This article was downloaded by: [129.241.231.

104] On: 13 November 2018, At: 00:42

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

Adaptive Distributionally Robust Optimization

Dimitris Bertsimas, Melvyn Sim, Meilin Zhang

To cite this article:

Dimitris Bertsimas, Melvyn Sim, Meilin Zhang (2018) Adaptive Distributionally Robust Optimization. Management Science

Published online in Articles in Advance 09 May 2018

. https://doi.org/10.1287/mnsc.2017.2952

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 [email protected].

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 © 2018, 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
Articles in Advance, pp. 1–15
http://pubsonline.informs.org/journal/mnsc/ ISSN 0025-1909 (print), ISSN 1526-5501 (online)

Adaptive Distributionally Robust Optimization

Dimitris Bertsimas,a Melvyn Sim,b Meilin Zhangc
a Massachusetts Institute of Technology, Cambridge, Massachusetts 02139; b Department of Analytics and Operations Research,

NUS Business School, National University of Singapore, Singapore 119077; c SUSS Business School, Singapore University of Social Science,
Singapore 599494
Contact: [email protected], http://orcid.org/0000-0002-1985-1003 (DS); [email protected],
http://orcid.org/0000-0001-9798-2482 (MS); [email protected] (MZ)

Received: March 5, 2016 Abstract. We develop a modular and tractable framework for solving an adaptive distribu-
Revised: February 16, 2017; July 30, 2017 tionally robust linear optimization problem, where we minimize the worst-case expected
Accepted: August 21, 2017 cost over an ambiguity set of probability distributions. The adaptive distributionally robust
Published Online in Articles in Advance: optimization framework caters for dynamic decision making, where decisions adapt to
May 9, 2018 the uncertain outcomes as they unfold in stages. For tractability considerations, we focus
https://doi.org/10.1287/mnsc.2017.2952 on a class of second-order conic (SOC) representable ambiguity set, though our results
can easily be extended to more general conic representations. We show that the adaptive
Copyright: © 2018 INFORMS distributionally robust linear optimization problem can be formulated as a classical robust
optimization problem. To obtain a tractable formulation, we approximate the adaptive
distributionally robust optimization problem using linear decision rule (LDR) techniques.
More interestingly, by incorporating the primary and auxiliary random variables of the
lifted ambiguity set in the LDR approximation, we can significantly improve the solutions,
and for a class of adaptive distributionally robust optimization problems, exact solutions
can also be obtained. Using the new LDR approximation, we can transform the distribu-
tionally adaptive robust optimization problem to a classical robust optimization problem
with an SOC representable uncertainty set. Finally, to demonstrate the potential for solving
management decision problems, we develop an algebraic modeling package and illustrate
how it can be used to facilitate modeling and obtain high-quality solutions for medical
appointment scheduling and inventory management problems.

History: Accepted by Noah Gans, optimization.

Funding: This research was funded by the Singapore Ministry of Education [Social Science Research
Thematic Grant MOE2016-SSRTG-059].
Supplemental Material: The electronic companion is available at https://doi.org/10.1287/mnsc.2017.2952.

Keywords: dynamic programming • linear programming • nonlinear programming • quadratic • uncertainty • hospitals

1. Introduction relatively new approach that deals with ambiguity

Addressing uncertainty in many real-world optimiza- in mathematical optimization problems. In classical
tion problems has often lead to computationally intrac- robust optimization, uncertainty is described by a dis-
table models. As a result, uncertainty is often ignored tribution free uncertainty set, which is typically a conic
in optimization models, and this may lead to poor representable bounded convex set (see, for instance,
or even unacceptable decisions when implementing El Ghaoui and Lebret 1997; El Ghaoui et al. 1998;
them in practice. We characterize uncertainty as risk, Ben-Tal and Nemirovski 1998, 1999, 2000; Bertsimas
whenever the probability distribution is known, or and Sim 2004; Bertsimas and Brown 2009; Bertsimas
otherwise as ambiguity (Knight 1921). Traditionally, et al. 2011). The key advantage of a robust optimiza-
mathematical optimization models such as stochastic tion model is its computational tractability, and it has
programming (see, for instance, Birge and Louveaux been successful in providing computationally scalable
1997, Ruszczyński and Shapiro 2003) are based on solutions for a wide variety of management-inspired
the paradigm of risk, and they do not incorporate optimization problems.
ambiguity in their decision criteria for optimization. In evaluating preferences over risk and ambiguity,
However, with the growing importance of ambigu- Gilboa and Schmeidler (1989) propose a decision crite-
ity in decision making (see, for instance, Ellsberg rion that is based on the worst-case expected utility or
1961, Hsu et al. 2005), research on ambiguity has gar- disutility over an ambiguity set of probability distribu-
nered considerable interest in various fields includ- tions. Scarf (1958) was the first to study a single-product
ing economics, mathematical finance, and manage- newsvendor problem where the precise demand distri-
ment science. In particular, robust optimization is a bution is unknown but is only characterized by its mean
1
 Bertsimas, Sim, and Zhang: Adaptive Distributionally Robust Optimization
2 Management Science, Articles in Advance, pp. 1–15, © 2018 INFORMS

and variance. Subsequently, such models have been deal with uncertainty such as stochastic optimiza-
extended to minimax stochastic optimization models tion and dynamic optimization have been less suc-
(see, for instance, Žáčková 1966, Dupačová 1987, Breton cessful. Software packages that facilitate robust and
and EI Hachem 1995, Shapiro and Kleywegt 2002, distributionally robust optimization modeling have
Shapiro and Ahmed 2004) and recently to distribution- begun to surface in recent years. Existing toolboxes
ally robust optimization models (see, for instance, Chen include YALMIP (Löberg 2012), AIMMS (http://www
et al. 2007, Chen and Sim 2009, Popescu 2007, Delage .aimms.com/), ROME (Goh and Sim 2009, 2010),
and Ye 2010, Xu and Mannor 2012). In terms of tractable and JuMPeR (http://jumper.readthedocs.org). Of these,
formulations for a wide variety of single-stage convex ROME, AIMMS, and JuMPeR have also incorporated
optimization problems, Wiesemann et al. (2014) pro- LDR approximation.
pose a broad class of ambiguity sets where the family Our contributions to this paper are as follows:
of probability distributions are characterized by conic 1. We propose a tractable and scalable framework
representable expectation constraints and nested conic for solving an adaptive distributionally robust linear
representable confidence sets. optimization problem, where we minimize the worst-
Dynamic optimization models, where decisions case expected cost over a second-order conic (SOC)
adapt to the uncertain outcomes as they unfold in representable ambiguity set. We show that the adap-
stages, typically suffer from the “curse of dimension- tive distributionally robust linear optimization prob-
ality” and are computationally intractable (see, for in- lem can be formulated as a classical robust optimiza-
stance,Shapiro and Nemirovski 2005, Dyer and Stougie tion problem.
2006, Ben-Tal et al. 2004). To yield tractable models, 2. To obtain tractable formulation, we approximate
the linear decision rule (LDR), where adaptive deci- the adaptive distributionally robust linear optimiza-
sions are restricted to affine functions of the uncer- tion problem using LDR techniques. Depending on
tain parameters, was proposed in the early literature the choice of ambiguity set, the resulting framework
of stochastic programming. However, the technique is either a linear optimization problem or a second-
had been abandoned because of suboptimality of the order conic optimization problem (SOCP), which can
decision rule (see Garstka and Wets 1974). Neverthe- be solved efficiently by general-purpose commercial-
less, LDR approximation was revived by Ben-Tal et al. grade solvers such as CPLEX and Gurobi.
(2004) in their seminal work on adaptive robust opti- 3. We show that we can significantly improve the
mization.1 Subsequently, Bertsimas et al. (2010) estab- LDR approximation by incorporating the auxiliary
lish the optimality of LDR approximation in some random variable associated with the lifted ambigu-
important classes of adaptive robust optimization prob- ity set. This approach outperforms more sophisticated
lems. Chen and Zhang (2009) further improve the LDR decision rule approximations developed in Chen and
approximation by extending the affine dependency to Zhang (2009), Chen et al. (2008), See and Sim (2009),
the auxiliary variables associated with the support set. Goh and Sim (2010). Using the new LDR approxima-
For solving adaptive distributionally robust optimiza- tion, we can transform the adaptive distributionally
tion problems, Chen et al. (2007) propose tractable robust optimization problem to a classical robust opti-
formulations using LDR approximation techniques. mization problem with an SOC representable uncer-
Henceforth, variants of piecewise-linear decision rule tainty set.
approximation have been proposed to improve the 4. We demonstrate our approach for addressing a
approximation while maintaining the tractability of medical appointment scheduling problem as well as a
the adaptive distributionally robust optimization mod- multiperiod inventory control problem. In these prob-
els. Such approaches include the deflected and segre- lems, we also show that by incorporating partial cross-
gated LDR approximation of Chen et al. (2008), the moments information in the ambiguity set, we can
truncated LDR approximation of See and Sim (2009), significantly improve the solutions over alternatives
and the bideflected and (generalized) segregated LDR
found in the recent literature where the ambiguity set
approximation of Goh and Sim (2010). Interestingly,
is only characterized by marginal moments.
there is also a revival of using LDR approximation
for solving multistage stochastic optimization problems Notations. We use [N], N ∈  to denote the set of run-
(Kuhn et al. 2011). ning indices, {1, . . . , N }. We generally use boldfaced
For broader impact, a general-purpose optimiza- characters such as x ∈ N and A ∈ M×N to represent vec-
tion framework to address a wide variety of opti- tors and matrices, respectively, and [x]i or x i to denote
mization problems should be implementable in soft- the ith element of the vector x. We use (x)+ to denote
ware packages where reliable solutions can be obtained max{x, 0}. Special vectors include 0, 1, and ei , which
with reasonable computational effort. Compared to are, respectively, the vector of zeros, the vector of ones,
deterministic optimization frameworks such as linear and the standard unit basis vector. We denote RN, M as
optimization, classical optimization frameworks that the space of all measurable functions from N to M
Bertsimas, Sim, and Zhang: Adaptive Distributionally Robust Optimization
Management Science, Articles in Advance, pp. 1–15, © 2018 INFORMS 3

