Professional Documents
Culture Documents
An Introductory Tutorial On Stochastic Linear PDF
An Introductory Tutorial On Stochastic Linear PDF
An Introductory Tutorial On Stochastic Linear PDF
Programming Models
airlines, energy, manufacturing, and tele- requirements between various node pairs
communications. Notwithstanding its suc- are known with certainty. Such determin-
cesses, however, the assumption that all istic network-design problems result in a
model parameters are known with cer- tree structure (Figure 2a). With a tree de-
tainty limits its usefulness in planning un- sign, all demand pairs have paths through
der uncertainty. When one or more of the which calls may be routed. However, the
data elements in a linear program is repre- design is rigid in that only one such path
sented by a random variable, a stochastic is available. During periods of high de-
linear program (SLP) results. mand, the lack of alternative routes results
In deterministic activity analysis, plan- in the rejection of calls and a reduced level
ning consists of choosing activity levels of service. Moreover, if a link should fail
that satisfy resource constraints while because of some catastrophic event, nodes
maximizing total profit (or minimizing to- will be disconnected from the network.
tal cost). All the information necessary for Attempts to counter these difficulties by
decision making is assumed to be avail- scaling the demand upward, for example,
able at the time of planning. Under uncer- will increase the capacity of the links used;
tainty, not all the information is available, it will not eliminate the rigidity of the de-
and some parameters should be modeled sign. To obtain a more flexibly designed
as random variables. We discuss here network (Figure 2b), one must incorporate
models that can include random variables the need for flexibility within the model.
within optimization problems. Since deter- One must construct a model that explicitly
ministic methodology has been prevalent considers the likelihood of periodic (and
in optimization models, it may be tempt- correlated) heavy loads on segments of the
ing to suggest that random variables network and the possibility of catastrophic
should be replaced by their means and the equipment failures. The improvement pos-
resulting optimization problem solved. In sible from the use of a stochastic model in-
general, this approach provides solutions creases with the size of the network. In
that are structurally different from those fact, in a case study conducted at Bellcore,
provided by stochastic optimization Sen, Doverspike, and Cosares [1994] re-
models. port a 75-percent reduction in the number
To understand this, consider a network of lost calls using stochastic LP models in
with n nodes, as in Figure 1, on which place of deterministic models.
demand for connections between the (n2 ) Methods for forecasting important
demand pairs must be accommodated. quantities, such as demand, are well
Networks, such as those in telecommuni- known and widely used. Moreover, the
cations systems, are complex and typically fields of statistics and simulation provide
include hundreds of nodes. In the design methods for obtaining distributional rep-
problem that we consider, the capacity of resentations of these quantities when point
each network link must be determined in estimators are inadequate. Although many
anticipation of future demand require- people routinely formulate LP models,
ments. It is customary to assume that the only recently have OR/MS practitioners
INTERFACES 29:2 34
STOCHASTIC LINEAR PROGRAMMING
Minimize 1x2
subject to x1 ` x2 ` x3 4 2
1x1 ` x2 ` x4 4 2 (1)
11 # x1 # 1
xj > 0, j 4 2, 3, 4.
(a21, a22)
MarchApril 1999 35
SEN, HIGLE
x2, x3, x4) 4 (0, 2, 0, 0)), is feasible under That is, if implemented, either solution
uncertainty. Under uncertainty, the con- would have a 50-percent chance of failing
straint corresponding to (1) is equally to satisfy a constraint.
likely to be either As illustrated by Examples 1 and 2, an
appropriate decision-making framework
3
x1 ` x2 ` x4 4 2 under uncertainty should explicitly con-
4
sider the consequences of future infeasibil-
or ity within the model. This aspect of mod-
eling responses to future infeasibility sets
5
13x1 ` x2 ` x4 4 2. stochastic programs apart from their de-
4
terministic counterparts. In the stochastic-
The vector (0, 2, 0, 0) does not satisfy ei- programming literature, two approaches
ther of these equations and thus is infeasi- are widely studied: one is based on mod-
ble under uncertainty! Under uncertainty, eling future recourse (response) and an-
the formulation in which random vari- other restricts the probability of infeasibil-
ables are replaced by their expected values ity (typically equivalent to system failures)
may not provide a solution that is feasible to be no greater than a prespecified
with respect to the random variables. threshold. The first approach yields the
Example 2: Wait and See so-called recourse problems, and the sec-
Another approach that practitioners ond approach yields problems with proba-
often adopt is based on a wait-and-see bilistic (or chance) constraints. While a
analysis (sometimes referred to as scenario specific application may call for both ap-
analysis or what-if analysis). This ap- proaches, we discuss them separately.
proach mimics the process of delaying all The stochastic-programming literature
decisions until the last possible moment, also considers another problem: the distri-
after all uncertainties have been resolved. bution problem. Researchers focus on
As a result, the LPs associated with all characterizing the distribution of the opti-
possible outcomes of the random quanti- mal value or optimal solutions of random
ties are solved. This yields a collection of LPs. As with wait-and-see problems, the
decision vectors, one for each possible out- distribution problem does not provide a
come of the random variable(s). In gen- decision-making framework. Nevertheless,
eral, none of these solutions may be it provides a mathematical common
worthwhile. For example, consider the ground between the second-stage random
two possible realizations of the problem in LP in recourse problems and the random
Example 1. The solution associated with LP of the wait-and-see approach. From a
(a21, a22) 4 (1, 3/4) is (11, 3, 0, 0.75), computational point of view, this problem
while the solution associated with (a21, a22) remains a major challenge [Prekopa 1995,
4 (13, 5/4) is (12/17, 32/17, 0, 0.75). As chapter 15].
