ARTICLE IN PRESS

Int. J. Production Economics 120 (2009) 607612

Contents lists available at ScienceDirect

Int. J. Production Economics

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

Dynamic lot sizing problem with continuous-time Markovian

production cost
Amir Azaron a,b,, Ou Tang b,d, Reza Tavakkoli-Moghaddam c
a

Michael Smurt Graduate School of Business, University College Dublin, Carysfort Avenue, Blackrock, Co. Dublin, Ireland
ping University, Linko
ping, Sweden
Department of Management and Engineering, Linko
Department of Industrial Engineering, College of Engineering, University of Tehran, Tehran, Iran
d
School of Economics and Management, Tongji University, Shanghai, China
b
c

a r t i c l e in fo

abstract

Article history:
Received 11 April 2007
Accepted 10 April 2009
Available online 21 April 2009

This paper develops a polynomial algorithm for obtaining dynamic economic lot sizes in
a single product multiperiod production system with the objective of minimizing total
production and inventory costs over T periods. It is assumed that production costs are
linear, inventory costs are concave, setup costs are zero and backlogging is not permitted
in all periods. Moreover, the unit production cost is a stochastic variable, which is
evolved according to a continuous-time Markov process over the planning horizon. The
model is formulated as a stochastic dynamic programming (DP) optimization with the
state variable being unit production cost. Then, it is solved using the backward dynamic
programming approach. To justify the application of the proposed model, two practical
cases are presented.
& 2009 Elsevier B.V. All rights reserved.

Keywords:
Dynamic lot sizing
Dynamic programming
Stochastic processes

1. Introduction
The classical dynamic lot sizing problem species a
discrete-time nite horizon single product inventory
management problem subject to deterministic timevarying demand that must be satised. This problem has
been the subject of extensive research since its introduction in the late 1950s until today. When the production
cost and the inventory cost of each period are linear,
several authors have presented theorems that can reduce
the computational effort required in solving the problem.
Wagner and Whitin (1958) and Zabel (1964) give results
for the no-backlogging case, while Zangwill (1969)
analyzes the backlog case. Many generalizations of the
basic model have been considered including introducing
bounds on inventory and/or production capacity, as well
 Corresponding author at: Michael Smurt Graduate School of
Business, University College Dublin, Carysfort Avenue, Blackrock, Co.
Dublin, Ireland.
E-mail addresses: amir.azaron@ucd.ie, aazaron@gmail.com
(A. Azaron).

0925-5273/$ - see front matter & 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.ijpe.2009.04.007

as generalizations to multiproduct settings. Brahimi et al.

(2006), Karimi et al. (2003) and Maes and Wassenhove
(1998) provide excellent reviews about single item lot
sizing, multiproduct capacitated lot sizing and multi item
single level capacitated dynamic lot sizing problems,
respectively.
In the last decade, three important papers, Aggarwal
and Park (1993), Federgruen and Tzur (1991) and Wagelmans et al. (1992) improved the time complexity for
obtaining an optimal solution from O(T2) to O(T log T), if T
represents the length of the planning horizon.
The existing materials in the area of stochastic lot
sizing consider mainly demand as the uncertain parameter. Silver (1978) applies a heuristic to compute the
actual lot size as the solution of a newsvendor problem.
Bookbinder and Tan (1988) convert the stochastic problem
to an equivalent deterministic problem that has the same
form as the deterministic dynamic lot sizing problem. Sox
(1997) developed an optimal algorithm for the single item
dynamic lot sizing problem with random demand and
non-stationary costs. Melo (1996) provides an excellent
review about the stochastic lot sizing in production

