-.

1.
Y

N A S A CONTRACTOR
REPORT

DECISION-MAKING I N A
FUZZY ENVIRONMENT

by R. E. Bellman und L. A. Zudeh

$3 Prepared by
8
-5 UNIVERSITY OF CALIFORNIA
,&
4
W!
Berkeley, Calif. 94720
'",, for

N A T I O N A LE R O N A U T I CA SNS DP A CA ED M I N I S T R A T I O N W A S H I N G T O N , D. C. MAY 1970

 TECH LIBRARY KAFB, NM

C&W'O'
J NASA CR-1594

/' DECISION-MAKING IN A FUZZY ENVIRONMENT

J L/
By R. E. Bellman and L. A. Zadeh

..
, 7 . !, ,

,i/' .( -.
" t
c -
Issued by Originator as Report No ERL-69-8
'i

pJ
Prepared under Grant No. NGL-05-003-016 by
ELECTRONICS RESEARCH LABORATORY ; J C ~ C
/ - . ,..iL:L'!4;i<c'L:&

Ltc,3'
UNIVERSITY OF CALIFORNIA I
Berkeley,
Calif. 94720 J,,i! za", F b ,

for
NATIONAL AERONAUTICS AND SPACE ADMINISTRATION

For sale by the Clearinghouse far Federal Scientific and Technical Information
Springfield, Virginia 22151 -
CFSTI price $3.00
 ABSTRACT

By decision-making in a fuzzy environment is meant a

decision process in which the goals and/or the constraints, but not

necessarily the system under control, are fuzzy in nature. This means

that the goals and/or the constraints Constitute classes

Of alternatives

whose boundaries are not sharply defined.

An example ofa fuzzy constraint is: "The cost of

A

a," where a is a specified

should not be substantially higher than

constant. Similarly, an example of a fuzzy goal

is: "x should be in

the vicinity of x ' I where x is a constant. The underlined words are

0, 0

the sources of fuzziness in these examples.

Fuzzy goals and fuzzy constraints can be defined pre-

cisely as fuzzy sets in the space of alternatives. A fuzzy decision,

as an intersection of the given goals and con-

then, may be viewed

straints. A maximizing decision is defined as the set of points in

the spaceof alternatives at which the membership function of a fuzzy

decision attains its maximum value.

The use of these concepts is illustrated by examples

involving multistage decision processes in which the system under

-iii-
control is either deterministic or stochastic.By using dynamic pro-

a maximizing decisionis reduced to the

gramming, the determination of

solution of a system of functional equations. A reverse-flow technique

is described for the

solution of a functional equation arising
in con-

nection witha decision process in

which the termination time is defined

implicitly by the condition that the process stops

when the system under

control enters a specified set of states in its state space.

-iv-
 CONTENTS
Page No.

1. INTRODUCTION 1

2. A BRIEF TO FUZZY SETS

INTRODUCTION 5

3. FUZZY GOALS, CONSTRAINTSAND DECISIONS 15

4. MULTISTAGE DECISION PROCESSES 25

5. A FUZZY ENVIRONMENT
STOCHASTIC SYSTEMS IN 31

6. SYSTEMS WITH IMPLICITLY DEFINED TERMINATION TIME 36

7. CONCLUDING REMARKS 51

REFERENCES 53

FIGURES

-V-
 1. INTRODUCTION

Much of the decision-making in the real world takes place

in an environment in which the goals, the constraints and the conse-

To deal quantita-
quences of possible actions are not known precisely.

tively with imprecision, we usually employ the concepts and techniques

of

probability theory and, more particularly, the tools provided by decisSon

In so doing, we are tac-

theory, control theory and information theory.

whatever
itly accepting the premise that imprecision - its -
nature c.an

be equated with randomness. This, in our v+ew, is a questionable as-

sumption.

Specifically, our contention is that there is a need for

differentiation between randomness and fuzziness, with’ the latter being

a major source of imprecision in many decision processes. By fuzziness,

we mean a type
of imprecision which is associated with the use of fuzzy
sets,’”that is, classes in which there is no sharp transition from

membership to non-membership. For example, the classof green objects

is a fuzzy set. So are the classes of objects characterized by such

commonly used adjectives as large, small, substantial, significant,

important, serious, simple, accurate, approximate, etc. Actually, in

sharp contrast to the notion of a class or a set in mathematics, most of

the classes in the real world do not have crisp boundaries which separate

those objects which belong to the classes in question from those which

do not. In this connection, itis important to note that, in the dis-

is several
course between humans, statements such as “John inches taller

.
than Jim," "x is much larger than y, "Corporation
X has a bright

future, "the stock market has suffered or-

a sharp decline, convey inf

of the underlined words. In

matlon despite the fuzziness of the meaning

fact, it may be argued that the main distinction between human intel-

ligence and machine inte1,ligence lies in the ability

of humans -- an
ability which present-day computers do not -- to manipulate
possess

fuzzy concepts and respond

to fuzzy instructions.

What is the distinction between randomness and fuzziness?

Essentially, randomness has to do with unkertainty concerning membership

or non-membershipof an object in
a non-fuzzy set. Fuzziness, on the

other hand, has to do with classes in which there may be grades of mem-

To
bership intermediate between full membership and non-membership.

X has a modern
illustrate the point, the fuzzy assertion "Corporation

outlook" is imprecise by virtue of the fuzziness of the terms "modern

outlook." On the other hand, the statement "The probability that Cor-

poration X is operating at
a loss is0.8," is a measure of the uncer-

X in the non-fuzzy class

tainty concerning the membership of Corporation

of corporations which are operating a at

loss.

Reflecting this distinction, the mathematical techniques

for dealing with fuzziness are quite different from those of probability

theory. They are simpler in many ways because to the notion of prob-

ability measure in probability theory corresponds the simpler notion of

membership function in the theory of fuzziness. Furthermore, the cor-

2
respondents ofa +b and ab, where
a and b are real numbers, are the

simpler operations Max(a,b) and Min(a,b). For this reason, even in

those cases in which fuzzinessa decision

in process can be simulated by

a probabilistic model, it is generally advantageous to deal with it

through the techniques provided by the theory of fuzzy sets rather than

through the employment of the conceptual framework of probability theory.

Decision processes in which fuzziness enters in one way

or another’canbe studied from many points of 3’4

view.
’5 In the present

note, our main concern is with introducing three basic concepts: fuzzy

goal, fuzzy constraint and fuzzy decision, and exploring the application
of these concepts to multistage decision processes in which the goals or

the constraints may be fuzzy, while the system under control may be eith

-
deterministic or stochastic but not fuzzy. This, however, is not an

intrinsic restriction on the applicability of the concepts and techniques

described in the following sections.

Roughly speaking, bya fuzzy goal we mean an objective

which can be characterized aasfuzzy set in an appropriate space.

To

illustrate, a simple example of

a fuzzy goal involving
a real-valued

variable x would be: “x should be substantially larger than

100.”

Similarly, a simple exampleof a fuzzy constraint would be: “x should

20-25.’’
be approximately in the range The sources of fuzziness in these

statements are the underlined words.

3
 A less trivial example is provideda deterministic
by

discrete-time system characterized by the state equations

xn+l = xn + un , n=0,1,2,. ..
where x and u denote, respectively, the state and input at ntime
and
n n
in which for simplicity xn and un are assumed to be real-valued. Here

a fuzzy constraint on the input may be

where the wavy bar under

a symbol plays the role a of
fuzzifier, that is,

a transformation which takes

a non-fuzzy set into
a fuzzy set which is

approximately equal to it. In this instance, u < 1, would read "un

:
n
should be approximately less than or equal 1, and
to the effect of the

fuzzifier is to transform the non-fuzzy set

n-
- -
1 < u < 1 into a fuzzy set
-1 - un 5 1. The way in which the latter set can be a precise
given mean-
w u

ing will be discussed in Section 2.

