European Journal of Operational Research 134 (2001) 309329

www.elsevier.com/locate/dsw

Theory and Methodology

Age-based vs. stock level control policies for a perishable

inventory system
 u G
Eylem Tekin a, Ulk urler b,*
, Emre Berk c

a
Department of Industrial Engineering and Management Science, Northwestern University, Evanston, IL 60208-3119, USA
b
Department of Industrial Engineering, Bilkent University, 06533 Ankara, Turkey
c
Faculty of Business Administration, Bilkent University, 06533 Ankara, Turkey
Received 31 August 1999; accepted 24 July 2000

Abstract

In this study, we investigate the impact of modied lotsize-reorder control policy for perishables which bases re-
plenishment decisions on both the inventory level and the remaining lifetimes of items in stock. We derive the ex-
pressions for the key operating characteristics of a lost sales perishable inventory model, operating under the proposed
age-based policy, and examine the sensitivity of the optimal policy parameters with respect to various system pa-
rameters. We compare the performance of the suggested policy to that of the classical Q; r type policy through a
numerical study over a wide range of system parameters. Our ndings indicate that the age-based policy is superior to
the stock level policy for slow moving perishable inventory systems with high service levels. 2001 Elsevier Science
B.V. All rights reserved.

Keywords: Inventory; Perishable; Lost sales

1. Introduction and literature review

In this paper, we consider a continuous review perishable inventory system operating under a modied
lotsize-reorder control policy which also takes into account the remaining lifetime of the items in stock. Due
to the importance of perishable inventories in food, chemical and pharmaceutical industries, mathematical
modeling of such systems have attracted the attention of many researchers.
Under periodic review, if the items cannot be retained in stock for more than one period, the perishable
inventory problem reduces to the well-known newsvendor problem. When the lifetime of the items is
greater than one period, the problem becomes more dicult. Van Zyl [30] provides the earliest formulation

*
Corresponding author. Tel.: +90-312-2901520; fax: +90-312-2664054.
E-mail addresses: [email protected] (E. Tekin), [email protected] (U. G
urler), [email protected] (E. Berk).

0377-2217/01/$ - see front matter 2001 Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 7 - 2 2 1 7 ( 0 0 ) 0 0 2 5 0 - 2
310 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329

for nite horizon, where the shelife is assumed to be exactly two review periods. Following Van Zyl,
Nahmias and Pierskalla [14], Fries [3] and Nahmias [15] investigate the properties of the optimal policy for
the two-period and the general m-period shelives, respectively. Given the complexity of the optimal policy
structure for lifetimes longer than two periods, research has henceforth shifted towards development of
approximate policies. Among them, we may cite Nahmias [16], Cohen [2], Nahmias [17], Nahmias [18] and
recently Nandsakumar and Morton [21]. A comprehensive survey of works under periodic review can be
found in Nahmias [19].
The existing work under continuous review can be categorized as models with zero, exponential and
constant leadtimes. With zero leadtime, Weiss [31] considers a model with Poisson external demand process
and constant shelives for the items in a batch, and shows that the optimal policy is a continuous review
s; S policy. Kalpakam and Arivarignan [5] consider the s; S model with exponential shelives and a zero
reorder point. Liu [8] studies a similar model with backorders. Recently, Lian and Liu [10] have studied a
discrete time s; S backordering model with geometric interdemand times and batch demands. Liu and
Lian [11] consider a similar model with constant shelives and a renewal demand process. Liu and Shi [9]
later investigate the exponential lifetime case in the same setting. Moorthy et al. [13] consider a dierent
model, where the lifetimes are Erlangian but the aging process of an item begins only after it is put on
display. With exponential leadtimes, Kalpakam and Sapna [6] analyze an s; S system with Poisson de-
mands and exponential shelives. Excess demand is lost and only one outstanding order is allowed at any
point in time. Liu and Yang [12] consider backordering and relax the restriction on the number of out-
standing orders in a model similar to Kalpakam and Sapna [6]. Kalpakam and Sapna [7] consider a lost
sales (S 1; S) model with exponential lifetimes and demands generated according to a general renewal
process. With instantaneous replenishment and exponential leadtimes, mostly, it is possible to analyze the
perishable inventory systems by means of Markov renewal techniques. With constant leadtimes, a
Markovian structure can not be attained and the analysis becomes signicantly dicult. For constant
leadtimes, Schmidt and Nahmias [27] study (S 1; S) policies when external demands are generated from a
Poisson process, items have xed shelives and excess demand is lost. The steady-state distribution of
stocking levels is obtained by means of the method of supplementary variables. This is the rst study that
provides the exact continuous review analysis of perishables with xed life in the presence of nonzero
leadtimes. Recently Perry and Posner [22], have considered an (S 1; S) model with constant leadtimes,
leadtime-dependent backordering and analyze it by means of level-crossing methodology. Nahmias and
Wang [20] provide the rst treatment of perishable inventory with xed ordering costs and a constant
leadtime. For Poisson demand, a heuristic (Q; r) model is developed under the assumption of at most one
order outstanding. The diculty of xed lifetimes is circumvented by means of a constant decay approx-
imation for the aging process. Chiu [1] re-examines the problem and provides another heuristic solution.
Ravichandran [25] considers a constant lifetime (s; S) system with Poisson demand and random leadtimes.
He introduces a specic aging process for each order batch similar to that in [13].
There is also a large body of work on continuously deteriorating inventory systems, which falls outside
the scope of this study. For an exhaustive review, we refer the reader to [24].
The form of the optimal policy for a continuous review perishable inventory system with positive
leadtime appears to be extremely complex for analysis and implementation. Hence, as summarized above,
the existing studies consider only the (s; S) type control policies, which are known to be optimal for
nonperishables facing Markovian demand. In perishable inventory systems, the eective demand is non-
Markovian due to decaying of items even with Markovian external demand. Schultz [28] has shown for
slow moving nonperishable items that time-based ordering policies dominate the conventional inventory
level-based policies when the interdemand distribution has an increasing hazard rate. Therefore, it is
worthwhile to investigate the performance of similar age-based ordering policies for perishables. The
contribution of this study is to introduce such a control policy which bases reordering decisions not only on
the stock levels but also on the remaining shelives of the items in stock. Our results show that the proposed
 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329 311

age-based policy dominates the classical control policies and suggest that policies of the proposed structure
would be promising in other perishable inventory settings as well. We hope that our ndings would initiate
research in this direction.
The rest of the paper is organized as follows. Section 2 states the assumptions of the model and in-
troduces the control policy. In Section 3, we develop the expressions for the key operating characteristics of
the inventory system, and discuss the special cases of the proposed model. In Section 4, numerical results
are presented on the sensitivity of the optimal policy parameters with respect to various system parameters
and on the performance of the proposed policy in comparison with the classical policies. The paper ends
with some concluding remarks and comments on future work.

