Professional Documents
Culture Documents
Traditional proceduresfor building modelsconsistessen- to summarize all relevant information arises. In the context
tially of the followingsteps: (1) choosinga goal and de- of environmentalpollution, attribute 1 may reflect the cost
fining an objectivefunction, (2) gatheringrelevant infor- of cleaninga polluted environment,attribute 2 may reflect
mation, (3) buildinga model, (4) validating and operating the degreeof discomfort,attribute 3 may refer to the cost
the model, (5) determininga feasiblecontrol policy, (6) of damagedone,and so forth. In the context of a public
applying the policy to. the system,and (7) reachingthe investmentproblem,Y, may representregionalbenefitsfor
stated goal. Implementationof all these steps is re.oreor i = 1 and national benefits for i = 2, or J, may simply
lessstraightforv•ardin a majority of simplecases,since representthe time stream of benefitsfor the first year
they often involvesatisfyingoneobjectivefunctionto reach (i = 1), secondyear (i = 2), and so forth. In the casein
the stated goal by usingone informationset and one con- whichJ, representselementsof a time stream,there exist
trol policy. However,a goodmany problemsdefy analysis well-definedproceduresfor reducingthe vector to a scalar
usinga single-objectivefunction.For example,in the design, quantity [Johnsen,1968]. Indeedthe presentvalue theory
operation,and planning of many complexsystemsa par- of economics dealspreciselywith this topic. In the general
ticularly vexing methodologicaldifficulty arises in the case,however,the procedureto combinethe various
context of choosingthe 'best' o.f severalalternative actions into a singleJ is by no meanstrivial.
when each action will result in one of several possibleout- Considera vectoroptimizationproblemsuchas 'minimize
comesor consequences [Raij•a, 1968]. For instance,in J,' whereJ is givenby
planninga largeriver valley projectan administrator
might
be concerned with the time stream of benefits and costs or(x)-- [J•(m),,1•(•), .--, or•(m)]
• (1)
and the incidence of these benefits and costs to different
wherethe decision variablex itselfis generallya vectorand
sectors.Quite often, someof thesebenefitsmay themselves
the superscriptT standsfor the transpose.
be incommensurable entities.Systematictreatment of these Method 1. An extremelysimpleminded, thoughnot
incommensurables is by no meansa trivial exercise.How-
gant,approachis to attemptto minimizeoneof the eora-
ever, if one considersthe casein which each consequence is
ponentsof J, say J,, but to allow all other components
describedin terms of a single value or a set of numerical
vary withincertainboundsor tolerances/•f.or j = 1 to
values, then the problem usually boris down to what is
and j .=/=i. Usingvariousvaluesof /• for all j on a trial
customarilyreferred to as a vector-valuedor multiple-
and error basis,the designergainssomeknowledge of the
objective'optimization problem [Johnsen, 1968; Major, trade offs involved. This method suffers from two major
1969; Geo#rion, 1970].
drawbacks.First, the analyst has to choosethe tolerance
SCALARIZATIOl•IPROCESS levels• purelyon a heuristicbasisand may very well
choosea combinationof /• for which no solutionexists.
A classicalmethodof treatingmultiple-objectiveproblems
Second,the methodplacesundueimportanceon the un-
has been to reduce them to an equivalent scalar-valued
boundedindex J, that is minimized,one index thus being
objectiveproblem,usingone techniqueor other, and apply
arbitrarily favored over the others.
standard optimizationprocedures.A perennial difficulty Method œ. A variation of the above procedureis
with this approachliesin the scalarizationconcept.In many
complexsystemsit is neither easy n.or desirableto sum-
optimize
J, for eachi sequentially,
all othercomponents
of
J beingheldwithincertaintolerances.
Therearetwo varia-
marize the essentialsof a consequence
by means of a single
numerical value. Thus the need to associate a set of numbers
ß
tionsto thismethod.Oneis to conduct
a one-pass
sequential
optimizationin a hierarchical
orderchosen
by the analyst.
The other is to o.ptimizethe J, Reratively using some
• Now at School of Advanced Technology, State UnivemSty arbitrarily chosenstoppingrule.
of New York, Binghamton, New York 13901. Method 3. Another method is based on the realizati0n
that it is relativelyeasyto extendthe scalarindexoptimiza-
Copyright • 1974 by the American GeophysicalUnion. tion theoryto the treatmentof a two-dimensional vector
VEM:UaI: MUL•rlrLE-OsJECTIVE OPTIMIZATIO• 45
index.Thus one can reso• to a sequenceof binary minimiza- N, if it exists,is a candidatefor the minimizationproblem
tions. In this procedureone attempls to minimize a scalar [Ho, 1970]. This point N is called the Nash equilibrium
index J definedby J = F(J•, J•) where F can take a form solutionto the nonzero-sum game.
such as Formally, the solutionset {z•, x?, ---, x•*} in an N
player (or N controller)nonzero-sum gameis calleda Nash
+ (2) equilibrium solution if
w• > 0 w2 > 0 w• + w2 = 1 -.. ,
As before, this method also demandsa selectionof w• and <
-- • • ... • x,_*
1 • 30i• Xi+l * • . ' ß •
w• to conductthe optimization.
