Multiobjective_statistical_method_for_in

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

WATER RESOURCES RESEARCH, VOL. 16, NO.

3, PAGES 465-475, JUNE 1980

Multiobjective
Statistical
Methodfor InteriorDrainageSystems
Y. Y. HAIMES AND K. A. LOPARO

Systems
Engineering
Department,
CaseWestern
Reserve
University,
Cleveland,
Ohio44106

S.C. OLENIK

DataSystems
Division,
IBM Corporation,
Poughkeepsie,
NewYork12602

S. K. NANDA

U.S.ArmyEngineer
District,RockIsland,Illinois61201

In thispaperthedesign
of a leveedrainage
system
is formulated
asa multiobjective
optimization
problem
inaprobabilistic
framework.
Thestatistical
nature
oftheproblem
isreflected
bytheprobabilis-
ticbehavior
ofrainfallandriverstage
events
in anygivenmonth.Themultiobjective
approach
allowsfor
theincorporation
ofnoncommensurable
objectives
suchasaesthetics,
economics,
andsocial
issuesinto
theoptimization
problem,
providing
a morerealistic
quantification
oftheimpact
ofa floodorhigh water
situation
in an interiorbasin.A newmethodreferredto asthemultiobjective
statistical
method,which
integrates
statistical
attributes
withmultiobjective
optimization
methodologies
suchasthesurrogate
worthtrade-off
method,
isdeveloped
in thispaper.A casestudyusingdatafromtheMolineareain Illi-
noissuggests
the useof the procedure.

1. INTRODUCTION flowsintotheriverby gravitywhenriverstages


arelow.They
areequippedwithgatesto preventriverflowsfromentering
The essence
of the methodology--multiobjective
statistical
thefloodplain
areaduringfloods.
Pumping stations
discharge
method (MSM)•presented hereforthedesign of leveedrain- interiordrainageflowsover the leveesof flood walls,or
agesystems istheintegration of a multiobjective optimizationthroughpressure lineswhengravityoutletsare blockedby
scheme (thesurrogateworthtrade-off (SWT)method)anda highriverstages. Pondingareasconsistof anylowareasnear
statistical
procedure to assess the differentcombinations of the inletsto gravitydrainsor pumpingstations that are in-
possible system configurations relativetoeconomic, aesthetic,tendedforusein temporarilystoring
interiordrainageflowsin
andsocialobjectives. Givena certainleveeheight,thereare excessof the rate at whichtheseflowscan be passedthrough
manydifferent possible configurations andcapacities for sys- thelineof protection
underspecifieddesign conditions.They
tem components to handleinteriorrunoffand to providea maybe stormdrains,expressly setasideareas,or evenstreets
certainlevelof protection for a givenarea.The levelof pro- and parkinglots,if temporary pondingof interiorrunoff
tectiondepends explicitly ontheriverstageandtheintensity wouldnot causeunacceptable damages[Haimes,1978].
and durationof rainfall, the two statisticalvariableswhich are
Thefollowing sections
detailthedevelopment of theMSM.
considered in theanalysis. Givena setof objectives forsystem
2. DEVELOPMENT OF THE MSM
performance anda finitesetofalternative strategies,
thereex-
istssomeconfigurationwhichwill be 'optimal,in a Paretoop- a. Mathematical Formulation of
timalsensein relationto otherpossibleconfigurations. the Interior Drainage Problem
Althoughthe methodology presented in thispaperis ge- DefineE(x; •/i,rm)asthemaximumpondelevation givena
neric,its development wasaimedat improvingthe interior decision vector,x - [X•, ..., X•,,-.., X•,],a riverstage•/i,anda
drainage systemin Moline.Moline,Illinois,islocated on the stormeventrm.We assumethat rmis associated with a storm
Mississippi
Riverandhasa floodplain
areaof about475acres. hyetograph
whichcanbecharacterized
bya sequence
of rain-
Muchof thefloodplainisheavilydeveloped by industry,espe- fall intensities
•,•anddurations
t,. The basicideais to con-
ciallymostof theriverfront
sites;85%of industrial
acreage in structthestormeventsrmfroma sequence of elementaryrain-
Molineis in the floodplain.Floodingcurrentlythreatensthe fall eventss•,,whereeachs•,is completely specified
by an
area.Generallywhensnowmelts,heavyrainsand frozen intensity
t5 anda duration
t,. If rm,m • {1,2, ..', M}, to be
groundcombine to givea highrunoffupstream of Moline, specifiedlater,is a rainfalleventgivenin termsof the storm
leadingto floodstagesat Molineitself.In general,thelargest hyetograph of Figure1,thenrm-' l,Jjej[-Jn•37
Sj.forappropri-
floodson theMississippiat MolineoccurbetweenMarchand ateindexsets• and• wheret5 is theintensity of a rainfall
late June. eventin thetimeperiod[ti_,,ti-l q-ti],andto---0. A setof de-
A floodcontrolsystemhasbeenproposed to reduceflood cisionsx includes(1) pumpcapacity,(2) pumpoperatingse-
damages in thecityof Moline,Illinois,usinga combinationof quence (3) gravitydrainconfigurations,and(4) others. The
differenttypesof structures
andlandusepolicies. The struc- stochastic
natureof the problemis reflectedin the statistical
tural modificationsinclude levees,flood walls, pumping sta- behavior of the random variables *h and rm.
tions,andgravitydrains;pondingareasalsoplayan integral LikewisedefineD(x; *h,rm)as the durationfor whichthe
role in the overall operations.Gravity outletsare openings pondelevation
exceeds
a specified
threshold
level,givena se-
throughthe leveesthat permitdischarge
of interiordrainage quence
of decisions
xt,,k - 1,2, ..., K, a riverstage
rh,anda
Copyright
¸ 1980by theAmerican
Geophysical
Union. rainfall event rm.

465
Papernumber80W0219.
0043-13977807080W-0219501.00
466 HAIMES ET AL.: MULTIOBJECTIVE STATISTICAL METHOD

visual aesthetics,(2) health and safetyhazards,(3) land use


INTENSITY
I losses,(4) businessand transportationinterruption,and (5)
others.
Formulation of the planningand operationof a leveedrain-
Z/3 agesystemin a multiobjectivesettingeliminatesthe difficulty
encounteredin combiningnoncommensurable objectives.
,, For the sake of analysis,all objectivesare assumedto be a
function of the variablesE( ) and D( ) describedearlier.
I More specifically,considerthe problem
t,+ t2 t,+ t2+ t3 ... DURATION
Fig. 1. Example of a stormhyetographintensityversusduration.

