Fuzzy Linear Programming Models and Extensions

by
Suresh Chandra

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Crisp Linear Programming Problem

Max Z = c T x
subject to
Ax
x

b,
0,

where x Rn , c Rn , b Rn , A Rmn . Let

S = {x Rn : Ax b, x 0}
denote its feasible set/feasible region. Then Solving LPP means :
Determining x S such that
c T x c T x, x S
. we call
x : Optimal Solution
Z = c T x : Optimal Value.
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Here we must note that

c Rn , b Rn , A Rmn are not just given mathematical
quantities, but rather they have some physical/economical meaning.
The given LPP is not just a mathematical formulation, it is the
mathematical model of a given real life situation.
Thus though c, b, A are given parameters, their interpretation
depends on the context in which the given problem is stated.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

CRISP MODELING/STOCHASTIC
MODELING/FUZZY MODELLING

A typical interpretation of b, A and c could be (refer to Resource

Allocation Problem)
Max availability of the i th resource,
(say raw materials/budget etc)
aij : units of i th resource being used in producing one
unit of j th product,
cj : per unit profit on the j th product.
bi :

Here i = 1, 2, . . . , m and j = 1, 2, . . . , n.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Resources
(RawM aterials)
(i = 1, 2, , m)

P roduction
System

Suresh Chandra

F inishedP roducts
(j = 1, 2, , n)

Fuzzy Linear Programming Models and Extensions

But now comes to the main question

From where these numbers bi , aij and cj

(i = 1, 2, . . . , m, j = 1, 2, . . . , n) will be known.
Will they be known precisely ?
Are these parameters certain/uncertain/vague?
Let the first resource b1 denote the budget.
While modeling, the modeller will probably ask the following question
to the decision maker: What is your budget ?

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

We may expect the following typical answer:

1

My buget is Rs. 200.

I am uncertain about it at present. But probably it is in the interval

[180,220]. One can probably estimate it from the historical data
where similar manufacturing is done.

Available budget is most likely to be Rs. 190 but certainly not more
than Rs. 220. Further, if things do not go the way I am expecting, it
could as low as Rs. 180. In any case I am very sure that it will not be
less than Rs 180, because I have this much amount in the bank for
this purpose.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

If it is Answer 1, then we say that the parameter b1 (budget) is

known precisely. (crisp optimization)
If it is Answer 2, then we say that b1 (budget) is a random variable
having some known probability distribution.(stochastic optimization)
If it is Answer 3, then we say that b1 (budget) is fuzzy set having a
membership function which looks like
b1 (x)

180

Suresh Chandra

190

220

Fuzzy Linear Programming Models and Extensions

Unlike the (crisp) optimization where we may be interested in the

maximization of the profit, in a fuzzy scenario, the decision maker may
specify his goal in the following manner:
Determining a production schedule so that the sales are significantly
more than that of last year.
Determining a production schedule so that the profit is essentially
more than Rs. 500, which we denote by Profit Z & Z0 (= 500).
i.e.
P rof it

490

Suresh Chandra

500

P rof it

Fuzzy Linear Programming Models and Extensions

Other parameters of the LPP may also be specified in a similar manner.

e.g., c1 = per unit profit for the 1st product. Then the decision maker may
specify
c1 = Rs 4 (Crisp optization)
c1 N( = 4, 2 = 2) (stochastic optimization)
c1 is close to Rs. 4, but more likely to be less than 4. My
experience and expert opinion guides me that it can not be more than
5 and less than 2. (Fuzzy Optimization)
There could be many other real life scenarios where goals/constraints are
prescribed vaguely in English/natural language e.g.,
In a marketing situation the decision maker may desire that at least 20
clients should be contracted. But it will be too bad if less than 20 clients
are contracted.
The above situation where parameters are prescribed vaguely in a natural
language clearly show the role of fuzzy modeling.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Why Fuzzy Optimization?

The traditional (crisp) optimization forces the decision maker to
prescribe the parameters precisely even if he does not wish to or does
not have mean to do so.
This forcing of the decision maker to give a precise numerical value to
the parameters result in a Wrong model of the real life situation at
hand.
Thus this approach of forcing the decision maker to prescribe the
parameters precisely solves a problem which has never been posed by
the decision maker.
We do not want a precise mathematically correct solution of a
Wrong problem and hence Fuzzy Optimization, which (provides)
more flexibility to the decision maker while prescribing the problem
parameters.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

But
This does not mean, we should not study (crisp) optimization. Because be
it stochastic optimization/ Fuzzy optimization, they will be appropriately
converted to a crisp optimization and then solved. The main challenge is
How to conceptualize the relevant fuzzy model for the given scenario ?
How to conceptualize the meaning of its solution ?
How to construct the appropriate (crisp) optimization problem whose
solution gives the desired solution in the fuzzy scenario and how to
interpret it?
Since in a given situation, some parameters could be prescribed
precisely, some stochastically and some vaguely (fuzzy), how to
aggregate such goals and constraints so that the resulting
mathematical model represents the reality ?

