Applied Soft Computing 11 (2011) 5341–5349

Contents lists available at ScienceDirect

Applied Soft Computing

journal homepage: www.elsevier.com/locate/asoc

Advanced sensitivity analysis of the fuzzy assignment problem

Chi-Jen Lin a,∗ , Ue-Pyng Wen b , Pei-Yi Lin b
a
Department of Industrial Engineering and Management, Ta Hwa Institute of Technology, Hsinchu, Taiwan, ROC
b
Department of Industrial Engineering and Engineering Management, National Tsing Hua University, Hsinchu, Taiwan, ROC

a r t i c l e i n f o a b s t r a c t

Article history: This paper concentrates on sensitivity analysis of the fuzzy assignment problem (FAP). Since most real
Received 30 November 2009 environments are uncertain, the FAP is more realistic than the assignment problem in application. Owing
Received in revised form 8 July 2010 to the high degeneracy of the FAP, as that of the assignment problem, traditional sensitivity analysis,
Accepted 2 May 2011
called Type I sensitivity analysis, which determines the range in which the current optimal basis remains
Available online 7 May 2011
optimal, is impractical. Hence, we attempt to perform other two types of advanced sensitivity analysis,
called Type II and Type III sensitivity analysis, to overcome this problem. A labeling algorithm is then
Keywords:
presented, where Type II sensitivity analysis is to determine the range of perturbation to keep the current
Assignment problem
Fuzzy theory
optimal assignment remaining optimal, and Type III sensitivity analysis is to determine the range for
Sensitivity analysis which the rate of change of optimal value function remains unchanged. The procedure of the labeling
Degeneracy algorithm is divided into two parts: one is when the unassigned cell is perturbed, and the other is when
Labeling algorithm the assigned cell is perturbed. An example is presented to demonstrate that the labeling algorithm is a
useful tool for determining the Type II and Type III sensitivity analysis of the FAP.
© 2011 Elsevier B.V. All rights reserved.

1. Introduction the cij ’s are deterministic numbers and usually denoted as an n × n

matrix C = [cij ]. An assignment problem is then defined to choose
Assignment problem (AP) has been widely used in manufactur- precisely one element from each row and column so that the total
ing and service systems as well as many indirect applications. The sum is minimized.
AP can be stated as: Let a number n of jobs be given that must However, in many real-world applications, costs are not deter-
be performed by n workers, where the costs depend on the specific ministic numbers. In recent years, many researchers have begun
assignments. Each job must be assigned to one and only one worker to investigate assignment problem and its variants under fuzzy
and each worker has to perform one and only one job. The problem environments. Belacela and Boulasselb [2] studied a multi-criteria
is to find such an assignment that the total cost becomes a mini- fuzzy assignment method applied to medical diagnosis. Ridwan
mum [1]. An AP can be formulated as a 0–1 integer programming [3] studied a fuzzy preference-based traffic assignment problem.
or linear programming problem as follows: Liu and Li [4] studied the fuzzy quadratic assignment problem
with penalty. Feng and Yang [5] studied a two-objective fuzzy k-

n 
n
cardinality assignment problem. Liu and Gao [6] studied the fuzzy
min cij xij
weighted equilibrium multi-job assignment problem.
i=1 j=1 Lin and Wen [7] investigated a fuzzy assignment problem (FAP)
 n
in which the cost depends on the quality of the job. The quality of
s.t. xij = 1 for i = 1, 2, . . . , n,
(1.1) job is regarded as the performance of worker. The manager of the
j=1 workers also has to limit the total cost with a range, and thus the
 n
total cost is related to the performance of the manager. Hence, they
xij = 1 for j = 1, 2, . . . , n,
define the membership function of cost, denoted as c̃ij , as the linear
i=1
monotone increasing function shown in (1.2). ˛ij is the minimum
xij ∈ {0, 1} or xij ≥ 0 for i, j = 1, 2, . . . , n;
cost for worker i to perform job j, and ˇij is the minimum cost asso-
where cij , the only input parameter, represents the cost associated ciated with worker i to perform job j at the highest quality qij . The
with worker i (i = 1,2,. . .,n) who has performed job j (j = 1,2,. . .,n). All greater the cost spent, between ˛ij and ˇij , the higher the quality is
attained. However, any expense exceeding ˇij is useless since the
quality can no longer be enhanced. Without loss of generality, it
∗ Corresponding author.
is assumed that 0 < ˛ij < ˇij and they define the quality matrix [qij ]
where 0 < qij ≤ 1. Condition xij = 1 is added to (1.2) because there is
E-mail address: lcj@thit.edu.tw (C.-J. Lin).

1568-4946/$ – see front matter © 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.asoc.2011.05.025
5342 C.-J. Lin et al. / Applied Soft Computing 11 (2011) 5341–5349

no real expense if xij = 0 in any feasible solution. Table 1

⎧ Parameters for cells of LW algorithm.
⎪ q if cij ≥ ˇij , xij = 1,
⎨ qij (c − ˛ ) If xij = 1, i.e., an assigned cell If xij = 0, i.e., an unassigned cell

