Journal of the Operational Research Society (2008) 59, 44 --53  2008 Operational Research Society Ltd.

Ltd. All rights reserved. 0160-5682/08 $30.00

www.palgrave-journals.com/jors

Raw material inventory solution in iron and steel

industry using Lagrangian relaxation
L Tang1∗ , G Liu1 and J Liu2
1 Northeastern University, Shenyang, China; and 2 Loughborough University, Leicestershire, UK
Because raw material inventories can compensate for the unexpected demand fluctuations as well as variability
in the replenishment process, large inventories are maintained to ensure the continuity of production. But
storing them increases the inventory cost. Therefore, how to best balance raw material inventory and production
demands under capacity constraints has become the serious subject faced by most large steel companies. This
paper studies the raw material inventory problem abstracted from the production of Shanghai Baoshan Iron
and Steel Complex (Baosteel) in a theoretical light, which provides the scientific foundation for practical
applications. Safety stock and safety lead time are first introduced to absorb the random fluctuations. Then, a
constrained optimization model is formulated to minimize the total cost attributed to raw material inventories.
By using a synergistic combination of Lagrangian relaxation, ordinal introduction of constraints algorithm
and heuristic algorithms, high-quality plans are obtained in a timely manner.
Journal of the Operational Research Society (2008) 59, 44 – 53. doi:10.1057/palgrave.jors.2602335
Published online 1 November 2006

Keywords: steel production; logistics; inventory; safety stock; Lagrangian relaxation

Introduction profitable steel enterprises in the world enjoying international

competence. Its annual production consists of over 20 million
Facing strong competition from both overseas steel indus-
tonnes of steel, and its auto sheet accounts for more than
tries and domestic mini-mills which take advantage of newer
60% of the domestic market share. Therefore, each year it
technology and lower costs, integrated steel companies exist
must consume large quantities of raw materials, for example,
within an economic environment suffused with strong compe-
about 17 960 000 tonnes of iron ore per year. Obviously, the
tition, shrinking markets and lagging long-term profitability.
replenishment of raw materials in Baosteel is a matter of great
This highly competitive global steel market makes the old
significance.
topic, cost reduction, generalized widely. Since carrying
After an investigation that dealt with the issue of managing
inventories for various reasons can cost the steel companies
raw material inventories in Baosteel has been made, we find
anywhere from 20 to 40% of their value a year, inventory
it convenient to describe the entire process of raw material
management, which has been called and quite fairly so, one
inventory management in Baosteel using Figure 1. When the
of the classics in operation management literature during
demand plan of raw materials submitted by the production
the last few decades, has recently been the focus of atten-
department is received, the raw material control centre, which
tion. Since raw material inventories account for the largest
is responsible for raw material inventory management, will
part of the inventories in the iron and steel industry, mini-
set corresponding inventory policies to provide guiding prin-
mizing them in a quantitative manner is of vital economic
ciples for the purchasing department. The main purpose of
significance. Thus, improving the management level of raw
our research is to help the raw material control centre deter-
material inventory is clearly of extreme importance.
mine the optimal order interval and inventory level for each
The raw material inventory problem studied in this paper
material.
is to determine the fixed order size and fixed interval of the
However, generating an efficient raw material inventory
replenishment process for each material based on the mini-
plan for a large steel company is not an easy task for several
mization of the total cost attributed to raw material inven-
reasons. First, the raw material inventory problems derived
tories. This inventory problem arises from the production of
from iron and steel production are new and important research
BaoSteel. BaoSteel is not only the largest and most advanced
topics, which are seldom considered in previous researches.
iron and steel enterprise in China, but also one of the most
Although similar problems have been studied and solved,
the effort required to find better plans for problems with
∗ Correspondence: L Tang, 135#, The Logistics Institute, Northeastern different backgrounds may vary considerably. Therefore, they
University, Shenyang, Liaoning 110004, China. still deserve studying carefully. Second, the problem is char-
E-mail: [email protected] acterized by large production demands, high inventory costs,
 L Tang et al—Raw material inventory solution 45

main raw auxiliary raw discharging chute

material dock material dock for trucks

auxiliary raw
coal yard ore yard
material yard

coke-making shop
chutes

power station
crushing and
screening

blending yard

iron making steel making slabbing-

sintering shop lime shop
shop shop blooming mill

Figure 2 Schematic diagram of the comprehensive stock yard.

