You are on page 1of 11

Decision Support

Robust regret for uncertain linear programs with application


to co-production models
q
Tsan Sheng Ng

Department of Industrial and Systems Engineering, National University of Singapore, Singapore 119260, Singapore
a r t i c l e i n f o
Article history:
Received 23 May 2011
Accepted 14 January 2013
Available online 23 January 2013
Keywords:
Uncertainty modelling
Linear programming
Minimax regret
a b s t r a c t
This paper considers the regret optimization criterion for linear programming problems with uncertainty
in the data inputs. The problems of study are more challenging than those considered in previous works
that address only interval objective coefcients, and furthermore the uncertainties are allowed to arise
from arbitrarily specied polyhedral sets. To this end a safe approximation of the regret function is devel-
oped so that the maximum regret can be evaluated reasonably efciently by leveraging on previous
established results and solution algorithms. The proposed approach is then applied to a two-stage co-pro-
duction newsvendor problem that contains uncertainties in both supplies and demands. Computational
experiments demonstrate that the proposed regret approximation is reasonably accurate, and the corre-
sponding regret optimization model performs competitively well against other optimization approaches
such as worst-case and sample average optimization across different performance measures.
2013 Elsevier B.V. All rights reserved.
1. Introduction
Linear programming is a widely used approach to model large-
scale industrial optimization problems, such as inventory, produc-
tion planning, and workforce planning problems. Given the re-
quired data and parameter inputs, linear programs can be solved
very efciently using off-the-shelf platforms to generate planning
solutions. However, in reality, there are often uncertainties associ-
ated with the data parameters required in the model. In particular,
production problems involve parameters such as demands, yields
coefcients, costs that can be uncertain at the time when decisions
need to be generated and implemented. A solution obtained based
on a single realization of the input data (e.g. average case data) can
be of inferior quality. To mitigate the effects of planning uncertain-
ties, stochastic optimization modelling is a well-known approach
adopted among researchers and practitioners, where the uncertain
parameters are modelled as random variables arising from some
assumed distributions. On the other hand, in many practical situa-
tions such as the case assumed in this paper, there may only be
very limited information regarding the uncertain parameters avail-
able. For instance, only the bounds or support sets of the uncertain
parameters may be known. In such cases, approaches such as ro-
bust optimization, which is essentially a worst-case planning ap-
proach, can be applied. However, when only support set
information is used, robust optimization solutions may turn out
to be overly-conservative and performs poorly in the many situa-
tions. As an alternative, in this work, rather than hedging against
worst-case realizations, another approach termed as regret optimi-
zation, (or minimax regret, or the Savage Criterion (Savage, 1951))
is considered for the uncertain linear programming problem.
1.1. Regret optimization in decision-making under uncertainty
The inuence of regret on human decision-making behavior has
received extensive research interest in various elds, including
experimental psychology (Kahneman and Tversky, 1982; Connolly
and Zeelenberg, 2002), economics (Loomes and Sugden, 1987),
consumer research (Simonson, 1992), organizational behavior (Ri-
tov, 1996) and operations (Bell, 1982). Regret in decision-making
under uncertainty is intimately related to the value of information,
i.e. the discrepancy between the actual cost of a decision made
using partial information, and the decision made with full knowl-
edge of the future outcome. In many important situations, policies
and decisions are often subjected to post-audits (Gulliver, 1987;
Neale and Holmes, 1990; Azzone and Maccarrone, 2001), and
achieving low regret in decisions are desirable since managers
want to avoid being viewed as having exercised poor judgment.
This motivates the use of prescriptive regret models as a decision
support tools for planning under uncertainties.
Minimax regret optimization has been applied to various prob-
lems in operations research, e.g. production scheduling (Daniels
and Kouvelis, 1995), portfolio selection (Inuiguchi and Tanino,
2000), knapsack problems (Conde, 2004), econometrics and pricing
(Manski, 2007; Bergemann and Schlag, 2008). Minimax regret
0377-2217/$ - see front matter 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ejor.2013.01.014
q
This work is sponsored by Grant NUS R-266-000-043-133.

Tel.: +65 65162562.


E-mail address: isentsa@nus.edu.sg
European Journal of Operational Research 227 (2013) 483493
Contents lists available at SciVerse ScienceDirect
European Journal of Operational Research
j our nal homepage: www. el sevi er . com/ l ocat e/ ej or
ordering in newsvendor problems with only known demand inter-
vals were studied by Kasugai and Kasegai (1961), and extensions to
multi-item newsvendor problems under budget constraints were
considered by Vairaktarakis (2000). More recently, Perakis and
Roels (2008) employ the minimax regret approach to derive robust
order quantities for the newsvendor problem with partial informa-
tion about the demand distribution (e.g., support, mean, variance,
symmetry, unimodality), and apply entropy maximization crite-
rion for selecting a probability distribution as an input to the news-
vendor problem. Multiple market newsvendor problems using the
minimax regret criterion is also recently studied by Lin and Ng
(2011). These works motivate the relevance of the regret optimiza-
tion models in operations research. On the other hand, most of
these works are conned to very specic structures and simplied
abstractions of planning problems. For instance, the newsvendor
problems only assume uncertainty in the demand information. In
the application examples in this paper, the newsvendor problems
that are considered can also have uncertainties in the supply and
product yield information. In particular, the use of the minimax re-
gret approach for generic linear programming models that may
have uncertainties in any of the data coefcients is proposed in this
work.
1.2. Regret optimization models for uncertain linear programs
The focus on this paper is the regret optimization criteria for
uncertain linear programs. Inuiguchi and Sakawa (1995) rst ad-
dressed a linear programming problem with interval objective
function coefcients using the minimax regret approach. The
authors propose a solution algorithm based on a relaxation proce-
dure developed by Shimizu and Aiyoshi (1980). The algorithm re-
quired solving a subproblem known as the candidate regret
maximizer problem repeatedly in order to obtain data realizations
that yielded the maximumregret for a given planning solution. The
subproblem is essentially a bilinear programming model in struc-
ture, and the authors adopted a naive solution approach that re-
quired the enumeration of all the vertices in the linear
programming feasible space. This was clearly computationally
infeasible even for problems with moderate size. For the same
problem, Mausser and Laguna (1998a); Mausser and Laguna,
1998b proposed a mixed integer formulation for the candidate re-
gret sub-problem, and the computational study indicates that such
an mixed integer formulation signicantly reduces solution time
by exploiting state-of-the-art integer solvers. The authors also de-
velop a greedy heuristic procedure for the candidate regret maxi-
mizer problem, and demonstrated its computational effectiveness
for large-scale problems. Averbakh and Lebedev (2005) showed
that the minimax regret linear programming problemwith interval
objective function coefcients is indeed strongly NP-hard.
This work extends the research of the above cited papers, and in
fact generalizes several of the past results. In particular, linear pro-
grams with both uncertain objective coefcients and constraint
coefcients are considered. Rather than adopting the robust opti-
mization approach that attempts to ensure that all constraints
are fully met, constraints are allowed to be violated at certain pen-
alty costs. This is common in applications such as inventory and
production planning problems where uncertainty in demands
and supplies are expected. However, as will be discussed in detail,
this complicates the regret optimization problem signicantly, so
that even the previous established results to evaluate the maxi-
mum regret and the solution algorithms are rendered unsuitable
for all practical purposes. The main contribution in this work is
the development of a safe approximation of the regret function
for this class of uncertain linear programs. While it is not the en-
deavor of this work to resolve the computational complexity issues
of the regret optimization problem (since the problem considered
here are even harder than those considered by Averbakh and Lebe-
dev (2005)), the proposed approximations retain the important
structural properties of the problems in the previous works. This
enables good leverage of the available results and solution algo-
rithms to solve the problems efciently.
Furthermore, this work expands on the uncertainty models
assumed in the previous works (that only considered interval
data). The expansion includes the consideration of the correla-
tion among uncertain parameters through the use of a random
factors model, and also the use of arbitrary support sets de-
scribed with polyhedrons. Because these uncertainty models
and the associated concepts are quite popular in the robust opti-
mization parlance, the author term this work robust regret
optimization.
1.3. Outline
The outline of the rest of the paper is as follows. The next sec-
tion introduces the uncertain linear programming problem of
interest in this study, and the corresponding regret optimization
model. The issues related to the computational challenges of eval-
uating and solving these regret models are discussed in detail. An
approximate regret function is then developed along with a
solution scheme of the proposed regret optimization model. To
showcase the application of the proposed model, in Section 3, a
co-production newsvendor model is introduced. Structural
properties of the co-production newsvendor are then developed
to enable the application of the regret optimization model. Sec-
tion 4 present the results of some computational experiments
to evaluate and compare the performances of the proposed ap-
proach with other commonly-used solution approaches such as
sample-average and worst-case optimization. Finally, Section 5
concludes this work.
2. Robust regret for linear programming
2.1. Linear programming planning model
Let c
i
R
N
, for all i = 0, . . . , I, c = [c
0
, . . . , c
I
], s R
I
, and
A R
MN
and a R
M
be the input data to the following linear pro-
gramming problem:
min
x
c
/
0
x
s:t: c
/
i
x s
i
P0 \i = 1; . . . ; I
Ax Pa; x P0
where x R
N
are the decision variables, and s
i
denote the ith ele-
ment of s. In the situation of interest, the data inputs (c, s) can be
uncertain at the point in time when decisions x need to be
determined (we assume that A and a does not involve uncertain-
ties). Hence, a solution obtained based on solving the above using
a given realization of the inputs may become infeasible. However,
in most practical problems, when there are uncertainties to be
expected, constraints are allowed to be violated at a penalty cost.
This is for instance in production-inventory models, where demand
shortages are allowed at a shortage penalty cost, or excess inven-
tory are stocked at penalty costs. The following problem is thus
considered:
f (c; s) = min
x
c
/
0
x

