This article was downloaded by: [130.75.172.

160] On: 11 March 2015, At: 09:47

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA

Transportation Science
Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org

Deterministic Order-Up-To Level Policies in an Inventory

Routing Problem
Luca Bertazzi, Giuseppe Paletta, M. Grazia Speranza,

To cite this article:

Luca Bertazzi, Giuseppe Paletta, M. Grazia Speranza, (2002) Deterministic Order-Up-To Level Policies in an Inventory Routing
Problem. Transportation Science 36(1):119-132. http://dx.doi.org/10.1287/trsc.36.1.119.573

Full terms and conditions of use: http://pubsonline.informs.org/page/terms-and-conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact permissions@informs.org.

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.

© 2002 INFORMS

Please scroll down for article—it is on subsequent pages

INFORMS is the largest professional society in the world for professionals in the fields of operations research, management
science, and analytics.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
 Deterministic Order-Up-To Level Policies
in an Inventory Routing Problem
Luca Bertazzi • Giuseppe Paletta • M. Grazia Speranza
Downloaded from informs.org by [130.75.172.160] on 11 March 2015, at 09:47 . For personal use only, all rights reserved.

Dip. di Metodi Quantitativi, Università di Brescia, Italy

Dip. di Economia Politica, Università della Calabria, Italy
Dip. di Metodi Quantitativi, Università di Brescia, Italy

W e consider a distribution problem in which a set of products has to be shipped from

a supplier to several retailers in a given time horizon. Shipments from the supplier to
the retailers are performed by a vehicle of given capacity and cost. Each retailer determines
a minimum and a maximum level of the inventory of each product, and each must be visited
before its inventory reaches the minimum level. Every time a retailer is visited, the quantity
of each product delivered by the supplier is such that the maximum level of the inventory
is reached at the retailer. The problem is to determine for each discrete time instant the
retailers to be visited and the route of the vehicle. Various objective functions corresponding
to different decision policies, and possibly to different decision makers, are considered. We
present a heuristic algorithm and compare the solutions obtained with the different objective
functions on a set of randomly generated problem instances.

Introduction and Griffin (1996). Three main management areas of

A company is a system with complex internal-
management problems and complex relations with
the suppliers and the customers. The internal man-
agement and the relations with the external world
are strongly interrelated, and correct management of
the company requires an understanding of these rela-
tions. In the last decade the importance of these rela-
tions has been widely recognized and the expression
“supply-chain management,” which emphasizes the
view of the company as part of the supply chain, has
become common. Sometimes the expression “coor-
dinated supply-chain management” is used with an
additional emphasis on the coordination among the
different components of the supply chain. The avail-
ability of data and information systems that derives
from the advances in technology and communication
systems has created the conditions for the coordina-
tion inside the supply chain.
On the modeling side, there is a strong need
for models that explicitly consider the coordination
issues. A review of models can be found in Thomas
the physical flow require coordination: production,
inventory, and transportation. An interesting intro-
duction to the problem of coordinating production
and transportation issues can be found in Hall and
Potts (2000). Sometimes, when the company is not
a factory, the production area may not exist. One
of the critical coordination problems that all compa-
nies face is the coordination of transportation issues
with inventory control issues. Typical examples are
given by the internal distribution systems, in which
the supplier and the retailers are different echelons of
a single company, and by external distribution sys-
tems, in which the supplier replenishes the retailers
with respect to a given service level. Models that
integrate vehicle routing with inventory control prob-
lems are called inventory-routing models (see Fed-
ergruen and Zipkin 1984 for the first optimization
model that combines inventory and routing costs and
Campbell et al. 1998 for a recent overview) and can
be classified as follows (see Bramel and Simchi-Levi
1997): single-period models with stochastic demand,

