error, information theory

[Human_Factors_Lab]

[From Rick Marken (930307.1500)]

Well, I think I can post to the net again; the secret seems
to be using the correct addess.

Bob Clark (930306.1930 EST) --

Thanks for the comments on human error.

I think your distinction between the "user's" and the "engineer's"
view of error is very important; and it was one of the points I was
planning to make in my discussion of error (I was going to call it
anthropocentric vs egocentric). But I think my approach to this
distinction is a bit different than yours. So rather than respond
to your post directly, let me first broach my ideas and then see if
we can converge (if we want to try).

I would begin my attempts to understand human error by pointing out
that perceptions are just perceptions; in themselves they are neither
right nor wrong; neither correct nor errors. They are just what they
are -- varying intensities, sensations, transitions, configurations,
sequences, relationships, programs, categories, principles, system
concepts. When I see a glass of wine (configuration) knocked off of
a table (relationship, transition) onto an expensive (category) white
rug (configuration), that's what I see. Error implies deviation from
a comaprative reference (specifying the way things should be). If the
perception of spilled wine seems like an error to an observer it
must be because that perception deviates from some specification IN THE
OBSERVER of what should be perceived. This seemingly obvious fact about
perception is completely missed in all discussions of human error that I
have read. In these discussios, one get's the impression that an error is
a perception itself -- corresponding to something "out there" in boss
reality. But we know that very often a perception that is an error
to one person is not an error at all to another. The spilled wine, for
example, might be just what someone wanted to see -- that's why they knocked
over the glass. So PCT enters the picture before it is even invited;
the fact that a person can see an event as an error implies that they
have a reference for how they want that perception to be; they might
not be actively trying to control that perception, but they have a
reference for it nonetheless. An engineer who studies human error must
realize that a perception may be an error from his or her perspective
but not necessarily from the perspective of "the user" -- the person who
is involved in contributing to that perception (like the person who
knocked over the glass of wine). So my version of Bob Clark's "engineers"
perspective on human error requires that the engineer know that the
"human error" he or she is talking about exists because the engineer
him or herself IS a control system.

The engineer must then realize that his or her perceptions (whether
they are considered errors are not) may be the side effects of the
efforts of other control systems to keep their perceptions matching
their own specifications for these perceptions. The engineer must realize
that the "user's" perspective is not anything like his or her own perspective.
The engineer who deals with human error must decide whether his or her goal
is to eliminate his or her own perceptual error (created as a side effect of
the user's actions) or the perceptual error of the user (if there is
any). If the engineers goal is the latter (which the humane and really
the only possibly achievable goal from a PCT perspective) then he or she must
learn 1) what perceptual variables the user is trying to control and 2)
what values of these perceptual variables are considered "right" by the user.
Eliminating human error from the user's perspective is then largely a matter
of figuring out how to design the feedback function (from user output
to input) so that the user has better CONTROL. Some human factors
engineers have successfully designed systems that help user's control
better -- but, since these engineers don't understand PCT, they are
not able to go about the process in a systematic manner (using the
test for the controlled variable, for example).

Martin Taylor (930307 00:50) --

There's no conflict in my mind between the information-theoretic approach
and "straight" PCT. I've said this over and over. The models are the same.

I believe that there is no conflict in your mind; I just think there
should be -- and a BIG one. Information theory has been part of
psychology since the start of the "cognitive revolution". If it is
really the same as "straight" PCT then why aren't ANY information
processing type psychologists aware of some of the fundemental facts
about living systems from a control system perspective; that they
control perceptual variables; that input-output transfer functions
depend on characteristics of the environment, not the internal "processing"
capabilties of the organism; that the use of statistics is an unnecessary
consequence of the failure to test for controlled variables; etc etc.
In other words, information theory based psychology should already be where
PCT was in the 1960s. In fact, psychology, with the benefit of information
theory, demonstrably has NO CLUE about the nature of the phenomenon of
control (purposive behavior) or how to study it.

I can see that you want to cling to this information theory thing;
apparently it's very important to you. And I am honestly willing to
be convinced of it's value (answering Powers' recent challenge
successfully would go along way toward convincing me). But as it sits,
it looks to me like you want to cling to infomation theory the way
other psychologists want to cling to their favorite theory -- even
while embracing PCT. I'm afraid that it just can't be done (and at
the same time get PCT right).