ij (cij ) =
ij ij ij
if ˛ij ≤ cij ≤ ˇij , xij = 1, (1.2) α ij α ij
⎪
⎩ ˇij − ˛ij Qij
0 otherwise.
γ ij γ ij
Moreover, notation ˛ij , ˇij  is employed to denote c̃ij . Matrix [c̃ij ] is
shown as follows:
range in Sections 3 and 4, respectively. Furthermore, two numeri-
[c̃ij ] = [˛ij , ˇij ]
cal examples are presented in order to demonstrate the proposed
In addition, all the ˛ij ’s form the matrix [˛ij ] and all the ˇij ’s form approaches. Section 5 concludes the study with a brief summary.
the matrix [ˇij ].
They also define the membership function of total cost c̃T , which 2. LW algorithm and sensitivity analysis
is related to the performance of the manager, as the linear mono-
tonically decreasing function in (1.3). Numbers a and b are the lower 2.1. Concepts of the LW algorithm
and upper bounds of total cost, respectively; and notation a, b
denotes the fuzzy interval c̃T . It is suggested that a number less 1
Let  = n n >0 (2.1)
than or equal to the minimum assignment of matrix [˛ij ] should b−a+  x
i=1 j=1 ij ij
be taken as a, and a number larger than or equal to the maximum
assignment of matrix [ˇij ] should be taken as b. and yij =  · xij for i, j = 1, 2, . . . , n (2.2)
⎛ ⎞
(1.4) can then be converted into a linear programming model as

n

n

T (cT ) = T ⎝ cij xij ⎠ follows:

i=1 j=1 
n 
n

max − ˛ij yij + b · 

⎧
⎪ 1, cT ≤ a,
⎪
⎨ n n
i=1 j=1

b− c x  n
= i=1 j=1 ij ij b − cT (1.3) s.t. yij −  = 0 for i = 1, 2, . . . , n
= , a ≤ cT ≤ b,
⎪
⎪ b−a b−a
⎩ j=1
0, cT ≥ b.  n (2.3)
yij −  = 0 for j = 1, 2, . . . , n
Assume that a team comprises all workers and the manager. The
i=1
performance of the team is determined by taking the lowest per- n 
 n
formance of members in the group, the company has to equally ij yij + (b − a) = 1
emphasize the performance of each member in the group. Hence,
i=1 j=1
the Bellman–Zadeh’s criterion [8] is used and the FAP can be derived
yij ,  ≥ 0 for i, j = 1, 2, . . . , n
as:
n n Suppose u1 ,. . .,un , v1 ,. . .,vn and f are the dual variables of this
b− i=1
˛ x
j=1 ij ij
max n n model, then its corresponding dual problem is as follows:
b−a+ i=1
 x
j=1 ij ij
min f

n

s.t. xij = 1 for i = 1, 2, . . . , n s.t. ui + vj + ij f ≥ −˛ij for i, j = 1, 2, . . . , n

j=1
(1.4)

n
xij = 1 for j = 1, 2, . . . , n
j=1
xij ∈ {0, 1} or xij ≥ 0 for i, j = 1, 2, . . . , n
where  ij = (ˇij − ˛ij )/qij for ∀ i, j, is the slope of the membership
function.
Furthermore, Lin and Wen [7] proposed an efficient labeling
algorithm, denoted as the LW algorithm, for solving (1.4), which is
a linear fractional programming problem and extreme-point opti-
mum exists. Note that the constraints of (1.4) are identical to those
of (1.1). Hence, the optimal solution of (1.4) is inherently highly pri-
mal degenerate and corresponds to several different optimal bases.
Consequently, changing the optimal basis does not ensure that the
optimal assignment will be changed. Degeneracy thus makes the
traditional sensitivity analysis, an important issue after obtaining
the optimal solution, impractical.
The main objective of this paper is to investigate the proper-
ties of the FAP and then propose a modified labeling algorithm to
determine two other types of practical sensitivity range. The rest of
this paper is organized as follows. Section 2 reviews the concepts of
the LW algorithm, types of sensitivity range and various perturba-
tions of the FAP. We then propose the algorithm which is extended
from the LW algorithm to determine the two types of sensitivity

