You are on page 1of 16

201

Invited Review

Using fuzzy sets in operational research


H.-J. Z I M M E R M A N There are other fields, however, in which the
R WTH Aachen. Germany, Fed. Rep. description of problems is also vague, but in which
the vagueness is of a different kind than random-
Received December 1982
ness. One of the most important areas of this kind
is probably that of decision making: here the
During the last four decades a large number of mainly h u m a n factor enters with all its vaguenesses of
mathematical tools have been developed in and for Operational
perception, of subjectivity, of attitudes of goals
Reserach. They are primarily devices to find optimal solutions
after a problem has been modelled formally. Hardly any new and of conceptions. To use a modelling language
tools have been developed for proper modelling of problems which is dichotomous in this area seems far from
which are not well-structured and which can not easily be cast adequate.
into classical mathematical models. Fuzzy Set Theory as pro- The language is, roughly speaking, the link be-
posed by L. Zadeh in 1965 promises to bridge part of this gap.
It is much more adaptable to different problem structures and tween the thinking of h u m a n beings, the object
better suited to model human evaluation and decision making system, which is to be modelled, and the model
processes, than classical mathematics. This paper will primarily itself. Statements in a model are formulated by
focus attention on four questions: means of linguistic expressions. That means that
Why Fuzzy Sets in OR?
the model language sematically and systematically
What are Fuzzy Sets?
Where can they be used? and has to be well suited to grasp those aspects of
How are they being used? reality which are included in the model; the per-
spective of a scientific theory is already contained
in its liguistic terms.
1. Why fuzzy sets in operational research? Let us look at the appearance of scientific theo-
ries and of practical problems which we want to
1.1. Vague phenomena, vague relations and the model: Most scientific systems do not have a clear
modelling of problems distinction between logical and descriptive con-
stants. In fact a science seems to be written in a
Operational Research covers a number of areas mixture of everyday language describing experi-
of quite different character. In all of these areas ments, mathematical equations describing calcula-
modelling is of prime concern to Operational Re- tions, and semiformulized deduction if the science
searchers. W h e n applying O R - m e t h o d s to en- is sufficiently advanced to contain a theoretical
gineering problems, for instance, the problems to part [2].
be modelled and solved are normally quite clear " T h e problems of management exist only in the
cut, well described, crisp. They can generally be minds of managers and in the minds of their
modelled by using classical mathematics which is advisers. There is, therefore, a completely subjec-
dichotomous in character. If vagueness enters, it is tive personal basis to our science. It is not the case
normally of the stochastic kind which can properly that the c o m p a n y has an inventory problem in the
be modelled by using probability theory. This is sense that if you visit it you can see it there in a
true for m a n y areas such as inventory theory, concrete form .... Even within our own minds as
traffic control, scheduling, etc. in which probabil- we appraoch a practical problem, we will be un-
ity theory is applied via queuing theory. sure how to show which form of model is likely to
be optimal .... Since the problems as seen in each
North-Holland of our minds will be different and mutual com-
European Journal of Operational Research 13 (1983) 201-216 parison will be impossible, it follows that we have

0 3 7 7 - 2 2 1 7 / 8 3 / $ 0 3 . 0 0 © 1983 Elsevier Science Publishers B.V. (North-Holland)


202 H.-J. Zimmermann / Fuzzy sets in OR

a significant, even dangerous, freedom in the way The usefulness of the mathematical language
in which we select the models which will represent for modelling purposes is undisputed. However,
decision making situations." [32] there are limits of the usefulness and of the possi-
On the other hand a mathematical model is bility of using the classical mathematical language,
defined to be a collection of statements about a set based on the dichotomous character of set theory,
of variables from which the truth or falsity of to models in particular systems and phenomena in
other statements can be deduced. the social sciences: "There is no idea or proposi-
The utter importance of modelling language is tion in the field, which can not be put into
recognized by Apostel, when he says: " T h e re- mathematical language, although the utility of
lationship between formal languages and domains doing so can very well be doubted" [10].
in which they have models must in the empirical Schwartz [36] brings up another argument
sciences necessarily be guided by two considera- against the unreflected use of mathematics, if he
tions that are by no means as important in the states: An argument which is only convincing if it
formal sciences: is precise loses all its force if the assumptions on
(a) The relationship between the language and which it is based are slightly changed, while an
the domain must be closer because they are in a argument, which is convincing but imprecise may
sense produced through and for each other; well be stable under small perturbations of its
(b) Extensions of formalism and models must underlying axioms".
necessarily be considered because everything intro- By now it should have become clear, that tradi-
duced is introduced to make progress in the de- tional (crisp) mathematics is quite well suited to
scription of the objects studied. Therefore, we model either crisp problems or problems in which
should say that the formalization of the concept of the vagueness is of the stochastic (probabilistic)
approximate constructive necessary satisfaction is type. These types of problems are very common in
the main tasks of semantic study of models in the natural sciences and engineering. They are by far
empirical sciences". [2] not as frequent in social sciences and in particular
Because we request that a modelling language is in decision making.
unequivocal and nonredundant on one hand and, Vagueness of the above mentioned type in-
at the same time, catches semantically in its terms dudes phenomena which are 'intrinsicly' vague,
all that is important and relevant for the model we such as "good labour relations", "acceptable prof-
seem to have the following problem: Human its", "dependable forecasts", "good schedules". It
thinking and feeling, in which ideas, pictures, also includes relationships such as " I f we have a
images, and value systems are formed, first of all good management, we can survive hard times",
has certainly more concepts or comprehensions " I n growing markets competition is less fierce
than our daily language has words. If one consid- than in a recession" or " T h e available capacities
ers, in addition, that for a number of notions we should not be excessively used". It is not difficult
use several words (synonyms) then it becomes to see that classical mathematics is not a proper
quite obvious that the power (in a set theoretic tool to model these kinds of phenomena or rela-
sense) of our thinking and feeling is much higher tionships properly and that a new modelling tool
than the power of a living language. If in turn we would be quite desirable.
compare the power of a living language with the
language of logic then we will find that logic is 1.2. Informational vagueness
even poorer. Therefore, it seems to be impossible
to guarantee a one to one mapping of problems Another crucial factor in modelling reality is
and systems in our imagination and a model using the quality and the quantity of data one can
a mathematical or logical language. One might expect to obtain about the problem to be mod-
object that logical symbols can arbitrarily be filled elled. The saying 'garbage in-garbage out' is well
with semantic contents and that by doing so the known and speaks for itself. The following quota-
logical language becomes much richer. It will be tion from Josiah Stamp [38] points in the same
shown later that it is very often extremely difficult direction: "Governments are very keen in amass-
to appropriately assign semantic contents to logical ing statistics. They collect them, add them, raise
symbols. them to the n th power, take the cube root and
H.-J. Zimmermann / Fuzzy sets in OR 203