that are bounded on compact sets. We use P 0 (I ) to rep- recourse problems, which ensure that no outcome
resent the set of all probability distributions on I . A can produce infeasible results. However, there are
random variable, z̃, is denoted with a tilde sign, and also problems that would generally not satisfy com-
we use z̃ ∼  ,  ∈ P 0 (I ) to define z̃ as an I dimen- plete recourse, such as a production planning problem
sional random variable with probability distribution  . where a manager determines a production plan today
We denote Ɛ [ · ] as the expectation over the probability to satisfy all uncertain demands for tomorrow instead
distribution  . For a set W ⊆ I ,  [z̃ ∈ W ] represents of incurring penalty. For more information, we refer
the probability of z̃ being in the set W evaluated on the interested readers to Birge and Louveaux (1997).
distribution  . Typically, a weaker condition is assumed in stochas-
tic programming to ensure that the second-stage prob-
2. An Adaptive Distributionally Robust lem is essentially feasible.
Linear Optimization Problem Definition 2. The second-stage problem (1) has rela-
We first focus on a two-stage adaptive distribution- tively complete recourse if and only if the problem is fea-
ally robust linear optimization problem where the first- sible for all x ∈ X and z ∈ W .
stage or here-and-now decision is a vector x ∈ N1 chosen
Given an ambiguity set of probability distribu-
over the feasible set X. The cost incurred during the
tions, F , the second-stage cost is evaluated based
first stage in association with the decision x is deter-
on the worst-case expectation over the ambiguity set
ministic and given by c0 x, c ∈ N1 . In progressing to the
given by
next stage, the random variable z̃ ∼  ,  ∈ P 0 (I1 ) with
β(x)  sup Ɛ [Q(x, z̃)]. (2)
support W ⊆ I1 is realized; thereafter, we could deter-  ∈F
mine the cost incurred at the second stage. Similar to
Corresponding, the here-and-now decision, x, is deter-
a typical stochastic programming model, for a given
mined by minimizing the sum of the deterministic first-
decision vector, x, and a realization of the random vari-
stage cost and the worst-case expected second-stage
able, z ∈ W , we evaluate the second-stage cost via the
cost over the ambiguity set as follows:
following linear optimization problem,

Q(x, z)  min d0 y Z ∗  min {c0 x + β(x)}

(3)
s.t. A(z)x + By ≥ b(z), (1) s.t. x ∈ X.
y ∈ N2 . 2.1. The SOC Ambiguity Set
It has well been known that the tractability of a ro-
Here, A ∈ RI1 , M×N1 and b ∈ RI1 , M are functions that bust linear optimization problem is highly dependent
map from the vector z ∈ W to the input parameters on the choice of the uncertainty set. We will focus
of the linear optimization problem. We assume that on an SOC representable set, which encompasses a
these functions are affinely dependent on z ∈ I1 and polyhedron as a special case, and we can describe
are given by a wide variety of common uncertainty sets includ-
X X ing the Euclidean ball and the budgeted uncertainty
A(z)  A0 + Ak z k , b(z)  b0 + bk z k ,
k∈[I1 ] k∈[I1 ]
set of Bertsimas and Sim (2004). For a general con-
vex ambiguity set, we would refer readers to Ben-Tal
with A0 , A1 , . . . , AI1 ∈ M×N1 and b0 , b1 , . . . , bI1 ∈ M . et al. (2015) for the techniques of reformulating robust
The matrix B ∈ M×N2 , also known in stochastic pro- optimization problems via Fenchel duality, which can
gramming as the recourse matrix, and the vector d ∈ N2 also be applied to the analysis in this paper. Although
are constants, which correspond to the stochastic pro- general-purpose solvers are not available for more gen-
gramming format known as fixed recourse. Note that eral convex problems, this may benefit readers working
problem (1) may not always be feasible and, as in the on problems whose structural properties would allow
case of complete recourse, the recourse matrix can influ- the resulting reformulation to be solved even without
ence the feasibility of the second-stage problem. general-purpose solvers.
The robust counterpart corresponding to an SOC
Definition 1. The second-stage problem (1) has com-
uncertainty set is at most an SOCP, which is an estab-
plete recourse if and only if for any t ∈ M , there exists
lished optimization format supported by commercial
y ∈ N2 such that By ≥ t.
solvers that scale well with the number of variables in
Complete recourse is a strong sufficient condition the problem. Similar to the classical robust linear opti-
that guarantees the feasibility of the second-stage prob- mization problems, the tractability of a distribution-
lem for all x ∈ N1 and z ∈ I1 . Many real-life stochas- ally robust linear optimization problem is also depen-
tic programming models, including the newsvendor dent on the choice of the ambiguity set. In particular,
problem and its variants, are modeled as complete the ambiguity set based on information of moments,
 Bertsimas, Sim, and Zhang: Adaptive Distributionally Robust Optimization
4 Management Science, Articles in Advance, pp. 1–15, © 2018 INFORMS

notwithstanding its popularity, may not necessarily An important class of ambiguity set that could be
yield tractable distributionally robust counterparts. We modeled in the ambiguity set of Wiesemann et al.
propose an SOC representable ambiguity set where (2014) but not ours is the cross-moment ambiguity set
we restrict only to SOC representation. For generaliza- as follows:
tion to the ambiguity set of Wiesemann et al. (2014),
z̃ ∼ 
we refer interested readers to EC.5 of the electronic

 

Ɛ [z̃]  µ


 


companion. F CM   ∈ P 0 ( ) I1
.
 Ɛ [(z̃ − µ)(z̃ − µ)0 ]  Σ 
Definition 3. An SOC ambiguity set, F , is an ambiguity  [z̃ ∈ W ]  1

 

 
set of probability distributions that can be expressed as
Observe that the semidefinite constraint
 z̃ ∼  
Ɛ [(z̃ − µ)(z̃ − µ)0 ]  Σ
 
Ɛ [Gz̃]  µ


 


F   ∈ P 0 (I1 )  (4)
Ɛ [g i (z̃)] ≤ σi , ∀ i ∈ [I2 ]
is equivalent to the following semi-infinite quadratic
 
 [z̃ ∈ W ]  1

 

  constraints:
with parameters G ∈ L1 ×I1 , µ ∈ L1 , σ ∈ I2 , support set
Ɛ [(f 0 (z̃ − µ))2 ] ≤ f 0Σf, ∀ f ∈  I1 . (5)
W ∈ I1 , and functions g i ∈ RI1 , 1 , i ∈ [I2 ]. The support
set W is an SOC representable set and the epigraph of Hence, as a conservative approximation of the cross-
each g i , i ∈ [I2 ], moment ambiguity set, we propose the partial cross-
moment ambiguity set, which is an SOC ambiguity set
epi g i  {(z, u) ∈ I1 ×  | g i (z) ≤ u} as follows:
is an SOC representable set. 
 z̃∼ 

Ɛ [z̃]µ
 
I1 +1

 

As an illustration, for some given (f, h) ∈  , the F PCM   ∈P 0 ( )  0 I1
,
function g(z)  ((f 0 z + h)+ )3 is an SOC representable 
 Ɛ [(fk (z̃−µ))2 ]≤fk0 Σfk , ∀ k ∈[K] 

function because its epigraph is SOC representable

  [z̃∈W ]1 

given by
for some choice of parameters f1 , . . . , fK ∈ I1 . Observe
0

 v1 ≥0, v1 ≥f z+h 
 that the approximation never deteriorates with the
addition of new vectors, fk , k > K. In our applica-

 q 

 
v12 + 14 (v2 −1)2 ≤ 21 (v 2 +1)

 

epi g  (z, u)∈I1 × ∃ v∈2 : . tions to inventory control and appointment schedul-
 q 
ing problems, we will demonstrate how the partial
 
v22 + 14 (v1 −u)2 ≤ 41 (v1 +u) 

 
 
  cross-moment ambiguity set can yield tractable models
The formulation of SOC representable functions is a and provide far less conservative solutions than those
process that can be automated in an algebraic model- obtained from the marginal-moment ambiguity set, an
ing software package. For more information, we refer ambiguity set that does not consider cross-moment
interested readers to Ben-Tal and Nemirovski (2001a) information.
for an excellent reference on the algebra of SOC repre- As in Wiesemann et al. (2014), we also define the
sentable functions. The SOC ambiguity set provides a lifted ambiguity set, G, that encompasses the primary
useful and interesting characterization of distributions random variable z̃ and the lifted or auxiliary random
including: variable ũ as follows:
• Bounds on mean values: Ɛ [z̃] ∈ [µ, µ̄].  (z̃, ũ) ∼  
• Upper bound on absolute deviation: ¯ Ɛ [|f 0 z + h|] ≤ σ, 


 Ɛ [Gz̃]  µ




I1 +1
for some (f, h) ∈  . G   ∈ P 0 (I1 × I2 )  , (6)
 Ɛ [ũ] ≤ σ 
• Upper bound on variance: Ɛ [(f 0 z + h)2 ] ≤ σ, for some  [(z̃, ũ) ∈ W¯ ]  1

 

(f, h) ∈ I1 +1 .
 
