Академический Документы
Профессиональный Документы
Культура Документы
2A
j
vD
j
=h
j
q
=
2A
j
v
X
i
l
i
x
ij
=h
j
q
: (1)
Moreover, the optimal reorder level and the safety stock level at each DC can be derived as
r
j
= D
j
L
j
z
a
C
j
L
j
q
(2)
and
SS
j
= z
a
C
j
L
j
q
= z
a
X
i
v
i
x
ij
X
k
l
jk
y
jk
q
= z
a
X
i
X
k
v
i
l
jk
x
ij
y
jk
q
; (3)
where a is called a service level, which means that stockouts occur during the lead time with a probability of a or less, and
z
a
is the standard normal deviate such that P(Z 6 z
a
) = a. For example, if a = 0.025, which means the lead time demands are
satised with a probability of 97.5%, z
a
= 1.96. According to Elsayed and Boucher (1994), the optimal inventory cost function
at each DC is derived as
C
INV+
j
=
2A
j
h
j
vD
j
q
h
j
z
a
C
j
L
j
q
=
2A
j
h
j
v
X
i
l
i
x
ij
q
z
a
h
j
X
i
X
k
v
i
l
jk
x
ij
y
jk
q
: (4)
By use of Eq. (4), the total annual cost in the whole supply chain network can be derived as
TC(x; y; z) =
X
k
g
k
z
k
X
j
F
j
u
j
X
i
X
j
X
k
vl
i
p
jk
x
ij
y
jk
X
i
X
j
vl
i
c
ij
x
ij
X
j
2A
j
h
j
v
X
i
l
i
x
ij
q
X
j
z
a
h
j
X
i
X
k
v
i
l
jk
x
ij
y
jk
q
; (5)
where the rst two terms represent the xed annual supplier and DC setup costs, respectively, and the third and fourth
terms represent the inbound and outbound network transportation costs, respectively, and the last two terms indicate
the on-hand inventory and the safety stock costs at DCs, respectively. Furthermore, the relation
P
k
y
jk
= u
j
holds for each
j, so that the total xed DC setup cost
P
j
F
j
u
j
can be transformed into
P
j
P
k
f
jk
y
jk
, by introducing new parameters f
jk
which
denote the annual xed cost of locating DC j and assigning it to supplier k. Finally, Eq. (5) can be expressed as in Eq. (6) of the
following problem formulation:
(P)
minimize TC(x; y; z) =
X
k
g
k
z
k
X
j
X
k
f
jk
y
jk
X
i
X
j
X
k
vl
i
p
jk
x
ij
y
jk
X
i
X
j
vl
i
c
ij
x
ij
X
j
2A
j
h
j
v
X
i
l
i
x
ij
r
X
j
z
a
h
j
X
i
X
k
v
i
l
jk
x
ij
y
jk
r
; (6)
subject to
X
j
x
ij
= 1; \i; (7)
x
ij
6
X
k
y
jk
; \i; j; (8)
X
i
l
i
x
ij
6 B
j
; \j; (9)
X
k
y
jk
6 1; \j; (10)
y
jk
6 z
k
; \j; k; (11)
x
ij
; y
jk
; z
k
0; 1; \i; j; k: (12)
4 S. Park et al. / Transportation Research Part E xxx (2010) xxxxxx
ARTICLE IN PRESS
Please cite this article in press as: Park, S., et al. A three-level supply chain network design model with risk-pooling and lead times. Trans-
port. Res. Part E (2010), doi:10.1016/j.tre.2009.12.004
Eqs. (7) and (10) represent the single-sourcing constraints for each retailer and each DC, respectively. Eqs. (8) and (11)
ensure that each retailer and each opened DC can only be assigned to one of their providers to be opened. Eq. (9) represents
the capacity for each DC.
Problem (P) is a nonlinear integer programming problem, which is known as hard to solve in general, due to two square-
root terms in the objective function. Moreover, the problem (P) is obviously NP-hard since it is an extension of the Capac-
itated Facility Location Problem (CFLP), which is well known as NP-hard (Mirchandani and Francis, 1990), into a nonlinear
three-level location-inventory problem. For the problem, a Lagrangian relaxation approach is adapted to provide a good fea-
sible solution in reasonable time.
3. Solution approach
3.1. Lagrangian subproblems
In Problem (P), Eqs. (7) and (11) are relaxed using Lagrangian multipliers, k and c (P0), respectively. The resulting relaxed
problem is derived as
(LR(k; c))
Z(k; c) = minimize
X
k
g
k
X
j
c
jk
!
z
k
X
j
X
k
(f
jk
c
jk
)y
jk
X
i
X
k
vl
i
p
jk
x
ij
y
jk
X
i
(vl
i
c
ij
k
i
)x
ij
"
2A
j
h
j
v
X
i
l
i
x
ij
r
z
a
h
j
X
i
X
k
v
i
l
jk
x
ij
y
jk
r
#
X
i
k
i
; (13)
subject to (8)(10), x
ij
, y
jk
{0, 1}, " i, j, k, and c
jk
P0, "j, k.
Then, a Lagrangian dual problem is derived as
(LD)
Z
D
= maximize
k;cP0
Z(k; c): (14)
Note that Z(k, c) is a lower bound for the original problem (P) for all k and c (P0). The relaxed problem (LR (k, c)) can then
be decomposed into two kinds of subproblems as follows:
(SP1)
h
+
SP1
= minimize
X
k
g
k
X
j
c
jk
!
z
k
; (15)
subject to
X
k
z
k
P1; (16)
z
k
0; 1; \k; and c
jk
P0; \j; k;
and
(SP2)
h
+
SP2
= minimize
X
j
X
k
(f
jk
c
jk
)y
jk
X
i
X
k
vl
i
p
jk
x
ij
y
jk
X
i
(vl
i
c
ij
k
i
)x
ij
"
2A
j
h
j
v
X
i
l
i
x
ij
q
z
a
h
j
X
i
X
k
v
i
l
jk
x
ij
y
jk
q i
; (17)
subject to (8)(10), x
ij
, y
jk
{0, 1}, " i, j, k, and c
jk
P0, "j, k.
The rst subproblem (SP1), is the problem of selecting the locations of suppliers to minimize the sum of xed supplier
setup costs. Note that Eq. (16) is a valid inequality. The second subproblem (SP2), is the problem of selecting the locations
of DCs, assigning each retailer to one of the opened DCs with capacity restrictions, and assigning each opened DC to one of
the suppliers to minimize the sum of the xed DC assignment, transportation and inventory costs. Note that the subproblem
(SP2) can be decomposed into a number of subproblems for each j.
3.2. Analysis of the subproblems
The rst subproblem (SP1) is easily solved by the following procedure. If there is any k at which the associated coefcient
is negative, then set at z
+
k
= 1, and otherwise set at z
+
r
= 0 for all r k. If all the associated coefcients have positive values,
then set at z
+
k
= 1, where
k is the index associated with the minimum value of the coefcients, and z
+
r
= 0 for all r
k. How-
ever, no one in the literature has dealt with the subproblem (SP2) yet. Therefore, in this paper, an optimal solution algorithm
for the subproblem (SP2) is derived based on some solution properties to be characterized in this section.
S. Park et al. / Transportation Research Part E xxx (2010) xxxxxx 5
ARTICLE IN PRESS
Please cite this article in press as: Park, S., et al. A three-level supply chain network design model with risk-pooling and lead times. Trans-
port. Res. Part E (2010), doi:10.1016/j.tre.2009.12.004
3.2.1. Analysis of the subproblem (SP2)
The optimal solution property of the subproblem (SP2) is now characterized.
Property 1. Denoted by (x
+
; y
+
) be the optimal solution vector of the subproblem (SP2). Then, x
+
ij
= 0 for all i if and only if y
+
jk
= 0
for all k.
Proof. If y
+
jk
= 0 for all k, it is obvious that x
+
ij
= 0 for all i. Only if x
+
ij
= 0 for all i, then y
+
jk
= 0 for all k due to minimization of
Eq. (17). h
By Property 1, the subproblem (SP2
/
) can be transformed into the subproblem (SP2
/
) as follows:
(SP2
/
)
h
+
SP2
= minimize
X
j
V
j
y
j;
~
k
j
; (18)
subject to
X
j
B
j
y
j;
~
k
j
P
X
i
l
i
; (19)
y
j;
~
k
j
0; 1; \j;
where
~
k
j
= arg min
k
V
k
j
n o
; V
j
= min
k
V
k
j
, and
V
k
j
= min
P
i
l
i
x
ij
6 B
j
;
x
ij
0; 1; \i;
c
jk
P0; \j; k
(f
jk
c
jk
)
X
i
(vl
i
p
jk
vl
i
c
ij
k
i
)x
ij
2A
j
h
j
v
q
X
i
l
i
x
ij
q
z
a
h
j
l
jk
q
X
i
v
i
x
ij
q
( )
:
Fig. 2. Flow chart for the improvement phase of the Lagrangian heuristic.