ARTICLE IN PRESS
608

A. Azaron et al. / Int. J. Production Economics 120 (2009) 607612

planning. Sox et al. (1999) survey the most current

research literature on the stochastic lot scheduling
problem, which deals with scheduling production of
multiple products with stochastic demand. Recently,
Haugen et al. (2001) used a meta-heuristic to solve
stochastic lot sizing problems.
This paper develops a polynomial algorithm for
obtaining the optimal production scheduled for each
period in single product multiperiod production systems
with stochastic production (or purchasing) costs. The
productions (purchases) must be made in markets
characterized by signicant cost (price) uctuations.
There is no assurance that the price tomorrow will be
the same as that today. Examples of such commodities are
rubber, sugar, copper, coffee and cereals. When the price is
low the purchasing ofcer would like to buy to build up
his stocks to meet at least part of the requirements for the
periods when the price is high. When the price is high he
will avoid purchasing and use his stocks to meet his
current needs.
Published paper in the eld of stochastic price
illustrating the use of operational research techniques is
rare, see Kingsman (1986). To our knowledge, only Fabian
et al. (1959) and Kingsman (1969) considered this type of
problem, but in both of their studies prices in successive
periods are independent, which is a restrictive assumption
in many practical cases.
In this study, we assume that the unit production cost
(price) is a stochastic variable, which is evolved according
to a continuous-time Markov process over the time
horizon. So, the actual price in each period will be
dependent on the price in the previous period. This
assumption is different compared with the ones in Fabian
et al. (1959), Kingsman (1969). Such an assumption also
reects better the price movement in practice and it is the
main contribution of this paper. To justify the application
of the proposed model, two practical cases are presented
below. In both, the price is varying over the time horizon
and in each period its change depends on the price in the
previous period.
First, consider the problem of purchasing oil by
governments, according to the variations of oil price over
the planning horizon. In this regard, oil price can be
considered as low, medium and high levels or even more
accurately in several levels. The length of planning
horizon is dependent on the political situation of each
country (normally 4 years). Clearly, the duration of time
that the oil price remains at its present level, before
proceeding to the next level, is not constant, rather it
would be a random variable and can be considered to
follow an exponential distribution. The government
purchases oil in discrete points during the planning
horizon. Moreover, if it seems that the price of oil, in
any period, is relatively inexpensive, it might be a good
decision to purchase oil in a high volume and store it for
the next periods needs.
To model this problem and to obtain the economic oil
ordering size (number of million gallons) per period
(possibly year), oil price can be supposed to evolve
according to a proper continuous-time Markov process.
In order to properly estimate the transition probabilities

(we explain more about this in Section 2), we ignore the

transitions of the oil price inside each level and only
consider the transitions of the oil price from each level to
other levels in this particular problem. In this regard, it is
reasonable to assume that after remaining the oil price at
any level for some time, which is stochastic, it will
eventually transit to one of the other levels. In this
problem, a transition matrix with zero diagonal elements
can properly reect the dynamic nature of oil price during
the planning horizon. As long as we decrease the range of
levels, the accuracy of the proposed model will be
increased.
Another important application arises in Iran where the
unit production costs are inuenced by the government
which frequently xes costs of basic raw materials. The
cabinet, one or two times per year, decides about xing
the unit costs for a number of strategic raw materials like
steel, wood, petroleum, etc., according to ination in the
country which is always upward. In this regard, the
cabinet submits the corresponding bill to the parliament
and the parliaments decision on xing the unit costs or
increasing them to one of the proposed unit costs, based
on the cabinets bill and the ination rate, is nal.
Increasing the raw materials costs clearly increases the
unit production costs, because for producing each product, the producer needs to purchase some related raw
materials from the government and then makes the nal
product. Moreover, the duration of time that the raw
materials costs remain at their present positions, before
changing to some new positions, is not constant, rather it
depends on the time interval for deciding about changing
the raw materials costs by the government (the cabinet
and the parliament).
To model this problem in order to obtain the economic
lot size in single product multiperiod production systems,
the unit production cost can be supposed to evolve
according to a proper continuous-time Markov process.
When the state variables (values of the unit production
cost) are sorted from the smallest to the largest one, an
upper triangular transition matrix with a single absorbing
state can properly reect the dynamic nature of the unit
production cost and its dependency on the governments
decision in this particular problem in Iran.
What makes a continuous-time Markov chain different
from a Markov chain is that the amount of time it spends
in each state, before proceeding to the next state, is
exponentially distributed, instead of a constant deterministic value. In both practical cases, the amount of time to
spend in each state is clearly a random variable. The cost
(price) in each period is also dependent on the cost in the
previous period. If the process of evolving the cost
coefcient over the time horizon is memoryless, then it
is possible to model the cost changes occurring at any
time over the planning horizon including the beginning of
each period, when the decisions about the economic lot
sizes are actually made, and to derive the proper
stochastic dynamic programming recursive functions.
The only distribution which poses the memoryless
property is exponential. Therefore, the above mentioned
problems and many similar ones in production and
business can be best modeled by continuous-time Markov

