Descriptions, theories, and facts

[Bruce_Abbott]

[From Bruce Abbott (961124.1140)]

Bill Powers (961123.1345 MST)

Bruce Abbott (961123.1405 EST)

That's why I want to leave out the "functional:" it asserts a causal
relationship by saying that operant rate is a function of Y.

More terminological confusion. One can state a mathematical function
describing a relationship between X and Y without asserting any particular
scheme of causality. The pressure of a gas in a confined area varies
according to a definite function with its temperature, but that doesn't
assert any particuar scheme of causality. It's just a description of the
observed relationship.

One of my objections to the Skinnerian formulation is that it is NOT
mathematically constructed.

I'm sorry, I missed the part where you indicated your agreement or
disagreement with the point I was arguing for. Instead, you launch another
argument. Do you agree or disagree with the above?

As to your new argument, Skinner may not have pursued a mathematical
formulation of the relationships he observed, but there are plenty of
examples of such formulations in the literature. But this gets us off on
another tangent. I'm still trying to get us to the point where you can
verbalize the distinction between a descriptive (functional) approach and a
mechanistic one. Thus far I've seen no evidence that you appreciate the
difference.

But I am not stating anything about how process X works -- it is only a name
for some unspecified something which I infer to be at work to produce,
somehow, the observed relationship.

That's the problem, isn't it? You have a feeling that there's an
"unspecified something" that somehow is producing the observed
relationships. It's just got to be in there somewhere. Yet no way of saying
what it is quite says what you're trying to get across.

The whole point, Bill, is that I don't have to. It's not a requirement of
this approach.

I'm perfectly willing to keep on considering your various ways of trying to
describe X, but I think you must seriously consider a possibility that would
short-cut the whole discussion: THE PROCESS X YOU'RE TRYING TO DESCRIBE
DOESN'T EXIST IN THE PLACE WHERE YOU'RE LOOKING.

No, Bill, I'm not looking for Process X -- I'm describing a widening set of
empirical relationships -- laws of behavior -- that Process X, if anyone
ever discovers what it is, will explain.

There is an inference that a process is at work which is responsible for
this relationship, but the causal structure of this process has been left
completely open.

It hasn't been left "completely open" if you call process X "reinforcement"
and say that the food pellets are "reinforcers." It's not open if even you
just say that the food pellets are "process-Xers." When you say that, you're
fastening on one of the observable variables as being responsible for
initiating process X, just because it's involved in a loop in which process
X is said to occur.

But Bill, I haven't called process X "reinforcement," and I haven't called
the pellets "reinforcers." (Those terms are holdovers from Pavlov and an
older theory in which all behavior was considered to consist of S-R
reflexes.) I haven't identified anything at all about "process X," except
that something about the delivery of the pellets is leading to a higher
sustained rate of lever-pressing than is observed when the pellets are
omitted, and that some property of the pellet itself appears to be
important, because other objects don't have this effect. I have merely
given a name to this mysterious phenomenon ("process X"). I have said no
more about this "process X" than I have said about the "process X" that
keeps the gasolone engine of my analogy running.

Suppose this process X that you're looking for doesn't take place in the
observable variables at all. Suppose it takes place in the organism, with
the observable variables being passive and totally dependent on this process
in the organism.

I haven't made any statements whatsover about where this process X takes
place, certainly not that it takes place in the observable variables. (But
with Skinner and probably everyone else in the field, if asked I would state
my certain belief that the process takes place in the organism.)

If, like Skinner, you're convinced that all causes of behavior must be
traceable back to the observable environment, then you have only one
reasonable candidate for the causes of the phenomena we see.

You have to understand that Skinner's "causes" and your "causes" are
different concepts. For Skinner (and many others), "casue" only means that
changing the value of some observable variable in the environment is
reliably followed by a particular change in observable behavior. The
physical linkages within the organism through which this observable
relationship emerges are assumed to be, well, within the organism. To
return to my analogy, my discovery that the engine's RPM depends on the
throttle setting does not lead me to believe that the position of the
throttle per se is the cause (as you understand that term) of the observed
RPM; I know very well that the throttle lever is altering something within
the engine to bring about this relationship, though I may have no idea how.

The only way to leave the question completely open is to describe the
observations and leave it at that until you have a model that can explain them.

Nothing I have said about "process X" presupposes a particular model of the
mechanism whereby "process X" works. "Process X" is a phenomenon to be
explained, and to be used to explain other observations.

I hope I can make this clear by analogy.

Your analogy works fine as long as it is only the engine you're considering.

Have I learned anything of value about the engine? Can I use this
information in practical applications?

When you bring in the driver, very little of what you have said about the
engine continues to have the same meaning.

You've changed the example without first telling me what "works fine" about
the analogy as presented. "Works fine" is very short admission to what I
see as a very important point -- one could almost miss it. Please elaborate
on it. (See questions immediately above.)

One of the things you might have
observed is that when the engine powers the car downhill, it consumes less
fuel than when the car goes uphill. You can show that this occurs because
the throttle varies its setting; less open when going downhill, more open
when going uphill. So you might come up with a process X which gives the
slope of the road an effect on the throttle setting -- purely a description,
of course, without attributing any causation to the slope. But if you then
started referring to uphill slopes as throttle-openers, you would be off on
a wild goose chase, because the real cause of this phenomenon is the
driver's intention (or the cruise-control's) to keep the speed constant. If
the driver liked to dawdle up hills and zoom down them, you'd see the
opposite relationship between slope and throttle opening. It has nothing to
do with the engine even though the engine is intimately involved in the
phenomenon.