6 S. Park et al. / Transportation Research Part E xxx (2010) xxxxxx
ARTICLE IN PRESS
Please cite this article in press as: Park, S., et al. A three-level supply chain network design model with risk-pooling and lead times. Trans-
port. Res. Part E (2010), doi:10.1016/j.tre.2009.12.004
Note that Eq. (19) is a valid inequality. Note also that the subproblem (SP2
/
) can be solved by a dynamic programming
algorithm, referring to Nemhauser and Wolsey (1988), since it can be transformed into a linear binary knapsack problem
by using the new variables y
j
(= 1 y
j;
~
k
j
) for all js. However, it is not trivial to obtain V
k
j
s in solving the subproblem
(SP2
/
), because a nonlinear knapsack problem with a non-separable objective function needs to be solved.
3.2.2. Obtaining V
k
j
s
Without loss of generality, consider the following nonlinear knapsack problem;
(SSP2x)
V
+
= maximize f (x) =
X
i
a
i
x
i
p
1
X
i
l
i
x
i
q
p
2
X
i
v
i
x
i
q
; (20)
subject to
X
i
l
i
x
i
6 B; (21)
x
i
0; 1; \i;
where l
i
s, v
i
s and B are assumed to be integers. Note that V
k
j
can be obtained as V
k
j
= (f
jk
c
jk
) V
+
, at a
i
= (vl
i
p
jk
vl
i
c
ij
k
i
); p
1
=
2A
j
h
j
v
p
, and p
2
= z
a
h
j
l
jk
p
, respectively.
The subproblem (SSP2x) is obviously NP-hard since it is a nonlinear extension of a linear binary knapsack problem, which
is well known as NP-hard. This paper derives two kinds of solution methods for solving the subproblem (SSP2x), including an
exact solution algorithm and an efcient approximate method which is useful for large-sized problems.
Some solution properties of the subproblem (SSP2x) are also characterized.
Table 2
Data generation.
Mean customer demands l
i
~ l
0
U(1, 5)
Customer demand variances v
i
~ v
0
U(1, 5)
DC capacities B
=
j
b
0
D=m 1 U(0:15; 0:15)
a
Fixed DC assignment costs f
jk
= f
0
c B
j
1 U(q
c
; q)
b,c
Fixed supplier setup costs
g
k
= 2
f 1 U(q; q)
d
Fixed ordering setup costs A
j
= 150 c l 1 U(q; q)
e
Per unit holding costs h
j
~ h
0
{1 + U(q, q)}
Order lead times l
jk
~ 15 U(1, 5)
a
D: sum of the mean customer demands.
b
c: the average value of the per unit outbound network transportations cost.
c
q: percentage of the cost variation.
d
f : the average value of f
jk
s for all j and k.
e
l: the average value of the mean customer demands.
Table 3
Data parameter conguration for the experiment cases.
Cases l
0
v
0
b
0
f
0
h
0
q z
a
1. Base case 10 3 2 300 10 0.15 1.645
2. Loose DC capacity 10 3 4 300 10 0.15 1.645
3. High xed facility setup costs 10 3 2 500 10 0.15 1.645
4. High holding costs 10 3 2 300 20 0.15 1.645
5. High cost variation 10 3 2 300 10 0.30 1.645
6. High mean customer demands 20 3 2 300 10 0.15 1.645
7. Highly variable customer demands 10 6 2 300 10 0.15 1.645
8. Low service level 10 3 2 300 10 0.15 0
9. High service level 10 3 2 300 10 0.15 2.325
Table 1
Set of the selected combinations for the problem instances in the numerical experiment.
# of suppliers (l) # of DCs (m) # of retailers (n) # of variables # of constraints
5 10 20 255 545
7 10 20 277 587
7 15 20 412 867
7 15 40 712 1487
10 15 20 460 960
10 15 40 760 1580
10 20 40 1010 2090
10 20 60 1410 2910
S. Park et al. / Transportation Research Part E xxx (2010) xxxxxx 7
ARTICLE IN PRESS
Please cite this article in press as: Park, S., et al. A three-level supply chain network design model with risk-pooling and lead times. Trans-
port. Res. Part E (2010), doi:10.1016/j.tre.2009.12.004
Property 2. If a
i
6 0, then f(x) is a non-increasing function of x
i
. If a
i
> 0, then it is a strictly convex function of x
i
.
Proof. The rst and second derivatives of f(x) are @f (x) @x
i
= = a
i
p
1
l
i
2
P
i
l
i
x
i
p
1=2
p
2
q
i
2
P
i
v
i
x
i
p
1=2
. .
and
@
2
f (x) @x
2
i
= (pl=4)
P
i
l
i
x
i
3=2
(pq
i
=4)
P
i
v
i
x
i
3=2
, respectively.
Since @f(x)/@x
i
6 0 if a
i
6 0, f(x) is a non-increasing function of x
i
. Otherwise, f(x) is a strictly convex function of x
i
since
@
2
f (x)=@x
2
i
> 0. h
Property 2 implies that there is an optimal solution for the subproblem (SSP2x) at one of the extreme points of a feasible
solution set, which means that x
+
i
0; 1 if a
i
> 0, but x
+
i
= 0 if a
i
6 0. Moreover, due to Property 2, the size of the subproblem
(SSP2x) can be reduced because it is sufcient to consider the variables with a
i
> 0. The size of the subproblem (SSP2x) can
also be reduced by Property 3.
Property 3. Let the problem (SSP2xU) be dened as V
0
= max
x0;1
n
f (x). Note that the problem (SSP2xU) is an unconstrained version
of the subproblem (SSP2x). Denoted by x
0
and x
*
be the optimal solution vectors of the problems (SSP2xU) and (SSP2x), respectively.
Then, x
+
i
= 0 if x
0
i
= 0.
Proof. Denoted by f
S
the objective function value for an arbitrary solution set S ( #{1, 2, . . . , n}), and by S
0
and S
*
the optimal
solution sets for the problems (SSP2xU) and (SSP2x), respectively. Suppose that S
*
= S
1
+ {j} where S
1
S
0
and j R S
0
. Then,
f
S
1 f
S
+
= f
S
1 f
S
1
j
= a
j
p
1
X
iS
1
l
i
l
j
q
p
2
X
iS
1
v
i
v
j
q
p
1
X
iS
1
l
i
q
p
2
X
iS
1
v
i
q
(22)
and
f
S
0 f
S
0
j
= a
j
p
1
X
iS
0
l
i
l
j
q
p
2
X
iS
0
v
i
v
j
q
p
1
X
iS
0
l
i
q
p
2
X
iS
0
v
i
q
: (23)
Therefore, f
S
1 f
S
+
> f
S
0 f
S
0
j
> 0, since
P
iS
1 l
i
<
P
iS
0 l
i
;
P
iS
1 v
i
<
P
iS
0 v
i
and the square-root function is concave and
increasing. h
Table 4
Performance of the proposed algorithm (Case 1 base case).
l m n T
0
(min) D
0 a
T
/
(s) D
/ b
X
LB
c
X
UB
d
5 10 20 Avg 4.9 1.81 Avg 2.5 12.91 Avg 8.51 1.45
Min 2.1 0.18 Min 1.0 9.44 Min 5.31 0.42
Max 6.5 4.58 Max 5.0 16.13 Max 12.98 5.12
7 10 20 Avg 5.1 1.82 Avg 4.0 11.32 Avg 7.06 1.56
Min 2.5 0.27 Min 2.0 4.62 Min 3.71 0.82
Max 8.6 4.41 Max 7.0 14.75 Max 10.77 6.34
7 15 20 Avg 6.4 2.63 Avg 8.1 16.92 Avg 11.35 0.96
Min 3.2 0.23 Min 5.0 10.83 Min 7.06 1.38
Max 11.9 5.71 Max 10.0 20.04 Max 13.68 2.47
7 15 40 Avg 71.4 3.98 Avg 44.9 14.48 Avg 8.05 1.21
Min 40.0 0.52 Min 29.0 8.63 Min 5.95 0.25
Max 105.9 8.57 Max 65.0 20.83 Max 11.23 2.24
10 15 20 Avg 10.4 2.73 Avg 13.5 19.29 Avg 11.91 2.24
Min 3.2 0.08 Min 9.0 12.59 Min 8.42 1.21
Max 24.7 6.61 Max 17.0 25.61 Max 17.05 9.52
10 15 40 Avg 78.0 5.70 Avg 64.0 15.96 Avg 7.66 1.31
Min 50.9 3.21 Min 35.0 11.90 Min 5.98 1.15
Max 108.2 8.80 Max 87.0 18.92 Max 9.34 4.18
10 20 40 Avg 66.3 4.39 Avg 95.7 18.35 Avg 9.42 2.12
Min 41.9 1.48 Min 74.0 12.77 Min 7.49 1.90
Max 114.1 7.52 Max 139.0 26.41 Max 11.13 10.55
10 20 60 Avg 270.3 4.81 Avg 325.8 17.37 Avg 8.21 2.35
Min 157.3 3.16 Min 246.0 12.78 Min 6.64 1.74
Max 391.9 6.21 Max 511.0 21.10 Max 9.57 5.57
a
D
0
= (UB
0
LB
0
)=LB
0
100(%).
b
D
/
= (UB
/
LB
/
)=LB
/
100(%).
c
X
LB
= (LB
0
LB
/
)=LB
0
100(%).
d
X
UB
= (UB
/
UB
0
)=UB
0
100(%).
