EFFICIENT MULTI-ECHELON INVENTORY

OPTIMIZATION MODELS FOR INDUSTRIAL

PRACTICE

V. G. Achkar1,2*, B. B. Brunaud3 and I. E. Grossmann1

1
Carnegie Mellon University, 2Universidad Nacional del Litoral, 3Johnson & Johnson
1
Pittsburgh, PA 15213, 2Santa Fe, Argentina, 3New Jersey, NJ 08901

Abstract
In this work, we propose a generalized Multi-Echelon Inventory Optimization (MEIO) model based on
the Guaranteed-Service approach for allocating safety stocks across the network, seeking to meet customer
service levels at minimum holding costs. This is especially challenging in multi-echelon supply chains,
which face supply and demand uncertainty, and the decisions on one echelon impact the other echelons.
The proposed model offers a more realistic representation of real-world supply chains, accounting for
several features such as non-normal demand distributions, specified time periods between reviews,
Minimum Order Quantities (MOQ), and different service level performance indicators (fill rate and cycle
service levels). To improve the computational efficiency and make the model effective to use in industrial
practice, we propose to reformulate the nonlinear programming model as a Mixed-Integer Quadratically
Constrained Program (MIQCP) reformulation, a stepwise function and a piecewise linear approximation.
This solution strategy leads to order of magnitude reductions in computational time. The performance of
the model is evaluated by solving several instances of a real-world case study.

Keywords
Multi-echelon inventory optimization, safety stock, non-normal demand, QCP, piecewise approximation.

Introduction
Multi-echelon inventory optimization (MEIO) seeks to The present paper relies on the Guaranteed-Service
optimize safety stock setting for the entire supply chain with Model (GSM) to optimize safety stocks (Graves & Willems,
centralized coordination (You & Grossmann, 2009). MEIO 2000; Simpson, 1958). The objective is to develop an MEIO
eliminates over-buffering of inventory, and builds model to address the problem of safety stocks in real-world
confidence in product availability. De Kok et al. (2019) complex systems. Managing a real-world supply chain
present a general typology and review of stochastic MEIO gives rise to two important challenges: (i) the model must
models in which they summarize the extensive research on represent industrial practice features and dynamics, and (ii)
multi-echelon inventory management. The authors state the optimal solution must be found with modest
that multi-echelon inventory systems are still a very active computational time. Over the years, many authors have
area of research because of their complexity and practical worked on extending the original GSM assumptions to
relevance. Recently, Gonçalves et al. (2020) present a enable real-world supply chain characteristics to be
systematic literature review, and they also highlight that the captured, as presented in the survey by Eruguz et al. (2016).
number of contributions to MEIO has seen a significant However, the GSM for multi-echelon supply chains
increase from the year 2005 onwards, and they list many proposed in the literature does not fully account for all the
potential directions and trends for future research. issues and characteristics arising in industrial practice. We

