CD

^ to

COHOtm, ORIOINS. AND USE

OV UNBAR PROORAIMINO

Ocorge B. Dsntslg

F-980

»«•ch 7, 1957

^
COPY /, JOF "3 Approved for OTS release
IARD COPY
«ICROFICHE

DDC

DDCIRA E
 P-980
>-7-57
-11-

SUMMARY
\

This paper discusses the concept, origins, and some of

the applications of linear programming. ( ^
 P-980
3-7-57
-1-

CONCEPTS, ORIGINS, AND USE OF LINEAR PROORAMMINO

George B. Dantzlg

1, CONCEPTS
Too often linear prognunmlng la Introduced by citing a
number of miscellaneous examples that appear to have little
in conmon except that their mathematical description requires
a solution of a system of linear Inequalities that minimizes
(or maximizes) a linear form. Conceptually linear programming
is concerned with building a model for describing the Inter-
relations of the components of a system. It has a certain
philosophy or approach to model building that has application
to a broad class of decision problems. Since this Important
aspect of linear programming has not been receiving enough
emphasis, we shall begin with a review.
The first step In building a model consists In regarding
a system under design as composed of a number of elementary
functions that are called "activities'* (von Neumann used the
term "process" while T. C. Koopmans coined the term "Activity
Analysis" to describe this approach). The multitude of different
type activities In which a system can engage constitutes Its tech-
nology; these are the representative building blocks that can
be recomblned In varying amounts for form a new system. The

1
J. von Neumann, "A Model of General Economic Equilibrium,"
Review of Economic Studies, Vol. XIII (l). No. 33, 19^5-^6,
PP. 1-9.
"Activity Analysis of Production and Allocation," T. C.
Koopmans editor, Wiley and Sons publisher, 1931. The formulation
of the linear programming model can be found In Chapter I,
"The Programming of Interdependent Activities—Mathematical
Discussion" by George B. Dantzlg.
•
 P-980
3-7-57

quantities of different types of activities to be performed

by the new system Is called Its program. It becomes a feasible
program. If the structure reared by the blocks Is mutually
self-supporting and consistent with the Inputs available from
outside the system and the outputs required to be produced by
the system.
As for the building blocks themselves, each activity Is
viewed as a kind of "black box" In which the quantities of
various Items, such as supply, people, money, and machinery,
that flow Into It are transformed into different quantities
of Items that flow out. The black ^ox, of course. Is a myth.
What we mean Is that, given the stated Input flows, we have
the technical know-how to combine them to produce the stated
output flows. For example, In the refining of crude oil, a
simple distillation activity of a certain type crude might be
characterized by

Inputs Distillation Outputs

* <
Activity
1 bbl crude/day .55 bbl. fuel oil
1 bbl heater capacity .20 bbl. dlesel oil
$1.80 .25 bbl. gasoline
components

It seems quite reasonable to suppose that If twice (or

any positive multiple) of the Input flows were made available
that the same technical know-how could be used to produce
 P-980
3-7-57

corresponding multiple of output flows. Thus distilling twice

as many barrels of crude per day would use twice as many barrels
of crude per day, would require twice the distillation capacity,
would cost twice as much, and would produce twice as many barrels
of fuel, diesel, and the lighter type oils (which are used as
components in the blending of gasoline stocks). It is this
proportionality aspect more than any other that characterizes
the linear programming model. Assuming its general truth
within a model, it is natural to consider all activities that
can be generated as a multiple of a fixed activity as forming
<m
*otivity type; each activity within the type is characterized
by the multiple of the fixed activity called the "activity level."
Associated with each item that can flow into or out of
eaoh activity is a material balance equation. To be precise,
for each Item» it is required that the total amount OP hand
equal the amount flowing into the various activities minus
the amount flowing out.
Finally, one of the items of the system is regarded as
precious in the sense that the total amount of it produced
by the system measures the payoff. The programming problem
consists In determining values for the activity levels which
are positive or zero such that the flows of each item (for
these activity levels) satisfies the material balance equations
and such that the value of the payoff is maximum.
These steps, if followed, lead to a well—defined mathematical
model of the system, called the linear programming model.
 P-980
>-7-67

