European Journal of Operational Research 115 (1999) 195214

Theory and Methodology

An investigation of lead-time eects in manufacturing/

remanufacturing systems under simple PUSH and PULL control
strategies
Erwin van der Laan a, Marc Salomon

b,*

, Rommert Dekker

Erasmus Universiteit Rotterdam, P.O. Box 1738, NL-3000 DR, Rotterdam, The Netherlands
Katholieke Universiteit Brabaat, Department of Econometrics, P.O. Box 90153, NL-5000 LE Tilburg, The Netherlands
Received 1 March 1996; accepted 1 March 1998

Abstract
In this paper we extend the PUSH and PULL control strategies dened by van der Laan et al. (E.A. van der Laan,
M. Salomon, R. Dekker, Production planning and inventory control for remanufacturable durable products, Working
paper 9531/A, Econometric Institute, Erasmus University Rotterdam, The Netherlands, 1995) to evaluate numerically
the eects of lead-time duration and lead-time variability on total expected costs in production/inventory systems with
remanufacturing. Although both strategies are non-optimal, they are relatively easy to analyse numerically and, more
importantly, they are actually used in practice. The most important outcomes of the study are, that for both control
strategies: (i) manufacturing lead-times have a larger inuence on system costs than remanufacturing lead-times; (ii) a
larger remanufacturing lead-time may sometimes result in a cost decrease; and (iii) a larger variability in the manufacturing lead-time may sometimes result in a cost decrease. 1999 Elsevier Science B.V. All rights reserved.
Keywords: Production planning and inventory control; Manufacturing; Remanufacturing; Statistical re-order point
models; Computational experiments

1. Introduction
In Ref. [1] one of the rst attempts has been made to investigate the eects of remanufacturing on costs
in hybrid manufacturing/remanufacturing systems. In [1] a single product system is considered, in which
demands for new products could either be fullled by manufacturing new products, or by remanufacturing
used products. The processes and goods ows in this system are shown in Fig. 1.
The production and inventory system of Fig. 1 is a simplied version of a system that has been implemented at a large Dutch manufacturer of photocopiers. Some years ago the manufacturer had developed
*

Corresponding author. E-mail: M.Salomon@KUB.nl

0377-2217/99/$ see front matter 1999 Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 7 - 2 2 1 7 ( 9 8 ) 0 0 1 0 8 - 8

196

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

Fig. 1. A hybrid system with manufacturing and remanufacturing operations, and stocking points for remanufacturables and serviceables.

a new generation of photocopiers consisting of easy-to-disassemble modules. After use those photocopier
modules that satisfy certain quality requirements are remanufactured and assembled with newly manufactured modules to obtain a copier that is sold in the market for new products. Modules that do not satisfy
the quality standards for remanufacturing are either used as spare-parts for the second-hand market or are
disposed of.
To plan and control manufacturing and remanufacturing operations simultaneously, the copier remanufacturer initially implemented a PUSH-strategy. By this strategy returned modules are `pushed'
through the remanufacturing process as soon as a sucient amount of modules is available. Only when the
joint inventory of new and remanufactured modules (the serviceable inventory) appears to be too low to
satisfy the future expected demands adequately, a manufacturing order is placed to produce new modules.
In principle, disposal of a disassembled module occurs only when the quality of the module is insucient
for remanufacturing.
The copier manufacturer had the impression that the eciency of their system could be improved by the
introduction of a control strategy which oers a higher level of coordination between manufacturing and
remanufacturing operations. For this purpose, the change to a PULL-strategy had been investigated.
Under this strategy disassembled modules are `pulled' through the remanufacturing process only when they
are actually needed to satisfy the demand for new modules. If the output of the remanufacturing process
appears to be too low to cover the future expected demands, a manufacturing order is placed.
The PUSH- and PULL-strategies that we consider in this paper rely on the strategies that we observed at
the copier manufacturer. More formally the strategies are as follows.
The sm ; Qm ; Qr PUSH-strategy (Fig. 2(a)): In this strategy remanufacturing starts whenever the inventory of remanufacturables contains Qr used products. In that case, all Qr products enter the remanufacturing process to be remanufactured. Manufacturing takes place in batches of size Qm , and starts
whenever the serviceable inventory position (serviceable inventory minus backlog plus all products in
(re)manufacturing work in process) drops to the level sm . The strategy is named PUSH-strategy since
used products are pushed into the remanufacturing process as soon as possible, independently from
the actual demands and from the on-hand serviceable inventory.
The sm ; Qm ; sr ; Sr PULL-strategy (Fig. 2(b)): In this strategy remanufacturing starts whenever the serviceable inventory position is at or below sr , and sucient remanufacturable inventory exists to increase
the serviceable inventory position to Sr . Manufacturing starts whenever the serviceable inventory posi-

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

197

Fig. 2. A schematic representation of (a) the PUSH-strategy and (b) the PULL-strategy.

tion drops to the level sm 6 sr . The manufacturing batch size is Qm . The strategy is named PULL-strategy since remanufacturable inventory is pulled into the remanufacturing process only when needed to
fulll customer demands for serviceables.
Although it has been shown in [1] that neither the PUSH-strategy, nor the PULL-strategy are optimal
within the class of all control strategies, our motivation to investigate these strategies is that they are not
too complex to be analysed numerically, and more importantly, they are actually used in practice. In [2]
Inderfurth proves for a much simpler system than ours, that a simple PULL-strategy is optimal. The