Identify the parametersf• and •/as the minimum and maxi-


minimize •(x) -
x•X / '

mum attainableriver elevations,respectively,to be considered whereX is the setof feasibledecisions; ft (E(x; •1,,rm),D(x; •b,
in a given study. Similarly, define •, fi, and T to be the mini- rm))is the value of the/th objectivefunction for each x, •b,and
mum and maximum intensities(in inchesper hour) and the rm.Since ft dependsexplicitly on the random variables •b and
duration (in hours), respectively,for the componentrainfall rm (the river stage and rainfall events),via the values of the
events(sj,,)to be considered
in a givenstudy.It is importantto elevation and duration functions E( ) and D( ), respec-
observethat not every river elevation is contained in [f/, •/], tively, the optimization problem must be posedto accountfor
nor can everyrainfall eventrmbe generatedas a combination this feature.
of rainfall intensitiesin the range [•, • and a set of durations To takecareof this,let ft (x) denotetheexpected
(average)
in the range [0, T]; i.e., characterizedby a sequence value of the function ft (E(x; T/i, rm), D(x; •b rm)). Mathemati-
However, in any given study,we assumethat the parameters cally,[t(x) -- •{f,(E(x; •/i,rm),D(x;•b,rm))},whereg( ) de-
•, •/, •, fi, and T are definedso as to includeall possibleevents notesexpectedvalue relative to the joint probability distribu-
of significance,and it is in this contextthat we can refer to the tions of the random variables •/, and rm. The set of feasible
aboveparametersasdefininga completesystemof river stages decisionsX is generally characterized by constraintsof the
and rainfall events. form
Define the integers I, J, and N which represent a dis-
cretization of the intervals [//, q/], [•, •, and [0, T] into a se- G(x) <_0
quenceof subintervalsof length H(x) -- 0

an -- - (Eachobjective
function• ( ), 1= 1,2, -.., L depends
implic-
itly on the decisionvectorx sincethe elevationand duration
A•, _=(•- •)/J relationshipsE( ) and D( ) dependexplicitlyon x via the
At = T/N simulation programs.
In orderto proceed,
thefunctions
[•(x), 1= 1,2, ..., L, must
respectively. be computed.The random variablesinvolved are •b and rm,
Define a river stageevent •/i as and they are introducedinto the problem formulationby
meansof the peak pondingelevationand pondingduration
•/,--- iA•/ i e {1, 2, -.., I} relationships.
and similarlya componentrainfall eventsj,,as Define the events(1) A•,,* equal to a componentrainfall
event(inchesof total rain) characterized by a rainfallintensity
s•, -- (/A•,) . (nAt) in the rangeI* 05) = 09-l, •9) and a stormdurationin the
(The componentrainfall eventss•,,are definedin terms of rangeI* (&) = (/n--l,In),j • l, 2, '", J andn = 1, 2, -.., N and
inchesof rain; there exist many combinationsof j and n (in- (2) Bi equal to a river elevationin the range
tensityand duration) which yield equivalentamountsof total Z*(•i) •- (•i--l, •i) i= 1, 2, ..., I
rain. By definings•,,as above,eachrainfall event(and hence
eachstormhyetograph)remainsdistincteven thoughthe total Fromthedefinitionof A•, it isclearthatthereexistsa total
amount of rainfall might be the same.) of JN distinctcomponentrainfall events.For any givenstudy,
Given a sequenceof decisionvariablesX•,, k = 1, 2, ..-, K, a let M denotethe number of stormhyetographsto be consid-
river stage•/i, and a rainfall event rm,E(x; •/i, rm) and D(x; •/i, ered.Associated with eachhyetograph is an eventAm,rn E {1,
rm) are computedvia simulation. 2, -", M} whichcharacterizes the hyetograph. That is to say,
for eachrn E {1, 2, -.., M} thereexistsequences of integers
j
b. Formulationof the OptimizationProblem and n, 1 • j _<J, 1 • n • N, suchthat Am= Uj•j On]9 Aj,*,
The planning and operation(long term) of a leveedrainage where• and• aresubsets of {1, 2, ..., J} and {1, 2, ..., N},
systemhave classicallybeen studiedusingbenefit-costanaly- respectively.
Remark
sis,where the various multiple noncommensurableobjectives
have been commensurated in monetary terms. The benefit-
cost analysisis primarily concernedwith comparisonof the j=l n=!
0 0 I*(•)XI*(t,)=[•,• X[0,TI
reduction in expectedannual damage versusthe costof the fa-
cility. Unfortunately, this type of analysis precludesthe ex-
plicit considerationof nonmonetary objectivesincluding (1) i=1
HAIMES ET AL.: MULTIOBJECTIVE STATISTICAL METHOD 467

In otherwords,{I* (vj) x I* (t,)}, j = 1, -.., J andn = 1, ..-, or, equivalently,


N is a disjoint sequenceof intervalswhich is collectivelyex- I M

haustive in the sensethat any component rainfall event with jr,(x)= Y'. Y'.f, (E(x;% r,., D(x;•,, r•))P(A•IB,)P(B,)
intensityin the range [•, • and duration in the interval [0, T] i-- I m-- I

is containedin a subintervalof I* (vj) x I* (t,) for somej • Steœ,l. Selecta different decisionvector x • X, increment
{1, 2, --., J} and n • {1, 2, -.., N}. Similarremarksapply for q by l, i.e., q -, q + l; if q •< Q, go to step2; otherwise,go to
the river elevation events.
step $.
From available data (discussedin detail in section2c), the
Step$. Giventhesetof orderedpairs{x(q),]?,n},q = 1,
conditional probabilitiesP(A,•IBi), m = 1, 2, ..., M, i = 1, 2, ..., Q, 1= 1, 2, -.-, L, a curvefittingtechniquesuchas least
2,-.-, I (i.e., the probability of attaining a rainfall event A,• squarescan be used to determinethe functional relationship
'given'that the eventB, hasoccurred)canbe computed.These f,:x -, f, (x).
can be usedto computethe conditionalexpectationoff ( )
by meansof the formula d. The MultiobjectiveStatisticalMethod
M
The completeprogram package(see Figure 2) which does
f, (')(x;7,)-- Y'.f, (E(x;•,, rm)
, D(x;•,, rm)
) P(AmIB,) the work necessaryfor this analysisconsistsof three modules:
m•---' I
(1) a hydrologicalsimulationprogram, (2) a regressionpro-
(The conditional
expectation
jr,(0(x;]•) canbe thoughtof as gram which also computesthe expectedvaluesof the objec-
the expected(average)valueof the functionf, ( ) giventhat tive functions,and (3) a separableprogramingroutine usedas
the fiver elevation is in the range (],_,, ],).) part of the surrogateworth trade-offanalysisto determineop-
Let P(B•), i = 1, 2, ..., I denotethe probabilityof the occur- timal designparameters.
rence of the river elevationevent B•, computedfrom available The simulation model Indran [U.S. Army Corpsof Engi-
data (referto section2c). The expectedvalueof f, ( )denoted neers,1975]givespeak pondingelevationsand pondingdura-
by f, ( ) relativeto thejoint probability
distribution
of 7, and tions above three index elevations for various setsof decisions,
rm is given by river stages,and storm events. Elevations and durations are
computed for all combinationsof river stages,storm events,
f,(x)= •.f,('•
(x;7,)P(B,) and decisionvariablesto be used in the study. Output from
this module givesthe resultsof the routings,and thesevalues
M are used as inputs to the regressionprogram.
= i==l
Y.f,(E(x;
n,,rm),
D(x;
n,, The second module of the program develops linear or