make wonderful diagrams. But what you must crisp set to a set with vague boundaries. The
never forget is that every one of these figures degree of membership of single objects to this set
comes in the first instace from the village watch- varies gradually. Partitioning such a set ap-
man who just puts down what he damn pleases". propriately into smaller subsets would certainly
It must, however, be born in mind that the decrease the number of descriptors which are nec-
effort put into deriving and obtaining numerical essary to obtain crisp subsets. For illustrative pur-
values or relations must be geared to the value of poses we shall use the term 'creditworthiness':
the model and that when data are scarce it may This 'subjective category' (one might prefer to
still be useful to draw conclusions from not fully think of creditworthy people) can be subdivided
satisfactory input data. In this case a tentative into two major determinants, financial basis and
look at the dependence of the solution from the personality, which again are fuzzy but probably
quality of the input data may be very advisable. less than creditworthiness. By further subdivisions
Probably nobody will object to the statement we might end up with a hierarchy of 'subjective
that poor or insufficient data are a barrier for categories' and 'subcategories' which in our minds
good modelling. The contrary, however, can also are the better described the lower in the hierarchy
be true. An abundance of information can also be they appear (see Fig. 1).
the cause for vagueness: In day-to-day conversa- Putting it the other way round: The higher in
tion people use very effectively terms such as the hierarchy a term appears the fuzzier is its
"pleasant personality", "comfortable houses" or description.
"creditworthiness" as labels for sets of objects
which could be described uniquely and crisply if a 1.3. Heuristic algorithms
large number of 'descriptors' were used. Since the
human information storage- and processing capac- Whenever a problem or the model of it are too
ity is very limited, however, all the necessary de- complex to be solved exactly with acceptable ef-
scriptors of such a set are not consciously used fort one turns to the use of heuristic algorithms.
when communicating with other persons. The de- Heuristics normally either reduce the solution
scription and conception of such a set--which we space which is considered in searching for an
will call 'subjective category'--changes form a optimal solution or make the algorithm faster and

Credit- 1
worthiness

Financial Personality
basis

Business
behaviour
I

Property minus Othernet Income minus I Continuity Economic Conformation with


long-term property expenses Motivations ] Lthinking social & economic
of margin
debts standards

Fig. 1. Concept hierarchy of creditworthiness.


204 H.-J. Zimmermann / Fuzzy sets in OR

by doing so lose the certainty that an optimal We will also refer to Zadeh's following basic
solution is really found. A human decision maker concepts:
without adequate EDP-software and hardware at Support: The support of a fuzzy set A is a set
his disposal very often resorts to such an approach S(A) such that
and achieves quite reasonable results. His results
x > 0.
are frequently at least good enough from a practi-
cal point of view and often the decision procedure Normality: It has already been mentioned that
is more stable than a sophisticated mathematical #(-) is not limited to values between 0 and 1. If
control algorithm [26]. Supx#A(X ) = 1 the fuzzy set A is called normal. A
One could also call this procedure 'approxi- not empty fuzzy set A can always be normalized
mate" rather than 'exact'. Classical mathematics is by dividing/~A(X) by Supx/~A(x ).
often badly suited to model these algorithms; ap- Equality: Two fuzzy sets A and B are equal iff
proximate reasoning seems to be more ap-
propriate, How these types of approaches can be ~A(x)=/~,(x) ( o r / ~ A = t ~ , ) , x ~ X.
implemented by use of fuzzy sets will be illustrated In the following we will for convenience always
in Section 3.2. write #a for/aA(X ), X ~ X; unless stated otherwise
and we will assume normalized fuzzy sets.
Containment: A is a subset of B iff ~a </.to.
2. What are fuzzy sets? Intersection: The membership function of A A B
2.1. Fuzzy set theory, possibility theory, probability /L4ne = min(/~A, ttB),
theory or using the conjunction symbol A :

Recognizing the shortcomings of traditional


mathematical models in some areas L. Zadeh sug- Union: The membership function of A U B is
gested in the mid-sixties [46] the notion of a fuzzy defined as
set. He defined a fuzzy set as follows: /~AuB = max(/~A, /~B)
or
Definition. If X = (x) is a collection of objects
denoted generically by x, then a fuzzy set A in X is ].LAUB~'~A V ~B"
a set of ordered pairs, Algebraic product: The membership function ~AB
A = (x, of the algebraic product of two fuzzy sets A and B
is defined as
~ ( x ) is called the membership function or grade
of membership of x in A which maps X to the
membership-space M. (When M contains only the Algebraic sum: The membership function of the
two points 0 and 1, A is nonfuzzy and #A(X) is algebraic sum of A and B is also defined by its
identical to the characteristic function of a non- membership function:
fuzzy set.)
/-tA.o = ttA + ~n --/~A/~O •
/z(. ) is a function the range of which is a subset
of the nonnegative real numbers and has the prop- Relation: A fuzzy relation, R, in the product
erty that the supremum of this set is finite. space X × Y = ((x, y)), x ~ X, y ~ Y, is a fuzzy set
in X × Y, the membership function #R of which
Example. Let X = (10, 20, 30, 40, 50, 60, 70, associates with each ordered pair (x, y ) a grade of
80, 90, 100, 110) be possible speeds (mph) at which membership i~R(x,y) in R. An n-ary relation in a
cars can cruise over long distances. Then the fuzzy product space X = X~× X 2 . - - X" is then char-
set A of "comfortable speeds for long distances" acterized by a corresponding n-variate member-
may be defined by a certain individual as ship function.
In 1965 Zadeh already mentioned that some of
A = ((30, 0.7), (40, 0.75), (50, 0.8), (60, 0.8),
his concepts--in particular the union and the in-
(70, 1.0), (80, 0.8), (90, 0.3)). tersection with their interpretations as 'inclusive
H.-J. Zimmermann / Fuzzy sets in OR 205