2. The model

We consider an inventory system facing unit external demands generated by a Poisson process with
mean k. All unmet demands are lost and there is a positive leadtime, L. We assume that the items in a batch
are perishable, with a constant lifetime s, under a special aging process as explained below. The associated
costs are the xed ordering cost K incurred with each order, the holding cost charged at h per unit held in
stock per unit time and the perishing cost p charged per unit that perishes in stock. There is no direct cost of
lost sales. Instead, we impose a service level constraint that requires the fraction of unmet demand not to
exceed a prespecied value, a.
We assume that the aging of items in a batch begins after all units of the previous batch are exhausted
either by demand or through decay. That is, the aging of a batch starts when the batch is `unpacked' for
consumption. This aging phenomenon was observed by the authors in the production processes of an
aircraft manufacturer. In this application, composite materials are used for the manufacture of aircraft
body panels. The composite fabrics in use are perishable and have a lifetime which depends on the
storage environment. Before being released to the shop oor for cutting and plastering, the fabric is kept
in a special freezer without any deterioration up to six months. Given the low volume of consumption
and the long lifetime in the freezers, composite fabrics are virtually nonperishable materials as long as
they are stored at very low temperatures. As the composite materials are released to the shop oor,
deterioration starts due to environmental exposure; hence, the composite fabrics have a very short
oorlife (life time on the shop oor) measured in weeks. Once the materials are sent out to the pro-
duction area, they stay in this severe environment until they perish or are consumed. Other examples of
perishable goods that experience a similar aging process are some of the adhesives and coatings used in
aircraft and shipbuilding industries, and foodstus in supermarkets that are kept in separate storage
areas prior to being put on shelves. (See also [13,25] for the same aging process.) Note that this special
aging assumption is essentially similar to the commonly made assumption that items in an order do not
deteriorate while in transit. Limited simulation studies we conducted, also indicate that the model de-
veloped herein is a good approximation to slow moving inventory systems in which items begin aging as
soon as they join the stock. With this aging structure, we propose the following modied lotsize reorder
control policy which bases the replenishment decision on both the inventory levels and the remaining
shelives of the items in stock.
Policy. A replenishment order of size Q is placed either when the inventory drops to r, or when T units of
time have elapsed since the last instance at which the inventory level hit Q, whichever occurs rst.
The policy described above is in spirit a modied reorder point-ordering quantity (Q; r) policy [23] and,
therefore, we refer to it as the (Q; r; T ) policy. Further, due to the specied aging process, T corresponds
directly to an age threshold for reorder, whereas, r is an inventory level threshold for reorder in the classical
sense. We assume in the sequel that r < Q, which results in at most one outstanding order at any time given
the assumption of lost sales.
312 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329

Under the above policy, it is easy to see that the instances at which the inventory level hits Q constitute
regenerative epochs. Hence, we dene a cycle as the time between two such consecutive instances, and
employ the renewal reward theorem [26] to derive the operating characteristics of the system. Our objective
is to nd the optimal parameters of the control policy that minimizes the expected total cost rate under the
service level constraint.

3. Operating characteristics

In this section, we derive the expressions for the expectations of cycle length, on hand inventory, number
of lost sales and number of items that perish in a cycle as a function of the policy parameters Q, r and T.
These expressions are then used to construct the average cost function with the service level criterion.
Before we proceed with the derivations, we introduce some notation. Let Xn ; n 1; 2; . . . ; denote the
random variable representing the arrival time of nth demand with distribution function Fn t and let f t be
the p.d.f. of the time interval between successive demands. Also, let N t denote the counting process as-
sociated with the demand process in 0; t, such that Fn t is the probability that N t 6 n, and let
F n t 1 Fn t.
As illustrated in Figs. 1 and 2, there exist eight possible realizations for a cycle, based on the relations
among the age threshold for reorder, the time of the (Q r)th demand, the time of the Qth demand, and the
shelife. Note here that XQ r and XQ are random variables, whereas T and s are nonnegative constants.
Fig. 1 illustrates the realizations where the order placement is triggered by the age threshold, that is,

Fig. 1. Possible inventory level realizations in a cycle T < XQ r .

 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329 313

Fig. 2. Possible inventory level realizations in a cycle T > XQ r .

when T < XQ r , whereas Fig. 2 illustrates the cases where an order is placed when the inventory level drops
to r before perishing, that is, T > XQ r . In the following analysis, we do not allow T to be greater than s,
since it would not be desirable to postpone the ordering decision until a batch has completely deteriorated.
For the sake of clarity, we consider the cases s > T L and T < s < T L separately. Derivations of the
expressions in this section are given in the Appendix A.1.
Case 1. T < s < T L.
Considering all possible realizations for a cycle and combining some events, we express the cycle length,
CL, as follows:
8
>
> XQ r L if XQ r < T ; XQ < minXQ r L; s
>
>
< or XQ r < T ; s < mins; XQ r L;
CL XQ if XQ r < T ; XQ r L < XQ < s; 1
>
>
>
> s if X Q r < s L; X Q > s;
:
T L if T < XQ r < s:

Then the expected cycle length is found as follows:

Q r 
ECL L T F Q r T FQ r1 T FQ r1 s L
h k i
 r
s LFr L Fr1 L FQ r s L gQ; r; T ; 2
k
314 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329

where
Z s L hr i
gQ; r; T Fr1 s t s tFr s t fQ r t dt
0 k
is the expected time that a received batch stays in stock until the previous one perishes or is depleted. We
denote the area under the inventory level curve within a cycle as OH. Considering the possible realizations
mentioned above, it can be expressed as
8 PQ
>
>
> i1 Xi if XQ r < T ; XQ < minXQ r L; s
>
>
>
> or XQ r > T ; XQ < s;
>P
>
>
> N s
< i1 Xi sQ N s
> if s < XQ ; s L < XQ r < T ;
OH or s < XQ ; T < XQ r ; 3
>
> PQ
>
>
>
> i1 Xi QXQ XQ r L if XQ r < T ; XQ r L < XQ < s;
>
> P
>
> N s
>
> i1 Xi sQ N s
:
Qs XQ r L if XQ r < s L; XQ > s:

Taking expectations, we obtain the following expression:

 h i
Q1 r
EOH Q gQ; r; T FQ1 s sF Q s s LFr L Fr1 L FQ r s L
2k k

Q r ks2
FQ r1 s L F Q 1 s: 4
k 2

Similarly, the number of lost sales in a cycle, LS, is

8
> N XQ r L XQ if XQ r < T ; XQ r < XQ < XQ r L; XQ < s;
>
>
< N X
Q r L s if s L < XQ r < T ; XQ > s;
LS 5
>
> N T L s if XQ r > T ; XQ > s;
>
:
N T L XQ if T < XQ r < XQ < s:

Then the expected number of lost sales in a cycle is

h r i
ELS kgQ; r; T sF Q s T F Q r T L k s LFr L Fr1 L FQ r s L
k
Q rFQ r1 T FQ r1 s L QFQ1 s: 6

Finally, the number of items that perish in a cycle, P, is given by


Q N s if XQ > s;
P 7
0 otherwise;

and

EP EQ N sIXQ > s
X
Q 1
QFQ s nP N s n QF Q s ksF Q 1 s: 8
n0
 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329 315

Case 2. s > T L.
The operating characteristics for this case are derived similarly and the following expressions are ob-
tained:
Q
ECL fQ; r; T sF Q s FQ1 s; 9
k
where
Z T L h i
r
fQ; r; T T L tFr T L t Fr1 T L t fQ r t dt
T k
h r i
LFr L Fr1 L FQ r T
k
is the expected duration of a stockout period. Also, we have
 
Q r 3Q 1
EOH Q fQ; r; T TFQ r T FQ r1 T FQ1 s 2sF Q s T L
k 2k
ks2
F Q 1 s 10
2
and
ELS kfQ; r; T ; 11

EP QF Q s ksF Q 1 s: 12

We can now formulate the optimization problem of minimizing the expected total cost rate, TCQ; r; T ,
subject to the service level constraint
K hEOH pEP
min TCQ; r; T 13
ECL

s.t.
ELS
6 a; 14
kECL

where a is the maximum permissible fraction of lost sales. Note that from the theory of Lagrange multi-
pliers, minimizing (13) subject to (14) is equivalent to minimizing a modied long-run average cost with
explicit shortage costs. The foregoing expressions regarding the cost functions render it almost impossible
to derive analytical results about the average cost function. The implementations in the rest of the
manuscript are, therefore, done via numerical tools.
Special cases.
(i) Q; r Model. If we let T s in the original model, we obtain a Q; r type model for perishables,
in which an order is placed when the inventory position hits r or when all items in stock perish at age
s (raising the inventory level to Q). Perishing would not be observed too often due to its high costs in
realistic settings; therefore, the reordering decisions under this policy would typically be based on
only stocking level. We use this model in our numerical study to compare the performances of the
age-based policy and the conventional Q; r type policy. As another case, note that our perishable
age-based policy reduces to the classical Q; r policy for nonperishables [4] as s ! 1, and
T ! 1, respectively.
316 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329

(ii) Q; T Model. Another two-parameter policy is obtained as a special case when we let r 0 in our
model. Noting again that perishing would not be common in realistic cost settings, the reordering deci-
sions under this policy would typically be based only on the age of the batch currently in use. We have
conducted a numerical study between the Q; r and Q; T policies and observed that the Q; r policy
dominated the Q; T over the experimental set we investigated [29]. Hence, we shall not elaborate further
on this special case.

4. Numerical results

We conducted numerical studies to examine the sensitivity of the optimal policy parameters with respect
to various system parameters, and to compare the performance of the Q; r; T policy vis-a-vis the classical
Q; r policy over a wide range of system parameter settings. To be consistent in our comparison of the two
classes of control policies, we maintain the restriction on r r < Q for both policies.

4.1. Sensitivity analysis

We rst discuss our ndings about how the optimal policy parameters and average cost rate change with
respect to ordering cost, perishing cost, permissible fraction of lost sales and shelife. The reported ex-
perimental points represent a broad range of cases from almost no lost sales with high ordering costs and
short lifetimes to lost sales of 10% with low ordering costs and long shelives (see Table 1). Disposal costs
of toxic raw materials like adhesives and composite materials are very high due to strict environmental
regulations; very high unit costs of perishing (p 50) in our numerical set are intended to include such
scenarios.
We present our results in Table 2 for xed values of k 5, h 1 and L 1. We note that most of the
results agree with what one would intuitively expect.
The optimal ordering quantity Q increases as K increases. However, this increase is not as signicant as
it is for nonperishables. For instance, when the shelife is short (s 2), there is almost no change in Q as K
increases. The overall eect of increasing K on reordering is to delay the order placement by either in-
creasing T  or decreasing r (or both).
As the perishing cost p increases, we observe that Q decreases in order to decrease the average number
of items that perish. At the same time, the reordering decisions get to be made sooner, that is at smaller T 
or larger r . When products are subject to decay in a short time and a small fraction of lost sales is desired
(a < 0:02), r is not sensitive to the changes in p, but T  decreases considerably as p increases. The expected
cost rate is observed to be more sensitive to T when a and s values are small.
The length of the shelife has a noticeable eect on the optimal ordering quantity. As the shelife in-
creases, Q increases considerably. At the same time, r tends to decrease slightly whereas T  tends to