Assume that the fuzzy goal to

is make x approximately
3
x = 1. Then, the problem is
equal to 5, starcing with the initial state
.

to find a sequence of inputsu ul, u2 which will realize the specified

0'

goal as nearly as possible, subject to the specified constraints

u on
0' ul,

In what follows,
we shall consider in greater detail
a few

It should be stressed that our

representative problems of this type.

limited objective in the present paper is to draw attention to problems

involving multistage decision processes in

a fuzzy environment and

4
suggest tentative ways of attacking them, rather than to a develop
gen-

eral theoryof decision processes in which fuzziness and randomness may

enter.in a variety of ways and combinations. In particular,

we shall not

concern ourselves with the application to decision-making of the concept

of a fuzzy -
algorithm3 a concept which may be of use in problems which

are less susceptible to quantitat'ive analysis than those considered in

the sequel.

a brief summary of the

For convenience of the reader,

is provided in the following section.

basic properties of fuzzy sets

2.
""
A Brief
~ .. .
Introduction
~- ~~~
to Fuzzy Sets
~~

Informally, a fuzzy set is

a class of objects in which

there is no sharp boundary between those objects that belong to the class

and those that do not.A more precise definition may be stated as fol-

lows.

Definition. Let X = (x} denote a collection of objects (points) denoted

generically by x. Than a fuzzy set A X is a set of ordered pairs

A = {(x,pA(x))} , x E X (1)

of x
where pA (x) is termed the grade of membership A, and vA: X +. M

is a function fromX to a space M called the membership space. When

M
contains only two points,
0 and 1, A is non-fuzzy and its membership

5
function becomes identical with the characteristic function
a non- of

fuzzy set.

In what follows, we shall assumeM is

that
the interval

[0,1], with 0 and 1 representing, respectively, the lowest and highest

grades of membership. (More generally,

M can bea partially ordered set

or, more particularly,

a lattice6” .) Thus, our basic assumption is that

a fuzzy setA - despite the unsharpness of its -

boundaries
can be de-

fined precisely by associating with each object

x anumber between0 and

1 which represents its grade of membership

A. in

Example Let X = (0,1,2,. .. I be the collection of non-negative integers.

of “several objects” may be defined (sub-

In this space, the fuzzy Aset

jectively) as the collection of ordered pairs

A = {(3,0.6) ,(4,0.8), (5,1.0), (6,1.0), (7,0.8), (8,0.6)) (2)

A (x))
( 2 ) we list only those pairs (x,v
with the understanding that in

in whichLI (x) is positive.

A

Comment, It should be noted that in many practical situations the member-

\
ship function,LIAy has to be estimated from partial information about it,

such as the values which it takes a over

finite set of sample points

xl, ...,%. When A is defined incompletely- and hence only approximately -

in this fashion,
we shall say that it is partially defined by exempli-

6
fication. The problem of pA from.
estimating knowledge of the
the. setof

pairs (xl,pA(xl)) ,...(%,pA(%)) is the problem of -

abstraction
a prob-

lem that playsa central rolein pattern recognition. ” We shall not

concern ourselves with the solution

of this problemin the present

paper and will assume except

throughout - where explicitly stated to the

contrary - x in X.
that p (x) is given for all
A

For notational purposes, it is convenient toa have

de-

vice for indicating a fuzzy setA .is obtained

that froma non-fuzzy set

by fuzzifying the boundaries of the latter set. For this purpose,

we

shall employ a wavy bar undera symbol (or symbols) which define x. For
example, if A is the set.of real numbers between2 and 5, i.e., x = {x I
2 <
”
x < 5}, then A = (x I 2 < X 5 5)
- is a fuzzy setof real numbers which
“3

are approximately between

2 and 5. Similakly, A = {x I -
x = 5) or simply

.. of numbers which are approximately equal 5to

5 will denote the set . The

symbol - will be referred to aasfuzzifier.

We turn next to the definition of several basic concepts

which we shall need in later sections.

Normality: A fuzzy setA is normal if and only if pA= 1, that is,

Sup
(x)
x
the supremumof vA(x) over X is unity. A fuzzy set is subnormal if it is

not normal. A non-empty subnormal fuzzy set can be normalized by

dividing eachpA(x) by the factor S;p pA(x). (A fuzzy set A is empty if

and only if
p,(x) i 0. )

7
Equality: as A = B, if and onlyif
Two fuzzy sets are equal, written

VA
-- VB, that is,vA(x) = pB(x) for allx in X. (In the sequel, to
simplify the notation we shall omit when an equality or
the xargument

inequality holds for all values inofX.) x

Containment: A fuzzy set A is contained or is a subset ofa fuzzy s e t

1-IA -< PB.

B, written asA C B , if and only if In this sense, the fuzzy
Eat

of very large numbers is a subset of the fuzzy set of large numbers.

Complementation: A' is said to be the complementA if

of and only if

PA' = - pA. For example, the fuzzy sets: A =' {tall men} and A' =
{not tall men) are complements of one another if the negation
is "not"

interpreted as an operation which replaces

1-1 (x) with l-vA(x) for each
A
x inX.

A and B is denoted byA n B and is de-

Intersection: The intersection of

and B.
fined as the largest fuzzy set contained in Aboth The membership

function ofA r
l B is given by

where Min(a,b) = a if azb and Min(a,b)= b if a>b. In infix form, using

the conjunction symbol

A in placeof Min,(3) can be written more simply as

a
 The n o t i o n of i n t e r s e c t i o n b e a r s a close r e l a t i o n t o t h e

n o t i o n of the connective "and."Thus, i f A is t h e class of t a l l men and

B i s t h e class of f a t men9 t h e n A n B i s t h e class of menwho are both t a l l

-
and f a t .

Comment, It s h o u l d b e n o t e d . t h a t o u r i d e n t i f i c a t i o n of "and" w i t h (4)

i m p l i e s t h a t we are i n t e r p r e t i n g "and" i n a "hard" sense, that i s , we do

n o t a l l o w any tradeoff between p (x) and p (x) so long as p (x) > (x).
~ . l '
A B A
For example, i f pA(x) = 0.8 and pB(x) = 0.!j9 then pA n ,(x) = 0.5 so long

as pA(x) 0.5. I n some cases, a s o f t e ri n t e r p r e t a t i o n of "and'! vhich

corresponds to forming the algebraic product of pA(x)and pB(x) - rather

than the conjunction 1-1 (x) A 1-1 (x)
A B
- may b e c l o s e r t o t h e i n t e n d e d mean-

i n g of "and." From the mathematical as well as p r a c t i c a l p o i n t s o f view,

the identification of "and" w i t h A is p r e f e r a b l e t o its identification

with the product, except where A c l e a r l y d o e s n o t e x p r e s s t h e s e n s e in

whichonewants "and" t o b e i n t e r p r e t e d . F o r t h i s r e a s o n , i n t h a t f o l -

10~7s"and" will b e u n d e r s t o o d t o b e a hard "and" u n l e s s e x p l i c i t l y s t a t e d

t h a t i t should be interpreted as a s o f t "and"(in t h e s e n s e of corre-

sponding t o t h e a l g e b r a i c p r o d u c t ofmembership functions).