But it has everything to do with the new system you've created. We're no
longer talking about the engine per se, but about a more elaborate system
including the cruise control. What your example brings out (quite
correctly) are the limitations of a purely descriptive, functional approach.
This is, of course, why a mechanistic approach is to be greatly preferred,
whenever someone has the experience and the insight to come up with a
testable mechanism that behaves in accordance with the laws discovered
through the empirical approach.

The same goes for process X. The relation we see between the contingent
event and the pressing is not produced by anything about that relationship;
it involves the organism doing the pressing, and a host of variables outside
the scope of these observations.

Of course! I'm getting the strong impression that you don't see why I agree
with you here, and are surprised that I do. But I'm not sure I can make it
any plainer.

Regards,

Bruce

[Bill_Powers1]

[From Bill Powers (961124.1500 MST)]

Bruce Abbott (961124.1140)--

One can state a mathematical function
describing a relationship between X and Y without asserting any particular
scheme of causality. The pressure of a gas in a confined area varies
according to a definite function with its temperature, but that doesn't
assert any particuar scheme of causality. It's just a description of the
observed relationship.

One of my objections to the Skinnerian formulation is that it is NOT
mathematically constructed.

I'm sorry, I missed the part where you indicated your agreement or
disagreement with the point I was arguing for. Instead, you launch another
argument. Do you agree or disagree with the above?

I agree with your generalization, but I do not agree that it applies to the
arguments concerning reinforcement, because those arguments are not stated
in sufficiently unequivocal terms to permit classing them with either
mathematical functions or physical theories. I should have made this clear
in my comment.

As to your new argument, Skinner may not have pursued a mathematical
formulation of the relationships he observed, but there are plenty of
examples of such formulations in the literature.

Yes, I have seen them. All of them assume the correctness of Skinner's basic
formulation. Perhaps there have been a few exceptions, but that doesn't
constitute a reform of behvioristic logic.

But this gets us off on
another tangent. I'm still trying to get us to the point where you can
verbalize the distinction between a descriptive (functional) approach and a
mechanistic one. Thus far I've seen no evidence that you appreciate the
difference.

It's my strong impression that what you call the functional approach
involves postulating a direction of dependence (not, as you say, a "mere
correlation") and thus a causal relationship. Perhaps this isn't necessary,
but anyone who speaks of "stimulus control of behavior" is definitely
proposing a causal relationship.

There is an inference that a process is at work which is responsible for
this relationship, but the causal structure of this process has been left
completely open.

It hasn't been left "completely open" if you call process X "reinforcement"
and say that the food pellets are "reinforcers." It's not open if even you
just say that the food pellets are "process-Xers."

But Bill, I haven't called process X "reinforcement," and I haven't called
the pellets "reinforcers." (Those terms are holdovers from Pavlov and an
older theory in which all behavior was considered to consist of S-R
reflexes.) I haven't identified anything at all about "process X," except
that something about the delivery of the pellets is leading to a higher
sustained rate of lever-pressing than is observed when the pellets are
omitted, and that some property of the pellet itself appears to be
important, because other objects don't have this effect. I have merely
given a name to this mysterious phenomenon ("process X"). I have said no
more about this "process X" than I have said about the "process X" that
keeps the gasolone engine of my analogy running.

Bruce, that's just not so: you specifically said in a post only two days ago
that you called the contingent events "process-Xers." Maybe you want to take
that back now; if so, OK. But then we also rule out calling pieces of food
"reinforcers." The idea that there is something about delivery of the
pellets that is causing -- excuse me, leading to -- the higher sustained
rate of pressing is not at all the only interpretation you can put on this.
For example, I would say that it is something about the organism that leads
it to focus its physical activities on the lever that produces pellets, and
when the production is confirmed, to cease to apply its efforts elsewhere.

Actually, you can rule out the _delivery_ of the pellets as your proposed
critical variable. They could be delivered in any number of ways with the
same effect, as long as the lever pressing was the signal for their
appearance. What you are really saying that it is something about the
_pellets_ that makes the difference. I say it is something about the
_organism_ that makes the difference.

Suppose this process X that you're looking for doesn't take place in the
observable variables at all. Suppose it takes place in the organism, with
the observable variables being passive and totally dependent on this
process in the organism.

I haven't made any statements whatsover about where this process X takes
place, certainly not that it takes place in the observable variables. (But
with Skinner and probably everyone else in the field, if asked I would state
my certain belief that the process takes place in the organism.)

But the implication is still (see my earlier post this afternoon) that you
can characterize process X entirely in terms of the variables observable in
operant conditioning experiments. That is clearly not possible.

If, like Skinner, you're convinced that all causes of behavior must be
traceable back to the observable environment, then you have only one
reasonable candidate for the causes of the phenomena we see.

You have to understand that Skinner's "causes" and your "causes" are
different concepts. For Skinner (and many others), "cause" only means that
changing the value of some observable variable in the environment is
reliably followed by a particular change in observable behavior. The
physical linkages within the organism through which this observable
relationship emerges are assumed to be, well, within the organism. To
return to my analogy, my discovery that the engine's RPM depends on the
throttle setting does not lead me to believe that the position of the
throttle per se is the cause (as you understand that term) of the observed
RPM; I know very well that the throttle lever is altering something within
the engine to bring about this relationship, though I may have no idea how.

The only way to leave the question completely open is to describe the
observations and leave it at that until you have a model that can explain

them.

Nothing I have said about "process X" presupposes a particular model of the
mechanism whereby "process X" works. "Process X" is a phenomenon to be
explained, and to be used to explain other observations.

"Process X" is ITSELF being offered as an explanation of the observable
relationships. The observable relationships do not contain any process X.
You have actions producing pellets in a certain way, under certain
conditions. The observations include the increase in behavior when it begins
producing pellets. There is nothing else going on that is observable. There
is no need, and no justification, for simply postulating a process X that is
capable of producing these observed results.

