IISE Transactions

ISSN: 2472-5854 (Print) 2472-5862 (Online) Journal homepage: https://www.tandfonline.com/loi/uiie21

Optimal Control Policies for Assemble-to-Order

Systems with Commitment Lead Time

Taher Ahmadi, Zumbul Atan, Ton de Kok & Ivo Adan

To cite this article: Taher Ahmadi, Zumbul Atan, Ton de Kok & Ivo Adan (2019): Optimal Control
Policies for Assemble-to-Order Systems with Commitment Lead Time, IISE Transactions, DOI:
10.1080/24725854.2019.1589658

To link to this article: https://doi.org/10.1080/24725854.2019.1589658

Accepted author version posted online: 03

Apr 2019.

Submit your article to this journal

Article views: 26

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=uiie21
Optimal Control Policies for Assemble-to-Order
Systems with Commitment Lead Time
Taher Ahmadi*, Zumbul Atan, Ton de Kok, Ivo Adan

*Email: serkewtin@gmail.com

Taher Ahmadi is a Ph.D. candidate in Supply Chain and Logistics at the

School of Industrial Engineering & Innovation Sciences of the Eindhoven
University of Technology. He received his bachelor's and master's degrees in
Industrial Engineering from Isfahan University of Technology and Amirkabir

t
University of Technology in Iran, respectively. His research interests are

rip
stochastic modeling and optimization of supply chain and logistic systems.

c
Zumbul Atan is an assistant professor at the School of Industrial Engineering
& Innovation Sciences of Eindhoven University of Technology. She got her

us
Ph.D. in Industrial Engineering in 2010 from Department of Industrial and
Systems Engineering at Lehigh University, PA, USA. She does research on
an
optimization of multi-echelon supply chains subject to demand and supply
uncertainty, revenue management and retail operations in collaboration with
multiple international universities and companies. Her research is published
M

in prestigious journals like Manufacturing & Service Operations

Management, IIE Transactions, Production and Operations Management and
European Journal of Operational Research.
ed

Ton de Kok received a Ph.D. in Mathematics in 1985 from the Free University
pt

of Amsterdam. Since 1992 he is a full-time professor Operations Management

at the same university. He has published over 100 articles in international
ce

scientific journals, such as Management Science, Operations Research and

Manufacturing and Operations Management. In 2004 he was Edelman Award
Finalist as a member of the Philips Semiconductors team. In 2014 he was
Ac

appointed as Fellow of the International Society for Inventory Research (ISIR).

Ton’s main research areas are Supply Chain Management and Concurrent
Engineering with emphasis on quantitative analysis. His research results have
been successfully tested and implemented in a multitude of projects with
industry.

Ivo Adan is a Full Professor at the School of Industrial Engineering &

Innovation Sciences of the Eindhoven University of Technology and holds the
Manufacturing Networks chair. His current research interests are in the area
of modeling, design and control of manufacturing systems, warehousing
systems and transportation systems, and more specifically, in the
mathematical analysis of multi-dimensional structured Markov processes and
queueing models. He is a senior fellow of the workshop centre Eurandom and
he is the research director of Beta, the Netherlands network of Operations
Management and Logistics.

Abstract

In this paper, we study a preorder strategy which requires customers to place

orders ahead of their actual need. We characterize the preorder strategy by a
commitment lead time. We define the commitment lead time as the time that

t
elapses between the moment an order is communicated by the customer and

rip
the moment the order must be delivered to the customer. We investigate the
value of using this preorder strategy in managing assemble-to-order systems.

c
For this purpose, we consider a manufacturer, who operates an assemble-to-

us
order system with two components and a single end product. The
manufacturer uses continuous-review base-stock policies for replenishing
an
component inventories. Customer demand occurs for the end product only
and unsatisfied customer demands are backordered. Since customers provide
M

advance demand information by preordering, they receive a bonus. We refer

to this bonus from the manufacturer’s perspective as a commitment cost. We
ed

determine the optimal component base-stock levels and the optimal length of
the commitment lead time, which minimize the sum of long-run average
pt

component inventory holding, backordering and commitment costs.

ce

We find that the optimal commitment lead time is either zero or equals the
replenishment lead time of one of the components. When the optimal
Ac

commitment lead time is zero, the preorder strategy is not beneficial and the
optimal control strategy for both components is buy-to-stock. When the
optimal commitment lead time equals the lead time of the component with the
shorter lead time, the optimal control strategy for this component is buy-to-
order and it is buy-to-stock for the other component. On the other hand, when
the optimal commitment lead time equals the lead time of the component with
the longer lead time, the optimal control strategy is the buy-to-order strategy
for both components. We find the unit commitment cost thresholds which
determine the conditions under which one of these three cases hold.

Keywords: Assemble-to-order systems; preorder strategy; commitment lead

time.

1 Introduction

The companies in high-tech, car manufacturing and white good industries aim
to benefit from the strategic advantages of mass customization and delayed
differentiation by attempting to delay the production of the end product until
they obtain better or complete demand information. Assemble-to-order (ATO)

t
rip
is a very popular strategy that enables companies to reduce their customer
response time by keeping component inventories and delaying the final

c
assembly of the end products until the arrival of customer demand (Atan

us
et al. 2017; Benjaafar & ElHafsi 2006). An ATO system with long component
supply lead times has high component availability uncertainty. This implies
an
high uncertainty in delivery time of the end products. Having high component
inventory levels can increase the responsiveness of the system but results in
M

high inventory holding cost. On the other hand, low component inventory
levels lower the inventory holding cost but can result in a less responsive
ed

system. In order to achieve simultaneous improvement in both cost and

responsiveness, companies need to decrease the demand and supply
mismatch. Having more accurate information on future customer demand, i.e.,
pt

advance demand information (ADI), helps to reduce this mismatch.

ce

One form of ADI is a preorder strategy in which customers place orders ahead
of their actual need. The preorder strategy is characterized by a commitment
Ac

lead time. We define the commitment lead time as the time that elapses
between the moment an order is communicated by the customer and the
moment the order must be delivered to the customer. The preorder strategy
provides advance demand information and hence, reduces the demand
uncertainty. Although in today’s competitive market firms cannot force their
customers to place orders before their actual need, they can entice the
customers to follow the preorder strategy by giving bonuses. Long
commitment lead times can be made acceptable and attractive if the bonuses
increase with the length of commitment lead times. From an uncertainty point
of view, a long commitment lead time implies less demand uncertainty and
subsequently lower inventory unavailability risk (Lutze & Özer 2008).

Although the preorder strategy is a form of ADI, it is different than the form of
ADI that is usually used in the existing literature. In the existing literature, ADI
helps to make better forecast of the future customer demand. On the other
hand, the preorder strategy, as another form of ADI, is utilized operationally to
reduce demand-supply mismatch by reducing the lead time demand

t
rip
uncertainty. This form of ADI works for service and custom-production
companies, where service customers can make reservations and customers

c
of custom products order in advance of their needs (Hariharan & Zipkin 1995).

us
In this study, we investigate the value of using the preorder strategy in
managing the assemble-to-order (ATO) systems. For this purpose, we
an
consider a manufacturer, who operates an ATO system with two components
and a single end product. The manufacturer needs one unit of each
M

component to assemble/produce a unit of end product. The manufacturer

keeps inventory of the components. The component inventories are
ed

replenished from two different uncapacitated suppliers. Continuous-review

base-stock policies with deterministic replenishment lead times are used to
pt

replenish component inventories. The assembly time is negligible. Customer

demand occurs for the end product only and unsatisfied customer demands
ce

are backordered. The manufacturer offers a preorder strategy to its customers

and consequently, they are paid a commitment cost. The commitment cost
Ac

function is strictly increasing in the length of the commitment lead time. The
manufacturer aims to find the optimal component base-stock levels and the
optimal length of the commitment lead time, which minimize the total long-run
average cost. This cost is the sum of the long-run average component
inventory holding, backordering and commitment costs. We formulate the total
long-run average cost and answer the following questions:
 1. When and how should the manufacturer use the preorder strategy?

We find that the optimal commitment lead time is either zero or equals the
replenishment lead time of one of the components. When the optimal
commitment lead time is equal to zero, the preorder strategy is not
beneficial. The manufacturer should choose a strategy that is similar in
spirit to a pure make-to-stock procurement strategy. In our context, this
strategy is called buy-to-stock (BTS) strategy. When the optimal
commitment lead time equals the lead time of the component with the
shorter lead time, the optimal strategy for this component is buy-to-order
(BTO) (similar in spirit to a make-to-order strategy) and it is BTS for the

t
rip
other component. On the other hand, if the optimal commitment lead time
equals the lead time of the component with the longer lead time, the

c
optimal strategy is the BTO strategy for both components. We determine

us
the conditions under which one of these three cases holds.

2. What are the optimal component base-stock levels?

an
We show that the optimal base-stock policies are of “all-or-nothing” type.
M

This means that when the commitment lead time is zero, the
corresponding base-stock levels are the solution of a well-known two-
ed

stage serial system with deterministic lead times. This results follows from
the fact that the assembly system can be reduced to an equivalent serial
pt

system (Rosling 1989). When the commitment lead time is longer than or
equal to the replenishment lead time of a component, the corresponding
ce

optimal base-stock level is zero.

Ac

Scholars have studied inventory management with ADI broadly from different
perspectives. Hariharan & Zipkin (1995) study a single location system with
perfect ADI, continuous-review, deterministic replenishment lead time and
Poisson demand. Assuming that the commitment lead time is the same for all
customers, the authors prove the optimality of a base-stock policy. A similar
setting has been studied in Ahmadi et al. (2017). Assigning a cost to
commitment lead time, the authors characterize the optimal preorder strategy
and the corresponding optimal replenishment strategy. More specifically, they
prove the optimality of the bang-bang preorder and all-or-nothing
replenishment strategies. Both studies consider a single location system. Our
work uses these two studies as building blocks for characterizing the optimal
preorder strategy and base-stock policy for a more complicated assemble-to-
order system. The optimal replenishment policy for our problem is unknown.
We use the base-stock policy due to its practical applicability and analytical
tractability.

The remainder of this paper is organized as follows. In Section 2, we provide

a brief review of related literature. In Section 3, we analyze the ATO system

t
rip
with commitment lead time and derive the expression for the long-run average
cost function. In Section 4, we determine the optimal replenishment and

c
preorder strategies. In Section 5, we provide numerical examples. Finally, in

us
section 6, we conclude the paper. We defer the proofs to the Appendix.

2 Literature Review
an
The literature on advance demand information (ADI) assumes either perfect
M

or imperfect demand information available ahead of the realization of actual

demand. This literature can be broadly classified into two categories based on
ed

the accuracy of the demand information. These categories are perfect ADI
and imperfect ADI.
pt

When the firm has perfect ADI, customers place orders ahead of time in
ce

specific quantities to be delivered at specified due dates. Hariharan

& Zipkin (1995) are the first to study the perfect ADI situation in a continuous-
Ac

review setting. After this seminal work, many researchers assume perfect ADI
and study different problems. We provide a summary of the most relevant
literature on perfect ADI in Table 1.

When a firm has imperfect ADI, customers place their orders in advance but
they provide only an estimate of either the actual due dates or order sizes
(Gayon et al. 2009). Imperfect ADI converges to perfect ADI as information
gets available over time. We provide a summary of the most relevant literature
on imperfect ADI in Table 2.