0041-1655/02/3601/0119$05.00 Transportation Science © 2002 INFORMS

1526-5447 electronic ISSN Vol. 36, No. 1, February 2002 pp. 119–132
 BERTAZZI, PALETTA, AND SPERANZA
Deterministic Order-Up-To Level Policies
multiperiod models with deterministic demand, and
infinite-horizon models with deterministic demand.
An example of single-period models with stochas-
Downloaded from informs.org by [130.75.172.160] on 11 March 2015, at 09:47 . For personal use only, all rights reserved.
tic demand can be found in Federgruen and Zipkin
(1984). In this model, the quantity of product to ship
to each retailer is determined on the basis of the level
of the inventory at the retailer. Then, the retailers
are assigned to the vehicles and the routes are deter-
mined. The multiperiod models with deterministic
demand are deterministic models in which several
shipments can be performed over a time horizon. In
these models, simpler models are often used as sub-
problems. For instance, in Dror and Ball (1987) single-
period models are used as subproblems, while in
Bertazzi et al. (1997) direct shipping models are used.
Finally, in the infinite-horizon models with determin-
istic demand the product is absorbed by each retailer
at a given constant rate and the problem is to deter-
mine an infinite-horizon shipping policy that mini-
mizes the sum of inventory and vehicle-routing costs.
Examples of these models can be found in Anily
and Federgruen (1990) and in Chan and Simchi-Levi
(1998).
We study a multiperiod model with deterministic
demand in which a set of products is shipped from
a common supplier to several retailers. Each prod-
uct is made available at the supplier and absorbed by
the retailers in a deterministic and time-varying way.
By deterministic way we mean that the quantity of
each product that will be made available at the sup-
plier and the quantity that will be absorbed by the
retailers are known in each discrete time instant. By
time-varying way we mean that the quantity of each
product made available and absorbed in each time
instant can be different from the one made available
and absorbed in a different time instant. The prod-
ucts are shipped from the supplier to the retailers by
a vehicle of given transportation capacity and cost.
Shipments can be performed only in the discrete time
instants (delivery time instants) that belong to a given
time horizon. For each product, a starting level of the
inventory is given both for the supplier and for each
retailer, and the level of the inventory at the end of the
time horizon can be different from the starting one.
Therefore, the problem is not periodic. Each retailer
120
determines a lower and an upper level of the inven-
tory of each product and can be visited several times
during the time horizon. Every time a retailer is vis-
ited the quantity delivered is such that the maximum
level of inventory is reached. This inventory policy
is inspired, in a deterministic setting, by the classical
order-up-to level policy, widely studied in inventory
theory (see Axsäter and Rosling 1994 for an overview
of inventory policies in multilevel systems). A unique
decision maker, typically, a logistics manager, has to
determine a shipping policy that minimizes the sum
of transportation costs and inventory costs both at the
supplier and at the retailers. A shipping policy con-
sists of determining for each delivery time instant the
set of retailers to visit, the quantity of each product
to ship to each retailer, and the route of the vehicle.
The constraints of the problem guarantee that in each
route the capacity of each vehicle is not exceeded and
that the level of the inventory both at the supplier
and at each retailer is never lower than the minimum
level. This problem, referred to as deterministic order-
up-to level inventory-routing problem, is obviously
NP-hard, because it reduces to the TSP in the class of
instances in which the time horizon is one, the inven-
tory costs are zero, the capacity of the vehicle is infi-
nite, and all the retailers need to be served.
The scope of this paper is twofold. On one hand,
we aim at solving the above problem. On the other
hand, we intend to study the impact of the objective
function on the problem solution. Because the prob-
lem is very complex and the exact solution would be
impractical, we propose a constructive heuristic algo-
rithm in which at each iteration a retailer is inserted
in the solution. For each retailer, the heuristic builds
a network to represent the incremental costs due to
the insertion of the retailer in the solution. The short-
est path of the network identifies a “min cost” policy
for adding the retailer to the solution. To evaluate
the impact of the objective function on the problem
solution, we consider different objective functions that
include some of the cost components only: the trans-
portation cost only, the inventory costs at the retailers
only, and the sum of the transportation cost and the
inventory cost at the supplier. In this way we obtain
three variants of the problem, never studied in the lit-
erature to the best of our knowledge. We try here to
Transportation Science/Vol. 36, No. 1, February 2002
 BERTAZZI, PALETTA, AND SPERANZA
Deterministic Order-Up-To Level Policies
give a partial answer to the question of how impor-
tant the coordination issue is, evaluating which of the
solutions is closer to the solution in which the total
Downloaded from informs.org by [130.75.172.160] on 11 March 2015, at 09:47 . For personal use only, all rights reserved.
cost is minimized, the impact of each cost on the total
cost, and how the objective function affects each type
of cost. To have a better understanding of the model,
in this paper we investigate the case of a single prod-
uct and a single vehicle. The model and the proposed
heuristic can be extended easily to the case of multi-
ple products and vehicles.
The paper is organized as follows. In §1 the prob-
lem is described and in §2 the heuristic algorithm is
presented. The variants of the problem, with different
objective functions, are introduced in §3. Finally, in §4,
the computational results obtained on randomly gen-
erated problem instances are shown and discussed.
1. Problem Description
We consider a logistic network in which a prod-
uct is shipped from a common supplier 0 to a set

=
1 2     n of retailers over a given time horizon
H . In each discrete time instant t ∈  =
1 2     H 
a quantity rit of the product is absorbed at the retailer
i ∈
and a quantity r0t is made available at the sup-
plier. Each retailer i ∈
defines a minimum level Li
and a maximum level Ui of the inventory of the prod-
uct. If retailer i is visited at time t, then the quantity xit
of product shipped to retailer i is such that the level
of the inventory in i reaches its maximum level Ui
(order-up-to level policy). More precisely, if we denote
by Iit the level of the inventory at retailer i at time t,
then xit is either equal to Ui − Iit if a shipment to i is
performed at time t, or equal to 0 otherwise.
Shipments from the supplier to the retailers can
be performed in each time instant t ∈  by a vehi-
cle with given capacity C, and routing is allowed. In
each route the vehicle visits the retailers that must be
served at the same time instant. We assume that any
combination of customers can be visited in a single
delivery time instant. The transportation cost cij from
i to j, with i j ∈
 =
∪
0, is known. Therefore,
given a route, the corresponding transportation cost
is simply obtained by summing up the cost of the arcs
that belong to the route.
Transportation Science/Vol. 36, No. 1, February 2002
An inventory cost is charged both at the supplier
and at the retailers. If we denote by hi the unit inven-
tory cost at node i ∈
 , then the total inventory cost
over the time horizon can be computed as follows.
At the supplier 0 the level of the inventory at time
t + 1 is given by the level at time t, plus the quantity
of product made available at time t, minus the total
quantity shipped to the retailers at time t, that is


Bt+1 = Bt + r0t − xit
i∈

where B0 (the starting level of the inventory) is given,
r00 = 0 and xi0 = 0, ∀ i ∈
. Therefore, the total inven-

tory cost at the supplier is t∈  h0 Bt , where   =  ∪

H + 1. At each retailer i ∈
, the level of the inven-
tory at time t + 1 is given by the level at time t plus
the quantity of product shipped from the supplier to
the retailer i at time t minus the quantity of product
absorbed at time t; that is,
Iit+1 = Iit + xit − rit
where Ii0 (the starting level of the inventory) is given,
and ri0 = xi0 = 0. Therefore, the total inventory cost at

the retailer i is t∈  hi Iit . The time instant H + 1 is
included in the computation of the inventory cost to
take into account the consequences of the operations
performed at time H .
The problem is to determine for each retailer i ∈

a set i of delivery time instants, and for each time
instant t ∈  a route t that visits all the retailers
served at time t such that the sum of the transporta-
tion cost and of the inventory cost both at the supplier
and at the retailers is minimized and the following
constraints are satisfied:
(1) Capacity Constraints. They guarantee that the
total quantity of the product loaded on the vehicle is
not greater than the transportation capacity:


xit ≤ C t ∈   (1)
i∈

(2) Stockout Constraints at the Supplier. They guaran-
tee that the level of the inventory Bt is nonnegative in
each time instant t ∈  :