n

n
(2.4)
− ui − vj + (b − a)f ≥ b
i=1 j=1
u1 , . . . , un , v1 , . . . , vn , f ∈R
The LW algorithm begins with a feasible solution of (2.3), i.e.,
of (1.4), and proceeds to obtain a feasible solution of (2.4) while
maintaining complementary slackness. At each stage of the solu-
tion procedure, a better solution of (2.3) is obtained until the
optimal solution is reached. The steps of the LW algorithm are
performed directly on a table as that of classical transportation sim-
plex method. Parameters corresponding to cell (i, j) are displayed
in Table 1. Lin and Wen [7] demonstrated that the LW algorithm
offers an effective and efficient method for solving FAP.
2.2. Three types of sensitivity range
Koltai and Terlaky [9] showed that managerial questions are
not answered satisfactorily with the mathematical interpretation
of the traditional sensitivity analysis when the solution of a lin-
ear programming model is degenerate. In order to perform proper
sensitivity analysis for (1.4), two other types of sensitivity analysis
must be considered depending on how the analysis is performed.
Here, we summarize three types of sensitivity analysis as follows:
[9–11]
 C.-J. Lin et al. / Applied Soft Computing 11 (2011) 5341–5349
• Type I (Basics Invariancy): Type I sensitivity is the traditional sen-
sitivity analysis, which determines Type I Sensitivity Range (Type
I SR) such that the current optimal basic solution remains optimal
in the perturbed problem.
• Type II (Support Set Invariancy): Type II sensitivity, depending on
the current optimal solution (not necessarily a basic solution), is
to determine Type II Sensitivity Range (Type II SR) so that vari-
ables with a zero and with a positive value in the given optimal
solution remain zero and positive in the optimal solution of the
perturbed problem, respectively.
• Type III (Optimal Partition Invariancy): Type III sensitivity, depend-
ing only on the problem data and perturbed parameter, is to
determine Type III Sensitivity Range (Type III SR) for which the
rate of change of the optimal value function remains unchanged.
In case that the range happens to be a breakpoint, Type III SR is
defined as [0,0].
Simplex-type algorithms divide variables of the solution into
basic variables and nonbasic variables. Type I SR focuses on keep-
ing the set of optimal basic variables unchanged and overlooks the
effects caused by degeneracy. When degeneracy occurs, Type I SR
provides incomplete information. Type II and Type III SRs that dis-
regard the impractically degenerate basic variables are suitable for
the degenerate problems. Type II SRs in FAP, in this paper, are to
determine the largest range of variations such that a given optimal
assignment remains an optimal assignment of the perturbed prob-
lem. Keeping the assignment unchanged makes the workers work
with the same machine, the machines settle on its original loca-
tions, etc. Focusing only on the sensitivity ranges of non-degenerate
basic variables makes Type II SR differ from Type I SR. The opti-
mal value function of (2.3) is composed of several linearity ranges
with different slopes. In mathematical insight, Type III SR is a lin-
ear range or breakpoint of the optimal value function and depends
only on the model parameters. In managerial insight, Type III SR
is to determine the largest range of variations such that all the
optimal assignments remain optimal assignments of the perturbed
problem [10,12]. It is then obvious that Type II and III Ranges are
μij
qij
0 αij βij cij
(a)
μij
qij
0 αij βij cij
(c)
Fig. 1. Membership function of various perturbations.
5343
identical when there is only one optimal assignment. Multiple opti-
mal assignments make Type II and III Ranges different and will give
more choices for decision-makers. More optimal solutions make
the decision-making more flexible. When decision-makers empha-
size the flexibility of optimal decision-making, Type III SR is the
proper display.
2.3. Various perturbations of the FAP
According to model (1.4), there are four groups of parameters
˛ij , ˇij , qij and  ij . We divide them into five different cases when
discussing the occurrences of perturbation follows:
Case 1: Perturbation occurs in ˛ij and ˇij while  ij and qij are
unchanged (see Fig. 1(a)). Hence, the variations of ˛ij and ˇij are
kept the same. In real-world applications, this case occurs when
the wages of workers are alerted or a new skill reduces the oper-
ating time of the workers.
Case 2: Perturbation occurs in ˛ij while ˇij and  ij are unchanged
(see Fig. 1(b)). This case happens when the managers adjust admis-
sible quality level while the rate of the worker unchanged. One has
to lower (raise) the highest quality qij when the admissible quality
level is raised (lowered).
Case 3: Perturbation occurs in qij with ˛ij and ˇij remain
unchanged; hence  ij changes with the perturbation of qij (as
shown in Fig. 1(c)). It describes the case when the worker is altered
or trained so as to change the highest quality qij and the rate of
quality progress  ij .
Case 4: Perturbation occurs in qij with  ij and ˛ij remain
unchanged; hence, ˇij changes with the perturbation of qij (as
shown in Fig. 1(d)). It describes the case when the worker can
enhance the highest quality qij because of the job training.
Generally, Case 1 frequently happens in real practice and sensi-
tivity ranges of Cases 2 and 3 can be determined by the approach of
Case 1 though computation may be somewhat tedious. In addition,
the perturbation of Case 4 will not influence the original optimal
μij
qij
0 αij βij cij
(b)
μij
qij
0 αij βij cij
(d)
5344 C.-J. Lin et al. / Applied Soft Computing 11 (2011) 5341–5349