8 S. Park et al. / Transportation Research Part E xxx (2010) xxxxxx
ARTICLE IN PRESS
Please cite this article in press as: Park, S., et al. A three-level supply chain network design model with risk-pooling and lead times. Trans-
port. Res. Part E (2010), doi:10.1016/j.tre.2009.12.004
Property 3 implies that it is sufcient to consider only the set of variables which belong to the optimal solution set for the
unconstrained problem (SSP2xU). Note that the optimal solution set for the problem (SSP2xU) can be obtained in polynomial
time by use of the algorithm of Shen et al. (2003).
Now, two solution algorithm alternatives for solving the subproblem (SSP2x) are derived. A dynamic programming algo-
rithm, say Alg0_SSP2x, to solve the subproblem (SSP2x) optimally s rst derived as follows;
Algorithm. Alg0_SSP2x
Step 1. Initialize f(i, s, p, q) = , for all i, s, p and q.
Step 2. for s = 0 to B
f(1,s,0,0) = 0
if s
1
Pl
1
then
f (1; s; l
1
; v
1
) = a
1
p
1
l
1
_
p
2
v
1
_
Step 3. for i = 1 to n
for s = 0 to B
for p = 0 to
P
i
l
i
for q = 0 to
P
i
v
i
if s Pl
i
, p Pl
i
and q Pv
i
if a
i
p
1
p
_
p
2
q
_
p
1
p l
i
_
p
2
q v
i
_
f (i 1; s l
i
; p l
i
; q v
i
) > f (i 1; s; p; q) then
f (i; s; p; q) = a
i
p
1
p
_
p
2
q
_
p
1
p l
i
_
p
2
q v
i
_
f (i 1; s l
i
; p l
i
; q v
i
)
else f(i, s, p, q) = f(i 1, s, p, q)
else f(i, s, p, q) = f(i-1, s, p, q)
end for q
end for p
end for s
end for i
Table 5
Performance of the proposed algorithm (Case 2 loose DC capacities).
l m n T
0
(min) D
0 a
T
/
(s) D
/ b
X
c
LB X
d
UB
5 10 20 Avg 5.6 2.29 Avg 2.5 6.15 Avg 4.79 0.21
Min 1.2 0.40 Min 1.0 0.11 Min 0.29 0.95
Max 12.5 6.63 Max 5.0 13.37 Max 19.02 1.83
7 10 20 Avg 8.7 2.67 Avg 3.5 4.06 Avg 1.29 0.20
Min 3.3 1.27 Min 0.0 1.19 Min 0.42 1.37
Max 19.2 5.26 Max 6.0 7.04 Max 2.41 2.68
7 15 20 Avg 7.0 5.02 Avg 8.9 10.66 Avg 4.01 1.17
Min 4.3 1.41 Min 4.0 5.10 Min 3.03 4.24
Max 9.7 7.44 Max 14.0 14.67 Max 6.23 8.25
7 15 40 Avg 72.0 4.85 Avg 45.3 8.27 Avg 1.99 0.60
Min 33.8 2.41 Min 9.0 3.50 Min 1.13 1.67
Max 117.9 8.39 Max 77.0 15.83 Max 3.21 2.71
10 15 20 Avg 12.0 3.97 Avg 13.0 9.69 Avg 3.75 0.93
Min 5.8 0.99 Min 10.0 4.57 Min 1.84 2.14
Max 19.6 5.87 Max 17.0 15.40 Max 6.43 5.16
10 15 40 Avg 103.6 5.47 Avg 48.0 8.54 Avg 2.12 1.69
Min 79.6 0.50 Min 9.0 3.33 Min 7.00 0.99
Max 119.3 9.17 Max 72.0 14.96 Max 12.54 6.72
10 20 40 Avg 80.6 6.89 Avg 106.7 12.62 Avg 3.63 1.08
Min 55.2 4.13 Min 75.0 8.19 Min 1.99 2.42
Max 102.5 9.64 Max 138.0 14.96 Max 4.64 4.33
10 20 60 Avg 331.3 5.75 Avg 419.9 11.34 Avg 2.52 2.18
Min 199.4 2.91 Min 194.0 7.69 Min 0.64 0.34
Max 527.0 8.14 Max 584.0 16.02 Max 4.06 4.66
a
D
0
= (UB
0
LB
0
)=LB
0
100(%).
b
D
/
= (UB
/
LB
/
)=LB
/
100(%).
c
X
LB
= (LB
0
LB
/
)=LB
0
100(%).
d
X
UB
= (UB
/
UB
/
)=UB
0
100(%).
S. Park et al. / Transportation Research Part E xxx (2010) xxxxxx 9
ARTICLE IN PRESS
Please cite this article in press as: Park, S., et al. A three-level supply chain network design model with risk-pooling and lead times. Trans-
port. Res. Part E (2010), doi:10.1016/j.tre.2009.12.004
Note that the algorithm Alg0_SSP2x can solve the subproblem (SSP2x) in O nB
P
i
l
i
P
i
v
i
time. In other words, the algo-
rithm Alg0_SSP2x is a pseudo-polynomial time algorithm so that too much time may be required to solve large-sized in-
stances. Therefore, this paper derives another solution method, say Alg
/
_SSP2x, which solves the subproblem (SSP2x)
approximately, but in polynomial time even for large-sized instances.
* Alg
/
_7SSP2x
In the algorithm Alg
/
_SSP2x, an upper bound for the exact solution value of the subproblem (SSP2x), V
*
, is to be obtained in
polynomial time by pursuing the following two candidates. First, the optimal value for the problem (SSP2xU), say
b
V
1
, ob-
tained by the algorithm of Shen et al. (2003) can be used as one of the upper bound candidates, since the knapsack constraint
of the subproblem (SSP2x) is relaxed in the problem (SSP2xU). Second, an upper bound of the general supermodular knap-
sack problem, say
b
V
2
, obtained by the algorithm of Gallo and Simeone (1988) may also be used as the other candidate, since
the objective function f(x) is supermodular (Shen et al., 2003). Thereupon, the minimum value between
b
V
1
and
b
V
2
is taken as
the upper bound for V
*
. This leads to the algorithm Alg
/
_SSP2x to be done in O(n
3
log n) time.
3.3. Solution procedure for the relaxed problem (LR(k, c))
For the Lagrangian multipliers k and c (P0) given, the optimal solution of the relaxed problem (LR(k, c)) is obtained based
on the solution properties and algorithms discussed above. The detailed procedure for solving the relaxed problem (LR(k, c))
is provided as follows:
Step 1. Obtain h
+
SP1
by solving the subproblem (SP1).
Step 2. The subproblem (SP2) is solved according to the following procedure.
Step 2.1. For each j and k, calculate V
k
j
(or
b
V
k
j
) by solving the subproblem (SSP2x).
Step 2.2. For each j, choose
~
k
j
and V
j
such that
~
k
j
= arg min
k
V
k
j
(or
b
V
k
j
)
n o
and V
j
= min
k
V
k
j
(or
b
V
k
j
)
n o
.
Step 2.3. Solve the subproblem (SP2
/
), and then obtain h
+
SP2
.
Step 3. Calculate Z(k; c) = h
+
SP1
h
+
SP2
P
i
k
i
.