I
i=1
g
i
s
i
c
/
i
x
_

; s:t: Ax Pa; x P0 (1)


where g
i
denotes the penalty cost per unit of the shortfall level [s
i
-
c
i
x]
+
. Denoting C = [c
1
; . . . ; c
I
[ R
NI
, the linear programming for-
mulation of (1) is:
484 T.S. Ng / European Journal of Operational Research 227 (2013) 483493
f (c; s) = min
x;y
c
/
0
x

I
i=1
g
i
y
i
(2)
s:t: y
i
c
/
i
x Ps
i
\i = 1; . . . ; I (3)
Ax Pa; x P0; y P0 (4)
Denoting the multipliers of (3) and (4) as k R
I
and l R
M
respec-
tively, the dual formulation is then stated as:
f (c; s) = max
k;l
s
/
k a
/
l (5)
s:t: Ck A
/
l 6 c
0
(6)
k 6 g; k P0; l P0 (7)
Since any dual feasible solution (5)(7) can be used to generate a
lower bound on (2)(4), the dual formulation plays a pivotal role
in the development of a safe regret approximation in Section 2.4.
Some basic characteristics of f as value functions of the inputs c
and s are rst stated in the following Proposition, the proof of which
follows from basic results of linear programming duality (see for in-
stance Bertsimas and Tsitsiklis, 1997) and is omitted here.
Proposition 1. For any given C and s, f is concave in c
0
. For any given
C and c
0
, f is convex in s.
2.2. Model for data uncertainties
To model the data uncertainties, the factor model approach com-
monly used in robust optimization literature is followed; see for
instance, Ben-Tal and Nemirovski (1998); Chen et al. (2008); and
Goh and Sim (2010). A factor model of uncertainty formulates each
uncertain parameter as an afne function of a set of random fac-
tors. Let
~
z = [~z
1
; . . . ; ~z
K
[ denote a K-vector of the random factors
(also termed as primitive uncertainties), with the support set Z de-
scribed by the polyhedron:
Z = z R
K
[Bz 6 b; z P0 (8)
Thus, all possible outcomes of z are contained in the assumed
bounded set Z. Note that the denition of Z in (8) includes the spe-
cial case of a hypercube set. The uncertain linear program parame-
ters c are dened as afne functions of z using a set of factor
coefcients. It is assumed that the factor coefcients have been esti-
mated (e.g. using past data studies) a priori. The factor model can
thus be regarded as a rst order approximation of the uncertain
parameter. Specically, for each i = 0, . . . , I, and k = 1, . . . , K, denote
the factor coefcients c
k
i;n
so that
c
i;n
=

K
k=1
c
k
i;n
z
k
c
0
i;n
\i = 1; . . . ; I; n = 0. . . ; N
and similarly, the coefcients s
k
i
for k = 1, . . . , K, so that:
s
i
=

K
k=1
s
k
i
z
k
s
0
i
\i = 1; . . . ; I
Thus, the use of the factor coefcients c
k
i;n
and s
k
i
allows correlation
among the uncertain variables c to be modelled.
2.3. Uncertain linear programs and regret function
The regret r(x, z) associated with a feasible solution x for some
given z is dened as that difference between the cost achieved
by a solution x and the minimum cost achieved when z is fully
known before solving the problem (1), i.e.
r(x; z) = c
/
0
x

I
i=1
g
i
s
i
c
/
i
x
_

f (z)
The maximum regret associated with a decision x is then:
R(x) = max
zZ
c
/
0
x

I
i=1
g
i
s
i
c
/
i
x
_

f (z) (9)
Thus, R(x) represents theworst cost differential incurredbyadecision
x due to the absence of complete exact planning data. The robust
regret linear programming seeks x that minimizes the worst regret:
min
xX
R(x) = min
xX
max
z
c
/
0
x

I
i=1
g
i
s
i
c
/
i
x
_

f (z) (10)
where X in the above denotes the feasible set of alternatives
{x:Ax Pa}. Basically, (10) is a minimax optimization problem,
where the inner problem maximizes the regret associated with a gi-
ven x, that is, equivalent to evaluating R(x) in (9). For this problem,
Inuiguchi and Sakawa (1995) considered the case when there are
only uncertainties in the objective coefcients c
0
. In particular,
the uncertainties arises from a hypercube set,
c
0
R
N
[c
0
6 c
0
6 c
0
. The authors then proposed the notions of
x-optimality and c-consistency to demonstrate that it sufces to con-
sider the set of all extrema of the hypercube to locate the regret
maximizer. Note that this is the consequence of the fact that f(c
0
)
is (piece-wise linear) concave in c
0
given the rest of the inputs
(Proposition 1), and hence r() conditioned on x, C, and s is convex
in c
0
. The result then follows by noting that the maximum of a con-
vex function over a convex set is always achieved at some extrema
of the convex set. The key insight here is that the convexity of the
regret function allows the reduction of the search space to the set
of all extreme points of Z, which is nite. This is termed as the prop-
erty of extrema sufciency, and as a result the search for the regret
maximizer can be cast as a discrete optimization problem, where
the solution of the discrete optimization model is an extreme point
of Z. In general the regret maximizer problem is known to be com-
putationally NP-hard. Mausser and Laguna (1998a); Mausser and
Laguna, 1998b proposed a linear mixed-integer programming
(MIP) reformulation of the regret maximizer problem. While this
does not resolve the computational issues entirely, state-of-the-
art large-scale MIP solvers are now easily available, and one can ex-
ploit the power of such solvers conveniently.
In the case when C and s are uncertain, while the realized cost
function c
0
x

