You are on page 1of 12

Performance Evaluation 125 (2018) 68–79

Contents lists available at ScienceDirect

Performance Evaluation
journal homepage: www.elsevier.com/locate/peva

Continuous-type (s, Q )-inventory model with an attached


M/M/1 queue and lost sales
Jung Woo Baek a , Yun Han Bae b, *, Ho Woo Lee c , Soohan Ahn d
a
Department of Industrial Engineering, Chosun University, Gwangju, Republic of Korea
b
Department of Mathematics Education, Sangmyung University, Seoul, Republic of Korea
c
Department of Systems Management Engineering, Sungkyunkwan University, Suwon, Republic of Korea
d
Department of Statistics, University of Seoul, Seoul, Republic of Korea

article info a b s t r a c t
Article history: In this study, we consider an M/M/1 queuing model with an attached continuous-type
Received 14 October 2017 inventory. Customers arrive in the system according to a Poisson process and are served
Received in revised form 17 May 2018 individually on a first-come, first-served basis. The service times of customers are assumed
Accepted 5 July 2018
to be independent and identically distributed exponential random variables. Along with
Available online 17 July 2018
the queue, there is an internal finite storage for the inventory and each service requires
an exponentially distributed random amount H of inventory from the storage. Therefore,
Keywords:
Queuing-inventory model a customer leaves the system with H amount of item at his/her service completion time.
(s, Q ) inventory control The inventory is replenished by an outside supplier with a random lead time under an
Lost sales (s, Q ) inventory control policy. We assume that the customers who arrive during stock-out
Continuous-type inventory periods are lost (lost sales). For this queuing-inventory system, we derive the stationary
Product form stationary distribution joint probability distribution of queue length and inventory level in explicit product form.
Numerical examples followed by a cost model are also presented.
© 2018 Elsevier B.V. All rights reserved.

1. Introduction

Over the past few decades, considerable research has been conducted on queuing-inventory models. This type of queuing
model originated from well-known assembly-like queues in which several types of parts were served at the same time to
produce a complete product [1–4]. In practice, this queuing model is found in manufacturing assembly systems, and thus
it would be beneficial for a production/manufacturing company to develop an analytic method. However, even for simple
cases, limited results have been presented because of the complexity caused by the intrinsic dependency in the system.
Early studies focused on cost optimization without computing the stationary probability distribution of queue length or
inventory level. Berman et al. [5] analyzed a queuing-inventory model in which the demands and service processes proceed
in a deterministic manner. Berman and Kim [6] analyzed an M/M/1 queuing-inventory model with a zero lead time and
Berman and Sapna [7] studied an extended model with generally distributed service times. Application models in production
industries were further investigated by [8,9].
Only recently has the joint probability of queue length and inventory level been investigated. Schwarz and Daduna [10]
and Schwarz et al. [11] conducted extensive studies of M/M/1 queuing-inventory systems with exponentially distributed
lead times. They derived a closed-form stationary joint probability of queue length and inventory level in product form
under various inventory control policies. More details on the queuing-inventory models can be found in [12–16]. We refer
interested readers to Marand et al. [17] for the comprehensive review on the queuing-inventory models.

* Corresponding author.
E-mail addresses: jwbaek@chosun.ac.kr (J.W. Baek), yhbae@smu.ac.kr (Y.H. Bae), hwlee@skku.edu (H.W. Lee), sahn@uos.ac.kr (S. Ahn).

https://doi.org/10.1016/j.peva.2018.07.003
0166-5316/© 2018 Elsevier B.V. All rights reserved.
J.W. Baek et al. / Performance Evaluation 125 (2018) 68–79 69

Fig. 1. Schematic diagram of the proposed queuing-inventory model.

Most studies of queuing-inventory models only consider a discrete-type inventory, which is unsuitable for modeling a
chemical production system or a 3D-printing manufacturing system in which a fluid-type raw material, such as chemicals
or a metal powder, is used in production. In this context, we present a new queuing-inventory system in which the type of
inventory is continuous. Customers arrive in the system according to a Poisson process and are served by a server with
an exponentially distributed service time. At the service completion time, customers have consumed an exponentially
distributed amount of inventory, and the inventory is replenished by an outside supplier under the well-known (s, Q )
policy. To the best of the authors’ knowledge, this study constitutes the first research on a queuing system integrated with a
continuous-type (s, Q ) inventory model with lost sales and general lead times. Moreover, we derive the explicit stationary
joint probability distribution of queue length and inventory level in product form similar to the solutions in Saffari et al. [16].
The remainder of this paper is organized as follows. In Section 2, we describe our model in more detail and define a new
modified model by assuming the service times in the original model to be zero. The modified model plays an important
role because we analyze the proposed model with the use of the modified model. In Section 3, we derive the product form
stationary joint probability distribution of queue length and inventory level similar to the result in [16]. The long-term
performance measures are also presented. In Section 4, we offer numerical examples and describe the cost model. Finally,
we conclude our study in Section 5 with generalization issues for future work.

2. Preliminaries

In this section, we describe the proposed model in more detail and introduce a modified model that can be used to analyze
the original model.