I hope I can make this clear by analogy.

Your analogy works fine as long as it is only the engine you're considering.

You've changed the example without first telling me what "works fine" about
the analogy as presented. "Works fine" is very short admission to what I
see as a very important point -- one could almost miss it. Please elaborate
on it. (See questions immediately above.)

I meant that it works fine as a description of studying the relation of fuel
consumption to the workings of a car engine. I didn't mean that I accept it
as a valid analogy of the situation with the organism.

In your tests of the car engine you were doing purely empirical research. If
you do this, that will happen. No Process X is required; you don't have to
say that the fuel "controls" the car engine's behavior or that it "increases
the probability" that car engine will go faster or slower. You simply say
"here is the chart of engine RPM in relation to throttle setting and fuel
consumption."

When I tried to point out that there may be relations among the variable
that are unsuspected when you consider only the car's engine, you said

We're no
longer talking about the engine per se, but about a more elaborate system
including the cruise control. What your example brings out (quite
correctly) are the limitations of a purely descriptive, functional approach.
This is, of course, why a mechanistic approach is to be greatly preferred,
whenever someone has the experience and the insight to come up with a
testable mechanism that behaves in accordance with the laws discovered
through the empirical approach.

But as I realized today, this also means that there is really no valid way
to talk about the "car engine per se" when it is always an inextricable part
of a larger system. What seem to be relationships among the observable
variables might not be due to the engine at all, but to effects which take
paths that go outside the universe of the car engine per se. What seems to
you a complete list of the necessary conditions is not actually a complete
list. You are looking at the system under very special circumstances without
knowing what is special about them.

The same goes for process X. The relation we see between the contingent
event and the pressing is not produced by anything about that relationship;
it involves the organism doing the pressing, and a host of variables
outside the scope of these observations.

Of course! I'm getting the strong impression that you don't see why I agree
with you here, and are surprised that I do. But I'm not sure I can make it
any plainer.

I feel like giving this particular argument a rest for a while. How about you?

As I look back over it, I see that a consistent thread running through it is
the assumption that if the rate of obtaining food pellets is increased by
lever pressing, the rate of lever pressing increases. In the light of what
we have been finding, is there any evidence that this phenomenon actually
occurs as you describe it? If there is any experiment in which lever
pressing actually changes in rate as a function of manipulated variables, we
should be studying it instead of all these cases where it doesn't seem to be
affected at all (contrary to the way these experiments are described in the
journals).

Best,

Bill P.

[T_B_Harms1]

From Tracy Harms (961124.19 MST)

Bill Powers (961124.1500 MST) in reply to Bruce Abbott (961124.1140)

As I look back over it, I see that a consistent thread running through it is
the assumption that if the rate of obtaining food pellets is increased by
lever pressing, the rate of lever pressing increases. In the light of what
we have been finding, is there any evidence that this phenomenon actually
occurs as you describe it? If there is any experiment in which lever
pressing actually changes in rate as a function of manipulated variables, we
should be studying it instead of all these cases where it doesn't seem to be
affected at all (contrary to the way these experiments are described in the
journals).

I'm pretty sure this assumption has *not* been made by Bruce. The limit of
his assertion was that there was a statistical increase in lever-pressing
when food pellet deliver is coupled with lever depression than when it is
not. Beyond this no additional association between pellet-delivery and
lever-pressing was claimed. And for good reason: things get rapidly less
tractable beyond this.

Tracy Bruce Harms tbh@tesser.com
Boulder, Colorado caveat lector!

[Bill_Powers1]

[From Bill Powers (961124.2000MST)]

Tracy Harms (961124.19 MST) --

I'm pretty sure this assumption has *not* been made by Bruce. The limit of
his assertion was that there was a statistical increase in lever-pressing
when food pellet deliver is coupled with lever depression than when it is
not. Beyond this no additional association between pellet-delivery and
lever-pressing was claimed. And for good reason: things get rapidly less
tractable beyond this.

I'll let Bruce explain my reference to our findings.

Best,

Bill P.

[T_B_Harms1]

[From Tracy Harms (961124.21 MST)]

Bill Powers (961124.2000MST)

[...]
I'll let Bruce explain my reference to our findings.

Best,

Bill P.

Fine. You've obviously poured a lot of time into your recent replies. But
I specifically thought that he meant what you say he meant, then changed my
mind on closer examination of his writings.

A claim of a general trend of increased lever-pushes with increase
food-pellet delivery would run contrary to some well-known Skinner-box
observations, as I'm sure you both know. Increased lever-pushing
correlates better with *decreases* in pellet delivery than it does with
increases. So to even identify a basic "process", as Bruce chooses to, the
comparison has to be between none vs some, not less vs more.

Tracy Bruce Harms tbh@tesser.com
Boulder, Colorado caveat lector!

[Bill_Powers1]

[From Bill Powers (961124.2355 MST)]

Tracy Harms (961124.21 MST) --

A claim of a general trend of increased lever-pushes with increase
food-pellet delivery would run contrary to some well-known Skinner-box
observations, as I'm sure you both know. Increased lever-pushing
correlates better with *decreases* in pellet delivery than it does with
increases.

I had that figured out long ago; the trouble is, I figured it out wrong. I
fitted a control model to a whole batch of data in which, over most of the
range, the measured rate of pressing decreased as the rate of reinforcement
increased. I was very pleased because this is just what control theory said
should happen and the model fit very well.

Early last year, Bruce Abbott became suspicious of some similar
pressing-rate data, because it seemed not to take into account the constant
time required to collect the reinforcer after completing the required number
of presses. These data purported to show how pressing rate depended on the
ratio requirement, the number of presses needed to get one reinforcer.