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Fuzzy Sets (Zadeh (1965))

X : Universe
= {(x, (x)) : x X }: Fuzzy Set
A
A

A : X [0, 1]: Membership Function

c = {(x, c (x)) : x X , where c (x) = 1 (x)}
A
A
A

A
=A
B,
(x) = Max (x), (x) .
D
D
A
B


=A
B,
(x) = Min (x), (x) .
E
D
A
B

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

The membership function A is the generalization of the characteristic

function A (x), where

1,
x A,
A (x) =
0,
x A,
Just as a crisp set A is completely characterized by its characteristic
is characterized by its membership function
function, a fuzzy set A
A (x), x X .

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

c, A
B
and A
B
follow very
Keeping this in mind, the definition of A
c
naturally if we understand the (crisp) A B, A B and A in terms of
the associated characteristic functions A and B .
Thus a fuzzy set is a generalization of the usual (crisp ) set A, where x
belongs to A with degree A (x). A : X {0, 1}, where as
A : X [0, 1]
Most of set theoretic operations hold for fuzzy sets, but Law of Excluded
Middle and Law of Contradiction do not hold, i.e.,
A
c need NOT be X.
A
A
c need NOT be (Empty set).
A

Note that X (x) = 1 x X and (x) = 0 x X

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Decision Making in Fuzzy Environment (Bellman and

Zadeh(1970))
Let m Fuzzy Goals be prescribed by Fuzzy Sets Gi (i = 1, , m)
respectively.
Let n Fuzzy Constraints be prescribed by Fuzzy Sets Cj (j = 1, , n)
respectively.
is defined as an appropriate
Then the Fuzzy Decision Set D

aggregation of fuzzy goals Gi and fuzzy constraints Cj .

Therefore
= (G1 G2 Gm C1 C2 Cn )
D
where


D (x) = min Gi (x), Cj (x)|x X .
i,j

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Fuzzy Decision
Once fuzzy decision D is known, we define x X to be an optimal
Decision if
D (x ) = max(D (x)|x
 X)

= max min Gi (x), Cj (x)|x X
i,j

G
1

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Decision Making Criteria

the above problem

In view of the definition of the fuzzy decision D,
reduces to
(LP)

Max
subject to
Gi (x) , i
Cj (x) , j
0 1, x X .

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Fuzzy Linear Programming

Two distinct type of FLP 0 s
LPP with Fuzzy Goals/Fuzzy Constraints (also called Flexible
Linear Programming)
LPP with Fuzzy Parameters (also called Fuzzy Number Linear
Programming)
Meaning of the fuzzy inequality: x &p a; x, a, p R
To understand the above formulations of fuzzy LPPs, we need to
understand the meaning of the given fuzzy statement. The given fuzzy
inequality is the fuzzy representation of the linguistic fuzzy statement x is
essentially more than or equal to a with tolerance level p. Let S be the
fuzzy set prescribed by this fuzz inequality. Then the membership
function, of the above fuzzy statement is represented as

1,
x a,

(a x)
S (x) =
,
(a p) x < a,
1

0,
x (a p),
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Membership Function of Fuzzy Inequality x &p a0

ap

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

LPP with Fuzzy Goals/Fuzzy Constraints :

Zimmermanns Symmetric Model (1978), (1985)
(FLP)

Find x Rn such that

cT x
Ai x
x

Here

&p0
.pi

Z0 ,
bi , (i = 1, 2, . . . , m, )
0,

Z0 : Aspiration level of the decision maker.

p0 : Tolerance level of the fuzzy goal.
p : Tolerance level of the flexible constraints.
i
we may choose

1,
c T x Z0 ,

(Z0 c T x)
0 (c T x) =
,
(Z0 p0 ) c T x < Z0 ,
1

p
0

1,
c T x < (Z0 p0 ),

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Zimmermanns (FLP) Model

1,

(Ai x bi )
i (Ai x) =
1
,

pi

0,

Ai x bi ,
bi < Ai x bi + pi ,
Ai x > bi + pi .

Now Bellman-Zadeh principle gives

Max
subject to
0 (c T x) ,
i (Ai x) ,
1, , x 0.
which on substitution gives the equivalent crisp (ELP)

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Equivalent Crisp Linear Programming Problem

(ELP)

Max
subject to
c T x Z0 (1 )p0 ,
Ai x bi + (1 )pi , (i = 1, , m)
0 1, x 0.

