Re: [UAI] Fuzzy sets vs. Bayesian Network

From: Lluis Godo (godo@iiia.csic.es)
Date: Thu Feb 24 2000 - 17:03:12 PST

  • Next message: Hansen, Peter Friis: "RE: [UAI] Fuzzy sets vs. Bayesian Network"

    Hello,

    I would like to comment a little bit on the relationship/differences
    between fuzzy sets and Bayesian nets.

    First of all, Bayesian networks is a model to reasoning under uncertainty.
    Fuzzy sets theory is a model, extension of classical set theory, to
    represent the fuzziness or vagueness inherent to gradual rather than crisp
    (well-defined) notions or properties.
    So, in principle the aims of both models are completely different and thus
    they can be considered as orthogonal as Pearl suggest.

    However fuzzy sets can be linked to a particular model of uncertainty
    which is called Possibility theory, which again has nothing (or little) to
    do with probability.
    The idea is very simple. Suppose you are interested in the likelihood of
    an event "the value of variable x belongs to a set A", where A is a
    (classical) subset of the (discrete) domain of the variable x. For
    instance, let x be "temperature of a certain device", A = [20, 30].
    Consider the following three scenarios:

    1) You are lucky and you are provided with a probability distribution p of
    variable x. Then you can evaluate the likelihood of the event simply as
    summing up the probability values p(t) for all t in A.

    2) Much worst, you are only provided with the information that only values
    of another subset B (e.g. [25, 35]) are feasible. In such case you can
    only evaluate about the possibility of the event:
            - "x in A" is possible event if A intersects B
            - "x in A" is an impossible event otherwise

    3) Still worst than 1. Suppose that you are told that only "high
    temperatures" are feasible for that device. What to do in a such a
    case?. Assume that "high
    temperatures" can be represented as a fuzzy set with a membership function
    m_high. Notice that m_high(t) is an evaluation to what degree a particular
    temperature t can be considered high in the current context.
    For instance we may have m-high(10) = 0, m_high(20) = 0.5, m_high(30) = 1,
    m_high(40) = 1.
    Now, POssibility theory postulates that the information "only high
    temperatures are feasible" induces a "possibility distribution" on which
    values the variable can take (this is uncertainty modelling!), and
    moreover it postulates which is this distribution, it is just
            poss(X=t) = m_high(t)
    Notice that, if you know that "only high temperatures are feasible", to
    have both poss(x= 30) = poss(x= 40) = 1 is perfectly reasonable.
    A possibility distribution allows to evaluate how "possible (or
    plausible)" and how "necessary (or certain)" is then any event:

            Poss("x in A") = max{ poss(t) : t is in A}
            Nec("x in A") = 1 - Poss( x is not in A) = inf{ 1-poss(t) : t is
    not in A}

    Thus the possibility degree of an event is determined by its highest
    plausible elementary event.

    If the fuzzy set "high temperatures" turns out to be defined as a
    classical (non-fuzzy) set B, then we come down to the above sceneario 2.

    In summary, Possibility theory is an (full flecthed) uncertainty model,
    related to fuzzy set theory, different from probability theory, which
    takes into account "weak" forms of information and thus with a qualitative
    flavour and can be used if "strong" numerical information is missing.

    Hope this might help you.
    There is plenty of (serious) bibliography about all this stuff if you are
    interested.

    Lluis Godo.



    This archive was generated by hypermail 2b29 : Thu Feb 24 2000 - 17:07:42 PST