198

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

characteristics of this system include periodic review, zero xed ordering costs, and equal non-stochastic
lead-times for manufacturing and remanufacturing. Additionally, he argues that under stochastic leadtimes and/or xed ordering costs the structure of the optimal control strategy will be very complex. Furthermore, suppose that the optimal structure could be identied. Even then it is highly questionable
whether the optimal strategy could ever be implemented in practice. Also, it should be noted that in a
situation with xed ordering costs a strategy without batching decisions (as in [2]) may not be a very good
approximation to the optimal strategy.
This paper extends the analysis of [1] to allow for stochastic (re)manufacturing lead-times. While [1]
presents a numerical study in which total average system costs are compared under dierent assumptions
with respect to demand processes, return processes, and cost structure, this paper extends the numerical
study to investigate the eects of lead-times, in particular lead-time duration and lead-time variability. The
motivations to devote a separate paper to the analysis of lead-times are, that the analysis requires a nontrivial modication of the cost calculation procedure, and that some of the eects of lead-times on costs
need further explanation since they appear somewhat counter-intuitive. To our knowledge no other paper
studies the eects of stochastic lead-times on system behaviour in a production environment with remanufacturing operations. Muckstadt and Isaac [3] study a similar model but do not consider stochastic
manufacturing lead-times nor holding costs for remanufacturables. Only a push strategy is considered and
the analysis is not exact. For a general qualitative overview of operations research models for reverse logistics we refer to the paper by Fleischmann et al. [4].
This paper is further organized as follows. Besides an introduction of the notation and the system assumptions, Section 2 denes the cost function to be minimized under the PUSH- and the PULL-strategy.
Furthermore, Section 2 outlines the procedures to numerically evaluate the cost function in presence of
stochastic lead-times. Section 3 presents the set-up of an empirical study and discusses the outcomes of the
study. Finally, concluding remarks are presented in Section 4.
2. System assumptions and system analysis
The hybrid manufacturing/remanufacturing system that is central here and in [1] is visualised in Fig. 1. It
consists of a remanufacturing process, a manufacturing process, and two stocking points to keep remanufacturable and serviceable inventory. We assume that one product is returned per return occurrence, and
that the return process is Poisson with rate kR .
Each returned product rst enters remanufacturable inventory, where it is stored against inventory
holding costs of chr per product per time-unit. Products from remanufacturable inventory enter the remanufacturing process. The remanufacturing costs consists of a xed component cfr per batch, and a
variable component cvr per product. The remanufacturing lead-time is modelled by a discretely distributed 1
stochastic variable Lr .
Demand that cannot be fullled from serviceable inventory is either fullled by remanufacturing, or by
manufacturing. The cost structure for manufacturing is analogous to the cost structure for remanufacturing, i.e., it consists of a xed component cfm per batch, and a variable component cvm per product. The
manufacturing lead-time is modelled by a discretely distributed stochastic variable Lm .
Manufactured and remanufactured products enter serviceable inventory. The inventory holding costs in
serviceable inventory are chs per product per time-unit. Customer demands are either satised immediately
from serviceable inventory, or backordered against backordering costs of cb per product per time-unit. At

1
Our analysis only applies to a situation in which the number of possible lead-time realisations for manufacturing and
remanufacturing is nite. It follows that the analysis is restricted to discrete lead-time distributions.

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

199

each demand occurrence one product is demanded. The demand occurrences are assumed to follow a
Poisson process with rate kD .
The total expected operating costs under strategy : is dened as
8 h
cs  time average on-hand serviceable inventory
>
>
>
>
>
h
>
>
> cr  time average on-hand remanufacturable inventory
>
>
>
>
cv  time average number of remanufactured products
>
< r
1
C: cfr  time average number of remanufacturing batches
>
>
>
v
> cm  time average number of manufactured products
>
>
>
>
>
>
cfm  time average number of manufacturing batches
>
>
>
:
cb  time average backordering position:
In the presence of stochastic lead-times, the analysis of C: diers from the analysis with deterministic leadtimes regarding a number of cost components. The expressions of the cost components in [1] that remain
unchanged in case of stochastic lead-times are given in Table 1.
The procedures to calculate the other cost components are outlined in Section 2.1 (PUSH-strategy) and
in Section 2.2 (PULL-strategy). Issues concerning the search procedure to determine the optimal control
parameters for PUSH- and the PULL-strategy are briey discussed in Section 2.3. In the remainder of this
paper we follow the notation dened in Table 2. Additionally, we use for the lead-time probabilities the
following notation.
Let the manufacturing lead-time Lm be a discrete valued stochastic variable with nm possible outcomes,
2
nm
say `1
m ; `m ; . . . ; `m . Similarly, let the remanufacturing lead-time Lr be a discrete valued stochastic variable
2
nr
1
with nr possible outcomes, say `1
r ; `r ; . . . ; `r . Without loss of generality we assume that `m <
2
nm
1
2
nr
`m <


