ORC 66-21

November 1966

00

MVjjJl OBJECTIVE LINEAR PROGRAMMING

by

David Savir

CLEARINGHOUSE
FOR FEDERAL SCIENTIFIC AND
TECHNICAL INFORMATION
Hardoopy { Microfiche
oo .66
1&
mmi mf
C*f-T%£- I

OPERATIONS RESEARCH CENTER

COLLEGE OF ENGINEERING

UNIVERSITY OF C A L I F 0 R N I A - B E R K E L E Y
■

MULT I OBJECTIVE LINEAR PROGRAMMING

by

David Savlr

November 1966 ORC 66-21

This research was supported by the Office of Naval Research under Contract
Nonr-222(83), the National Science Foundation under Grant GP-4593 with the
University of California. Reproduction in whole or in part is permitted for
any purpose of the United States Government.

t ~
 ABSTRACT

Let us consider a linear program with several

objective functions. The traditional approach
has been either to "tradeoff" by weighting each
function, or if a "trade off" vector cannot be
provided to ignore all but the most significant.
We are interested in classes of programs whose
.Tiembers possess some common characteristics.
Examples are sequences of production, refining,
inventory problems over time at one installation.
If sufficient conditions exist an estimate of
a "trade-off" vector can be made. This estimate
Improves over the sequence.

A set X exists which contains the solutions

obtained by optimizing with respect to all
nonnegative combinations of objective functions.
A decision maker is not indifferent to these
solutions but can characterize preferred solutions
A method is presented whereby he can direct a
finite sequence of solutions, (x.), over X
towards a preferred solution. As the estimate
of the "trade-off" vector improves, the expected
length of the sequence (x.) diminishes, and
the efficiency of solution Increases.
 1. Introduct ion

In this paper we will investigate linear programs with more than one

objective function:

0) Objective function z = Q^

subject to Ax = b

x > 0

where 2 is a p-vector

x an n-vector

b an m-vector
n . • _ / 1 k p.'
Q a p x n matrix = (c ,..,c ,..,c )

A an n x m matrix.

We will defi.ie a preferred set S such that only if x is in S will the

objective be satisfied. Under certain conditions a semipositive linear transformation

(x) on z cai be made c>uch that XQx will be minimized and the minimizer x

will be in S . We will define a set Z of weakly undominated z and show

that under arbitrary X

min XQx and z = Qx , 2 s Z

x

will be attained by the same x . Also for any x such that Qx 'S not in
1 2
Z , Qx can be strictly decreased to a point Qx which Is in Z . Hence the

simplex method will always yield points in Z as optimal basic feasible solutions.

Having established that arbitrary X will produce an extreme point x from

a set of extreme points which also contains a preferred extreme point (assuming

hopefully that S contains an extreme point!) we will find a X which will

yield the desired x in S . For this we make certain assumptions, in particular

that X is statistically estimable. If an immediate solution is not found an

' - "^ v.jte.

algorithm is used to search for the preferred extreme point while minimizing

the squared deviation of X from its estimate. .

The assumption that X is estimable is not unrealistic if the approach

taken in this paper is applied to a class of problems with certain similarities,

for example: production planning over sequential periods where decisions are

made at each period.

2. The Presentation of the Problem.

The objectives of many linear programs can be defined adequately only as

vector functions. For example, in addition to minimizing cost, a decision

maker might also wish to minimize execution time of a project; or to maximize

employment, etc., or any combination of these.

Suppose that there exist p objective functions z. ■ c x, k»l,. . . ,p , to

a linear program subject to the constraint x e T where T ■ fx|Ax « b, x > 0, x e R }

We will assume throughout this paper that T is nonempty and bounded. In general

no x e T exists such that z. = c x is an optimum for all k . We construct

the multiple objective vector function z = (z.,. . . ,z. z ) € R . After

we exclude the case just mentioned, it is obvious that z cannot be minimized.

Although other approaches can no doubt be taken, it seems both legitimate and

consistent, since the subject is linear programming, to transform z to a

scalar function which lends itself to minimization.

We define a pair of both physical and mathematical significance: the decision

maker (DM) and the linear program problem solver (LP) . First, mathematically,

DM is a function from RP to R with value Xz transforming the components

of z by semi positive linear combinations to a scalar objective function. LP

is a correspondence from R to T : given a real valued linear objective function

toptimum is used to mean maximum or minimum. In subsequent sections the minimum

problem only is treated.
DM(z) , x" e LP(DM) cT such that DM Is minimized. We call such x* an

optimal solution. If the decision maker is satisfied with some optimal solution

and terminates the problem, the solution is defined to be preferred. If he is

unsatisfied but no feasible optimal solution satisfies him more, then the solution

is almost preferred. Clearly, the latter definition implies that:

Given the existence of one of the solutions,

either a preferred solution or an almost
preferred solution exists, but not both.

We shall prove in Section k that under certain conditions one of these solutions

does indeed exist. A solution which is either preferred or almost preferred

will be abbreviated to a POA^P solution. No implication has been made as to the

uniqueness of any solution, only to the uniqueness of a set of solutions.

Secondly, physically, the problem thus far stated (l) can be decomposed

between DM and LP . We can consider that DM generates by some means a

vector of weights by which he constructs a linear combination of objective

functions. He has then produced a linear program for LP to solve. LP ,

characteristically a computer, can produce an optimal feasible solution by

conventional linear programming techniques.

The concept "satisfied" is unmathematical intuitively and undefined explicitly.

In the light of the assumption DM2 appearing at the end of the section we can
define the following:

(i) DM is sat isfied if he does not specify any vector d , given x .

(ii) DM is unsat isfied if he specifies some vector d , given x .
(iii) DM is unsatisfied but no x' satisfies him more if he specifies some
vector d , given x ; but no x' exists such that d^' ,S) < d(x,S) .

Both the decision maker and the decision maker function will be called DM for
notational simplicity. Similarly, both the linear program problem solver and
its correspondence will be called LP . No ambiguity will occur.

ß w
 (Recall that T is nonempty and bounded.) DM now evaluates the solution.

If he is satisfied he terminates the problem; otherwise he tries a new vector

and sends the data back to LP .