main wharf

Figure 1 The institution of the replenishment plan for raw

materials. coal yard E
auxiliary wharf

coal yard A
no stockout, limited inventory capacities, finite storage time
for some kinds, etc. These features that are often disregarded coal yard B

add extra difficulties to the inventory problem under our

coal yard C
consideration. Moreover, in order to ensure the continuity of
production, large inventories usually have to be maintained to coal yard D
compensate for the unexpected demand fluctuations as well
blending yard A
as variability in the replenishment process. In view of these
facts, how to best balance raw material inventory and produc- blending yard B
tion demands under capacity constraints in the iron and steel
auxiliary raw
industry becomes a difficult task. material yard A
For the convenience of research, a logical grouping auxiliary raw
material yard B
concept is introduced in this paper to divide raw materials
auxiliary raw
according to their own properties, which is different from material yard C
physical grouping that concerns actual storage positions. This
ore yard D
grouping principle aims to centralize the scattered inventory
capacities of stock yards. It is very helpful to determine the discharging chute for trucks ore yard E

total capacity of these stock yards for each raw material group
ore yard F
which consists of raw materials with common properties. To
understand it better, the schematic diagrams about the compre- ore yard G
hensive stock yard of Baosteel are given in Figures 2–4.
ore yard H
This comprehensive stock yard, including stock yard phase I
and II and stock yard phase III, is in charge of the central- carried in by trucks miscellaneous
auxiliary yard I
ized management and handling of raw materials, aiming at
smooth supplying to the complex. According to the principle Figure 3 Schematic diagram of stock yard phase I and II.
of logical grouping, one imaginary auxiliary raw material
yard is defined to include all the real auxiliary raw material
yards in the Baosteel comprehensive stock yard. The inven- In the remaining part of the paper, we first review the related
tory capacity of this imaginary yard is determined to be the works as well as the current state of research concerning
total capacities of the real yards included in it. The materials inventory. A novel mathematical programming model is then
stored in those real yards compose the ‘Auxiliary Raw formulated to solve the inventory problem. The solution
Material Group’. Since out-of-stock (OOS) will bring enor- methodology and the computational results are presented in
mous losses to Baosteel, it is not allowable here. the following part. The final section concludes our study.
46 Journal of the Operational Research Society Vol. 59, No. 1

the order-up-to level and the indifference inventory level.

wharf phase III
Hoshino (1996) proposed two theoretical criteria allowing for
selection from fixed-size ordering and fixed-interval ordering
policies and the choice between the pull-type ordering and
main & auxiliary dry shed coal A the push-type ordering policies. Ishii and Imori (1996)
wharf phase I & II
dry shed coal B
proposed an effective production ordering system for two-
item, two-stage, capacity-constraint production and inventory
coal yard I systems, which reduced fluctuations in the total workload
and inventory levels. Maia and Qassim (1999) presented an
coal yard H
analytical solution for an optimized model that determined
coal yard G when it was preferable to incur inventory or opportunity
cost. Ganeshan (1999) proposed a near-optimal (s, Q)-type
coal yard F
inventory policy for a production/distribution network with
blending yard D
multiple suppliers replenishing a central warehouse, which
in turn distributed to a large number of retailers. Nechval and
blending yard C Nechval (1999) investigated the effect of estimation risk on
auxiliary raw
the simplest of inventory problems. Matheus and Gelders
material yard R (2000) considered an inventory problem subject to a proba-
auxiliary raw
material yard Q
bilistic non-unit sized demand pattern and proposed an exact
and an approximate reorder point calculation method for the
auxiliary raw
material yard P (R, Q) inventory policy. Relph and Barrar (2003) argued that
auxiliary raw excess was important because there was evidence that even
material yard O
in well- managed businesses a significant proportion of the
ore yard J inventory existed in excess at any given time.
ore yard K
Problem description and formulation
ore yard L
The EOQ model is a classical approach to determining order
ore yard M and production lot sizes as well as setting order intervals for
a single product with an unconstrained capacity. Countless
ore yard N
modifications have been developed in order to adapt the model
to various real-world situations. The approach has two major
Figure 4 Schematic diagram of stock yard phase III. advantages: one is its simplicity and the other is the flatness
of the function around the curve’s minimum. Our raw mate-
Literature review rial inventory management is a multi-item problem requiring
Various inventories exist at practically every stage of the considerations of inventory capacities, demand rates, set up
production process as raw materials, work-in-process semi- costs as well as inventory costs. In general, this problem
finished or finished goods. During the last few decades, can be viewed as an economic lot size problem. The classical
inventory control has received considerable attention in liter- economic lot size model (ELSM) making simple tradeoffs
ature. Therefore, a large collection of results is available. In between ordering and inventory costs is applicable to uncon-
spite of that, few of the previous research, however, dealt strained cases. To make it more realistic, some capacity
with the raw material inventory problem derived from iron restrictions must be imposed. Inspired by the idea of ELSM,
and steel production. Bowman (1956) and Manne (1958) in this paper, we establish a constrained optimization model
solved lot sizing problems by linear programming, respec- to solve the inventory management problem arising from
tively. Wagner and Whitin (1958) proposed the WW-model Baosteel taking into account all the features summarized in
for the uncapacitated lot sizing problem and gave a corre- Table 1. The objective is to determine the optimal order
sponding polynomial-time algorithm. Newhart et al (1993) interval and inventory level for each raw material.
quantified the effects of inventory required for locating parts
Assumptions
of the supply chain in different geographic areas by using
a two-phased approach. Chiang and Gutierrez (1996) anal- The problem characteristics described above form the basis
ysed a periodic review inventory system in which there are for developing the mathematical models. The modelling
two modes of re-supply, namely a regular mode and an assumptions are outlined below.
emergency mode, within the framework of an order-up-to-
R inventory control policy. They simultaneously proposed (1) The distribution functions of demands and order lead
procedures to determine the two policy parameters, that is, times are given.
 L Tang et al—Raw material inventory solution 47