Indeed, let the subscript J ■ I, 2, ..., n denote the J

type activity and x. its quantity; let the subscript i denote
the i type item and a. denote the input of the i type
item into the J type activity per unit quantity of activity
(a,j, if negative, denotes output instead of input); let b-
denote the given available input to the system from exogeneous
sources (b., if negative, denotes the required outputs from
the system) • The material balance equations then become

a11x1 4 a12x2 ^ ... ♦ alnxn - ^

a
21xl * a
22X2 ♦ • • • ♦ a
2nxn ' b
2

(1)

+ a
^l*! m2x2 * •'• + a
mnxn " b
m

where the term a^.x. denotes the total flow of the i item
into the J type activity if its activity level is x,. For
the precious item let c. denote the "cost" per unit level of
Jth activity; then

(2) c^j ♦ c2x2 ♦ ... ♦ c^ - z

 P-980
>-7-57

measures the total system costs (—z measures total payoff).

The linear programming problem then consists In choosing
numbers

(5) x1 > 0, Xg 2 o, ...f xn > 0

satisfying (l), such that the value of ' given by (2) is

minimum.
 P-980
>-7-57

2. ORIGINS

ECONOMIC ORIQIMS OP LINEAR PROQRAMWIWQ

Linear programming is an anachroniam. Logically it should
have begun around 1760 when economists first began to describe
economic systems in mathematical terms* Thus a crude example
of a linear programming model is found in the "Tableau Economiquc"
of Queanay, who attempted to interrelate the roles of the land-
lord, the peasant and the arclean. During the next 173 years
there was, however, little in the way of exploitation of a
linear-type model although it did appear as part of the Walrasian
system in the 1870's« For the most part, mathematical economists
made use of general functions whose parameters were as a rule
unspecified. In their defense, it should be remembered that
very little in the way of facts were available on income,
quantities of production, investments, savings, distribution,
etc., so that the purpose of the mathematical equations was to
describe in a qualitative rather than a quantitative way the
assumed interrelations within a system while the manipulation of
equations corresponded to logical deductions from the assumptions.
The first major impetus to the construction of practical
mathematical models to describe an economy, came about in the
1930's as a result of the great depression for, with the advent
of the "New Deal," there was a serious attempt on the part of
the government to support certain activities which, it was hoped,
would speed recovery.
 P-980
-7-

Wasslly Leontief, professor at Harvard, brought out hie

book "The Structure of the American Economy*1 during this period.
This marked a sharp departure from earlier model building be-
cause a large number of industries (or sectors) were inter-
related in a single model that permitted estimation of the
parameters in a practical and quantitative way. Each "sector11
represented a number of related functions that were lumped
together and called an industry, such as steel. By estimating
the dollar sales of each sector to every other sector it was
possible to develop a so-called input-output model. This was
done by assuming the dollar input of products into sector A
from B was proportional to the dollar output of sector A. The
scope, accuracy, and area of application of Leontief-type models
were greatly extended by the Bureau of Labor Statistics in the
years that followed. It was this work that was generalized in
the post war period into a form suitable for dornamic Air Force
•pplications.
About the same time that Leontief was producing his model,
John von Neumann, world—famous mathematician, published his
"Model of Economic Equilibrium," (1937). His model (like the
Leontief model) was a linear programming model. It was more
general, as it allowed for alternative activities. It was
completely general except it did not possess a general objective
form. His purpose was limited to proving a relation between the
expansion rate and the interest rate on money for a constantly
expanding economy. As far as influence is concerned, this paper,
like many others, had for purpose, an interesting mathematical
■

P-980
3-7^7

theorem. There appears to have been little interest in the model

itself among economists # for, according to T. C. Koopmans, "To
many economists, the term linearity is associated with narrowness,
restrictiveness, and inflexibility of hypothesis.
However, it was the broad applicability of the linear
programming approach to military planning that stimulated
economists during the post war period to recognize the poten-
tialities of this type model for studying economic problems.