Multiple studies consider both perfect and imperfect ADI. These are listed in
Table 3.

Our work belongs to the category of perfect ADI. We study the impact of
commitment cost as a function of the commitment lead time in a two-
component, single end product ATO system with perfect ADI. We consider
continuous-review base-stock policies, deterministic replenishment lead times
and Poisson customer demand. The study that is most closely related to ours

t
rip
is Ahmadi et al. (2017). Different from us, the authors study a single-location
inventory system. They prove the optimality of bang-bang and all-or-nothing
policies for the commitment lead time and replenishment policy, respectively.

c
They introduce a unit commitment cost threshold to make a decision on cost-
effectiveness of ADI.
us
an
We contribute to the literature on ADI by providing the first results for using a
preorder strategy in ATO systems. The benefits of acquiring and providing
M

information about future demand are undeniable. We identify the conditions

under which having information on future customer demand helps ATO
ed

systems in reducing component inventory levels without sacrificing high

service levels. Not only the manufacturers of these systems but also their
pt

customers, who provide information on the timing and quality of their future
demand, benefit from the preorder strategy. Manufacturers of ATO systems
ce

can reduce their inventory levels but still provide high-quality service to their
customers. Hence, many companies in high-tech, car manufacturing and
Ac

white good industry can benefit from the results of this study.

3 Problem Formulation

We consider a manufacturer managing an ATO system with two components

and a single end product. One unit of each component is needed to produce
one unit of the end product. We use index j = 1, 2 for the components. The
manufacturer uses continuous-review base-stock policies to replenish the
component inventories from uncapacitated suppliers. The base-stock level
and the deterministic replenishment lead time for component j are sj and lj,
respectively. Since the component replenishment lead times are relatively
longer than the assembly time, we neglect the assembly time and assume
that assembly is instantaneous. Customer orders/demands occur for the end
product only and describe a stationary Poisson process with a rate λ. Each
customer orders a single unit.

The manufacturer aims to investigate the profitability of a preorder strategy,

which requires that customers place their orders ψ time units before their
actual need. We say that the demand occurs ψ time units after the

t
rip
corresponding order. We call ψ commitment lead time and assume that ψ can
take any value in [ 0, ).

c
The manufacturer pays a commitment cost c per commitment time unit to

us
each customer. Hence, commitment cost per customer is c . In addition to
the commitment cost, the manufacturer pays an inventory holding cost of hj
an
per unit of component j per time unit. The preorder strategy implies that there
is a commitment to deliver each customer order by the end of the commitment
M

lead time ψ, otherwise, demand is backordered and a backordering cost of p

per unit per time unit is paid to the customer. The customer does not accept
ed

delivery before the end of commitment lead time since the product is not
needed before that. The manufacturer’s objective is to find the commitment
pt

lead time, ψ, and the base-stock levels, s1 and s2, which minimize the total
long-run average cost.
ce

In the rest of the paper, without loss of generality, we name the component
Ac

with the longer replenishment lead time as component 1, i.e., j = 1 and the
component with the shorter replenishment lead time as component 2, i.e., j =
2. Refer to Figure 1 for the graphical representation of the ATO system with
l1 l2
.

Next, we derive the expression for the total long-run average cost. This
requires multiple preliminary derivations. First, we define the equivalent serial
system. Then, we derive the expression for the component net inventory
levels and the remnant inventory level. We defer the definition of the latter to
Section 3.3.

3.1 Commitment Lead Time

The manufacturer uses the commitment lead time to improve the operational
efficiency of the system. As stated above, the commitment lead time is the
time difference between the customer order and the corresponding demand,
i.e., the customer places her order ψ time units before her arrival, i.e,
demand. According to Hariharan & Zipkin (1995) and Ahmadi et al. (2017) a

t
single-location inventory system with commitment lead time is equivalent to a

rip
single-location inventory system where the commitment lead time is
subtracted from the component replenishment lead time. The same is valid for

c
the ATO system under consideration.

We define intervals 1 [0, l2 ], 2 [l2 , l1 ] and us 1 2 [0, l1 ] . The

an
lengths of intervals 2 and 1 represent the replenishment lead times of
components 1 and 2 in the equivalent serial system, respectively. By
M

considering commitment lead time and subtracting it from the replenishment

lead time of each component, we can find the updated replenishment lead
ed

times L1 and L2 in the equivalent serial system as follows;

max(0,(l1 l2 ) max(0, l2 )),

pt

L1
ce

and

L2 max(0, l2 ).
Ac

The ATO system and its corresponding equivalent serial system are shown in
l1 l2 , i.e., L1 L2 0
Figure 2. Notice that when , we have . This is
an ideal service situation with zero holding and backordering costs. On the
other hand, when , there are two options for the pair (L1, L2) depending

on whether 1 or 2. We have
 (l1 l2 , l2 ), 1

( L1 , L2 ) (l1 ,0), 2

(0,0), .

From the first two options we observe that, depending on the length of
commitment lead time , one of the component replenishment lead

times depends on ψ. If 1, L2 depends on ψ and if 2, L1 depends on

ψ. Accordingly, we divide our analysis into two cases; when 1 and

2. In the former case we have a two-stage serial inventory system, while

in the latter case we have a single-location inventory system, i.e.,

L2 0 . It is

t
rip
important to remember that ψ is a decision variable. Hence, the optimal case
depends on the system parameters.

c
3.2 Net Inventory Levels

The manufacturer manages the replenishment processes of both

us
components. Hence, the inventory system is managed centrally; the
an
manufacturer has all the information on component net inventory levels and
can coordinate the replenishment processes to avoid having excess inventory
M

of one component and out-of-stock situation for the other component. The
manufacturer would like to avoid such imbalances to the extent possible.
ed

Zipkin (2000) provides a detailed explanation on adjusting component

pt

inventories for avoiding imbalances. For an assembly system with two

components, the component with the longer lead time uses an ordinary base-
ce

stock policy while the policy of the component with the shorter lead time
should be adjusted to avoid the imbalances mentioned above. We define
Ac

D(t1 , t2 ]
as the cumulative customer demand from time t1 to t2. Accordingly,
the net inventory of component 1 at any time t can be written as
IN1 (t ) s1 D(t ( L1 L2 ), t ]
. The net inventory of component 2 should be
adjusted such that the net inventories of both components are equal at time
t L2 IN1 (t L2 ) IN2 (t L2 )
, i.e., . This can be achieved by adjusting the
IOP2 (t ) min(s2 , IN1 (t ) IO1 (t ))
inventory order position for component 2 as ,
where
IO1 (t ) is the portion of the outstanding component 1 replenishment

orders at time t that arrives before time

t L2 .

As defined above,
IO1 (t ) is the portion of the outstanding component 1

replenishment orders at time t that arrives before time

t L2 . Accordingly,
IO1 (t ) is the portion of the outstanding component 1 replenishment orders at

time t that has not arrived before time

t L2 . Refer to Figure 3 for a graphical

representation of these two portions.

t
We define
IO1 (t ) as the outstanding component 1 orders at time t. Since

rip
IO1 (t ) IO1 (t ) IO1 (t ), IOP2 (t ) can alternatively be written as

IOP2 (t ) min(s2 , IN1 (t ) IO1 (t ) IO1 (t )) . Both components are supplied by

c
us
suppliers with infinite capacity. Therefore, every replenishment order is
fulfilled by the suppliers immediately. This implies that the inventory order
an
position, defined as
IOP1 (t ) IN1 (t ) IO1 (t ) , satisfies the equality
IOP1 (t ) s1 , t . Hence, we have IOP2 (t ) min(s2 , s1 IO1 (t )) . In addition,
M

since each outstanding order corresponds to a demand,

IO1 (t ) is equal to the

cumulative demand during

(t ( L1 L2 ), t L2 ] . Hence,
ed

IOP2 (t ) min(s2 , s1 D(t ( L1 L2 ), t L2 ])

.
pt

Using the stationary and independent increments property of the Poisson

ce

process, we can rewrite the net inventory of component 1 at time t as

IN1 (t ) s1 (D(t ( L1 L2 ), t L2 ] D(t L2 , t ]) . As a result, in steady-state the
Ac

component net inventory levels are;

IN1 s1 ( X Y ) (1)

and

IN2 min(s2 , s1 X ) Y , (2)

where X and Y are cumulative demand during L1 and L2, respectively.

3.3 Remnant Inventory Levels

In an ATO system, when a customer order arrives, the ability of the

manufacturer to satisfy the corresponding demand depends on the net
inventory levels of the components. The manufacturer is able to meet the
demand at a specific point of time if the minimum of the net inventories of both

components is positive, i.e.,

min( IN1 , IN2 ) 0 ; otherwise, the demand is

backordered.

t
By definition of the base-stock policy, a customer order for the end product

rip
implies orders for each individual component. If component inventory levels
are positive, one unit of each component is dedicated to the specific customer

c
order. We call the component inventory that is dedicated to specific customer

orders remnant inventory and use

Ij
us
to represent the remnant inventory
level of component j. In our setting, we have the remnant inventory only when
an
one component is available and another one is missing (De Kok 2003).
Hence, at any point in time, there can be at most one component type in
M

remnant inventory. The steady-state remnant inventory level of each

component can be calculated as
ed

I1 max(0, B2 B1 ) max(0,max(0, IN2 ) max(0, IN1 )),

pt

and
ce

I 2 max(0, B1 B2 ) max(0,max(0, IN1 ) max(0, IN2 )).

Ac

Here, B1 and B2 are steady-state backorder levels for components 1 and 2,

respectively. It is obvious that a delayed delivery of one component does not
only result in backordering cost of the end product but also can result in
inventory holding cost for the other component if it is already in stock as
remnant inventory. Next, we determine the expected remnant inventory levels
for both components.
Proposition 1. The expected remnant inventory levels for a centralized ATO
system with two components and Poisson customer demand for the end
product are

s1 s2 s1 x
{ I1} ( y s2 )P2 ( y ) ( s1 s2 x)(1 F2 ( s1 x)) P1 ( x),
x 0 y s2 1

{ I 2 } 0.

Fj ( x) Pj ( x)
Here, and are the cumulative distribution and probability mass
Lj
functions of a Poisson random variable with mean j
, respectively.
According to Proposition 1, when component inventories are controlled

t
rip
centrally by the manufacturer, the orders for component 2 can be
{ I2} 0
synchronized with the orders for component 1 such that .

c
Component 2 has a shorter replenishment lead time than component 1. The
manufacturer needs one unit from both components to satisfy a unit of
demand. The manufacturer can place orders with the supplier of component 2us
an
such that the component 2 never needs to wait for the arrival of component 1.
This is why there is no remnant inventory for component 2. In addition, if
M

s1 s2
, the expected remnant inventory level of component 1 is zero, i.e.,
{ I1} 0 s1 s2
. When , if there is a demand and if there is at least one unit
ed

of component 1, there is also at least one unit of component 2 in the

inventory. This is why there is no remnant inventory for the components.
pt

s1 s2
Except this case with , even under a centralized control, the expected
ce

remnant inventory level of component 1 is positive. Under a decentralized

control scheme both expectations can be positive.
Ac

3.4 Long-run Average Cost

Having the expressions for the net inventory levels and the expected remnant
inventory levels, we are ready to derive the expression for the long-run
average cost. For that purpose, we define Ij as on-hand inventory level of
component j and B as the number of backorders of the end product. We use s

to represent the base-stock level pair (s1, s2). Then, for an arbitrary base-
stock level pair s and commitment lead time ψ we can write the long-run

average cost
Ca ( s, ) as follows

2
Ca ( s, ) h j (I j I j ) pB c .
j 1

Note that we append the subscript a to indicate that this is the conventional

cost of an ATO system.