Table 1 Features of the raw material inventory the safety lead time of material i is presented on condition
problem in Baosteel that the distribution function, mathematical expectation and
• Too many categories of raw materials are stored in the depots variance of its order lead time are given as Fi (t) = P(T
t),
• The demand for the products is relatively stable
i2 and i2
2
, respectively. The safety lead time of material i
• The demand for each raw material is large and relatively can be determined as follows.
stable
• The lead time needed for replenishment is relatively long
• OOS is not allowable (1) Calculate ti1 , which is deduced from P(T 
i2 + ti1 ) =
• The raw material inventory is costly i1 . Let i1 be a small positive real number which is
• Some raw materials have storage lives predetermined by the planners of Baosteel.
(2) Calculate ti2 , the minimum lead time needed for Baosteel
to prepare for the replenishment.
(3) Calculate ti3 , the shortest time needed for purchasing the
(2) All the raw materials are bought in tonnes. material from others.
(3) The raw materials are consumed in large quantities. (4) The safety lead time of raw material i, t0i , is defined as
(4) The demand for each raw material is relatively stable. t0i = max{ti1 , ti2 + ti3 }.
(5) The decision horizon is suggested to be one year by the
planners of Baosteel. Notations
(6) The time horizon is divided into T time units.
(7) OOS is not allowable. Parameters
T decision horizon
Safety stock and safety lead time N set of raw materials stored in Baosteel comprehensive
stock yard, index i
To absorb the random fluctuations in demands and transporta- Di demand rate of raw material i, i ∈ N
tion times, safety stock and safety lead time are introduced in hi inventory holding cost per unit of material i, that is,
this section. In the following part, we will give the estimation the capital occupied by keeping one unit of material
methods for them. i in stock, i ∈ N
Ki fixed set-up cost incurred every time Baosteel places
an order of material i, i ∈ N
Estimation of safety stock
vl i storage life, that is, the longest storage time, of mate-
A method for determining the safety stock of each raw mate- rial i, i ∈ N
rial is proposed in this part. Here material i is used as an Ik raw material group k (according to logical grouping),
example. Other materials could be treated in the same way. 1
k
n; I1 ∪ I2 ∪ · · · ∪ In = N ; Ii ∩ I j = , i, j =
If the distribution function, mathematical expectation and 1, . . . , n, i = j
variance of the demand for material i are given as Fi (s) = Vk the largest inventory capacity of Baosteel comprehen-
P(S
s),
i1 and i1
2
, respectively, the safety stock of mate- sive stock yard for the raw materials included in Ik ,
rial i can be determined as follows. 1
k
n.