Table 6
Performance of the proposed algorithm (Case 3 high xed facility setup costs).
l m n T
0
(min) D
0 a
T
/
(s) D
/ b
X
c
LB X
d
UB
5 10 20 Avg 6.3 0.97 Avg 2.7 8.14 Avg 5.39 1.31
Min 3.2 0.23 Min 2.0 4.91 Min 3.64 0.47
Max 10.7 1.52 Max 4.0 11.17 Max 7.18 4.29
7 10 20 Avg 6.8 0.66 Avg 3.9 7.40 Avg 5.54 0.77
Min 2.9 0.19 Min 2.0 5.28 Min 3.80 0.09
Max 13.2 2.12 Max 6.0 12.22 Max 9.11 2.44
7 15 20 Avg 8.9 1.77 Avg 9.5 13.14 Avg 8.85 1.34
Min 5.4 0.50 Min 6.0 10.33 Min 7.37 0.89
Max 19.3 4.85 Max 13.0 16.57 Max 10.09 5.62
7 15 40 Avg 104.3 1.87 Avg 45.3 11.20 Avg 5.15 3.54
Min 66.2 0.25 Min 32.0 6.49 Min 3.80 0.52
Max 181.1 5.42 Max 55.0 17.57 Max 6.89 5.84
10 15 20 Avg 14.4 2.41 Avg 14.8 11.60 Avg 7.35 0.95
Min 4.7 0.00 Min 12.0 8.65 Min 6.07 1.80
Max 26.4 7.38 Max 20.0 18.18 Max 9.28 2.24
10 15 40 Avg 101.9 2.12 Avg 74.5 10.94 Avg 5.27 2.89
Min 64.0 0.56 Min 51.0 6.45 Min 3.24 0.50
Max 182.3 5.77 Max 125.0 18.39 Max 7.74 8.38
10 20 40 Avg 116.9 3.10 Avg 112.7 14.30 Avg 6.71 3.42
Min 47.3 1.01 Min 74.0 10.95 Min 5.86 1.22
Max 184.3 6.09 Max 158.0 18.51 Max 9.16 5.97
10 20 60 Avg 365.3 4.10 Avg 332.4 12.13 Avg 5.01 2.34
Min 277.2 1.38 Min 190.0 10.28 Min 3.37 1.06
Max 552.8 7.85 Max 501.0 15.46 Max 6.56 6.26
a
D
0
= (UB
0
LB
0
)=LB
0
100(%).
b
D
/
= (UB
/
LB
/
)=LB
/
100(%).
c
X
LB
= (LB
0
LB
/
)=LB
0
100(%).
d
X
UB
= (UB
/
UB
/
)=UB
0
100(%).
10 S. Park et al. / Transportation Research Part E xxx (2010) xxxxxx
ARTICLE IN PRESS
Please cite this article in press as: Park, S., et al. A three-level supply chain network design model with risk-pooling and lead times. Trans-
port. Res. Part E (2010), doi:10.1016/j.tre.2009.12.004
Step 1 can obviously be done in O(l) time. Step 2.1 can be done in O(lmn
2
B
2
C) time when the algorithm Alg0_SSP2x is ap-
plied for solving the subproblem (SSP2x), where B and C denote the maximum DC capacity and the sum of customer demand
variances, respectively, while only O(lmn
3
logn) time is necessary when the algorithm Alg
/
_SSP2xis applied. Steps 2.2 and 2.3
can be done in O(lm) and Om(mB nd) time, respectively, where d denotes the minimum of the mean customer demands.
When the algorithm Alg0_SSP2x is applied, the total time complexity for solving the relaxed problem (LR(k, c)) is in the order
of O(lmn
2
B
2
C), since Step 2.1 is dominant among all the steps. On the other hand, it takes O(maxlmn
3
log n; m(mB nd))
time in total to solve the relaxed problem (LR(k, c)) when the algorithm Alg
/
_SSP2x is applied.
3.4. Finding a feasible solution
Since the solution of the relaxed problem (LR(k, c)) may be infeasible to the original problem (P), a heuristic procedure to
obtain a good feasible solution to the original problem (P) is derived in this paper. The proposed heuristic is composed of two
phases construction phase and improvement phase. In the construction phase, a feasible solution is constructed based on the
solution to the relaxed problem. In the improvement phase, the initial feasible solution is improved by using a local search
technique in order to obtain a tighter upper bound. In this paper, the scheme of tabu search is employed as the local search
technique in the improvement phase.
3.4.1. Construction phase
In this phase, the solution of the relaxed problems, which is infeasible, is modied to satisfy the relaxed constraints. First,
retailers with an infeasible assignment are reassigned. The retailers assigned to one or more DCs are assigned to only one DC
incurring minimum cost increase. Next, unassigned retailers are assigned to a capacity-available DC with the minimum cost
increase. If no opened DC is available, then the retailer will be assigned to a new DC with minimum cost increase (including
the xed assignment cost). For the next step, if the solution violates the DC capacity constraints (especially when the algo-
rithm Alg
/
_SSP2x is applied), the DC having the largest excess demand is selected and the retailer with the largest demand is
reassigned to the other available DC with the minimum cost increase, which is repeated until all the DC capacity constraints
Table 7
Performance of the proposed algorithm (Case 4 high holding costs).
l m n T
0
(min) D
0 a
T
/
(s) D
/ b
X
c
LB X
d
UB
5 10 20 Avg 6.5 1.64 Avg 3.3 12.53 Avg 8.82 0.91
Min 2.1 0.03 Min 1.0 7.01 Min 5.95 0.48
Max 11.0 6.60 Max 7.0 19.53 Max 11.99 3.69
7 10 20 Avg 7.8 1.92 Avg 4.5 10.50 Avg 7.08 0.70
Min 2.2 0.01 Min 3.0 6.01 Min 4.31 0.16
Max 11.8 4.34 Max 6.0 17.41 Max 9.75 1.65
7 15 20 Avg 8.5 2.69 Avg 8.4 23.08 Avg 15.18 1.57
Min 3.9 0.41 Min 6.0 16.17 Min 12.58 2.09
Max 13.9 6.09 Max 11.0 32.06 Max 19.27 5.54
7 15 40 Avg 86.2 2.63 Avg 43.3 17.41 Avg 8.65 4.48
Min 31.3 0.79 Min 18.0 9.76 Min 5.45 0.00
Max 123.6 4.64 Max 68.0 22.43 Max 11.84 11.51
10 15 20 Avg 12.5 2.07 Avg 13.7 19.34 Avg 13.33 1.30
Min 8.0 0.16 Min 8.0 12.85 Min 10.25 2.10
Max 16.5 5.35 Max 19.0 26.13 Max 16.17 5.92
10 15 40 Avg 111.2 4.44 Avg 65.5 16.63 Avg 9.15 1.47
Min 56.6 0.52 Min 47.0 12.35 Min 7.76 1.02
Max 193.9 9.57 Max 88.0 19.67 Max 11.85 5.19
10 20 40 Avg 127.6 2.44 Avg 91.9 20.89 Avg 11.51 4.42
Min 81.5 0.18 Min 58.0 13.63 Min 7.71 0.35
Max 178.2 5.92 Max 117.0 28.49 Max 15.80 7.49
10 20 60 Avg 391.3 4.93 Avg 282.9 17.90 Avg 9.65 1.48
Min 194.7 1.65 Min 225.0 12.96 Min 7.13 0.87
Max 595.9 9.60 Max 431.0 24.71 Max 12.13 5.45
a
D
0
= (UB
0
LB
0
)=LB
0
100(%).
b
D
/
= (UB
/
LB
/
)=LB
/
100(%).
c
X
LB
= (LB
0
LB
/
)=LB
0
100(%).
d
X
UB
= (UB
/
UB
/
)=UB
0
100(%).
S. Park et al. / Transportation Research Part E xxx (2010) xxxxxx 11
ARTICLE IN PRESS
Please cite this article in press as: Park, S., et al. A three-level supply chain network design model with risk-pooling and lead times. Trans-
port. Res. Part E (2010), doi:10.1016/j.tre.2009.12.004
are satised. Once such reassignments of the infeasibly assigned retailers are nished, the synchronization between y- and z-
variables is made so as to satisfy Eq. (11) for the next step. In other words, since the relaxed problem (LR(k, c)) may end up
with the solution where some DCs may be assigned to suppliers which are not to be opened, such DCs will be reassigned to
an opened supplier with the minimum cost increase.
3.4.2. The improvement phase by tabu search
The objective of the improvement phase is to obtain a tighter upper bound by searching the solution area as widely as
possible. The local search technique used in this paper is based on the scheme of tabu search, referring to Glover (1989)
and Osman and Kelly (1996), which have been extensively applied in many complex facility location problems for nding
a good feasible solutions (Cortinhal and Captivo, 2003; Correia and Captivo, 2006; Sun, 2006; Prins et al., 2007). The scheme
of tabu search applied in the improvement phase and the related parameter settings are described in Appendix A.
Since the improvement phase is more time-consuming than the construction phase, the improvement phase is performed
only when both the lower bound and the upper bound in the current Lagrangian iteration are different from those in the
previous iteration. The detailed procedure of the improvement phase is described in a ow chart as in Fig. 2.
3.5. Updating the Lagrangian multipliers
To nd a good lower bound on the optimal value of Problem (P), the Lagrangian multipliers are updated by a subgradient
optimization method, as described in Fisher (1981). At iteration t, subgradient vector x
t
= (n
t
; f
t
) is determined by
n
t
i
=
P
j
~x
ij
1 and f
t
jk
= ~ y
jk
~
z
k
, where (~x; ~ y;
~
z) is the optimal solution vector of the relaxed problem (LR(k, c)) at that iteration.