• Upper bound on p-ordered deviation: Ɛ [(|f 0 z + h|)p ] ≤ where W¯ is the lifted support set defined as
σ, for some (f, h) ∈ I1 +1 and some rational p ≥ 1.
W¯  (z, u) ∈ I1 × I2 | z ∈ W , g(z) ≤ u ,

• Upper bound on semivariance: Ɛ [((f 0 z + h)+ )2 ] ≤ σ, (7)
for some (f, h) ∈ I1 +1 .
with g(z)  (g 1 (z), . . . , g I2 (z)). Observe that the lifted
• Approximate upper bound on entropy: Ɛ [exp(f 0 z)]
ambiguity set has only linear expectation constraints
≤ σ, for some f ∈ I1 . We refer readers to Ben-Tal and
and that the corresponding lifted support set is SOC
Nemirovski (2001a) for the approximate SOC representation.
representable.
• Upper bound on convex piecewise linear function:
Ɛ [maxp∈[P] {fp0 z + h p }] ≤ σ, for some (fp , h p ) ∈ I1 +1 , Proposition 1. The ambiguity set, F , is equivalent to the
p ∈ [P]. set of marginal distributions of z̃ under  , for all  ∈ G.
Bertsimas, Sim, and Zhang: Adaptive Distributionally Robust Optimization
Management Science, Articles in Advance, pp. 1–15, © 2018 INFORMS 5

Proof. The proof is similar to Wiesemann et al. (2014). The explicit formulation is given by
z̃ G ⊆ F , where
Q Q
We first show that z̃ G repre-
β(x)  min r + s 0µ + t 0 σ

sents the set of marginal distributions of z̃ under
any  ∈ G. Indeed, for any  G ∈ G, and  F is the s.t. r ≥ πi 0 h + pi 0 b0 − pi 0 A0 x, ∀ i ∈ [P]
marginal distribution of z̃ under  G , then we have  p0i (b1
− A x)  1

Ɛ G [Gz̃]  Ɛ F [Gz̃]  µ. Moreover, since  G [(z̃, ũ] ∈ W¯ ) C0 πi  

 ..  − G0s, ∀ i ∈ [P],
.
 1, we have  G [z̃ ∈ W ]  1 and  G [g(z̃) ≤ ũ]  1.

 p0 (bI1 − AI1 x) 
Hence,  F [z̃ ∈ W ]  1 and  i 
D πi  −t, ∀ i ∈ [P],
0

Ɛ F [g(z̃)]  Ɛ G [g(z̃)] ≤ Ɛ G [ũ[≤ σ. E0 πi  0, ∀ i ∈ [P],

πi K 0, ∀ i ∈ [P],
Conversely, suppose  F ∈ F ; we observe that
t ≥ 0,
 F [(z̃, g(z̃)) ∈ W¯ ]  1. Hence, we can then construct a
probability distribution  G ∈ P 0 (I1 × I2 ) for the ran- r ∈ , s ∈ L1 , t ∈ I2 ,
dom variable (z̃, ũ) ∼  G so that π i ∈ L 2 , ∀ i ∈ [P], (10)
where p1 , . . . , pP are the extreme points of the polyhedron, P,
(z̃, ũ)  (z̃, g(z̃))  -a.s.
given by
P  p ∈ +M : B0 p  d .

Observe that
Proof. Refer to EC.1 of the electronic companion. 
Ɛ G [ũ]  Ɛ G [g(z̃)] ≤ σ Remarks. Note that although similar results have been
derived in Isii (1962), Bertsimas and Popescu (2005),
and Shapiro and Kleywegt (2002), among others, our proof
 G [(z̃, ũ) ∈ W¯ ]  1. is based on conic duality and avoids analysis of func-
Hence, F ⊆
Q
G.  tional spaces. The assumption of relatively complete
z̃
recourse ensures that the dual of problem (1) is feasi-
2.2. Reformulation as a Classical Robust ble and finite so that it suffices to focus on the extreme
points of the polyhedron P.
Optimization Problem
Theorem 1 shows that we can transform the two-
For simplicity and clarity of the exposition, we will
stage adaptive distributionally robust optimization
focus on deriving the exact reformulation of β(x) for
problem to a well-known classical robust optimiza-
a relatively complete recourse problem. We can eas-
tion problem with an SOC representable uncertainty
ily incorporate the result to problem (3) to obtain the
set, W¯ . Hence, if W¯ is a polyhedron, then problem (9)
corresponding here-and-now decision, x ∈ X.
will become a linear optimization problem (LP). This
Assumption 1 (Slater’s Condition). Suppose that the lifted is the case, for instance, when the ambiguity set has
support set has the following explicit formulation: a hyper-rectangular support set and expectation con-
straints over absolute deviations. On the other hand, if
W¯  (z, u) ∈ I1 ×I2 | ∃ v ∈ I3 : Cz+Du+Ev K h , (8) the ambiguity set has expectation constraints over vari-

ances, the adaptive distributionally robust optimiza-
with C ∈ L2 ×I1 , D ∈ L2 ×I2 , E ∈ L2 ×I3 , h ∈ L2 , and K ⊆ L2 tion problem will become an SOCP. Nevertheless, we
is a Cartesian product of second-order cones. There exists note that problem (9) is generally an intractable prob-
z† , u† , v† such that lem unless the number of extreme points of P is small.
Hence, in the next section, we will show how we can
Gz†  µ, tractably compute an upper bound of β(x) using linear
u† < σ, decision rule (LDR) approximation techniques.
Cz† + Du† + Ev† ≺K h.
3. Linear Decision Rule (LDR)
Theorem 1. Suppose the second-stage problem (1) has rel- Approximation
atively complete recourse, then under Assumption 1, prob- Observe that we can express problem (2) as a mini-
lem (2) is equivalent to the following robust optimization mization problem over a measurable decision function
problem: map, y ∈ RI1 , N2 , as follows:
β(x)  min {r + s 0µ + t 0 σ} β(x)  min sup Ɛ [d0 y(z̃)]
s.t. r + s 0 (Gz) + t 0 u ≥ Q(x, z), ∀ (z, u) ∈ W¯ ,  ∈F

t ≥ 0, s.t. A(z)x + By(z) ≥ b(z), ∀z ∈ W ,

I1 , N2
r ∈ , s ∈ L1 , t ∈ I2 . (9) y∈R . (11)
 Bertsimas, Sim, and Zhang: Adaptive Distributionally Robust Optimization
6 Management Science, Articles in Advance, pp. 1–15, © 2018 INFORMS

However, problem (11) is generally an intractable opti- where

mization problem because y is a function map instead  ∃ y0 ,y1i ,y2j ∈N , ∀ i ∈[I1 ], j ∈[I2 ]: 
of a finite vector of decision variables. Nevertheless,
 
I1 +I2 , N

 X 2  
L̄  y∈R
N
yj u j  .
X 1
we can obtain an upper bound of the problem by  y(z,u)y + 0
yi z i +
i∈[I1 ] j∈[I2 ]
 
restricting y to a smaller class of function map or so-  
called decision rule, which can be characterized by a
polynomial number of decision variables. In a regular Theorem 2.
LDR approximation, the decision rule is restricted to β(x) ≤ β E (x) ≤ β L (x)
one that is affinely dependent on the primary random
variable z̃. Proof. Given the flexibility of the new LDR approxi-
Under the regular LDR approximation, we obtain an mation, it follows trivially that β E (x) ≤ β L (x). Let y∗ be
upper bound of β(x) by solving the following problem: an optimal decision rule of problem (15), and we define
the decision rule y† ∈ RI1 , N2 such that
β L (x)  min sup Ɛ [d0 y(z̃)]
 ∈F y† (z)  y∗ (z, g(z)).
s.t. A(z)x + By(z) ≥ b(z), ∀z ∈ W ,
Observe that y† (z̃) is feasible in problem (11) and,
y ∈ L N2 , (12) hence,
where β(x) ≤ sup Ɛ [d0 y† (z̃)]  sup Ɛ [d0 y∗ (z̃, g(z̃))]
 ∈F  ∈G

 ∃ y0 , y1i ∈ N , i ∈ [I1 ]: 
 0 ∗
≤ sup Ɛ [d y (z̃, ũ)]  β E (x). 
y ∈ R I1 , N

 

LN  yi z i  .
X 1
 y(z)  y + 0  ∈G
i∈[I1 ]
 
  Proposition 2. Problem (15) is equivalent to the affinely
Unfortunately, despite its simplicity, the LDR ap- adaptive robust optimization problem of Ben-Tal et al. (2004)
proximation may lead to infeasibility even if the prob- as follows:
lem has complete recourse (see, for instance, Chen et al.
β E (x)  min {r + s 0µ + t 0 σ}
2008, Goh and Sim 2009). As an illustration, we con-
sider the following complete recourse problem: s.t. r + s 0 (Gz) + t 0 u ≥ d0 y(z, u), ∀ (z, u) ∈ W¯ ,
A(z)x + By(z, u) ≥ b(z), ∀ (z, u) ∈ W¯ ,
β  min sup Ɛ [y(z̃)]
 ∈F t ≥ 0,
s.t. y(z) ≥ z, ∀ z ∈ , r ∈ , s ∈ L1 , t ∈ I2 ,
y(z) ≥ −z, ∀ z ∈ , y ∈ L̄ N2 . (16)
y∈R 1, 1
, (13)
Proof. We omit the proof as it follows straightfor-
where wardly from Theorem 1. 
F  { ∈ P 0 () | Ɛ [| z̃|] ≤ 1}. As a consequence of Proposition 2, we can trans-
Clearly, y(z)  |z| is the optimal decision rule that form an adaptive distributionally robust optimization
yields β  1. However, under the regular LDR approx- problem under the LDR approximation to a classical
imation, we require y(z)  y0 + y1 z for some y0 , y1 ∈ , affinely adaptive robust optimization problem. Hence,
which will be infeasible in the following set of semi- software packages for handling adaptive robust opti-
infinite constraints: mization problems such as AIMMS or JuMPeR can
also be used to address adaptive distributionally robust
y0 + y1 z ≥ z, ∀ z ∈ , optimization problems. Moreover, by incorporating the
(14)
y0 + y1 z ≥ −z, ∀ z ∈ . auxiliary random variables in the LDR approximation,
we are able to resolve the issue of infeasibility in com-
3.1. Incorporating Auxiliary Variables in the LDR
plete recourse problems.
We propose a powerful extension of the LDR ap-
proximation by incorporating the auxiliary random Theorem 3. Suppose problem (11) has complete recourse
variable ũ associated with the lifted ambiguity set. and the objective is bounded from below. For any ambiguity
Specifically, under this LDR approximation, we solve set, F , such that
the following problem:
Ɛ [| z̃ i |] < ∞, ∀ ∈ F ,
β E (x)  min sup Ɛ [d0 y(z̃, ũ)]
 ∈G there exists a lifted ambiguity set, G, whose corresponding
s.t. A(z)x + By(z, u) ≥ b(z), ∀ (z, u) ∈ W¯ , LDR is feasible in problem (15).
y ∈ L̄ N2 , (15) Proof. Refer to EC.2 of the electronic companion. 
Bertsimas, Sim, and Zhang: Adaptive Distributionally Robust Optimization
Management Science, Articles in Advance, pp. 1–15, © 2018 INFORMS 7