Ca ( s, ) is the sum of long-run average holding,

backordering and commitment costs. Holding cost of each component is

comprised of on hand inventory holding cost and remnant inventory holding

t
cost. The manufacturer incurs backordering cost if a customer demand is

rip
satisfied more than ψ time units after the corresponding order is given. Finally,
the manufacturer incurs commitment cost. It is important to note that the

c
Ij, Ij
random variables and B depend on and ψ. Our purpose in the

us
s

Ca ( s, )
remainder of this section is to rewrite the expression for such that
an
these dependencies are made explicit.

Firstly, we rewrite
Ca ( s, ) by using the following equalities; I j [ IN j ] , for j =
M

1, 2,
I1 max(0,[ IN2 ] [ IN1 ] ), I 2 max(0,[ IN1 ] [ IN2 ] ) and

B max([ IN1 ] ,[ IN2 ] ) , where [ x]

ed

max(0, x) and [ x] max(0, x) . We have

Ca ( s, ) h1 (
{[ IN ] } {max(0,[ IN ] [ IN ] )})
pt

1 2 1

h ( {[ IN ] } {max(0,[ IN ] [ IN ] )})
2 2 1 2
(3)
ce

p {max([ IN ] ,[ IN ] )} c
1 2

The first and second rows are expected holding costs for on-hand and
Ac

remnant inventory levels of components 1 and 2, respectively. The third row is

the sum of expected backordering and commitment costs. Next, we rely on
Ca ( s, )
simple algebra to obtain the following equivalent expression for . We
refer the reader to the Appendix for the derivations.

2
Ca ( s, ) h j {IN j } ( p h1 h2 ) {max([ IN ] ,[ IN ] )}
1 2 c (4)
j 1
Rosling (1989) and Chen & Zheng (1994) show that a two-component, single
end product ATO system can be reduced to an equivalent two-stage serial
system with modified replenishment lead times Li (refer to Figure 2). In the
equivalent serial system, the local holding costs for stages 1 and 2 are h1 and
h1 h2 , respectively. According to Shang & Song (2003) the long-run average

cost of the serial system can be written as

2
Cs ( s, ) h j {IN j } ( p h1 h2 ) {[ IN ] }
2 c
j 1

Note that we use subscript s to indicate the long-run average cost of the

t
rip
equivalent serial system. Under a centralized control scheme, the expression

for
B , as derived in the Appendix, is as follows:

c
{max([IN ] ,[IN ] )} {[IN ] }. (5)

us
B 1 2 2

This result is consistent with Proposition 1. According to Proposition 1, we

an
{ I2} 0
have , which means that component 2 never waits for the arrival of
component 1 to satisfy a customer demand. Hence, if there is a backorder, it
M

is due to a missing component 2. This is why the expected number of

customer backorders equals the expected number of component 2
ed

backorders.
pt

Although all the terms of

Cs ( s, ) and Ca ( s, ) look exactly the same, there is

a small difference that results from the fact that in the ATO system there is no
ce

holding cost for in-transit inventory, while in the serial system the holding cost
for items in-transit to stage 2 is included. Hence, the total cost of the serial
Ac

h1 L2
system exceeds the total cost of the assembly system by (Zhang 2006).
Ca ( s, ) Cs ( s, ) h1 L2
Then, for an arbitrary , we have . It is a well-
known result that, for an arbitrary , the total cost of the equivalent serial
Cs ( s, )
system can be derived using the recursive method proposed by
h1 L2
Shang & Song (2003). By subtracting and applying some simple algebra
we can rewrite the expression for
Ca ( s, ) in terms of system parameters and

decision variables. We refer the reader to the Appendix for the details.

s1 1
Ca ( s , ) 1 p ( p h1 ) F1 ( s1 s2 1) ( p h1 h2 ) F2 ( s1 t 1) P1 (t )
t s1 s2

s1
s1 p ( p h1 ) F1 ( s1 s2 ) ( p h1 h2 ) F2 ( s1 t ) P1 (t )
t s1 s2 1
(6)
s1 1

2 p ( p h1 h2 ) F2 ( s2 1) F1 ( s1 s2 1) F2 ( s1 t 1) P1 (t )
t s1 s2

p h1
s2 ( p h1 h2 ) F1 ( s1 s2 ) F2 ( s2 ) c

t
p h1 h2

rip
, Fj ( x ) Pj ( x)
When j
and are independent of ψ.

c
4 Characterization of Optimal Control Policies

In this section, we analyze the properties of the cost function us Ca ( s, ) and

an
characterize the optimal control policies.
M

4.1 Optimization of the Component Base-Stock Levels

For an arbitrary , determining the base-stock levels which minimize

ed

Ca ( s, )
is a well-known optimization problem in the multi-echelon inventory
theory literature. We have
pt

Ca ( s , ) min Ca ( s, ),
ce

where
Ca ( s , ) is a tight lower bound for Ca ( s, ) . s represents the optimal
Ac

pair (
s1 , s2 ) corresponding to a . To compute the optimal echelon base-
stock levels, two nested convex functions should be minimized. More
specifically, for a given , we need to solve recursive optimization
equations which have been proposed by Shang & Song (2003). Knowing that
the optimal base-stock levels of a two-component ATO system are equal to
the optimal echelon base-stock levels of the equivalent serial system, we can
formulate the following theorem, which states the inequalities that need to be
satisfied by the optimal base-stock levels.

Theorem 1. In a two-component ATO system with a given commitment lead

time , the optimal base-stock levels of components 1 and 2 are two

non-negative integers
s1 and s2 , which satisfy the inequalities

p h1
F2 ( s2 1) F2 (s2 ),
p h1 h2

and

t
rip
p
1 (s1 1, s2 ) 1 (s1 , s2 ),
p h1 h2

c
Fj ( x) Pj ( x)
where and
are the cumulative distribution and probability mass

functions of a Poisson random variable with mean j

Lj
and us
an
x
p h1
1 ( x, y ) F1 ( x y) F2 ( x n) P1 (n).
p h1 h2 n x y 1
M

Theorem 1 states that the optimal component base-stock levels depend on

the commitment lead time. For each value of ψ, obtaining the optimal base-
ed

stock levels needs two consecutive steps. First, the optimal base-stock level
s2 should be calculated, and then by using this value in the
pt

of component 2,

second inequality, the optimal base-stock level of component 1,

s1 is
ce

calculated. colorredThe optimal base-stock levels of a two-component ATO

system are equal to the optimal echelon base-stock levels of the equivalent
Ac

serial system (Rosling 1989; Shang & Song 2003). Our ATO system and its
equivalent serial system are depicted in Figure 2. It is known that the optimal
base-stock policy for a two-location serial systems can be computed through
minimizing two nested convex functions recursively, starting from the location
closest to the customer, i.e., location where component 2 is stored (Shang
& Song 2003). Theorem 1 is consistent with this result and it provides the
inequalities that need to be satisfied by the optimal component base-stock
levels.

We would like to note that, when 2, we have 2 0 and s2 0 . Hence,

when 2 the manufacturer relies on a buy-to-order (BTO) strategy for

component 2. In this case, Expression (6) reduces to

Ca ((s1 ,0), ) 1 ( p ( p h1 ) F1 (s1 1)) s1 ( p ( p h1 ) F1 (s1 )) c .

Ca ((s1 ,0), ) represents the long-run average cost of a single-location

t
inventory system. We refer to Ahmadi et al. (2017) for detailed analysis of the

rip
single-location system with commitment lead time.

c
With the next result, we prove the behavior of the optimal base-stock levels
with respect to the commitment lead time.

Theorem 2. The optimal base-stock level of component j,

us sj
, is a piecewise-
an
constant and non-increasing function of for j = 1, 2.
M

According to Theorem 2, the manufacturer can operate with less inventory if

commitment lead time gets longer. In fact, for a sufficiently long commitment
ed

lead time, the manufacturer can optimize his long-run average cost by setting
both component base-stock levels to zero. This would imply using a buy-to-
pt

order (BTO) strategy. In Figure 4, we illustrate how the optimal component

base-stock levels change with respect to the commitment lead time.
ce

Next, we are interested in the behavior of the average cost function with
Ac

respect to the commitment lead time. In Figure 5, for a specific parameter

setting with 0.4, h [5, 7], p 20, l [8,5] and c = 6, we draw the average

cost function for multiple base-stock level pairs over 1 2, where

[0,5] [5,8]
1 and 2 . The vertical dotted line at ψ = 5 shows the

separation of the intervals 1 and 2. We also indicate the tight lower bound

of these cost functions. This tight lower bound is

Ca ( sψ , ) . This is the
function that needs to minimized for obtaining the optimal commitment lead

time. Note that the function

Ca ( sψ , ) does not have a well-defined behavior.

This is why finding the optimal commitment lead time requires a different
approach.

4.2 Unit Commitment Cost Thresholds

In the previous section, we analyzed the properties of the average cost

function and the optimal base-stock levels when the commitment lead time is
fixed to a certain value. In this section, we define unit commitment cost
thresholds which are used to identify the optimal commitment lead time.

t
rip
First, we define a set, which helps us with providing a better explanation for
the subsequent analysis. The set is related to the end points of the sub-

c
domains of , which are
0,l2 and l .
1

Definition 1. Let 1 [0, l2 ]

and 2 [l2 , l1 ] us
be the sub-domains of the
an
commitment lead times in a two-component ATO system. Then, we define a
{0, l2 , l1}
terminal set as .
M

Using the data in Figure 4, we have {0,5,8} .

ed

Next, we define three unit commitment cost thresholds. Each definition uses
two elements of the set . We use cik, i < k, to represent the threshold that
pt

uses ith and kth elements of the set . We have

ce

Ca ( s 0 , 0) Ca ( s l2 , l2 )
c12 c
l2
Ac

Ca ( s 0 , 0) Ca ( s l1 , l1 )
c13 c
l1
Ca ( s l2 , l2 ) Ca ( s l1 , l1 )
c23 c
(l1 l2 )

Although it seems like the unit commitment cost thresholds depend on unit
commitment cost, c, – c comes out of the fractions in each expression and it
cancels out c . Remember that s is the optimal base-stock level vector for
commitment lead time ψ. Note that we have defined a unit commitment cost
threshold for each pair of points in the terminal set. The unit commitment cost
thresholds ensure the equality of the long-run average costs when the
commitment lead time is set to the corresponding pair of points. For example,

when
c c12 , we have Ca ( s0 ,0) Ca ( sl , l2 ) . In Figure 6, we provide three
2

numerical examples with the same parameter settings where the unit
commitment costs are set to c12, c13 and c23.

