This article was downloaded by: [212.175.193.

57] On: 06 April 2016, At: 04:35

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA

Management Science
Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org

The Impact of Modeling on Robust Inventory Management

Under Demand Uncertainty
Oğuz Solyalı, Jean-François Cordeau, Gilbert Laporte

To cite this article:

Oğuz Solyalı, Jean-François Cordeau, Gilbert Laporte (2016) The Impact of Modeling on Robust Inventory Management Under
Demand Uncertainty. Management Science 62(4):1188-1201. http://dx.doi.org/10.1287/mnsc.2015.2183

Full terms and conditions of use: http://pubsonline.informs.org/page/terms-and-conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact permissions@informs.org.

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.

Copyright © 2015, INFORMS

Please scroll down for article—it is on subsequent pages

INFORMS is the largest professional society in the world for professionals in the fields of operations research, management
science, and analytics.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
 MANAGEMENT SCIENCE
Vol. 62, No. 4, April 2016, pp. 1188–1201
ISSN 0025-1909 (print) — ISSN 1526-5501 (online) http://dx.doi.org/10.1287/mnsc.2015.2183
© 2016 INFORMS

The Impact of Modeling on Robust Inventory

Management Under Demand Uncertainty
Downloaded from informs.org by [212.175.193.57] on 06 April 2016, at 04:35 . For personal use only, all rights reserved.

Oğuz Solyalı
Business Administration Program, Middle East Technical University, Northern Cyprus Campus, Kalkanlı, Mersin 10, Turkey,
solyali@metu.edu.tr

Jean-François Cordeau, Gilbert Laporte

Interuniversity Research Centre on Enterprise Networks, Logistics and Transportation (CIRRELT),
Montreal, Quebec H3C 3J7, Canada; and HEC Montréal, Montréal, Quebec H3T 2A7, Canada
{jean-francois.cordeau@hec.ca, gilbert.laporte@cirrelt.ca}

T his study considers a basic inventory management problem with nonzero fixed order costs under interval
demand uncertainty. The existing robust formulations obtained by applying well-known robust optimization
methodologies become computationally intractable for large problem instances due to the presence of binary
variables. This study resolves this intractability issue by proposing a new robust formulation that is shown to
be solvable in polynomial time when the initial inventory is zero or negative. Because of the computational
efficiency of the new robust formulation, it is implemented on a folding-horizon basis, leading to a new heuristic
for the problem. The computational results reveal that the new heuristic is not only superior to the other
formulations regarding the computing time needed, but also outperforms the existing robust formulations in
terms of the actual cost savings on the larger instances. They also show that the actual cost savings yielded by
the new heuristic are close to a lower bound on the optimal expected cost.
Data, as supplemental material, are available at http://dx.doi.org/10.1287/mnsc.2015.2183.
Keywords: robust optimization; inventory management; lot sizing; integer programming
History: Received September 26, 2010; accepted February 2, 2015, by Dimitris Bertsimas, optimization.
Published online in Articles in Advance October 14, 2015.

1. Introduction problems with uncertain parameters taking their

Robust optimization (RO) has recently emerged as a
powerful approach to optimization problems involv-
ing uncertain parameters with ambiguous probabil-
ity distributions. It incorporates uncertain parameters
into the optimization models in a tractable way, does
not require full knowledge of the probability distri-
bution of uncertain parameters, and finds the best
solution considering all possible realizations of uncer-
tain data while ensuring feasibility regardless of their
realization.
Several modeling frameworks are available within
the RO paradigm because uncertainty can be han-
dled in a number of ways. Since RO has proved to be
an effective approach to optimization problems with
uncertain parameters, the aim of this study is not to
compare RO approaches against dynamic program-
ming and stochastic programming, but rather to study
the effect of modeling in robust optimization on the
practical performance (i.e., actual cost savings and
computing time) yielded by the robust formulations.
Our analysis will be performed on a basic inventory
management problem using different RO approaches.
The work of Soyster (1973) was the first robust opti-
mization study to address linear programming (LP)
1188
worst-case values in a given uncertainty set. Later,
Ben-Tal and Nemirovski (1998, 1999, 2000), El-Ghaoui
and Lebret (1997), and El-Ghaoui et al. (1998) consid-
ered ellipsoidal uncertainty sets for uncertain convex
optimization problems (e.g., second-order cone pro-
gramming (SOCP) and semidefinite programming
(SDP) problems). However, this approach does not
easily extend to discrete optimization problems be-
cause solution methods for SOCP and SDP problems
with integer variables are not well developed. To fill
this gap, besides controlling the degree of conserva-
tiveness, Bertsimas and Sim (2004) proposed the bud-
get of uncertainty approach, which has the desirable
property that the robust counterpart of the uncertain
problem preserves the complexity of its nominal (i.e.,
without uncertainty) problem. Thus, this approach
naturally extends to discrete optimization problems
(see Bertsimas and Sim 2003) with only additional
variables and constraints in the robust counterpart
of an uncertain problem. Bertsimas and Sim (2003)
showed that 0–1 discrete optimization problems with
uncertain objective coefficients can be tackled by solv-
ing m + 1 nominal problems, where m is the num-
ber of binary variables. For polynomially solvable 0–1
 Solyalı, Cordeau, and Laporte: Impact of Modeling on Robust Inventory Management
Management Science 62(4), pp. 1188–1201, © 2016 INFORMS
discrete optimization problems (e.g., the shortest path
problem or the matching problem), this means their
robust counterparts are also polynomially solvable.
Atamtürk (2006) proposed strong robust formulations
for mixed 0–1 programming problems under objective
coefficient uncertainty. Instead of solving m + 1 nomi-
nal problems as in Bertsimas and Sim (2003), he devel-
Downloaded from informs.org by [212.175.193.57] on 06 April 2016, at 04:35 . For personal use only, all rights reserved.
oped a tight robust formulation for the polynomially
solvable 0–1 discrete optimization problems yielding
integral solutions. Kouvelis and Yu (1997) considered
a robust decision-making framework for discrete opti-
mization problems in which the realization of uncer-
tain data was represented by a set of scenarios, and
the aim was to minimize either the worst-case perfor-
mance or the maximum regret under the given set of
scenarios. However, in this robustness approach, the
robust counterparts of many polynomially solvable
0–1 discrete optimization problems become NP -hard
problems.
All of the above-mentioned solution frameworks
were developed for single-stage decision-making prob-
lems, ignoring the fact that some decisions can be
made after observing the realization of some uncertain
parameters in multistage decision-making problems.
Considering this fact, Ben-Tal et al. (2004) developed
the adjustable robust counterpart (ARC) for uncer-
tain multistage LP problems in which some decision
variables, called nonadjustable, are determined a pri-
ori (i.e., at time 0), whereas others, called adjustable,
are set after having observed the realization of some
uncertain parameters. Since the ARC of an uncertain
LP problem is intractable except for some restricted
cases, Ben-Tal et al. (2004) proposed, as a tractable
approximation to the ARC, the affinely adjustable
robust counterpart (AARC), in which adjustable vari-
ables are expressed as an affine function of real-
ized uncertain parameters. Recently, Chen and Zhang
(2009) improved the AARC by proposing the so-called
extended AARC (EAARC). However, the EAARC
of an uncertain LP problem is an SOCP problem,
so it does not fit well when integrality restrictions
exist. For network flow and design problems under
demand uncertainty, Atamtürk and Zhang (2007)
proposed a two-stage robust discrete optimization
approach in which the first-stage decisions are made
a priori, whereas the second-stage decisions are made
after observing the demand. Unlike Ben-Tal et al.
(2004), Atamtürk and Zhang (2007) did not restrict the
second-stage decision variables to be affine functions
of uncertain demand.
Robust inventory management problems under
demand uncertainty, being intrinsically multistage
problems due to their multiperiod planning hori-
zons, have been solved by many researchers using
the robust optimization approaches just described.
Ben-Tal et al. (2004, 2005) formulated two different
1189
inventory management problems as uncertain LP
problems and considered their AARCs, which give
better objective function values than their pure robust
counterparts. Bertsimas and Thiele (2006) modeled
a single-installation and a multi-installation inven-
tory management problem with a tree network struc-
ture as mixed integer linear programming problems
(resp., linear programming problems) in the presence
(resp., absence) of fixed order costs using the robust-
ness approach developed by Bertsimas and Sim (2004).
Bienstock and Özbay (2008) solved the same single-
installation inventory management problem without
fixed order costs as Bertsimas and Thiele (2006), in
addition to the problem with base-stock policy, using
a Benders-like two-phase decomposition algorithm.
Ben-Tal et al. (2009) modeled a multi-installation
inventory management problem with a serial structure
as an uncertain LP problem and considered its global-
ized robust counterpart, which is an extension of the
AARC. Unlike the above-mentioned studies, See and
Sim (2010) considered a single-installation inventory
management problem with nonzero order lead times
and characterized uncertain demand by covariance
and directional deviations besides the usual mean
and support information. They considered static, lin-
ear, and truncated linear replenishment policies and
formulated the problem as an SOCP model, which
approximates the objective function with good upper
bounds (UBs). Note that all these papers, except
Bertsimas and Thiele (2006), consider zero fixed order
costs (i.e., there are no integer decision variables).
In this study, we consider the same basic single-instal-
lation inventory management problem as Bertsimas
and Thiele (2006) with nonzero fixed order costs,
where the aim is to determine when and how much
to order under interval demand uncertainty over a
planning horizon of T time periods, so that the worst-
case total cost across all feasible demand scenar-
ios is minimized. Note that this problem is called
the uncapacitated lot-sizing problem with backlog-
ging (ULSB) in the literature when the demand is
known (see Pochet and Wolsey 2006, Chap. 10). We
introduce a new modeling approach to robust inven-
tory management by making use of a well-known
reformulation of the ULSB. Aiming to study the
impact of modeling in robust inventory management
under demand uncertainty, with and without bud-
get of uncertainty, we present several robust formu-
lations constructed by using the main known robust
optimization approaches (i.e., pure robust, budget
of uncertainty robust, and affinely adjustable robust
counterparts, as well as robust counterpart with an
approximate objective function) and compare these
with the new modeling approach.
 Solyalı, Cordeau, and Laporte: Impact of Modeling on Robust Inventory Management
1190
The contributions of this study are as follows:
• Whereas, to the best of our knowledge, all stud-
ies in the robust inventory management literature use
a variant of standard inventory flow balance equa-
tions in their robust formulations, we consider a
reformulation approach and propose a new robust
formulation with budget of uncertainty.
Downloaded from informs.org by [212.175.193.57] on 06 April 2016, at 04:35 . For personal use only, all rights reserved.
• In contrast to the existing robust formulations
obtained by applying prominent RO methodologies,
the proposed robust formulation is computationally
efficient even for large problem instances. Specifically,
we show that the new robust formulation with a bud-
get of uncertainty can be solved in polynomial time
by solving O4T 2 ) LP problems when the initial inven-
tory is zero or negative. We further propose another
reformulation-based robust formulation, which we
prove to be polynomially solvable in O4T 4 5 time.
Because of the computational efficiency of the new
robust formulation, we propose a folding-horizon
implementation of it that results in a new heuristic for
the problem.
• We adapt to our problem the robust formula-
tion for the linear replenishment policy of See and
Sim (2010), which works with an approximate objec-
tive function and results in an SOCP formulation with
binary variables. We computationally test this formu-
lation for the first time in the RO literature.
• The computational results on test instances
reveal that the new heuristic not only outperforms the
others in terms of computing time needed, but also
exhibits a superior performance in terms of the actual
cost savings compared to pure, budget of uncer-
tainty, and affinely adjustable robust formulations.
Our approach also outperforms the robust formula-
tion with an approximate objective function on the
larger instances. The computational results also show
that the actual cost savings yielded by the new heuris-
tic are close to a lower bound (LB) on the optimal
expected cost.
• Our computational results provide insights into
the performance of different RO formulations.
The remainder of this paper is organized as fol-
lows. We give the detailed problem description in §2.
We present a summary of the existing robust formu-
lations in §3, and the new robust formulation and
resulting folding-horizon heuristic in §4. We present
the computational results obtained by all formula-
tions and heuristics on test instances and conclude the
paper in §§5 and 6, respectively.
2. Problem Description: The Robust
Single-Installation Inventory
Management Problem
To show the impact of modeling in robust inventory
management, we chose the single-installation inven-
tory management problem, which is a basic building
Management Science 62(4), pp. 1188–1201, © 2016 INFORMS
block of more complex inventory management prob-
lems. The single-installation (with a single item) in-
ventory management problem can be described as
follows.
Over a finite horizon, the installation faces dynamic
stochastic demand dk in each period k ∈ T = 8110001T 9.
The probability distribution of the random variable dk
is unknown. The only information available concern-
ing dk is that it can take a value from the interval
[d¯k − dˆk 1 d¯k + dˆk ], where d¯k is the point estimate (nom-
inal value) and dˆk is the maximum deviation for the
demand during period k. The installation can place
an order uk to its supplier at the beginning of each
period k to fulfill some of the demand. Whenever an
order is placed, a variable order cost ck for each unit
ordered and a fixed order cost fk independent of the
order size are incurred. It is assumed, without loss of
generality, that the orders arrive instantaneously (i.e.,
the lead time is zero). The inventory level Ik+1 at the
beginning of period k + 1 is equal to the inventory
level Ik at the beginning of period k, plus the ordered
amount uk , minus the demand dk in period k. If Ik +
uk > dk , then the excess amount is held in inventory
with a holding cost hk for each unit held in inventory
at the end of period k. If Ik + uk < dk , then the excess
demand is backlogged (i.e., it is not met on time) and
is either satisfied when the stock becomes available or
never satisfied. Each unit backlogged at the end of a
period incurs a unit backlogging cost pk . We assume
that pk > ck so that it is likely to be more economical to
order stock until the last period. We also assume that
all parameters are nonnegative. Note that the inven-
tory level Ik+1 at the beginning of period k + 1 can be
rewritten as
k
X
Ik+1 = I1 + 4ut − dt 5 k ∈ T1 (1)
t=1
where I1 is the known initial inventory level. Equa-
tion (1) is the standard equation used in the robust
inventory management literature to determine the
inventory levels and the inventory holding and back-
logging costs.
There are no capacity limits on the order sizes that
can be placed to the supplier, and no limits on the
amount of inventory that can be carried. The problem,
referred to as the robust single-installation inventory
management problem (RSP), is to determine order
quantities and their timing. The objective is min-max,
i.e., minimizing the worst-case total cost composed of
inventory holding and backlogging as well as fixed
and variable order costs across all feasible demand
scenarios.
 Solyalı, Cordeau, and Laporte: Impact of Modeling on Robust Inventory Management
Management Science 62(4), pp. 1188–1201, © 2016 INFORMS 1191