or' and 'and' respectively--might have to be de- for breakfast". The possibility distribution /~(x)
fined differently depending on the context of their and the probability distribution p(x) might then
semantic interpretation. In 1978, Zadeh [48] speci- be as shown in Table 1.
fied one possible interpretation of fuzzy sets as Decision models might contain probabilistic as
possibility distributions. Since there is a certain well as possibilistic components. In addition to
resemblance between possibility theory and proba- those they might, however, contain components
bility theory some comments might be useful: and relationships which are vague but neither in
Probabilistic models are models which describe the probabilistic nor in the possibilistic sense. By a
situations in probabilistic terms, i.e. probabilities, 'decision model', as considered here, we mean a
distribution function etc., and which for their model which can be descriptive or prescriptive and
transformations and relations use the axiomatic which shall be factual in the sense that human
system of probability theory. Hamacher [17] fol- beings can behave as required by the model
lowed Carnap and Stegmi~ller [11] when he dis- (prescriptive case) or do behave as asserted by the
tinguished 12 different kinds of probabilities. They model (descriptive case). I.e. those models have to
differ from each other with respect to their be formulated in a way that they can be tested
semantic interpreation (for instance Kolmogoroff- empirically. This holds for the modelling of vaguely
type probabilities vs Koopman-type probabilities), described entities (corresponding to probability or
with respect to their mathematical properties (the possibility functions) as well as for the computa-
former is a Boolean ring, satisfying the 8 Koop- tional or aggregation rules.
man's axioms, the latter is a o-algebra with specific Suggestions for axiomatic and empirical justifi-
properties) and with respect to their linguistic form cations of concepts for 'and' and 'or' will be
(explicite vs implicite probabilities). Probability presented in Section 4 of this paper.
theory of whichever kind normally assumes that
the 'events' are well defined (crisp). Concerning 1.2. Semantic interpretations of fuzzy sets
the aggregation of probabilities (random numbers
or truth sets) in the sense of 'and', 'or', 'not', the Even though many different kinds of probabili-
operations used (addition, multiplication, sub- ties exist, only two basic interpretations can be
straction from 1) are uniquely defined. Even found: degrees of belief or truth on one side and
though the systems are generally formal and not the frequentistic interpretation on the other side.
factual theories, their empirical validity with re- For fuzzy sets few semantic interpretations have
spect to 'stochastic' situations can be considered so far been put forward:
as satisfactorily established. The above mentioned notion of a possibility
Possibility theory [48] is essentially also a for- distribution can be considered as one interpreta-
mal theory which was developed in the framework tion. The concept of a subjective category as men-
of the theory of fuzzy sets. The semantic interpre- tioned in Section 1.2 is another. Rather similar is
tation of 'possibility' differs, however, from that the concept of a 'linguistic variable'. In fuzzy
of 'probability' and so do the underlying axioms clustering notions of similarity are represented by
and aggregation rules. Possibility theory corre- fuzzy sets. In reliability theory, the term 'safety' or
sponds formally to the min-max version of fuzzy 'reliability' is sometimes considered to be repre-
set theory. The different meanings of a probability sentable by a fuzzy set and in project management
distribution and a possibility distribution might the project time. It can be expected that many
become apparent in the following example: more context-dependent interpretations will be
Let us consider the statement "Paul eats x eggs suggested and used. In all these cases the question

Table 1
Possibility d i s t r i b u t i o n vs. p r o b a b i l i t y d i s t r i b u t i o n

x 1 2 3 4 5 6 7 8 9 10
~(x) 1 1 1 1 0.8 0.6 0.4 0.2 0 0
p(x) 0.1 0.8 0.1 0 0 0 0 0 0 0
206 1-1.-J. Zimmermann / Fuzzy sets in OR

of interpretation can be reduced to a measurement In conventional nonfuzzy decision making un-


problem: How is 'Fuzziness' or the 'degree of der certainty we are used to thinking of a decision
membership' measured, on which scale level is the as consisting of
information available and which operators are ap- (a) a set of possible activities (decision varia-
propriate to model the situation or problem under bles),
consideration appropriately. The flexibility of fuzzy (b) a set of constraints limiting the choice be-
set theory is one of its strengths and weaknesses at tween the alternatives (solution space), and
the same time. (c) the objective functions which assigns a
Two essential features of the theory of fuzzy 'value' to each result due to a certain choice of
sets should have become obvious by now: activities according to their 'desirability'.
(1) Membership functions of fuzzy sets and opera- The optimal decision is then the selection of the
tions with and on them play a crucial role in fuzzy activity with the highest 'desirability' (for instance,
set theory and both have been of concern to many the alternative which results in minimum cost,
scientists in the meantime. maximum profit etc.).
(2) Fuzzy set theory is essentially a very general, In a fuzzy environment this picture of a deci-
flexible, formal theory. If it is to be applied to real sion has to be revised: The fuzzy objective func-
problems it can and has to be adapted carefully. tion is characterized by its membership function
Neither the concept of membership nor the 'opera- and so are the constraints. Since we want to satisfy
tors' have a unique sematic interpretation. The (optimize) the objective function as well as the
context-dependent semantic interpretations will constraints, a decision in a fuzzy environment is
lead to different mathematical definitions of mem- defined by analogy to nonfuzzy environments as
bership functions and appropriate operators. the selection of activities which simultaneously
In the following we will focus our attention on satisfy objective function(s) and constraints.
contributions of fuzzy set theory to Operations If one accepts Bellman and Zadeh's [7] defini-
Research. We will, therefore, exclude some fea- tion of a decision as the intersection of all fuzzy
tures and parts of this theory, in particular the sets involved, i.e. the congruence of 'and' and
extensive mathematical development during the 'intersection', and if one also accepts the mathe-
recent past, from our considerations. matical model of the minimum operator for the
intersection of fuzzy sets, then the 'decision' in a
fuzzy environment can be viewed as the intersec-
3. Where can they be used? tion of fuzzy constraints and fuzzy objective func-
tion(s). The relationship between constraints and
Fuzzy set theory can primarily be used in OR in objective functions in a fuzzy environment is,
three different ways: therefore, fully symmetric, i.e. there is no longer a
(a) As a language to model appropriately prob- difference between the former and the latter.
lems and situations which contain fuzzy phenom- For illustration the following example:
ena or relationships; The board of directors has to decide on the
(b) As a tool to analyze such models in order to dividend which should be offered to the share-
get a better insight into structures of problems and holders at the next meeting. Some members of the
models; board argue that " T h e divident should be substan-
(c) As an algorithmic tool to make solution tially larger than 10% in order to make our shares
procedures more stable or faster. attractive".
In the following three subsections we shall pre- Let us call this demand the 'objective'. The
sent some examples for each of these potential membership function of the fuzzy set "Dividends
application areas. substantially larger than 10%" might look like Fig.
2.
3.1. Fuzzy sets for modelling Other members of the board argue that for
reasons of stability " T h e dividend should be in the
One of the areas of OR to which fuzzy set vincinity of 11%" (This would be the 'constraint').
theory has been applied most is that of decision The membership function of this fuzzy set is shown
modelling. in Fig. 3.
H.-J. Zimmermann / Fuzzy sets in OR 207

u(x) x ~(×)'
I-

0
10 15 x[%]
o 4. ~ _ 10 11 11,75 x[%]
)
Fig. 2. "Dividens substantially larger than 10%". Fig. 4. A fuzzy decision.