ARTICLE IN PRESS
A. Azaron et al. / Int. J. Production Economics 120 (2009) 607612

processes. We think the memoryless assumption is quite

reasonable. All price information in early periods is
included in the last period, from which the current price
will departure.
Finally, the model is formulated as a stochastic DP
optimization with the state variable being unit production
cost, and the corresponding DP functional equation is
derived. Then, it is solved in the backward DP way. We
also discuss about the complexity of the proposed
algorithm to solve the problem.

609

Another way to estimate the parameters is using the

judgments of experts in this area, especially in unstable
processes. A combination of both approaches, depending
on how much the process is unstable, can also be used to
estimate the parameters. In the combination approach, as
much as the instability of the process is increased, more
weight is assigned to the experts points of views for
estimating the parameters.
Assumption 3. Let fim t represent the conditional probability that the unit production cost changes to cm at time
t, given that it was ci at time zero.

2. The dynamic lot sizing problem

Notations:
T
Xt
Dt
It
Ht(It)
ct

number of periods in the planning horizon

production scheduled for period t (t 1,2,y,T)
expected demand in period t
net inventory at the end of period t (I0 IT 0)
inventory cost in period t (concave function)
unit production cost at the beginning of period t

Assumption 1. Let the number of states of the unit

production cost be equal to N (these states are in this
order: c1,c2,y,cN), and pim represents the probability of
changing this cost from ci to cm. These probabilities form
an N  N transition matrix P.
Assumption 2. Let tim represent the duration of time that
the unit production cost is ci, before changing to cm. tim is a
random variable with the exponential density function
fim(t) and the parameter lim, because the unit production
cost is evolved according to a continuous-time Markov
process. Clearly wi(t) or the density function of staying
time in state ci is given by
wi t

N
X

pim f im t

(1)

m1

The parameters of the continuous-time Markov process

can be estimated from available historical data, based on
the structure of the problem. For example, in the problem
of purchasing oil, the historical data related to the staying
times in different price levels and the number of
transitions from each level to other levels during the
previous L periods can be easily collected. Since changing
the oil price over the time horizon is not a stable process,
very old data should not be used to estimate the
parameters. Then, the average staying times in each level
i 1=li and also the probabilities of changing the price
from each level i into any new level m (pim) are all
estimated by using collected historical data. Finally, the
values of lim are estimated as follows:

lim pim li

(2)

The readers may also refer to Boguslavskiy (2005) to nd

out about using the method of maximum of likelihood for
estimating the parameters of a continuous-time Markov
process.

How can a process that started by entering state ci at time

zero be in state cm at time t? One situation in which this is
possible is when ci and cm represent the same state and
the process does not leave state ci throughout the period
(0, t). This requires that the process makes its rst
transition after time t. Every other way to get from state
ci to state cm in the interval (0, t) requires that the process
makes at least one transition during that interval. For
example, the process could have made its rst transition
from state ci to some state cl at a time t, 0otot, and then
by some succession of transitions have made its way to
state cm at time t. These considerations lead us to Eq. (3)
for computing fim t

fim t dim
(

dim

wi t dt

N
X
l1

if i m

otherwise

pil

t
0

f il tflm t  t dt
(3)

