Universit du Qubec
cole de technologie suprieure
DPARTEMENT DE GNIE DE LA PRODUCTION AUTOMATISE
GES802
Analyse de faisabilit
NOTES DE COURS
Par
Ali Gharbi
Professeur
Rvis : Aot 2009
CHAPITRE 1:
CHAPITRE 2:
CHAPITRE 3:
CHAPITRE 4:
CHAPITRE 5:
CHAPITRE 6:
CHAPITRE 7:
CHAPITRE 8:
CHAPITRE 9:
ANNEXE 1:
Table des matires
ESTIMATION DES MARCHS, REVENUS ET COTS
INTRODUCTION LA NOTION DE RISQUE ET D'INCERTITUDE
ANALYSE CONOMIQUE TRADITIONNELLE ET AVANCE SOUS
RISQUE ET INCERTITUDE
TECHNIQUES DE DCISION STATISTIQUE AVEC INFORMATIONS
ADDITIONNELLES
L'ARBRE DE DCISION
LMENTS DE BASE DE LA SIMULATION DES SYSTMES
VALUATION DES NOUVELLES TECHNOLOGIES : UNE APPROCHE
MULTICRITRE
DCIDER FACE LA COMPLEXIT: ANALYSE HIRARCHIQUE DES
PROCDS (AHP)
LMENTS DE CONCEPTION D'UN PROJET
MANUEL D'UTILISATION DU LOGICIEL EXPERT CHOICE
estimating markets,
revenues,
and costs
chapter 1
5.1 INTRODUCTION
In any planning analysis for strategie manufacturing investments, there is
great need for reasonably ace urate estimates of market conditions (e.g.,
market size and share) : pricing structure, and cost patterns. Severa! tech
niques for forecasting the size of the market for new or m_odified product s
are presented in .. Quantitative forecasting l)let_hods covered are
regression techn_i_qyes in
elude the Once an estimate of
market demand for a new or modified product has been preparea, the priing
policy and revenues associated with that product become a matter of con
to ... .. toplc iii the .. subject of SectiQJJ $.3. Numerous
for capital, material, and .. overhead costs are
dis,c.ussed in Section 5.4. Finally, the impact of inflation on. the cost estimat
add_r:essed in Section 5.5.

Estimating Markets, Revenues, and Costs Chap. 5
5.2 FORECASTING THE MARKET
A forecast is useful if it reduces the uncertainty surrounding an event, and by
doing so results in a decision that has increased value in excess of the cost
required to produce the forecast. Forecasts are also useful to givc manage
ment an image of the firm it guides. Companies are often what they want
themselves to be. An openminded, optimistic vision of the future can lead
to growth and profit; a narrow. pessimistic view can lead to stagnation or
worse.
5.2.1 Forecasting Procedure
A conscientious forecaster will follow a logical stepbystep procedure
in developing and revising forecasts . The first step is to decide what re
sponse, or quantity, to forecast. Th en pertinent numerical data are gathered
and summarized in graphical form, whenever possible. Often the data repre
sent sorne response as a function of time, such as sales per quarter. Such
data are referred to as comprising a rime series. The time series should be
examined for underlying patterns, and the forecaster should attempt to ex
plain these patterns . For example, a large sudden increase in sales m ~ y ~ e
explained by increased government spending. A longlerm steady nse m
sales may be a result of graduai public acceptance of a new product.
Typical forecasts based on limeseries data merely ex tend historical
patterns into the future and do not result in predicting abrupt changes (tum
ing points) in activity levels . ln this respect, the forecaster should search for
causal relationships between the response variable (e.g., sales) and other
variables . For example, a rise in interest rates is usually tollowed by a
decline in new construction. This is a causal relationship thal is useful in
forecasting major upswings and downturns in the construction industry.