With the following lemma, we define multiple different relationships among

unit commitment cost thresholds.

t
rip
Lemma 1. Consider a two-component ATO system and unit commitment cost
thresholds c12, c13 and c23. The following results hold;

c
(l1 l2 )c23 l1c13 l2c12 ,
1.

2. either
c12 c13 c23 or c23 c13 c12 holds. us
an
Remark 1. When the component replenishment lead times are the same, i.e.,
c12 c13 and c23
M

l1 = l2, we have . In this case, there is a unique unit

c0 c12 c13
commitment cost threshold which is exactly the same as unit
ed

commitment cost threshold in the single-location problem (see Ahmadi

et al. (2017) for more details).
pt

Next, using the unit commitment cost thresholds, we define lower bounds.
ce

These bounds are used for obtaining the optimal commitment lead time.

ti , t j CLBij ( )
Ac

Definition 2. For all and ti < tj, define as

CLBij ( ) Ca ( sti , ti ) (cij c)( ti ),

CLBij ( ) ti , Ca ( sti , ti )
where is a linear function which connects the points and
tj
t j , Ca ( s , t j )
. We call this function a linear lower bound.
Based on this definition, in a two-component ATO system, there exists three
CLB12 ( ), CLB23 ( ) CLB13 ( )
linear lower bounds; , and . The slope of each linear
lower bound depends on the corresponding unit commitment cost thresholds.
In a two-component ATO system, the lower bounds form a triangle. From

Lemma 1, we know that either

c12 c13 c23 or c23 c13 c12 holds. When
c12 c13 c23 holds the triangle points up and when c23 c13 c12 holds the

triangle points down. Refer to Figure 7 for examples. Since in the former case
there are two potential minimum points and in the latter case there are three
potential minimum points we call them Case 2 and Case 3, respectively. Note

t
that when
c12 c13 c23 , we have a straight line with two potential minimum

rip
points.

c
Given an ATO system with a specific parameter setting, we can determine the
Case number by calculating two of the three unit commitment cost thresholds

only. Suppose that c12 and c13 are calculated. Then we know that if us c12 c13
an
holds, we have Case 2; otherwise, we have Case 3.
M

4.3 Optimization of the Commitment Lead Time

In this section we characterize the optimal preordering strategy. First, we

ed

formulate a conjecture which identifies a lower bound on

Ca ( s , ) . If the

lower bound is known, we can also indicate which ψ values are candidates for
pt

being optimal. Then, it is enough to evaluate the cost

Ca ( s , ) only at these
ce

points.

Conjecture 1. In a two-component ATO system with commitment lead time ψ

Ac

CLBij ( ) (ti , Ca ( sti , ti ))

and lower bounds that connect the points and
tj
(t j , Ca ( s , t j )) ti , t j
, where such that ti < tj, for all 1 we have

CLB13 ( ) Ca ( s , ) c12 c13 c23

1. if or
CLB12 ( ) Ca ( s , ) c23 c13 c12
2. if .
We present the above result as a conjecture since we do not have a proof for
it. We test the correctness of the conjecture by generating random instances.
Our belief on the correctness of the conjecture is confirmed in every single
instance. We provide more detail on our numerical setup and results in
Section 5.

Theorem 3. In a two-component ATO system with commitment lead time ψ

CLB13 ( ) Ca ( s , ) c12 c13 c23 and CLB ( ) Ca ( s , ) when
with when 12

*
c23 c13 c12 , the optimal commitment lead time is either zero or equal to
*
the replenishment lead time of one of the components, i.e.,
0, l2 , l1 .

t
rip
Theorem 3 characterizes the optimal commitment lead time. When the
optimal commitment lead time is equal to zero, the preorder strategy is not

c
beneficial to the ATO system and the optimal component ordering strategy is

us
a buy-to-stock (BTS) strategy. When the optimal commitment lead time is
equal to l2, the optimal strategy of components 1 and 2 are BTS and buy-to-
an
order (BTO) strategies, respectively. When the optimal commitment lead time
is equal to l1, the optimal strategy for both components is an BTO strategy.
M

Next, we use the unit commitment cost thresholds to provide a guideline on

ed

when to set the commitment lead time to

0,l2 or l . For an arbitrary parameter
1

setting, the optimal commitment lead time can be determined using the results
pt

in Theorem 4.
ce

Theorem 4. In a two-component ATO system with commitment lead time ψ

CLB13 ( ) Ca ( s , ) c12 c13 c23 and
and unit commitment cost c, with when
Ac

*
CLB12 ( ) Ca ( s , ) c23 c13 c12
when , the optimal commitment lead time
is

0 if Case 2, c c13 or Case 3, c c12

*
l1 if Case 2, c c13 or Case3, c c23
l2 if Case 3, c23 c c12
where, we have Case 2 if
c12 c13 c23 and Case 3 if c23 c13 c12 .

Recall that for any parameter setting we end up with either Case 2 or Case 3.
In Cases 2 and 3 there are two and three potential optimal solutions,
respectively. In addition to indicating the optimal commitment lead time,
Theorem 4 tells us when the preordering strategy is beneficial or not. The
*
preordering strategy is not beneficial if 0 . Hence, if c c13 in Case 2, the

preordering strategy is not beneficial. In Case 3, if

c c12 , the preordering

strategy should not be preferred.

t
Theorem 4 specifies the optimal component ordering strategy for different

rip
values of the unit commitment cost. In Figure 8 we provide a visual
representation of the results in Theorem 4. More specifically, we identify the

c
values of unit commitment cost for which BTO or BTS strategy is optimal. In

us
Case 2, for any value of unit commitment cost the manufacturer should use
the same strategy for both components, i.e., the switch from BTO to BTS
an
happen at the same threshold value c13. In Case 3, based on the value of unit
commitment cost, the manufacturer may use the same or different strategies
M

c23 c c12
for the components. For example, when , the optimal strategy is
BTS and BTO for component 1 and 2, respectively.
ed

5 Numerical Results
pt

In this section, we conduct a numerical experiment to confirm the correctness

ce

of Conjecture 1. A sample of examples is randomly generated with parameter

values drawn from uniform distributions as follows;
~ U [0.1,10.1], h1 ~ U[0.1,10], h2 ~ U[0.1,10], p ~ U[15,50], l1 ~ U [.5,20], l2 ~ U [0.1, ll ]
Ac

, and c ~ U [0.1, h 3] , where x ~ U [a, b] means that the values of parameter x

are chosen from a Uniform distribution in interval [a, b] . We consider 8000
different instances by generating different combinations of parameters
, h1 , h2 , p, l1 , l2
and c.
 * * *
We rely on enumeration for calculating s1 , s2 and . We choose a step size

of 0.01 for ψ. We change the value of ψ and calculate

s1 , s2 and C (sψ , ) . We

use Theorem 1 for calculating the base-stock levels. The enumeration stops

when both
s1 and s2 equal zero. We choose the solution that gives the

lowest cost. The enumeration results and the results obtained through our
analytical derivations are exactly the same. The output of all 8000 instances
are consistent with Conjecture 1. Hence, the correctness of the conjecture is
confirmed.

We report multiple results, in case other researchers would like to reproduce

t
rip
them. Based on the Case number and the optimal commitment lead time, we
organize the results in five main categories. The first two categories are Case
* *

c
2 with
{0, l1} and the other three categories are Case 3 with {0, l2 , l1} .

us
Three samples of each category are presented in Table 4. For each sample,
parameter values, three unit commitment cost thresholds, Case number, cost
an
value at each point in the terminal set, the optimal base-stock levels, optimal
commitment lead time and minimum cost are reported.
M

At the bottom of Table 4, we add two special cases when the replenishment
*
lead times of both components are equal,
l1 l2 l and {0, l} . For each
ed

special cases two samples are reported. Note that when

l1 l2 l , the
pt

problem could be presented as a single-location problem with replenishment

lead time l.
ce
Ac

6 Conclusion

We consider a manufacturer who operates a two-component single end

product ATO system. We investigate a centralized control scheme with a
preorder strategy. The time from a customer’s order until the date the end
product is actually needed is called the commitment lead time. Under the
preorder strategy, a commitment cost should be paid to the customer. This
cost is increasing in the length of the commitment lead time. The
manufacturer uses a base-stock policy to replenish component inventories.

The manufacturer aims to find the optimal component base-stock levels and
the optimal commitment lead time such that the long-run average cost
consisting of component inventory holding cost, backordering cost and
commitment cost is minimized. For an arbitrary commitment lead time, we
determined the optimal base-stock levels through optimizing two nested
convex functions recursively. We find that the optimal base-stocks are
piecewise constant and non-increasing in the commitment lead time.

t
rip
We conjecture that the optimal commitment lead time is either zero or equal to
the replenishment lead time of one of the components. When the optimal
commitment lead time is zero, the preorder strategy is not beneficial. The

c
manufacturer should choose a strategy that is similar in spirit to a make-to-

us
stock procurement strategy. In our context, this strategy is called buy-to-stock
(BTS) strategy. When the optimal commitment lead time equals the lead time
an
of the component with the shorter lead time, the optimal strategy for this
component is buy-to-order (BTO) and it is buy-to-stock for the other
M

component. On the other hand, if the optimal commitment lead time equals
the lead time of the component with the longer lead time, the optimal strategy
ed

is the BTO strategy for both components. We determine the conditions under
which one of these three cases holds. We define three unit commitment cost
pt

thresholds. These thresholds enable the manufacturer to evaluate the benefit

of the preorder strategy under different unit commitment costs. By calculating
ce

these thresholds once for an arbitrary parameter setting, we can find the
optimal solutions under different commitment unit commitment costs without
Ac

solving the optimization problem again.

Our work can be extended in multiple different ways. An obvious extension is

to consider an ATO system with more than two components. Our result that
the optimal commitment lead time is either zero or equal to one of the
components lead times will still hold. Another extension can consider serial
systems with commitment lead times among all stages. Analyzing the effect of
commitment lead time on the customer waiting time and the optimization of
commitment lead time under service level constraints are other possible
research directions. Studying other forms of the commitment cost (e.g., non-
linear in commitment lead time) could be another interesting extension.

t
c rip
us
an
M
ed
pt
ce
Ac
References

Ahmadi, T., Atan, Z., de Kok, T. & Adan, I. (2017), ‘Optimal control policies in
a production firm with commitment cost’, Working paper.

Atan, Z., Ahmadi, T., Stegehuis, C., de Kok, T. & Adan, I. (2017), ‘Assemble-
to-order systems: A review’, European Journal of Operational Research.

Benbitour, M. H. & Sahin, E. (2015), ‘Evaluation of the impact of uncertain

advance demand information on production/inventory systems’, IFAC-
PapersOnLine 48(3), 1738–1743.

t
rip
Benjaafar, S., Cooper, W. L. & Mardan, S. (2011), ‘Production-inventory
systems with imperfect advance demand information and updating’, Naval
Research Logistics 58(2), 88–106.

c
us
Benjaafar, S. & ElHafsi, M. (2006), ‘Production and inventory control of a
single product assemble-to-order system with multiple customer classes’,
an
Management Science 52(12), 1896–1912.
M

Bernstein, F. & DeCroix, G. A. (2014), ‘Advance demand information in a

multiproduct system’, Manufacturing & Service Operations Management
ed

17(1), 52–65.