The Lagrangian multipliers for the next iteration are generated by k
t1
i
= k
t
i
s
t
n
t
i
and c
t1
jk
= c
t
jk
s
t
f
t
jk
h i
, where s
t
is a posi-
tive scalar step size calculated as s
t
= a
t
UB Z(k
t
; c
t
)=|x
t
|
2
where a
t
is a positive scalar (0 < a
t
< 2),UB is an upper bound
on the optimal value of (P), and | | denotes the norm of a vector. The detailed description of the overall procedure of the
proposed Lagrangian relaxation solution approach is provided in Appendix B.
Table 8
Performance of the proposed algorithm (Case 5 high cost variation).
l m n T
0
(min) D
0 a
T
/
(s) D
/ b
X
c
LB X
d
UB
5 10 20 Avg 5.4 2.32 Avg 2.7 12.70 Avg 7.96 1.32
Min 1.7 0.01 Min 1.0 7.64 Min 1.52 0.78
Max 9.1 5.17 Max 4.0 17.63 Max 12.33 5.01
7 10 20 Avg 7.4 2.62 Avg 4.0 14.41 Avg 8.53 1.95
Min 3.1 0.54 Min 2.0 9.12 Min 5.10 1.36
Max 11.1 8.33 Max 6.0 20.71 Max 11.76 5.67
7 15 20 Avg 7.4 2.96 Avg 8.1 18.16 Avg 12.14 0.93
Min 5.0 0.49 Min 6.0 9.08 Min 6.66 1.19
Max 12.3 6.46 Max 10.0 33.00 Max 19.29 6.21
7 15 40 Avg 80.1 3.54 Avg 45.6 17.32 Avg 9.04 3.06
Min 48.3 1.06 Min 11.0 12.16 Min 6.84 0.44
Max 113.4 5.23 Max 74.0 28.05 Max 11.05 13.76
10 15 20 Avg 10.8 3.03 Avg 13.4 19.23 Avg 12.63 1.08
Min 4.2 0.03 Min 9.0 13.85 Min 10.08 2.12
Max 15.8 7.13 Max 32.0 25.59 Max 15.78 4.67
10 15 40 Avg 117.4 5.11 Avg 62.6 15.20 Avg 8.10 2.35
Min 45.4 1.54 Min 16.0 2.07 Min 4.54 0.99
Max 225.9 11.02 Max 106.0 22.38 Max 9.57 4.79
10 20 40 Avg 113.3 4.34 Avg 88.5 20.93 Avg 10.78 3.41
Min 42.7 0.05 Min 16.0 16.27 Min 7.86 2.92
Max 210.9 7.48 Max 129.0 34.54 Max 14.50 11.82
10 20 60 Avg 397.4 5.65 Avg 234.6 18.29 Avg 8.51 5.51
Min 273.9 2.38 Min 46.0 0.91 Min 1.72 1.07
Max 543.7 10.85 Max 445.0 26.39 Max 11.37 11.10
a
D
0
= (UB
0
LB
0
)=LB
0
100(%).
b
D
/
= (UB
/
LB
/
)=LB
/
100(%).
c
X
LB
= (LB
0
LB
/
)=LB
0
100(%).
d
X
UB
= (UB
/
UB
/
)=UB
0
100(%).
12 S. Park et al. / Transportation Research Part E xxx (2010) xxxxxx
ARTICLE IN PRESS
Please cite this article in press as: Park, S., et al. A three-level supply chain network design model with risk-pooling and lead times. Trans-
port. Res. Part E (2010), doi:10.1016/j.tre.2009.12.004
4. Numerical experiments
This section conducts the numerical experiments for verifying the performance of the proposed Lagrangian relaxation
solution algorithm. To evaluate the performance of the proposed Lagrangian relaxation solution algorithm, the experiment
is conducted with 1440 randomly generated problem instances for 9 different experimental cases according to the different
data parameter congurations. In each case, 160 random instances are tested in total, and 20 instances for each set of
selected combinations of the number of suppliers (l), the number of facilities (m), and the number of retailers (n), respec-
tively. The selected combinations for the experiment are given in Table 1.
All required data including locations, costs, and capacities are randomly generated as follows: The locations of the sup-
pliers, DCs, and retailers are uniformly distributed over the square of (0, 10], and per unit (inbound) outbound delivery costs
are calculated to be proportional to Euclidean distance in the plane. Per unit per distance transportation costs are set at the
value 0.5 for an inbound network, and the value 1 for an outbound network, respectively. The rest of the data are generated
in the way as shown in Table 2, and data parameter setting for each experimental case is shown in Table 3. The number of
working days per year (v) set at the value 250. Note that the data are generated to reect the situation where the inventory
control decision may affect the network design decision. Especially, the costs are generated as referred to the cost ratio infor-
mation in US industries as stated in Nozick and Turnquist (2001).
In the experiment, two alternatives of Lagrangian relaxation procedures discussed in Section 4 are tested and compared
with each other; one concerned with the algorithm Alg0_SSP2x being used for solving the subproblem (SSP2x) optimally
(say, LR
0
) and the other concerned with the algorithm Alg
/
_SSP2x being used to nd the upper bound of the subproblem
(SSP2x) (say, LR
/
), respectively. The parameters for the solution approaches are set as follows: The Lagrangian multipliers
are initially set at k
0
= 40,000 and c
0
= 9000. For the approach LR
0
, a
t
is initially set at the value 1.5 (a
0
= 1.5) and halved
if there is no improvement for 10 iterations for fast convergence rate. For the approach LR
/
, a
t
is initially set at the value
1.9 and halved for every 100 iterations without the improvement in the feasible solution. T
MAX
is set at the values 500
and 800 for the approaches LR
0
and LR
/
, respectively. Finally, e
1
and e
2
are set at the values 0.0001 and 10, respectively,
for the both approaches. The proposed solution algorithm are coded in C language and all tests are performed in a worksta-
tion running Windows XP Professional Edition with Core 2 Duo 2.13 GHz CPU and 2.0 G Byte RAM.
Table 9
Performance of the proposed algorithm (Case 6 high mean customer demands).
l m n T
0
(min) D
0 a
T
/
(s) D
/ b
X
c
LB X
d
UB
5 10 20 Avg 4.2 2.17 Avg 2.4 8.02 Avg 5.27 0.14
Min 1.8 0.24 Min 1.0 5.46 Min 3.29 2.04
Max 7.9 8.43 Max 4.0 14.84 Max 8.14 1.55
7 10 20 Avg 5.1 1.76 Avg 4.3 6.94 Avg 4.85 0.12
Min 1.7 0.03 Min 3.0 4.65 Min 3.02 2.13
Max 7.7 4.47 Max 5.0 10.02 Max 6.51 0.74
7 15 20 Avg 5.3 3.45 Avg 7.9 15.85 Avg 8.92 2.02
Min 3.2 0.43 Min 5.0 13.97 Min 6.63 2.43
Max 7.4 7.49 Max 10.0 19.44 Max 10.88 5.38
7 15 40 Avg 96.6 3.29 Avg 46.2 11.44 Avg 5.13 2.34
Min 29.3 1.62 Min 9.0 6.40 Min 0.02 0.88
Max 331.3 5.26 Max 72.0 17.07 Max 7.50 9.77
10 15 20 Avg 7.4 2.99 Avg 12.8 16.11 Avg 10.14 1.26
Min 2.7 1.81 Min 10.0 10.63 Min 5.94 0.89
Max 25.7 4.82 Max 18.0 20.74 Max 14.26 4.30
10 15 40 Avg 122.4 4.68 Avg 70.3 13.32 Avg 5.23 2.54
Min 64.3 2.05 Min 18.0 9.00 Min 0.94 0.52
Max 190.6 7.63 Max 112.0 18.41 Max 9.27 5.84
10 20 40 Avg 114.3 4.20 Avg 93.7 16.93 Avg 8.57 2.53
Min 31.3 1.30 Min 14.0 8.75 Min 4.04 1.40
Max 203.0 10.11 Max 146.0 25.36 Max 12.57 7.35
10 20 60 Avg 367.1 6.60 Avg 334.5 17.16 Avg 7.28 1.90
Min 83.9 4.14 Min 219.0 14.04 Min 5.80 0.35
Max 761.1 8.92 Max 470.0 21.40 Max 10.37 4.62
a
D
0
= (UB
0
LB
0
)=LB
0
100(%).
b
D
/
= (UB
/
LB
/
)=LB
/
100(%).
c
X
LB
= (LB
0
LB
/
)=LB
0
100(%).
d
X
UB
= (UB
/
UB
/
)=UB
0
100(%).
S. Park et al. / Transportation Research Part E xxx (2010) xxxxxx 13
ARTICLE IN PRESS
Please cite this article in press as: Park, S., et al. A three-level supply chain network design model with risk-pooling and lead times. Trans-
port. Res. Part E (2010), doi:10.1016/j.tre.2009.12.004
Tables 412 show the performance of the approaches LR
0
and LR
/
for each experimental case dened as shown in Table 2.
