Você está na página 1de 9

Production, Manufacturing and Logistics

A primaldual algorithm for the economic lot-sizing problem with


multi-mode replenishment
Sandra Duni Eks ioglu
*
Department of Industrial and Systems Engineering, Mississippi State University, P.O. Box 9542, Mississippi State, MS 39762, USA
a r t i c l e i n f o
Article history:
Received 19 February 2007
Accepted 29 April 2008
Available online 7 July 2008
Keywords:
Inventory
Economic lot-sizing
Primaldual algorithm
Multi-setups costs
Network ows
a b s t r a c t
The classical economic lot-sizing problem assumes that a single supplier and a single transportation
mode are used to replenish the inventory. This paper studies an extension of this problem where several
suppliers and transportation modes are available. The decision-making process in this case involves iden-
tifying (i) the timing for an order; (ii) the choice of shipment modes; and (iii) the order size for each
mode. The problem is dened as a network ow problem with multiple setups cost function and addi-
tional side constraints. This study provides an MIP formulation for the problem. We also provide an addi-
tional formulation of the problem by redening its decision variables and show that the dual of the
corresponding LP-relaxation has a special structure. We take advantage of the structure of the dual prob-
lem to develop a primaldual algorithm that generates tight lower and upper bounds. Computational
results demonstrate the effectiveness of the algorithm.
2008 Elsevier B.V. All rights reserved.
1. Introduction
The classical economic lot-sizing problem is well studied in the
literature of inventory control, production planning and capacity
planning. In 1958, Wagner and Whitin (1958) solved the problem
efciently, but a continuing interest in the problem remains be-
cause of its practical applications. For example, economic lot-sizing
is the core problem in aggregate production planning in MRP sys-
tems (Nahmias, 1997). For an extensive review of literature that
discusses this problem, see: Aggarwal and Park (1993); Bahl
et al. (1987); Belvaux and Wolsey (1998, 1999, 2001); Nemhauser
and Wolsey (1988); and Wolsey (1995).
The classical economic lot-sizing problem is dened as follows:
Given the demand, the unit production cost, the unit inventory
holding cost for a commodity, and the set-up costs for each time
period over a nite and discrete-time horizon nd a production
schedule that satises demand at minimum costs. This model as-
sumes a xed and a variable component of production costs as well
as a deterministic time-varying demand. Wagner and Whitin
(1958) developed a dynamic programming algorithm that solves
the problem in O(T
2
), where T is the length of the planning horizon.
Almost thirty years later, Wagelmans et al. (1992); Aggarwal and
Park (1993), and Federgruen and Tzur (1991) independently stud-
ied and improved the time complexity for obtaining an optimal
solution to O(T log T) in the general case and to O(T) when the costs
have a special structure.
A number of studies have generalized the classical model with
various considerations. These extensions include nite production
capacity models (Florian and Klein, 1971; Bitran and Yanasse,
1982), multi-echelon models (Kaminsky and Simchi-Levi, 2003)
and multi-item models (Manne, 1958; Barany et al., 1984). The
general purpose of these extensions was to replicate real problems
faced by manufacturing companies. However, because of todays
changing logistics practices, the new prevailing extensions of this
problem need to be studied. This is one of the motivations for this
research.
This paper studies the economic lot-sizing problem with multi-
mode replenishment. The classical problem assumes that items
can be purchased from a single supplier and/or delivered using a
single transportation mode. This assumption holds true when a
particular supplier or transportation mode is always superior to
others or only one mode is feasible due to restrictions or require-
ments such as using tank trucks or using company-owned trucks.
However, studies have shown that a more cost-effective practice
considers the availability of several suppliers and transportation
modes when making inventory replenishment decisions.
A number of studies in the area of supply chain management
have discussed coordination of inventory and distribution deci-
sions (Burns et al., 1985; Gupta, 1992; Hahm and Yano, 1992,
1995; Eks ioglu et al., 2006, 2007; Chopra and Meindl, 2006). These
studies analyze the trade-offs between inventory and transporta-
tion costs, however, they consider that a single transportation
mode is available, and corresponding transportation costs are
either assumed to be linear with respect to distance and volume
or to have a xed charge cost structure. Researchers have also
0377-2217/$ - see front matter 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2008.04.043
* Tel.: +1 662325 9220; fax: +1 662325 7012.
E-mail address: sde47@ise.msstate.edu
European Journal of Operational Research 197 (2009) 93101
Contents lists available at ScienceDirect
European Journal of Operational Research
j our nal homepage: www. el sevi er . com/ l ocat e/ ej or
analyzed the impact of multi-mode delivery opportunities on
transportation costs. For example, Baumol and Vinod (1970) iden-
tify factors that impact the selection of a particular transportation
mode. They show that faster, more dependable transportation
modes reduce shippers or receivers inventories, including the
safety stock and in-transit. These studies propose a mathematical
model that identies the single, best transportation mode when gi-
ven the annual demand forecasts. Other studies consider transpor-
tation related problems with multi-mode delivery options;
however, they do not relate to the economic lot-sizing literature.
To our knowledge, Jaruphongsa et al. (2005) are the only ones
who have studied the economic lot-sizing problem with multi-
mode replenishment and cargo-capacity considerations. However,
they analyze and provide exact solution algorithms to the special
case of the problem when two transportation modes are available.
Our paper studies the multi-mode version of this problem and pro-
vides an efcient solution approach.
The model and the solution method we discuss can be used as a
sub-module in MRP systems for requirements planning with mul-
ti-mode replenishment options, or as a decision tool to determine
whether a particular replenishment mode, or a combination of dif-
ferent modes, should be employed to replenish the inventory. The
main contribution of this paper is developing an algorithm that, as
demonstrated by our computational results, provides good quality
solutions in a reasonable amount of time. Jaruphongsa et al. (2005)
describe an application of this model in a third-party logistics (TPL)
setting.
Initially, the problem is dened as a network ow problem with
multiple setups cost function and additional side constraints, and
formulated as a mixed-integer program (MIP) (Section 2). Next,
an additional formulation is proposed by redening the decision
variables (Section 3). The dual of the corresponding LP-relaxation
has a special structure. A primaldual algorithm has been devel-
oped that takes advantage of the structure of the dual problem
(Section 3.2). Computational results demonstrate the effectiveness
of the algorithm (Section 4). The special case of the economic lot-
sizing problem with multi-mode replenishment where the cost
function has a xed charge structure has been analyzed, and a dy-
namic programming algorithm that solves the problem in O(IT
2
)
has been developed (Section 2.1), where I is the number of replen-
ishment modes available.
2. Problem description
The economic lot-sizing problem is dened as a network ow
problem with multiple setups cost function and additional side
constraints. The structure of this network for a problem with two
replenishment modes and three time periods is presented in
Fig. 1. The network consists of a source node s that supplies the to-
tal demand and T demand nodes. The arcs connecting the source
node to the same demand node represent the different replenish-
ment options available. The demand nodes in two consecutive time
periods are connected through the inventory arcs.
Because of the assumption that no capacity limitations on stor-
age space or replenishment quantities occur, the total replenish-
ment cost in each period is the sum of shipment costs of the I
modes. Each replenishment mode is characterized by a different
cost function. For replenishment mode i (i = 1; . . . ; I), let s
i
, C
i
, A
i
and p
i
respectively denote the xed setup costs, the capacity of a
cargo container, the xed cost of a cargo container, and the unit
procurement cost. These parameters are assumed to be stationary.
The model identies: (i) the timing for an order; (ii) the choice of
shipment modes; and (iii) the order size for each mode. Character-
istics required to satisfy demand at minimum cost.
For t = 1; . . . ; T, we use the following notations:
h
t
unit inventory holding cost during period t.
b
t
demand in period t.
H
t
amount of inventory at the end of period t.
q
it
amount received using mode-i in period t.
The cost of receiving q
it
units is given by:
Q
i
(q
it
) =
s
i
y
it
A
i
q
it
C
i
_ _
p
i
q
it
if q
it
>0
0 otherwise
for i =1; . . . ; I; t =1; . . . ; T;
_

