Re: [UAI] Fuzzy sets vs. Bayesian Network

From: Ronen Brafman (brafman@indigo.CS.BGU.AC.IL)
Date: Tue Feb 29 2000 - 05:53:14 PST

  • Next message: Dietrich Eherler: "Re: [UAI] Graphical representation of CPTs in belief nets"

    Regarding possible interpretation of fuzzy propositions. At least
    with regards to vague concepts, such as Tall, what we have is
    uncertainty, not about the state of the "objective" world, but rather
    uncertainty with regards to the semantics of the term. We simply do
    not know precisely what Tall means.

    If you believe that uncertainty should (at least ideally) be captured
    via a probability distribution, this justify a view of fuzzy
    membership as a probability distribution over possible interpretations
    of this term (perhaps what Pearl had in mind). In the case of Tall,
    different interpretations correspond to different intervals. If
    accepts the subjective approach to probability theory, then the
    decision-maker/analyst should determine what is his/her subjective
    interpretation and define a probability distribution.

    Since our interpretations of various concepts are not generally
    independent, but could have arbitrary relations, Bayes nets could be
    useful for describing this relationship, and the truth-functional
    operations of fuzzy logic will not, in general apply (as suggested by
    Kathy). I have never had the energy to really see how this compares
    with the practice and theory of fuzzy logic.

    With regards to introducing utility here, that's interesting, but
    I would be cautious, as this might complicate the picture
    once we want to make decisions (in a clean and justified manner).
    Plus, we need a good semantic reason to do so.

    Ronen

    ------- End of Forwarded Message



    This archive was generated by hypermail 2b29 : Tue Feb 29 2000 - 05:53:31 PST