xit ≤ Bt  t ∈   (2)
i∈

121
 BERTAZZI, PALETTA, AND SPERANZA
Deterministic Order-Up-To Level Policies
(3) Stockout Constraints at the Retailers. They guar-
antee that for each retailer i ∈
the level of the inven-
tory Iit in each time instant t ∈   is not lower than
the minimum level Li :
Downloaded from informs.org by [130.75.172.160] on 11 March 2015, at 09:47 . For personal use only, all rights reserved.
Iit ≥ Li  t ∈   i ∈
 (3)
Note that the problem in which several products
have to be shipped from the supplier to the retailers
can be handled easily by this model. It is sufficient to
replace each retailer i with a set of retailers, one for
each product absorbed by the retailer i, setting to zero
the transportation cost between nodes of the same
set and setting to the transportation cost between the
retailers the transportation cost between nodes of dif-
ferent sets.
2. A Heuristic Algorithm
We propose a two-step heuristic algorithm to solve
the order-up-to level inventory-routing problem. In
the algorithm the retailers are ranked in the nonde-
creasing order of the average number of time units
needed to consume the quantity Ui − Li and the retail-
ers with the same number of time units are ranked in
the nonincreasing order of Ui − Li . In the initialization
phase of the algorithm, referred to as Start, a feasible
solution of the problem is built by an iterative pro-
cedure that inserts a retailer at each iteration. When
retailer i is considered, a set of delivery time instants
is determined by solving a shortest-path problem on
an acyclic network in which every node is a possi-
ble delivery time instant (procedure Assign). Then, for
each of the selected delivery time instants, the retailer
is inserted in the route traveled by the vehicle by
applying the well-known rule of insertion at cheap-
est cost (procedure Insert). In the second phase of the
algorithm, referred to as Improve, the current solu-
tion is improved iteratively. At each iteration, a pair
of retailers is temporarily removed from the current
solution by using the procedure Remove. Then, each of
the removed retailers is inserted in the current solu-
tion as in the procedure Start. Finally, if this reduces
the total cost, then the solution is modified accord-
ingly. The second step of the algorithm is repeated as
long as an improvement in the total cost is reached.
The algorithm can be formally described as follows.
122
Heuristic Algorithm
0.— Sort the set of retailers
in the nondecreas-
ing order of the ratio between Ui − Li and
1 
H t∈ rit . If there exist retailers with the same
ratio, sort them in the nonincreasing order of
Ui − Li . Rename the retailers accordingly.
1. Start.
For s = 1 2     n
— Determine for the retailer s ∈
a set s of
delivery time instants by using the procedure
Assign.
— For each time instant t ∈ s insert the retailer
s in the route t traveled by the vehicle at
time t by using the procedure Insert.
2. Improve.
Let TC be the total cost of the solution.
(a) For s = 1 2     n
For i = n n − 1     1 and i = s
— t = t , ∀ t ∈  .
— Remove the retailer s from the routes
t by using the procedure Remove. Do
the same for the retailer i.
— Determine for i a new set of deliv-
ery time instants by using the proce-
dure Assign and insert the retailer in the
routes t by using the procedure Insert.
Do the same for the retailer s.
— Let TC be the cost of the obtained solu-
tion. If TC < TC, then adopt the new
solution.
(b) If a new solution has been adopted for at
least a pair of retailers, then go to (a).
2.1. Determining the Delivery Time Instants
In this section we describe the procedure Assign
which is used to determine a feasible set of delivery
time instants for each retailer s. This procedure works
on an acyclic network Gs s  s  s  s , in which
each element of the set s is a node that corresponds
to a discrete time instant between 0 and H + 1 and
each element askt of the set s is an arc that exists if no
stockout occurs in s whenever s is not visited between
k and t. Therefore, each path on the network between
0 and H + 1 represents a set of delivery time instants
Transportation Science/Vol. 36, No. 1, February 2002
 BERTAZZI, PALETTA, AND SPERANZA
Deterministic Order-Up-To Level Policies

for s that satisfy the stockout constraints (3). Each ele- to be inserted between two of the nodes of the route
s
ment qkt of the set s is a weight on the arc askt that t ; given that in the algorithm the rule of insertion at
represents the quantity of product to deliver at time t, cheapest cost is used (see, for instance, Rosenkrantz
s
et al. 1977), then s would be inserted between the
Downloaded from informs.org by [130.75.172.160] on 11 March 2015, at 09:47 . For personal use only, all rights reserved.

and each element pkt of the set s is a weight on the

s
arc akt used in order to determine the shortest path node i∗ ∈ t and its successor sui∗  ∈ t such that
between 0 and H + 1 on the network, i.e. a “good” set
of delivery time instants for s. Let us describe in more i∗ = arg min
cis + cssui − cisui 
i∈t
detail the sets s , s , and s . The set s has for ele-
therefore, the estimate c̃ts of the variation in the trans-
ments the arcs that satisfy the stockout constraints (3)
portation cost is ci∗ s + cssui∗  − ci∗ sui∗  . Obviously, if
at the retailer s; in particular, the arc as0t , 1 ≤ t ≤ H + 1,
 the capacity constraint (1) at time t is violated, then
exists if tj=1 rsj−1 ≤ Is0 − Ls and the arc askt , 1 ≤ k < t ≤
 c̃ts = +. The second component of the weight pkt s
is
H + 1, exists if tj=k+1 rsj−1 ≤ Us − Ls . Note that if the  s
the estimate Bkt of the variation in the inventory cost
arc as0H +1 exists, then a feasible policy is to not visit
at the supplier. This estimate is computed by consid-
the retailer during the time horizon. The set s is a set s
s ering that if a quantity qkt of product is shipped to
of weights in which each element qkt , associated with
s the retailer s at time t, then the level of the inventory
the arc akt , represents the quantity of product to ship s
Bt of the supplier decreases by a quantity qkt for all
to s at time t. Given that an order-up-to level policy
s the time instants between t + 1 and H + 1. Therefore,
is adopted, then the quantity qkt is such that the max- s = −h0 H + 1 − tq s . Obviously, if the stockout con-
B
imum level of the inventory Us is reached in s, that is, kt kt
s = +.
 straints (2) at the supplier are violated, then B
s
qkt = tj=k+1 rsj−1 for each arc askt with 1 ≤ k < t ≤ H and kt
 Finally, the third component of the weight pkt s
is the
q0ts = Us − Is0 + tj=1 rsj−1 for each arc as0t with 1 ≤ t ≤ H .  s
s estimate Ikt of the variation in the inventory cost at the
Note that qkH +1 , 0 ≤ k ≤ H is obviously equal to 0,
retailer s. This estimate is computed by considering
given that a shipment cannot be performed in H + 1.
that every time the retailer s is visited the level of the
Finally, the set s is a set of weights in which each
s inventory in s reaches its maximum value Us and that
element pkt associated with the arc askt represents the
it then decreases during the time on the basis of the
estimate of the variation in the total cost obtained by
quantities absorbed in s. Therefore, if a shipment to s
including in the current solution a visit of the retailer
is performed in t and the previous shipment has been
s at time t, given that the previous visit has been at
in k, then the estimate  I s of the variation in the inven-
time k. For each arc askt the weight pkt s
is computed t kt j
tory cost in s is hs j=k+1 Us − l=k+1 rsl−1 , while, if
on the basis of the partial solution generated by the
the shipment performed at time t is the first shipment
algorithm before applying this procedure, that is, on 
to s during the time horizon, then  Ikts = hs tj=1 Is0 −
the basis of the route t traveled by the vehicle at j s
time t ∈  , of the level of the inventory Bt at the sup- l=1 rsl−1 . In conclusion, the weight pkt associated to
s
plier, and of the level of inventory Ist at the retailer s the arc akt is
s
pkt s + I˜s 
= c̃ts + B
at time t ∈   . If this partial solution does not include kt kt