Remarks. Note that the class of ambiguity set depicted Hence, for all z ∈ W , y(z) ≤ −|z1 + z2 | + 2, and it is
in Theorem 3 encompasses any random variable with therefore feasible in the last two sets of constraints of
finite deviation—i.e., Ɛ [| z̃ i − µ i | p i ] < ∞ for some µ i ∈ , problem (17). We also note that there does not exist a
p i ≥ 1, i ∈ [I1 ], since feasible LDR in which y is affinely dependent on z—
i.e., y(z)  y0 + y1 z 1 + y2 z2 for some y0 , y1 , y2 ∈ .
Ɛ [| z̃ i |] ≤ Ɛ [| z̃ i − µ i |]+ |µ i | ≤ (Ɛ [| z̃ i − µ i | p i ])1/p i + |µ i | < ∞. Indeed, when substituting the extreme points of W in
problem (17), we will obtain the following set of infea-
More interestingly, we show in the following result that
sible equations:
the new LDR approximation can attain the optimal
objective values for a class of adaptive distributionally z1  z2  1 ⇒ y0 + y1 + y2  0
robust linear optimization problems.
z 1  z2  −1 ⇒ y0 − y1 − y2  0
Theorem 4. Suppose that problem (11) is a complete re- z 1  −z2  1 ⇒ y0 + y1 − y2  2
course problem with only one second-stage decision variable
(i.e., N2  1), then z 1  −z2  −1 ⇒ y0 − y1 + y2  2.

β(x)  β E (x). For any lifted ambiguity set, the corresponding lifted
support set W¯ is defined in (6), where the parameter u
Proof. Refer to EC.3 of the electronic companion.  is associated with the auxiliary random variable. Incor-
porating the auxiliary random variable in the LDR,
Remarks. Note that for complete recourse problem
we have
with N2  1, problem (9) becomes a tractable prob-
y(z, u)  y0 + y1 z + y2 u.
lem since the number of extreme points of the polyhe-
dron P equals to M. Nevertheless, notwithstanding the Since u is unbounded from above in the lifted ambigu-
simplicity, we are not aware of other types of decision ity set W¯ , the constraints
rules that would yield tight results for this instance.
A natural question is whether we could extend the y(z, u) ≥ z1 − z2 , ∀ (z, u) ∈ W¯
results of Theorems 3 and 4 to the case of relatively
imply y2 ≥ 0. Similarly, the constraints
complete recourse. However, this is not the case as
depicted in the following negative result even for the y(z, u) ≤ z 1 + z 2 + 2, ∀ (z, u) ∈ W¯
case of N2  1.
Proposition 3. There exists a relatively complete recourse
imply y2 ≤ 0. Hence, the LDR is reduced to one that is
problem with N2  1 for which problem (15) is infeasible only affinely dependent on z, which, as we have shown,
under any LDR that incorporates both the primary and aux- is infeasible in problem (17). 
iliary random variables associated with the lifted ambigu- Quite surprisingly, by simply extending the LDR
ity set. approximation to include the auxiliary random vari-
Proof. Consider the following problem: able of the lifted ambiguity set, we are able to attain
the optimum objective values for the class of complete
min 0 recourse problems described in Theorem 4. In con-
s.t. y(z) ≥ z 1 − z2 , ∀ z ∈ W , trast, the proposed deflected LDR (DLDR) approxima-
y(z) ≥ z 2 − z1 , ∀ z ∈ W , tions of Chen et al. (2008), Goh and Sim (2009), albeit
(17)
y(z) ≤ z 1 + z2 + 2, ∀ z ∈ W , more complex, do not necessarily recover the optimum
y(z) ≤ −z1 − z2 + 2, ∀ z ∈ W , objective values for this class of complete recourse
y ∈ R2, 1 , problems. In EC.4 of the electronic companion, we will
further demonstrate that the new LDR approximation
where can indeed improve over the more sophisticated DLDR
W  {z ∈ 2 | kzk∞ ≤ 1}. approximations.
We verify that problem (17) is one with relatively com- As a useful consequence of Theorem 4, the worst-
plete recourse by showing that y(z)  |z1 − z2 | is a fea- case expectation of a convex piecewise linear function,
sible solution. Indeed, y(z) is feasible in the first two h i
sets of constraints of problem (17). Moreover, for all Ɛ max{a0i (z̃)x + b i (z̃)} ,
i∈[K]
kzk∞ ≤ 1,
can also be expressed as
|z1 − z2 | + |z1 + z2 |
 max z1 − z2 + z1 + z2 , −z1 + z2 + z1 + z2 ,
 min sup Ɛ [y(z̃, ũ)]
 ∈G
z1 − z2 − z1 − z2 , −z1 + z2 − z 1 − z 2

s.t. y(z, u) ≥ ak (z)0 x + b k (z), ∀ k ∈ [K], (z, u) ∈ W¯ ,
 max{2z1 , 2z2 , −2z2 , −2z1 } ≤ 2. y ∈ L̄ , 1
 Bertsimas, Sim, and Zhang: Adaptive Distributionally Robust Optimization
8 Management Science, Articles in Advance, pp. 1–15, © 2018 INFORMS

which can be modeled directly using an algebraic mod- with the support set as proposed in Chen et al. (2008),
eling software package. In fact, this technique can Chen and Zhang (2009), and Goh and Sim (2010).
be applied straightforwardly to obtain exact solutions However, these approaches do not provide systematic
in adaptive distributionally robust optimization prob- ways of refining the approximations toward optimality.
lems found in recent applications such as Meng et al. More recently, Zhen et al. (2016) demonstrate how an
(2015) and Qi (2015). We will use the case study of med- adaptive robust or distributionally robust optimization
ical appointment scheduling to show how we could problem can be transformed to a static robust optimiza-
easily apply our results to studying various types of tion problems via Fourier–Motzkin elimination (FME).
ambiguity sets. For instance, without imposing complete recourse, if
N2  1, we can eliminate the only recourse variable via
3.2. Generalization to Multistage Problems FME, and solve the static robust optimization prob-
Another important feature of the LDR approximation lem to optimality in polynomial time. Although this
is the ability to easily enforce nonanticipativity in mul- approach would generally create an exponential num-
tistage decision making. For given subsets S i ⊆ [I1 ], ber of constraints, to keep the model tractable, we can
i ∈ [N2 ], that reflect the information dependency of the perform partial FME and apply our LDR approxima-
adaptive decisions, yi , we consider the generalization tion to improve the solutions. Hence, this generic refor-
of problem (11) as follows: mulation technique enhances our LDR approximation
and enables us to solve adaptive distributionally robust
γ(x)  min sup Ɛ [d0 y(z̃)] optimization problems to the level of optimality within
 ∈F the limits of the available computational resources.
s.t. A(z)x + By(z) ≥ b(z), ∀z ∈ W ,
yi ∈ RI1 (S i ), ∀ i ∈ [N2 ], (18) 3.4. On Interpreting Decision Rule as Policy and
the Issue of Time Consistency
where we define the space of restricted measurable For a given x ∈ X, let y∗ be the optimal decision rule
functions as of problem (11), and y∗E be the optimal decision rule of
problem (15). Consider a policy based on the decision
RI (S )  y ∈ RI,1 | y(v)  y(w), ∀ v, w ∈ I : rule y† ∈ RI1 , N2 such that


vj  wj, ∀ j ∈ S . y† (z)  y∗E (z, g(z)).

Problem (18) solves for the optimal decision rule Observe that y† (z̃) is feasible in problem (11), and from
y ∈ RI1 , N2 that minimizes the worst-case expected ob- the proof of Theorem 2, it follows that
jective taking into account the information depen- β(x) ≤ sup Ɛ [d0 y† (z̃)] ≤ β E (x).
dency requirement. Under the new LDR approxima-  ∈F

tion, we have Suppose β(x)  β E (x), which is the case for complete
recourse problems and N2  1; there is a tendency to
γE (x)  min sup Ɛ [d0 y(z̃, ũ)] infer the optimality of y† vis-à-vis y∗ so that
 ∈G
s.t.A(z)x + By(z, u)) ≥ b(z), ∀ (z, u) ∈ W¯ , d0 y† (z)  d0 y∗ (z), ∀z ∈ W .
yi ∈ L̄(S , T ), i i
∀ i ∈ [N2 ], (19) However, this is not the case, and we will demonstrate
this fallacy in the following simple example. Consider
where the following complete recourse problem:

 ∃ y 0 , yi1 , y 2j ∈, ∀ i ∈S , j ∈T : 
 β  min sup Ɛ [y( z̃)]
I1 +I2 ,1

 

L̄(S ,T ) y ∈R  ∈F
X 1 X 2
y(z,u) y 0 + yi z i + yj uj 

 i∈S j∈T
 s.t. y(z) ≥ z, ∀ z ∈ ,
 
y(z) ≥ −z, ∀ z ∈ ,
and the subsets T i ⊆ [I2 ], i ∈ [N2 ], are consistent with
y ∈ R1, 1 , (20)
the information restriction imposed by S i ⊆ [I1 ]. In Sec-
tion 5, we will illustrate how we can use this model where
to formulate a multiperiod inventory control problem.
F   ∈ P 0 () | z̃ ∼  , Ɛ [ z̃]  0, Ɛ [ z̃ 2 ] ≤ 1 .

