Professional Documents
Culture Documents
With the above notations and denitions, the model for the capacitated plant selection problem is for-
mulated as below.
Model 1
min
y2S
xy
z
1
X
m
i1
f
i
y
i
X
m
i1
p
i
Cap
i
y
i
X
j2IS
i
d
j
a
ij
x
ij
!
1
s:t: min
x2S
xy
z
2
X
m
i1
w
i
X
j2IS
i
d
j
a
ij
x
ij
!
X
m
i1
X
j2IS
i
d
j
R
ij
x
ij
(
2
s:t:
X
i2JS
j
x
ij
1; j 1; . . . ; n; 3
X
j2IS
i
d
j
a
ij
x
ij
6Cap
i
y
i
; i 1; . . . ; m; 4
X
j2IS
i
x
ij
6ny
i
; i 1; . . . ; m; 5
x
ij
P0; y
i
2 f0; 1g; i 1; . . . ; m; j 2 IS
i
)
: 6
100 D. Cao, M. Chen / European Journal of Operational Research 169 (2006) 97110
In the above model, S
xy
in Eqs. (1) and (2) is the feasible region determined by Eqs. (3)(6). Eq. (1) is to
minimize the cost to open the plants and the cost related to the unused capacities of the selected plants. This
is the objective of the principal rm. The rst term in Eq. (1) is the xed charge cost. The second term is the
opportunity cost for extra production capacities of the selected plants. The objective function is decided by
both y
i
and x
ij
. The selection of y
i
is the direct decision of the principal rm. The determination of x
ij
is the
results of the selected plants operations. Eq. (2) is the objective function to minimize the total operation
cost of the selected plants. In this second objective function, the rst term is the cost related to production
capacity consumptions. The second term is transportation cost for shipping components or subassemblies
to the principal rm. Eq. (3) ensures that demand for each type product be satised. Eq. (4) presents the
opened plant production capacities. Eq. (5) means that no product can be produced if a plant is not se-
lected. Eq. (6) gives the binary constraints on plant selection variables and the non-negative constraints
on production quantities. The production quantities of product j produced in plant i are approximated
by d
j
x
ij
.
Model 1 is a two-level mixed integer nonlinear programming model. The principal rm, or the leader,
controls variables y
i
to optimize its objective function z
1
at the upper level. In this paper, we assume that
the selected plants cooperate with each other to minimize the total production cost in satisfying the produc-
tion requirement set by the principal rm. In the model proposed in Sakawa and Nishizaki (1999), it was
assumed that decision makers at dierent levels can coordinate their decisions. In our model, the cooper-
ation is assumed to exist between the selected plants at the lower level only. Similar models exist in practice
and have been summarized in literature. In a warehouse planning process (Miller, 2001), for example, the
facility selection (or design) and the activity planning can be done using a hierarchical planning methodol-
ogy. The hierarchical structure identies and integrates the functional relationships between the facility
selection level and the activity planning level. The facility selection decision is made at strategic level while
the decision variables at operational level are grouped into a hierarchical order. Such decision making
structure is the same as the one presented in this paper. In the model proposed in this paper, the two dif-
ferent objective functions at dierent levels explicitly represent the eventual conict between the strategic
and operational levels.
One may notice that this model does not t into a multi-attribute decision making model (Steuer, 1987).
The objective function of this model is not based on a mutually agreed criterion with compromised interests
between the principal rm and the opened plants. In the past twenty years, various models and solution
methods have been developed to solve this type of two-level mathematical programming problems (Bard
and Falk, 1982; Bard, 1998; White and Anandalingam, 1993; Wen and Huang, 1996). In theory, two-level
optimization problems are NP hard even in the simplest linear cases (Audet et al., 1997). In this research,
we investigate and utilize specic features of the plant selection problem to derive an equivalent single level
model. The linearized version of the single level model can be solved by available optimization software as
shown in the next section and in Section 4.
3. Model transformation and linearization
In this section, we rst transform the two-level optimization model into an equivalent single level model.
The nonlinear mixed integer programming problem is then linearized to be solved by available optimization
software. We rst present the two-level to one-level model transformation.