_
(1)
where y
it
= 1 if q
it
> 0, and y
it
= 0 otherwise. a| denotes the small-
est integer that is greater than or equal to a. As in Lee (1989), we
refer to the structure of the replenishment cost function Q
i
(q
it
), as
the multiple setups cost function. Each replenishment mode may
have a different but constant lead time. Thus, items received
through a particular replenishment mode may spend a constant
time as in-transit inventory and may be subject to a holding cost.
The problem can easily be transformed into a problem with zero
lead time, as follows: (a) an order that is needed at time t and
replenished using mode i (with lead time of k periods, k = 0; . . . ; t)
is made at t k. As a result, the arrival of this order will correspond
with the scheduled arrival time. (b) The unit procurement cost is in-
creased by

t1
l=tk
h
l
, which corresponds to the unit inventory hold-
ing cost for the time periods t k to t, when the item is in-transit.
Therefore, the reminder of this paper assumes that lead times are
negligible.
The following is an MIP formulation of the economic lot-sizing
problem with multi-mode replenishment:
minimize

I
i=1

T
t=1
s
i
y
it
A
i
z
it
p
i
q
it
h
t
H
t
(P)
subject to

I
i=1
q
it
H
t1
H
t
= b
t
t = 1; . . . ; T (2)
q
it
6 b
tT
y
it
i = 1; . . . ; I; t = 1; . . . ; T (3)
q
it
6 C
i
z
it
i = 1; . . . ; I; t = 1; . . . ; T (4)
H
0
= H
T
= 0 (5)
y
it
0; 1 i = 1; . . . ; I; t = 1; . . . ; T (6)
z
it
Z

i = 1; . . . ; I; t = 1; . . . ; T (7)
q
it
; H
t
P0 i = 1; . . . ; I; t = 1; . . . ; T (8)
where b
tT
presents the demand during the time periods t (t =
1; . . . ; T) to T (i.e. b
tT
=

T
s=t
b
s
). Variable z
it
is such that, z
it
=
q
it
C
i
_ _
.
S

=
3
1 t
t
b
q
11
q
21
b
1
b
2
H
1
b
3
H
2
q
12
q
22
q
13
q
23
Fig. 1. Network representation of ELS with I = 2 and T = 3.
94 S.D. Eks ioglu/ European Journal of Operational Research 197 (2009) 93101
The following lemmas provide properties of an optimal solution
to problem (P).
Lemma 1 Jaruphongsa et al. (2005). There exists an optimal solution
such that, for any t = 1; . . . ; T,

I
i=1
q
it
> 0 if and only if H
t1
< b
t
.
Lemma 2 Jaruphongsa et al. (2005). There exists an optimal solu-
tion such that if t
+
1 and s are two consecutive regeneration points
then there is at most one partial replenishment during periods t
+
through s.
Jaruphongsa et al., dene time period t (1 6 t 6 T) as a regener-
ation point, if H
t
= 0. Let q
t
(= q
1t
q
2t
q
It
) denote the
amount shipped in period t. They dene q
t
to be a full replenishment
if q
it
is a full-truck-load shipment (q
it
= mC
i
for some positive inte-
ger m) or zero for all i = 1; . . . ; I; otherwise q
t
is referred to as a par-
tial replenishment.
2.1. Fixed-charge replenishment cost structure
In the special case when A
i
= 0 for i = 1; . . . ; I, the multi-setups
cost structure given in (1) simplies to a xed setup cost structure.
This special case is important because inventory replenishment
problems with such cost structures arise often in practice. For
example, consider the problem when the inventory can be replen-
ished using more than one supplier, and each supplier charges a
different xed cost, different unit procurement costs, and has dif-
ferent delivery times. In this case
Q
i
(q
it
) =
s
i
y
it
p
i
q
it
if q
it
> 0
0 otherwise for i = 1; . . . ; I; t = 1; . . . ; T:
_
(9)
Using the network ow interpretation, we establish the re-
quired properties of optimal solutions for this special case of the
uncapacitated economic lot-sizing problem with multi-mode
replenishment.
Proposition 1. There exists an optimal solution to P such that:
v q
+
it
q
+
lt
= 0, for i; l = 1; . . . ; I, il and t = 1; 2; . . . ; T.
v q
+
it
H
+
t1
= 0, for i = 1; . . . ; I, t = 1; . . . ; T .
v H
+
t1
(

I
i=1
q
+
it
) = 0, for t = 1; . . . ; T.
Proof. The uncapacitated economic lot-sizing problem with
multi-mode replenishment minimizes a concave cost function
over a bounded convex set; therefore, its optimal solution cor-
responds to a vertex of the feasible region (Zangwill, 1968). Let
(q
+
; H
+
) be an optimal solution. In an uncapacitated network
ow problem, a vertex is represented by a tree solution. The
tree representation of the optimal solution implies that demand
in every time period will be satised by exactly one replenish-
ment mode; in other words, q
+
it
q
+
lt
= 0, for il and t = 1; 2; . . . ; T.
Furthermore, if the inventory level in a particular time period is
positive, no replenishment occurs, and vice versa. This implies
that H
t1
(