The next step of the forecasting process is to apply expert judgment to
forecasts obtained by utilizing causal relationships and/or timeseries analy
sis . The result of this evaluation, whether it takes the form of a Del phi study
or simply a group's or an individual's opinion, should be documented and
included in the forecast. Thus forecasting involves a coordinated approach
to collecting data, identifying and mathematically testing causal relation
ships , and/or selecting a procedure for projecting timeseries data, and
choosing a structured approach for eliciting exper1 opinion concerning the
problem under consideration.
One of the most effective forecasting strategies is first to use mathe
matical techniques based on past data and then t introduce judgment in
tt mpting 1 d cid h w th futu r pr bobly wi ll di ff r fr m th p st. A
pr irn udva nr n of this opproa h is th at 1t r lu th l\111 11 r f f11 t M ~ 1
whi h h11111 tn
1
ri n 11111 1 1 ' 1ppl 11 Il wl ly ' ll nll 11 11 1111lh 11111
Sec. 5.2 , Forecasting the Market
83
cal techniques and informedjudgment, both methods can serve as checks on
each other and tend to minimize gross errors.
To evaluate any quantitative forecasting technique, the procedure first
should be defined and theo applied to representative data. For example, if
one's goal is to forecast sales one quarter in advance, one could start by
applying the proposed forecasting technique to, say. the last three years of
quarterly data. Begin with the first quarter and apply the procedure to take
each new quarter's actual sales into account when developing the next quar
ter's forecast. Theo for the historical data one can measure the difference
between the forecast and actual sales. The difference is the forecasting er
ror, and severa! techniques should be evaluated in an effort to minimize this
err or.
5.2.2 Forecastlng New Products
Because CIM usually requires rethinking the way a firm does business,
modified existing products and new products form the basis of forecasting
and eventually justifying the enabling technology. When a company wishes
to forecast with reference to a particular product, it must consider the stage
of the product's li fe cycle for which it is making the forecast. The availabil
ity of data and the possibility of establishing relationships between the fac
tors depend directly on the maturity of a product, and hence the !ifecycle
stage is a prime determinant of the forecasting method to be used (e.g .. see
Ref. [1]).
At each phase in the li fe of a product, from conception to maturity, the
decisions that management must make are characteristically quite different,
and they require different kinds of information as a base. ln the earl y stages
of product development, the manager wants answers to questions such as
these:
What are the alternative growth opportunities of pursuing product X?
How have established products similar to product X fared?
Should we enter this business, and if so, in what segments?
How should we allocate R&D efforts and funds?
How successful will different product concepts be?
How will product X fit into the markets 5 or W years from now?
'
!
1
J
1
i.
1
1
' 1
: ~
EstimatlnJ.! Markets, Reuenues , and osts Chap. 5
product that is still "only a gleam in the eye," information about its likely
performance can be gathered in a number of ways, provided that the market
in which it is to be sold is a known entity. For example, one can compare a
proposed product with competitors' present and planned products, ranking
J it on quantitative scales for different factors. If this approach is to be suc
cessful, it is essential th at the persons (hopefully. experts) who pro vide the
basic data come from different disciplinesmarketing, R&D, manufactur
ing, legal, and so onand that their opinions be unbiased.
Frequently, however, the market for a new product is weakly defined,
the product concept is still fluid, and history seems irrelevant. This has been
the typical case for gas turbines, modular housing, pollution measurement
deviees, and computer terminais. Before a product can enter its (hopefully)
rapid penetration stage, the market potential must be and the .product
must be introduced. Then market testing may be advtsable. At thts stage,
management needs answers to such questions as:
What shall our marketing plan bewhich markets should we enter,
and with what production quantities and priees?
How much manufacturing capacity will the early production stages
require?
As demand grows, where should we build this capacity?
How shall we allocate our R&D resources over time?
Significant profits depend on finding the right answers, and it is there
fore economically feasible to expend relatively large amounts of effort and
money on obtaining good forecasts (short, medium, and long range) .
A sales forecast at this stage should provide three points of informa
tion: the date when rapid sales will begin, the rate of market penetration
du ring the growth phase, and the ultimate levei of penetration, or sales rate,
during the maturity phase.