Wanting to stick with a grand old theory is a very common phenomenon -- and
it's the reason why 1) most psychologists don't get into PCT and 2) if
they do, they don't get PCT right. It's why we say PCT is revolutionary.
I know that it seems impossible that all the old, revered theories in
psychology are invalidated by the work of a nice engineer from Chicago
who doesn't even have a PhD in psychology -- but that's the fact Jack.
I know that psychologists in particular are used to clinging to some
remant of the past while moving on to new verbalizations (theories). But
PCT is a whole new enchilada -- it is not like anything else that
has been dreamed of before in your philosophies (except possibly by James
who did say that purposefully produced results were intended SENSORY
CONSEQUENCES of action -- but he had NO IDEA how this worked or what
it meant for the study of behavior).

If you don't want to let go of the old stuff, that's OK. I understand.
But if info theory really has something important to contribute to
the PCT model then SHOW ME WHAT IT IS. I don't want to hear that it just
DOES. Heck, I can go to Agre's conference and hear about how important it
is for me to have a "principled" (I hate that word) understanding of
the interaction between agents and their environments. I don't want
philosophical BS -- I want to see precisely how info theory fits into
my models.

Best

Rick

[_Martin_Taylor]

[Martin Taylor 930308 12:00]
(Bill Powers 930307.0930)

I now understand what Bill Powers wants, which is different from what I
thought. I will try to do it, though it will lack the background I am
trying to develop in the paper. See the end of this posting.

(Rick Marken 930307.1500)

There's no conflict in my mind between the information-theoretic approach
and "straight" PCT. I've said this over and over. The models are the same.

I believe that there is no conflict in your mind; I just think there
should be -- and a BIG one. Information theory has been part of
psychology since the start of the "cognitive revolution". If it is
really the same as "straight" PCT then why aren't ANY information
processing type psychologists aware of some of the fundemental facts
about living systems from a control system perspective; that they
control perceptual variables; that input-output transfer functions
depend on characteristics of the environment, not the internal "processing"
capabilties of the organism; that the use of statistics is an unnecessary
consequence of the failure to test for controlled variables; etc etc.
In other words, information theory based psychology should already be where
PCT was in the 1960s. In fact, psychology, with the benefit of information
theory, demonstrably has NO CLUE about the nature of the phenomenon of
control (purposive behavior) or how to study it.

You are right on target here, except, I think, in that line about "should be."
(And, I think though I won't argue the point, in rejecting statistics entire).

As I understand the problem, the failure of information theorists has been
that of most other theorists, that they take a Newtonian view rather than
an Einsteinian view. They take "probability" as being some kind of a limit
after an infinite number of replications of the frequency with which one
event happens rather than another. This leads to the idea of information
as something that is transmitted through a channel whose size can be
exactly determined. The reason I am taking so long about the information
theory PCT paper is that I am trying to go to fundamental principles to
show how wrong and unworkable that idea is. In essence, the error is
exactly the same as the error of a cognitive planning system that works
only if the environment and the effects of the actuators are precisely
known.

Probability is something relating only to the information available at the
point where that information is used. It relates to individual possible
events that might be detected at that point, not to frequency. Frequencies
of past events that happened when conditions at the point of interest
were "for all practical purposes" identical may well be used in assessing
the probability of some future event, but they are not its probability.

I can see that you want to cling to this information theory thing;
apparently it's very important to you. And I am honestly willing to
be convinced of it's value (answering Powers' recent challenge
successfully would go along way toward convincing me). But as it sits,
it looks to me like you want to cling to infomation theory the way
other psychologists want to cling to their favorite theory -- even
while embracing PCT. I'm afraid that it just can't be done (and at
the same time get PCT right).

Well, you may be right. I don't think so. I find that every aspect of PCT
is illuminated and makes sense if I think from an informational point of
view. I have had my disagreements with Bill Powers on technical aspects
of PCT, but those have usually been based on a (possibly incorrect) belief
that Bill either is asssuming that information is available somewhere it
isn't, or that he is ignoring a source of information that could be used.
Sometimes we come to an agreement when I see that I am wrong in my assumption,
sometimes (as in the reorganization style issue) Bill changes. But the
issue is almost always "where is information available, and is it used."