3. Robust Formulations: Summary of Because (2) is a semi-infinite optimization prob-

Existing Methods lem that is intractable, we equivalently reformulate it
In this section, we present first the pure and the by ensuring feasibility for any possible realization of
budget of uncertainty robust formulations for the demands in the given uncertainty set as follows:
RSP, and then the affinely adjustable robust formula-   k
X

tion and the robust formulation with an approximate yk ≥ max hk I1 + 4ut − dt 5 1 k ∈ T1 (3)
d∈Dk
objective function for the same problem. t=1
Downloaded from informs.org by [212.175.193.57] on 06 April 2016, at 04:35 . For personal use only, all rights reserved.

  k 
3.1. Pure and Budget of Uncertainty
X
yk ≥ max −pk I1 + 4ut − dt 5 1 k ∈ T0 (4)
Robust Formulations d∈Dk t=1
Here we present the pure robust formulation (PR) and
In the case of the uncertain pure formulation,
the closely related budget of uncertainty robust for-
inequalities (3) and (4) yield the pure robust
mulation (BUR) of Bertsimas and Thiele (2006).
formulation:
In the pure robust formulation, the optimal solu-
tion is certainly feasible provided that the realized T
X
demands are within their given supports. Whereas (PR) min 4ck uk +fk vk +yk 5
k=1
such a solution is quite robust, it may also be over-
conservative. To control the degree of robustness
 k 
s.t. yk ≥ hk I1 + 4ut − d¯t + dˆt 5 1
X
and conservatism, Bertsimas and Thiele (2006) define k ∈ T1
t=1
“budgets of uncertainty” ât , obeying the relation ât ≤  k 
ât−1 + 1 for t ∈ T, with â0 = 0, which allows only some
yk ≥ −pk I1 + 4ut − d¯t − dˆt 5 1
X
k ∈ T1
of the demand figures to simultaneously deviate from
t=1
their nominal values. This approach guarantees the
feasibility of the optimal solution if no more than ât 0 ≤ uk ≤ Mvk 1 vk ∈ 80119 k ∈ T0 (5)
(t ∈ T) demands of periods 1 through t jointly deviate
In the case of the budget of uncertainty formula-
from their nominal values, and it provides a proba-
tion, Bertsimas and Thiele (2006) must deal with the
bilistic guarantee of feasibility (with respect to a con-
following auxiliary problem in (3) and (4) for each
straint) depending on the value ât and on the number
k ∈ T when deriving the budget of uncertainty robust
of uncertain parameters (see Bertsimas and Sim 2004).
formulation:
Note that when ât = t for t ∈ T, the budget of uncer-
tainty robust formulation is equivalent to the pure k
dˆt zt
X
robust formulation. All the decisions are given a pri- Ak = max
t=1
ori (i.e., at time 0) in both robust formulations.
k
Let yk be the inventory holding/backlogging cost X
at the end of period k ∈ T, and let vk be 1 if an order s.t. zt ≤ âk 1
t=1
is placed in period k ∈ T and 0 otherwise. Then, the
uncertain formulation for the RSP is as follows: 0 ≤ zt ≤ 11 1 ≤ t ≤ k0 (6)
T
X Note that a “budget of uncertainty” âk is consid-
min 4ck uk + fk vk + yk 5
ered for each cumulative demand up to period k ∈ T
k=1
in (6). Using the robust optimization methodology of
 k 
Bertsimas and Sim (2004) derived by associating the
k ∈ T1 d ∈ Dk 1
X
s.t. yk ≥ hk I1 + 4ut − dt 5 1
t=1
first two sets of constraints in (6) with dual variables
 k  q and r, respectively, and using strong duality, we
yk ≥ −pk I1 +
X
4ut − dt 5 1 k ∈ T1 d ∈ Dk 1 obtain the budget of uncertainty robust formulation,
t=1 which was introduced by Bertsimas and Thiele (2006):
0 ≤ uk ≤ Mvk 1 vk ∈ 801 191 k ∈ T1 (2) T
X
(BUR) min 4ck uk +fk vk +yk 5 (7)
where Dk is the uncertainty set denoting the vector of k=1
demands d = 4d1 1 0 0 0 1 dk 5, and M = Tk=1 4d¯k + dˆk 5.
P
s.t.
The above formulation with Dk = DPR k
= 8d¯t + dˆt zt :
k k
—zt — ≤ 11 ∀ 1 ≤ t ≤ k9, where zt denotes the scaled devi-
 