* vachkar@fiq.unl.edu.ar
extend the work of Achkar et al. (2022) to account for a
novel approach to include stochastic lead times, and we
specially focus on new solution strategies. In this work, the
following characteristics are included: (i) non-normally
distributed demands, (ii) manufacturing plants placed at any
location in the network, (iii) fill rate as an alternative key
customer service performance indicator apart from the
Cycle Service Level (CSL), and (iv) Minimum Order
Quantities (MOQ) for replenishment orders and the period
between reviews. Furthermore, this work presents solution
strategies to increase computational efficiency. From an
optimization perspective, inventory decisions in multi-
echelon systems are challenging because the problems are
nonlinear and nonconvex. We introduce a solution approach
that allows solving real-world size problems in modest
computational time. The NLP model is reformulated as an
MIQCP by exploiting the structure of the constraints of the
base model. In addition, the problem involves piecewise
and stepwise functions, reducing additional complexities.
Computational examples for a real industrial system are
presented to illustrate the application of the proposed model
and its resulting improved computational performance. Figure 1. Guaranteed-service model elements
The outline of the paper is as follows. The problem
statement is described first, followed by the model Demand is assumed to be bounded. If the demand in a
formulation. Next, an extension to account for nonnormal certain time period exceeds the bound, it is assumed that
distributions for the demand is described. The paper ends other extraordinary measures are used to satisfy the excess
with the application of the model to illustrate its application demand. In addition, each stage of the supply chain operates
to real-world case studies. Conclusions are drawn in the under a periodic review inventory policy with a basestock
final section. level. The demand is independent and identically
distributed at each demand node. The statistical
distributions of the demand are not necessarily normal.
Problem Statement
Similarly, lead times are variable but assumed to be
Given is a supply chain with a fixed design for normally distributed.
locations j ∈ J and a set of materials p ∈ P that can be either Demand is propagated upstream considering the risk
raw materials or finished goods. Holding costs are incurred pooling assumptions described in You and Grossmann
at all the nodes. If the customer places an order of size djp(t) (2009). At plant locations, there is a coefficient ϕpq that
on node j at time t, this order will be fulfilled by time t + represents the bill of materials for product transformation
Sjp, with Sjp being the guaranteed-service time of node j and depends on material-finished good relationships. We
(Graves & Willems, 2000). Moreover, each node j receives assume divergent networks: a node that holds a material p
a service commitment from its upstream node i ∈ J, called can only receive this material from a single node and can
inbound service time SIjp, and has an order processing time distribute it to one or more locations, as is usual in finished
or lead time of LTj. The Net Lead Time (NLT), as shown in goods supply chains. The same node can be supplied with
Figure 1, represents the time period that is not covered another material q ∈ P from another location, but this last
within the guaranteed service time, and must be covered one should be the only supplier of q for that node.
with safety stock, defined as NLTjp = SIjp + LTjp - Sjp. If Furthermore, a Minimum Order Quantity moqjp may be
NLTjp = 0, the node works under a make-to-order policy required for the replenishment orders. Each node makes its
without storing any inventory. On the other hand, if NLTjp own replenishment decisions and has no delay in ordering.
> 0, there is a time period that should be covered with safety For each node, there is a safety stock factor kjp related to the
stock. The objective is to determine the guaranteed-service Cycle Service Level (CSL), which reflects the percentage
times (SIjp and Sjp) for each material at each location, and of replenishment cycles with non-stocking out.
consequently how much safety stock to maintain to Alternatively, the modeler can also ask for a fill rate to be
minimize the total holding costs and satisfy a specified considered as a target service measure, which equals the
customer service level. The service times at the initial and proportion of the demand met over the total demand placed.
the final nodes are given. This option is assumed to be active only if demand is
normally distributed.
Model Formulation given a safety stock level for continuous review policies.
From this formula, we can derive Eq. (7) that links the fill
The following formulation involves four positive
rate (frjp) to the safety factor kjp, and consequently to the
continuous variables: the guaranteed-service time Sjp of a
CSL. The safety factor, in this case, becomes a positive
product on a given node, the inbound service SIjp, the safety
̂𝑗𝑝 related to CSL, and the Net Lead Time NLTjp. The continuous variable, and the aim is to find the lower
factor 𝐾
required CSL that is able to meet the desired fill rate. We
objective function is to minimize safety stock holding cost assume that Qjp is the average replenishment quantity of
as defined by Eq. (1), where hjp is the coefficient that product p on location j, with Qjp = max{µjp rjp, moqjp}for
represents holding cost for each material p at each location periodic-review policies. Fs(kjp) and fs(kjp) correspond to the
j, and the square root represents the “sigma-combination” cumulative and normal density distribution functions,
formula, known from the stochastic-service approach respectively.
(Clark & Scarf, 1960). We introduce stochastic lead times
into the GSM by extending the approach proposed by 2 2 2
√𝑁𝐿𝑇𝑗𝑝 𝜎𝑗𝑝 + 𝜇𝑗𝑝 𝜎𝑁𝐿𝑇𝑗𝑝
Minner (1998) because it faces all the variability sources 𝑓𝑟𝑗𝑝 ≤ 1 + ̂𝑗𝑝 [1 − 𝐹𝑠 (𝐾
(𝐾 ̂𝑗𝑝 )] − 𝑓𝑠 (𝐾
̂𝑗𝑝 ))
using safety stock. The mean demand and the standard 𝑄𝑗𝑝 (7)
deviation are represented by µjp and σjp, respectively. We ∀ 𝑗 ∈ 𝐽, 𝑝 ∈ 𝑃𝑗 , (𝑗, 𝑝) ∈ 𝐹
introduce a new parameter σNLTjp that refers to the NLT
variability instead of using the lead time variability. On the other hand, if the service level target is CSL, the
safety factor is given as an input kjp, as in Eq. (8):
̂𝑗𝑝 √𝑁𝐿𝑇𝑗𝑝 𝜎𝑗𝑝
𝑚𝑖𝑛 ∑ ∑ ℎ𝑗𝑝 𝐾 2 2
+ 𝜇𝑗𝑝 2
𝜎𝑁𝐿𝑇 𝑗𝑝 (1) ̂𝑗𝑝 = 𝑘𝑗𝑝
𝐾 ∀ 𝑗 ∈ 𝐽, 𝑝 ∈ 𝑃𝑗 , (𝑗, 𝑝) ∉ 𝐹 (8)
𝑝 𝑗 ∈ 𝐵𝑗𝑝