Let (x , ) be an optimal solution of equivalent crisp (ELP). Then

x is an optimal solution of (FLP) and is the highest common
degree up to which the aspiration level Z0 of the decision maker is
met and the flexible constrained are satisfied.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Example

Find (x1 , x2 ) R2 such that

x1 + x2 .2 14.5,

x1 + 3x2 .3 21,

x1 + 3x2 .6 27,

4x1 + 3x2 .6 45,

3x1 + x2 30,
x1 , x2 0,

Here Z0 = 14.5.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

The Equivalent (crisp) LPP

(ELP)

Max
subject to
2 x1 x2
3 x1 + 3x2
6 + x1 3x2
6 + 4x1 + 3x2
0
x1 , x2

12.5,
24,
33,
51,
1,
0,

An optimal solution of above(ELP) is x1 = 6, x2 = 7.75, z = 13.75 and

= 0.625

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

A Revisit to Zimmermanns Approach

Remark
1 Zimmermanns formulation of FLP is symmetric as it does not
distinguish between objective and constraints. This is in contrast with the
(crisp) LPP formulations, where a solution has to be first feasible before it
is optimal.
Recently there have been attempts to re-visit original Zimmermanns
model and present a two-phase approach which differentiates between the
constraints and the objectives. Some of these studies are
1

S.Chandra and A.Aggarwal, On solving fuzzy linear programming

problems:A revisit to Zimmermannsapproach, Journal of Intelligent
and Fuzzy Systems(2014).

D.Dubey, S.Chandra and A. Mehra, Computing a Pareto-optimal

solutions for multi-objective flexible linear programming in a bipolar
framework, Int. Journal of General Systems (2014).
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

A Two Phase Approach for Flexible Linear

Progarmming (S.Chandra /A.Aggarwal : On solving
fuzzy linear programming problems:A revisit to
Zimmermannsapproach:Journal of Intelligent and
Fuzzy Systems(2014)

There are two objectives in FLP, namely satisfaction of constraints

and attainment the aspiration level of the objective function
In (crisp) LPP, the solution has first to be feasible before it is optimal.
While solving (crisp) LPP by two phase method, Phase-I deals only
with the constraints. Once a basic feasible solutions has been
obtained, Phase-II deals with the optimality of the given objective
function. We wish to follow the same approach for solving FLPs.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Here we solve the given (FLP) in two phases. Phase-I aggregates

constraints only via Zimmermanns approach and obtains feasible solutions
of maximum permissible degree. Thus Phase-I, problem is as follows and is
termed as (FLP : Phase-I)
(FLP : Phase-I) Max

subject to
Ai x bi + (1 )pi , (1 = 1, 2, . . . , m),
x

0,

[0, 1].

Let (x, ) be a solution of (FLP : Phase-I). Then x is termed as

-feasible solution of (FLP). Let S be the collection of all -feasible
solution of (FLP). This is analgous to the set of feasible solutions in crisp
LPP.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

We next construct Phase-II problem termed as (FLP : PhaseII ).

Find x Rn such that

(FLP : Phase II )

cT x
x

&p
Z0 ,
S,

Therefore the crisp equvalent of the above problem is

Max

subject to
cT x
Ai x
x

Z0 (1 )p0 ,

bi + (1 )pi , (i = 1, 2, . . . , m),
0,

[0, 1].

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Let (b
x,
b) be an optimal solution of (FLP : Phase-II). Then b
x is identified
as an optimal solution of (FLP). In fact we denote it by (b
, )-optimal
solution.
Here

b : the degree up to which the aspiration level of the decision maker

is met.
: the common highest degree up to which b
x satisfies the constraints
Ax . b.
Remark
Let (x , ) be a solution of the given (FLP) by Zimmermanns approach.
Then it is obvious that Max(b
, ) > Therefore this two phase approach
does a better trade-off between the twin objectives of meeting the
aspiration level and satisfaction of constraints better than
Zimmermanns approach.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

A Bi-objective Programming Approach for (FLP)

As there are two basic goals associated with (FLP),satisfaction of
constraints and meeting the aspiration level, it is natural to consider the
following bi-objective programming problem (BOP)
(BOP)

Max
(, )
subject to
0 (c T x) ,

i (Ai x) , (1 = 1, 2, . . . , m),
x

0,

[0, 1],

We can relate solution of (FLP) in terms of efficient solutions of the

problem (BOP) as stated in the following lemma.
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Lemma
The solution (b
x,
b, ) is an efficient solution of the bi-objective
optimization problem (BOP).
Remark
In view of the above lemma, there will be no unique solution of (FLP) in
general. But rather there will be entire efficient frontier of (BOP) giving
all solutions of (FLP).
Remark
One can choose a trade-off parameter 0 1 and solve
Max
+ (1 )
subject to
c T x Z0 (1 )p0 ,
Ai x

bi + (1 )pi , (i = 1, 2, . . . , m),
0, , [0, 1].