Union: The n o t i o n of the union of A of B i s d u a l t o t h e n o t i o n of i n t e r -

section. Thus, the union of A and B, denoted as A u B , i s d e f i n e d as t h e

smallest fuzzy set containing both A and B. The membership function of

A u B i s given by

9
where Max (a,b) = a if a b and Max (a,b) = b if a < b. In infix form,

using the disjunction symbol

veinplace of Max, we can (5) more
write

simply as

of the intersection, the union A of

As in the case and B

bears a close relation to the connective "or.Thus, if A = (tall men)

and B = (fat men), then A u B = (tall or fat men}. Also, we can differen-

tiate betweena hard "or:'1, which corresponds(6),toand a s o f t "or",

which corresponds to the algebraic of A and B, which is denoted by

A e B and is defined by
(9).

u and are related to one

It is easy to verify that

another by the identity

A and B is denoted by A3 and

Algebraic product: The algebraic product of

is defined by

Algebraic sum: The algebraic of A and B is denoted byA +3 B and is

defined by

10
 It is easyto verify that

1
A 0 B = (A'B')

v and II are associative

Comment. .It should be noted that the operations

and distributiverover on the

one$nother. other hand,
e (product) and+3(sum) are

associative but not distributive,

Convexity and concavity. LetA be a fuzzy setin X = R". Then A is

If f o r every pairof points (x,y) i n X, the membership

convex if and only

. for 0 X 51. Dually, A is concave if its complementA' is convex. It

I s easy to show that if A and B are convex,so is A!%. Dually, ifA and

B are concave,so is A m ,

Relation. A fuzzy relation,R, in the product XxY = { (x,y) 1, x

space E X,
y E Y, is a fuzzy seti n X x Y characterized bya membership function%

which associates with each ordered pair (x,y)

a grade of membership

pR(x,y) in R. More generally, an n-ary fuzzy relationa product

in space

11
x = x1 X X2 X.. ..xx"is a fuzzy set X characterized by an n-variate
in

membership function%(x1, ..., x,), xi E X

i
,i = 1, ..., n.
1
Example. Let X = Y = R1, where R is the real line(-my ). Then x>>y

is a fuzzy relationin R
2
. A subjective % in this case
expression for

might be: %(x,y) -

= 0 for x<y; uR(x,y) = (1+ (x-y) -2 )-1 for x>y.

Fuzzy sets induced by :X+Y be a mapping

mappings. fLet fromX = {x) to

Y = cy), with the image xofunder f denoted byy = f (x). Let A be a

X.
fuzzy set in Then, the mappingf induces a fuzzy setB in Y whose

membership function is given by

of points f
is taken over the set
where the supremum
-1(y) in X which are

mapped byf into y.

Conditioned fuzzy sets:A fuzzy setB(x) in Y = cy) is conditionedon x

on x as a parameter. This dependence

if its membership function depends

is expressed by uB(ylx) .

x ranges
Suppose that the parameter overa space X, so

x in X corresponds a fuzzy setB ( x ) in Y.

that to each Thus, we have a

mapping - characterized bypB(y Ix) - from X to the spaceof fuzzy sets

12
i n Y. Through t h i s mapping, any given fuzzy set A i n -X induces a ' fuzzy

set B i n Y which is defined by

where ll and y denotethe membership f u n c t i o n s of A and B, r e s p e c t i v e l y .

A B
In terms of A and v (13) may b e w r i t t e n more simply as

Note t h a t this equation i s analogous - but not equivalent - t o t h e ex-

pression for the marginal probability distribution of t h e j o i n t d i s t r i b -

t i o n of two random v a r i a b l e s , w i t h p B ( y ]x) p l a y i n g a r o l e a n a l o g o u s to

t h a t of a c o n d i t i o n a l d i s t r i b u t i o n .

Decomposability: L e t X = {x), Y = {y)and l e t C b e a fuzzy set i n t h e

product space Z = XxY defined by a membership function yc(x,y). Then C

is decomposable a l o n g X and Y i f and only i f C admits of the represen-

tation

c = A ~ B

or e q u i v a l e n t l y
where A and B are fuzzy sets w i t h membership f u n c t i o n s o f t h e form

pA(x) and p g ( y ) ,r e s p e c t i v e l y . (Thus, A and B are c y l i n d r i c a lf u z z y

sets i n Z.) The same h o l d s f o r a fuzzy set i n t h e p r o d u c t ofany finite

number of spaces.

P r o b a b i l i t y offuzzyevents: L e t P be a p r o b a b i l i t y measureon Rn. A

fuzzy eventlo A i n Rn i s d e f i n e d t o b e a fuzzy subset A of Rn whose mem-

bership function, llA, i s measurable. Then, t h e p r o b a b i l i t y of A is de-

f i n e d by t h e L e b e s g u e - S t i e l t j es i n t e g r a l

Equivalently,

where E denotes the expectation operator. In the case of a non-fuzzy s e t ,

(16) reduces to t h e c o n v e n t i o n a l d e f i n i t i o n of t h e p r o b a b i l i t y o f a non-

fuzzy event.

This concludes our brief introduction to some of t h e b a s i c

c o n c e p t sr e l a t i n gt of u z z y sets. I nt h ef o l l o w i n gs e c t i o n , we s h a l l u s e

these concepts as a b a s i s f o r d e f i n i n g t h e c e n t r a l n o t i o n s of g o a l , con-

s t r a i n t and d e c i s i o n i n a fuzzy environment.

14
 3. Fuzzy Goals,'Constraints and Decisions

In the conventional approach to decision-making, the

principal ingredients of
a decision process (a) A set of alterna-
are

tives; (b) a set of constraints on the choice between different alter-

natives; and (c)a performance function which associates with each

alternative the gain (or loss) resulting from the choice of that alter-

native .

When t7e view a decision process from the broader perspec-

tive of decision-making ainfuzzy environment,a different and perhaps

more natural conceptual framework suggests itself. The most important

Eeature of this framework is its symmetry with respect and goals

to

constraints - a symmetry, which erases the differences between them and

makes it possible to relate ainrelatively simple way the concept

a of

decision to those of the goals and constraints of

a decision process.

More specifically, X = {x) be a given

let setof alterna-

a &,
tives. Then, a fuzzy goal or simply G, in X will be identified

G in X.
with a given fuzzy set X = R'
For exampleg if (the real line),

in words as
then the fuzzy goal expressed x should be substantially
I'

1
larger than10" might be represented abyfuzzy setin R whose member-

ship function is (subjectively) given by

I

PG(X) = 0 , x<10

15
 = (1 + (x-10) -2)-1 , XLlO.

Similarly, the "x should

goal be in the vicinity
of 15" might be repre-

sented bya fuzzy set whose membership function is of the form

Note that bothof these sets are convex in the of (11).

sense

In the conventional approach, the performance function

associated with a decision process serves to define
a linear ordering

pG(x),
on the set of alternatives. Clearly, the membership function,

of a fuzzy goal serves the

*
same purpose and, in fact, may be derived

from a given performance function aby

normalization which leaves the

linear ordering unaltered. In effect, such normalization provides

a

common denominatorf o r the various goals and constraints and thereby

make's it possible to treat them alike. This, as

we shall see, is one

of the significant of regarding

advantages the concept aofgoal - rath-
er than that of
a performance function as- oneof the principal compo-

nents of a conceptual framework for i n a fuzzy

decision-making environ-

ment.

* Assuming, of course, that

p takes values in
a linearly ordered set.
G
T-
I'.

In a similar manner, a fuzzy constraint or simply a

constraint, C, in X is defined to be a fuzzy setX.in For example, in

R1, the constraint"x should be approximately between
2 and lo," could

be represented by a fuzzy set whose membership function might be of the

fo m

where a is a positive number and m is a posftive even integer chosen in

such a way as to reflect the sense in which the approximation to the in-

= 4 mand a =
terval [2,10] is to be understood. For example, if we set

5-4, then at x
= 2 and x= 10 we have approximately
1~ (x) = 0.8, while
C

at x = 1 and x = 11, IJC (x) = 0.5; and at x= 0 and x= 12, uc(x) is ap-
proximately equal to0 . 3 .

An important aspectof the above definitions of the con-

cepts of goal and constraint is that both are defined as fuzzy sets in

the space of alternatives and thus, as will be elaborated upon below,

of a decision. By con-
can be treated identically in the formulation

trast, in the conventional approach to decision-making, a constraint

set is taken X, where-

to be a non-fuzzy set in the space of alternatives

as a performance function is a functionX to

fromsome other space.

of the conventional approach, the use of

Nevertheless, even in the case

Lagrangian multipliers and penalty functions makes it apparent that

there is an intrinsic similarity between performance functions and con-

17
 *
straints. This similarity - indeed identity - is made explicit
in our

formulation,
I

As an illustration, suppose that we have a fuzzy

G goal

and a fuzzy constraint

C expressed as follows:

G: x should be substantially larger than

10, with pG(x)

given by (17)

c: x should be in the viclnity 15,

of with p C (X) ex-

pressed by (18).