with the solution to the expected-value Objectives Under Uncertainty
problem, neither of these solutions is feasi- A great deal of research revolves
ble with respect to the alternate outcome. around the choice of objectives in decision
INTERFACES 29:2 36
STOCHASTIC LINEAR PROGRAMMING
making under uncertainty. One of the with prosperous and recessionary times
more common objectives is to optimize ex- may be decidedly different. To meet such
pected costs (or returns). However, as de- managerial objectives under uncertainty, it
cision makers, we might be interested in may be appropriate to minimize expected
the variability of costs (or returns) associ- absolute deviations from set goals.
ated with a plan. More generally, a deci- (3) Vector optimization under uncertainty
sion makers choices may be guided by a is a class of models that generalize the sto-
utility function. In decision-making mod- chastic goal-programming approach. An
els for an individual, the concept of a util- example of a multiobjective model would
ity function has many merits, although its be a travelers advisory system to recom-
specification can be elusive. The notion of mend routes from origin to destination in
a utility function can become even more Tucson, Arizona. Because flash floods oc-
elusive in large-scale applications of LP. In cur during the monsoon season in Tucson,
the following, we discuss four of the more the advisory system must include low wa-
common objectives for large-scale LPs un- ter level on the roads as one of the objec-
der uncertainty: tives. In addition, it should incorporate the
(1) Minimization of expected costs is by traditional objective of minimal travel
far the most common objective used in time. Because these objectives are essen-
large-scale optimization under uncer- tially noncommensurate, it is appropriate
tainty. For such applications as planning to adopt the vector-optimization frame-
power generation, average seasonal cost work. Furthermore, since the route is to be
per day reflects the repetitive cost of sup- recommended before a potential down-
plying electricity. For some applications in pour, water levels are random variables
telecommunications systems, system per- (as are travel times). This results in a
formance is often measured in terms of vector-optimization problem under
average unserved demand. Finally, in uncertainty.
production-and-inventory systems, it is (4) Minimization of maximum costs is an
common to use average production and alternate class of models. There are vari-
holding costs in evaluating the cost effec- ous interpretations of the term minimax in
tiveness of a system. For such systems, the stochastic-programming models. In one in-
expected cost criterion is easily justified. terpretation, no distributional information
(2) Minimization of expected absolute de- is available, and all that is known is the
viations from goals is a class of objectives set of possible outcomes. In this case, the
that results from extending goal- minimax objective minimizes the maxi-
programming techniques to account for mum loss among all possible outcomes of
uncertainty. In some cases, it may be ad- the random variable. A similar class of
vantageous to specify goals that depend problems arises in the case of partial infor-
upon particular scenarios. For example, mation regarding the probability distribu-
production goals may depend upon eco- tions. For instance, one may have informa-
nomic factors that are modeled as uncer- tion regarding some characteristics of the
tain quantities. Thus, the goals associated distribution (for example, support, mean,
MarchApril 1999 37
SEN, HIGLE
and variance), and the set of probability sponse for each outcome of the random
measures of interest may be those that elements that might be observed. In gen-
share those characteristics. A worst-case eral, this response will also depend upon
approach under partial information is one the first-stage decisions. In practice, this
in which we choose a decision that mini- type of planning involves setting up poli-
mizes maximum expected loss, regardless cies that will help the organization adapt
of the distribution (from among the class to the revealed outcome. For example, in
with the specified characteristics). When production and inventory systems, the
the class of distributions can be character- first-stage decision might correspond to
ized as a polyhedral set, this class of prob- production quantities, and demand might
lems can be solved using generalized LP. be modeled using random variables.
This minimax approach is known to be When demand exceeds the amount pro-
conservative and may be appropriate in duced, policy may dictate that customer
models that plan to avoid catastrophes. demand be backlogged at some cost. This
Thus, models associated with environmen- policy constitutes a recourse response. The
tal planning may appropriately use this exact level of this response (the amount
objective [Pinter 1991]. backlogged) depends on the amounts pro-
In this tutorial, we discuss primarily duced and demanded. Under uncertainty,
models with the expected value objective. it is essential to adopt initial policies that
Two-Stage Recourse Models will accommodate alternative outcomes.
In two-stage recourse models, we explic- Consequently, modeling under uncer-
itly classify the decision variables accord- tainty requires that we incorporate a
ing to whether they are implemented be- model of the recourse policy.
fore or after an outcome of the random In some applications, it is possible to de-
variable is observed. Decisions that are im- viate from prescribed limits, although with
plemented before are known as first-stage a penalty cost. For example, in production
decisions while those after are second-stage and inventory management, a backlogging
decisions. The first-stage decision variables policy leads to shortage costs whenever
can be regarded as proactive and are often the demand exceeds the amount in stock.
associated with planning issues, such as Such a policy is called a simple recourse
capacity expansion or aggregate produc- policy, which we illustrate using the data
tion planning. Second-stage decision vari- from Example 1.
ables can be regarded as reactive and are Example 3: A Simple Recourse Model
often associated with operating decisions. Consider the data for Example 1 and
These second-stage decisions allow us to suppose that the variables (x1, x2, x3, x4)
model a response to the observed out- are all first-stage (planning) variables.
come, which constitutes our recourse. Suppose that the recourse policy allows
When outcomes are revealed sequentially, one to compensate for second-stage dis-
decision making involves a multistage crepancies by incurring a penalty cost of 5
planning problem. per unit of deviation from the right-hand-
In recourse planning, we model a re- side value 2. With this added flexibility,
INTERFACES 29:2 38
STOCHASTIC LINEAR PROGRAMMING
MarchApril 1999 39
SEN, HIGLE
Minimize
5 ` 5 1 5 ` 5 1
1x2 ` y1 ` y1 ` y2 ` y2
2 2 2 2
subject to
x1 `x2 `x3 42
3
x1 ` x2 `x4 ` y`
1 1 y1
1 42
4
5
13x1 ` x2 `x4 ` y`
2 1 y1
2 42
4
11 # x1 # 1.
All other variables are nonnegative.
