Safety stock allocation in a multi item multi-echelon

inventory system with quantity-based order

consolidation
G. P. Kiesm
uller , A. G. de Kok
Faculty of Technology Management,
Technische Universiteit Eindhoven,
P.O. Box 513,
5600 MB Eindhoven,
The Netherlands
Abstract
The efficient management of a distribution systems requires coordination between transportation planning and inventory control. Small and frequent shipments
can reduce inventory levels while maintaining high customer service levels, but they
increase unit transportation costs due to inappropriate utilization of the vehicles.
On the other hand shipment consolidation policies reduce unit shippping cost because of economies of scale but they require higher inventory levels to maintain
high service levels. Hence, it is of considerable interest to understand the trade-off
between shipment consolidation policies and inventory levels.
In this paper we provide an algorithm which enables us to allocate the safety
stock in the distribution network under given parameters of the shipment policy. We
consider a quantity based shipment consolidation policy which means all orders for a
particular destination are held and shipped when a predetermined weight or volume
is reached. The studied system is a divergent multi item multi-echelon network
with stochastic demand at the lowest echelon. The stockpoints are controlled by
continuous review (s, Q) installation stock policies where the replenishment leadtime
is dependent on the shipment consolidation policy. We derive approximations for
the leadtimes, the reorder levels, the inventory levels and the service levels which
are the base for our algorithm. In a numerical study we test our approximations
using computer simulation and we illustrate that the use of our algorithm leads to
estimated service and inventory levels closed to the actual ones.

keywords: stochastic, inventory, multi-echelon, multi item, shipment consolidation

1

Introduction

The management of multi-echelon inventory systems is a crucial part of supply chain

management. In order to reduce costs and to improve efficiency an integrated approach
is needed. For instance, decreasing inventory levels and thereby lowering inventory costs
induces more frequent shipments of smaller sizes, which in general increases transportation
unit costs. We believe that careful coordination of shipment and replenishment policies
can lead to substantial cost savings.
In this paper we propose an algorithm to allocate safety stocks in a multi item multiechelon distribution network where several deliveries are dispatched as a single load, realizing economies of scale in transportation costs. We consider a quantity-based shipment
policy which means that orders for a destination are held and shipped when a minimum
consolidated amount is reached (Higginson and Bookbinder (1994)). The required amount
to be consolidated has an influence on the truck departure time and therefore, on the replenishment leadtime of the supplied stockpoints. We model this interaction considering
the replenishment leadtime as an endogenous variable. Moreover, we explicitly model the
waiting time due to shipment consolidation and the waiting time due to a lack of stock
at the higher echelon.
We consider a divergent network of warehouses keeping stock of different items. Inventories are controlled using an (s, Q) installation stock policy, which means a replenishment
order is placed at the moment where the local inventory position, which is defined as
the physical stock plus the stock on order minus backorders, is equal or smaller than the
reorder level s. Customer demand is stochastic, only observed at the lowest echelon, and
modeled as a compound renewal process which means that inter-arrival time of orders as
well as demand sizes are modeled as random variables. Demand which cannot be satisfied is backordered. We assume the batchsizes Q to be given and identical for the same
stockpoints at an echelon.
Target service levels, in this case the fillrate, must be met and the problem arises to
determine the reorder levels, respectively the safety stock needed. Our proposed algorithm
determines successively in a first step the aggregate demand processes at all upper echelons
starting at the lowest echelon, and in a second step the waiting times and the leadtimes are
determined successively, starting at the highest echelon, which enables the calculation of
2

the reorder levels. Moreover, the average inventory levels are calculated. The algorithm is
based on approximate expressions which we derive using fitting techniques (Tijms (1994)),
a superposition technique proposed by Whitt (Whitt (1982)), and approximations based
on renewal theory (Cox (1962)).
This paper is an extension of Kiesm
uller et al. (2004), where a multi-echelon distribution
network is studied for a single item and each replenishment order induces an immediate
shipment which is the general assumption in most of the literature on single item multiechelon inventory models (see for example Axsater (1993)). Moreover, the shipment policy
is not explicitly modeled. Since the amount of literature in the field of single item multiechelon inventory systems is enormous we only refer to surveys as given in Axsater (2003),
Federgruen (1993), Axsater (1993), Diks et al. (1996) and Houtum et al. (1996).
Multi item multi-echelon stochastic inventory systems are mainly studied for spare parts.
In this field a series of papers, as for example Muckstadt (1973), followed upon the famous
METRIC model of Sherbrooke (1968). But in contrast to our problem, consolidation of
orders is not an issue in spare parts management. The studied problem is the allocation
of a given budget over expensive spare parts such that service levels or system availability
are maximized. A similar problem is studied in Masters (1993) who considers a multiechelon retail inventory/distribution system for multiple items with high value and low
level of stochastic demand and a one-for-one resupply logic and minimizes lost sales and
inventory costs subject to a budget limit on total inventory investment.
There is also a large body of literature available where inventory and transportation
problems are solved in an integrated way using deterministic modeling of logistic networks.
We can distinguish between two main modeling approaches. The first approach has its
bases in the EOQ model (see for example Burns et al. (1985) or C
etinkaya and Lee
(2002)) while the other approach is using mathematical programming formulations of the
problem (see for example Popken (1994)). We also refer to an excellent overview in this
field by Bertazzi and Speranza (2002). However, all these deterministic models cannot
cover with the effects of stochastic demands and consolidation that arise in the problem
setting studied in this paper.
An integrated approach for the coordination of inventory and transportation decisions in
VMI systems can be found in C
etinkaya and Lee (2000). A renewal theoretic model is

developed for the case of Poisson demand to compute the optimal replenishment quantity
and dispatch frequency simultaneously. In contrast to our paper, the authors consider
a time-base policy, where an accumulated load is shipped every T periods. Axsater
(2001) provides a note on the model of C
etinkaya and Lee (2000) and he suggests an
approximation and an adjustment that can be used to improve the proposed heuristic of
C
etinkaya and Lee.
Related to our study is also the work done by Cachon (2001). He balances transportation,
inventory and shelf space costs for a single link between one retail store and one warehouse.
Demand is modeled as a Poisson process and the replenishment orders of the retailers are
generated for each product by a one-for-one replenishment policy where the order-upto level is equal to the shelf space. Besides a time based consolidation policy he also
investigates a quantity based policy. He proposes among other things a heuristic to
compute the order-up to levels and the amount to be consolidated before the truck can
depart.
The model under study in this paper is an extension of the above mentioned model since
we assume a compound renewal model for the demand process, which comprehends the
Poisson Process as a special case. We additionally consider (s, Q) replenishment policies
which take into account that order sizes are often restricted to material handling units,
such as pallets and boxes. Moreover, our presented algorithm can also by applied in
multi echelon networks with more than 2 echelons, several warehouses and non-identical
customers.
The sequel of the paper is organized as follows. In section 2 a detailed model and problem
description is given. In section 3 we explain the used approximation techniques and in
section 4 we derive approximated expressions for the reorder level, the average inventory
level, the first two moments for the waiting time due to a lack of stock and the waiting
time due to shipment consolidation. The algorithm to allocate the safety stock is proposed
in section 5. In a numerical study we test the performance of the approximations and the
performance of the distribution network when the reorder levels are determined by our
computation procedure. The paper ends with a short summary and conclusions.