In Tables 412, the rst three columns indicate the problem sets classied by the problem size (the number of suppliers,
DCs and retailers). T
0
and D
0
indicate the average CPU times (min) and the average dual gaps by the approach LR
0
, respec-
tively. Similarly, T
/
and D
/
indicate the average CPU times (s) and the average dual gaps by the approach LR
/
, respectively. The
rightmost two columns marked X
LB
and X
UB
indicate the average percentage differences of the best lower and upper bounds
between the approaches LR
0
and LR
/
, respectively.
The efciency of the proposed algorithm is discussed. The CPU time by the approach LR
0
over all instances is about
69.8 min in average, while that by the approach LR
/
is only 68.8 s (about 1.15 min) in average, which is much smaller than
the approach LR
0
. This is because the pseudo-polynomial time algorithm, Alg0_SSP2x, is used in the approach LR
0
, while the
polynomial time algorithm, Alg
/
_SSP2x is used in the approach LR
/
.
As a matter of fact, general-purpose nonlinear optimization softwares including LINGO solver and MATLAB could be tried
to solve the proposed problem. However, they can only solve very small-sized problems but not large-sized problems in rea-
sonable time. For example, LINGO solver could not solve even the problem instance (4, 7, 10) within 7 h (420 min), which is
much smaller than the smallest problem instance, (5, 10, 20), tested in the experiment. From this aspect, it can be said that
the proposed solution algorithm is efcient.
It is observed that the CPU times by the both approaches tend to increase with problem sizes in all the cases. Especially,
the CPU times are much dependent on the number of DCs (m) and retailers (n), which is due to the fact that the number of
subproblems (SSP2x) to be solved increases with the number of DCs (m) and that the size of the subproblem (SSP2x) in-
creases with the number of retailers (n).
Some observations are also made on trends in the CPU time according to different experimental cases (see Fig. 3). When
DC capacities are loose (Case 2; see Table 5), the CPU times of the approaches LR
0
and LR
/
tend to increase, since the time
complexity of the subproblem (SSP2x) being dependent on total DC capacity increases in the approach LR
0
, and the longer
CPU time is required due to the increase in the xed DC assignment cost being proportional to DC capacities. When the xed
facility setup costs (for suppliers and DCs) are high (Case 3; see Table 6), the CPU times for the approaches LR
0
and LR
/
get
greater, due to the increasing scale of the cost so as to take more time to converge to the termination level xed at the value
10. A similar trend in the CPU time of the approach LR
0
is also found when the inventory cost is relatively high (Cases 4 and 6;
Table 10
Performance of the proposed algorithm (Case 7 highly variable customer demands).
l m n T
0
(min) D
0 a
T
/
(s) D
/ b
X
c
LB X
d
UB
5 10 20 Avg 3.7 2.25 Avg 2.9 10.65 Avg 7.25 0.35
Min 2.4 0.12 Min 2.0 6.00 Min 5.08 1.06
Max 6.3 8.35 Max 4.0 16.22 Max 11.61 4.10
7 10 20 Avg 6.1 3.59 Avg 4.4 11.11 Avg 6.51 0.28
Min 4.1 0.42 Min 3.0 7.94 Min 4.18 1.46
Max 11.5 8.46 Max 6.0 15.42 Max 9.31 3.45
7 15 20 Avg 6.2 3.52 Avg 7.8 15.94 Avg 10.26 0.47
Min 3.5 0.10 Min 6.0 9.44 Min 7.83 2.07
Max 12.9 6.74 Max 12.0 21.21 Max 12.61 5.30
7 15 40 Avg 56.5 3.63 Avg 45.1 13.35 Avg 7.04 1.63
Min 39.0 0.80 Min 28.0 7.21 Min 4.87 1.34
Max 80.9 7.09 Max 67.0 18.84 Max 8.79 3.37
10 15 20 Avg 7.8 2.10 Avg 11.9 17.25 Avg 11.07 2.00
Min 6.2 0.58 Min 8.0 8.30 Min 5.13 0.48
Max 10.8 4.62 Max 15.0 26.02 Max 15.79 6.53
10 15 40 Avg 79.0 4.72 Avg 68.5 16.11 Avg 7.62 2.43
Min 35.4 2.76 Min 42.0 9.62 Min 4.99 1.31
Max 126.2 9.15 Max 94.0 21.61 Max 10.14 6.95
10 20 40 Avg 79.3 3.86 Avg 89.0 19.15 Avg 10.19 3.04
Min 46.4 0.60 Min 56.0 14.35 Min 8.19 0.36
Max 173.2 8.59 Max 117.0 24.21 Max 14.53 5.86
10 20 60 Avg 236.4 6.72 Avg 300.7 17.72 Avg 7.32 2.28
Min 97.1 1.88 Min 101.0 14.60 Min 1.16 1.98
Max 365.4 10.75 Max 484.0 20.81 Max 9.63 13.79
a
D
0
= (UB
0
LB
0
)=LB
0
100(%).
b
D
/
= (UB
/
LB
/
)=LB
/
100(%).
c
X
LB
= (LB
0
LB
/
)=LB
0
100(%).
d
X
UB
= (UB
/
UB
/
)=UB
0
100(%).
14 S. Park et al. / Transportation Research Part E xxx (2010) xxxxxx
ARTICLE IN PRESS
Please cite this article in press as: Park, S., et al. A three-level supply chain network design model with risk-pooling and lead times. Trans-
port. Res. Part E (2010), doi:10.1016/j.tre.2009.12.004
see Tables 7 and 9), because the magnitude of the negative solution values of the subproblem (SSP2x) gets greater and then
the corresponding lower bound gets weaker. In a situation with high cost variations (Case 5; see Table 8), the CPU time also
Fig. 4. Average dual gaps for the experimental cases.
Fig. 3. Average CPU times by the proposed solution algorithm alternatives (a) LR
0
; (b) LR
/
.
S. Park et al. / Transportation Research Part E xxx (2010) xxxxxx 15
ARTICLE IN PRESS
Please cite this article in press as: Park, S., et al. A three-level supply chain network design model with risk-pooling and lead times. Trans-
port. Res. Part E (2010), doi:10.1016/j.tre.2009.12.004
tends to become greater, since the problem instances get more difcult due to high variation of cost parameters. Under the
situation with low service level (50%), the CPU time for the approach LR
0
is signicantly small, because a square-root term
indicating the sum of the customer demand variances is eliminated in the objective function of the subproblem (SSP2x) and
its solution algorithm has much less time complexity.
The quality of the solutions by the proposed solution approach is now discussed. The dual gaps for the approach LR
0
are
very small (3.46% on the average) and the approach LR
0
provides the exact optimal solution in some instances. On the other
hand, the dual gaps for the approach LR
/
are somewhat larger than the approach LR
0
(14.28% on the average). This is due to
the weaker lower bound for the approach LR
/
(see the columns, marked X
LB
, in the tables), which results from the fact that
the bound value for the subproblem (SSP2x) is only used in the approach LR
/
instead of solving it exactly.
While the approach LR
0
tends to provide the better feasible solutions than the approach LR
/
, the difference between the
best feasible solution values of the both approaches is quite small (1.72% on the average). Note that the approach LR
/
some-
times gives better feasible solutions than the approach LR
0
(see the columns, marked X
UB
, in the tables). The reason is that
the solution of the relaxed problem (LR(k, c)) in the approach LR
/
may violate more constraints (especially DC capacity con-
straints) than that in the approach LR
0
, so that the feasible solution search procedure in the approach LR
/
can begin with a
wider solution area than in the approach LR
0
.
Some observations are made on trends in the dual gaps. First, the average dual gaps in the approach LR
0
tend to increase
slightly with the size of the problem, which may be due to the number of relaxed constraints being increased with the prob-
lem size. On the other hand, the average dual gaps in the approach LR
/
tend to decrease as the number of retailers (n) in-
creases, which may be due to the effect of each x-variable on the objective value for the subproblem (SSP2x) getting
smaller and the difference between the bound value and the exact optimal value getting smaller.
