International Journal of Production Research, 2019

Vol. 57, No. 13, 4148–4165, https://doi.org/10.1080/00207543.2018.1518606

Optimising installation (R,Q) policies in distribution networks with stochastic lead times: a
comparative analysis of guaranteed- and stochastic service models
a,b∗ a,c
Niels De Smet , El-Houssaine Aghezzaf and Bram Desmetb
a Department of Industrial Systems Engineering and Product Design (ISyEPD – EA18), Faculty of Engineering and Architecture,
Ghent University, Zwijnaarde, Belgium; b Solventure NV, Gent, Belgium; c Flanders Make, Lommel, Belgium
(Received 19 April 2018; accepted 24 August 2018)

This paper studies two modelling approaches to the multi-echelon inventory optimisation problem in a distribution network
with stochastic demands and lead times. It compares the performance of a novel guaranteed-service model (GSM), using
an installation (R, Q) inventory control policy, with a stochastic service model (SSM) considering ordering, holding and
flexibility costs. From both cycle service level and fill rate perspectives, our numerical analysis of the 1-warehouse 2-
retailer network shows that cost difference between both models is driven by the internal service level at the warehouse.
The GSM outperforms the SSM for over 80% of the simulated instances and realises an average total cost improvement of
approximately 10%. This analysis goes against earlier results that showed a relatively low-cost difference between the two
approaches, and demonstrates that it is worthwhile to evaluate competing models for multi-echelon inventory optimisation
in real-world supply chains with batch ordering and variable lead times.
Keywords: inventory control; distribution systems; multi-echelon; guaranteed-service; stochastic service

1. Introduction
The subject and objective of this paper was inspired by the following question that was posed to the authors of this work by
a large manufacturing company:
Which multi-echelon inventory optimisation model should we apply to our distribution network, if we want to include
the variability of our lead times and setup costs through lot sizing decisions in the model, as we believe these are the main
cost drivers for our supply chain?
Academic literature on safety stock optimisation models for multi-echelon supply chains can be divided into two
categories: the guaranteed-service models (GSM) and thestochastic service models (SSM). The GSM was introduced by
Simpson Jr (1958), a few years later, Clark and Scarf (1960) introduced the SSM, one of the most seminal papers in the field
of multi-echelon inventory optimisation.
These two models differ in how they treat demand propagation and service times. The original GSM assumes that each
stage can quote an outbound service time to its downstream stage(s) in which it guarantees to deliver. This service time
is deterministic because the GSM bounds the demand at every stage, which means that it will only commit to satisfy a
portion of the demand variability, up to a pre-specified demand bound. For demand that exceeds this bound, expediting,
subcontracting, premium freight transportation and/or overtime are assumed to be used, a cost that is typically not explicitly
modelled. The objective function of the model is concave, combined with a polytope as feasible region, this leads to the well-
known ‘all or nothing’ property. The latter states that the optimal solution will be an extreme point of the feasible region,
thus a stage will either keep zero safety stock, or enough safety stock to cover the maximum of the net replenishment lead
time, which is the sum of the incoming service time at the stage and its nominal lead time, the time it takes to complete the
process at that stage given that all required components are available.
The SSM does not put a bound on demand, nor assumes a deterministic outgoing service time. Extraordinary measures
like expediting can’t be used, allocating more inventory is the only measure against variability in this approach. If an
upstream stage has insufficient inventory, the items are backordered, the downstream stage will hence have to wait for
one of the next shipments before the demand can be fulfilled. The characterisation of the backorder service time is what
complicates the model and makes the lead times stochastic.

*Corresponding author. Email: [email protected], [email protected]

© 2018 Informa UK Limited, trading as Taylor & Francis Group

 International Journal of Production Research 4149

Historically, the mathematical elegance and (computational) simplicity of the GSM have led to a larger adoption rate
compared to the SSM. However, multi-echelon inventory optimisation has been around for almost 60 years now, we argue
that the recent advancements in computing should have lessened the computational effort required to apply the SSM to
real-world supply chains, and thus should be seen as a viable candidate for use in practice. According to Walker (2009), and
recent price quotes from services such as Google’s and Amazon’s cloud platform, the cost per CPU hour is less than $0.10.
Knowing that many of the computations can be parallelised, leads to the conclusion that running more complex models
which can potentially lead to better results, are definitely justified from a cost perspective, especially for larger companies
with millions of dollars of inventory and significant setup costs.
More recent works on stochastic inventory optimisation include Khan, Hussain, and Cárdenas-Barrón (2017),
who studied the impact of learning in production and screening errors in a vendor–buyer supply chain with defec-
tive items where the demand during lead time is normally distributed, and Minoux (2018), who showed how one
can construct state–space representable uncertainty sets at any probability level that lead to efficient resolution of
both the stochastic and robust versions of various multi-echelon inventory and production problems under Markovian
uncertainty.
Unfortunately, to date there is very little research on how the performance of both models compare to each other (Graves
and Willems 2003; Klosterhalfen and Minner 2010; Li 2013; Li and Wu 2018), and there seems to be no conclusive evidence
on which approach is preferred when batch ordering and lead time variability should be included in the model. This is why
we derive two mathematical models in this work to test both approaches for various parameter sets on a simple 1-warehouse
2-retailers network. With this analysis, one can make an informed recommendation on the modelling approach for larger
distribution networks, such as the one from the aforementioned company.
As for the academic contribution, to the best of our knowledge, the numerical study we present that compares both
models is the first one that takes into account lot sizing decisions through the installation (R, Q) inventory control policy,
stochasticity of the lead times, ordering, holding and flexibility costs, and that compares performance both for cycle service
levels and fill rates. A second contribution is the development of the first GSM that can optimise inventory in distribution
networks with installation (R, Q) inventory control policies and stochastic lead times.
This paper is organised as follows. Section 2 presents the literature review related to the GSM, SSM and comparison
between both models, Section 3 derives the mathematical models that will be used for the numerical study in Section 4,
finally, we summarise our main conclusions and provide future research suggestions in Section 5.

2. Literature review
2.1. Guaranteed-service model
A recent review paper (Eruguz et al. 2016) notes that 80% of the work on the GSM has been published in the last decade,
which is remarkable since the approach was first developed more than 50 years ago. This paper presents a detailed overview
of how the several base assumptions regarding external demand, lead times, capacity constraints, service times, replenish-
ment policies, review periods and extraordinary measures of the GSM have been challenged since the work of Simpson
Jr (1958) and Graves and Willems (2000).
In this work, we zoom in on the assumptions regarding lead times which are considered deterministic and the base-stock
ordering policy. Both are significant oversimplifications for many real-world supply chains and have hampered the adoption
of this modelling approach.
Inderfurth (1993) was the first to challenge the deterministic lead time assumption by assuming that the lead time at
each stage in the network is normally distributed.
 His approach lead to an expression similar to the single-echelon safety
stock (SS) formula, which is known as SS = k μL σd2 + μ2d σL2 , where k is a safety factor, and demand and lead times are
normally distributed, d ∼ N (μd , σd2 ) and L ∼ N (μL , σL2 ) respectively. In the work of Minner (2000), it is shown that the
former approach leads to an overestimate of the required safety stocks, and an alternative formulation for serial systems is
presented. Humair et al. (2013) derive an expression for the safety stock based on the positive part of the net replenishment
lead time, and early-arrival stock based on the negative part of the net replenishment lead time. A comparison between this
approach and two heuristics for 12 real-world supply chains shows that the former leads to the most accurate inventory
levels and a different inventory stocking strategy.
In many industries such as the air conditioning industry, ordering in batches and the associated (fixed) ordering cost that
comes with it is an inherent part of the inventory optimisation problem, see Desmet (2009). Li and Chen (2015) extended
the GSM model and consider a (guaranteed-service) formulation with a continuous-review batch ordering policy instead of
a periodic-review base stock policy.
4150 N. De Smet et al.