Model and problem description

2.1

The distribution network

In this paper we study a divergent multi-echelon distribution network consisting of several

warehouses and external suppliers. At the warehouses are stockpoints for different items
which are uniquely numbered with arabic numbers, while the warehouses are numbered
with roman numbers. The set of warehouses is denoted by W, the set of all stockpoints
by M, and the set of stockpoints located at warehouse m W is denoted by Lm . We
assume that each stockpoint is supplied exactly by one stockpoint and without loss of
generality a warehouse is supplied by exactly one warehouse or an external supplier.
Let k be an arbitrary stockpoint (k M). Then Sk is defined as the set of all immediate
successors of k, where j M is a successor of k if and only if j and k represent the same
item and replenishment orders for j are part of the demand process of k. Further, Pk is
the set of immediate predecessors of k M, i.e. replenishment orders of k are part of the
demand process of j Pk . Note that |Pk | = 1 or for all k M due to the divergent
network structure.
We denote N as the number of echelons in the network and En as the set of stockpoints
of the n-th echelon.
E1 := {k M|Sk }
n1
S

En := {k M|Sk (

n2
S

Ei )}\{k M|Sk (

i=1

where

0
S

Ei )}

n = 2, ..., N

i=1

Ei = . Further, we assume that customers are replenished by k E1 , a

i=1

stockpoint j En1 by k En , (n = 2, ..., N ) and k EN by an external supplier.

In Figure 1 we illustrate a three echelon distribution network with four warehouses and
four different items.

Transportation rates are often not only dependent on the origin-destination distance but
also on the quantity shipped. In many situations increased per-shipment volume results
in economies of scale and lower per unit transportation cost. Therefore, orders from
the same warehouse are consolidated and shipped together. In this paper we assume

Customer
Demand

10

Customer
Demand
Customer
Demand

11

12
1

5
III

External
Supplier

8
I

Customer
Demand

13

Customer
Demand

14

Customer
Demand

15

Customer
Demand

II

16
IV

Customer
Demand

Figure 1: A typical distribution network

a quantity-based consolidation policy (Higginson and Bookbinder (1994)) which means
orders are dispatched until a target quantity Qcm is reached. Further, transportation times,
defined as the time between the truck departure at a warehouse and the moment that the
items are available on stock at the supplied warehouse, are denoted by Ldk . We model the
transportation time as a random variable and assume the first two moments to be known.
The stockpoints in the warehouses are continuously reviewed and replenishment orders
are triggered according to an (sk , Qk )-installation stock policy. This means that as soon
as the local inventory position, which is expressed as the physical inventory plus the stock
on order minus backorders, is equal to sk or drops below sk an amount Qk is ordered such
that the inventory position is raised to a value between sk and sk + Qk . We call sk the
reorder level and Qk the batch size of stockpoint k M.
In order to keep material handling costs low managers of warehouses try to use standardized mechanical material handling equipment. Moreover, a standardization of the
replenishment lotsize can be observed. Therefore, we assume in this paper that all
n , k En ,) and
stockpoints at one echelon use the same size batch sizes (Qk = Q
that the batch size of a stockpoint k is a multiple of the batch size of its successor
n , k Pj , j En ). As a consequence we also restrict the values of Qc to
(Qk = m Q
m
multiples of Qj , (j Sk , k Lm ) as well as the reorder levels sk .
Customer demand is only observed at the most downstream stockpoints at the first echelon
(k E1 ) and can be modeled as a compound renewal process. Let Ak , k E1 , denote the
6

length of an arbitrary inter-arrival time and Dk , k E1 , the demand size of an arbitrary

customer at the stockpoint. We assume the first two moments of these random variables
to be known. Further, we assume that all unsatisfied demand is backordered.
Many inventory managers work with target service levels in order to be able to satisfy
customer demand. A reasonable service level for measuring the availability of stock is the
fillrate, defined as the fraction of demand delivered directly from shelf. We assume that
for all stockpoints k En , (n = 1, 2, . . . , N ) a target fillrate ktarget is given.
Summarizing, the following random variables are exogenous to the system under consideration and we assume the first two moments to be known.
Dk
Ak
Ldk

Demand size for k E1

Time between two subsequent arrivals of orders k E1
Transportation time to k M

Additionally we assume that the following parameters are given.

Qk
ktarget
Qcm

2.2

Batchsize at k M
Target fill rate at k M
Predetermined consolidation quantity towards m W

The problem description

Since forecasted demand in general differs from actual demand safety stock is hold in
order to be able to satisfy customer demand. One of the major aims of inventory management is to balance service requirements with the costs related to inventory availability.
Therefore, it is important to know how to allocate safety stocks among different stockpoints at minimal costs, respectively which reorder levels sk the stockpoints should use.
In this paper we provide an algorithm to compute these reorder levels in a multi-echelon
multi-item distribution network under a quantity based consolidation policy such that
prespecified service requirements can be met.
It is well known that the leadtime Lk (k M), defined as the time between the placement
and the arrival of an order at stockpoint k, has an influence on the required safety stock
at stockpoint k. However, the leadtime of a stockpoint in a multi-echelon network is
dependent on the stock availability of its predecessor. If there is not enough stock at Pk
to fulfill the entire order, then the order has to wait until Pk is replenished and afterwards
the order becomes available for consolidation. If more than one order is waiting due to a
7

lack of stock a FIFO (first in, first out) rule is applied. The time between the arrival of
the replenishment order at the stockpoint Pk and the time where the order is available
for consolidation is called the waiting time due to a lack of stock and is denoted with Wks .
When an order is available for consolidation it has to wait until the target quantity Qcm is
dispatched and the truck departs. We call this part Wkc of the leadtime the waiting time
due to shipment consolidation. In our algorithm these components of the leadtime are
explicitly considered using relation (1) which enables us to study the relationship between
safety stocks and the shipment consolidation policy.
Lk = Ldk + Wks + Wkc

(1)

We assume that the waiting times for replenishment orders at the external suppliers are
equal to zero (Wks = 0, Wkc = 0, k EN ).
Besides the reorder levels sk our algorithm also computes the average inventory levels
E[Ik ] at the stockpoints and the actual service levels k as performance measures of the
distribution network.
sk
Reorder level of stockpoint k M
E[Ik ] Long run expected physical inventory level at k M
k
Actual fill rate at stockpoint k
Below we define the used functions and operators.
E[Y ]
2 (Y )
cY