- • •.f,(E(x;
i-•l
%rm),D(x;
m== I
hi, STATE
NEEDS
]
where the last equality follows from the multiplicationfor-
mula of probability for any fixed x • X and l = 1, 2, ... , L. VERBAL
OBJECTIVES,
CONSTRAINTS,
DECISIONS
The methodologydevelopedhere is valid in a generalsense,
sinceit doesnot require any a priori assumptionregardingthe
HYDROLOGICAL
RECORDS
QUANTIFY OBJECTIVE
statisticaldependenceor independenceof rainfall and river SPECIFY I E• M FUNCTIONS,
fk(E,O)
stage events.

c. Algorithmto Determinethe
Quantitiesf l ( ), 1= 1, -.., L
COMPUTE SIMULATION
Step 1. Following the procedureoutlined in Appendix A, PROBABILITIES GENERATE E(') 8, O(')
the required probabilities P(A,.i) or equivalently
and P(B•),m = 1, 2, ..-, M, i -- 1,.2, -.., I are computed.Set q
•--' I.
Step 2. Given a fixed set of feasibledecisions,x • X, for
eachm • {1, 2, ..., M} and i • {1, 2, ..-,/}, determinethe
values of the elevation and duration E( ) and D( ) using GENERATE
fk(') ASFUNCTIONS
OF_x, fk(x)
hydraulic simulation programs such as Indran [U.S. Army
Corpsof Engineers.1975] (a total of M- I values for each).
Step 3. Let fl(E, D) denote the lth objective function SURROGATE WORTH TRADE-OFF
which dependsexplicitly on the pond elevation E and the GENERATE:
pond duration D specifiednumericallyin step2. Using the re- PARETO OPTIMAL SOLUTIONS

sultsgiven by stepsI and 2, computethe expectedvalue (rela- TRADE-OFF


VALUES,
)•ij
tive to the joint probability distribution of the random vari- SURROGATE
WORTHVALUES,
Wij
DETERMINE PREFERRE'D SOLUTION(S)
ablesr,.(rainfall)and7, (riverstage))jr1of thefunctionfl, 1--
1, 2, --., L.
More specifically,
I EVALUATE
RESULTS
I
jrt(x)-- • Y'.
Mf, (E(x;•,, r,., D(x;% r,.)P(A,.,)
i== ! m---- !
Fig. 2. Schematicdiagramof the multiobjectivestatisticalmethod
(MSM).
468 HAIMES ET AL..' MULTIOBJECTIVE STATISTICAL METHOD

to the expectedvaluesof the objec- the cost[Haimesand Hall, 1974;Haimeset al., 1975, Cohon
quadraticapproximations
tive functions which are used in the multiobjective opti- and Marks, 1975; Hall and Haimes, 1976;Haimes, 1977; Co-
mization. Given a set of objectiveswhich are functions of hon, 1978; Haimes and Chankong, 1979].
peak pondingelevationand duration,a setof decisionvectors
and probabilitiesfor a collectivelyexhaustiveand mutually e. Useof the SurrogateWorth
exclusiveset of stage/stormevents,the program computesthe Trade-Off (S WT) Method
expectedvaluesof the objectivefunctionsfor eachsetof deci- The use of the surrogateworth trade-off method with re-
sionsand fits a linear or quadratic function of the decision gardto the interiordrainageproblemis examinedin detailin
variables to the expectedvalue of the objective functions,us- this section.Becauseof the forms of the problem and the ob-
ing a leastsquaresalgorithm.The resultingequationsfor the jectivesthe SWT analysisreducesto a muchsimplerprocess
expectedvaluesof the objectivesas functionsof the decision than in the generalcasewheremanyobjectivesare in conflict
variablesmay then be usedin the multiobjectiveoptimization. with each other.
Linear and quadratic approximationswere made to allow Since all damagesassociatedwith ponding elevation and
us to use a separableprogramingoptimizationroutine. Other duration will increasewith increasingpeak ponding elevation
fitting techniquescould alsobe used,alongwith nonlinearop- and pondingduration,then improvingoneobjectivefunction
timization techniquessuited to the problem. will improveall of the otherobjectives(exceptfor the cost)at
All objectivesare assumedto be functionsof the peak pon- the sametime. Naturally, increasingthe flood protectionlevel
ding elevationE and the pondingduration D. E(x; r/i, rm)is as measured by any of the multiple objectivescosts some-
defined as the peak ponding elevation for given decisionsx, thing.Thus,in the end,the problemis essentially
reducedto a
river stage*b,and stormeventrm.Likewise,D•(x; •,, rm)is de- bicriterion (two objective)optimizationproblem. All of the
fined as the duration for which ponding elevation exceedsa multipleobjectives havetrade-offsassociatedwith the overall
specifiedthresholdlevelj, for givendecisionsx, river stage*b, cost,but trade-offsamongother multiple objectiveswould be
and storm rm.Becausethree different durations (measuredin meaningless(i.e., a trade-off betweenman-hourslost and
relation to three different index damage elevations) are deter- flooddamagecouldnot be madebecausetheir behaviorwith
mined in Indran; any of thesedurationscan be used in con- respectto floodinglevel and duration is similar). The SWT
structingthe objectivefunctions. analysisis most useful when objectivesare in conflict (i.e.,
Two different curve-fittingroutinescan be usedhere. If the when the level of one objectivecan be improved only at the
linear option is chosen,then the/th objectivefunction is as- expenseof others).The trade-offanalysiscan then be con-
sumed to have the form ducted between the cost and each of the other objectives.
In the E constraint formulation [Haimes et al., 1971], in the
[, (x) = bo•ø+ b?X l q- b2q)X2
+ '" + b•øX• first iteration, one or more of the objectiveswill be binding,
and the routine usesa least squarestechniqueto determine while the others remain nonbinding. A dual trade-off value
the coefficientsb?, which give the best fit to the available for the binding objectiveswill be generatedin the process,but
data. in order to obtain trade-offsfor the other objectivesit is neces-
If the quadratic option is chosen,then the lth objective saryto changethe right-handsideE levelsfor the nonbinding
function is assumed to have the form constraintsin such a way as to make them binding. For ex-
ample, if one of the objectivee constraintsis f3(x) •< 10 and if,
/•(x) = bo•ø+ bl½øX
! q-b2q)X2
+ "' + b•q)X•+ bn+l(O(Xl)
2 after an initial optimization,f3(x) has a value of 9, then re-
+ b•+•(ø(x•)• + ... + b?(x•) • formulating the constraintas f3(x) •< 9 (i.e., e3 = 9) would
make this constraintbinding in the neighborhoodof e3= 9 at
The higher-order terms give a closerfit to the inherently non- the next iteration. The printed values for the dual trade-offs
linear objective functions, so the quadratic fit would be ex- give the marginal costsassociatedwith varying the constraint
pected to give more accurateresults.The function of the re- levels by one unit. These trade-off values can be used to vary
gressionpart of the program is simply to carry out the matrix the original e constraintlevels systematicallyuntil a preferred
operationsneededto evaluate thesecoefficients. Pareto optimal solution is reached via the use of the surrogate
Program 'output' consistsof a list of the decisionvectorswith worth functions [Haimes and Hall, 1974].
the correspondingexpectedvalues of the multiobjectives.Be- A stepwiseprocedure correspondingto an algorithm out-
fore performing the curve fitting, the values of the decision lined by Haimes et al. [1975] but specializedfor this problem
vectors may be changed to allow a more reliable fit to be is outlined below.
made (i.e., to reduce the possibility of having an ill-condi- Step 1. Find minimum and maximum values for each of
tioned matrix). Finally, the regressioncoefficientsare used the multiobjectives.These can be easily found by examining
with the correspondingdecision vector values to compare ac- the output from the regressionmodule of the program. This is
tual and modeled objective function values. done to find the approximate range of each objective, as a
The resultsof this program module are then used as inputs function of the decisions to be examined.
to the optimization routine, which usesseparableprograming Step 2. Set initial right-hand sidevalues(i.e., E values) for
to obtain the optimal solution. The first requirement for the each of the constraintscorrespondingto the multiobjectives.
routine involves reformulating the multiobjective problem so Eache•> fjminbut alsoei< fjmax
forj = 2, 3, -.., n, wheren is
that it can be handled by the separableprograming routine. the numberof objectivesandfim,nandfjmax are the minimum
Cost is taken to be the primary objective, and the other mul- and maximum values of the jth objective function.
tiple objectivesare put into • constraint form for the purpose Step 3. Solve
of generating Pareto optimal solutionsand their correspond-
ing trade-offvalues [Haitneset al., 1971].As part of the overall min fl(X)
optimization cycle, the surrogate worth trade-off (SWT)
method is used to solve for the e levels as well as to minimize subjectto f(x) < e xG T
HAIMES ET AL.: MULTIOBJECTIVE STATISTICAL METHOD 469