The decision, i.e. the fuzzy set including the The concept of a decision in a fuzzy environ-
dividens in the vincinity of 11% and substantially ment has been applied to special types of decisions
larger than 10%, is indicated in Fig. 4. such as linear programming.
Algebraically the membership function of the We shall consider the problem
objective could be
minimize Z = cx,
0 for x < 10% subject to A x <~b, (1)
#°(x)= (l + (x - 1 0 ) - 2 ) -I forx>~ 10%. x>~0
in which the objective function as well as the
That of the constraint could be
constraints are fuzzy sets. We can then write the
,c(x)=((l+(x_ll) 4) ' problem as

Find x such that


The membership function of the decision is then
cx<Z,
~D(X) = min{~t0(x), /~c(x)}. Ax z b, (2)
Generally a crisp decision is desired rather than x>~O.
a fuzzy one. It seems appropriate to consider as a Here c is the vector of the coefficients of the
(crisp) 'optimal decision' the solution with the objective function, b is the vector of the constants
highest degree of membership in the fuzzy set of the constraints, A is the usual coefficient matrix
decision. and Z is an aspiration level which calibrates the
Thus for an optimal decision fuzzy set 'objective funtion'. The symbol < de-
~ ( X * ) = max min{/lc(X ),/~0(x)}. notes the fuzzified version of '~< ' and reads for
instance "essentially smaller than or equal to".
In the above example We now define a function f : R " + 1 ~ [0, 1] such
that
X*=11.755 and ~(X*)=0.755.
0 if A x <~b and cx <~Z is strongly
F(Ax) =
violated,
1 if A x <~b and cx <~Z is satisfied.
(3)
Using the simplest version of the functionf(ax, cx)
we assume it to be linear and the intersection of
the (fuzzy) constraints ~tnd the (fuzzy) objective
function.
11 x[%]
Thus

Fig. 3. "Dividens in the vincinity of 11%". f ( A x , cx) = f ( B x ) = M i n ~ ( ( B x ) , x> 0


208 H . - J . Zimmermann / Fuzzy sets in OR

with tive functions [48] and which accommodate several


objective functions [51]. Other types of operators
1 for ( Bx ) i <~b i, can also be included in the models. It is more
1 -- ( B x ) i - bi difficult, however, to solve the latter type of model
di for b~ < (Bx)i numerically. This, of course, does not decrease the
/,((ax),) = modelling quality of fuzzy set theory.
<~b i + di,
In other areas of decision theory, such as game
for ( Bx ) i theory, group decision theory [29], multi-criteria
> bi + d i , analysis, fuzzy set theory has been used as a
modelling tool similarly. One of the original
(4) motivations for the development of fuzzy set the-
where d i are subjectively chosen constants of ad- ory was pattern recognition. It is therefore not
missible violations of the constraints, f~(Bx)i is the surprising, that this area has remained a field of
membership function of the i th row of the linear intensive use of fuzzy set theory as modelling tool.
system ( Bxi), Three books on fuzzy clustering are witnesses of
this potential [3,5,21]. Clustering is one of the
Minf,. ( ( B x ) , ) areas in which structures which are hidden behind
masses of data are made visible. Fuzziness can
is the 'fuzzy' decision, and enter in a number of ways: Fuzzy objective func-
Max Minf~ ((Bx) ,) tions can be used rather than crisp ones, the
x>~O i gradual membership of an element to many clus-
the decision with the highest degree of member- ters can be expressed rather than the dichotomous
ship. membership to a single cluster. Algorithms have
Substituting been developed which in certain cases turn out to
be more efficient than classical cluster-algorithms.
b;=z, 8;= If the interested reader does not want to start
right away with a book on fuzzy clustering the two
papers by Bezdek and Harris and by Roubens are
componentwise and simplifying by dropping the
a recommendable introducton [34].
'1' (which does not change the problem!) we arrive
Another good and typical example of fuzzy
at the following problem:
modelling is presented by Prade [31] when he
Max Min (b" - ( Bx);) (5) presents a version of 'fuzzy PERT'. He bases his
x>~0 i approach on possibilistic considerations and char-
or acterises the flow times of activities by fuzzy num-
bers which are defined by socalled L - R member-
MaX#D(X). ship functions. These membership functions are
x>~0
uniquely determined by the three-tuples (m, a,/3)
As is well known, this problem is equivalent to
as follows:
solving the following linear program:
Maximize X, [L[(m- x)a-'] for x ~<m,
subjectto X~bi-(Bx)~, i = 0 , 1 . . . . . m, t~(x)=tR[(x-m)fl-' ] forx>~m,
x ~<0. (6) a >..-0,/3 >f 0 (7)
The optimal solution to (5) is also the optimal where m is the mean value (or characteristic value)
solution to (6). of ~ , a and/3 are the left and right spreads of m,
So far linear membership functions and the and L(0) = R(0) = 1.
appropriativeness of the minimum operator have Basing on Zadeh's extension principle which
been assumed. Models, however, have already been defines the membership function of the sum of
suggested in the literature [25] which consider non- two fuzzy numbers ~ + ff as
linear membership functions, which take into con-
sideration crisp and fuzzy constraints and objec- /l,~_~(z) = sup m i n [ / l , ~ ( x ) , / ~ ( x ) ] ,
Z~X4"y
H.-J. Zimmerrnann / Fuzzy sets in OR 209