So Bruce recalculated the pressing rate after removing the time-delay due to
the collection and consumption of the reinforcer. It turned out to be
constant; the apparent relationship to the schedule disappeared. The rats
were pressing at a constant rate regardless of the ratio. The reinforcements
and the contingency neither increased nor decreased the pressing rate. This
was discussed on the net before you arrived.

That, of course, blew the control system model up; it was applied to a
non-phenomenon.

Bruce has found other data in which this same constancy of pressing rate
seems to exist; he is replicating a well-known experiment and so far finding
the same result in his own data, although the original experimenter never
discovered this effect. It's not known yet how far this finding will
generalize, particularly to interval schedules and variable schedules.

In the experiments Bruce is now carrying out and I am helping to analyze, we
have continuous data for the operant chambers from which we can time every
response and every reinforcement to 1/30 second. We have looked at the data
within sessions, although most of our effort has gone into a more global
analysis using daily totals. What we find in each daily experimental session
is that at the beginning, the pressing rate is very high, declining during
the session and reaching zero after something like 25 minutes out of the
hour, after which the rat does other things away from the lever. During this
initial 25 minutes, as video recordings reveal, the rats' method of
collection varies from hanging onto the bar while still chewing the previous
pellet and reaching into the food cup to pick the next pellet out of the
cup, to removing the pellet, sitting back, and nibbling until it's gone. So
any changes in rate of pressing within a session are due primarily to
changes in the speed of collection.

I feel that this result is sufficiently well-established that it requires
re-examining all experimental operant conditioning data. Unfortunately, raw
data have not been retained in most experiments in a form that would permit
removing the collection time effect from the reported average rates of
behavior. This means that an enormous body of experimental results must now
be called into question, and the experiments must be re-done to see if the
results are actually what they were reported to be. This isn't a question of
cheating; it's just that the possibility of a significant collection-time
effect on apparent rates or responding was not quantitatively investigated.
Unfortunately, the standard method of reporting behavior-rate data is to
divide total responses by total time during a session of one hour or more, a
method that not only includes collection time, but time during which the
animal is not even at the bar or key -- as I mentioned some time ago, it
included time, for all rats in our experiments, when a rat in the operant
chamber was actually asleep.

This, of course, applies to ALL the experiments Skinner used as the basis
for formulating his laws of reinforcement. They must all be done over again.

Best,

Bill P.

[Jeff_Vancouver]

[from Jeff Vancouver 961125.09:30 EST]

I am trying fairly successfully to not read or become involved in the
rewards debate. I am sure it would just frustrate me (that is what my
model says anyway). But I was curious about the literature noted below.
Is there some cite that can lead me to a review of this finding?

[From Tracy Harms (961124.21 MST)]

A claim of a general trend of increased lever-pushes with increase
food-pellet delivery would run contrary to some well-known Skinner-box
observations, as I'm sure you both know. Increased lever-pushing
correlates better with *decreases* in pellet delivery than it does with
increases. So to even identify a basic "process", as Bruce chooses to, the
comparison has to be between none vs some, not less vs more.

Thanks

Jeff

[T_B_Harms1]

From Tracy Harms (96;11,25.09:00 MST)

Jeff Vancouver (961125.09:30 EST)

As you might have noticed, Bill Powers' latest (961124.2355 MST) reports
that the findings I mentioned, which have become treated as something
rather close to "common knowledge" in psychological research, are now
highly suspect. Or, to quote Bill:

: I feel that this result is sufficiently well-established that it requires
: re-examining all experimental operant conditioning data. Unfortunately,
: raw data have not been retained in most experiments in a form that would
: permit removing the collection time effect from the reported average rates
: of behavior. This means that an enormous body of experimental results must
: now be called into question, and the experiments must be re-done to see if
: the results are actually what they were reported to be. [...]
: This, of course, applies to ALL the experiments Skinner used as the basis
: for formulating his laws of reinforcement. They must all be done over again.

My response, in a word: Wow!

Back to the drawing board...

Tracy Bruce Harms tbh@tesser.com
Boulder, Colorado caveat lector!

[Jeff_Vancouver]

[from Jeff Vancouver 961125.11:40 EST]

>From Tracy Harms (96;11,25.09:00 MST)
Jeff Vancouver (961125.09:30 EST)

As you might have noticed, Bill Powers' latest (961124.2355 MST) reports
that the findings I mentioned, which have become treated as something
rather close to "common knowledge" in psychological research, are now
highly suspect. Or, to quote Bill:

: I feel that this result is sufficiently well-established that it requires
: re-examining all experimental operant conditioning data. Unfortunately,
: raw data have not been retained in most experiments in a form that would
: permit removing the collection time effect from the reported average rates
: of behavior. This means that an enormous body of experimental results must
: now be called into question, and the experiments must be re-done to see if
: the results are actually what they were reported to be. [...]
: This, of course, applies to ALL the experiments Skinner used as the basis
: for formulating his laws of reinforcement. They must all be done over again.

My response, in a word: Wow!

Back to the drawing board...

The relationship you described that I found interesting was between a
decrease in pellets leading to an increase in behavior. This seems like
it is easily interpretable via a control theory model, not an S-R model.
This is the basis of my curiousity. That is, what are S-R people saying
to explain it and whether it can be used as data supporting the basic
control theory model.

I have learned to take the extreme statements in stride. I do wonder if
the interpretation of the data is reasonable given the design of the
studies. But questioning the conclusions for a series of studies
arising from a paradigm is status quo.