where T -- {x: x •> 0, and all constraints,


otherthan the E con- will lead to a certain lossof production for the business.If the
straintsare satisfied}and flooding becomesseriousenough,total evacuationof the busi-
ness may become necessary,and both workers' wages and
f(x) = [f2(x), '", fn(X)]
normal productionwill be lost. In this case,duration seemsto
x = IX,, ..., Xn] be a likely candidatefor a strictlylinear type of measure,since
a business which closes when flooding reaches a certain
Each solution also givesthe trade-off vector, A• where A• --
thresholdlevel is likely to remain closedthroughoutthe entire
[3•2,3•3, -", 3•n].If all of the E constraintsare binding, then e
time that flooding is above that level. The business inter-
= f(x), and the outputsare f•*(f) and A,*(f) at f(x) -- e. Trade-
ruption function could then have the form
off values of 3•i•correspondingto nonbinding constraints
should be ignored.
bid O•< E•< 1
Step 4. If enough information has been generated from
previous iterations, then proceed to step 5. Otherwise, select a2 4- b2D 1 •< E •< 1.5
new valuesfor e and return to step3. To selectnew e values,
(1) if a censtraint
j is not binding,thensetE•-- fi (x*), or (2) a f ,(E, D)= < a3 4- b3D 1.5 •< E •< 2.0 -
binding constraintmay be varied to make it more binding, ß

ß
o

etc. If decisionmaker interactionscan be made during this


process,then searchtechniqueslimiting the values to be ex- an + bnD En•< E
amined in finding the preferred solution can be used to ad-
vantage (i.e., fewer choicesfar from the indifference band f l ( ) wouldbe themeasure
(in man-hours
lost)of thebusi-
need be examined). nessinterruptionobjective.The coefficientsai and bi are func-
Step 5. Develop the surrogateworth functiens--W•2(f), tions of the elevation, and they would be determinedfrom a
W•3(f), "-, W•n(f)--by generatingdecisionmaker input as fol- detailed examinationof the businesses affected,basedon the
lows. For each set of valuesf, A•(f), and f•*(f) at which the answersfrom a questionnaire.
value of the worth is desired,ask the decisionmaker (DM) for 2. Drowning as a function of elevation and duration will
his assessment of how much3• (f) additionalunitsof objec- havethe sameform asbusiness
interruption.Onceagain,the
tive f• are worthin relationto oneadditionalunit of objective elevationwill determinesomedegreeof hazard, in this case
f•. The DM's assessment on an ordinal scaleof-10 to +10 the number of drowningsper unit time. The duration will
(W• > 0 meansthe DM preferssucha trade, W• < 0 means once again enter as a multiplicativefactor which givesthe
he doesnot, and W• = 0 impliesindifference)then provides numberof drowningsfor the total time duringwhich the ele-
the value of W•/f), for eachj = 1, 2,-.- n. vationexceeds the thresholdlevel.The drowningversuseleva-
Step 6. Repeat step5 until a value f *, the preferredsolu- tion curve could be in the form of a discontinuousfunction,
tion, is found which correspondsto a point at which all of the sincethe numberof drowningsfor the low pendinglevelsan-
W•i(f*), j '-- 2, 3, -.., n equalzero.Oncethisis achieved,other ticipatedwould certainly be small, and the number of deaths
values near f* can be tried to determine the extent of the in-
is restrictedto integervalues.There is little correlationof past
difference band.
fatality data with a putative postleveesituation, and so the
Step 7. The preferred decision vector can be found by functionwould have to be a subjectiveindex with personal
solvingthe sameproblem,with the e constraintsset equal to opinionsweighingheavilyin the final quantificatien.Here
the f * values.
again,the generalform of the drowningfunctionf2(E, D) is:
Step 8. Stop.
This processgives marginal costs (trade-offs) associated f2(E,D) = {a,+ b,D,1• •<E •<E•,i= 1,2, ..., n}
with Pareto optimal solutionsfor improvingany of the objec-
3. General aesthetics would includevisual,olfactory,and
tives at given levels for all of the functions. Decision maker other considerations. An ordinal scale would be established
preferencesare examinedto seewhetherfurther changeis de-
with a rangefrom zero (worst)to 10 (best).The form of the
sirable.
objectivefunctionwould determinea valuein this range,de-
pendingprimarily on the person'sopinionaboutthe aesthetic
f. Multiple Objectives
appealof a certainpendinglevel or pendingduration.For a
Severalmultiple objectivesare suggested belowwithout im- certain elevation range, coefficientscould be establishedto
plying that they constitutea completelist. Possiblefunctional match the ordinal best/worst scale for all the durations which
forms for the multiple objectivesin termsof pending eleva- are possible.The constantswould be such as to increasethe
tion and duration are also considered in this section. Some ob- value of the function for low elevations and short durations.
jectives may be implicitly containedin the NED account,but This functionwould be definedby meansof inputswhich
the strengthof this methodologyis that it allows them to be would give a rating for differentpendingelevationsand dif-
examined independentlyat the same time as objectivesin ferentdurationsasmeasuredat any of threepossibledamage
their own right. elevations.Theseinputswould be usedas grid pointsfor lin-
1. Businessinterruption can be measuredeither in terms ear interpolation(but linearizingonly betweendurationsmea-
of man-hourslost or monetaryloss.For any pending area, suredrelativeto the sameindexheightand not betweenmore
elevation of the pended water can be related to its inundated than one).
area, so that it is alwayspossibleto find which businesses
and 4. Health hazardswould includemosquitobreeding,wa-
other structureswill be affectedby a certain pending level. ter contamination,or similarconsiderations.
Mosquitobreed-
Each businesswill probablyhave a flood damageprevention ing would becomeimportant only if pended water were ex-
plan, which may involve using the employeesin flood pre- pected to remain for about 2 weeks, which is about the
ventionefforts.orin movingcompanyinventoryto saferleca- minimum time neededfor the cyclefrom egg to adult mes-
tiensratherthanhavingthemworkat theirnormaljobs.This ' quiteto be completed.
In thatcase,thenumberof mosquitoes
470 HAIMES ET AL.: MULTIOBJECTIVE STATISTICAL METHOD