We also refer readers to He et al. (2017) for modeling
and optimizing multiperiod repositioning decisions in Clearly, y ∗ (z)  |z| is the optimal solution, and it is also
vehicle sharing problems. the optimal objective value for all z ∈ . However, under
the LDR approximation, we obtain y † (z)  (1 + z 2 )/2,
3.3. Enhancements of LDR Approximations which is essentially greater than the optimum pol-
As further enhancement to the LDR approximation, icy y ∗ (z) except at z  1 and z  −1. Incidentally,
we can incorporate the auxiliary variables associated the worst-case distribution  ∈ F corresponds to the
Bertsimas, Sim, and Zhang: Adaptive Distributionally Robust Optimization
Management Science, Articles in Advance, pp. 1–15, © 2018 INFORMS 9

two point distributions with  [z̃  1]   [ z̃  −1]  1/2, reformulation of standard and distributionally robust
which explains why their worst-case expectations coin- counterparts using the techniques described in this
cide. This is similar to the observation of Delage and paper. Interestingly, XProg (http://xprog.weebly.com),
Iancu (2015) that the worst-case policy generated by a new MATLAB-based algebraic modeling package
decision rule can be inefficient, and such degeneracy is that implements our proposed framework, has inde-
common in robust multistage decision models. pendently emerged. The design of XProg is similar
Another issue with using the optimal decision rule to ROC. Since the MATLAB platform is a more user-
as a policy is the potential violation of time consistency. friendly environment, XProg can be used for rapid pro-
In dynamic decision making, time inconsistency arises totyping of models, while ROC would be better suited
when an optimal policy perceived in one time period to deployment of the solutions. The examples in our
may not be recognized as optimal in another. Delage numerical studies below can easily be implemented in
and Iancu (2015), Xin et al. (2015) show that in address- both ROC and XProg.
ing multiperiod robust or distributionally robust opti-
mization problems, time consistency may be affected 4. An Application in Medical
by how the ambiguity sets are being updated dynam-
Appointment Scheduling
ically. While time consistency is a desirable feature in
For the first application, we consider a medical appoint-
rational decision making, policies that may violate time
ment scheduling problem where patients arrive at their
consistency have also been justified in the literature
stipulated schedule and may have to wait in a queue
(see, for instance, Basak and Chabakauri 2010, Kydland
to be served by a physician. The patients’ consulta-
and Prescott 1977, Richardson 1989, Bajeux-Besnainou
tion times are uncertain, and their arrival schedules are
and Portait 1998).
determined at the first stage, which can influence the
Consequently, when solving the adaptive distri-
waiting times of the patients and the overtime of the
butionally robust optimization problem, we caution
physician.
against using the optimal decision rule as a policy.
To formulate the problem, we consider N patients
In many practical applications of dynamic decision
arriving in sequence with their indices j ∈ [N], and
making, it suffices to implement the here-and-now deci-
the uncertain consultation times are denoted by z̃ j ,
sion without having to commit to a policy that dic-
j ∈ [N]. We let the first-stage decision variable x j rep-
tates how the solution might change as uncertainty resent the interarrival time between patient j to the
unfolds. For a two-stage problem, the second-stage adjacent patient j + 1 for j ∈ [N − 1] and x N to denote
decision should be determined by solving a linear opti- the time between the arrival of the last patient and
mization problem after the uncertainty is resolved. In the scheduled completion time for the physician before
addressing a multistage decision problem, we advocate overtime commences. The first patient will be sched-
using the LDR approximation to obtain the here-and- uled to arrive at the starting time of zero, and subse-
now decision, x ∈ X, which accounts for how decisions quent patients i, i ∈ [N], i ≥ 2 will be scheduled to arrive
might adapt as uncertainty unfolds over the stages. P
at j∈[i−1] x j . Consequently, the feasible region of x is
As we proceed to the next stage, we should adopt the given by
folding horizon approach and solve for new here-and-  
X

now decision using the latest available information as X  x ∈ +N  xi ≤ T ,
inputs to another adaptive distributionally robust opti- i∈[N]

mization problem. where T is the scheduled completion time for the phy-
sician before overtime commences.
3.5. Software Packages A common decision criterion in the medical appoint-
As a proof of concept, we have develop the soft- ment schedule is to minimize the expected total cost
ware package ROC (http://www.meilinzhang.com/ of patients waiting and physician overtime, where the
software) to provide an intuitive environment for for- cost of a patient waiting is normalized to one per unit
mulating and solving our proposed adaptive distri- delay and the physician’s overtime cost is γ per unit
butionally robust optimization models. ROC is devel- delay. For a given arrival schedule x ∈ X, and a realiza-
oped in the C++ programming language, which is tion of consultation times z ∈ +N , the total cost can be
fast, highly portable, and well suited to deployment determined by solving the following linear optimiza-
in decision support systems. A typical algebraic mod- tion problem:
eling package provides the standardized format for  
declaration of decision variables, transcription of con- X
Q(x, z)  min yi + γ yN+1
straints and objective functions, and interface with i∈[N]
external solvers. ROC has additional features including
s.t. yi − yi−1 + x i−1 ≥ z i−1 , ∀ i ∈ {2, . . . , N + 1},
declaration of uncertain parameters and linear decision
rules, transcriptions of ambiguity sets, and automatic y ≥ 0, (21)
 Bertsimas, Sim, and Zhang: Adaptive Distributionally Robust Optimization
10 Management Science, Articles in Advance, pp. 1–15, © 2018 INFORMS

where yi denotes the waiting time of patient i, i ∈ [N], where

and yN+1 represents the overtime of the physician.   
z≥0
Note that the appointment scheduling problem is one W̄ MM  (z, u) ∈  ×  
 N N
Case[N]
that has complete recourse. From Theorem 1, we can
compute the worst-case expectation over an ambiguity with
s
set F , 2
ui + 1

2 ui − 1
min sup Ɛ [Q(x, z̃)], Case[N] : (z i − µ i ) + ≤ , ∀ i ∈ [N].
x∈X  ∈F 2 2
by enumerating all of the extreme points of the corre- In the proof of their Theorem 1, Mak et al. (2014)
sponding dual feasible set, show that the unconstrained dual variables that are

p i − p i−1 ≥ −1, ∀ i ∈ {2, . . . , N }
 associated with the second-moment equality con-
P p∈ +N . straints must be positive at optimality. Incidentally,
pN ≤ γ
when the problem is replaced by the relaxed ambigu-
However, given the exponentially large number of ity set, these variables would have nonnegative con-
extreme points of P, it would be generally intractable straints. Hence, optimizing over the relaxed ambiguity
to obtain exact solutions. set FMM would yield the same set of optimal solutions
Kong et al. (2013) is first to propose a distributional as those obtained via the ambiguity set F̄MM .
robust optimization for the appointment scheduling However, since the marginal-moment ambiguity set
problem. They consider a cross-moment ambiguity set is incapable of incorporating covariance, despite the
that characterizes the distributions of all nonnegative ease of computing the optimal solution, it may lead to
random variables with some specified mean values µ conservative solutions. As a compromise, we propose
and covariance Σ as follows: the following partial cross-moment SOC ambiguity set:

 z̃ ∼  
  z̃ ∼  
Ɛ [z̃]  µ
 
Ɛ [z̃]  µ


 

 
 

F̄CM   ∈ P 0 ( ) 
N
. (22)
 
Ɛ [(z − µ)(z − µ)0 ]  Σ

 

  F PCM   ∈ P 0 (N ) Ɛ [( z̃ i − µ i )2 ] ≤ σi2 , ∀ i ∈ [N] .
 [z̃ ∈ +N ]  1
 
Ɛ [(10 (z̃ − µ))2 ] ≤ 10Σ1
  
 

   
 [z̃ ∈ +N ]  1

 

As the problem is intractable, they formulate an SDP  
(25)
relaxation that solves the problem approximately.
The corresponding lifted ambiguity set G PCM is
To obtain a tractable formulation, Mak et al. (2014)
ignore information on covariance and consider a mar- 
 (z̃, ũ) ∼  

Ɛ [z̃]  µ
 
ginal-moment ambiguity set as follows: 







G PCM   ∈ P 0 (N × N+1 ) Ɛ [ ũ i ] ≤ σi2 , ∀ i ∈ [N] ,

 z̃ ∼  
 
 Ɛ [ ũ N+1 ] ≤ 10Σ1


Ɛ [z̃]  µ
   
 [(z̃, ũ) ∈ W¯ PCM ]  1

 
  
F̄MM   ∈ P 0 ( ) 
N
,
 
Ɛ [( z̃ i − µ i )2 ]  σi2 , ∀ i ∈ [N]
 
 
 [z̃ ∈ +N ]  1 where

 

 
(23) 
 z≥0 

where σi2 , i ∈ [N] is the variance of z̃ i . Surprisingly, Mak (z i − µ i )2 ≤ u i , ∀ i ∈ [N]

 



 


et al. (2014) show that the model has a hidden tractable W¯ PCM  (z, u) ∈ N × N+1  X 2 ,
reformulation, which they have cleverly exploited to (z i − µ i ) ≤ u N+1

 


 

obtain exact solutions. Observe that due to equality 
 i∈[N]


constraints on variances, F̄MM is not a SOC ambiguity (26)
set. Nevertheless, by relaxing the equality constraints
to inequality constraints, we will obtain the following or equivalently as a second-order cone representable
SOC ambiguity set: feasible set,

  z≥0 
z̃ ∼ 


  


W̄ PCM  (z, u) ∈ N ×N+1 ,

 
  Case[N]
Ɛ [z̃]  µ


 


FMM   ∈ P 0 ( ) 
N
.

