You are on page 1of 19

Available online at www.sciencedirect.

com

Transportation Research Part E 44 (2008) 665683


www.elsevier.com/locate/tre

Integrated inventory and transportation mode selection:


A service parts logistics system
Erhan Kutanoglu *, Divi Lohiya
Graduate Program in Operations Research and Industrial Engineering, Department of Mechanical Engineering,
The University of Texas at Austin, United States
Received 9 August 2006; received in revised form 4 January 2007; accepted 21 February 2007

Abstract
We present an optimization-based model to gain insights into the integrated inventory and transportation problem for a
single-echelon, multi-facility service parts logistics system with time-based service level constraints. As an optimization
goal we minimize the relevant inventory and transportation costs while ensuring that service constraints are met. The
model builds on stochastic base-stock inventory model and integrates it with transportation options and service responsiveness that can be achieved using alternate modes (namely slow, medium and fast). The results obtained through dierent
networks show that signicant benets can be obtained from transportation mode and inventory integration.
 2007 Elsevier Ltd. All rights reserved.
Keywords: Inventory management; Transportation mode selection; Service parts logistics; Integer programming

1. Introduction
The availability of technically advanced systems like heavily automated production systems, military systems, medical equipments and computer systems increasingly aects daily operations as they play an important role in society. Downtime of critical equipments due to part failures may have serious consequences, e.g.
in terms of loss of production, ineective military missions, or quality reduction in health care. Various measures are therefore taken to reduce the amount of system downtime, such as providing system redundancy,
providing appropriate preventative maintenance and eective corrective maintenance. Especially with respect
to the latter, it is essential to have fast supply of service parts, where we dene a service part as the part needed
to replace the failed part.
In todays marketplace, when selecting suppliers of technically advanced systems, high quality after-sales
service through availability of spare parts and quick response time has become important criteria. A survey
by Cohen et al. (1997) on 14 companies with service logistics operations indicated that after-sales service revenues are equal to 30% of the product sales and the service parts inventories equal to 8.75% of the value of the
*

Corresponding author. Tel.: +1 512 232 7194; fax: +1 512 232 1489.
E-mail address: erhank@mail.utexas.edu (E. Kutanoglu).

1366-5545/$ - see front matter  2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.tre.2007.02.001

666

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

product sales. The results of this survey indicated following characteristics and trends common in the service
delivery organizations:

There are large number of service parts to be stocked, varying between 2500 and 300,000.
The cost of these service parts is increasing due to increasing complexity and modularity of the products.
The service parts have high obsolescence rate caused by short product life cycles.
The fraction of slow moving service parts is large and increasing. This is caused by increased system customization and design improvements. Average inventory turnover for such parts equals .87 parts per year,
hence there are many parts with demand rates less than one per year.

In the realm of service part logistics, a service provider establishes relations with its customers through service agreements that extend over a period, usually in years. These agreements typically specify the level of service as well as the time window within which the service will be provided. Customers agreements may involve
multiple customer locations, hence an overall service level to be achieved across multiple locations. A service
provider therefore faces a problem of determining optimal stocking levels of service parts while providing the
demanded parts to the customers within the specied time windows. The overall service parts logistics problem
therefore brings forth several tradeos as insucient stock can lead to extended customer system downtime
and hence the lost customer goodwill and sometimes heavy penalties, while maintaining excessive service parts
can lead to large inventory carrying costs. The time dimension which is usually called response time adds
another unique feature to the problem, which is the interaction between inventory and transportation. Since
the response time is a function of distance to the customer, which in turn depends on which facility in the network the demand is satised from and which transportation mode is used to satisfy the demand, the transportation decisions and associated costs are aected by inventory decisions.
Service parts are often supplied via a multi-echelon distribution network. One reason to have a multi-echelon network is the need to achieve quick response time and the need for stock centralization to reduce holding
costs. There is a trend however to reduce the number of echelons and the number of locations per echelon in
order to reduce xed location costs and service parts obsolescence costs. This striving for an ecient network
is facilitated by stocking essential parts close to the customers and by using the fast transportation modes.
However, when multiple modes with dierent costs and speeds (say fast, medium, slow) or competing alternatives for transportation are available, the integrated problem becomes rather unique in the sense that there
are many opportunities to investigate for total cost savings and the overall eciency.
In this paper, our objective is to develop a tactical optimization-based model to gain insights into the integrated inventory/transportation problem for a single-echelon multi-facility service parts logistic system with
time-based service level constraints. We view the use of this model at the tactical level, in which service levels
and demands of individual customer locations are aggregated according to overall service agreements to control the size and the granularity of the model. In this model, the service levels are dened for a group of customers (which in turn could be a collection of individual customer sites). As an optimization goal, we
minimize the relevant inventory, and normal and emergency transportation costs. The time-based service level
is dened as the portion of total customer demand that is satised within a specied period of time (service time
window), and the model seeks the minimum-cost solution that satises the target service levels for each service
region. (Note that in this tactical model, aggregation may lead to solutions in which some customers are not
satised within the time window due to the transportation mode assignment and/or stocking level. The model
tries to keep the demand portion unsatised within the time window under control.) Finally, note that the
model is capable of handling multiple modes in the overall system (typically more than 2). However, due to
the cost and service structure inherent in the problem, the choices for individual customers eectively boil down
to selecting between the least costly mode that delivers parts within the time window and the least costly mode
that does not deliver within the time window, although these two can be dierent across dierent customers.
The paper is organized as follows: Section 2 discusses the literature on inventory management for service
parts with specic focus on base-stock models for lost sales and integrated inventory and transportation models with service level considerations. In Section 3, we formulate the problem and present a detailed description
of the model. The experimental design used to test the model and the computational results are discussed in
Sections 4 and 5, respectively. Section 6 lists some extensions and improvements possible on the current model.