(1) Calculate si1 , which is deduced from P(S 
i1 + si1 ) =

Decision variables
i1 . Let i1 be a small positive real number which is
predetermined by the planners of Baosteel. Qi fixed maximum inventory level of raw material i,
(2) Calculate si2 , the minimum safety stock level specified i∈N
by these planners. ti fixed order interval of raw material i, i ∈ N
(3) Calculate si3 =
i1 ti2 , where ti2 is the shortest time needed mi the number of the orders placed in the decision
for purchasing material i from the nearest supplier. horizon for raw material i, i ∈ N
(4) The safety stock of raw material i, s0i , is defined as
s0i = max{si1 , si2 , si3 }. The decision model
In this section, the decision model for determining the optimal
Estimation of safety lead time order interval and inventory level of each raw material is
formulated. The objective is to minimize the manufacturer’s
Since using safety stock to absorb the random fluctuations total cost attributed to raw materials inventories. The details
in transportation times will lead to overstock, in this section, of the model are presented below.
we introduce the concept of safety lead time and give the
estimation method for it. Safety lead time is the additional

1
time, beyond the expected lead time, needed to guarantee Minimize C, with C ≡ Ki mi + h i (Q i +s0i ) (1)
receipt of the materials ordered. The method used to determine i∈N i∈N
2
48 Journal of the Operational Research Society Vol. 59, No. 1

Lagrangian relaxation
The problem formulation presented previously shows that
only constraints (4) result in the coupling of different raw
materials. By relaxing these constraints, the original problem
can be decomposed into smaller and easier subproblems, each
for one material. For that reason, we form the following
relaxed problem by introducing constraints (4) to the objec-
tive function through Lagrangian multipliers {uk}:
LR
Figure 5 Inventory level of material i as a function of time.

1
Minimize CLR , with CLR ≡ Ki mi + h i (Q i + s0i )
i∈N i∈N
2
 
Subject to
n

+ uk  Q i − Vk  (9)
m i ti = T, i∈N (2) k=1 i∈lk

Q i = s0i + ti Di , i ∈ N (3) Subject to constraints (2), (3), (5), (6), (7), (8).

Let {{m i∗ }, {Q i∗ }, {ti∗ }} be an optimal solution of LR for
Q i
Vk , 1
k
n (4) a given set of multipliers. The dual problem of LR is to
i∈Ik maximize the dual function C D (u):
LD
Q i
vl i Di , i∈N (5)
Maximize C D (u), with
0 < ti
T, i∈N (6) C D (u) = CLR ({u k }, {{m i∗ }, {Q i∗ }, {ti∗ }}) (10)

m i is a positive integer, i∈N (7) Subject to constraints (2), (3), (5), (6), (7), (8) and

u k  0, 1
k
n (11)
Q i > 0, i∈N (8)
Here u is a vector of non-negative Lagrangian multipliers
The two terms in objective function (1) represent setup cost
with elements {u k }. Since constraints (2), (3) and (5)–(8) are
and inventory holding cost, respectively. Constraints (2) reveal
independent of each other, the relaxed problem can be decom-
the relationship among m i , ti and T . Constraints (3) indicate
posed into the following material-level subproblems:
that the inventory level of each material is determined by its
LRi
safety stock and the consumption during its order interval.
Constraints (4) are the inventory capacity restrictions. The Minimize CLR (i), with
maximum inventory level, Q i , is forced to be not greater than  
1
the consumption of material i during its longest storage time, CLR (i) ≡ K i m i + h i (Q i +s0i )+ u k Q i (12)
vl i , by (5). Constraints (6)–(8) define the value ranges of the 2 k=(i)

variables. The relationship between Q i and ti can be seen

Subject to constraints (2), (3), (5), (6), (7), (8).
clearly from Figure 5.
Here (i) denotes the raw material group that material i
belongs to.
Solution methodology
Ordinal introduction of constraints algorithm for
Owing to the integer decision variables included, the inventory subproblems (L R i )
problem described above is difficult to solve directly. As we
know, Lagrangian relaxation is one of the most widely used Based on constraints (2) and (3), m i and Q i can be represented
and efficient techniques for performing constrained optimiza- using ti :
tion. The central idea of Lagrangian relaxation is to formu-
m i = T /ti and Q i = s0i + ti Di
late a decomposable relaxed problem which is easier to solve
than the original one. Based on the ideas of Luh et al (1990), Constraints (5) can then be expressed as ti
vl i − s0i /Di .
Diaby et al (1992) and Tang et al (2002), we use Lagrangian This can be combined with constraints (6) into the following
relaxation as the main part of our solution methodology. For constraints.
a minimization problem, the solution to the Lagrangian dual 
problem provides a lower bound, while each feasible solution s0i
0 < ti
min vl i − ,T
provides an upper bound. Di
 L Tang et al—Raw material inventory solution 49

Note that when ti > 0, constraints (8) are automatically ordinal introduction of constraints algorithm. The steps of the
satisfied. With these manipulations, problem (LRi ) becomes algorithm to determine ti are detailed below.