LP1: Linear programming problem associated with Example 3.
state a prototypical model allowing a sim- two cases. More generally, the per unit
ple recourse policy as follows: cost of bi 1 aix may be g`
i
for positive val-
ues (of this random variable) and g1
i for
Minimize cx ` o giE[|bi 1 aix|]
[
i R negative values. In this case, the costs
subject to A1x 4 b1 used for compensating variables (y` 1
i , yi )
L1 # x # U1. are g` 1
i and gi and the objective function
INTERFACES 29:2 40
STOCHASTIC LINEAR PROGRAMMING
bility translates into greater responsive- clude penalty costs for infeasibility. An
ness and greater profitability. optimal solution for this problem is x1 4
Example 4: A General Recourse Model 0.1176, while x21 4 x22 4 1.8824, x41 4
Consider the situation described in Ex- 0.4706, and all other variables are zero.
ample 1 with the following modification: it The cost-effectiveness of this flexibility is
is now possible to postpone decisions re- reflected in the optimal values of the gen-
garding x2, x3, and x4 until after an out- eral and simple recourse problems. The
come of the random variable is observed. optimal value for the general recourse
Thus the only first-stage variable is x1, problem is 11.8824, which is better than
which must be implemented right away. the optimal value obtained for the simple
This greater flexibility leads to greater prof- recourse case discussed in Example 2,
its, as shown below. To formulate the 11.7772. Because of the added flexibility
decision-making problem, we will con- of our recourse policy, this particular
tinue to assume that minimizing expected model need not include penalty costs for
cost is an appropriate objective. infeasibility.
Since x2, x3, and x4 are implemented af- As with the formulation of a simple re-
ter an outcome is observed, we define one course model, we will present the general
set of second-stage (recourse) variables for recourse formulation as an extension of an
each outcome. Thus let (x21, x31, x41) de- LP problem:
note the second-stage (recourse) variables
Minimize cx
when the outcome is (1, 3/4) and let (x22,
subject to Ax 4 b
x32, x42) denote the recourse variables
L # x # U.
when the outcome is (13, 5/4). Recalling
that each outcome occurs with probability Suppose that the decision maker speci-
1/2, we can formulate the two-stage pro- fies a subvector of x, say x1, as the first-
gram with general recourse as in LP2. stage decision variables. As in Example 3,
Because of the added flexibility of our these variables cannot be postponed until
recourse policy, this model need not in- better information is available. The re-
Minimize 1 1
1 x 1 x
2 21 2 22
subject to x1 `x21 `x31 42
x1 3 `x41 42
` x21
4
x1 `x22 `x32 42
13x1 5 `x42 4 2.
` x
4 22
11 # x1 #1
All other variables are nonnegative.
LP2: Linear programming problem for the two-stage program with general recourse.
MarchApril 1999 41
SEN, HIGLE
maining variables, say x2, can be post- unique realization of these quantities (Bs,
poned. Naturally, with this temporal divi- A2s, c2s, b2s). If S is a discrete set, then for
sion of the problem, two types of each s [ S, let ps 4 P{(B, a2, c2, b2) 4 (Bs,
constraints arise: constraints that involve A2s, c2s, b2s)}. Also, let x2s denote the re-
only the first-stage variables (x1), and con- course response associated with scenario s.
straints that may involve both sets of vari- The two-stage program with general re-
ables. Thus, it is convenient to think of a course may now be written as follows:
submatrix A1 (of A) and a subvector b1 (of
b) yielding a subset of the constraints, A1x1 Minimize c1x1 ` opsc2sx2s
s[S
(2)
4 b1. The remaining constraints involve x1
subject to A1x1 4 b1
and x2, which we write as Bx1 ` A2x2 4
b2. Finally, the cost vector c is partitioned Bsx1 ` A2sx2s 4 b2s s [ S
as (c1, c2) so that we may rewrite the for- L1 # x1 # U1, L2 # x2s # U2 s [ S. (3)
mulation as
This formulation is unlike the simple re-
Minimize c1x1 ` c2x2
course formulation, in that some (or per-
subject to A1x1 4 b1
haps all) choices of x1 that satisfy (2) can
Bx1 ` A2x2 4 b2
render (3) infeasible for some scenarios. It
L1 # x1 # U1, L2 # x2 # U2.
is possible to include compensating vari-
It is convenient to think of this deter- ables (with positive penalty costs) to en-
ministic LP as the core problem from sure that the resulting problem is feasible.
which the stochastic LP will be derived. It Furthermore, it can be shown that this ex-
models the time-staged dynamics of the tended formulation always has a lower
interactions among the decision variables. optimal value than a formulation in which
The constraints A1x1 4 b1 include the the decision maker restricts all decision
immediate constraints, those that involve variables in x (the vector from the deter-
only the variables that cannot be delayed. ministic LP) to be first-stage decisions and
As such, there are no random variables in only a simple recourse policy is allowed in
the immediate data (c1, A1, b1). The ran- the second stage.
dom variables appear in the second stage The stochastic program with general re-
of the problem, which includes the vari- course is also referred to as a problem
ables x2 and can be postponed until the with random recourse, since the matrices
uncertainties are realized. Thus, we con- A2s are allowed to depend on the outcome
sider the second-stage data to include ran- s [ S. However, since the term random re-
dom variables, so that we express them as course might be misconstrued as a case in
(c2, B, a2, b2) (here, we use ; to indicate a which the decision maker has no control
random entity). over the recourse policy, we use the term
To formulate the stochastic LP, let S de- general recourse. When the matrices A2s are
note an index set of all possible outcomes the same for all s [ S (that is, A2 is not
of the second-stage quantities (B, a2, c2, b2) random), the stochastic program is said to
such that each s [ S corresponds to a have fixed recourse. Even in such cases, the
INTERFACES 29:2 42
STOCHASTIC LINEAR PROGRAMMING
random right-hand-side vector, b2, causes dependent on any outcome of the random
the recourse decision itself to vary with s, variable and hence are nonanticipative. An
and hence the fixed-recourse formulation alternate statement of this requirement is
retains the variables x2s, s [ S. Finally, the given in the scenario formulation below:
special case of fixed recourse, in which A2s
4 [I, 1I] (where I denotes an identity ma-
Minimize ops [c1x1s ` c2sx2s]
s[S
trix) yields the simple recourse model dis- subject to A1x1s 4 b1 s [ S
cussed earlier. Bsx1s ` A2sx2s 4 b2s s [ S
A general recourse problem is said to x1 1 x1s 4 0 s [ S (4)
have complete recourse if for any choice of L1 # x1s # U1, L2 # x2s # U2 s [ S.