Suppose a common unit could be established in which the value of z. would

be expressed. Then, in the above example, DM could assign #. units of value

to one unit of cost, #- units of value to one unit of execution time, etc.,

and all the components would be transformed to a unique additive measure p,

s
in units of value, p. cy'z , where a is well defined and constant. We shall

require that X be normalized so DM(z) - a'z/jlall » which would then be

completely determined. The problem would then reduce to a conventional linear

program. We will assure that the decision maker cannot establ ish such a unit

of value.

We can now define (l) more explicitly:

(2) Find (x, DM) , x e T , DM e <{2k]>+t such that x e LP(DM) and x POAP .

Equivalently:

(2') min DM = XQx

x
subject to Ax = b
x > 0
x POAP
P
for each X > 0 , I X. = 1
k=l K

where Q = (c1,. . .,ck,. . .,cP)' . t

"For each X" seems to imply that the set of X is at most countable, which is
not true. However in Section 6 we will show that there exist a finite number of
sets A. , and any X e A. will generate the same solution set {x} as any
X' e A- • Thus "for each X" is to be interpreted "for any X from each A-'1
tt,.'<[*.}>" reads "the convex hull of the set of all zk."
We will approach (2') eventually as a two stage problem: firstly, for some X ,

minimizing DM ; secondly, determining the given so'ution's POAP properties,

and revising X if necessary. Call then the first stage, i.e. (2'), neglecting

the x POAP condition: U11).

In Section 3 we will prove that solutions to (2,,) exist under certain

conditions, and that the simplex method will find a subset of these solutions.

The metiiod proposed for finding a solution has little practical significance

for a single problem, since the value cf X needed to obtain a POAP solution

is known only after the POAP solution has been found. (This is clear from

the definitions of preferred and almost preferred.) However, if we have a

class (? of problems possessing certain properties, and the first T-l sequential

problem can be solved more efficiently when LP has available the values

X , t=l,...)T-l , used to obtain the respective POAP solutions, than otherwise.

The required properties are:

th
Let C »_ be
e C? oe the
tne tr sequential problem of form (2') possessing a

POAP solution.

a) Let T £ T for C • Then 0 T »* 0 where t is the index set

t t t. t
{1.2,...} .

b) Let x = x e T for C . Then x e Rn V t and Z x e Rn •

L L L L. . L

c) If x is preferred for some C , and if x is optimal for some

C^i > t'^t, then x is preferred for C , . Equivalently

n ^
c') There exists a domain Sc R such that if optimal x e S it is preferred,

d) z is common to C , V t

e) The distribution of T is such that a maximum likelihood estimate of

X exists. This estimate is the mean of X over t .

ß m
 Now we require some assumptions on LP and on DM . Our recognition of

LP as a linear programming problem solver should be sufficient: we know that

it will read and interpret submitted data without error, that it will produce

an optimal feasible solution with certainty, and that it will yield DM the

results with typographic perfection. As for DM we assume the following:

DM1 (Recognition). DM can always recognize an optimal solutions x e S to be

preferred. He can always recognize an optimal solution x ^ S to be not

preferred.

DM2 (Direction). Given an optimal solution x ^ S , let the distance between x

and the closest point in S be d(x,S) . Then DM can specify a vector d

such that d(x + ed, S) < d(x,S) for arbitrarily small e > 0 . No

knowledge of any point s e S is assumed.

Using the assumptions DM will be able to direct LP towards S and

terminate if LP produces x e S .

3. The Existence of a Solution.

We define a vector z to be weakly undominated if z »^ z in the domain

of z . This definition will enable us to replace the scalar "minimize" by the

vector "find weakly undominated" in the vector optimization problem:

(3) Find weakly undominated z = Qx

subject to Ax = b j
! or x e T
x > 0)

where symbols are defined in (l)

"z £ z0" reads "it is not true that z. < z.0 , k=l ,p ."
 We can P.TW find a set of x which will solve (3).

Let z be any vector z such that z. < z. , for any z in

the range of x . Then z is weakly undominated. Thus we can produce

p such weakly undominated vectors, each of which is minimal in one component.

How can we produce the inverse images of such z. ? Consider one such Problem-

en) min z = Qx , x e T
w.r.t. k

Clearly z, , , k' ^ k , can adopt any value and is independent of the minimizing

process. Hence (4) is equivalent to

(5) min z. = c x , x e T

which is a linear program. Hence by solving p linear programs of form ^5) we

can find p basic feasible solutions to (3). each of which has the minimal
/ \ k*
property in a different component. Call a feasible solution to (5) x

("Solution" is hrre used according to the linear programming convention to

mean a value of x at which the objective is minimized. No ambiguity results.

k" k*
Al' solutions are assumed feasible.) Let X = {x } . From wel 1-knovsn results
!<*
X is convex.

Now to dispose of the simplest case:

** k*
Theorem 1 (3) has a minimum solution x if and only if fl X ^ U 1 and any

point in the nonempty intersection is such a solution.

Proof Suppose x is a minimum solution to (3). Then z = Qx is

not only weakly undominated but also a minimum. Hence z < z , x e T .

It follows that z ' ' <i 2. , k=l p . so x' solves (5) for
,'—1. KJL

each of the components, and x = x , k=l p . Consequently

x,W e Xk'" for all k , and x"' e 0 Xk,% . Suppose HX " ^ 0 . Let
k

f m
8

x0 e (IX . Then x solves (5) for each of the components,

Qx0 ■ z0 < z , x e T , z0 is a minimum vector and x solves (3)

to a minimum, x was chosen arbitrarily from HA , so any

x0' e OX1** solves (3). ///

Now a minimum solution to (3) seems the best we can hope for, and indeed

will solve (2') provided we can show that it is POAP . Theorem 3 will give us

that answer. Meanwhile let us consider the case where QX = £) . We have

shown that there is no minimum solution to (3).

k*
Now we have shown that a set UX of solutions to (3) exists, and we

would like to discover, if possible, a maximal such set from which to select

some preferred or almost preferred x . Also, since we have the simplex method

in mind, we would like to know if we can move from x to a better x' .