to get an efficient solutionSuresh

of (FLP).
Chandra

Fuzzy Linear Programming Models and Extensions

Example 2

Consider the following fuzzy linear programming problem

Find x R2 such that
x1 + x2
x1 + x2
x1 + 10x2
x1 + 3x2
3x1 x2
x1 , x2

&
.
&
.
&

9.3,
2,
15,
12,
20,
0.

with p0 = 1.8, p1 = 4, p2 = 3, p3 = 5, and p4 = 6. By using

Zimmermann approach, we have to solve the following problem

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Max

subject to
x1 + x2 1.8 7.5,
x1 + x2 + 4 6,

x1 + 10x2 3 12,

x1 + 3x2 + 5 17,
3x1 x2 6 14,

x 0, [0, 1].

The optimal solution is (x1 = 6.88, x2 = 2.11, = 0.754). Then

= 0.754 is the highest common degree of satisfaction up to which all
the constraints and the aspiration level of the decision maker can be
satisfied.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Now we solve the same problem by the proposed two phase method. The
Phase-I problem is
Max
subject to
x1 + x2 + 4 6,

x1 + 10x2 3 12,

x1 + 3x2 + 5 17,

3x1 x2 6 14,
x 0,

[0, 1].

The optimal solution of Phase-I problem is

(x 1 = 6.88, x 2 = 2.11, = 0.754). Therefore = 0.754 is the highest
common degree up to which all the constraints are satisfied.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Next the Phase-II problem is

Max
subject to
x1 + x2 1.8 7.5,
x1 + x2 + 4 6,

x1 + 10x2 3 12,
x1 + 3x2 + 5 17,
3x1 x2 6 14,
x

0,

[0, 1].
The optimal solution of the above Phase-II problem is
(b
x1 = 6.88, b
x2 = 2.11,
b = 0.844).
Here we note that Max(b
, ) = Max(0.844, 0.754) = 0.844 is greater

than = 0.754 as obtained via Zimmermanns approach.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

5. LPP with Fuzzy Parameters (Also called Fuzzy

Number Linear Programming

(FLP)

Max
cT x
subject to

Ax
x

()b,
0.

Here
x Rn
= (aij ) is a fuzzy matrix with aij as a fuzzy numbers, c and b
A
are vectors in Rn and Rm respectively, with entries as fuzzy numbers.

() is an appropriate fuzzy order to compare fuzzy numbers.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

5. LPP with Fuzzy Parameters (Also called Fuzzy

Number Linear Programming

There are three types of Models for LPP with fuzzy Parameters.
Tanaka and Asai Approach (1984)
Ranking Function Approach (Yager 1981).
Standard Fuzzy Order Appraoch (Clemente, Fernandez,
Puerto(2012)).
There are numerous variants of the original Ranking Function Approach of
Yager(1981), e.g. Bector, Chandra and Vijay (2004), Li(1999).
Here we shall discuss the details of Tanaka and Asai. We shall also make
certain observation on the ranking function approach of Clemente et al.
(2012).

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Fuzzy Number
We now define the following
Definition
A fuzzy number a is a fuzzy set in R such that
(1) d R such that a (d) = 1 (The fuzzy set a is normal).
(2) Every -cut,

a = {x R : a (x) > }
is a closed-bounded 0 < 1 interval. For = 0, a0 is defined
separately as the closure of the set {x R : a (x) > 0}
(Condition(2) is equivalent to the facts that a is upper semi-continuous,
quasi concave and its support is compact).
The arithmetic of fuzzy numbers is understood in terms of the interval
arithmetic of its cuts or by resorting to Zadeh Extension Principle.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Triangular Fuzzy Number

a = [al , a, au ] : TFN
a = [a1 , a2 ] : cut of a

a (x)

al

a1

Suresh Chandra

a2

au

Fuzzy Linear Programming Models and Extensions

Trapezoidal Fuzzy Number

a = [al , a, a, a, au ] : TrFN

a (x)
1

al

a1

a2

au

Figure:
fuzzy
SureshTrapezoidal
Chandra
Fuzzy
Linearnumber
Programming Models and Extensions

Tanaka and Asais Approach for FLP (1984)

Here all parameters of FLP are prescribed by symmetric triangular
fuzzy numbers.
Thus aij , bi , cj , (i = 1, 2, . . . , m), (j = 1, 2, . . . , n) are symmetric
TFN.
Definition
a F (R) is called a symmetric TFN if it is a TFN whose left and right
spreads are equal, i.e. for a = (al , a, au ), au a = a al = p(say). Then
we write a = (a, p) and call a as its center and p as its spread.
Definition
Let a1 , a2 F (R) be symmetric TFN. A function of type
y = a1 x1 + a2 x2 , (x1 , x2 ) R2 is called a fuzzy linear function.
For the input x T = (x1 , x2 )T = (1, 1)T and a1 = (4, 1), a2 = (1, 2), we
get y = (5, 3).
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

In a similar manner, we can understand

y = a1 x1 + a2 x2 + . . . + an xn ,
(x1 , x2 , . . . , xn ) Rn and its generalization to the system
y1
yi
....
..