  CaseN+1 
Ɛ [( z̃ i − µ i )2 ] ≤ σi2 ∀ i ∈ [N]
 
   
 [z̃ ∈ +N ]  1
 
where
 
 
(24) s 2
u i +1

Correspondingly, the lifted ambiguity set, G MM , is u −1
2
Case[N] : (z i − µ i ) + i ≤ , ∀ i ∈ [N];
2 2

 (z̃, ũ) ∼  
 v
Ɛ [z̃]  µ
t 2  2
u N+1 +1
 
u N+1 −1

 

G MM   ∈ P 0 (N × N )  ,
X
Ɛ [ ũ i ] ≤ σi2 , ∀ i ∈ [N] CaseN+1 : (z i − µ i ) + ≤ .
  2 2
 [(z̃, ũ) ∈ W¯ MM ]  1 i∈[N]

 

 
Bertsimas, Sim, and Zhang: Adaptive Distributionally Robust Optimization
Management Science, Articles in Advance, pp. 1–15, © 2018 INFORMS 11

For these SOC ambiguity sets, we can obtain approx- For each approach and α ∈ {0.00, 0.25, 0.50, 0.75,
imate solutions to the appointment scheduling prob- 1.00}, we obtain the objective values of seven randomly
lem based on the new LDR approximation as follows: generated instances. We report the results in Table 1.
  Observe that Regular LDR performs extremely
poorly. Indeed, as noted in Model (12), the regu-
X
min sup Ɛ yi (z̃, ũ) + γ yN+1 (z̃, ũ)
 ∈G i∈[N] lar LDR approximation is unable to incorporate most
s.t. yi (z, u) − yi−1 (z, u) + x i−1 ≥ z i−1 , of the information of the ambiguity set other than
∀ (z, u) ∈ W¯ , ∀ i ∈ {2, . . . , N + 1}, the mean, µ, and the nonnegative support set, which
leads to the ultraconservative result. We note that
y(z, u) ≥ 0, ∀ (z, u) ∈ W¯ , Exact MM improves marginally over Approx MM, and
x ∈ X, Approx PCM improves over Approx MM. For uncor-
y ∈ L̄ N+1 . (27) related or mildly correlated random variables—i.e.,
α ∈ {0, 0.25}—Approx PCM would then yield a lower
In our numerical study, we investigate the perfor- objective. Under perfect correlation—i.e., α  1—the
mance of the appointment scheduling problem among objective values of Approx PCM and Approx MM coin-
the following approaches: cide, and they are marginally higher than the objective
• (Regular LDR): Solutions based on the regular values of Exact MM. Hence, while there are benefits
LDR approximation. Note that regardless of whether
partial cross moments or marginal moments are used
to define the ambiguity set, the reformulation under Table 1. Objective Values Under Different Instances with
the regular LDR is the same as one with an ambiguity Varying Correlation Coefficient α
set that has only information on the mean and support
Objective value
of the random variable.
• (Exact MM): Exact solutions (Mak et al. 2014) Instance Regular Exact Approx Approx Approx
under the marginal-moment ambiguity set, F̄MM . α index LDR MM MM PCM CM
• (Approx MM): Solutions based on the new LDR 0.00 1 1,987.66 22.12 22.67 20.71 19.27
approximation under the relaxed marginal-moment 2 2,089.22 46.32 47.21 43.06 33.52
ambiguity set, FMM . 3 1,987.66 37.03 37.79 34.51 27.40
• (Approx PCM): Solutions based on the new LDR 4 2,089.22 66.32 66.88 61.00 49.28
approximation under the partial cross-moment ambi- 5 2,009.91 89.47 91.62 83.59 72.25
6 2,100.27 121.33 123.65 112.33 93.23
guity set, F PCM . 7 1,979.56 219.35 222.36 202.51 158.43
• (Approx CM): Solutions based on Kong et al. (2013) 0.25 1 1,987.66 22.12 22.67 21.76 20.17
conservative SDP approximation of the cross-moment 2 2,089.22 46.32 47.21 45.27 41.38
ambiguity set, F̄CM . Because of the instability of these 3 1,987.66 37.03 37.79 36.27 32.08
models, we use three different SDP solvers—namely, 4 2,089.22 66.32 66.88 64.13 56.56
SDPT3 (Tutuncu et al. 2003), SeDuMi (Sturm 1999), and 5 2,009.91 89.47 91.62 87.80 77.48
6 2,100.27 121.33 123.65 118.27 110.38
MOSEK—and report only the results with confirmed 7 1,979.56 219.35 222.36 213.09 188.86
status of optimality by at least one of the solvers. 0.50 1 1,987.66 22.12 22.67 22.39 21.12
The numerical settings of our computational experi- 2 2,089.22 46.32 47.21 46.59 43.20
ments are similar to Mak et al. (2014). We first consider 3 1,987.66 37.03 37.79 37.31 35.05
N  8 patients. The unit overtime cost is γ  2. For each 4 2,089.22 66.32 66.88 66.01 61.89
patient i ∈ [N], we randomly select µ i based on uni- 5 2,009.91 89.47 91.62 90.43 84.90
6 2,100.27 121.33 123.65 121.90 120.69
form distribution over [30, 60] and σi  µ i · , where  is 7 1,979.56 219.35 222.36 219.43 206.34
randomly selected based on uniform distribution over 0.75 1 1,987.66 22.12 22.67 22.66 21.11
[0, 0.3]. The covariance matrix is given by 2 2,089.22 46.32 47.21 47.17 43.86
( 3 1,987.66 37.03 37.79 37.76 36.45
ασi σ j if i , j, 4 2,089.22 66.32 66.88 66.82 64.19
[Σ]i j  5 2,009.91 89.47 91.62 91.55 88.16
σ2j otherwise, 6 2,100.27 121.33 123.65 123.53 115.36
7 1,979.56 219.35 222.36 222.18 213.55
where α ∈ [0, 1] is the correlation coefficient between 1.00 1 1,987.66 22.12 22.67 22.67 —
any two different random consultation times. The eval- 2 2,089.22 46.32 47.21 47.21 —
uation period, T, depends on the instance parameters 3 1,987.66 37.03 37.79 37.79 —
as follows: 4 2,089.22 66.32 66.88 66.88 —
s
N N 5 2,009.91 89.47 91.62 91.62 —
X X
T µ i + 0.5 σi 2 . 6 2,100.27 121.33 123.65 123.65 —
i1 i1
7 1,979.56 219.35 222.36 222.36 —
 Bertsimas, Sim, and Zhang: Adaptive Distributionally Robust Optimization
12 Management Science, Articles in Advance, pp. 1–15, © 2018 INFORMS

Table 2. Problem Scales Under Different Models Where N Jobs Presents

Regular LDR Exact MM Approx MM Approx PCM

(N + 1)N
Linear inequality constraints 2N 2 + N + 1 +1 2N 2 + N + 1 2N 2 + N + 1
2
(N + 1)N
SOC constraints 0 2N 2 + N 2N 2 + 3N + 1
2
1 2 1
Decision variables 2N 2 + 4N + 1 N +5 N +1 6N 2 + 8N + 1 6N 2 + 12N + 5
2 2

from the new LDR approximation, it does not replicate may not necessarily be optimally verified by the SDP
the optimal solution of Mak et al. (2014). solvers as the problem size increases. In contrast, the
In Table 2, we show how the size of the four tractable Approx PCM can be computed quickly and reliably,
models (Regular LDR, Exact MM, Approx MM, and and its solution consistently improves over Exact MM
Approx PCM) scales with the number of jobs, N. For and Approx MM for the case when α  0. Hence, this
Approx CM, unlike the previous approaches, we are underscores the importance of having a stable opti-
unable to obtain its optimally verified solutions by all mization format and reaffirms our restriction to the
three SDP solvers when the correlation is high. Among SOC ambiguity set.
the instances that the solutions of Approx CM could Similar to Mak et al. (2014), we also compare the
be optimally verified by at least one of the solvers, performance of the approaches in an out-of-sample
we observe that the corresponding objective values study using truncated normal and log normal prob-
attained are the lowest values among the approaches. ability distributions. We assume that the underly-
In Tables 3 and 4, we increase the number of jobs ing random variables are independently distributed
from N  12 to N  30, and report the objective values and that the parameters of the distributions corre-
and computational times for the different approaches. spond to the respective ambiguity sets for which
In this numerical study, we set α  0 and   0.15. the distributionally robust solutions are obtained. On
The results indicate that the approach with a tighter obtaining the solutions from the various methods, we
approximation also incurs a longer computational compare the performance and present the results in
time. Observe that it takes significantly longer time Table 5. These values are evaluated via Monte Carlo
to solve Approx CM and, more seriously, its solution simulation with 10,000 trials under each specific distri-
bution. Interestingly, despite the differences in objec-
tive values attained, the out-of-sample study alludes
Table 3. Objective Values Under Different Instances with to the closeness of results between Exact MM and
Increasing Jobs, N
Approx MM as well as between Approx PCM and
Models N  12 N  14 N  16 N  18 N  20 N  25 N  30 Approx CM. Since the random variables are indepen-
dent, and hence uncorrelated, as we have expected,
Exact 216.74 278.75 368.62 364.84 520.24 736.24 1,129.32 incorporating covariance or partial covariance infor-
MM
Approx 219.65 288.48 374.71 380.50 538.52 796.55 1,206.43
mation in the ambiguity would lead to improvement
MM in the out-of-sample performances.
Approx 196.64 255.07 329.23 333.67 459.48 608.02 889.51
PCM
Approx 143.87 179.88 226.84 224.29 301.04 419.72 — 5. An Application in Multiperiod
CM Inventory Control
For the second application, we illustrate how the
adaptive distributionally robust linear optimization
Table 4. Computation Time Under Different Instances with framework can model a multistage decision problem.
Increasing Jobs, N
We consider a finite-horizon, T period single-product
Models N  12 N  14 N  16 N  18 N  20 N  25 N  30 inventory control problem. Demands are filled from
on-hand inventory, and unfilled demands are fully
Exact <1 <1 <1 <1 <1 <1 <1 backlogged. At the beginning of period t, a quantity of
MM
Approx <1 <1 <1 <1 <1 <1 1
x t ∈ [0, x̄ t ] is ordered, which will arrive immediately to
MM replenish the stock before the demand is realized. The
Approx <1 <1 <1 <1 <1 <1 2 unit ordering cost is c t , excess inventory will incur a
PCM per-unit holding cost of h t , while backlogged demand
Approx 7 15 31 57 123 378 — will be penalized with a per-unit underage cost of b t .
CM
At the last period, t  T, lost sales could be accounted
Bertsimas, Sim, and Zhang: Adaptive Distributionally Robust Optimization
Management Science, Articles in Advance, pp. 1–15, © 2018 INFORMS 13

