Solving Chance-Constrained Stochastic Programs via

Sampling and Integer Programming

Shabbir Ahmed and Alexander Shapiro

H. Milton Stewart School of Industrial & Systems Engineering
Georgia Institute of Technology, Atlanta, GA 30332

June 6, 2008

Abstract
Various applications in reliability and risk management give rise to optimization
problems with constraints involving random parameters, which are required to be
satisfied with a pre-specified probability threshold. There are two main difficulties
with such chance-constrained problems. First, checking feasibility of a given candi-
date solution exactly is, in general, impossible since this requires evaluating quantiles
of random functions. Second, the feasible region induced by chance constraints is,
in general, non-convex leading to severe optimization challenges. In this tutorial we
discuss an approach based on solving approximating problems using Monte Carlo
samples of the random data. This scheme can be used to yield both feasible solutions
and statistical optimality bounds with high confidence using modest sample sizes.
The approximating problem is itself a chance-constrained problem, albeit with a fi-
nite distribution of modest support, and is an NP-hard combinatorial optimization
problem. We adopt integer programming based methods for its solution. In partic-
ular, we discuss a family valid inequalities for a integer programming formulations
for a special but large class of chance-constraint problems that have demonstrated
significant computational advantages.

1 Introduction
A large class of optimization problems arising from important planning and design ap-
plications in uncertain environments involve service level or reliability constraints. Con-
sider, for example, the problem of locating service centers for responding to medical
emergencies. Requiring 100% coverage over all possible emergency scenarios is physi-
cally and economically impractical and so typically emergency preparedness plans calls
for some minimum response reliability [1, 4]. Service level agreements in telecommu-
nication contracts require network providers to guarantee, with high probability, that
packet losses will not exceed a certain percentage [20, 32]. In financial portfolio plan-
ning, investors often require that, with high probability, portfolio losses do not exceed
some threshold (value-at-risk) while maximizing expected returns [11, 25]. Mathemati-
cal models for planning/designing reliability constrained systems such as these lead to
optimization with chance constraints or probabilistic constraints.
A generic chance-constrained optimization problem can be formulated as

min f (x) subject to Pr G(x, ξ) ≤ 0 ≥ 1 − ε, (1)
x∈X

1
where X ⊂ Rn represents a deterministic feasible region, f : Rn → R represents the ob-
jective to be minimized, ξ is a random vector whose probability distribution is supported
on set Ξ ⊂ Rd , G : Rn × Rd → Rm is a constraint mapping, 0 is an m-dimensional vector
of zeroes, and ε ∈ (0, 1) is a given risk parameter (significance level). Formulation (1)
seeks a decision vector x from the feasible set X that minimizes the function f (x) while
satisfying the chance constraint G(x, ξ) ≤ 0 with probability at least 1 − ε. It is assumed
that the probability distribution of ξ is known.
By way of illustration, consider the following simple facility sizing example. We need
to decide capacities of n facilities servicing uncertain customer demand. The cost-per-
unit capacity installed for each facility is given, as is the joint demand distribution.
The goal is to determine the cheapest capacity configuration so as to guarantee that
the installed capacity exceeds demand with probability 1 − ε. This chance-constrained
problem can be formulated as follows.
n
X 
min ci xi subject to Pr ξi − xi ≤ 0, i = 1, . . . , n ≥ 1 − ε. (2)
x≥0
i=1

Here xi , ci and ξi denote the capacity, cost, and random demand for facility i, respectively.
It is assumed that the (joint) probability distribution of the random vector ξ = (ξ1 , ..., ξn )
is known (otherwise the probabilistic constraint in (2) is not defined). Note that the
probabilistic (chance) constraint of (2) can be considerably weaker than trying to satisfy
the demand for all possible realizations of ξ. Note also that (2) is an example of (1) with
G(x, ξ) = ξ − x.
In this example, we require that the reliability requirement be applied
 to all facilities
jointly. One could also consider the individual chance constraints Pr ξi ≤ xi ≥ 1−εi , i =
 n, applied to each facility separately. This leads to a much simpler problem, since
1, ...,
Pr ξi ≤ xi ≥ 1−εi is equivalent to Fi−1 (xi ) ≥ 1−εi , where Fi is the cumulative distribution
function (cdf) of ξi . Note, however, that in order to ensure the joint chance constraint
by enforcing the individual chance constraints, the corresponding risk parameters εi
should be considerably smaller than ε especially when n is large.
Beginning with the work of Charnes, Cooper and Symonds [7], chance-constrained
stochastic programs have been extensively studied. In addition to the facility location,
telecommunication and finance examples cited earlier, chance constrained models have
been used in numerous other applications including production planning [22, 16], chem-
ical processing [14, 15] and water resources management [29, 31]. See [28] for back-
ground and an extensive list of references. Despite important theoretical progress and
practical importance, chance-constrained stochastic problems of the form (1) are still
largely intractable except for some special cases. There are two primary reasons for this
difficulty.