The first set of constraints comes from the classic

GSM. Equation (2) defines the first inbound service time Solution Approach
for the starting nodes in the network J0, where si0 is a given
The model (1)-(8) corresponds to a nonconvex
input. Equation (3) links the inbound guaranteed-service
nonlinear problem (NLP). These problems can be in
time SIjp and the guaranteed-service time of upstream nodes
principle be solved with nonlinear global optimization
Siq. A is a set with elements (i,j,p) indicating that there is an
solvers. However, the computational time required to find a
enabled route for material p from i to j. Note that q = p and
global solution with these solvers may be very expensive.
i ≠ j if it is a distribution link (i to j) of the same product p,
Eq. (7) includes the normal distribution density and
and q ≠ p and i = j if node j is a plant location that produces
cumulative functions. Moreover, Eqs. (1) and (7) include
p from q. The service time Sjp is bounded by Eq. (4), and
the value σNLTjp, whose calculation significantly affect the
Eq. (5) becomes active if there is a maximum accepted
computational burden. To overcome this difficulty, we
service time. The NLT is defined in Eq. (6), with ltjp being
propose three solution strategies: (i) an exact quadratically
the lead time and rjp the time period between reviews.
constrained reformulation, (ii) a stepwise function to obtain
0 the value of σNLTjp, and (iii) a piecewise linear
𝑆𝐼𝑗𝑝 = 𝑠𝑖𝑗𝑝 ∀ 𝑗 ∈ 𝐽0 , 𝑝 ∈ 𝑃𝑗 (2) approximation of Eq. (7) to replace normal density and
cumulative functions.
First, we reformulate the NLP as a MIQCP, that is
𝑆𝐼𝑗𝑝 ≥ 𝑆𝑖𝑞 ∀ (𝑖, 𝑗, 𝑝) ∈ 𝐴, (𝑞, 𝑝) ∈ Φ, 𝑝 ∈ 𝑃𝑗 (3) guaranteed to obtain the global optimum. Solvers like
CPLEX and Gurobi can solve MIQCPs quite effectively in
reasonable computational times and we avoid to use general
𝑆𝑗𝑝 ≤ 𝑆𝐼𝑗𝑝 + 𝑙𝑡𝑗𝑝 + 𝑟𝑗𝑝 ∀ 𝑗 ∈ 𝐽, 𝑝 ∈ 𝑃𝑗 (4) nonlinear solvers such as BARON. To account for this
reformulation, a new positive continuous variable, Ujp,
represents the square root, i.e. 𝑈𝑗𝑝 ≥ √𝜏𝑗𝑝 , as shown in
𝑆𝑗𝑝 ≤ 𝑚𝑎𝑥𝑆𝑗𝑝 ∀ 𝑗 ∈ 𝐽, 𝑝 ∈ 𝑃𝑗 (5) Figure 2, being τ the argument of the square root. We
2
include Eq. (9) to define 𝑈𝑗𝑝 :