2.1. The model

The motivation of our model comes from a chemical or a 3D-printing make-to-order manufacturing system that combines
the production and inventory control processes. Customers randomly arrive in the system and request that an item
be manufactured (an order). In this type of manufacturing system, the manager monitors all inventory (raw material)
movements and decides the timing and size of an order according to a predetermined rule.
To model this type of manufacturing system, we consider an M/M/1 queuing model with an attached continuous-type
inventory. Fig. 1 shows the schematic diagram of the proposed model and Fig. 2 shows a sample path of the M/M/1 queuing
model with a continuous-type inventory under the (s, Q ) policy. N(t) and I(t) represent queue length and inventory level at
time t, respectively.
In our model, customers arrive in the system according to a Poisson process at a rate of λ. A customer can join the queue
only when the inventory level is positive. If an arriving customer finds an empty inventory, the customer immediately leaves
the system without being served (lost sales). The server serves the customers on a first-come, first-served basis. The service
times are independent and identically distributed with exponential distribution with a mean of 1/µ. We assume that the
system is stable, i.e., the supplemented queuing-inventory state process is ergodic. At the service completion time, each
customer requires an exponentially distributed random amount H of inventory (with a mean of 1/β ); we call this the stock
requirement of the customer. Therefore, the inventory level decreases by the stock requirement (H) at the service completion
time if the on-hand inventory is sufficient. For later use, we define h(x) = β e−β x , (x ≥ 0) and H(x) = 1 − e−β x , (x ≥ 0) as the
probability density function and the distribution function of H, respectively. We assume that if a customer finds insufficient
inventory at his/her service completion time, he/she leaves the system with all the remaining stock and an extra cost is
incurred (see and in Fig. 2).1 Once the inventory is empty, the customers in the queue wait until the inventory has been
replenished before being served.

1 We make this assumption for analytic convenience. However, while changing this assumption (e.g., the customer waits for the inventory to be
replenished) allows us to generalize the findings of the presented analysis, such a generalization changes the entire analysis of the system.
70 J.W. Baek et al. / Performance Evaluation 125 (2018) 68–79

Fig. 2. Sample path of the joint process Q = {(N(t), I(t)) : t ≥ 0}. (For interpretation of the references to color in this figure legend, the reader is referred
to the web version of this article.)

We consider an (s, Q ) inventory control policy, which is a popular continuous review control policy. Specifically, under
this control policy, an order of size Q is placed with an outside supplier as soon as the inventory level falls to or below s.
After a random lead time, the supplier replenishes the inventory with size Q . By convention, we assume that 0 ≤ s < Q . We
assume that the lead time of an order follows a general continuous probability distribution with the distribution function
V (x) and probability density function v (x). We assume that the lead time is independent of queue size and inventory level.
For analytical convenience, we denote by WD the inventory level at which an order is placed (see the red circles in Fig. 2).
Let WR be the inventory level at which the storage is replenished by the supplier with size Q (see the blue circles in Fig. 2).
We call the first passage time to level WD from the end of the previous lead time a regular period. Next, we call the ensuing
lead time a delivery period. We denote the lengths of a regular period and a delivery period by LR and LD , respectively.
We note that N(t) and I(t) are strictly dependent. However, we deal with the inventory level process as if it were a
regenerative process wherein a delivery period starting point plays the role of the regeneration point. This concept of
regenerative process is beneficial for the clear understanding of the stochastic behavior of the inventory level process. We
call the length between the starting point of a delivery period and the end of a regular period a cycle (denoted by C = LD + LR ).

2.2. The modified model

This study aims to derive the stationary joint probability distribution of the queue length and inventory level of the
process Q = {(N(t), I(t)) : t ≥ 0} in closed form. For this purpose, we first define a modified model that can be effectively
and efficiently used to analyze the original process.

Definition 2.1 (Modified Model). The modified model is the modification of the original process Q in which all the service
times are assumed to be zero, i.e. the downward jumps of size Hi occur at arrival times (see Fig. 3).

Remark 2.1. Krishnamoorthy and Viswanath [15] was the first to consider the modified model in which the service times are
assumed to be zero for obtaining the joint probability of queue length and the discrete-type inventory level. They derived the
joint probability by comparing the steady-state equations for the original and the modified models. This approach well-fits
to the model of our interest even though the type of inventory is continuous.
Fig. 3 shows the modified model in the same timescale of the original system. The upper part is the queue length process
of the original model, and the associated arrival times are presented on A(t). We assume that the sizes of Hi ’s and the length
of the delivery periods are the same to those of the original model in Fig. 2 except for the last delivery period. We denote
by Qmod = {I(t) : t ≥ 0} the modified process. No customer queue can exist in the modified model because of the zero
service time assumption; therefore, the modified process Qmod can be regarded as an (s, Q ) inventory model with general
lead times, lost sales and Poisson demand with mean number of arrivals of λ per unit time.
J.W. Baek et al. / Performance Evaluation 125 (2018) 68–79 71

Fig. 3. Sample path of the modified process Qmod = {I(t) : t ≥ 0}. (For interpretation of the references to color in this figure legend, the reader is referred
to the web version of this article.)

In the next section, we compare the steady-state system equations of the original and modified models and prove that
the stationary joint probability of the original model can be decomposed into two parts. Therefore, we first present the
steady-state equations of the modified model.
We first note that Qmod = {I(t) : t ≥ 0} is not a Markov process because the inventory level process during a delivery
period is dependent on the elapsed lead time (or remaining lead time). To construct a Markovian structure for the modified
model, we define V+ (t) as the remaining lead time at time t in a delivery period and an indicator function ξ (t) as

1, if the system is in the delivery period at time t ,


{
ξ (t) =
0 if the system is in the regular period at time t .