Table 1
Test parameters
Parameter Symbol Values tested
Frac. of lost sales a 0.005, 0.01, 0.02, 0.05, 0.1
Ordering cost K 50, 100
Perishing cost p 1, 10, 50
Shelf life s 2, 4, 6, 12, 15, 20
Demand rate k 0.25, 0.5, 1, 2, 5, 10
Table 2
Sensitivity results w.r.t. K; s; p; a
k5 p1 p 10 p 50
a Q1 r1 T C1 Q1 r1 T C1 Q1 r1 T C1
K 50 s2 0.005 13 9 1.00 37.24 10 9 0.23 45.01 10 9 0.23 73.46
0.01 12 8 0.94 36.22 9 8 0.11 43.28 9 8 0.11 62.42
0.02 12 7 1.00 35.22 9 7 0.30 42.09 8 7 0.05 54.60
0.05 11 5 0.98 33.45 9 6 0.84 39.53 7 6 0.10 49.18
0.1 11 2 1.05 31.48 9 5 1.16 36.90 7 5 0.50 45.79
s4 0.005 21 8 2.95 27.56 16 9 4.00 29.63 13 9 4.00 31.78
0.01 20 7 2.81 26.70 16 8 4.00 28.54 13 8 4.00 30.69
0.02 20 6 3.01 25.61 16 7 4.00 27.41 13 7 4.00 29.52
0.05 19 4 2.99 23.96 16 5 3.02 25.34 13 5 1.79 27.81
0.1 18 2 3.02 22.32 16 3 2.89 23.58 13 4 3.19 25.48
s6 0.005 24 8 4.79 26.04 22 8 3.65 26.48 20 8 2.88 27.16
0.01 24 7 4.57 25.17 22 7 3.70 25.61 20 7 2.99 26.29
0.02 23 6 4.61 24.08 22 6 4.10 24.47 20 6 3.30 25.15
0.05 24 4 4.76 22.57 22 4 4.05 22.97 19 5 6.00 23.32
0.1 20 3 6.00 20.92 20 3 6.00 21.02 19 3 4.61 21.37

K 100 s2 0.005 13 9 1.00 62.96 10 9 0.23 73.44 10 9 0.23 101.90

0.01 14 7 1.00 61.92 10 8 0.38 72.29 9 8 0.11 92.58
0.02 13 6 1.00 60.67 10 7 0.57 70.72 8 7 0.05 87.11
0.05 12 2 0.99 58.59 10 6 1.03 67.47 8 6 0.44 83.37
0.1 12 4 1.15 55.49 10 3 0.97 63.35 8 5 1.03 78.13
s4 0.005 21 8 2.95 40.89 18 8 2.14 44.92 15 9 4.00 49.92
0.01 21 7 3.00 40.01 18 8 3.20 43.87 15 8 4.00 48.74
0.02 20 6 3.01 39.06 18 6 2.54 42.57 15 7 4.00 47.41
0.05 20 4 3.05 37.36 17 5 3.11 40.20 15 5 2.57 44.82
0.1 19 2 3.13 35.26 17 3 3.10 37.63 15 3 2.55 42.04
s6 0.005 27 8 5.03 36.09 25 8 5.02 37.19 22 8 3.65 39.13
E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329

0.01 27 7 5.05 35.14 25 7 5.01 36.18 22 7 3.70 38.19

0.02 26 6 5.11 34.15 24 6 5.03 35.06 22 6 4.10 36.92
0.05 26 4 5.13 32.29 25 4 5.04 33.19 22 4 4.05 35.05
0.1 26 1 5.04 30.33 24 2 5.05 31.06 21 3 6.00 32.64
317
318 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329

Table 3
Sensitivity results w.r.t. L; p; a
k5 p1 p 10 p 50
s2 a Q1 r1 T C1 Q1 r1 T C1 Q1 r1 T C1
L 0:5 0.005 12 5 1.50 36.12 9 5 0.82 43.09 7 5 0.28 52.93
0.01 12 4 1.50 35.39 9 5 1.59 42.18 7 5 2.00 51.84
0.02 12 3 1.50 34.77 9 4 1.52 40.93 7 4 0.77 50.76
0.05 11 2 1.52 33.40 9 3 1.58 39.10 7 3 1.11 48.36
0.1 11 0 1.58 31.49 9 1 1.45 36.64 7 2 1.34 45.30
L1 0.005 13 9 1.00 37.24 10 9 0.23 45.01 10 9 0.23 73.46
0.01 12 8 0.94 36.22 9 8 0.11 43.28 9 8 0.11 62.42
0.02 12 7 1.00 35.22 9 7 0.30 42.09 8 7 0.05 54.60
0.05 11 5 0.98 33.45 9 6 0.84 39.53 7 6 0.10 49.18
0.1 11 2 1.05 31.48 9 5 1.16 36.90 7 5 0.50 45.79
L 1:5 0.005 13 12 0.04 38.41 13 12 0.04 53.79 13 12 0.04 122.12
0.01 14 11 0.46 37.06 12 3 0.03 49.17 12 3 0.03 102.09
0.02 13 10 0.49 35.56 11 10 0.04 45.05 11 10 0.04 84.27
0.05 12 7 0.50 33.50 10 9 0.25 40.68 10 9 0.25 67.86
0.1 11 5 0.54 31.50 9 8 1.13 37.13 9 8 1.13 54.53

increase. It seems that the eect of the age threshold decreases with increasing shelife. As expected, the
average costs are increasing in p and K but decreasing in s and a.
For the inventory system under consideration, the ordering decision depends not only on the demand
rate but also on the lifetime of items. In order to avoid lost sales during the leadtime, the perishing time is as
important as the number of demands during the leadtime. Thus, we are particularly interested in the in-
teraction of the leadtime and the product lifetime. For this purpose we report the case with K 50; s 2
with dierent choices for L which take the values of 0:25s, 0:5s and 0:75s. Table 3 displays our results for
this setting.
For xed a, as the leadtime increases, Q and r increase, and T  decreases. This is to be expected because
a larger leadtime increases the risk of having lost sales during this leadtime. The reordering parameters are
observed to be more sensitive to the changes in leadtime. The average cost rate is also increasing in
leadtime.