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

667

2. Literature review
We focus our attention to the literature on inventory policies used for spare/service parts management, and
models that seek to integrate transportation mode decisions into inventory policies.
Optimization of inventory systems to nd stock levels of service parts has been a subject of research for
many years. Of special interest and assumed in this report are high cost, low demand, critical parts. Due to
their nature, a base-stock policy (also called one-for-one replenishment), usually denoted by the (S  1, S)
model where S represents the base-stock level, is an appropriate replenishment policy. Here, the base-stock
level S is dened as the desired maximum number of parts or the inventory stocking level for the part. It is
also common to assume demand to have Poisson distribution, which is quite accurate for real service parts
systems. Early papers such as Feeney and Sherbrooke (1966) considered (S  1, S) continuous review
inventory policies and derived steady state probabilities for various state variables of the system. The
famous METRIC model developed by Sherbrooke (1968) derived the stock levels for a multi-echelon
system with multiple parts, with complete backlogging for stock-outs. Sherbrooke and others later
improved and generalized the method to consider more complex issues (see, e.g. Muckstadt, 1973; Muckstadt and Thomas, 1980; Sherbrooke, 1986). Muckstadt (2005) is a reference for inventory research in
service parts logistics. The literature on service parts logistics inventory management is extensive, but other
selected service parts-related studies include Cohen et al. (1990), Caglar et al. (2004) and Wong et al.
(2006).
Smith (1977) demonstrated how to evaluate and nd optimal (S  1, S) policies for an inventory system
with zero replenishment costs and generally distributed stochastic lead times with lost sales. He assumed, if
a nominal stock S is exhausted before the necessary replacement part arrives, a penalty cost L is incurred
for each unsatised demand and will have to be lled on an emergency basis (corresponding to penalty
incurred due to lost sales). The emergency handling procedure corresponds to priority shipping using premium
mode of transport like special air shipment. He used known formulas for the steady state probabilities of the
queuing system associated with the inventory system to formulate a cost function for the inventory system in
the steady state. Silver and Smith (1977) used the result from this model to develop a methodology to construct a set of indierence curves which permitted the exact determination of optimal inventory level in a system. Andersson and Melchiors (2001) presented a heuristic method for evaluation and optimization of
(S  1, S) policies for a two-echelon inventory model with lost sales. Their heuristic uses the METRIC approximation as framework for nding cost eective base-stock policies.
Although limited, a more relevant line of research focuses on incorporating multiple transportation and
delivery modes into inventory models. A recent example is Jaruphongsa et al. (2005) that considers the availability of multiple modes within a dynamic lot sizing model. Another study by Klincewicz and Rosenwein
(1997) considered multiple transportation modes in the context of warehouse outbound logistics planning with
the goal of making timely and consolidated deliveries to customers and keeping the warehouse workload balanced. This is not exactly an inventory and transportation mode integration, but the underlying mode-based
planning decisions aect the inventory levels. There are similar studies that incorporated mode-selection issues
into a higher level supply chain design and strategy models, some of which include long term inventory policies
and costs, but their focus and the underlying goals are dierent from ours as they focus on strategic design
choices and high-level planning decisions. In this category, Monahan and Berger (1977) analyzed the tradeo
between the transportation cost and the maximum time until a shipment reaches a consolidation point or a
central warehouse within a high-level distribution system. Van Roy (1989) considered transportation options
(truck eet or outside carriers) within a facility location planning model with an approximate inventory cost
function, as applied to petrochemical industry. Kiesmuller et al. (2005) investigated the value of delaying
transportation mode decisions and the value of using a slow mode along with a fast mode, especially with
an explicit consideration of manufacturing lead times and inventory policies. Eskigun et al. (2005) explicitly
considered transportation mode choices along with lead times and inventory costs within a supply chain network design model, as applied to an outbound vehicle distribution system. Rao et al. (2000) presented a novel
decomposition-based optimization application redesigning the supply chain of a production line with routing
and inventory decisions, with choices that can be viewed regular and expedited transportation modes.
Also, several inventory-focused papers considered expediting individual deliveries to increase service levels