Minimize CLR (i), with Step 1: Let y  (t)=0, calculate the optimal solution of uncon-

T 1 strained problem (13) and denote it as t1∗ .
CLR (i) ≡ K i + Di h i + u k Di ti Step 2: If 0 < t1∗
min{vl i − s0i /Di , T }, set t2∗ = t1∗ ; other-
ti 2 i∈Ik
wise set t2∗ = min{vl i − s0i /Di , T }. Here t2∗ stands
+h i s0i + u k s0i (13) for the optimal solution of constrained problem (13)
i∈Ik and (14).
Step 3: If T /t2∗ is an integer, let t ∗ = t2∗ ; otherwise, set c1 =
Subject to
T /t2∗ , t1 = T /c1 , t2 = T /(c1 + 1). Calculate y(t1 )
 and y(t2 ). If y(t1 ) < y(t2 ), set t ∗ = t1 ; otherwise set
s0i
0 < ti
min vl i − ,T (14) t ∗ = t2 . t ∗ obtained here is the optimal solution of
Di
constrained problem (13)–(15).
T
is an integer (15) Construction of a feasible solution to the original problem
ti
The solution to the relaxed problem is generally infeasible
The objective function takes the form of for the original problem because the inventory capacity
constraints (4) might be violated. To construct a feasible
y(t) = a1 /t + a2 t + b
solution, a heuristic approach is proposed. The detailed
where algorithmic steps of this heuristics are described as follows.

1 Step 1: Search for the first raw material group k, 1
k
n,

a1 = K i T ; a2 = Di · h i + u k ·Di ;
2 i∈Ik which violates the corresponding inventory capacity
constraint. If there exists no such group, set flag =
b = h i · s0i + u k ·s0i 1 and go to Step 6; otherwise calculate sum =
i∈Ik
i∈Ik (Q i − s0i ), sum1 = i∈Ik Q i .
Obviously, a1 , a2 and b are all positive. It is easy to see, by Step 2: Calculate sub = sum1 − Vk . If sub > sum, this
checking its first- and second-order derivatives, that y(t) is a heuristics is of no effect for this problem, set flag=0
strictly convex function. A graph of function y for a1 = 100, and go to Step 6.
a2 = 70 and b2 = 29 is shown in Figure 6 as an example. Step 3: Create a material sequence for Ik by indexing
Therefore, the subproblem is to minimize a strictly convex the materials in the ascending order of K i −
function of one variable in a fixed interval with additional 0.5Di h i (T /m i − T /(m i + 1)), i ∈ Ik .
integer constraints, which can be solved easily using the Step 4: Take the first unselected material i from the above
sequence. Set a = Q i , m i = m i + 1, update ti = T /m i
and Q i = s0i + Di ti , then calculate sub = sub −
(a − Q i ).
Step 5: If sub > 0, go to Step 3; otherwise, go back to
Step 1.
Step 6: Stop. If f lag = 1, {(m i , Q i , ti )|i ∈ N } obtained
above defines a feasible replenishment plan satis-
fying all the constraints. If f lag = 0, the manufac-
turer must reduce the safety stocks of the materials
included in Ik or increase the value of Vk .

Updating Lagrangian multipliers

It is well known that subgradient algorithm is commonly used
to solve Lagrangian dual problems. In order to solve the dual
problem with respect to (10), the algorithm is adopted to
update the vector of Lagrangian multipliers, u, by

u m+1 = Max{0, u m + t m m (u m )}, 1
k
n (16)

Figure 6 Plot of y as a strictly convex function of t at a1 = 100, where t m is the step size at the mth iteration and m (u mk )=
a2 = 70 and b2 = 29. i∈Ik Q i − Vk , 1
k
n is the subgradient component related
50 Journal of the Operational Research Society Vol. 59, No. 1