¯
X X
ation of dt , yields the uncertain pure formulation, yk ≥ hk I1 + 4ut − dt 5+qk âk + rtk 1 k ∈ T1 (8)
t=1 t=1
k ¯ ˆ
whereas Pkthe one with D = 8dt + dt zt : —zt — ≤ 11 ∀ 1 ≤  k k 
t ≤ k1 t=1 —zt — ≤ âk 9 yields the uncertain budget of
yk ≥ pk −I1 − 4ut − d¯t 5+qk âk + rtk 1
X X
k ∈ T1 (9)
uncertainty formulation. t=1 t=1
 Solyalı, Cordeau, and Laporte: Impact of Modeling on Robust Inventory Management
1192 Management Science 62(4), pp. 1188–1201, © 2016 INFORMS

letting uk = uk0 + k−1

Pk
qk +rtk ≥ dˆt 1
P
k ∈ T1 1 ≤ t ≤ k1 (10) t=1 ukt dt and yk = yk0 + t=1 ykt dt .
Then, the affinely adjustable counterpart of (2) is
qk ≥ 01 rtk ≥ 01 k ∈ T1 1 ≤ t ≤ k1 (11) as follows:
0 ≤ uk ≤ Mvk 1 vk ∈ 801191 k ∈ T0 (12) XT   k−1
X

Xk 
min ck uk0 + ukt dt +fk vk +yk0 + ykt dt
Note that qk âk + kt=1 rtk is the objective function of
P
k=1 t=1 t=1
the dual of (6), and constraints (10) and (11) are the k  k  i−1 
Downloaded from informs.org by [212.175.193.57] on 06 April 2016, at 04:35 . For personal use only, all rights reserved.

constraints of the dual of (6). We formulate the fol-

X X X
s.t. yk0 + ykt dt ≥ hk I1 + ui0 + uit dt −di 1
lowing remarks on the BUR. t=1 i=1 t=1
k
k ∈ T1 d ∈ DPR 1
Remark 1. As also stated by Bienstock and Özbay
(2008), one can replace qk âk + kt=1 rtk with Ak in
P k  k  i−1 
X X X
constraints (8) and (9) and eliminate constraints (10) yk0 + ykt dt ≥ −pk I1 + ui0 + uit dt −di 1
t=1 i=1 t=1
and (11) from the BUR, which leads to a robust for- k
mulation without any additional variables and con- k ∈ T1 d ∈ DPR 1
straints. This is because the optimal objective function k−1
X
value Ak of formulation (6) for each k ∈ T can be cal- 0 ≤ uk0 + ukt dt ≤ Mvk 1 vk ∈ 801191 k ∈ T0 (14)
t=1
culated in advance independently of the optimal solu-
tion values of the u, v, and y variables. Also note that Since (14) is a semi-infinite optimization problem, it
the optimal solution of formulation (6) for each k ∈ T is reformulated as a tractable mixed integer program-
gives the realization of the uncertain demands, which ming (MIP) problem, in a way similar to what was
is independent of the optimal solution of the BUR. done in (3) and (4), by ensuring feasibility of its con-
straints for any demand realization. Unlike (3) and (4),
Remark 2. For the case of static cost parameters,
where uncertain parameters have coefficients of −1
Bertsimas and Thiele (2006) show that the optimal
or 1, because some variables are defined as affine
robust policy can be found by solving an equivalent
functions of the demands, some uncertain parame-
nominal problem with modified demand:
ters have continuous variables as coefficients in (14),
p−h which may lead to additional variables and con-
dk0 = d¯k + 4A − Ak−1 51 k ∈ T1 (13) straints in the tractable MIP problem. Ben-Tal et al.
p+h k
(2004) reformulate an inequality of the form
where A0 = 0, Ak = qk∗ âk + kt=0 rtk∗ for k ∈ T, and q ∗
P
‹0 + ‹t dt ≤ 01 d ∈ DTPR 1
X
(15)
and r ∗ are the optimal q and r variables in the BUR
t∈T
formulation. The objective function value of the opti-
mal robust policy is equal to that of the nominal where ‹0 is a constant and ‹t 4t ∈ T5 denotes an
problem withPmodified demand, plus the extra term expression involving some variables, as
62ph/4p + h57 Tk=1 Ak . Furthermore, base-stock levels
‹0 + 4‹t d¯t + Œt dˆt 5 ≤ 01
X
(4s1 S5 if f > 0 and 4S1 S5 if f = 0) defining the optimal
robust policy depend on modified demand values dk0
t∈T (16)
4k ∈ T5. Thus, the associated robust formulation can −Œt ≤ ‹t ≤ Œt 1 t ∈ T0
be solved in closed form.
Using this reformulation methodology, we obtain
Note that Remarks 1 and 2 can easily be extended the following affinely adjustable robust formulation:
to the network case (i.e., several installations with
an arborescent/distribution structure) addressed in (AAR)
T   k−1 
Bertsimas and Thiele (2006). X X
¯
min ck uk0 + ukt dt + fk vk + yk0
k=1 t=1
3.2. Affinely Adjustable Robust Formulation k 
ykt d¯t + dˆk nk
X
Here we present the affinely adjustable robust for- + (17)
mulation derived using the methodology of Ben-Tal t=1
et al. (2004). The affinely adjustable robust formula- T
X
tion decides on the time of orders a priori but adapts s.t. −nt ≤ ytt + 4ck ukt + ykt 5 ≤ nt 1 t ∈ T1 (18)
the order quantities based on the realized demand. k=t+1

Specifically, we take the binary vk variables as nonad- k  k  i−1 

¯ ¯ ¯
X X X
justable variables determined at time 0 (as also dis- yk0 + ykt dt ≥ hk I1 + ui0 + uit dt − di
cussed by See and Sim 2010 in the context of an t=1 i=1 t=1
k
SOCP) and the uk and yk variables as adjustable vari-
dˆt skt 1
X
+ k ∈ T1 (19)
ables defined as affine functions of the demands by t=1
 Solyalı, Cordeau, and Laporte: Impact of Modeling on Robust Inventory Management
Management Science 62(4), pp. 1188–1201, © 2016 INFORMS 1193

k
X In contrast to See and Sim (2010), where the trun-
−skt ≤ hk uit − hk − ykt ≤ skt 1 cated linear replenishment policy (TLRP) performs
i=t+1
k ∈ T1 1 ≤ t ≤ k1 (20) best in terms of average-case performance, the robust
formulation for the linear replenishment policy (LRP)
k  k  i−1  always yields better results than the robust formula-
¯ ¯ ¯
X X X
yk0 + ykt dt ≥ −pk I1 + ui0 + uit dt − di tion for the TLRP in our computational experiments.
t=1 i=1 t=1
This is because there are much more second order
k
Downloaded from informs.org by [212.175.193.57] on 06 April 2016, at 04:35 . For personal use only, all rights reserved.