x1, a feasible recourse decision is possible
for all outcomes s [ S. The simple re- In this formulation, the variables x1s are
course formulation possesses complete re- dependent on the outcome s. However,
course. A slightly less restrictive property constraint (4) explicitly enforces imple-
is that of relatively complete recourse mentability by requiring that all outcomes
whereby one requires that a feasible re- agree on the same planning decision x1.
course decision be possible for all out- We can obtain a slightly more compact
comes s provided the first-stage decision representation of this formulation by re-
(x1) satisfies the first-stage constraints (A1x1 quiring first A1x1 4 b1 and then requiring
4 b1, L1 # x1 # U1). By using penalty (4). By doing so, we avoid replicating the
costs for deviations from constraint satis- first set of constraints for each outcome.
faction, one can ensure complete recourse Both of these are equivalent representa-
in any problem. tions of the two-stage SLP with general re-
One of the more important notions in- course. The particular representation used
corporated within a stochastic program- typically depends on the algorithm being
ming formulation is that of implementabil- used to solve the problem.
ity (or nonanticipativity). This term reflects Note that the general recourse problem
the requirement that under uncertainty, is a finite-dimensional linear program
the planning decisions (x1) must be imple- whenever S is a finite set. However, when-
mented before an outcome of the random ever the random variable is continuous
variable is observed. That is, the planning these formulations lead to infinite dimen-
decision is made while the random vari- sional problems. Under these circum-
able is still unknown, and therefore it can- stances, it is more convenient to state the
not be based on any particular outcome of model in the following decomposed form:
the random variable. Thus the wait-and- 1)]
Minimize cx1 ` E[h(x
see approach, which is anticipative, is not
subject to A1x1 # b1
an appropriate decision-making frame-
L1 # x1 # U1
work for planning. On the other hand, the
here-and-now approach embodied in the where each outcome hs(x) of the random
two-stage SLP with general recourse pro- variable h(x) is a function of the LP de-
vides planning decisions (x1) that are not fined by the outcome (c2s, A2s, Bs, b2s) of
MarchApril 1999 43
SEN, HIGLE
the random variable (c2, a2, B, b2). That is, Although the recourse function is con-
vex, it is not, in general, differentiable. It is
hs(x1) 4 Minimize c2sx2s (5)
well known from LP theory that the value
subject to A2sx2s 4 b2s 1 Bsx1
of a linear program as a function of its
L2 # x2s # U2.
right-hand side is piecewise linear and
This decomposed formulation is con- convex (when the LP is stated as a mini-
venient when the sample space S contains mization problem). Hence every outcome
either a large number of atoms (in the case of h() is a piecewise linear function. It fol-
of discrete random variables) or a contin- lows that, if the number of outcomes of
the random variable is finite, then E[h(x 1)]
uum (in the case of continuous random
1)] is referred
variables). The function E[h(x is a convex combination of finitely many
to as the recourse function. This formulation piecewise linear functions and conse-
emphasizes the time-staged nature of the quently piecewise linear. It is therefore
decision problem. That is, the selection of clear that for problems with discrete ran-
x1 is followed by the selection of x2, which dom variables, the recourse function is
is undertaken in response to the scenario piecewise linear and therefore nondiffer-
that unfolds. Thus, the first decision, x1, entiable in general. Indeed conditions re-
represents the immediate commitment quired to ensure differentiability of the
made, while the second decision is de- recourse function are quite stringent, re-
layed until additional information is ob- quiring absolutely continuous random
tained. For this reason, when solving a re- variables for all elements of the right-hand
course problem, one typically reports only side in (5) [Kall 1976].
the first-stage decision vector. Scenario Construction
Much of the difficulty associated with Each scenario corresponds to a particu-
recourse models may be traced to difficul- lar outcome of the random quantity (c2, a2,
ties with evaluating and approximating B, B2). It is largely a matter of notational
the recourse function. In essence, the diffi- convenience that we refer to these vectors
culty in solving the recourse problem may and matrices as being random. In most
be attributed to the evaluation of the ex- cases, only a small number of the elements
pectation of the random linear-program are actually random; the rest are constant
value function, h(x1), which involves mul- (that is, degenerate random variables). In
tidimensional integration. Notwithstand- the examples weve presented (Examples
ing the impracticality of the multidimen- 1-3), only two coefficients are random. In
sional integration of this particular defining the set of scenarios, it is necessary
function, the recourse function possesses to identify all possible outcomes of (c2, a2,
one of the most sought-after properties in B, B2). This is equivalent to identifying the
all of mathematical programming, namely values of those elements that are fixed and
convexity. the set of all possible outcomes of those
Theorem 1 [Wets 1974]: The recourse elements that vary. In undertaking this
1)], is convex over its effec-
function, E[h(x last task, it is important to note the dis-
tive domain D 4 {x [ X | E[h(x 1)] , `}. tinctions between models of dependent
INTERFACES 29:2 44
STOCHASTIC LINEAR PROGRAMMING
MarchApril 1999 45
SEN, HIGLE
vices, it is customary to plan for a grade of Theorem 2 [Prekopa 1971]: Suppose that
service based on the probability of an- the matrix A2 is deterministic, p [ (0, 1) is
swering a call within a prespecified time given, and the vector b2 has a log-concave
limit. In such cases, there is an implicit ac- multivariate probability density function.
ceptance of the inability to meet system re- Then {x | P{A2x > b2} > p} is a convex set.
quirements at all times. Hence the system For the sake of completeness, we in-
is designed in such a way as to meet crite- clude the following definition: A function f
ria most of the time. Such models lead to is said to be log-concave if for all k [ [0, 1]
mathematical programs with probabilistic and z1, z2,
constraints.