"Introduction," Activity Analysis of Production and

Allocation, T. C. Koopmans (ttditor), page t>. —
 P-980

MIUTARY ORIOINS OF PROGRAM PLANNIKO

The following quotation from a paper by Wood and Oelsler

traces developments In the military area through World War XI:
"It was once possible for a supreme commander to plan
operations personally. As the planning problem expanded In
space, tlmef and general complexity, however, the inherent
limitations in the capacity of any one man were encountered.
Military histories are filled with instances of commanders who
failed because they bogged down in details, not because
i
they
could not eventually have mastered the details, but because
they could not master all the relevant details in the time
available for decision.
*Gradually, as planning problems became more complex, the
supreme commander came to be surrounded with a general staff
of specialists, which supplemented the chief In making decisions.
The existence of a general staff permitted the subdivision of
the planning process and the assignment of experts to handle
each part. The function of the chief then became one of selec-
ting objectives, coordinating, planning, and resolving conflicts
between staff sections."
_

The quotation that follows is taken from a paper by Marshall

K. Wood and Murray A. Oelsler, "Development of Dynamic Models for
Program Planning, Chapter XII, Activity Analysis of Production and
Allocation, pp. 189-192.
 P^980
>-7-57
-10-

In the U.S. Air Force, for example, during the war a

group under the guidance of Profesaor E. P. Learned of Harvard
Buslneas School evolved a scheme which was a program for pro-
gramming by means of which the work of many staff and command
agencies participating in the process was carefully scheduled.
In order to obtain consistent programming the ordering of the
steps in the schedule were so arranged that the flow of in-
formation from echelon to echelon was only in one direction.
Even with the most careful scheduling, it took about seven
months to complete the process.
After the war the Air Force consolidated the statistical
control, programming monitoring, and budgeting functions under
the Air Force Comptroller^—General E. W. Rawlings. It became
clear to members of this organization that coordinating quickly
and well the energies of whole nations in a total war required
scientific programming techniques. Undoubtedly this need had
occurred many times in the past but this time the means of
accomplishment were at hand—these were the electronic computer
developments and a new way to describe the interrelations of
a complex organization so as to permit computation on electronic
computers. Here, of course, I am referring to linear programming.
Intensive work began in June 19^7 in a group that later
(October 19^8) was given the official title of ProjectTSCOOP"
(Scientific Computation of Optimum Programs).
While the initial proposal was to use a linear programming
model to develop Air Force programs, it was recognized at an
 P-980
>-7-57
-11-

early date that even the moat optimistic estimates of the

efficiency of future computing procedures and facilities would
not be powerful enough to prepare detailed Air Force programs.
Accordingly, in the Spring of 19^8, there was a second SCOOP
proposal that special linear programming models, called tri-
angular models, be developed whose computational solution
would parallel the stepwise staff procedure. Since 19^8 the
Air Staff has been making more and more active use of mechani-
cally computed programs. The triangular models are used by
them for detailed programs, while the general linear programming
*

models are used in certain areas.

'
P-980
3-7-57
-12-

3. APPLICATIONS

It is of course impossible in a brief talk to do more

than give a short list of military and industrial applications
and to illustrate briefly. Military applications include:
1. Optimal routing for air transport
2. Time phased distribution of supply from factories
and depots to bases
3* Allocation of electronic equipment to naval vessels
4. Production smoothing problems
3« Contract award problems
6. Communications system design and message routing
7* Personnel assignment problems
8. Maximal flow in transportation networks
9* Minimal tanker fleets to meet a fixed schedule

Some of these which I have cited represent areas where

the linear programming tool has been proposed or has been used
from time to time. Others« like the production smoothing«
contract awards« and supply distribution« have found frequent
application.
In the spring of 1951* four years after the birth of
linear programming« Chames and Cooper at Carnegie began to
pioneer industrial applications. With Bob Mellon of Oulf Oil
Company« they were the first to develop the application of
linear programming to refinery problems. This has since proved
to be the most fruitful of industrial applications. At the
 P-980
3-7-57
-15-

recent Management Science meeting« Oarvin, Crandall, John, and

Spellmann reported on a survey of applications in the oil
industry. So intense has been the development in this area
that it now covers every phase of their worki (1) exploration,
(2) production, (3) refining, (4) distribution.
Next to the petroleum industry, the food industry is
perhaps the second most active customer. Heinz Company uses
it to determine which of a half a dozen plants should ship
ketchup to seventy warehouses located in all parts of the
country. A major milk producer has been considering it as a
means of determining which of its dozen or so milk producing
centers should ship canned milk to thousands of warehouses in
this country and to foreign markets. Armour and Company use
it to determine the most economical mixture of feeds for
animals.
In metal processing applications, the president of Argus
Cameras reports a savings of 13^,000 per quarter because the
price implications of linear programming told them whether to
make or to buy a part. The production planning manager of
SXF says that the linear programming helped them save 6,000
machine hours each month, or |100,000 per year over the Oantt
Chart Control System used by them prior to 1933*
Important applications are being made in other countries.
Thus the Atitibi Power and Paper Company of Canada and the
Australian Paper Manufacturing, Ltd. use linear programming
methods to minimise paper machine trim losses. It is also
 F-980
3-7-57
-14-