assignment without increasing the upper bounds of total cost b. Labeling procedure: Suppose (h, k) is the pivot cell,
Therefore, we will focus on investigating Case 1 with perturbation
√
in one pair of ˛ij and ˇij in this paper. Suppose ˛st and ˇst is the (1) Label column k with ( ).
√
pair of parameters perturbed with variation st , then the perturbed (2) If column q is labeled, then we label row i with ( ) where xiq = 1.
√
problem can be shown as follows: (3) If row p is labeled, then we label column j with ( ) where Qpj = 0
n n and xpj = 0. In case column j has already been labeled, then we
b− i=1
˛ x
j=1 ij ij
− st xst keep its label.
max n n (2.5) (4) Repeat (2) and (3) until further labeling is impossible.
b−a+ i=1
 x
j=1 ij ij
s.t. the constraints of (1.4)
Revaluing procedure: (Expression m ← n stands for “m being
In addition, we assume that any variations of parameter will replaced by n” in the procedure.)
always satisfy the conditions about a and b, as suggested and Let J = {(i,j)|Qij > 0 and row i is labeled, but column j is not labeled}
derived in Lin and Wen [7]. and ε = min {Qij }. If row i is labeled, then ui ← ui − ε. If column j is
(i,j) ∈ J
labeled, then vj ← vj + ε. We then update related Qij and erase all
3. Type II sensitivity analysis labels.
We outline the procedure of Type II SA algorithm for determin-
Assume that the current optimal assignment E* of (1.4) has ing Lst , (s, t) ∈
/ E*, as follows:
been determined by the LW algorithm, and the decision-maker is
going to identify the sensitivity ranges. The proposed LW algorithm Step 1: Update Qst ← Qst + st .
of sensitivity analysis for Type II SR, called Type II SA algorithm Step 2: Assume all Qij ≥ 0, we have (or update) the lower bound
for short hereafter, will be discussed in two parts. The first part st ≥ Lst . Choose (h, k), where Lst occurs, as the pivot cell and pro-
is when the perturbation occurs in a cell (s, t) ∈ / E*, i.e., cell of ceed with the Labeling procedure. If row h has no label, go to Step
∗ = 0, the other is when the perturbation occurs in a cell (s,
xst 3; otherwise, stop, and the current inequality st ≥ Lst is the Type
∗ = 1. In addition, some theorems which help
t) ∈ E*, i.e., cell of xst II SR.
simplify the determination of Type II SRs are derived in this Step 3: Perform the Revaluing procedure and return to Step 2.
section.
3.2. Type II SA algorithm for perturbation in the assigned cell
3.1. Type II SA algorithm for perturbation in the unassigned cell
To determine the Type II SR of an assigned cell is more compli-
Let (2.5) be the perturbed problem. Further, denote Ust and Lst cated than that of an unassigned cell.
as the upper and lower bounds of st ’s Type II SR, respectively. Theorem 3.3. Let E* be the current optimal assignment of (1.4). Fur-
thermore, suppose E* is also the minimal assignment of matrix [ ij ]. If
Theorem 3.1. Let E* be the current optimal assignment of (1.4). If
(s, t) ∈ E* and st ∈ (−∞, 0], then E* remains the optimal assignment
(s, t) ∈
/ E* and st ∈ [0, ∞), then E* remains the optimal assignment of
of model (2.5).
(2.5).
Proof. Let E# , Z0 (E*), Z0 (E# ), Zp (E*) and Zp (E# ) be defined as that
Proof. Let E# be any feasible assignment of (1.4). Denote Z0 (E*)
in Theorem 3.1. We obtain (3.1)–(3.3).
and Z0 (E# ) as the objective values of E* and E# on (1.4), respectively. ∗ = 1, we have
Since (s, t) ∈ E*, i.e., xst
We obtain
(−st )
Z0 (E ∗ ) ≥ Z0 (E # ) (3.1) Zp (E ∗ ) = Z0 (E ∗ ) +  (3.5)
b−a+ 
(i,j) ∈ E ∗ ij

We further denote Zp (E*) and Zp (E# ) as the objective values of E* Furthermore, since E* is the minimal assignment of matrix [ ij ], we
and E# on perturbation of (2.5), respectively. Hence, then have
 
∗)
(−st · xst ij ≤ ij (3.6)
∗ ∗
Zp (E ) = Z0 (E ) +  (3.2)
b−a+ 
(i,j) ∈ E ∗ ij (i,j) ∈ E ∗ (i,j) ∈ E #

# ∈ {0, 1},  ∈ (−∞, 0] and (3.6), we have

By conditions xst
and st

#)
(−st · xst Zp (E ∗ ) ≥ Zp (E # ) (3.7)
Zp (E # ) = Z0 (E # ) +  (3.3)
b−a+ 
(i,j) ∈ E # ij (3.7) shows that E* remains an optimal assignment of (2.5).
The next corollary follows directly from Theorem 3.3.
Since (s, t) ∈ ∗ = 0,  ∈ [0, ∞) and x# ∈ {0, 1}, we have
/ E*, i.e., xst st st
Corollary 3.4. Let E* be the current optimal assignment of (1.4).
Zp (E ∗ ) = Z0 (E ∗ ) ≥ Z0 (E # ) ≥ Zp (E # ) (3.4) Furthermore, suppose E* is also the minimal assignment of matrix [ ij ].
If (s, t) ∈ E*, then the lower bound of st ’s Type II SR Lst → −∞.
Therefore, (3.4) shows that E* remains an optimal assignment of
We outline the procedure of Type II SA algorithm, applying the
(2.5).
Labeling procedure and Revaluing procedure, for determining Lst
The following corollary is a direct result of Theorem 3.1.
and Ust for (s, t) ∈ E*, as follows:
Corollary 3.2. Let E* be the current optimal assignment of (1.4). If Phase 0:
(s, t) ∈
/ E*, then the upper bound of st ’s Type II SR Ust → ∞.
Step 1: Suppose st is the variation, then update f value as
As for determining the lower bound of st ’s Type II SR Lst , (s,
t) ∈
/ E*, we may apply Type II SA algorithm. However, we must

b− (i,j) ∈ E ∗
˛ij − st
define two required procedures, namely the Labeling procedure and f ← 
b−a+ 
(i,j) ∈ E ∗ ij
Revaluing procedure in advance.
 C.-J. Lin et al. / Applied Soft Computing 11 (2011) 5341–5349 5345

Table 2
After f is updated, we need to update u1 ,. . .,un , v1 ,. . .,vn . One may ij ’s Type I SRs of Example 1.
keep all ui unchanged, and update values vs ← vs + · st − st , and
Type I SR
vj ← vj + · ij , for j = 1,2,. . .,n and (i,j) ∈ E*, i =
/ s, where  
[0, ∞) [−8, 0] [0, 20/3]
(−∞, 0] [0, 8] [−8, ∞)