4.2. Comparison with Q; r policy

As we propose a new policy for controlling perishable inventories, it is of interest to compare the
performance of the Q; r; T policy with that of the conventional Q; r policy. For this purpose, we tested
the two policies over a wide range of parameter settings. We report our ndings on the representative cases
of low and relatively high demand rates k 0:25; 5 for dierent values of s, p and a. The xed parameters
are K 50; L 1 and h 1. In Tables 4 and 5, we present the optimal parameters, the corresponding
expected cost rates of the two policies, and the percentage improvement in the average cost rate under the
Q; r; T policy over the optimal cost under the Q; r policy. Note that the improvement obtained by the
Q; r; T policy is higher when a is low a < 0:02 which corresponds to high service levels. When the service
level constraint is tight, the Q; r policy increases r in order not to take the risk of losing sales during the
leadtime. However, the Q; r; T policy operates with smaller r and chooses to order proactively at time T
which results in lower average costs. As the service level constraint gets looser, superiority of the Q; r; T
policy diminishes.
 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329 319

Table 4
Q; r; T vs. Q; r k 0:25
a Q1 r1 T C1 Q2 r2 C2 D%
p1 s 12 0.005 4 1 9.84 8.19 5 4 11.11 26.28
0.01 4 1 11.09 7.93 3 2 9.01 11.99
0.02 4 1 11.33 7.85 3 2 9.01 12.87
0.05 4 0 11.34 7.28 4 1 7.64 4.71
0.1 3 0 12.00 6.87 3 0 6.87 0.0
s 15 0.005 5 1 14.02 7.53 3 2 8.53 11.72
0.01 4 1 14.14 7.244 3 2 8.53 15.12
0.02 4 1 14.56 7.17 4 2 8.08 11.26
0.05 4 0 14.30 6.51 4 1 7.10 8.31
0.1 4 0 15.00 6.34 4 0 6.34 0.00
s 20 0.005 5 1 19.00 6.80 4 2 7.67 11.34
0.01 4 1 19.31 6.70 5 2 7.59 11.73
0.02 4 1 20.00 6.66 4 1 6.66 0.00
0.05 4 0 19.11 7.85 4 1 6.66 12.16
0.1 4 0 20.00 5.76 4 0 5.76 0.00
p 10 s 12 0.005 4 1 9.84 9.29 5 4 12.78 27.31
0.01 3 1 11.02 8.68 3 2 9.66 10.14
0.02 3 1 11.43 8.60 3 2 9.66 10.97
0.05 3 0 10.03 7.97 3 1 8.49 6.12
0.1 3 0 12.00 7.45 3 0 7.45 0.00
s 15 0.005 4 1 11.04 8.29 3 2 8.88 6.64
0.01 3 1 14.02 7.86 3 2 8.88 11.49
0.02 3 1 14.88 7.78 4 2 8.72 10.78
0.05 4 0 14.30 7.13 4 1 7.73 7.76
0.1 3 0 15.00 6.68 3 0 6.68 0.00
s 20 0.005 5 1 19.00 7.28 4 2 7.95 8.43
0.01 4 1 19.31 6.98 4 2 7.95 12.20
0.02 4 1 20.00 6.93 4 1 6.93 0.00
0.05 4 0 19.11 6.11 4 1 6.93 11.83
0.1 4 0 20.00 6.01 4 0 6.01 0.00
p 50 s 12 0.005 2 1 1.12 11.84 5 4 20.02 41.39
0.01 2 1 3.47 11.42 3 2 12.53 8.86
0.02 2 1 11.58 10.91 3 2 12.53 12.93
0.05 2 0 4.65 10.57 2 1 10.88 2.85
0.1 2 0 10.92 9.42 3 0 10.06 6.36
s 15 0.005 3 1 5.77 10.22 3 2 10.47 2.39
0.01 3 1 14.02 9.44 3 2 10.47 9.84
0.02 3 1 14.88 9.34 2 1 9.68 3.51
0.05 3 0 10.94 8.77 3 1 9.33 6.00
0.1 3 0 15.00 8.13 3 0 8.13 0.00
s 20 0.005 3 1 6.29 8.59 3 2 8.89 3.37
0.01 3 1 19.08 7.85 3 2 8.89 11.70
0.02 3 1 20.00 7.82 3 1 7.82 0.00
0.05 4 0 19.11 7.28 3 1 7.82 6.91
0.1 3 0 20.00 6.63 3 0 6.63 0.00

When a is low, parameters of the Q; r policy are not sensitive to changes in p. On the other hand,
those of the Q; r; T policy depend highly on the value of p. The exible nature of the Q; r; T policy
exploits better the trade-o between the tight service level constraint and the risk of incurring high
perishing costs.
320 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329

Table 5
Q; r; T vs. Q; r k 5
a Q1 r1 T C1 Q2 r2 C2 D%
p1 s2 0.005 13 9 1.00 37.24 11 10 38.67 3.70
0.01 12 8 0.94 36.22 12 10 38.15 5.06
0.02 12 7 1.00 35.22 10 8 36.74 4.14
0.05 11 5 0.98 33.45 10 7 34.95 4.29
0.1 11 2 1.05 31.48 12 7 32.41 2.87
s4 0.005 21 8 2.95 27.56 17 9 28.16 2.13
0.01 20 7 2.81 26.70 16 8 27.38 2.48
0.02 20 6 3.01 25.61 16 7 26.26 2.48
0.05 19 4 2.99 23.96 18 6 24.58 2.52
0.1 18 2 3.02 22.32 17 4 22.50 0.80
s6 0.005 24 8 4.79 26.04 22 9 26.90 3.20
0.01 24 7 4.57 25.17 22 8 25.88 2.74
0.02 23 6 4.61 24.08 22 7 24.86 3.14
0.05 324 4 4.76 22.57 21 5 22.85 1.23
0.1 20 3 6.00 20.92 20 3 20.92 0.00
p 10 s2 0.005 10 9 0.23 45.01 11 10 47.65 5.54
0.01 9 8 0.11 43.28 10 9 44.58 2.92
0.02 9 7 0.30 42.09 9 8 42.47 0.89
0.05 9 6 0.84 39.53 9 7 40.75 2.99
0.1 9 5 1.16 36.90 9 6 38.55 4.28
s4 0.005 16 9 4.00 29.63 16 9 29.63 0.00
0.01 16 8 4.00 28.54 16 8 28.54 0.00
0.02 16 7 4.00 27.41 16 7 27.41 0.00
0.05 16 5 3.02 25.34 16 6 26.22 3.36
0.1 16 3 2.89 23.58 15 4 23.72 0.59
s6 0.005 22 8 3.65 26.48 21 9 27.11 2.32
0.01 22 7 3.70 25.61 21 8 26.09 1.84
0.02 22 6 4.10 24.47 21 7 25.06 2.35
0.05 22 4 4.05 22.97 20 5 22.99 0.09
0.1 20 3 6.00 21.02 20 3 21.02 0.00
p 50 s2 0.005 10 9 0.23 73.46 11 10 87.54 16.08
0.01 9 8 0.11 62.42 10 9 72.99 14.48
0.02 8 7 0.05 54.60 9 8 61.50 11.22
0.05 7 6 0.10 49.18 8 7 53.15 7.47
0.1 7 5 0.50 45.79 7 6 47.73 4.06
s4 0.005 13 9 4.00 31.78 13 9 31.78 0.00
0.01 13 8 4.00 30.69 13 8 30.69 0.00
0.02 13 7 4.00 29.52 13 7 29.52 0.00
0.05 13 5 1.79 27.81 13 6 28.25 1.56
0.1 13 4 3.19 25.48 14 4 25.80 1.24
s6 0.005 20 8 2.88 27.16 19 9 27.51 1.27
0.01 20 7 2.99 26.29 19 8 26.48 0.72
0.02 20 6 3.30 25.15 19 7 25.44 1.14
0.05 19 5 6.00 23.32 19 5 23.32 0.00
0.1 19 3 4.61 21.37 20 3 21.46 0.42

