Assumptions and justifications

[Martin Taylor 951103 13:10]

Bill Powers (951103.0745 MST)

On the cat:

...I would proceed to apply disturbances that
ought to alter the consequence in question (concurrently, or on
succeeding trials) and see whether the cat's actions varied in just the
way needed to maintain the consequence the same.

I have a puzzlement, which seems to have fairly wide ramifications. In
Thorndike's cat box, Thorndike knows that the movement of the stick is
what causes the door to open (and I can use "cause" here because that
linkage is open-loop). But the cat can't know it, and no matter how
unimaginably clever a super-cat Thorndike put in the box, the super-cat
couldn't find out. The cat is in the position of a player in the coin
game who cannot alter the pattern. Thorndike didn't, so far as I remember,
move the cat from one box to another, or from one kind of box to another
in which only the stick remained a constant feature. And I'm sure he didn't
ask the cat to specify the kind of box to use.

Given that the cat can't, _in principle_, determine that pushing the stick
is the mechanism whereby the presumed controlled perception of "being at the
food dish" was achieved, we are left with the everyday statement of PCT--
randomized reorganization until the (uncontrolled) actions that occur
with a positive-going output signal just happen to bring about the desired
perception. At that point, reorganization stops.

So, the cat can and does control lots of low-level perceptions, many of
which are used in the escape procedure, but only Thorndike knows which
actions (not behaviours) are the effective ones. On each introduction of
the cat to the box, it does lots of random things, changing because its
reference to be at the food does not match its perception of the current
state. As it does these things, one side-effect of some behaviour is the
action of moving the pole. That serves to bring the "at the food" perception
to its reference value. But it is only a side-effect, and the cat is
"doing" something quite different.

As the trials go on, the reorganizations always succeed when a side effect
of some random behaviour has the right result. If the cat thinks of it at
all, it probably thinks something like "if I mark my territory and then
lie on my left side facing the clock, the door will open and I can get my
food." The same behaviour will always have the same side effect, because
the environment hasn't changed. "Marking the territory" requires a certain
attitude, and after that, lying on the left facing the clock just happens
to involve brushing the stick. But there's no reason at all to suppose that
the cat notices the stick. The fact that the cat gets out quicker and
quicker on successive trials has no bearing on whether the stick appears
in any of the cat's controlled perceptions.

Now consider:

No, if I were seriously proposing any particular consequence that the
cat is "doing", I would not just be guessing. I would have to guess at
first, of course, but then I would proceed to apply disturbances that
ought to alter the consequence in question (concurrently, or on
succeeding trials) and see whether the cat's actions varied in just the
way needed to maintain the consequence the same. If that part of the
Test were failed, I'd have to guess again, but if it were passed, I
would verify that the cat's actions alone are causing the resistance to
disturbance, and that interrupting the cat's perceptions of the variable
would destroy the stability of the variable.

If what I said above is true, then what the cat is "doing" has no connection
with the door opening, except by coincidence unknowable to the cat. There are
myriads of other things that coincide with "marking the territory and lying
on the left facing the clock." So if you introduce disturbances, you will
be testing for controlled variables that will change unpredictably if "mttal-
otlftc" no longer involves brushing the stick. It will appear to you as if
those variables had not been controlled. How then could one study what the
cat is "doing" in escaping from the box?

The analogy with human superstition and religion is far too close to avoid
examination, so I won't make the examination explicit.

When the environment is relatively unchanging, the side effects of our
actions are also relatively stable, and we may achieve our ends by means
that are radically different from the way we think we are achieving them.
If, to take an example I assure you to be hypothetical, I habitually bring
my wife flowers on Fridays, and she expresses pleasure and acts generally
as I would wish, I may think she likes flowers, and try bringing more if
I think she is in a bad mood. But maybe she just enjoys the silly smile
I have when I hand her the flowers, and a toy train might serve just as
well. I could never tell, since there are an infinite number of side-effects
of my behaviour, and any of those actions might be having the desired

Anyway, I hope you see my puzzlement: when you set up an experiment in a
fixed environment, you cannot be sure whether the actions you observe are
behaviour or side-effect, but when you change the environment you may
change the behaviour that you are trying to observe. Only if you happen
to hit on a perception whose control is unaffected by the change of environment
will you even begin to be able to apply the Test.



When the theoretical assertions are made clear in this way,
we can see that they require justification, and that in turn forces us
to search for other experimental manipulations that will subject them to
the usual tests of science. You can't just point to the original data as
evidence; you have to find some _other_ means of justifying the
assumption. When you express the original data in terms of the
assumption, you have used up those data as far as justification is
concerned. You need _different data_, and not just a repeat of the
original results in a similar experiment. You need a _new experiment_.

This is what I was taught to do as a grad student in psychology. It was then
called "converging operations," following (I think) P.W.Bridgeman. My
professor (Tex Garner) thought that it was the only legitimate way to
investigate nature. So I don't understand why you follow up with:

I far as I have ever been able to see, psychology is simply not
conducted in this way.


Shannon Willyams (951103)

What is worse is: a hypothesis that does not have a causal mechanism
is not subject to error. It is infallible.

Are you sure?


What is the causal mechanism for the hypothesis about which you are so sure?
Are you sure that your answer IS the correct causal mechanism? If so, what
is the causal mechanism for this hypothesis? etc...