applied to minimizing transportation costs from mills to

customers. Recently I received the following communication
from a friend visiting In England:
There are some very nice applications around —
I have counted a dozen myself, although only two
or three have appeared in the literature. In
general, firms are keeping their activity quiet
for the usual reason that Macy's does not tell
Olmbel's.
Elsewhere he says,
A very lovely application was developed by a
statistician-economist for planning the operations
of one of the divisions of Shell Chemical. It
seems to Increase the planning profits (based on
sales estimates) by about 5£*
There are two interesting fields where linear programming
is just beginning to be used:
1. The first was discussed at the 1936 Managern«r . Science
meeting in Los Angeles by Dr. Pierre Masse, Executive Vice
President of the Nationalized Electric Power Industry of
Prance, and Dr. R. Qibrat. They showed how linear programming
could be used to decide how much should be Invested in new
hydroelectric, steam, tidal, and atomic plants.
2. The second was discussed in the October issue of
Management Science by R. Kalaba and M. Juncosa in their paper
on the application of linear programming to optimal routing of
messages in a communications network and to the problem of
investment in communication facilities. This work began with
the military application but recently research workers in the
Bell System have become Interested. It seems to offer an
 P-980

approach to over-all systems studies which are too complicated

to handle at present by waiting line theory.
For Air Force applications there is an excellent brochure
prepared by the United States Air Force Comptroller that lists
a number of typical examples. From this source I have picked
two:
The first typical example is a transportation problem.
This military example is similar to the Heinz Ketchup example.
Assume that Lockboume Air Force Base at Columbus« Ohio, has
been testing a large item of equipment, weighing a ton, for
the B-47. It is now desired that this equipment be tried at
other bases. Suppose that five each are required by three
Air Force bases and three each are required by the two others.
To supply these twenty-one items, Columbus and Oklahoma City
can ship eight each and Warner-Robins at Macon, Oeorgia, has
five. The items of equipment are to be airlifted to their
destinations. The problem is to determine the routing which
fulfills the requirements and minimizes the total ton-miles.
There are, of course, many possible routings. The mini-
mum solution can be obtained by one of several methods of
linear programming that are no* available.
The second typical problem concerns production versus
requirements. This military example has its counterpart in
many industrial examples.

I.
"The Applloatlon of Linear Frogramaing Techniques to
Air Force Frobleaui - A Mon-Technlcal Discussion," Report
AFANA 3-^-52, prepared by Computation Division, Directorate
of Nanagesent Analysis, 0C8/Coaptroller, Headquarters U. S.
Air Force, December 17« 195*.
1
N-980
3-7-57
-16-

Training schools within the Air Porce are concerned with

the efficient production of graduates. In order to meet a
schedule of future requirements# each school must make
available the required number of trained men at the right
time.
Since each training activity has its own limitations and
needs, an efficient method of production for one will not
necessarily be efficient for another. Some training activities
can readily increase their rates of output while others can do
so only within limits and at substantial cost. Where, for
example, equipment and personnel must be released when
production drops and reinstated as it expands, large fluctuations
in the level of production are undesirable. When faced with
fluctuating requirements, such an activity will ordinarily
overproduce in periods of low requirements, pool the surplus,
and use the excess In periods of high requirements. The
production pattern can thus be made quite stable. However, a
solution may be undesirable if it yields in turn comparatively
large monthly surpluses of trained men.
Problems of this nature illustrate the difficulties that
arise whenever conflicting objectives are inherent in a problem.
Here the desire is to determine a schedule that tries to reduce
both large fluctuations in output and large monthly surpluses
of men. Por these problems, efficient planning means the
determination of a middle ground lying between the two extremes.
 P-980
5-7-57
. /

The optimum production schedule will depend, of course, on the

relative importance assigned to the conflicting objectives.