1. In general, for a given x ∈ X, computing Pr G(x, ξ) ≤ 0 accurately, i.e., check-
ing whether x is feasible to (1), can be hard. In multidimensional situations this
involves calculation of a multivariate integral which typically cannot be computed
with a high accuracy.

2. The feasible region defined by a chance constraint generally is not convex even if
G(x, ξ) is convex in x for every possible realization of ξ. This implies that even
if checking feasibility is easy, optimization of the problem remains difficult. For

2
 example, the facility sizing example (2) with n facilities and m equiprobable realiza-
tions of the demand vector ξ is equivalent to a maximum clique problem on a graph
with n nodes and m edges, and is therefore strongly NP-hard, [19].
In light of the above difficulties, existing approaches for chance-constrained stochas-
tic programs can be classified as follows. First are the approaches for problems where
both difficulties are absent, i.e., the distribution of ξ is such that checking feasibility
is easy, and the resulting feasible region is convex. A classical example of this case is
when G(x, ξ) = v − ξ > x and ξ has a multivariate normal distribution with mean µ and
covariance matrix Σ. Then for ε ∈ (0, 0.5),
n o n √ o
x ∈ Rn : Pr ξ > x ≥ v ≥ 1 − ε = x ∈ Rn : v − µ> x + zε x> Σ x ≤ 0 ,


where zε = Φ−1 (1 − ε) is the (1 − ε)-quantile of the standard normal distribution. In this

case, under convexity of X, the chance-constrained problem reduces to a deterministic
convex optimization problem. The second class of approaches are for problems where
only the second difficulty is absent, i.e., the feasible region of the chance constraint is
guaranteed to be convex. The best known example of this case is when G(x, ξ) = ξ − Ax,
where A is a deterministic matrix and ξ has a log-concave distribution. In this case
the chance constraint feasible set is convex [26]. However it may still be difficult to
compute Pr G(x, ξ) ≤ 0 exactly. Solution methods in this class are primarily based
on classical nonlinear programming techniques adapted with suitable approximations of
the chance constraint function and its gradients (see [27]). The third class of approaches
are for problem where the first difficulty is absent, i.e., computing Pr G(x, ξ) ≤ 0 is
easy, e.g., when ξ has a finite distribution with a modest number of realizations (in
this case the feasible region is typically non-convex). A number of approaches based
on integer programming and global optimization have been developed for this class of
problems [8, 10, 30]. Finally, more recently, a number of approaches have been proposed
to deal with both difficulties [5, 6, 23, 24, 3]. The common theme in these approaches is
that they all propose convex approximations of the non-convex chance constraint that
yield solutions which are feasible, or at least highly likely to be feasible, to the original
problem. Thus the difficulty of checking feasibility as well as non-convexity is avoided.
Unfortunately, often, the solutions produced by these approaches are quite conservative.
In this tutorial we consider an approximation of the chance constraint problem (1)
where the true distribution of ξ is replaced by an empirical distribution corresponding to
a Monte Carlo sample. The resulting sample average approximation problem can be used
to provide good candidate solutions along with optimality gap estimates. The sampled
approximation problem is a chance-constrained problem with a discrete distribution and
can be quite difficult. We discuss integer programming based approaches for solving it.