Then, we can model Qmod as a continuous time Markov process Mmod = {(I(t), V+ (t), ξ (t)) : t ≥ 0}.
We define the following steady-state probabilities:

X̃ D (x, w )dw = lim Pr [I(t) ≤ x, w < V+ (t) ≤ w + dw, ξ (t) = 1], (w ≥ 0, 0 ≤ x < s),
t →∞

X R (x) = lim Pr [I(t) ≤ x, ξ (t) = 0], (s ≤ x < s + Q ),


t →∞

B̃R = lim Pr [I(t) = Q , ξ (t) = 0].


t →∞

In the above equation, we define BR because X R (x) has a point mass at x = Q (see in Fig. 3). By defining χ R (x) = d R
dx
X (x)
and χ̃ D (x, w ) = dx
d D
X̃ (x, w ), the steady-state equations for the modified model are given as
Q +s
( ∫ )
d
− X̃ (0, w ) =λ B [1 − H(Q )] +
D R
χ (y) [1 − H(y)] dy v (w)
R
dw s
∫ s
+λ χ̃ D (y, w) [1 − H(y)] dy, (w ≥ 0), (1)
0
72 J.W. Baek et al. / Performance Evaluation 125 (2018) 68–79

( )
∂ D
∫ Q +s
− χ̃ (x, w) = −λχ̃ (x, w) + λ B h(Q − x) +
D R
χ (y)h(y − x)dy v (w)
R
∂w s
∫ s
+λ χ̃ D (y, w)h(y − x)dy, (w ≥ 0, 0 < x < s), (2)
x

Q
( ∫ )
0 = −λχ (x) + λ B h(Q − x) +
R R
χ (y)h(y − x)dy , (s ≤ x ≤ Q ),
R
(3)
x

∫ Q +s
0 = −λχ R (x) + χ̃ D (x − Q , 0) + λ χ R (y)h(y − x)dy, (Q < x < Q + s), (4)
x

0 = −λBR + X̃ D (0, 0). (5)

3. Main results

3.1. Product form stationary joint distribution

We now prove that the joint probability distribution of the inventory level and the queue length in the original process can
be written in a product form in steady-state. To be more specific, we show that the stationary joint probability distribution
can be decomposed into two parts: one part represents the queue length probability, and the other part represents the
inventory level distribution.
The original process Q = {(N(t), I(t)) : t ≥ 0} is not Markovian because the queue length and the inventory level
processes during a delivery period are dependent on the elapsed lead time (or remaining lead time). Similar to the modified
model, we construct a Markovian structure by adding V+ (t) and ξ (t) to the original process. We then model Q as a continuous
time Markov process M = {(N(t), I(t), V+ (t), ξ (t)) : t ≥ 0} and define the following steady-state probabilities:
Ψ̃nD (x, w )dw = lim Pr [N(t) = n, I(t) ≤ x, w < V+ (t) ≤ w + dw, ξ (t) = 1], (n = 0, 1, 2, . . . , w ≥ 0, 0 ≤ x < s),
t →∞

ΨnR (x) = lim Pr [N(t) = n, I(t) ≤ x, ξ (t) = 0], (n = 0, 1, 2, . . . , s ≤ x < Q + s),


t →∞

ARn = lim Pr [N(t) = n, I(t) = Q , ξ (t) = 0], (n = 0, 1, 2, . . .),


t →∞

where we define ARn because ΨnR (x) has a point mass at x = Q (see in Fig. 2).
We define ψnR (x) = dx
d
ΨnR (x) and ψ̃nD (x, w ) = dx
d
Ψ̃nD (x, w ) to obtain the following steady-state equations:
( ∫ Q +s )
d
− Ψ̃nD (0, w ) =µ ARn+1 [1 − H(Q )] + ψnR+1 (y) [1 − H(y)] dy v (w)
dw
∫ ss
+µ ψ̃nD+1 (y, w) [1 − H(y)] dy, (n = 0, 1, 2, . . . , w ≥ 0), (6)
0

( )
∂ D
∫ Q +s
− ψ̃ (x, w) = −λψ̃0 (x, w) + µ A1 h(Q − x) +
D R
ψ1 (y)h(y − x)dy v (w)
R
∂w 0 s
∫ s
+µ ψ̃1D (y, w)h(y − x)dy, (w ≥ 0, 0 < x < s), (7)
x

∂ D
− ψ̃ (x, w) = λψ̃nD−1 (x, w) − (λ + µ)ψ̃nD (x, w)
∂w n ( )
∫ Q +s ∫ s
+ µ An+1 h(Q − x) +
R
ψn+1 (y)h(y − x)dy v (w) + µ
R
ψ̃nD+1 (y, w)h(y − x)dy,
s x

(n = 1, 2, . . . , w ≥ 0, 0 < x < s), (8)

Q
( ∫ )
0 = −λψ0R (x) + µ AR1 h(Q − x) + ψ1R (y)h(y − x)dy , (s ≤ x ≤ Q ), (9)
x
J.W. Baek et al. / Performance Evaluation 125 (2018) 68–79 73