= st [−5, ∞) [0, 5] [−12, ∞)
b−a+ 
(i,j) ∈ E ∗ ij

Then we update each Qij ← ui + vj + ˛ij + f· ij , for (i,j) ∈

/ E*. Table 3
Initial table for determining the lower bound L11 .
Step 2: Assume all Qij ≥ 0, and then we have some inequali-
ties of st . Derive intersection of these inequalities, denoted as v1 = −9.5 v2 = −6.5 v3 = −5
Lst ≤ st ≤ Ust , where Lst occurs at cell (h1 , k1 ) and Ust occurs at 4+ Δ 11 3 2
cell (h2 , k2 ). Update Lst and Ust . If Lst → −∞, then go to Phase 2; u1 = 0 0.5+ Δ 11 5.5
otherwise, go to Phase 1. 10 15 5
4 6 7
Phase 1: This phase begins with the table obtained in Phase 0.
u 2 = −0.5 5 7.5
10 10 10
Step 3: If E* is the minimal assignment of matrix [ ij ], we have
7 4 6
Lst → −∞ and go to Phase 2; otherwise, go to Step 4.
Step 4: Choose (h1 , k1 ), where Lst occurs, as the pivot cell and pro- u3 = −0.5 0 6.5
ceed with the Labeling procedure. If row h1 has not been labeled, 5 5 10
then go to Step 5; otherwise, stop and the current Lst is the lower
bound of st ’s Type II SR.
section, we first discuss the situation that the perturbation of Case 1
Step 5: Perform the Revaluing procedure. Assume all Qij ≥ 0 and
occurs in an unassigned cell, and then discuss that when it occurs in
derive intersection of these inequalities, we update the lower
an assigned cell. For comparison, Table 2 shows Type I SRs solving
bound Lst . Go to Step 4.
(2.3) with y12 , y13 , y21 , y22 , y32 ,  as the basic variables obtained by
the commercial software LINGO 10.
Phase 2: This phase also begins with the table obtained in Phase
0. Example 1. Assume that the perturbation of Case 1 occurs in unas-
signed cell (1, 1) and 11 denotes the variation of ˛11 and ˇ11 .
Step 6: If Ust is unbounded, then stop and the current Lst and Ust Tables 3 and 4 show the initial table and final table for determining
form the st ’s Type II SR; otherwise, go to Step 7. the lower bound of Type II SR in the unassigned cell, respectively.
Step 7: Choose (h2 , k2 ), where Ust occurs, as the pivot cell and pro- We have 11 ≥ −8 as Type II SR. The procedure of Example 1 is
ceed with the Labeling procedure. If row h2 has not been labeled, detailed in Appendix A.
then go to Step 8; otherwise, stop and the current Lst and Ust form
Example 2. Assume that the perturbation of Case 1 occurs in
the st ’s Type II SR.
assigned cell (1, 3) with the variation 13 of ˛13 and ˇ13 , and then
Step 8: Perform the Revaluing procedure. Assume all Qij ≥ 0 and
update ˛13 ← ˛13 + 13 in the table. Tables 5 and 6 show the initial
derive intersection of these inequalities, we update the upper
table and final table for determining the upper bound of Type II SR
bound Ust . Go to Step 7.
in the assigned cell, respectively. We have 13 ≤ 80/11 as Type II
SR.
3.3. Numerical example

Consider the following 3 × 3 fuzzy assignment problem adopted 3.4. Result of Type II sensitivity analysis
from example 1 of Lin and Wen [7]. The matrices [c̃ij ] and [qij ] are:
⎡ ⎤ ⎡ ⎤ According to Type II SA algorithm, we can summarize ij ’s
4, 13 3, 12 2, 6 0.9 0.6 0.8 Type II SRs of the perturbation of Case 1 in Table 7. Comparing
⎢ ⎥ ⎢ ⎥ Tables 2 and 7 shows that Type I SR is a subset of Type II SR since
[c̃ij ] = ⎣ 4, 13 6, 14 7, 15 ⎦ , [qij ] = ⎣ 0.9 0.8 0.8 ⎦
changing the optimal basis does not ensure that the optimal assign-
7, 10 4, 8 6, 12 0.6 0.8 0.6 ment will be changed.

and one can then obtain

⎡ ⎤ ⎡ ⎤ Table 4
4 3 2 13 12 6 Final table for determining the lower bound L11 .
⎢ ⎥ ⎢ ⎥
[˛ij ] = ⎣ 4 6 7 ⎦ , [ˇij ] = ⎣ 13 14 15 ⎦ and [ij ]
v1 = −2 v2 = −4 v3 = −5
7 4 6 10 8 12
4+ Δ 11 3 2
⎡ ⎤ u1 = 0 8+ Δ 11
10 15 5 8
⎢ ⎥ 10 15 5
= ⎣ 10 10 10 ⎦
4 6 7
5 5 10 u 2 = −8 0 0
10 10 10
Further specify the sum of maximum elements in each row of
7 4 6
matrix [ˇij ] as b = 40, and the sum of minimum elements in each
row of matrix [˛ij ] as a = 10. The fuzzy assignment problem was u3 = −3 5 4
solved with the LW algorithm. The optimal assignment is E* = {(1,
3), (2, 1), (3, 2)} and the associated global maximum is 0.6. In this 5 5 10
5346 C.-J. Lin et al. / Applied Soft Computing 11 (2011) 5341–5349