any retailer, then each route t , t ∈  , is empty and
the level of the inventory Bt at the supplier is equal to
the level obtained if no shipments occur up to time t,

that is, Bt = B0 + tj=1 r0j−1 , for each time instant t ∈   .
s
Given the partial solution, the weight pkt is computed
as the sum of three components. The first one, c̃ts , is
the estimate of the variation in the transportation cost
obtained if the retailer s is served at time t. This esti-
mate is 2c0s if no retailers are visited at time t in the
partial solution, i.e., if t = . Otherwise, c̃ts is com-
puted by taking into account that the retailer s has
Once the weight pkt s
is computed for each arc
askt∈ s , the procedure determines the shortest path
between 0 and H + 1 by using an algorithm for acyclic
networks (see, for instance, Hu 1982) to obtain a set
of delivery time instants for s that allows us to “min-
imize” the total cost. Finally, the procedure includes
in the set s of the selected delivery time instants for
s the intermediate nodes that belong to the shortest
path.
The procedure Assign can be formally described as
follows.

Transportation Science/Vol. 36, No. 1, February 2002 123

 BERTAZZI, PALETTA, AND SPERANZA
Deterministic Order-Up-To Level Policies
Procedure Assign
• Build the acyclic network Gs s  s  s  s .
• Determine the shortest path between 0 and H + 1
on the basis of the weights in s .
Downloaded from informs.org by [130.75.172.160] on 11 March 2015, at 09:47 . For personal use only, all rights reserved.
• If the length of the shortest path is finite, then
include in the set s the intermediate nodes that
belong to the shortest path; otherwise, the algorithm
is not able to find a feasible solution.
Note that the total quantity Qts of product shipped
to s up to each delivery time instant t, 1 ≤ t ≤ H , is
independent of the path between 0 and t selected on
the network Gs s  s  s  s . In fact, if t1  t2      tn
are the delivery time instants selected up to time t,
with t = tn , then the total quantity Qts shipped to s up
to time t is q0ts 1 + qts1 t2 + · · · + qtsn−1 tn , that is equal to Us −
 t 1
Is0 + tj=1 rsj−1 , as q0ts 1 = Us − Is0 + j=1 rsj−1 and qtsm tm+1 =
tm+1 s
j=1+tm rsj−1 , 1 ≤ m < n. Therefore, Qt is independent
of the selected delivery time instants t1  t2      tn−1 .
Moreover, note that the total quantity K s of prod-
uct shipped to s during the time horizon depends on
the last delivery time instant t̂ selected for s; in fact,
K s = Qt̂s . Therefore, K s can be different from the total