Expectation of the random variable Y

Variance of the random variable Y
)
)
Coefficient of variation of Y (cY := (Y
E[Y ]

(a)+
P (A)

Maximum of the real number a and zero

Probability of event A

2.3

The processes

In order to be able to compute the reorder levels, the service levels and the average
inventory levels for all stockpoints in the distribution network we have to consider several
processes and we need some variables to describe these processes.
The demand processes at higher echelons are also modeled as compound renewal processes
with random inter-arrival times of orders Ak and fixed demand sizes Dk equal to the
batchsizes of the successors.
8

m
(R k, Qk)

Pk

(Ck, Qk )

(C m * ,Q k)

Consolidation dock

Figure 2: The replenishment processes

In the following we will distinguish between two types of arrival processes (see Figure 2):
the arrival process (Rk , Qk ) of replenishment orders from stockpoint k at stockpoint Pk
and the arrival process of replenishment orders from stockpoint k at the consolidation
dock. We define Ck as the stationary inter-arrival time of the latter process. Both processes are equivalent when there are enough items on stock. Otherwise the replenishment
orders at the consolidation dock are delayed and also clustering of the orders is possible,
which means that more than one replenishment order of the same item can arrive simultaneously at the consolidation dock. The inter-arrival time of arbitrary replenishment

orders from warehouse m at the consolidation dock is denoted with Cm

.

The waiting time of an order due to shipment consolidation is dependent on the amount
of goods already consolidated when the order arrives at the consolidation dock Yk . We
are further interested in the consolidated quantity m (t) for warehouse m at time t and
the number of arrivals of replenishment orders from warehouse m between the arrival of a
replenishment order at the consolidation dock and the departure of the truck N (Qcm Yk )
The distribution of the following random variables is unknown:

Time between subsequent replenishment orders of stockpoint k M

Time between two subsequent arrivals of orders at stockpoint k M\E1
Demand sizes at stockpoints k M\E1
Inter-arrival time of replenishment orders from stockpoint k M
at the consolidation dock

Cm
Inter-arrival time of an arbitrary replenishment order from
warehouse m at the consolidation dock for m W
m (t)
Consolidated quantity for warehouse m W at time t
Yk
Amount consolidated between the last truck departure
towards m W and the arrival of an arbitrary order from k Lm
c
N (Qm Yk ) Number of arrivals of replenishment orders from m between the
arrival of a replenishment order at the consolidation dock
from k Lm and the departure of the truck towards m W.
Dk (T )
Demand at stockpoint k M during a time interval with length T
s
Wk
Waiting time due to a lack of stock for a replenishment order from
stockpoint k M
c
Wk
Waiting time due to shipment consolidation for a replenishment order
from stockpoint k M
Rk
Ak
Dk
Ck

The approximation techniques

In general, the exact expressions of the distribution functions or of the moments of the
above mentioned random variables are difficult to obtain or intractable. Therefore, our
proposed algorithm is mainly based on approximated expressions. Before we will derive
these expressions we will explain the used techniques. We can distinguish between three
different approaches.

3.1

Fitting mixed Erlang Distributions

When exact expressions for the distribution function of a random variable are difficult to
obtain or intractable, then we can resort to the following approximation technique. We
evaluate the first two moments of the random variable and we approximate the distribution
of it by a mixed-Erlang distribution with the same first two moments.
A mixed-Erlang distribution, denoted as Ek1 ,k2 ((1 , 2 ), (p1 , p2 )), is here defined as the
mixture of two-Erlang distributions. This means a mixed-Erlang distributed random
variable X is with probability p1 (resp. p2 = 1 p1 ) the sum of k1 1 (resp. k2 1) inde10

pendent exponentials with mean

1
1

(resp.

1
).
2

The density of an Ek1 ,k2 ((1 , 2 ), (p1 , p2 ))

distribution is therefore given by

fX (x) :=

2
X

pi ki i

i=1

xki 1 i x
e
(k1 1)!

x>0

(2)

The parameters pi , ki and i (i = 1, 2) can be found in the following way if the first two
moments of X are known.
- If c2X < 1 then a mixture of two Erlang distributions with the same scale parameter
is used. Hence,
k1 := c12 , k2 := k1 + 1
X

p
p1 := 1+c1 2 k2 c2X k2 (1 + c2X ) k22 c2X , p2 := 1 p1
X

1 :=

k2 p1
,
E[X]

2 := 1

- If c2X 1 then a mixture of two exponential distributions used.

k1 := 1, k2 := 1
r

c2X 21
2
1 := E[X] 1 + c2 +1 , 2 :=
X

p1 :=

1 (2 E[X]1)
,
2 1

4
E[X]

p2 := (1 p2 )

This fitting technique is common, see for example Tijms (1994).

3.2

Superposition of renewal processes

The demand process of a stockpoint k is the superposition of the replenishment order

processes of the stockpoints j Sk . We assume that the superposition process is again
a renewal process and for the computation of the first two moments of the inter-arrival
times of orders we use the stationary interval method developed by Whitt (1982). But
instead of superposing hyper-exponential and shifted exponential distributions, we superpose mixtures of Erlang distributions. A detailed description of the algorithm is given in
Appendix A.

11

3.3

Approximations based on renewal theory

One of the problems when analyzing the distribution network is to determine the number
of events N (L) in a given time interval with length L, when the events follow a renewal
process. If the moments of the inter-arrival time X of the events are known, then the
first two moments of N (L) can be computed using asymptotic relations from standard
renewal theory and we obtain (see, for example Cox (1962))
L
E[X]

(3)

E[X 2 ]
1 E[X 2 ]2
E[X 3 ]
L2
+
L

E[X]2
E[X]3 E[X]
2E[X]4 3E[X]3

(4)

E[N (L)]
and
E[N (L)2 ]

if time epoch zero is an arbitrary point in time. Otherwise, if at time zero an arrival
occurs, we get:
L
E[X 2 ]
E[N (L)]
+
1
E[X] 2E[X]2

(5)

and
2E[X 2 ]
3 3E[X 2 ]2 2E[X 3 ] 3E[X 2 ]
L2
+L

+1
E[N (L) ]
E[X]2
E[X]3
E[X]
2E[X]4
3E[X]3 2E[X]2
2

(6)

If L is a random variable then we take the expectations of the right hand side of the
equations (3) - (6).