We cannot directly compute fim t from Eq. (3), but

since the second integral of Eq. (3) is a convolution of two
functions, fim t can be computed by using the Laplace
e
transform. Let f il s represent the Laplace transform of fil(t)
e
and fim s represent the Laplace transform of fim t,
which is given by

feim s dim

1
0

1
t

est wi t dt dt

N
X

pil f il sflm s

(4)

l1

Now, we can compute fim t by taking the inverse Laplace

e
of fim s in Eq. (4), see Howard (1971) for more details.
When the unit production cost in each period is
deterministic and dynamic over the planning horizon,
the appropriate model for obtaining the optimal values of
Xt is
Min

T
X

ct X t

t1

s:t:

T 1
X

Ht It

t1

X 1  D1 I1
It1 X t  Dt It

t 2; 3; . . . ; T  1

IT1 X T  DT 0
It X0;

X t X0

(5)

It is proved, in this case, that an optimal program has

the property that production in any period t must be one
Pk
of these values: 0 or
jt Dj, for k t, t+1,y,T, because
It1Xt 0, see Johnson and Montgomery (1974) for more
details.

ARTICLE IN PRESS
610

A. Azaron et al. / Int. J. Production Economics 120 (2009) 607612

So, if we make a decision to supply the requirements

for periods j+1, j+2,y,k by production at the beginning of
period j+1, then in an optimal program we will have

3. Numerical examples

X j1 Dj1 Dj2    Dk

Consider the following four-period production planning problem involving raw material purchasing in Iran.
There is no initial inventory and the nal inventory level is
to be zero. The inventory cost is linear. Assuming ht as the
unit holding cost in period t, the inventory cost has the
form as Eq. (9). Estimates of ht and Dt are given in Table 1:

It X j1 

t
X

k
X

Dr

Dr

t j 1; j 2; . . . ; k  1

rt1

rj1

Ij Ik 0

(6)

and the total production and inventory costs from period

j+1 to k will be as Eq. (7), and the DP functional equations
can be easily written in this case to obtain the optimal
production scheduled in all periods:
k
X

cj1

Dr

rj1

k1
X

Ht

tj1

k
X

!
(7)

Dr

rt1

Now, let us get back to our model. The only difference

between the classical dynamic programming model and
ours is that in our model, the production cost in period t is
based on the value of the cost coefcient at the beginning
of the period. In this case, subject to the Markovian
assumption, ct plays the role of the state variable of a
stochastic sequential decision process.
Let Vj(ci) represent the minimum total costs from
period j+1 to T pertaining to zero inventory level at the
end of period j, if the unit production cost at the beginning
of period j+1 (decision point) is ci or cj1 ci.
The probability that after kj periods, the unit
production cost at the beginning of period k+1 changes
to cm, given that at the beginning of period j+1 this cost
was ci, is equal to fim k  j, because the continuous-time
Markov process corresponding to the transitions of the
cost coefcient over the time horizon is memoryless.
Then, by conditioning on the state of the unit production
cost after kj periods, the stochastic DP functional
equation is written as follows:
i

V j c min

8
<

jokpT :

N
X

ci

k
X
rj1

Dr

k1
X
tj1

fim k  jV k cm

Ht

k
X

!
Dr

rt1

i 1; 2; . . . ; N

3.1. Numerical example 1

Ht It ht It

The unit production cost is inuenced by the governments decision on xing or increasing the required raw
materials costs over the planning horizon. The cost
coefcient is assumed to evolve according to a continuous-time Markov process with the following ve states:
c1 100, c2 120, c3 150, c4 180 and c5 200. The
upper triangular transition matrix is estimated as
3
2
0 0:6 0:4 0
0
6 0 0 0:5 0:5 0 7
7
6
7
6
7
P6
6 0 0 0:2 0:5 0:3 7
7
6
0 0:8 0:2 5
40 0
0 0
0
0
1
It should be noted that the transition matrix in this
particular case should be in upper triangular form, as
explained in Section 1. But, it can technically have any
general form such as the transition matrix related to the
oil purchasing problem, which will be solved later.
It is also assumed that the time interval for deciding
about changing the raw materials costs by the government is distributed according to an exponential distribution with the average of one period. Accordingly, fim(t) for
all i and m would be
f im t et ;

t40

(10)