As with the recourse models, we can f[kz1 ` (1 1 k)z2] > f(z1)kf(z2)11k.
view this formulation as an extension of When A2 is fixed, the probabilistically
deterministic LP formulations. Suppose constrained model may be stated as
that the constraints of a deterministic LP follows:
are represented in the form Ax > b. Under
uncertainty, suppose that we partition Minimize cx
these constraints as inequalities that con- subject to A1x > b1
tain only deterministic parameters and A2x 1 z 4 0
those that contain at least one random F(z) > p,
variable. The former (deterministic) con- L#x#U
straints will be denoted A1x > b1, and the
where F(z) denotes the joint cumulative
latter will be stated as a probabilistic con-
distribution function of the right-hand-
straint as follows: side vector, b 2 (that is, F(z) 4 P{b 2 # z}).
Minimize cx Prekopa [1989] has introduced a type of
subject to A1x > b1 polynomial multivariate distribution func-
P(A2x > b2) > p, tion that has a product form. This distribu-
L#x#U tion has been shown to be log-concave
and is particularly well suited for geomet-
where p [ (0, 1) denotes the reliability ric programming problems.
with which the system is required to Next we illustrate a case in which a
operate. probabilistic constraint leads to a noncon-
The probabilistic constraint in this for- vex feasible set.
mulation is known as a joint probabilistic Example 5: Nonconvex Feasible Set in
constraint because there may be multiple Probabilistically Constrained Problems
inequalities in the system A2x > b2. In gen- Consider the following problem:
eral, the set of points x that satisfy the con-
straint may be nonconvex. However, when Minimize x1 ` x2
the matrix A2 is known with certainty, subject to P(2x1 ` x2 > b1;
Prekopa [1971] provides conditions under x1 ` 2x2 > b2) > 0.5,
which convexity (of the feasible set) is
assured. where b 1 and b 2 are dependent random
INTERFACES 29:2 46
STOCHASTIC LINEAR PROGRAMMING
variables with joint probability mass func- function of b2 and let Kp be chosen so that
tion given by F(Kp) 4 p. The constraint P(a2x > b2) > p
can be written as F(a2x) > p, or equiva-
b1 b2 P(b 1 4 b1, b 2 4 b2)
lently, a2x > Kp.
0 1 0.5
Other special cases for which a probabi-
1 0 0.5.
listic constraint can be easily converted to
The feasible region for this example is a more standard type of constraint have
shown in Figure 3. Clearly, this set is not been studied. Prekopa [1995] provides an
convex. excellent summary of this subject. In clos-
One of the early stochastic- ing this section, we reiterate that probabil-
programming models studied by Charnes istically constrained models and recourse
and Cooper [1959] was based on multiple models need not be treated as mutually
probabilistic statements, such as P(a2ix > exclusive approaches for modeling uncer-
b2i) > pi, where i is a row index. In some tainty. In certain applications, it is worth-
applications, this notion of feasibility may while to combine the two approaches.
be appropriate. For example, in some ap- Alternative Models
plications within telecommunications- We have outlined the more popular ap-
network planning, analysts often specify proaches in stochastic programming. To
the grade of service for each type of cus- extend the scope of stochastic program-
tomer. Hence the grade-of-service require- ming models, researchers have proposed
ment for each type of customer may be some alternative models. We shall com-
stated in the form of a single probabilistic ment on these more recent approaches.
constraint. To ensure that a meaningful Integrated Chance Constraints
model results, one must carefully capture Prekopa [1973] and Klein Haneveld
the various customers competition for the [1986] have proposed models with so-
network resources. For example, in addi- called integrated chance constraints (ICC).
tion to the probabilistic constraints, one ICCs can be thought of as offering a bal-
may use a network-flow model to capture ance between recourse models and chance
the manner in which the network will be (probabilistic) constraint models. That is,
loaded and thus the potential for blocked ICCs can be used to constrain the expected
calls. In such cases, one can write probabi- or average behavior of some phenomena.
listic constraints involving a single in- In contrast to probabilistic constraints,
equality using the inverse of the cumula- which are interpreted as imposing reliabil-
tive distribution function. ity requirements, ICCs may be used to
Consider a single probabilistic con- constrain availability, average perfor-
straint, in which b2 is a one-dimensional mance, and other similar measures. One of
random variable, the vector a2 is determin- the main advantages of this approach is
istic, and we wish to satisfy that, unlike probabilistically constrained
models that may result in nonconvex fea-
P(a2x > b 2) > p.
sible sets, models based on ICCs are often
Let F denote the cumulative distribution convex.
MarchApril 1999 47
SEN, HIGLE
Figure 3: This is an illustration of the feasible region for Example 5. The shaded region depicts
the set of points that satisfy the probabilistic constraint. The lack of convexity is readily appar-
ent.