Note that G and

C are connected to one another by the

connective e.Now, as was pointed out in Sec.2, corresponds to

the intersection of fuzzy sets. This implies that in the example under

consideration the combined effect

of the fuzzy goal G and the fuzzy

constraint C on the choiceof alternatives may be represented by the

intersection G n C . The membership -function

of the intersectionis given

by

or more explicitly

-2 -1 -4 -1
(x) = Min ((1
l’lGr?C:
+ (x-10) ) , (1 + (x - 15) ) ) for x > 10
-
12.
*For a very thorough discussionof these points see Ref.

18
Note thatG n C is a convex fuzzy set since both
G and C are convex

fuzzy sets.

Turning to the concept of a decision, we observe that,

intuitively, a decision is basically a choice or a set of choices drawn

from the available alternatives. The preceding example suggests that a

fuzzy decision or simply a decision be defined as the fuzzy set of alter-

natives resulting from the intersection of the goals and constraints.

We formalize this idea in the following definition.

and a fuzzy con-

Definition. Assume that we are given a fuzzy Ggoal
straint C in a spaceof alternatives X. Then, G and C combine to form a

G and
decision, D, which is a fuzzy set resulting from intersection ofC.

In symbols,

D = G n C

and correspondingly

Lf) = ?JG A PC

The relation between

G, C and D is depicted in Fig.1.

19
 More generally, suppose that wen goals
have G190.09 G
n
and m constraints C190.e3 Cm. Then, the resultant decision is the inter-

section of the given goals

G~ 9 G and
~ the
~ given
~ 9 constraints C1OO..s
n
Cm. That is,

D = ~~nG~ n... n G n
n
cp c2n o o o n cm

and correspondingly

Note that in the above definition

a decision,
of the

D inforexactly
goals and the constraints enter into the expression

the same way. This is the basis for our earlier statement concerning

the identity of.the roles of goals and constraints in our formulation

of decision processes ainfuzzy environment.

Coment. The definition of

a decision as the intersection of the goals

our interpretation of "and" in the "hard"

and constraints, reflects

sense of (4) If the interpretation of "and" is left open, we shall

say thata decision - viewed asa fuzzy -
set is a confluence of the

goals and the constraints. Thus, "confluence" acquires the meaning of

"intersection" when "and" is interpreted in the of (4); the mean-

sense

ing of "algebraic product" when "and" is interpreted in the sense of

20
(8); and may be assigned other concrete meanings when a need f o r a spe-

cialinterpretation of "and" arises. Inshort, a b r o a dd e f i n i t i o n of t h e

concept of d e c i s i o n may b e s t a t e d as:

Decision = ConfluenceofGoals and C o n s t r a i n t s

As a n i l l u s t r a t i o n o f ( 2 1 ) ' we s h a l l c o n s i d e r a very

simple example i n which X = €1'2 $ o . ,101and G19 G2, C and C2 are de-
l
f i n e d below.

2 3 4 5 6 7 8 9 10

0.1 0.4 0.8 1.0 0.7 0.4 0.2 0 0

0.6 1.0 0.9 0.8 0.6 0.5 0.3 0 0

0.6 0.9 1.0 0.8 0.7 0.5 0.3 0.2 0.1

0.4 0.6 0.7 0.9 1.0 0.8 0.6 0.4 0 .'2

Forming the conjunction of p and

pc we o b t a i n
the
G1
$ 'G2 ' C' 1 2
following table of values for pD(x)

9 10
X I 1 2 3 4 5 6 7 8
0 0.1 0.4 0.7 0.8 0.6 0.4 0.2 0 0

Thus the d e c i s i o n i n t h i s c a s e i s t h e f u z z y set

D = {(2,O.1),(3,O.4),(4,O.7)9.(5,O.8)9(6yO~6)y(7,O~4)~(8,O~2~~

21
 I

Note that nox in X has full (that is, unity grade) membership
D. in

This reflects,of course, the fact that the specified goals and con-

straints conflict with one another, ruling out the existence

of an

alternative which fully satisfies all of them.

The concept ofa decision asa fuzzy set in the space of

alternatives may appear at first to be somewhat artificial. In fact it

is quite natural, since

a fuzzy decision may be viewed as an instruction

whose fuzziness is
a consequence of the imprecision of the given goals

and constraints. Thus, in our example, G1,

G2, C
1 and C 2 may be respec-
''x should be closeto 5," "x should be
tively expressed in words as:

4" and "x should be close to

close to 3 , " "x should be close to 6". The

decision, then, is to choose

x to be close to5. The exact meaning of
11
close" in each case is given by the of
values
the corresponding member-

ship function.

How shoulda fuzzy instruction such as "x should be close

to 5" be executed? Although there does not appear to

a universally
be
*
valid answer to questions of this type, it is reasonable in many in-

stances to choose thatx or x's which have maximal grade

of membership

in D. In the case of our example, this would 5.

x = be

* 3.
The executionof fuzzy instructions is discussed in Ref.

22
 M
More generally, let D be a fuzzy decision and let D be

is, the non-fuzzy set of points X in

its core, that at which the maximum

of pD(x) over X, if it exists, is attained. Then, we shall say that any

x in DM constitutes a maximizing decision. In other words,

- a maximizing

decision is simply any alternativeX which

in maximizespD(x), e.g., x = 5

in the foregoing example. Note that ina sufficient

Rn condition for the

uniqueness of a maximizing decision is that D be a strongly convex fuzzy

set.
*

D as the intersection or
In defining a fuzzy decision -
- of the goals and constraints, we are
more generally, as the confluence

tacitly assuming that all of the goals and constraints that enter into

D are,in a sense,of equal importance. These are some situations, how-

ever, in which some of the goals and perhaps some of the constraints are

D might be expressed
of greater importance than others. In such cases,

as a convex combinationof the goals and the constraints, with the

weighting coefficients reflecting the relative importance of the con-

stituent terms. More explicitly, we may express

1-1D (x) as

n m

* A fuzzy set is strongly convex if it is convex and its membership

function is unimodal.

23
where the Q
i
and 8J. are membership functions such that

n m
x . Qi(X) + c Bj(X) =1
i=1 j=1

of and 8 . (x) can be

Subject to this constraint, then, the valuesa (x)
i J
chosen in such
a way as to reflect the relative importance ...,
ofG G1,
n
and C1, ...,Cm' m = n = 1, it is easy to verify that
In particular, if

(22) can G u C in
generate any fuzzy set which is contained and contains

GnC. Note that (22) resembles the familiar artifice

of transforming a

vector-valued criterion intoa scalar-valued criterion by forming

a linear

of the vector-valued objective function.

combination of the components

So far, we have restricted our attention

to situations in

which the goals and the constraints are fuzzy sets in space of
X, the

alternatives. A more general case which is of practical interest is one

in which the goals and the constraints are fuzzy sets in different

spaces. Specifically, let f be a mapping from X = {x) to Y = {y), with

x representing an input (cause) and

y, y= f(x), representing the corre-

sponding output (effect).

Suppose that the goals are defined as fuzzy

G sets
1'""'
n m
...,
G in Y while the constraints C1, C are defined as fuzzy sets X.in
*
Now, given a fuzzy set G in Y, one can readily finda fuzzy setG in
i i

24
X which induces Gi in
Y. Specifically, the membership function
of Gi
*
is given by the equality

,i = l,...,n

The decisionD; then, can be expressed as the intersection

of
* *
G1 ,...,G and C1,. ..,Cm* Using (23), we can express
1-1 (x) more ex-
n D
plicitly as

where f: X-tY. In this way, the case where the goals and the constraints

are defined as fuzzy sets in different spaces can be reduced to the case

where they are defined in the same space. We shall ( 2find

4 ) of use in

the analysis of multistage decision processes in the following section.

4. i.-
Multistage Decision
~ ~ Processes

As an application of the concepts introduced in the pre-

ceding sections, we shall consider a few basic types

of problems involv-

ing multistage decision-making in a fuzzy environment. It should be

stressed that, in what

follows, our main purpose is to illustrate the use

of the concepts of fuzzy goal, fuzzy constraint and fuzzy decision, rather

than to develop a general theory of multistage decision processes in

25
which fuzziness enters in one way or another.