The model from Li and Chen (2015) will serve as the starting point for the GSM model in our numerical study, as it
already includes the optimisation of lot sizes via the echelon (R, Q) policy, and has proven to be effective for distribution
systems.

2.2. Stochastic service model

De Kok et al. (2018) presented a comprehensive survey on (stochastic) multi-echelon inventory models. Over 350 papers
on the SSM were reviewed and classified using a new typology. Simchi-Levi and Zhao (2012) provide another classification
for SSM, and discuss the strengths and weaknesses of the queueing-inventory, lead-time demand, and flow-unit method
for single-echelon, distribution, assembly and general networks. We also refer to the review papers of Diks, de Kok, and
Lagodimos (1996); Axsäter (2003); Van Houtum (2006) on SSM.
We focus on four sets of papers here that inspired the development of the SSM further used in this work.
A first set of papers by Deuermeyer and Schwarz (1981); Schwarz, Deuermeyer, and Badinelli (1985); Schwarz (1989);
McGavin, Schwarz, and Ward (1993); McGavin, Ward, and Schwarz (1997) concentrates on distribution systems with
continuous review installation (R, Q) policies, first-come-first-served (FCFS) allocation and backordering, but assumes that
lead times are constant.
On the characterisation of the actual lead time and the associated demand during lead time, we owe many of our insights
to Lee and Billington (1993); Lin et al. (2000); Ettl et al. (2000). Based on an inventory-queue model, Ettl et al. (2000)
optimises continuous review installation base-stock policies (S-1,S) with FCFS allocation and backordering that can be
applied to generic networks.
A third set of papers (Diks and de Kok 1999; de Kok and Visschers 1999; de Kok and Fransoo 2003; de Kok
et al. 2005) on the synchronised base stock policy show how a generic network can be decomposed into diver-
gent systems and subsequently in single-echelon system to determine the optimal order-up-to levels for the entire
network.
The work of Desmet (2009); Desmet, Aghezzaf, and Vanmaele (2010a, 2010b) draws on the insights from
all the previous sets of papers described in this section. The model presented by Desmet optimises continuous
review installation (R, Q) policies with FCFS allocation and backordering for generic networks with stochastic lead
times.
The academic research on batch ordering policies is relatively thin compared to the base stock policy, as shown by de
Kok et al. (2018). We chose the model from Desmet, Aghezzaf, and Vanmaele (2010a) as the basis for our SSM model
because it takes into account lot sizing through the continuous review installation (R, Q) policy and stochasticity of the lead
times, and can be applied to generic networks.

2.3. Comparison between GSM and SSM

Most of the academic research works focus on one of two approaches and try to expand its formulation such that it can tackle
more complex problems, or try to relax one of the assumptions made in order to find a mathematically elegant solution. In
this work, however, we aim to address performance differences between the GSM and the SSM, similar to the work of
Graves and Willems (2003); Klosterhalfen and Minner (2010); Li (2013); Li and Wu (2018).
To the best of our knowledge, the work presented in Klosterhalfen and Minner (2010) is the only paper that compares
the performance of the GSM to the SSM for a 1-warehouse 2-retailer distribution network. The simulation study presented
by Klosterhalfen and Minner shows that the cost difference between the two approaches is quite small, amounting to 4% at
most.
We want to verify whether this relatively low-cost difference between both approaches holds when inventory is con-
trolled through an (R,Q) installation ordering policy with continuous review instead of an echelon base stock policy (S-1,S)
with a common review period for a similar distribution network.
In Klosterhalfen and Minner (2010), it is also assumed that the lead times in the distribution network are all deterministic,
which is a major simplification of actuals business situations. Verstraete has confirmed the conclusions from Minner (2000)
and has shown in Verstraete (2010) that the safety stock costs can easily triple for the GSM when lead times are considered
stochastic instead of deterministic. For this reason, we will also assess the impact of the stochasticity of the lead time on the
total cost for a distribution system.
For this analysis, we will start by deriving a mathematical formulation for the GSM with an (R,Q) installation ordering
policy and continuous review based on the model from Li and Chen (2015). This model will consequently be compared to
the performance of an SSM with normal approximations for the retailer replenishment lead times and an (R,Q) installation
policy with continuous review inspired by Desmet, Aghezzaf, and Vanmaele (2010a).
 International Journal of Production Research 4151

3. Mathematical model formulations

3.1. Guaranteed-service model
The GSM originally formulated by Simpson Jr (1958) assumed a base stock (S-1,S) installation policy with a common
review period. The GSM formulation was extended by Li (2013) to optimise continuous review (R, Q) echelon policies for
serial, assembly and distribution networks. For a detailed discussion on the model presented below, we refer the reader to
Li (2013); Li and Chen (2015), we will however introduce and reuse some of the notations from this work.
The model from Li and Chen (2015) assumes there is only external (bounded) demand at the nodes which are furthest
downstream, and that the demand follows a Poisson distribution with mean λi . The internal demand at the warehouse is
solely generated by the retailers. The lead time is considered deterministic and is defined as μLi , which represents the time it
takes until the process at stage i is carried out, given that all components are available. Furthermore we will use αi to denote
the cycle service level, and βi to represent the fill rate.
The lot size here is defined as Qi and is not linked to the lead time, mi is an integer multiplier that enforces the integer
ratio constraint among the lot sizes, to put it differently, the lot size of the warehouse has to be an integer multiple of the lot
size of the downstream retailers. E[·] denotes the expected value operator.
Symbol Iie denotes the echelon on-hand inventory at stage i, which is equal to the sum of the physical inventory at stage
i and its downstream stages, and all physical inventory that is in-transit between these locations.
We also have the outgoing service time Si , the time that is quoted to the downstream stage in which we guarantee to
deliver, the incoming service time SIi , which is the time a stage has to wait until it receives the orders from the upstream
stage, and si that denotes the time frame during which the customer has to be serviced.
Index 0 represents the warehouse, N denotes the number of retailers. For the costs we respectively use ci to express the
fixed order cost, hei for the echelon holding cost and pi for the operating flexibility cost.
It is worth highlighting that the fixed order costs are applied to the average demand that is normally fulfilled λi βi ,
therefore the average fixed order cost for stage i is equal to ci λi βi /Qi .
Similarly, on average we will have to resort to extraordinary measures for 1 − βi percent of the demand, the cost for this
(operating) flexibility per unit is denoted by pi , the total flexibility cost thus becomes pi λi (1 − βi ).
The model defined as Pgsm aims to minimise three types of costs at the same time: inventory holding costs, fixed order
costs and operating flexibility costs that have to be incurred when demand exceeds a certain bound.
N  
ci λi βi
Pgsm minimise + hei · E[Iie ] + pi λi (1 − βi ) (1)
i=0
Qi

subject to Q0 = mi Qi for i = 1, . . . , N (2)