3.1. Model transformation
In Model 1, y
i
at the upper level are binary variables and x
ij
at the lower level are continuous variables.
These types of decision variables allow us to transform the two-level problem into an equivalent single level
D. Cao, M. Chen / European Journal of Operational Research 169 (2006) 97110 101
model, if the decision making at the lower level partially cooperates with that at the upper level. Partial coop-
eration means that if there exist multiple optimal solutions at the lower level, the one favorable to the upper
level objective function will be chosen and implemented. This is a practical assumption since the production
costs associated with multiple optimal solutions are the same for the selected plants. From the study of two-
level decision making process in Bialas and Karwan (1984) and from the partial cooperation assumption, we
can transform Model 1, the two-level decision making model, into the following model.
Model 2
min
y
z
1
X
m
i1
f
i
y
i
X
m
i1
p
i
Cap
i
y
i
X
j2IS
i
d
j
a
ij
x
ij
y
!
7
s:t: y
i
2 S
xy
; 8
x
ij
y arg min
x2S
xy
z
2
X
m
i1
w
i
X
j2IS
i
d
j
a
ij
x
ij
!
X
m
i1
X
j2IS
i
d
j
R
ij
x
ij
; 9
where S
xy
is the feasible region formed by the constraints in Eqs. (3)(6).
With the partial cooperation assumption, both Model 1 and Model 2 are optimistic two-level optimiza-
tion models as discussed in Mallozzi and Morgan (1996) and in Cao and Leung (2002). Further, consider
the following single level nonlinear integer programming model.
Model 3
min
y;x;t;u;v
z
1
X
m
i1
f
i
y
i
X
m
i1
p
i
Cap
i
y
i
X
j2IS
i
d
j
a
ij
x
ij
!
10
s:t: y
i
; x
ij
2 S
xy
; 11
X
m
i1
w
i
X
j2IS
i
d
j
a
ij
x
ij
!
X
m
i1
X
j2IS
i
d
j
R
ij
x
ij
6
X
n
j1
t
j
X
m
i1
Cap
i
y
i
u
i
ny
i
v
i
; 12
t
j
d
j
a
ij
u
i
v
i
6w
i
d
j
a
ij
d
j
R
ij
; i 1; . . . ; m; j 2 IS
i
; 13
t
j
2 R
1
; v
i
60; u
i
60; i 1; . . . ; m; j 1; . . . ; n: 14
In this Model 3, t
j
, u
i
, v
i
are dual variables corresponding to Eqs. (3)(5) in Model 1 when decision variables
y
i
are xed at the upper level. The left hand side of Eq. (12) is the primal objective value of the selected
plants at the lower level. The right hand side of Eq. (12) is its dual objective value. This transformation
to Model 3 is similar to that presented in Bard (1998) in converting a bi-level optimization problem to a
nonlinear integer programming problem using KarushKuhnTucker conditions. For the single level opti-
mization problem presented in Model 3, we have the following conclusion.
Theorem 1. With the partial cooperation assumption, Model 3 is equivalent to Model 1.
Proof. With the partial cooperation assumption, the inequality constraint in Eq. (12) can be satised only
in the situation where the equality holds and when the dual gap vanishes. In this situation, x
ij
are the lower
level reaction decisions to optimize objective function z
2
when the upper level decision variables are xed by
the values of y
i
. Hence Model 3 is optimized with
x
ij
2 arg min
x2S
xy
z
2
X
m
i1
w
i
X
j2IS
i
d
j
a
ij
x
ij
!
X
m
i1
X
j2IS
i
d
j
R
ij
x
ij
:
102 D. Cao, M. Chen / European Journal of Operational Research 169 (2006) 97110
Therefore, Model 3 is equivalent to Model 2. Solving Model 3 is equivalent to solving the original two-level
problem Model 1 from the equivalence between Model 1 and Model 2. h
By Theorem 1, solving the plant selection problem in a decentralized environment is equivalent to solv-
ing the single objective nonlinear integer programming problem formulated in Model 3. The nonlinear
terms are those quadratic terms in Eq. (12). They are the products of continuous variables u
i
or v
i
with bin-
ary variables y
i
. These terms can be linearized as shown below.