Table 5 analysis, is to determine the linearity range of fst which include

Initial table for determining the upper bound U13 . st = 0. There are three types of graphs of fst around st = 0 as shown
Δ13 Δ13 9Δ13 in Fig. 2.
v1 = −9.5 + v2 = −6.5 + v3 = −5 −
5 10 10
The following three theorems, derived from Hadigheh and
4 3 2+ Δ13 Terlaky’s [11] results of linear programming perturbing cost coef-
Δ13 ficients, show the relationships between Type II and Type III SRs,
u1 = 0 0.5 5.5 −
5 which can be applied to (2.3), and help determine Type III SR when
10 15 5 Type II SR is known.
4 6 7 Theorem 4.1. Suppose that fst at st = 0 is not a breakpoint, then
Δ13 11Δ13 st ’s Type II and Type III SRs are identical.
u 2 = −0.5 5− 7.5 −
10 10
Theorem 4.2. Either the lower bound or the upper bound of st ’s
10 10 10
Type II SR is 0, then fst at st = 0 is a breakpoint, i.e., st ’s Type III
7 4 6
SR = [0,0].
Δ13 11Δ13
u3 = −0.5 6.5 −
Theorem 4.3. Suppose that fst at st = 0 is a breakpoint, then st ’s
10 10

5 5 10
Type II SR will either be the immediate left-side or right-side linear
range of fst at st = 0.

Table 6 Further denote m− +

st and mst as the left-side slope and right-side
Final table for determining the upper bound U13 . slope of fst at st = 0, respectively.

Theorem 4.4. Suppose there are k optimal assignments of

Δ13 11Δ13
v1 = −4.5 v2 = −6.5 + v3 = 0.5 − model (1.4), denoted as E1∗ , E2∗ , . . . , Ek∗ . Define value  = 1/(b − a +

10 10
 )
(i,j) ∈ E ∗ ij
as that in (2.1).
4 3 2+ Δ13

u1 = −5.5 +
Δ13
0 0 (1) If (s, t) ∈ ∀Ei∗ , i ∈ {1, 2,. . .,k}; then m− +
st = mst = −. As shown in
5
Fig. 2(c).
10 15 5 / ∀Ei∗ , ∀i ∈ {1, 2,. . .,k}; then m−
(2) If (s, t) ∈ +
st = mst = 0. As shown in
4 6 7 Fig. 2(a).
Δ13 Δ13 11Δ13 (3) If there exist i and j such that (s, t) ∈ Ei∗ and (s, t) ∈/ Ej∗ , then m−
st =
u2 = −5.5 + 8−
5 10 10 − and m+
st = 0, as shown in Fig. 2(b).
10 10 10
7 4 6 Proof. Since xij in model (1.4) are 0–1 variables and yij = ·xij , (1),
Δ 13Δ13 (2) and (3) of this theorem can easily be proved by m−
st = −max{yst :
i
u3 = −0.5 5 − 13 12 − ∀ i
10 10 i ∈ E ∗ } and m+ = −min{yi : yi ∈ E ∗ } [13].
yst i st st st i
5 5 10 ∀i

It is trivial to get the following results from Theorem 4.4.

Table 7
ij ’s Type II SRs of Example 1.
Corollary 4.5. Suppose an optimal assignment E* of model (1.4) is
known,
Type II SR
  (1) If (s, t) ∈ E*, then m+
[−8, ∞) [−12, ∞) (−∞, 80/11] st = 0, as shown in Fig. 2(a) or (b).
(−∞, 5] [−5, ∞) [−8, ∞) / E*, then m−
(2) If (s, t) ∈ st = −, as shown in Fig. 2(b) or (c).
[−5, ∞) (−∞, 5] [−12, ∞)
After ij ’s Type II SRs are obtained using Type II SA algorithm,
Type III SRs can be determined almost at the same time. If the lower
4. Type III sensitivity analysis bound of ij ’s Type II SR is negative and the upper bound is posi-
tive, i.e., fij at ij = 0 is not a breakpoint, by Theorems 4.1 and 4.2,
4.1. Type III sensitivity ranges derived by using Type II sensitivity Type III SR is coincident with Type II SR. Otherwise, fij at ij = 0
ranges is a breakpoint and ij ’s Type III SR is [0,0]. In the case that ij ’s
Type II and Type III SRs are identical, m− ij
= m+
ij
occurs. If (i,j) ∈ E*,
Denote fst as the optimal objective function with variation st .
then m−
ij
= m+
ij
/ E*, then m−
= −; if (i, j) ∈ ij
= m+
ij
= 0. In the case that
Type III sensitivity analysis, or optimal partition type sensitivity

fst fst fst

Type III R Type III R
Type III R
= [0,0] = (−∞,Ust]
= [Lst,∞)

Δst Δst 0 Ust Δst

Lst 0 0

(a) (b) (c)

Fig. 2. Graphs of fst around st = 0.
 C.-J. Lin et al. / Applied Soft Computing 11 (2011) 5341–5349 5347