Si ≤ si for i = 1, . . . , N (3)

Si ≤ μLi + SIi for i = 0, . . . , N (4)

S0 ≤ SIi for i = 1, . . . , N (5)

Qi , mi ≥ 0 and integer for i = 1, . . . , N (6)

Si , SIi ≥ 0 and integer for i = 0, . . . , N (7)

Constraint 2 models the integer ratio restriction, which demands that the lot size at the warehouse is an integer multiple of
the lot sizes at the retailers. In Constraint 3, we impose that the retailers should be serviced within si , the service time that
was quoted to or is expected by the retailers.
Constraint 4 ensures that the net replenishment lead time, which is equal to SIi + μLi − Si , is positive for each stage.
Recognise that it would not make sense to have a negative net replenishment lead time (given that the lead time μLi is
deterministic), as this would mean that the outgoing service time Si is larger than the sum of the incoming service time
SIi and lead time μLi , in which case the upstream stage would hold on to a shipment (that is already complete) for the
downstream stage deliberately, and thus increase its own inventory holding costs unnecessarily. Constraint 5 is included
to make sure that the incoming service time at the retailers SIi is at least as large as the outgoing service time S0 that was
quoted to the retailers by the warehouse. Finally, the lot sizes, integers multiplier and different service times are assumed to
be integers.
In what follows, we derive a novel GSM that more adequately models real-world supply chains and can be compared to
a more complex SSM, inspired by the work of Graves and Willems (2000); Humair et al. (2013); Li and Chen (2015).
4152 N. De Smet et al.

3.2. Extended GSM

3.2.1. External demand
High volume articles typically have the largest impact on supply chain costs and are thus the most frequently selected
candidates for inventory optimisation. As the demand for fast movers often comes from many independent customers, we
can assume via the central limit theorem that the external demand is independent and normally distributed, and modify the
GSM accordingly.

3.2.2. Demand bound

We remind the reader that the GSM assumes that demand is bounded, and that extraordinary measures will be used for
demand that is outside of a pre-specified limit. In this section, we provide insight on how this demand bound is calculated
by characterising the maximum demand that can be expected during a certain time period.
If we define the maximum reasonable lead time demand at stage i over τ units of time as Di (τ ) and let t denote a time
index, we have
Pr [di [t − τ , t) ≤ Di (τ )] ≥ αi . (8)
In other words, the probability that the demand during τ time units is smaller than the demand bound, should be greater than
or equal to the service level αi . From this one can easily recognise that
√
Di (τ ) ≥ τ μdi + −1 (αi )σdi τ for i = 1, . . . , N, (9)

where −1 is the inverse of the cumulative

 distribution of the standard normal distribution. Graves and Willems (2000)
have shown that setting D0 (τ ) = Ni=1 Di (τ ) overestimates the maximum reasonable lead time demand at the warehouse
because it ignores risk pooling at the retailer level, therefore we define D0 (τ ) as
√
D0 (τ ) ≥ τ μd0 + −1 (α0 )σd0 τ (10)
 
with μd0 = Ni=1 μdi and σd20 = Ni=1 σd2i .

3.2.3. Impact of lead time stochasticity on the demand during the net replenishment lead time
We can further extend the former inequalities by including the stochasticity of the lead time in the demand bound, which
will lead to increased (safety) stock. We will follow the approach presented in Humair et al. (2013) to integrate this lead
time variability into Pgsm . We introduce the following expressions for the lead time at stage i with a continuous distribution
 Ti
M (Ti ) =
1
fLi (l)dl, (11)
l=0

 ∞
M 2 (Ti ) = lfLi (l)dl, (12)
l=Ti

 ∞
M (Ti ) =
3
l2 fLi (l)dl, (13)
l=Ti

where Ti = max(Si − SIi , 0) and fLi denotes the probability density function of the lead time at stage i. Let the expected
value of the positive part of the net replenishment lead time be

J(Ti ) = M 2 (Ti ) − Ti (1 − M 1 (Ti )). (14)

The variance of the positive part of the net replenishment lead time is defined as

K(Ti ) = Ti2 M 1 (Ti )(1 − M 1 (Ti )) − 2Ti M 1 (Ti )M 2 (Ti ) + M 3 (Ti ) − (M 2 (Ti ))2 . (15)

Based on these expressions, Humair et al. (2013) show that the demand during the net replenishment lead time Dnrlt ∼
N (μDnrlt , σD2nrlt ) can be characterised as follows:
μDnrlt
i
= μdi J(Ti ) (16)
 International Journal of Production Research 4153

σDnrlt
i
= J(Ti )σd2i + μ2di K(Ti ) (17)
The form of this expression closely relates to the single-echelon safety stock equation that was mentioned in the introduction
of this paper. Note that the average and standard deviation of the demand during the positive part of the net replenishment
lead time will always be higher compared to when lead times are deterministic and no distinction is made between the
positive and negative parts of the net replenishment lead time.
Finally, if τ is equal to expected value of the positive net replenishment lead time of stage i, we have τi = J(Ti ) and the
new demand bound becomes
Di (J(Ti )) ≥ μDnrlt
i
+ −1 (αi )σDnrlt
i
for i = 0, ..., N. (18)

3.2.4. Early arrival stock

Humair et al. (2013) separate the positive part of the net replenishment lead time from the negative part. If we allow the net
replenishment lead time to go negative in the model, we have to account for stock that is ready to be shipped in advance of
the actual shipping date to the customer. This kind of stock is called the early arrival stock and is defined as follows:
early 
Ii = μdi J(Ti ) − μLi + Ti . (19)
With a slight abuse of notation, we recognise from Equation (14) that J(0) = μLi , hence there will only be early arrival
stock when the outgoing service Si is larger than the incoming service time SIi .

3.2.5. From echelon to installation inventory

The objective function from Pgsm minimises the echelon inventory and thus assumes that the upstream stages take into
account the inventory levels downstream for ordering decisions. The Pssm , however, does not assume such visibility, there-
fore we have to alter the objective function of Pgsm to make sure that it minimises the installation stock instead of the echelon
inventory, that way each stage can take ordering decisions only based on local information.
Lemma A.1 proves that the expected on hand inventory at stage i can be characterised as follows:


Di (J(Ti )) − μdi βi J(Ti ) + μdi J(Ti ) − μLi + Ti + Qi2−1 for i = 1, . . . , N,
E[Ii ] =   (20)
D0 (J(T0 )) − μd0 β0 J(T0 ) + μd0 J(T0 ) − μL0 + T0 + Ni=0 Qi2−1 for i = 0.

