Chance-Constrained Programming

Author(s): A. Charnes and W. W. Cooper

Source: Management Science, Vol. 6, No. 1 (Oct., 1959), pp. 73-79
Published by: INFORMS
Stable URL: http://www.jstor.org/stable/2627476
Accessed: 27/10/2010 04:27

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=informs.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Management Science.

http://www.jstor.org
 CHANCE-CONSTRAINED PROGRAMMING*tt
A. CHARNES2 AND W. W. COOPER.3

A new conceptual and analytical vehicle for probleins of temporal planning

under uncertainty, involving determination of optimal (sequential) stochastic
decision rules is defined and illustrated by means of a typical industrial ex-
ample. The paper presents a methoc. of attack which splits the problem into
two non-linear (or linear) programming paits, (i) determining optimal prob-
ability distribtitions, (ii) approximating the optimal distribuitions as closely
as possible by decision rules of prescribed form.

Introduction
The problem of stochastic (or better, chance-constrained) programming is
here defined as follows: Select certain random variables as functions of random
variables with known distributions in such a manner as (a) to maximize a func-
tional of both classes of random variables subject to (b) constraints on these
variables which must be maintained at prescribed levels of probability. More
loosely, the problem is to determine optimal stochastic decision rules under
these circumstances. An example is supplied in [2]. Temporal planning in which
uncertainty elements are present, but in which management has access to "con-
trol variables" with which to influence outcomes, is a general way of character-
izing these problems. Thus, queuing problems in which the availability of
servers, customers, or both are partly controllable fall within this classification.
It should be noted, that the constraints to be maintained at the specified
levels of probability will typically be given in the form of inequalities.
The method of attack which will be outlined in this paper consists of splitting
the problem into two parts: (i) determining distributions which maximize the
functional, subject to the probability constraints; (ii) approximating the distri-
butions so determined as closely as possible (in some sense) by functions of the
known random variables of some prescribed or admissible class. The functions
so determined can be regarded as approximations to the optimal stochastic
decision rule from the admissible class of such rules.
Specifically, for discrete distributions, and piecewise linear functionals along
* Received November, 1958
t Part of the research underlying this paper was uridertaken for the project Temporal
Planning and Management Decision under Risk and Uncertainty at Northwestern University
and part of the project Planning and Control of Industrial Operations at Carnegie Institute
of Technology. Both projects are under contract with the U. S. Office of Naval Research.
Reproduction of this paper in whole or in part is permitted for any purpose of the United
States Government., Contract Nonr-1228 (10), Project NR 047-021 and Contract Nonr-
760 (01), Project NR 047011.
1 This paper provides the background for the invited lecture presented by the authors
at the Fifth International Meeting, The Institute of Management Sciences, Philadelphia,
October 17, 1958.
2 The Technological Institute, Northwestern University
Carnegie Institute of Technology
73
74 A. CHARNES AND W. W. COOPER

with linear inequalities involving the random variables-to be maintained at

prescribed levels of probability-we factor the original problem into two new
problems: (1) a problem which determines the coefficients of the step functions
comprising the optimal discrete distributions-or, alternately, the discrete
probability frequencies-and (2) another problem which determines the pa-
rameters of the optimal decision rule in the sense of a "best approximation".
To fix the ideas, we present first a formulaiion of a partly controllable situa-
tion which has heretofore been formulated and treated (inadequately) by queuing
models. We consider terminal tankage facilities supplied by a refinery and with
pickup by tankers.
Let
Rj-= amount of oil sent to the tankage facilities in the jth period
Sj' = amount of oil which is picked up from the facilities in the jth period
by the ath type of tanker.
Ij -inventory on hand at the beginning of the jth period.
TA = tankage (i.e., storage facilities) available.
Thus,

(1) Ij = Io + R - S, )]

relates the inventory Ij to the initial inventory, Io, and the rates of input and
withdrawal.
Trheobjective is to minimize the expected total cost of input and withdrawal
including such features as the cost of changing input rates, demurrage, charter
and dispatch. Formally this may be stated as
minimize EfC(R, S)]
where the vectors R and S indicate that cost is to be considered over all periods.
The expectation sign, E, indicates that these vector variables are to be considered
stochastically in assessing the expected value of the total costs, C.
This general formulation can be given a variety of specific forms and inte-
pretations. For example,

(2a) (R, S)- Cj (Rj) + Ewj Rj-Rj_l + EdjAJ-

j j j

where
Ij+ Rj- =Sja Aj = An+-Aj-

Ai+) A j- > 0.
and

A = Aj - AjO when Aj > O

2 l-Aj when Aj < O
A + Aj l+ Aj _f0 when Ai < 0
2 "Aj when Aj>O
 CHANCE-CONSTRAINED PROGRAMMING 75