= a11 x1 + a12 x2 + + a1n xn ,

= ai1 x1 + ai2 x2 + + ain xn ,

ym = am1 x1 + am2
x2 + + amn
xn .
If we write

Y =

y1
y2
..
.
ym

a11

..
,A = .

am1

a12

a1n

am2

amn

,
X
=

x1
x2
..
.

xn

Then the above system can be written as

.
Y = AX
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

LPP with Symmetric TFN Data

(FLP)

Max
cT x
subject to

Ax
x

b,
0.

= (aij ) are the

where cT = (c1 , c2 , , cn ), b = (b1 , b2 , , bn )T and A
symmetric TFN for all i and j.
Thus aij = [aij pij , aij , aij + pij ], cj = [cj j , cj , cj + j ] and
bi = [bi i , bi , bi + i ]

|r aij |

,
r [aij pij , aij + pij ]
1
aij (r ) =
pij

0,
otherwise

|r cj |

1
,
r [cj j , cj + j ]
cj (r ) =
j

0,
otherwise
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

bi (r ) =

|r bi |
,
i

0,

r [bi i , bi + i ]
otherwise

Let yi = nj=1 aij xj . Then using the membership function of the symmetric
TFN aij . We can show that

T (x)|

|r a(i)

, |x| =
6 0,

T |x|

p(i)
yi (r ) =
T (x) p T |x|, aT (x) + p T |x|]
r [a(i)

(i)
(i)
(i)

0,
x = 0, r = 0,
T = [a , a , , a ] p T = [p , p , , p ]
Here a(i)
i1 i2
in
i1 i2
in
(i)
|x| = (|x1 |, |x2 |, , |xn |)
T (x) = n a x ,
T |x| = n p |x |
a(i)
p(i)
j=1 ij j
j=1 ij j

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Almost positive Symmetric TFN

Let a = (a, p) be a symmetric TFN. Then a is said to almost positive with
degree h, denoted by a &h 0 if a > 0 and a (1 h). The following
figure illustrates a &h 0

1
h

a
= (a, p)
1h

(a p)

(a + p) x

a
p

Figure: a &h 0

The larger the values of h, the stronger is the meaning of almost

positive a
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Almost positive Symmetric TFN

Next let us try to understand a &h 0 in terms of area. For a

symmetric TFN, let
R
Positive area of a
a (r )dr
0
S = R
=
Total area of a

(r
)dr
a

Then S = 0.5 + h 0.5h2 . Therefore if we choose h = 0.6, we get
S = 0.92 . In that case we may say that a & 0 with grade o.92
We may now extend the above definition to a P
sum of symmetric TFN.
let ai P
= (ai , pi ) be
symmetric TFNs and a = ki=1 wi ai . Then
P
a = ( ni=1 wi ai , ni=1 wi pi ) is also a symmetric TFN.
P
Then a &h 0 if ni=1 wi ai > 0 and a (1 h).

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Reformulation of the given Fuzzy No. LPP

Find x Rn such that

cT x
T x
A
x

& Z0 , (Goal)
(Constraints)
. b,
0.

Without loss of generality, the above problem can be written as

y0
y1
....
..

= a00 x1 + a01 x2 + + a0n xn & 0,

= a10 x1 + a11 x2 + + a1n xn & 0,

ym = am0 x1 + am1
x2 + + amn
xn & 0.
x0 = 1.
Here yi & 0, (i = 0, 1, , m) is understood in the sense of almost
positive.
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Reformulation of the given Fuzzy No. LPP

& 0, x 0, where
The above system can be written as Y = Ax

A0
y0

y1

A1
= .
Y = . , A
,
..