cone constraints in the formulation for the TLRP than

dˆt skt
X 0
+ 1 k ∈ T1 (21)
t=1
for the LRP, which means that obtaining a good TLRP
k
X is very difficult within the time limit using a state-
0 0
−skt ≤− pk uit + pk − ykt ≤ skt 1 of-the-art solver in the presence of binary variables.
i=t+1 Hence we adapt the robust formulation for the LRP
k ∈ T1 1 ≤ t ≤ k1 (22) to our problem in the following.
k−1 In the same way as AAR, the robust formulation
4d¯t ukt − dˆt xkt 5 ≥ 01
X
uk0 + k ∈ T1 (23) with an approximate objective function decides on the
t=1 time of orders a priori, but adapts the order quan-
−xkt ≤ ukt ≤ xkt 1 k ∈ T1 1 ≤ t < k1 (24) tities based on the realized demand. Defining yk0 as
k−1
the inventory level at the beginning of period k, the
4d¯t ukt + dˆt xkt 5 ≤ Mvk 1 binary vk variables are nonadjustable variables deter-
X
uk0 +
t=1 mined at time 0, whereas the uk and yk0 variables are
vk ∈ 801 191 k ∈ T1 (25) adjustable variables defined as affine functions of the
factors (i.e., uk = uk0 + k−1 0
P
random demand t=1 ukt zt and
where M = k=1 4d¯k + dˆk 5.
PT
0 0
P k−1 0 0
yk = yk0 + t=1 ykt zt ). Then, the robust formulation for
Note that objective (17) and constraints (18) are the LRP is as follows:
obtained by treating the objective function PT of (14) T  3 
as
Pk−1 a constraint (i.e., min
Pk Ž s.t. Ž ≥ k=1 4c k 4uk0 + min
X X 1 2
ck uk0 +fk vk + 4hk ƒi1 k+1 +pk ƒi1 k+1 5 (28)
T
t=1 ukt dt 5 + fk vk + yk0 + t=1 ykt dt 5, d ∈ DPR ). k=1 i=1
0 0
3.3. The Robust Formulation with an s.t. yk+11 t = ykt +ukt −dkt 1 k ∈ T1 0 ≤ t ≤ k1 (29)
Approximate Objective Function k−1
X
We now present the robust formulation we derived 0 ≤ uk0 + ukt z0t ≤ Mvk 1
from See and Sim (2010). Unlike the RSP, See and Sim t=1

(2010) considered an inventory management prob- vk ∈ 801191 k ∈ T1 z0t ∈ 6−dˆt 1 dˆt 71 (30)
lem that involves nonzero lead time, zero fixed cost, k
Y11i k+11 0 +max Y11i k+11 t z0t ≤ ƒ11
i
X
a capacity over order quantities, and a factor-based 0 k+1 1
zt
demand model. As a result, we adapt this robust t=1
k ∈ T1 i = 1121 (31)
formulation to our problem. The main idea of this
i
formulation is to approximate the expected value 0 ≤ ƒ11k+1 1 k ∈ T1 i = 1121 (32)
functions in the objective function of the stochastic k
problem with good upper bounds. z0t 4−Y21i k+11 t 5 ≤ ƒ21
i
X
max
0 k+1 1 k ∈ T1 i = 1121 (33)
The factor-based demand model adopted by See zt
t=1
and Sim (2010) is an affine function of zero mean ran- Y21i k+11 0 ≤ ƒ21
i
k+1 1 k ∈ T1 i = 1121 (34)
dom factors z0t defined as follows:
 k 1/2
1 i
+ 12 4Y31i k+11 0 52 + ‘z20t 4Y31i k+11 t 52
N
X
X Y
2 31 k+11 0
dk = dk0 + dkt z0t 1 k ∈ T1 (26) t=1
t=1 i
≤ ƒ31 k+1 1 k ∈ T1 i = 1121 (35)
where N is the number of random factors, which are
realized in sequence over time. Y111 k+11 t +Y211 k+11 t +Y311 k+11 t = yk+11
0
t1
In our case of independently distributed demand, k ∈ T1 0 ≤ t ≤ k1 (36)
the factor-based demand model takes the form
Y112 k+11 t +Y212 k+11 t +Y312 k+11 t = −yk+11
0
t1
k−1
X k ∈ T1 0 ≤ t ≤ k1 (37)
dk = dk0 + dkk z0k + dkt z0t 1 k ∈ T1 (27)
where M = k=1 4d¯k + dˆk 5, y10 = I1 , dk0 = d¯k , dkt = 0 for
PT 0
t=1

where dk0 = d¯k , dkk = 1, dkt = 0 for 1 ≤ t < k, z0t ∈ 1 ≤ t < k, dkk = 1, ykk 0
= ukk = 0, and ‘z20 = 42dˆt 52 /12.
t
2
6−dˆt 1 dˆt 7 for t ∈ T, and the random factors z01 1 0 0 0 1 z0t Note that ‘z0 is the variance of z0t , which is uniformly
t

are known at the end of period t. distributed between −dˆ and dˆ . t t

 Solyalı, Cordeau, and Laporte: Impact of Modeling on Robust Inventory Management
1194 Management Science 62(4), pp. 1188–1201, © 2016 INFORMS

As done for AAR, using the reformulation method-
ology (15) and (16), we reformulate model (28)–(37) as
a tractable SOCP formulation with binary constraints:
(SS) min 4285
s.t. (29), (32), (34)–(37),
Downloaded from informs.org by [212.175.193.57] on 06 April 2016, at 04:35 . For personal use only, all rights reserved.
Thus, the positive initial inventory should explicitly
be considered in the formulation. On the other hand,
in the case of backlogged demand (i.e., negative ini-
tial inventory, I1 < 0), we can easily treat initial inven-
tory as zero by defining a dummy demand d0 = −I1 at
period 0, which is going to be backlogged, and solv-
ing an equivalent problem with I1 = 0 and periods

k−1
dˆt xkt ≥ 01 k ∈ T ∪ 809.
X
uk0 − k ∈ T1 (38)
t=1 Using the disaggregation idea of the FL formula-
−xkt ≤ ukt ≤ xkt 1 1 ≤ t < k1 k ∈ T1 (39) tion, let w0k be the fraction of the customer demand in
k−1 period k that is supplied by the positive initial inven-
dˆt xkt ≤ Mvk 1
X
uk0 + tory, let wtk be the fraction of the customer demand
t=1 in period k that is ordered in period t, and let wT +11 k
vk ∈ 801 191 k ∈ T1 (40)
be the fraction of customer demand in period k that
k
is unsatisfied. Defining gtk as the unit cost of satis-
Y11i k+11 0 + dˆt mik+11 t ≤ ƒ11
i
X
k+1 1
t=1
fying customer demand in period k by ordering in
k ∈ T1 i = 11 21 (41) period t, the new uncertain formulation we propose
for the RSP (assuming I1 ≥ 0) is as follows:
−mik+11 t ≤ Y11i k+11 t ≤ mik+11 t 1
1 ≤ t ≤ k1 k ∈ T1 i = 11 21 (42)  T
X TX T
+1 X
k min fk vk + gtk dk wtk
dˆt nik+11 t ≤ ƒ21
i
X
k+1 1 k ∈ T1 i = 11 21 (43) k=1 t=1 k=1
t=1 T
X

−nik+11 t ≤ −Y21i k+11 t ≤ nik+11 t 1 + 4hk I1 − HkT dk w0k5 (45)
k=1
1 ≤ t ≤ k1 k ∈ T1 i = 11 20 (44)
TX
+1
s.t. wtk = 11 k ∈ T1 (46)
Note that (38)–(40), (41) and (42), and (43) and (44) t=0
are the robust counterparts of (30), (31), and (33),
respectively. wtk ≤ vt 1 t ∈ T1 k ∈ T1 (47)
T
dk ∈ 6d¯k − dˆk 1 d¯k + dˆk 71
X
dk w0k ≤ I1 1 (48)
4. A New Robust Formulation k=1
We now propose a new robust formulation that rep-
wtk ≥ 01 vk ∈ 801 191
resents the RSP differently from the PR, BUR, AAR,
0 ≤ t ≤ T + 11 k ∈ T1 (49)
and SS formulations. As mentioned earlier, the RSP is
actually the robust counterpart of the ULSB for which
tight reformulations, such as facility location (FL) where HkT = Tl=k hl 1gtk = ct + k−1
P P
h if t ≤ k, gtk = ct +
Pt−1 PT l=t l
and shortest path (SP) reformulations, are known. We p
l=k l if k < t, and g T +11k = p
l=k l .
make use of the FL formulation because of its simplic- The objective (45) is the total of ordering costs, in-
ity compared with the SP formulation, in proposing ventory holding/shortage costs, and the cost due to
a new robust formulation for the RSP. The main idea the initial inventory. Constraints (46) ensure that the
is to disaggregate ut variables for the demand of each sum of the fraction of demand in period k that is sup-
period by specifying when to place an order to satisfy
plied by the initial inventory, the fraction of demand
the demand of a specific period.
in period k that is ordered from period 1 through T ,
An important difference between the robust prob-
plus the fraction of demand in period k that is left
lem and the deterministic ULSB is the impact of the
positive initial inventory. In the case of the determin- unmet is equal to one. Constraints (47) stipulate that a
istic ULSB, the positive initial inventory can be treated fixed order cost is incurred in period t if any amount
as zero because one can obtain an equivalent prob- is ordered in that period. Constraints (48) ensure that
lem with zero initial inventory by deducting the initial the total amount supplied by the initial inventory
inventory from the deterministically known demands does not exceed its level. Constraints (49) impose the
until the initial inventory is depleted, whereas this nonnegativity and integrality of the variables.
cannot be performed in the case of the robust problem Since uncertain demand parameters appear both in
since the demand realizations are not known a priori. the objective (45) and constraints (48), we apply the
 Solyalı, Cordeau, and Laporte: Impact of Modeling on Robust Inventory Management
Management Science 62(4), pp. 1188–1201, © 2016 INFORMS 1195

robust optimization methodology of Bertsimas and each period t by inserting the values of the wtk vari-
Sim (2004) to these, and obtain the following nonlin- ables into (50) and solving the inner maximization
ear robust formulation problem of (50), which can easily be done by solving
T T T +1  a continuous knapsack problem. Note that the solu-
¯ tion of the inner maximization problem for a given w,
X X X
(NR) min fk vk + dk gtk wtk −HkT w0k +H1T I1
k=1 k=1 t=1 denoted by z∗0k , yields the worst-case demand realiza-
T
T +1
  tion. In particular, the worst-case demand realization
Downloaded from informs.org by [212.175.193.57] on 06 April 2016, at 04:35 . For personal use only, all rights reserved.