quantity t∈ rst of product absorbed from s during
the time horizon.
2.2. Inserting and Removing a Retailer
In this section we describe the procedures Insert and
Remove which are used during the algorithm to insert
and to remove, respectively, a retailer s from the route
t traveled by the vehicle at time t. Let us first con-
sider the procedure Insert. As described in the previ-
ous section, two different situations can happen when
the retailer s has to be inserted in the route t . The
first one happens when the route t is empty; in this
case, the insertion of the retailer gives a route com-
posed of only the arcs 0 s and s 0. The second one
happens when the route t already contains some
retailers; in this case, the rule of insertion at cheap-
est cost described in the previous section is used. The
procedure can be formally described as follows.
Procedure Insert
If t = , then t =
0 s 0.
Else
—Select the retailer i∗ such that
i∗ = arg min
cis + cssui − cisui 
i∈t
124
—Remove from t the arc i∗  sui∗ .
—Introduce in t the arcs i∗  s
and s sui∗ .
The insertion of the retailer s in the route t implies
an increase in the total quantity of the product loaded
on the vehicle equal to qkt s
, a variation in the trans-
portation cost equal to either 2c0s if the route t was
empty before inserting s or ci∗  s + cs sui∗  − ci∗  sui∗  oth-
erwise, and a reduction of the level of the inventory
at the supplier equal to
s
Bj+1 = Bj+1 − qkt  j = t     H 
Let us now describe the procedure Remove used
during the algorithm to remove the retailer s from
the route t . Two different situations can happen,
depending on the fact that s is the only retailer visited
in the route or not before removing it. Let prs be the
predecessor to the retailer s in the route t .
Procedure Remove
If t =
0 s 0, then remove the arcs 0 s
and s 0.
Else
—Remove the arcs prs s and s sus.
—Introduce the arc prs sus.
The decrease in the total quantity of the product
s
loaded on the vehicle is qkt , while the variation in the
transportation cost is −2c0s if only the retailer s was in
the route before to remove it, and is cprs sus −cprs s −
cs sus otherwise. Finally, the level of the inventory at
the supplier becomes
s
Bj+1 = Bj+1 + qkt  j = t     H 
3. Variants of the Problem
In this section we consider variants of the problem
described in §1, in which the aim of the decision
maker is not to minimize the overall cost but to min-
imize some of the cost components only. All the vari-
ants share the same set of data and the same set of
Transportation Science/Vol. 36, No. 1, February 2002
 BERTAZZI, PALETTA, AND SPERANZA
Deterministic Order-Up-To Level Policies
constraints of the problem described above. The only
difference is in the objective function.
In the first variant the aim of the decision maker is
Downloaded from informs.org by [130.75.172.160] on 11 March 2015, at 09:47 . For personal use only, all rights reserved.
to minimize the sum of the inventory cost at the sup-
plier and of the transportation cost only, without any
regard to the inventory costs at the retailers. This hap-
pens when the shipping policy is defined by the sup-
plier who organizes the routing and determines at its
discretion when to visit each retailer, under the con-
dition that the level of the inventory of each retailer
is never lower than the given minimum. In this case,
the supplier has a global view of his own internal
costs but does not care about the retailers’ costs. This
may be due to a lack of interest in the retailers’ costs
when supplier and retailers are different companies;
instead, it may be simply due to a lack of coordina-
tion in the case in which the supplier and the retailers
are part of the same company. This first variant is
NP-hard because it has the TSP as a particular prob-
lem and can be heuristically solved by a modified
version of the algorithm described in the previous sec-
tion. The only required change is in the procedure
s TC accordingly.
Assign where the weights pkt have to be computed
here as the sum of the estimate c̃ts of the variation in
s of the
the transportation cost and of the estimate B kt
variation in the inventory cost at the supplier only.
The total cost TC is computed accordingly.
A second interesting variant is obtained when the
objective function includes only the transportation
cost. This may happen when the transportation is out-
sourced and the agreement between the supplier and
the transportation company defines the basic princi-
ples of the service, namely that the stockout situa-
tions have to be avoided. However, the transportation
company can organize the deliveries on the basis of
its own costs. Then, the inventory costs at the sup-
plier and at the retailers follow from the shipping
policy of the transportation company. In the case in
which the whole system is owned by the same com-
pany, the focus on the transportation cost only may
again be simply due to a lack of global view and
to a lack of coordination. This variant remains NP-
hard since the routing problem remains to be solved
here and can be heuristically solved by modifying the
s
weights pkt of the procedure Assign. Here, the weights
Transportation Science/Vol. 36, No. 1, February 2002
include the estimate c̃ts of the variation in the trans-
portation cost only, and the total cost TC is computed
accordingly.
The third variant puts the focus on the retail-
ers’ costs and considers in the objective function the
inventory costs of the retailers only. The optimal solu-
tion here tends to serve each retailer when the inven-
tory reaches the minimum level, as stated by the clas-
sical order-up-to level policy. The only reason why
the optimal solution may happen to serve a retailer
before it is necessary is the capacity of the vehicle or
the stockout constraints at the supplier. Because an
optimal solution only determines the delivery time
instants and leaves the routing undetermined, we
consider the transportation cost as a second objec-
tive. This variant corresponds to the situation where
the retailers can determine the delivery times and
the routing is optimized afterwards. This hierarchi-
cal problem is again NP-hard and can be heuristi-
cally solved by computing in the procedure Assign the
weights pkts
as the estimate Ikts of the variation in the
inventory cost at the retailer s only and the total cost
The original problem and the above variants model
different organizational situations of a company, with
a different emphasis on what is important in the
objective of the decision. In the following section, we
compare the solutions obtained in the various cases,
trying to derive some conclusions about the impact
of the objective on the solution, both in terms of total
cost and in terms of single components of the total
cost.
4. Computational Results
The heuristic algorithm described in §2 has been
implemented in Fortran and used to solve the prob-
lems described in §§1 and 3 in a set of computa-
tional experiments. We first describe how the random
instances have been generated. Then we show a
series of computational results to analyze the solu-
tion obtained for the problem in which the total cost
is minimized and to compare the solutions obtained
by the different variants of the problem. Finally, we
computationally evaluate the quality of the solution
generated by the heuristic algorithm.
125
 BERTAZZI, PALETTA, AND SPERANZA
Deterministic Order-Up-To Level Policies

4.1. Generation of the Instances Two hundred and forty instances have been gener-
The instances have been generated on the basis of the ated on the basis of the data presented in §4.1; in
following data: particular, 10 instances have been randomly gener-
Number of retailers n: 50;
Downloaded from informs.org by [130.75.172.160] on 11 March 2015, at 09:47 . For personal use only, all rights reserved.

ated for each of the 24 combinations obtained on the

Time horizon H : 30; basis of the different values or intervals of the last
Quantity of product rit absorbed by retailer i at four parameters. The computational results allow us
time t: Constant over time, i.e., rit = ri , ∀ t, and ran- to evaluate the distribution of the total cost on the cost
domly generated as an integer number in the interval components (transportation cost, inventory cost at the
[10, 100]; retailers, and inventory cost at the supplier) and to
Quantity of product r0t made available at the sup- carry out a sensitivity analysis on the total cost. The

plier at time t: i∈
ri ; obtained results are shown in Tables 1 and 2 and in
Minimum level Li of the inventory at retailer i: Figure 1.
Randomly generated as an integer number in the Table 1 allows us to evaluate how the total cost is
interval [50, 150]; distributed on the different cost components in three
Maximum level Ui of the inventory at retailer i: pairs of classes of instances, where a pair of classes
Li + ri gi , where gi is randomly selected from the set covers all the tested instances. Each class is composed

2 3 5 6 10 15 30 and represents the number of of all the instances for which a given parameter has
time units needed in order to consume the quantity been generated in the same interval or has the same
U i − Li ; value. Three different parameters have been consid-
Starting level Ii0 of the inventory at the retailer i: ered: the transportation cost cij , the inventory cost at
U i − ri ; the retailers hi , and the inventory cost at the supplier
Starting level B0 of the inventory at the supplier: h0 . For each of the parameters, two intervals or values

i∈
Ui − Li ; are considered: [0, 500] and [0, 1000] for the trans-
Inventory cost at retailer i ∈
, hi : Randomly gen-
portation cost, [0.1, 0.5] and [0.6, 1] for the inventory
erated in the intervals [0.1, 0.5] and [0.6, 1];
cost at the retailers, and, finally, 0.3 and 0.8 for the
Inventory cost at the supplier h0 : 0.3 and 0.8;
  inventory cost at the supplier. In the following, the
Transportation capacity C: i∈
ri , 2 i∈
ri , and
 first interval or value of each of the parameters will
3 i∈
ri ;  be called “low” and the second “high.” The table is
Transportation cost cij :  xi − xj 2 + yi − yj 2 ,
organized as follows. Each row corresponds to a class
where the points xi  yi  and xj  yj  are obtained by of instances; the first column gives the parameter that
randomly generating each coordinate as an integer defines the class, the second one the interval in which
number in the interval [0, 500] and in the interval the parameter has been generated or its value, and
[0, 1000]. the remaining three columns show the correspond-
In all cases, random selections have been performed
ing average transportation cost, inventory cost at the
in accordance with a uniform distribution. The com-
putations have been carried out on an Intel Pentium
II personal computer. Table 1 Average Costs
Other experiments have been carried out on other
Average Costs
sets of instances, such as instances with Li = 0 and ri
randomly generated as an integer number in the inter- Parameters Transportation Inv. Retailers Inv. Supplier
val [5, 25]. Because the results obtained were similar, cij [0, 500] 357608 3655402 1465967
we do not report them. cij [0, 1000] 676215 3630848 1527730
hi [0.1, 0.5] 551851 2241609 1056990
4.2. Computational Analysis of the Solution hi [0.6, 1] 481972 5044641 1936707
We now focus on the original problem, referred to h0 0.3 482038 3234416 1161151
h0 0.8 551785 4051834 1832546
as IS + IR + T , in which the total cost is minimized.