For simplicity we shall assume that the system under con-

trol, A, is a time-invariant finite-state deterministic system in which

X =
t = 0,1,2,..., ranges over a finite set
the state, x at time t,
ty
{o1,.. .,u
n
1 , and the inputut, ranges over a finite Uset
= {al,.. ,a 1 .
m
.
The temporal evolution Aofis described by the state equation

x
t+l
= f (xt, Ut) , t = 0,1,2,. .. (25)

fromto X.
in which f is a given function XxU Thus, f(x u ) represents
t' t
the successor state of x for input u Note, that if f is a random
t t'
function, thenA is a stochastic system whose state at time
is at+l

X, P ( x ~ + ~Xt,ut), which is conditioned

probability distribution over
I
A is a fuzzy
on x and u Analogously, if f is a fuzzy function, then
t t'
2
t + 1 is a fuzzy set conditioned on x and u
system whose state at time
t ty
which means that it is characterizeda membership
by functionof the form

I xt,ut).
~ ( x ~ + ~
*
Since we will not be concerned with such systems in

the sequel, it will be understood that f is non-fuzzy unless explicitly

stated to the contrary.

We assume that at each time

t the input is subjected to

a fuzzy Ct , which
constraint is a fuzzy setU characterized
in by a

membership function 1-1 (u

t t
). Furthermore, we assume that the goal is a

* It should be noted that when we speak of a fuzzy environment, we mean

that the goals and/or the constraints are fuzzy, but not necessarily the
system which is under control.

26
 fuzzy setGN in X, which is characterized by a membership' function

p ,(x ), where N is the time of termination of the process. These

Gh
assumptions are common to most of the problems considered in the sequel.

1;
1':

/:I:
g
J Problem 1. In this case, the system is assumed to be characterized by
(25), with fa given non-random function. The termination time
N is
2 assumed to be fixed and specified, The initial state, xo, is assumed

a maximizing decision.
to be given. The problem is to find

Applying (20), the decision- viewed asa decomposable

I '8
fuzzy set'in U x U X ... X U, may be expressed at once as
1
;i

explicitly, in terms of membership functions, we have

where % is expressible as a function of.

x and u0,. .., uK1 through the
iteration of(25).

27
 Our problem, then,is to find a sequence of inputs

U u which maximizes p as given by (27). As is usually the case

0’”” N-1 D
in multistage processes, itis expedient to express the solution
In the

fo m

u = IT (x,)
t t
,t = 0,1,2,. ..,N-1 ,

where T is a policy function. Then, we can employ dynamic programming

t
M M
to give us both the
IT and the maximizing decision
u
t * 9 uN-l’

More specifically, using( 2 6 ) and (25), we can write

Now, if y is a constant andy is any function of uGl,

we have the identity

Consequently, (28) maybe rewritten as

28
 where

’!

may be regarded
as the membership function
of a fuzzygoal at time t =

N-1 which is induced by the given goal GN att =time

N.

On repeating this backward iteration, which is a simple

instance of dynamic programming,

we obtain the set
of recurrence equations

which yield the solution to the problem. Thus, the maximizing decision
M is given by the successive maximizing values
U
Of”” UN-19 %-v ofin
(311, with s-vdefined as a function
of xN.-V ’ V = l,..., N.

29
Example. As a simple illustration, consider
a system with three states

u1, cJ2, u3 and two inputs a1 and a 2 . Assume N = 2 for simplicity. Let

the fuzzy goal at t = 2 be defined bya membership

time function
G2
whose values are given by

at0 and t = 1 be defined

Furthermore, let the fuzzy constraintst =

respectively by

f in (25) is
The state transition table which defines the function

assumed to be

a1 U
3
U
1

a2 *2
U
3

Using ( 3 0 ) , the membership function of the fuzzy goal

induced att = 1 is found to be

30
and the Corresponding maximizing decision
is given by

Similarly, for t = 0

and

Thus, i f the initial state (at

t = 0) is GI, then the maximizing deci-

sion is cc ,a and the corresponding valueof )J is 0.8.

2 1 G2

Next, we turn to a more general multistage decision

is stochastic, while the goal

process in which the system under control

and the constraints are fuzzy.

5. Stochastic Systems ina Fuzzy Environment

As in the proceding problem, assume that the termination

time N is fixed and that an initial state
x o i s specified. The system

by a conditional probability function

is assumed to be characterized

P(X~+~
(xt,ut). The problem ist o maximize the probability
of attainment

31

"
 C ,
N, subject to the fuzzy constraints
of the fuzzy goal at time
0
...,
cN-l

GN is regarded as a fuzzy
If the fuzzy goal event'' in X,

then the conditional probability of this event givenXN-1 s-1is

expressed by

where E denotes the conditional expectation and

1-1 is the membership
G
function of the given fuzzy goal.

We observe that (32) expresses Prob (GN I xNml, or,

equivalently, E lJ (x ), as a function of and s-~, just as in the

GN
preceding problem1-1 N ( x N ) was expressed as a function
of 3 - 1 and
G
via (25). This implies that E 1-1 N(xN) can be treated in the same
G
way as N ( ~ N >was treated in the non-stochastic case, thus making it
G
possible to reduce the solution of the problem under consideration to

that of the preceding problem.

More specifically, the recurrence equations (31) are

replaced by

32
where, as b e f o r e , 1-1 (x ) . d e n o t e tsh e membership f u n c t i o notfh e
G N-V P-v
fuzzy goal a t t = N -
V induced by t h e f u z z y g o a l at t = N - v + 1 ,
V = 1, ..., N . Theseequationsyield a s o l u t i o n t o theproblem, as i s

i l l u s t r a t e d by t h e f o l l o w i n g example.

Example. As i n t h e p r e c e d i n g example, we assume t h a tt h es y s t e mh a s

t h r e e states
1’
u2, d
3
and two i n p u t s a a
1.’ 2’
N i s assumed t o be
equal

t o 2 , and t h e p r o b a b i l i t y f u n c t i o n p(x u ) i s given by the follow-

t + J X t ’t
i n g two tables,correspondingtout = al, and u = a respectively.
t 2’

I. u = a 11. u = a
t 1 t 2

U U
Ol 2 u3 u1- 2 u3
%x
0.8 0.1 0.1 0.1 0.9 0

0 0.1 0.9 0.8 0.1 0.1

O2 O2

0.8 0,1 0.1 0.1 0 0.9

“3 u3

The entries i n these t a b l e s are t h e v a l u e s of P ( X ~Ixt

+ ~, u t ) .

33
Thus , t h e e n t r y 0.8 i n t h e p o s i t i o n (o1, a2) . i n t h e f i r s t t a b l e s i g n i f i e s
t h a t i f thesystem i s i n s t a t e 0 a t time t andinput cx1 is a p p l i e d ,t h e n
1
with probability 0.8 t h e s t a t e a t t i m e t + 1w i l l b e a2.

The fuzzy goal a t t = 2 i s assumed t o b e t h e same as i n

thepreceding example, t h a t i s

L i k e w i s e ,t h ec o n s t r a i n t s are assumed t ob et h e same. Thus

Using ( 3 3 ) , we compute Ep 2(x2) as a f u n c t i o n of x and

1.
G
Tabulatingtheresults, we have
u1

a2
I 0.93 0.42 0.75

Next, using (33) with v = 1 andcomputing p (x ) w e

G1
obtain

34
 which correspond to the following values
of the maximal policy function

!
The final iteration v = 2 yields
with

a 0.62 0.62 0.62

1

a 0.8 0.62 0.60

2

The values of l~ in (61) represent the probabilitiesof

GO
attaining the given goal tat= 2 starting with 0 , a2 and 0
3
, respec-
tively, assuming that the inputs are determined by the maximal policy

t
function IT
t’
that is, u = IT^ (x,) ( t = 0, 1, xt:= 01, 02, a,, ut =

CLCL ) whose values are givenin (60) and ( 6 2 )

1’ 2

Comment. It should be noted that when the fuzzy goal at N time

is

defined in such a way that the probability of attaining it is small

for all values of%-1 and N-1

u it’ may be necessary to normalize the

35

I: .
fuzzy goal induced at N-1
timebefore finding its intersection with CN-l,

for otherwise the decision would be unifluenced by the constraints.