𝑁𝐿𝑇𝑗𝑝 ≥ 𝑆𝐼𝑗𝑝 − 𝑆𝑗𝑝 + 𝑙𝑡𝑗𝑝 + 𝑟𝑗𝑝 ∀ 𝑗 ∈ 𝐽, 𝑝 ∈ 𝑃𝑗 (6) 2 2 2 2

𝑈𝑗𝑝 ≥ 𝑁𝐿𝑇𝑗𝑝 𝜎𝑗𝑝 + 𝜇𝑗𝑝 𝜎𝑁𝐿𝑇 𝑗𝑝
∀ 𝑗 ∈ 𝐽, 𝑝 ∈ 𝑃𝑗 (9)
As mentioned above, the GSM uses the CSL as the
Hence, the objective function becomes a linear function
customer service performance indicator when setting safety
given by Eq. (10):
stocks. As mentioned before, the fill rate is another measure
that represents the fraction of demand that was met on-time
from the inventory. We extend the GSM to allow specifying ̂𝑗𝑝 𝑈𝑗𝑝
𝑚𝑖𝑛 ∑ ∑ ℎ𝑗𝑝 𝐾 (10)
fill rates if desired (j,p ∈ F). Chopra and Meindl (2013) 𝑝 𝑗 ∈ 𝐵𝑗𝑝
propose a formula to obtain the corresponding fill rates
 𝑁𝐿𝑇𝑗𝑝 ≤ ∑ 𝑢𝑏𝑗𝑝𝑠 𝑉𝑗𝑝𝑠 ∀ 𝑗 ∈ 𝐽, 𝑝 ∈ 𝑃𝑗 (13)
𝑠∈𝑆1

∑ 𝑉𝑗𝑝𝑠 = 1 ∀ 𝑗 ∈ 𝐽, 𝑝 ∈ 𝑃𝑗 (14)
𝑠∈𝑆1

Figure 2. Reformulation from NLP to QCP

𝑋𝑗𝑝 = ∑ 𝑉𝑗𝑝𝑠 𝑐𝑗𝑝𝑠 ∀ 𝑗 ∈ 𝐽, 𝑝 ∈ 𝑃𝑗
In second place, we introduce a stepwise function to (15)
𝑠∈𝑆1
represent the NLT standard deviation, σNLTjp. We assume
that the variability of the lead time is pushed downstream if
Finally, the last solution strategy involves the fill rate
the NLT of node j for product p is not enough to cover the
constraint. We aim to improve the model tractability by
total replenishment time inherent to node j. In other words,
replacing ℎ(𝐾 ̂) = 𝐾̂ [1 − 𝐹𝑠 (𝐾
̂ )] − 𝑓𝑠 (𝐾
̂ ) with a piecewise
if NLTjp ≤ SIjp +ltjp + rjp, then, σLTjp should be propagated
downstream. We introduce a stepwise function as depicted linear approximation, as shown in Figure 4. This function
in Figure 3 to represent the value of σNLTjp, dependent of the will be the same for all nodes on all locations, while demand
value of NLTjp. is normally distributed. In the example there are four
Equation (9) is replaced by Eq. (11), with Xjp being a breakpoints and three segments. If the number of
positive continuous variable that represents σNLTjp. breakpoints increases, the precision of the estimation also
does, but the computational efficiency decrease. For the
2
𝑈𝑗𝑝 2
≥ 𝑁𝐿𝑇𝑗𝑝 𝜎𝑗𝑝 2
+ 𝜇𝑗𝑝 𝑋𝑗𝑝 ∀ 𝑗 ∈ 𝐽, 𝑝 ∈ 𝑃𝑗 (11) current formulation we propose 7 breakpoints.
0.00
-0.05
-0.10
-0.15
-0.20
h(K)