We will show that this maximal set X exists and x e X is a solution

to min X Ox , x e T , for soc^e X > 0 . We would like also to be able to

x

specify X and thus generate a specific solution. If the rows of Q are not

positively linearly independent we can find some X such that X Ox = 0 for

all x e T ! Legislating this trivial solution out of existence we require:

k k
Positive linear independence of c , i.e. Z X|C =0,Xk>0=»X"0 .
k

We can now construct a closed convex cone Y « { <k(i,> , k > 0 , a scalar } ,

k * * . * X
the minimal cone containing c . Let X = fx j x minimizes c x ■ z , x c T ,

for some c e Y ] . Clearly x e X for all k . Also for any c e Y

there exists the corresponding x by the compactness of T .

Analogously to linear programming we can say that x' is better than x if

z' = Qx' < Ox = z .

 :
Theorem 2 If there exists x ^ X' , x e T , and z = Qx ; where Q

is a matrix of p positively linearly independent rows, the k

. . k , . 0 00 /. 0 00 »
row being c , there exists z < z (i.e. z. < z. , V k)

such that x e X',

00 *'-
Proof x ^ X .
00 k 00 , ,
zk = c x , k = 1,. . . ,p .

(6) Referring to (5): z^0 > zj"' else x00 e X^ c x" .

k 00 00
u i if .
p hyperplanes c x = z, pass thiough x . Hence there
, . , ,. k 00 . ,
exist p closed hall-spaces c x < z. with nonempty intersection

d? , and, by the positive linear independence of cx and the

compa ctness of T , «Jc. has an interior. (Intc^) 0 7^0 by (6).

It is necessary and sufficient to show that there exists x e

(intc^) P T satisfying x e X since x e (intcif) H T =* z < z

Select a point x in (intt>Ä) 0 T . Defined = fx|c x < z.

Vk } . Then-Jt H T c (intu^) 0 T . Construct a sequence

x , x ,. .x ' such that (intc^) 0 T ^J? 0 13. . . zxxl1 fl T 3

• • 'vL , T are closed, so**- 0 T is closed. rJL P T is a

decreasing monotone sequence of sets so 1 imc>C. p T = 0 (<xl. 0 T) =

1
0 0 , '"" o^
x. . x, lies on a supporting hyperplane, since x. lies on the

boundary of T , for if not a smaller set could be found containing

points o^ "oi and T . The boundary of T Is a set of hyperplanes.

0
So x. must lie on at least one.

Construct p - 1 other different sequences yielding x. (k~l,..,p)

arbitrarily close and lying on the same hyperplane. Elther all x.

are different or 1 imcA. P T is unique over all sequences. If

i-«aD
O = O = O OOj. o,_oA o
x. *9 • • -x, . . .x = x , for any x +AX€T,z +Az>z

Hence at x , z attains a local minimum. By the linearity of the

* m
 10

space and every objective function, a local minimum is global. Hence

x0 e Xk* (for all k) c X" .

If all x. are different we will find a such that

i o
a- z. = a constant

cy'Z - z , a constant vector

where Z = (z.1 ,. . . ,r 0)
P
.,,0/0 o o%
Let Zk - (zk , zk ,. . .,zk ) .

Then Q," (Z0 - Z^) = 0 .

Transposing, and by Parkas1 Lemma

(Z -Zk)y-0,0.y<0 has no solution

therefore (Z -Z, )la-0,astQ has a solution

0 - a* (Z0 - Zk0) = a' (OX0 - QXk0)

0
r[where
u / 0 , x. 0
Xv = (x. x
0
\)

X k 0 = (x k 0 * x k 0 » • • • »Ä
xu 0)1
'J

- a'QCx0 - xk0) .

But (X " X. ) is a (p - l) - tuple of vectors in one supporting

hyperplane, so v'CX - X. ) = 0 is the equation of the hyperplane,

v »* 0 . Therefore Q'o = v ^ 0 , and & t 0 . a ^ 0 t a 4 0 * a > 0 .

0 - 0
Now o'Z = z implies that at each x. is attained the value

z ■ Q-'QX , a constant on the boundary hyperplane; therefore at

x. is attained a local minimum. By the previous argument It is

k
also global. a'Q. 'S a semipositive linear combination of c , hence
0 ■>'.-
c/Q e V . Consequently x. e X . ///

from the conditions of Theorem 2 falls out that

Corollary 1 X" is the set of all solutions to (2M)

Proof X = {x |x minimizes c x = z , for any c e Y , x e T}

""'"' A.
 11

By the convexity of Y , cX = r X.ck where ck is an extreme ray of

k
0 k
Y and >-|< > . ^ > 0 . But c are the independent rows of Q ,
P
X k k
so c = Z X.c where X. , = 0 if c is dependent. So x"
K K
k=1

minimizes XQx , X > 0 , x" also minimizes aXQx where a is a

scalar > 0 . Let a = , and x" is a solution to (2") .

I X.
k=1 K

Hence X" is the set of all solutions to (2'') . ///

We have now found X" , the set of all solutions to (2' ') . We show that

(2'') then is equ:valent to (3) and a search for a weakly undominated z replaced

by a simple linear program. Let Z be the set of all weakly undominated z .

Corollary 2 x e X" if and only if z e Z .

Proof Suppose x e X" . Then either

(i) at least one component of z is minima'. Then z is weakly

undominated; or

(ii) cy'z is minimal for some o > 0 . Suppose there exists some

z < z . Then ty'?. < ^'z which contradicts the hypothesis.

00
Suppose z e Z . Then z < z ^ Z and x e X by Theorem 2. ///

Thus far we have produced a set of solutions X" , each element of which is

equally satisfactory. The choice will be narrowed by restricting (or attempting

to restrict) x to a preferred set S . Now that we know that the search for

x e S can be done by linear programming techniques we must assume:

LP1 (Finiteness). Let x e X " be an optimal solution. Then x e X" , another

optimal solution, can be produced in a finite number of steps.

k. X' and the Preferred Set

In Section 2 we have assumed the existence of a preferred set S . We also

f w
 12

constructed an interaction between DM and LP , whereby LP proposes optimal

solutions to a linear program, and DM accepts them or tries to steer them

towards S (assumptions DM1 , DM2 ). In Section 3 we showed that the set of

points that LP can propose is indeed that set of points from which DM wishes

to select a preferred point. Now what about S ?

In this paper we choose not to make any assumptions on S . Such assumptions

would involve economic concepts such as insatiability of preferences, convexity

of preferences, etc., which are not appropriate here. However, it is important

to grasp that c.1 though LP has a linear objective, to find some x e X , DM

tias no such thing. Now suppose that we have a set X obtained from solving

(2'') or from solving (3), we will show that a complete solution to (2') exists.