I
i=1
q
it
) = 0. Thus, zero inventory policy holds for this
problem (q
+
it
H
+
t1
= 0, for i = 1; . . . ; I, t = 1; . . . ; T). h
Proposition 1 implies properties 1 and 2 of an optimal solutions
to our problem.
Property 1. In an optimal solution, demand in period t (t = 2; . . . ; T)
is either satised from inventory or by a replenishment received in
period t.
This property of the optimal solutions is often referred to in the
literature as the Zero Inventory Property (ZIP). ZIP applies to the
classical single-item lot-sizing problem and some of its general-
izations (Wagner and Whitin, 1958).
Property 2. In an optimal solution, at most one replenishment mode
is used in a period.
Property 3. In an optimal solution, a positive replenishment is large
enough to satisfy demand for a number of consecutive time periods
(q
it
= 0 or q
it
=

s
l=t
b
l
for t 6 s 6 T).
Theorem 1. There exists a dynamic programming algorithm that
solves the economic lot-sizing problem with multi-mode replenish-
ment in O(IT
2
) when replenishment costs have a xed charge
structure.
Proof. Let P(q
t
) denote the replenishment cost in period t. Based
on Property 2, only one replenishment mode is used in a given time
period (q
t
= q
1t
; q
2t
; . . . ; q
It
). Thus, the cost of receiving a ship-
ment in period t is
P(q
t
) = minQ
1
(q
t
); Q
2
(q
t
); . . . ; Q
I
(q
t
): (10)
Let t 1 and s be two successive regeneration points
(H
t1
= H
s
= 0). Based on Property 1, there will be a single replen-
ishment during the time interval t; . . . ; s that happens in time t.
The replenishment amount q
t
= b
ts
and q
l
= 0 for l = t 1; . . . ; s.
Thus, for a given 1 6 s 6 T, such that b
s
> 0, the total cost C(t; s)
during the time interval t; . . . ; s is equal to
C(t; s) = P(b
ts
)

s1
l=t
h
l
b
l1;s
; (11)
where P(b
ts
) is the minimum cost replenishment given by Eq. (10),
and

s1
l=t
h
l
b
l1;s
calculates the cost of the inventory held to satisfy
demand b
ts
.
Let F(s) denote the minimum total cost of satisfying demand
b
1s
. Based on the properties of an optimal solution dened above
and the principle of optimality, the optimal solution can be com-
puted using the following recursive functions:
F(0) = 0;
F(s) = minF(t 1) C(t; s) : t 6 s; for s = 1; . . . ; T:
(12)
The optimal solution is F(T). For a given t and s (1 6 t 6 s 6 T) it
takes O(I) to calculate C(t; s). This is due to calculating P(b
ts
)
using Eq. (10). Thus, it takes O(IT
2
) to nd the optimal solution
F(T). h
3. Problem reformulation
This section presents an equivalent formulation for the eco-
nomic lot-sizing problem with multi-mode replenishment. LP-
relaxation of this new formulation could be used to generate lower
bounds. We show that these lower bounds are tighter than the
lower bounds from the LP-relaxation of P. In addition, a primal
dual algorithm is developed using this reformulation. The algo-
rithm gives optimal or close to optimal solutions.
The LP-relaxation of P is not very tight, because of constraints
(3) and (4). In the LP-relaxation of P, variable y
it
determines the
fraction of demand from periods t to T satised from replenish-
ment mode i in period t. b
tT
is usually a very high upper bound,
since the amount received in a period rarely equals this amount.
Variable z
it
determines the number of cargos received in period t
using mode i. Rarely, q
it
could be a multiple of C
i
.
One way to tighten formulation P is to split variables q
it
by des-
tination into variables q
its
(s = t; . . . ; T). q
its
presents the amount of
inventory replenished in period t by using mode i to satisfy de-
mand in period s. For the new variables, a trivial and tight upper
bound is the demand in period s (i.e. b
s
). This leads to the following
identities:
S.D. Eks ioglu/ European Journal of Operational Research 197 (2009) 93101 95
q
it
=

T
s=t
q
its
; (13)
H
t
=

I
i=1

t
s=1

T
s=t1
q
iss
: (14)
After replacing variables q
it
and I
t
with the new variables q
its
, for-
mulation P becomes:
minimize

I
i=1

T
t=1

T
s=t
c
its
q
its
s
i
y
it
A
i
z
it
_ _
(Q)
subject to

I
i=1

s
t=1
q
its
= b
s
s = 1; . . . T; (15)
q
its
6 b
s
y
it
i = 1; . . . ; I; 1 6 t 6 s 6 T; (16)

T
s=t
q
s
6 C
i
z
it
i = 1; . . . ; I; t = 1; . . . T; (17)
y
it
0; 1 i = 1; . . . ; I; t = 1; . . . ; T; (18)
z
it
Z

i = 1; . . . ; I; t = 1; . . . ; T; (19)
q
s
P0 i = 1; . . . ; I; 1 6 t 6 s 6 T (20)
where c
its
= p
i

s1
s=t
h
s
. In the special case when A
i
= 0, this refor-
mulation resembles the uncapacitated facility location problem.
When A
i
= 0, in an optimal solution, the following conditions
should be satised: z
it
Z

and z
it
P

T
s=t
q
its
C
i
. Since z
it
is not in the
objective function, it can be set to a very large integer. Thus, con-
straints (17) and (19) become redundant, and they can be elimi-
nated from the formulation. Now, consider that I T is the
number of facilities from which demand (inventory) can be served
(replenished). The decision to be made is whether to open or not to
open a certain facility it I T (whether to replenish the inventory
in period t using transportation mode i). The cost for opening a facil-
ity (mode i set-up cost) is denoted by s
i
. The cost of serving cus-
tomer (replenishing the inventory in period) s (s Pt) is c
its
per
unit of demand. Although this formulation resembles the facility
location problem, it is not exactly the same. In our problem, the fol-
lowing relationship exists among replenishment costs: c
its
=
c
itl