The echelon holding cost hei can be transformed into the installation holding cost hi by the following equation:

hi = hei + hj (21)
j∈P(i)

with P(i) the set of immediate predecessors of stage i.

3.2.6. Service times

Similar to Klosterhalfen and Minner (2010), we will use the common assumption that the supplier of the warehouse has
ample stock SI0 = 0, and that the customer expects to be delivered immediately, si = 0, for i = 1, . . . , N. Since the lead
times are now variable, we no longer require the net replenishment lead time to be positive and drop the Si ≤ μLi + SIi
constraint.

3.2.7. Model description

If we apply the modifications presented in the previous sections to the Pgsm , we arrive at the following mathematical
formulation:
 N  
ci μdi βi 
ext
Pgsm minimise + hi Di (J(Ti )) − μdi βi J(Ti ) + μdi J(Ti ) − μLi
i=0
Qi

Qi − 1  N
Qi − 1
+ Ti + + pi μdi (1 − βi ) + h0 (22)
2 i=1
2
4154 N. De Smet et al.

subject to Q0 = mi Qi for i = 1, . . . , N (23)

SI0 = 0 (24)

Si = 0 for i = 1, . . . , N (25)

S0 ≤ SIi for i = 1, . . . , N (26)

Qi , mi ≥ 0 and integer for i = 1, . . . , N (27)

Si , SIi ≥ 0 and integer for i = 0, . . . , N (28)

This model can be further decomposed into the Q-problem and the R-problem, as suggested by Li (2013), and shown in
Appendix A.

3.2.8. Optimisation procedure

Although the mathematical formulation of the Q-problem and R-problem has been adjusted, we can still follow an opti-
misation procedure similar to BETA procedure described in Li (2013) if we modify the calculation of the real fill rate for
the different stages in the network. The adjusted algorithm goes as follows (note that we assume that αi for i = 0, . . . , N is
known)
(1) For i = 0, . . . , N set βi = αi
(2) Solve the Q- and R-problems to get Qi and Ri for each stage
(3) Calculate the real fill rate βi∗ for each stage i given the (Ri , Qi ) policy

(4) If Ni=0 |βi∗ − βi | ≤  with ε a very small positive number, stop. Otherwise set βi = βi∗ for i = 0, . . . , N and go
back to 1.
For the calculation of the real fill rate, we introduce the following formula based on Axsäter (2015):

1 
βi∗ = 1 − G(Ri , μDnrlt , σ nrlt ) − G(Ri + Qi , μ nrlt , σ nrlt )
Di Di Di (29)
Qi i

with the help function G

 
G(r, μDnrlt , σDnrlt ) = (μDnrlt − r) 1 − (r, μDnrlt , σD2nrlt ) + σD2nrlt φ(r, μDnrlt , σD2nrlt ) (30)

where Dnrlt represents the demand during the net replenishment lead time and φ refers to the probability density function of
the normal distribution.

3.3. Stochastic service model

Recall that the SSM does not put a bound on the demand, hence service times are no longer deterministic. This implies that
the SSM can only allocate more inventory to cope with demand uncertainty. In the case of a stock out at an upstream stage,
the replenishment lead time of the downstream stage will be stochastic even when the processing time at the downstream
stage itself is deterministic. The complexity of the mathematical formulation and computation of this replenishment lead
time increases together with the size and complexity of the supply chain considered.
External demand at the most downstream stages is assumed to be normally distributed di ∼ N (μdi , σd2i ), the demand at

the warehouse which consists of the retailer orders is assumed to be known and normally distributed with μd0 = Ni=1 μdi

and σd20 = Ni=1 σd2i .
The stages follow an installation (R,Q) inventory control policy. The orders that are received in the warehouse are
fulfilled on an FCFS basis. If there is insufficient inventory in the warehouse to satisfy the request of the retailer, the
entire order is backordered, and thus the retailer has to wait until enough inventory is replenished again in the warehouse.
This introduces a delay in the lead time from the retailer’s perspective and warrants the definition of what we will call the
 International Journal of Production Research 4155

‘actual’ retailer replenishment lead time that accounts for this potential delay due to stock-outs in the warehouse. In Desmet,
Aghezzaf, and Vanmaele (2010a), this actual retailer replenishment lead time is defined as follows:
Li ∼ N (μL0i + μS0B (1 − β0 ), σL20 + σS2B (1 − β0 )) (31)
i 0

where Li represents the actual lead time for stages i for i = 1, . . . , N, L0i ∼ N (μL0i , σL20 ) is the nominal lead time assuming
i
sufficient goods are available in the warehouse and μSB0 characterises the exponential approximation that is used for the
service times for backorders in the warehouse, we refer the reader to Desmet, Aghezzaf, and Vanmaele (2010a) for more
details.
The reorder point Ri of a stage i ∈ N is defined as (analogous to a single-echelon setting)

Ri = μdi μLi + ki μLi σd2i + μ2di σL2i = μDddlt
i
+ ki σDddlt
i
= μDddlti
+ SSi (32)

with Dddlt representing the demand during lead time for which Dddlt ∼ N (μDddlt , σD2ddlt ). The safety factor ki can be found for
a given βi via Equation (30). The total safety stock (SStot ) can be calculated by
 
ssm
SStot = ki μLi σd2i + μ2di σL2i . (33)
i∈N

Desmet, Aghezzaf and Vanmaele also present a formulation for the system inventory optimisation problem that aims to
minimise the total average stochastic inventory


N N 

Pssm minimise E[Ii ] = H(Ri + Qi , Qi , μDddlt
i
, σDddlt
i
)
i=0 i=0

− H(Ri , Qi , μDddlt
i
, σDddlt
i
) (34)

where H is a help function, specified in Appendix A.

3.4. Extended SSM

3.4.1. Lot sizing
Q,ext ext
The lot size is considered given in Pssm , we will relax this assumption and let Pgsm propose the lot sizes to use in Pssm based
on the fill rate, ordering cost, average demand and holding costs.

3.4.2. Ordering, holding and flexibility costs

Model Pssm optimises the expected installation inventory. We will compare the performance of the two models based on the
ordering, holding and flexibility costs, therefore we introduce the ordering cost ci , holding cost hi and flexibility cost pi .
Analogous to the GSM, we incur ordering costs for the demand that will be fulfilled without backordering, μdi βi .
Although we assume that excessive demand will be backordered in the SSM, it is reasonable to state that the costs
associated with the demand that will be backordered on average, μdi (1 − βi ), are linearly increasing with the number of
units backordered and comparable to the operating flexibility costs in the GSM, hence the reuse of notation pi .