2 Sample Average Approximation

In order to simplify the presentation we assume, without loss of generality, that the
constraint function G : Rn × Rd → R in (1) is scalar valued. Of course, a number of
constraints Gi (x, ξ) ≤ 0, i = 1, . . . , m, can be equivalently replaced by one constraint
G(x, ξ) := max1≤i≤m Gi (x, ξ) ≤ 0. The chance-constrained stochastic program (1) can be
rewritten as
min f (x) subject to q(x) ≤ ε, (3)
x∈X

3
 
where q(x) := Pr G(x, ξ) > 0 .
Now let ξ 1 , . . . , ξ N be an independent identically distributed (iid) sample of N realiza-
tions of random vector ξ. Given x ∈ X let
N
X
−1
1l(0,∞) G(x, ξ j ) ,


q̂N (x) := N
j=1

where 1l(0,∞) : R → R is the indicator function of (0, ∞). That is, q̂N (x) is equal to the
proportion of realizations with G(x, ξ j ) > 0 in the sample. For some given γ ∈ (0, 1)
consider the following optimization problem associated with a sample ξ 1 , . . . , ξ N ,

min f (x) subject to q̂N (x) ≤ γ. (4)

x∈X

We refer to problems (3) and (4) as the true and sampled average approximate (SAA)
problems, respectively, at the respective risk levels ε and γ.
The SAA problem is a chance-constrained stochastic problem with a different (dis-
crete) distribution and a different risk level than (3). Unless N is prohibitively large,
the chance-constrained problem SAA does not suffer from the first difficulty (computing
q̂N (x)) mentioned in Section 1, however it may still be difficult to solve. Assuming we
have a scheme for solving SAA, what can we say about an optimal solution and the op-
timal value of SAA in relation to that of the true problem (3)? Intuitively, assuming N is
large enough, if γ ≤ ε, then a feasible solution of SAA is likely to be feasible to the true
problem, and if γ ≥ ε then the optimal value of SAA is likely to be a lower bound to that of
the true problem. Thus the SAA problem can be used to obtain both candidate feasible
solutions to the true problem as well as optimality gap estimates. Next we discuss these
concepts slightly more rigorously.
We assume that X is compact, f (·) is continuous, G(x, ·) is measurable for every
x ∈ Rn , and G(·, ξ) is continuous for almost every ξ. Then the functions q(x) and q̂N (x) are
lower-semicontinuous, and the true problem (3) and the SAA problem (4) are guaranteed
to have optimal solutions if they are feasible. Let X ∗ (ε) and X̂N (γ) denote the set of
optimal solutions of the true and SAA problems, respectively, v(ε) and v̂N (γ) denote the
optimal value of the true and SAA problems, respectively.

2.1 Convergence Properties

Under reasonable regularity assumptions it can be shown that for γ = ε the optimal
value v̂N (γ) and optimal solutions of set X̂N (γ), of the SAA problem, converge to their
true counterparts v(ε) and X ∗ (ε) with probability one as N approaches infinity [25].
Suppose now that γ > ε. Then we may expect that with increase of the sample size N ,
an optimal solution of the SAA problem will approach an optimal solution of the true
problem with the risk level γ rather than ε. Of course, increasing the risk level leads to
enlarging the feasible set of the true problem, which in turn may result in decreasing of
the optimal value of the true problem. For a point x̄ ∈ X we have that q̂N (x̄) ≤ γ, i.e., x̄
is a feasible solution to the SAA problem, if and only if no more than γN times the event
“G(x̄, ξ j ) > 0” happens in N trials. Since the probability of the event “G(x̄, ξ j ) > 0” is q(x̄),
it follows that 

Pr q̂N (x̄) ≤ γ = B bγN c; q(x̄), N , (5)

4
where Pk  
N
B(k; q, N ) := i=0 i q i (1 − q)N −i , k = 0, ..., N, (6)

denotes the cdf of binomial distribution. Recall that by the Chernoff inequality [9] for
k > N p,
B(k; q, N ) ≥ 1 − exp −N (k/N − q)2 /(2q) .