Later

Jeff

[Bill_Powers1]

[From Bill Powers (961125.1118 MST)]

Jeff Vancouver 961125.11:40 EST --

The relationship you described that I found interesting was between a
decrease in pellets leading to an increase in behavior. This seems like
it is easily interpretable via a control theory model, not an S-R model.
This is the basis of my curiousity. That is, what are S-R people saying
to explain it and whether it can be used as data supporting the basic
control theory model.

Actually, what has been found is that pressing rate neither increases nor
decreases with the various conditions. Apparently the rats just press as
fast as they can when they're pressing at all, or at some rate that is
comfortable. The rest of the time they're doing something else that isn't
being measured. There are other behavior measures that are more promising,
for example total food consumed per meal and number of meals per day.

Best,

Bill P.

[Bruce_Abbott]

[From Bruce Abbott (961125.1450 EST)]

Re: Bill Powers (961124.1500 MST), etc.

I'm not sure we're getting anywhere in this discussion. Perhaps it might
help if I boiled down what I am trying to get across to its essence:

1. A functional approach answers the question, what does it do?

2. A mechanistic approach answers the question, how does it do it?

3. When you have an answer to the first question, you have a limited
    ability to explain and predict. For example, if I can barely hear
    the audio on my TV set, I may explain this by noting that the volume
    control has been set to the low end, and I can predict that if I move
    the volume control toward the high end, that the sound will get louder.
    I say that this is limited because, for example, if moving the control
    does not help, I may be at a loss to explain why it doesn't help. On
    the other hand, this failure will itself become the impetus for further
    manipulations designed to discover what other variables are involved
    from which this outcome might have been successfully predicted.

4. When you have an answer to the second question, then you can construct
    a model of the system, enter starting parameters and conditions, crank
    the model, and it will behave as described in the answer to the first
    question. The model will probably also indicate conditions under which
    the answers given by the first approach no longer hold. Once you have
    an adequate answer to the second question, there is no further need for
    the first approach, except perhaps as a check on the model's predictions.

5. The answers to the first question assume no particular model of the
    underlying mechanism. The fact that rotating a particular little knob
    on the TV normally changes the sound volume in predictable ways presupposes
    no particular mechansim by which this linkage is accomplished. (It could
    operate through a variable resistor or a servomechanism, for all I know.)
    A proper model of the mechanism will, however, have to explain how rotating
    the knob changes the volume.

6. I am specifically NOT arguing that the functional approach is as good as
    the mechanistic one in explaining the observations (it isn't). I am NOT
    arguing that EAB researchers have always strictly followed the functional
    approach (they haven't). I am NOT arguing that the functional approach
    will identify all the crucial system variables (it won't). I am NOT
    arguing that the terms adopted to refer to specific identified phenomena
    (e.g. "reinforcement") or analytic components of the system (e.g.,
    "stimuli," "reponses") do not bias one's thinking toward a particular
    theory of causation (they do). I AM asserting that it is the only
    available way to proceed before there is enough information available to
    develop a model that has a reasonable chance of being close enough to
    reality to succeed, and that the results of such functional explorations
    are useful within the range of conditions under which the relationships so
    uncovered generally hold. I AM asserting that to undertake to treat a
    purely functional system of variables as if it represented a specific
    proposal of mechanism is a mistake.

Regards,

Bruce

[Bill_Powers1]

[From Bill Powers (961125.1750 MST)]

Bruce Abbott (961125.1450 EST)--

I'm not sure we're getting anywhere in this discussion.

Me too.

Perhaps it might
help if I boiled down what I am trying to get across to its essence:

1. A functional approach answers the question, what does it do?

Fine. I call this a description of observations. But we crash in flames
right at this point, because we disagree about what constitutes a
description. You introduce terms like "maintain" and "support" which to me
go beyond mere description.

Suppose we have the following two plots:

···

*
                                      *
Behavior *
rate *

           *
    ******
   --------------------------------------------------------------
           \
             contintency established

                                                               *
food *
rate *
            *
    -----*--------------------------------------------------------
          \
            contingency established

We get these curves when the animal is deprived of food, and when the lower
curve measures food intake and not sand intake. Some other conditions also
are required.

It seems to me that this is as far as we can go without a model. We can see
how the behavior causes the food to appear. We can't see any OTHER relation
between the behavior and the food intake. To claim that the food intake does
something to the behavior rate is totally unjustified by the observations.
We have only the one relationship to look at and we've already used it in
saying that the behavior is producing the food. As they say in systems
analysis, we've "used up" that relationship; we can't use it again, for
example by looking at the same relationship backward or in a different
mathematical form, like the inverse. We're stuck.

The reason we're stuck, as I tried to point out a while ago, is that we're
looking at only part of the physical system. We don't have all the data
needed for a complete description of the system. Our only recourse is to
physically investigate the rest of the system, to tie up all the loose ends,
or to propose a model of the rest of the system for testing. There's nothing
more we can do with the above observations alone.

In the behaviorist literature I've seen, time and time again, attempts to
set up system equations and solve them, using only observations like those
above. Without exception they amount to nothing more than deriving one
mathematical identity after another -- or else, equally trivially, using
illegal operations or simply making mistakes. Such manipulations are futile,
because they are trying to do the impossible: solve for the variables in a
system that has not been completely described. Both behavior rate and food
intake rate are dependent variables. You can't solve for their values until
you have the full description of the system, including the true independent
variables.

Skinner, whether he knew it or not, was going beyond his data when he said
that a reinforcer produces an increment in behavior. That relationship is
not observable. The only relationship that is observable is that an
increment in behavior produces an increment in reinforcement.

If you can't see this, Darth, then I'm afraid that the old Emperor has
clouded your mind.

Best,

Bill P.