< `m , and `r < `r <


< `r . The probability mass function (p.m.f) of Lm and the p.m.f. of
Lr are then given by




j
j
j
pLm Pr Lm `m
; 1 6 j 6 nm ; and pLr Pr Lr `rj ; 1 6 j 6 nr :
Furthermore, dene

def 
[  1 2

2
nr
m
L `1 ; `2 ; . . . ; `n `1
`m ; `m ; . . . ; `n
r ; `r ; . . . ; `r
m
with `1 < `2 <


< `n . Then we have for 1 6 k 6 n;
k
P Lm

Pr Lm < `

maxfj X
j `m <`k g
i1

pLm ;

Table 1
Components of C: under the PUSH- and PULL-strategy
Component in C:

PUSH-strategy

PULL-strategy

Time-average
Time-average
Time-average
Time-average
Time-average
Time-average
Time-average

See Section 2.1

Qr 1=2
kR
kR =Qr
kD kR
kD kR =Qm
See Section 2.1

See Section 2.2

See Section 2.2
kR
See Section 2.2
kD kR
kD kR =Qm
See Section 2.2

on-hand serviceable inventory

on-hand remanufacturable inventory
number of remanufactured products
number of remanufacturing batches
number of manufactured products
number of manufacturing batches
backordering position

200

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

Table 2
Notation
Bt
B
IsOH t
OH
Is
IrOH t
OH
Ir
Wr t
Wm t
Isnet t
Is t
Dt0 ; t1
Enet t t0 ; t t1

Total number of products in backorder at time t

Time-average number of products in backorder
Number of products in on-hand serviceable inventory at time t
Time-average number of products in on-hand serviceable inventory
Total number of products in remanufacturable on-hand inventory at time t
Time-average number of products in on-hand remanufacturable inventory
Total number of products in remanufacturing work-in-process at time t
Total number of products in manufacturing work-in-process at time t
IsOH t Bt, the net serviceable inventory at time t
Isnet t Wr t Wm t, the serviceable inventory position at time t
Total demand in the time-interval t0 ; t1
Total number of products that are ordered for (re)manufacturing in the time-interval t t0 ; t t1 and will
enter serviceable inventory at or before time t, minus the total demand in the interval t t0 ; t t1
Total number of remanufacturing batches processed in the time-interval t0 ; t1
Time-average number of remanufacturing batches
Total number of manufacturing batches processed in the time-interval t0 ; t1

Or t0 ; t1
Or
Om t0 ; t1

k
P Lr

Pr Lr < `

maxfj X
j `r <`k g
i1

pLr :

We need the cumulative probabilities in Eqs. (2) and (3) for the numerical analysis of the PUSH- and
PULL-strategies.

2.1. Analysis for the (sm ,Qm ,Qr ) PUSH-strategy

This section outlines a numerical procedure to calculate the time-average on-hand serviceable inventory
OH
(I s ) and the time-average backordering position (B) for the PUSH-strategy. The procedure relies on the
calculation
of the distribution for the net serviceable inventory in the steady-state situation

). From this distribution, the time-average on-hand serviceable inventory and the time( Pr Isnet inet
s
average backordering position are derived as
OH

Is

 net

;
inet
inet
s Pr Is
s

 net

inet
inet
:
s Pr Is
s

inet
s >0

inet
s <0



, we use the relation that the net serviceable inventory at time t equals the net
To calculate Pr Isnet inet
s
serviceable inventory at time t `n plus the remanufacturing and manufacturing work-in process at time
t `n plus all products entering the manufacturing or remanufacturing processes in the interval t
`n ; t `1 that will be nished at or before time t minus the demand that occurred in the interval
t `n ; t, i.e.,
Isnet t Isnet t `n Wr t `n Wm t `n Enet t `n ; t `1 Dt `1 ; t
which is rewritten as

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

Isnet t Is t `n Enet t `n ; t `1 Dt `1 ; t:

201

Analogous to [1], (p. 11) (i.e., taking into account the correlations with the on-hand remanufacturable
inventory), we derive from Eq. (6) the steady-state distribution of the net serviceable inventory
r 1
X QX



Pr Is t `n is ; IrOH t `n
Pr Isnet inet

lim
s

t!1

X1 iOH
r 0

n
1
1
iOH
r ; Enet t ` ; t ` enet ; Dt ` ; t d

lim

r 1
X QX

t!1



Pr Enet t `n ; t `1 enet j Is t `n is ; IrOH t `n iOH
r

X1 iOH
r 0

kD `1 d
;
d!


where X1 is ; enet ; d j is enet d inet
. The steady-state probability
s


n
OH
n
OH
p1 is ; iOH
r lim Pr Is t ` is ; Ir t ` ir
1

kD `
 p1 is ; iOH
r exp

t!1

is numerically obtained from the continuous time Markov chain M1 with two-dimensional state space
S1 fsm 1; . . . ; 1g  f0; . . . ; Qr 1g (see [1], p. 11).