It follows that if q(x̄) ≤ ε and γ > ε, then 1 − Pr q̂N (x̄) ≤ γ approaches zero at a rate
of exp(−κN ), where κ := (γ − ε)2 /(2ε). Of course, if x̄ is an optimal solution of the true
problem and x̄ is a feasible point of the SAA problem, then v̂N (γ) ≤ v ∗ (ε). That is, if γ > ε,
then the probability of the event “v̂N (γ) ≤ v ∗ (ε)” approaches one exponentially fast. By
similar analysis it can be shown that if q(x̄) ≤ γ, i.e., x̄ is a feasible solution of SAA and
γ < ε, then the probability that x̄ is a feasible solution of the true problem approaches
one exponentially fast (see [18]). Based on this analysis, we can compute a priori the
sample size required in the SAA problem so that it produces a feasible solution to the
true problem with high probability (typically such estimates of a required sample size
are quite conservative). Next we discuss techniques for assessing the quality (feasibility
and optimality gap) of an arbitrary candidate solution to the true problem.

2.2 Solution Validation

For a given candidate point x̄ ∈ X, say obtained as a solution of an SAA problem, we
would like to validate its quality as a solution of the true problem. This involves two
questions, namely whether x̄ is a feasible point of the true problem, and if yes, then
what is the optimality gap f (x̄) − v(ε). Of course, if x̄ is a feasible point of the true
problem, then f (x̄) − v(ε) is nonnegative and is zero if and only if x̄ is an optimal solution
of the true problem.
Let us start with verification of feasibility of x̄. For that we need to estimate the
probability q(x̄). We proceed by employing again the Monte Carlo sampling technique.
0
Generate an iid sample ξ 1 , ..., ξ N and estimate q(x̄) by q̂N 0 (x̄). Note that this random
sample should be generated independently of a random procedure which produced the
candidate solution x̄, and that we can use a very large sample, of size N 0 , since we do
not need to solve any optimization problem here. The estimator q̂N 0 (x̄) of q(x̄) is unbiased
and for large N 0 and not “too small” q(x̄) its distribution can be approximated reasonably
well by the normal distribution with mean q(x̄) and variance q(x̄)(1 − q(x̄))/N 0 . This leads
to the following approximate (1 − β)-confidence upper bound on q(x̄):
p
Uβ,N 0 (x̄) := q̂N 0 (x̄) + zβ q̂N 0 (x̄)(1 − q̂N 0 (x̄))/N 0 (7)

where zβ = Φ−1 (1 − β). We can now compare Uβ,N (x̄) to ε to check feasibility of x̄ (see [24]
for a slightly more accurate confidence upper bound).
In order to get a lower bound for the optimal value v(ε) we proceed as follows. Let us
choose two positive integers M and N , and let


θN := B bγN c; ε, N

and L be the largest integer such that

B(L − 1; θN , M ) ≤ β. (8)

5
Next generate M independent samples ξ 1,m , . . . , ξ N,m , m = 1, . . . , M , each of size N , of
random vector ξ. For each sample solve the associated SAA problem and record the
corresponding optimal objective values v̂N m (γ), m = 1, . . . , M . It may happen that the

SAA problem is either infeasible or unbounded from below, in which case we assign
its optimal value as +∞ or −∞, respectively. We can view v̂N m (γ), m = 1, . . . , M , as an

iid sample of the random variable v̂N (γ). Next we sort the calculated optimal values
(1) (M )
in nondecreasing order, i.e., v̂N (γ) ≤ · · · ≤ v̂N (γ). It is possible to show that with
(L)
probability at least 1−β, the random quantity v̂N (γ) is a lower bound of the true optimal
value v(ε) [18, 24].
Extensive computational results on the performance of the above mentioned val-
idation schemes on various classes of chance-constrained problems are reported in
[18, 24, 25].