[Bruce_Abbott]

[From Bruce Abbott (961126.1750 EST)]

Bill Powers (961125.1750 MST) --

Bruce Abbott (961125.1450 EST)

Perhaps it might
help if I boiled down what I am trying to get across to its essence:

1. A functional approach answers the question, what does it do?

Fine. I call this a description of observations. But we crash in flames
right at this point, because we disagree about what constitutes a
description. You introduce terms like "maintain" and "support" which to me
go beyond mere description.

I would use other terms, but my Newspeak vocabulary fails me: there appear
to be no words in Newspeak that convey these concepts. The best I can do is
invoke analogies. In the engine analogy, if I wished to maintain the engine
RPM above idle, I would need to open the throttle. In my old vocabulary, I
might have said that a higher throttle setting maintains the engine RPM at a
higher level, or that the higher throttle setting supports a higher RPM.
(This will be true only under given conditions, of course.) Because I have
shown this relationship empirically, I would hold that it just describes
what is observed to happen: the engine RPM tracks the throttle setting,
other things being equal. It says that throttle setting (A) and engine RPM
(B) are linked in some way such that, when A is set to a particular value, B
will change to a particular value, under those conditions in which the
relationship holds (e.g., the engine is running).

Suppose we have the following two plots:

                                                                *
                                     *
Behavior *
rate *

          *
   ******
  --------------------------------------------------------------
          \
            contintency established

                                                              *
food *
rate *
           *
   -----*--------------------------------------------------------
         \
           contingency established

We get these curves when the animal is deprived of food, and when the lower
curve measures food intake and not sand intake. Some other conditions also
are required.

It seems to me that this is as far as we can go without a model. We can see
how the behavior causes the food to appear. We can't see any OTHER relation
between the behavior and the food intake. To claim that the food intake does
something to the behavior rate is totally unjustified by the observations.

Now wait a minute. What you describe is apparent in the graphs _after_ the
contingency has been established, _by itself_. But we can see more. We can
see how the behavior causes food to appear in the contingent part of the
Graph 2, we can see that behavior does NOT cause food to appear in the
noncontingent part, and we can see that behavior occurs at a higher rate
during the contingency than before. The relationship between no contingent
food/contingent food and baseline rate of lever pressing/above baseline rate
of lever pressing is NOT used up by the observed relationship between press
rate and food rate shown in the contingent parts of the graphs; it is an
independent finding.

The reason we're stuck, as I tried to point out a while ago, is that we're
looking at only part of the physical system. We don't have all the data
needed for a complete description of the system. Our only recourse is to
physically investigate the rest of the system, to tie up all the loose ends,
or to propose a model of the rest of the system for testing. There's nothing
more we can do with the above observations alone.

There are a few additional things we can do with it. For one, if we see
behavior being maintained (sorry) at rates above those expected in the
absence of reinforcement, we can infer that there is probably a contingency
between the behavior and _something_ that is sustaining (sorry) these rates.
Another thing we can do is develop hypotheses about how these various
relationships might interact, and test them. For example, one might
hypothesize that pellet amount and level of deprivation will combine
multiplicatively in their joint relationship with rate of lever pressing.
We can come up with a whole network of such relationships that will enable
us to predict relative press-rates across a variety of situations.

What we _can't_ do merely from observing these relationships is understand
_why_ the observed relationships hold under the conditions in which they do,
or why they break down when they do. For that we need to specify a mechanism.

In the behaviorist literature I've seen, time and time again, attempts to
set up system equations and solve them, using only observations like those
above. Without exception they amount to nothing more than deriving one
mathematical identity after another -- or else, equally trivially, using
illegal operations or simply making mistakes. Such manipulations are futile,
because they are trying to do the impossible: solve for the variables in a
system that has not been completely described. Both behavior rate and food
intake rate are dependent variables. You can't solve for their values until
you have the full description of the system, including the true independent
variables.

Yes, I've seen those mistakes, too, and I understand what is missing. In
those cases the authors are usually attempting to move beyond a merely
descriptive system to one that specifies some sort of mechanism; they get
the mechanism wrong and end up doing something silly like deriving the
inverse of the EFF and concluding that this describes the organism.

Skinner, whether he knew it or not, was going beyond his data when he said
that a reinforcer produces an increment in behavior. That relationship is
not observable. The only relationship that is observable is that an
increment in behavior produces an increment in reinforcement.

If by that term "produces" Skinner meant a linea1 chain of causality running
from reinforcer to behavior rate, then he was wrong. If by "produces" he
meant only that the response-contingent delivery of certain events supports
a rate of responding above that observed when such events are not
response-produced, then he was just stating an empirically confirmed
relationship.

If you can't see this, Darth, then I'm afraid that the old Emperor has
clouded your mind.

Maybe so, Bill, maybe so. But who's that standing behind you .. why, it's
the Shadow . . .

Darth

[Bill_Powers1]

[From Bill Powers (961126.1730 MST)]

Bruce Abbott (961126.1750 EST)--

It seems to me that this is as far as we can go without a model. We can see
how the behavior causes the food to appear. We can't see any OTHER relation
between the behavior and the food intake. To claim that the food intake
does something to the behavior rate is totally unjustified by the
observations.

Now wait a minute. What you describe is apparent in the graphs _after_ the
contingency has been established, _by itself_. But we can see more. We can
see how the behavior causes food to appear in the contingent part of the
Graph 2, we can see that behavior does NOT cause food to appear in the
noncontingent part, and we can see that behavior occurs at a higher rate
during the contingency than before. The relationship between no contingent
food/contingent food and baseline rate of lever pressing/above baseline rate
of lever pressing is NOT used up by the observed relationship between press
rate and food rate shown in the contingent parts of the graphs; it is an
independent finding.