To be

consistent, such nomalization may have to be carried out at each stag

the decision process. Althoughwe shall not dwell further upon this aspect

of the problem in the present paper, it should be emphasized that it is

no means a trivial one and requires a more thorough analysis.

6. Systems With Implicitly Defined"Te-rgination Time

In the preceding cases,

we have assumed that the termi-

nation time,N, is fixed a priori. In the more general case we

which

shall consider in this section, the termination time is assumed to be

determined implicitly by a subsidiary condition of thex &form

C , where
*
of X termed the terminationset. Thus,
T is a specified non-fuzzy subset
the process terminates when the state
of the system under control enters,

for the first time, a specified subset of the state space. In this case,

setT
the goal is defined as a fuzzyG in , rather X.
than in

A,
More concretely, assume that the system under control,

is a deterministic system characterized by a state equation of the form

X
t+l
= f(x u )
t' t
, t = 0,1,2,. .. (34)

where x
t
ranges overX = {u1, .. ,G ,O
R R/bla***' n
u 1, in which =T{at+l,. ..,
0 }>constitutes the termination set.
As before, f is assumed to
be a
n
*In its conventional (non-fuzzy) formulation,
this case plays an important
role in the theoryof optimal control and Markoffian decision processes.
Some of the more relevant papers on this subject are cited in of the list
references.
36
given function from
XxU to X, where U = {al, ...,am} is the range of

u t = 0,1,2,.
t’
.. . Note that if(Si is an absorbing state, that
is, a

state inT, then we can f aj) = ai for

write(0i, alla in U.
j

The fuzzy goal is assumed to

a subset
be ofT character-

1-1 (x ), where
ized by a membership function Nis the time at which
x ET,
G N t
with xt$ T for t< N . As for the constraints on the input, we assume for

simplicity that they are independent of time but not necessarily the

state, Thus, if A is in stated at time t, then the fuzzy constraint

i
on u is assumed to be represented by
a fuzzy setC(Oi) (orC(xt)) in U
t
ai. The membership functionof this setwill
which is conditioned on

be denoted by vC(utlxt).

Let x0 be an initial state in TI, where

T’ = {dl,..,,dk}

is the complement of T X.
in To each such initial state will correspond

a decision,D(xo), given by

where the successive states x ..., XN-~,

1’
5 can be expressed as iter-
ated functionsof .x and uo, ..., %-1 through the state equation

(34). Thus

37
 ....... e . . . . . . . . . .

(26), the C’s in (35) should be regarded

Note that, as in

as fuzzy sets in the product space

U XU x...x Ux T . Another point

D(x 1 is uniquely determined by(35) for

that should be noted is that
0

each x with the understanding that D(xo) is empty if there

is no
0’

finite sequence of inputsuo, ...?uN-1which takes the initial state

x into T. T is not reachable from the

In this event, we shall say that
0
\
initial state.

From (35), we can readily derive a simpler implicit equa-

D(xo).
tion which is satisfied by Specifically, in virtue
of the time-

invariance of A and the time-independence of the goal and constraint

sets, (35) implies

for t = 0,192¶. .. . In particular,

and hence (37) can be written as

38
or, using ( 3 4 ) ,

which is the desired implicit equation. Expressed in terms

of the member-

ship functionsof the sets in question, this equation assumes the follow-

ing form (for =t0 )

where the termination time

N is also a function of
.x and uo, ul, u2, ...
T
( 3 4 ) and the termination condition ExN
through the state equation ,
with 6 T, ..., %-1 6 T .

Now suppose that the successive inputs u uls * * . , u

N-1
IT , IT :
are determined by a stationary (time-invariant) policy function

T' +v , which x in T' an input u which

associates with each state
t t
should be applied toA when it is in state
x Thus,
t'

Since uo, ..., uN-1are determined by x

0
and n through

( 4 2 ) and the state equation (34), the membership functionD(xof)

0

'can be written
as p (x
D o
!IT>.Similarly, pc(u0 I x,) can be written as

vc(n(xo 1 /xo1, and pD(u1, ., %-llf(xo, uo)> as v,-,(f(xo,.rr(xo))IIT)

39
With these substitutions,(41) assumes the more compact form

issystem ofR equations (one for each value xo)

which in effect a of

in the %. This system of equations determinesFC as a function of .

x
D
for each-IT , with % = 0 if underIT the process
the understanding that

finite N such that

does not terminate, that is, there does not a exist

% E T. 1 - 1 ~= uG for states inT.

Furthermore, it is understood that

( 4 3 ) has a unique solution.

It is easy to demonstrate that

of states T' = {U1,

Specifically, by decomposing the set ..., u }
!L
into

disjoint subsets T'l,..,, Tk , where Ti , x = l,..., K, represents the

set of states from which A steps, it is readily seen

T is reachable in

( 4 3 ) corresponding to thex
that the equations in which are in T1 yield
0

uniquely the respective values

of 1-1 In terms of these, the equations
D o
in ( 4 3 ) corresponding to the x in T yield uniquely the values of
1-1
0 2 D
for x in T2. Continuing in this manner, all the
uD1s can be determined
0

(43)
uniquely by successively solving subsets of the system of equations

for the blocks of variablesT i ,in ..., T k .

For our purposes, it will be convenient to represent

a

policy -IT as a policy vector

40
whose ith component,- = 1,
i ...,ti, is the input which must be applied

when A is in state
0 . U = {al,.
Note that Tr((Ji) ranges over the set ..,
~1
m1 and thus that there are
mR distinct policiesin the policy space.

( 4 3 ) , let
With reference to the system of equations

be an n-vector, termed goal

the attainment vector, whose components are

an (corresponding
the valuesof the membership function of D at ul, ...,
to policy T ). It is naturalto define a partial ordering in the policy

space by the inequality

T ' is better than

which means that a policy or equal to a policyT" if

and only if p
D
(Oili'r') 2 pD(aiI?') for i = 1, ..., n. Then, a policy

T will be said to be maximal if and only

T is if
better than or equal

to every policy in the policy space.

Does there exist a maximal policy for the problem under

consideration? The answer to this question is in the affirmative. This

assertion can be proved rigorously,* but it will suffice for our purposes

*A proof for the case of a stochastic finite-state system is given in

Ref. 13.

41
to regard it as -a
a consequence of the alternation principle14. principle

of broad validity which in concrete cases can be asserted as a provable

'theorem.

Specifically, letIT'and IT" be two arbitrary policy vectors,

with p (IT')and p (IT")

being the Corresponding goal attainment vectors.
D D
Using IT'and IT" , let us construct a policy vector
IT in accordance with

the following rules:

for each component IT.of IT

1
,i= 1, ..., R . Then, according to the

alternation principle,IT IT'and T -

> IT" , that is, IT is better than

or equal to both IT'and IT" From this and the finitenessof the

policy space it follows at once that there exists a maximal policy.

From ( 4 3 ) it is a simple matter to derive a functional

equation satisfied by the goal attainment vector corresponding to the

maximal policy. Thus, let

= Max p (IT)
IT
D

x n matrix of zeros and ones whose ij th element is

and letP(IT) be an n

42
one i f and o n l y i f D = f (ai, n ( o i ) ) , t h a t is, t h e state u is the
j. j
immediate successorof 0
i
underpolicy T .

which i s t h e d e s i r e d f u n c t i o n a l e q u a t i o n f o r
VD
. A l t h o u g hd i f f e r e n t

indetail,equation ( 4 9 ) i s of t h e same g e n e r a l form as t h e f u n c t i o n a l

equations arising in the theory of Markoffian decision processes.

12 -27

Its s o l u t i o n , however, i s considerablysimplertoobtainbecauseofthe

d i s t r i b u t i v i t y of Max and A .