3 Solving Sample Approximations

We have seen that we can generate as well as validate candidate solutions to the chance
constrained problem (3) by solving (several) sampled approximations (4). In this section
we explore approaches for solving these problems.
If γ = 0 then the SAA problem reduces to

min f (x) subject to G(x, ξ j ) ≤ 0 j = 1, . . . , N. (9)

x∈X

When the functions f (·) and G(·, ξ j ) for j = 1, . . . , N are convex (linear) and the set X is
convex (polyhedral) then (9) is a convex (linear) program, and can usually be solved effi-
ciently using off-the-shelf software. From the convergence discussion in Section 2.1 we
can see that the candidate solutions generated from (9) can often be overly conservative
in terms of the objective function. We can then consider increasing the risk level γ in the
SAA problem. However with γ > 0 the SAA problem is a chance constrained optimization
problem (with a finite distribution) and is NP-hard even in very simple settings (such
as the facility sizing example discussed in Section 1) [19]. A wide variety of approaches
have been proposed to solve different classes of chance-constrained optimization prob-
lems under finite distributions (cf. [8, 10, 28] and references therein). In this tutorial we
consider an integer programming based approach.
The SAA problem (4) can be formulated as the following mixed-integer problem (MIP)

min f (x)
subject to
G(x, ξ j ) ≤ Mj zj j = 1, . . . , N
PN (10)
j=1 zj ≤ γN
zj ∈ {0, 1} j = 1, . . . , N
x∈X

where zj is a binary variables and Mj is a large positive number such that Mj ≥ maxx∈X G(x, ξ j )
for all j = 1, . . . , N . Note that if zj is 0 then the constraint G(x, ξ j ) ≤ 0 corresponding to
the realization j in the sample is enforced. On the other hand zj = 1 does not pose any
restriction on G(x, ξ j ). The cardinality constraint N
P
j=1 zj ≤ γN requires that at least γN
j
of the N constraints G(x, ξ ) ≤ 0 for j = 1, . . . , N are enforced.

6
 Even in a linear setting (i.e., the functions f and G are linear in x and the set X is
polyhedral) moderate sized instances of the MIP (10) are typically very difficult to solve
as-is by state-of-the-art MIP solvers. The difficulty is due to the fact that the continuous
relaxation of (10) (obtained by dropping the integrality restriction on the z variables)
provides a weak relaxation, and hence slows down the branch-and-bound algorithm
that is the work-horse of MIP solvers. This difficulty can be alleviated by strengthening
the formulation (10) by addition of valid inequalities or reformulation. Such improved
formulations have tighter continuous relaxation gaps and can serve to significantly cut
down solve times.
A variety of approaches for strengthening special classes of the MIP (10) have been
proposed. Here we discuss an approach for the case of joint probabilistic constraints
where the uncertain parameters only appear on the right-hand side, i.e.,
G(x, ξ) = max {ξi − Gi (x)}.
i=1,...,m

Note that the facility sizing example (2) is of this form. By appropriately translating, we
assume that ξji ≥ 0 for all i and j. The MIP (10) can then be written as
min f (x)
subject to
Gi (x) ≥ vi i = 1, . . . , m
vi + ξij zj ≥ ξij i = 1, . . . , m, j = 1, . . . , N (11)
P N
j=1 zj ≤ γN
zj ∈ {0, 1} j = 1, . . . , N
x ∈ X, vi ≥ 0 i = 1, . . . , m.
Note that we have introduced the auxiliary variables vi for i = 1, . . . , m to conveniently
represent Gi (x). As before, if zj is 0 then the constraints Gi (x) ≥ ξij for i = 1, . . . , m
corresponding to the realization j in the sample is enforced. Consider now the following
subsystem corresponding to the i-th row of the probabilistic constraint system
n N
Fi := (vi , z1 , . . . , zN ) ∈ R+ × {0, 1}N : vi + ξij zj ≥ ξij j = 1, . . . , N,
X
zj ≤ γN .
j=1

Note that the set of feasible vectors (vi , z1 , . . . , zN ) for the MIP (11) are contained in Fi and
hence in its convex hull conv(Fi ). Thus any valid inequality for conv(Fi ) will also be valid
for (11) and can potentially strengthen its continuous relaxation. Next we describe a
family of such inequalities. Without loss of generality we can assume that ξji are indexed
such that ξi1 ≥ ξi2 ≥ · · · ≥ ξiN . Let p := bγN c. The cardinality constraint N
P
j=1 zj ≤ γN
p+1
implies that we cannot have zj = 1 for all j = 1, . . . , p + 1 thus vi ≥ ξi . This also implies
that the constraints vi + ξij zj ≥ ξij for j = p + 1, . . . , N are redundant. We can therefore
tighten the formulation of Fi to
n N
j p+1 j
X
Fi0 N
:= (vi , z1 , . . . , zN ) ∈ R+ × {0, 1} : vi + (ξi − ξi )zj ≥ ξi j = 1, . . . , p, zj ≤ γN .
j=1
PN
The subsystem obtained by dropping the cardinality constraint j=1 zj ≤ γN from Fi0 is
the well-studied mixing set [2, 13, 12, 21]
n
Mi := (vi , z1 , . . . , zp ) ∈ R+ × {0, 1}p : vi + ξij zj ≥ ξij j = 1, . . . , N, .