All right, let's consider that. Let's suppose that the contingency is F =
K*P, in any appropriate units for Food and Pressing.

What you have are two observed relationships (approximating the imaginary
curves I drew):

1. F = 0[zero] * P (before the contingency is established), and

2a. P := Po*(1 - e^-at), after it's established,
  from which it follows that
2b. F := K*Po*(1 - e^-at)

There is nothing in these descriptive equations to indicate that the Food
influences, supports, maintains, sustains, encourages, determines,
regulates, elicits, generates, promotes, provokes, actuates, brings about,
creates, effects, invokes, makes, produces, induces, inaugurates,
instigates, initiates, launches, originates, precipitates, or in short
causes the Pressing (or any other term you can find in a bigger thesaurus
that comes from "cause").

You might find that a or K is dependent on time of deprivation according to
some relationship. In that case you could substitute expressions containing
deprivation time (and whatever other variables might enter) for the values
of a and K. That would complicate the equations, but it still wouldn't yield
any expression for the dependency of P on F. You might find other variables
that depend on P, so you could add new relationships to the collection.
There might be, as you suggest, multiplicative relations so one variable
affected by P influences the effect of P on other variables. You might even
find some surprising relationships among the variables.

Once you have been engaged in this shuffling of mathematical relations long
enough, you might run across an expression containing P and F which no
longer bears any obvious relationship to the two equations above, and it
might occur to you that you can now solve the new expression for P as a
function of F. And indeed you could, with a little luck. You might think
that you have finally established the way in which the contingent food
affects, influences, directs, instigates .... causes the pressing rate. This
would be a complete mathematical blunder. All you will have done will be to
have found a very complex way of writing the inverse of equations 1 and 2.

When you're all done with this shuffling, there will still be an unexplained
observation: Equation 2a, and its companion, Equation 2b. Unless a new
explicit relation between P and F, independent of these equations, can be
brought into the system, there is simply no way to account for the behavior
of P. That exponential rise of P to an asymptote is caused by a part of the
physical system that hasn't been represented by any equations derived from
observations outside the organism. Without knowledge of what is going on
inside the organism (or an informed guess), the behavior of P is simply
inexplicable.

There are a few additional things we can do with it. For one, if we see
behavior being maintained (sorry) at rates above those expected in the
absence of reinforcement, we can infer that there is probably a contingency
between the behavior and _something_ that is sustaining (sorry) these rates.

Saying "sorry" doesn't justify using unjustifiable language. Yes, something
is "sustaining" the behavior. But it's not the food: according to PCT, it's
a reference signal. The reference signal is one of the true independent
variables in this system of which we see only one part from outside it. And
the real explanation of the apparent effect of creating a contingency has to
go beyond even that, because why, when food is not produced by a small
amount of behavior, does the behavior not increase enormously in the attempt
to produce it? Only because some still higher system recognizes that there
is no relation between that action and that desired result, and turns to
some other action that might produce it. The only valid ways to characterize
the relation between the pressing and the food require describing the
remainder of the whole physical situation, which lies inside the organism.

If by that term "produces" Skinner meant a linea1 chain of causality running
from reinforcer to behavior rate, then he was wrong. If by "produces" he
meant only that the response-contingent delivery of certain events supports
a rate of responding above that observed when such events are not
response-produced, then he was just stating an empirically confirmed
relationship.

I'm not getting across, Bruce. There's no way to use the word "produces" or
any of the other synonyms for "causes" that doesn't assert an unobserved
connection from the food to the pressing. By using such a term, you are
asserting the existence of such an unobserved connection, and this takes you
out of the world of pure empiricism and into the world of model-building --
whether you want to be there or not, and whether you admit being there or
not. By the rules of system analysis, you are then obligated to propose a
form for this connection, because merely to say that there IS a connection
is to say nothing of any use.

Skinner, in fact, did so. He tried to say what that connection was in a
number of ways. He said that an increment in reinforcement causes an
increment in behavior, which is wrong. He said that a reinforcement
increases the probability of the behavior that caused it, which is wrong. He
said that behavior is controlled by its consequences, which is wrong. Not
one of his statements about the effect of reinforcers on behavior can be
turned into a specific model that works; there's something wrong with each
of them, which can be cured only by bringing in auxiliary assumptions, if at
all. Perhaps someone who understands modeling could find a way to make one
or more of his propositions work. But he never did, and he never admitted
that he was doing anything but presenting the simple objective facts. What
Skinner left behind was a bag of linguistic tricks by which you can assert
causation while claiming that you are only reporting observations. The way
you use words like support and maintain is only a small fragment of the
complex, interlocking network of word-associations that Skinner developed,
through which it is possible to interpret behavior as being under the
control of the external world, while insisting that this is simply a
description of how things are.

Best,

Bill P.

[Bruce_Abbott]

[From Bruce Abbott (961127.1750)]

Bill Powers (961126.1730 MST) --

All right, let's consider that. Let's suppose that the contingency is F =
K*P, in any appropriate units for Food and Pressing.

What you have are two observed relationships (approximating the imaginary
curves I drew):

1. F = 0[zero] * P (before the contingency is established), and

2a. P := Po*(1 - e^-at), after it's established,
from which it follows that
2b. F := K*Po*(1 - e^-at)

There is nothing in these descriptive equations to indicate that the Food
influences, supports, maintains, sustains, encourages, determines,
regulates, elicits, generates, promotes, provokes, actuates, brings about,
creates, effects, invokes, makes, produces, induces, inaugurates,
instigates, initiates, launches, originates, precipitates, or in short
causes the Pressing (or any other term you can find in a bigger thesaurus
that comes from "cause").