4
4.1

Analytical approximations
The reorder level

Our algorithm determines the reorder levels such that given target fillrates ktarget can be
met. This can be done using the following relationship between the reorder levels sk and
the fillrates k , (k M)

k = 1

E[(Dk (Lk ) + Uk sk )+ ] E[(Dk (Lk ) + Uk sk Qk )+ ]

Qk

(7)

For the evaluation of (7) we have to distinguish between stockpoints k E1 at the first
echelon where demand sizes are random and the stockpoints k M\E1 at upper echelons
where demand sizes are equal.
12

For the stockpoints k at the first echelon E1 we use the approximation technique as described in section 3.1 and determine the first two moments of the demand during the leadtime Dk (Lk ) and the undershoot Uk and fit a mixed Erlang distribution. The expectations
in (7) can easily be computed due to the relation given in (8) which holds under the assumption that the random variable X is mixed Erlang distributed Ek1 ,k2 ((1 , 2 ), (p1 , p2 ))
and z R.
Z
(x z)dFX (x)
E[(X z) ] =
+

1
X
j=0

(z)j

2
X
i=1

ki i X z l (ki j)! i z
pi
e
(ki 1)! l=0 l!ikj lj+1

(8)

Formulas for the moments of the demand during the leadtime Dk (Lk ) and the undershoot
Uk are obtained using asymptotic relations based on renewal theory. They are given in
Appendix B.
j , (j Sk ).
Stockpoints k at upper echelons k M\E1 only observe one demand size Q
Since the reorderlevel is also a multiple of this demand size the undershoot Uk at these
stockpoints is always equal to zero. Further, the demand during the leadtime is a discrete
random variable and the expectations in (7) can be computed as
+

E[(Dk (Lk ) c) ] =

j P (Dk (Lk ) = iQ
j ) =
(i r)Q

j P (N (Lk ) = i)
(i r)Q

(9)

i=r

i=r

j . How to compute the probability function of N (Lk ) can be found in

with c = rQ
Appendix C.
It remains to explain how to compute the first two moments of the leadtime. As mentioned
in section 2.2 the leadtime is composed of the transportation time Ldk , the waiting time
due to a lack of stock Wks and the waiting time due to shipment consolidation Wkc . In
order to be able to use the approximation technique as described in section 3.1 we have
to determine the first two moments of Wks and Wkc .

4.2

The waiting time due to a lack of stock

From standard probability theory it is well-known that the first two moments of a continuous random variable W which can only have values with positive probability on the
13

interval [0, ] is given as

Z
E[W ] =
P (W x) dx

E[W ] =

2xP (W x) dx

(10)

In order to determine the cumulative probability function P (Wk x) of a replenishment

order from stockpoint k we use
P (Wk x) =

P (Wk x | Lj = ) fL () d

(11)

with j Pk , and we first compute the conditional probability function P (Wk x | Lj =

).
In the sequel we assume s 0. If the leadtime is equal to then the waitingtime Wk can
never be larger than . Further, the waitingtime of an order of size Qk arriving at time
t at stockpoint j Pk is smaller than x if the inventory position Zj (t l + x) at time
t + x minus the demand D(t + x , t) during the interval (t + x l, t) is larger than
the ordersize Qk .
Yj (t + x) Dj (t + x , t) Qk Wk x

(12)
Q

Since the inventory position is uniformly distributed on sj + iQk , i = 1, 2, . . . Qkj we get

P (Wk x | Lj = )
Qj
Qk

P (Yj (t + x) Dj (t + x , t) Qk | Yj (t + x) = sj + iQk )P (Yj = sj + iQk )

i=1

Qj
Qk

sj + (i 1)Qk
Qk X
P N ( x)
=
Qj i=1
Qk
How to obtain the distribution function of N (l) can be found in the Appendix C and the
integrals in (10) and (11) are computed numerically.

4.3

The consolidation process

Since the arrival process of replenishment orders at the consolidation dock has a large
impact on the truck departure process and therefore also on the waiting time due to
14

(13)

consolidation, we investigate in the sequel the process in more detail and we derive approximated formulae for the first to moments of the inter-arrival time Ck of an individual
replenishment process at the consolidation dock.
As already mentioned in Section 2.3 the inter-arrival times Ck of the replenishment orders
from stockpoint k at the consolidation dock are different from the inter-arrival times Rk
of the replenishment processes at the stockpoint Pk due to the clustering of arrivals at
the consolidation dock. See for an illustration Figure 3.
Arrivals of orders
at the stockpoint

Arrival of a
replenishment order
Wk (2)s
Out of stock

Rk

(1)

Wk (3)s

Arrival process of
replenishment orders
at a stockpoint

time

R k(2)

Arrival process of
replenishment orders
at the consolidation dock

time

Ck (1)

Arrivals of orders
to be consolidated

Figure 3: The clustering of arrivals at the consolidation dock

The following equation holds, denoting the inter-arrival time between the n-th and (n+1)(n)

th order of stockpoint k at the consolidation dock with Ck :

(n)

Ck

(n)

(n)s

= Rk Wk

(n+1)s

+ Wk

k M

(14)

Assuming that all random variables in (14) are identically distributed for all n leads to
E[Ck ] = E[Rk ]

k M
(n)s

The second moment of Ck is difficult to obtain because Wk

(15)
(n+1)s

and Wk

are dependent

as illustrated in Figure 3. But if the number of stockpoints allocated to a warehouse is

large and the fill rate at Pk is not too small then the probability that two replenishment
orders of stockpoint k have to wait for the same replenishment order that refills stockpoint
Pk is small and thus the probability that replenishment orders at the consolidation dock
for the same stockpoint are clustered is small. Therefore, we neglect the clustering effect
15

of the replenishment orders for stockpoint k at the consolidation dock. Then the following
formula can be obtained:
2 (Ck ) = 2 (Rk )

k M

(16)

Further, we are not only interested in the arrival process of replenishment orders of one
stockpoint k at the consolidation dock but in the aggregated process of arbitrary orders

from warehouse m. The first two moments of Cm

are derived using the superposition

algorithm as given in Appendix A.

4.4

The waiting time due to shipment consolidation

Replenishment orders at the consolidation dock for warehouse m W are collected until
the amount is equal to a predetermined quantity Qcm . Then the consolidation process
starts all over again. The collected quantity at the consolidation dock to be shipped
to warehouse m at time t is denoted with m (t). The departure times of the trucks
are random, because they depend on the arrival process of replenishment orders at the
consolidation dock (see Figure 4).
Consolidation
proces

Truck departure
Q cm

time

Figure 4: Evolution of the consolidated quantity

The waiting time due to shipment consolidation for an order of size Qk from stockpoint k
Lm is dependent on the order process after the arrival of this order and the amount already
consolidated. We define Yk as the collected amount immediately after a replenishment
order of stockpoint k has arrived at the consolidation dock and N (Qcm Yk ) as the number
of arrivals between an arrival of a replenishment order of stockpoint k at the consolidation
dock and the departure of the truck.