To motivate the discussion, consider a hand sides are random, some of his results
situation in which a target budget b [ 5 may be extended to convex functions,
is given, and the cost associated with whose parameters may be random vari-
plan x is given by the random variable cx. ables. A particularly relevant convex func-
A probabilistic constraint that restricts the tion that arises in stochastic programming
probability of exceeding the budget to be is the recourse function, and it gives rise
at most 1 1 p may be written as to models that can be called recourse con-
strained models. To understand this class
P{cx # b} > p. of models, recall that in a standard two-
stage stochastic program with recourse the
That is, cost overruns may be permissible
first-stage decision often denotes a strate-
in extraordinary circumstances. This can
gic plan, while the recourse function asso-
be represented using an ICC by restricting
ciated with the second stage denotes the
the long-term-cost overrun to be at most
expected cost of future operations. Such
a:
recourse models do not explicitly ac-
E[Max {cx 1 b, 0}] # a. knowledge a decision makers attitude to-
ward variability in costs associated with
While Klein Haneveld focuses on linear the second stage. For example, in a
constraints whose coefficients or right- capacity-planning study for a large auto-
INTERFACES 29:2 48
STOCHASTIC LINEAR PROGRAMMING
mobile manufacturer, Eppen, Martin, and makers often wish to investigate trade-offs
Schrage [1989] initially studied a pure between means and variances of costs (or
two-stage stochastic program with re- profits) associated with their decisions.
course. In this application, l(x) denoted the In an attempt to model such trade-offs,
amount of lost revenue associated with Mulvey, Vanderbei, and Zenios [1995]
capacity plan x, which varied by scenario propose a model referred to as the robust
and thus was a random quantity. The ini- optimization (RO) model. Assuming that
tial application of stochastic programming the random variable is discrete, they sug-
used the term E[l(x)] in the objective func- gest that an apparent mean-variance type
tion, which resulted in a two-stage sto- of model may be stated as follows:
chastic linear program with recourse. An Minimize c1x1 ` o pszs ` h soS ps(zs 1 z)2
s[S [
examination of the results of this model
revealed that the minimization of expected subject to A1x1 4 b1
losses yielded inadequate solutions. There Bsx1 ` A2sx2s 4 b2s s [ S
was a clear need to guide the choice of ca- c2sx2s 4 zs
pacity plans toward those that held lost o pszs 4 z
s[S
revenues below a given target, b. To con- L1 # x1 # U1, L2 # x2s # U2 s [ S.
strain the downside risk, Eppen, Martin,
and Schrage [1989] used a recourse- In this formulation, the parameter h . 0
constrained model to successfully restrict may be interpreted as the weight assigned
the decision space to plans that would be to the variance of the random variable z
considered acceptable. Higle and Sen whose outcomes are {zs}, each occurring
[1995] discuss statistical algorithms for this with probability {ps}. It is typically in-
class of problems. tended as a measure of the decision mak-
Robust Optimization ers aversion to objective function variabil-
Stochastic programming has had several ity. A solution to this formulation depends
on the choice of h and the units used in
successes in portfolio-planning models
the formulation. While it is reminiscent of
[Carino et al. 1994; Kusy and Ziemba
the Markowitz mean-variance portfolio-
1986]. While these models optimize an
optimization model, we caution that the
expected-value criterion, they often in-
objective differs from the more appropri-
clude constraints on downside risk that
ate objective
can be modeled using convex functions
[Carino et al. 1994; Dembo 1989]. How- 1)] ` hVar[h(x
Minimize c1x1 ` E[h(x 1)].
ever, financial planners are often inclined
to model variance as a measure of risk. As in previous sections, h(x1) denotes a
This approach has its roots in Markowitz random variable representing the cost of
[1959], which was based on such assump- the optimal second-stage response. The
tions as normally distributed returns. discrepancy between the two models is at-
While these assumptions may not neces- tributed to the fact that zs need not reflect
sarily hold in some applications, decision the optimal second-stage cost for scenario
MarchApril 1999 49
SEN, HIGLE
s. That is, the random variables z and h though x22 is not. Table 1 summarizes the
need not be identical. Once an outcome of failure of the RO model to achieve cost
the random variable has been revealed, minimization.
the appropriate response in the second The data in Table 1 illustrate the dra-
stage is one that yields the least cost. matic differences between the second-
Hence because z is generally different stage response assumed by the RO model
from h, the RO model paints a misleading and the least-cost second-stage response.
picture of the variance of the second-stage For example, when x1 4 10, x21 is the
objective. The following example illus- same in both cases. However, x22 varies
trates this discrepancy. dramatically between the two models. The
Example 6: A Comparison of the Robust RO model uses the suboptimal response x22
and Mean-Variance Models 4 10.5. This artificially increases the cost
Consider a two-stage problem in which of scenario 2 to bring it closer to that of
the first-stage decision is to be chosen in scenario 1, thereby providing the appear-
the range 0 # x1 # 10 with c1 4 16. Sup- ance of less variability. In our example, the
pose that the second-stage data are uncer- ineffiency induced by the RO model re-
tain, with scenarios described as follows: sults in a cost increase of more than 400
For scenario 1, p1 4 0.1 and c21 4 1, A21 percent for the most likely scenario!
4 11, B1 4 3, b21 4 4, so that the con- For the given value of the first-stage
straint is an inequality of the form 3x1 1 x21 variable, x1, the robust model yields
# 4. The cost-minimizing response is x21 second-stage costs that are at least as large
4 Maximize {0, 3x1 1 4}. For scenario 2, p2 as those produced by the least-cost model,
4 0.9 and c22 4 2, A22 4 12, B2 4 1, b22 with probability one. That is, the least-cost
4 5, so that the constraint is an inequality responses, which one obtains from re-
of the form x1 1 2x22 # 5. As in scenario course models, dominate the responses
1, the form of the cost-minimizing re- from the RO model. This is always the
sponse is x22 4 Maximize {0, 0.5x1 1 2.5}. case for the RO model, which provides a
With these data and h 4 1, we solve the strong argument against its use.
robust optimization model and obtain x1 To further illustrate the pitfalls associ-
4 10, x21 4 26, and x22 4 10.5. In this so- ated with the RO model, we solve the
lution, x21 is a cost-minimizing value, al- mean-variance problem with h 4 1 and
1
obtain x*1 4 6 o6 and x21
* 4 14.5 and
7
2nd-Stage Solutions x22 4 12. The mean and variance of h(x
* o 1*)
are 2.45 and 15.8, respectively. Thus, we
Robust Least Cost
see that the solutions obtained from the
so-called robust models are, in general,
x21 26 (26) 26 (26)
x22 10.5 (21) 2.5 (5) structurally unrelated to the solutions ob-
expected cost 21.5 7.1 tained from the mean-variance recourse
variance 2.25 39.69 model and are dominated by the solutions
Table 1: Output from robust and least-cost obtained from the least-cost model.
models.