126 Transportation Science/Vol. 36, No. 1, February 2002

 BERTAZZI, PALETTA, AND SPERANZA
Deterministic Order-Up-To Level Policies
Downloaded from informs.org by [130.75.172.160] on 11 March 2015, at 09:47 . For personal use only, all rights reserved.
Figure 1
retailers, and inventory cost at the supplier, respec-
tively. In each class of instances the inventory cost
at the retailers is the main part of the total cost, fol-
lowed by the inventory cost at the supplier and then
by the transportation cost. In the first two rows of
the table the average costs obtained in the classes of
instances with low and high transportation cost are
compared. The instances with high transportation cost
have slightly lower inventory cost at the retailers but
higher inventory cost at the supplier. This is an effect
that can be explained by observing that an increase of
the transportation cost tends to reduce the number of
times a retailer is served and, therefore, to reduce the
transportation cost and to increase the inventory cost
at the supplier. Rows 3 and 4 allow us to compare
the average costs obtained in the classes of instances
with low and high inventory cost at the retailers. The
results show that the instances with high inventory
cost at the retailers have lower transportation cost,
but higher inventory cost at the supplier. This is due
to the order-up-to level policy; in fact, contrary to
the classical inventory models, here the inventory cost
at the retailers is minimized by serving the retailers
rarely over time. Finally, in rows 5 and 6 the aver-
age costs obtained in the classes of instances with low
and high inventory cost at the supplier are compared.
The results show that the instances with high inven-
tory cost at the supplier have higher transportation
cost and higher inventory cost at the retailers. This is
Transportation Science/Vol. 36, No. 1, February 2002
due to the fact that the supplier tends to ship more
frequently.
Figure 1 allows us to evaluate the impact of each
cost component on the total cost in the classes of
instances described above. The figure is composed
of three figures: The first one (1a) compares the classes
of instances with low and high transportation cost, the
second one (1b) the classes of instances with low and
high inventory cost at the retailers, and the third one
(1c) the classes of instances with low and high inven-
tory cost at the supplier. In each figure the comparison
between the two classes of instances is performed by
computing, for each class, the percentage of each cost
component in the total cost. The first figure shows
that in the class of instances with high transportation
cost the impact of the inventory cost at the retailers
on the total cost is significantly lower with respect
to the class with low transportation cost; the second
figure shows that in the class of instances with high
inventory cost at the retailers the impact of the trans-
portation cost on the total cost is significantly lower
with respect to the class with low inventory cost at
the retailers; finally, the third figure shows that the
impact of the inventory cost at the retailers and of the
transportation cost does not significantly change in
the classes of instances with low and high inventory
cost at the supplier.
In Table 2 a sensitivity analysis on the total cost in a
given instance is presented. Each row of the table cor-
responds to a parameter of the instance and shows the
127
 BERTAZZI, PALETTA, AND SPERANZA
Deterministic Order-Up-To Level Policies

Table 2 Sensitivity Analysis of the retailers. This behavior reflects the fact that the dif-
Total Cost ferent cost components have a significantly different
 − 04  + 04 importance in the total cost. The last two rows of the
table show the variations obtained in the total cost by
Downloaded from informs.org by [130.75.172.160] on 11 March 2015, at 09:47 . For personal use only, all rights reserved.

cij −195 191

varying by about 80% and 26.7% the minimum level
hi −6169 3719
h0 −4092 1636 of the inventory at the retailers and the transporta-
Li −503 497 tion capacity, respectively. The results show that, as
C 012 −005 expected, these parameters do not significantly affect
the total cost and that the total cost increases when
percent variation of the total cost obtained by reduc- the former parameter increases and when the latter
ing and increasing, respectively, the parameter of a one decreases.
given percentage of its original value. The parame-
4.3. Computational Comparison of the Solutions
ters of the instance have been generated as in §4.1,
Obtained with Different Objective Functions
with the exception of the ones that are involved in the
We now compare the solutions obtained for the prob-
sensitivity analysis, namely the transportation cost,
lem IS + IR + T with the solutions obtained for its
the inventory cost at the retailers, the inventory cost
variants, namely, problem IS + T , in which the sum
at the supplier, the minimum level of the inventory
of the inventory cost at the supplier and of the trans-
at the retailers and, finally, the transportation capac-
portation cost only is minimized; problem T , in which
ity. These parameters have been obtained as follows:
the transportation cost is minimized only; and prob-
First, the transportation costs have been generated in
lem IR, in which first the inventory cost at the retail-
the interval [0, 500], the inventory cost at the retailers
ers and then the transportation cost are minimized.
and at the supplier are set equal to 1, the minimum
The analysis has been carried out on the instances
level of the inventory at the retailers is generated in
generated in §4.2.
the interval [50, 150] and, finally, the transportation
 The obtained results are shown in Tables 3–5 and
capacity is set equal to i ri . Then, each of these val-
in Figure 2. The first row of Table 3 presents the aver-
ues has been multiplied by a factor (, which takes
age total cost of the different problems on all the
value 0.5 for all the parameters, with the exception
instances. The results show that the average total cost
of the transportation capacity for which ( takes value
of problem T , in which only the transportation cost is
1.5. The total cost of this instance is 634,014. The sensi-
minimized, is closer to the average total cost of prob-
tivity analysis has been carried out by computing the
lem IS + IR + T than the total cost generated by prob-
percent increase of the cost obtained by changing, for
lem IS + T and by problem IR. The last three rows of
one parameter at a time, the value of ( with respect to
Table 3 show the importance of each type of cost in
the cost of the basic instance. In particular, ( has been
the total cost.
set equal to ( ± 04, corresponding to an increase and
While Table 3 is especially interesting when read
a decrease, respectively, of 80% for the transportation
column by column, Table 4 has to be read row by row.
costs, the inventory costs, and the minimum level of
It gives, for each type of cost (total cost, transportation
the inventory at the retailers, and of about 26.7% for
cost, inventory cost at the retailers, and, finally, inven-
the transportation capacity. The first three rows of the tory cost at the supplier), the average percent increase
table show the variations obtained in the total cost by
varying the transportation cost, the inventory cost at Table 3 Average Costs
the retailers, and, finally, the inventory cost at the sup-
IS + IR + T IR T IS + T
plier by 80% of their original value, respectively. The
variations in the transportation cost have a very low Total cost 5656883 6286500 6064238 6165471
impact on the total cost, while the contrary happens Transportation cost 516911 664567 455388 625712
for the same variations in the inventory cost at the Inv. cost retailers 3643124 3041648 3537898 4758684
Inv. cost supplier 1496848 2580285 2070952 781075
supplier and, in particular, in the inventory cost at the