Q
( ∫ )
0 = λψnR−1 (x) − (λ + µ)ψnR (x) + µ ARn+1 h(Q − x) + ψnR+1 (y)h(y − x)dy ,
x
(n = 1, 2, . . . , s ≤ x ≤ Q ), (10)
∫ Q +s
0 = −λψ0R (x) + ψ̃0D (x − Q , 0) + µ ψ1R (y)h(y − x)dy, (Q < x < Q + s), (11)
x

∫ Q +s
0 = λψnR−1 (x) − (λ + µ)ψnR (x) + ψ̃nD (x − Q , 0) + µ ψnR+1 (y)h(y − x)dy,
x
(n = 1, 2, . . . , Q < x < Q + s), (12)

0 = −λAR0 + Ψ̃0D (0, 0), (13)

0 = λARn−1 − (λ + µ)ARn + Ψ̃nD (0, 0), (n = 1, 2, . . .). (14)


We then have the following theorem.
( )( )n
λ λ
Theorem 3.1. Let PM /M /1 (n) = 1 − µ µ
, (n = 0, 1, 2, . . .) be the stationary queue length probability of the conventional
M/M/1 queue. Then, the stationary joint probability distribution of the original model is given as

Ψ̃nD (x, w ) = PM /M /1 (n) · X̃ D (x, w ), (n = 0, 1, 2, . . . , w ≥ 0, 0 ≤ x < s), (15)

ΨnR (x) = PM /M /1 (n) · X R (x), (n = 0, 1, 2, . . . , s ≤ x ≤ s + Q ), (16)

ARn = PM /M /1 (n) · BR , (n = 0, 1, 2, . . .). (17)

Proof. The proof is shown in Appendix. □


Theorem 3.1 implies that the joint probability distribution for Q can be decomposed into two parts: one part represents
the queue length probability, and the other part represents the inventory level distribution. It also implies that the first part
is the same as the stationary queue length distribution of the normal M/M/1 queue and the second part is the same as the
stationary inventory level distribution of the modified system. Therefore, we can determine the joint probability distribution
of Q (i.e., the original model) by obtaining the stationary probability distribution of the inventory level for Qmod (i.e., the
modified model).

3.2. Analysis of the modified system

In this section, we derive the stationary probability distribution of the inventory level of the modified model to complete
Eqs. (15)–(17). We first have the following theorem.

Theorem 3.2. The modified model Qmod is a regenerative process wherein a delivery period starting point plays the role of the
regeneration point.

Proof. For convenience, we assume that the modified process starts with Q amount of inventory in a regular period at time
0. For k = 1, 2, . . . , we define Hi(k) as the stock requirement of ith arrival in kth regular period and Xk as the overshoot from
level s at the kth delivery period starting point. We define wk (≥ Q ) as the inventory level at the start of kth regular period,
(k) (k)
and let τi = H1(k) + · · · + Hi(k) , (i = 1, 2, . . . , τ0 = 0).
(k) (k)
We now consider a counting process {Ψ (k) (t), t ≥ 0} in which Ψ (k) (t) is defined as Ψ (k) (t) = i when τi ≤ t < τi+1 .
Then, the process {Ψ (k) (t), t ≥ 0} is none other than the Poisson process because Hi(k) ’s are sampled from i.i.d. exponential
(k) (k)
distribution. Let us define Mk as the discrete random variable that satisfies τM −1 ≤ wk − s ≤ τM . Then, for all k, it is not
k k
(k)
difficult to have that Xk = τM − (wk − s), and Xk is identical to H due to the memoryless property. Consequently, we note
k
that the inventory level at the kth delivery starting point is given as max(s − Xk , 0) = max(s − H , 0) for all k to complete
the proof. □

Remark 3.1. Theorem 3.2 holds due to the memoryless property of the exponential distribution because once the inventory
level falls to or below level s, the vertical length from level s to the level at the downward jump ends (i.e., the vertical lengths
from level s to the points and to the point in Fig. 3) is stochastically identical to H by the memoryless property.
74 J.W. Baek et al. / Performance Evaluation 125 (2018) 68–79

Since we prove that Qmod is a regenerative process, we can apply the conventional approach base on renewal theory. The
expected length E(LD ) of the delivery period is the same as the mean E(V ) of the lead time. The expected length E(LR ) of an
arbitrary regular period is the mean of the first passage of time to the level less than s from the inventory level WR (≥ Q )
at the start of the regular period. The inventory level decreases by the exponentially distributed amount (at a rate of β ) for
each arrival time during the regular period. Let Λi be the ith inter-arrival time from the start of the arbitrary regular period
and Ñ(t) be the number of Poisson arrivals at a rate of β during t. Then, we use Wald’s equation and obtain
[ ] [ ] 1[ [ ]
E(LR ) = E Λ1 + Λ2 + · · · + ΛÑ(WR −s)+1 = E (Λ1 ) E Ñ(WR − s) + 1 = β E(WR ) − s + 1 ,
]
(18)
λ
where E(WR ) can be determined using Eqs. (24)–(25) by direct computation.
The mean length E(C ) of a cycle is given as
1[ [ ]
β E(WR ) − s + 1 .
]
E(C ) = E(LD ) + E(LR ) = E(V ) + (19)
λ

3.2.1. Analysis of the delivery period


In this section, we derive the stationary probability distribution of the inventory level in a delivery period for the modified
system. We first define H (k) (y) and h(k) (y) as the k-fold convolution of H(x) and h(x), respectively.
As before, WD denotes the inventory level at the start of an arbitrary delivery period (see the red circles in Figs. 2 and 3). We
define UD (x) = Pr(WD ≤ x), (0 ≤ x < s). According to Theorem 3.2 and Remark 3.1, we have UD (x) = Pr (max(0, s − H) ≤ x)
and obtain