I was about to say how impressed I was at your ability to dredge up synonyms
from memory, but I see you've consulted Roget's.

I agree: there is nothing in these descriptive equations to indicate any of
those things.

Once you have been engaged in this shuffling of mathematical relations long
enough, you might run across an expression containing P and F which no
longer bears any obvious relationship to the two equations above, and it
might occur to you that you can now solve the new expression for P as a
function of F. And indeed you could, with a little luck. You might think
that you have finally established the way in which the contingent food
affects, influences, directs, instigates .... causes the pressing rate. This
would be a complete mathematical blunder. All you will have done will be to
have found a very complex way of writing the inverse of equations 1 and 2.

Yes, and I've seen that done. Go on . . .

When you're all done with this shuffling, there will still be an unexplained
observation: Equation 2a, and its companion, Equation 2b. Unless a new
explicit relation between P and F, independent of these equations, can be
brought into the system, there is simply no way to account for the behavior
of P. That exponential rise of P to an asymptote is caused by a part of the
physical system that hasn't been represented by any equations derived from
observations outside the organism. Without knowledge of what is going on
inside the organism (or an informed guess), the behavior of P is simply
inexplicable.

Again, I agree. Bill, you seem to think that I don't get it, but I do. I
really do.

There are a few additional things we can do with it. For one, if we see
behavior being maintained (sorry) at rates above those expected in the
absence of reinforcement, we can infer that there is probably a contingency
between the behavior and _something_ that is sustaining (sorry) these rates.

Saying "sorry" doesn't justify using unjustifiable language. Yes, something
is "sustaining" the behavior. But it's not the food: according to PCT, it's
a reference signal.

Yes, but the reference signal would have no effect without the rest of the
components in the system. It's the control _mechanism_ that sustains
behavior at whatever level is observed. Given that mechanism and a
particular reference level, it would be just as valid to say that the
disturbance "sustains" the behavior. But that is not quite right, because
the behavior rate on the lever falls if the delivery and consumption of food
does not reduce the error that this disturbance induces.

The reference signal is one of the true independent
variables in this system of which we see only one part from outside it. And
the real explanation of the apparent effect of creating a contingency has to
go beyond even that, because why, when food is not produced by a small
amount of behavior, does the behavior not increase enormously in the attempt
to produce it? Only because some still higher system recognizes that there
is no relation between that action and that desired result, and turns to
some other action that might produce it. The only valid ways to characterize
the relation between the pressing and the food require describing the
remainder of the whole physical situation, which lies inside the organism.

Yes, Bill, I know all that. I was simply describing another, more limited
approach that can take one part of the way (with, as you have noted, the
danger of being mislead into inappropriate conclusions about mechanism),
something I have been labeling as the functional approach. It has its
limitations, but also can provide some useful results and conclusions that
can guide the development of a theory of mechanism. I have been suggesting
that there are things that can be inferred about the operation of the
mechanism even though the mechanism itself remains unknown and unspecified.
Apparently, I have failed somehow to get that across, but in retrospect, I
don't think it really matters in the end. Time to give it a rest.

Regards,

Bruce

[Bill_Powers1]

[From Bill Powers (961128.0145 MST)]

Bruce Abbott (961127.1750)--

Yes, something
is "sustaining" the behavior. But it's not the food: according to PCT, it's
a reference signal.

Yes, but the reference signal would have no effect without the rest of the
components in the system. It's the control _mechanism_ that sustains
behavior at whatever level is observed. Given that mechanism and a
particular reference level, it would be just as valid to say that the
disturbance "sustains" the behavior. But that is not quite right, because
the behavior rate on the lever falls if the delivery and consumption of food
does not reduce the error that this disturbance induces.

This is why words like "sustain" should be avoided. There is a whole class
of terms like this which seem almost deliberately ambiguous; they are a way
of seeming to say something precise without committing yourself to any
particular meaning. It's hard to call them to mind because I normally avoid
using them when being precise matters. Another example is "affect," which
doesn't say whether there are any other variables also capable of affecting
something, and it doesn't say whether there is any _effect_, because if
several variables simultaneously affect something in equal and opposite
directions, nothing will happen; there will be no _effect_. Another term,
beloved of some neurologists, is "modulate," which doesn't indicate whether
the effect is in the same or the opposite direction, or whether an additive
or a multiplicative relationship is involved (in electronics, the meaning of
modulate is always "multiply."). Then there's "modify," which lets you
imagine any kind of effect whatsoever. There are lots more; as I say, I have
trouble calling them to mind because I don't often use them (except when I
have to bluff my way through an incomplete explanation). I call them
weasel-words; no matter how your prediction turns out, you can always weasel
out of being wrong by saying you intended the meaning that fits the actual
outcome.

I have been suggesting
that there are things that can be inferred about the operation of the
mechanism even though the mechanism itself remains unknown and unspecified.
Apparently, I have failed somehow to get that across, but in retrospect, I
don't think it really matters in the end. Time to give it a rest.

I feel the same way, but it's hard to let go of this subject. I certainly
agree that there are things that can be inferred from incomplete knowledge
of a system, but there are legitimate, unambiguous, public ways of making
inferences, and there are ways that depend mainly on vehement assertions and
subjective feelings of conviction. The latter type, on which Skinner relied
heavily, are not very useful or very likely to prove right.

In the back of my mind there is a nonverbal picture of the difference
between the way Skinner approached organisms and the way I do. All this
discussion has brought me closer to being able to describe it, but I'm not
there yet. In the meantime it helps to keep nibbling at the problem. It
seems to me that Skinner started on the wrong foot with his very first step,
but just what that step was I can't put my finger on yet.

Best,

Bill P.