Powers (970107.1150 MST) revising conditional probabilities

[CCHERPAS]

[From Chris Cherpas (970108.1049 PT)]
  [re: Bill Powers (970107.1150 MST)]

Sorry to ask for another copy, but my email problems were apparently not
completely solved. Bill, could you resend your (970107.1150 MST) post?
If you don't want a duplicate cluttering up csgnet you could send it
directly to ccherpas@cccpp.com. I would greatly appreciate it.

One thing I have observed in the recent probability discussion between
Bill Powers and Martin Taylor is that although frequentist and classical
notions of probability have been more or less explicitly addressed, Bayesian
revision has not. Repeatedly Bill has said how he tries to eliminate
uncertainty with some perception-controlling action, and at that moment,
the probabilistic moment, is over (or perhaps, for Bill, never existed):
either one outcome or another is perceived.

However, if one is in a situation where one's uncertainty is
changed, but not sufficiently, one will act again, but now with
a new _conditionalized_ basis for perceiving the probabilities
associated with a set of potential outcomes. As one repeats the
conditionalization process, one is revising one's perception of
probabilities -- with each cycle, so to speak.

If one's attempts at controlling any perception simply aren't sufficiently
successful, the reference for that perception will get varied by some
higher-up perceptual control, including, for those who have been so trained,
by controlling the perception of a program -- I call it a "Bayesian revision
cycle" for the moment (perhaps involving calculations). Good control of
this program-level perception also means achieving certainty among
constituents of the whole perception you were trying to control in the
first place, providing control of perceptions of _conditional probabilities_.

Until you get into revising through conditionals, I doubt if you are using
probability crucially in your decision-making. Working from a frequency
distribution, per se, is the lowest level of competence you can start
with for improving your odds -- doing experiments to re-conditionalize
the probabilities is when you start to actually learn something useful.
Working with the algebra of Bayesian probability provides a way of
administering and analyzing a process of revising your beliefs, but isn't
equivalent to your perception of the probability of X; the program-level
perception you are controlling presumably makes the process of learning
about the probability of X something verifiable, so that you and others can
agree to move on to the next set of doubts, questions, and uncertainties
of life.

Best regards,
cc

[Bill_Powers1]

[From Bill Powers (970108.1630 MST)]

Chris Cherpas (970108.1049 PT)--

One thing I have observed in the recent probability discussion between
Bill Powers and Martin Taylor is that although frequentist and classical
notions of probability have been more or less explicitly addressed, Bayesian
revision has not. Repeatedly Bill has said how he tries to eliminate
uncertainty with some perception-controlling action, and at that moment,
the probabilistic moment, is over (or perhaps, for Bill, never existed):
either one outcome or another is perceived.

However, if one is in a situation where one's uncertainty is
changed, but not sufficiently, one will act again, but now with
a new _conditionalized_ basis for perceiving the probabilities
associated with a set of potential outcomes. As one repeats the
conditionalization process, one is revising one's perception of
probabilities -- with each cycle, so to speak.

This is an interesting point. I think Martin actually said something similar
(although not so specific), but I was on a different wavelength.

Of course, when you put it this way, I do have a perception of probability
or uncertainty, and I control it. Instead of simply accepting the
probabilities and then trying to make forecasts or plans on the basis of
them, my strong preference is to try to get the probabilities up into the
range where I can can behave as if there isn't any uncertainty. Of course we
always _behave_ as if there's no uncertainty, but I like to eliminate the
errors that result from doing that when the uncertainty is actually very
high. I've been talking about that for years, but never thought of it this
way. I've accused psychologists of giving up on finding models that really
work, so they've simply accepted the huge noise level in their data as
"natural variability" about which nothing can be done. You might say that my
whole reason for getting into this subject in the first place was that I
thought the statistical approach of psychology worthless and was confident
(at one is the age of 27) that I could do better -- given Wiener's big insight.

I can't comment on the Bayesian approach -- don't know enough about it.

Best,

Bill P.

If one's attempts at controlling any perception simply aren't sufficiently
successful, the reference for that perception will get varied by some
higher-up perceptual control, including, for those who have been so trained,
by controlling the perception of a program -- I call it a "Bayesian revision
cycle" for the moment (perhaps involving calculations). Good control of
this program-level perception also means

achieving certainty among

···

constituents of the whole perception you were trying to control in the
first place, providing control of perceptions of _conditional probabilities_.

Until you get into revising through conditionals, I doubt if you are using
probability crucially in your decision-making. Working from a frequency
distribution, per se, is the lowest level of competence you can start
with for improving your odds -- doing experiments to re-conditionalize
the probabilities is when you start to actually learn something useful.
Working with the algebra of Bayesian probability provides a way of
administering and analyzing a process of revising your beliefs, but isn't
equivalent to your perception of the probability of X; the program-level
perception you are controlling presumably makes the process of learning
about the probability of X something verifiable, so that you and others can
agree to move on to the next set of doubts, questions, and uncertainties
of life.

Best regards,
cc