-0.25
-0.30
-0.35
f(z)
-0.40 Piecewise approximation
-0.45
0 0.5 1 1.5 2 2.5 3
K

Figure 4. Piecewise linear approximation

Figure 3. Stepwise function to define σNLTjp Variable Yjp is added to the model to represent the
piecewise linear function as in Eq. (16). The value of 𝐾 ̂𝑗𝑝
The lower and upper bounds for each step s ∈ S1 are
defines which segment s ∈ S2 is active, and it forces the
defined as lbjps and ubjps, respectively. For example, for the
first step (s1) of product p at node j, lbjps1 = 0 and ubjps1 = ltjp activation of the binary variable Wjps and the continuous
̃𝑗𝑝𝑠 as in Eqs. (17), (18). Equations (19) and (20)
variable 𝐾
+ rjp. As NLTjp increases, the bounds increase by adding the
lead times and the review periods of upstream nodes. The force only one segment to be active.
new binary variable Vjps defines what the active step in Eqs.
(12) and (13) is according to the value of NLTjp. Equation ̃𝑗𝑝𝑠 + 𝑏𝑗𝑝𝑠 𝑊𝑗𝑝𝑠 )
𝑌𝑗𝑝 = ∑ (𝑎𝑗𝑝𝑠 𝐾 ∀ 𝑗 ∈ 𝐽, 𝑝 ∈ 𝑃𝑗 (16)
(14) ensures that only one step is active for each product on 𝑠∈𝑆2
each location. Eq. (15) assigns the value of σNLTjp associated
with the step s, given by the parameter cjps, which represents ̃𝑗𝑝𝑠
𝑙𝑏𝑗𝑝𝑠 𝑊𝑗𝑝𝑠 ≤ 𝐾 ∀ 𝑗 ∈ 𝐽, 𝑝 ∈ 𝑃𝑗 , 𝑠 ∈ 𝑆2 (17)
the sum of lead time variances according to Figure 3. In the
2 2 2
example, cjps1 = 0, cjps2 = 𝜎𝐿𝑇 𝑗𝑝
, and cjps3 = 𝜎𝐿𝑇 𝑗𝑝
+ 𝜎𝐿𝑇 𝑖𝑝
. It
is worth to mention that this coefficient also includes the ̃𝑗𝑝𝑠 ≤ 𝑢𝑏𝑗𝑝𝑠 𝑊𝑗𝑝𝑠
𝐾 ∀ 𝑗 ∈ 𝐽, 𝑝 ∈ 𝑃𝑗 , 𝑠 ∈ 𝑆2 (18)
lead time variance of raw materials if they are involved in
the production of product p.