Table 5. Results of Monte Carlo Simulation Under as follows:

Truncated Normal and Log Normal Distributions  T
X

min sup Ɛ (c t x t (z̃t−1 ) + v t (z̃t ))
Instance Objective Truncated Log  ∈F t1
index Models value normal normal
s.t.yt+1 (zt )  yt (zt−1 ) + x t (zt−1 ) − dt (zt ),
1 Regular LDR 1,987.66 1,306.37 (0.58) 1,306.36 (0.61) ∀ z ∈ W , t ∈ [T],
Exact MM 22.12 11.12 (0.25) 11.07 (0.3)
Approx MM 22.67 11.08 (0.30) 11.16 (0.31) v t (zt ) ≥ h t yt+1 (zt ), ∀ z ∈ W , t ∈ [T],
Approx PCM 20.71 10.99 (0.26) 10.10 (0.27) v t (zt ) ≥ −b t yt+1 (zt ). ∀ z ∈ W , t ∈ [T],
Approx CM 19.27 10.68 (0.25) 10.07 (0.26)
0 ≤ x t (zt−1 ) ≤ x̄ t , ∀ z ∈ W , t ∈ [T],
2 Regular LDR 2,089.22 1,328.41 (1.20) 1,328.40 (1.17)
Exact MM 46.32 22.17 (0.50) 22.12 (0.51) x t ∈ Rt−1, 1 , yt+1 ∈ Rt, 1 , v t ∈ Rt, 1 , ∀ t ∈ [T]. (28)
Approx MM 47.21 22.27 (0.51) 22.48 (0.52)
Approx PCM 43.06 22.09 (0.49) 20.19 (0.51) We consider the following partial cross-moment
Approx CM 33.52 22.12 (0.51) 19.86 (0.52) ambiguity set,
3 Regular LDR 1,987.66 1,307.89 (0.96) 1,307.88 (1.01)
Exact MM 37.03 18.75 (0.41) 18.25 (0.42)

 z̃ ∼  

Ɛ [z̃]  0

 

Approx MM 37.79 18.47 (0.40) 18.65 (0.41)

 


 2 

Approx PCM 34.51 16.72 (0.41) 16.90 (0.42)  
≤ φ 2st ,

  Pt 

Approx CM 27.40 16.67 (0.40) 16.76 (0.41) F PCM   ∈ P 0 ( ) Ɛ T
rs z̃ r , (29)
4 Regular LDR 2,089.22 1,329.58 (1.70) 1,329.54 (1.68)

 

∀ s ≤ t, s, t ∈ [T] 

 
Exact MM 66.32 31.13 (0.76) 31.36 (0.78) 
 

 [z̃ ∈ [z, z̄]]  1
 
Approx MM 66.88 31.55 (0.73) 31.96 (0.76) 
 

Approx PCM 61.00 28.32 (0.74) 28.70 (0.78)  ¯ 
Approx CM 49.28 28.20 (0.76) 28.60 (0.78) where the lifted ambiguity set, G, is
5 Regular LDR 2,009.91 1,323.02 (2.52) 1,323.23 (2.56)
Exact MM 89.47 42.08 (0.96) 43.93 (1.03) G PCM
Approx MM 91.62 43.10 (0.93) 44.09 (0.99) 
 (z̃, ũ)∼ 

Approx PCM 83.59 39.32 (0.97) 40.26 (1.05) 

 Ɛ [z̃]0



Approx CM 72.25 38.94 (0.98) 39.15 (1.03) 
 

  ∈P 0 (T ×(T+1)T/2 ) Ɛ [ ũ st ]≤ φ 2st , ,
6 Regular LDR 2,100.27 1,357.65 (3.03) 1,357.59 (3.15)
Exact MM 121.33 55.69 (1.23) 56.15 (1.28)



 ∀ s ≤ t, s, t ∈[T] 



Approx MM 123.65 56.45 (1.19) 57.75 (1.25)

  [(z̃, ũ)∈ W̄ PCM ]1 

Approx PCM 112.33 50.66 (1.18) 51.87 (1.32) ¯
W PCM
Approx CM 88.42 48.23 (1.22) 51.07 (1.31)
2
,
Pt
7 Regular LDR 1,979.56 1,271.30 (5.87) 1,272.27 (6.13)




 u st ≥ rs z r





Exact MM 219.35 103.36 (2.51) 107.89 (3.05)  (z,u)∈T ×(T+1)T/2 ∀ s ≤ t, s, t ∈[T] . (30)
Approx MM 222.36 105.61 (2.49) 109.97 (2.98)  
z∈[z, z̄]
 
Approx PCM 202.51 96.49 (2.52) 100.54 (3.11)
 
Approx CM 156.15 95.46 (2.50) 99.26 (3.14)
 ¯ 
This partial cross moment incorporates the variances
Note. Standard errors are reported in parentheses. of the sum of factors leading to the time period t. We
let ũt  ( ũ rs )1≤r≤s≤t , t ∈ [T], which, together with z̃t , are
associated with the information available at the end of
for via the backorder cost. We use yt to indicate the net period t. Consequently, we use ROC to formulate the
inventory level at the beginning of period t. The initial problem via the new LDR approximation and solve it
net inventory level of the system is y1  0. using CPLEX.
As in Graves (1999) and See and Sim (2009), we Following the numerical study of See and Sim (2009),
model the demand process as an integrated moving we set the parameters x¯t  260, c t  0.1, h t  0.02, for
average (IMA) process of order (0, 1, 1) as follows: all t ∈ [T], b t  0.2, for all t ∈ [T − 1], and b T  2. We
assume that random factors z̃ t are uncorrected random
dt (z̃)  z̃ t + α z̃ t−1 + α z̃ t−2 + · · · + α z̃1 + µ variables in [− √z̄, z̄] with standard deviations bounded
 dt−1 (z̃) − (1 − α) z̃ t−1 + z̃ t , below by (1/ 3) z̄. In characterizing the partial cross-
moment ambiguity set, we have µ t  µ, t ∈ [T] and
for t ∈ [T], where α ∈ [0, 1] and the uncertain factors, z̃ t , t −s +1 2
φ 2st  z̄ , ∀ s ≤ t, s, t ∈ [T].
are zero means and uncorrelated random variables. 3
Hence, with α  0, the demands are uncorrelated. As Observe that i.i.d. uniformly distributed random vari-
α grows, the variances and correlation of the demands ables in [− z̄, z̄] would be a feasible distribution in
also increase. the ambiguity set, and we use this to obtain a lower
The objective of the problem is to minimize the bound to the inventory control problem. Specifically,
worst-case expected total cost over the entire horizon we investigate the performance of the multiperiod
 Bertsimas, Sim, and Zhang: Adaptive Distributionally Robust Optimization
14 Management Science, Articles in Advance, pp. 1–15, © 2018 INFORMS