16

Then the waitingtime can be calculated as

N (Qcm Yk )

Wkc

Cm,i

(17)

i=1

where Cm,i
is the inter-arrival time between the i-th and i+1-th order at the consolidation

dock.

Assuming that, N (Qcm Yk ) and Cm,i

are independent random variables, and Cm,i
is

identically distributed for all i, it follows from (17) that

]
E[Wkc ] = E[N (Qcm Yk )]E[Cm

(18)

]
E[(Wkc )2 ] = E[N (Qcm Yk )] 2 (Cm
) + E[N (Qcm Yk )2 ]E 2 [Cm

(19)

Since Qcm is a multiple of Qk the truck consolidation process, m (t), follows a discrete distribution. Moreover, in the steady state m (t) is uniformly distributed over
c

{0, Qk , 2Qk , ..., ( QQmk 1)Qk } with the different possibilities having the same probability
namely

Qk
Qcm

and it easily follows that

(

E[N (Qcm

Qc
m 1)
Qk

Qk X
1 Qcm
Yk )] = c
j=
1
Qm j=0
2 Qk
(

E[N (Qcm Yk )2 ] =

(20)

Qc
m
1)
Q

Qc
1
Qk X 2 1 Qcm
m

j
=

Qcm j=0
3 Q
2
Q

(21)

We can fill these expressions for the first two moments of N (Qcm Yk ) together with the

expressions for the first two moments of Cm

in equations (18) and (19) to get approxima-

tions for the first two moments of the waiting time due to shipment consolidation.

4.5

The average inventory level

In order to measure the performance of the distribution network we additional compute

the average inventory levels. We have to distinguish between the situation with random
demand sizes, as given at the first echelon, and the situation with deterministic demand
sizes as given at the upper echelons. For the first echelon, which means for all k E1 ,
we can use formula (22) and the approximation technique described in Section 3.1 to fit
17

the first two moments of the undershoot and the demand during the leadtime on a mixed
Erlang distribution. Denoting with Ak the asymptotic residual lifetime of Ak and with

k ]E[Ak ]
we get
ck := E[Dk ] E[AE[A
k]

1
+ 2
+ 2
E[ (sk + Qk Dk (Lk )) ] E[ (sk Dk (Lk )) ]
E[Ik ] =
2Qk
h
+ i
h
+ i
E sk + Qk Uk Dk (Lk )
E sk Uk Dk (Lk )
+ck
Qk

(22)

(for a proof of (22) see Kiesm

uller et al. (2004)).
For the upper echelons we have to take into account that the demand sizes are deterministic and that the inventory level can only have discrete values. Then formula (23) can be
used for computing the average inventory level for the stockpoints k M\E1 .

iX
s+Q

Q2j
1

(is+Q i)E[Ak ] + (is+Q i)(is+Q i 1)E[Ak ] P (N (Lk ) = i)

E[Ik ] =
Qk E[Ak ] i=0
2

X
is

Q2j
1

+
(is i)E[Ak ] + (is i)(is i 1)E[Ak ] P (N (Lk ) = i)
Qk E[Ak ] i=0
2

(23)

with j Sk , is =

s
Qj

and is+Q =

s+Qk
.
Qj

For a proof of (23) see Appendix E.

The algorithm

The algorithm can be split in two main parts. In the first part we determine the replenishment order processes, the aggregate demand processes and the consolidation processes.
The demand process at a stockpoint can be translated in the replenishment order process
independently of the value of the reorder level. The superposition of all replenishment
order processes of the stockpoints j Sk describes the aggregate demand process at
stockpoint k, and the superposition of the replenishment order processes of all stockpoints j Lm describes the consolidation process. We start at the lowest echelon and
compute successively the moments of the processes. A sketch of the first part of the
algorithm is given below.
1. Set n = 1.
18

2. Compute the first two moments of the inter-arrival time of replenishment orders
Rk , k En using the formulas (41) and (42) as given in Appendix D.
3. Compute the first two moments of the inter-arrival time of orders Aj for all j En+1
by superposing the replenishment processes of all stockpoints k Sj using the
algorithm given in Appendix A.

4. Compute the first two moments of the inter-arrival time Cm

of an arbitrary re-

plenishment order from warehouse m at the consolidation dock by superposing the

replenishment processes of all stockpoints j En , j Lm using the algorithm as
given in Appendix A.
5. Set n := n + 1 and repeat step 2,3, and 4 until n = N 1
After determining the moments of the waiting time due to a lack of stock and of the waiting
time due to shipment consolidation the moments of the leadtime can be computed. Since
the demand process is already determined in the first part of the algorithm and target
service levels are given, the reorderlevels can be computed such that a service level of at
least ktarget can be met. The second part of the algorithm starts at the highest echelon and
computes successively the reorder levels echelon for echelon until all values are determined.
We outline the second part of the algorithm as follows:
6. Set n := N, Wks := 0 and Wkc := 0 k En .
7. Compute the first two moments of the leadtime Lk k En using relation (1).
8. Compute the reorder levels sk k En using equation (7).
9. Set n := n-1
10. Compute the moments of the waiting time due to a lack of stock Wks k En as
described in section 4.2.
11. Compute the moments of the waiting time due to shipment consolidation Wkc k
En using the formulas (18) and (19).
12. Compute the first two moments of the leadtime Lk k En using relation (1).
13. Compute the reorder levels sk k En using equation (7).
19

14. Repeat step 9-13 until n = 1

For the application of the algorithm data about the demand process at the lowest echelon
(Ak , Dk ), (k E1 ), the transportation times Lkd , the batchsizes Qk and the target fillrates
ktarget at all stockpoints and the target dispatch quantity Qcm for all warehouses are needed
as input parameters.

Numerical study

In this paragraph we investigate the performance of the approximations used in our algorithm by means of discrete event simulation. The performance of the approximations
for a single item distribution network without order consolidation is tested in Kiesm
uller
et al. (2004) and therefore, we only study the performance of the approximations of the
moments of the inter-arrival times of the replenishment orders at the consolidation dock
and the moments of the waiting time due to shipment consolidation. We further report on
the performance of the distribution network in terms of actual service level and average
inventory when reorder levels are used, computed with our algorithm.
As a base case we consider a two echelon distribution network with one central warehouse
I and 4 decentral warehouses {II, III, IV, V }. At k E1 demand arrives according to a
compound renewal process. The inter-arrival time and the order size of k E1 are mixedErlang distributed with known first two moments. In the first part of the numerical study
we assume all customers to be identical. For the simulation of the base case the input
parameters are chosen as given in Table 1.
To investigate the effects of the input parameters on the output parameters we vary one
input parameter while the other parameters are fixed. If not mentioned something else,
we have chosen the input parameters as given in Table 1.
For measuring the performance of the approximations, we define (Zk ) the relative error
of Zk where k W or M depending of the context.
(Zk ) :=

|Zksimu Zk |
100%
Zksimu

(24)