Before proving the next theorem we require a lemma:

Lemma X* » <X"> 0 bdry T

Comment We proved in Theorem 2 that X c bdry T . We now require that

JL,

X be simply connected over the boundary i.e. that there be no

holes. In Theorem 3 we will have sequences of solutions starting

in X and ending In X . These sequences must stay in X

Proof Let x e bdry T . Then

(A) x ■ E y..x{ , where fi} is the index set of all extreme points

of T contained in the supporting hyperplane containing x , and

p. > 0 , E p,. = 1 .

Let x e < X > . Then

(B) x = r X. x. . where {j} 'S the index set of all extreme points

of T in X' , and X. > 0 , I x. = 1

But the extreme point representation of any x e T is unique, and

if x0 e < X > (1 bdry T then if |i. > 0 . X. = pi. . Similarly

 13

if ^b > ^ » ^L- ~ M-b • Thus the extreme points of (A) are the
%
same as those of (B) and x = Z^,. x." . But all x. , i e fi}
t i

minimize 0/ Qx for some & . Hence x also minimizes o- Qx ,

o * o " o * o
and x e X . Suppose x e X . Then x e <X > and x e bdry T .

So x e <X > 0 bdry T and X" = <x"> 0 bdry T . ///

Theorem 3 will show under what conditions a preferred solution can be found.

The effects of pathological circumstances, those in which the linear program

and the decision maker are opposed, will become clear.

The lemma has shown that X is a simply connected part of the boundary

of T , intuitively a partly open convex set without an interior. |f then we

required conditions for a preferred solution to be produced, it would seem

immediate that would suffice firstly: convexity of S , and secondly:

X 0 S ^ 0 . However Figure 1 shows a case where these conditions are insufficient

Using the decision maker's axioms, and starting from the left of the Figure, almost

preferred solutions are encountered twice, at x. and at x_ before a preferred

solution is reached at x_ . Clearly the choice between x. , x. , or x_ is

determined by the original trial solution and the direction of search. The

situation is pathological because we do not expect a decision maker to be satisfied

with a solution that is nonoptiiiaK

Fig. 1

* ■m
 1^

Even if the diagram is less pathological (Fig.2) there is still and almost

preferred solution at x. . The third condition on S which will guarantee

a preferred solucion is that the asymptotic cone of S, A (S) , contains Y

Fig. 2,

This implies insatiability of preferences in the direction of every possible

optimizing vector (Fig.3). That it is insufficient merely to have nonempty

intersection of the cones is shown in Fig. k , where there is an almost preferred

solution at xc . It is evident that these three conditions are not necessary

for a preferred solution - merely sufficient.

Fig. 3
 15

Fig. ^4.

Theorem 3 Let X be the set of solutions to (3). Let x e X" be

any such solution. Then

(I) If S is convex, X' OS ^ 0 and Y c A(S) a preferred x

can be produced,

(ii) If X OS = 0 an almost preferred x can be produced,

(iii) Otherwise, either a preferred or an almost preferred x can

be produced.

Proof (i) Suppose x ^ S . The vector d can be specified by DM2 ,

and k > 0 is a scalar. Let x + k d e S be the closest

point of S to x . If x + k d ^ X let x be the

* O OO OiOO*
closest point ofX tox+kd . Ifx+kdgX

let x = x + k d . Suppose x ^ S . By the convexity

of S there exists a unique point x + k d e S which

1 .1 .1 o .0
is the closest point of § to x . Ifx + kd^x + kd

a hyperplane separating X from S could be defined, for

by the lemma X is a simply connected subset of a convex

body either wholly contained in S or closed with respect

to S since Y C A(S) . Hence the points x + k d are

16

not in T . Also dCx1 + kV, x1) <d(x1, x0 + kV)

< d(x + k d , x ) . Consequently a sequence of pairs

• ■ 1 * 1 ' 1

(x , x + k d ) for i > 1 can be defined. Suppose

I * * 1
x ^ S . Then the sequence is countable and d(x , x +
* 1 * 1 *
k d ) is strictly decreasing. Hence 1 im x e §
l-co

and some x arbitrarily close to the limit is in X OS.

Let x=x ifxeS.orx otherwise. Then by

LP1 x can be produced in a finite number of iterations,

and by DM1 it is recognized as preferred.

(ii) Let x + k d be defined as in case (i). Construct a

. •i / i ' -1 ■ i-1 . i-1\
simrlar sequence (x , x + k d ). Since there is
it
no convexity assumption on S and since X OS * 0 ,
k
„. . . i^M k+1 , . k+1 .k+1 _..
there may be some x+kd-x +k d .This

implies that x = x -...= x =..., i > 1 . Let

ko k k+1 k+i ko ,
■ x
1JF
x =x =...= x =.... If x does not
. ^L J/ ' i+l i ' -1 J '-K . .i
exist then d(x , x + k d ; is a sequence strictly

decreasing to some limit«^ > 0 . As d approaches^*-, x

approaches some limit X since X is bounded. Let

x = x or x whichever is defined. Then by LP1 x

can be produced in a finite number of steps, and by DM1

it is recognized to be not preferred.

(iii) This case follows case (ii) except that either

a) x e S or

b) <A = o may occur since X CIS ^ 0 . Referring to

case (i) , if either a) or b) occurs a preferred

solution is the outcome; otherwise, as in case (ii)

the solution is almost preferred. ///

 17

The converse to Theorem 3 is trivial:

•X.

Corollary 3 If a solution to (2') exists then it is some x e X" .

Before attacking the method of solution proposed in Section 5, the reader

is cautioned to recall that any x e X Is a solution to (2'') , given the

appropriate X . However, the simplex method produces only extreme point

solutions^ If S contains an extreme point LP will find it, but if S contains

no ex reme point and S P X" / 0 the simplex method will have LP oscillating

across S indefinitely. Clearly DM must take some linear combination of

extreme points himself in this case.

5. The Method of Solution

If we accept the assumptions and conditions on T , S , Q , DM , and LP

so far discussed we have shown that a solution x POAP can Indeed be found.

Without loss of generality we can restrict ourselves to the case in which S

contains an extreme point of T . If S does not contain an extreme point DM

can terminate when by DM2 he must reverse direction; and If an almost preferred

solution is produced DM can terminate when LP can propose no solution In

the direction specified.