We observe the highest percentage dierence in the average costs of the two policies when the lifetime of
items is short and the service level is high. For these cases, parameters of the Q; r policy do not change as p
increases. The reason for this phenomenon is the one order outstanding assumption. The Q; r policy keeps
r as high as possible in order to satisfy the constraint on service level and hence operates with large Q. But
 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329 321

the Q; r; T policy can easily satisfy this assumption by keeping r low and placing orders proactively. Thus
it can decrease the value of Q in order to avoid the risk of perishing especially when p is large. For large p,
an improvement of more than 40% may be obtained.
We do not observe a monotonic behavior in the percentage dierences between the policies with respect
to s; p and a. The reason for this is that the Q; r policy selects the optimal values of Q and r from a
discrete space. For instance, when s 20 and a 0:02, the average costs for the two policies are the same.
When we increase a to 5%, the percentage deviation between the two policies becomes 12.16%. Here we
observe that the Q; r policy is the same for both a values, that is it cannot nd any better values for Q and
r. However, as T is a continuous variable Q; r; T policy attains a lower average cost by adjusting T. The
main conclusions of our experimental study can be summarized as follows. The Q; r; T policy outperforms
the classical Q; r policy in 140 experimental points out of 180. The maximum savings are observed for the
cases where the demand for items is low and the service level is high. The mean improvement for these cases
is approximately 12%. Some cases of high desired service levels in the experimental setup are not even
feasible under the Q; r policy. The introduction of an age threshold for reorder becomes less eective in
reducing costs as the shelife of products gets longer as expected.

5. Conclusions and future work

In this paper, we have proposed a control policy for perishables which bases reordering decisions not only
on the current level of inventory but also explicitly on the remaining shelife of the items in stock. Under this
policy, we have developed a lost sales perishable inventory model that is suitable for environments where the
items in a batch are preserved without deterioration until the batch is unpacked for usage.
The proposed Q; r; T policy has been observed to be superior to the conventional Q; r policy which
ignores the lifetime status of items in stock. The proposed policy performs particularly well under tight
service level constraints for slow moving items with short lifetimes.
A generalization of our work would be to allow the aging of the batches to start as soon as they join the
stock. In this case, since the items start to age immediately after the order is received, at the instance when
the inventory level hits Q, the items in the batch may not be fresh with the remaining shelives being
random due to demand realizations after the order placement. Hence, such instances do not constitute
regeneration points as in our model and a single age threshold policy may not be appropriate. An age-based
policy in the spirit of the one proposed herein would consist of a surface of age thresholds that depends on
the remaining lifetimes of the items as a new batch is unpacked. Similarly, the introduction of backordering
into our model would call for a more complex policy structure.

Acknowledgement

This research is partially supported by TUBITAK MISAG-104 Grant.

Appendix A

A.1. Derivation of (2)

Letting I
denote the indicator function of the event stated in the argument, the expected cycle length is
given by
322 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329

ECL EXQ r LIXQ r < T ; XQ < minXQ r ; L; s

or XQ r < T ; s < minXQ ; XQ r L

EXQ r LIs L < XQ r < T ; XQ > s

EXQ IXQ r < T ; XQ r L < XQ < s

EsIXQ r <s L; XQ > s ET LIT < XQ r < s:

Then, noting that we can express XQ as the sum of two independent components as XQ XQ r Xr , we have
Z s L Z L
ECL x LfQ r xfr y dy dx
x0 y0

Z T Z 1
x LfQ r xfr y dy dx
xs L y0

Z s L Z s x
x yfQ r xfr y dy dx
x0 yL

Z s L Z 1 Z s
s fQ r xfr y dy dx T L fQ r x dx;
x0 ys x xT

which can be written as

Q r
ECL Fr LFQ r1 s L LFr L FQ r s L
k
Z T Z T
Q r
Fr s xfQ r1 x dx L Fr s xfQ r x dx
k s L s L

Z T
Q r Q r
FQ r1 T FQ r1 s L Fr s xfQ r1 x dx
k k s L

LFQ r T FQ r s L
Z T Z s L
Q r
L Fr s xfQ r x dx Fr s xfQ r1 x dx
s L k 0

Z s L
Q r r
Fr LFQ r1 s L Fr1 s xfQ r x dx
k k 0

r
Fr1 LFQ r s L sFQ r s L
k
Z s L
s Fr s xfQ r x dx T L1 FQ r T :
0
 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329 323

Finally, we can simplify it as

Z s L
Q r
ECL L T F Q r T Fr s xfQ r1 x dx
k 0
Z s L Z s L
r
Fr1 s xfQ r x dx s Fr s xfQ r x dx
k 0 0
h r i
L Fr L Fr1 L s L FQ r s L
k
Q r
FQ r1 T FQ r1 s L:
k

The result is obtained after some algebraic manipulation.