Zk is the calculated and Zksimu is the simulated value. Further, we define (Zk ) as the
20

Parameter
for
value
E[Dk ]
k E1
50
2
cDk
k E1
1
E[Ak ]
k E1
1
2
c Ak
k E1
1
Qk
k E1
500
Qk
k E2
4000
d
Lk
k E1
2
Lk
k E2
4
target
k
k E1
0.95
target
k
k E2
0.75
c
Qm
m {II, III, IV, V } 2000
Table 1: Input parameter values for the base case
average relative error
(Zk ) :=

K
P

k=1

(Zk )

(25)

and max( (Zk ) ) as the maximum relative error

max( (Zk ) ) :=

6.1

(Zk )

max
k
k = 1, ..., K

(26)

Approximations for the second moment of Cm

In order to compute the second moment of the inter-arrival times of replenishment orders at the consolidation dock we have used the approximation technique based on the
superposition of renewal processes as developed by Whitt (1982). We first investigate
the influence of the number of renewal processes which are superposed on the quality of
the approximations. Therefore, we vary the number of different items at a warehouse
m (|Lm | {4, 8, 12, 16, 24, 32}, m {II, III, IV, V }) and analyze the errors. In order to
exclude the clustering effect we have assumed a high service level at the upper stockpoints
and we have set ktarget = 0.90 k E2 . In Figure 5 we illustrate the average relative error
2
(E[(Cm ) ]) computed over all stockpoints at the first echelon for different values for the

squared coefficient of variation of the inter-arrival times of customer demand at the first
echelon and for different number of items.
It can be seen that the average relative error is quite small for all examples and decreases
21

Figure 5: The influence of the number of different items on the performance of the ap 2
proximation of E[(Cm
)]

with increasing squared coefficient of variation of the inter-arrival time of the demands at
the first echelon. We can further observe that after a short increase the error is decreasing
with increasing number of items.

Figure 6: The influence of the demand size variation on the performance of the approxi 2
mation of E[(Cm
)]

In Figure 6 the influence of the demandsize variability at the first echelon is illustrated.
Different values of the coefficient of variation lead to different average errors, but we can
see, that the differences are quite small and that other parameters have a larger impact
on the performance, as for example the number of different items.
In order to investigate the influence of the clustering effect on the performance of the
2
approximations we have computed the second moment E[(Cm
) ] of the inter-arrival time

of replenishment orders at the consolidation dock for different target service levels ktarget =
22

betaW, k E2 . We have chosen a small number of items (4) to have a high probability of
cluster forming and we study the performance of the approximations as a function of the
target service level at the second echelon betaW . The results are illustrated in Figure 7.

Figure 7: The influence of the target service level at the upper echelon on the performance
2
of the approximation of E[(Cm
)]

It can be seen that for smaller target service levels at the upper echelon larger errors of
the approximation can be observed due to the cluster effect. It is obvious that in case of
a lower target service level it happens more often that orders have to wait due to a lack of
stock, and therefore it can also happen more often that two orders for the same item and
the same warehouse have to wait for a replenishment. Since in many practical situations
target service levels are chosen quite high, we conclude that the clustering effect can be
neglected.

6.2

The waiting time due to shipment consolidation

We are interested in the influence of the shipment consolidation process on the leadtime,
respectively on the waiting time due to shipment consolidation. Although formulae (20)
and (21) are exact, formulae (18) and (19) only provide approximations since the moments of the inter-arrival times of the replenishment orders at the consolidation dock are
approximations. In the following we study the impact of these approximations on the
moments of the waiting time due to shipment consolidation.
We can observe a similar behavior as for the approximations for the inter-arrival time for
the replenishment orders at the consolidation dock. With increasing number of different
23

items the average relative errors are decreasing and the impact of the variability of the
demand process is quite small as well as the impact of the target service level at the
second echelon (see Table 2 in Appendix E).

6.3

The performance of the distribution network

In the last part of the numerical study we investigate the performance of the distribution
network when the reorder levels are computed with our algorithm. Moreover, we compute
the actual service levels observed by the customers and the average inventory levels with
the formulae (7),(22), and (23) and compare our results with results obtained by simulation. Note that due to the discrete demand distribution at the upper echelon we have to
find
:= min{sk | k (sk ) ktarget ,
smin
k

k E2 }

(27)

to determine the reorder level for stockpoint k M\E1 . We compare the computed service
level k (smin
k ) and the service level obtained by simulation and measure the absolute error
by
(k )

:= |ksimu k (smin
k )| 100%

k M\E1

(28)

and
(k )

:= |ksimu ktarget | 100%

k E1

(29)

(k ) and max( (k ) ) are computed similar to (25) and (26) .

We have considered one leadtime scenario Ldk = 2, k E1 and Lk = 4, k E2 and two
scenarios of the lotsizes and the capacity of the vehicles.
1. Qk = 200, k E1 , Qk = 800, k E2 and Qcm = 800, m {II, III, IV, V }
2. Qk = 500, k E1 , Qk = 4000, k E2 and Qcm = 2000, m {II, III, IV, V }
The other parameters are chosen as given in Table 2. The number of different items at
each warehouse are equal and we have computed all possible combinations, which leads
in total 532 examples.
Additionally, we considered another set of 532 examples where customers are non-identical.
For the average demand sizes we have drawn values randomly from a uniform distribution
24

Parameter
for
E[Dk ]
k E1
2
cDk
k E1
E[Ak ]
k E1
2
c Ak
k E1
target
k
k E1
target
k
k E2
Lm
m I, II, III, IV

values
50
0.4,1,1.6
1
0.4,1,1.6
0.90,0.95, 0.99
0.75, 0.9
4,8,12,16

Table 2: Input parameter values

of the interval (30,70) and for the average arrival times we have chosen randomly from a
uniform distribution of the interval (0.5,2.5)
In Figure 8 the average of the absolute deviation of the service level at the first echelon
is illustrated. It can be seen that our procedure for allocation the safety stock leads to
a performance very closed to the required. We can further observe smaller deviations
in case of non-identical demand streams. This is in line with the results obtained in
Kiesm
uller et al. (2004) where it is also illustrated that the approximations perform best
in heterogeneous distribution structures, i.e. non-identical demand streams.

average service level

350

identical demand streams

different demand streams

300

frequency

250
200
150
100
50
0
0.00%

1.00%

2.00%

3.00%

absolute deviation

Figure 8: The absolute mistake of the service level at the first echelon
The performance of the approximations for the average inventory levels are illustrated in
Figure 9. On the left hand side a histogram for the relative mistake of the approximation
at the first echelon is given while on the right hand sight the relative mistakes for the
approximations for the second echelon are given. Again, we can observe a very good
25

performance and only quite small deviations between the approximations and the values
obtained by simulation.