By solving (2') we are trying to define an optimum in terms of the decision

maker's preference, and then to find a function which will yield this optimum,

rather than to find an optimizing function and insist that the decision-maker

like It. We are granting him intuition and knowledge that he has withheld from

the computer, unwittingly or otherwise—after all. It was this intuition and

knowledge that made him a decision maker In the first place.

Why then bother with X at all? A very simple approach could be select

' " —- «to.

 18

1 2
any extreme point x , rejectin g it and selecting X and some direction
1 1 2 1
of improvement d such that d (x - x) > 0 , and continuing until no
n+l
X
such that n
can be found. Then X is clearly optimal

by the convexity of T •

If we follow this approach, though, we find that the solution to the tth

sequential problem is a lengthy in computation and as tedious and demanding

on the decision maker as the first problem , even thou gh the solut ions may be

very close. For example ; in January a manu f acturer or automobiles decides,

after investigating many extre me point solu t ion, that i t is preferred to

produce 700 Rolls Royces and 400 Bentleys. Now in Fe brua r y he notices that

management policy remains the sa me a nd that all his unit costs and unit profits

are in the same proportion. Consequently , he decides that rather than reinvestigating
1
all the extreme points anew he wi ll again select X = (700,400) . However,
1
he d i scovers very qu ickly t hat either X i s infeasible for February, or t hat
2
it is not optima 1 : he could fi nd some better X which is preferred. He

realizes that conditions (a) throu gh (e) of Section 2 hold. He sees that the
0
January problem could have been solved by (2 1 ) for some "A. , and if he knew

Ao , (2 1 ) us i ng Ao a nd T for February would offer him a so l ution close to

x 1 that would be both feas i bl e and opti n~l. Further reflection, and an
0
unde r standing of linear programmi ng , tells him that in general "A. is a point

in a closed convex set A and a ny A€A would have solved the January problem.

Also, ca I I i ng A in January AI , there must be some A2 for February, and it

i s i med ia t e t ha t AI and A2 are not necessarily equal. It would be folly

to select arb i trar ily AI from and to hope that AI would also be in
AI A2
But now he recal l s condition (e) wh ich tells him that there is some ~ which

is mos t 1 i kely to be in the intersection of many of the At , and that ~

 19

is the mean of all successful At to date.

We can find a way to estimate ~ Suppose we are solving the first

1
sequential problem. We start with Al , an arbitrary estimate of A , and
1 1
produce an x which minimizes A.!Qx Suppose x is not POAP. We modify
2 n
to and produce a n X which min imizes unt i 1 x 1 which
n
1
minimizes A Qx is POAP When we solve the second problem clearly the
1
nl
best value of with whi c h to begi n is Al , since all the sequential
problems are of arne class. If the n the solution to the second problem is
n
attained using , we will beg i n the third problem using A~ = HAl 1

1
t: n.
Similarly the t+lth problem would be entered using At+l = ~ L: )
l:o: A.
t.1= 1 t
(If we were to waive the as s umption of time invariance of s the mean J:
I
might be a weighted mean.) As t becomes large ~t approaches \ in Cesaro

limit, and ~ is the maximum 1 ikelihood estimate of A which will minimize

A.Qx and produce a POAP x •

th 1
Consider the t+l sequential problem. The first solution offered X

minimizes AtQx over Henceforth the subscript on t will be dropped

Tt+l •
1
whe re no a mbi guity can result. If X is preferred the problem is s olved and
t +2th 1 2
remains unchanged for the problem. If X is not preferred, X such
1 2 1
that d (x - x ) > 0 is to be found .

i+l
6. The Generation of X

i ~ i .
The problem to which X is the solution, using 1\ , IS

min z
Subject to Ax =b
X ::::_ 0 ,
A.iOx =z
 20

Rearranging columns so that the first m columns are the basis B. we have

C.

_z.

where subscript B corresponds to basic variables and subscript C corresponds

to nonbasic variables. The solution x is found by premult iply ing both sides by

B. 0
i
yielding
1
-X'k B^ 1
i

r "I
(7) I B
i C
i v I
B. b
i

0 XY -xV B^C z-X'Qg B^b

i i
L 'J i

x' = (B^b , 0) .

Now suppose that DM specifies direction vector d = (dD , dr ) . That is to

. . . I I
say that d(x -x ) > 0 can be considered an additional constraint to the

problem. Of course, If we were to try to resolve the problem using this constraint

we would succeed merely in shifting x an infinitesimal distance In the required

direction; but we will use the constraint to determine which variables can enter

the bas is (B.^.)

x for x
i+l

There is no further need to maintain the iteration affix (i) everywhere,

since henceforth the analysis will be based on solution x . We rewrite (7):

(7') I B-'C X
B
V'b 1
B

0 x^-Qg e-'c). -xc.

_z-)iQBB''bJ
 21

I+1 e
Theorem h If a solution x exists x will enter the basis (B- + i) ft>r
1+1 . e -1 e
x only If d- -dB C > 0 , whert e denotes a column of C .

Proof Rewrite d(x -x ) > 0

as , i+1 > dx
dx , i .

• i

we have x = (B- b,0); yielding

dx - s = dRB b + e (e>n) , where s is a nonnegative slack variable,

Add the constraint into (7') , yielding

(8) I B^C 0 x„n

B b

k d
c -1 dpB'Ve

|o XCCL-C^B^C) 0 >- -j _i Lz-XQ^B^bJ

^

Reducing (8,1 to canonical form by premul t ipl y ing both sides by

-d, 0 we have

(9) B C o" X B b
" B1
c
o W 1 x
c
0 x(Qc- QßB c) 0_
oj _s J Lz-XQßB'^J

The m+1 row of the system (second row of the matrix (9)) is infeasible since

the value of the nonnegative slack s is -e , (e>0) . To restore feasibility

Ke
x must enter the basis with positive value; hence d- - dB C > ///
22

e -1 e i+l
Corollary k If d- - d-B C < 0 for all e=m+l,...,n then no feasible x

can be found that will minimize (2 ') subject to d(x -x ) > 0 .

