Learning;Red Sox; theories and stories

[From Bill Powers (960607.0600 MDT)]

Jeff Vancouver 960606.12:30 EST --

RE: first question

     This is a nice example of a positive feedback loop in a results
     (system dynamic's) model. It is also an example of several
     negative feedback loops in a control theory sense, although, these
     can only be speculations. ... I could do this with every one of
     your questions, but I think you get the point.

Good, an extra-credit answer. I do get the point, which is that your
understanding of the control model goes past the words you happened to
use in the post that made me wonder.

     Assuming that a person wants to achieve/maintain a goal (usually a
     performance goal provided by some organization/researcher), the
     output function must be organized such that the person can reduce
     an error. Over the long run, for any system in the perceptual
     hierarchy (not the intrinsic one), information about the state of
     the environmental variable will have different influences on the
     control system. Once a control system is established and well-
     functioning (unlike John's above) information that indicates a
     disturbance, results in an error which is corrected by some
     behavior(s).

I think it will pay to maintain a clear distinction between learning and
performance. Learning is a control process of a _different kind_ from
the control process that is learned. The underlying goal of learning, I
suggest, is not the same as the goal of the process that is learned.

As a simple example, consider a rat learning to press a lever to obtain
food pellets. The rat would have no natural interest in making little
brown objects rattle into a dish. The hunger does not represent a
deficiency in food pellets; it represents an error in some critical
internal state, for example a state of fullness or of the available
energy levels. The rat is not trying to learn how to obtain food
pellets; the error that motivates the learning is not specific to food
pellets. ANYTHING that resulted in reduction of the fullness or
nutritional state error would be learned (if possible). To say that the
naive rat is learning how to obtain food pellets is to imply that the
rat already knows that if it could make those little brown things appear
in the food cup, it could reduce the (for short) hunger error.

As Hugh Petrie pointed out (1981, _The dilemma of enquiry and learning_,
U. of Chicago Press), learning involves the Meno Paradox of Socrates.
Socrates said to Meno

     You argue that a man cannot enquire either about that which he
     knows or about that which he does not know; for if he knows, he has
     no need to enquire, and if not, he cannot; for he does not know the
     very subject about which he is to enquire.

What the rat already knows it does not have to learn. If the rat has
already had experience with food pellets, then the goal of learning
could be said to involve learning how to find food pellets. But then
what is learned would not be that getting food pellets would help
assuage hunger, but that performing a specific kind of act would cause
food pellets to appear. The rat does not know in advance what act will
produce food pellets; it must be able to try different acts until it
sorts out which acts (if any) produce food pellets and which do not. The
reason that learning is needed is that the rat does not know what act
will have the desired result. Thus when we say that the rat is learning
the acts necessary to produce food pellets, we imply a contradiction:
that the rat already knows what acts it must learn to perform, and yet
is trying to learn them.

Consider all this in the context of your comment:

     On the other hand, when the system is not well-functioning (e.g.
     novel task). That is, outputs do not reduce error, reorganization
     must take place. In this case information that some behavior
     resulted in a reduction in error would, I hypothesis and model,
     result in halting or at least slowing down reorganization such that
     the output function holds the values that "work." Information that
     results in a reduction in error is what you (and me) would call
     negative feedback, but if that information comes from an external
     agent, it is called positive feedback in our literature (because
     the agent is able to say "yes, that's it" or something
     like that).

What you seem to be saying here is that some error exists, but in the
case of interest that the system has no way of knowing that the error
has been reduced until told by an external agent. This implies that the
error is not an error between the learning system's own perceptions and
its own goals, but something existing in its environment. Only in that
way could an external agent know that there has or has not been a
reduction in error, while the acting system does not know. In fact, what
you are describing is control organized around an error that the
external agent experiences. In other words, it is the external agent
that is learning what to tell the acting system in order to control the
external agent's perceptions.

I think we need to be careful when using the term "reorganization." What
we commonly call learning does not always entail reorganization.

Suppose you're the one who is learning. You would like to make money,
and you do not have as much money as you want (both being prerequisites
to what follows). You are told that if you press the correct lever, you
will get $100, but if you press the wrong one you get nothing. In either
case, you get only one try. Since there are only two levers, what is to
be learned (which lever to press) is trivially simple, yet you can't be
sure of learning it correctly.

If an external agent informs you that lever A will produce the money,
your problem is solved: you press lever A, and get the money. I think
this is more or less the scenario you are thinking of.

But is this learning? Here the only thing that prevents you from
producing the right act is a lack of information about the effects of
the two levers. When the missing information is provided, learning is
instantly complete (if you believe it, and it is right). In fact, you
didn't have to learn HOW to do any act you didn't already know how to
do. No reorganization was involved.