Fig. 4 shows the average dual gaps (D
0
and D
/
) for the approaches LR
0
and LR
/
according to nine experimental cases. The
dual gap for the approaches LR
0
tends to increase slightly under loose DC capacity (Case 2), because feasible solutions under
the loose capacity have smaller number of DCs, each of which is more cost-sensitive so that changing any assigned DCs can
cause greater cost-change during the heuristic procedure. In addition, the dual gaps for the approaches LR
0
and LR
/
tend to
decrease when the xed facility setup cost is high (Case 3; see Table 6). This may be due to the portion of the xed facility
assignment cost in the total cost value getting greater and the impact of the subproblem (SSP2x) with a negative solution
Table 11
Performance of the proposed algorithm (Case 8 low service level; 50%).
l m n T
0
(min) D
0 a
T
/
(s) D
/ b
X
c
LB X
d
UB
5 10 20 Avg 3.0 2.00 Avg 2.5 8.44 Avg 5.75 0.19
Min 1.0 0.19 Min 2.0 4.91 Min 4.45 3.64
Max 5.0 5.38 Max 3.0 14.11 Max 7.24 3.05
7 10 20 Avg 4.8 1.98 Avg 3.2 10.05 Avg 6.55 0.80
Min 2.0 0.11 Min 2.0 4.33 Min 3.59 0.71
Max 8.0 4.09 Max 5.0 13.07 Max 10.19 2.88
7 15 20 Avg 5.5 1.74 Avg 7.5 16.24 Avg 11.54 0.99
Min 3.0 0.29 Min 5.0 8.51 Min 6.80 1.44
Max 10.0 3.31 Max 11.0 21.67 Max 15.18 4.60
7 15 40 Avg 37.5 3.24 Avg 38.8 13.81 Avg 7.05 2.44
Min 27.0 0.43 Min 6.0 8.29 Min 5.31 0.52
Max 54.0 6.03 Max 62.0 21.19 Max 9.81 8.44
10 15 20 Avg 9.1 3.01 Avg 10.9 18.17 Avg 12.08 0.69
Min 6.0 0.81 Min 8.0 8.85 Min 7.19 1.11
Max 14.0 7.46 Max 13.0 27.69 Max 16.19 2.92
10 15 40 Avg 50.8 5.04 Avg 60.3 14.58 Avg 7.06 1.32
Min 40.0 0.62 Min 37.0 6.37 Min 4.37 2.12
Max 68.0 9.06 Max 90.0 22.37 Max 11.03 5.60
10 20 40 Avg 50.5 5.09 Avg 89.9 19.01 Avg 9.92 2.00
Min 41.0 2.31 Min 64.0 11.79 Min 6.40 0.86
Max 64.0 9.33 Max 105.0 22.94 Max 11.76 6.19
10 20 60 Avg 134.5 5.82 Avg 273.2 18.95 Avg 8.59 2.82
Min 105.0 0.73 Min 179.0 14.35 Min 6.63 2.35
Max 174.0 10.30 Max 415.0 23.55 Max 9.89 12.45
a
D
0
= (UB
0
LB
0
)=LB
0
100(%).
b
D
/
= (UB
/
LB
/
)=LB
/
100(%).
c
X
LB
= (LB
0
LB
/
)=LB
0
100(%).
d
X
UB
= (UB
/
UB
/
)=UB
0
100(%).
16 S. Park et al. / Transportation Research Part E xxx (2010) xxxxxx
ARTICLE IN PRESS
Please cite this article in press as: Park, S., et al. A three-level supply chain network design model with risk-pooling and lead times. Trans-
port. Res. Part E (2010), doi:10.1016/j.tre.2009.12.004
value on the lower bound getting relatively smaller. Finally, the average dual gaps for the approach LR
0
are all within 5%
while there are little uctuations found in them with the different cases. This implies that the proposed solution approach
is very effective under various situations.
From the experimental results, some remarks on the proposed solution approaches can be made. First, the proposed solu-
tion approaches LR
0
and LR
/
can solve the proposed problems very efciently, which cannot be solved by any general-purpose
nonlinear optimization softwares within reasonable time. The approach LR
0
provides tight lower bounds and good feasible
solutions within reasonable time, and the approach LR
/
solves the problem very fast and also provides a good feasible solu-
tion. Moreover, the approach LR
/
could be applied to nd a good near-optimal solution for large-sized problems in business
practice within reasonable time.
5. Conclusion
This paper deals with a network design problem for a three-level supply chain consisting of multiple suppliers, distribu-
tion centers (DCs) and retailers, where risk-pooling strategy is incorporated and lead times are dependent on DC-to-supplier
pairs. In the problem, an integrated issue of network design and inventory control is addressed in consideration of DC-to-
supplier dependent lead times, and safety stock to be maintained at DCs for achieving risk-pooling benets. The proposed
problem is formulated as a nonlinear integer programming model, for which a two-phase heuristic solution algorithm based
on the Lagrangian relaxation solution approach is derived. Two efcient/effective methods are also derived for solving the
associated nonlinear subproblem during the solution approach. Experimental results show that the proposed Lagrangian
relaxation approach is both effective and efcient, even for large-sized problem instances which may not be solved by
any general-purpose nonlinear optimization softwares within reasonable time.
In the real situation, especially in global-sourcing environment where a number of suppliers/plants are located overseas,
it takes long supply/order lead times. Moreover, such lead times are dependent on DC-to-supplier pairs. However, as sur-
veyed in the literature, no one has yet considered such a real situation of dealing with the associated integrated network
design and inventory control models. In this regard, the contribution of this paper is made as in the proposed efcient solu-
tion algorithm which is derived for the integrated (simultaneous) decisions on multi-level network design and inventory
control.
Table 12
Performance of the proposed algorithm (Case 9 high service level; 99%).
l m n T
0
(min) D
0 a
T
/
(s) D
/ b
X
c
LB X
d
UB
5 10 20 Avg 3.8 1.82 Avg 2.9 10.25 Avg 6.43 1.26
Min 0.3 0.02 Min 1.0 4.85 Min 3.26 0.07
Max 7.9 5.40 Max 5.0 15.92 Max 10.04 3.05
7 10 20 Avg 5.6 1.63 Avg 4.8 10.15 Avg 6.36 1.48
Min 3.2 0.07 Min 3.0 7.56 Min 2.84 1.24
Max 8.4 4.09 Max 6.0 14.02 Max 9.30 4.29
7 15 20 Avg 5.6 1.98 Avg 7.7 17.53 Avg 11.37 2.14
Min 3.2 0.00 Min 5.0 12.35 Min 9.42 0.09
Max 7.1 4.85 Max 10.0 21.36 Max 13.02 5.55
7 15 40 Avg 56.0 4.82 Avg 43.7 13.38 Avg 6.71 0.92
Min 35.9 0.46 Min 26.0 9.92 Min 4.76 2.24
Max 102.3 7.95 Max 63.0 16.29 Max 8.51 4.77
10 15 20 Avg 9.2 1.89 Avg 12.7 17.25 Avg 11.46 1.79
Min 4.9 0.65 Min 10.0 12.05 Min 7.01 0.14
Max 13.2 3.99 Max 20.0 23.62 Max 17.57 4.48
10 15 40 Avg 72.4 4.77 Avg 71.0 15.82 Avg 7.49 2.25
Min 27.6 1.23 Min 50.0 7.59 Min 4.94 0.75
Max 134.6 8.42 Max 90.0 20.35 Max 10.97 5.90
10 20 40 Avg 77.1 5.73 Avg 102.4 18.95 Avg 10.13 1.12
Min 54.3 2.54 Min 72.0 13.14 Min 8.16 2.54
Max 110.1 10.38 Max 166.0 23.90 Max 13.63 4.01
10 20 60 Avg 246.1 4.97 Avg 345.8 17.13 Avg 8.13 2.55
Min 147.7 0.22 Min 232.0 14.46 Min 7.04 1.71
Max 409.1 9.89 Max 514.0 21.76 Max 10.88 6.00
a
D
0
= (UB
0
LB
0
)=LB
0
100(%).
b
D
/
= (UB
/
LB
/
)=LB
/
100(%).
c
X
LB
= (LB
0
LB
/
)=LB
0
100(%).
d
X
UB
= (UB
/
UB
/
)=UB
0
100(%).
S. Park et al. / Transportation Research Part E xxx (2010) xxxxxx 17
ARTICLE IN PRESS
Please cite this article in press as: Park, S., et al. A three-level supply chain network design model with risk-pooling and lead times. Trans-
port. Res. Part E (2010), doi:10.1016/j.tre.2009.12.004
The proposed problem may be immediately applied in various industrial situations including distribution network design
for products incurring high inventory holding costs such as frozen food products or having random demands such as sea-
sonal fashion products, distribution network design for new products requiring a decision on either new facility (sup-
plier/DC) investment (incurring low operating/inventory costs) or any given facility adaptation (incurring high operating
costs), and global supply chain network design for wholesalers having a number of global-sourcing suppliers/plants
overseas.
Further interesting research issues may include the development of an exact solution algorithm such as a branch-and-
bound algorithm, and the design of networks incorporating multi-echelon inventory control issues or allowing random sup-
ply along with the random demand.