The procedure to evaluate the conditional probability limt!1 Pr Enet t `n ; t `1 enet j Is t `n
is ; IrOH t `n iOH
r g is outlined in Appendix A.
2.2. Analysis of the (sm ,Qm ,sr ,Sr ) PULL-strategy
To evaluate the time-average on-hand serviceable inventory and the time-average backordering position
the same procedure can be applied as for the PUSH-strategy, except that the Markov chain M1 should be
k
replaced by the Markov chain M2 described in [1], (p. 13), and the Markov chain M1 should be replaced
k
k
by the Markov chain M2 . The Markov chain M2 has state space
S02 fsm 1; . . . ; maxfsm Qm ; Sr gg  f0; . . . ; 1g  f1; . . . ; 1g;
k

k
k1
Is is , IrOH iOH
and state-vector X2 s is ; iOH
r ; enet , indicating that at time s 2 t ` ; t `
r , and
k
Enet enet . Expressions for the transition rates of M2 are given in Appendix B.
OH
The time-average on-hand remanufacturable inventory (I r ) is obtained from the relation
OH

Ir

maxsX
1
m Qm ;Sr X
is sm 1

OH
iOH
r p2 is ; ir ;

iOH
r 1

where p2 is ; iOH
r is the steady-state distribution, which is numerically obtained from M2 . Finally, the timeaverage number of remanufacturing batches (Or ) is written as
Or

1
X
iOH
r Sr sr

p2 sr 1; iOH
r kD

( PSr sm

iOH
r Sr sm Qm

0;

sr
X

p2 is ; Sr is 1kR

is sm 1

p2 sm 1; iOH
r kD ;

if sr P sm Qm ;
otherwise:

202

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

2.3. Search procedures to nd optimal control parameters

For the case with deterministic lead-times we reported in [1], (p. 14), that we neither succeeded for the
PUSH-strategy, nor for the PULL-strategy in identifying a `nice' structure in the cost function C:.
Identication of a `nice' structure in C: could be used to derive non-trivial lower bounds and upper
bounds on the control parameters. These bounds could be used to speed-up the search for optimal control
parameters, i.e., the set of control parameters for which the costs function (1) is minimized. Unfortunately,
the cost function in case of stochastic lead-times is even more complex than with deterministic lead-times.

Therefore, as in [1] we were committed to apply an enumerative search procedure to calculate C PUSH

minsm ;Qm ;Qr Csm ; Qm ; Qr and C PULL minsm ;Qm ;sr ;Sr Csm ; Qm ; sr ; Sr .
3. Numerical study
In a recent study, Song [5] investigated the eect of manufacturing lead-times on total expected costs in a
stochastic production/inventory system. The system diers from ours, in that no product returns occur
(kR 0) and, consequently, no remanufacturing operations exist. Furthermore, demands are modelled by a
compound Poisson process, which is a generalisation of the Poisson demand process considered in our
system. Finally, manufacturing costs consist of a linear component only, i.e., xed manufacturing costs are
zero (cfm 0).
Under these assumptions, a base-stock control policy is optimal (see [2], p. 604). Under base-stock
control, Song showed the following two results to hold:
an increase (decrease) in the duration of the manufacturing lead-time results in higher (lower) total expected costs,
an increase (decrease) in the variability of manufacturing lead-times results in higher (lower) total expected costs.
To express the variability of stochastic variables, we follow the denition used by Song (see [2], p. 609).
Denition 1. Consider two random variables X and Y with the same mean E(X) E(Y), having distributions
F and G with densities f and g. Suppose X and Y are either both continuous or both discrete. We say, X is
more variable than Y, denoted by X P var Y , if
Sign changef g 2; with sign sequence ; ;
That is, f crosses g exactly twice, rst from above and then from below.
Note that when kR 0 and cfm 0, the above results apply to our PUSH and PULL control strategies as
well, since both strategies are in this case equivalent to a base-stock control policy with base-stock level
sm 1.
Main objective in this numerical study is to investigate whether the above results also apply to our
PUSH and PULL strategies when kR > 0. For this purpose we have subdivided our empirical study in two
parts. In Section 3.1 we analyse the inuence of the manufacturing and remanufacturing lead-time duration
on total expected costs, whereas in Section 3.2 we investigate the inuence of variability in the manufacturing and remanufacturing lead-times on total expected costs. The experiments in both sections start out
from a base-case scenario dened in Table 3. The inuence of alternative assumptions regarding lead-times
and other input parameters are briey discussed in Section 3.3.

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

203

Table 3
Base-case scenario
Problem component

(Numerical) value

Unit of measure

Fixed remanufacturing costs (cfr )

Variable remanufacturing costs (cvr )
Fixed manufacturing costs (cfm )
Variable manufacturing costs (cvm )
Remanufacturable inventory costs (chr )
Serviceable inventory costs (chs )
Backordering costs (cb )

)
)
)
)
0.5
1.0
50.0

$/remanufacturing set-up
$/product remanufactured
$/manufacturing set-up
$/product manufactured
$/product per time-unit
$/product per time-unit
$/product per time-unit

Expected remanufacturing lead-time (ELr )

Distribution of remanufacturing lead-time
Expected manufacturing lead-time (ELm )
Distribution of remanufacturing lead-time

2
Deterministic
2
Deterministic

Time-units

Demand process
Demand intensity (kD )
Return process
Return intensity (kR )