In this case
Ej Cj(Rj) refers to costs of handling due to the amount Rj,
(2b) >j w jI - Rj-l I is the component of total costs due to changing the
rate in period j,
j djA7- are demurrage costs,
and other details may be added as necessary.
In this model the Sj are random variables. However the random components
are stated in terms of deviations from scheduled amounts. In addition, the model
is of a conditional stochastic variety so that past data, developing experience
and forecasts of the future all enter into determining the optinrnm Rj .
The direct problem is stated in terms of meeting the objective specified in (2)
subject to the following constraints.
(3a) Pr{Ij+Rj > ZSj} I> oj j = 1,2, ...
ci

(3b) Pr {Ij+ Rj< TA} > Y'v, j = 1,2,* .

where "Pr" means "probability" and fj and :yj are the prescribed confidence
levels desired for each period j = 1, 2, .
The long range problem is concerned with evaluating TA in terms of the
effects on expected cost C. Notice, however, the two significant features asso-
ciated with this evaluation in (3b): One, there is a valuation element associated
with varying TA while the :yj are fixed. Two, there is a valuation element asso-
ciated with cost effects on the risks of not meeting schedules or varying R as
different levels of confidence are specified. The former is associated with cost
reductions (or increases) arising from varying TA (and hence R and S in response
thereto) at given levels of risk. Hence both risk and service may be evaluated
in various combinations when studying the alteration of tankage.
We next illustrate the factoring procedure by means of the following simpler
example:
max E {Pi(Si)Si - (c3 + Tj)R3-K,
(I-+Kii+)}
subject to
(i) PrJIj+ Rj _ Sj+ Imin} I ai
(1) (ii) Pr fIi + Rj - iSj ? Imax} > P

(iii) R >_ 0
where, in period j = 1, 2, * , N,
Ij = Inventory on hand at start
Rj = Production rate to be scheduled
S1 = Sales demand
= Minimum inventory to be maintained
= Storage capacity
Kj = Inventory carrying charge
cj, Tj = Production and Transport Cost (per unit), respectively,
Pj(Sj) = Unit sales price as function of sales demand
76 A. CHARNES AND W. W. COOPER

and E indicates expectation. I.e., the objective is to maximize the expected net
return over an N period planning horizon subject to the probabilistic constraints
which are to be honored in each of the j = 1, 2, ***, N periods.
This problem-indeed, a more general one involving a convex functional-
was treated in [2] by means of a restricted class of decision rules which made it
possible to transform the problem into a deterministic one (involving certainty
equivalents) which could be solved by a specially developed convex programming
algorithm. The purpose of the present paper is, by contrast, to suggest a new
analytic method which offers the possibility of handling a much wider class of
decision rules.
For this example--as well as for the class of rules which will be considered-
the observable stochastic variables are independent. Because the decision on R?
must be made before Sj is observed, the admissible class of decision rules for RI
(or A j = Ij + Rj) can involve, as random variables, only S1, Sj.. ***, .
The Aj may thus be considered statistically independent of the respective Sj .
This means that the distribution of A i - Sj is given by a convolution of the
distributions of A1 and -Sj.
In considering the first of the two parts into which the problem is to be split
we shall transform it into a mixed integer programming problem for determining
the relative frequencies of distributions for the Aj . By definition of Ij and Aj,

(2) Rj = Aj -Aj-1 + S-1 and

Ij = A j1 - Sji-
Thus, by well-known properties of the expectation operator, E, the maximand is
reduced to a linear function of the (yet-to-be determined) relative frequencies
of the Aj.
Similarly, the expression

(3) Pr {A; - Sj _ x} = f(y)gj(x - y) dy,

where Ajand gj are the density functions respectively, for -SSj and Aj, is a
linear function of the Xj Xthe (unknown) relative frequency of the rth possible
amount for Aj . Thus, (1 i) and (1 ii) go over into linear inequalities involving
the Xjr. To these we must append the conditions
E Xjr = 1
(4) r
Xjr
i 0
so that the Xjr may be interpreted as relative frequencies. Also, since (1liii)
may be rewritten
(5.1) A > Aj1 -Sjl
we may interpret it as
(5.2) min Aj > max (Aj-Sj_1)
 CHANCE-CONSTRAINED PROGRAMMING 7t

in order to transform the non-negativity requirement to a condition on the

frequency functions.
As is known, the density function for Aj, - Sj-l has its relative frequencies
as linear functions of the Xj1, r's; let these be denoted by Xj-l,k . The require-
ment (5.2) can then be expressed by

hj EX
,j-1)k i =1, 2, ... n
k=l

j 1, 2, AN
(6)