INTERFACES 29:2 50
STOCHASTIC LINEAR PROGRAMMING
MarchApril 1999 51
SEN, HIGLE
Figure 4: The scenario tree is a useful mechanism for depicting the manner in which events
may unfold. It can also be used to guide the formulation of a multistage SLP model.
scenarios might be completed before oth- mechanism for formulating multistage re-
ers. For this reason, it is convenient to in- course problems. With the evolution of
dex the nodes of the tree, a, b, c, . . ., as time, outcomes are revealed sequentially,
depicted. Of course, each node has a cor- and one can trace a sample path through
responding stage index (for example, node the tree, as indicated by the bold line in
c appears in the second stage), so that one Figure 4. An underlying tree structure ex-
can recover stage information easily if ists even when one uses continuous ran-
necessary. dom variables. However, in this case, the
The scenario tree provides a convenient branches span a continuum, rather than
INTERFACES 29:2 52
STOCHASTIC LINEAR PROGRAMMING
MarchApril 1999 53
SEN, HIGLE
INTERFACES 29:2 54
STOCHASTIC LINEAR PROGRAMMING
MarchApril 1999 55
SEN, HIGLE
INTERFACES 29:2 56
STOCHASTIC LINEAR PROGRAMMING
MarchApril 1999 57
SEN, HIGLE
assume that flooding occurs when the uct i using process j in period t,
stream flow on a tributary exceeds its ca- aijk 4 the number of units of resource k re-
pacity. Reservoirs can be used on certain quired to produce a unit of product i by
using process j,
tributaries to contain stream flow and pre-
hit 4 the cost for each unit of product i
vent it from continuing to a downstream
held in inventory at the end of period t,
location. This leads to a system of linear and
inequalities pit 4 the cost for each unit of product i on
back-order at the end of period t.
T n # Rx,
The uncertain parameters are the
which indicate that at each point of inter- following:
est stream flow is contained. That is, T n dit 4 the demand for product i in period t;
models accumulated upstream flows and dits denotes the value of dit associated with
Rx models accumulated capacities. Thus, if scenario s.
we let p denote the desired reliability of bkt 4 the amount of resource k available in
the reservoir system, the following formu- period t;
lation results: bkts denotes the value of of bkt associated
Minimize ocjxj
with scenario s, and
j ps 4 the probability with which scenario s
subject to 0 #xj # uj j 4 1, . . ., J occurs. Note that
P{T n # Rx} > p. ps 4 P{d it 4 dits, b ikt 4 bikts, t 4 1,. . .,T}.
Finally, the decision variables are as
A Multistage Production-and-Inventory follows:
Model Xijts 4 the number of units of product i
This model is an extension of a deter- produced by process j in period t under
ministic process-selection model presented scenario s, and
by Johnson and Montgomery [1974, Exam- Iits 4 the inventory of product i in period t
ple 4-11, pp. 243244]. Within the model, under scenario s.
each product can be produced by several As time progresses, the collection of out-
alternative processes. However, product comes that may unfold can be organized
demand and resource availability are into a scenario tree. In addition, a node n
modeled as random variables. The objec- in the scenario tree corresponds to a par-
tive is to minimize the expected produc- ticular time period, t(n) and summarizes a
tion costs, including inventory and back- unique unfolding of the random events
order costs, over multiple periods. Given from the initial period until period t(n). To
the nature of inventory and back-order ensure that the model yields solutions that
quantities, a multistage model with simple are implementable, one must ensure that
recourse results. at any given time, scenarios that share a
Let common history are constrained to yield a
T 4 the number of time periods under common production-and-inventory plan at
consideration, that time. Thus, let 1 denote the set of
cijt 4 the per-unit production cost of prod- nodes in the scenario tree. For each n [ 1,
INTERFACES 29:2 58
STOCHASTIC LINEAR PROGRAMMING
MarchApril 1999 59
SEN, HIGLE
area may try to obtain information on the possible. As one might expect, a model
usage of this new service in multiple with few random variables is easier to rep-
ways. It may look at usage data from a resent for computational algorithms and
similar demographic region in a different may be more amenable to exact solution
part of the country. It could also obtain using deterministically motivated algo-
surrogate data from a computer simula- rithms, such as the method developed by
tion model. And finally, it could carry out Rockafellar and Wets [1991]. In many in-
a market survey or perform a test within a stances, it may also be possible to derive
small segment of the region. All of these deterministic upper and lower bounds on
approaches provide estimates of market the value of the stochastic program, as in
demand for the new service, and they are Birge [1982]. Nevertheless, one can easily
likely to be different. With a stochastic- run up against very large-scale stochastic-
programming model, the company can in- programming models for which determin-
clude these alternative forecasts within istic methods soon become inadequate. In
one decision-making model to produce a such instances, sample-based algorithms,
more robust plan. such as the stochastic decomposition
Data Requirements method [Higle and Sen 1991], provide a
Many of the data requirements for practical solution approach.
stochastic-linear-programming models are For any of the approaches mentioned
similar to those of linear-programming above, data on the random variables are
models. The additional data in stochastic usually provided to the algorithms via the
programming are needed to represent un- SMPS format developed by Birge et al.