Table 8 lem data and perturbing parameter, it is independent of Type I and

ij ’s Type II SRs of Example 3.
Type II sensitivity analysis; (3) if there is only one optimal assign-
Type II sensitivity analysis ment, then Type II SR and Type III SR are identical; and (4) when
  optimal value function happens to be a breakpoint at st = 0, Type
(−∞, 0] [−247/51, ∞) [0, ∞)
III SR = [0,0] will be a subset of Type II SR.
[0, ∞) [−5, ∞) (−∞, 0]
There are several application areas of this study, such as human
[−247/51, ∞) (−∞, 247/56] [−588/51, ∞)
resource management, job assignment planning, transportation
planning, fuzzy network planning, etc. Besides, the further exten-
Table 9 sions and applications of this research are suggested in following
ij ’s Type III SRs of Example 3. directions:
Parameter Type III sensitivity analysis
  1. Consider the perturbation more than one parameter. In the real
[0, 0] [−247/51, ∞) [0, 0]
world applications, it is usual that perturbation occurs on more
ij [0, 0] [−5, ∞) [0, 0]
than one parameters, e.g., perturbing ˛ij and  ij at the same time,
[−247/51, ∞) (−∞, 247/56] [−588/51, ∞)
one column of matrix ˛ij , or one row of matrix qij , etc.
2. Extend the two-dimensional fuzzy assignment problem to
Table 10 multi-dimensional fuzzy assignment problem and develop
fij ’s slopes at ij = 0. a solution algorithm and sensitivity analysis procedure for
ij Left-side slope Right-side slope the multi-dimensional fuzzy assignment problem. Multi-
dimensional assignment problems have been widely used in
11 −1/51 0
12 0 0 scheduling problem, capital investment and facility location
13 −1/51 0 problem, etc.
21 −1/51 0
22 0 0
Appendix A. Detailed procedures of Example 1
23 −1/51 0
31 0 0
32 −1/51 −1/51 Replace ˛11 with ˛11 + 11 = 4 + 11 .
33 0 0
Step 1:
Update Q11 ← 0.5 + 11 . Please refer to Table A1.
ij ’s Type Range = [0,0], the right-side and left-side slopes of fij will
Step 2:
be different. It must be that m+ = 0 and m− = − whether (i, j) ∈ E*
ij ij Set 0.5 + 11 ≥ 0, we get 11 ≥ −0.5 as the transient lower
or not.
bound. Choose (1, 1) as the pivot cell, make the assumption
11 < −0.5, proceed with the Labeling procedure by assigning label
4.2. Numerical example √ √
( ) to column 1, and label ( ) to row 2 due to x21 = 1. Row 1 has
no label, go to Step 3. Please refer to Table A2.
Example 3. Suppose we change element c̃23 of [c̃ij ] in Example 1
into 2, 6,  23 = 5 and let a = 8, b = 39. (Note that the new problem Table A1
has alternative optimal assignment.)
v1 = − 9.5 v2 = − 6.5 v3 = −5
We can solve Example 3 by the LW algorithm, and get the opti-
mal assignment E1∗ = {(1, 1), (2, 3), (3, 2)}. The associated global 4+ Δ 11 3 2
performance equals 29/51. Table 8 shows ij ’s Type II SRs with u1 = 0 0.5+ Δ 11 5.5
respect to the optimal assignment E1∗ . Note that if the current opti- 10 15 5
mal assignment is E2∗ = {(1, 3), (2, 1), (3, 2)}, ij ’s Type II SRs may 4 6 7
be different. u 2 = − 0.5 5 7.5
Table 9 shows the ij ’s Type III SRs of Example 3. This table can 10 10 10
be determined by Table 8. Type III SRs are the same whether the 7 4 6
current optimal assignment is E1∗ or E2∗ . Table 10 shows the left-side
u3 = − 0.5 0 6.5
and right-side slope of the optimal objective function.
5 5 10

5. Conclusions

Table A2
The constraints of FAP are identical to those of AP, which is inher-
ently a highly primal degenerate LP model. Degeneracy negatively
influences the computation of an optimal solution and sensitivity
v1 = −9.5 v2 = −6.5 v3 = −5
analysis of LP. This paper proposes a Type II SA algorithm for iden-
tifying Type II SR, which can also be extended for identifying Type 4+ Δ 11 3 2
III SR of the FAP. The Type II SA algorithm generates a streamlined u1 = 0 0.5+ Δ 11 5.5
procedure from searching the optimal solution till performing the 10 15 5
sensitivity analysis of assignment problem. Commercially available 4 6 7
software can only help to determine Type I SRs, which is impractical u 2 = −0.5 5 7.5
in degenerate problems. However, Type II and Type III SRs pro- 10 10 10
vide more precise and accurate information and help make correct
7 4 6
decisions. We summarize the relationships between three types of
u3 = −0.5 0 6.5
sensitivity ranges as follows: (1) Type I SR is always a subset of
Type II SR; (2) since Type III sensitivity depends only on the prob- 5 5 10
5348 C.-J. Lin et al. / Applied Soft Computing 11 (2011) 5341–5349

Table A3 Table A6

v1 = −4.5 v2 = −6.5 v3 = −5 v1 = −2 v2 = −4 v3 = −5
4+ Δ 11 3 2 4+ Δ 11 3 2
u1 = 0 5.5+ Δ 11 5.5 u1 = 0 8+ Δ 11 8
10 15 5 10 15 5
4 6 7 4 6 7
u 2 = −5.5 0 2.5 u 2 = −8 0 0
10 10 10 10 10 10
7 4 6 7 4 6
u3 = −0.5 5 6.5
u3 = −3 5 4
5 5 10
5 5 10
Table A4