he develops specific rules for the above mentioned 3. 3. Fuzzy algorithmic approaches
type of membership functions. He then solves
PERT-type sequencing problems in which the So far fuzzy sets were used to model a problem
scheduling is constrained by crisp and by fuzzy and crisp traditional mathematics was used to
constraints. either find solutions to equivalent crisp models or
By contrast to the fuzzy linear programming to investigate their structural sensitivity. Now we
approaches described earlier the solution to the shall consider the algorithmic use of fuzzy set
above scheduling program is not a crisp but a theory.
fuzzy schedule. Two major positions of departure can be recog-
nized: One is a generalization of two-valued logic,
3.2. Fuzzy set theory as an analytic tool which leads to 'approximate' or 'fuzzy' reasoning.
The other could probably best be characterized by
Let us use the linear programming as a vehicle the term 'fuzzy controller'. The idea behind this
for illustration again: approach is that human controllers are very often
Sensitivity analysis in linear programming in- much more stable and efficient in controlling for
vestigates changes of the optimal value of the instance semi-automatic production processes than
objective function or the optimal solution as a are mathematically formulated and EDP-sup-
function of changes of the right-hand side or the ported control devices. One, therefore, tries to
coefficients of the objective function. Since all model human controlling behaviour by fuzzy set
coefficients are crisp, only crisp changes can be theory and after it has been formalized as a 'fuzzy
considered. This is obviously already additional control algorithm' substitute it for control systems
structural information about the problem or model which are very often mathematically very sophisti-
compared to the knowledge of the optimal solu- cated but tend to overreact and 'explode' unter
tion only. If a problem description is vague, how- certain constellations.
ever, it can be modelled as a fuzzy LP-model. In between these two approaches are others,
Then changes of stability as a function of varying such as fuzzy decision tree methods [1], fuzzy
fuzziness are of interest and can be obtained. branch and bound methods etc. which, however,
What consequences do increased tolerance inter- are still at the very beginning of their develop-
vals have? What changes can be expected if the ment.
aspirations and hereby the membership functions Here we shall elaborate some more about 'fuzzy
of the constraints or objective functions change? reasoning' and ' fuzzy controllers'.
What effects on the solutions can be expected if "Informally, by approximate or, equivalently,
crisp constraints turn into fuzzy constraints? These fuzzy reasoning we mean the process by which a
are questions asked and answered in sensitivity possibly imprecise conclusion is deduced from a
analysis of fuzzy linear programs [18]. Similar collection of inprecise premises. Such reasoning is,
holds for duality theory in fuzzy linear program- for the most part, qualitative rather than quantita-
ming: Traditional crisp duality theory in linear tive in nature and almost all of it falls outside of
programming is uniquely defined and considers the domain of applicability of classical logic" [47].
optimal pairs of primal-dual solutions only. In Zadeh's system for fuzzy logic essentially con-
fuzzy duality theory [33] the pairs of primal-dual sists of sets of rules which together constitute a
models can be formulated differently. In this re- consistent system for approximate reasoning. He
spect certain analogies exist between stochastic distinguishes: 'Translation Rules', i.e. rules that
linear programming and fuzzy linear program- yield the translation of a modified composite pro-
ming. Fuzzy duality also does not restrict con- position from the translations of its constituents.
siderations only to optimal pairs of solutions but The translation of a proposition must be under-
also considers suboptimal solutions (decision be- stood as its associated possibility distributions.
haviour), which is probably more realistic than There are rules pertaining to modification (x is
focussing on optimal behaviour only. Another in- very small. Therese is highly intelligent.), rules
teresting type of sensitivity analysis via fuzzy sets pertaining to composition (X is small and Y is
is presented by Wang and Togai in a very recent large. If X is small then Y is large.), rules pertain-
article [42]. ing to quantification (Most Swedes are tall. Many
210 H.-J. Zimmermann / Fuzzy sets in OR

men are taller than most men) and rules pertaining which otherwise could not be defined satisfactorily
to qualification (John is tall is very true. John is at all.
tall is quite possible). One application of fuzzy logic is fuzzy control.
A typical elementary example of the composi- For the purpose of illustration let us regard one of
tional translation rules is for example [47]: X and the first experiments in this direction [26]:
Y are variables taking values in U and IS, respec- The 'plant' to be controlled was a steam engine
tively. F and G are fuzzy subsets of U and V and boiler combination. Two inputs were consid-
(possibility distributions/~y and H x). ered: heat input to the boiler and throttle opening
If at the input of the engine cylinder. The outputs are
the steam pressure in the boiler and the speed of
XisF ~ Hx=F,
the engine. Tests showed, that the relationships
YisG ~ Hy=G, (8) were highly nonlinear with both magnitude and
then polarity of the input variables. Thus an installed
digital controller had to be retuned each time the
X i s F a n d Y i s G --, H<x.y)=F(qG operating point was altered. Six variables were
=F×G, defined: PE (Pressure Error), SE (Speed Error),
CPE (Change in Pressure Error), CSE (Change in
X is F or Y is G ~ H(x.y) = F + G. (9) Speed Error), HC (Heat Change) and TC (Throttle
Change), the last two of which are the action
Here H(x,y ) is the possibility distribution of the
variables. Between 5 and 8 fuzzy sets were defined
binary variable (x, y), ff and G are the cylindrical
for each of these variables: PB (Positive Big), PM
extensions of F and G, F × G is the cartesian
(Positive Medium), PS (Positive Small), NO (Nil).
product of F and G.
PO (Just above zero), NS (Negative Small), N M
If, for instance, U = V = (1, 2, 3) and
(Negative Medium) and NB (Negative Big). For
Fi small =a ((1, 1), (2, 0.6), (3, 0.1)), the variable TC (Throttle Change) five of these
fuzzy sets (i.e. their membership functions) are
G i large ~ ((1, 0.1), (2, 0.6), (3, 1)),
shown in Table 2.
then (8) and (9) yield In addition to the rules for the union and
intersection of fuzzy sets mentioned earlier the
x is small and Y is large ~ 1-l<x.y) complement rA of a fuzzy set A was defined by
H<x,y )= ([(1, 1), 0.1], [(1, 2), 0.6], [(1, 3), 0.1], /lrA(x ) = 1 -/~A(x). This corresponded to the ne-
gation of the statement behind fuzzy set A. A
X [(2, 1), 0.1], [(2, 2), 0.6], [(2, 3), 0.61, heater control algorithm and a throttle control
× [(3, 1), 0.1], [(3, 2), 0.1], [(3, 3), 0.1]), algorithm was designed and implemented. For
illustration we show a part of the heater algorithm:
x is small or Y is large
If PE = NB
II<x.y )= ([(1, 1), 1], [(1, 2), 1], [(1, 3), 1], and CPE = not (NB or NM)
X [(2, 1), 0.6], [(2, 2), 0.6],[(2, 3), 1], and SE = ANY
× [(3, 1), 1], [(3, 2), 0.6], [(3, 3), 1]).

In addition to the 'Translation Rules' there are Table 2


'Modifier Rules', and 'Rules of Inference' which
shall not be discussed here. The major use of all Fuzzy set Setting
these rules seems to be the analysis of complex
-2 -1 0 +1 +2
systems of statements, their transformation into
either simple but equivalent or less redundant PB 0 0 0 0.5 1
systems or the derivation of conclusions from im- PS 0 0 15 1 0.5
precise premises. This might sometimes amount to NO 0 0.5 1 0.5 0
NS 0.5 1 0.5 0 0
the 'more precise definition' of vague terms or
NB 1 0.5 0 0 0
relations by means of their membership functions
H.-J. Zimmermann / Fuzzy sets in OR 211