..

ym
Am
Ai = [ai0 , ai1 , , ain ], (i = 0, 1, , m).
Our aim is to find x = (x0 = 1, x1 , x2 , , xn ) Rn+1 satisfying the above
system. The above system is equivalent to
y0

= 1

..
.
ym

= 1

AT
(0) x
T x
P(0)

AT
(m) x
T x
P(m)

(1 h),

(1 h),

Suresh Chandra

T x 0
P(0)

T x 0.
P(m)

Fuzzy Linear Programming Models and Extensions

Reformulation of the given Fuzzy No. LPP

Note that ai0 = (ai0 , pi0 ), aij = (aij , pij ), i = 1, 2, , m, j = 1, 2, , n
and A(m) , P(m) are obtained accordingly.
These can be rewritten as
(A(0) hP(0) )T x 0
(A(i) hP(i) )T x 0, (i = 1, 2, , m)
x 0, (x0 = 1).
The resulting (crisp) optimization problem is
(EFNLP)

Max h
subject to
(A(i) hP(i) )T x 0, (i = 0, 1, 2, , m)
x0 = 1, xj 0 (j = 1, 2, , n)
0 h 1.
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Note that
Here x and h both are variables, so it is a non-linear programming
problem. Tanaka and Asai suggested the following algorithm to solve
it.
Let for a given 0 h 1. The set Xh = (x0 = 1, x Rn ) satisfies
the constraints of (EFNLP) is called the admissible set of (EFNLP)
for given h. Then h1 < h2 Xh2 Xh1 . Thus larger h is, the smaller
is its admissible set.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Algorithm for (FNLP)

Step 1: Determine a small initial value of h so that there exists an

admissible set of (EFNLP).
Step 2: Let > 0 be small increment. Find h + (k + 1) which is the
smallest value not combatable with the constraints of (EFNLP), in the
process of increasing h at , 2, , . i.e. continue till the admissible set
of nonempty and choose the appropriate k.
Step 3: Choose the most important or useful inequalities (as determined
by some user specified equation ) from (EFNLP). If it is the k th
inequalities, consider the resulting (crisp) optimization problem is

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Max
AT
(j) x
subject to
(A(i) (h + k)P(i) )T x
x0 = 1, xj

0, (i = 0, 1, 2, , m)
0 (j = 1, 2, , n)

0 h 1.

Let x = (x0 = 1, x1 , x2 , . . . , xn ) be an optimal solution of above (LPP).

We the take (x0 = 1, x1 , x2 = 2, . . . , xn ) as an appropriate solution of
(FNLP).

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Example
Consider the following (FNLP)

6x1 +
5x2

2x1 +
9x2

11x1
5x2
1+
2x
3x2
x1 0, x2

30,

45,

44,

12,
0.

Equivalently
y0
y1
y2
y2
x1

=
=
=
=

+
2 0,
30
6x1 + 5x

45 2x1 9x2 0,
11x
1+
44
5x2 0,

12 + 2x1 3x2 ,
0, x2 0.

yi = Ai x 0, (i = 0, 1, 2, 3), where x0 = 1, x1 and x2 to be

determined.
Suresh Chandra

Fuzzy Linear Programming Models and Extensions


A0 = (30,
6,
5), P(0) = (6, 2, 1), therefore
= (30, 6), 6 = (6, 2),
30
5 = (5, 1) are symmetric TFN

A1 == (45,
2,
9), P(1) = (8, 2, 3)
11,

A2 = (44,
5), P(2) = (4, 2, 1)

A3 = (12,
2,
3), P(0) = (4, 1, 2)
This gives the following (EFNLP)
Maxh
subject to
30 6h + (6 2h)x1 + (5 h)x2 0,
45 8h (2 + 2h)x1 (9 + 3h)x2 0,

44 4h (11 + 2h)x1 + (5 h)x2 0,

12 4h + (2 h)x1 (3 + 2h)x2 0.
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

y0

y1
0.5

0.4
7.5

27.7

10.1

12.5

37.3

12.4

1
1

y2

y3
0.5
0.4
8.0

7.6

23.2

Suresh Chandra

0.6 0

13.1

26.8

Fuzzy Linear Programming Models and Extensions

(3)

(3)
(1)

(2)

(2)

h = 0.3

h=0

(1)
(0)

(0)

(2)
(3)

2.6
(1)

h = 0.5
(0)
4.5

From the figure we obtain that approximate solution of FNLP is

(x1 = 4.5, x2 = 2.6) with h = 0.5.
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

From the above figure, we infer that y1 and y2 provide restriction on

h, while there is a leeway from y0 and y3 .
The solution x ensures that fuzzy inequalities hold with a degree
more than h . Therefore it might be called the max-min solution.
Tanaka, Ichihashi and Asai (1984,1985) extended the formulation to
symm TrFNs.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Ranking Function Approach for solving FNLP

Let N(R) be the set of all fuzzy numbers in R and a, b N(R).