So corollary k can be the criterion by which the decision maker determines

x to be almost preferred and terminates.

e
Now suppose that some x can be found to enter B_, . As e increases
i+l
Q
so does x until some x ,tA = l,...,m , is reduced to zero and is driven out

of the basis.

x is the solution to (2'') using a given X . Since we are going to modify

X to produce our next solution x , we must have some properties of X .

P
Let A = {X|X.>0 , Z X.=l}
K k
k=l

Let X be some value of X such that x minimizes X Ox i x eX

V ^L V p,

Let A = fXlx minimizes XQx , x eX ] ; X eA

Theorem 5 The number of distinct A isfinite. A is closed and convex.

^ v v
A = UA • No A is disjoint.
v v
v

Proof Suppose x is an extreme point of T . The number of extreme

points is finite. Suppose x is not an extreme point, i.e.,

x is a linear combination of extreme points. From the proof of

Theorem 2, given a linear combination of extreme points x

U
supporting T , (DB * . ß >0 , IB =0 . the set of a as defined

in Theorem 2 is independent of the positive weights (ß ) in the

combination. The number of combinations of x with positive ß

is finite. X= ]i^i[ yields the same minimizer as a • Suppose

 23

X.Qx
1 is minimized by x and X.Qx is also minimized by x .
H 2 V
Let X1 and >.2 eA^ , X^Q = c1 , ^2^=c2 ' 't 's a
well-known
result that [a- c. + (l-ry) c_ ]x is also minimized by x , where

0 < Q/ < 1 . Hence [Q-X, + (l-o')X0]Qx is minimized by x and

- - 12 ^
A^ is convex. Suppose the sequence x.Qx , ^2^ ^^

is minimized by x , where X, , X- ,... .X, , ■ - .-^X ; X, , X»,...

|i I 2 t 12

X..«.-eA • Then c.x , c0x,... ,c^x,... is minimized by x

t v \ I t [J, .

If c, , c„,...,c »...-»c t^en ex is minimized by x since

I 2 t \x
both T and z = ex are closed. Hence A is closed.
v

From the definitions A ' A • Suppose there exists some

v
X e An(UA ) • Then no x is minimized by X Qx . which
V Ll
V ^

contradicts the compactness of T . Suppose some A is disjoint.

Then A ~ A is not closed; but the union of closed sets is closed.///

v

Suppose that X is an element of A . We must find some Xe/V which

will generate x subject to the conditions of Theorem ^4. We know that

X is the most likely value to use to achieve our aim, but we also know that

X will not work because we have already tried it. Consequently, we try some

XeAV 1 such that i1 1 X-X|I

1 1
is minimized, considering only those X which

generate x subject to the constraints of Theorem k.

Let us formalize the conditions on the new X • Let E be the set of

columns of C satisfying the conditions of Theorem k. For notational simplicity

we drop the superscript of X ; where no affix exists we refer only to the new

X that we are seeking. Suppose that we can generate x using XeA 1 • Then

the canonical set of equations is still (7'), although the value of X is not

identical. Let the value of X used to generate x originally be in A

V

What can we say from this? Firstly, the solution is x since x=(B b,0)

ß m
2U

Secondly, since x' is optimal XC^-Q^B" C)>0 . Thirdly, XeA . But from

Theorem 5 we know that there exists some XeA HA , . Consequently, this

v v
choice of X from A HA i is possible and justified. We can say then that

(10) XC^-OgB"^)^

x>o
p
E >.. = 1
k
k=l

and also (a) that there is at least one equality in the first row of (10),

and further (b) that equality corresponds to a column in E . In addition we

have an objective -- to minimize E(X, -^iJ* • So we have something that looks

like a quadratic program. From the optimal ity properties of x It is clear

that there is a one to one correspondence between the columns of C and the rows

of (10). It follows that if column j is in E then row j of the first

1ine of (10) Is eligible for equal ity. For simplicity of notat ion we will

write that row j Is In E .

Let us simplify (10) into a quadratic program by removing the additional

constraints, (a) and (b):

kV
(11) min TA\

subject to RX<0

x>o

ö.k=i

where R = -(QC-(LB'1C)

By inspection the optimal solution is >.=X , the point from which we started.
p
The constraint set has an interesting geometrical interpretation. In R all
 25

solutions X 1 ie on the intersection of the hyperplane EX.=1 and the nonnegative

orthant.. A solution to (10) is found when a point X lies on one of the

admissible feasible RX=0 (row e E) while remaining as close to X as possible,

if no such point can be found then case (ii) of Theorem 3 holds. If X lies

on the axis of the orthant a solution (in general) does not exist there since

n
A Av, i 's not
necessary at the boundary of A , hence a change of basis in (?)

does not occur.

We set up the system of equations to solve (ll) by Wolfe's method of

quadratic programming, addina slack variables c>p , multipliers TT and

complementary slack variables v , complementary to (X,a) :

p-
- — mm «4

(12) R iI 0 0 0 0 X 0 i

1 0 0 0 0 0 a = 1

-21 0 R' -i' i 0 TT -2X

0 0
lo 1 r 0 0 I V . J
The matrix contains 2(n-m)+p+l rows and 3(n-m)+2p+l columns. We have

already established that there exists a solution in nonnegative X , a ,

unrestricted TT and zero v which is optimal both a fortiori and by the Wolfe

optimality criterion of (X,a),v= 0 . Hence by Gauss-Jordan elimination or

simplex phase 1 we make the first n-m+k columns basic, and immediately obtain

the tableau corresponding to the solution \=\

Y 1 (s+1 Y
l ,s+n-m+p
1 ?!
B2
(13) •J
Y s,s+l 's.s+n-nH-p
S
26

where s*2(n-m)+p+l

B are nonnegative values of \

B are nonnegative values of a

3
B are unrestricted values of TT

7. The Computation of X

We recall in (11) that it is Sufficient for one inequality to be an

equation, provided its row e E . to yield a solution for X . Suppose row

jeE , nnd the j inequality of (10) is an equation. Then Q. = 0 and

v +
_L. enters the basis. From the last row of (12):
P J

TT. = V , .
J P+J

a. and v +,. cannot simultaneously be in the basis hence the pivot operation
J P J
entering
y v and expelling3 a must be on vY +. • . .• • From Kuhn-Tucker theory
P J.s+p+j
we obtain that the multipliers rr. are nonpositive, hence v .. < 0 (since
J P+J -
the solution will be nonoptimal with respect to the Wolfe algorithm). Since

a.>0 before the pivot and v +. . <0 after the pivot

J P J -

vY . . < 0 .
p+j ,s+p+j

From the arithmetic of pivot operations, referring to (13)