UD (0) = e−β s , (20)


and
d
uD (x) = UD (x) = β e−β (s−x) , (0 < x < s). (21)
dx
We now have the following theorem.
∫∞
Theorem 3.3. Let X D (x) = 0
X̃ D (x, w )dw, (0 ≤ x < s) and χ D (x) = d D
dx
X (x). We then have
[ ∞
]
∫ s
E(V ) ∑
χ (x) =
D
uD (x)Φ0 + uD (y) Φk h (y − x)dy
(k)
E(C ) x k=1
[∞ )]
β k+1 (s − x)k e−β (s−x)
(
E(V ) ∑
= Φk , (0 < x < s), (22)
E(C ) k!
k=0

[ ∞
]
∫ s
E(V ) ∑
D
Φk 1 − H (y) dy ,
(k)
( )
X (0) = UD (0) + uD (y)
E(C ) 0 k=1
∞ k
∑ ∑ (β s)n e−β s
= Φk , (23)
n!
k=0 n=0

(λt)n e−λt
∫∞
where Φn = t =0 n!
dV− (t) and V− (t) is defined as
∫t
[1 − V (x)] dx
V− (t) = 0
, (t ≥ 0).
E(V )

Proof. E(V )/E(C ) is the probability that the modified system is in a delivery period at an arbitrary time in steady-state. By
using this probability, we can obtain the stationary inventory level distribution during a delivery period by conditioning on
the level at the start of the delivery period.
For Eq. (22), we suppose that the inventory level at the start of a delivery period is x and that no arrival occurs during
the elapsed lead time. Then, the first term in the brackets is immediately obtained. Next, we consider the case in which a
delivery period starts with y (x < y < s) amount of inventory and k customers arrive and require (y − x) amount of inventory
during the elapsed lead time to make the inventory level x. By summing all possible k and y, we obtain the second term in
the brackets.
A similar approach can be used for Eq. (23). Suppose a delivery period starts with zero inventory. Since no arrival is
accepted when the inventory storage is empty, we obtain the first term in the brackets. For the second term, we consider
the case in which a delivery period starts with inventory amount y (0 < y < s) and k arrivals make the inventory storage
empty during the elapsed lead time. By summing all possible k and y, we obtain the second term in the brackets.
The second equalities in Eqs. (22) and (23) are obtained by direct algebraic manipulation. □
J.W. Baek et al. / Performance Evaluation 125 (2018) 68–79 75

3.2.2. Analysis of the regular period


Next, we derive the stationary probability distribution of the inventory level during a regular period in the modified
model. The inventory level during a regular period depends on the inventory level at the start of the period. Therefore, we
first derive the probability distribution of the inventory level at the start of an arbitrary regular period.

Theorem 3.4. Let us define UR (x) = Pr(WR ≤ x), (Q ≤ x ≤ Q + s) and uR (x) = d


U (x).
dx R
We then have
∫ s ∞

uR (x) = uD (x − Q )V0 + uD (y) Vk h(k) (Q + y − x)dy
x−Q k=1

β k+1 [s − (x − Q )]k e−β[s−(x−Q )]
∑ ( )
= Vk , (Q < x < Q + s), (24)
k!
k=0

∫ s ∞

Vk 1 − H (k) (y) dy
( )
UR (Q ) = UD (0) + uD (y)
0 k=1
∞ k
∑ ∑ (β s)n e−β s
= Vk , (25)
n!
k=0 n=0

(λt)n e−λt
∫∞
in which, Vk = t =0 n!
dV (t), (k ≥ 0).

Proof. Similar to the proof of Theorem 3.3, we obtain the inventory level distribution at the start of a regular period by
conditioning on the level at the start of the previous delivery period. For Eq. (24), if a delivery period starts with x − Q amount
of inventory and no arrivals occur during the previous delivery period, then the first term in the brackets is immediately
obtained. Next, we consider the case in which a delivery period starts with inventory amount y (x − Q < y < s) and the
customers arriving during the delivery period spend y − (x − Q ) on inventory. Then, we obtain the second term.
Eq. (25) is obvious because the regular period starts with exactly Q amount of inventory only when the inventory level
becomes zero during the previous delivery period.
The second equalities in Eqs. (24) and (25) are obtained by direct algebraic manipulation. □
We are now ready to derive the probability distribution of the inventory level during the regular period.

Theorem 3.5. We have


1/λ
BR = UR (Q ) (26)
E(C )

1/λ Q +s
⎧ [ ∫ ]

⎪ u R (x) + β uR (y)dy , (Q < x < Q + s),
E(C )

⎨ x
χ (x) =
R
(27)
⎩ β/λ ,

(s ≤ x ≤ Q ).