Set-point his problem• Also many authors base their models


on the min- and max- operators as they are being
used in possibility theory without any proof why
p/ d'
xc there are adequate in a specific situation•
There are two ways to justify the use of particu-
lar membership functions or operators: Axiomatic
justification, similar to the approaches used in
utility theory and empirical proofs. We shall com-
ment on these two approaches in the next para-
I I I I I I I I I I I graphs.
IT 2T 3T 4T ..... T 80see
Time 4.1. Axiomatic systems
Fig. 5. Fixed controller (DDC-algorithm) ( × , D) vs. Fuzzy
controller (O). The first axiomatic justification for the min-op-
erator to be a model for the intersection of fuzzy
sets and therefore a model for the (logical) 'and'
and CSE = A N Y was given by Bellman and Giertz [6].
then HC = PB Starting from ten plausible axioms Hamacher
Else [17] derived four basic nonredundant axioms for
the connective 'and' (conjunction) interpreted as
If PE = N B o r N M the intersection operator, which also seemed to be
and CPE = NS
sensible from the point of view of rational decision
and SE = ANY
making:
and SCE = A N Y
then HC = PM
Axiom 1. The connective C (conjunction) is as-
The performance of the fuzzy controller compared sociative, i.e. A A (B A C) = (A A B) A C (neces-
to other automatic controllers turned out to be sary to extend C from 2 to n statements)•
quite good as can be seen from Fig. 5.
Axiom 2. C is continuous (secures stability of C
for small changes of/~A and/ta).
4. How are they being used?
Axiom 3. C is injective in each argument, i.e.
By 'how' we mean: are fuzzy sets used as a (A A B)~:(A A C) i f B : ~ C.
purely formal tool or language or do they assert to
factually model problems, situations or be- Axiom 4. / ~ x~<o.ll" C ( x , x ) = x ,~ x = l .
haviours?
Fuzzy set theory is obviously a formal tool A, B and C denote fuzzy sets and x is the
when used algorithmicly or as a language for for- degree of membership of an element with respect
mal models• to the appropriate fuzzy set.
If, however, fuzzy set models claim to factually Hamacher proved that if C is an algebraic
describe in subjective categories, decision making function, a rational function, and a polynom in x
behaviour of human beings or the aggregation and y then only the following mathematical de-
process in decision making then some more proof scription of the operator is possible:
and justification is needed• In fuzzy set theory
C(x,y) = x .y
there are primarily two areas in which such justifi- v+ ( 1 - v ) ( x + y - x y ) '
cations are desirable: membership functions and
operators. Most publications still proceed by "as- 7 is an arbitrary constant• For Y = 1 the above
suming a membership function of that and that formula reduces to
type". They certaintly are suggesting formal mod- C(x,y)=x.y,
els, and the reader is left with the problem of
determining appropriate membership functions for which obviously contradicts the assertion that the
212 H . - J. Zirnmerrnann / F u z z y sets in O R

minimum operator is used to model the 'and'. types of managerial decisions. Therefore, first of
It should be stressed that Hamacher's justifica- all a new 'and' had to be introduced; we called it
tion was primarily aiming at the use of 'and' in the 'compensatory and' in contrast to the 'logical
decision problems! It does not seem possible to and'. As mathematical models for this 'compensa-
justify any model for the 'and' contextfree, i.e. tory and' a number of possible operators were
valid for all possible problem areas. Dombi [13] tested: Max, Min, arithmetic mean, geometric
pointed to the fact that it might even be difficult mean and a special 'y-operator'. Detailed descrip-
to derive and justify models for connectives without tions of the tests and the results can be found in
taking into consideration the models used for [40]. It turned out, that the operator which by far
membership functions and weighting of fuzzy sets. performed best was the 3,-operator, which incorpo-
For shapes of membership functions hardly any rated the following idea:
axiomatic justifications have become known so far. When people aggregate subjective categories in
Schwab [35], again for fuzzy sets representing fuzzy their minds in order to evaluate or decide they use
criteria in decision models, justified spline func- a great variety of operators. These operators vary
tions as proper membership functions. He bases with respect to the degree of compensation and
his proof on 10 'rationality axioms' and derives a with respect to the importance attached to the
type of function which is unique and which can be criteria expressed in the subjective categories. Then
calibrated by the decision maker by setting three expressing their thoughts verbally they are forced
interpretable parameters. An empirical validation to choose one of the only two words available to
of his axioms is, however, not yet available and them: 'and' or 'or'. These words have been seman-
might be difficult to obtain. tically defined, 'and' as the 'logical and' without
any compensation and 'or' as the 'inclusive or'
4. 2. Empirical approaches with full compensation. Thus already by ex-
pressing their thoughts due to the 'poorness' of
When testing empirically the product operator our language people have to approximate their real
and the minimum operator whether they really problems or thought processes more or less well by
modelled the 'and' used in human decisions it was selecting the verbal connective which is closest to
found in the second half of the seventies that both their real connnective.
operators had to be rejected as adequate models Thus the real approximation takes place already
[40]. when verbally expressing evaluation models and
There appeared even doubts on whether not so much when turning this description into a
managerial decisions could be looked at as 'inter- mathematical model.
sections' of goals and constraints: Intersection in Since the real connective with respect to com-
the classical sense means that there is no com- pensation has to be between the 'logical and' and
pensation between the membership to different the 'inclusive or' we defined the 'compensatory
sets: If somebody is looking for a black-and-red and' as a combination of the intersection and the
skirt, then a green (not black)-and-red skirt will union of fuzzy sets:
not be acceptable, even though it might be " v e r y
red". (In two characteristic functions a zero of one ]XASB ~ ~ A ~ B t ' A U B" (lO)
can never be compensated by a one in the ap-
If the intersection and the union are algebraically
propriate position of another characteristic func-
tion when intersecting them.) represented by the product and the algebraic sum,
In managerial decisions, however, in most cases respectively, then (10) becomes
compensatory effects between objectives are pre-
sent. Two portfolios might, for instance, be of /to= /x 1- (1-/1,) , (11)
i=1 i=1
equal worth even though one has a slightly higher
expected return and the other a lower risk in- 0</~<l,0<y< 1,
volved! If full compensation was present, the un- i = 1, 2 , . . . , m, m = number of sets to be connected.
ion, modelled by the max-operator might be an
acceptable model for aggregation. This, however, If it is desired to introduce different weights for
seems unlikely in the light of the large variety of the sets in question. /~ and 1 - / ~ , could for in-
Credit-
II worthiness
Weight

f
i
I
y _ 0.59

Financial
Personality
basis
8 = 0.95
N

I e~
II
y = 0.62 7 = 0.60
II
/
3' - ' 0.78

Liquidity Business I
Potential
I behavior [

1
7 = 0.98 ]( y = 0.84 3' = 0.57 J V = 0.90 7 = 0.55 "r = 0.77 "y = 0.55

Property minus
I
Conformation with
Other net Income minus Continuity Ph) ical Economic" }
long-term property expenses ,ntal Motivation thinking social & economic
of margin & ~
pott
debts ztial 8 = 0.80 standards
3 = 2.03 8 = 1.01
8 = 0.23 8 = 0.92

Fig. 6. Concept hierarcy of creditworthiness together with individual weights and y-alues for each level of aggregation.