I
i=1
g
i
s
i
c
/
i
x
_

is a convex function in the uncer-


tain parameters, the optimal value function f(z) is generally not
concave, and consequently the regret function r(x, z) is non-convex.
Even in the case where only the right-hand side vector s is uncer-
tain, and f(s) convex (from Proposition 1), the regret function is
still neither concave nor convex in general. This can be veried
even in very simple problems, for instance in the co-production
problem (Section 3) for a single and two product grade problem,
as shown in Figs. 1 and 2 respectively. It can be noted that in both
cases the maximum regret does not occur at the extrema of the
uncertain parameter space. In summary, without extrema suf-
ciency, evaluating the maximum regret for problems of even mod-
est dimensions becomes practically impossible, and the regret
maximizers generally has to be located using global optimization
approaches. However, global optimization methods typically only
locate a lower bound on the maximum regret and cannot provide
a safe performance guarantee about the true maximum regret.
Also, the regret maximization typically needs to be solved a large
number of times iteratively in a solution algorithm.
In this work, C and s can uncertain in general, and the objective
is to develop a safe approximation to the regret function that pre-
serves extrema sufciency, thus circumventing the abovemen-
tioned issues.
2.4. Approximate regret function
In this section, an approximate regret function to the actual
maximum regret is proposed. The approximation is essentially
T.S. Ng / European Journal of Operational Research 227 (2013) 483493 485
based on creating convex upper bounds on the actual regret using
Lagrangian duality. This is presented formally in the following
proposition.
Proposition 2. Let
R
k;l
(x) = max
zZ
c
0
x

I
i=1
g
i
s
i
c
/
i
x
_

s
/
k a
/
l (11)
where k = [k
1
, . . . , k
I
], l = [l
1
, . . . , l
N
] and p = [p
1
, . . . , p
N
]. Dene the
set X to be the polyhedron dened as follows:
X = (k; l; p) P0[b
/
p
n
A
/
l 6 0; B
/
p
n
PD
n
k c
0;n
\n = 1; . . . ; N (12)
where D
n
denotes the K I coefcient matrix so that, for each
i = 1; . . . ; I; D
n
k;i
= c
k
i;n
, and c
0;n
= c
1
0;n
; . . . ; c
K
0;n
_ _
. We then have that
R
k;l
() PR() for any (k, l, p) X.
Proof. For any k and l feasible in (6) and (7), i.e. the dual linear
programming model, we have
r(x; z) =c
/
0
x

I
i=1
g
i
s
i
c
/
i
x
_

f (z) 6c
/
0
x

I
i=1
g
i
s
i
c
/
i
x
_

s
/
k a
/
l
(13)
where the inequality follows from the weak duality of linear pro-
gramming. Since k and l must be feasible in (6), and since C is a
function of the random factors z, a given k and l feasible for some
Fig. 1. Regret prole for the co-production newsvendor for a single product problem (J = 1).
Fig. 2. Regret prole for the co-production newsvendor for a two product problem (J = 2).
486 T.S. Ng / European Journal of Operational Research 227 (2013) 483493
realization of z might not be feasible in another. Consequently, in
the following we enforce the worst-case or robust constraint of (6)
in considering the candidature of k and l. Let C
n
denote the nth
row of C, i.e. C
n
= [c
1,n
, . . . , c
I,n
]. The robust version of (6) is then
written as:
A
/
l max
zZ
C
n
k c
0;n
6 0 \n = 1; . . . ; N (14)
Clearly any l and k that are feasible in (14) is also feasible in (6) for
any z Z. For a given k, the maximization term on the left-hand-
side of (14) can be evaluated as follows:
max
z

K
k=1

I
i=1
k
i
c
k
i;n
c
k
0;n
_ _
z
k
(15)
s:t: Bz 6 b; z P0 (16)
This is a linear programming problem and the dual problem is:
minb
/
p
n
s:t: B
/
p
n
PD
n
k c
0;n
p
n
P0
where p
n
RI are the dual variables, and D
n
and c
0,n
are as dened
in the statement of the proposition. Hence for a given k and l, if
there exists p so that:
A
/
l b
/
p
n
6 0 and B
/
p
n
PD
n
k c
0;n
; and p P0 (17)
we then have
0 PA
/
l bp
n
PA
/
l max
zZ
C
n
k c
0;n
(18)
where again the second inequality in (18) follows from weak dual-
ity. In fact, since there exists p

that achieves the dual optimal solu-


tion, by the strong duality of linear programming the inequality
becomes tight by choosing p

. Combining the above, k, l and p must


be feasible in (17), which is the set X dened in (12). h
A direct corollary of Proposition 2 is that for a given l and k, the
approximate regret function
r
l;k
(z; x) = c
0
(z)
/
x

I
i=1
g
i
[s
i
(z) c
i
(z)
/
x[

s(z)
/
k a
/
l (19)
is convex in z, since it is simply the sum convex functions in z. Con-
sequently, the maximum approximate regret R
l,k
(x) can be evalu-
ated by solving a discrete optimization problem of investigating a
nite set of extrema of Z. This will be elaborated in Section 2.5.
The tightest maximum regret approximation for a given x is then gi-
ven by:
R(x) = min
(k;l;p)X
max
zZ
r
k;l
(20)
Since the evaluation of max
zZ
r
l;k
(z; x) is the point-wise maximum
of a set of functions afne in l and k, the minimization over l and k
in the above is a (piece-wise linear) convex optimization problem
over the polyhedron X and can be resolved easily using linear pro-
gramming techniques. Finally, the minimax regret problem that we
solve is:
min
xX
R(x) = min
xX;(k;l;p)X
max
zZ
r
k;l
(z; x) (21)
2.5. Candidate maximum regret problem
Consider now the resolution of the inner regret maximization
problem, that is:
R
k;l
(x) = max
zZ
c
0
(z)
/
x

I
i=1
g
i
c
i
(z)
/
x s
i
(z)
_

s(z)
/
k a
/
l (22)
For the case when only c
0
is assumed to be uncertain, and that c is a
hypercube, Mausser and Laguna (1998a); Mausser and Laguna,
1998b proposed an integer programming formulation for the candi-
date regret maximizer problem. While this does not resolve the
computational issues entirely, state-of-the-art large-scale MIP solv-
ers are now easily available, and one can exploit the power of such
solvers conveniently. However, in many practical situations, the
support sets Z may consist of several intersections of polyhedra that
describe known relationships of the parameters. For instance, in
production processes, while product yields may be uncertain data,
they have to add up to be one to be physically logical. In such cases
it is generally not possible to directly identify and represent all the
extrema explicitly, and hence the integer programming model of
Mausser and Laguna (1998a); Mausser and Laguna, 1998b is not di-
rectly applicable. Instead, in this work the following mixed integer
programming reformulation to solve the candidate regret maxi-
mizer problem is proposed.
Proposition 3. .Dene y = [y
1
, . . . , y
I
] and s = [s
1
, . . . , s
I
] as a set of
real-valued variables and binary-valued variables respectively. For a
given x a candidate regret maximizer can be obtained by solving the
following mixed integer programming problem:
R
k;l
(x) = max
y;s;z

I
i=1
g
i
y
i
c
0
(z)
/
x s(z)
/
k a
/
l (23)
s:t: y
i
6 c
i
(z)
/
x s(z) Ms
i
\i = 1; . . . ; I (24)
y
i
6 M(1 s
i
) \i = 1; . . . ; I (25)
y R
I
; s 0; 1; z Z (26)
where M is some large enough number.
Proof. For each c, we can re-write, for each i = 1, . . . , I,
g
i
c
/
i
x s
i
_