So when would reorganization be involved? Suppose the two levers were
marked with symbols: a circle and a plus mark. The external observer
says "Pressing the time-out symbol will produce the money." This is like
a certain word-processor manual, which says "Press the Help key to get
context-sensitive help." The problem is, there is no help key on a
standard keyboard; the help key is actually (as a search through the
manual eventually shows) labeled "F1". So being told to press the time-
out symbol to get the money is of no use at all because you don't know
whether the circle or the plus mark is called the time-out symbol. Now,
if it were possible, you would have to reorganize to solve the problem
-- start pressing levers at random until the money appeared, and
discover for yourself how to correct the error. Of course under the
rules as stated you'd have only one try, and a 50% chance of getting the
money.

     Thus, I am trying to show how, why and when positive and negative
     feedback (in psychology's use of the terms) is useful using a
     negative feedback model of behavior (perception control).

That's a worthy objective, but considering the nuances above, it may be
that the psychological use of these terms rests on a rather sketchy
analysis of the phenomenon to be explained. And anyway, I still feel
that there's a certain legitimacy in saying that positive and negative
feedback already had specific technical meanings before the terms were
picked up by psychologists, and that to alter those meanings arbitrarily
is an unwise and unfriendly act. Put the shoe on the other foot: What if
I had decided that "motivation" was to mean an action that causes
something to move? Don't you think that psychologists might have
objected to this usurpation of one of their favorite words to mean
something that is not at all what they intended the word to mean? And
wouldn't we have a terrible problem in communication if both camps
continued to use the word as they define it, without regard for its
meanings to the other side?

"The rat motivated the lever to produce food." That's roughly how "the
teacher provided positive feedback to the student" sounds to me.

···

-----------------------------------------------------------------------
Rick Marken (960606.1030) --

Your answers to how PCT might be applied to masochistic Red Sox fans
were more useful than mine, and also far more interesting.
-----------------------------------------------------------------------
Bruce Gregory (960606.1325 EDT) --

     If the control theory dilettante answer given above is the best PCT
     can do, PCT doesn't have anything very interesting to say about Red
     Socks Fans. The persistence of masochism may be perfectly
     consistent with PCT, but then so is Copernican model of the solar
     system.

That's the point. PCT by itself is just a tool, like algebra. What does
algebra have to say about the problem of conserving fossil fuels? To
answer "Nothing" would be quite incorrect, but it would also be foolish
to expect to find the answer by looking in an algebra textbook. Algebra
is not about fossil fuels, but about relationships among variable
numerical quantities. Somebody has to formulate an analysis and get some
data before algebra can help with this problem. Same with PCT and the
Red Sox: somebody has to propose some controlled variables, interview
fans, manipulate what is manipulable, and so forth -- and THEN apply the
principles of PCT. PCT is about relationships among variables in a
system with strong stable negative feedback. Those relationships say
nothing about the Red Sox or their fans -- until someone does the work.

     Applying PCT to interpersonal conflicts can be illuminating and can
     suggest solutions of the type Ed Ford has developed. Apparently
     applying PCT to understanding the appeal of spectator sports is not
     nearly so fruitful. This confirms my initial reaction, but is
     disappointing never the less.

Ed started with a considerable data base from his practice with
individuals and youthful offenders. He applied the principles of PCT to
his experiences to help him understand what was going on. Studying the
appeal of spectator sports would require an equal amount of study if PCT
were to be applied profitably. This surely could be done. But it
probably wouldn't help the Red Sox much.
-----------------------------------------------------------------------
Martin Taylor 960606 14:30 --

Remember the motto of the National Reorganization Association (NRA):
Theories don't explain phenomena; People explain phenomena.

     Very wise. But like the motto of the other NRA, a half-truth.
     People without theories don't explain phenomena nearly as easily as
     do people with theories, and some theories explain more readily
     than others. People with some theories can divine implications that
     people with other theories cannot, just as some of the weapons
     favoured by that other NRA can do their work much more easily than
     can others.

My point, which I didn't get around to making, is that while any theory
can provide a plausible-sounding explanation of any phenomenon (in the
hands of the verbally adept), there are ways of testing theories that
can show that some theories are a lot better explanatory tools than
others. The beauty and persuasivess of a theory are seldom an adequate
test of its usefulness.

     "What stories we tell ourselves, of origins and endings, of form
     and transformation, of gods, the word, and law."

Indeed.
-----------------------------------------------------------------------
Martin Taylor 960606 16:00 --

     If PCT is correct, all purposive action is to control some
     perception(s).