Chen, F. & Zheng, Y.-S. (1994), ‘Lower bounds for multi-echelon stochastic
pt

inventory systems’, Management Science 40(11), 1426–1443.

ce

Claudio, D. & Krishnamurthy, A. (2009), ‘Kanban-based pull systems with

advance demand information’, International Journal of Production Research
Ac

47(12), 3139–3160.

De Kok, T. G. (2003), Evaluation and optimization of strongly ideal assemble-

to-order systems, in ‘Stochastic modeling and optimization of manufacturing
systems and supply chains’, Springer, pp. 203–242.
Gallego, G. & Özer, Ö. (2001), ‘Integrating replenishment decisions with
advance demand information’, Management Science 47(10), 1344–1360.

Gallego, G. & Özer, Ö. (2003), ‘Optimal replenishment policies for

multiechelon inventory problems under advance demand information’,
Manufacturing & Service Operations Management 5(2), 157–175.

Gayon, J.-P., Benjaafar, S. & De Véricourt, F. (2009), ‘Using imperfect

advance demand information in production-inventory systems with multiple
customer classes’, Manufacturing & Service Operations Management
11(1), 128–143.

t
rip
Hariharan, R. & Zipkin, P. (1995), ‘Customer-order information, leadtimes, and
inventories’, Management Science 41(10), 1599–1607.

c
us
Iida, T. (2015), ‘Benefits of leadtime information and of its combination with
demand forecast information’, International Journal of Production Economics
an
163, 146–156.

Iida, T. & Zipkin, P. (2010), ‘Competition and cooperation in a two-stage

M

supply chain with demand forecasts’, Operations Research 58(5), 1350–1363.

ed

Iyer, A. V., Deshpande, V. & Wu, Z. (2003), ‘A postponement model for

demand management’, Management Science 49(8), 983–1002.
pt

Karaesmen, F., Liberopoulos, G. & Dallery, Y. (2004), ‘The value of advance

ce

demand information in production/inventory systems’, Annals of Operations

Research 126(1-4), 135–157.
Ac

Li, C. & Zhang, F. (2013), ‘Advance demand information, price discrimination,

and preorder strategies’, Manufacturing & Service Operations Management
15(1), 57–71.

Liberopoulos, G. (2008), ‘On the tradeoff between optimal order-base-stock

levels and demand lead-times’, European Journal of Operational Research
190(1), 136–155.
Liberopoulos, G. & Koukoumialos, S. (2008), ‘On the effect of variability and
uncertainty in advance demand information on the performance of a make-to-
stock supplier’, MIBES Transactions International Journal 2(1), 95–114.

Lutze, H. & Özer, Ö. (2008), ‘Promised lead-time contracts under asymmetric

information’, Operations Research 56(4), 898–915.

McCardle, K., Rajaram, K. & Tang, C. S. (2004), ‘Advance booking discount

programs under retail competition’, Management Science 50(5), 701–708.

Özer, Ö. (2003), ‘Replenishment strategies for distribution systems under

t
advance demand information’, Management Science 49(3), 255–272.

rip
Özer, Ö. & Wei, W. (2004), ‘Inventory control with limited capacity and

c
advance demand information’, Operations Research 52(6), 988–1000.

us
Papier, F. (2016), ‘Supply allocation under sequential advance demand
information’, Operations Research 64(2), 341–361.
an
Papier, F. & Thonemann, U. W. (2010), ‘Capacity rationing in stochastic rental
M

systems with advance demand information’, Operations Research 58(2), 274–

288.
ed

Rosling, K. (1989), ‘Optimal inventory policies for assembly systems under

random demands’, Operations Research 37(4), 565–579.
pt

Shang, K. H. & Song, J.-S. (2003), ‘Newsvendor bounds and heuristic for
ce

optimal policies in serial supply chains’, Management Science 49(5), 618–

638.
Ac

Szekli, R. (2012), Stochastic ordering and dependence in applied probability ,

Vol. 97, Springer Science & Business Media.

Tan, T., Güllü, R. & Erkip, N. (2007), ‘Modelling imperfect advance demand
information and analysis of optimal inventory policies’, European Journal of
Operational Research 177(2), 897–923.
Tang, C. S., Rajaram, K., Alptekinoğlu, A. & Ou, J. (2004), ‘The benefits of
advance booking discount programs: Model and analysis’, Management
Science 50(4), 465–478.

Thonemann, U. W. (2002), ‘Improving supply-chain performance by sharing

advance demand information’, European Journal of Operational Research
142(1), 81–107.

Wang, T. & Toktay, B. L. (2008), ‘Inventory management with advance

demand information and flexible delivery’, Management Science 54(4), 716–
732.

t
rip
Wang, Y. & Tomlin, B. (2009), ‘To wait or not to wait: Optimal ordering under
lead time uncertainty and forecast updating’, Naval Research Logistics

c
56(8), 766–779.

us
Wijngaard, J. & Karaesmen, F. (2007), ‘Advance demand information and a
an
restricted production capacity: On the optimality of order base-stock policies’,
OR Spectrum 29(4), 643–660.
M

Zhang, F. (2006), ‘Competition, cooperation, and information sharing in a two-

echelon assembly system’, Manufacturing & Service Operations Management
ed

8(3), 273–291.
pt

Zhu, K. & Thonemann, U. W. (2004), ‘Modeling the benefits of sharing future

demand information’, Operations Research 52(1), 136–147.
ce

Zipkin, P. H. (2000), Foundations of inventory management, McGraw Hill,

Ac

Boston, Massachusetts.
Appendix

For concision, define the following mathematical notations: [ x] max

( x,0), [ x] max ( x,0), x y max(x, y) and x y min(x, y).

Proof of Proposition 1

Under centralized control scheme, the component net inventory levels are
IN1 s1 ( X Y ) and IN2 (s2 (s1 X )) Y , where X and Y are two

independent Poisson random variables with mean

L1 and L2 , respectively.

t
The remnant inventory levels for components 1 and 2 are defined as

rip
I1 [ IN 2 ] [ IN1 ] I2 [ IN1 ] [ IN 2 ]
and , respectively. We have

c
X Y s1 , X Y s1 ,
[ IN1 ] [ s1 Y] (A1)

us
X
0, X Y s1.

s2 , s1 s2 , Y s2 ,
an
Y X
X Y s1 , X s1 s2 , X Y s1 ,
[ IN 2 ] [( s2 ( s1 X )) Y ] (A2)
0, X s1 s2 , Y s2 ,
M

0, X s1 s2 , X Y s1.

Then, the component 1 remnant inventory level is

ed

s1 s2 X, X s1 s2 , X Y s1 , Y s2
pt

I1 [ IN 2 ] [ IN1 ] Y s2 , X s1 s2 , Y s2 , X Y s1 (A3)
0, otherwise
ce

Based on Equation (A3), the expected component 1 remnant inventory level

Ac

can be calculated as

s1 s2 s1 x s1 s2
{ I1} ( y s2 )P2 ( y ) P1 ( x) ( s1 s2 x)P2 ( y ) P1 ( x)
x 0 y s2 1 x 0 y s1 x 1

s1 s2 s1 x
( y s2 )P2 ( y ) ( s1 s2 x)P2 ( y ) P1 ( x)
x 0 y s2 1 y s1 x 1

s1 s2 s1 x
( y s2 )P2 ( y ) ( s1 s2 x)(1 F2 ( s1 x)) P1 ( x)
x 0 y s2 1
 I2 [ IN1 ] [ IN 2 ] 0
Similarly, using Equations (A1) and (A2), we get .

Hence, the expected component 2 remnant inventory level is

{ I2} 0 .

Simplification of Equation (3)

First, we rewrite Equation (3) using the notation defined at the beginning of
the Appendix. We obtain the following expression:

Ca ( s, ) h1 IN1 IN 2 IN1

h2 IN 2 IN1 IN 2

t
rip
p [ IN1 ] [ IN 2 ] c

c
[ IN1 ] [ IN1 ] [ IN2 ] [ IN2 ]
By adding and to the first and
second round brackets, respectively, and applying linearity property of the
expectation operator, we obtain us
an
Ca ( s, ) h1 [ IN1 ] [ IN1 ] [ IN 2 ] [ IN1 ] [ IN1 ]
M

h2 [ IN 2 ] [ IN 2 ] [ IN1 ] [ IN 2 ] [ IN 2 ]

p [ IN1 ] [ IN 2 ] c
ed

For all x, y , [ x] [ x] x and [ x y] y x y (Szekli 2012). Then the

pt

last expression can be rewritten as follows:

ce

Ca ( s, ) h1 IN1 [ IN1 ] [ IN2 ] h2 IN 2 [ IN1 ] [ IN2 ] p [ IN1 ] [ IN2 ] c

h1 IN1 h2 IN 2 ( p h1 h2 ) [ IN1 ] [ IN 2 ] c
Ac

Derivation of Equation (5)

The number of backorders depends on the component stock-out levels, i.e.,

[ IN1 ] [ IN 2 ]
and . We have

B [ IN1 ] [ IN2 ] .
From Equations (A1) and (A2),
[ IN1 ] [ IN2 ] can be derived based on its

sub-domains as follows:

Y s2 , X s1 s2 , Y s2
[ IN1 ] [ IN 2 ] X Y s1 , X s1 s2 , X Y s1 (A4)
0, otherwise

By comparing Equations (A2) and (A4), we obtain

[ IN1 ] [ IN2 ] [ IN2 ] . As a

result,
B [ IN2 ] .

Derivation of Expression (6)

t
rip
In this derivation, we rely on the recursive procedure proposed by Shang
Cs ( s, ) and

c
& Song (2003) for serial systems. We use their expression for

us
given that
Ca ( s, ) Cs ( s, ) h1 2 , we define Ca ( s, ) Ca ( s, ) c and

write
Ca ( s, ) as follows:
an
Ca ( s, ) h1 ( s1 1 ) h2 s2 1 s1 s2 1 F1 s1 s2 1 2
M

( p h1 h2 ) 2 ( s2 ) F1 s1 s2 2 (s1 n) P1 (n) h1 2
n s1 s2 1
ed

From Shang & Song (2003), we know that

m m mFj (m) F (m 1) m 1 m Fj (m) 1
and .
pt

j j j j j

Hence, we can rewrite

Ca ( s, ) as
ce
Ac
Ca ( s , ) h1 ( s1 1 2 ) h2 s2 ( 1 ( s1 s2 ) F1 ( s1 s2 ) 1) F1 ( s1 s2 ) 1 2

( p h1 h2 )(( 2 s2 s2 F2 ( s2 ) 2 F2 ( s2 1)) F1 ( s1 s2 ))

( p h1 h2 ) ( 2 ( s1 n) ( s1 n) F2 ( s1 n) 2 F2 ( s1 n 1)) P1 (n)
n s1 s2 1

h1 ( s1 1 2 ) h2 s2 1 ( s1 s2 ) 2

( p h1 h2 )(( 2 s2 s2 F2 ( s2 ) 2 F2 ( s2 1)) F1 ( s1 s2 ))

( p h1 h2 ) ( 2 ( s1 n) ( s1 n) F2 ( s1 n) 2 F2 ( s1 n 1)) P1 (n)
n s1 s2 1

h1 ( s1 1 2 ) h2 s1 1 2 ( s1 s2 ) F1 ( s1 s2 ) F ( s1 s2 1)
1 1

( p h1 h2 )(( 2 s2 s2 F2 ( s2 ) 2 F2 ( s2 1)) F1 ( s1 s2 ))