128 Transportation Science/Vol. 36, No. 1, February 2002

 BERTAZZI, PALETTA, AND SPERANZA
Deterministic Order-Up-To Level Policies
Table 4 Percent Increase of the Costs
IS + IR + T IR T IS + T
Total cost 01 140 86
Downloaded from informs.org by [130.75.172.160] on 11 March 2015, at 09:47 . For personal use only, all rights reserved.
Transportation cost 145 467 00
Inv. cost retailers 254 00 162
Inv. cost supplier 1065 2222 1581
of the cost of each problem with respect to the mini-
mum corresponding cost obtained by solving the four
problems. The different problems, with substantially
different objective functions, have obtained total costs
which are not very different from each other, with a
maximum increase of 14% corresponding to problem
IR. The percentage of 0.1% corresponding to problem
IS + IR + T is due to the fact that in a small sub-
set of the tested instances the best solution has been
obtained by a variant of the problem instead of by
the problem IS + IR + T . This is because the problems
are heuristically solved. The last three rows of Table 4
allow us to evaluate how the goal of the decision
maker affects each type of cost. The problems IR and
T have obtained, as expected, the minimum inven-
tory cost at the retailers and the minimum transporta-
tion cost, respectively, on each instance. Moreover, the
problem IS + T has obtained the minimum inventory
cost at the supplier on almost all the instances. Con-
trary to the situation of the total cost, which is similar
in the different problems, the different cost compo-
nents give substantially different contributions to the
total cost in the different problems.
Table 5 gives the average number of visits and
the average quantity of product delivered during the
time horizon obtained in the different problems. The
results show that problem IR gives the minimum
number of visits, followed by problem T , then by
problem IS + IR + T and, finally, by problem IS +
T . The reason can be found in the trade-off exist-
ing between the inventory cost at the retailers and
Table 5 Average Number of Visits and Delivery Quantity
IS + IR + T IR T IS + T
Number of visits 4611 3002 3480
Delivery quantity 708291 628009 693575
Transportation Science/Vol. 36, No. 1, February 2002
the transportation cost on one hand and the inven-
tory cost at the supplier on the other hand. When the
inventory cost at the supplier is not included in the
84 objective function (see problems IR and T ) the retail-
400
ers are more rarely visited than in the other cases.
560
01
The problems are ranked in the same order when the
total quantity of product delivered during the time
horizon is considered. Because in the problem formu-
lation no condition is imposed on the inventory at the
end of the horizon, the different problems create dif-
ferent ending situations. Since a visit brings the level
of the inventory of the visited retailers to the maxi-
mum level, when the visits are frequent the level of
the inventory at the retailers tends to be high at the
end of the time horizon. This implies that, as the ini-
tial inventory situation and the quantities absorbed by
the retailers on the horizon are identical for all prob-
lems, in the problems with more frequent visits the
total delivered quantity is higher.
In Figure 2 we consider four classes of instances
obtained on the basis of the combinations of “low”
and “high” values of inventory cost at the supplier
and of the inventory cost at the retailers, as defined in
the previous section. Figures 2a and 2d show that in
the instances in which the inventory cost at the retail-
ers and the inventory cost at the supplier are both low
or high, the results are very similar to the ones shown
in Table 3. Figure 2b shows that when the inventory
cost at the retailers is low and the inventory cost at the
supplier is high the problem IS + T and the problem
IS + IR + T generate, as expected, very similar solu-
tions. Moreover, as expected, the inventory cost at the
supplier is a major part of the total cost. A different
situation can be observed in Figure 2c.
Figure 3 shows the percent distribution of the total
cost on the various cost components obtained by the
problems IR T , and IS + T on the class of instances
with “low” transportation cost (in the interval [0, 500])
with respect to the distribution obtained in the class of
instances with “high” transportation cost (in the inter-
val [0, 1000]). In each of these problems the increase
in the transportation cost implies a reduction of the
impact of the inventory cost at the retailers, as shown
in Figure 1a for the problem IS + IR + T .
Finally, in Table 6 we present a sensitivity anal-
6878 ysis on the number of visits obtained in the prob-
796758
lems IS + IR + T and IS + T in a given instance. Each
129
 BERTAZZI, PALETTA, AND SPERANZA
Deterministic Order-Up-To Level Policies
Downloaded from informs.org by [130.75.172.160] on 11 March 2015, at 09:47 . For personal use only, all rights reserved.

