solving zadeh's magnus problem
TRANSCRIPT
Mohammad Reza Rajati1, Jerry Mendel1, Dongrui Wu2
1University of Southern California2GE Global Research
Kolmogorov→Dempster→ZadehZadeh: “…[Various theories of uncertainty such as]
fuzzy logic and probability theory are complementary rather than competitive”
� Most Swedes are tall. Most tall Swedes are
blond. What is the probability that Magnus (a
Swede picked at random) is blond?
� Involves linguistic quantifiers (most) and
linguistic attributes (tall, blond)
� An implicit assignment of the linguistic value
“Most” to:“Most” to:
the portion of Swedes who are tall
the portion of tall Swedes who are blond.
� Therefore categorized as a prototypical
advanced CWW problem.
� Q1 A’s are B’s
� Q2(A and B)’s are C’s
� Q1 x Q2A’s are (B and C)’s
At least (Q xQ ) A’s are C’s� At least (Q1 xQ2 ) A’s are C’s
� x is the multiplication of two fuzzy sets via:
� At least is the following operation:
1 1 22 ( ) sup(min( ( ), ( )))Q Q Q Qz xy
z x yµ µ µ×=
=
( ) ( ) sup( ( ))At least Q Qy x
x yµ µ≤
=
� 50% of the students of the EE Department at
USC are graduate students.
� 80% of the graduate students of the EE
Department at USC are on F1 visa.Department at USC are on F1 visa.
� 50% ×80% of the graduate students of the EE
Department at USC are on F1 visa.
� At least 40% of the students of the EE
Department at USC are on F1 visa
� In Magnus problem:
Q1= Most, Q2=Most, A= Swede, B= tall, C=blond
Therefore, At least (MostxMost)=Most2� Therefore, At least (MostxMost)=Most2
Swedes are both tall and blond.
� Most is modeled as a monotonic quantifier
and therefore At least (Most2)=Most2
� Zadeh interprets a linguistic constraint on the
portion of a population as a linguistic
probability (LProb), and directly concludes
that:that:
� LProb(Magnus is blond)=MostxMost=Most2
� We construct a MF for Most:
� We construct a vocabulary of type-1 fuzzy
probabilities to translate the solution to a
word:
Absolutely improbable, Almost improbable, � Absolutely improbable, Almost improbable,
Very unlikely, Unlikely, Moderately likely,
Likely, Very likely, Almost Certain, Absolutely
Certain
� MF of the words are shown here:
� The MF of Most2 is depicted in the following:
� We compute the Jaccard’s similarity between
Most2 and the members of the vocabulary
� It is concluded that “It is Likely that Magnus is
tall”
� Most Swedes are Tall
� A few Swedes are not Tall
� We generally have the following syllogism:
Q A’s are B’s� Q A’s are B’s
� ¬Q A’s are not B’s
( ) (1 )
( ) 1 ( )
Q Q
not B B
u u
u u
µ µ
µ µ¬ = −
= −
� Similarly:
� Most tall Swedes are blond
� A few tall Swedes are not blond
� However, we do not know about the
distribution of blonds among those few
Swedes who are not tall.
� All of them or none of them can be blond
� The available information is summarized in
the following tree:
� In the pessimistic case, none of Swedes who
are not tall is blond, so:
LPr obMost Most Few �one− × + ×
=
� In the optimistic case, all of Swedes who are
not tall is blond, so:
LPr obMost Most Few �one
Most Few
− × + ×=
+
LPr obMost Most Few All
Most Few
+ × + ×=
+
� LProb(blond|Swede) =LProb(tall|Swede) ×
LProb(blond|tall and Swede)+
LProb(¬tall|Swede) × LProb(blond|¬tall and
Swede)Swede)
� Assuming LProb(blond|¬tall and Swede) is
either None or All yields LProb- (Magnus is
blond) or LProb+(Magnus is blond).
� All and None are modeled as singletons:
1 0( )
0 otherwise�one
uuµ
==
� We also construct models for Most and Few,
and a vocabulary of linguistic probabilities
1 1( )
0 otherwiseAll
uuµ
=
=
� MF’s of T2FS models of Most and Few:
� We construct a vocabulary of linguistic
probabilities to decode the solution to a
word:
� The pessimistic and optimistic linguistic
probabilities are depicted here:
� The Jaccard’s similarities between the
solutions and the members of the vocabulary
are demonstrated in the following table:
� “The probability that Magnus is blond is
between Likely and Very Likely”
Using the average centroids of the solutions � Using the average centroids of the solutions
we can also say that:
� “The probability that Magnus is blond is
between around 80% and around 89%.”
� Linguistic Approximation is similar to
rounding numeric values
� The resolution of the vocabulary is important
When vocabularies are small, the pessimistic � When vocabularies are small, the pessimistic
and optimistic probabilities may map to the
same word
� We studied the effect of the size of
vocabularies on the decoded solution
� Vocabularies with different sizes:
� Tables show the similarities of the solutions
with members of each of the vocabularies
� Using all of these vocabularies, both the
pessimistic and the optimistic solutions map
to the same word, which is Likely for the first
vocabulary, and is Very Likely for others.vocabulary, and is Very Likely for others.
� For small vocabularies, the total ignorance
present in the problem does not affect the
outcome.
� Novel Weighted Averages are promising
when dealing with linguistic probabilities
� Our solution builds a probability model for
the problem which obeys a set of axiomsthe problem which obeys a set of axioms
� Is the problem really reduced to calculating
the belief and plausibility of a Dempster-
Shafer Model?