The three methodsrenewed aboveaH suffer,moreor less, for all x, and i -- 1, 2, ..-, N. As a parentheticalcomment,
from the same fundamental drawbacks. AR involve scalar- note that if J• -- -J.• in a two-personcase,the problem
izationof J usingsubjectiveor ad hoccriteria. Furthermore, in Figure I degenerates to a zero-sumgame,and the Nash
oncethe vectorJ is scalarized,the exactmeaningof multipl•- solutiondegenerates to the well-knownsaddlepoint solution.
objectiveoptimizationis no longerclear. In resourcemanagementproblemsit is assumed,albeit
The proceduredescribedin this paper still usesthe con- optimisticalIy,that the noncooperativeaspectis not quite
ceptof scalarizationas it was presentedin method3 above prevalent. If A and B wish to cooperate,then any point
but attempts to postponethe selectionof weights w, that in the set representedby the shadedregion P represents
goeswith eachJ, until after the performancecharacteristics a solutionthat is superiorto N for both controllers.In-
of the problem are well understood.Thus one need not spectionrevealsthat a subsetof P is of particular signifi-
assignpriorities to the various componentsof J in an ad cance.The set of all pointslying on the locusof the tangent
hoc manner. The proceduregeneratesa set of solutions, points between the J• and the J2 contourshas a special
of whichare equallyviable,and the task of makinga final property. A deviation from any point on this line O•0.•
deci•on is left to the decision maker. cannot result in a simultaneousimprovement of both J•
and J.•. The set A of all points on this line is calledthe
NONINFERIOR INDEX
Pareto optimal set, and the solutionsrepresentedby the
The price one pays for a scalarizationis that it is no points of A are called noninferior solutionsor noninferior
longerclearwhat onemeansby optimization.This ambiguity indexJ• [Zadeh, 1963; Ho, 1970]. ]Ztshouldbe emphasized
is best illustrated by •reating a simple two-criteria two- that one componentof J• cannot be decreasedin value
controlleroptimization problem. Such a problem is some- without increasingthe value of the other componentif one
timesalso referred to as a two-persongame. Consider,for wishesto stay in the Paretooptimalset.That is, the optimal
instance,the problem of two adm•istrators faced with the set alonecontainssolutionsthat are worthy of consideration
task of coordinatingthe use of a co--on resourcesuchas for cooperst. ion or negotiati'on.
•he use of water in a river that cross• a •urisdictional Formally, the solutionset {x?, •, ..-, x• •} in an N
boundary.One administrator,say A, wouldlike to •nimize player (or N controller)nonzero-sumgame is callednon-
J• • J•(x•, x•) by adjustingthe controlx•, whereasthe inferior if there doesnot exist anotherset {x•, x.•, ---, x•}
other, say B, attempts •o minimize such that
GeometricalInterpretation of Noninferiority
An optimizationis more interestingif there are con-
straints on the decisionvariables. The nature of the par-
ticular proceduredependsto a good extent on the
mathematical structure of the constraints. If the constraints
canbe expressed as a setof linearor nonlinearinequalities,
then one would have to posethe multiple-objective opti-
mizationproblemas a mathematical programing problem.
If the constraintsare expressedas a set of differential
CONSTANT
JICONTOURS equations,then one wouldhave to posethe multiple-ob-
jective optimizationproblemas a controlproblemor a
problemin differentialgames.In any event, the important
Fig. 1. An exampleshowingthe Paretooptimalset. pointis that the setof constraintsdefines• feasibleregion
46 ' VEMURI' 1V•ULTIPLE-OBJECTIVE
OPTIMIZATIOlq'
in an appropriately
defined
space.Anyoptimization
prob-
lem,scalaror vectorvalued,
is generally
tractable
if the
feasibleregionis convex.
Consideran N-dimensional
performance
vectorJ whose
components
arethescalarindexes
J•,J•, "', J•. Withno
lossof generality this vectorcanbe assumed to haveall
positive
components (i.e.,J• > 0, i -- l, 2, -' ', N). If the
components of Y are regarded asthe coordinates, thenthe
vectorY lies in the positivequadrant,octant,or orthant,
respectively,
of two-, three-,'or N-dimensional Euclidean
space.Thereforethe set of all pointsdefiningthe feasible
vectorslies in a region,shownby the shadedarea, in the
positiveorthant (Figure 2). A little reflectionshowsthat
the noninferior vectors must lie on the boundary 8f•, in
particularon that portion of 8f• shownby the heavyline.
(Here we are implicitlyassuming fim.tY is beingminimized.)