Figure 2

Figure 3

130 Transportation Science/Vol. 36, No. 1, February 2002

 BERTAZZI, PALETTA, AND SPERANZA
Deterministic Order-Up-To Level Policies

Table 6 Sensitivity Analysis of the Number of Visits each discrete time instant all the retailers on the basis
IS + IR + T IS + T
of an optimal route. The second policy, referred to as
Latest, is to visit in each delivery time instant t the
 − 04  + 04  − 04  + 04
Downloaded from informs.org by [130.75.172.160] on 11 March 2015, at 09:47 . For personal use only, all rights reserved.

set of retailers that will have stockout at time t + 1 if

cij 402 337 1500 873 not served at time t. The retailers served at time t are
hi 1022 297 1185 1185
visited on the basis of an optimal route.
h0 299 1006 634 1402
Li 325 325 1185 1185
Eight instances have been generated according to
C 328 343 1227 975 the different intervals and values of the parameters
described in §4.1, with the only exception of the
transportation capacity C, which we assumed to
row corresponds to a parameter of the instance and be large enough to guarantee a feasible solution
gives the number of visits obtained in the different for the policies Every and Latest, and the trans-
problems by decreasing and then increasing, respec- portation costs cij , which have been generated as
tively, the parameter of a given percentage of its origi- 
 xi − xj 2 + yi − yj 2 + 05, since the optimal solu-
nal value. The sensitivity analysis has been carried out
as described in §4.2 on the same instance. The num- tion of the TSP problems has been obtained by
ber of visits obtained in the original instance is 325 applying an existing exact algorithm in which
for the problem IS + IR + T and 1,185 for the problem the transportation costs are computed as indicated
IS + T . The table clearly says that the number of visits above.
strongly depends on the values of the parameters and The computational results are shown in Table 7.
on the objective function. In particular, the number The first three columns give the intervals in which
of visits in the problem IS + T is substantially greater the transportation costs cij and the inventory costs
than in the problem IS + IR + T . The reason is that, as at the retailers hi have been generated and the value
shown in particular in the second row of the table, the of the inventory cost at the supplier h0 , respectively;
number of visits decreases when the inventory cost the last two columns give the percent increase error of
at the retailers increases. This is due to the fact that the optimal cost of the policies Every and Latest with
the application of the order-up-to level policy implies respect to the cost of the solution generated by the
that the inventory cost at the retailers is minimized heuristic algorithm.
when the retailers are served rarely over time. More- The results show that the solution obtained by the
over, as expected, the number of visits increases when heuristic algorithm always outperforms the optimal
the transportation cost decreases and when the inven- solution of the two intuitive policies. In fact, the pol-
tory cost at the supplier increases, has no variations icy Every, which requires the optimal solution of one
with respect to the minimum level of the inventory at TSP per instance, gives an average error of about
the retailers, and has relatively small variations with 14% with respect to the cost of the heuristic solution,
respect to the transportation capacity.
Table 7 Percentage Increase Error of Two Intuitive Policies
4.4. Performance Evaluation
cij hi h0 Every Latest
Finally, we show the results obtained in a computa-
tional experiment carried out to evaluate the quality 0 500 01 05 03 1385 189
0 500 01 05 08 250 1419
of the solution generated by the heuristic algorithm in
0 500 06 1 03 2668 002
the problem IS + IR + T . Since the complexity of the 0 500 06 1 08 460 163
problem precludes the possibility of solving it opti- 0 1000 01 05 03 1742 494
mally and no lower bounds are known, we compare 0 1000 01 05 08 745 1308
the cost of the solution generated by the heuristic 0 1000 06 1 03 3227 042
0 1000 06 1 08 738 140
algorithm with the optimal cost of two intuitive poli-
1402 470
cies. The first policy, referred to as Every, is to visit in

Transportation Science/Vol. 36, No. 1, February 2002 131

 BERTAZZI, PALETTA, AND SPERANZA
Deterministic Order-Up-To Level Policies
while the policy Latest, which gives an average error
of about 5%, requires the solution of 30 exact TSPs
per instance. Note that the maximum error generated
Downloaded from informs.org by [130.75.172.160] on 11 March 2015, at 09:47 . For personal use only, all rights reserved.
by the policy Every is obtained when the inventory
cost at the retailers and the transportation cost are
large and the inventory cost at the supplier is small;
instead, the maximum error generated by the pol-
icy Latest is obtained when the inventory cost at the
retailers and the transportation cost are small and the
inventory cost at the supplier is large.
Conclusions
We studied a problem in which a deterministic order-
up-to level policy is adopted for the minimization
of the costs in a distribution system. We considered
different problems obtained by different goals of the
decision maker and we used a heuristic algorithm
to solve each problem and to compare the solutions
obtained by the different problems on randomly gen-
erated problem instances. The results show how rel-
evant the goal is on the obtained solution. A future
research issue consists of studying the production pol-
icy of the supplier and its influence on the system
inventory.
Acknowledgments
The authors wish to thank two anonymous referees for their useful
comments and suggestions.
Received: December 2000; revision received: September 2001; accepted: September 2001.
132
References
Anily, S., A. Federgruen. 1990. One warehouse multiple retailer sys-
tems with vehicle routing costs. Management Sci. 36 92–114.
Axsäter, S., K. Rosling. 1994. Multi-level production-inventory con-
trol: Material requirements planning or reorder point policies?
Euro. J. Oper. Res. 75 405–412.
Bertazzi, L., M. G. Speranza, W. Ukovich. 1997. Minimization of
logistic costs with given frequencies. Transportation Res. B 31
327–340.
Bramel, J., D. Simchi-Levi. 1997. The Logic of Logistics. Springer,
New York.
Campbell, A., L. Clarke, A. Kleywegt, M. Savelsbergh. 1998. The
inventory routing problem. T. G. Crainic, G. Laporte, eds.
Fleet Management and Logistics. Kluwer Academic Publishers,
London, UK. 95–113.
Chan, L. M. A., D. Simchi-Levi. 1998. Probabilistic analyses and
algorithms for three-level distribution systems. Management
Sci. 44 1562–1576.
Dror, M., M. Ball. 1987. Inventory/routing: Reduction from an
annual to a short-period problem. Naval Res. Logist. Quart. 34
891–905.
Federgruen, A., P. Zipkin. 1984. A combined vehicle routing and
inventory allocation problem. Oper. Res. 32 1019–1032.
Hall, N. G., C. N. Potts. 2000. Supply chain scheduling: Batching
and delivery. Working paper, University of Southampton.
Hu, T. C. 1982. Combinatorial Algorithms. Addison-Wesley, Reading,
MA.
Rosenkrantz, D. J., R. E. Stearns, P. M. Lewis. 1977. An analysis
of several heuristics for the traveling salesman problem. SIAM
J. Comput. 6 563–581.
Thomas, D. J., P. M. Griffin. 1996. Coordinated supply chain man-
agement. Euro. J. Oper. Res. 94 1–15.
Transportation Science/Vol. 36, No. 1, February 2002