3.4.3. Model description

ext
Taking into account the modifications described earlier, we present Pssm
N  
ci μdi βi
ext
Pssm minimise + hi · E[Ii ] + pi μdi (1 − βi ) (35)
i=0
Qi
Q,ext
where Qi is found through Pgsm , and the expected installation inventory at stage i is determined by Equation (34).
Observe that this is an unconstrained optimisation problem due to the introduction of the actual retailer replenishment
Q,ext
lead time concept, and because the lot size is proposed by an Economic Order Quantity (EOQ) like calculation in Pgsm that
accounts for the integer ratio constraint.
4156 N. De Smet et al.

3.4.4. Optimisation procedure

Model Pssmext
assumes that βi for i = 1, . . . , N is given. The fill rate of the warehouse β0 is thus the decision variable
ext
that the model tries to optimise. There is no strict optimisation procedure for Pssm suggested by Desmet, Aghezzaf, and
Vanmaele (2010a) to come to a solution we perform a one-dimensional grid search to find β0 that minimises the total cost
in the network.

3.5. Summary of model extensions in relation to other literature

In this section, we have integrated and extended the work of various authors in our models. For the GSM, we zoom in on
the following.
Based on the insights from Graves and Willems (2000), we were able to tighten the demand bound at the warehouse to
allow for risk pooling at the retailer level.
The lead time variability (and early arrival stock) was included in the model via the work of Humair et al. (2013), which
has a significant impact on the required safety stock in network.
We have derived an expression for the expected on hand inventory when only local information is available, in other
words, we extended the model of Li and Chen (2015) from echelon (R,Q) policies to installation (R,Q) inventory control
policies, which we believe is the second most important contribution of this paper, aside from the insights from the numerical
study.
Less extensions were necessary to the SSM to come to a ‘fair’ comparison of both models, as its formulation already
satisfied most of our requirements.

4. Numerical study
4.1. Design
4.1.1. Network
We compare the performance of the GSM and SSM for a 1-warehouse 2-retailer system, similar to Klosterhalfen and
Minner (2010). The parameters of the distribution system that will be used for the numerical study are given in Table 1,
note that the retailers are assumed to be identical. Excluding the service levels, we simulate 64 different parameter sets for
the network.

4.1.2. Service levels

 
We assume that αi for i = 1, . . . , N is given and that α0 = Ni=1 μdi αi / Ni=1 μdi , similar to Li (2013). The adjusted BETA
procedure described in Section 3.2.8 then allows us to find the real fill rates βi∗ for the network. After six iterations all

Table 1. Parameters numerical study 2-echelon distribution system with 1 warehouse and 2
identical retailers.
Description Notation Values
Number of retailers nrt 2
Cycle service level retailers αi [0.50, 0.99]
Fill rate retailers βi [0.50, 0.99]
Holding cost warehouse h0 10
Holding cost retailers hi 15, 20
Ordering cost warehouse c0 300
Ordering cost retailers ci 30
Backorder/flexibility cost retailers and warehouse pi 25, 50
Average demand retailers μdi 100
σdi
Coefficient of variation demand retailers μdi 0.20, 0.40
Nominal lead time warehouse μL0 5, 30
σL00
0
Coefficient of variation nominal lead time warehouse μL0 0.20, 0.40
0
Nominal lead time retailers μL0 1
σL0i
i
Coefficient of variation nominal lead time retailers μL0 0.20, 0.40
i
 International Journal of Production Research 4157

simulations reached the terminal state, α0 varies between 50% and 99%, we thus solve 3200 instances with the GSM or
19,200 if you include the iterations.
Model Pssm assumes that βi for i = 1, . . . , N is known. For Pssm
ext
, we use the same assumption and vary β0 between 50%
and 99% to find the fill rate at the warehouse that minimises the total inventory costs for each fill rate at the retailer. Through
the inversion of Equation (30), we can find the safety factor ki that is equal to −1 (αi ), from which αi can be derived. In
total, we solved 160,000 instances of the problem with the SSM.
Despite the different formulations and solving techniques, each model thus gives us the α and β service levels of all
stages in the network.

4.1.3. Computational environment

This numerical study was carried out using the Stevin Supercomputer Infrastructure at Ghent University on a cluster with
2 × 10-core Intel E5-2660v3 (2.6 GHz) processors and 128 GB of memory per node. Both models were implemented in
Python 3.6, a walltime of 2 hours and 4 GB of memory on (in total) 16 cores was sufficient to carry out the experiments.
Each specific problem instance can be solved in less than a minute via either of the two approaches.

4.2. Results
We study the performance difference between the GSM and SSM for six different parameters: holding costs at the retailers,
flexibility costs, demand variability at the retailers, nominal lead time at the warehouse, lead time variability at the warehouse
and lead time variability at the retailers. Figures A1–A3 show the numerical results for the cycle service level, Figures A4–
A6 show similar plots for the fill rate.
Per type of service level, each figure shows the same 3200 data points. A single data point represents one cost comparison
between the best solution of the GSM and SSM, for one of the 64 parameter sets, and one of the 50 different retailer service
levels (from 50% to 99%, in increments of 1%). By colouring the data points with a respectively low (red) or high (blue)
value for the parameter studied, we can assess whether for example a low or high holding cost at the retailers shifts the cost
improvement of one model over the other up- or downwards. Through visual inspection of the plots, we conclude that none
of the aforementioned parameters seem to have a meaningful impact on the performance difference between the models.
However, both the cycle service level and fill rate figures show that the retailer service level is an important driver for the
cost improvement, for lower service levels the GSM consistently outperforms the SSM, for the higher service levels we see
that the SSM starts dominating the GSM.
In Table 2, we calculate descriptive statistics for the entire service level range 50–99%, and the higher service levels
in the 85–99% interval. For the full service level range, we find that the GSM leads to a better solution in more than 80%
of the cases and had a total cost advantage of 10% on average, for the higher service level range the GSM is still superior,
although the difference is less significant.
We would also like to draw attention to the fact that there is a significant spread on the cost difference, as shown by the
min. and max. improvements of −32% and 28% respectively for the cycle service level, and −42% and 26% for the fill
rate.

Table 2. Percentage total cost improvement GSM compared to SSM: cycle service level
and fill rate.
Cycle service level Fill rate
50–99%
No. of instances with positive improvement 2893 (90.46%) 2706 (84.56%)
Mean percentage improvement 10.66% 9.30%
Mean absolute percentage improvement 12.07% 13.36%
Max. improvement 27.90% 26.17%
Min. improvement − 31.60% − 41.68%
85 – 99%
No. of instances with positive improvement 653 (68.16%) 781 (81.35%)
Mean percentage improvement 2.73% 6.66%
Mean absolute percentage improvement 7.73% 8.72%
Max. improvement 27.90% 24.13%
Min. improvement − 31.60% − 23.03%
4158 N. De Smet et al.