Appendix A. The scheme of the tabu search in the improvement phase of the heuristic
v Solution: a set of the indices of the DC and the supplier providing services to each retailer.
v Neighborhood solution: a solution in which a certain retailer is assigned to a DC (supplier) different from that in the
current solution. If a neighborhood satises the DC capacity constraints, it is called feasible. Otherwise, it is called
infeasible. Note that a current solution has (lmn 1) neighborhoods in theory.
v Tabu: a set of the indices of a retailer, the corresponding DC and supplier of a current solution, which will be updated
when a move from the current solution to a neighborhood solution occurs.
v Aspiration condition: if the objective value of a neighborhood is better than the current best objective value, then the
move to that neighborhood from the current solution will be allowed even if the tabu corresponding to that move is in
the tabu list.
v Diversication strategy: a move to an infeasible solution is allowed only if there is no improvement in the current best
objective value for p
1
iterations. When a move to an infeasible neighborhood occurs, the objective value is calculated
as its total costs plus a penalty cost V with large amount.
v Intensication strategy: the size of the tabu list is dynamically allocated. That is, the tabu list size will decrease if the
best neighborhood solution has the better objective value than the current solution. Otherwise, it increases.
v Termination criteria: The improvement phase stops if either of the following conditions is satised: (i) if the number of
iterations exceeds over p
max
or (ii) if the best solution is not improved for p
2
iterations.
In preliminary experiments, the best performance was shown when p
1
, p
2
, and p
max
are n, 2 n, and n
2
, respectively. And
the penalty cost for an infeasible neighborhood solution, V, is set at the value 10
9
. The initial size of the tabu list is set at n/3|
and its maximum size allowed is set at n/2|.
Appendix B. The whole procedure of the Lagrangian relaxation approach
Step 1. t 0; (k
t
; c
t
) (k
0
; c
0
); LB ; UB .
Step 2. t t + 1.
For (k
t
; c
t
) given, solve the relaxed problem, (LR(k
t
; c
t
)).
LB maxLB; Z(k
t
; c
t
):
If the solution of (LR(k
t
; c
t
)) is feasible, stop the procedure.
Otherwise, nd a feasible solution by a two-phase heuristic. Let Z
t
be the objective function value of the feasible solution.
UB minUB;
Z
t
:
Step 3. Update k
t
and c
t
.
Step 4. If one of the following termination conditions are satised, stop the procedure.
(i) t PT
MAX
(ii) {(UB LB)/LB 100(%)} < e
1
(iii) s
t
< e
2
Otherwise, go to Step 2.
References
Axster, S., 1996. Using the deterministic EOQ formula in stochastic inventory control. Management Science 42 (6), 830834.
Chen, M.S., Lin, C.T., 1989. Effects of centralization on expected costs in a multi-location newsboy problem. Journal of Operations Research Society 40 (6),
597602.
Correia, I., Captivo, M.E., 2006. Bounds for the single source modular capacitated plant location problem. Computers and Operations Research 33 (10), 2991
3003.
Cortinhal, M.J., Captivo, M.E., 2003. Upper and lower bounds for the single source capacitated location problem. European Journal of Operational Research
151 (2), 333351.
18 S. Park et al. / Transportation Research Part E xxx (2010) xxxxxx
ARTICLE IN PRESS
Please cite this article in press as: Park, S., et al. A three-level supply chain network design model with risk-pooling and lead times. Trans-
port. Res. Part E (2010), doi:10.1016/j.tre.2009.12.004
Daskin, M.S., Coullard, C.R., Shen, Z.M., 2002. An inventory-location model: formulation, solution algorithm and computational results. Annals of Operations
Research 110 (1), 83106.
Elsayed, E.A., Boucher, T.O., 1994. Analysis and Control of Production Systems, second ed. Prentice-Hall, Englewood Cliffs, NJ.
Eppen, G., 1979. Effects of centralization on expected costs in a multi-location newsboy problem. Management Science 25 (5), 498501.
Eppen, G.D., Martin, R.K., 1988. Determining safety stock in the presence of stochastic lead time and demand. Management Science 34 (11), 13801390.
Erlebacher, S.J., Meller, R.D., 2000. The interaction of location and inventory in designing distribution systems. IIE Transactions 32 (2), 155166.
Federgruen, A., Zheng, Y.S., 1992. An efcient algorithm for computing an optimal (r, Q) policy in continuous review stochastic inventory systems.
Operations Research 40 (4), 808813.
Fisher, M.L., 1981. The Lagrangian relaxation method for solving integer programming problems. Management Science 27 (1), 118.
Gallo, G., Simeone, B., 1988. On the supermodular knapsack problem. Mathematical Programming 45 (2), 295309.
Glover, F., 1989. Tabu search, part 1. ORSA Journal on Computing 1, 190206.
Huang, W., Romeijn, H.E., Geunes, J., 2005. The continuous-time single-sourcing problem with capacity expansion opportunities. Naval Research Logistics
52 (3), 193211.
Karmarkar, U., 1993. Manufacturing lead times, order release and capacity loading. In: Graves, S.C., Rinnooy Kan, A.H.G., Zipkin, P. (Eds.), Handbooks in
Operations Research and Management Science, vol. 4. Elsevier, Amsterdam, pp. 288291.
Lin, L., Gen, M., Wang, X., 2009. Integrated multistage logistics network design by using hybrid evolutionary algorithm. Computers and Industrial
Engineering 56 (3), 854873.
Miranda, P.A., Garrido, R.A., 2004. Incorporating inventory control decisions into a strategic distribution network design model with stochastic demand.
Transportation Research Part E, Logistics and Transportation Review 40 (3), 183207.
Mirchandani, P.B., Francis, R.L. (Eds.), 1990. Discrete Location Theory. Wiley, New York.
Nemhauser, G.L., Wolsey, L.A. (Eds.), 1988. Integer and Combinatorial Optimization. John Willey and Sons, New York.
Nozick, L.K., Turnquist, M.A., 2001. Inventory, transportation, service quality and the location of distribution centers. European Journal of Operational
Research 129 (2), 362371.
Osman, I.H., Kelly, J.P. (Eds.), 1996. Meta-heuristics: Theory & Applications. Kluwer Academic, Boston.
Ozsen, L., Coullard, C.R., Daskin, M.S., 2008. Capacitated warehouse location model with risk pooling. Naval Research Logistics 55 (4), 295312.
Prins, C., Prodhon, C., Ruiz, A., Soriano, P., Calvo, R.W., 2007. Solving the capacitated location-routing problem by a cooperative Lagrangian relaxation-
granular tabu search heuristic. Transportation Science 41 (4), 470483.
Romeijn, H.E., Shu, J., Teo, C.P., 2007. Designing two-echelon supply networks. European Journal of Operational Research 178 (2), 449462.
Shen, Z.M., Coullard, C.R., Daskin, M.S., 2003. A joint location-inventory model. Transportation Science 37 (1), 4055.
Shu, J., Sun, J., 2006. Designing the distribution network for an integrated supply chain. Journal of Industrial and Management Optimization 2 (3), 339349.
Shu, J., Teo, C.P., Shen, Z.J.M., 2005. Stochastic transportation-inventory network design problem. Operations Research 53 (1), 4860.
Snyder, L.V., Daskin, M.S., Teo, C., 2007. The stochastic location model with risk pooling. European Journal of Operational Research 179 (3), 12211238.
Sourirajan, K., Ozsen, L., Uzsoy, R., 2007. A single-product network design model with lead time and safety stock considerations. IIE Transactions 39 (5),
411424.
Sourirajan, K., Ozsen, L., Uzsoy, R., 2009. A genetic algorithm for a single product network design model with lead time and safety stock considerations.
European Journal of Operational Research 197 (2), 599608.
Sun, M., 2006. Solving the uncapacitated facility location problem using tabu search. Computers and Operations Research 33 (9), 25632589.
Teo, C.P., Shu, J., 2004. Warehouseretailer network design problem. Operations Research 52 (3), 396408.
Teo, C.P., Ou, J., Goh, M., 2001. Impact on inventory costs with consolidation of distribution centers. IIE Transactions 33 (2), 99110.
ster, H., Keskin, B.B., entinkaya, S., 2008. Integrated warehouse location and inventory decisions in a three-tier distribution system. IIE Transactions 40
(8), 718732.
Vidyarthi, N., elebi, E., Elhedhli, S., Jewkes, E., 2007. Integrated production-inventory-distribution system design with risk pooling: model formulation and
heuristic solution. Transportation Science 41 (3), 392408.
Yang, B., Geunes, J., 2007. Inventory and lead time planning with lead-time-sensitive demand. IIE Transactions 39 (5), 439452.
You, F., Grossmann, I.E., 2008. Mixed-integer nonlinear programming models and algorithms for large-scale supply chain design with stochastic inventory
management. Industrial and Engineering Chemistry Research 47 (20), 78027817.
Zheng, Y.S., 1992. On properties of stochastic inventory systems. Management Science 38 (1), 87103.
Zheng, Y.S., Chen, F., 1992. Inventory policies with quantized ordering. Naval Research Logistics 39 (3), 285305.
S. Park et al. / Transportation Research Part E xxx (2010) xxxxxx 19
ARTICLE IN PRESS
Please cite this article in press as: Park, S., et al. A three-level supply chain network design model with risk-pooling and lead times. Trans-
port. Res. Part E (2010), doi:10.1016/j.tre.2009.12.004