certainty. In some applications, one repre- [1987]. This data format is based on the
sents uncertainty by subjectively assessing MPS format of mathematical-
weights to assign to possible future sce- programming systems and provides a con-
narios. In such cases, one can build the venient representation of random vari-
stochastic-programming model using few ables in a stochastic-linear program. A
scenarios and set up the model as a large- more recent framework for multistage sto-
scale linear program. Such models are of- chastic programs is available within the
ten solved using off-the-shelf LP software. OSL system marketed by IBM. Finally, the
The case study (from GM) reported by stochastic-programming community is
Eppen, Martin, and Schrage [1989] is such working toward an object-oriented stan-
a model. In many applications, however, dard for representing this class of prob-
econometric models and forecasting sys- lems. We expect it to develop such a stan-
tems provide detailed information regard- dard over the next several years.
ing some of the random variables. Under Acknowledgment
these circumstances, it is difficult to cap- This work was supported in part by
ture the randomness via a handful of sce- Grant No. NSF-DMII-9414680 from the
narios. Nevertheless, it is advantageous to National Science Foundation.
be able to represent the uncertainty in References
terms of only a few random variables, if Birge, J. R. 1982, The value of the stochastic
INTERFACES 29:2 60
STOCHASTIC LINEAR PROGRAMMING
solution in stochastic linear programs with University Press, New Haven, Connecticut.
fixed recourse, Mathematical Programming, Mulvey, J. M.; Vanderbei, R. J.; and Zenios, S.
Vol. 24, No. 3, pp. 314325. 1995, Robust optimization of large scale sys-
Birge, J. R.; Dempster, M. A. H.; Gassmann, tems, Operations Research, Vol. 43, No. 2, pp.
H. I.; Gunn, E. A.; King, A. J.; and Wallace, 264281.
S. W. 1987, A standard input format for Murty, K. 1983, Linear Programming, John Wiley
multiperiod stochastic linear programs, and Sons, New York.
IIASA working paper, IIASA, Laxenburg, Pinter, J. 1991, Stochastic modeling and opti-
Austria. mization for environmental management,
Carino, D. R.; Kent, T.; Myers, D. H.; Stacy, C.; Annals of Operations Research, Vol. 31, pp.
Sylvanus, M.; Turner, A.; Watanabe, K.; and 527544.
Ziemba, W. T. 1994, The Russell-Yasuda Prekopa, A. 1971, Logarithmic concave mea-
Kasai model: An asset/liability model for a sures with applications to stochastic pro-
Japanese insurance company using multi- gramming, Acta Scientifica Mathematica
stage stochastic programming, Interfaces, (Szeged), Vol. 32, No. 3, pp. 301316.
Vol. 24, No. 1, pp. 2949. Prekopa, A. 1973, Contributions to the theory
Charnes, A. and Cooper, W. W. 1959, Chance of stochastic programming, Mathematical
constrained programming, Management Sci- Programming, Vol. 4, No. 4, pp. 202221.
ence, Vol. 6, No. 1, pp. 7379. Prekopa, A. 1989, Numerical solution of prob-
Dembo, R. S. 1989, Scenario optimization, abilistic constrained programming prob-
Technical report, Algorithmics, Inc., Toronto, lems, in Numerical Techniques for Stochastic
Canada. Optimization, Eds. Yu. Ermoliev and R. J-B.
Eppen, G. D.; Martin, R. K.; and Schrage, L. Wets, Springer-Verlag, Berlin, pp. 123139.
1989, A scenario approach to capacity plan- Prekopa, A. 1995, Stochastic Programming, Klu-
ning, Operations Research, Vol. 37, No. 4, pp. wer Academic Publishers, The Netherlands.
517527. Prekopa, A. and Szantai, T. 1978, Flood con-
Ferguson, A. R. and Dantzig, G. B. 1956, The trol reservoir system design using stochastic
allocation of aircraft to routes, Management programming, Mathematical Programming
Science, Vol. 3, No. 1, pp. 4573. Study 9, pp. 138151.
Higle, J. L. and Sen, S. 1991, Stochastic decom- Rockafellar, R. T. and Wets, R. J-B. 1991, Sce-
position: An algorithm for two-stage linear narios and policy aggregation in optimiza-
programs with recourse, Mathematics of tion under uncertainty, Mathematics of
Operations Research, Vol. 16, No. 3, pp. 650 Operations Research, Vol. 16, No. 1, pp. 119
669. 147.
Higle, J. L. and Sen, S. 1995, Recourse con- Sen, S.; Doverspike, R. D.; and Cosares, S. 1994,
strained stochastic programs, Annals of Network planning with random demand,
Operations Research, Vol. 56, pp. 157175. Telecommunications Systems, Vol. 3, No. 1, pp.
Johnson, L. A. and Montgomery, D. C. 1974, 1130.
Operations Research in Production Planning, Talluri, K. and van Ryzin, G. 1996, An analy-
Scheduling and Inventory Control, John Wiley sis of bid-price controls for network revenue
and Sons, New York. management, Technical report, Columbia
Kall, P. 1976, Stochastic Linear Programming, University, New York.
Springer-Verlag, Berlin. Wets, R. J-B. 1974, Stochastic programs with
Klein Haneveld, W. K. 1986, Duality in Stochas- fixed recourse: The equivalent deterministic
tic Linear and Dynamic Programming, Lecture program, SIAM Review, Vol. 16, No. 3, pp.
notes in economics and mathematical sys- 309339.
tems, No. 274, Springer-Verlag, Berlin. Williamson, E. L. 1992, Airline network seat
Kusy, M. I. and Ziemba, W. T. 1986, A bank control, PhD dissertation, MIT.
asset and liability management model,
Operations Research, Vol. 34, No. 3, pp. 356
376.
Markowitz, H. 1959, Portfolio Selection, Yale
MarchApril 1999 61