bred would be the mostrelevantmeasure.This quantitycould min f2(x) = man-hourslost


be found fairly accuratelyfor any particular area, given the
amountof pondedwater and the typesof mosquitoes. max f3(x) = aesthetics
Contaminationof a municipalwater supply could be im- subjectto constraints.
portant in certainlocations.If pondedrunoff reacheda point The man-hourslost objective,f2( ), was formulatedas a
whereit couldleachinto groundwatersuppliesor perhapsen- functionof pondingelevationE andpondingdurationDi (i.e.,
ter through the tops of wells, a health hazard for the entire the duration during which the pondedwater exceedsan index
community would be created.The seriousness
of this would of elevation Er (/)) as follows:
course depend on whether an alternate water supply was
available, and what percentageof normal needscould be han- f,( )- 0 for 0 _<E_<Er (1)
dled by an alternate system.Generally, it would be a situation f,( )= blD,E for Er(l)--< E_<Er(2)
whereeitherpondedwaterexceededa thresholdlevel (which
would lead to contamination),or it did not. The expressionof f2( ) = b2D2E for Er (2) _<E_<Er (3)
the objectivein this casewould be a value which would have a
f ,( ) = b3D3E for E > Er (3)
zero value up to a certainpondingdepth wherecontamination
would first start. As ponding depth increases,increased bl = I b2 = 10 b3 = 100
amounts of drinking water would become contaminated, and
Er(1) = 564 Er(2) = 565 Er(3) = 566
the objectivewouldreflectthisasa functionof pondingdepth.
5. There seemto be a very few ecologicalconsiderations all in feet above msl.
which would be relevantto the pondingscenarioexaminedin The aestheticsobjectivef3( ) was formulated on a scale
this case study. Environmental considerationswould have to from zero to 10, with 10 being most satisfactoryand zero
be investigatedin detail at each pondingsite. A measureof being least satisfactory.All durationshere were taken relative
this objectivecould be acresof grassdestroyedor number of to the lowestindex elevationEr (1).
treeskilled, etc. by differentpondingconditions. Separableobjectivefunctionsand constraintsin the form of
6. Land use lossesspecificallyincludeplaygrounds,recre- f•(X) are commonin the literature.In particular,overallob-
ational area, and parks. Ponded water will tend to restrict the jectivefunctionswhichare the sumof the subobjectives con-
use of these areas, and the amount of curtailment will be a stitutea classof problemsoften referredto as separablepro-
functionof pondingelevationand pondingduration.The cur- blems. The general form is
tailmentcouldbe measuredin the form of userdayslost.This
would require an examination of the normal use of these rec-
reational areas and a determination of user curtailment versus
ponding elevation and duration.
A portionof a samplequestionnaire is givenin AppendixB
wherethejth subobjective
function
depends
onlyonXj, where
for one sampleobjective.Similar formswould be usedto eval-
X• can be a vector of decisionvariables.This format is as-
uate other objectives.
sumed for the aestheticsfunction, with two components
3. APPLICATION OF THE MSM f3,,( ) and f3b( ), where

A case study using three decision variables (pump size, f3a( )= 10 for D=0hours
pump-on elevation,and gate closureelevation)and three ob- f3a( )= 7 for D= 12 hours
jectives(pumpingcost,job man-hourslost,and aesthetics) is
givenhere.Hydrologicaldatafor area2-A of the Moline proj- f3•( )= 4 for D=24hours
ectareain Illinoiswereusedin the problem.In a normalproj- f3•( )= 0 for D _>36hours
ect evaluation,inputsfrom affectedpersonsin the projectarea
would be important in determining the forms of the multi- f3o( )= 10 for E = 0
objectivefunctions.The projectengineerswould usethesein-
f3t,( ) = 9 for E = ET(1)
putsto developobjectiveswhich dependupon pondingeleva-
tion and duration. In the case study here, however, this f3•,( )-- 6 for E--Er(2)
processwas bypassed,and functional forms which approxi-
mated the functional form of the objectiveswere used. f3•,( )•- 0 for E_•Er(3)
Four different pump sizes(0, 65,000, 90,000, and 150,000 Thesepoint valueswereusedas grid pointsfor a piecewise
gpm), three different pump-on elevations(564, 565, and 566 linearfit, sothatelevations
anddurations
betweenthosegiven
feet abovemean sealevel (msl)) and three differentgate clo- above could have a functional value associated with them.
sureelevations(565, 566, and 567 feet abovemsl) were exam- Finally,f3( ) wasdefinedasf3( ) -- f3a( ) + f3b(), so
ined. Costswere associatedwith each pump size and these that the value of 20 wasits maximumpossiblevalue and zero
values were used as grid points for a piecewiselinear fit, so was its lowest possiblevalue.
that intermediatepump sizescould have costsassociatedwith
them by linear interpolationbetweentwo of the grid points. a. Formulationof Linear and QuadraticForms
The primarycostobjectivefunctionf,( ) is of this piecewise Define the following three decisionvariables:
linear form.
The overall problem was Xl pump sizein gallonsper minute times 104;
X2 pump-on elevation in feet above 563 feet above msl;
min f l(X) = cost X3 gate closure elevation in feet above 564 feet above msl.
HAIMES ET AL.: MULTIOBJECTIVE STATISTICAL METHOD 471

Pump-onelevationis the water elevationat which the pump Linear regression.


will turn on. Gate closure elevation is the river height at
]2( )= 10,100- 610X, + 279X2+ 15.2X3
whichgravitydrainsmustbe closedto preventbackflow from
the river. ]3( ) = 0.00606+ 0.0341X,- 0.0473X2
- 0.0126X3
Four differentpump sizeswere examined:0, 65,000,90,000,
and 150,000gpm, with correspondingcostsof $0, $280,000, Quadraticregression.
$350,000, and $520,000. The three pump-on elevationsused ]2( ) = 15,700+ (43.7X! 2 - 1570X!)
were 564, 565, and 566 feet above msl, while the three gate
closure elevations were 565, 566, and 567 feet above msl. + (243X22
- 573X2)+ (135X32
- 527X3)
Sincewe have alreadyexpressed the costfunctionfl(X!) in ]3( ) -- -0.857 + (0.126X!- 0.00418X! 2)
termsof the pump sizealone,it is not necessaryto perform a
regression on fl(X!) in termsof the othertwo variables.That + (0.0525X2 - 0.0313X22)
+ (0.425X3- 0.11X3•)
is, the cost is assumedhere to depend on the pump size X• A formulation of the overall multiobjectiveproblem was
aloneand not on the pump-onelevationor gateclosureeleva- carriedout usingthe E constraintapproach,with the costob-
tion. The four pump sizesand associatedpump costswere jectiveconsidered asthe primaryobjectiveand the othertwo
used as grid points for determiningthe cost function f!(X!) objectives enteringtheproblemasEconstraints [Haimeset aL,
througha piecewiselinear fit. Samplepump sizesother than 1971; Haimes and Hall, 1974].
the four given above can then have a cost associatedwith The form of the overall problem then becomes
them by linearizing between the grid points immediately
above and below the samplepump size.The equationfor de- mthjr, (x)
termining this cost is
subjectto
samplepump size
sample pump cost= ]': (x) <_
higher pump - lower pump
x (higherpump cost- lower pump cost)
(x) >_
g(x) _<b
For example,given a samplepump size of 32,500 gpm, the
associatedpump costhere would be given by x_>0