( p h1 h2 ) ( s1 ) P1 (n) nP1 (n) s1 F2 ( s1 n) P1 (n)

t
2

rip
n s1 s2 1 n s1 s2 1 n s1 s2 1

( p h1 h2 ) nF2 ( s1 n) P1 (n) 2 F2 ( s1 n 1) P1 (n)

c
n s1 s2 1 n s1 s2 1

(h1 h2 )( s1 ) h2 ( s1 s2 ) F1 ( s1 s2 ) F ( s1 s2 1)

us
1 2 1 1

( p h1 h2 )(( 2 s2 s2 F2 ( s2 ) 2 F2 ( s2 1)) F1 ( s1 s2 ))
an
( p h1 h2 ) ( 2 s1 )(1 F1 ( s1 s2 )) 1 (1 F1 ( s1 s2 1)) s1 F2 ( s1 n) P1 (n)
n s1 s2 1

( p h1 h2 ) F2 ( s1 n 1) P1 (n) F2 ( s1 n 1) P1 (n) F2 ( s2 1) P1 ( s1 s2 )
M

1 2 2
n s1 s2 n s1 s2
ed
pt
ce
Ac
 (h1 h2 )( s1 1 2 ) h2 ( s1 s2 ) F1 ( s1 s2 ) F ( s1 s2 1)
1 1

( p h1 h2 )(( 2 s2 s2 F2 ( s2 ) 2 F2 ( s2 1)) F1 ( s1 s2 ))
( p h1 h2 ) ( s1 1 2 ) ( 2 s1 ) F1 ( s1 s2 ) F ( s1 s2 1)
1 1 2 F2 ( s2 1) P1 ( s1 s2 )

( p h1 h2 ) s1 F2 ( s1 n) P1 (n) 1 F2 ( s1 n 1) P1 (n) 2 F2 ( s1 n 1) P1 (n)

n s1 s2 1 n s1 s2 n s1 s2

p ( s1 1 2 ) h2 ( s1 s2 ) F1 ( s1 s2 ) F ( s1 s2 1)
1 1

( p h1 h2 )(( 2 s2 s2 F2 ( s2 ) 2 F2 ( s2 1)) F1 ( s1 s2 ))
( p h1 h2 ) ( 2 s1 ) F1 ( s1 s2 ) 1 1F ( s1 s2 1) 2 F2 ( s2 1) P1 ( s1 s2 )

( p h1 h2 ) s1 F2 ( s1 n) P1 (n) 1 F2 ( s1 n 1) P1 (n) 2 F2 ( s1 n 1) P1 (n)

n s1 s2 1 n s1 s2 n s1 s2

p ( s1 ) h2 ( s1 s2 ) F1 ( s1 s2 ) F ( s1 s2 1)

t
1 2 1 1

rip
( p h1 h2 ) ( s1 s2 s2 F2 ( s2 )) F1 ( s1 s2 ) F ( s1 s2 1)
1 1 2 F2 ( s2 1) F1 ( s1 s2 1)

( p h1 h2 ) s1 F2 ( s1 n) P1 (n) F2 ( s1 n 1) P1 (n) F2 ( s1 n 1) P1 (n)

c
1 2
n s1 s2 1 n s1 s2 n s1 s2

p ( s1 1 2 ) ( p h1 ) ( s1 s2 ) F1 ( s1 s2 )
( p h1 h2 ) s2 F2 ( s2 )) F1 ( s1 s2 ) 2
1 1

F2 ( s2 1) F1 ( s1 s2 1) us
F ( s1 s2 1)
an
( p h1 h2 ) s1 F2 ( s1 n) P1 (n) 1 F2 ( s1 n 1) P1 (n) 2 F2 ( s1 n 1) P1 (n)
n s1 s2 1 n s1 s2 n s1 s2
M

By categorizing in terms of μ1, s1, μ2, and s2, and adding the commitment cost
ed

c we obtain
pt
ce
Ac
Ca ( s , ) p 1 ( p h1 ) 1 F1 ( s1 s2 1) ( p h1 h2 ) 1 F2 ( s1 n 1) P1 (n)
n s1 s2

ps1 ( p h1 ) s1 F1 ( s1 s2 ) ( p h1 h2 ) s1 F2 ( s1 n) P1 (n)
n s1 s2 1

p 2 ( p h1 h2 ) 2 F2 ( s2 1) F1 ( s1 s2 1) ( p h1 h2 ) 2 F2 ( s1 n 1) P1 (n)
n s1 s2

( p h1 h2 ) s2 F2 ( s2 ) F1 ( s1 s2 ) ( p h1 ) s2 F1 ( s1 s2 ) c
s1 1

1 p ( p h1 ) F1 ( s1 s2 1) ( p h1 h2 ) F2 ( s1 n 1) P1 (n)
n s1 s2

s1
s1 p ( p h1 ) F1 ( s1 s2 ) ( p h1 h2 ) F2 ( s1 n) P1 (n)
n s1 s2 1

t
rip
s1 1

2 p ( p h1 h2 ) F2 ( s2 1) F1 ( s1 s2 1) F2 ( s1 n 1) P1 (n)
n s1 s2

c
p h1
s2 ( p h1 h2 ) F1 ( s1 s2 ) F2 ( s2 ) c
p h1 h2

us
an
PROOF OF THEOREM 1
M

For an arbitrary 1, a two-component ATO system can be reduced to a

two-stage serial system with modified lead times (Rosling 1989). In the two-
ed

stage serial system, the lead times are Lj and the installation holding costs for
h1 h1 h2 h1 h2
stages 1 and 2 are and , respectively. Since in the
pt

assembly system, there is no in-transit inventory cost for the components, the
total cost of the serial system exceeds the total cost of the ATO system by
ce

h1 2 , where μ is mean lead time demand of stage 2. Note that for any
2

1 , h1 2 is a constant with respect to s , it has no impact on optimization of

Ac

base-stock levels. To find optimal base-stock levels corresponding to an

arbitrary 1, we use the recursive optimization equations proposed by

Shang & Song (2003).

C3 ( x) ( p h2 )[ x] C j 1 ( x)
Set . For j = 2, 1, given , compute
Cˆ j ( x) h j x C j 1 ( x)
C j ( y) Cˆ j ( y D j )
sj argmin C j ( y )
C j ( x) C j s j x

C j (.)
Each is a convex function with respect to y and has a finite minimum
point. Let j = 2, then

Cˆ 2 ( x) h2 x C3 ( x) h2 x ( p h2 )[ x]
C2 ( y ) Cˆ 2 ( y D2 ) h2 ( y D2 ) ( p h2 )[ y D2 ] h2 ( y 2 ) ( p h2 ) [ D2 y]

t
rip
h2 ( y 2 ) ( p h2 ) (x y )P2 ( x) h2 ( y 2 ) ( p h2 ) 2 ( y)
x y

c
where
j ( y)
x y
(x y )Pj ( x)
. Then, we can calculate us C2 ( y 1) as follows.
an
C2 ( y 1) h2 ( y 1 2 ) ( p h2 ) 2 ( y 1) h2 h2 ( y 2 ) ( p h2 )( 2 ( y) F2 ( y) 1)
C2 ( y) h2 ( p h2 )( F2 ( y) 1)
M

C2 ( y) is convex with respect to y. Hence, we find the optimal value of y by

ed

C2 ( y)
calculating the difference function y
. The optimal y is the smallest y
pt

C2 ( y) 0
that satisfies the inequality y
. Hence, we have
ce

y C2 ( y) C2 ( y 1) C2 ( y) h2 ( p h2 )( F2 ( y) 1) 0
Ac

The optimal y is the smallest integer number that satisfies

p h2 h2 p h1
F2 ( y)
p h2 p h1 h2 . In other words,

p h1
s2 y 0 : F2 ( y 1) F2 ( y )
p h1 h2

In addition, we have
C2 ( x) C2 ( s2 x) Cˆ 2 ( s2 x) D2

h2 ( s2 x) D2 ( p h2 ) ( s2 x) D2

h2 ( s2 x) 2 ( p h2 ) D2 ( s2 x)

h2 ( s2 x) 2 ( p h2 ) 2 s2 x

We derive similar expression for j = 1. We refer the reader to Shang

( y)
& Song (2003) for the alternative expressions for j
.

Cˆ1 ( x) h1 x C2 ( x) h1 x h2 ( s2 x) 2 ( p h2 ) 2 s2 x
C1 ( y ) Cˆ1 ( y D1 ) h1 ( y D1 ) h2 ( s2 ( y D1 )) ( p h2 ) ( s2 ( y D1 ))

t
2 2

rip
h1 ( y 1 ) h2 ( s2 ( y D1 )) 2 ( p h2 ) 2 s2 ( y D1 )

h1 ( y 1 ) h2 s2 1 y s2 1 F1 y s2 1 2

c
( p h2 ) 2 ( s2 ) F1 y s2
n y s2 1
2

us
( y n) P1 (n)
an
Then, we have
M

C1 ( y 1) h1 ( y 1 1 ) h2 s2 1 y 1 s2 1 F1 y 1 s2 1 2
ed

( p h2 ) 2 ( s2 ) F1 y 1 s2 2 ( y 1 n) P1 (n)
n y 1 s2 1
pt

C1 ( y ) h1 h2 1 F1 ( y s2 ) ( p h2 ) F2 ( y n) P1 (n) 1 F1 ( y s2 )
n y s2 1
ce

C2 ( y) C1 ( y)
Ac

Similar to y
, we calculate the difference function y
as
 y C1 ( y ) C1 ( y 1) C1 ( y )

h1 h2 1 F1 ( y s2 ) ( p h2 ) F2 ( y n) P1 (n) 1 F1 ( y s2 )
n y s2 1

h1 h2 1 F1 ( y s2 ) ( p h2 h1 ) F2 ( y n) P1 (n) 1 F1 ( y s2 )
n y s2 1

p ( p h1 ) F1 ( y s2 ) ( p h1 h2 ) F2 ( y n) P1 (n)
n y s2 1
y
p ( p h1 ) F1 ( y s2 ) ( p h1 h2 ) F2 ( y n) P1 (n)
n y s2 1

t
C ( y) 0

rip
The optimal y is the smallest integer to satisfy the inequality y 1
.
Hence, the optimal y satisfies

c
y
p h1 p
p h1 h2
F1 ( y s2 )
n y s2 1
F2 ( y n) P1 (n)
p
us
h1 h2
.
an
Define 1 ( y, s2 ) as

y
M

p h2
1 ( y, s2 ) F1 ( y s2 ) F2 ( y n) P1 (n).
p h1 h2 n y s2 1
ed

Hence, we have

p
pt

s1 y 0 : 1 ( y 1, s2 ) 1 ( y, s2 ) .
p h1 h2
ce

Proof of Theorem 2
Ac

We need to show that for all

y 0 , , F2 ( y) is non-decreasing in y and ψ,
y , , ( y, s2 )
and for all 0 1 is non-decreasing in y and ψ.