0 ? hi ? 1
and hj shall be an integer.
We have thus transformed the first part of the problem into the form: maxi-
mize a linear function of the Xjr subject to the linear conditions given by (4)
through (6), plus the requirement that the hj's shall be integers. General methods
for such mixed-integer problems have been provided by E. M. L. Beale [1]
and R. Gomory [3].
The solution to the first part thus leaves us with a solution to a problem which
is less restricted than the one originally stated. It should also be obvious that
more general piecewise linear functionals can be comprehended via this mode of
attack (with at worst mixed integer requirements) and that more complicated
linear stochastic constraints may be handled where suitable variable transforma-
tions-e.g., analogous to those from the Rj, I, to the AZ-permit a translation
into convolutions (hence linear inequality conditions) for the unknown relative
frequencies.
Knowing now the solution to the first part, we next seek to approximate as
closely as possible the distributions for the A j by means of functions
(7) Aj = Aj(S, *** , Si-,)
where the functions are from some specified admissible class; the class of possible
stochastic decision rules any one of which will prescribe the value of Aj given
S1, * , Sj_l .4 For example,the functions
i-1
(8) A j = ajo + air Sr

comprise the class of linear decision rules.

For this class-i.e., (8)-the problem of approximation is probably best carried
out in terms of characteristic functions. E.g., the characteristic function corre-
sponding to the decision rule for Aj-i.e., D(Aj)-is given by [4]
j-.1
(9) 'OAD()(t) = II 3 (Ojr t),
r=0

4This class may possibly be extended to include still other variables with known dis-
tributions that might improve the fit.
78 A. CHARNES AND W. W. COOPER

a product of the known characteristic functions of the Sr involving the unknown

aej in their argument. By choice of the a,j we seek to approximate OA1(t), the
characteristic function for the distribution obtained as a solution to the first
problem, in such a manner that the distribution (or density) function corre-
sponding to 4D(Ai) (t) has (as closely as possible) the desired characteristics of
the distribution of A j .
Evidently there are many ways of specifying the latter problem. For example,
one may approximate some subset of the cumulants, or semi-invariants; this is
equivalent to approximating the mean and (or) other selected moments (and
associated characteristics) of the distribution of A j as closely as possible. Another
possibility is to minimize

f
0o
[ A (t) - Os
ckr (a?r t) dt.
r=0

By Parseval's Theorem this integral is equal to

00
2', [rPAi(X)
-
PD(A,) (X)]2 dx
00

where PA, (x) is the density function corresponding to OkA,(t) and PD(A3)(X) is
the density function corresponding to the decision rule. Still another possibility
would be to weight the dispersion of PD(A,) (x) from PA2(x) by the relative fre-
quencies of PA3 (x)-e.g., to minimize
00

27r f pAi(X)[pAi(X)
-
PD(A3j)(X)] dx.
00

By Parseval's Theorem, and the fact that the Fourier transform of the product
of two functions is the convolution of their individual transforms, this is equiva-
lent to
minimizing f [{0Aj((A -
ID(Ai)) }()]2 dt,
00

where the * denotes the convolution operation. All three of these possibilities
are classical non-linear minimization problems since the air are completely
unrestricted.
As should now be clear, the problem of stochastic (chance-constrained) pro-
gramming involves difficulties of an order incommensurate to that of "certainty"
programming. These difficulties stem fundamentally from the probabilistic con-
straints, which experience (let alone theory) has made clear, are not adequately
represented as some have done by applying the expectation operator to the
stochastic form. It is hoped that the conceptual framework and approximation
ideas above will stimulate additional research on models and methods of this
character which are essential to insight into and progress on management prob-
lems of a temporal nature involving conditional decisions.
References
1. BEALE, E. M. L., "A Method for Solving Linear Programming Problems when Some but
not All of the Variables Must Take Integral Values", Technical Report No. 19,
July, 1958, Statistical Techniques Research Group, Department of Mathematics,
Princeton University.
 CHANGE-CONSTRAINEDPROGRAMMING 79
2. CHARNES,A., W. W. COOPERand G. H. SYMONDS,"Cost Horizons and Certainty Equiv-
alents: An Approach to Stochastic Programming of Heating Oil", Management
Science, 4, No. 3 April, 1958.
3. GOMORY,R., Oral Presentation at American Mathematical Society, Summer Meeting,
1958.
4. GNEDENKO, B. W., "Lehrbuch der Wahrscheinlichkeitsrechnung", Akademie-Verlag
Berlin, 1957.