This is a good way to put it, although as Rick explained "purposive
action" and "control" (or at least "attempted control") are pretty much
synonymous in PCT. A commonly misunderstanding of "all behavior is aimed
at controlling perceptions" is "all perceptions are controlled by
behavior." Your way of saying it puts the emphasis on accounting for
actions in terms of the perceptions they are produced to control, which
is better.

The application to Red Sox fans is obvious. What isn't clear is what
perceptions are under control (or are at least intended to be
controlled, if not controlled very well). As you say, some experiments
would be in order.
---------------------------
P.S. Very sorry you won't be at the meeting, but thanks from all of us
to you for your most generous contribution to the CSG!
----------------------------------------------------------------------
Ken Hacker (960607) --

     Would it be possible for Bill to help someone study why a voter
     selects on presidential candidate over another? If yes, I would
     love to get involved with that kind of PCT application into another
     field of inquiry.

An ambitious and worthy undertaking. Yes, I would be glad to help. Maybe
you'd like to come up to my place (a day's drive) for a few days to talk
it over.
-----------------------------------------------------------------------
Best to all,

Bill P.

[From Bruce Gregory (960607.1220 EDT)]

Bill Powers (960607.0600 MDT)

Responding to (Jeff Vancouver 960606.1330 EST)

What you seem to be saying here is that some error exists, but in the
case of interest that the system has no way of knowing that the error
has been reduced until told by an external agent. This implies that the
error is not an error between the learning system's own perceptions and
its own goals, but something existing in its environment. Only in that
way could an external agent know that there has or has not been a
reduction in error, while the acting system does not know. In fact, what
you are describing is control organized around an error that the
external agent experiences. In other words, it is the external agent
that is learning what to tell the acting system in order to control the
external agent's perceptions.

I have never seen a more forceful or succinct statement of the
paradox of teaching. We developed a high school text in which each
chapter began by asking students to make a prediction. The chapter
revealed that often that prediction was wrong. The approach was
taken because a co-worker has superb intuitions about what works and
what does not in school settings. I now see that we were trying to
create a situation in which the student's predictions and subsequent
reading would establish an error signal (in the student) which the
student could seek to reduce. I also see why it is futile to try to
convince students of something (such as the validity of Newton's
third law). We are trying to reduce _our_ error signals, not theirs.

To me:

Ed started with a considerable data base from his practice with
individuals and youthful offenders. He applied the principles of PCT to
his experiences to help him understand what was going on. Studying the
appeal of spectator sports would require an equal amount of study if PCT
were to be applied profitably. This surely could be done. But it
probably wouldn't help the Red Sox much.

Oh ye of little faith.... Nonetheless, I take your point. I wanted
to run before I learned to crawl...

Regards,

Bruce G.

[from Jeff Vancouver 960610.10:30 EST]

First, let me say that I come into the office (and thus check e-mail) two
days week. Just if anyone wondered about my lag.

[From Bill Powers (960607.0600 MDT)]

I think it will pay to maintain a clear distinction between learning and
performance. ...

Yes, this perhaps the fundamental point I make.

Consider all this in the context of your comment:

me:

     On the other hand, when the system is not well-functioning (e.g.
     novel task). That is, outputs do not reduce error, reorganization
     must take place. In this case information that some behavior
     resulted in a reduction in error would, I hypothesis and model,
     result in halting or at least slowing down reorganization such that
     the output function holds the values that "work." Information that
     results in a reduction in error is what you (and me) would call
     negative feedback, but if that information comes from an external
     agent, it is called positive feedback in our literature (because
     the agent is able to say "yes, that's it" or something
     like that).

Bill:

What you seem to be saying here is that some error exists, but in the
case of interest that the system has no way of knowing that the error
has been reduced until told by an external agent. This implies that the
error is not an error between the learning system's own perceptions and
its own goals, but something existing in its environment. Only in that
way could an external agent know that there has or has not been a
reduction in error, while the acting system does not know. In fact, what
you are describing is control organized around an error that the
external agent experiences. In other words, it is the external agent
that is learning what to tell the acting system in order to control the
external agent's perceptions.

Yes and no. There is definitely an error in the external system that is
responsible for the external system's action. But that does not preclude
that there is error in the focal system - the system I am studying.
(Given that in my experiment the external system is a computer program
with an IF-THEN statement, I _know_ that the external system is a control
system attempting to reduce its "error."). It seems you got distracted
by this external system. The point is that I set up a situation in which
the focal system will detect a difference between the current state and
the desired state, the system will be allowed to act, and the system will
be allowed to get new information about the current state, following that
act. That some external system was involved in providing that
information should be of little interest to understanding the behavior of
the focal system (as I have so far manipulated the situation).

I think we need to be careful when using the term "reorganization." What
we commonly call learning does not always entail reorganization.