668

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

as part of an overall problem. Most relevant ones are by Aggarwal and Moinzadeh (1994) and Moinzadeh and
Aggarwal (1997).
3. Model formulation
3.1. Problem description
The inventory/transportation system under consideration is a single-echelon system with multiple facilities
for stocking service parts. These facilities also serve the purpose of dispatching parts to their customers which
are dispersed within a geographic area. If any of the facilities are out of stock, the demand for that part is
fullled through an emergency shipment from the central warehouse (which is assumed to be outside the time
window). The facilities inventories are controlled by a one-for-one replenishment policy (also called (S  1, S)
policy). We assume that the central warehouse has innite stock, i.e., can satisfy any replenishment or direct
emergency shipment request without any extra delay due to stock-outs (hence shipments from the central
warehouse take constant time). The replenishment lead time is assumed to be the same for all facilities.
The emergency shipment transportation time between the central warehouse and any customer in need is
assumed to be outside the service time window hence does not count toward the time-based service level.
As these shipments typically handled as the next day delivery and are handled by outsourced service providers
with a at rate, we assume that these shipments cost the same across dierent customers. The local deliveries
(facilities to customers) introduce cost and time dierentials that are critical to the overall problem, and hence
handled more explicitly in the model.
The demand for the parts from the customers follows independent Poisson distributions, which is a common assumption in almost all service parts models as reviewed in the literature review. From the managements point of view, the service provider may have further divided the geographic area into service
regions. Each service region must respect the time-based service levels that are derived from the service agreements signed with the customers in that region. The time-based service level is dened as the percentage of
total service region demand satised within a time window. We require the actual percentage satised within
the time window (after considering the ll rates of the facilities within the region) to be at least a certain target
service level. If the overall geographic area represents a single region, then a system-wide service level should
be satised. To meet these constraints, the company uses three dierent transportation modes for delivery
(called slow, medium and fast) to customers which are distinguished based on their speed and cost. A tradeo
therefore exists between the cost savings using the less costly modes and the responsiveness achieved using the
faster modes. Furthermore, in order to achieve the required time-based service level, it is essential that the
facility has the requested part when a customer needs it. Hence, inventory stocking decisions should be integrated into the mode choice decisions. The problem would therefore involve two types of decisions: The rst
type is the stocking of the service parts that the facilities should maintain. The second type is the assignment of
transportation modes to the customers, so as to satisfy the time-based service constraints. As a rst step to
analyze the general problem with multiple parts, we consider one service part version in this paper although
the multi-part extension is relatively straightforward. We develop a model to study the behavior of the system,
i.e., how the optimal stocking policies and mode choice decisions change with the changes in various parameters of the system, including cost structure and service levels.
Note that in our model ll rates are decision variables which could be varied to achieve the overall timebased service level. Hence, the dened (time-based) service level is a function of ll rates which were traditional
service measures in the literature. In this sense, the time-based service levels are similar in sprit and form to
the demand-weighted composite ll rates in Muckstadt (2005) and Hopp et al. (1997). As all demands are
eventually satised within the system (either directly from stock on hand at a facility or with an emergency
shipment from the central warehouse), the level of demand satisfaction is 100%.
3.2. Model
The mathematical models objective is to minimize the total cost of the system. The objective function
therefore comprises of the following components:

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

669

Holding cost (HC) Cost of stocking the service part at all stocking facilities
Transportation cost (TC) Cost of transporting the part from facilities to the customers.
Emergency shipment cost (EC) Cost of fullling the demand from the central warehouse through direct
emergency shipments, needed when the main facility responsible from the customers demand is out of
stock at the time of the demand.
This is subject to the constraints that ensure:
Demand of each customer is met and is fullled by one mode, and
Target time-based service levels are met.
First, we introduce the notation used in the model (with a little abuse of notation, we use one symbol to
denote both the set and its size, e.g., J is the set of customers and the number of customers).
Sets and indices
I
set of modes, indexed by i = 1, 2, 3
J
set of customers, indexed by j = 1, 2, . . . , J
K
set of regions, indexed by k = 1, 2, . . ., K
M
set of stocking facilities, indexed by m = 1, 2, . . ., M
Jk
set of all customers located in service region k
Jm
set of all customers assigned to facility m
Demand and service parameters
kj
mean demand rate (per unit time) of customer j
P
km
mean demand rate at location m, km j2J m kj
P
k
total mean demand rate from all customers in the overall area, k j2J kj
s
replenishment lead time from the central warehouse to a facility, same for all facilities
ak
time-based service level for region k
h
time within which customers must be served to be considered part of the time-based service level
dij
1 if customer j can be reached within h time units from its assigned location using mode i, 0
otherwise
Cost parameters
cij
unit transportation cost for delivering the part to customer j using mode i from its assigned facility
hm
holding cost per unit time at facility m
e
unit emergency shipment cost for direct deliveries from the central warehouse to customers (assumed
to be the same for all customers)
Decision variables
Xij
1 if mode i is used to deliver the part to customer j from its assigned location, 0 otherwise
Sm
base-stock level at facility m
bm
long run ll rate of facility m (percentage of demand satised directly from stock on hand at location
m)
The following is the formulation for the minimization of the steady state total cost per unit time subject to
time-based service level constraints:

670

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

Minimize

hm S m  bm km s

XXX
m

cij kj bm X ij

j2J m

subject to bm 1  gS m =GS m 8m
X
XX X
dij kj bm X ij P ak
kj
m

j2J k \J m

X ij 1

ekm 1  bm

2
8k

j2J k

8j

1 8i; j

X ij 0

or

Sm P 0

and integer

8m

Objective function (1) is the total cost, which is composed of the three components mentioned earlier, namely,
the total holding cost as the sum of average holding cost at all the facilities, the total transportation cost for
delivering the part to customer locations using the assigned mode, and the cost for satisfying demands passed
from the facilities and satised by direct shipment from the central warehouse. Note that the second term in
the objective is non-linear due to the multiplication of ll rates and mode-selection variables.
Constraints (2) calculate the ll rates for all facilities, using the lost sales case formula. Here, gm(Sm) and
Gm(Sm) are respectively the probability mass function and the cumulative distribution function of the lead time
demand at facility m. The lead time demand at each facility has also Poisson distribution (with mean kms),
because individual customer demands are independent Poisson random variables (see Zipkin (2000) for more
discussion and derivation of the lost sales ll rate formula).
Constraints (3) satisfy the time-based service level constraints, each written for a service region. Here, the
left hand side of the constraint calculates the total demand satised directly from stock from a facility within
the time window (h) using the corresponding transportation mode. This amount divided by the total system
demand must be at least ak. Note that the left hand side is non-linear because of multiplication between the ll
rates bms (which are themselves functions of stock levels that are in turn decision variables) and mode-selection variables, Xij.
Constraints (4) ensure that all customers are served and only one mode is assigned to each customer.
Finally, constraints (5) are the binary restrictions on mode-selection variables, and constraints (6) are the
non-negativity and integrality restrictions on the stock levels.
3.3. Special case: one facility, one mode setting
We can obtain more direct insights on the model and its solution when we analyze the special case with one
facility (hence one service region) and one transportation mode. Note that in this case, all the demand is
assigned to a single facility using one mode of transportation. Hence, the only real decision variable is the
facilitys base-stock level that minimizes the total costs such that the service level is satised. This special case
can be represented with the following model (omitting the indices for location (m), service region (k), and
mode (i)):
X
Minimize hS  bks b
cj kj ek1  b
7
j2J