Bob Clark's Engineer's viewpoint is fine, but in using it the engineer must
try to empathize with the many viewpoints that occur at all places within
the system. If point A is a perceptual signal that has as part of its input
a sensory signal B, the engineer cannot assume that every variation in B
is reflected exactly in A. The question must be "what does A see of the
variation in B" before the engineer can properly assess what will happen
at A. None of Bob Clark's viewpoints seem to me to be of the class
that I might call "internal." They are all "external," even the User's.
An information theoretic approach must take the internal view, even at
second-hand, as an empathetic engineer.

In other words, information theory based psychology should already be where
PCT was in the 1960s. In fact, psychology, with the benefit of information
theory, demonstrably has NO CLUE about the nature of the phenomenon of
control (purposive behavior) or how to study it.
....
Wanting to stick with a grand old theory is a very common phenomenon -- and
it's the reason why 1) most psychologists don't get into PCT and 2) if
they do, they don't get PCT right. It's why we say PCT is revolutionary.

Quite right. I don't know if I get PCT right. I can't know that,
ever. What I can know is that I get what seems like a considerable
error reduction, or subjectively, the first satisfying feeling, after
30 years as a professional psychologist, about a theory that fits what
I believe about the world.

Some 10 years ago I had a conversation with someone who was in graduate
school with me, in which we bemoaned the fact that there seemed to have
been nothing new learned in psychology since we were at school. All that
seemed to have happened was a cyclic change in fashions. We looked for
the Newton of psychology to show us how all these little fashionable
micro-theories could be superseded by a unifying viewpoint. Then 2
years ago I discovered Bill Powers, and although it took me a year to
really feel that I had a good enough handle on what was going on that I
could contribute to the discussion in ways other than simply asking
questions, it took much less time than that to see that the unifying
viewpoint was at hand.

It is still the case that (like poor Bruce Nevin) every moment I put into
PCT is a moment stolen. But I have been trying to shift the constructs
of human-computer interaction theory into a PCT framework, which turns
out to be relatively easy, given that my existing Layered Protocol Theory
can be seen as PCT applied to two interacting hierarchies, if I fudge a
little on the requirement for scalar perceptual variables in classical PCT.
That work is not on stolen time.

But you are right that I have a reference for sticking with information
theory. It has served me very well for nearly 40 years in understanding
economics, aesthetics (from my bachelor's thesis, in which I see little
need for change even now), perceptual psychology, and now PCT. Maybe I
would be better if I dropped it. But I see no evidence pointing in that
direction other than arguments that sound quite analogous to those who
say "Believe in Christ and You Shall Be Saved."

If you don't want to let go of the old stuff, that's OK. I understand.
But if info theory really has something important to contribute to
the PCT model then SHOW ME WHAT IT IS.

I tried, many times, last year. The result was that some people thought
I was just trying to puff up my ego. So I swore off trying to show you
until I could do a more thorough job, which is why I'm trying to find the
odd moments to work on the information->PCT paper, and make it something
that works from principles everyone can agree on, rather than from higher
level constructs that different people read in different ways. I have to
redevelop information theory using subjective probability, in such a way
that I do not have to put up with claims such as "Shannon information has
been shown not to relate to the everyday notion of information or meaning."

I don't want
philosophical BS -- I want to see precisely how info theory fits into
my models.

That's what I want to give you.

···

------------------
Back to Powers:

I am asking that you take the
specific experiment I described, analyze it in terms of
information theory, and tell me what information theory says
about the expected result. If information theory leads directly
to PCT, then obviously information theory will lead to the right
prediction. Just show me how it does, in this specific case.

I'll try to do that, ASAP.

I'm a bit puzzled at the claim that information theory would
allow fitting the model to the data with fewer arbitrary
parameters. In the model we use most often, we obtain an
excellent fit for an individual by changing just one parameter,
the integration factor of the output function. It is hard to
imagine fitting the data with fewer parameters.

If you change, say, the contrast of a tracked target (or the tracking
cursor), or the bandwidth of the disturbance, or put a grid on the
visual surface...is the integration factor the only thing that changes?

I once had an information-based theory of the displacement of the figural
aftereffect. It fitted many different aftereffects in different perceptual
dimensions with only two parameters. Later, a study was published in
which the contrast (a parameter not included in the earlier studies) was
changed. The theroy fitted it with ZERO parameter estimation from the data.
The effect of contrast on information rate was determined from studies
that looked as if they were looking at something entirely different.
That's the sort of thing I mean by extending the range of prediction.