s1
k=l
h
k
for s > l. Such a relationship does not exist in the facil-
ity location problem. This fact makes ours a special case of the facil-
ity location problem.
A special case of problem P (when I = 1 and A = 0) is the classi-
cal economic lot-sizing problem. In this case, the LP-relaxation of
formulation Q has an integer solution. This result is due to Krarup
and Bilde (1977). Although this result does not hold true for cases
when I > 1 and A P0, the lower bounds generated when solving
the LP-relaxation of Q are very good, and the solutions are close
to optimal. In fact, the lower bounds found solving the LP-relaxa-
tion of Q are tighter than the lower bounds generated by solving
the LP-relaxation of P.
Proposition 2. The optimal cost generated by the LP-relaxation of
formulation Q of the economic lot-sizing problem with multi-mode
replenishment is at least as high as the optimal cost from the LP-
relaxation of formulation P.
Proof. Every feasible solution to the LP-relaxation for formulation
Q of the economic lot-sizing problem with multi-mode replenish-
ment can be transformed into a solution of LP-relaxation of for-
mulation P by using Eqs. (13) and (14). Therefore, the optimal
solution of the LP-relaxation of Q can be transformed to a feasible
solution (not necessarily the optimal solution) for the LP-relaxa-
tion of P. h
3.1. Understanding the dual problem
The dual of the LP-relaxation of Q has a special structure that al-
lows us to develop a primaldual based algorithm. The formulation
of the dual problem appears below:
maximize

T
t=1
b
t
v
t
(D)
subject to

T
s=t
b
s
w
its
6 s
i
i = 1; . . . ; I; t = 1; . . . ; T; (21)
v
s
w
its
h
it
6 c
its
i = 1; . . . ; I; 1; 6 t 6 s 6 T; (22)
C
i
h
it
6 A
i
i = 1; . . . ; I; t = 1; . . . ; T; (23)
w
its
P0 i = 1; . . . ; I; 1 6 t 6 s 6 T; (24)
h
it
P0 i = 1; . . . ; I; t = 1; . . . ; T: (25)
In an optimal solution for D, both constraints h
it
P0 and h
it
6
A
i
C
i
should be satised. h
it
is not in the objective function, and
A
i
C
i
P0.
Since this is a maximization problem, the value of v
s
will be in-
creased by setting h
it
to the maximum value that it can have,
h
it
=
A
i
C
i
. In this case, constraints (22) become v
s
w
its
6 c
its

A
i
C
i
.
This leads to the following dual formulation:
maximize

T
t=1
b
t
v
t
subject to (21) and (24)
(D
/
)
v
s
w
its
6 c
its
i = 1; . . . ; I; 1; 6 t 6 s 6 T; (26)
where c
its
= c
its

A
i
C
i
. In addition, in an optimal solution to D
/
, con-
straints w
its
P0 and w
its
Pv
s
c
its
should be satised. Since w
its
is not in the objective function, it can be replaced with w
its
= max
(0; v
s
c
its
), and leading to the following dual formulation:
maximize

T
t=1
b
t
v
t
subject to
(D
+
)

T
s=t
b
s
max(0; v
s
c
its
) 6 s
i
i = 1; . . . ; I; t = 1; . . . ; T: (27)
Suppose the LP-relaxation of Q has an optimal solution (q
+
; y
+
; z
+
)
that is integral. Let C = (i; t)[y
+
it
= 1, and let (v
+
; w
+
; h
+
) denote an
optimal dual solution. The complementary slackness conditions
for this problem appear as follows:
(C
1
)y
+
it
[s
i

T
s=t
b
s
w
+
its
[ =0 for i = 1; . . . ; I; t = 1; . . . ; T;
(C
2
)q
+
its
[c
its
v
+
s
w
+
its
[ =0 for i = 1; . . . ; I; 1 6 t 6 s 6 T;
(C
3
)z
+
it
[A
i
C
i
h
+
it
[ =0 for i = 1; . . . ; I; t = 1; . . . ; T;
(C
4
)w
+
its
[q
+
its
b
s
y
+
it
[ =0 for i = 1; . . . ; I; 1 6 t 6 s 6 T;
(C
5
)v
+
t
[b
t

t
s=1
q
+
its
[ =0 for t = 1; . . . ; T;
(C
6
)h
+
it

T
s=t
q
+
its
C
i
z
+
it
_ _
=0 for i = 1; . . . ; I; t = 1; . . . ; T:
By conditions (C
1
) and (C
3
), if a replenishment mode is used in a
particular time period, the xed set-up cost and xed cargo con-
tainer cost must be fully paid (i.e., if (i; t) C, then s
i
=

T
s=t
b
s
w
+
its
96 S.D. Eks ioglu/ European Journal of Operational Research 197 (2009) 93101
and A
i
= C
i
h
it
). Consider conditions (C
4
). Now, if replenishment
mode i is used in period t (y
it
= 1), but demand in period s (t < s)
is not satised by this replenishment, then q
+
its
= 0 and
q
+
its
b
t
y
+
it
0. In this case, w
+
its
= 0. This implies that only the peri-
ods whose demand was satised by a replenishment made in period
t using mode i will contribute to the corresponding set-up cost.
By conditions (C
2
), if q
+
its
> 0, then v
+
s
= c
its
w
+
its
. Thus, we can
think of v
+
s
as the total cost (per unit of demand) in period s. Of this
amount, c
its
pays for production and inventory holding costs;
A
i
C
i
pays for the xed cost of the cargo container; and w
+
its
contributes
to the set-up cost.
3.2. Description of the Algorithm
The simple structure of the dual problem can be exploited to
obtain near optimal feasible solutions by inspection. Initially, dual
variables v
t
= 0 for t = 1; . . . ; T. Next, we increase the value of each
dual variable maintaining the feasibility of the problem. For t = 1,
constraints (27) in formulation (D
+
) become: b
1
max(0; v
1
c
i11
) 6
s
i

T
s=2
b
s
max(0; v
s
c
i1s
) = s
i
for i = 1; . . . ; I. Note that, the dual
variable v
1
appears only on this set of constraints. Therefore, given
b
1
> 0, in order to maximize the dual problem we should assign v
1
the largest value that will satisfy these constraints. As a result,
v
1
= min
i
c
i11

S
i
b
1
.
Suppose that the optimal values of the rst k 1 dual variables
of (D
+
) are known. Then, to be feasible, the k-th dual variable (v
k
)
must satisfy the following constraints for all i = 1; . . . ; I and
t = 1; . . . ; k:
b
k
max(0; v
k
c
itk
) 6 M
it;k1
= s
i

k1
s=t
b
s
max(0; v
+
s
c
its
): (28)
When b
k
> 0, this value is
v
k
= min
i;t6k
c
itk