subject to b 1  gS=GS
X
dj kj P ak
b

8
9

j2J

SP0

and integer

10

P
First, for this problem to be feasible, we must have a 6 j2J dj kj =k, because the demand rate within the time
window of the facility divided by the total demand rate provides an upper bound on the maximum fraction
that can be satised within the time window after considering the facilitys ll rate. We now show that the total
cost function is convex under most conditions, hence itPis not sucient to nd the smallest stock level that
satises the service level (i.e., nd S such that b P ak= j2J dj kj ):

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

Proposition 1. The total cost (7) is (discretely) convex function of stock level S if ek P

671

j2J cj kj

 hks.

Proof. Note that the lost sales ll rate is a (strictly) concave function of S because the Erlang loss probability (g(S)/G(S))
is strictly convex in S (Karush, 1957). The objective function can be rewritten as
P
hS b j2J cj kj  ek  hks  ek. Hence, showing that the objective is convex reduces to showing that coefcient of b in the cost function is non-positive which leads to the condition in the proposition. h
Note that under our assumptions (and most realistic conditions) this condition is already satised because
the unit emergency shipment cost (e) from the central warehouse is greater than any of the unit transportation
costs from a stocking facility (cj). Then, to nd the optimal stock level (denoted by S*) subject to the service
level constraint, we rst nd the stock level that minimizes the cost function (7) P
(denote this stock level by S 0
0 0
0
0
and its ll rate by b (S )), and compare its service level with the target: if b S j2J dj kj P ak, then S* = S 0
0 0
0
(from convexity of the
P cost function, S is* the smallest non-negative integer such that b S 1
0 0
b S 6 h=hks ek  j2J cj kj . Otherwise,
the service level. Fig. 1
P S is the smallest stock level that satisesP
shows an illustrative example where k = 4, j2J dj kj 3.6, h = 100, e = 150, s = 1, and j2J cj kj 20. Here, if
the target service level a is set to a level larger than 90%, the problem is infeasible; for any a 6 72%, S* = 5 is
optimal. For a = 85%, S* = 8.
3.4. Special case: multiple facilities, one mode
A multi-facility extension of this model (still with one region and one mode) can be solved with a Lagrangian-relaxation method. The model in this case reads as follows:
X
X X
X
Minimize
hm S m  bm km s
bm
c j kj
ekm 1  bm
11
m

subject to

j2J m

bm 1  gS m =GS m 8m
X X
X
bm
dj kj P a
kj
m

j2J m

Sm P 0

12
13

j2J

and integer 8m

Fig. 1. Costs and service levels for an illustrative example with one facility and one mode.

14

672

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

Relaxing constraint (13) with multiplier u, we obtain the relaxed problem:


Minimize

hm S m  bm km s

bm

c j kj

j2J m

subject to bm 1  gS m =GS m 8m
S m P 0 and integer 8m

X
m

ekm 1  bm u ak 

X
m

bm

!
dj kj

15

j2J m

16
17

This problem decomposes into subproblems, one for each facility, very similar to the one introduced earlier
(Eqs. (7)(10)). The main dierence here is that the transportation costs are now modied with respect to
u: Instead of cij, we use cij  udj. Suppose the stock level that solves the subproblem for location m for given
u is S 0m u. Starting with u = 0, if S 0m 0 values lead to ll rates that satisfy the service level constraint (13), then
these stock levels are optimal. Otherwise, we seek to nd the smallest value for u such that the service level
constraint is satised. We can use bisection search or subgradient optimization to nd the optimal u.
3.5. Special case: One facility with a given stock level
For the problem with one facility with multiple transportation modes available, we rst make the stock
level a parameter. Hence, given a stock level S = s, and its ll rate b, we formulate the problem as follows
(Note that in this case the inventory and emergency shipment costs are constants):
XX
Minimize
cij kj X ij
18
subject to

j2J

j2J

XX
X

dij kj X ij P ak=b

X ij 1

19

8j

20

1 8i; j

21

X ij 0

or

For the setting with two transportation modes (mode 2 being faster and more expensive), we can further
simplify the problem with substitution X1j = 1  X2j:
X
X
Minimize
c1j kj
c2j  c1j kj X 2j
22
j2J

subject to

j2J

X
d1j kj
d2j  d1j kj X 2j P ak=b

j2J

X 2j 0

23

j2J

or

8j

24

Here, one can view the reformulation as all the customers are (temporarily) assigned the slow (and less costly)
mode (note that d2j P d1j). Assuming that this is not sucient for the target service level, the remaining problem is to decide which of the customers (that can take advantage of the fast mode, i.e., d2j = 1, and d1j = 0)
should be switched to the fast mode such that the switching cost is minimized and the target service level
is satised. This reduced problem can be as hard as the Knapsack problem in general (when there is no clear
pattern in transportation costs (or distances) and customer demand rates), which is known to be NP-hard.
One can extend this analysis to include more transportation modes, but it is clear that even the problem with
one facility, two transportation modes, and given stock level is not trivial.
3.6. General case
Note that even though the single-mode problem (with single or multiple facilities) is directly handled, the
problem with multiple modes (even for a single facility with a given stock level) does not lend itself to such
methods, mainly due to discrete mode choices available to customers. To remove non-linearity in the original
model in such general cases and make it amenable to be solved by direct optimization solvers such as CPLEX
(ILOG, 2005) or Xpress-MP (Dash Optimization, 2006), we introduce the following additional notation:

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

673

Parameter
bms
ll rate level at location m with stock level s
Decision variable
Yms
1 if location m uses stock level s, 0 otherwise
Here, bms represents the tabulated values of ll rates, calculated at potential stock levels s = 1, . . . , Smax, for
each location m, where Smax is the stock level that practically gives approximately 100% ll rate. This value
is usually a small number, between 3 and 10, for most of the practical cases, due to extremely low demand
levels experienced in SPL systems. Hence, bms values are preprocessed ll rates for potential stock levels, evaluated with known lead time demand for each facility, i.e., bms = 1  gm(s)/Gm(s), for all s and m, each evaluated using the locations corresponding mean lead time demand kms. With potential ll rates in parameters, we
can calculate the actual ll rate variables at facilities (depending on the stock levels chosen) as
X
bm
bms Y ms 8l
s

P
We can also represent the stock level variable with the expression S l s sY ls 8l. By substituting these two
into the model, we obtain the following version:
!
!
X
X
X
XXXX
X
X
Minimize
hl
sY ls 
bls Y ls kl s
cij kj bls X ij Y ls
ekl 1 
bls Y ls
s

subject to

XXX X
l

dij kj bls X ij Y ls P ak

j2J k &
j2J l

j2J l

25
kj

8k

26

j2J k

X ij 1

8j

27

Y ls 1

8l

28

X
s

X ij 0 or 1 8i; j
Y ls 0 or 1 8l; s

29
30

Here, constraints (28) make sure that only one of the available stock levels is chosen for each facility. This new
version of the model internalizes the potential ll rates as parameters, but it still has non-linear terms in the
objective and the constraints, but these terms are eectively handled by another substitution easily: Dene Zijls
as the decision variable that is 1 if customer j is assigned to mode i and the customers location l uses stock
level s, 0 otherwise. This way, Zijls = XijYls. To obtain the correct representation, we add the following constraints into the model (25)(30):
Z ijls 6 X ij

8i; j; l; s

31

Z ijls 6 Y ls

8i; j; l; s

32

Z ijls P X ij Y ls  1 8i; j; l; s
Z ijls 0 or 1 8i; j; l; s

33
34

Now with the model fully linearized, one can use any of the standard integer programming solvers.
4. Experimental settings
To investigate if a signicant interaction exists between mode decisions and inventory and to obtain insight
into how the behavior of the system changes with the changes in the system parameters, we designed four networks by randomly locating 100 customers and 3 facilities on a grid of 100 100 units, representing an area to
be served by the facilities. For each of these customers and facilities, two random numbers were generated

674

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

between 0 and 100 and these were taken as the x and y coordinates of customer/facility locations. Using the
coordinates, the Euclidean distances of the customers from facilities were calculated. These networks were also
solved for the single facility case. The single facility was located at the centroid of the three facilities. In real
life, this would represent the case, when the decision has to be made between opening one large facility to
cover the demand of the entire area or opening three smaller facilities to meet the demand of the segregated
customers in the area. We focused on the single service region instances.
In generating the rest of the data, we base our choices on the characteristics of a real life industrial data set.
For each of these networks, ve demand scenarios were randomly simulated. The demand range was chosen
so as to incorporate slow moving characteristic of service parts. Each customer is assumed to be assigned to its
nearest facility. The behavior of the system was analyzed by observing how the costs of the system, mode
usage, and stocking decisions change with the changes in the following factors:
mean customer demand rates,
time-based service levels,
emergency shipment cost.
In the experimental study, we used the data shown in Table 1.
The system was analyzed for low and high value of penalty level dened as a scalar multiple of xed holding
cost. For our experiments we chose low and high levels to be 5 and 10 times the holding cost, respectively.
Increasing the value of penalty level indicates higher level of cost resulting from stock-outs at the facilities
and consequently paying more for direct emergency shipments from the central warehouse to the customers.
For networks 13, we generated low demands, while for network 4 was characterized by relatively high
demand. Networks 13 are generated to represent randomness in facility locations and unequal demand pattern at the facilities. Network 4 was generated in a more controlled way, such that the facilities are centrally
located and each of the facilities shared similar demand pattern.
The holding cost was kept constant at $20/unit/week, while the mode speed factors were xed at 4, 10 and
20 units of distance per unit time (hour) for the slow, medium and fast modes, respectively. We assigned shipment cost from the facilities to the customer location as $1, $1.5 and $2.5 per unit distance for slow, medium
and fast modes, respectively. Average replenishment lead time from the central warehouse to the facilities was
xed at 1 week and the service delivery time window from the facilities to the customer locations was chosen as
4 h. From the locations of customers and facilities and their distances, and speeds of the modes, we compute
dijs.
Note that there is a certain bound on how much of demand one can cover within a time window (which is
our dened service level, with one or more transportation
modes) even if we assume 100% ll rate (part availP
ability) at every facility. This bound, given by j2J dij kj =k, is a function of how the customers are geographically dispersed, what their demand rates are, and where the facilities are located (with respect to customer
locations). For example, requiring 95% service level may be infeasible given that maximum one can cover
within the time window is 85%, because only 85% of the total demand is achievable within the time window
from all the facilities. Using multiple modes gives some exibility as one may be able to cover more of the
demand by using faster modes. In any case, in the random problems generated, the high time-based service
levels (80%) are demanding given the maximum coverage bounds. The low levels (60%) are considered
for comparison.
One expectation is that for slow moving service parts, one can save substantially by reducing inventory at
the facilities by using appropriate combination of modes while maintaining the same service level with the