i'O
t~
214 H. -.I. Zimmermann / Fuzzy sets in OR

stance be replaced by Some of those real applications have not been


published. One of the reasons for the relatively
/.l, = $1, 1 -p,=(l -v,p (12) small number of real applications is probably that
a new theory or tool has to be widely accepted by
where V, are the (raw)membership values and S,
practitioners before it is applied to real problems.
their corresponding weights. In order to preserve
Thus most ‘applications’ so far are applications to
the structure of our model, the sum of weights S,
‘model problems’. We shall cite some of them:
should be equal to the number of sets connected.
Control of processes has been a widespread
In the laboratory type tests this connective per-
concern of people working in fuzzy set theory. The
formed very well. Formally it also has the ad-
most information about applications of fuzzy set
vantages that it is
theory to control problems can be found in Gupta,
- pointwise injective,
- continuous, Saridis and Gaines [15]. Here one finds several
monotonic and commutative,
_ in accordance with truth tables of dual logic. descriptions of applications to warm water plants,
to a sinter making plant in which the standard
It is also ‘context adaptable’ by its two parame-
deviations was reduced by 40% compared to nor-
ters which can be easily determined empirically
mal control, to a basic oxygen steel making pro-
whenever desired.
cess and to cement Kiln control.
This connective was then tested in a field study
Applications to inventory problems are de-
type experiment:
scribed by Sommer [37] and by Kacprzyk and
45 credit clarks of 5 different banks were asked
Staniewsky [20], very interesting discussions of
to rate 50 fictious credit applications as to their
applications of fuzzy set theory to problems of
credit worthiness. A detailed description of experi-
civil engineering are presented by Blockley [8], to
mental procedures, results and data can be found
queuing models by Dubois and Prade [ 141, and to
in [53] and [54].
project management by Chanas and Kamburow-
Fig. 6 shows the values of y and 6 which were
sky [ 121 and by Dubois and Prade [ 141.
determined for the creditworthiness hierarchy of
A book recently edited by Yager [45] also in-
Fig. 1.
cludes descriptions of the interesting application
In order to test the above mentioned models for
of fuzzy set theory to data display problems, to
operators empirical membership functions had to
water resource planing and earth quake engineer-
be determined empirically before.
ing, to regional development and to aggregate
Computer programs were developed which de-
production scheduling and to a number of medical
rived from observations informations on a scale
problems.
level appropriate for degrees of membership. These
Of course, fuzzy linear programming has been
referred to fuzzy sets only which consisted of a
applied to real problems in logistics (truck fleet
finite number of elements. Later models were de-
scheduling, container scheduling), and to blending
veloped and tested [40] which were suited to de-
problems and to model problems of multi criteria
scribe membership functions for continuous fuzzy
analysis [51], media selection [43], etc.. At present
sets such as “tall men”, “large profits”, etc. These
membership functions can also be calibrated con- an interactive decision support system on the basis
of fuzzy linear programming with several objective
text dependently by setting certain parameters.
Focus was again the modelling of subjective cate- functions and crisp and fuzzy constraints is being
gories by fuzzy sets. Other types of membership developed. Publications in this direction can be
functions might have to be developed for other expected in 1983/84.
application areas.

5. Conclusions
4.3. Applications
Fuzzy set theory-even though still young-has
The theory of fuzzy sets is still quite young. Its already proved to be a very useful modelling and
fast growth really started in the second half of the algorithmic tool in Operational Research. The the-
seventies. It is therefore not surprising that the oretical mathematical development is quite ad-
number of ‘real’ applications is still not very high. vanced but not yet finished by far. Kandel lists in
H.-J. Zimmermann / Fuzzy sets in OR 215

his recent book [21] 3064 references which is cer- Induced Fuzzy Sets (Delftse Universitaire Pers, Delft, 1978).
[4] J.F. Baldwin, A new approach to approximate reasoning
tainly a considerable number of publications. In
using fuzzy sets, Fuzzy Sets and Systems 2 (1979) 309-325.
particular, if most of them have appeared within [5] J.C. Bedzek, Pattern Recognition with Fuzzy Objective
the last ten years. Two Journals are published in Function Algorithms (Plenum, New York, 1981).
this area and looking at the contribution in these [6] R. Bellman and M. Giertz, On the analytic formalism of
journals one gets the impression t h a t - - t o an in- the theory of fuzzy sets, Information Sci. 5 (1973) 149-156.
creasing degree--the theoretical papers outnum- [7] R. Bellman and L.A. Zadeh, Decision-making in a fuzzy
environment, Management Sci. 17 (1970) 141-164.
ber the applied papers. This is certainly regretta-
[8] D.I. Blockley, The role of fuzzy sets in civil engineering,
ble, but it is a tendency which can also be recog- Fuzzy Sets and Systems 2 (1979) 267-278.
nized in other areas, for instance, Operational [9] D.I. Blockley, The Nature of Structural Design and Safety
Research! (Wiley, Chicester, 1980).
Fuzzy Sets are so far primarily used in Opera- [10] H.W. Brand, The fecundy of mathematical methods,
Dordrecht (1961).
tional Research [11] R. Carnap and W. Stegmiiller, Induktive Logik und
- as a modelling language. Wahrscheinlichkeit, Wien (1959).
All cases in which fuzzy set theory is properly [12] S. Chanas and J. Kamburowski, The use of fuzzy variables
used as a modelling tool are characterized by in PERT, Fuzzy Sets and Systems 5 (1981) 11-19.
three features: [13] J. Dombi, A general class of fuzzy operators, the De
Morgan class of fuzzy operators and fuzziness measures
(a) Fuzzy phenomena, relations or evaluations
induced by fuzzy operators, Fuzzy Sets and Systems 8
are modelled by a well defined and founded (1982) 149-164.
theory (there is nothing fuzzy about fuzzy [14] D. Dubois and H. Prade, Fuzzy Sets and Systems: Theory
set theory!); and Applications (Academic Press, New York, 1980).
(b) By doing so a better approximation of real [15] M.M. Gupta, G.N. Saridis and B.R. Gaines (Eds.), Fuzzy
Automata and Decision Processes (North-Holland, Amster-
phenomena by formal models is achieved;
dam, 1977).
(c) A better modelling of real phenomena nor- [16] M.M. Gupta, R.K. Ragade and R.R. Yager, (Eds.), Ad-
mally requires more and more detailed in- vances in Fuzzy Sets Theory and Applications (North-Hol-
formation than is needed for rather rough land, Amsterdam, 1979).
dichotomous modelling. [ 17] H. Hamacher, (]ber logisehe Aggregationen nicht - bini~r ex-
plizierter Entscheidungskriterien (Fischer, Frankfurt/Main,
- as an algorithmic tool. The main advantages are
1978).
that heuristic algorithms can often be made [18] H. Hamacher, H. Leberling and H.J. Zimmermann, Sensi-
more efficient by use of fuzzy sets and that tivity analysis in fuzzy linear programming, Fuzz),, Sets and
human controlling or problem solving be- Systems 1 (1978) 269-281.
haviour can often be modelled better by fuzzy [19] H.M. Hersh and A.A. Caramazza, A fuzzy set approach to
modifiers and vagueness in natural language, J. Exp.
set theory than by traditional mathematics.
Psyschol. (General) 105 (1976) 254-276.
There exists a strong need for more empirical [20] J. Kacprzyk and P. Staniewski, Long-term inventory policy
research, in particular on membership functions. making through fuzzy decision making models, Fuzzy Sets
This is probably one reason for the lack of more and Systems 8 (1982) 117-132.
real applications and their publication. Some pro- [21] A. Kandel, Fuzzy Techniques in Pattern Recognition (Wi-
ley, New York, 1982).
gress in this direction will hopefully be made in
[22] A. Kaufmann, Theory of Fuzzy Subsets, Vol. 1 (Academic
the near future. This would certainly make fuzzy Press, New York, 1975).
set theory even more useful for Operational Re- [23] W.J.M. Kickert, Fuzzy Theories on Decision-Making
search. (Martinus Nyhoff, Boston, 1978).
[24] G.E. Lasker (Ed.), Applied Systems and Cybernetics, Vol.
V1 (Academic Press, New York, 1981).
[25] H. Leberling, On finding compromise solutions in multi-
References criteria problems using the fuzzy min-operator, Fuzzy Sets
and Systems 6 (1981) 261-269.
[1] J.H. Adamo, Fuzzy decision trees, Fuzzy Sets and Systems [26] E.H. Mandani and S. Assilian, An experiment in linguistic
4 (1980) 207-219. synthesis with a fuzzy logic controller, J. Exp. Psychol.
[2] L. Apostel, Formal study of models, in: H. Freudenthal, (General) 105 (1976) 311-325.
Ed., The Concept and the Role of the Model In Mathematics [27] E.H. Mandani and B.R. Gaines (Eds.), Fuzzy Reasoning
and Natural and Social Sciences (D. Reichel, Dordrecht, and its Applications (Academic Press, London, 1981).
Holland, 1961). [28] C.V. Negoita and D.A. Ralescu, Applications of Fuzzy Sets
[3] E. Backer, Cluster Analvsis by Optimal Decomposition of to Systems Analysis (Wiley, New York, 1975).
216 H.-J. Zimmermann / Fuzzy sets in OR