Y,
Y
p+j ,s+p+j

! .B^.. - Vj'.s^j B2 where

)+ +
J ? J Yn+; c+n+: P+J J J

2
2 ^ B. p-Kj
B
P+J Y
p+j.s+p+j
 27

To maintain feasibility 0 < B. < 1

2
and 0 < B
- k

We have thus found the conditions by which -. may leave the basis

The steps are:

1) Identify jeE

2) El imi na*-« j » y
p+j ,s+p+j > 0

3) El iminate j » B < 0 or B > 1

2
k) Eliminate j > B. . < 0 , Ji V j
J

Those that survive the test are feasible and we wish to minimize the

deviation from X . It
is sufficient to minimize EJX. -X.|1 since the expressi on
k k
2
is monotone with respect to E(X. -A )

We have found values of X , if they exist, which will make a new basis

in (ll) as desired as the current basis. In order to precipitate the change we

replace X by X+e(X-X ) where e is small and positive. Clearly e(X-X )

sums to zero. Now suppose another iteration is required. X computed on

i+1
the first iteration becomes X and a new value of X is desired. R will

of course be different since it is derived from the x , We follow the exact

procedure as in the first iteration, forcing a basic solution in the first

n-m+p columns. This will not be optimal since X is excluded from the constraint

set. Consequently there will be at least one value of a < 0 . Again using

Theorem ^4 we find that value of v , if any, which satisfies the conditions for

p ivot i ng.

8. Termi nat ion

There are three ways of terminating the procedure.

28

Case I. The decision maker accepts x produced by minimizing X Qx • Then

xt+, Mtxt + xJ+,)/(t+i)

Case M. The program can find no v to bring into the basis which satisfies

the pivoting rules. Then case ii of Theorem 3 applies and the last x is the

best that can be provided.

=
Vi (txt-x;+1)/(t+i)

as in case I.

Case 111. The decision maker causes cycling between a number > 2 of solutions.

This implies that S does not contain an extreme point and the decision maker

must interpolate points of the cycle I .

Xt+1 = (tx"t + EXJ+1 / ||?||)/(t+l)

9- Numerical Example see figure 5«

A very simple example is illustrated:

A constraint set T cC c(?

-xl + 2x
2 £ ^

x + l/2x < 6

2x] - x2 < 4

yielding a minimum at (^.M under

c x = -x1

2
ex- -X-

X1 = X = (1/2,1/2)

The t+1 sequential problem contains T . c C

 29

+ 2x
"xl 2 -Li

x1 + 2x2 < 6

2x x
l ~ 2 £ ^

The optimal solution is

1 0 0 1/5 2/5 14/5

0 1 0 2/5 -1/5 8/5

0 0 1 -3/5 V5 18/5

0 0 0 3/10 1/10 .n/5.

producing x = (1V5, 8/5) which is not in S .We wish to increase at least x,

and write c!,: = x. > 8/5 .

e -I e
We compute d- - dRB C for x, and x

d B } k
d^ - B ' t = 0 - [0 1 0] "1/5

2/5 -2/5
.-3/5.

d5 - dgB^C5 = 0 - [0 1 0] 2/5

-1/5 1/5

V5J

Only x is permitted to enter the basis.

-1 c
We calculate Qß8 - ^

1 0 0 ' 1/5 2/5- 0 0

0 -1 0 2/5 -1/5 0 0

.-3/5 V5.

-1/5 -2/5

-2/5 1/5

ß m
 30

We add slacks (J > 0 yielding ( 11 )

min (A 1 -~ )2 + (A 2 -~) 2

[-1/5
[~]
subject to

]
-215 Al
-215 115 0 A2 =

1 0 (J l

(J2

We have seen that only x5 can enter the next basis. x5 is the second
column of c ' so ~ = (2} • We reach the tableau (13); showing y .. and the
IJ

RHS:

-114 114 1120 -3120 A 112

114 - 114 -112 0 3120 (J 112
=
1120 -1 12u -11100 31100 TT 3110

-3120 3120 31100 -91100 v 1110

0 0 -1 0 0
0 0 0 -1 0
1 1
.1 -2 -2 -311 0 -1110 0

p=2 , j=2 so we inspect = -91100

So far a pivot is feasible.

=
-3I 2 o]·( 111o \ = [ : ] + ~
1 -312J] = [1 13]
[
3120 -911oo 1 2 [
3/20 213

s2 = 3110- (31100) ( -1019) = 113

2
So A = (113 - £16 , 213 + £16)
We re-enter the primal problem with objective min ((-113 + el6x 1+(-213-el6)x 2 )

and obtain x = (l,512)eS.

 31

dT

T
t+l

Fig. 5.
32

REFERENCES

[1] Dantaig, G. B., Linear Programming and Extensions, Princeton University

Press (1963).

[2] Hadley, G., Nonlinear and Dynamic Programming, Addison-Wesley (1964).

 Unclass ified
Security Classification
DOCUMENT CONTROL DATA • R&D
(Saajnly cfaiii/icadon ot Ulla, body of abstract and indexing annotation muat be antarad whan tha ovarmll report la claaailiad)

I ORIGINATING ACTIVITY (Corporata author) 2a REPORT StCU«l TY CLASSIFICATION

Unclass i fied
University of California, Berkeley 2b CROUP

3 REPORT TITLE

MULTI OBJECTIVE LINEAR PROGRAMMING

4 DESCRIPTIVE NOTES (Tvoa ol raporl and mcluaiva dataa)

Research Report
S A UTMORfS.) Ci-aaf name first nam« initial)

Savir, David

6 REPORT DATE 7a TOTAL NO OF PACES 7b NO OF REFS

November 1966 36
8a CONTRACT OR GRANT NO 9a ORIGINATOR'S RFPORT NUMBERCSJ

Nonr-222(83)
b PROJBC T NO
ORC 66-21
NR 0^7 033
c 9b OTHER REPORT NOC5; (A ny othat numbara iiat may ba a$*l0^ta
thla raporl)
Research Project No. RR 003-07-01

10 AVAILABILITY/LIMITATION NOTICES

Available without limitation on dissemination

11 SUPPLEMENTARY NOTES 12 SPONSORING MILITARY ACTIVITY

Also supported by the National Mathemat'cal Science Division

Science Foundation under Grant
GP-4593
13 ABSTRACT

Let us consider a linear program with several objective functions.