Poisson
1.0
Poisson
0.7

Time-units

Product per time-unit

Product per time-unit

3.1. The eects of lead-time duration

To analyse the inuence of the lead-time duration on total expected costs, we have conducted a set of
computational experiments in which either Lm or Lr is varied, while keeping the other lead-time variable
xed. The parameter settings in our experiments are summarized in Table 4.
The computational experiments led to the following observations.
Observation 1. An increase in manufacturing lead-times results for both strategies in much larger cost
increases than equivalent increases in remanufacturing lead-times (see Figs. 3 and 4). The reason why this
eect occurs is, that our PUSH- and PULL-strategies give priority to remanufacturing and use
manufacturing as a last resort to avoid backorderings. Therefore, the serviceable inventory position level
at which manufacturing orders are placed is lower than or equal to the serviceable inventory position level
at which remanufacturing orders are placed. Consequently, when Lr  Lm , the probability that during an
outstanding manufacturing order a backordering position occurs is larger than the probability that during
an outstanding remanufacturing order a backordering position occurs. This causes that longer manufacturing lead-times require higher safety stocks to protect against costly stock-out events than equivalently
longer remanufacturing lead-times. As a consequence, it can be expected that the dierence between the
eects of increasing the manufacturing lead-time and the eects of increasing the remanufacturing lead-time
Table 4
Summary of computational experiments
Experiment

Control strategy

Lead-times

Reference

1
2

PUSH
PUSH

Lm has been varied, while keeping Lr 0

Lr has been varied, while keeping Lm 0

see Fig. 3(a)

see Fig. 3(b)

3
4

PULL
PULL

Lm has been varied, while keeping Lr 0

Lr has been varied, while keeping Lm 0

see Fig. 4(a)

see Fig. 4(b)

5
6

PUSH
PUSH

Lm has been varied, while xing Lr 0, or Lr 2, or Lr 4

Lr has been varied, while xing Lm 0 or Lm 2 or Lm 4

see Fig. 5(a)

see Fig. 5(b)

204

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

Fig. 3. C PUSH as function of (a) the manufacturing lead-time with Lr 0, and (b) the remanufacturing lead-time with Lm 0:

should increase with increasing backordering costs. Additional experiments indeed conrm this
expectation.
Observation 2. Increases in remanufacturing lead-times sometimes result in cost decreases. This eect is
visualised in Fig. 5. Note that the cost decreasing eect of an increase in remanufacturing lead-times occurs
in particular when Lr is small compared to Lm . The arguments on why this eect may occur are related to
the arguments used in Observation 1. Since our PUSH- and PULL-strategies give priority to
remanufacturing, the serviceable inventory position level at which a remanufacturing order is placed is
higher than or equal to the level at which a manufacturing order is placed. Suppose that the
remanufacturing lead-time is much shorter than the manufacturing lead-time, then:

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

205

Fig. 4. C PULL as function of (a) the manufacturing lead-time with Lr 0; and (b) the remanufacturing lead-time with Lm 0:

1. We may set the manufacturing reorder point sm at a relatively low level to allow that the remanufacturing orders are placed when the inventory position is low to minimize holding costs. However, this will
result in high backorder costs due to the long manufacturing lead-time.
2. We may set the manufacturing reorder point sm at a relatively high level to protect against the high manufacturing lead-time. However, this will result in high holding costs since remanufacturing orders are
delivered long before they are really needed to satisfy demands.
The above indicates that our PUSH- and PULL-strategies are not very ecient if the remanufacturing
lead-time is much smaller than the manufacturing lead-time. If the remanufacturing lead-time would be
increased to a value closer to the manufacturing lead-time, the strategies would be less inecient and result
in lower optimal costs. Note that this ineciency only occurs if Lr < Lm , so intuitively it is expected that an

206

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

Fig. 5. C PUSH as function of (a) the manufacturing lead-time with Lr 0 (curve 1), Lr 2 (curve 2), and Lr 4 (curve 3), and (b) the
remanufacturing lead-time with Lm 0 (curve 1) Lm 2 (curve 2), and Lm 4 curve 3).

increase in manufacturing lead-times will always result in cost increases. To make this more clear suppose
that Lm 6 Lr . In this case the policies are relatively ecient, so according to intuition we expect that an
increase in the manufacturing lead-time results in a costs increase. If Lm > Lr the policies are relatively
inecient and an increase in manufacturing lead-times will only increase this ineciency, resulting in increasing costs. Therefore, an increase in manufacturing lead-times will always result in cost increases. The
gures conrm this.
It should be noted that in the above we did not consider holding costs associated with work in process in
the remanufacturing facility. If remanufacturing lead-times increase also the amount of work in process
holding costs increases. This cost increasing eect could partly neutralize the cost decreasing eect just
described. However, it is interesting to observe that an increase in remanufacturing lead-times may have a
smaller impact then expected.