to the kth stock yard. The step size t m is given by let f lag = 0; otherwise change the order of the
materials in Ik and go back to Step 1.
t m = m (C U − C L )/m (u m )2 , Step 10: Stop. If f lag = 1, update up bound = C1 , let
m+1 = m exp(−0.52m ), {(m i(1) , Q i(1) , ti(1) )|i ∈ N } → {(m i , Q i , ti )|i ∈ N };
Otherwise this heuristics is not effective for this
m = 0.15 [m/20] (17) problem.
where C U is the best objective value found so far and C L is
equal to the value of C D (u) at the mth iteration. The algorithm Phase II:
terminates, if one of the following conditions is satisfied. Step 1: Take the current best feasible solution as the initial
solution, that is, set m i(2) = m i , Q i(2) = Q i , ti(2) = ti ,
(1) (C U − C L )/C L < , where  > 0 is a very small number; for each i ∈ N .
(2) m > the maximum iteration number given by the user. Step 2: Search for the first unselected material l ∈ N which
satisfies the conditions: T /(m l(2) − 1)
min{vl i −
Improvement method (s0l /Dl ), T } and m l(2)  2; if there exist no such
When 10 consecutive iterations fail to improve the current materials, stop.
optimal solution, a three-phase improvement method will be Step 3: Search for the raw material group k that l belongs
activated. Let {(m 1i , Q 1i , t1i )|i ∈ N }, {(m i , Q i , ti )|i ∈ N }, to. If we can obtain a better feasible solution 
only
up bound and lo bound denote the relaxed solution, the by decreasing the value of m l(2) to m l(2 ) , let C2
current best feasible solution, upper bound and lower bound, denote the corresponding objective  
value,
update
respectively. The detailed description of this method is the up bound = C2 , let {(m i(2 ) , Q i(2 ) , ti(2 ) |i ∈ N } →
outlined below. {(m i , Q i , ti )|i ∈ N } and then go back to Step 1.
Step 4: Search for the material r that satisfies the conditions:
Phase I: r ∈ Ik and r = l. If we can get a better feasible
solution by both decreasing the value of m l(2) to
Step 1: Take the solution obtained from solving the relaxed 
m l(2 ) and increasing the value of m r(2) to m r(2 ) , let

problem as the initial solution, that is, set m i(1) =
m 1i , Q i(1) = Q 1i , ti(1) = t1i , for each i ∈ N . C2 denote the corresponding objective 
value,

update
Step 2: Search for the first raw material group k, up bound = C2 , let {(m i(2 ) , Q i(2 ) , ti(2 ) |i ∈ N } →
1
k
n, which violates the corresponding inven- {(m i , Q i , ti )|i ∈ N }. Finally go to Step 1.
tory capacity constraint. If there exists no such
group, set f lag = 1 and go to Step 7; other- Phase III:
wise calculate sum = i∈Ik (Q i(1) − s0i ), sum1 = Step 1: Take the current best feasible solution as the initial
(1)
i∈Ik Q i . solution, that is, set m i(3) = m i , Q i(3) = Q i , ti(3) = ti ,
Step 3: Calculate sub = sum1 − Vk . If sub > sum, let for each i ∈ N .
f lag = 0 and go to Step 10. Step 2: Search for the first unselected material l ∈ N ; if
Step 4: Search for the first unselected material l ∈ Ik which there exists no such material, stop.
satisfies the conditions: Q l(1) − s0l  Dl (T /m l(1) − Step 3: Search for the raw material group k that material l
T /(m l(1) + 1)) and m l(1) + 1 < T . If there exists no belongs to.
such material, set f lag = 0 and go to Step 10. Step 4: Search for the material r that satisfies the condi-
Step 5: Set m l(1) = m l(1) + 1, tl(1) = T /m l(1) , Q l(1) = s0,l + tions: r ∈ Ik and r = l. If we can get a better
Dl · tl(1) , sub = sub −
Q l , where
Q l represents feasible solution by both increasing the value of m l(3)

to m l(3 ) and decreasing the value of m r(3) to m r(3 ) , let

the decrement of Q l(1) .
Step 6: If sub > 0 and material l still satisfies the condition C3 denote the corresponding 
objective
 
value, update
stated in Step 4, go back to Step 5; If sub > 0 but up bound = C3 , let {(m i(3 ) , Q i(3 ) , ti(3 ) )|i ∈ N } →
material l cannot meet the condition, go back to {(m i , Q i , ti )|i ∈ N }. Finally go to Step 1.
Step 4; If sub
0, go back to Step 2.
Step 7: Calculate the objective value C1 = i∈N K i m i(1) +
(1) Computational results
i∈N 1/2h i (Q i + s0i ), if C 1 < up bound, go to
Step 10; otherwise set f lag = 0. Generation of problem instances To study the performance
Step 8: Form set S composed of all the raw material groups of the algorithms described above, two types of experiments
that violate inventory capacity constraints, based are performed across a range of test problems. In the first
on the relaxed solution. group of the tests, the number of raw material groups is
Step 9: Search for raw material group k ∈ S in which there fixed, and in the second group, the number of raw materials
exist unused exchanges. If there is no such group, is fixed. To generate representative problem instances, we
 L Tang et al—Raw material inventory solution 51