A.2. Derivation of (4)

Expected value of (3) is written as

" N s # " #
X X
Q
EOH E Xi IXQ > s E Xi IXQ < s
i1 i1

" N s #
X  
E Xi IXQ > s E sQ N sIXQ > s
i1

 
E QXQ XQ r LIXQ r < T ; XQ L < XQ < s
r

 
E Qs XQ r LIXQ r <s L; XQ > s :

Expanding each of the terms separately, we get

" N s # " N s #
X X
EOH 1 E Xi IXQ > s E Xi IN s < Q
i1 i1

" #
X
N s
E E Xi IN s < Q j N s n
i1

X
Q 1
s ks
n
ks2  
ks
n e 1 FQ 1 s :
i1
2 n! 2

Noting that the joint density of X1 ; X2 ; . . . ; xQ is given by

f x1 ; x2 ; . . . ; xQ ke kx1
ke kx2 x1
. . . ke kxQ xQ 1
kQ e kxQ
;
324 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329

we have
" #
X
Q
EOH 2 E Xi IXQ < s
i1
Z s Z xQ Z x2



x1 x2


xQ kQ e kxQ
dx1 dx2


dxQ
xQ 0 xQ 1 0 x1 0

Q Z
X s Z xQ Z xi1 Z x2






xi kQ e kxQ
dx1


dxi


dxQ
i1 xQ 0 xQ 1 0 xi 0 x1 0

XQ
i QQ 1
FQ1 s FQ1 s;
i1
k 2k

EOH 3 EsQ N sIXQ > s

Qs1 FQ s sEN sIN s < Q

Qs1 FQ s ks2 1 FaQ 1 s;

EOH 4 EQXQ XQ r LIXQ r < T ; XQ r L < XQ < s

Z s LZ s x
Q y LfQ r xfr y dy dx
x0 yL
Z
r s L
Q Fr1 s xfQ r x dx
k x0
r
Q Fr1 LFQ r s L QLFr LFQ r s L
k
Z s L
QL Fr s xfQ r x dx:
x0

Finally

EOH 5 EQs XQ r LIXQ r < s L; XQ > s

Z s LZ 1
Qs L fQ r xfr y dy dx
x0 ys x
Z s L Z 1
Q xfQ r xfr y dy dx
x0 ys x
Z s L
Qs LFQ r s L Qs L fQ r xFr s x dx
x0
Z s L
Q r Q r
Q FQ r1 s L Q fQ r1 xFr s x dx:
k k x0

The result is obtained by summing the above terms.

 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329 325

A.3. Derivation of (9)

We start with evaluating the expected cycle length. Observe that we can write
8
> XQ X Q r < T ; X Q r L < XQ < s
>
>
>
> or XQ r > T ; T L < XQ < s;
<
CL XQ r L XQ r < T ; XQ r < XQ > XQ r L; A:1
>
>
>
> s XQ > s;
>
:
T L T < XQ r < XQ < T L;

so that, after integration we have

Z T Z s x
ECL x yfQ r xfr y dy dx
x0 yL
Z T L Z s x
x yfQ r xfr y dy dx
xT yT L x
Z s Z s x
x yfQ r xfr y dy dx
xT L y0
Z T Z L
x LfQ r xfr y dy dx
x0 y0
Z T L Z T L x
sEIXQ > s T L fQ r xfr y dy dx A:2
xT y0

and evaluation of the integrals results in

Z
Q Q r T L
ECL FQ1 s s1 FQ s Fr T L xfQ r1 x dx
k k T
Z T L h i
r
T L Fr T L xfQ r x dx LFr L Fr1 L FQ r T
T k
Z
r T L
Fr1 T L xfQ r x dx: A:3
k T

A.4. Derivation of (10)

Writing
8 PQ
>
> i1 Xi QXQ XQ r L XQ r < T ; XQ r L < XQ < s;
>
> PQ
>
>
> i1 Xi
>
>
XQ r < T ; XQ > XQ r L
>
< or XQ r > T ; XQ < T L;
OH PN s A:4
>
> i1 Xi sQ N s Qs XQ L XQ < T ; XQ > s;
>
>
r r
>
> PN s
>
> Xi sQ N T Qs T L XQ > T ; XQ > s;
>
> Pi1
r
: Q
i1 Xi QXQ T L XQ r > T ; T L < XQ < s;
326 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329

we have
" N s # " #
X X
Q
EOH E Xi IXQ > s E Xi IXQ < s
i1 i1
 
E QXQ XQ r LIXQ r < T ; XQ r L < XQ < s
 
E sQ N sIXQ > s
 
E Qs XQ r LIXQ r < T ; XQ > s
 
E Qs T LIXQ r > T ; XQ > s
 
E QXQ T LIXQ r > T ; T L < XQ < s :

We evaluate the components of the above expression separately as follows:

" N s # " " N s ##
X X
EOH 1 E Xi IXQ > s E E Xi IN s < Q j N s n
i1 i1

X
Q 1
s s X
Q 1
n P N s n nP N s n
n0
2 2 n0
" #
s X
1
ks2  
ks ks P N s n 1 FQ 1 s ; A:5
2 nQ 1
2

" #
X
Q
QQ 1
EOH 2 E Xi IXQ < s FQ1 s; A:6
i1
2k

 
EOH 3 E QXQ XQ r LIXQ r < T ; XQ r L < XQ < s
Z T Z s x
Q y LfQ r xfr y dy dx
x0 yL
Z T Z s x
Q yfQ r xfr y dy dx
x0 yL
Z T Z s x
QL fQ r xfr y dy dx; A:7
x0 yL

EOH 4 EsQ N sIXQ > s

X
Q 1
QsFQ s s nP N s n
n0
" #
X
1
QsFQ s s ks nP N s n
nQ

QsFQ s ks2 FQ 1 s; A:8

 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329 327

EOH 5 EQs XQ r LIXQ r < T ; XQ > s

Z T Z 1
Qs L fQ r xfr y dy dx
x0 ys x
Z T Z 1
Q xfQ r xfr y dy dx
x0 ys x
Z T
Qs LFQ r T Qs L fQ r xFr s x dx
x0
Z T
Q r Q r
Q FQ r1 T Q fQ r1 xFr s x dx; A:9
k k x0

EOH 6 EQs T LIXQ r > T ; XQ > s

Z s Z 1
Qs T L fQ r xfr y dy dx
xT ys x
Z 1 Z 1
Qs T L fQ r xfr y dy dx
xs y0

Qs T L Qs T LFQ r T
Z s
Qs T L fQ r x1 Fr s x dx: A:10
xT

Collecting these terms yields the result.