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

207

3.2. The eects of lead-time variability

To investigate the eects of lead-time variability, we assume in the initial set of experiments Bernoulli
distributed lead-times. 2 Associated to the Bernoulli distribution, we dene the parameters ` and P` , reecting the situation that either an `-period lead-time will occur (with probability P` ), or an instantaneous
delivery will occur (with probability 1 P` ). In the computational experiments, the parameters ` and P`
have been determined, such that the average lead-time E(L) is kept constant at the base-case level of 2 time
units, whereas the variance r2L has been varied. Note that for Bernoulli distributed random variables, a
higher variance is equivalent to a higher variability according to Denition 1.
Fig. 6 shows the relation between lead-time variability (expressed in terms of r2L ) and total expected
costs, for the PUSH-strategy (Fig. 6(a)) and for the PULL-strategy (Fig. 6(b)). The curves labelled with (1)
correspond to Bernoulli distributed manufacturing lead-times and deterministic remanufacturing leadtimes, whereas the curves indicated by label (2) correspond to Bernoulli distributed remanufacturing leadtimes and deterministic manufacturing lead-times. From Fig. 6 the following observations can be made.
Observation 3. An increase in the variability of manufacturing lead-times results for the base-case in lower
total expected costs, both under the PUSH-strategy and under the PULL-strategy (curves (1) are
decreasing). In this context, it should be noted that Song proved that this eect will never occur with
kR 0. To explain why the eect still may occur with kR > 0, note that with Bernoulli distributed
manufacturing lead-times a more variable lead-time (given a xed average lead-time) implies, (i) an increase
in `, and (ii) a decrease in P` and, consequently, an increase in the probability of a zero lead-time
occurrence. Apparently, if kR > 0 and remanufacturing lead-times are deterministic, the cost increasing
eect of (i) is partly neutralised by a frequent arrival of remanufacturing orderings, compensating for
manufacturing orderings with a long `-period lead-time. Consequently, the impact of the cost increasing
eect of (i) is dominated by the cost decreasing eects of (ii), resulting in the observed net cost decrease. As
a consequence of the arguments above and as a consequence of Song, the magnitude of the eect will not
only depend on r2L , but also on the return intensity. If the return intensity is low the cost increasing eect of
(i) is more dominant, since not all long (`-period) manufacturing lead-times are automatically compensated
by the arrival of remanufacturing orderings. On the other hand, if the return intensity is high, there will
only be a small number of manufacturing orders and the manufacturing lead-time will not have much
inuence (see Section 3.3).
Since the cost decreasing eect of manufacturing lead-times only occurs for kR > 0, one may intuitively
conclude that this eect will only occur in systems where multiple sources exist to satisfy demands. However
in [6] it is shown that a similar eect occurs in the Newsboy Problem, which relates to a single source
system.
Observation 4. An increase in the variability of remanufacturing lead-times results in an increase in total
expected costs, both under a PUSH-strategy and under a PULL-strategy (curves (2) are increasing).
Although we cannot analytically prove the observed eect, we were unable to nd datasets for which the
opposite eect occurs.
The experimental results also indicate that lead-time variability has not much inuence on the relative
performance of PUSH and PULL-strategy. Analogous to the situation with deterministic lead-times (see

2
We have chosen Bernoulli distributed lead-times, since the Bernoulli distribution is one of the few commonly used discrete
probability distributions of which the parameters provide sucient freedom to allow for a sensitivity analysis on lead-time variability
(variance) at a constant average lead-time E(L).

208

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

Fig. 6. (a) C PUSH and (b) C PULL as functions of the variance in manufacturing lead-times (curves 1), and as functions of the variance in
remanufacturing lead-times (curves 2).

Figs. 3 and 4, and [1]) the PULL-strategy outperforms the PUSH-strategy in the base-case scenario, where
chr < chs .
3.3. Sensitivity analysis
To investigate the extend to which the above observations are specic to the assumptions and parameter
settings in the base-case scenario, we also study the sensitivity of the total expected costs for changes in
demand and return intensity, cost structure, and lead-time distributions. Here we briey discuss the outcomes of this study.

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

209

Demands and returns: As can be expected, the results of the sensitivity analysis indicate that when kR is
small compared to kD , the inuence of changes in the duration and variability of remanufacturing lead-times
on costs are limited, since most demands are satised by manufacturing rather than by remanufacturing.
Also, if kR  kD , costs are rather insensitive to changes in the variability and duration of manufacturing
lead- times, since most of the demands are satised by remanufacturing. Generally, the magnitude of the
cost decreasing eect related to an increase in the duration of remanufacturing lead-times and the magnitude of the cost decreasing eect related to an increase in the variability of manufacturing lead-times both
increase when increased coordination is required between the remanufacturing and the manufacturing
processes. The coordination needs to be maximal if the time-average number of remanufactured products is
close to the time-average number of manufactured products, i.e., when kR  kD kR .
Cost structure and cost parameters: The occurrence and magnitude of the observed eects depend very
much on the cost structure and cost parameters. For instance, work-in-process costs have been disregarded
in our study. Taking into account these costs would imply that an additional cost component (which is
usually linear in the average lead-time duration), must be added to the cost function. Depending on the
numerical value of the work-in-process costs, the cost decreasing eect of an increase in remanufacturing
lead-times may be reduced, or even completely disappear. It should however be noted that the observations
related to lead-time variability will not be aected by the inclusion of work-in-process costs.
Non-zero xed costs may also inuence the total expected operating costs, since non-zero xed manufacturing (remanufacturing) costs imply larger manufacturing (remanufacturing) batches. We observed
that a larger manufacturing or remanufacturing batch will reduce the cost decreasing eect of an increase in
the variability of manufacturing lead-times (Fig. 7). The reason why this eect occurs is, that with larger
remanufacturing batches the inter-arrival time between subsequent remanufacturing batches will become
more erratic. Consequently, long manufacturing lead-times may not be compensated immediately by the
arrival of remanufacturing orderings.
Finally, analogous to the situation with deterministic lead-times the experiments for the case with
stochastic lead-times indicate that when remanufacturable inventory is valued lower than serviceable inventory (chr < chs ), the PULL-strategy outperforms in most cases the PUSH-strategy. Otherwise, the PUSHstrategy yields in most cases lower total expected costs. This is consistent with the results of Inderfurth [2]
who proved for a similar but periodic system, that a PULL-strategy is optimal if chr < chs .
Lead-time distributions: The results of a simulation study by Van der Kruk [7] indicate that Observations
34 are not specic to Bernoulli distributed lead-times only. The eects related to lead-time variability also
occur for continuously distributed lead-times, such as for Lognormal distributed lead-times (Fig. 8(a)), and
for Coxian-2 distributed lead-times (Fig. 8(b)). Both for Lognormal and Coxian-2 distributed random
variables a larger variance (with constant expectation) is equivalent to a higher variability according to
Denition 1.
4. Conclusions
We have investigated the inuence of lead-time duration and lead-time variability on total expected costs
in a system with manufacturing and remanufacturing operations. To control the system, we have applied
simple PUSH- and PULL-strategies. Although these strategies are non-optimal, they have the advantage
that they are easy to implement and actually used in practice. Also, we have introduced a novel and nontrivial procedure to include stochastic lead- times in our analysis.
A numerical study shows that changes in the duration of manufacturing lead-times have larger inuences
on total expected costs than changes in the duration of remanufacturing lead-times. In some cases, an increase in remanufacturing lead-times may even result in cost decreases. In this context, it has been stressed
that one should strive for a situation in which the remanufacturing lead-time and the manufacturing

