field of learning and
behavior is that S-R psychology--the notion that all behavior consists
of responses to prior stimuli--died a quiet death a long time ago.
Nobody believes in it, least of all Skinnerian radical behaviorists.
Bruce, I claim that this is a myth, at least outside radical behaviorism
(I'll take your word for that). Take practically any experimental
journal in the life sciences, and you will find experiments in the form
of "the effect of manipulated variable A on behavior B." That is what we
mean when we talk about S-R theory. It doesn't matter if people have
stopped using the terms "stimulus" or "response" or both. What matters
is that they think behavior is caused by antecedent events or inputs.
They don't see that behavior is _variable_ and that it varies _as
required_ to produce a consistent outcome. They think that regular
outcomes necessarily imply regular actions. So when they study behavior,
they vary things that presumably affect the organism, and look for
correlated changes in the behavior. Convince me that 90% or more of
behavioral studies are not done that way and I'll agree that S-R theory
is dead.
-----------------------------
One of the goals of EAB is to establish more effective and humane
control over behavior. I sense that you have strong objections to this
goal because the methods developed thus far appear to require some kind
of deprivation (i.e., disturbance) to achieve that control.
If you'll look at that summary of the contents of Closed Loop that Mary
posted a little while ago, you'll find one issue on Social Control. I
advise reading it. The stance of PCTers concerning control of other
people is not based on moral grounds or personal preferences, but simply
on an analysis of what is involved in one person controlling the
behavior of another. I think that you'll find complete agreement with
your scenario about having your car stolen: as a control system, you try
to prevent that from happening. You will try to control anyone else who
tries or threatens to do that. That's how control systems work. If you
desire to control others in a humane way, that is what you will try to
do. If you desire to control them in an inhumane way, you will try to do
that. If you desire not to control them at all, you will not control
them at all, although you will experience certain consequences as a
result that may induce error into other control systems within you.
The point of PCT is not to make recommendations about controlling the
behavior of others, people or animals, but to deduce from basic
principles what will happen under various conditions where control of
others might occur. There are many conditions under which one person can
control the behavior of another very accurately, with no objection from
the other. Tom Bourbon has several two-person experiments showing
various ways in which this can happen. There are also cases in which two
people can each control the other's behavior quite freely, again without
causing any problems; Tom has illustrated such cases as well, in several
forms. Let me run through some of the logic of PCT as it concerns
controlling other organisms.
--------------------------------------
The basic principle is that one person can freely control the behavior
of another person if the actions the other person is made to produce do
not cause errors in _any_ of the other person's control systems. When
you understand the difference between variable actions and controlled
outcomes, you can easily think of ways of doing this, if seeing another
person behave in a certain way is your desire.
Of course it is possible to force another person to behave in a certain
way even if doing so causes errors in the other person. To do this one
must simply be faster or stronger than the other person, or be able to
call on greater resources than the other person has. This mode of
control generates conflict, but if one has means of output that can
overwhelm the other person's means of output, the stronger system can
have its way and the weaker system can only go along.
The greatest problems of control arise when the systems involved are
similar in control capacities. Then, when control by one system causes
error in the other, conflict arises and neither system can easily
prevail. In that case, the two systems become locked in opposition,
devoting large amounts of resources to cancelling the resources of the
other, and being unable to use them in the normal way. That's neither
bad nor good; it's just a fact. If you love conflicts of that sort, take
up wrestling or football.
As to methods of operant conditioning, they will work quite well under
certain circumstances. The primary requirement is that the controller of
behavior must have exclusive control over whatever is to be used as a
reinforcer. An equal requirement is that the controllee must want, and
continue to want, the reinforcer. The best way to guarantee the latter
is to make sure the reinforcer is something required for survival; the
controllee is almost certain to continue to want it.
The next step is to establish the conditions under which the controllee
can obtain the reinforcer. This is usually done by establishing a
contingency: if you do this, the reinforcer will be given. It is
essential, for this to work, that the controller be able to establish
the contingency and maintain it in force, and to make sure that there is
no alternative means by which the controllee can obtain the same
reinforcer. This is not hard to do when the controller is a human being
and the controllee is a rat or a pigeon. The controller can say "This is
the environment you are given: learn to live in it."