Let F : N(R) R be a suitable ranking function, i.e. we define

a b if F (a) F (b).
Yager proposed the following ranking function
R au
a (x)dx

a x
1 F1 (
a) = R lau (x)dx , where al and au are the lower and upper limits
al

u
of the support of a. If a = (al , a, au ) is TFN, then F1 (a) = al +a+a
3
R max
F2 (a) = 0
m[aL , aR ]d, where max is the height of a,
a = [aL , aR ] is its cut and m[aL , aR ] is its mean value of the
u
cut. For the a = (al , a, au ), F2 (a) = al +2a+a
4

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Remark
does not necessarily means a = b and the ordering
Since F (a) = F (b)

being a partial order only,

(F (a) F (b)) (a b)
(a = b)
is to be considered in the sense of equivalence
(F (a) = F (b))
classes only.
Remark
For a = (al , a, au ), a is called its mode and (au al ) is called its
divergence. For TFNs these three quantities namely (F1 or F2 ), mode,
divergence) are enough to get a total order. The main difficulty here is
that which order should really be followed to get the total order.
This suggests that the best option is to conceptualize the solution of
FNLP as an appropriate efficient solution rather than stitching to a
specified total order.
In fact one can show that there could not add any concept of unique
fuzzy optimal value of FNLP (Kaur and Kumar, JOTA (2013)). The
correct approach is to think in terms of efficient frontier of solutions.
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

A Standard Fuzzy Order Approach (Clemente,

Fernandez and Puerto (2010))

Consider the fuzzy number LP denoted by (FLP)

(FLP)

Max
cT x
subject to

Ax
x

()b,
0.

We now have the following definitions.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Standard Ranking Function

Definition
Let = {1 , . . . , r } [0, 1] with 1 < . . . < r 1 < r = 1 be a set of
cuts. A standard fuzzy ranking function f : FN(R) Rr 2 , such that


f (
a) = (pij (
a)) = aj i , (i = 1, . . . , r , j = 1, 2) ,
Definition
A fuzzy order () is standard fuzzy order if there exists a standard ranking
function f such that
f (

()b
a
a) 5 f (b).
where () is the componentwise order on Rr 2 . Any fuzzy order defined
by a standard ranking function f introduces the following equivalence
relationship on N(R)
f (

(=)b
a
a) = f (b).
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Ordering of TFNs
Therefore () is a partial order and (=) is to be understood in terms of an
rather than saying that
indifference relation. Thus when f (
a) = f (b)

= b, we say that a
and b are equivalent, and denote the same as a
(=)b.
a
For TFN and TrFN, in Definition 2 we need only two values of , namely
= 0 and = 1.
Let a = (al , a, au ) and b = (bl , b, bu ) be two TFN 0 s. Then

 1 

a
(a al ) + al
f (a) =
=
(a au ) + au
a2
Similarly
=
f (b)

(b bl ) + bl
(b bu ) + bu

As per our definition, a()b if


 

(a al ) + al
(b bl ) + bl

(a au ) + au
(b bu ) + bu
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Ordering of TFNs

As a and b are TFN 0 s, we need to consider only = 0 and = 1. Thus

we have a()b if




bl
al
()
bu
au
and

a
a


()

b
b

This gives a()b if and only if

al bl , au bu , a b

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Equivalence Between Multi-objective Linear

Programming Problem (MLP) and (FLP)()
Ordering of TrFNs can be understood similarly. For a general piecewise
linear fuzzy number, 1 , . . . , r in [0,1] can be chosen appropriately and
fuzzy ordering can be defined.
Consider FNLP (denoted by (FLP) ) as
(FLP)

Max
cT x
subject to

Ax
x

. b,
0.

Now we can construct,

(MLP)

Max (pij (
c )x (i = 1, , r , j = 1, 2))
subject to
pij (b),

pij (A)x
(i = 1, , r , j = 1, 2)
x 0.
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Remark
As (FLP) is defined w.r.t the standard fuzzy order (), solving (FLP)
amounts to determination of its maximal elements. This is because

() is a partial order only. Here a()b is the same as as b()

a.
Finding Maximal elements of (FLP) is equivalent to finding the set of
efficient (Pareto-optimal) solutions of (MLP).
(MLP) can be efficiently solved by using the theory/algorithms of
Multi-objective programming.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Optimization under fuzzy IF Then RULES (Carlsson

and Fuller, 1998/2000)
1

Unlike the usual FLP model (e.g., Zimmermanns formulation) the

fuzzy value of the objective function may not be known directly for
any x R n .
Rather we are only able to describe the casual link between x and
f (x) linguistically using fuzzy if then rules.
Here the optimization problem is
Max/Min f (x)
subjectto
(R1 (x), R2 (x), . . . , Rm (x)) : x X ,
where x1 , x2 , . . . , xn are linguistic variables, X R n is a (crisp or
fuzzy) set of constraints on the domain of x1 , x2 , . . . , xn and for
i = 1, 2, . . . , m
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Example

e i1 , x2 is A
e i2 , . . . , xn is A
e in ), Then f (x) isC
ei .
Ri (x) : If (x1 is A
e ij and Ci are fuzzy numbers (i = 1, 2, . . . , m; j = 1, 2, . . . , n).
Here A
Generalizing the fuzzy reasoning approach (say TSK model). We
determine the crisp value of f at y X and obtain the optimal solution to
(FOP) by solving the resulting (usually non linear optimization problem)
Max/Min f (y )
subject to
y X.
We illustrate this methodology for the following example.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Example