M
it;k1
b
k
_ _
(29)
Note that if M
it;k1
P0 implies v
k
Pc
itk
. A dual feasible solution can
be obtained simply by sequentially calculating the value of the dual
variables by using Eq. (29) (Fig. 2). A backward construction
algorithm generates primal feasible solutions (Fig. 3). The running
time of the algorithm is O(IT
2
). The primal and dual solutions gen-
erated may not necessarily satisfy the complementary slackness
conditions. Hence, in order to determine if the solution obtained
by using the primal algorithm is optimal, one can check to see if
the corresponding objective function value equals the objective
function from the dual algorithm. Since the algorithms generate
feasible solutions, at optimality, primal and dual solutions are
equal.
Proposition 3. The solutions obtained with the primal and dual
algorithms are feasible, and they always satisfy the complementary
slackness conditions (C
1
), (C
2
), (C
3
), and (C
5
).
Proof. Clearly, the primal and dual solutions generated are feasi-
ble by construction. Since the primal algorithm sets q
its
> 0 only
when c
its
v
s
w
its
= 0, the solution satises conditions (C
2
).
The dual algorithm constructs solutions by making sure that Eq.
(28) is satised. Therefore, the dual solutions are always such that
b
s1
w
it;s1
6 M
its
. If M
its
= 0, then w
it;s1
= 0, and, since the dual
algorithm sets M
it;s1
= M
its
b
s1
w
it;s1
, we also have M
it;s1
= 0.
Continuing this way, it is clear that if at some point in the calcula-
tion we get M
its
= 0, we subsequently obtain
M
its
= M
it;s1
= . . . = M
itT
= 0
and
w
its
= w
it;s1
= . . . = w
itT
= 0:
The primal algorithm sets y
it
= 1 only when M
its
= 0; this implies
that conditions (C
1
) will always be satised. Also, conditions (C
3
)
are always satised as the dual algorithm sets h
it
=
A
i
C
i
for
i = 1; . . . ; I and t = 1; . . . ; T. Given that h
it
is positive (for A
i
0), for
conditions (C
6
) to be satised, z
it
should equal