E(C )

Proof. If the inventory level process during a regular period reaches level x, it remains there for an exponentially distributed
random time with a mean of 1/λ. Then, we directly obtain Eq. (26) because an arbitrary regular period starts with Q amount
of inventory at a probability of UR (Q ) and remains at that level for 1/λ on average.
We derive Eq. (27) by conditioning on the inventory level at the start of an arbitrary regular period. Let m(t) be a renewal
density function with exponentially distributed renewal intervals with a mean of 1/β ; thus, m(t) = β . The inventory level
during a regular period decreases by the exponentially distributed amount with a mean of 1/β at each arrival. Therefore,
the inventory level process during
( ) a regular period reaches level (x, x + dx], (Q < x < Q + s) at a probability of
∫ Q +s
uR (x) + x
uR (y)m(y − x)dy dx. Since the mean sojourn time at level x during a regular period is 1/λ, we obtain the
first
( case in Eq. (27). Similarly, the process during
) a regular period reaches level (x, x + dx], (s ≤ x ≤ Q ) at a probability
∫ Q +s
of UR (Q )m(Q − x) + Q uR (y)m(y − x)dy dx. By using this probability, we obtain the second case after simple algebraic
manipulation. □

3.2.3. Mean performance measures


Let us define E(IR ) and E(ID ) as the expected inventory levels in a regular and in a delivery period both in the original and
in the modified systems. Then, by direct integration, we obtain

E(I) = E(I D ) + E(I R ), (28)


76 J.W. Baek et al. / Performance Evaluation 125 (2018) 68–79

∫ Q +s
E(I R ) = QBR + xχ R (x)dx
s
1/λ Q +s
β (Q 2 − s2 ) Q +s Q +s
[ ∫ ∫ ∫ ]
= QUR (Q ) + xuR (x)dx + x+β x uR (y)dydx
E(C ) Q 2 Q x

1/λ β (Q 2 − s2 ) Q +s y
[ ∫ ∫ ]
= E(WR ) + +β uR (y) xdxdy
E(C ) 2 Q Q

1/λ β (Q 2 − s2 ) E(WR2 ) − Q 2
[ ( )]
= E(WR ) + +β
E(C ) 2 2
1/λ E(WR2 ) − s2
[ ( )]
= E(WR ) + β (29)
E(C ) 2
and
∫ s
D
E(I ) = xχ D (x)dx
0
{ ∞ )]}
Γ (k + 1, sβ ) Γ (k + 2, sβ )
)[ ( (
E(V ) ∑ 1
= Φk s 1 − − k+1− ,
E(C ) k! β k!
k=0
{ ∞ ]}
λE(V 2 ) ∑ β Γ (k + 2, sβ ) − sΓ (k + 1, sβ )
[ −1
E(V ) 1
= s− − + Φk , (30)
E(C ) β 2E(V )β k!
k=0
∫∞
where Γ (a, x) is the incomplete gamma function defined as Γ (a, x) = x t a−1 e−t dt, and we can obtain E(WR ) and E(WR2 )
using Eqs. (24)–(25) by direct computation.

4. The cost model

In the cost model, we consider the following coefficients for the cost function.

(1) Holding cost (cH ): the cost incurred to keep a unit amount of an item in the system per unit time,
(2) Ordering cost (cO ): the cost incurred to place an order (a fixed cost spent once in a cycle),
(3) Lost sales cost (cL ): the cost incurred by the loss of customers during stock-out periods,
(4) Inventory deficiency cost (cD ): the cost incurred by inventory deficiency at the service completion time of a customer
(at most once in a cycle), and
(5) Waiting cost (cW ): the cost per unit time incurred to hold a customer in the system.

Similar to Schwarz et al. [11], we can construct the average operating cost function C (s, Q ) per unit time as follows:
1 UR (Q )
C (s, Q ) = cH · E(I) + cO · + λ · cL · X D (0) + cD · + LM /M /1 · cW (31)
E(C ) E(C )
RU (Q )
where E(C )
is the rate (mean number of occurrences per unit time) at which a customer faces insufficient inventory in both
the original and the modified systems (see and in Figs. 2 and 3). This is because the rate at which a regular period starts
with exactly Q is the same as the rate at which a customer finds insufficient inventory in steady-state.
By using the above cost function, Fig. 4 shows the optimal values (s∗ , Q ∗ ) and optimal costs under various CoV settings
for the lead time. We considered three probability distributions, namely exponential, Erlang, and hyper-exponential, to vary
the CoVs. We set the parameters to be the same as the expectations of these three probability distributions and the CoV of
the exponential distribution (× in Fig. 4) to be double that of the Erlang distribution (+ in Fig. 4). The CoVs of the hyper-
exponential distribution (see the triangles in Fig. 4) were set to be 2, 4, 8, and 16 times that of the exponential distribution.

5. Conclusion

In this study, we considered an M/M/1 queuing model with an attached continuous-type inventory, motivated by
chemical and 3D-printing manufacturing systems. For inventory control, we considered an (s, Q ) policy with general lead
times. We then derived the stationary joint distribution of queue length and inventory level in product form. By using the
mean performance measures, we presented the cost model and numerical examples. From a supply chain management
point of view, our queuing-inventory model can be a useful tool for optimal inventory control for make-to-order production
systems integrated with a continuous-type inventory.
J.W. Baek et al. / Performance Evaluation 125 (2018) 68–79 77

Fig. 4. The optimal costs under various coefficient of variation (CoV) settings for the lead time (cH = 1, cO = 200, CL = 300, CD = 50, CW = 1, β =
0.5, λ = 1, µ = 10/9, E(V ) = 7).