7
It is known that the convex hull of Mi is completely characterized by the so-called star
inequalities
l
j
(ξijk − ξi k+1 )zjk ≥ ξij1 ∀ J := {j1 , . . . , jl } ⊆ {1, . . . , p},
X
vi + (12)
k=1
j
where ξi l+1
:= ξip+1 .
Clearly the star inequalities (12) are valid for Fi and hence for (11).
It turns out that the above star inequalities define, some but not all, facets of conv(Fi ).
Even though there are exponentially many such star inequalities (one for each subset
J) these can be separated very efficiently [2]. It has been observed that addition of the
above star inequalities within a branch-and-cut framework for solving linear instances
of (11) can have tremendous computational benefits [19]. A number of additional classes
of valid inequalities and an extended reformulation for Fi which also provide significant
computational advantage in solving (10) have also been developed [17].

4 Conclusion
The stochasticity and non-convexity associated with chance-constrained stochastic pro-
grams make these extremely hard to solve. In this tutorial we discussed a sampling
based approach wherein we approximate the problem by replacing the distribution of
the uncertain parameters by an empirical distribution corresponding to a Monte Carlo
sample. The approximate problem, called a sample average approximation (SAA), is still
a chance-constraint problem, albeit with a finite distribution of modest support. The
SAA problem can serve to provide both candidate solutions as well as solution quality
estimates. Since the SAA problem is an NP-hard combinatorial problem, we adopt in-
teger programming based methods for its solution. In particular, we discuss a family
valid inequalities for a integer programming formulations for a special but large class of
chance-constraint problems that have demonstrated significant computational advan-
tages.

References
[1] A. Aly and J. White. Probabilistic formulation of the emergency service location
problem. The Journal of the Operational Research Society, 29:1167–1179, 1978.
[2] A. Atamtürk, G. L. Nemhauser, and M. W. P. Savelsbergh. The mixed vertex packing
problem. Mathematical Programming, 89:35–53, 2000.
[3] A. Ben-Tal and A. Nemirovski. Robust convex optimization. Mathematics of Opera-
tions Research, 23:769–805, 1998.
[4] P. Beraldi, M. Bruni, and D. Conforti. Designing robust emergency medical service
via stochastic programming. European Journal of Operational Research, 158:183–
193, 2004.
[5] G. C. Calafiore and M. C. Campi. Uncertain convex programs: randomized solutions
and confidence levels. Mathematical Programming, 102:25–46, 2005.
[6] G. C. Calafiore and M. C. Campi. The scenario approach to robust control design.
IEEE Transactions on Automatic Control, 51:742–753, 2006.

8
 [7] A. Charnes, W. W. Cooper, and G. H. Symonds. Cost horizons and certainty equiva-
lents: an approach to stochastic programming of heating oil. Management Science,
4:235–263, 1958.

[8] M.-S. Cheon, S. Ahmed, and F. Al-Khayyal. A branch-reduce-cut algorithm for the
global optimization of probabilistically constrained linear programs. Mathematical
Programming, 108:617–634, 2006.

[9] H. Chernoff. A measure of asymptotic efficiency for tests of a hypothesis based on

the sum observations. Annals of Mathematical Statistics, 23:493–507, 1952.

[10] D. Dentcheva, A. Prékopa, and A. Ruszczyński. Concavity and efficient points of

discrete distributions in probabilistic programming. Mathematical Programming,
89:55–77, 2000.

[11] A. Gaivoronski and G. Pflug. Value-at-risk in portfolio optimization: Properties and

computational approach. The Journal of Risk, 7:1–33, 2005.