Oh, I see. Your "Yes" is not subject to error:-)

In the end, all hypotheses rest on observations of the world, none of which
are assured, and on probabilistic or logical relations inferred among these
observations. There's no "causal mechanism" in PCT; there's only a
reference to accepted relationships at another level of description.
And that's the point of Hans's question that you finessed:

How about quantum mechanics with its predictions that
are accurate to 1 part in 10^9 or better? What is the _cause_ that
everything is quantized?

I do not know what is the cause that everything is quantized. Why do you

You can, in principle, derive just about everything mechanistic to very
high accuracy if you accept that everything is quantized, and not otherwise.
The various computational formulae work exceedingly well and provide
"causes" for why some atoms are unstable, why some elements form compounds
with others, and so forth. But we know no "cause" of why the universe
works this way, and to say "God willed it" is only to push the question
back a level.

Theories that sound as if they present a cause are really only saying that
they are grounded in theories of a different kind, or in observations.
All theories are only compact descriptions of nature, and the more compact
they are, the more they seem to represent "causes."


Rick Marken (951102.1550) to Bruce Abbott (951102.1505 EST)

Methods not covered

How to determine the:
absolute threshold for a sensation
threshold for the difference between two sensations
size of an illusion
controlled perception
D' and Beta in signal detection
duration of short-term memory
serial-position curve
magnitude of the Stroop effect
optimal interstimulus interval in classical conditioning
simple reaction-time
flicker-fusion frequency
onset and duration of REM sleep

Every one of the "methods" in the list above, except for The Test, is based
on the assumption that behavior is _caused output_. Every one of these
methods, except for The Test, assumes that the relationship between an
independent variable and a dependent variable tells us something about the
organism that exists "between" these variables -- about sensory thresholds,
memory storage, attentional filters, etc. In all cases, the assumption is
that the organism is "open loop"; that response (dependent) variables depend
on input (independent) variables but that input variables do not
simultaneously depend on response variables.

I hope this response was just an aberration, given that we sorted out this
whole topic two or three years ago to (what I thought was) everyone's

To summarize:

(1) A closed loop consists of a set of input-output links, and acts as a
closed loop only when all links are connected.

(2) The I/O properties of the links in a closed loop affect the performance
of the loop when it is closed, but they can be studied in an open-loop

(3) In a typical psychophysical experiment such as a detection task, there
is no closed loop whereby the subject's response affects the presentation
to which the response is made; that aspect of the experiment is stimulus-
response, in the sense that the subject has a reference level derived from
a higher closed loop (next item).

(4) In a typical psychophysical experiment, there is a closed loop in which
the subject has a reference perception something like "the experimenter
is satisfied with my performance." In this loop, the open-loop S-R
results of the overt experiment form a link. Usually, it is the perceived
consistency of results when compared with other related results that
satisfies the experimenter's reference perception, and thus the experimenter.

(5) In many of the items to which you object, the open-loop performance
being studied is a limiting capability. The subject cannot improve
performance beyond some limit, and it is that limit that provides the
consistency that the experimenter is seeking. Good experiments of this
kind make it very hard for a subject to achieve consistency according to
any internal criterion. In other words, they explicitly try to avoid the
possibility that the subject might be controlling anything that could
affect the thing being measured. Experimenters know this, and work hard
to achieve it. Some subjects do try to make things easier for themselves
by setting a low criterion for capability, but these are seldom very
consistent, and they usually fail to "satisfy the experimenter."

(6) If the experiment is properly designed, so that the subject has no control
of the measured results, the capacity of that link will be effectively
measured, and can be used in other studies of control loops that involve
the link. Studies of the control loops in action are required, but cannot
be used by themselves to examine the properties of individual links in
the loop.

Capsule: The subject controls a perception of the experimenter's satisfaction.
The experimenter controls a perception of the subject's consistency. That
consistency is attainable by the subject most readily at a limiting condition
such as "doing my best." At the limit, the properties of a link are optimum,
and those optima can be used in any other link or loop of which this link
forms a part.

Which of the excoriated study types conform?

absolute threshold for a sensation -- Yes, but it's a bad measure anyway
threshold for the difference between two sensations -- ditto
size of an illusion -- consistency, but not limiting. Depends on technique.
controlled perception -- see first part of this posting
D' and Beta in signal detection -- Yes, and these are better measures
duration of short-term memory -- questionable
serial-position curve -- Yes, if you mean testing memory limitations
magnitude of the Stroop effect -- problem of interpreting averages as

                                     representing individuals. Otherwise OK.

optimal interstimulus interval in classical conditioning -- ??
I.Q. -- ??
simple reaction-time -- questionable
flicker-fusion frequency -- Yes.
onset and duration of REM sleep -- This examines what is presumably a side

                                    effect of some ongoing control. Maybe
                                    it means something and maybe it doesn't.
                                    Not clear how it could be a link in a loop.

I don't follow your characterization at all of the relation between
stimuli and responses in these open-loop studies. Or perhaps I should say
that to the degree to which I can interpret it, I don't accept it.

All you have done in the list above is named a bunch of phenomena (except
controlled perceptions) that are observed in the context of studying organisms
as cause- effect systems; all these phenomena (except for controlled
variables) are studied using the methods described in your text; methods
designed for the study of open-loop, cause-effect systems; methods that
reveal nothing about purposes of purposeful systems.

All true. And all missing the point, that to predict the behaviour of control
loops (as opposed to observing it) you need to know the characteristics of
the component parts. Those characteristics may be best studied in open-loop
situations. If you don't want to predict, then don't bother with those kinds
of experiments. But don't say they are useless in a PCT context, either.