= max g
i
y
i
s:t: y
i
6 c
/
i
x s
i
_ _
s
i
; s
i
0; 1
which can be linearized into the form of (24) and (25). The result
then follows by combining each i = 1, . . . , I to formulate the objec-
tive function, and noting that the regret maximizer is obtained by
searching over all z Z. h
2.6. Master regret problem
Although the results in Sections 2.4 and 2.5 provide optimiza-
tion formulations for evaluating the approximate maximum regret
R
k;l
(x), there is still no tractable formulation of (22) in the decision
variables x for evaluating the minimax regret solution in (21). This
issue is circumvented by using the notion of regret cuts, which is
similar to the approaches based on using sub-gradients in solving
stochastic linear programming models (Birge and Louveaux,
1997). Since for a given z, the approximate regret (19) is convex
in x, and the point-wise maximum of a set of convex functions re-
tains convexity, lower bounds functions linear in x (termed as re-
gret cuts here) can be used to approximate R
k;l
(x) successively.
The following is a procedure to generate valid regret cuts.
2.6.1. Regret cut generation
Procedure GenCut
1. Solve (23)(26) to generate a regret maximizer
^
z E(Z), where
E(Z) is dened as the set of all extreme points of Z.
2. Fix c(
^
z) in the problem:
min
y

I
i=1
g
i
y
i
s:t: y
i
Pc
i
(
^
z)
/
x s
i
(
^
z) \i = 1; . . . ; I; y P0
T.S. Ng / European Journal of Operational Research 227 (2013) 483493 487
The dual formulation of the above, with decision variables a, can be
written as:
max
a

I
i=1
(c
i
(
^
z)
/
x s
i
(
^
z))a
i
(27)
s:t: a
i
6 g
i
\i = 1; . . . ; I (28)
3. Solve for ^ a and formulate the inequality:
r Pc
0
(
^
z)
/
x

I
i=1
^
a
i
(c
i
(
^
z) s(
^
z))
/
x s(
^
z)
/
k a
/
l (29)
Note that while the dual formulation (27) and (28) is trivial to
solve, it is a more general approach for certain classes of two-stage
problems like the co-production problem in Section 3. Further-
more, note that the set of all extrema in (27) and (28), denoted here
as C are fully characterized by the known parameters and does not
depend on the uncertain parameters z and the decisions x. Also,
since the regret maximizers are located in the extreme point set
E(Z) of Z, in theory the minimax regret problem in (21) can be
re-formulated through a complete enumeration of E(Z) and C as
follows:
min r (30)
s:t: rPc
0
(
^
z)
/
x

I
i=1
^ a
i
(c
i
(
^
z)s
i
(
^
z))
/
xs(
^
z)
/
ka
/
l \
^
zE Z ( ); \^ aC (31)
(k;l;p)X; xA; rP0 (32)
The above problem is a linear program with the auxiliary variable r
modelling the approximate maximum regret. Unfortunately, even
for Z dened as a hypercube by Inuiguchi and Sakawa (1995),
[E(Z)[ grows at 2
K
, and consequently solving (31) directly is highly
impractical even with powerful computational resources. The situ-
ation is further exacerbated with the I non-negative penalty terms
g
i
c
/
i
x s
i
_

, so that at worst the number of constraints in (31)


are [E(Z)[ [C[. The proposed approach is to solve the problem
using an iterative relaxation and cut generation approach, summa-
rized as follows.
1. Initialization: Let 1 be a subset of the set of tuples
(z; a) E(Z) C.
2. Solve the following relaxed minimax regret problem:
f =min r (33)
s:t: r Pc
0
(
^
z)
/
x

I
i=1
^ a
i
(c
i
(
^
z) s
i
(
^
z))
/
xs(
^
z)
/
ka
/
l \(
^
z; ^ a) 1
(34)
(k; l; p) X; x A; r P0
Let the solution to the above be
^
x;
^
k and ^ l.
3. Using
^
x,
^
k and ^ l, apply Procedure GenCut. If R
k;l
(
^
x) 6 f , the cur-
rent solution
^
x is the optimal solution, and the procedure termi-
nates. Otherwise, append the generated cut (29)(34), and
update 1 := 1 (
^
z; ^ a). Go to Step 1.
Note that in any iteration, the objective function value f in (33)
is a valid lower bound to the minimax regret, while the evaluated
objective function value R
k;l
(
^
x) in the candidate regret maximizer
problem (23)(26) is an upper bound. Furthermore, the procedure
is guaranteed to terminate since the sets E(Z) and C are nite.
3. Application to a co-production newsvendor problem
3.1. Problem description
The objective of this section is to apply the robust regret ap-
proach to a single period ordering problem under uncertainty. In
the problem, there are customer demands for a product that can
have different grades. Prior to learning the exact demands, an order
needto be placed for the input material. This is thenprocessedinto a
set of graded products, the proportions of which can also be uncer-
tain. This is known as a co-production process (Bitran and Dasu,
1992; Bitran and Leong, 1992; Bitran and Gilbert, 1994), motivated
by characteristics in semiconductor manufacturing. In semiconduc-
tor production, processor chips are produced fromthe same stock of
wafer. Because not all chips cut fromthe wafers have identical oper-
ating characteristics, they are separated into different grades
according to their speeds and other performance characteristics.
In co-production, it is often the case that some form of product
substitution is allowed during the demand-fulllment process. In
particular, the practice of down-substitution is assumed in the
works of Bitran and Dasu (1992); Bitran and Leong (1992) and Bi-
tran and Gilbert (1994). That is, products of a higher grade can be
used to ll the demands of a lower grade (but not vice versa). For
instance, in semiconductor manufacturing, excess supply of high
clock-speed chips can be used to ll demands for lower end chips
if necessary. Generally, product down-substitution can improve
exibility in the demand fulllment process.
The following notation pertains to the problem data. Other
notation will be introduced as and when necessary during the
model development.
3.1.1. Data and parameters
j index for product grade, j = 1, . . . ,J
J the number of product grades and demands
c
0
per unit production cost
g
j
per unit revenue, or shortage penalty of grade j
u vector of grade fraction realizations, u = [u
1
, . . . ,u
J
]
d vector of demand realizations, d = [d
1
, . . . ,d
J
]
3.1.2. Decision variables
x order quantity
q
j;j
/ level of grade j product used to fulll grade j
/
demand
y
j
shortage level of grade j
The co-production newsvendor problemwith downward substi-
tution considered is described in Hsu and Bassok (1999). The se-
quence of events and actions of the co-production newsvendor is
as follows. At the beginning of the planning, the co-production
newsvendor needs to determine the order quantity x prior to learn-
ing the actual demands d and grading fractions u of the products.
The order arrives, and the levels of each graded product u
j
x and cus-
tomer demands d
j
for all j = 1, . . . , J are observed. In the following it
is assumed that a product indexed j is always of higher (better)
grade than all products indexed j
/
> j. The allocation q
j;j
/ fromsupply
of grade j to demand for grade j
/
is then executed. The co-production
newsvendor seeks to maximize his total prots, that is the total rev-
enues less purchasing costs. It is assumed that the revenue per unit
g
j
is non-increasing in j, i.e. g
j
Pg
j
/ whenever j < j
/
. The problem for
a given data realization z can be formulated as the following:
max
x;qP0

J
j=1
g
j
d
j
(z) g
j
d
j
(z)

j
j
/
=1
q
j
/
j
_
_
_
_

c
0
x (35)
s:t:

J
j
/
=j
q
jj
/ 6 u
j
(z)x \j = 1; . . . J (36)
488 T.S. Ng / European Journal of Operational Research 227 (2013) 483493
In the objective function (35), the term d
j
(z) d
j
(z)

j
j
/
=1
q
j
/
j
_ _

corresponds to the total sales volume of each product j, which is


the minimum of the demand realized and allocated quantity

j
j
/
=1
q
j
/
j
. Note that the demand for grade j can only be fullled with
grades 1, . . . , j, and (36) imposes that product of grade j can only be
used to fulll demands for grades j
/
Pj. The above formulation can
also account for salvage values, by augmenting to the set of product
grades a dummy product grade J + 1 with u
J+1
= 0, and that d
J+1
and
g
J+1
represent the salvage demand market size and salvage rate of
all unsold products. The salvage demand market size can be mod-
elled as the uncertain parameter d
J+1
(z), or can be set sufciently
high (e.g. to be the sum of all maximum possible demand levels)
if desired. (35) and (36) can be re-written in minimization format,
and by introducing the demand shortfall variables y
j
, is equivalent
to the following linear programming model:
f (z) = min
x;y;qP0
c
0
x

J
j=1
g
j
y
j

J
j=1
g
j
d
j
(z) (37)
s:t: (36) and

j
j
/
=1
q
j
/
j
y
j
= d
j
(z) \j = 1; . . . J (38)
The following exposition focuses on the minimization format for
consistency with the development in Section 2, and g
j
is referred
to as the shortage penalty rate of product j for convenience. Also,
because the term

J
j=1
g
j
d
j
in the objective function (37) is indepen-
dent of the decision variables, it does not play an interesting role in
the optimization procedure and is hence omitted in the rest of the
paper of simplicity. For a given x, the second-stage allocation problem
is dened as:
w(x; z) = min
y;qP0

J
j=1
g
j
y
j
s:t: (36) and (38):
(39)
The problem (36)(38) can then be re-written compactly as:
f (z) = min
xP0
cx w(x; z) (40)
The robust regret optimization problem for the co-production
newsvendor with downward substitution is then:
min
xP0
R(x) = min
xP0
max
zZ
cx w(x; z) f (z) (41)
The following development will lead to demonstrate how the pro-
posed approximate minimax regret approach in Section 2 can be
applied to co-production newsvendor problem with downward
substitution under uncertainty. Note that because the co-produc-
tion with downward substitution is a two-stage decision-making
problem with non-trivial recourse, the results in Section 2 might
not be easily extended. To exploit these results, some additional
structural properties of the allocation problem (39) will need to
be established rst.
3.2. Characterizing the second-stage allocation problem
This section presents the results that allow the characterization
of the optimal solutions of the downward substitution problem di-
rectly without the allocation variables q
j;j
/ . As mentioned, a product
with a higher grade is assumed to always have shortage penalty
rate no less than lower grade products. In such a situation, it is
intuitively clear that an optimal allocation always follows by ful-
lling a higher grade demand as much as possible rst. Any out-
standing supply is then down-substituted to feed the next lower
grade demand. The allocation procedure is then iterated for the
next lower grade, etc. The following proposition states this result
formally, the proof of which is provided in the Appendix section.
Proposition 4. An optimal solution y

of the downward substitution


allocation problem is as follows:
y
+
1
= [d
1
u
1
x[

(42)
y
+
j
= d
j

j1
j
/
=1
y
+
j
/ u
j
/ x d
j
/
_ _
u
j
x
_
_
_
_

\j = 2; . . . ; J (43)
Proof. See Appendix. h
The optimal shortages y

in (42) and (43) can be evaluated by


solving a linear programming model for the down-substitution
problem without the allocation variables q
jj
/ . This is presented in
the next Proposition.
Proposition 5. The optimal solution for the down-substitution
problem in Proposition 4 is an optimal solution in the below linear
program.
w(x; u; d) = min
yP0

J
j=1
g
j
y
j
(44)
s:t:

j
j
/
=1
y
j
/ P

j
j
/
=1
d
j
/

j
j
/
=1
u
j
/ x \j = 1; . . . ; J (45)
Proof. See Appendix. h
The next result shows how to formulate the down-substitution
problem in Proposition 5 as a maximization problem instead, using
the techniques of integer programming. The purpose of this is to
obtain a single mixed integer programming formulation of the can-
didate maximumregret problem similar to the approach presented
in Proposition 3.
Proposition 6. .Consider the following mixed-integer programming
problem, where y are real variables, and s are dened as binary
variables.
max
y;s

J
j=1

J
j=1
g
j
y
j
(46)
s:t:

j
j
/
=1
y
j
/ 6

j
j
/
=1
d
j
/

j
j
/
=1
u
j
/ x Ms
j
\j = 1; . . . ; J (47)
y
j
6 M(1 s
j
) \j = 1; . . . ; J (48)
y P0; s 0; 1 (49)
The optimal objective function value of (46)(49) is then equal to that
of the linear program (44) and (45). Furthermore, there always exist an
optimal solution to (46)(49) with y

taking the values of (42) and (43)


in Proposition 4.
Proof. See Appendix. h
3.3. Robust regret for the co-production newsvendor
The co-production newsvendor desires to minimizes his maxi-
mum regret in the decision-making:
min
xP0;k;p
R
k
(x);
s:t: (k; p) X
(50)
where the upper bound R
k
(x) on the maximum regret is stated as:
T.S. Ng / European Journal of Operational Research 227 (2013) 483493 489
R
k
(x) = max
yP0;zZ

J
j=1
g
j
y
j
c
0
x k
/
d; s:t:(47) (49)
This can be obtained by following through the derivation in Propo-
sition 2, and then applying the results obtained in Proposition 6. As
before, k and p refer to the dual multipliers of the linear program-
ming models (36)(38), (15) and (16) respectively, and X the corre-
sponding feasible space of (k, p). The technical development
essentially uses the same approach presented in Sections 2.3 and
2.4, and will not be repeated in detail here.
3.3.1. Regret cut generation
Let a = [a
1
, . . . , a
J
] be the dual multipliers corresponding to the
demand shortage (45) of the second-stage allocation linear pro-
gramming model (44) and (45) respectively. Thus using the gener-
ated
^
z obtained by solving the mixed integer model (46)(49), the
following is a valid regret cut:
r Pc
0
x kd(
^
z)

J
j=1

J
j
/
=1
a
j

j
j
/
=1
d
j
/ (
^
z) u
j
/ (
^
z)x (51)
3.3.2. Master robust regret problem
The regret cut generation procedure in Section 2.4 is applied to
solve the robust regret optimization problem. In each pass of the
solution process, the following relaxed master problem is solved:
min r
s:t: r Pcx k
/
d

J
j
/
=1
a
j

j
j
/
=1
^
d
j
/ ^ u
j
/ x
_ _
z E
/
(Z); a C
/
(k; p) X; x A; r P0
where E
/
(Z) and C
/
are subsets of E(Z) and C respectively. These are
updated in each pass by applying Procedure CutGen in Section 2.4
to obtain a candidate regret maximizer
^
z and ^ a, so that
E
/
(Z) := E
/
(Z)
^
z, C
/
:= C
/
^ a.
4. Computational study
In this section, computational studies of the co-production
newsvendor problem with full downward substitution are per-
formed. Section 4.1 rst present some studies on small-size simu-
lated problems, the objective of which is to evaluate the
performance of the robust regret model in approximating the true
regret function. Section 4.2 then expands the application of the ro-
bust regret approach for a co-production system motivated by the
characteristics of a semiconductor production problem. The objec-
tive is to compare the performance of the robust regret approach
with other optimization approaches such as worst-case optimiza-
tion and sample average approximation models. All the computa-
tional experiments done are run in Intel Core Duo, 2.13 GHz CPU
with 2.99 GB of RAMdesktops under Windows. The solution proce-
dures are all coded in Java, and the linear and integer programming
models are solved by calling the CPLEX 11 libraries (www-01.ibm.-
com/software/integration/optimization/cplex-optimizer).
4.1. Performance of regret approximation
The experimental setup for the computational study will be as
follows. Eight test problems of J = 3 are randomly generated. For
each test problem, given a value of the order quantity x, the regret
maximizer problem to evaluate R(x) is rst solved. The solution ^x
that minimizes R is then located by performing a line search of a
set of values of x. Similarly, the actual maximum regret R(x) at each
of these values of x are also evaluated by an extensive grid search
over the space of Z. Clearly, this approach of evaluating the maxi-
mum regret and then obtaining the minimax regret order quanti-
ties is only computationally feasible for small test problems, and
is performed here only for comparison purpose. Table 1 presents
the results for the eight test instances for values of x [0, 1800],
where the ratios of R(x)=R(x) are tabulated in the cells.
The results in Table 1 shows that the gap between the upper
bound and the evaluated maximum regret is on average 7.3%
across the instances. The largest gap is also no more than 17.5%.
This veries that the obtained upper bound approximation is rea-
sonably tight. Next, Table 2 compares the results of the optimal
solutions of the actual minimax regret problem, x