dˆk  gtk wtk −HkT w0k z0k 2

X 
in period k would be d¯k + sk dˆk z∗0k , where sk is equal
X
+max
to 1 if Tt=1
P +1
k=1 t=1 gtk wtk ≥ HkT w0k , and −1 otherwise. When
T
X
 z∗0k are inserted into the z1k variables in the maximiza-
z0k ≤ âT 1 0 ≤ z0k ≤ 11 k ∈ T (50) tion problem of (51), this is feasible to the maximiza-
k=1
tion problem of (51), which means that we have a
s.t. (46), (47), (49), lower bound for the maximization problem of (51).
T  T T
Thus, this solution is also feasible for (51). Using the
d¯k w0k +max dˆk w0k z1k 2 obtained worst-case demand realization, one can eas-
X X X
z1k ≤ âT 1
k=1 k=1 k=1 ily compute the amount to order in period t as ut =
Pt−1 ˜ PT ¯ ˆ ∗ ˜
k=1 dk wtk + k=t 4dk + sk dk z0k 5wtk , where dk denotes

0 ≤ z1k ≤ 11 k ∈ T ≤ I1 0 (51) the realized demand in period k.
Second, NR is a computationally attractive formu-
Using the strong duality theorem, we obtain the fol- lation since it can be solved in polynomial time when
lowing equivalent linear model: the initial inventory level is zero or negative, as dis-
cussed in the sequel. Third, unlike the BUR formu-
 T T T +1 
(NR) min
X
fk vk +
X
d¯k
X
gtk wtk −HkT w0k lation, NR requires only a global budget of uncer-
k=1 k=1 t=1 tainty âT , instead of a funnel of budgets of uncertainty
T
X
 4â1 1 0 0 0 1 âT 5. Although a funnel of budgets could be
+H1T I1 +ˆ0 âT + 0k (52) used in formulating NR, we could not extend the
k=1
polynomial solvability property of the current NR to
s.t. (46), (47), (49), the resulting formulation, and we have not observed
ˆ0 +0k ≥ dˆk ‚k 1 k ∈ T1 (53) any advantage to using a funnel of budgets over a
global budget of uncertainty regarding the average-
TX
+1
case performance.
−‚k ≤ gtk wtk −HkT w0k ≤ ‚k 1
t=1
In the presence of nonzero fixed costs, the formu-
k ∈ T1 (54) lation for the RSP in Bertsimas and Thiele (2006)
T T is a mixed integer programming problem, which is
d¯k w0k +ˆ1 âT +
X X
1k ≤ I1 1 (55) not polynomially solvable. This is because demand
k=1 k=1 uncertainty normally implies data uncertainty in
ˆ1 +1k ≥ dˆk w0k 1 k ∈ T1 (56) the constraints of a formulation. However, we have
transferred demand uncertainty from the constraints
ˆ0 ≥ 01 ˆ1 ≥ 01 0k ≥ 01 1k ≥ 01 to the objective function of the formulation, and there
k ∈ T0 (57) can thus be an opportunity to find a polynomial time
algorithm when the initial inventory is nonpositive.
We have three important observations to formu-
Next, we show how NR can be solved in polynomial
late on the NR formulation. First, the solution of
time when I1 ≤ 0. (As discussed before, the case with
NR defines a policy like AAR; that is, the solution
I1 < 0 can be treated as the case with I1 = 0.) Letting
explicitly defines the time of orders at time 0 and
I1 = 0 in NR, it is easy to show that the w variables
the fraction of each period’s demand that is to be
automatically take a binary value, and thus we obtain
satisfied regardless of the demand realization. Thus,
the following robust formulation:
the vk variables indicating the time of orders are
decided a priori, whereas the wtk variables are adapt-  T TX T
+1 X
d¯k gtk wtk
X
able to demand realization if t > k. Unlike the wtk (NR-0) min fk vk +
variables for t > k, where the demand in period k has k=1 t=1 k=1
T 
already realized, meaning that the amount to order X
+ ˆ0 âT + 0k (58)
in period t is known, the wtk variables for t ≤ k do
k=1
not explicitly give the amount to order in period t
TX
+1
since demand in period k is still unknown at t. Nev-
s.t. wtk = 11 k ∈ T1 (59)
ertheless, we can compute the amounts to order in t=1
 Solyalı, Cordeau, and Laporte: Impact of Modeling on Robust Inventory Management
1196 Management Science 62(4), pp. 1188–1201, © 2016 INFORMS

wtk ≤ vt 1 t ∈ T1 k ∈ T1 (60) The results obtained for the RSP with a budget of
TX
+1 uncertainty can easily be extended to the robust prob-
ˆ0 + 0k ≥ dˆk gtk wtk 1 k ∈ T1 (61) lem without backlogging. Hence, a reformulation-
t=1 based robust formulation of the latter problem is also
ˆ0 ≥ 01 0k ≥ 01 wtk ∈ 801 191 vk ∈ 801 191 polynomially solvable in O4T 4 5 time.
As will be seen in the sequel, the NR formulation
1 ≤ t ≤ T + 11 k ∈ T0 (62)
can be solved to optimality within a few seconds even
Downloaded from informs.org by [212.175.193.57] on 06 April 2016, at 04:35 . For personal use only, all rights reserved.

for large-scale instances with positive initial inven-

Note that 0k = max8dˆk t=1
PT +1
gtk wtk − ˆ0 1 09 in an
tory. Thus, to improve its performance, NR can be
optimal solution. Thus, we can rewrite the expression
implemented through a folding horizon approach,
for 0k as follows:
which we denote by NR-FH. Under this approach, the
 TX
+1  NR formulation is sequentially solved for each period,
0k = max dˆk gtk wtk − ˆ0 1 0 and only the decisions made for the current period are
t=1 implemented, whereas the decisions concerning the
TX
+1 following periods are reoptimized when solving the
= max8dˆk gtk wtk − ˆ0 1 09 (due to (59)) next problem. In this approach, when NR is solved
t=1 for a specific period i (covering periods from i to T ),
TX
+1 a fixed order cost is incurred if an order is placed in
= max8dˆk gtk − ˆ0 1 09wtk 0 (63) period i and the amount to order in period i is calcu-
t=1
lated as ui = −I˜i wi1 i−1  + Tk=i 4d¯k + sk dˆk z∗0k 5wik , where
P

We can then restate NR-0 using (63) and letting ĝtk =  = 1 if the realized inventory level I˜i at the beginning
dˆk gtk as of period i is negative (i.e., I˜i = I˜i−1 + ui−1 − d˜i−1 < 0),
and 0 otherwise. We next present the NR formulation
 T TX T
+1 X that is sequentially solved for each period i (1 ≤ i ≤ T )
d¯k gtk wtk + ˆ0 âT
X
(NR-0) min f k vk + in NR-FH:
k=1 t=1 k=1
T TX
X +1   T T T +1 
max8ĝtk − ˆ0 1 09wtk i
d¯k
X X X
+ 4NR 5 min f k vk + gtk wtk −HkT w0k
k=1 t=1 k=i k=i t=i

s.t. (59), (60), T TX

+1 
I˜i gt1 i−1 wt1 i−1
X
+ˆ0 âT −i+1 + 0k −
ˆ0 ≥ 01 wtk ∈ 801 191 vk ∈ 801 191 k=i t=i