T
s=t
q
its
C
i
. However,
the primal algorithm rounds up the value of z
it
. As a result, for prob-
lems in which the optimal solution contains partial replenishment
(Lemma 2), conditions (C
6
) will not be satised. h
Hence, one can determine whether the solution obtained with
the primal and dual algorithms is optimal by checking if conditions
(C
4
) and (C
6
) are satised, or if the objective function values from
the primal and dual algorithms are equal. The running time of this
primaldual algorithm is O(IT
2
).
4. Computational results
This section describes our computational experiences in solving
the economic lot-sizing problem with multi-mode replenishment.
We generate a wide range of test problems to provide some indica-
tion of both the quality and the computing time of the solutions
generated using the primaldual algorithm. Fig. 2. Dual algorithm.
Fig. 3. Primal algorithm.
S.D. Eks ioglu/ European Journal of Operational Research 197 (2009) 93101 97
The quality of the upper and lower bounds for the problems we
generate compares to optimal solutions from solving formulation
Q by using CPLEX 10.1 callable libraries. CPLEX callable libraries
also solve the LP-relaxations of formulations (P) and (Q). The algo-
rithms are written in C-programming language and are compiled
and executed on an HP personal computer with a 2.33 GHz Intel
Core 2 processor and 2 GB RAM.
We have organized our results into two groups: First, we dem-
onstrate how the primaldual algorithm works by using a simple
example; second, we implement the primaldual algorithm on a
set of randomly generated problems and present the results.
4.1. Example
This section demonstrates how the primaldual algorithm
works by using a simple example. We consider a lot-sizing prob-
lem with 3 periods and 2 replenishment modes. The data for this
problem are presented in Table 1. Tables 7 and 8 (see Appendix)
present the steps of the dual algorithm, and Table 9 (see Appendix)
presents the steps of the primal algorithm for this example.
4.2. Experimental analysis
We generate a nominal case problem as follows:
v Fixed set-up costs s
i
- U[200; 600[.
v Fixed cost of a cargo container A
i
- U[200; 400[.
v Unit procurement costs p
i
- U[25; 75[.
v Holding costs h
t
- U[5; 15[.
v Capacity of a cargo container C
i
- U[50; 60[.
Most of the parameters are the same as the ones used by Wu
and Golbasi (2002) and Eks ioglu et al. (2006, 2007) for a related
problem. We consider demands to show seasonality pattern, and
they are generated as follows:
b
t
= 200 rz
t
asin
2p
d
(t d=4)
_ _
where
r is the standard error of demand.
z
t
i.i.d. standard normal random variable.
a amplitude of the seasonal component.
d is the length of a seasonal cycle in periods.
These demands are generated in the same way as in Baker et al.
(1978) and Chen et al. (1994). In our test problems, we take
r = 3:2, a = 6:2 and d = 12.
Varying the number of replenishment modes and the number of
time periods, we generated 15 different problems. For each prob-
lem we generate 20 problem instances, and we report on the aver-
age error bounds and average running times. Table 2 presents the
characteristics of the problems generated, and Table 3 presents the
results from solving these problems. The quality of the lower
bounds found using the dual algorithmcompare to the correspond-
ing optimal solutions from CPLEX. The tightness of the lower
bounds is measured as follows:
Dual Error (%) =
CPLEX Lower bound
CPLEX
+ 100:
The quality of the upper bounds that are found by using the primal
algorithm compare to the corresponding optimal solutions from
CPLEX. The tightness of the upper bounds can be measured as
follows:
Primal Error (%) =
Primal Solution CPLEX
CPLEX
+ 100:
Table 1
Data for economic lot-sizing problem with T = 3 and I = 2
b s p h A C
b
1
= 10 s
1
= 0 p
1
= 1 h
1
= 0 A
1
= 50 C
1
= 25
b
2
= 15 s
2
= 25 p
2
= 2 h
2
= 1; 000 A
2
= 10 C
2
= 100
b
3
= 100
Table 2
Problem characteristics
Problem I T Nodes Arcs
1 20 24 505 6,480
2 30 24 745 9,720
3 40 24 985 12,960
4 20 48 1,009 24,480
5 30 48 1,489 36,720
6 40 48 1,969 48,960
7 20 96 2,017 95,040
8 30 96 2,977 142,560
9 40 96 3,937 190,080
10 20 192 4,033 374,400
11 30 192 5,953 561,600
12 40 192 7,873 748,800
13 20 384 8,065 1,486,080
14 30 384 11,905 2,229,120
15 40 384 15,745 2,972,160
Table 3
Summary of experimental results
Problem Error (in %) CPU (in sec)
Dual Primal CPLEX Dual Primal
1 0.53 0.33 15.05 0.00 0.00
2 0.60 0.40 439.15 0.00 0.00
3 0.61 0.32 129.75 0.00 0.00
4 0.61 0.29 956.13 0.00 0.00
5 0.58 0.25 4853.00 0.00 0.00
6 0.41 0.61 18406.67 0.00 0.00
7 1.05 1.05 N/A 0.00 0.00
8 1.03 1.03 N/A 0.00 0.00
9 0.99 0.99 N/A 0.00 0.00
10 0.93 0.93 N/A 0.00 0.05
11 1.02 1.02 N/A 0.00 0.00
12 0.97 0.97 N/A 0.00 0.10
13 0.95 0.95 N/A 0.00 0.10
14 0.94 0.94 N/A 0.00 0.10
15 0.91 0.91 N/A 0.00 0.20
Table 4
Summary of experimental results
Problem Error (in %) CPU (in sec)
CPLEX Primal CPLEX Primal
1 0.66 0.04 0.05 0.00
2 0.81 0.01 0.20 0.00
3 0.74 0.00 0.30 0.00
4 0.66 0.01 0.50 0.00
5 0.74 0.00 0.95 0.00
6 0.76 0.02 1.20 0.00
7 0.78 0.03 4.20 0.00
8 0.80 0.01 5.95 0.00
9 0.74 0.02 8.60 0.00
10 0.67 0.02 24.00 0.05
11 0.74 0.02 39.25 0.00
12 0.76 0.02 52.20 0.10
98 S.D. Eks ioglu/ European Journal of Operational Research 197 (2009) 93101
While solving problem number 4, CPLEX ran out of memory in 5 of
the 20 probleminstances generated. CPLEX ran out of memory in 13
of the 20 instances generated for problem 5 and in 17 of the 20 in-
stances generated for problem 6. CPLEX ran out of memory while
solving all instances generated for problems 715. Therefore, the er-
ror gaps presented for problems 715 are calculated as follows:
Error(%) =
Primal Solution Dual Solution
Primal Solution
+ 100:
The primaldual algorithm performed very well while solving prob-
lems 115 as the maximum error gap presented is 1.05%, and the
maximum running time is 0.20 cpu seconds. The results show that
increasing the number of time periods and transportation modes,
while they do not really impact the running time of the primaldual
algorithm, they do impact the running time of CPLEX. For problems
715 CPLEX ran out of memory mainly in the search for the optimal
solution. In order to have a fair comparison of our algorithm to
CPLEX, we performed the following experiment: We fed the error
gap found by the primaldual algorithm for problems 115 to
CPLEX, and used that gap as a terminating condition. As a result,
any time that CPLEX found a solution that gave the same or smaller
error gap than the one generated by the primaldual algorithm, it
terminated and reported the best lower and upper bound found
so far. The results of the experiment are presented in Table 4. CPLEX
error reported is the duality gap found by CPLEX, and the Primal er-
ror is calculated as presented above. CPLEX was not able to solve
problems 1215 because it ran out of memory due to the problem
size (not in the search for an optimal solution). The error gap pro-
vided by CPLEX and the primaldual algorithm was almost the
same. However, the running time of the primaldual algorithm
was smaller than CPLEX. We wanted to provide more insights about
how problem characteristics affect the performance of the primal
dual algorithm. It has been shown in other studies (Hochbaum and
Segev, 1989; Eks ioglu et al., 2002) that in network ow problems
with xed charge costs, the problem difculty depends on the val-
ues of the xed costs. We alter the nominal case problem by chang-
ing the distribution of the xed setup cost s
i
and generate additional
problems. The distribution of s
i
is changed to U[50; 100[, U[100; 200[,
U[200; 400[, U[200; 600[, U[200; 900[, U[600; 900[, U[900; 1200[, and
U[1200; 1500[, problems 1623. For each problem we generate 20
instances and report the average over these instances. We change
the distribution of the xed cost of a cargo container A
i
(of the
nominal case problem) to U[50; 100[, U[100; 200[, U[200; 400[,
U[200; 600[, U[200; 900[, U[600; 900[, U[900; 1200[, and
U[1200; 1500[, problems 2431. We change the distribution of de-
mand b
t