Future work could include various generalizations to make the system more realistic. One direct extension would be to
generalize the distributions of the inter-arrival and service times. Further, it would be beneficial to consider lost sales due
to customers’ will, such as reneging and balking, which may seriously affect the management of the system. Finally, future
work could consider the case when the service time is dependent on the stock requirement of a customer. This generalization
would be valuable to model a system in which the server continuously consumes the inventory item in practice. The model
presented herein could be extended by applying the stochastic fluid-flow model [18–21] in which the fluid inventory varies
depending on the background Markov process.

Acknowledgments

Jung Woo Baek was supported by Basic Science Research Program through the National Research Foundation of Korea
(NRF) funded by the Ministry of Education (NRF-2015R1D1A1A01057116).
Yun Han Bae was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF)
funded by the Ministry of Education (NRF-2015R1D1A1A01056962).
Soohan Ahn was supported by the Basic Science Research Program of the National Research Foundation of Korea
(NRF-2015R1D1A1A09059663).

Appendix. Proof of Theorem 3.1

λ
We prove the theorem by directly substituting the equations given by (15)–(17) into Eqs. (6)–(14). Let ρ = µ
. Then, using
Eqs. (15)–(17), we can rewrite Eq. (6) as

Q +s
( ∫ )
d
− (1 − ρ )ρ n
X̃ (0, w ) =µ(1 − ρ )ρ
D n+1 R
B [1 − H(Q )] + χ (y) [1 − H(y)] dy v (w)
R
dw s
∫ s
+ µ(1 − ρ )ρ n+1 χ̃ D (y, w) [1 − H(y)] dy, (n = 0, 1, 2, . . . , w ≥ 0). (A.1)
0

By simple algebraic manipulation, we obtain


Q +s
( ∫ )
d
− X̃ D (0, w ) =λ BR [1 − H(Q )] + χ R (y) [1 − H(y)] dy v (w)
dw
∫s s
+λ χ̃ D (y, w) [1 − H(y)] dy, (n = 0, 1, 2, . . . , w ≥ 0). (A.2)
0

Therefore, we confirm that Eq. (A.2) is the same to Eq. (1) irrespective of n.
78 J.W. Baek et al. / Performance Evaluation 125 (2018) 68–79

Next, we respectively rewrite Eqs. (7) and (8) using Eqs. (15)–(17) as
∂ D
− (1 − ρ ) χ̃ (x, w)
∂w ( )
∫ Q +s
= −λ(1 − ρ )χ̃ (x, w) + µ(1 − ρ )ρ B h(Q − x) +
D R
χ (y)h(y − x)dy v (w)
R

s
∫ s
+ µ(1 − ρ )ρ χ̃ D (y, w)h(y − x)dy, (n = 0, w ≥ 0, 0 < x < s), (A.3)
x

and
∂ D
− (1 − ρ )ρ n χ̃ (x, w) = λ(1 − ρ )ρ n−1 χ̃ D (x, w) − (λ + µ)(1 − ρ )ρ n χ̃ D (x, w)
∂w ( )
∫ Q +s
+ µ(1 − ρ )ρ n+1 R
B h(Q − x) + χ (y)h(y − x)dy v (w)
R

s
∫ s
+ µ(1 − ρ )ρ n+1 χ̃ D (y, w)h(y − x)dy, (n = 1, 2, . . . , w ≥ 0, 0 < x < s), (A.4)
x

which yield
( )
∂ D
∫ Q +s
− χ̃ (x, w) = −λχ̃ (x, w) + λ B h(Q − x) +
D R
χ (y)h(y − x)dy v (w)
R
∂w s
∫ s
+λ χ̃ D (y, w)h(y − x)dy, (n = 0, 1, 2, . . . , w ≥ 0, 0 < x < s), (A.5)
x

and we note that Eq. (A.5) is the same to Eq. (2) irrespective of n.
Similar manipulation for Eqs. (9) and (10) yields
Q
( ∫ )
0 = −λ(1 − ρ )χ R (x) + µ(1 − ρ )ρ BR h(Q − x) + χ R (y)h(y − x)dy , (n = 0, s ≤ x ≤ Q ), (A.6)
x

and

0 = λ(1 − ρ )ρ n−1 χ R (x) − (λ + µ)(1 − ρ )ρ n χ R (x)


Q
( ∫ )
+ µ(1 − ρ )ρ n+1 BR h(Q − x) + χ R (y)h(y − x)dy , (n = 1, 2, . . . , s ≤ x ≤ Q ), (A.7)
x

Therefore, we can confirm that the following equation is the same to Eq. (3) for all n.
Q
( ∫ )
0 = −λχ R (x) + λ BR h(Q − x) + χ R (y)h(y − x)dy , (s ≤ x ≤ Q ), (A.8)
x

For Eqs. (11) and (12), we can have the rewritten forms as

0 = −λ(1 − ρ )χ R (x) + (1 − ρ )χ̃ D (x − Q , 0)


∫ Q +s
+ µ(1 − ρ )ρ χ R (y)h(y − x)dy, (n = 0, Q < x < Q + s). (A.9)
x

and

0 = λ(1 − ρ )ρ n−1 χ R (x)−(λ + µ)(1 − ρ )ρ n χ R (x) + (1 − ρ )ρ n χ̃ D (x − Q , 0)


∫ Q +s
+ µ(1 − ρ )ρ n+1 χ R (y)h(y − x)dy, (n = 1, 2, . . . , Q < x < Q + s), (A.10)
x