1 ≤ t ≤ T + 11 k ∈ T0 (64) TX
+1
s.t. 41−5w0k + wtk = 11 i≤k≤T 1
t=i
Similar to Bertsimas and Sim (2003), it is easy to
TX
+1
show that ˆ0 takes a value in 801 ĝ11 1 0 0 0 1 ĝT +11 T 9 in
wt1 i−1 = 1 if I˜i < 01
an optimal solution. Therefore, solving NR-0 for each t=i
distinct value of ˆ0 and selecting the one with the best
wtk ≤ vt 1 i≤t ≤T 1 i≤k≤T 1
objective function value yields the optimal solution.
wt1 i−1 ≤ vt 1 i≤t ≤T 1
Theorem 1. NR-0 can be solved in polynomial time by
solving at most T 4T + 15 + 1 LP problems. ˆ0 +0k ≥ dˆk ‚k 1 i≤k≤T 1
TX
+1
Proof. Since the LP relaxation of the FL formula-
−‚k ≤ gtk wtk −HkT w0k ≤ ‚k 1 i≤k≤T 1
tion for the ULSB is integral, the proof follows from t=i
the above discussion. ƒ T T
d¯k w0k +ˆ1 âT −i+1 + 1k ≤ 41−5I˜i
X X
To propose a polynomial time algorithm for the RSP
with a budget of uncertainty, which does not require k=i k=i

solving LP problems, we make use of the SP formula- ˆ1 +1k ≥ dˆk w0k 1 i≤k≤T 1
tion proposed for the ULSB and propose a new robust
formulation referred to as the SPR formulation (see ˆ0 ≥ 01 ˆ1 ≥ 01 0k ≥ 01 1k ≥ 01 w0k ≥ 01
the appendix for details). The result is given in the vk ∈ 801191 i≤k≤T 1
following theorem.
wtk ≥ 01 i ≤ t ≤ T +11 i −1 ≤ k ≤ T 1 (65)
Theorem 2. The SPR formulation is polynomially solv-
able in O4T 4 5 time. where wt0 = 0 for i ≤ t ≤ T + 1.
 Solyalı, Cordeau, and Laporte: Impact of Modeling on Robust Inventory Management
Management Science 62(4), pp. 1188–1201, © 2016 INFORMS
5. Computational Results
We performed computational experiments on test
instances to assess the average-case performance of
solutions yielded by the PR, BUR, AAR, SS, NR,
and NR-FH formulations. The average-case perfor-
mance of formulations was assessed by considering
the average performance of the solution yielded by
Downloaded from informs.org by [212.175.193.57] on 06 April 2016, at 04:35 . For personal use only, all rights reserved.
each formulation for an instance over a given number
of simulations of realized demand. To evaluate how
far the average-case performance of different formu-
lations is from the optimal expected objective function
values, we also computed a lower bound on the opti-
mal expected objective function value for each test
instance. We also tested the computational effective-
ness of these formulations with respect to the CPU
time needed to solve them to optimality. All these for-
mulations were solved by CPLEX 12.5 with its default
settings using a single thread, and all computational
experiments were performed on a 2.4 GHz Worksta-
tion with 48 GB RAM running on Windows 7 (64 bit).
For the computational experiments, we generated
test instances using settings similar to those used by
Bertsimas and Thiele (2006) as follows. The length of
the planning horizon was set to two different values:
T ∈ 8201 509. The unit order cost was taken as ck = 1.
The fixed cost was set to fk = 11000. Two opposite
relations were used for unit holding and unit back-
logging costs: hk < pk and hk > pk . The unit backlog-
ging cost was set to pk ∈ 851 71 99 when hk = 4, and
the unit holding cost was set to hk ∈ 851 71 99 when
pk = 4. Both zero initial inventory I1 = 0 and nonzero
initial inventory I1 = 130 were considered. The uncer-
tain demand has a mean d¯k = 100 and standard devi-
ation ‘ = 10 or 20. In our robust optimization frame-
work, the uncertain demand dk can take any value
from the interval 6d¯k − 2‘1 d¯k + 2‘7, i.e., dˆk = 2‘.
To simulate the performance of solutions yielded by
the PR, BUR, AAR, SS, NR, and NR-FH formulations,
different demand scenarios with different underlying
demand distributions and different degrees of cor-
relation among demands were considered. Specifi-
cally, realized demand values were generated as (see,
e.g., Federgruen and Tzur 1999) d1 = …1 and dt =
dt−1 + 41 − 5…t , where …t is an independent ran-
dom variable that is uniformly or truncated normally
distributed from the interval 6d¯k − 2‘1 d¯k + 2‘7, and
 ∈ 801 00251 00759. Note that the demands are inde-
pendent when  = 0, whereas demands are weakly
and strongly correlated, respectively, when  = 0025
and  = 0075. Thus, combining all parameter settings,
we generated 48 core instances and 288 instances
(48 instances with two probability distributions and
three correlation coefficients) for the average-case
performance assessment, in total. To estimate the
average performance of solutions provided by the
different formulations, we generated 100 simulations
1197
of realized demand (see, e.g., Ben-Tal et al. 2004)
for each instance. Note that these 100 sample paths
for demand are identical across all formulations. For
BUR, NR, and NR-FH, the budgets of uncertainty âk
were generated using the algorithm by Bertsimas and
Thiele (2006), which attempts to compute the optimal
budgets of uncertainty minimizing the expected value
of the objective function value. For NR and NR-FH,
we also tested the method of Bertsimas and Sim (2004)
to determine the budgets of uncertainty âk . Specifi-
cally, we chose âT −k+1 for the formulation at period k
(k ∈ T) such that the realized cost will be larger than
the objective function value of the formulation with
probability of at most „%, where „ ∈ 811 51 109. Note
that such a â value corresponds to at most „% proba-
bility of infeasibility of constraint (48) for a given opti-
mal solution of the formulation (see Bertsimas and
Sim 2004, Theorem 3). However, the average-case per-
formance obtained with this method was worse than
that obtained by the former method on the major-
ity of the instances. Thus, we did not report the
results obtained with the method of Bertsimas and
Sim (2004). The LB on the optimal expected objective
function value was found by solving a deterministic
version of the problem for each simulation of realized
demand, assuming all demand is known in advance,
and then taking the average objective function value
over 100 simulations of realized demand.
We first tested the computational effectiveness of
the formulations. We present the average results ob-
tained by all formulations in Table 1. We imposed
a time limit of two hours for all formulations. The
seven columns in Table 1 show the formulation, the
length of the planning horizon, the initial inventory
level, the time spent to solve the instances in seconds,
the number of nodes explored after the root node
in the branch-and-bound tree, the percentage gap
between the best known upper (zUB ) and lower (zLB )
bounds (i.e., Gap% = 1004zUB − zLB 5/zLB 5 obtained by
each formulation, and the number of optimally solved
instances out of 72 instances for each combination
of T and I1 , respectively. Note that the results pro-
vided in “Seconds,’’ “Nodes,’’ and “Gap%’’ columns
were found by taking the average over 72 instances
for each combination of T and I1 .
Table 1 shows that all instances with T = 20 could
be solved to optimality within a few seconds by PR,
BUR, and NR, and within a few minutes by AAR,
whereas SS could optimally solve just 8.33% of the
instances with I1 = 0 within 1.97 hours and less than
half of the instances with I1 > 0 within 1.85 hours.
NR could optimally solve larger instances within a
second on average (at most two seconds), whereas
AAR and SS (resp., PR and BUR) could not solve any
(resp., many) of the instances with T = 50 to optimal-
ity within the two-hour time limit. Also, the remain-
ing percentage gap between the best UB and LB is still
 Solyalı, Cordeau, and Laporte: Impact of Modeling on Robust Inventory Management
1198 Management Science 62(4), pp. 1188–1201, © 2016 INFORMS

Table 1 Effectiveness of the Formulations We compared all the formulations with each other
with respect to their best solutions (the ones found
Formulation T I1 Seconds Nodes Gap% #Opt
within the two-hour time limit) and their practical
PR 20 0 005 2123807 0000 72 performance with respect to the same demand real-
>0 006 2165804 0000 72 izations. The key results presented in Table 2 are as
50 0 4110403 141456149105 0034 42 follows:
>0 4131701 151076180807 0027 36
• NR-FH and SS yield the best average-case per-
BUR 20 0 007 2181603 0000 72
Downloaded from informs.org by [212.175.193.57] on 06 April 2016, at 04:35 . For personal use only, all rights reserved.

>0 006 1196501 0000 72 formance and outperform all other formulations on
50 0 3191301 131786164002 1041 36 most instances. In particular, NR-FH outperforms all
>0 3137709 101704162405 0067 54 formulations on instances with T = 50 and all formu-
AAR 20 0 42404 71170808 0000 72 lations except SS on instances with T = 20. SS, on the
>0 30605 55180704 0000 72 other hand, outperforms all formulations on instances
50 0 7120000 12162706 74016 0
>0 7120000 13122209 74097 0
with T = 20.
SS 20 0 7107501 95189203 13046 6 • AAR, being the third best, outperforms PR, BUR,
>0 6164409 91175505 7022 31 and NR on most instances and performs slightly bet-
50 0 7120000 77701 276093 0 ter than SS on instances with T = 50.
>0 7120000 73802 281000 0 • BUR performs better than PR on most instances
NR 20 0 003 709 0000 72
and better than NR on the majority of instances. NR,
>0 003 5507 0000 72
50 0 005 3006 0000 72 on the other hand, performs better than PR on most
>0 008 29401 0000 72 instances.
NR-FH 20 0 7501 — — — • PR performs poorly compared to other formula-
>0 7403 — — — tions on almost all instances, regardless of the initial
50 0 83702 — — — inventory level or of the length of the time horizon.
>0 86403 — — —
• The presence of a nonzero initial inventory
adversely affects the performance of PR, BUR, and
quite large for AAR, and especially for SS. Because of NR-FH compared to the cases without an initial
the computational inefficiency of the PR, BUR, AAR, inventory. On the other hand, the performance of
AAR, SS, and NR becomes slightly better in the pres-
and SS formulations, only NR was implemented
ence of an initial inventory.
through a folding horizon approach, which is denoted
• The main benefit of NR is computational in that
by NR-FH. For PR, BUR, AAR, SS, and NR, the for-
it does not perform well when not executed on a fold-
mulations were solved only once, and the obtained ing horizon, whereas with a folding-horizon imple-
policies or solutions were implemented directly. As mentation (i.e., NR-FH) it is very attractive and yields
can be seen in the last four rows of Table 1, NR-FH the best average-case performance on most large-scale
can be executed quite fast even for the instances with instances.
T = 50, where NR-FH was solved in 15 minutes on Next we determined how far from LB the average-
average. case performance is by simply deriving the efficiency