Table 2 Average duality gap (%) of our method

Number of raw Number of raw material groups
materials
2 3 5 8 10
4 0.016664
6 0.737173 0.096731
8 0.013334 0.040305
10 0.009391 0.078186 0.476428
20 0.004203 0.004371 0.094818 0.812137 0.722313
30 0.003634 0.005042 0.504785 0.455958 0.514772
40 0.003676 0.004381 0.023636 0.493823 0.369827
50 0.003889 0.009497 0.008296 0.212416 0.788269
60 0.004054 0.003773 0.004827 0.336061 0.331341
70 0.003267 0.003783 0.031582 0.098539 0.147511
80 0.003597 0.004074 0.049304 0.161772 0.106913
90 0.003757 0.004349 0.004813 0.106769 0.232951
100 0.004077 0.003907 0.004521 0.005471 0.040730
110 0.003526 0.023673 0.004109 0.052178 0.035269
120 0.003329 0.003908 0.004793 0.004641 0.007379
130 0.003471 0.004540 0.003941 0.005099 0.024059
140 0.003757 0.003871 0.003800 0.005493 0.057286
150 0.003760 0.003186 0.079443 0.072584 0.054176
160 0.003135 0.003575 0.068496 0.013035 0.011364
170 0.003265 0.003091 0.004296 0.004849 0.114604
180 0.003730 0.003458 0.063535 0.024556 0.259759
190 0.003401 0.003532 0.004213 0.013025 0.017685
200 0.003420 0.003078 0.004280 0.007403 0.011008

analyse the actual production data from Baosteel. A day is Testing results The algorithms have been implemented using
taken as the basic time unit. The decision horizon is set to Visual C++ and the experiments are carried out on a Pentium-
be one year. The number of raw material groups is chosen IV 2.4 GHz PC. Because Lagrangian relaxation cannot guar-
from n ∈ {2, 3, 5, 8, 10}, while the number of raw materials antee optimal solutions, the relative dualilty gap (C UB −
changes from 4 to 200. For each combination of the number of C LB )/C LB was used as a measure of solution optimality,
materials and the number of raw material groups, 10 instances where C UB is the upper bound to the original problem and
are randomly generated, therefore resulting in a total of 1030 C LB is the lower bound. The maximum iteration number 4000
test problems used in this research. The parameters of raw is imposed on the stopping criterion. As mentioned earlier,
materials are generated based on the actual production data for each problem structure we randomly generate and solve
from Baosteel. Let G denote the set composed of raw mate- 10 problem instances. All the values shown in the tables
rials whose practical data have been obtained. In what follows, below represent the average of performance measures and
material j is used as an example to illustrate the generation running times for these 10 instances of the corresponding
of those parameters. problem structure. The results of these experiments are shown
in Tables 2 and 3. It can be observed that the average dual-
(1) One material i is chosen from G randomly for material j. ity gap obtained from the Lagrangian relaxation algorithm is
(2) The demand rate of material j is randomly generated 0.090648% in about 25 s of computation time. From the com-
from [0.8, 3] times of Di . putational results, we also have the following observations.
(3) The inventory holding cost per unit of material j is
generated uniformly from [0.5 ∗ h i , 1.5 ∗ h i ]. (1) When the number of raw material groups is fixed, the
(4) The safety stock and the fixed set-up cost of material j is duality gap decreases but the computation time increases
randomly generated from [0.5, 1.5] times of s0i and K i , as the number of raw materials increases.
respectively. (2) When the number of raw materials is fixed, the duality
(5) To be closer to the practical production, the storage life gap and the computational time increase as the number
of material j is randomly generated from a uniform of raw material groups augments.
distribution [60, 365].
(6) The largest inventory capacity of Baosteel comprehensive
Conclusions
stock yard for the materials included in each raw material
group is randomly generated from [1.05, 1.1] times of Since the steel industry is resource intensive, effective inven-
their total safety stocks. tory management in this industry assumes even greater
52 Journal of the Operational Research Society Vol. 59, No. 1