∑ 𝑊𝑗𝑝𝑠 = 1 ∀ 𝑗 ∈ 𝐽, 𝑝 ∈ 𝑃𝑗 (19)
∑ 𝑙𝑏𝑗𝑝𝑠 𝑉𝑗𝑝𝑠 ≤ 𝑁𝐿𝑇𝑗𝑝 ∀ 𝑗 ∈ 𝐽, 𝑝 ∈ 𝑃𝑗 (12) 𝑠∈𝑆2
𝑠∈𝑆1
 and the one obtained using the empirical distribution, as
̃𝑗𝑝𝑠 = 𝐾
∑𝐾 ̂𝑗𝑝 ∀ 𝑗 ∈ 𝐽, 𝑝 ∈ 𝑃𝑗 stated in Eq. (22):
(20)
𝑠∈𝑆2 −1
ℎ𝑗𝑝 (𝐶𝑆𝐿𝑗𝑝 ) − 𝜇𝑗𝑝
𝑘̂𝑗𝑝 = 𝑚𝑎𝑥 {𝛷 −1 (𝐶𝑆𝐿𝑗𝑝 ), } (22)
Finally, Eq. (7) is redefined including the new variables 𝜎𝑗𝑝
related to the piecewise and stepwise functions:
Table 2 presents the results of using Eq. (22). All CSLs
𝑈𝑗𝑝 are achieved or surpassed. Note that this extension only
𝑓𝑟𝑗𝑝 ≤ 1 + 𝑌 ∀ 𝑗 ∈ 𝐽, 𝑝 ∈ 𝑃𝑗 , (𝑗, 𝑝) ∈ 𝐹 (21) applies when the target service level is the CSL, and the
𝑄𝑗𝑝 𝑗𝑝
main disadvantage is that it can over-buffer (as in Table 2,
The new mathematical reformulation is a MIQCP row 80%). Future work will extend it for fill rate targets.
given by Equations (2)-(6), (8), (10)-(20).
Table 1. Effective CSL for large CVs with
original and adapted safety factors
Extension for non-normal demand
In industrial practice, it is frequently found that demand CV = 0.56 CV = 2 CV = 5
data histograms do not fit the shape of a normal distribution. Expected
Normal Empirical
This is generally detected when the coefficient of variation CSL N E N E
(N) (E)
(CV) increases. For distributions with large CV, the model 50% 55% 55% 68% 68% 68% 68%
predicts a slightly lower CSL than expected when targets 60% 64% 58% 75% 68% 75% 68%
are large, as shown in Figure 5. The plot presents the results 70% 73% 68% 81% 68% 81% 68%
of several simulations in Excel for 10,000 periods, 80% 81% 79% 86% 75% 86% 75%
assuming a deterministic lead time equal to 1, period 90% 89% 90% 91% 88% 91% 88%
between reviews equal to 1, and normally distributed 96% 95% 96% 94% 96% 94% 96%
demand datasets with different CVs. 98% 97% 98% 95% 98% 95% 98%

Table 2. Effective CSL for proposed safety factor

Expected CSL CV=0.56 CV=2 CV=5

50% 55% 68% 68%
60% 64% 75% 75%
70% 73% 81% 81%
80% 81% 86% 86%
90% 90% 91% 91%
96% 96% 96% 96%
98% 96% 98% 98%

Application
In order to illustrate the application of the proposed
solution strategy, we carry out a computational experiment
Figure 5. Expected vs Effective CSL based on a real-world case study with 800 products and 18
locations, as shown in Figure 6. The numbers next to each
In this work, we aim to find an alternative to adapt the node refer to the number of products that can be stored in
model to set safety stocks for cases that are able to achieve that location, because not all products follow the same
at least the desired customer service level. Mirzaee (2017) route. Nodes with people icons mean that those nodes
proposes an alternative way to set safety stocks by adjusting receive external demand. The demand is different for each
the service level, using the equivalent k safety factor that is product, all of them are all stochastic, independent and
equal to the empirical cumulative distribution function identically distributed. Lead times are normally distributed.
(h(x)) value. We simulate several datasets of gamma- Figure 7 displays the computational time required to
distributed demands. In Table 1 we present the results solve different instances ranging from 100 to 800 products.
obtained for CSL using the k safety factors obtained from All the instances are modeled with Pyomo and solved with
the normal distribution, and the k equivalent from the Gurobi 9.5.0, on an Intel Core i7 16 MB RAM. The same
empirical distribution function h(x). Note that the empirical instances were tested using both targets: the orange line
correlation fits better than the normal for large CSL but it corresponds to the problem with CSL targets of 97% for all
does poorly for low values. We propose a new approach that products, and the blue line represents the results fixing a fill
selects the maximum value between the classic safety factor rate (FR) target of 98%. Note that real-world cases can be
solved to optimality within few seconds of CPU time. The model that brings together an MIQCP reformulation with
computational burden increase when the fill rate is the target piecewise linear functions to improve the computational
measure, because more constraints and variables become efficiency of this optimization model. Future work will
active. Table 3 shows the detailed model sizes and the address an extension by including responsive characteristics
objective values for some instances. It is worth to mention to account for supply chain disruptions and by including
that using BARON to solve the smallest instance for both storage capacity limitations.
the NLP and the QCP formulations could not yield a
feasible solution within 1000 seconds. This demonstrates
Acknowledgments
that the proposed MIQCP reformulation solving with
Gurobi yields order of magnitude increases in efficiency. The authors gratefully acknowledge the financial
support from Johnson and Johnson and the Center for
Advanced Process Decision-making (CAPD) from
Carnegie Mellon University.