Table 6. Objective Values of the Various Models Under choice of such an ambiguity set appears ad hoc, and it
Different Instances begs an interesting question as to how we can system-
atically adapt and improve the partial cross-moments
Lower Approx Approx
ambiguity set. Chen et al. (2016) have recently pro-
T µ z̄ α bound MM PCM
posed a new class of infinitely constrained ambigu-
5 200 20 0 108.0 167.3 115.7 ity sets where the number of expectation constraints
10 200 10 0 206.0 272.5 214.9 could be infinite. To solve the problem, they consider
20 240 6 0 486.0 583.2 499.2
30 240 4 0 725.0 838.6 740.8
a relaxed ambiguity set with a finite number of expec-
5 200 20 0.25 108.0 181.0 124.8
tation constraints, as in the case of the partial cross-
10 200 10 0.25 206.0 303.7 232.8 moments ambiguity set. More interestingly, for static
20 240 6 0.25 487.0 684.7 543.6 robust optimization problems, the “violating” expec-
30 240 4 0.25 725.0 1028.8 811.2 tation constraint can be identified and added to the
5 200 20 0.50 109.0 195.1 133.6 relaxed ambiguity set to improve the solution. While the
10 200 10 0.50 207.0 335.2 250.7 approach works for static distributionally robust opti-
20 240 6 0.50 496.0 795.1 588.4
30 240 4 0.50 732.0 1232.2 882.9
mization problems, the extension to adaptive problems
has not been studied. We believe that this is an impor-
inventory control problem among different approaches tant extension of this framework that will help us model
as follows: and solve adaptive distributionally robust optimization
• (Lower Bound): A lower bound obtained by using problems for a larger variety of ambiguity sets.
i.i.d. uniformly distributed random factors and solving
Acknowledgments
the dynamic inventory control problem to optimality The authors thank department editor Noah Gans, the anony-
(for the dynamic programming implementation, see mous associate editor, and the reviewers for their valuable
See and Sim 2009). and insightful comments. Any opinions, findings, and con-
• (Approx MM): Solutions based on the new LDR clusions or recommendations expressed in this material are
approximation under the marginal-moment ambiguity those of the authors and do not reflect the views of the Sin-
set (i.e., known mean values, upper bound on vari- gapore Ministry of Education or the Singapore Government.
ances, and nonnegative support).
• (Approx PCM): Solutions based on the new LDR Endnote
1
approximation under the partial cross-moment ambi- Note that we prefer the term “adaptive” over “adjustable” as used
in Ben-Tal et al. (2004).
guity set, F PCM .
In Table 6, we report the objective values attained for
References
the different approaches under various parameters. As
Alizadeh F, Goldfarb D (2003) Second-order cone programming.
in the previous computational study, we observe that Math. Programming 95(1):2–51.
by incorporating partial cross-moment information, we Bajeux-Besnainou I, Portait R (1998) Dynamic asset allocation in
can significantly improve the objectives of the adaptive a mean-variance framework. Management Sci. 44(11, Part 2):
distributionally robust optimization problems. More- S79–S95.
Basak S, Chabakauri G (2010) Dynamic mean-variance asset alloca-
over, the objectives of Approx PCM are reasonably tion. Rev. Financial Stud. 23(8):2970–3016.
close to the lower bounds. It has well been known Ben-Tal A, Nemirovski A (1998) Robust convex optimization. Math.
that despite the gaps from the lower bounds, numeri- Oper. Res. 23(4):769–805.
Ben-Tal A, Nemirovski A (1999) Robust solutions of uncertain linear
cal experiments on robust inventory control problems
programs. Oper. Res. Lett. 25(1):1–13.
have demonstrated that the actual objectives attained Ben-Tal A, Nemirovski A (2000) Robust solutions of linear program-
in out-of-sample analysis are often closer to the optimal ming problems contaminated with uncertain data. Math. Pro-
values than what the model objectives have reflected gramming Ser. A 88(3):411–424.
Ben-Tal A, Nemirovski A (2001a) Lectures on Modern Convex
(see, for instance, Bertsimas and Thiele 2006, See Optimization—Analysis, Algorithms, and Engineering Applications
and Sim 2009). Moreover, the benefit of distributional (Society for Industrial and Applied Mathematics, Philadelphia).
robustness arises when there is disparity between the Ben-Tal A, Nemirovski A (2001b) On polyhedral approximations of
actual demand distribution and the demand distribu- the second-order cone. Math. Oper. Res. 26(2):193–205.
Ben-Tal A, den Hertog D, Vial J (2015) Deriving robust counter-
tion in which the optimal policy is derived. In such parts of nonlinear uncertain inequalities. Math. Programming
cases, the robust solution could perform significantly 149(1):265–299.
better than the misspecified optimum policy (see the Ben-Tal A, Goryashko A, Guslitzer E, Nemirovski A (2004) Adjust-
numerical experiments of Bertsimas and Thiele 2006). able robust solutions of uncertain linear programs. Math. Pro-
gramming 99(2):351–376.
Bertsimas D, Brown DB (2009) Constructing uncertainty sets for
6. Future Work robust linear optimization. Oper. Res. 57(6):1483–1495.
Bertsimas D, Popescu I (2005) Optimal inequalities in probability
In our numerical studies, we show the benefits of theory: A convex optimization approach. SIAM J. Optim. 15(3):
the partial cross-moments ambiguity set. However, the 780–804.
Bertsimas, Sim, and Zhang: Adaptive Distributionally Robust Optimization
Management Science, Articles in Advance, pp. 1–15, © 2018 INFORMS 15

Bertsimas D, Sim M (2004) The price of robustness. Oper. Res. 52(1): Knight FH (1921) Risk, Uncertainty and Profit (Hart, Schaffner &
35–53. Marx, Boston).
Bertsimas D, Thiele A (2006) A robust optimization approach to Kong Q, Lee CY, Teo CP, Zheng Z (2013) Scheduling arrivals to a
inventory theory. Oper. Res. 54(1):150–168. stochastic service delivery system using copositive cones. Oper.
Bertsimas D, Brown DB, Caramanis C (2011) Theory and applications Res. 61(3):711–726.
of robust optimization. SIAM Rev. 53(3):464–501. Kuhn D, Wiesemann W, Georghiou A (2011) Primal and dual lin-
Bertsimas D, Iancu DA, Parrilo PA (2010) Optimality of affine poli- ear decision rules in stochastic and robust optimization. Math.
cies in multistage robust optimization. Math. Oper. Res. 35(2):
Programming 130(1):177–209.
363–394.
Birge JR, Louveaux F (1997) Introduction to Stochastic Programming Kydland FE, Prescott EC (1977) Rules rather than discretion: The
(Springer, New York). inconsistency of optimal plans. J. Political Econom. 85(3):473–491.
Breton M, El Hachem S (1995) Algorithms for the solution of stochastic Löberg J (2012) Automatic robust convex programming. Optim. Meth-
dynamic minimax problems. Comput. Optim. Appl. 4(4):317–345. ods Software 27(1):115–129.
Chen W, Sim M (2009) Goal-driven optimization. Oper. Res. 57(2): Mak HY, Rong Y, Zhang J (2014) Appointment scheduling with lim-
342–357. ited distributional information. Management Sci. 61(2):316–334.
Chen X, Zhang Y (2009) Uncertain linear programs: Extended affinely Meng F, Qi J, Zhang M, Ang J, Chu S, Sim M (2015) A robust opti-
adjustable robust counterparts. Oper. Res. 57(6):1469–1482. mization model for managing elective admission in a public
Chen X, Sim M, Sun P (2007) A robust optimization perspective on hospital. Oper. Res. 63(6):1452–1467.
stochastic programming. Oper. Res. 55(6):1058–1071. Popescu I (2007) Robust mean-covariance solutions for stochastic
Chen Z, Sim M, Xu H (2016) Distributionally robust optimization optimization. Oper. Res. 55(4):98–112.
with infinitely constrained ambiguity sets. Optim. Online. Qi J (2015) Mitigating delays and unfairness in appointment systems.
Chen X, Sim M, Sun P, Zhang J (2008) A linear decision-based
Management Sci. 63(2):566–583.
approximation approach to stochastic programming. Oper. Res.
56(2):344–357. Richardson HR (1989) A minimum variance result in continuous
Delage E, Iancu D (2015) Robust multistage decision making. trading portfolio optimization. Management Sci. 35(9):1045–1055.
INFORMS Tutorials Oper. Res. (2015):20–46. Ruszczyński A, Shapiro A, eds. (2003) Stochastic programming.
Delage E, Ye Y (2010) Distributionally robust optimization under Handbooks in Operations Research and Management Science, Vol. 10
moment uncertainty with application to data-driven problems. (Elsevier Science, Amsterdam).
Oper. Res. 58(3):596–612. Scarf H (1958) A min-max solution of an inventory problem.
Dupačová J (1987) The minimax approach to stochastic program- Arrow K, ed. Studies in the Mathematical Theory of Inventory and
ming and an illustrative application. Stochastics 20(1):73–88. Production (Stanford University Press, Stanford, CA), 201–209.
Dyer M, Stougie L (2006) Computational complexity of stochas- See C-T, Sim M (2009) Robust approximation of multiperiod inven-
tic programming problems. Math. Programming Ser. A 106(3): tory management. Oper. Res. 58(3):583–594.
423–432. Shapiro A, Ahmed S (2004) On a class of minimax stochastic pro-
El Ghaoui L, Lebret H (1997) Robust solutions to least-squares prob- grams. SIAM J. Optim. 14(4):1237–1249.
lems with uncertain data. SIAM J. Matrix Anal. Appl. 18(4):
Shapiro A, Kleywegt A (2002) Minimax analysis of stochastic pro-
1035–1064.
grams. Optim. Methods Software 17(3):523–542.
El Ghaoui L, Oustry F, Lebret H (1998) Robust solutions to uncertain
semidefinite programs. SIAM J. Optim. 9(1):33–53. Shapiro A, Nemirovski A (2005) On complexity of stochastic pro-
Ellsberg D (1961) Risk, ambiguity and the savage axioms. Quart. J. gramming problems. Jeyakumar V, Rubinov A, eds. Continuous
Econom. 75(4):643–669. Optimization (Springer, New York), 111–146.
Garstka SJ, Wets RJ-B (1974) On decision rules in stochastic program- Sturm JF (1999) Using SeDuMi 1.02, a MATLAB toolbox for optimi-
ming. Math. Programming 7(1):117–143. zation over symmetric cones. Optim. Methods Software 11(1–4):
Gilboa I, Schmeidler D (1989) Maxmin expected utility with non- 625–653.
unique prior. J. Math. Econom. 18(2):141–153. Tutuncu RH, Toh KC, Todd MJ (2003) Solving semidefinite-quadra-
Goh J, Sim M (2009) Robust optimization made easy with ROME. tic-linear programs using SDPT3. Math. Programming Ser. B 95(2):
Oper. Res. 59(4):973–985. 189–217.
Goh J, Sim M (2010) Distributionally robust optimization and its Wiesemann W, Kuhn D, Sim M (2014) Distributionally robust convex
tractable approximations. Oper. Res. 58(4):902–917. optimization. Oper. Res. 62(6):1358–1376.
Graves SC (1999) A single-item inventory model for a nonstation- Xin L, Goldberg DA, Shapiro A (2015) Time (in)consistency of mul-
ary demand process. Manufacturing Service Oper. Management
tistage distributionally robust inventory models with moment
1(1):50–61.
constraints. https://arxiv.org/abs/1304.3074.
He L, Hu Z, Zhang M (2017) Robust Repositioning for Vehi-
cle Sharing. Working paper, National University of Singapore, Xu H, Mannor S (2012) Distributionally robust Markov decision pro-
Singapore. cesses. Math. Oper. Res. 37(2):288–300.
Hsu M, Bhatt M, Adolphs R, Tranel D, Camerer CF (2005) Neural sys- Žáčková J (1966) On minimax solution of stochastic linear program-
tems responding to degrees of uncertainty in human decision- ming problems. Časopis pro Pěstování Matematiky 91:423–430.
making. Science 310(5754):1680–1683. Zhen J, Sim M, den Hertog D (2016) Adjustable robust optimization
Isii K (1962) On sharpness of Tchebycheff-type inequalities. Ann. via Fourier-Motzkin elimination. Optim. Online, http://www
Institute Statist. Math. 14(1):185–197. .optimization-online.org/DB_HTML/2016/07/5564.html.