32,500 A separable programingroutineisusedto performtheopti-


x ($280,000- $0) -- $140,000
65,000 - 0 mization, so the problemmust be manipulatedslightlyto
satisfyitsinputrequirements.
Levelsfor e• ande3mustalsobe
The costfunctionf,(X!) wasconstructed on the basisof the
specified.
Thesee valuesaremanipulated
aspartof theoverall
four providedgridpoints.In orderto retainthe linearcharac- surrogate
worthtrade-off(SWT) analysisto givethe overall
teristicsof the model but still reflect the nonlinearity in pump
preferredsolution.
costversuspump size,separableprogramingwasused.
Reformulatingboth problemformsaccordingto standard
The form of the piecewiselinear costfunctionis then given
procedures for usingseparable programing (usinga grid of
by
fivepointsfor X• and X3 in the quadraticcase)givesthe fol-
fl (X,) '- 0X,o + 2.8X,! + 3.5X,2 + 5.2X!3 lowingresults(settingE2= 4000,•3 -- 0.35):

where the cost coefficients0, 2.8, etc. have been divided by min(0X,o + 2.8X,, + 3.5X,: + 5.2X,3) (1)
105,and wherethe X,• are specialvariables.The pump sizeis
definedin terms of thesespecialvariablesby the equation subjectto the constraints
X, -- 0X, o + 6.5X,, + 9.0X,2 + 15.0X13 10,100- (0X,o + 3965X,, + 5490X,2 + 9150X,3)
It is also required that + 279x• + 15.2x3_<4000 (2)
Xlo + Xii -•- X12+ Xi3 -" 1 0 _<X v _< 1 0.00606+ (OX,o + 0.2216x,, + 0.3169x,• + 0.5115x,3)
- 0.0473X• - 0.0126X3 > 0.35 (3)
insuringthat onlytwoadjacentspecialvariables
canbe non-
zero at once.
1 _<X2 < 3 1 _<X3_<3 (4)
Expectedvaluesfor man-hours lostf•( ) and aesthetics
f3( ) weredetermined usingthemethodof section 2d for a X,o + X,, + X, 2 + X, 3-- 1 (5)
complete setof riverstagesandstormevents. Theseexpected Pump size -- 0X,o+ 6.5x,,+ 9.0x,2 + 15.0x,3
valueswerethenregressed in termsof thethreedecisionvari-
ablesX,, X2, and X3,determiningthe regression as Equation(1) isf,, thepiecewise
coefficients linearformof thecostobjec-
outlinedin section2e. Both linear and quadraticfits were tive.Equation(2)is f•, theman-hours lostobjective.
Equation
performed
on thesamedata,allowing
a comparison
to be (3) is f•, the aesthetics
objective.
Equations
(4) restrictthe
madebetweenthe resultsfor each.The outputfromthe re- pump-onelevation
andgateclosure
elevation
to valueswhich
gression
modulethengivesthe following formsof the ex- havebeenexamined in theregression
andsimulation.Equa-
pected
values
forf•( ) andf3( ) interms
ofthedecisionsX•, tion(5) is therestriction
on thespecial
variableassociated
X•, and X3. with the pumpsize
472 HAIMES ET AL.' MULTIOBJECTIVE STATISTICAL METHOD

Using the optimization routine to achievea Pareto optimal


solutiongave the following results:

Pump size

Xl -- 11.7 - 117,000 gpm


Pump-on elevation

X2 = 1 = 564 feet above msl

Gate closure elevation

X3 = 1 = 565 feet above msl


Man-hours lost

jr2= 3268
Aesthetics index

f3 = 0.35
Pump cost

$426,000
Trade-offs

•'12-" $0.0046/man-hour

•'13= $8.7/units of aesthetics

The trade-off value for man-hours lost is 0.0046. Namely,


an expenditure of $1000 would reduce the number of man-
hours lost by 4.6. Other Pareto optimal solutionscan be ob-
tained for different levels of man-hours lost.
Table 1 summarizesa sample of Pareto optimal solutions
and their associatedtrade-off values. This example using the
linear caseis very useful for illustrating the way in which Pa-
reto optimal solutionscan be generatedand compared.How-
ever, the assumptionsused in formulating the linear model
would not lead to as gooda solutionas the more closelyfitted
quadratic form.
b. Evaluationand Comparisonof Results
The quadratic form approximates the trend of the data
more accurately, so it would be expectedto yield more realis-
tic solutionsto the problem,provided that the grid usedin the
separableprograming routine is fine enough to take advan-
tageof the increasein accuracy.The quadraticform generally
requiresmore computationsto determine grid point valuesfor
the objectivesbut is generallypreferred, sincethe linear form
has a tendency to oversimplify the problem.
For more accurate solutions it may be advantageousto
solvea problem initially with a rather coarsegrid (grid points
widely spread)and then re-solveit, using a finer grid only in
the neighborhoodof the solution of the first approximation.
This would allow a better approximationwithout increasing
the number of grid points (and thus the computationalbur-
den).
The linear form of the objectiveregressionsis X•l.,x•+ B.
Becauseonly a unique trend stipulatedby Ai is given,xi will
either be assignedto its upper bound or its lower bound if it
has a zero costcoe•cient in the objectivefunction.Becauseof
this property the systemsolution may becomeoversimplified
and unrealistic.If the formsof the objectivesafter a linear re-
gressiondo not closely correspondto the actual values as a
functionof elevationand duration,especiallyin the neighbor-
hood of the upper and lower bounds,a quadratic regression
would be more suitable.
HAIMES ET AL.: MULTIOBJECTIVE STATISTICAL METHOD 473

Quadratic regressionapproximatesfunction trends much (2) there existsan integerM suchthat the sequenceof events
better and does not force variables to assumean upper or {Am}m--•
s•includesall possible
combinations
of railafalldepth
lower bound but rather allows them to take on discrete valuesand storm duration required for the analysis,where P(Am) is
within the interval of definition. The discrete values corre-given by available Weather Bureau data, are adjusted(on a
spondto the grid pointswhich have beenusedto approximate monthly or annual basis)to satisfy
the forms of the separable functions of that variable. The M

value which will be assignedto that variable will correspond


to eitherthe maximumor minimumof the quadratic(Ax? + msi

Bxi + C) form provided that xi has zero costcoefficients.