average inventory at the second echelon

350

300

300

250

250

200

identical demand streams

different demand streams

150

frequency

frequency

average inventory at the first echelon

350

identical demand streams

different demand streams

200
150

100

100

50

50
0

0
0,00%

1,00%

2,00%

3,00%

0.00%

4,00%

1.00%

2.00%

3.00%

4.00%

relative mistake

relative mistake

Figure 9: The relative mistake of the approximation of the inventory levels

Summary and conclusions

In this paper we have provided an algorithm to determine reorder levels, respectively safety
stocks, for stockpoints in a multi item multi-echelon stochastic inventory system where
replenishment orders are consolidated based on a quantity-based shipment policy. Our
computational procedure is based on approximated expressions for the leadtime, inventory
levels and service levels. We have tested the quality of the approximations by means
of simulation and we have illustrated that the performance of a distribution network,
where policy parameters are determined with our algorithm, is closed to the required
one. Moreover, the algorithm enables us to analyze the trad-off between inventory levels
and transportation frequency. Combined with a cost model, which is respect to future
research, the proposed algorithm is an excellent planning tool to control inventory and
transportation cost in a distribution network.

References
Axsater, S., 2003. Supply Chain Operations: Serial and Distribution Inventory. Handbooks in Operations Research and Management Science, vol 11, Supply Chain Management: Design, Coordination and Operation. Editors: de Kok, A.G., and Graves, S.C.
Elsevier North-Holland.
26

Axsater, S., 1993. Continuous review policies for multi-level inventory systems with
stochastic demand. Handbooks in Operations Research and Management Science, vol
4, Logistics of Production and Inventory. Editors: Graves, S.C., Rinnooy Kan, A.H.G.
and Zipkin, P.H. Elsevier North-Holland.
Axsater, S., 1998. Exact and approximate evaluation of batch-ordering policies for twolevel inventory systems. Operations Research 41, 777-785.
Axsater, S., 2001. A Note on Stock Replenishment and Shipment Scheduling for VendorManaged Inventory Systems. Management Science, Vol. 47, No. 9, 1306-1310.
Bertazzi, L., and Speranza, M.G., 1999. Models and Algorithms for the Minimization of
Inventory and Transportation Costs: A survey, Lecture Notes in Economics and Mathematical Systems vol 480, New Trends in Distribution Logistics. Editors: Speranza,
M.G. and Staehly, P., eds., Springer-Verlag, Berlin Heidelberg.
Burns, L. D., Hall, R.W., Blumenfeld, D.E. and Daganzo,C.F. 1985. Distribution Strategies that Minimize Transportation and Inventory Costs. Operations Research, Vol. 33,
469-490.
Cachon, G., 2002. Managing a Retailers Shelf Space, Inventory, and Transportation.
Manufacturing and Service Operations Management, Vol. 3, No. 3, 211-229.
C
etinkaya, S. and Lee, C.Y., 2002. Optimal Outbound Dispatch Policies: Modeling Inventory and Cargo Capacity. Naval Research Logistics, Vol. 49, 531-556.
C
etinkaya, S. and Lee, C.Y., 2000. Stock Replenishment and Shipment Scheduling for
Vendor-Managed Inventory Systems. Management Science, Vol. 46, No. 2, 217-232.
Cox, D.R., 1962. Renewal theory. Methuen, London
Diks, E.B., Kok, A.G. de and Lagodimos, A.G., 1996. Multi-echelon systems: a measure
perspective. European Journal of Operational Research, vol 95, 241-264.
Federgruen, A., 1993. Centralized planning models for multi-echelon inventory systems
under uncertainty, Handbooks in Operations Research and Management Science, vol
4, Logistics of Production and Inventory. Editors: Graves, S.C., Rinnooy Kan, A.H.G.
and Zipkin, P.H. Elsevier North-Holland.

27

Higginson, J.K. and Bookbinder, J.H., 1994. Policy Recommendations for a shipment
consolidation program. Journal of Business Logistics 15, 87-112.
Houtum, van G.J., Inderfurth, K., and Zijm, W.H.M, 1996. Material coordination in
stochastic multi-echelon systems. European Journal of Operational Research, vol 95,
1-23.
Kiesm
uller, G.P., de Kok, A.G., Smits, S.R., Van Laarhoven, P.J.M., 2004. Analysis of
Divergent N-echelon (s,nQ)-policies Under Compound Renewal Demand, OR Spectrum,
vol 26/4, 547-577.
Masters, J.M., 1993. Determination of near optimal stock levels for multi-echelon distribution inventories, Journal of Business Logistics 14(2), 165-195.
Muckstadt, J.A., 1973. A model for a multi- item, multi-echelon, multi-identure inventory
sytem, Management Science 20(4), 472-481.
Popken, D.A., 1994. An algorithm for the multiattribute, multicommodity flow problem
with freight consolidation and inventory costs. Operations Research 42(2), 274-286.
Ross, S.M., 1993. Introduction to probability models, Fifth edition, Academic Press.
Sherbrooke,C.C., 1968. METRIC: A Multi-Echelon Technique for Recoverable Item Control, Operations Research 16, 122-141.
Silver, E.A., Pyke, D.F., and Peterson, R.,1998. Inventory Management and Production
Planning and Scheduling, third edition, Wiley, New York.
Smits, S.R., de Kok, A.G., 2002. Approximations for the waiting time in (s,nQ)-inventory
models for different types of consolidation policies. Lecture notes Economics and Mathematical Systems 519, Springer Verlag.
Tijms, H.C., 1994. Stochastic models: an algorithmic approach, John Wiley & Sons.
Whitt, W., 1982. Approximating a point process by a renewal process, I: Two basic
methods, Operations Research 30, 125-147.