For all , the optimal component 2 base-stock level is

p h1
s2 y 0 : F2 ( y 1) F2 ( y )
p h1 h2
 F2 ( y) is the cumulative distribution function of a
We would like to note that

Poisson random variable with mean 2

L2 , where L depends on ψ. Hence,
2

although
F2 ( y) seems to be a function of y only, it is also a function of ψ by

definition. For a single location system Ahmadi et al. (2017) prove that 2
F ( y) is

non-decreasing in y and ψ. This result is enough to conclude that

s2 is non-

increasing in ψ. We refer to Ahmadi et al. (2017) for the details.

Next, we prove the result for

s1 . We use the following lemma.

t
Lemma 2. 1 ( y, s2 ) is non-decreasing in y and ψ for all y 0 , .

rip
Proof. We show that the first order difference function of 1 ( y, s2 ) with

c
( y, s2 ) ( y, s2 ) with
respect to y, defined as , and the first derivative of

us
y 1 1

d
1 ( y, s2 )
respect to ψ, defined as d are non-negative. We start with
an
( y, s2 )
y 1
. For all , we have
M

y 1 ( y, s2 ) 1 ( y 1, s2 ) 1 ( y, s2 )
y 1
p h1
ed

1 ( y 1, s2 ) F1 ( y 1 s2 ) F2 ( y 1 n) P1 (n)
p h1 h2 n y 1 s2 1

y 1
p h1
( F1 ( y s2 ) P1 ( y s2 1)) ( F2 ( y n) P2 ( y n 1)) P1 (n)
pt

p h1 h2 n y s2 2

y 1
p h1
ce

( y, s2 ) P1 ( y s2 1) F2 ( s2 1) P1 ( y s2 1) P2 ( y n 1) P1 (n)
p h1 h2 n y s2 2

y 1
h1 p
Ac

( y, s2 ) P1 ( y s2 1) F2 ( s2 1) P2 ( y n 1) P1 (n)
p h1 h2 n y s2 2

( y, s2 )
Hence, y 1
is

y 1 ( y, s2 ) 1 ( y 1, s2 ) 1 ( y, s2 )
y 1
h1 p
P1 ( y s2 1) F2 ( s2 1) P2 ( y n 1) P1 (n)
p h1 h2 n y s2 2
 h1 p
F2 ( s2 1)
We know that for all
s2 , we have p h1 h2 , then

s1 1

y 1 ( y, s2 ) P2 (s1 n 1) P1 (n) 0
n s1 s2 2

Therefore, for all

, 1 ( y, s2 ) is non-decreasing in y.

d
1 ( y, s2 )
Next, we derive the expression for d . We consider the regions 1

and 2 , separately. For all 1 , F1 (.) and P1 (.) do not depend on ψ. Then

t
rip
s
for all 2 0 and 1 we have;

c
y x
d d
1 ( y, s2 ) (p h1 h2 ) F2 ( y n) P1 (n) (p h1 h2 ) P2 ( y n) P1 (n) 0

us
d n y s2 1 d n y s2 1
an
For all 2 , F2 (.) and P2 (.) do not depend on ψ, and s2 0 . Then, for all

we have;
M

y
p h1 p h1
( y, 0) F1 ( y 1) ( p h1 h2 ) F2 ( y n) P1 (n) F1 ( y 1)
ed

1
p h1 h2 n y 1 p h1 h2
pt

Then,
ce

d p h1 d p h1
1 ( y, 0) F1 ( y 1) P1 ( y 1) 0
d p h1 h2 d p h1 h2
Ac

Hence, for all

x 0 and
, 1 ( y, s2 ) is non-decreasing in ψ. Then, for
y , ( y, s2 )
all 0 and 1 is non-decreasing in y and ψ. □

For all , the optimal component 1 base-stock level is

p
s1 y 0 : 1 ( y 1, s2 ) 1 ( y, s2 )
p h1 h2
From Lemma 2, we know that for all
y 0 , , 1 ( y, s2 ) is non-

decreasing in y and ψ. As it is shown in Ahmadi et al. (2017), having

1 ( y, s2 ) being non-decreasing in both y and ψ implies that s1 is non-

increasing in ψ.

Combining the facts that i) the inequalities in Theorem 1 hold for the optimal

base-stock levels and ii)

F2 ( y) and 1 ( y, s2 ) are non-decreasing in y and ψ

for all
y 0 , implies that when we increase the commitment lead time,
each component’s optimal base-stock level either stays the same or it
decreases by one unit, i.e., the jump size is 1.

t
rip
Proof of Lemma 1

c
The equality
(l1 l2 )c23 l1c13 l2c12 is the same as (l1 l2 )c23 l1c13 l2c12 0 . We

use the definitions of the thresholds to prove the latter expression. We have
us
an
Ca ( s l2 , l2 ) Ca ( s l1 , l1 ) Ca ( s 0 , 0) Ca ( s l1 , l1 ) Ca ( s 0 , 0) Ca ( s l2 , l2 )
(l1 l2 )c23 l1c13 l2c12 (l1 l2 ) c l1 c l2 c
(l1 l2 ) (l1 0) (l2 0)
Ca ( s l2 , l2 ) Ca ( s l1 , l1 ) Ca ( s 0 , 0) Ca ( s l1 , l1 ) Ca ( s 0 , 0) Ca ( s l2 , l2 )
M

(l1 l2 ) l1 l2
(l1 l2 ) (l1 0) (l2 0)
1
(Ca ( s l2 , l2 ) Ca ( s l1 , l1 ) Ca ( s 0 , 0) Ca ( s l1 , l1 ) Ca ( s 0 , 0) Ca ( s l2 , l2 )) 0
ed

(l1 l2 )c23 l1c13 l2c12 0 (l1 l2 )c23 l1c13 l2c12

As a result, and .
pt
ce

Next, we prove the second result. Suppose

c23 c12 . We define 0 such

that
c23 c12 . Then, using the first result, we obtain
Ac

l2 l2 l2 l2 l2
c13 1 c23 c12 1 (c12 ) c12 c12 1
l1 l1 l1 l1 l1

l2
1
l2 l1 0 , we have l1
Since and . Hence,

l2
c13 c12 1 c12 .
l1
Using
c23 c12 and
c13 c12 , we obtain
c13 c23 . Combining this

inequality with
c12 c13 , we get the desired result, i.e., c12 c13 c23 .

Next, we consider the case with c23 < c12. We define 0 such that
c23 c12 . Then, using the first result, i.e.,
(l1 l2 )c23 l1c13 l2c12 , we calculate

c13 as follows:

l2 l2 l2 l2 l2
c13 1 c23 c12 1 (c12 ) c12 c12 1
l1 l1 l1 l1 l1

l2

t
1
l2 l1

rip
0 , we have l1
Since and . Hence,

l2

c
c13 c12 1 c12 .
l1

Using
c23 c12 and
c13 c12 us
, we obtain c13 > c23. Combining this
an
c23 c13 c12
inequality with c12 > c13, we get the desired result, i.e., .
M

c12 c13 c23 c23 c13 c12

As a result, either or holds.

Proof of Theorem 3
ed

To prove this theorem, we consider each Case separately.

pt

1. Case 2: For all 2, the problem is equivalent to a single-location

ce

CLB23 ( )
problem. Based on Ahmadi et al. (2017), is a tight linear lower
Ac

Ca ( s , ) over CLB23 (l2 ) Ca ( s , l2 )

l2
bound for 2 (i.e., and
CLB23 (l1 ) Ca ( s , l1 )
l1
, CLB13 ( ) CLB23 ( )
). Also, we know that for all 2

c13 c23 CLBij ( )

(since in Case 2 we have and the slope of is equal to

its corresponding cij). Then, for all 2, we can write

CLB13 ( ) Ca ( s , )
. Moreover, for all 1, based on Conjecture 1 we
CLB13 ( ) Ca ( s , ) , CLB13 ( ) Ca ( s , )
have . As a result, for all .
 CLB13 ( ) Ca ( s , ) over
Since, is a tight linear lower bound for (i.e.,
0 *
CLB13 (0) Ca ( s ,0) CLB13 (l1 ) Ca ( s , l1 ) l1
{0, l1} .
and ) then
CLB12 ( )
2. Case 3: For all 1, based on Conjecture 1, is a tight linear
0
Ca ( s , ) over CLB12 (0) Ca ( s ,0)
lower bound for 1 (i.e., and
*
CLB12 (l2 ) Ca ( sl2 , l2 ) , {0, l2 } . Also, For all
). Then, for all 1 2

, the problem is equivalent to a single-location problem. Based on

CLB23 ( )
Ahmadi et al. (2017), is a tight linear lower bound for
Ca ( s , ) over CLB23 (l2 ) Ca ( sl2 , l2 ) CLB23 (l1 ) Ca ( sl1 , l1 )
2 (i.e., and ).
*

t
Hence, for all 1 , {l2 , l1} . As a result, for all

rip
* *
1 2 , {0, l2 } {l2 , l1} . Therefore, for all , {0, l2 , l1} .

c
Proof of Theorem 4

We need the following lemma to prove the theorem. us

an
Lemma 3. In a two-component ATO system with three unit commitment cost
ti t j , i, j {1, 2,3} c cij
thresholds cij and three terminal points such that
M

, if ,
* *
tj c cij ti .
then and if , then
ed

ˆ c cij
Proof. We define cost function Ca ( s , ) with unit commitment cost ,
0 . Based on the definition of cij’s, Ca ( s , ) has the same value at
pt

where
c cij
the ti and tj for . Hence
ce

t
Ca ( sti , ti ) Ca ( s j , t j ) (A5)
Ac

Also, a
Cˆ ( s , ) can be rewritten in terms of Ca ( s , ) as

Cˆ a ( s , ) Ca ( s , )
. Then

Cˆ a ( sti , ti ) Ca ( sti , ti ) ti ,
Cˆ a ( s j , t j ) Ca ( s j , t j )
t t
tj,
 Cˆ a ( sti , ti ) Cˆ a ( s j , t j )
t
From Equation (A5) and ti < tj, we can conclude that .
Cˆ ( s , ) at t has higher value, then t can not be a candidate for
Since a j j
*
tj
optimal solution and .

Cˆ a ( s j , t j ) Cˆ a ( sti , ti )
t
c cij
Similarly, when , we can show that and then ti
*
can not be a candidate for optimal solution and
ti . □

We prove Case 2 and Case 3 separately.

c12 c13 c23

1. Case 2 ( ):

t
rip
From Theorem 3, we know that the possible set for
*
is
{t1 , t2 , t3} . We

c
consider all possible values for c and update the possible set using
Lemma 3.

if
c c12 , then c c12 , c c13 , and c c23 . So, us *
{t1 , t2 } and
an
*
t3 ,
M

if
c c12 , then c c13 , and c c23 . So, * {t1 , t2 } and * t3 ,
c c c13 , then c c12 , c c13 and c c23 . So, * {t1 , t2 } and
if 12
ed

*
t3
,
*
if
c12 c c13 , then c c12 , and c c23 . So, t2 and
pt

*
{t1 , t3} ,
ce

if
c13 c c23 , then c c12 , c c13 , and c c23 . So, *
{t2 , t3} and
*
t1 ,
Ac

* *
if
c13 c c23 , then c c12 , and c c13 . So, {t2 , t3} and t1

,
c c23 c c12 , c c13 c c23 *
{t2 , t3}
if , then and . So, and
*
t1 .

t1 0, t2 l2 , t3 l3 c12 c13 c23

We know . Hence, when we have
 *
l1 , c c13
0, c c13 .