A.5. Derivation of (11)

(
N XQ r L XQ if XQ r < T ; X Q < XQ r L;
LS A:11
N T L XQ if T < XQ r ; XQ < T L:

Taking expectations, we have

ELS1 EN XQ r L XQ IXQ r < T ; X Q < XQ r L;

ELS2 EN T L XQ IT < XQ r ; XQ < T L
Z T L Z T L x
kT L fQ r xfr y dy dx
xT y0
Z T L Z T L x
k xfQ r xfr y dy dx
xT y0
Z T L Z T L x
k yfQ r xfr y dy dx
xT y0
328 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329

Z T L
kT L Fr T L xfQ r x dx
xT
Z T L
Q r Fr T L xfQ r1 x dx
xT
Z T L
r Fr1 T L xfQ r x dx:
xT

Thus
Z T L
ELS FQ r T kLFr L rFr1 L kT L Fr T L xfQ r x dx
xT
Z T L  
rFr1 T L xfQ r x Q rFr T L xfQ r1 x dx: A:12
xT

A.6. Derivation of (12)

EP EQ N sIXQ > s
X
Q 1
Q1 FQ s nP N s n Q1 FQ s ks1 FQ 1 s:
n0

Further details of the derivations in this section may be found in Tekin [30].

References

[1] H.N. Chiu, An approximation to the continuous review inventory model with perishable items and leadtimes, European Journal
of Operations Research 87 (1995) 93108.
[2] M.A. Cohen, Analysis of single critical number ordering policies for perishable inventories, Operations Research 24 (1976)
726741.
[3] B. Fries, Optimal ordering policy for a perishable commodity with xed lifetime, Operational Research 23 (1975) 4661.
[4] G.J. Hadley, T.M. Whitin, Analysis of Inventory Systems, Prentice-Hall, Englewood Clis, NJ, 1963.
[5] S. Kalpakam, G. Arivarignan, A continuous review perishable inventory model, Statistics 19 (1988) 389398.
[6] S. Kalpakam, K.P. Sapna, Continuous review s; S inventory system with random lifetimes and positive leadtimes, Operations
Research Letters 16 (1994) 115119.
[7] S. Kalpakam, K.P. Sapna, A lost sales S 1; S perishable inventory system with renewal demand, Naval Research Logistics 43
(1996) 129142.
[8] L. Liu, s; S continuous review inventory models for inventory with random lifetimes, Operations Research Letters 9 (1990)
161169.
[9] L. Liu, D. Shi, An s; S model for inventory with exponential lifetimes and renewal demands, Naval Research Logistics 46.
[10] Z. Lian, L. Liu, A discrete-time model for perishable inventory systems, Annals of Operation Research 87 (1999) 103116.
[11] L. Liu, Z. Lian, s; S continuous review model for products with xed lifetimes, Operations Research 47 (1999) 150158.
[12] L. Liu, T. Yang, An s; S random lifetime inventory model with a positive leadtime, European Journal of Operations Research
113 (1999) 5263.
[13] A.K. Moorthy, Y.C. Narasimhulu, I.R. Basha, On perishable inventory with Markov chain demand quantities, International
Journal of Information and Managment Sciences 3 (1992) 2937.
[14] S. Nahmias, W.P. Pierskalla, Optimal ordering policies for a product that perishes in two periods subject to stochastic demand,
Naval Research Logistics Quarterly 20 (1973) 207229.
[15] S. Nahmias, Optimal ordering policies for perishable inventory, Operations Research 23 (1975) 735749.
 E. Tekin et al. / European Journal of Operational Research 134 (2001) 309329 329

[16] S. Nahmias, A comparison of alternative approximations for ordering perishable inventory, INFOR 13 (1975) 175184.
[17] S. Nahmias, On ordering perishable inventory when both demand and lifetime are random, Management Science 24 (1977) 8290.
[18] S. Nahmias, The xed-charge perishable inventory problem, Operations Research 26 (1978) 464481.
[19] S. Nahmias, Perishable inventory theory: A review, Operations Research 30 (1982) 681707.
[20] S. Nahmias, S. Wang, A heuristic lot size reorder point model for decaying inventories, Management Science 25 (1979) 9097.
[21] P. Nandsakumar, T.T. Morton, Near myopic heuristics for the xed-life perishable problem, Management Science 39 (1993)
14901498.
[22] D. Perry, M.J.M. Posner, An S 1; S inventory system with xed shelife and constant leadtime, Operations Research S46
(1998) 565571.
[23] R. Peterson, E.A. Silver, Decision Systems for Inventory Management and Production Planning, Wiley, New York, 1979.
[24] F. Raafat, Survey of literature on continuously deteriorating inventory models, Operational Research Society 42 (1991) 2737.
[25] N. Ravichandran, Stochastic analysis of a continuous review perishable inventory system with positive leadtime and poisson
demands, European Journal of Operational Research 84 (1995) 444457.
[26] S. Ross, Stochastic Processes, Wiley, New York, 1983.
[27] C.P. Schmidt, S. Nahmias, S 1; S policies for perishable inventory, Management Science 31 (1985) 719728.
[28] C.R. Schultz, Replenishment delays for expensive slow-moving items, Management Science 35 (1989) 14541462.
[29] E. Tekin, A time-based control policy for a perishable inventory system with lost sales, M.Sc. thesis, Department of Industrial
Engineering, Bilkent University, Ankara, Turkey.
[30] G.J.J. Van Zyl, Inventory control for perishable commodities, unpublished Ph.D. dissertation, University of North Carolina,
1964.
[31] H. Weiss, Optimal ordering policies for continuous review perishable inventory models, Operations Research 28 (1980) 365374.