Just to up
the ante a bit: until my challenge is met, I will feel free to
claim that information theory is incapable of predicting ANY
behavior that PCT can predict.

I'll use your circuits, then. And if PCT predicts behaviour that seems to
demand information that cannot be obtained at the point it is used (the
perceptual signal compared to the reference signal), and nevertheless
makes good predictions, then I'll allow your claim.

Look, in any one experiment, you will get as good prediction as you can
achieve with any circuit by varying the circuit parameters. You say that
for all experiments only the integration factor has to be changed. Fine.
If that's true, the challenge is really for me to show whether information
theory could predict the change of integration factor across experiments,
taking into account the conditions of the experiment. I don't know how
I would do that. The integration factor is not a construct that I have
so far analyzed in my thinking. I would assume it relates to motor
information rates as seen by the environment. But I don't know.

In making an analysis for a specific experiment, I would use other experiments
to obtain data--perceptual discrimination experiments, for instance, or
control experiments that use presumed lower-level control systems common
to the specified experiment.

Enough for now. I have a book to edit. Analysis this evening if I can.
At the weekend if I get the #%$@! book review out of my hair that's been
taking up every weekend for the last goodness knows how many. (Rick: that
book exemplifies what you are talking about, but I don't know how to
review it in a way that the audience that believes in its basic principles
will understand. And I find it very difficult to read more than 20 pages
at a sitting, without stopping to mull over why they seem so horribly wrong
and yet so erudite).

Martin

[_Cliff_Joslyn]

I posted something on information a few days ago, which Gary INFORMED
;-> me didn't make it out to the list. Since then the argument's
developed well beyond my post, and I'm enjoying it immensely. This
issue is of key importance to me, as both an information scientist and
control philosopher (if not theorist).

First a comment on semantic information:

[Martin Taylor 930308 12:00]
I have to
redevelop information theory using subjective probability, in such a way
that I do not have to put up with claims such as "Shannon information has
been shown not to relate to the everyday notion of information or meaning."

What do you think of Fred Dretske's Semantic Information Theory? His
view takes similar stances: that information IS about individual
messages, not ensembles with frequencies, and that it relates directly
to common language senses of information.

In the end, I suspect that waht's at stake is a philosophical
dispute about the nature of CAUSAL vs. INFORMATIONAL processes.
Clearly D is CAUSALLY linked to O, but the INFORMATIONAL link is
hardly (!) clear. This is a distinction that Dretske also clearly
makes, finding only a contingent relation between the two kinds of
links (as Bill does). This is contrary to received wisdom, which makes
causal and informational links essentially equivalent.

The whole issue is fascinating, espeically Bill's observations about
the "virtual communication channels" between (RL and CEV) and (D and
O). But I also agree w/Randall that IT and PCT are at qualitatively
different levels of description. wrt/the original challenge, there
seems to be agreement that no matter how you wire up the diagrams,
that IT can be used to describe the informational properties of ANY
communications channel. And that's all that the links in a PCT diagram
are: links in a communications network among environment, controlled
variable, perceived variable, reference level, error signal, output
fnuction, etc. Shannon's theorem applies at each link, and Ashby's Law
(which is isomorphic, I believe) as well: that each link must have
sufficient variety to encode the signal without error in the presence
of expected noise.

The question is of the SIGNIFICANCE of this point: just because IT is
LOGICALLY PRIOR to PCT, doesn't necessarily mean that it's FUNDAMENTAL
to it, but rather it may be IRRELEVANT to it. This is what I
understand Bill's position to be now. IT makes no prediction about one
particular network topology over another, only what the bandwidths of
each link must be to maintain good control. IT strikes me as a very
modest and local theory, nothing like the grand sweep of PCT or
Ashby-en (or anyone else's) Cybernetics in general.

Please don't stop, people. Even if it's "just" an argument about
semantics, those tend to be far more important than they're given
credit for.

O----------------------------------------------------------------------------->

Cliff Joslyn, Cybernetician at Large, 327 Spring St #2 Portland ME 04102 USA
Systems Science, SUNY Binghamton NASA Goddard Space Flight Center
cjoslyn@bingsuns.cc.binghamton.edu joslyn@kong.gsfc.nasa.gov

V All the world is biscuit shaped. . .