The traditional approach has been either to "trade off" by weighting
each function, or if a "trade-off" vector cannot be provided
to ignore all but the most significant. We are interested in
clasi.es of programs whose members possess some cormon characteristics,
Examples are sequences of production, refining, inventory problems
over time at one installation. If sufficient conditions exist
an estimate of a "trade-off" vector can be made. This estimate
improves over the sequence.
A set X exists which contains the solutions obtained by
optimizing with respect to all nonnegative combinations of
objective functions. A decision maker is not indifferent to
these solutions but can characterize preferred solutions. A
method is presented whereby he can direct a finite sequence of
solutions, (x.), over X towards a preferred solution. As the
estimate of the "trade-off" vector improves, the expected length
of the sequence (x .) diminishe>, and the efficiency of solution
increases.

DD I
FORM
JAN «4 1473 Unclass i fied
Security Classification
 Unclass if ted
Security Classification
M LINK A LINK B LINK C
KEY WORDS
ROLE

Multi Objective Functions

Linear Programming
Semi-Pos i t ive Linear Transformation

IMS1KUCTIONS
I OKIGINATING »CTIVITY: Enter th* name -nd address imposed by security r'.ussification, using standard statements
nf the coiitriictor, subcontractor, grantee. Department f De- such as:
(enae activity or other organization (corp-rate author) issuing
(1) "Qualifiri irq'j^steri may obtain copies of this
the report.
report from DDC "
2a. REPORT SECURITY CLASSIMC A': U.N: Enter the c. er-
(2) "Foreign announcement »nd disseiriination of this
all security claasificalion of the repn-,. indicate wheth/r
report t y V'OC is not authorised."
"Reatricted Data" is imluded. Marking is to be in uncord
ance with appropriate security reg Uatiuns. (3) "U. S. Government agencies mny obtain copies of
thiu report dire<-tly from DDC. Other qualified DDC
2b. "IROUP: Automatic downgrading is specified in DuD Di- users shall request through
rective 5200,10 and Armed Forces Industrial Manual, Etilt
the group number Also, when apph able, show that optional
markings have been used for Group 3 and (jroup 4 as luthor- (4) "U. S. military agencies may obtain copies of this
ized.
report directly from DDC Other qualified users
3. RETORT TITLE; Enfei the comple'e report title in all shall request through
< apilal letters. Titles in alt cases ahoiild be unclassified.
If a meaningful title cannot be selected without lussifica-
tiun, show title classification in all capitals in parenthesis (5) "All distribution of this report is controlled Qual-
immediately following the tnle. ified DDC users shall request through
4. DESCRIPTIVE NOTES; If appropriate rnter the type of
report, e.g., interim, progress, summary, an mal. ot final. If the report has been furnished to the Office of fechnical
Give the inclusive dates when a apecifn ri-porting penou is Services, Department of Commerce, for sale to the public, indi-
rovered, cate this fact and enter the price, if known.
5. ALrTHOR(S). Enter the name(s) of auihoKs1 us ^hc.wn on 11. SUPPLEMENTARY NOTES: Use for additional explana-
or in th" report. F.itei 'as« name, first name, midiiie initial. tory notes.
If T.ilitarv, show rank an'l bran« h of servii e. 7 ht- name of
ihe pnnciphl « 'thor is an absolute minirmim t. q'l'rement. 12. SPONSORING MILITARY ACTIVITY: Enter the name of
-
the departmental project office or laboratory sponsoring (pay-
h. REPORT DAT ^ Enter (he \nr ai u. rn« »:' as da, , ing for) the research and development. Include address.
month, ynar, or month, year. If " ^r(. th'.n ■>"? •'ate appears
in the ro^iort, use date <■( pub'iLwtion 13 ABSTRACT: Enter an abstract giving a brief and factual
0 summary of the document indicative of the report, even though
7a. TOTAL NUMBER Of" PA'.R . The tola' pf/e count it may also appear elsewhere in the body of the technical re-
should follow nomia! r.vination , .occlurer, i.e., '-nter the port. If additional space is required, a continuation sheet shall
number of ?*$,' s i vitainmg infcmaliora. be attached
7b. NUMBER OF kt ^•«RFNCFS Enter ihc total nur.bar of It is highly desirable that the abstract of classified reports
'cferpm es cileil ,n ne r-^ ■ t be unclassified. Each paragraph of the abslrsct shsll end with
8u. CONTKAt : ■. ^ ■.'- ANT M'MBF.K J I appropriate, enter I an indication of the military security classification of the in-
Ihe «pplvcable tiu";( - ■: Ihf rontract or grunt under which formation in the paragraph, represented as (TS). (S), (C), or (U)
•he report was wf t'"p. ! There is no limitation on the length of the abstract, How
8b, SL, flt 8d. PROJECT NUMBEK. Enter the appi >pri«te ever, the sugges ed length is from 150 to 225 words.
military department i ientifi "ation, such u» project number. |
14 KEY WORDS: Key words are technically meaningful terms
subproject number, oyster.- number«, task numbe , etc.
or short phrases that characterize a report and may be used as
-
4« ORIGINATOR'S R KP OP ; NUMBER(S): Enter the of fi index entries for cataloging the report. Key words must be
cial report number JV vSu-f; 'he i ^current will he Identified selected so that no security classification is required. Identi-
und rontrolird by Ihe >r ^matins activity. Dus number fiiust , fiers, such at equipment model designation, trade name, military
be unique to this report. I project code name, geographic location, may be used as key
')', >TliEK REPORT NUMBF.RtS; :i the report has been , words but will be followed by an indication of technical con-
text. The assignment of links, roles, and weights is optional.
assi^neri dnv other repor' nunbers ~,thfr bv the originator
or h\ (^e sponsor , els^. enter this nLiniber(g;. .
10, AVAILABILITY LIMJ'TATION NOTICES: Enter any Um- j
itu'.icns on further dissemination of tue report, other than thosej

DD I
FORM
J • N 64 1473 (BACK) Unclass i fled
Security Classification