4.3. Discussion
Our numerical study verifies, extends and also challenges some of the conclusions from Klosterhalfen and Minner (2010),
who suggested that the GSM benefits from moderate flexibility costs, large warehouse processing times, high retailer service
levels, and that the cost difference with the SSM is at most 4%. To explain this apparent discrepancy with the results from
this study, we start by providing more details on the assumptions of the latter paper.
Klosterhalfen and Minner explain that the GSM has two ways of dealing with variability, inventory and extraordinary
measures, which is why its cost superiority decreases with respect to the SSM, as the additional cost for such measures
increases.
The internal service level in the GSM they study is only dependent on the flexibility and holding costs, while the internal
service level of their SSM does increase if the lead time at the warehouse is longer, which results in more inventory (holding
costs).
High retailer service levels lead to a better solution for the GSM because the internal service level is not affected by the
external service levels, on the other hand, the SSM takes this into account and will redistribute the inventory between the
warehouse and retailers accordingly.
Returning to this study, based on Figures A1–A6, we can conclude that the cost advantage of our models is not dependent
on the flexibility cost, warehouse processing time or any of the other four parameters that we have examined in our study,
but only on the service level at the retailers.
This means that we have succeeded in aligning the various assumptions of both models, and have derived two mathemat-
ical formulations for which the cost difference between the GSM and SSM is not (significantly) influenced by the external
characteristics of the network.
In what follows, we clarify how the service level at the retailers does have an impact on the results, and how this is
linked to the real driver of the cost difference, which is the service level at the warehouse.
The optimal solution found by our SSM for distribution networks typically allocates most of the safety stock to the
retailers, while only keeping little inventory in the warehouse. For small multi-echelon networks with relatively long lead
times between the warehouse and retailers (and thus minimal opportunity for risk pooling) and high retailer service levels, it
is intuitive to keep most of the inventory as close to the customer as possible, and thus set the service level at the warehouse
low. The apparent suboptimal performance of the SSM can thus not be attributed to the way the internal service levels are
determined, it can however be explained by the exponential approximation of the backorder service time, which has been
shown in Desmet, Aghezzaf, and Vanmaele (2010a) to be conservative, the resulting inventory control policy could thus be
seen an upper bound to achieve the desired retailer service levels, hence the larger total cost of the solution.
The GSM on the other hand assumes that the internal service level is an input, dependent on the holding and flexibility
cost at the warehouse, or a demand weighted average of the external service levels. In the latter case, which is an assumption
that is common (Li and Chen 2015) and was therefore used in our model, this means that the internal service level will be
equal to external service level for identical retailers, which in turn will lead to a high (unnecessary) safety stock at the
warehouse for high retailer service levels, and explains the decreasing performance of the model in this service level zone.
From Table 2, we learned that the min. and max. values of the cost advantage for the GSM/SSM can be much larger
than the 4% that was found in Klosterhalfen and Minner (2010). On top of that, for both the cycle service level and fill rate,
the GSM outperformed the SSM in more than 80% cases, and realised an average cost improvement of approximately 10%.
These results suggest that care should be taken with regard to the selection of multi-echelon inventory model approaches
when some of the basic assumptions regarding lead times and inventory control policy are adjusted/relaxed to allow for more
realistic assumptions that better model real-world supply chains, as the cost difference between competing approaches can
be very significant depending on the (characteristics of the) network or model that is considered.

5. Conclusion and future research

In this work, we presented two mathematical formulations for inventory optimisation in distribution networks that account
for lot sizing decisions through an installation (R, Q) inventory control policy, stochasticity of lead times, ordering, holding
and flexibility costs. To the best of our knowledge, the adjusted mathematical formulation of the GSM we developed, and
use in our numerical study, is the first guaranteed-service approach that is capable of optimising installation (R, Q) policies
in distribution networks with stochastic lead times.
The numerical study in which we compared the performance of the aforementioned models on total costs shows that
low retailer (and therefore also warehouse) service levels increase the cost advantage of the GSM over the SSM. Lead times
at the warehouse, holding costs at the retailers, flexibility costs, and the variability of the demand and lead times did not
affect the performance of the two competing models.
 International Journal of Production Research 4159

Based on both the cycle service level and fill rate comparison, we have shown that the GSM outperforms the SSM in
more than 80% of the cases and realises an average percentage improvement of over 10%, with cost differences up to 42%.
This challenges the conclusions from Klosterhalfen and Minner (2010), who found that the cost difference between the
guaranteed service and stochastic service is not more than 4% for a similar 1-warehouse 2-retailer network.
Our results show that care must be taken when deciding on a modelling approach when it must be applied to real-
world supply chain networks where some of the basic assumptions from the GSM or SSM, for example regarding lead time
variability or inventory control policies, are not valid. We hope that this work inspires more researchers to study how the
performance of the GSM and SSM differs.
Ample further research opportunities can be identified from this work. The lot sizing problem, for example, has been
separated from the safety stock problem in this work, joint optimisation of the reorder points and lot sizes would lead to a
better solution for both models.
In practice, service levels are typically set to more than 85%, in this service level range the SSM might still prove to
be the better pick when lot sizes and safety stocks will be jointly optimised. However, additional research is necessary to
replace the grid search by a formal optimisation procedure to make the model more tractable. Also, further refinement of
the backorder service time approximation will help decrease the conservativeness of the SSM.
It would be interesting to assess how the tweaking of the internal service level α0 in the GSM affects its performance,
and how this can be translated into rules for setting the internal service levels in various supply chain configurations.
Only 1-warehouse 2-retailer networks have been studied, numerical and simulation studies on more complex
distribution, assembly, spanning tree and generic systems will help to further generalise the conclusion.

Funding
This work was supported by the Flanders Innovation and Entrepreneurship Agency [grant HBC.2017.0561] and Solventure NV.

Disclosure statement
No potential conflict of interest was reported by the authors.

ORCID
Niels De Smet http://orcid.org/0000-0002-3261-1530
El-Houssaine Aghezzaf http://orcid.org/0000-0003-3849-2218

References
Axsäter Sven. 2003. “Supply Chain Operations: Serial and Distribution Inventory Systems.” Handbooks in Operations Research and
Management Science 11: 525–559.
Axsäter Sven. 2015. Inventory Control. Vol. 225. http://link.springer.com/10.1007/978-3-319-15729-0.
Clark Andrew J, and Scarf Herbert. 1960. “Optimal Policies for a Multi-Echelon Inventory Problem.” Management Science 6 (4): 475–
490.
Crowston Wallace B, Wagner Michael, and Williams Jack F.. 1973. “Economic Lot Size Determination in Multi-Stage Assembly
Systems.” Management Science 19 (5): 517–527.
de Kok, Ton, and Jan Fransoo. 2003. “Planning Supply Chain Operations: Definition and Comparison of Planning Concepts.” Handbooks
in Operations Research and Management Science 11: 597–675.
de Kok, Ton, Christopher Grob, Marco Laumanns, Stefan Minner, Jörg Rambau, and Konrad Schade. 2018. “A Typology and Literature
Review on Stochastic Multi-Echelon Inventory Models.” European Journal of Operational Research 269 (3): 955–983.
de Kok Ton, Janssen Fred, Doremalen Jan Van, Wachem Erik Van, Clerkx Mathieu, and Peeters Winfried. 2005. “Philips Electronics
Synchronizes its Supply Chain to End the Bullwhip Effect.” Interfaces35 (1): 37–48.
de Kok, Ton, and Jeremy Visschers. 1999. “Analysis of Assembly Systems with Service Level Constraints.” International Journal of
Production Economics 59 (1–3): 313–326.
Desmet Bram. 2009. “Normal Approximations for Safety Stock Optimization in Multi-echelon Supply Chains.” .
Desmet Bram, Aghezzaf El Houssaine, and Vanmaele Hendrik. 2010a. “A Normal Approximation Model for Safety Stock Optimization
in a Two-Echelon Distribution System.” Journal of the Operational Research Society 61 (1): 156–163.
Desmet Bram, Aghezzaf El-Houssaine, and Vanmaele Hendrik. 2010b. “Safety Stock Optimisation in Two-Echelon Assembly Systems:
Normal Approximation Models.” International Journal of Production Research 48 (19): 5767–5781.
Deuermeyer, Bryan L, and Leroy B Schwarz. 1981. A Model for the Analysis of System Service Level in Warehouse-Retailer Distribution
Systems: The Identical Retailer Case. West Lafayette: Purdue University.
4160 N. De Smet et al.