210

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

Fig. 7. (a) C PUSH and (b) C PULL as functions of the variance in manufacturing lead-times for cfm cfr 0 (curves 1), cfm cfr 5 (curves
2), and cfm cfm 10 (curves 3).

lead-time do not dier too much in expectation. Also, the numerical study indicates that cost increases are
more sensitive to a larger variability in remanufacturing lead-times than to a larger variability in manufacturing lead-times. In some cases, a higher variability in manufacturing lead-times may even result in cost
decreases.
The experiments also indicate that with stochastic lead-times the PULL-strategy in most cases outperforms the PUSH-strategy when chr is smaller then chs . However, further experiments also indicate that when
remanufacturable inventory is valued as high as serviceable inventory, the PUSH-strategy may outperform
the PULL-strategy. These observations correspond to the observations made in [1] in relation to deterministic lead-times.

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

211

Fig. 8. Costs as function of the variance in manufacturing lead-times (curves 1), and remanufacturing lead-times (curves 2), while leadtimes follow (a) a Lognormal distribution and (b) a Coxian distribution.

Acknowledgements
The authors thank the anonymous referee for the many helpful comments that she/he made on an earlier
version of this paper.

212

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

Appendix A
The conditional probability limt!1 PrfEnet t `n ; t `1 enet jIs t `n is ; IrOH t `n iOH
r g
that appears in Section 2.1, is calculated as
Q
1
r 1
X
X

qi ;iOH ;e
s r

is sm 1 iOH 0
r

net

j is ;iOH
r ;0

`n `1 ;

A:1

where qs1 j s0 `n `1 is the conditional probability that during the interval t `n ; t `1 the initial
system state changes from the state s0 into state s1 . Here, the state-vectors s0 and s1 are elements of the
state-space S01 fsm 1; . . . ; 1g  f0; . . . ; Qr 1g  f1; . . . ; 1g.
Unfortunately, probabilities qs1 j s0 `n `1 cannot be obtained directly from a Markov chain model,
because the transitions of Enet t `n ; t `1 are time-dependent: a change in Enet t `n ; t `1 due to a
(re)manufacturing order depends on the probability that that particular order will arrive before time t, and
this probability depends on the dierence between the time of ordering and time t. However, the transitions
every
in each sub-interval t `k ; t `k1 for k 2; . . . ; n are time-independent. For instance,

manufacturing order that is ordered in the interval t `k; t `k1 has probability Pr Lm < `k that the
order will arrive before time t, and a probability 1 Pr Lm < `k that the order will arrive after time t (see
Fig. 9).
k
Thus, we dene for each sub-interval t `k ; t `k1 , 1 < k  n, a Markov chain M1 with state
k
k
0
OH
OH
space S1 and state variable X1 s Is ; Ir ; Enet u; u s. By denition, X1 s is ; ir ; enet whenever
k
`k1 .
Is s is , and IrOH s iOH
r , and Enet u; s enet , with s 2 u; u `
k
Using expressions (2) and (3) we arrive at the following time independent transition rates of M1 :
k

iOH
< Qr 1;
r

mis ;iOH ;enet ;is ;iOH 1;enet kR ;

r

mis ;iOH ;enet ;is Qr ;0;enet Qr PLr kR ;

r

k
mis ;iOH ;enet ;is Qr ;0;enet
r

9
=

k
PLr kR ; ;

iOH
Qr 1;
r

mis ;iOH ;enet ;is 1;iOH ;enet 1 kD ;

r

9
k
k
mis ;iOH ;enet ;sm Qm ;iOH ;enet Qm 1 PLm kD ; >
=
r
r
m

OH
is ;iOH
r ;enet ;sm Qm ;ir ;enet1

k
1 PLm kD ; >
;

is > sm1 ;

is sm1 :

Fig. 9. An
in the interval t `k ; t `k1 with stochastic lead-time L will arrive before or at time t with probability
 order placed

k
k
PL Pr L < `k , and will arrive after time t with probability 1 PL .