Suppose you're the one who is learning. You would like to make money,
and you do not have as much money as you want (both being prerequisites
to what follows). You are told that if you press the correct lever, you
will get $100, but if you press the wrong one you get nothing. In either
case, you get only one try. Since there are only two levers, what is to
be learned (which lever to press) is trivially simple, yet you can't be
sure of learning it correctly.

If an external agent informs you that lever A will produce the money,
your problem is solved: you press lever A, and get the money. I think
this is more or less the scenario you are thinking of.

Less than more, but I would not argue that the experiment is trivially
simple. Part of the business I am in involves demonstrating trivially
simple things. THe other part is interpreting the results of a mass of
these trivially simple things.

>

But is this learning? Here the only thing that prevents you from
producing the right act is a lack of information about the effects of
the two levers. When the missing information is provided, learning is
instantly complete (if you believe it, and it is right). In fact, you
didn't have to learn HOW to do any act you didn't already know how to
do. No reorganization was involved.

I do not agree. An output function looks different after the learning
you describe has taken place. It is your word, but is not changes to
functions reorganization?

>

So when would reorganization be involved? Suppose the two levers were
marked with symbols: a circle and a plus mark. The external observer
says "Pressing the time-out symbol will produce the money." This is like
a certain word-processor manual, which says "Press the Help key to get
context-sensitive help." The problem is, there is no help key on a
standard keyboard; the help key is actually (as a search through the
manual eventually shows) labeled "F1". So being told to press the time-
out symbol to get the money is of no use at all because you don't know
whether the circle or the plus mark is called the time-out symbol. Now,
if it were possible, you would have to reorganize to solve the problem
-- start pressing levers at random until the money appeared, and
discover for yourself how to correct the error. Of course under the
rules as stated you'd have only one try, and a 50% chance of getting the
money.

What you describe is reorganization of the input function and some higher
level output function (or is it the comparator function?). I have found
this type of reorganization much more difficult to understand (so I guess
it is not trivial), because it seems that two functions must reorganize
simultaneously and in sync. I would like to move on to it some day.

me [sorry I messed up the formatting]: > Thus, I am trying to show how,

why and when positive and negative

feedback (in psychology's use of the terms) is useful using a > negative

feedback model of behavior (perception control).

Bill: > That's a worthy

objective, but considering the nuances above, it may be > that the
psychological use of these terms rests on a rather sketchy > analysis of
the phenomenon to be explained. And anyway, I still feel > that there's a
certain legitimacy in saying that positive and negative > feedback already
had specific technical meanings before the terms were > picked up by
psychologists, and that to alter those meanings arbitrarily > is an unwise
and unfriendly act. ...

new me:
I have no quarrel with who defined the terms first. But you seem to imply
that I am responsible for defining them in psychology - lo to have that
much power. I am merely using the terms that they use to describe what I
am doing in relation to their research questions. If they had other
words, I would have used them (it would have saved us much of this
discussion for instance). Terminology is one of the most common sources
of problems as we (scientists, people) try to communicate with each
other. But you know that.

I am still working on the simulation. It would facilitate discussion if
you could see this. Once I do get it working, it requires Windows 3.1 or
higher (and would be better Zipped). Do you have these programs?

I will also post a good deal of the code later today if I can find a
reasonable way to do that (cut and paste does not seem to be working).
But the program is meant to demonstate the model, so it is serially run
and will be difficult to follow. Maybe you all have a better mind for
this. I find I cannot follow the PASCAL programs you all post.

Later

Jeff

[from Jeff Vancouver 960613.13:45]

[From Bill Powers (960607.0600 MDT)]

What you seem to be saying here is that some error exists, but in the
case of interest that the system has no way of knowing that the error
has been reduced until told by an external agent. This implies that the
error is not an error between the learning system's own perceptions and
its own goals, but something existing in its environment. Only in that
way could an external agent know that there has or has not been a
reduction in error, while the acting system does not know. In fact, what
you are describing is control organized around an error that the
external agent experiences. In other words, it is the external agent
that is learning what to tell the acting system in order to control the
external agent's perceptions.

I think Bruce G.'s issues nicely illustrates exactly the kinds of
situations that meet the criteria of an external agent as critical for
connecting the loop, that is, an educational one. The basic question is
nicely summed up in your last statement. We need to figure out how to set
up the environment so that the student learns how to control it
perceptions. I am that external agent, or at least wanting to give
advice to that external agent.

I would add one more very sticky piece, the educator also tries to make
sure that the perceptions that he/she is trying to get the student to
learn to control is the perception her/she wants the student to control.
But that is another problem we can discuss at another time.

I also want to return to the other points you brought up in this post
(e.g., reorganization), but I think it is prudent to wait until you get
the simulation up and running.

Later

Jeff