, and the approx-


imate minimax regret ^x. The actual maximum regret achieved by ^x,
normalized by the true minimax regret values R(x

) are also tabu-


lated. The evaluated maximum regret achieved by ^x was on aver-
age 1.08 times of the minimax regret R(x

), and was no worse


than 1.13 times the minimax regret, across the eight test problems.
The error in the optimal quantities (^x x
+
)=x
+
was within 12.3%
and 10% and across the test problems. In the results, the instances
with order quantities ^x deviating further from the actual optimal
solution also had poorer regret performance (e.g. Instances 3 and
5). On the other hand, the solutions in Instances 1, 2, and 8
achieved more accurate approximations of the minimax regret
solution and correspondingly better regret performance. The
experiment thus provides validation that the approximate mini-
max regret model is effective in producing good quality solutions
that are also close to the actual minimax regret solutions
consistently.
4.2. Performance comparison
In this section, computational experiments are performed using
a larger scale problem of the co-production problem to compare
the performance of the robust regret model (50) with two other
optimization modelling approaches described below. Comparisons
are made in terms of maximum realized regret, worst-case perfor-
mance and computation times. For the computational experi-
ments, data realizations were simulated based on estimates of
the supports of the grading fractions and demands. Different prob-
ability distributions were used to generate the data realizations.
Table 1
Regret evaluation R(x)=R(x) for test problems J = 3.
Instancex 0 300 600 900 1200 1500 1800 Average Maximum
1 1.078 1.144 1.127 1.175 1.120 1.140 1.136 1.13 1.175
2 1.125 1.16 1.082 1.101 1.071 1.035 1.074 0.94 1.125
3 1.102 1.125 1.091 1.072 1.153 1.144 1.120 1.12 1.153
4 1.063 1.085 1.104 1.112 1.045 1.081 1.126 1.08 1.126
5 1.117 1.124 1.112 1.030 1.048 1.045 1.086 1.08 1.124
6 1.104 1.077 1.041 1.021 1.012 1.032 1.113 1.06 1.113
7 1.121 1.118 1.097 1.045 1.083 1.084 1.112 1.09 1.121
8 1.085 1.081 1.078 1.082 1.073 1.024 1.105 1.07 1.105
490 T.S. Ng / European Journal of Operational Research 227 (2013) 483493
4.2.1. Average-case-data optimization model (AOP)
In this model, the uncertain parameters are assumed to be real-
ized at the mid-point values, and the solution

x based on that is ob-
tained by solving the linear programming model (36)(38).
4.2.2. Worst-case-data optimization model (WOP)
In this case, the uncertain parameters are assumed to be real-
ized at their worst-case values, so that any costs achieved by a
solution is as high as possible within the space of all possible real-
izations. The optimization problem is then to obtain the solution x
/
that minimizes the maximum cost of the problem.
min
xP0
max
zZ
g
j
[d
j
(z) u
j
(z)x[

c
/
x (52)
The problem in (52) is also known as the robust optimization prob-
lem and can be easily re-formulated as a linear programming prob-
lem using strong duality results, see for instance (Bertsimas and
Sim, 2004).
4.2.3. Sample average approximation model (SAA)
This approach assumes knowledge of the probability distribu-
tions of the uncertain parameters, which are used to generate a
set of Q random sample realizations, i.e. z
q
, where q = 1, . . . , Q.
The optimization problem is then to obtain the solution that min-
imizes the sample average cost of the problem.
min
xP0
c
0
x 1=Q

Q
q=1
w(x; z
q
) (53)
The formulation (53) cab be easily re-formulated and solved as a
linear programming model. In the computations the sample size
is set to be R = 50.
4.2.4. Experiment set-up
Table 3 summarizes the input parameters used in the computa-
tional study. There are 30 different input stocks whose levels need
to be determined. After the testing stage, these stocks are sorted
into a total of 20 nal product grades. The uncertainty ranges of
the demand and grading fractions were estimated from a sample
set of past data of the semiconductor production. The cost rate
parameters c were chosen from the interval [1, 10] to model the
relative per unit costs of the input stocks. Each uncertain variate
~z [z; z Dz[ with Dz P0 is modeled as a zero-mean random var-
iable as follows:
~
z = z Dz B(c
1
; c
2
)
where B is a beta-distributed random variable with shape parame-
ters c
1
and c
2
. The beta random variable is used to facilitate the
computational experiments because of its exibility in modeling
different bounded distributions. The mean of ~z is then
E(
~
z) = z Dz
c
1
c
1
c
2
The shape parameters c
1
and c
2
are then chosen with the above is
set to zero. In the problem case instances, many input stocks could
only produce a subset of the nal products. This can be accounted
for in our modelling framework by setting the appropriate Dz = 0
above.
For performance comparison, the following performance mea-
surements were evaluated, the sample maximum regret (SMR), sam-
ple maximum costs (SMC) and sample average costs (SAC). These
measures are aligned with the objective of each of the correspond-
ing optimization models. The following procedure was used in the
performance evaluation of each model:
1. Solve planning model for optimal input stock levels x

.
2. Generate K demand and grading fractions realizations according
to the assumed distribution.
3. For each realization z
k
:
v Solve the following second-stage allocation problem (39) to
evaluate the realized cost w(x

,z
k
).
v Solve the co-production newsvendor (36)(38), with full
information, i.e., z
k
for the optimal costs f(z
k
).
v Evaluate the realized regret R(x

, z
k
) = c
/
x

+ w(x

, z
k
) f(z
k
).
4. Evaluate the performance measures of:
v Sample maximum regret SMR(x

) = max
k=1, . . .,K
R(x

, z
k
).
v Sample maximum costs SMC(x

) = max
k=1,. . ., K
c
/
x

+ w(x

, z
k
).
v Sample average costs SAC(x
+
) = c
/
x
+
1=K

K
k=1
w(x
+
; z
k
).
4.2.5. Results and discussion
Table 4 shows the results of the computational experiments for
the comparison of the robust regret optimization with the worst-
case optimization (WOP), the sample-average approximation
(SAA) and the average-case optimization (AOP) models. The tabu-
lated values are the SMR, SMC and SAC of these models, normalized
with respect to the corresponding performance measures of the ro-
bust regret model. First, for the SMR, the average performance of
the WOP, SAA, AOP across the eight test problems are 1.92, 1.24
and 1.41 times that of the robust regret model respectively. At
worst, the solutions from these model solutions were 2.7, 1.6 and
2.3 times that of the robust regret model solution respectively. This
clearly demonstrates that while WOP and AOP are simple models to
formulate and solve, their regret performance were very much
Table 2
Comparison of minimax regret for test problems J = 3.
Instance x