Min
f (x)
subject to
x1 + x2 = 1/2,
0 x1 1,
0 x2 1,
Where f (x) given linguistically as

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

R1 (x) : If x1 is small and x2 is small then f (x) is small

R2 (x) : If x1 is small and x2 is big then f (x) is big
Compute the firing levels of the rules R1 and R2 respectively by the
product t-norm.
Let (y1 , y2 ) be the input vector to the fuzzy system. Then the firing levels
of the rules are
1 = (1 y1 )(1 y2 ), 2 = (1 y1 )y2 .
Notice that for y1 = 1, 1 = 0, 2 = 0, so no rule applies. So we can
exclude y1 = 1 from the set of feasible solutions.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

x1 is small

x2 is small

f (x) is small
1

z1

x1 is small

f (x) is big
x2 is big

y1

y2

Suresh Chandra

z2

Fuzzy Linear Programming Models and Extensions

The individual rule outputs are (refer to the figure here)

z1 = 1 (1 y1 )(1 y2 ),

z2 = (1 y1 )y2 .

Therefore the overall system output is (TSK model)

(1 y1 )(1 y2 )(1 (1 y1 )(1 y2 )) + ((1 y1 )y2 )
(1 y1 )(1 y2 ) + (1 y1 )y2
= y1 + y2 2y1 y2

f (y ) =

Therefore our original fuzzy optimization problem

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Min
f (x)
subjectto
{R1 (x), R2 (x)) : x X },
results into
Min
(y1 + y2 2y1 y2 )
subjectto
{Y1 + y2 = 1/2, 0 y1 1, 0 y2 1}.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

Which has optimal solution y1 = y2 = 1/4 and its optimal value is

f (y ) = 3/8.
It is clear that if there were no constraints on crisp values of x1 and x2 ,
then the optimal solution would be y1 = y2 = 0 with f (y ) = 0.

Suresh Chandra

Fuzzy Linear Programming Models and Extensions

References
(1) A. Aggarwal, A. Mehra and S. Chandra, Application of linear
programming with I-fuzzy sets to matrix games with I-fuzzy goals,
Fuzzy Optimization and Decision Making, Vol. 11(4), 465-480,
(2012)
(2) C. R. Bector, S. Chandra and V. Vidyottama, Duality in linear
programming with fuzzy parameters and matrix games with fuzzy
payoffs, Fuzzy Sets and Systems, Vol. 146, pp. 253-269, (2004).
(3) C. R. Bector and S. Chandra, Fuzzy Mathematical Programming and
Fuzzy Matrix Game, Berlin Heidelberg; Springer -Verleg, (2005).
(4) R. E. Bellman and L. A. Zadeh, Decision Making in a fuzzy
environment, Management Science, Vol. 17, 141-164, (1973).
(5) M. Clemente, F. R. Fernandez and J. Puerto, Pareto optimal security
strategies in matrix games with fuzzy payoffs. Fuzzy Sets and
Systems, Vol 176, 3645 (2011)
Suresh Chandra

Fuzzy Linear Programming Models and Extensions

(6) Kaur and Kumar, JOTA (2013)

(7) Christer Carlsson, Robert Fuller, Optimization under fuzzy if then
rules, Fuzzy Sets and Systems 119 (2001) 110-120.
(8) S. Chadra and A. Aggarwal, On Solving Fuzzy Linear Programming
Problems: A Revisit to Zimmermanns Approach, Journal of
Intelligent and Fuzzy system. Accepted (2014)
(9) D. Dubey, S. Chandra and A. Mehra, Computing a Pareto-optimal
solutions for multi-objective exible linear programming in a bipolar
framework, Int. Journal of General Systems (2014).
(10) R. Yager, A procedure for ordering fuzzy numbers of the unit interval,
Information Sciences, Vol. 24, pp. 143-161, (1981).
(11) H. J. Zimmerman, Fuzzy Programming and linear Programming with
several objective functions, Fuzzy Set and Systems, Vol. 1, 45-55,
(1978).

Suresh Chandra

Fuzzy Linear Programming Models and Extensions