Even thoughthe sequence
of events{Am}m--•
s• is not a com-
APPENDIX A: plete systemof eventsin general,sinceit is not collectivelyex-
COMPUTATION
OFTHEPROBABILITIES
P(A•IB•
) ANDP(BJ haustivein the spaceof all rainfall events,it is a completesys-
tem of events with respect to the (reduced) space SR of
One method suggestedfor computiug the probabilities 'relevant rainfall eventsconsideredin the analysis'
P(A,,,IB•
) and P(B•) relieson the assumption
that a historical
M
recordis available. When this is the case,the probabilitiesare
Sn= l,.J Am
computedusing the standard method of frequency analysis. m=l

However, this method may be unmanagableor undesirable


for several reasons: For eachmonth k, k = 1, 2, -.., 12, de•e Pmkas the proba-
1. A record of sufficientlength is unavailable. bi•ty that the event•m occurs• month k, m = 1, 2, ..., M, k
2. Significantinfrequent events,i.e., eventsof intenserain- = 1, 2, -.., 12;q•nas the probability that the event B• occurs•
fall and high fiver stagewhich occur with small, but nonzero month k, i = 1, 2, ..., I, k = 1, 2, -.., 12. Then for each k E
probability, may not be capturedin the record. {1, 2, --., 12} the probabilitiesPinkand qinare no•al•ed on
3. The computationalburden may be excessivedue to rec- an annual basis, i.e.,
ords of long length, i.e., those records which are extensive 12 M

enough to satisfy the probabilistic assumptionthat the 'fre- • EPmn=I


k•l
quency' of occurrenceis a good approximationof the 'proba-
bility' of occurrence. 12

To circumventthe problemsdiscussedabove in computing k•l i•i

the probabilitiesP(AmlB•)and P(B•) when usingan historical


record, a procedurecan be developedusing regional rainfall The random exper•ent consistsof 24 boxes a•anged • a
data (provided by the Weather Bureau) and the stage-dura- mat• format with 12 rows and 2 colu•s. Each row repre-
tion curves for the river. sentsa month • the year, the first colu• consistsof 12 boxes
The followingdata are availableon a monthly basis:(1) for each conta•g the events A•, A:, -.-, A•, and the second
a storm of a given duration, i.e., 1 hour, 1 day, etc., rainfall column consistsof 12 boxeseach conta•g the eventsB,, B:,
depthin inchesversusprobabilityof occurrence,and (2) fiver ß.., B•. For each row k and colu•s one and two, respec-
elevation (in feet) versusduration (in hours). The idea is to tively, associate the probabilitiesPmk,m E {1, 2, '", • and
developa probabilisticmodel basedon a sequenceof random q•n,i E {1, 2, ..., •. The randomtrial canbe described• two
trials which retainsthe statisticaldependenceof storm occur- steps.
rence and river elevation (when applicable).The objectiveis Step1. A row k, k E {1, 2, -.-, 12}, is chosenwhereit is
to computethe probability P(AmBi),i.e., the simultaneousoc- assumed that prob {rowk} = 1/12, k = 1, 2, ..-, 12,that is, a
currence of a rainfall event and a river elevation event, to be uniform distribution.
usedin computingthe mathematicalexpectationof the com- Step 2. Once a row has been selectedfrom step 1, a ra•-
ponent objective function. Two approachesare considered: fall event•m and a river stageeventB• are selecteds•ultane-
(1) computingP(AmB•)directly and (2) computingP(AmlB•) ouslyat randomfrom the boxes• colu•s one and two, re-
and P(B•) and then applying the multiplication formula specfively.
P(AmB•) = P(AmlB•)P(B•).If approach 1 or approach 2 is Us•g the theoremof total probability,
adopted,it still retainsthe flexibility of allowing assumptions 12

regardingthe statisticaldependenceor independenceof rain- P(•mB,)= • P(•mBJrow


k)P(rowk)
k•i
fall and fiver stageeventsto be employed.For example,if the
statisticalindependenceassumptioncan be justified (in a par- i= 1,2, ... ,I m= 1,2, ..-,M
ticular month), then P(AmB•) = P(Am)P(B,) and similarly,
P(AmlB•)= P(Am),simplifyingthe computationsrequired.
where P(row k) -- 1/12, k -- 1, 2, --., 12. If, in any given
Define the following events:,4,,; a rainfall event, i.e., a
month, the assumptionof statistical independencebetween
rainfall depthin inchescorresponding to a storm(or sequence
rainfall eventsand river stageeventsis valid, the computation
of storms)of given duration(s)and B•; river elevationevent;
simplifiesto
i.e., the river elevationis in the range(B•_•,Bi).
Considerthe followingprobabilisticmodel relatingthe oc- 12

currenceof a rainfall eventA mand a fiver stageevent B•. The


followingassumptions are required:(1) the sequenceof events
{Bi}•=•• representsa completesystemof events,where the i-- 1,2, -..,I m-- 1,2, .-.,M
probabilitiesP(Bi) are normalized to satisfy
and the requiredjoint probabilitiesare completelyspecified.
• P(B,)
i= i
=1 Considerthe situationwhen the statisticalindependenceas-
sumptionis not valid. In this case,the problem can be stated
474 HAIMES ET AL.: MULTIOBJECTIVE STATISTICAL METHOD

as follows.Let Pk(•4mBi)
----P(•4mBilrow
k), then usingthe mul- postleveesituationis what mustbe dealt with. This is a situa-
tiplicationtheoremwe compute tion with which no one has had any direct experience.So a
greatdealof thinkingaboutprobablecompanypoliciesin this
P(•4mB•)
= 1••12
Pk(,4mIB,)P(B,)
k----I
situationmustbe done to give the bestpossibleanswers.In all
probabilityit would be impossibleto specifyexactlywhat a
i= 1,2, -.. ,I m= 1,2, -.. ,M company'spolicy will be so early in the developmentof the
project, but your best estimate would be appreciated. It
wherePn(AmlB•) is availablefrom the analysisof data for any shouldbe emphasizedthat pondinglevelsreachingelevations
month k (theseconditionalprobabilitiespresenta computa- which will be mentioned here are rare events.Heavy rains, es-
tional problem comparableto that requiredby the coincident peciallyduring timeswhen gravity drain outletsare blocked,
frequencymethod [U.S. Army Corpsof Engineers,1975]),and will be the primary reasonfor having pondedwater. The sys-
the probabilitiesP(B•) are alsoknown. tem as designedwould handle ordinary stormswithout very
Similarly, much ponding.
12 Some of the questionswill seem very difficult to answer
P(•4mlBi)
-- • Pn(•4mlB,)P(row
k) quantitatively,but your bestand mostinformedanswerswill
k•l
be very helpful. Consultationswith districtrepresentatives of
i-- 1,2, ... ,I m-- 1,2, ... ,M the U.S. Army Corpsof Engineerswill be mosthelpful in es-
tablishinglimits of pondingin certainsituationsand also in
from the theoremof total probability. providingother technicalinformation.Thank you for your
In this section,a computationallyfeasible procedurefor help.
computingthe requiredprobabilitiesP(AmBi),rn -- 1, 2, ...,
M, i = 1, 2, -.., /, or equivalentlyP(AmIB•)using regional The Questionnaire
Weather Bureau data and fiver elevation versus duration data(Note: elevationsand durationslistedhere are not necessarily
on a monthly (hourly) basisis developed.If assumptionsre- thosewhichwouldbe usedin everyprojectarea.Pondinglev-
garding the statistical independenceof rainfall events and els are in feet abovethe zero damagelevel, and durationsare
fiver stageeventsare valid, for example,in the caseof a large in days.)
fiver and a largebasinin a shorttime horizon,a simplification 1. Past experiencewithflooding:general.
in the computationsrequiredis observed[Haimes, 1978]. Has floodingbeen a problem for you in the past?List the
flood years,the maximum floodingdepths,and the flooding
duration for each.
APPENDIX B: SAMPLE QUESTIONNAIRE
Introduction
Was it ever necessaryto suspendnormal businessopera-
tions, or to modify the normal work routine becauseof flood-
A sample questionnairewith the type of questionsthat ing?If so, at what floodingdepth did this becomenecessary?
would have to be answered in order to formulate reasonable If more changesoccur as a function of how high the water
objectivefunctionsfor severalobjectivesappearsbelow. In gets,try to list thesechanges,alongwith the floodingdepth at
addition, a short write-up explaining the potential flooding which they occur.
problemasit wouldexistafter leveeconstruction is given,and 2. Possiblebusinessinterruption.
this would probably be put as a prefaceto the questionnaire, (A) Briefly outline your company'sflood fighting proce-
to make sure that everyone would be addressingthe same dures.What stepsare taken to reducefloodingdamagein a
problem as they answerthe questions.The form has been normal flood(i.e., high-riverstage):Is adequatemanpoweral-
made as generalas practicable,but this doesnot imply in any ways available or are volunteersusedalso?
way that all of the objectiveslisted will be relevantin every Try to list theseproceduresas they dependon the water's
case.Any objectiveswhich are not consideredto be relevant maximumdepth, along with the appropriatenumberof man-
by the personfilling out the questionnairedon't have to be hourslostor dollarsof productionlost. For example,at a river
consideredin the multiobjectiveformulation [Haimes, 1978]. stage1 foot overthe minimumdamagelevel,perhaps40 man-
hourswill be lostper day, but at a depthof 2 feet,perhaps140
Introduction to Questionnaire
man-hoursmight be lost per day.
A flood protectionproject consistingof levees,flood walls,
and associatedinterior drainage facilitieshas been approved man-hours lost or dollars at I foot over minimum
for the area. Previously,topographicmaps having a contour damage level.
interval of 2 feet and a scale of 100 ft/in. were available for man-hours lost or dollars at 2 feet over minimum