Table 3 Average running time (s) of our method

Number of raw Number of raw material groups
materials
2 3 5 8 10
4 0.160800
6 0.982600 0.271900
8 0.543700 0.312500
10 0.157900 3.307600 14.292101
20 0.128000 0.260800 9.601601 38.751501 44.201398
30 2.307900 0.693800 11.720500 27.578000 44.878201
40 0.545300 2.326500 9.861000 17.979601 33.843802
50 1.361100 15.449899 7.347000 22.758000 25.220099
60 0.515600 3.718800 3.909700 21.403398 35.228403
70 2.646900 1.203100 17.971901 16.618900 31.476599
80 2.240400 3.779800 15.758002 38.637302 35.910999
90 3.940701 10.295199 9.584401 36.884598 26.009302
100 12.762801 2.112400 3.364100 18.796701 28.615698
110 15.373401 12.764200 8.557701 53.254700 63.378003
120 17.917201 5.374899 8.740400 13.289001 27.435999
130 31.987500 13.568600 8.134500 20.856200 33.331198
140 41.909302 8.482699 2.472000 45.303296 33.029602
150 30.010999 12.267200 39.753000 37.211002 56.367004
160 10.154501 50.643796 54.053101 162.504700 74.253101
170 15.651401 10.577800 14.679500 19.775000 60.817102
180 38.757898 9.690700 81.965704 51.501703 54.181201
190 39.093698 22.845200 4.315500 34.504398 69.220398
200 231.968896 25.743701 17.623300 26.140601 60.149902

importance. This research studied the raw material inven- National Natural Science Foundation of China (Grant No. 60274049 and
tory problem arising from Baosteel and formulated it as a Grant No. 60674084).
mixed integer programming model. A Lagrangian relaxation
algorithm was developed to solve it. In this algorithm, a
References
‘separatable’ relaxed problem was constructed by relaxing
inventory capacity constraints. The subproblems derived from Bowman H (1956). Production scheduling by the transportation
this relaxed problem were solved optimally by the ordinal method of lp. Opns Res 4: 1001–1003.
introduction of constraints algorithm, while the Lagrangian Chiang C and Gutierrez GJ (1996). A periodic review inventory system
with two supply modes. Eur J Opl Res 94: 527–547.
multipliers were iteratively updated along a subgradient
Diaby M, Bahl HC, Karwan MH and Zionts S (1992). A Lagrangean
direction. Numerical results indicated that our algorithms relaxation approach for very-large-scale capacitated lot-sizing.
generated high-quality solutions in a short computation time. Mngt Sci 38: 1329–1340.
Although this study is only a theoretical research, it provides Ganeshan R (1999). Managing supply chain inventories: A multiple
the scientific foundation for practical applications. Future retailer, one warehouse, multiple supplier model. Int J Prod Econ
59: 341–354.
efforts may be focused on the following lines of research.
Hoshino K (1996). Criterion for choosing ordering policies between
fixed-size and fixed-interval pull-type and push-type. Int J Prod
(1) Modifying the model to describe the inventory problems Econ 44: 91–95.
in other industries. Ishii K and Imori S (1996). A production ordering system for two-item,
two-stage, capacity-constraint production and inventory model.
(2) Extending the model to solve the multi-echelon inventory Int J Prod Econ 44: 119–128.
problems in the iron and steel industry. Luh PB, Hoitomt DJ and Pattipati KR (1990). Schedule generation
(3) Studying inventory control problems from the perspective and reconfiguration for parallel machines. IEEE Trans R&A 6(6):
of supply chain management, such as VMI. 687–696.
Maia LOA and Qassim RY (1999). Minimum cost safety stocks for
frequent delivery manufacturing. Int J Prod Econ 62: 233–236.
Acknowledgements — We would like to thank the Production Manufac- Manne AS (1958). Programming of economic lot sizes. Mngt Sci
turing Center in Baosteel for providing a lot of production information and 4(2): 115–135.
data. We also thank the anonymous referees for the valuable suggestions. Matheus P and Gelders L (2000). The (R, Q) inventory policy subject
This research is partly supported by National Natural Science Founda- to a compound Poisson demand pattern. Int J Prod Econ 68:
tion for Distinguished Young Scholars of China (Grant No. 70425003), 307–317.
 L Tang et al—Raw material inventory solution 53

Nechval NA and Nechval KN (1999). Applications of invariance to Tang LX, Luh PB, Liu JY and Fang L (2002). Steel-making process
estimation of safety stock levels in inventory model. Comput Ind scheduling using Lagrangian relaxation. Int J Prod Res 40(1):
Eng 37: 247–250. 55–70.
Newhart DD, Stott KL and Vasko FJ (1993). Consolidating product Wagner HM and Whitin TM (1958). A dynamic version of the
sizes to minimize inventory levels for a multi-stage production and economic lot size model. Mngt Sci 5: 89–96.
distribution system. J Opl Res Soc 44: 637–644.
Relph G and Barrar P (2003). Overage inventory-how does it occur Received June 2005;
and why is it important? Int J Prod Econ 81–82: 163–171. accepted August 2006 after four revisions