First, we compute fim t for i 1, 2, 3, 4, 5 and mXi, by

e
taking the inverse Laplace of fim s in Eq. (4), as follows
(clearly, fim t 0 for i 1, 2, 3, 4, 5 and moi):

f55 t 1;
(8)

(9)

f44 t e0:2t ;

0:8t

f33 t e

f45 t 1  e0:2t ,

f34 t 0:833e0:2t  0:833e0:8t

m1

The functional equation (8) is similar to the equation

associated with the classicaldeterministicWagner/
Whitin model, except that here future Vk(cm) values are
weighted by the conditional probabilities fim Dt.
Finally, the DP functional equation (8) is solved in a
backward manner for j T, T1,y,0. Clearly, VT(ci) 0 for
i 1; 2; . . . ; N. In order to solve (8), we need to compute
fim Dt by the inverse Laplace transformation in Eq. (4).
This is obtained by simultaneous linear equations, for
given s, in which computation time is not constant,
usually O(N3). Computing Vj(ci) in the recursive functions
(8) is carried out on O(T2N), because for each
j 0,1,y,T1, there are Tj combinations for all values
PT1
of Vj(ci), for i 1,2,y,N, and
j0 T  j TT 1=2.
Therefore, the time complexity of the proposed algorithm
would be O(N3+T2N).

f35 t 1  0:833e0:2t  0:167e0:8t ,

f22 t et ;

f23 t 2:5e0:8t  2:5et

f24 t 0:938et 1:146e0:2t  2:084e0:8t ,

f25 t 1 0:56et  1:146e0:2t  0:414e0:8t ,
f11 t et f12 t 0:6tet
Table 1
Data for the numerical example 1.
t

Dt

ht

1
2
3
4

20
30
40
30

20
20
30
30

ARTICLE IN PRESS
A. Azaron et al. / Int. J. Production Economics 120 (2009) 607612

f13 t 9:5et  1:5tet 9:5e8t ,

f14 t 6:641e

t

t

 0:562te

0:2t

1:276e

0:8t

 7:917e

f15 t 1 1:859et 0:338tet  1:583e0:8t  1:276e0:2t

611

Table 3
Data for the numerical example 2.
t

Dt

ht

1
2
3
4

10
8
15
20

20
20
30
40

Then, the DP functional equation (8) is solved in the

backward manner to obtain the optimal production
scheduled in each period. The results are summarized in
Table 2, where each cell of the table displays the
corresponding V t1 c and Xt values.
The information in Table 2 then provides optimal
decisions. For example, if the unit production cost at the
beginning of the planning horizon is equal to 100, then the
optimal production scheduled for period 1 would be 120,
or we should supply the requirements for all periods by
production at the beginning of period 1. The reason is that
the unit production cost at the beginning of the planning
horizon has the minimum value in this case. The total
costs in this case is equal to 16 300.
For another example, if the unit production cost at
the beginning of the planning horizon is equal to 150,
then the optimal production scheduled for period 1 is
equal to 50, which covers the demands for both the
rst and the second periods. Then, we jump to the third
period. According to the columns 3 and 4 of Table 2, the
optimal production scheduled for period 3 is equal to 40
either the unit production cost at the beginning of period
3 does not change or changes to 180 or 200, and then for
period 4 is equal to 30. This is due to the higher level of
the variable cost at the beginning of the planning horizon,
comparing with the previous case, which avoids producing a lot.