estipaating
elevation-damage
relationsfor eachaffectedarea. damagelevel.
man-hours lost or dollars at 3 feet over minimum
Thesesamemapscan now be usedfor evaluatingother objec-
damage level.
tive functions, to help visualize various possible situations. man-hours lost or dollars at 4 feet over minimum
The basic idea is to try to develop theseobjectivesin as de- damage level.
tailed a form as possible,as a function of peak ponding eleva- man-hours lost or dollars at higher flood stages.
tion and ponding duration.
Ponding levels in the questionnaireare given in terms of Detailed answerswould be appreciated.If thereis no previous
feet abovethe zero damagelevel,a levelwhichmustbe speci- experienceto fall back on, pleaseindicateso, and your best
fled for each different project area; the same topographic estimateswould be helpful.
maps that were used to compute elevation-damagecurves (B) What standard proceduresmight be developedin a
shouldbe usedhere. postleveesituation?Floods that come from pondedwater are
At the same time, it is important to understandthat the likely to give much lessadvancedwarningthan river floods,
HAIMES ET AL.: MULTIOBJECTIVE STATISTICAL METHOD 475

sinceall that would be necessaryfor floodingwould be a suffi- REFERENCES


ciently severe storm event. While this question involves a Cohon, J. L., Multiobjective Programming and Planning, Academic,
greatdeal of conjecture,try to makesomereasonableassump- New York, 1978.
tionsusingyour expertiseto make your bestguessat a policy Cohon, J. L., and D. H. Marks, A review and evaluation of multi-
which could be standardized.Remembertoo that implemen- objectiveprogramingtechniques,Water Resour.Res., 11(2), 208-
220, 1975.
tation time of any policy would have to be a factor here, since
Haimes, Y. Y., Hierarchical Analysesof Water ResourcesSystems:
pondingcould occurquite quickly during and after a heavy Modeling and Optimizationof Large Scale Systems,McGraw-Hill,
rain. Try to developa set of man-hoursand dollars lost fig- New York, 1977.
ures,given differentpondingdepthsas in part A. Haimes, Y. Y., Multiobjective framework for interior drainage sys-
tems:Multiobjectivestatisticalmethod (MSM), report submittedto
man-hours lost or dollars at ponding of 1 foot. the U.S. Army Corps of Engineers, Environ. SystemsManage.,
man-hours lost or dollars at ponding of 2 feet. Inc., University Heights, Ohio, March 1978.
man-hours lost or dollars at ponding of 3 feet. Haimes, Y. Y., and V. Chankong, Kuhn-Tucker multipliers as trade-
man-hours lost or dollars at ponding of 4 feet. offs in multiobjectivedecision-makinganalysis,Automatica, 15(1),
man-hours lost or 59-72, 1979.
dollars at ponding greater than 4
feet. Haimes, Y. Y., and W. A. Hall, Multiobjectives in water resources
systemsanalysis:The surrogateworth trade-offmethod, Water Re-
sour. Res., 10(6), 1051-1059, 1974.
(C) Do you anticipateany interruptionsof transportation Haimes, Y. Y., D. A. Wismer, and L. S. Lasdon, On bicriterion for-
when sufficientfloodingfrom pondingcould occurto cut off a mulation of the integrated systemidentification and systemopti-
plant from normal accessor from delivery of finishedprod- mization, IEEE Trans. SystemsMan Cybern., SMC-1, 296-297,
1971.
ucts,etc.?Correlatenumber of dollarslost with pondingele-
vations of 1, 2, 3, 4, and more than 4 feet for various durations Haimes, Y. Y., W. A. Hall, and H. T. Freedman, Multiobjective Opti-
mization in Water ResourcesSystems:The Surrogate Worth Trade-
such as 12 hours, 24 hours, etc. Off Method, Elsevier, New York, 1975.
Hall, W. A., and Y. Y. Haimes, The surrogateworth trade-off method
with multiple decision-makers,in Multiple Criteria Decision-Mak-
Acknowledgments.The authorswould like to thank Gilmer Blan-
ing: Kyoto 1975, edited by M. Zeleny, pp. 207-233, Springer,New
kenship,Bruce Evans, John Todhunter, and Greg Utrecht for their York, 1976.
contributionto the study summarizedin this paper; they were part of
the team that conductedthe study. The authors would also like to U.S. Army Corps of Engineers,Rock Island District, Draft of mate-
rial for Moline, Illinois, general design memorandum, hydrology
thank JosephRaoul, Jr. for his assistanceduring the conductof the
appendix, 6-21, Rock Island, I11., 1975.
study and Ven Te Chow for his valuablecommentsand suggestions
Wagner, H., Principlesof OperationsResearch,Prentice-Hall, Engle-
made during an early phase of the study. This paper is based on a
wood Cliffs, N.J., 1969.
project titled, 'Multiobjective framework for interior drainage sys-
tems:Multiobjective statisticalmethod (MSM),' conductedby Envi- (Received October 20, 1978;
ronmentalSystemsManagement,Inc. for the U.S. Army Corpsof En- revised November 14, 1979;
gineers,North Central Division. acceptedNovember 21, 1979.)

You might also like