3.2. Model linearization
The nonlinear terms Cap
i
y
i
u
i
+n y
i
v
i
in Eq. (12) of Model 3 can be linearized with added linear con-
straints. After these terms are linearized, the problem can easily be solved by available optimization soft-
ware. To linearize these terms we follow the procedure similar to that used in Chen (2001) and Fortuny and
McCarl (1981). Let
b
i
y
i
fCap
i
u
i
nv
i
g:
This implies that
b
i
Cap
i
u
i
nv
i
; if y
i
1;
0; if y
i
0:
Then Eq. (12) can be replaced by the following set of linear inequalities:
X
m
i1
w
i
X
j2IS
i
d
j
a
ij
x
ij
!
X
m
i1
X
j2IS
i
d
j
R
ij
x
ij
6
X
n
j1
t
j
X
m
i1
b
i
; 12:1
b
i
6Cap
i
u
i
nv
i
My
i
M; i 1; . . . ; m; 12:2
b
i
PCap
i
u
i
nv
i
; i 1; . . . ; m; 12:3
b
i
P My
i
; i 1; . . . ; m; 12:4
b
i
60; i 1; . . . ; m; 12:5
where in Eq. (12.2), M is a large positive number. Eqs. (12.2) and (12.3) enforce b
i
=Cap
i
u
i
+nv
i
when y
i
is
set to 1; Eqs. (12.4) and (12.5) make b
i
=0, when y
i
is set to 0. Eqs. (12.2) and (12.3) will have no eect on b
i
when y
i
is 0, since u
i
, v
i
6 0 and {Cap
i
u
i
+nv
i
} is always non-positive. Similarly, Eqs. (12.4) and (12.5) will
have no eect on b
i
when y
i
is 1. With the nonlinear constraint Eq. (12) replaced by Eqs. (12.1)(12.5), we
have the following approximate mixed integer programming model.
Model 4
min
y;x;t;u;v;b
z
1
X
m
i1
f
i
y
i
X
m
i1
p
i
Cap
i
y
i
X
j2IS
i
d
j
a
ij
x
ij
!
15
s:t: y
i
; x
ij
2 S
xy
; 16
X
m
i1
w
i
X
j2IS
i
d
j
a
ij
x
ij
!
X
m
i1
X
j2IS
i
d
j
R
ij
x
ij
6
X
n
j1
t
j
X
m
i1
b
i
; 17
b
i
6Cap
i
u
i
nv
i
My
i
M; i 1; . . . ; m; 18
b
i
PCap
i
u
i
nv
i
; i 1; . . . ; m; 19
b
i
P My
i
; i 1; . . . ; m; 20
t
j
d
j
a
ij
u
i
v
i
6w
i
d
j
a
ij
d
j
R
ij
; i 1; . . . ; m; j 2 IS
i
; 21
t
j
2 R
1
; b
i
; v
i
; u
i
60; i 1; . . . ; m; j 1; . . . ; n: 22
D. Cao, M. Chen / European Journal of Operational Research 169 (2006) 97110 103
The following theorem states the relationship between Model 3 and Model 4.
Theorem 2. If Model 3 has feasible solutions, then there exists a positive number M
, Model 4 has optimal solutions which are also optimal for Model 3.
Proof. It is only necessary to show that there exists a positive number M
:
b
i
Cap
i
u
i
nv
i
; if y
i
1;
0; if y
i
0:
max
i1;...;m
fCap
i
u
i
nv
i
g 23
If y
i
=0, from Eq. (20) and the non-positivity requirement on b
i
, we have b
i
=0 for any M P M
. In fact,
when y
i
=0, setting M P M
enforces b
i
to be 0 while Eqs. (18)(20) always hold. If y
i
=1, setting
M P M
, enforces b
i
to be uniquely determined by Eqs. (18)(20) while Eq. (20) is automatically satised.
That is, b
i
=Cap
i
u
i
+n v
i
, when y
i
=1. Since the original problem expressed in Model 3 is assumed to have
optimal solutions, the feasible region of decision variable pair (u
i
,v
i
) is bounded. This means that M