to U[5; 15[, U[10; 20[, U[50; 100[, U[100; 200[, U[200; 400[,
and U[200; 600[, problems 3237. We change the distribution of
holding cost h
t
to U[10; 20[, U[50; 100[, U[100; 200[, U[200; 400[,
and U[200; 600[, problems 38-42. We nally change the capacity
of the cargo container C
i
to U[50; 100[, U[100; 200[, U[150; 250[,
U[200; 400[, and U[400; 600[, problems 4347. The results from
these experiments are presented in Table 5.
The running time of the primaldual algorithm is less than 0.01
CPU seconds. Increasing the xed cost of a cargo container seems
to have the highest impact on the performance of the primaldual
algorithmand on CPLEX. The maximum error found by the primal
dual algorithm is 2.19% and corresponds to problem 31. Problems
4347 indicate that increasing the capacity of a cargo container
makes the problems easier. As the cargo capacity increases, de-
mand can be replenished using a single cargo container most of
the time. If only one cargo container is used, the xed cost of
replenishing demand will consist of the xed setup cost (s
i
) and
only one payment of the xed cargo cost (A
i
). In this case the
replenishment cost function can be described by a xed charge
function, with the xed charge being equal to s
i
A
i
. As demon-
strated in Section 2.1, the economic lot-sizing problem with mul-
ti-mode replenishment and xed charge cost function is an
easier problem to solve.
Finally, Table 6 presents the error gap of LP-relaxation of formu-
lations (P) and (Q). The results indicate that the lower bounds
found using LP-relaxation of Q are tighter. This is because the
upper bound on variables q
its
in formulation Q is tighter than the
upper bound on variables q
it
in formulation P. In formulation Q,
due to constraints (16), the upper bound for variables q
its
is the de-
mand in period s. In formulation P, the upper bound for variables
q
it
is b
tT
(constraints (3)), the demand from period t to the end of
the horizon. b
tT
is a very high upper bound as the amount received
in a period rarely equals this amount.
Table 5
Summary of experimental results
Problem Error (in %) CPU (in sec) Problem Error (in %) CPU (in sec)
Dual Primal CPLEX Dual Primal CPLEX
16 0.58 0.39 270.76 32 0.07 1.07 0.55
17 0.67 0.36 433.00 33 0.37 0.72 0.85
18 0.57 0.35 125.42 34 1.77 0.96 192.00
19 0.64 0.44 102.55 35 0.97 0.76 297.84
20 0.59 0.32 353.75 36 0.50 0.27 37.45
21 0.59 0.29 108.15 37 0.36 0.25 79.05
22 0.50 0.23 29.70 38 0.63 0.26 13.40
23 0.42 0.20 3.40 39 0.89 0.08 1451.72
24 0.22 0.03 0.25 40 0.93 0.04 303.65
25 0.41 0.13 4.45 41 1.05 0.06 734.78
26 0.59 0.33 54.15 42 0.89 0.03 11.39
27 0.70 0.50 114.26 43 0.58 0.26 3.60
28 0.64 0.48 392.90 44 0.89 0.17 392.45
29 0.92 1.10 2010.38 45 0.90 0.16 1.85
30 0.93 1.56 3017.64 46 0.59 0.14 0.35
31 0.96 2.19 8709.57 47 0.09 0.02 0.20
Table 6
Comparison of LP-relaxations
Problem Formulation P Formulation Q
Error(in %) CPU(in sec) Error(in %) CPU (in sec)
16 1.04 0.10 0.58 0.00
17 1.63 0.00 0.67 0.10
18 2.31 0.00 0.57 0.10
19 3.31 0.00 0.64 0.10
20 3.54 0.00 0.59 0.10
21 5.16 0.00 0.59 0.10
22 6.12 0.10 0.50 0.50
23 7.48 0.00 0.42 0.60
S.D. Eks ioglu/ European Journal of Operational Research 197 (2009) 93101 99
5. Conclusions
In this paper we discuss the economic lot-sizing problem with
multi-mode replenishment. This problem is an extension of the
classical economic lot-sizing model that accounts for the fact that
more than one supplier or transportation mode can be available to
replenish the inventory. In this case, managers will have to decide
how much and when to order, as well as which supplier and what
transportation mode to use.
We dene the problem as a network ow problem with addi-
tional side constraints and provide an MIP formulation. This for-
mulation takes care of the special structure of the inventory
replenishment costs by using a multiple setups cost function. We
provide an additional MIP formulation for this problem by reden-
ing the decision variables. The LP-relaxation of the reformulation
gives lower bounds that are at least as high as the lower bounds
from the LP-relaxation of theoriginal problem formulation. How-
ever, tighter lower bounds come to the expense of the problem
size. The number of decision variables changes from (2 + T + I
T 1) to (T + I +
T3
2
). We show that the dual of the LP-relaxation
of the reformulation has a special structure, and take advantage
of this structure to design a primaldual algorithm. The primal
dual algorithm described in Section 3.2 can be used to generate
lower and upper bounds for the economic lot-sizing problem with
multi-mode replenishment and cargo capacity constraints. The
running time of the algorithm is O(IT
2
) times.
We tested the performance of the algorithms on a wide range of
randomly generated test problems. Computational results show
high quality solutions are generated using the primaldual algo-
rithm. The maximum error gap was 2.19%, and the maximum run-
ning time was less than 1 cpu seconds. The running time of CPLEX
was affected most by the increase in the number of time periods
and replenishment modes as well as by the xed cost of a cargo
container. We identied a number of problems that CPLEX could
not solve because it ran out of memory. However, for all problem
classes, the primaldual algorithm gave good quality solutions in
a reasonable amount of time. These results clearly demonstrate
the superiority of the primaldual algorithm to handle, in a very
reasonable amount of time, large size problems.
Appendix
See Tables 1,2,3,4,5,6,7,8,9.
References
Aggarwal, A., Park, J., 1993. Improved algorithms for economic lot size problems.
Operations Research 41, 549571.
Bahl, H.C., Ritzman, L.P., Gupta, J.N.D., 1987. Determining lot sizes and resource
requirements: A review. Operations Research 35 (3), 329345.
Baker, K.R., Dixon, P., Magazine, M.J., Silver, E.A., 1978. An algorithmfor the dynamic
lot-size problem with time-varying production capacity constraints.
Management Science 24, 17101720.
Barany, I., van Roy, J., Wolsey, L., 1984. Strong formulations for multi-item
capacitated lot sizing. Management Science 30, 12551261.
Baumol, W.J., Vinod, H.D., 1970. An inventory theoretic model of freight transport
demand. Management Science 16 (7), 413421.
Belvaux, G., Wolsey, L.A., 1998. Lotsizelib: A library of models and matrices for lot-
sizing problems. Core discussion paper, Universite Catholique de Louvain.
Belvaux, G., Wolsey, L.A., 1999. Lot-sizing problems: Modeling issues and a
specialized branch- and-cut system bc-prod. Core discussion paper dp9849,
Universite Catholique de Louvain.
Belvaux, G., Wolsey, L.A., 2001. Modelling practical lot-sizing problems as mixed-
integer programs. Management Science 47 (7).
Bitran, G.R., Yanasse, H.H., 1982. Computational complexity of the capacitated lot
size problem. Management Science 28, 11741186.
Burns, L.D., Hall, R.W., Blumenfeld, D.E., Daganzo, C.F., 1985. Distribution strategies
that minimize transportation and inventory costs. Operations Research 33 (3),
469490.
Chen, H.-D., Hearn, D.W., Lee, C.-L., 1994. A new dynamic programming algorithm
for the single item dynamic lot size model. Journal of Global Optimization 4,
285300.
Chopra, S., Meindl, P., 2006. Supply Chain Management: Strategy Planning and
Operation, third ed. Prentice-Hall, Inc, Upper Saddle River, NJ.
Eks ioglu, B., Eks ioglu, S.D., Pardalos, P.M., 2002. Solving large-scale xed charge
network ow problems. In: Daniele, P., Giannesi, F., Maugeri, A. (Eds.),
Variational Inequalities and Equilibrium Models. Kluwer Academic Publisher.
Eks ioglu, S.D., Eksioglu, B., Romeijn, H.E., 2007. A lagreangean heuristic for
integrated production and transportation planning problems in a dynamic,
multi-item, two-layer supply chain. IIE Transactions in Scheduling and Logistics
39 (2).
Eks ioglu, S.D., Romeijn, H.E., Pardalos, P.M., 2006. A cross-facility production and
transportation planning problem. Computers and Operations Research 33 (11).
Table 7
Example: Implementing dual algorithm
Initially: M
110
= M
121
= M
132
= 0, M
210
= M
221
= M
232
= 25
s t i
1 v
1
= minc
111