Table 2 Average-Case Comparison of Formulations

T = 20 T = 50

I1 = 0 PR BUR AAR SS NR NR-FH Best I1 = 0 PR BUR AAR SS NR NR-FH Best

PR — 0 1 0 18 0 0 PR — 0 0 0 12 0 0
BUR 72 — 6 0 48 5 0 BUR 72 — 0 1 49 6 0
AAR 71 66 — 0 69 16 0 AAR 72 72 — 29 70 17 11
SS 72 72 72 — 72 67 67 SS 72 71 43 — 72 18 12
NR 54 24 3 0 — 3 0 NR 60 23 2 0 — 6 0
NR-FH 72 67 56 5 69 — 5 NR-FH 72 66 55 54 66 — 49

PR — 4 1 0 13 0 0 PR — 0 0 0 0 0 0
BUR 68 — 3 0 36 5 0 BUR 72 — 0 3 39 8 0
AAR 71 69 — 1 72 17 0 AAR 72 72 — 44 72 27 19
SS 72 72 71 — 72 69 69 SS 72 69 28 — 68 20 14
NR 59 36 0 0 — 2 0 NR 72 33 0 4 — 6 0
NR-FH 72 67 55 3 70 — 3 NR-FH 72 64 45 52 66 — 39

Notes. Except for the column “Best,” each entry represents the number of times the formulation indicated in the row yielded a better average-case performance
than the one in the column. Each entry in the column “Best” indicates the number of times the formulation written in the row yielded the best average-case
performance.
 Solyalı, Cordeau, and Laporte: Impact of Modeling on Robust Inventory Management
Management Science 62(4), pp. 1188–1201, © 2016 INFORMS 1199

of a robustness approach as LB/Z4 · 5, where Z4 · 5 is Figure 1 Standard Deviation of Objective Function Values on
the average-case performance yielded by a robust- Instances with T = 20
ness approach. Note that a higher efficiency value, 7,000
which can be at most one, indicates a better per- PR
BUR
formance. We compared all the formulations with 6,000
AAR
SS
each other with respect to their efficiency in Table 3, NR
5,000

Standard deviation
where the first column shows the instances included, NR-FH
Downloaded from informs.org by [212.175.193.57] on 06 April 2016, at 04:35 . For personal use only, all rights reserved.

the second column shows the length of the planning 4,000

horizon, the third column shows the initial inven-
tory level, the fourth column shows the variability of 3,000

demand, and the remaining columns show the aver- 2,000

age efficiencies of PR, BUR, AAR, SS, NR, and NR-FH,
respectively. Because NR-FH does not perform so well 1,000
on all instances with h = 4 and p = 9, we provide the
average efficiencies on all instances excluding those 0
0 20 40 60 80 100 120 140 160
with h = 4 and p = 9 in the last eight rows of Table 3. Instance
The key results of Table 3 are as follows:
• NR-FH, SS, and AAR exhibit a performance
superior to that of the other formulations with regard increases. In other words, the NR, BUR, and PR for-
to efficiency. They mostly provide actual cost savings mulations are more adversely affected by an increase
that are very close to the optimal expected cost on in the variability of demand than are the NR-FH, SS,
most instances. and AAR formulations with respect to efficiency.
• NR-FH shows a superior (resp., competitive) per- • Excluding instances with h = 4 and p = 9 pos-
formance on the instances with T = 50 (resp., T = 20) itively affects the average efficiency of all formula-
tions.
except for those with h = 4 and p = 9 compared to all
• The initial inventory, demand distributions, and
formulations (resp., SS).
correlation values do not significantly affect the effi-
• SS performs the best on the instances with T = 20.
ciency results.
The efficiency of SS is the most stable of all.
To see the variability of the formulations, the stan-
• The AAR formulation also performs very well.
dard deviations of objective function values (over
Specifically, it performs the best when T = 50 and
100 simulations of realized demand) are plotted and
‘ = 10.
presented in Figure 1. These results clearly reveal that
• When the variability of uncertain demand in-
AAR has the smallest variability, followed by NR-FH
creases (i.e., when ‘ increases from 10 to 20), the
and SS. The remaining formulations yield high vari-
efficiency gap between the NR-FH, SS, and AAR for-
ability, with NR being slightly less variable than BUR
mulations and the NR, BUR, and PR formulations
and PR.

Table 3 Average Efficiency of the Formulations

6. Conclusions
Instance T I1 ‘ PR BUR AAR SS NR NR-FH We have studied the impact of modeling in robust
inventory management problems under interval de-
All 20 010 00836 00912 00955 00971 00904 00949
20 00693 00819 00915 00944 00777 00927
mand uncertainty. We have derived robust formu-
>0 10 00827 00909 00957 00974 00904 00948 lations by applying prominent robust optimization
20 00690 00814 00911 00948 00820 00926 methodologies, and we have proposed a new robust
50 0 10 00676 00871 00944 00943 00857 00914 model obtained by reformulating the problem. Our
20 00485 00731 00910 00913 00705 00898 results show that the new robust formulation is poly-
>0 10 00672 00866 00942 00932 00864 00909
nomially solvable when the initial inventory is zero
20 00485 00730 00909 00915 00712 00897
or negative, and also indicate that the heuristic in
All0 20 0 10 00864 00921 00958 00975 00917 00971
20 00724 00831 00919 00949 00787 00942
its folding-horizon implementation of the new robust
>0 10 00855 00917 00962 00977 00918 00969 formulation is superior to the pure, budget of uncer-
20 00724 00826 00916 00953 00834 00942 tainty, and affinely adjustable robust formulations. It
50 0 10 00712 00881 00945 00945 00870 00966 is also competitive with the robust formulation using
20 00517 00742 00913 00915 00718 00935 an approximate objective function with respect to the
>0 10 00709 00876 00944 00933 00882 00964 average-case performance. Moreover, the new formu-
20 00517 00741 00914 00917 00748 00936
lation and the new heuristic are far more efficient than
Note. All0 indicates all instances except those with h = 4 and p = 9. the others in terms of computing time: instances with
 Solyalı, Cordeau, and Laporte: Impact of Modeling on Robust Inventory Management
1200 Management Science 62(4), pp. 1188–1201, © 2016 INFORMS

50 periods could be optimally solved within a few sec- Figure A.1 The Shortest Path Network Representation for an
onds with the new formulation, or within 15 minutes Example Instance with T = 3
with the new heuristic, whereas the majority of these 13
instances could not be solved to optimality within two
hours with other formulations, which also exhibited 12 23

large gaps between the best known upper and lower

bounds.
11 11
22 22
33 33
Downloaded from informs.org by [212.175.193.57] on 06 April 2016, at 04:35 . For personal use only, all rights reserved.

The reformulation idea is important since all inven- 1 1′ 2 2′ 3 3′ 4

tory management problems considered in the literature
43

21
32
can be reformulated using the same facility loca-

42
tion idea (see Pochet and Wolsey 2006). Consider-
31
ing that the problem studied in this paper is a basic
building block of more complex inventory management
41

problems, like problems with a network structure as in

Bertsimas and Thiele (2006), the reformulation-based
robust formulations can provide well-performing poli- by the variables ”tk 1 –tk for t ≤ T , and –T +11 k . Variables ”tk
cies for more complex inventory management prob- represent the flow from node t 00 to k + 1 41 ≤ t ≤ k ≤ T 5 with
lems within reasonable computing times. Another cost etk , –tk the flow from node k to t 00 41 ≤ k ≤ t ≤ T 5 with
important implication of this study is the realiza- cost btk , and –T +11 k the flow from node k 41 ≤ k ≤ T 5 to T + 1
tion that reformulation-based robust formulations, as with cost bT +11 k . A unit flow is sent to the network through
opposed to standard robust formulations obtained by node 1. An example shortest path network representation
applying classical robust optimization methodologies, for an instance with T = 3 is presented in Figure A.1. The
can yield different and better-performing policies or objective function of the SPR formulation for the nomi-
nal ULSB is Tk=1 fk vk + Tt=1
P +1 Pt−1 PT PT
solutions to problems under parameter uncertainty.
P
k=1 btk –tk + t=1 k=t etk ”tk ,
which involves expressions where dk (k ∈ T) is encountered
Supplemental Material several times. As a result, we should extract dk to be able to
Supplemental material to this paper is available at http://dx write the robust counterpart. After some algebraic manip-
ulations, we rewrite the objective function as Tk=1 fk vk +
P
.doi.org/10.1287/mnsc.2015.2183. PT Pk PT PT +1 Pk
k=1 dk 4 r=1 t=k grk ”rt + r=k+1 t=1 grk –rt 5. Then, the SPR
Acknowledgments formulation for the RSP with budget of uncertainty is
This work was partly supported by the Canadian Natural as follows:
Sciences and Engineering Research Council [Grants 227837-
09 and 39682-10]. This support is gratefully acknowledged.  T T  k X
T TX
+1 k 
d¯k
X X X X
min fk v k + grk ”rt + grk –rt
Appendix k=1 k=1 r=1 t=k r=k+1 t=1