[29] H. Nogiri, A model of fuzzy team decisions, Fuzzy Sets [42] P.P. Wang and M. Togai, Fuzzy sensitivity analysis and
and Systems 2 (1979) 201-212. synthesis technique, in: E. Nagel, P. Suppes and A. Tarski,
[30] H. Nojiri, Decision processes in new product develop- Eds., Logic, Methodology and Philosophy of Science (Stan-
ment, Fuzzy Sets and Systems 7 (1982) 227-241. ford University Press, Standford, 1962) 589-601.
[31] H. Prade, Using fuzzy set theory in a scheduling problem, [43] G. Wiedey and H.J. Zimmerman, Media selection and
Fuzzy Sets and Systems 2 (1979) 153-165. fuzzy linear programming, J. Operational Res. Soc. 29
[32] P. Rivett, Principles of Model Building, (Wiley, New York, (1978) 1071-1084.
1972). [44] R.R. Yager, Fuzzy decision making including unequal
[33] W. ROdder and H.J. Zimmermann, Duality in fuzzy linear objectives, Fuzzy Sets and Systems 1(1978) 3-28.
programming, in: A.V. Fiacco and K.D. Kortanek, Eds., [45] R.R. Yager (Ed.), Fuzzy Sets and Possibility Theory (Per-
Extremal Methods and Systems Analysis (Springer, Berlin, gamon, New York, 1982).
1980) 415-427. [46] L.A. Zadeh, Fuzzy sets, Information and Control 8 (1965)
[34] M. Roubens, Pattern classification problems and fuzzy 338-353.
sets, Fuzzy Sets and Systems 1 (1978) 239-253~ [47] LA. Zadeh, A theory of approximate reasoning, Mem-
[35] K.D. Schwab, Ein auf dem Konzept der unscharfen orandum No. 4CB/ERL M 77/58, Berkeley, CA (1977).
Mengen basierendes Entscheidungsmodell bei mehrfacher [48] LA. Zadeh, Fuzzy sets as a basis for a theory of possibil-
Zielsetzung, Dissertation RWTH Achen (1982). ity, Fuzzy Sets and Systems 1 (1978) 3-28.
[36] J. Schwartz, The permicious influence of mathematics in [49] L.A. Zadeh, K.S. Fu, K. Tanaka and M. Shimura (Eds.),
science, in: E. Nagel, P. Suppes and A. Traski, Eds., Logic, Fuzzy Sets and Their Applications to Cognitive and Decision
Methodology and Philosophy of Science, (Stanford Univer- Processes (Academic Press, New York, 1975).
sity Press, Stanford, 1962). [50] H.J. Zimmermann, Description and optimization of fuzzy
[37] G. Sommer, Fuzzy inventory scheduling, in: [24] sets, lnternat. J. General Systems 2 (1976) 209-215.
3052-3060. [51] H.J. Zimmermann, Fuzzy programming and linear pro-
[38] J. Stamp, quoted by D. White, in: Decision Methodology gramming with several objective functions, Fuzzy Sets and
(Wiley, New York, 1975). Systems 1 (1978) 45-55.
[39] P.J. Sticha, J.J. Weiss and M.L. Donnell, Evaluation and [52] H.J. Zimmermann, Testability and meaning of mathemati-
integration of imprecise information, M.S. 2093, Journal cal models in social sciences, Math. Modelling 1 (1980)
Supplement Abstract Service, Amer. Psychol. Ass., 123-139.
Washington, DC 20036 (1979). [53] H.J. Zimmermann and P. Zysno, Latent connectives in
[40] U. Thole, H.J. Zimmerman and P. Zysno, On the suitabil- human decision making, Fuzzy Sets and Systems 4 (1980)
ity of minimum and product operators for the intersection 37-51.
of fuzzy sets, Fuzzy Sets and Systems 2 (1979) 167-180. [54] H.J. Zimmermann and P. Zysno, Decisions and evalua-
[41] P.P. Wang and S.K. Chang (Eds.), Fuzzy Sets (Plenum, tions by hierarchial aggregation of information, Fuzzy Sets
New York, 1980). and Systems 10 (1983).

You might also like