Set Q11 ≥ 0, we update 11 ≥ −8. Choose (1, 1) as the pivot cell,
make the assumption 11 < −8, proceed with the Labeling proce-
v1 = −4.5 v2 = −6.5 v3 = −5 √ √
dure by assigning label ( ) to column 1, and label ( ) to row 2 due
√
4+ Δ 11 3 2 to x21 = 1. Assign label ( ) to column 2 and column 3 due to Q22 = 0
√ √
u1 = 0 5.5+ Δ 11 5.5 and Q23 = 0. Assign label ( ) to row 3 due to x32 = 1, and label ( )
10 15 5 to row 1 due to x13 = 1. Please refer to Table A6.
4 6 7
u 2 = −5.5 0 2.5 Since row 1 has been labeled, we stop. When 11 ≥ −8, the
optimal assignment remains optimal.
10 10 10
7 4 6
References
u3 = −0.5 5 6.5
5 5 10 [1] H. Haken, M. Schanz, J. Starke, Treatment of combinatorial optimal problems
using selection equations with cost terms. Part I. Two-dimensional assignment
problem, Physica D 134 (1999) 227–241.
[2] N. Belacela, M.R. Boulasselb, Multicriteria fuzzy assignment method: a useful
Step 3: tool to assist medical diagnosis, Artificial Intelligence in Medicine 21 (2001)
Find J = {(2,2), (2,3)}, ε = min{5, 7.5} = 5, and update u2 ← −5.5, 201–207.
v1 ← −4.5, Q11 ← 5.5 + 11 , Q22 ← 0, Q23 ← 2.5, Q31 ← 5 as shown [3] M. Ridwan, Fuzzy preference based traffic assignment problem, Transportation
Research Part C 12 (2004) 209–233.
in the following table. Erase all labels and return to Step 2. Please
[4] L. Liu, Y. Li, The fuzzy quadratic assignment problem with penalty: new mod-
refer to Table A3. els and genetic algorithm, Applied Mathematics and Computation 174 (2006)
Step 2: 1229–1244.
[5] Y. Feng, L. Yang, A two-objective fuzzy k-cardinality assignment problem, Jour-
Set Q11 ≥ 0, we update the transient lower bound 11 ≥ −5.5.
nal of Computational and Applied Mathematics 197 (2006) 233–244.
Choose (1, 1) as the pivot cell, make the assumption 11 < −5.5, [6] L. Liu, X. Gao, Fuzzy weighted equilibrium multi-job assignment problem
√
proceed with the Labeling procedure by assigning label ( ) to col- and genetic algorithm, Applied Mathematical Modelling 33 (2009) 3926–
√ 3935.
umn 1, and label ( ) to row 2 due to x21 = 1. Further assign label
√ √ [7] C.J. Lin, U.P. Wen, The labeling algorithm for the fuzzy assignment problem,
( ) to column 2 due to Q22 = 0, and label ( ) to row 3 due to x32 = 1. Fuzzy Sets and Systems 142 (2004) 373–391.
Row 1 has no label, go to Step 3. Please refer to Table A4. [8] R.E. Bellman, L.A. Zadeh, Decision-making in a fuzzy environment, Management
Step 3: Science 17B (1970) 141–164.
[9] T. Koltai, T. Terlaky, The difference between managerial and mathematical
Find J = {(2,3), (3,3)}, ε = min{2.5, 6.5} = 2.5, and update u2 ← −8, interpretation of sensitivity analysis results in linear programming, Interna-
u3 ← −3, v1 ← −2, v2 ← −4, Q11 ← 8 + 11 , Q12 ← 8, Q23 ← 0, tional Journal of Production Economics 65 (2000) 257–274.
Q33 ← 4 as shown in the following table. Erase all labels and return [10] C.J. Lin, U.P. Wen, Sensitivity analysis of the optimal assignment problem, Euro-
pean Journal of Operational Research 149 (2003) 35–46.
to Step 2. Please refer to Table A5. [11] A.G. Hadigheh, T. Terlaky, Sensitivity analysis in linear optimization: invariant
Step 2: support set intervals, European Journal of Operational Research 169 (2006)
1158–1175.
[12] C.J. Lin, U.P. Wen, Sensitivity analysis of objective function coefficients of the
Table A5 assignment problem, Asia–Pacific Journal of Operational Research 24 (2007)
203–221.
[13] J.K. Ho, Computing true shadow prices in linear programming, Informatica 11
(2000) 421–434.
v1 = −2 v2 = −4 v3 = −5
4+ Δ 11 3 2 Chi-Jen Lin is an Associate Professor in the Department of
u1 = 0 8+ Δ 11 8 Industrial Engineering and Management at Ta Hwa Insti-
tute of Technology, Hsinchu, Taiwan, ROC. He received his
10 15 5 Ph.D. in Industrial Engineering and Engineering Manage-
4 6 7 ment from National Tsing Hua University. His research
interests are operations research, fuzzy theory and MCDM.
u 2 = −8 0 0 His papers have appeared in the European Journal of Oper-
10 10 10 ational Research, Fuzzy Set and Systems, Asia-Pacific Journal
of Operational Research, Applied Mathematical Modelling,
7 4 6 Operations Research Letters, and Expert Systems with Appli-
cations.
u3 = −3 5 4

5 5 10
 C.-J. Lin et al. / Applied Soft Computing 11 (2011) 5341–5349 5349

Ue-Pyng Wen is a Professor in the Department of Pei-Yi Lin received her M.S. degree in Industrial Engineer-
Industrial Engineering and Engineering Management at ing and Engineering Management from National Tsing
National Tsing Hua University, Hsinchu, Taiwan, ROC. Hua University. Her research interests are operations
He received his Ph.D. in Industrial Engineering from the research and production management.
State University of New York at Buffalo, USA. His primary
research interests are in the area of multi-criteria decision
making, production management, and applied operations
research. He has published his research in journals such as
IIE Transactions, Journal of the Operational Research Society,
European Journal of Operational Research, Computers and
Operations Research, Fuzzy Sets and Systems and Interfaces.