Diks E. B., and De Kok A. G.. 1999. “Computational Results for the Control of a Divergent N-Echelon Inventory System.” International
Journal of Production Economics 59 (1–3): 327–336.
Diks E. B., De Kok A. G., and Lagodimos A. G.. 1996. “Multi-Echelon Systems: A Service Measure Perspective.” European Journal of
Operational Research 95 (2): 241–263.
Eruguz Ayse Sena, Sahin Evren, Jemai Zied, and Dallery Yves. 2016. “A Comprehensive Survey of Guaranteed-Service Models for Multi-
Echelon Inventory Optimization.” International Journal of Production Economics 172: 110–125. doi:10.1016/j.ijpe.2015.11.017.
Ettl Markus, Feigin Gerald E, Lin Grace Y, and Yao David D. 2000. “A Supply Network Model With Base-Stock Control and Service
Requirements.” Operations Research 48 (2): 216–232.
Graves Stephen C, and Willems Sean P. 2000. “Optimizing Strategic Safety Stock Placement in Supply Chains.” Manufacturing & Service
Operations Management 2 (1): 68–83.
Graves Stephen C, and Willems Sean P. 2003. “Supply Chain Design: Safety Stock Placement and Supply Chain Configuration.”
Handbooks in Operations Research and Management Science 11: 95–132.
Hadley G, and Whitin T. M.. 1961. “A Family of Inventory Models.” Management Science 7 (4): 351–371.
Humair Salal, Ruark John D, Tomlin Brian, and Willems Sean P. 2013. “Incorporating Stochastic Lead Times into the Guaranteed Service
Model of Safety Stock Optimization.” Interfaces 43 (5): 421–434.
Inderfurth Karl. 1993. “Valuation of leadtime reduction in multi-stage production systems.” In Operations Research in Production
Planning and Control, 413–427. Springer.
Khan Mehmood, Hussain Matloub, and Cárdenas-Barrón Leopoldo Eduardo. 2017. “Learning and Screening Errors in an EPQ Inventory
Model for Supply Chains with Stochastic Lead Time Demands.” International Journal of Production Research 55 (16): 4816–4832.
Klosterhalfen, Steffen, and Stefan Minner. 2010. “Safety Stock Optimisation in Distribution Systems: A Comparison of Two Competing
Approaches.” International Journal of Logistics Research and Applications 13 (2): 99–120.
Lee Hau L, and Billington Corey. 1993. “Material Management in Decentralized Supply Chains.” Operations Research 41 (5): 835–847.
Li Peng. 2013. “Optimization of (R, Q) Policies for Multi-echelon Inventory Systems with Guaranteed Service.” PhD diss., Troyes.
Li Peng, and Chen Haoxun. 2015. “Optimization of (R; Q) Policies for Two-level Distribution Inventory Systems with Operating
Flexibility.” In IEEE 12th International Conference on Networking, Sensing and Control (ICNSC), 603–606. IEEE.
Li Peng, and Wu Di. 2018. “A Comparison Study of Stochastic-and Guaranteed-Service Approaches on Safety Stock Optimization for
Multi Serial Systems.” In IOP Conference Series: Earth and Environmental Science, Vol. 108, 042070. IOP Publishing.
Lin Grace, Ettl Markus, Buckley Steve, Bagchi Sugato, Yao David D, Naccarato Bret L, Allan Rob, Kim Kerry, and Koenig Lisa. 2000.
“Extended-Enterprise Supply-Chain Management at IBM Personal Systems Group and Other Divisions.” Interfaces 30 (1): 7–25.
McGavin Edward J, Schwarz Leroy B, and Ward James E. 1993. “Two-Interval Inventory-Allocation Policies in a One-Warehouse
N-Identical-Retailer Distribution System.” Management Science 39 (9): 1092–1107.
McGavin Edward J, Ward James E, and Schwarz Leroy B. 1997. “Balancing Retailer Inventories.” Operations Research 45 (6): 820–830.
Minner, Stefan. 2000. “Strategic Safety Stocks in Supply Chains, volume 490 of Lecture Notes in Economics and Mathematical Systems.”
Minoux, Michel. 2018. “Robust and Stochastic Multistage Optimisation Under Markovian Uncertainty with Applications to Produc-
tion/Inventory Problems.” International Journal of Production Research 56 (1–2): 565–583.
Schwarz Leroy B. 1989. “A Model for Assessing the Value of Warehouse Risk-Pooling: Risk-Pooling Over Outside-Supplier Leadtimes.”
Management Science 35 (7): 828–842.
Schwarz Leroy B, Deuermeyer Bryan L., and Badinelli Ralph D.. 1985. “Fill-Rate Optimization in a One-Warehouse N-Identical Retailer
Distribution System.” Management Science 31 (4): 488–498.
Simchi-Levi, David, and Yao Zhao. 2012. “Performance Evaluation of Stochastic Multi-echelon Inventory Systems: A Survey.” Advances
in Operations Research 2012. doi:10.1155/2012/126254
Simpson Jr Kenneth F. 1958. “In-Process Inventories.” Operations Research 6 (6): 863–873.
Van Houtum Geert-Jan. 2006. “Multi-Echelon Production/Inventory Systems: Optimal Policies, Heuristics, and Algorithms.” Tutorials in
Operations Research INFORMS 2006: 163–199.
Verstraete Nick. 2010. “Optimalisatie van veiligheidsvoorraden in multi-echelon toeleverketens met stochastische vraag en
doorlooptijden.”
Walker Edward. 2009. “The Real Cost of a CPU Hour.” Computer 42 (4): 35–41.