Table 1
Data used for experimental study
Factor

Unit

Low

High

Demand/customer
Time-based service level
Penalty level factor

Units/week
% Demand satised in the region

0.000.02
60%
5

0.000.06
80%
10

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

675

customers. We also investigate the cases in which it may not be benecial to stock an item at the facilities. The
performance is measured in terms of the total cost of the system, the base-stock levels at the facilities, and the
mode usage for the given service level. The mode usage is dened as the percentage of demand satised by each
mode.
5. Computational results
The model was solved for single facility and three-facility cases in GAMS modeling language (Brooke et al.,
2000; Rosenthal, 2006) using CPLEX integer programming solver (ILOG, 2005) for the four networks generated. Eectively this resulted in 200 problem instances, with each of the four networks being tested for two
levels of the emergency transportation cost, two levels of the time-based service level, and ve demand scenarios, solved for a single facility model and a three-facility model. These problem instances were run on Pentium
M 1.3 GHz Centrino processor, with Windows based operating system and the solver was able to converge to
the solution for the specied instances of the problem within 110 min per instance. For the comparison purpose and to see the eectiveness of integrating mode choice decisions with the inventory decisions on the overall cost savings, we also used the models version with only one mode choice available.
The results based on the average of ve demand scenarios for the single facility and three-facility models,
both with multiple modes are presented in groups of four charts. Figs. 210 illustrate how stocking decisions,
mode usage and total cost structure change with changes in service level and penalty level for the four networks. More detailed results in tabular form are available from authors.
5.1. Network 1
From Fig. 2, it is clear that at xed service level of 60%, in the one-facility network, an increase in penalty
level is handled by increasing the base-stock level at the facility, which leads to increase in the holding cost.
Higher base-stock levels lead to decrease in emergency shipment costs because we have more parts available to
satisfy demands directly from the facility. At the low service level, the model suggests using slow mode to satisfy 6068% of the demand. The fast mode is not used at all. However, when the service level is set at 80%, the

Fig. 2. Single facility network #1.

676

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

Fig. 3. Three-facility network #1.

model suggests an increase in both the base-stock level as well as in the use of medium mode and the fast mode
to satisfy customers. Intuitively this makes sense, as we need to reach more customers in time to achieve 80%
service level, which calls for faster modes and stocking more parts at the facility.
In Fig. 3, the results for the three facilities are presented. The customer assignment based on nearest distance leads to satisfying approximately 82% of the demand from facility 2. The model suggests that for the
base case (low service level/low penalty level), we are better o by not stocking the part at facilities 1 and
3, but making more direct shipments from the central warehouse. Also, facility 2 (with location coordinates
(45, 30) being closer to the majority of customers) is able to satisfy 80% of the demand by using the slow mode.
For the base case (low service/low penalty level), the overall cost performance is closer to the one-facility location case. The dierence in the transportation cost is due to the location of facility 2.
An increase in penalty level deteriorates the performance as we incur higher emergency shipment cost. Also
to remain feasible, we are forced to stock at facility 3. The total cost goes up with both the higher penalty level
and the higher target service level, when compared to the single facility case. Higher penalty levels force the
model to start looking at the options of stocking the part at all the facilities while increasing service level forces
the model to satisfy more demand by using faster modes. Thus, for such a network we are better o by stocking one facility rather than three facilities.
5.2. Network 2
Fig. 4 presents a network when we have facilities located far from the customers. Since the distances
between the facilities and the customers are large, the model suggests using the fast mode to satisfy the
time-based service level constraint. Hence, transportation costs become signicant portion of the total cost.
The base-stock level follows a similar pattern; an increase in service or penalty level makes the facility stock
more. However, when we switch the service level to 80%, the model suggests turning both levers (maintain
larger stock and use faster modes) to satisfy greater demand within the time window for optimal performance.
It is interesting to note that the emergency shipment cost forms the small portion of the overall cost for such a
network.

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

677

Fig. 4. Single facility network #2.

Fig. 5 summarizes the results obtained for network 2 with three facilities. The customer assignments based
on nearest distance lead to a high demand at facilities 1 and 2, jointly covering 96% of the demand. Very low
demand at facility 3 indicates an average part requirement of less than one. In the three-facility case, although
we stock more parts than the single facility case, the total costs are about the same. For low penalty level, the
three-facility network gives slightly lower cost while with high penalty level the single facility network has
lower total cost. This is because we are closer to customers with three facilities; hence we can use cheaper medium mode instead of fast, and still satisfy the time-based service constraint. Since we do not stock at facility 3,
an emergency shipment cost is incurred for all demand assigned to this facility, leading to an increase in total
cost of the system. For such a network, the variable cost is the same and the decision to operate three facilities
or one facility should be driven by xed cost of opening a facility, customer demands, and cost of transportation modes.