Therefore, we have
∫ Q +s
0 = −λχ R (x) + χ̃ D (x − Q , 0) + λ χ R (y)h(y − x)dy, (n = 0, 1, 2, . . . , Q < x < Q + s), (A.11)
x

Hence, we can confirm that Eq. (A.11) is the same to Eq. (4) for all n.
Finally, we use Eqs. (15)–(17) into Eqs. (13) and (14) to obtain the following rewritten forms as

0 = −λ(1 − ρ )BR + λ(1 − ρ )X̃ D (0, 0), (n = 0), (A.12)


J.W. Baek et al. / Performance Evaluation 125 (2018) 68–79 79

and

0 = λ(1 − ρ )ρ n−1 BR − (λ + µ)(1 − ρ )ρ n BR + (1 − ρ )ρ n X̃ D (0, 0), (n = 1, 2, . . .). (A.13)


Therefore, we have

0 = −λBR + X̃ D (0, 0), (n = 0, 1, 2, . . .), (A.14)


and confirm that Eq. (A.14) is the same to Eq. (5) for all n.
Therefore, the proof is complete by noting that the steady-state equations for inventory level of the modified model is
the same to the steady-state equations for inventory level distribution of the original model derived by the above direct
substituting irrespective of n.

References

[1] Y.A. Bozer, L.F. McGinnis, Kitting versus line stocking: A conceptual framework and a descriptive model, Int. J. Prod. Econ. 28 (1992) 1–19.
[2] H. Brynzér, M.I. Johansson, Design and performance of kitting and order picking systems, Int. J. Prod. Econ. 41 (1995) 115–125.
[3] J.M. Harrison, Assembly-like queues, J. Appl. Probab. (1973) 354–367.
[4] E. Lipper, B. Sengupta, Assembly-like queues with finite capacity: Bounds, asymptotics and approximations, Queueing Syst. 1 (1986) 67–83.
[5] O. Berman, E.H. Kaplan, D.G. Shevishak, Deterministic approximations for inventory management at service facilities, IIE Trans. 25 (1993) 98–104.
[6] O. Berman, E. Kim, Stochastic models for inventory management at service facilities, Stoch. Models 15 (1999) 695–718.
[7] O. Berman, K. Sapna, Inventory management at service facilities for systems with arbitrarily distributed service times, Stoch. Models 16 (3–4) (2000)
343–360.
[8] Q.M. He, E. Jewkes, Performance measures of a make-to-order inventory-production system, IIE Trans. 32 (2000) 409–419.
[9] Q.M. He, E.M. Jewkes, J. Buzacott, Optimal and near-optimal inventory control policies for a make-to-order inventory–production system, European J.
Oper. Res. 141 (2002) 113–132.
[10] M. Schwarz, H. Daduna, Queueing systems with inventory management with random lead times and with backordering, Math. Methods Oper. Res. 64
(2006) 383–414.
[11] M. Schwarz, C. Sauer, H. Daduna, R. Kulik, R. Szekli, M/M/1 Queueing systems with inventory, Queueing Syst. 54 (2006) 55–78.
[12] J.W. Baek, S.K. Moon, The M/M/1 queue with a production-inventory system and lost sales, Appl. Math. Comput. 233 (2014) 534–544. http://dx.doi.
org/10.1016/j.amc.2014.02.033. URL. http://www.sciencedirect.com/science/article/pii/S0096300314002720.
[13] A. Krishnamoorthy, N. Anbazhagan, Perishable inventory system at service facilities with n policy, Stoch. Anal. Appl. 26 (2007) 120–135.
[14] A. Krishnamoorthy, B. Lakshmy, R. Manikandan, A survey on inventory models with positive service time, OPSEARCH 48 (2011) 153–169. http:
//dx.doi.org/10.1007/s12597-010-0032-z. URL. http://dx.doi.org/10.1007/s12597-010-0032-z.
[15] A. Krishnamoorthy, N.C. Viswanath, Stochastic decomposition in production inventory with service time, European J. Oper. Res. 228 (2013) 358–366.
http://dx.doi.org/10.1016/j.ejor.2013.01.041. URL. http://www.sciencedirect.com/science/article/pii/S037722171300088X.
[16] M. Saffari, S. Asmussen, R. Haji, The M/M/1 queue with inventory, lost sale, and general lead times, Queueing Syst. 75 (2013) 65–77.
[17] A.J. Marand, H. Li, A. Thorstenson, Joint inventory control and pricing in a service-inventory system, Int. J. Prod. Econ. (2017). http://dx.doi.org/10.
1016/j.ijpe.2017.07.008. URL. http://www.sciencedirect.com/science/article/pii/S0925527317302256.
[18] Y. Barron, An (s, k, s) fluid inventory model with exponential leadtimes and order cancellations, Stoch. Models 32 (2016) 301–332.
[19] Y. Barron, D. Hermel, Shortage decision policies for a fluid production model with map arrivals, Int. J. Prod. Res. 55 (2017) 3946–3969.
[20] O. Berman, D. Perry, W. Stadje, An (s, r, s) diffusion inventory model with exponential leadtimes and order cancellations, Stoch. Models 24 (2008)
191–211. http://dx.doi.org/10.1080/15326340802016985.
[21] V. Kulkarni, K. Yan, Production-inventory systems in stochastic environment and stochastic lead times, Queueing Syst. 70 (2012) 207–231.

You might also like