Appendices
Appendix 1. Expected on-hand inventory for a distribution system with guaranteed-service, variable lead times and
an installation (R, Q) ordering policy
The expected on-hand inventory for a distribution system with guaranteed service, variable lead times and an installation (R, Q) ordering
policy is equal to


Di (J(Ti )) − μdi βi J(Ti ) + μdi J(Ti ) − μLi + Ti + Qi2−1 for i = 1, . . . , N,
E[Ii ] =   Qi −1 (A1)
D0 (J(T0 )) − μd0 β0 J(T0 ) + μd0 J(T0 ) − μL0 + T0 + N i=0 2 for i = .
 International Journal of Production Research 4161

Proof If one starts from the inventory balance equation with ILei (t) defined as the echelon inventory level of stage i at time t and IPie (t)
as the echelon inventory position.
ILei (t) = IPie (t − Li ) − di [t − L, t) for i = 0, . . . , N. (A2)
The GSM assumes that all demand within the demand bound can be satisfied, thus we have ILei (t) = Iie (t) and the equation becomes
Iie (t) = IPie (t − Li ) − di [t − L, t) for i = 0, . . . , N. (A3)
For the retailers (stages most downstream), the echelon on-hand inventory is equal to the installation on-hand inventory
Ii (t) = Iie (t) for i = 1, . . . , N. (A4)
We are minimising the long run average inventory holding cost, the above equation thus implies
E[Ii ] = E[Iie ] for i = 1, . . . , N. (A5)
Using the echelon interpretation, we can express the installation on-hand inventory at the warehouse as follows:


N
I0 (t) = I0e (t) − IPie (t). (A6)
i=1

For the expected value, we have


N
E[I0 ] = E[I0e ] − E[IPie ]. (A7)
i=1
From Hadley and Whitin (1961), we know that

1 
Qi
1 + Qi
E[IPie ] = (Ri + j) = Ri + for i = 0, . . . , N. (A8)
Qi 2
j=1

Li derives the following expression for the expected value of E[Iie ]:

D (NRLTi ) − μdi βi NRLTi + Qi2−1 for i = 1, ..., N,

E[Iie ] = i N Q0 −1 N (A9)
i=0 [Di (NRLT i )] − μd0
β 0 NRLT 0 + 2 + i=1 Qi − N for i = 0.

note that we replace λi by μdi because we assume normally distributed demand. Combining Equations (A.7), (A.8) and (A.9) leads to

 N  
Q0 − 1  
N N
1 + Qi
E[I0 ] = [Di (NRLTi )] − μd0 β0 NRLT0 + + Qi − N − Ri +
2 2
i=0 i=1 i=1
N 
  
N
Qi − 1
= Di (NRLTi ) + − μd0 β0 NRLT0 − N − Ri (A10)
2
i=0 i=1

This can be further simplified by using the reorder point equation from Axsäter (2015)
Ri = Di (NRLTi ) − 1 for i = 1, ..., N. (A11)
Substituting Ri in Equation (A.10) gives


N
Qi − 1
E[I0 ] = D0 (NRLT0 ) − μd0 β0 NRLT0 + . (A12)
2
i=0

We can now describe the expected on-hand inventory as

Di (NRLTi ) − μdi βi NRLTi + Qi2−1 for i = 1, ..., N,

E[Ii ] =  Qi −1 (A13)
D0 (NRLT0 ) − μd0 β0 NRLT0 + N i=0 2 for i = 0.

To account for the stochasticity of the lead time, we recognise that only the positive part of the net replenishment lead time should be
considered for the calculation of the safety stock, and that we have to introduce an expression to account for the early arrival stock. If we
replace NRLTi by J(Ti ) and include early arrival stock, we prove that


Di (J(Ti )) − μdi βi J(Ti ) + μdi J(Ti ) − μLi + Ti + Qi2−1 for i = 1, ..., N,
E[Ii ] =   Qi −1 (A14)
D0 (J(T0 )) − μd0 β0 J(T0 ) + μd0 J(T0 ) − μL0 + T0 + N i=0 2 for i = 0.


4162 N. De Smet et al.

Appendix 2. Decomposition GSM into Q- and R-problem

ext , we recognise that the constant term −(1/2)[N h + Nh ] − N [h μ μ − p μ (1 − β )] can be excluded from the
From Pgsm i=0 i 0 i=0 i di Li i di i
objective functions of the Q-problem and R-problem. We can formulate the Q-problem as follows:
N 
 
Q,ext ci μdi βi Qi
Pgsm minimise + Hi (A15)
Qi 2
i=0

subject to Q0 = mi Qi for i = 1, . . . , N (A16)

Qi , mi ≥ 0 and integer for i = 1, . . . , N (A17)

by introducing Hi

h0 for i = 0
Hi = (A18)
h0 + hi for i = 1, . . . , N
such that it can be solved by a variation of the Crowston–Wagner algorithm that was initially introduced in Crowston, Wagner, and
R,ext
Williams (1973) for assembly systems. The R-problem Pgsm becomes


N  
R,ext
Pgsm minimise hi Di (J(Ti )) − μdi βi J(Ti ) + μdi (J(Ti ) + Ti ) (A19)
i=0

subject to SI0 = 0 (A20)

Si = 0 for i = 1, . . . , N (A21)
S0 ≤ SIi for i = 1, . . . , N (A22)

Si , SIi ≥ 0 and integer for i = 0, . . . , N. (A23)

Appendix 3. Help function SSM

From Desmet, Aghezzaf, and Vanmaele (2010a), we know that the help function for the calculation of the average stochastic inventory is
defined as follows:

1
H(r, Q, μDddlt , σDddlt ) = (r2 − (σD2 ddlt + μ2Dddlt ))(r, μDddlt , σD2 ddlt )
2Q

+ σD2 ddlt (r + μDddlt )φ(r, μDddlt , σD2 ddlt )

μDddlt
− (r − μDddlt )(r, μDddlt , σD2 ddlt )
Q

+ σD2 ddlt φ(r, μDddlt , σD2 ddlt )

σD2 ddlt
+ (r, μDddlt , σD2 ddlt ) (A24)
Q
 International Journal of Production Research 4163

Appendix 4. Performance comparison GSM and SSM based on cycle service level

Figure A1. Performance comparison GSM and SSM based on cycle service level: impact holding cost retailers and flexibility costs.

Figure A2. Performance comparison GSM and SSM based on cycle service level: impact demand variability retailers and nominal lead
time warehouse.
4164 N. De Smet et al.

Figure A3. Performance comparison GSM and SSM based on cycle service level: impact lead time variability warehouse and retailers.

Appendix 5. Performance comparison GSM and SSM based on fill rate

Figure A4. Performance comparison GSM and SSM based on fill rate: impact holding cost retailers and flexibility costs.
 International Journal of Production Research 4165

Figure A5. Performance comparison GSM and SSM based on fill rate: impact demand variability retailers and nominal lead time
warehouse.

Figure A6. Performance comparison GSM and SSM based on fill rate: impact lead time variability warehouse and retailers.
Copyright of International Journal of Production Research is the property of Taylor & Francis
Ltd and its content may not be copied or emailed to multiple sites or posted to a listserv
without the copyright holder's express written permission. However, users may print,
download, or email articles for individual use.