28

Appendix A: Algorithm to compute the first two moments of a superposed

compound renewal process
The inter-arrival time of the single processes is denoted with Xi and of the superposed
process it is denoted with X0 . The following iterative algorithm can be used to compute
the first two moments of X0 .
1. Order the N inter-arrival processes from the largest to the smallest first moment.
2. Compute
1

(1)

E[X0 ] :=

2
P

i=1
(1)

(1)

E[(X0 )2 ] := 2E[X0 ]

Z Y
2
i=1

1
E[Xi ]

(30)

1
E[Xi ]

! 2 Z
Y

(1 FXi (y))dy dx

(31)

i=1 x

(1)

3. Fit a mixed-Erlang distribution to the first two moments of X0 .

4. Initially set n:=2 and i:=3.
5. Compute
1

(n)

E[X0 ] :=

1
(n1)
]
E[X0

(32)

1
E[Xi ]

(n)

E[(X0 )2 ] :=
(n)

2E[X0 ]

Z
0

1
(n1)
E[X0
]E[Xi ]

Z
Z
(1 F (n1) (y))dy (1 FXi (y))dy dx (33)
X
0

(n)

Fit a mixed-Erlang distribution to the first two moments of X0

(N )

6. If n < N then n := n + 1, i := i + 1 and go to step 5 else E[X0 ] := E[X0 ] and

(N )

E[X02 ] := E[(X0 )2 ]

29

Appendix B: The demand during the leadtime for random demand sizes
The moments of the demand during the leadtime can be computed using the following
equations:
E[Dk (Lk )] = E[N (Lk )]E[Dk ],

k E1

E[Dk2 (Lk )] = E[N (Lk )] 2 (Dk ) + E[N (Lk )2 ]E[Dk ]2 ,

(34)
k E1

(35)

The moments of the number of demand arrivals can be determined using the approximations based on renewal theory.
E[N (Lk )] =

E[A2k ]
E[Lk ]
1,
+
E[Ak ] 2E[Ak ]2

k E1

2E[A2k ]
3
E[L2k ]
+ E[Lk ]

E[N (Lk ) ] =
E[Ak ]2
E[Ak ]3
E[Ak ]
2 2
3
3E[Ak ]
2E[Ak ]
3E[A2k ]
+

+ 1,
2E[Ak ]4 3E[Ak ]3 2E[Ak ]2

(36)

k E1

(37)

The moments for the undershoot are given as

E[Uk ] =

E[Dk2 ]
2D[Dk ]

E[Uk2 ] =

E[Dk3 ]
,
3E[Dk ]

k E1

(38)

Appendix C: The probability function of N (Lk )

P (N (Lk ) = i | Lk = ) = P (N () i) P (N () i + 1)
i
i+1
X

= P
Aj P
Aj
j=1

(39)

j=1

We use the approximation technique described in 3.1 to fit a mixed Erlang distribution
i
P
on the moments of
Aj .
j=1

P (N (Lk ) = i) =

P (N (Lk ) = i | Lk = )fLk () d

(40)

Appendix D: The first two moments of the of the time between replenishments
Rk
E[Rk ] =
E[Rk2 ] =

Qk
E[Ak ] k M
E[Dk ]

(41)

Q2
c2Dk
E[Dk2 ]2 2E[Dk3 ]
Qk 2
k
[Ak ]+E 2 [Ak ]
+
+Q

k M (42)
k
E[Dk ]
E[Dk ]2
E[Dk ] E[Dk ]4 3E[Dk ]3

30

Appendix E: The average inventory level

Define H(nQj ) as the expected area between the physical inventory level and the zero
level, given that the physical level is nQj at time zero, there are no outstanding orders
and no orders are placed in the future. Further we consider the situation that a customer
arrives at time zero. Then we get
H(nQj ) = E[Ak ]Qj +

n(n + 1)
,
2

k M\E1 , j Sk

(43)

For the situation that time zero is an arbitrary point in time we get for the expected area

between the physical inventory level and the zero level H(nQ
j)
n(n 1)

E[Ak ])
H(nQ
j ) = Qj (nE[Ak ] +
2

(44)

where Ak is the residual lifetime of Ak .

Now we consider the average stock on hand per time unit in an arbitrary replenishment
cycle, defined as the time between two successive arrivals of replenishment orders. It is
given as

b )] E[H(I
e )]
E[H(I
(45)
E[length of a replenishment cycle]
where Ib (Ie ) denotes the stock on hand at the beginning (end) of a replenishment cycle.
E[I] =

We get
E[Ik ] =

k + Qk D(Lk ))] E[H(s

k D(Lk ))]
E[H(s
Qk E[Ak ]
Qj

(46)

If is+Q is defined as
is+Q :=

sk + Qk
Qj

(47)

then we get
is+Q

k + Qk D(Lk ))] =
E[H(s

s+Q Qj iQj )P (N (Lk ) = i)

H(i

i=0

is+Q

= Qj

X
i=0

(is+Q i)E[Ak ] + (is+Q i)(is+Q i 1) P (N (Lk ) = i) (48)

2

Similar we get with is :=

k D(Lk ))] =
E[H(s

sk
Qj

is
X

s Qj iQj )P (N (Lk ) = i)
H(i

i=0

is

X
1

(is i)E[Ak ] + (is i)(is i 1) P (N (Lk ) = i) (49)

= Qj
2
i=0

The formulas (48) and (49) together with (45) proof our formula (23).
31

Appendix F:
ktarget , k E2

items

c2D
0.4

1.0

1.6
0.75
0.4

16

1.0

1.6

0.4

1.0

1.6
0.90
0.4

16

1.0

1.6

c2A
0.4
1.0
1.6
0.4
1.0
1.6
0.4
1.0
1.6
0.4
1.0
1.6
0.4
1.0
1.6
0.4
1.0
1.6
0.4
1.0
1.6
0.4
1.0
1.6
0.4
1.0
1.6
0.4
1.0
1.6
0.4
1.0
1.6
0.4
1.0
1.6

max((E[W c ])
1.08
0.57
0.70
0.76
1.01
0.65
0.61
0.61
0.74
1.16
1.13
1.30
1.67
1.14
1.09
1.23
0.97
1.31
0.91
0.53
0.66
0.74
0.92
0.72
0.49
1.10
0.63
1.12
1.25
1.43
1.56
0.94
1.21
1.18
0.90
1.54

(E[W c ])

0.44
0.19
0.30
0.39
0.41
0.31
0.25
0.22
0.32
0.36
0.36
0.42
0.44
0.36
0.34
0.36
0.36
0.36
0.41
0.22
0.26
0.38
0.37
0.26
0.25
0.24
0.34
0.43
0.41
0.49
0.43
0.36
0.36
0.25
0.24
0.34

max((E[(W c )2 ])
7.28
5.24
5.13
5.39
5.23
4.41
4.62
3.70
3.18
2.97
2.68
3.10
2.98
2.21
2.09
2.46
1.87
2.20
7.54
6.08
4.95
6.42
5.64
4.86
4.97
4.21
3.87
2.91
3.20
3.20
2.86
2.64
2.59
4.97
4.21
3.87

(E[(W c )2 ])

6.41
4.77
3.90
4.84
3.86
3.17
3.83
2.89
2.41
0.91
1.02
0.97
0.87
0.86
0.75
0.79
0.68
0.68
6.75
5.26
4.39
5.30
4.24
3.75
4.31
3.41
2.92
1.00
1.08
1.18
1.03
1.10
0.95
4.31
3.41
2.92

Table 3: Performance of the approximations for the first and second moment of the
waiting time due to shipment consolidation

32