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

213

Related to this Markov chain we dene the conditional probability qs2 j s0 `k `k1 , which is the
k
probability that during the interval t `k ; t `k1 the initial system state of X1 changes from state s0
1
n
1
into state s . Then, the probabilities qs1 j s0 ` ` can be expressed in terms of the recursive relation
(k 2; . . . ; n)
X
k
qs1 j s `k1 `1 qs j s0 `k `k1 :
A:2
qs1 j s0 `k `1
s2S01

To calculate qs2 j s0 `k `k1 we evaluate the transient behaviour of the Markov chain M1 . To do so
k
we use the discretization technique, which is discussed in more detail in [8]. This technique transforms M1
k
k
1;k
into an equivalent discrete time Markov chain M1 . The one-step transition probabilities qs1 j s0 in M1
are as follows:
8 k
< ms0 ;s1 =mk ;
s0 6 s1 ;
1;k


A:3
qs1 j s0
: 1 mk
=mk ;
s0 s1 ;
s0
P
k
k
k
where ms0 s1 2S0 ms0 ;s1 , and mk maxs0 2S01 fms0 g.
1
m;k
The m-step transition probabilities qs1 j s0 are derived from the recursive relation
X m1;k 1;k
m;k
qs1 j s0
qs2 js0 qs1 j s2 :
A:4
s2 2S01

Finally, the conditional probabilities qs1 j s0 `k `k1 are calculated as

k

qs1 j s0 `k `k1

1
X

expm

k `k `k1

m0

mk `k `k1 m;k
qs1 j s0 :
m!

A:5

Relations (7), (A.1)(A.5) now enable to calculate the distribution of net inventory in Eq. (7) and consequently the average on-hand inventory in Eq. (4) and the average backorder position in Eq. (5).
Appendix B
k

The transition rates of the Markov chain M2 are as follows:

k

is > sr or iOH
< Sr is 1;
r

mis ;iOH ;enet ;is ;iOH 1;enet kR ;

r

mis ;iOH ;enet ;Sr ;0;enet Sr is PLr kR ;

k

mis ;iOH ;enet ;Sr ;0;enet 1 PLr kR ;

is 6 sr and iOH
Sr is 1;
r

mis ;iOH ;enet ;i

r

OH
s-1 ;ir ;enet 1

is > sr 1 or fis > sm1 and iOH

6 Sr is g;
r

kD ;
k

mis ;iOH ;enet ;Sr ;iOH x;enet x1 PLr kD ;

k

)
is sr 1 and iOH
P x;
r

mis ;iOH ;enet ;Sr ;iOH x;enet 1 1 PLr kD ;

k

mis ;iOH ;enet ;sm Qm ;iOH ;enet Qm 1 PLm kD ;

k

mis ;iOH ;enet ;sm Qm ;iOH ;enet 1 PLm kD ;

)
is sm 1 and sr P sm Qm and iOH
P y;
r

214

E. van der Laan et al. / European Journal of Operational Research 115 (1999) 195214

9
k
mis ;iOH ;enet ;Sr ;iOH y;enet Sr sm 1 P11 kD ; >
>
>
r
r
>
>
k
P12 kD =
m OH
OH
is ;ir
k

;enet ;Sr ;ir

y;enet Qm 1

mis ;iOH ;enet ;Sr ;iOH y;enet y1 P21 kD ;

k

mis ;iOH ;enet ;Sr ;iOH y;enet 1 P22 kD ;

>
>
>
>
>
;

is sm 1 and iOH
< Sr sm and fiOH
< y or sr < sm Qm g;
r
r

where x Sr sr , y Sr sm Qm , P11 PLm PLr , P12 PLm 1 PLr , P21 1 PLm PLr , and
k
k
P22 1 PLm 1 PLr .
References
[1] M. Fleischmann, J.M. Bloemhof-Ruwaard, R. Dekker, E. van der Laan, J.A.E.E. van Nunen, L.N. Van Wassenhove, Quantitative
models for reverse logistics: A review, European Journal of Operational Research 103 (1997) 117.
[2] E. van der Kruk, Inventory control models with remanufacturing: A simulation study, Unpublished Masters Thesis, Erasmus
University Rotterdam, The Netherlands, 1995.
[3] E.A. van der Laan, M. Salomon, R. Dekker, Production planning and inventory control for remanufacturable durable products,
Working Paper 9531/A, Econometric Institute, Erasmus University Rotterdam, The Netherlands, 1995.
[4] J.A. Muckstadt, M.H. Isaac, An analysis of single item inventory systems with returns, Naval Research Logistics Quarterly 28
(1981) 237254.
[5] K. Inderfurth, Simple optimal replenishment and disposal policies for a product recovery system with leadtimes, OR Spektrum 19
(1997) 111122.
[6] A. Ridder, E.A. van der Laan, M. Salomon, How larger demand variability may lead to lower costs in the newsboy problem,
ERASM Management Report Series 265, Erasmus University Rotterdam, The Netherlands, to appear in Operations Research,
1996.
[7] J.-S. Song, The eect of lead-time uncertainty in a simple stochastic inventory model, Management Science 40 (5) (1994) 603613.
[8] H.C. Tijms, Stochastic Models: An algorithmic Approach, Wiley, Chichester, UK, 1994.