^x ^x=x
+
R(^x)=R(x
+
)
1 1210.1 1200.3 0.99 1.008
2 305.2 323.6 1.06 1.023
3 1250.7 1097.2 0.87 1.123
4 1640.3 1520.9 0.92 1.115
5 890.1 979.2 1.10 1.127
6 830.8 787.5 0.94 1.106
7 720.2 787.8 1.09 1.092
8 1090.4 1063.7 0.97 1.069
Table 3
Computational study input parameters.
Parameter Values
Number of input stocks I 30
Number of nal products J 20
Purchase cost rate c
0
[1, 10]
Out-of-sample performance sample size K 300
Estimated demand means [20, 100]
Estimated grading fractions means [0, 1]
Random factors ~z Beta (c
1
, c
2
)
Table 4
Performance comparison of WOP, SAA and AOP models to robust regret model.
Instance SMR SMC SAC
WOP SAA AOP WOP SAA AOP WOP SAA AOP
1 1.47 1.07 1.08 1.25 1.10 1.13 1.33 0.96 0.97
2 2.39 1.12 1.09 1.23 1.21 1.17 1.28 1.01 1.23
3 2.31 1.16 1.21 1.06 1.13 1.14 1.18 0.95 1.17
4 1.11 1.08 1.2 1.07 1.15 1.16 1.10 0.87 0.98
5 2.69 1.32 1.65 1.15 1.21 1.23 1.26 1.05 1.12
6 2.77 1.65 2.31 1.14 1.08 1.11 1.41 0.98 1.24
7 1.6 1.48 1.51 1.21 1.17 1.17 1.23 0.97 1.31
8 1.06 1.17 1.26 0.95 1.06 1.14 1.09 1.10 0.98
Average 1.92 1.24 1.41 1.13 1.14 1.15 1.24 0.98 1.12
T.S. Ng / European Journal of Operational Research 227 (2013) 483493 491
unstable and inferior compared to the robust regret. The SAA model
performed relatively better than these two models, but there is still
a signicant gap from the regret achieved by the robust regret
model. This motivates the need for a model that explicitly accounts
for regret performance.
Next, for the SMC performance, we note that the WOP model at-
tained a worst-case cost of on average 1.13 times that of the robust
regret. This is surprising since WOP model explicitly hedges against
worst case performance. One possible reason is that the WOC is
based on optimizing around a rare event realization to hedge
against worst-case outcomes. For instance, in some realizations
the WOP newsvendor is observed to take extreme positions such
as not to purchase and produce any products when variability of
the yield and demand values are high. In fact, the performance of
the WOP was only marginally better than the AOP (1.15 on average)
and not signicantly better than the SAA model (1.13 on average).
The robust regret model on the other hand out-performed all three
models in the SMC, and is hence relatively effective in hedging
against bad outcomes. This could be due to the nature of the regret
optimization, that tends to avoid solutions with large performance
gaps across different data realizations.
For the performance of sample average costs, the WOC model
averaged 1.23 times the sample average costs of the robust regret
solution. As before, this indicates the inferior performance of the
WOC solution on average across many data realizations. The simple
average-data model AOP on the other hand performed relatively
better, and in three test cases achieved an SAC lower than the ro-
bust regret solution. However, it is also more exposed to severely
bad outcomes, for instance in test Problems 2, 6 and 7. The SAA
model outperformed all the other models and averaged 0.98 times
the cost of the robust regret solution. This is not surprising since
the SAA explicitly optimizes the expected costs of the problem.
To put the results in perspective, however, the loss in performance
of the robust regret model in terms of the sample average costs is
marginal compared to the cost savings achieved in the measures of
SMR and SMC.
Finally, Table 5 shows the computational times of the eight test
cases of the four models. The AOP and WOP averaged about 22.9
and 34.1 seconds of solution time, while the SAA model averaged
68.4 seconds, due to the larger problem formulation of the model.
The robust regret model on average took 493.4 seconds to solve,
and at worst about 960.3 seconds (~16 minutes) across the eight
test problems. Thus, it is clear that robust regret model requires
signicantly more computational time to solve compared to the
other models. This is expected, since solving the robust regret
model requires the resolution of the candidate maximum regret
problem repeatedly, which is a mixed integer programming prob-
lem. The computational studies seem to suggest that the candidate
regret problem indeed poses as a bottleneck in the solution pro-
cess, since instances that require long solution times are also asso-
ciated with those that require more regret cuts to be generated
(e.g. Instances 3 and 5). However, in view of the good performance
of the robust regret solutions, the expended computational times
may still be considered acceptable if the planning cycles are per-
formed on a daily or weekly basis, or if real-time re-generation
of planning solutions are not required.
5. Conclusion
In this paper the regret optimization criterion for linear pro-
gramming problems under uncertainty is proposed. This work
builds on and extends the results of previous works by consider-
ing linear programs with not only interval objective coefcients
but also in the other data parameters. A safe approximation of
the regret function is developed. This approximation preserves
the extrema sufciency property so that the maximum regret
can be evaluated efciently. The approach is further applied to
a co-production newsvendor problem. Computational experi-
ments rst demonstrated that the regret approximation is reason-
ably accurate. Next, comparison with other optimization
approaches such as worst-case and sample average optimization
validate that the proposed model achieves a signicantly superior
performance in regret, and performs competitively across mea-
sures such as average and worst case costs. In this work, several
special structural results were established and exploited in the
co-production model for the application of the regret optimiza-
tion approach. Future research includes extending the robust re-
gret optimization approach to uncertain general two-stage
linear programs. Furthermore, recent developments in robust
optimization termed as distributionally robust optimization al-
lows the incorporation of distributional information such as mean
and variance in the specication of the uncertainties. Ongoing re-
search includes incorporating these distributional information in
the robust regret modelling framework.
Appendix A
A.1. Proof of Proposition 4
In the following, dene F
j
for all j = 1, . . . , J as the total supply of
grades 1, . . . , j 1 NOT used to fulll demands for grades
j = 1, . . . , j 1, i.e. the outstanding supply after fullling demands
for grades j = 1, . . . , j 1. For the case of j = 1 we dene F
1
= 0. An
optimal shortage level for grade j, for all j = 1, . . . , J, can then be
written as y
+
j
= [d
j
F
j
u
j
x[

, since g
j
Pg
j
/ for all j
/
> j, and hence
one always do no worse by satisfying as much of d
j
as possible with
F
j
compared to satisfying demands of grades j
/
> j. Extending the
same reasoning to the next grade j + 1,
F
j1
= [u
j
x F
j
d
j
[

That is, in the optimal solution, F


j+1
is that outstanding supply after
fullling d
1
, . . . , d
j
as much as possible. Consequently, to show the
expressions for y
+
j
in (42) and (43), it sufces to show that
F
j
=

j1
j
/
=1
y
+
j
/ u
j
/ x d
j
/
_ _
\j = 2; . . . ; J
First, note that for product grade j = 1, (42) is clear, i.e.
y
+
j
= [d
1
u
1
x[

, since it is always no worse to use u


1
x ll demands
for j = 1 as much as possible. Now suppose by hypothesis that the
above is true for some j = 2, . . . , J. Indeed, based on y
+
1
; F
2
can be
written as
F
2
= y
+
1
u
1
x d
1
Table 5
Computational times of planning models (number of regret cuts in parenthesis).
Instance AOP model WOP model SAA model Robust regret model
1 15.6 30.8 72.8 215.5 (6)
2 12.8 38.5 70.3 477.6 (9)
3 29.1 44.3 65.6 960.3 (17)
4 21.4 32.7 56.8 300.4 (6)
5 28.7 29.4 72.5 819.8 (14)
6 25.5 31.2 68.5 468.5 (10)
7 31.6 37 63.8 274.6 (8)
8 18.7 29.1 77.4 430.9 (10)
Average 22.9 34.1 68.4 493.4 (9.8)
492 T.S. Ng / European Journal of Operational Research 227 (2013) 483493
where it is noted that F
2
= u
1
x d
1
if y
+
1
= 0, and F
2
= 0 if y
+
1
> 0. For
the general case, the following can be written:
F
j1
= [u
j
x F
j
d
j
[