In this appendix, we show how the SPR formulation can be  T  k X

T TX
+1 k 
dˆk zk
X X X
solved in polynomial time. We simplify the SPR formula- + max grk ”rt + grk –rt 2
k=1 r=1 t=k r=k+1 t=1
tion given in Pochet and Wolsey (2006) and define new vari-
ables to account for the unmet demand case, as was done T
X

in NR. We first define the parameters and variables used in z0k ≤ âT 1 0 ≤ z0k ≤ 11 k ∈ T (66)
k=1
the SPR formulation, and we then describe the associated
shortest path network representation. Let ”tk be 1 if the total TX
+1
customer demand from periods t through k is satisfied in s.t. –k1 = 11 node 11 (67)
period t 41 ≤ t ≤ k ≤ T 5 and 0 otherwise, and let –tk be 1 if k=1

the total demand of the customer from periods k through TX

+1 t−1
X
t − 1 is satisfied in period t 41 ≤ k ≤ t ≤ T 5 and 0 otherwise. –kt − ”k1 t−1 = 01 node t1 2 ≤ t ≤ T 1 (68)
Also, let –T +11 k be 1 if the total demand of the customer from k=t k=1

periods k (k ∈ T) through T is not satisfied and 0 otherwise. t T

node t 00 1 1 ≤ t ≤ T 1
X X
Define Dtk as the total demand from periods t through k, − –tk + ”tk = 01 (69)
i.e., Dtk = kj=t dj ; define etk as the total cost of satisfying
P
k=1 k=t

the customer demand from periods t through k in period t, T

i.e., etk = ct Dtk + k−1
P X
j=t hj Dj+11 k ; and let btk be the total cost of ”tk ≤ vt 1 1≤t≤T1 (70)
satisfying demand of the customer from period k through k=t
t − 1 in period t, i.e., btk = ct Dk1 t−1 + t−1
P
j=k pj Dkj . In the asso- ”tk ∈ 801 191 1≤t≤k≤T1 (71)
ciated network, nodes represent the time periods, whereas
arcs represent the order decisions. There are two nodes for –tk ∈ 801 191 1 ≤ k ≤ t ≤ T + 11 k ≤ T 1 (72)
each period t, denoted by t and t 00 , and a dummy node T + 1.
There are three types of arcs on which the flow is represented vt ∈ 801 191 1≤t≤T0 (73)
 Solyalı, Cordeau, and Laporte: Impact of Modeling on Robust Inventory Management
Management Science 62(4), pp. 1188–1201, © 2016 INFORMS 1201

Using the strong duality theorem, we obtain the follow- Proof of Theorem 2. For any ˆ0 ≥ 0 value, SPR can be
ing equivalent model: recast as a shortest path problem on a network with O4T 2 5
T T  k T TX k
+1 X  nodes and O4T 2 5 acyclic directed arcs, which can be solved
d¯k
X X XX
min fk vk + grk ”rt + grk –rt in O4T 2 5 time. As already discussed, ˆ0 takes a value in
k=1 k=1 r=1 t=k r=k+1 t=1 801 ĝ11 1 0 0 0 1 ĝT +11 T 9 in an optimal solution. Thus, solving SPR
T
X
 as a shortest path problem for each distinct value of ˆ0
+ ˆ0 âT + 0k (74) (at most T 4T + 15 + 1 distinct values) and selecting the one
k=1 with minimum cost gives the optimal solution. ƒ
Downloaded from informs.org by [212.175.193.57] on 06 April 2016, at 04:35 . For personal use only, all rights reserved.

s.t. (67)–(73),
 k X
T TX
+1 k 
ˆ0 + 0k ≥ dˆk
X X
grk ”rt + grk –rt 1
r=1 t=k r=k+1 t=1 References
1≤k≤T1 (75) Atamtürk A (2006) Strong formulations of robust mixed 0-1 pro-
gramming. Math. Programming, Ser. B 108:235–250.
ˆ0 ≥ 01 0k ≥ 01 k ∈ T0 (76) Atamtürk A, Zhang M (2007) Two-stage network flow and design
under demand uncertainty. Oper. Res. 55:662–673.
Note that summing up constraint (67), constraints (68)
Ben-Tal A, Nemirovski A (1998) Robust convex optimization. Math.
from t = 2 to t = k, and constraints (69) from t = 1 to t = k
Oper. Res. 23:769–805.
gives
k XT TX
+1 Xk
Ben-Tal A, Nemirovski A (1999) Robust solutions to uncertain pro-
X
”rt + –rt = 11 1 ≤ k ≤ T 0 (77) grams. Oper. Res. Lett. 25:1–13.
r=1 t=k r=k+1 t=1 Ben-Tal A, Nemirovski A (2000) Robust solutions of linear pro-
gramming problems contaminated with uncertain data. Math.
We can rewrite 0k using (71), (72), and (77) as
Programming 88:411–424.
  k T TX
+1 Xk  
Ben-Tal A, Golany B, Shimrit S (2009) Robust multi-echelon multi-
0k = max dˆk
XX
grk ”rt + grk –rt − ˆ0 1 0 period inventory control. Eur. J. Oper. Res. 199:922–935.
r=1 t=k r=k+1 t=1
Ben-Tal A, Golany B, Nemirovski A, Vial JP (2005) Retailer-
 k T  supplier flexible commitments contracts: A robust optimization
= max dˆk
XX
grk ”rt − ˆ0 1 0 approach. Manufacturing Service Oper. Management 7:248–271.
r=1 t=k Ben-Tal A, Goryashko A, Guslitzer E, Nemirovski A (2004) Adjust-
 T +1 k  able robust solutions of uncertain linear programs. Math. Pro-
+ max dˆk
X X
grk –rt − ˆ0 1 0 gramming 99:351–376.
r=k+1 t=1 Bertsimas D, Sim M (2003) Robust discrete optimization and net-
k X
T
work flows. Math. Programming, Ser. B 98:49–71.
max8dˆk grk ”rt − ˆ0 1 09 Bertsimas D, Sim M (2004) The price of robustness. Oper. Res. 52:
X
=
r=1 t=k 35–53.
Bertsimas D, Thiele A (2006) A robust optimization approach to
TX
+1 k
inventory theory. Oper. Res. 54:150–168.
max8dˆk grk –rt − ˆ0 1 09
X
+
Bienstock D, Özbay N (2008) Computing robust basestock levels.
r=k+1 t=1
Discrete Optim. 5:389–414.
k X
T
Chen X, Zhang Y (2009) Uncertain linear programs: Extended
max8dˆk grk − ˆ0 1 09”rt
X
= affinely adjustable robust counterparts. Oper. Res. 57:1469–1482.
r=1 t=k
El-Ghaoui L, Lebret H (1997) Robust solutions to least-square prob-
TX
+1 k lems to uncertain data matrices. SIAM J. Matrix Anal. Appl.
max8dˆk grk − ˆ0 1 09–rt 0
X
+ (78) 18:1035–1064.
r=k+1 t=1 El-Ghaoui L, Oustry F, Lebret H (1998) Robust solutions to uncer-
Using (78), we can rewrite the SPR formulation as tain semidefinite programs. SIAM J. Optim. 9:33–52.
follows: Federgruen A, Tzur M (1999) Time partitioning heuristics: Appli-
T T  k T TX k
+1 X  cation to one warehouse, multi-item, multi-retailer lot-sizing
d¯k problems. Naval Res. Logist. 46:463–486.
X X XX
(SPR) min fk v k + grk ”rt + grk –rt
k=1 k=1 r=1 t=k r=k+1 t=1 Kouvelis P, Yu G (1997) Robust Discrete Optimization and Its Applica-
tions (Kluwer, Norwell, MA).
T k X
T 
+ ˆ0 âT +
X X
max8ĝrk − ˆ0 1 09”rt Pochet Y, Wolsey LA (2006) Production Planning by Mixed Integer
k=1 r=1 t=k
Programming (Springer, New York).
See CT, Sim M (2010) Robust approximation to multiperiod inven-
TX
+1 k 
X tory management. Oper. Res. 58:583–594.
+ max8ĝrk − ˆ0 1 09–rt
Soyster AL (1973) Convex programming with set-inclusive con-
r=k+1 t=1
straints and applications to inexact linear programming. Oper.
s.t. (67)–(73)1 ˆ0 ≥ 00 (79) Res. 21:1154–1157.