Specifically, l e t n
1
,..., r
7~ where r = m
R , d e n o t et h e

mR d i s t i n c tp o l i c yv e c t o r s . Then,on using v i np l a c e of Max, ( 4 9 )

becomes

Takingadvantage of t h e d i s t r i b u t i v i t y of v and A , and factoringlike

terms, we canput (50) i n t o a much simpler form which, written as a

system of e q u a t i o n s i n t h e components of pM
D
, reads
 M
where cc = r(ai) = input under policy IT in state oi ; pD (ai) = i th
j
component of the maximal goal attainment vector; f(a a ) = successor
i' j
state* of ai for inputa with f (ai,
j'
a.) = ai for i= ra+l,. ,n (that
J
..
is, fora in the termination set T); pC ( a j lai) = value of the member-
i
C in statea
ship function of the constraint for input a , with uc(a. loi) = 1
i 1 J
for i= R+l, ...,n; and for i= J!,+l,...,n, M
pD (ai) = pG(ai) =

value of the membership function of the given goal G 0at Thus, the:
i *
M .. uDM (ai) , I =
pD (cri), i = 1,. > R , are the unknowns in (51), while the

R+1 D . , ,PI,and the 1-1 ( a . I oi), i

c 3
= 1, ..., n ,j = 1, ..., m , are given

constants.

To make the solution of

(51) more transparent, it is help-

(51) by letting the unknowns (51)

ful to simplify the notation in in be

denoted by wi M
that is, wi = pD (ai) for i = 1, ... Furthermore,

let the product and plus symbols denote A and v, respectively. Then,

(51) can be written more compactly

in matrix form as

* in ( 4 9 ) are defined byP ( r ) .

Note that the successor states

44
 bik = of ui ;
0 if crk is not an immediate successor

bik = V b C(aP 10i1, where thea

P
ai into uk ;
are inputs which take
a
P

and

with the understanding that 1-1 (d. ) = 0 for states outside the termina-
G i
tion setT.

Having put (51) into the form of a linear equation

(52),

it is easy to show that(53) and hence (51) can be solved by iteration.

Specifically, let w O = (o,...,o> and

w = B w
S
+y , s = 0,1,2,. .. (54)

Then, by induction, the sequencew0 , w1 w

2
, .. is monotone non-decreas-

ing. For, assume that wk+l wk for some k. Using ( 5 4 ) , we have

and noting that'a 1. w

0
= 0 , it follows that us+' >
- Us for s = 0,1,2,. .. .

Since the sequence w

0
, w
1
, ... is monotone non-decreasing

and bounded froni above byw = (l,...,l), it follows that it converges to

45

I
the solutionof ( 5 2 1 , that is, to the first
fi components* of the maximal

goal attainment LI,,M .

vector Actually, a more detailed argument shows

that ( 5 4 ) yields the solution of

( 5 2 ) in not more than
fi iterations. In

T isif reachable from a state

essence, this follows from the fact that

in T ’ , then it is reachable inR or fewer steps.

To gain an intuitive insight into the above solution, it

( 4 9 ) to (51) with the aid of

is helpful to interpret the transition from

the state diagram of

A. A has five
Thus, for concreteness assume that

states, with transitions corresponding to various inputs shown in Fig.2.

In this diagram, the number associated with the branch leading

0
i from
a is the value of p ( a . 10 )
to its successor state via input
j C J i
States .
a4 and a6 are in the termination set and the corresponding values
of

p p ( a . l a ) cor-
(a.1) are shown alongside. The indicated values of the
G C J i
respond to the constraint sets

For the system in question, the state transition function

f(o i’aj) is given by the following table

46
 I
CLfi Ul “2 (5 3 “4 U5

CL a4 “3 U5 U4 U
1 5

a2 O2 (3 1 U4 ‘(55
cL2

From t h i s t a b l e , it is easytoconstructthe matrix P(IT) f o r anygiven

1: 1
policy.Forexample,for IT = ( a2, al, a2 ), w e have

P ( a2 , a1, C 2L ) =
0 1 0

0
1
0

The systemofequations(51) i s obtained by r e v e r s i n g t h e

d i r e c t i o n of f l o w i n e a c h b r a n c h ( s e e F i g . 3 ) a n d t r e a t i n g t h e s t a t e s i n

T, t h a t is, u and u as s o u r c e s ,w i t ht h e states i n T ’ , that is, U

4 5 1’
a2 and u
3
, p l a y i n gt h er o l e of r e c e p t o r s( s i n k s ) . From thediagram

shown i nF i g . 3 ,t h ee q u a t i o n si n (51)canbe w r i t t e n by inspection.Thus,

47
 in which A
Employing the simplified notation and v are
M
= pD
replaced by the product and sum, respectively, wiand (ai), i = 1,

2,3, the systemof equations (56) becomes

u = B w + y (57)

where

0 1 0

B = 0 1 0.8

0.7 0 0

Letting uo = ( o , o , o ) , we o5tain on first interationw1 =

(0.6, 0 , 0.8).

Subsequent iterations yield

u2 = (0.6, 0.8, 0 . 8 )

u3 = (0.8, 0.8, 0.8)

u4 = (0.8, 0.8, 0.8)

Thus, u3 = (0.8, 0.8, 0 . 8 ) is the solution of (57).

To visualize the iteration process, imagine that of

each

the sources in Fig.3 (which are the absorbing states in Fig.2) generates

balls of various diameters, witha

i y
i =R+l, ..., n, generating balls

48
of diameters ranging from0 to pG(ai). Furthermore, imagine that a

branch in Fig.2 which leaves state

(5 via input a , i s a pipe of diam-
i j
eter p (a),.I
c j
which can carry ballsof diameter 2~ (a Iu ) along the
c j i
reverse direction, that is, along the direction shown in Fig.3. Thus,

the diagram of Fig.3

may be visualized as a network
of pipes whose diam-

eters are indicated in the diagram and which can carry balls
of lesser

or equal diameter in the indicated directions. The states in the termi-

nation set (a and

4
0
5
1 play the roleof sources of balls of diameters up
to p (a ) and p (a ) respectively, while the remaining states
(a1’ a2
G 4
and a ) act as receptors. Because the absorbing states act as sources,
. 3
we shall refer to the method of solution described above as a reverse-

flow technique.

Now assume that it takes one unit of time for the balls

to travel froma node of the network of Fig.3 to another node.If we

start with no balls at al, 0 and 0 at time 0, then at time t = 1

2 3
a1’ 2 and
the maximum diameters of balls at will
be,
respec-
3
1 1 w2,
1 w 1) is the first iter-
tively , 3’
w
1’ w12 and w1 where w = (wl, 3
ate of (57). At time t = 2, the maximum diameters of balls will be

given by w2 and at time t = 3 by w

3 . Since it takes no more than

three units of time for any ball to travel from its source to any node

in the network, there will be no further increase in the size of balls

3
at each source upon further iteration. Thus, w gives the maximum

diameter of balls at each receptor node and hence is the desired

solution of (57).

49
 of ( 4 3 ) and the alternation
Turning to the illustration

principle, consider the policy vector 7~ = (al, al, all. For this 7~ ,
the systemof equations ( 4 3 ) becomes

In this case, u1 and a3 are in T'1 and 0

2
is in Ti .
Noting that vD (a4 1 IT)= vG (a,) = 1 and 1-1 (a5 I n) = vG (a5) = 0.8,

we find at once vD (01 I .rr) = 0.6 ; pD(u21 = 0.8 ; vD

IT) ('J31 IT)= 0.8
which is thed e s i r e d solution.

same computation for other

Carrying out the policy vectors,

we obtain the results tabulated below

(5 U
1 O2 3

0.6 0.8 0.8

0.6 0.6 0.6

0.6 0 0.8

0.6 0 0.6

0.8 0.8 0.8

0 0 0

0 0 0.8

0 0 0

50
1
1;

I'
Ascheck

IT'= (a,; al, a2) and

a

IT"
on the

=
alternation
principle,
take
let
us

(a1, a2, all. Using (47) leads toIT = (a1, a1,

1,;
!I Note that IT 2 IT'and IT 2 ITTT".
From inspection of the table, the maximal
policy is seen tobe (a,, al, al). which agrees with the result obtained

by iteration.