It is also assumed that the oil price remains at each

level, before proceeding to the next level, for an exponential time with the average of one period. F matrix with the
elements fim t for i 1, 2, 3 and m 1, 2, 3 is computed
as below:

3.2. Numerical example 2

6
F6
4

Oil is to be purchased for a four-period horizon. The

inventory has the same form as the numerical example 1.
Estimates of ht and Dt (in millions) are given in Table 3.
It is also assumed that the oil price is changing over the
planning horizon according to a continuous-time Markov
process with three states: c1 low level (40 Dollars per
gallon), c2 medium level (80 Dollars per gallon) and
c3 high level (120 Dollars per gallon). The transition

Table 2
Results of the numerical example 1.
c

100
120
150
180
200

t
1

16 300
120
18 434.8
90
21 018.98
50
22 242.38
20
24 000
20

12 300
100
14 300
100
16 763
70
18 342.34
30
20 000
30

7900
70
9300
70
11141.7
40
12 708.6
40
14 000
40

3000
30
3600
30
4500
30
5400
30
6000
30

Table 4
Results of the numerical example 2.
c

40 (low)
80 (medium)
120 (high)

3675.5
18
4387.8
10
5000.2
10

2771.2
23
3569.8
8
4172.8
8

1937.6
15
2868.8
15
3726.4
15

800
20
1600
20
2400
20

matrix is estimated as
2
3
0
1
0
6 0:4 0 0:6 7
P4
5
0
1
0

0:2 0:6et 0:2e2t

0:2  0:2e2t
0:2  0:4et 0:2e2t

0:5  0:5e2t
0:5 0:5e2t
0:5  0:5e2t

3
0:3  0:6et 0:3e2t
7
2t
7
0:3  0:3e
5
0:3 0:4et 0:3e2t

Then, the DP functional equation (8) is solved in

backward to obtain the optimal oil ordering size (in
million gallons). The results are summarized in Table 4.
According to Table 4, if for example the oil price at the
beginning of the planning horizon is low, then 18 gallons
is to be purchased at the beginning, which covers the
demands for both the rst and the second periods. Then,
the optimal oil ordering size in each of the last two
periods will be as equal as its expected demand, in all
states.
We also used Matlab to solve the problems up to 14
states for the unit production cost, analytically. Fig. 1
shows the computational times. Even for the higher
dimensions, since the Laplace transform and its inverse
transform are linear and the proposed algorithm is
polynomial, we do not have much problem in computation time. In this case, we can at least calculate fim t
numerically and then easily solve the DP functional
equation using the backward approach.
4. Conclusion
In this paper, we proposed a polynomial algorithm
based on semi-Markovian decision processes to solve the

ARTICLE IN PRESS
612

A. Azaron et al. / Int. J. Production Economics 120 (2009) 607612

optimal production scheduled for each period, in this new

problem.
The proposed model can also be extended to solving
the equipment replacement problem in dynamic setting.
Suppose that the unit procurement cost of facility is
changing over the time horizon according to a continuoustime Markov process. Then, the optimal strategy of
replacement can be easily found by extending the
proposed model in this paper. Hence the only difference
between the dynamic economic lot size model and the
standard equipment replacement model is the calculation
of the costs, that is the maintenance and procurement
costs.

350

Time (second)

300
250
200
150
100
50
0
0

5 6 7 8 9 10 11 12 13 14 15
Dimention Of Matrix

Acknowledgment

Fig. 1. Computational times versus the number of states.

This research is supported by Swedish Institute (SI).

dynamic lot sizing problem, in which the production and
the inventory costs of each period are concave, and the
unit production cost evolves according to a continuoustime Markov process over the planning horizon.
As explained, we need to consider continues-time
Markov process in our analysis, because we need the
memoryless property of evolving the unit production cost
over the time horizon to derive the DP functional Eq. (8).
The main computational difculty of the proposed
algorithm is to compute fim Dt by the inverse Laplace
transformation in Eq. (4). The usual method for solving
such equations is to approximate them by a discrete-time
analog or determine the inverse Laplace transform
numerically, in moderately large problems.
Apart from the two cases mentioned in Section 1, this
model can be applied in many practical production and
business cases for instance in global supply chains. Due to
the comparative advantages of different countries, nowadays rms are developing manufacturing and sourcing
strategies in a global supply chain. Material purchasing,
manufacturing and market often locate in different
regions. One challenge is then to deal with nancial risk.
Firms have an impressing need to take into consideration
of price and exchange rate uctuation in production
systems. The proposed model can be easily used to
manage sourcing decision at the strategic and tactical
level for purchasing decisions dealing with raw materials.
The proposed model can be extended to some more
general settings. For example, some other parameters of
the model, like the unit holding cost per period, might
also evolve according to continuous-time Markov processes. In this case, ht would be equal to ict+wt, where i is
the inventory carrying cost rate and wt is the unit storage
cost at the beginning of period t. It can be assumed that
the unit storage cost evolves according to another
continuous-time Markov process, which is independent
of the continuous-time Markov process corresponding to
the evolution of the unit production cost. There are two
state variables in this new problem. The rst is the unit
production cost, the same as that of the original problem,
and the second is the unit storage cost. Finally, the DP
functional equation (8) can be extended to obtain the