2. Case 3 (
c23 c13 c12 ):

if
c c23 , then c c23 , c c13 , and c c12 . So, *
{t1 , t2 } and
*
t3 ,

if
c c23 , then c c13 , and c c12 . So, * t1 and *
{t2 , t3} ,
c c c13 , then c c23 , c c13 and c c12 . So,
if 23
*
{t1 , t3} and
*
t2 ,

t
rip
*
if
c23 c c13 , then c c23 , and c c12 . So, {t1 , t3} and
*
t2 ,

c
if
c13 c c12 , then c c23 , c c13 , and c c12 . So, *
{t1 , t3} and
*
t2 ,
us
an
* *
if
c13 c c12 , then c c23 , and c c13 . So, t3 and {t1 , t2 }

,
M

if
c c12
, then
c c23 , c c13 and c c12 . So, *
{t2 , t3} and
*
t1 .
ed

t1 0, t2 l2 , t3 l3 c23 c13 c12

We know . Hence, when we have
pt

l1 , c c23
ce

*
l2 , c23 c c12
0, c c12 .
Ac
Fig. 1 An ATO system with two components and a single end product

t
c rip
us
an
M
ed
pt
ce
Ac
Fig. 2 ATO system and its equivalent serial system

t
c rip
us
an
M
ed
pt
ce
Ac
Fig. 3 Inventory-on-order for component 1;
IO1 and IO1

t
c rip
us
an
M
ed
pt
ce
Ac
Fig. 4 An illustration of the optimal base-stock levels in terms of commitment
lead time

t
c rip
us
an
M
ed
pt
ce
Ac
 t
rip
Ca ( s, ) Ca ( s , ) with respect to ψ
Fig. 5 The behavior of and

c
us
an
M
ed
pt
ce
Ac
 t
crip
us
an
Fig. 6 Unit commitment cost thresholds
M
ed
pt
ce
Ac
Fig. 7 An illustration of Case 2 and Case 3

t
rip
c
us
an
M
ed
pt
ce
Ac
Fig. 8 Optimal strategies for the components in terms of the unit commitment
cost

t
c rip
us
an
M
ed
pt
ce
Ac
Table 1 Literature on perfect ADI

Authors System Characteristics Results

A multi-product, single- Both the seller and the

location periodic-review customers benefit from
system with two types of sharing ADI, but sharing ADI
Thonemann (2002) ADI increases the bullwhip effect

Identify the conditions under

Karaesmen A capacitated single- which ADI may bring
et al. (2004) location system significant benefits

t
rip
Two retailers with an “ Establish conditions under
Advance Booking which the unique equilibrium
Discount (ABD) ” calls for launching the ABD

c
McCardle et al. (2004) program

us
program at both retailers

A single-location system Evaluate the benefits of the

an
with two customer ABD program and
demand classes and characterize the optimal
M

Tang et al. (2004) ABD program discount price

Introduce a threshold value

ed

A capacitated single- that allows order aggregation

Wijngaard location continuous- to find the optimal production
pt

& Karaesmen (2007) review system policy

Provide a sufficient condition

ce

A single-location under which the tradeoff

continuous-review between inventory and ADI
Ac

Liberopoulos (2008) system is linear

Determine an optimal state-

A single-location dependent policy and a
periodic-review system tractable approximation for
with homogeneous and homogeneous and
heterogeneous exogenous customers,
Wang & Toktay (2008) customers respectively
Authors System Characteristics Results

A rental company with

Papier two customer demand The optimal admission policy
& Thonemann (2010) classes is a threshold policy

A single-location system Accurate demand

with two customer information may improve
Li & Zhang (2013) demand classes product availability

The benefits of demand

A single-location forecast information may
periodic-review system significantly decrease as

t
rip
with uncertain lead lead time uncertainty
Iida (2015) times increases

c
Determine the optimal

us
solution and an efficient
heuristic policies under
an
A capacitated single- relaxed conditions and
location system with general conditions,
M

Papier (2016) different markets respectively

ed
pt
ce
Ac
Table 2 Literature on imperfect ADI

Authors System Characteristics Results

A single item, single-location
Gallego & Özer (2001) periodic-review system
State-dependent
policies are optimal
for systems with and
without fixed ordering
costs

State-dependent,
echelon base-stock

t
policies for finite and

rip
A multi-stage periodic-review infinite horizons are
Gallego & Özer (2003) inventory system optimal

c
us Demand
postponement
an
A two-stage capacity planning enables cost and
problem with two customer capacity requirement
Iyer et al. (2003) demand classes reductions
M

Develop a lower
A centralized system with one bound and propose a
ed

warehouse and multiple close-to-optimal

Özer (2003) retailers heuristic
pt

Information cost and

ce

demand correlation
are important factors
Ac

A single-retailer, multiple for determining the

Zhu demand classes with future optimal extent of FDI
& Thonemann (2004) demand information (FDI) sharing

Characterize the
behavior of optimal
A capacitated single-location policies with respect
Özer & Wei (2004) periodic-review system to capacity and ADI
Authors System Characteristics Results

The optimal policy is

of order-up-to type,
A single-location periodic- where the order level
Tan et al. (2007) review system is a function of ADI

A single
capacitated/uncapacitated Investigate the impact
Liberopoulos supplier with two demand of ADI on the optimal
& Koukoumialos (2008) classes decisions

The optimal

t
rip
production and
allocation policies are

c
state-dependent

us
A capacitated supplier with
base-stock and
multilevel rationing
an
Gayon et al. (2009) multiple demand classes policies, respectively

The firm becomes

M

less sensitive to lead

A single-location periodic- time variability as the
ed

review system with forecast forecast updating

updating and lead time process becomes
pt

Wang & Tomlin (2009) uncertainty more efficient

The impact of sharing

ce

forecasts on profit
could be negative in
Ac

the competitive
A two-echelon serial system setting, but it is
in competitive and always positive in the
Iida & Zipkin (2010) cooperative settings cooperative setting

A capacitated supplier with A state-dependent

Benjaafar et al. (2011) stochastic production times base-stock policy is
Authors System Characteristics Results

optimal

t
c rip
us
an
M
ed
pt
ce
Ac
Table 3 Literature on both perfect and imperfect ADI

Characteristics of the
Authors system Findings and Results

A three-echelon
capacitated serial system
integrating ADI with
Claudio Kanban-based pull
& Krishnamurthy (2009) strategy
Integrating ADI and
Kanban-based pull
strategy corrects
inefficiencies

A multi-product, single- Explore the impact of

t
location system with perfect and imperfect

rip
Bernstein capacitated resource ADI on optimal
& DeCroix (2014) selection capacities and profit

c
A capacitated single-
us The imperfectness of
demand information
an
Benbitour location periodic-review reduces the benefits of
& Sahin (2015) system ADI
M
ed
pt
ce
Ac
Table 4 Optimal results

(
C C C
* *
*
s1 , s2 ,
( , h1 , h2 , p, c) (l1, l2) (c12 , c23 , c13 ) Case s 0 ,0 )
*
s l2 , l2 s l1 , l1 C (sψ , *
)
(2.7, 13.2,
22.7, 48.6, (5.0, (8.45,
13.5) 3.8) 10.89, 9.03) 2 119.7 170.6 178.9 (14, 12, 0) 119.7

(6.6, 25.4, 6.0, (10.3, (2.60, 4.20,

41.8, 20.6) 4.6) 3.49) 2 235.7 780.1 1394.8 (69, 38, 0) 235.7

t
rip
(10.1, 12.0, (17.1, (0.76, 1.46, (176, 109,
2.1, 23.0, 1.7) 9.3) 1.08) 2 185.4 271.1 287.6 0) 185.4

c
(9.23,
(0.9, 21.8,
12.3, 44.9, 6.4) 4.4)
(10.8, 10.55,
10.01) 2 92.9 82.4 us 59.8 (0, 0, 10.8) 59.8
an
(3.6, 8.9, 28.1, (5.4, (6.94,
44.4, 6.6) 5.1) 13.76, 7.32) 2 142.3 136.4 128.7 (0, 0, 5.4) 128.7
M

(10.0, 26.3, (16.9, (2.50, 6.24,

22.0, 47.5, 0.5) 14.9) 2.94) 2 496.3 201.4 87.0 (0, 0, 16.9) 87.0
ed

(1.2, 11.0, 7.6, (18.6, (4.11, 3.77,

32.2, 11.4) 6.9) 3.90) 3 87.9 148.7 256.7 (25, 11, 0) 87.9
pt

(5.8, 3.8, 5.5, (19.2, (1.80, 0.56, (119, 17,

16.9, 4.5) 2.5) 0.72) 3 80.5 119.2 497.3 0) 80.5
ce

(9.2, 3.3, 2.2, (16.6, (0.57, 0.55, (163, 77,

Ac

18.0, 1.6) 7.2) 0.56) 3 85.4 156.6 251.0 0) 85.4

(1.6, 26.4, (11.5, (13.75,

28.4, 39.5, 8.6) 1.6) 6.46, 7.48) 3 137.0 123.8 157.1 (17, 0, 1.6) 123.8

(3.1, 8.6, 10.2, (5.0, (8.13, 2.99,

24.2, 4.7) 0.6) 3.61) 3 55.3 49.0 72.1 (16, 0, 0.6) 49.0

(8.6, 4.3, 11.2, (3.5, (10.91,

32.3, 8.4) 0.2) 1.41, 1.95) 3 58.8 54.5 254.1 (35, 0, 0.2) 54.5
 (
* *
C C C s1 , s2 , *

( , h1 , h2 , p, c) (l1, l2) (c12 , c23 , c13 ) Case s 0 ,0 )

*
s l2 , l2 s l1 , l1 C (sψ , *
)
(0.8, 19.7, (15.2, (17.96,
17.7, 32.0, 2.5) 0.8) 5.74, 6.39) 3 80.4 70.2 31.6 (0, 0, 15.2) 31.6

(4.6, 16.0, (1.5, (11.58,

17.8, 36.9, 5.5) 0.6) 9.23, 10.17) 3 70.4 53.6 38.2 (0, 0, 1.5) 38.2

(9.0, 28.4, (4.2, (6.79, 5.36,

20.3, 45.1, 1.4) 1.1) 5.74) 3 216.0 162.8 52.7 (0, 0, 4.2) 52.7

t
(3.6, 13.1, 2.2, (14.9, (2.42, NaN,

rip
35.4, 16.1) 14.9) 2.42) 2* 131.7 877.2 877.2 (58, 67, 0) 131.7

(8.1, 10.9,

c
17.7, 38.1, (1.0, (9.31, NaN,
16.0)

(3.4, 2.6, 3.9,

1.0)

(2.6,
9.31)

(2.79, NaN,
2*
us
75.4 129.4 129.4 (8, 10, 0) 75.4
an
18.3, 2.5) 2.6) 2.79) 2* 24.8 22.6 22.6 (0, 0, 2.6) 22.6

(1.8, 8.0, 11.7, (0.9, (16.55,

M

37.0, 2.1) 0.9) NaN, 16.55) 2* 26.4 3.3 3.3 (0, 0, 0.9) 3.3
ed
pt
ce
Ac