References
Achkar, V. G., Brunaud, B. B., Pérez, H. D., Musa, R., Méndez,
C. A., & Grossmann, I. E. (2022). Extensions to the
Guaranteed Service Model for Industrial Applications of
Multi-Echelon Inventory Optimization. Submitted for
Publication.
Chopra, S., & Meindl, P. (2013). Supply Chain Management. In
Pearson.
Figure 6. Case-study network http://www.doiserbia.nb.rs/Article.aspx?ID=0013-
32640670067A
Clark, A. J., & Scarf, H. (1960). Optimal Policies for a Multi-
Echelon Inventory Problem. Management Science,
50(12_supplement), 1782–1790.
https://doi.org/10.1287/mnsc.1040.0265
de Kok, T. (2019). Inventory Management: Modelling Real-life
Supply Chains and Empirical Validity. Foundations and
Trends in Technology, Information and Operations
Management, 12(4), 349–433.
http://www.nowpublishers.com/article/Details/TOM-057
Figure 7. Instance size vs. computational time Eruguz, A. S., Sahin, E., Jemai, Z., & Dallery, Y. (2016). A
comprehensive survey of guaranteed-service models for
multi-echelon inventory optimization. International
Table 3. Model sizes and optimal solutions Journal of Production Economics, 172, 110–125.
https://doi.org/10.1016/j.ijpe.2015.11.017
Continuous Integer Total cost Gonçalves, J. N. C., Sameiro Carvalho, M., & Cortez, P. (2020).
Items Constraints
variables variables ($) Operations research models and methods for safety stock
200 40 502 22 581 30 171 2 652 585 determination: A review. Operations Research
400 78 284 43 499 58 422 5 655 908 Perspectives, 7(April), 100164.
600 111 471 63 139 84 507 6 814 902 https://doi.org/10.1016/j.orp.2020.100164
800 153 123 86 819 115 877 13 738 148 Graves, S. C., & Willems, S. P. (2000). Optimizing Strategic
Safety Stock Placement in Supply Chains. Manufacturing
Conclusion & Service Operations Management, 2(1), 68–83.
https://doi.org/10.1287/msom.2.1.68.23267
The proposed model integrates several features Minner, S. (1998). Strategic Safety Stocks in Supply Chains.
commonly found in industrial practice that have a strong Mirzaee, A. (2017). Alternative Methods for Calculating.
impact on inventory levels. Moreover, the model with the University of Missouri-Columbia.
proposed solution strategy outperforms the original NLP Simpson, K. F. . (1958). In-Process Inventories. INFORMS, 6(6),
formulation, by finding optimal solutions at minimal 863–873.
computational expense for large scale problems. Therefore, You, F., & Grossmann, I. E. (2009). Integrated multi-echelon
the proposed tool is able to accurately represent real supply chain design with inventories under uncertainty:
systems, and to set tight safety stock levels in order to MINLP models, computational strategies. AIChE Journal,
achieve target customer service levels with minimum 59(4), 420–440. https://doi.org/10.1002/aic.12010
inventory cost. To the best of our knowledge, this is the first