But when the controllee is another person, it is hard to conceal from
the other person who is establishing this contingency, and for what
purpose. This does not mean that the controllee would necessarily resist
going along. If the contingency is that the controllee must type 100
words per day with 90 percent accuracy to receive a $1000 reinforcer,
the controllee will probably be perfectly willing to do that for as long
as the controller wishes.
Problems arise when the contingency established by the controller does
not enable the controllee to produce the kinds of reinforcers, or as
much of them, as the controllee wants. In short, there is a problem when
living under the given contingency causes errors in the controllee to
increase. If the increase is moderate, the controllee will simply
resist. If it is large, the controllee will begin to reorganize, and
then there is no predicting what will happen. The controllee might go on
strike, or escape to a different environment, or attack the controller.
So if the controller wants control of the behavior of another person of
equal capabilities, and for whatever reason wishes to avoid conflict,
this control must be done with extensive knowledge of the other's
reference signals. If the controller can establish contingencies that
entail a net decrease in error in the controllee, the chances for
successful control are best.
--------------------------------------
But what method do you offer in its place? Do we just let everyone
happily go about controlling their own perceptual variables?
As far as I know, there is no alternative to controlling: that is what
behavior is. We can't keep others from controlling their own perceptual
variables, because that's how they work, all of the time.
The simplest way to control another person's behavior is to apply a
disturbance to a variable that the other person is controlling. The
action that the other person takes to resist the disturbance and keep
the controlled variable from being altered away from its momentary
reference state is the behavior you can control in this way. If you
apply a small disturbance, the action will change by a small amount so
as to have an equal and opposite effect on the controlled variable. If
you want a larger action, you increase the amount of disturbance. If you
want an action in a different direction, you change the direction of the
disturbance. As long as producing the action doesn't cause the other
person any problems, you can create any amount and direction of action
by the other person that you please. You can use the rubber-band demo in
B:CP to illustrate this effect. Just don't try to make the person move
the hand into contact with something hot or sharp. To make this control
precise, don't apply so much disturbance that the other person loses
control of the controlled variable. The point is not to change the
controlled variable, but to elicit the action that keeps the controlled
variable from changing.
There are other methods, some innocuous and some not, but another method
of interest is to establish a contingency. You say "If you'll give me
your tickets to the Bears game, I'll give you $100 each for them" ($125,
$150, ... sold). This is explicit bargaining, in which each participant
attempts to satisfy his own goals by offering to help satisfy the
other's. There is also implicit bargaining: if you'll just be nice to
me, I'll stop complaining about how early I have to go to bed. This is
done simply by behaving as if the contingency were in effect; nobody has
to say anything about it. This is how the experimenter bargains with the
rat, without explicitly discussing it. The experimenter says, by
programming the apparatus, "If you'll press the bar, I'll give you one
pellet of food for every 20 presses." This doesn't cause the rat to
press the bar, or guarantee that there will be either presses or
delivered pellets, but it establishes one side of the bargain, in case
the rat finds it advantageous to sign on.
As you can see, the PCT way of looking at behavior control does not take
anyone's side: it simply recognizes that all organisms are controllers
and act to make their own experiences of the world match what they want
them to be. It recognizes that some controllers are smarter and more
powerful than others, but it points out the consequences of causing too
much error in other control systems. It reminds us that if control
entails the creation of important and persistent errors in the
controllee, the controllee will begin to reorganize and may learn how to
counteract the control; at the very least, the reorganization will make
the controllee progressively less predictable.
The real question is not whether there will be control, but how it will
be brought about.
Precisely. And what the consequences of succeeding will be.
Actually, I have great respect for the New Guinea forest native. So
give him (or her) a lab coat--I could use the help! (:->]
Damn, why didn't we think of that?
-----------------------------------------------------------------------
Best,
Bill P.