Fig. 2. Example of a noninferior index in a two-dimensional
This propositioncan be proved through the use of contra- performanceindex space.
diction by assumingthat a vector, say Y•, lying in the
interior of 12is noninferior and by showingthat somecom-
ponents of Y• are not less than the correspondingcom- wherez• > 0 for all i are decisionvariables and b•i are real
ponentsof Y*, the noninferiorvector (Figure 2). Therefore numbers.The Cobb-Douglasfunctions are chosenfor this
the noninferior vector cannot be contained in the interior purposeprimarilybecause of the fundamentalduality between
of the feasibleregion!2; it has to lie on the lower portion them and cost functions.
of the boundary c•12. The nextstepis to definean auxiliaryscalarindexY as
OptimalityVersusNot,inferiority Y= 0
Having establishedthe importance of the noninferior • (4)
set A, we sh, ould emphasizethe differencebetween an ;= i=l Iii
optimal vector index and a noninferiorvector index before
embarkingon methodsto determinesome or all members Thus the useof Cobb-Douglastype functionsin (3) leads
of A.
an auxiliary indexthat is a posynomial.A necessarycondition
By definitiona vectorindex J* • A is optimalif J•* _• J• for the minimization of 3 is
for all i. A vector index J* • A is noninferior if there exists
no other vector in A that is optimal with respect to J*.
These two definitionsimply that an optimal vect,or by ay = =o
definition is also a. noninferior vector. The converse is not
necessarily true. Indeed whenevera vector that is optimal
in the senseof the above definiti, on exists,then the non- The solution of the above defines the v•lue of x that minimizes
inferior vector index problem degeneratesinto one with •. Equation 5 reduces to
a uniquesolution,namely,the optimal vector itself. Phys-
ically,this situationimpliesthat the variouscomp'onents of
J are compatibleand a decrease in one signifiesa decrease
in the other.
where an asteriskdenotesthe minimizingvalue. By sub-
DESCRIPTIONOF THE ALGORITHI• stituting
ß = z* in (4) anddividing
it throughout
by
it can be seen that
Central to this algorithmare three basic features.First
of all, an unconstrained
optimization
will be performed
in =
terms of an auxiliaryperformance
index Y rather than J
itself.Second,the auxiliaryindexY is requiredto be a wherethe performanceweightshave beendefinedas
posynomial(i.e.,positivepolynomial)
[Duj•n et al., 1967].If
the abovetwo requirements are met, then the noninferior = (s)
indexelementsJi* (J = 1, 2, .-. , N) canbe relatedto the
optimum
auxiliaryindexY* via a simplefunctional
form In viewof the relationin (7) the minimumvalueof Y*(x*)
involving
weighting
coefficients
c•i(j - 1, 2, ... , N). This may be written as
finding
becomes
apparent
asthedevelopment ofthealgorithm
proceeds
[Reidand Vemuri,1971].
A firststepin thisprocessis to express
J• in termsof the
controlvariables •;. A particularlyconvenient way of ex-
by •rtue of (8). After • certain amount of algebrathe
pressingthe • is to assume that the oriare of the Cobb-
latter ,equality reducesto [Reid and Vem,uri, 1971]
Douglastype.That is, eachY• is assumed to be of the form
(10)
VEMtIRI: •VITJLTIPLE-OBJECTIVE
OPTIMIZATIO• 47
a = [e•(•,)•]•:" • > 0, •o=•t (•2) Step 3. Rewrite (19) for the optimum caseas
Data
Set J•* J2* 1/Js* x•* xs*
1 10 ø 10 • 106 1.06 X l0 s 1.06 X 10s 4.73 X 10s 1.51 X 10s 1.45 X 10•
2 10 ø 10 • 108 3.34 X 10• 3.34 X 10•' 1.50 X 102 1.51 • l0 s 8.18 X 10ø
3 10 ø 10 • 10 ? 3.34 X l0 s 3.34 >< 10s 1.50 X l0 s 1.51 • 10s 2.59 X 10•
4 10 ø 10 • l0 s 1.06 X 10• 1.06 X 104 4.73 X l0 s 1.51 X 10s 4.60 X 10•
5 10 ø 5 X 10 • 106 3.16 • 10• 1.58 • 10• 3.16 X l0 s 3.10 X l0 s 7.95 X 100
6 10 ø 10 •' 106 1.88 • 10• 1.88 X 10• 2.66 X 10s 3.79 • 10s 6.13 • 10ø
7 10 ø lO s 10 • 3.34 X 10ø 3.34 X iOs 1.50 X 10•' 6.09 X 10• 2.59 X 10ø
8 10 -• 10 • 10 • 5.95 X 10• 5.95 X 10'• 8.41 X lO s 3.79 X 10s 1.09 X 10•
9 5 X 10 -• 10 • 106 8.89 X 10• 1.78 X 10'• 5.62 X 10•' 2.19 X lO s 1.33 X 10•
10 10 -2 10 • 10 6 3.34 X 10• 3.34 X 10• 1.50 X 10• 6.09 X lO s 8.18 X 10ø