5.3. Network 3
Fig. 6 summarizes results for the single facility version of network 3, which shows similar results to the single facility version of networks 1 and 2 in general behavior. At high penalties and service levels, we increase
our average base-stock level from 2.4 to 3.2 and use faster modes to satisfy 54.72% of demand as compared to
39.79%.
Fig. 7 presents the results for the three-facility case for network 3. Again, there is an imbalance in the distribution of demand: only 6% of the total demand goes to facility 1, while the remaining demand is assigned to
facilities 2 and 3. For the 60% service level, the model suggests not to stock at facility 1 but supply its customers through emergency shipments from the central warehouse. The costs for the three facilities are almost same
as those of the one-facility model for this network. The holding cost increases while the transportation cost
goes down due to proximity to the customers. At higher service levels, we stock even at facility 1. This is
because the amount saved on transportation while satisfying demand from this facility more than compensates

678

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

Fig. 5. Three-facility network # 2.

Fig. 6. Single facility network # 3.

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

679

Stock Distribution-N3-3L

Fig. 7. Three-facility network # 3.

the higher emergency shipment costs from the central warehouse or the cost of transportation to meet other
customers demands from the other two facilities with faster modes.
5.4. Network 4
Network 4 was designed such that facilities were located centrally around the customers and each facility
faces about the same level of demand from its customers. The demand from a customer ranges from 0 to 3
parts per year. Due to higher demand rates, the model suggests maintaining higher base-stock levels. Since
the inventory moves faster, the average inventory costs do not increase in proportion to the increase in the
base-stock level as shown in Fig. 8.
Even though the facility is centrally located, we see that the transportation cost is very high portion of the
total cost. This is because when the demand rate is high, we incur higher frequency of shipments to the customers. The mode usage pattern indicates that at the 60% service level we tend to use the fast mode to satisfy
only 13.76% of the demand. At the 80% service level, this goes up to 30.42%. The medium mode usage remains
unchanged with changes in service and penalty levels. This is because at the 60% service level some customers
which are far o and are not served within the time window are now served using faster modes to reach the
80% service level.
In the three-facility scenario (Fig. 9), the total cost of the system is less than that of the single facility case.
With more facilities close to customers, the transportation cost reduces by approximately 5660% as facilities
can use slower modes and still satisfy the time-based service constraint. The reduction in transportation cost
more than compensates increase in emergency shipment cost as well as holding cost due to stocking inventory
at all three locations. In all above cases, the optimal solutions do not suggest use of the fastest mode. Hence,
understanding the mode options available and using the right one in conjunction with the inventory stocking
decisions would result in otherwise hidden cost savings for a company.
5.5. Comparative analysis
For comparative analysis, we rerun the three-facility models with integrated inventory and transportation
decisions and service constraints but with making only single mode available for delivery. To remain feasible

680

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

Mode Usage

X Cordinate

Penalty Cost, SL

Penalty Cost, SL

Fig. 8. Single facility network # 4.

Fig. 9. Three-facility network # 4.

and obtain a gross estimation of potential improvement, the fastest mode was assumed to be the only available
mode. Fig. 10 presents the costs for all the networks and scenarios for the single and multi-modal cases. From

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

681

Fig. 10. Cost comparison single mode and multiple modes.

the relative costs, it seems that the maximum benet from the multiple modes occurs when the network has
high service levels and low penalty levels. Also in all the cases, we nd that integrating mode choice decisions
with the inventory decisions helps us operate our network for lower cost, i.e. we are able to meet our network
constraints at lower cost. Note that the single-mode solution can be improved by selecting the modes for individual customers more carefully (slower and less costly modes that deliver the parts within the time window).
Hence, the improvement due to the usage of multiple modes here represents a gross estimation.
5.6. Discussion
In our analysis, we designed four networks to study how inventory decisions and mode choices are aected
by changes in required time-based service levels and in penalty level in service parts logistics. Integrating multiple mode options in the inventory and transportation decisions for all the networks resulted in lower cost,
while ensuring service level constraints. Our model determines optimal inventory and customer-mode assignment decisions. While all four networks are dierent with dierent solutions obtained, they share some common results and observations:
An increase in penalty level leads to stocking more in all four networks. However it is not always optimal to
stock all parts at all facilities. Our model provides insights whether it is better to stock at a particular facility
or ship directly from the central warehouse. This would help in determining major and non-major facilities
for dierent parts. Stocking only major facilities while letting faster modes satisfy demands such that service
levels are achieved will reduce both inventory costs and warehouse capacity pressures.
Increase in time-based service levels requires adjusting both inventory and mode assignment decisions.
Operating three smaller facilities or one large facility with fast mode to satisfy service constraints depends
on customer demand rates and the given network. While for networks 13, one could chose between opening one or three facilities, network 4 clearly suggests that three-facility design performed better.
6. Conclusions
Service parts are characterized as parts with extremely sporadic and low demand, and high costs. In this
paper, we developed a tactical optimization model that builds on the base-stock inventory model and integrates it with transportation options and service responsiveness that can be achieved using alternate modes.