M110
b1
; c
211

M210
b1

v
1
= min3
0
10
; 2:10
25
10
= 3
1 1 M
111
= max0; M
110
b
1
+ max0; v
1
c
111

M
111
= max0; 0 10 + max0; 3 3 = 0
2 M
211
= max0; M
210
b
1
+ max0; v
1
c
211

M
211
= max0; 25 10 + max0; 3 2:10 = 16
s i
2 v
2
= minc
112

M111
b2
; c
212

M211
b2
; c
122

M121
b2
; c
222

M221
b2

v
2
= min3
0
15
; 2:10
16
15
; 3
0
15
; 2:10
25
15
= 3
1 1 M
112
= max0; M
111
b
2
+ max0; v
2
c
112

M
112
= max0; 0 15 + max0; 3 3 = 0
2 M
212
= max0; M
211
b
2
+ max0; v
2
c
212

M
212
= max0; 16 15 + max0; 3 2:10 = 2:5
2 1 M
122
= max0; M
121
b
2
+ max0; v
2
c
122

M
122
= max0; 0 15 + max0; 3 3 = 0
2 M
222
= max0; M
221
b
2
+ max0; v
2
c
222

M
222
= max0; 25 15 + max0; 3 2:10 = 11:50
Table 8
Example: Implementing dual algorithm (cont)
s t i
3 v
3
= minc
113

M112
b3
; c
213

M212
b3
; c
123

M122
b3
; c
223

M222
b3
; c
133

M132
b3
,
c
233

M232
b3

v
3
= min1003
0
100
; 1002:10
2:5
100
; 1003
0
100
; 1002:10
11:50
100
;
3
0
100
; 2:10
25
100
= 2:35
1 1 M
113
= max0; M
112
b
3
+ max0; v
3
c
113

M
113
= max0; 0 100 + max0; 2:35 1003 = 0
2 M
213
= max0; M
212
b
3
+ max0; v
3
c
213

M
213
= max0; 2:5 100 + max0; 2:35 1002:10 = 2:5
2 1 M
123
= max0; M
122
b
3
+ max0; v
3
c
123

M
123
= max0; 0 100 + max0; 2:35 1003 = 0
2 M
223
= max0; M
222
b
3
+ max0; v
3
c
223

M
223
= max0; 11:50 100 + max0; 2:35 1002:10 = 11:50
3 1 M
133
= max0; M
132
b
3
+ max0; v
3
c
133

M
133
= max0; 0 100 + max0; 2:35 3 = 0
2 M
233
= max0; M
232
b
3
+ max0; v
3
c
233

M
233
= max0; 25 100 + max0; 2:35 2:10 = 0
The lower bound generated is equal to:3 + 10 3 + 15 2:35 + 100 = 310:
Table 9
Example: Implementing primal algorithm
Start j = 3 and t = 1
Step 1 M
233
= 0 and c
233
v
3
max0; v
3
c
233
= 0, therefore y
23
= 1
Step 2 c
233
v
3
max0; v
3
c
233
= 0, thus q
233
= 100
Start j = 2 and t = 1
Step 1 M
112
= 0 and c
112
v
2
max0; v
2
c
112
= 0, therefore y
11
= 1
Step 2 c
112
v
2
max0; v
2
c
112
= 0, thus q
112
= 15
c
111
v
1
max0; v
1
c
111
= 0, thus q
111
= 10
Step 3 z
11
=
q
111
q
112
C1
_ _
=
25
25
_ _
= 1
z
23
=
q
233
C2
_ _
=
100
100
_ _
= 1
The upper bound generated is equal to:1 + 25 1 + 50 2 + 100 1 + 25
1 + 10 = 310:
100 S.D. Eks ioglu/ European Journal of Operational Research 197 (2009) 93101
Federgruen, A., Tzur, M., 1991. A simple forward algorithm to solve general dynamic
lot sizing models with n periods in o(n log n) or o(n) time. Management Science
37, 909925.
Florian, M., Klein, M., 1971. Deterministic production planning with concave costs
and capacity constraints. Management Science 18 (1), 1220.
Gupta, O.K., 1992. A lot size model with discrete transportation costs. Computers
and Industrial Engineering 22, 397402.
Hahm, J., Yano, C.A., 1992. The economic lot and delivery scheduling problem: the
single item case. International Journal of Production Economics 28, 235252.
Hahm, J., Yano, C.A., 1995. The economic lot and delivery scheduling problem: The
common cycle case. IIE Transactions 27, 113125.
Hochbaum, D.S., Segev, A., 1989. Analysis of a ow problem with xed charges.
Networks 19, 291312.
Jaruphongsa, W., etinkaya, S., Lee, C.Y., 2005. A dynamic lot-sizing model with
multi-mode replenishments: Polynomial algorithms for special cases with dual
and multiple modes. IIE Transactions 37, 453467.
Kaminsky, P., Simchi-Levi, D., 2003. Production and distribution lot sizing in a two
stage supply chain. IIE Transactions 35, 10651075.
Krarup, J., Bilde, O., 1977. Plant location, set covering, an economic lot-size: An
o(mn) algorithm for structured problems. Numerische Methoden bei
Optimierungsaufgaben. Band 3: Optimierung bei Graphentheorietischen und
Ganzzahlieng Problemen, 155186.
Lee, C.-Y., 1989. A solution to the multiple set-up problem with dynamic demand.
IIE Transactions 21, 266270.
Manne, A.S., 1958. Programming of economic lotsizes. Management Science 4, 115
135.
Nahmias, S., 1997. Production and Operations Analysis, 3rd ed. Irwin Book Team.
Nemhauser, G.L., Wolsey, L.A., 1988. Integer and Combinatorial Optimization. Wiley
John & Sons, Incorporated.
Wagelmans, A., van Hoesel, S., Kolen, A., 1992. Economic lot sizing: An o(n log n)
algorithm that runs in linear time in the wagner-whitin case. Operations
Research 40 (1), 145156.
Wagner, H.M., Whitin, T.M., 1958. Dynamic version of the economic lot size model.
Management Science 5, 8996.
Wolsey, L.A., 1995. Progress with single-item lot-sizing. European Journal of
Operational Research 86, 395401.
Wu, S.D., Golbasi, H., 2002. Manufacturing planning over alternative facilities:
Modeling, analysis, and algorithms. In: Geunes, J., Pardalos, P.M., Romeijn, H.E.
(Eds.), Supply Chain Management: Models, Applications, and Research
Directions. Kluwer Academic Publisher, pp. 279316.
Zangwill, W.I., 1968. Minimum concave cost ows in certain networks.
Management Science 14, 429450.
S.D. Eks ioglu/ European Journal of Operational Research 197 (2009) 93101 101

Você também pode gostar