[12] Y. Guan, S. Ahmed, and G. L. Nemhauser. Sequential pairing of mixed integer

inequalities. Discrete Optimization, 4:21–39, 2007.

[13] O. Günlük and Y. Pochet. Mixing mixed-integer inequalities. Mathematical Program-

ming, 90:429–457, 2001.

[14] R. Henrion, P. Li, A. Möller, M. C. Steinbach, M. Wendt, and G. Wozny. Stochastic

optimization for operating chemical processes under uncertainty. In M. Grötschel,
S. Krunke, and J. Rambau, editors, Online Optimization of Large Scale Systems,
pages 457–478. Springer, 2001.

[15] R. Henrion and A. Möller. Optimization of a continuous distillation process under

random inflow rate. Computers and Mathematics with Applications, 45:247–262,
2003.

[16] M. A. Lejeune and A. Ruszczyński. An efficient trajectory method for probabilis-

tic inventory-production-distribution problems. Operations Research, 55:378–394,
2007.

[17] J. Luedtke. Integer programming approaches to two non-convex optimization prob-

lems. PhD thesis, Georgia Institute of Technology, Atlanta, GA, 2007.

[18] J. Luedtke and S. Ahmed. A sample approximation approach for optimization with
probabilistic constraints. To appear in SIAM Journal on Optimization, 2008.

[19] J. Luedtke, S. Ahmed, and G. Nemhauser. An integer programming approach for

linear programs with probabilistic constraints. In M. Fischetti and D. Williamson,
editors, IPCO 2007, Lecture Notes in Computer Science, pages 410–423, Berlin,
2007. Springer-Verlag.

[20] V. Marianov and M. Rios. A probabilistic quality of service constraint for a loca-
tion model of switches in ATM communications networks. Annals of Operations
Research, 96:237–243, 2000.

9
[21] A. J. Miller and L. A. Wolsey. Tight formulations for some simple mixed integer pro-
grams and convex objective integer programs. Mathematical Programming, 98:73–
88, 2003.

[22] M. R. Murr and A. Prékopa. Solution of a product substitution problem using

stochastic programming. In S. P. Uryasev, editor, Probabilistic Constrained Optimiza-
tion: Methodology and Applications, pages 252–271. Kluwer Academic Publishers,
2000.

[23] A. Nemirovski and A. Shapiro. Scenario approximation of chance constraints. In

G. Calafiore and F. Dabbene, editors, Probabilistic and Randomized Methods for
Design Under Uncertainty, pages 3–48. Springer, London, 2005.

[24] A. Nemirovski and A. Shapiro. Convex approximations of chance constrained pro-

grams. SIAM Journal on Optimization, 17:969–996, 2006.

[25] B. Pagnoncelli, S. Ahmed, and A. Shapiro. Computational study of a chance con-

strained portfolio selection problem. Submitted for publication, 2008. Preprint
available at www.optimization-online.org.

[26] A. Prékopa. On probabilistic constrained programmming. In H. W. Kuhn, editor,

Proceedings of the Princeton Symposium on Mathematical Programming, pages 113–
138, Princeton, N.J., 1970. Princeton University Press.

[27] A. Prékopa. Stochastic Programming. Kluwer Academic Publishers, Dordrecht, The

Netherlands, 1995.

[28] A. Prékopa. Probabilistic programming. In A. Ruszczyński and A. Shapiro, edi-

tors, Stochastic Programming, volume 10 of Handbooks in Operations Research and
Management Science, pages 267–351. Elsevier, 2003.

[29] A. Prékopa and T. Szántai. Flood control reservoir system design. Mathematical
Programming Study, 9:138–151, 1978.

[30] A. Ruszczyński. Probabilistic programming with discrete distributions and prece-

dence constrained knapsack polyhedra. Mathematical Programming, 93:195–215,
2002.

[31] A. K. Takyi and B. J. Lence. Surface water quality management using a multiple-
realization chance constraint method. Water Resources Research, 35:1657–1670,
1999.

[32] A. N. Ventetsanopoulos and I. Singh. Topological optimization of communication

networks subject to reliability constraints. Problem of Control and Information The-
ory, 15:63–78, 1986.

10