682

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

The model captures the cost tradeos at dierent service levels and penalty levels. We also analyzed several
special cases (such as single facility, single-mode problems) that can be solved eciently by taking advantage
of the special structure of these problems. We showed that a slight generalization toward multiple modes (even
for a single facility with a known stock level) leads to a hard Knapsack type problem. Hence, we proposed and
showed the feasibility of a linearization scheme for the general case, which makes it possible to obtain optimal
integrated solutions using a standard integer programming solver. We tested our approach on four dierent
networks for a single region with up to three facilities, 100 customers, and three modes.
There are several extensions possible for this model:
Extending the model beyond single region. This would require determining partitioning of the overall service
area based on customer demands and potentially dierent service levels in each region. Once the data for
partitioning the area is known, the model can be easily extended to handle this extension.
Extending the model for multiple-product case. In such a model, each product can have its own desired service level and penalty cost depending on criticality. This modeling extension too is relatively easy.
Integrating facility location decisions into the model. In our model, we assumed the locations of facilities
were xed. This extension can be used to analyze the cost changes when we optimally decide the location
of facilities. (See Candas and Kutanoglu (2007) for such an integrated location and inventory model with
one transportation mode.)
Integrating customer assignment decisions into the model. Here, the transportation modes would play even
more crucial role in determining the inventory at each facility.
Extending the model to a multi-echelon structure. Our model is valid for single echelon and assumes innite
capacity at the central warehouse. From a practical perspective, most service providers use at least two echelons in their distribution network where both levels are capacitated in terms stocking levels.
These extensions would make the model richer (and harder to solve due to the increased complexity and
size) and would show how eective it would be in solving real world problems. The maximum time to solve
the model for an instance of the problem was around 10 min of run time. It would be interesting to see how the
computational time increases as we try to solve larger and more complex networks.
References
Aggarwal, P.K., Moinzadeh, K., 1994. Order expedition in multi-echelon production/distribution systems. IIE Transactions 26, 8696.
Andersson, J., Melchiors, P., 2001. A two echelon inventory model with lost sales. International Journal of Production Economics 69 (3),
307315.
Brooke, A., Kendrick, D., Meeraus, A., Raman, R., 2000. GAMS A User Guide. GAMS Development Corporation, Washington, DC.
Caglar, D., Li, C., Simchi-Levi, D., 2004. Two-echelon spare parts inventory system subject to a service constraint. IIE Transactions 36,
655666.
Candas, M.F., Kutanoglu, E., 2007. Benets of considering inventory in service parts logistics network design problems with time-based
service constraints. IIE Transactions 39 (2), 159176.
Cohen, M.A., Kamesam, P., Kleindorfer, P., Lee, H., 1990. OPTIMIZER: IBMs multi-echelon inventory system for managing service
logistics. Interfaces 20, 6582.
Cohen, M.A., Zheng, Y.-S., Agrawal, V., 1997. Service parts logistics: a benchmark analysis. IIE Transactions 29, 627639.
Dash Optimization, 2006. XPRESS-Optimizer Reference ManualRelease 17.
Eskigun, E., Uzsoy, R., Preckel, P.V., Beaujon, G., Krishnan, S., Tew, J.D., 2005. Outbound supply chain network design with mode
selection, lead times and capacitated vehicle distribution centers. European Journal of Operational Research 165 (1), 182206.
Feeney, G.L., Sherbrooke, C.C., 1966. The (S  1,S) inventory policy under compound Poisson demand. Management Science 12 (5),
391411.
Hopp, W.J., Spearman, M.L., Zhang, R.Q., 1997. Easily implementable inventory control policies. Operations Research 45 (3), 327340.
ILOG, I., 2005. ILOG CPLEX 10.0 Reference Manual.
Jaruphongsa, W., Cetinkaya, S., Lee, C.Y., 2005. A dynamic lot-sizing model with multi-mode replenishments: polynomial algorithms for
special cases with dual and multiple modes. IIE Transactions 37, 453467.
Karush, W., 1957. A queuing model for an inventory problem. Operations Research 5, 693703.
Kiesmuller, G.P., de Kok, A.G., Fransoo, J.C., 2005. Transportation mode selection with positive manufacturing lead time.
Transportation Research Part E: Logistics and Transportation Review 41 (6), 511530.
Klincewicz, J.G., Rosenwein, B.M., 1997. Planning and consolidation shipments from a warehouse. Journal of the Operational Research
Society 48, 241246.

E. Kutanoglu, D. Lohiya / Transportation Research Part E 44 (2008) 665683

683

Moinzadeh, K., Aggarwal, P.K., 1997. An information based multiechelon inventory system with emergency orders. Operations Research
45 (5), 694701.
Monahan, J.P., Berger, P.D., 1977. A transportation mode selection model for a consolidation warehouse system. Mathematical Methods
of Operations Research 21 (5), 211222.
Muckstadt, J.A., 1973. A model for multi-item, multi-echelon, multi-indenture inventory system. Management Science 20, 472481.
Muckstadt, J.A., 2005. Analysis and Algorithms for Service Parts Supply Chains. Springer, New York, NY.
Muckstadt, J.A., Thomas, L.J., 1980. Are multi-echelon inventory methods worth implementing in systems with low-demand-rate items?
Management Science 26, 483494.
Rao, U., Scheller-Wolf, A., Tayur, S., 2000. Development of a rapid-response supply chain at caterpillar. Operations Research 48 (2), 189
204.
Rosenthal, R., 2006. GAMSA Users Guide.
Sherbrooke, C.C., 1968. METRIC: multi-echelon technique for recoverable item control. Operations Research 16, 122141.
Sherbrooke, C.C., 1986. VARI-METRIC: improved approximation for multi-indenture, multi-echelon availability models. Operations
Research 34, 311319.
Silver, E.A., Smith, S.A., 1977. A graphical aid for determining optimal inventories in a unit replenishment inventory system. Management
Science 24 (3), 358, 359.
Smith, S.A., 1977. Optimal inventories for an (S  1,S) system with no backorders. Management Science 23 (5), 522528.
Van Roy, T.J.V., 1989. Multi-level production and distribution planning with transportation eet optimization. Management Science 35
(12), 14431453.
Wong, H., van Houtum, G.J., Cattrysse, D., van Oudheusden, D., 2006. Multi-item spare parts systems with lateral transshipments and
waiting time constraints. European Journal of Operational Research 171, 10711093.
Zipkin, P., 2000. Fundamentals of Inventory Management. McGraw Hill-Irwin, Boston, MA.

You might also like