The approach to the solution

of problems involving implic-

itly defined termination time which we have described

in this section can

be extended to more complex decision processes in a fuzzy environment. In

particular, the technique employed (49)

for solving the functional equation

can readily be extended to fuzzy systems in a fuzzy environment. Furthermore,

- as in section4
( 4 3 ) and ( 4 9 ) can be extended also - to stochastic finite-

state systems. Because of limitations on space, we shall not consider these

cases in the present paper.

7. Concluding Remarks

The task of developing a general theory of decision-

making in a fuzzy environment is one of very considerable magnitude and

complexity. Thus, the results presented in this paper should be viewed

as merely a first attempt at constructing a conceptual framework for

such a theory.

There are many facets of the theory

of decision-making

in a fuzzy environment which require more thorough investigation, Among

51
 o f execution of fuzzy decisions; the way in
these are the question

which the goals and the constraints must be combined when they are of

unequal importance or are interdependent; the control of fuzzy systems

and the implementation of fuzzy algorithms; the notion of fuzzy feed-

on decision-making; controlof systems in which the

back and its effect

fuzzy environment is partially defined by exemplification; and decision-

making in mixed environments, that

is, in environments in which the

imprecision stems from both randomness and fuzziness.

52
 1:
r

References

1. t. A . Z a d e h", F u z z y S e t s , " I n f o r m a t i o na n dC o n t r o l ,v o l . 8 ,

p p . 3 3 8 - 3 5 3 J, u n e , 1965.

i:
> .
2. L. A . Zadeh,
"Toward a T h e o r y of F u z zSyy s t e m s ,I1

11; NASA CR-1432, September 1969

J
i

3. L. A. Z a d e h , " FIunzfzoyr m
C ao tnidtorno l ,

vo1.12, pp.99-102F
, e b r u a r y 1, 9 6 8

4. s. 5. L. C h a n gI, I F u z z yD y n a m i cP r o g r a m m i n ga n dt h eD e c i s i o n

M a k i n gP r o c e s s , "P r o c .3 dP r i n c e t o nC o n f e r e n c eo nI n f o r m a t i o n

S c i e n c e se n dS y s t e m s ,p p . 2 0 0 - 2 0 3 , 1969.

5. 5. FU a n d T . J . L i , "On t h e B e h a v i oorLf e a r n i n A
g utomata

a n d i t s A p p l i c a t i o n s , " T e c h . Rep. TR-EE 68-20, P u r d u eU n i v e r s i t y ,

L a f a y e t t e ,I n d i a n a , Aug. 1968.

6. J . Goguen, " L - f u z z yS e t s , "J o u r . Math. A n a la. n dA p p l . , vol.18,

p p . 1 4 5 - 1 7 4A
, pril, 1967.

7. J . G . Brown, l t F u z z yS e t so nB o o l e a n L a t t i c e s , " Rep. No. 1957,

B a l l i s t i c Research L a b o r a t o r i e s ,A b e r d e e n ,M a r y l a n d ,J a n u a r y , 1969.

53
8. A. B e l l m a n , R. K a l a b aa n d L. A . Z a d e h ," A b s t r a c t i o na n dP a t t e r n

C l a s s i f i c a t i o n , "J o u r .M a t h .A n a l .' a n dA p p l . ,v o 1 . 1 3 ,p p . 1 - 7 ,

J a n u a r y , 1966.

9. W. G. Wee, "On G e n e r a l i z a t i o no fA d a p t i v eA l g o r i t h m sa n dA p p l i c a -

t i o n o f t h eF u z z yS e tC o n c e p tt oP a t t e r nC l a s s i f i c a t i o n , "T e c h n i c a l

R e p o r t TR-EE-67-7, P u r d u eU n i v e r s i t y ,L a f a y e t t e ,I n d i a n a ;J u l y , 1 9 6 7 .

10. L. A. Z a d e h ," P r o b a b i l i t yM e a s u r e so fF u z z yE v e n t s , "J o u r . Math.

A n a la. n dA p p l .v, o l . 1 0p, p . 4 2 1 - 4 2 7A

, u g u s t1, 9 6 8 .

11. C. L. C h a n g ," F u z z yT o p o l o g i c a lS p a c e s o "J o u r .M a t h .A n a l y s i sa n d

Appl.~
, 0 1 . 2 4 p, p . 1 8 2 - 1 9 0 , 1968.

13. J . H. E a t o na n d L . A. Z a d e h ,' @ O p t i m a lP u r s u i tS t r a t e g i e si n Discrete-

S t a t eP r o b a b i l i s t i cS y s t e m s , "J o u r . Basic E n g i n e e r i n g (ASME), v o l . 8 4 ,

S e r i e s D, p p . 2 3 - 2 9M
, a r c h1, 9 6 2 .

14. J. H. E a t o na n d L. A. Z a d e h , "An A l t e r n a t i o nP r i n c i p l e f o r Optimal

C o n t r o l ," A u t o m a t i c na n dR e m o t eC o n t r o l , v o l . 2 4 ,p p . 3 2 8 - 3 3 0 , M a r c h , 1 9 6 3 .

15. R . E. B e l l m a n ," D y n a m i cP r o g r a m m i n g , l lP r i n c e t o nU n i v .P r e s s ,

P r i n c e t o n , N.J., 1957.

54
 1'ij 16. V. G. E o l t y a n s k ' i i , R. V. Gamkrelidze,
and L. S. P o n t r y a g i n , "On

1 t h eT h e o r yo fO p t i m a lP r o c e s s a s , I lI z v . Akad. Nauk SSSR, vo1.24,

I! 8' pp.3-42, 1960.

19. R. Hellman, "A M a r k o f f i a nD e c i s i o nP r o c e s s , " Jour. o f Math. and

r
Mechanics,vo1.6,pp.679-684,1957.

20. A. A. Howard,"DynamicProgrammingandMarkoffProcesses," M.I.T.

Press and J. Wiley,Inc.,Cambridge, Mass. and New Y o r k , N.Y.,

1960.

21. P. W o l f e and G. B. D a n t z i g ," L i n e a rP r o g r a m m i n gi n a Markoff

Chain,"Oper. Res., vol.10,pp.702-710,1962.

22. C. Derman and M. K l e i n , "Some Remarkson F i n i t eH o r i z o nM a r k o f f i a n

D e c i s i o nM o d e l s , "O p e r a t i o n s Research, vo1.13,pp.272-278, 1965.

0

23. C. Derman, " M a r k o f f i a nS e q u e n t i a lC o n t r o lP r o c e s s e s , - Denumerable

S t a t eS p a c e , " J o u r . Math. Anal.andAppl.,vol.10,pp.295-302,

1965.

I. -
24. D. B l a c k w e l l", D i s c o u n t e dD y n a m i cP r o g r a m m i n g , ' l Ann. M a t hS
, tat.,

~ 0 1 . 3 6 , pp.226-235,
1965.

25. A. F. V e i n o t t , J r . , "On F i n d i n gO p t i m a lP o l i c i e s i n Discrete

DynamicProgramming w i t h n oD i s c o u n t i n g , ' ' Ann. Math.Stat.,vo1.37,

pp.1284-1294,
1966.

26. R. M. Harpand M. H e l d ," F i n i t e - S t a t eP r o c e s s e sa n dD y n a m i c Pro-

grarnming,ll SIAM Jour. onAppl. Math.,vo1.15,pp.693-718,1967.

27. E. V . D e n a r d o ", C o n t r a c t i o nM a p p i n g s i n t h eT h e o r yU n d e r l y i n g

DynamicProgramming," SIAM Review, vo1.9, pp.165-177,1967.

56
 constraint goal

. . . . .. .. .. .

Fig. 1. Relation between thegoal,theconstraint and thedecision.

57
 Q~ 0.6

0.8

Fig. 2. State diagram for A.

58
 0.8
I
8
l
c5 0.8

Fig. 3. Reversed-flow s t a t e diagram

.
NASA-Langley, 1970 - 19 CR-1594 59