References
Aggarwal, A., Park, J.K., 1993. Improved algorithms for economic lot size
problems. Operations Research 41, 549571.
Boguslavskiy, J.A., 2005. Estimation of a matrix of intensities for model of
Markov process. Applied Mathematics and Computation 166,
282290.
Bookbinder, J.H., Tan, J.Y., 1988. Strategies for the probabilistic lot-sizing
problem with service-level constraints. Management Science 34,
10961108.
Brahimi, N., Dauzere-Peres, S., Najid, N.M., Nordli, A., 2006. Single item
lot sizing problems. European Journal of Operational Research 168,
116.
Fabian, T., Fisher, J.F., Sasieni, M.W., Yardieni, A., 1959. Purchasing raw
material on a uctuating market. Operations Research 7, 107122.
Federgruen, A., Tzur, M., 1991. A simple forward algorithm to solve
general dynamic lot sizing models with n periods O(n log n) or O(n)
time. Management Science 37, 909925.
Haugen, K.K., Lokketangen, A., Woodruff, D.L., 2001. Progressive hedging
as a meta-heuristic applied to stochastic lot-sizing. European Journal
of Operational Research 132, 116122.
Howard, R.A., 1971. Dynamic Probabilistic Systems. Wiley, New York.
Johnson, L.A., Montgomery, D.C., 1974. Operations Research in Production
Planning Scheduling and Inventory Control. Wiley, New York.
Karimi, B., Fatemi Ghomi, S.M.T., Wilson, J.M., 2003. The capacitated lot
sizing problem: a review of models and algorithms. Omega 31,
365378.
Kingsman, B.G., 1986. Purchasing raw materials with uncertain uctuating prices. European Journal of Operational Research 25, 358372.
Kingsman, B.G., 1969. Commodity purchasing. Operational Research
Quarterly 20, 5979.
Maes, J., Wassenhove, L.N., 1998. Multi-item single-level capacitated
dynamic lot sizing heuristics: a general review. Journal of the
Operational Research Society 39, 9911004.
Melo, R.M., 1996. Stochastic lot-sizing in production planning. Ph.D.
Thesis, Erasmus University, The Netherlands.
Silver, E., 1978. Inventory control under a probabilistic time-varying,
demand pattern. AIIE Transactions 10, 371379.
Sox, C.R., 1997. Dynamic lot sizing with random demand and nonstationary costs. Operations Research Letters 20, 155164.
Sox, C.R., Jackson, P.L., Bowman, A., Muckstadt, J.A., 1999. A review of the
stochastic lot scheduling problem. International Journal of Production Economics 62, 181200.
Wagelmans, A., Hoesel, S.V., Kolen, A., 1992. Economic lot sizing: an
O(n log n) algorithm that runs in linear time in the WagnerWhitin
case. Operations Research 40, 145156.
Wagner, H.M., Whitin, T.M., 1958. Dynamic version of the economic lot
size model. Management Science 5, 8996.
Zabel, E., 1964. Some generalizations of an inventory planning horizon
theorem. Management Science 10, 465471.
Zangwill, W.I., 1969. A backlogging model and a multi-echelon model of
a dynamic economic lot size production systema network
approach. Management Science 15, 506527.