= u
j
x

j1
j
/
=1
y
+
j
/ u
j
/ x d
j
/
_ _
d
j
_
_
_
_

= u
j
x d
j

j1
j
/
=1
y
+
j
/ u
j
/ x d
j
/
_ _
d
j
u
j
x

j1
j
/
=1
y
+
j
u
j
/ x d
j
/
_ _
_
_
_
_

j
j
/
=1
y
+
j
/ u
j
/ x d
j
/
_ _
Hence by induction the hypothesis holds true for all j = 2, . . . , J.
Thus, the solution presented in (42) and (43) is at least as good as
any optimal solution, and must also be optimal. h
A.2. Proof of Proposition 5
In the following it is shown that there exists an optimal solution
to the linear program in (44) and (45) with
y
+
j
= d
j

j1
j
/
=1
y
+
j
/ u
j
/ x d
j
/
_ _
u
j
x
_
_
_
_

for all j = 1; . . . ; J (54)


First consider (45) re-written as follows:
y
j
Pd
j

j1
j
/
=1
y
j
/ u
j
/ x d
j
/
_ _
u
j
x \j = 1; . . . ; J
Clearly y
+
j
P d
j

j1
j
/
=1
y
+
j
/ u
j
/ x d
j
/
_ _
u
j
x
_ _

by feasibility for any


optimal solution y

. Furthermore, there is always an optimal solu-


tion with the above inequality in (54) tight. This is so since any
y
+
j
> d
j

j1
j
/
=1
y
+
j
/ u
j
/ x d
j
/
_ _
u
j
x
_
_
_
_

canalways be decreasedby some non-zerolevel, andy


+
j
/ for some j
/
> j
can be increased by the same amount, resulting in a nonnegative cost
decrease of (g
j
g
j
/ ) y
+
j
d
j

j1
j
/
=1
y
+
j
/ u
j
/ x d
j
/
_ _
u
j
x
_ _

_ _
. Since
this is at least as good as the optimal solution, there must be an opti-
mal solution with y
+
j
as dened in (54). Finally, applying the result of
Proposition4, the optimal solutiontothe down-substitutionproblem
is hence also an optimal solution to the linear program (44) and
(45). h
A.3. Proof of Proposition 6
It sufces to show that the optimal solution of (46)(49),
y
j
=

j
j
/
=1
d
j
/

j
j
/
=1
u
j
/ x
_ _

for all j. First in (47), if the right-hand


side terms

j
j
/
=1
d
j
/

j
j
/
=1
u
j
/ x < 0, then we must have s
j
= 1 by fea-
sibility. Otherwise, it is always optimal to set s
j
= 0, and set
y
j
=

j
j
/
=1
d
j
/

j
j
/
=1
u
j
/ x, since because 1T is maximizing in the
objective function. h
References
Averbakh, I., Lebedev, V., 2005. On the complexity of minimax regret linear
programming. European Journal of Operational Research 160, 227231.
Azzone, G., Maccarrone, P., 2001. The design of the investment post-audit process in
large organisations: evidence from a survey. European Journal of Innovation
Management 4, 7387.
Bell, D.E., 1982. Regret in decision making under uncertainty. Operations Research
30, 961981.
Ben-Tal, A., Nemirovski, A., 1998. Robust convex optimization. Mathematics in
Operations Research 23, 769805.
Bergemann, D., Schlag, K.H., 2008. Pricing without priors. Journal of the European
Economic Association 6, 560569.
Bertsimas, D., Tsitsiklis, J.N., 1997. Introduction to Linear Optimization. Athena
Scientic.
Birge, J.R., Louveaux, F., 1997. Introduction to Stochastic Programming. Springer,
New York.
Bitran, G.R., Dasu, S., 1992. Ordering policies in an environment of stochastic yields
and substitutable demands. Operations Research 40, 9991017.
Bitran, G.R., Leong, T.Y., 1992. Deterministic approximations to co-production
problems with service constraints and random yields. Management Science 38,
724742.
Bitran, G.R., Gilbert, S.M., 1994. Co-production processes with random yields in the
semiconductor industry. Operations Research 42, 476491.
Chen, X., Sim, M., Sun, P., Zhang, J., 2008. A linear decision-based approximation
approach to stochastic programming. Operations Research 56, 344357.
Conde, E., 2004. An improved algorithm for selecting p items with uncertain returns
according to the minimax-regret criterion. Mathematical Programming 100,
345353.
Connolly, T., Zeelenberg, M., 2002. Regret in decision making. Current Directions in
Psychological Science 11, 212216.
Daniels, R.L., Kouvelis, P., 1995. Robust scheduling to hedge against processing time
uncertainty in single-stage production. Management Science 41, 363376.
Goh, J., Sim, M., 2010. Distributionally robust optimization and its tractable
approximations. Operations Research 58, 902917.
Gulliver, F.R., 1987. Post-project appraisals pay. Harvard Business Review March
April, 128132.
Hsu, A., Bassok, Y., 1999. Random yield and random demand in a production system
with downward substitution. Operations Research 47, 277290.
Inuiguchi, M., Sakawa, M., 1995. Minimax regret solution to linear programming
problems with an interval objective function. European Journal of Operational
Research 86, 526536.
Inuiguchi, M., Tanino, T., 2000. Portfolio selection under independent possibilistic
information. Fuzzy Sets and Systems 115, 8392.
Kahneman, D., Tversky, A., 1982. The Psychology of Preferences, Scientic American.
Kasugai, H., Kasegai, T., 1961. Note on minimax regret ordering policy static and
dynamic solutions and a comparison to maximin policy. Journal of the
Operations Research Society of Japan 3, 155169.
Lin, J., Ng, T.S., 2011. Robust multi-market newsvendor models with interval
demands data. European Journal of Operational Research 212, 361373.
Loomes, G., Sugden, R., 1987. Testing for regret and disappointment in choice under
uncertainty. The Economic Journal 97, 118129.
Manski, C.F., 2007. Minimax-regret treatment choice with missing outcome data.
Journal of Econometrics 139, 105115.
Mausser, H.E., Laguna, M., 1998a. A new mixed integer formulation for the
maximum regret problem. International Transactions in Operational Research
5, 389403.
Mausser, H.E., Laguna, M., 1998b. A heuristic to minimax absolute regret for linear
programs with interval objective function coefcients. European Journal of
Operational Research 117, 157174.
Neale, C.W., Holmes, D.E.A., 1990. Post-auditing capital projects. Long Range
Planning 23, 8896.
Perakis, G., Roels, G., 2008. Regret in the newsvendor model with partial
information. Operations Research 56, 188203.
Ritov, I., 1996. Probability of regret: anticipation of uncertainty resolution in choice.
Organizational Behavior and Human Decision Processes 66, 228236.
Savage, L.J., 1951. The theory of statistical decision. Journal of the American
Statistical Association 46, 5567.
Shimizu, K., Aiyoshi, E., 1980. Necessary conditions for minmax problems and
algorithms by a relaxation procedure. IEEE Transactions on Automatic Control
25, 6266.
Simonson, I., 1992. The inuence of anticipating regret and responsibility on
purchase decisions. The Journal of Consumer Research 19, 105118.
Vairaktarakis, G.L., 2000. Robust multi-item newsboy models with a budget
constraint. International Journal of Production Economics 66, 213226.
T.S. Ng / European Journal of Operational Research 227 (2013) 483493 493

You might also like