5.2.3 Quantitative Forecasting Techniques
Two relatively simple, yet extremely useful, techniques for obtaining
initial time series forecasts of demand for a new or modified product are
described in this section: ( 1) regression analysis and (2) exponential smooth
ing. Results of quantitative forecasting are coupled with subjective forecasts
to develop the final projection of market conditions.
5.2.3.1 Linear Regression Analysis
Regression is a statistical method of fitting a Ii ne through data to mini
mize squared error. It is exact; however, graphing might be used to.provide
a satisfactory approximation. With linear regression, statistics can be used
to obtain a forecast and a confidence interval for the forecast . For example,
Sec. 5.2
Forecasting tite Market
85
a person using regression might forecast sales for a particular product for the
next month to be, say, $425,000 and also can state that she is, for example,
95% confident that actual sales will be between $390,000 and $460,000. This
assumes that basi.c market conditions and trends will remain unchanged .
Confidence mtervals are useful in identifying a change in the trend of a
set of data. If two or three consecutive observations fall outside these timits,
the can be fairly sure that the basic relationship has changed since
the htstoncal were gathered. This information may prove financially
valuable by enabhng one to re act quickly to a change in trend. If the trend
changes, however, the forecaster is advised to discard or substantiafly dis
count data regarding the previous trend.
In linear regression involving one independent and one dependent vari
able, the relationship that is used to fit n data points ( 1 :s; ; :s; n) is
y= a+ bx
(51)
A mathematical statement of expressions used to estimate a and b in
the simple linear regression equation 51 is as follows:
i X;Yai  Xi Yai
b = i = l i = l
i XJ xi X;
(52)
; ,. , i = l
a= Ya bi
(53)
Here x and J? are averages of the independent variable and dependent vari
able, respecttvely, for the n data points .
Example 5l
A durable goods manufacturer has found persona! disposable income in its market
region in a given quarter to be strongly related to sales in the following quarter.
data are listed and summarized in Table 51. Because a plot of these data
md1cates an approximately linear relationship between the dependent variable (on
. the Y axis) and the independent variable (on the x axis), linear regression is used to fit
an equation to the data. The data and calculations summarized in Table 5 J are
utilized below to determine the linear regression equation:
n n
L X;Yai  XL Yai
b = i l i:l _ 2,626,817  250.6(9,788)
 1,416,926  250.6(5,01 2)
2: xl  X 2: x;
1=1 1= 1
173,944.2
160,918.8 = 1.
081
a = Ya  bi= 489.4  1.081(250.6) = 218.5
y= 218.5 + 1.081x
'
Table 5l
CALCULATIONS FOR S IMPLE LINEAR REGRESSION
Data Point, i
(period) Y ai
X; X;Yai x!
360 121 43,560 14,641
2 260 118 30,680 13,924
3 440 271 119,240 73,441
4 400 190 76,000 36, 100
5 360 75 27,000 5,625
6 500 263 131,500 69, 169
7 580 334 193,720 111,556
8 560 368 206,080 135,424
9 505 305 154,025 93,025
10 480 210 100,800 44, 100
t l 602 387 232,974 149,769
12 540 270 145,800 72,900
13 415 2 18 90,470 47,524
14 590 342 201,780 116,964
15 492 173 85, 11 6 29,929
16 660 370 244,200 136,900
17 360 170 61,200 28,900
18 410 205 84,050 42,025
19 680 339 230,520 114,921
20 594 283 168, 102 80,089
Totals 9,788 5,012 2,626,81 7 1,416,926
" "
LX; "' 5,0 12 2: y., = 9,788
i ::. l i l
"
LX;
"
< = !.:..!__ = 5,012 = 250 6 Lx!= 1,416,926
. " 20 .
i l
"
2: y.,
"
y.= i = l, =
= 489.4 2: x,y.1 = 2,626,817
il
y.,  actual quarterl y sales ($ 10
1
) for period i
x
1
di sposnble lncome in precedl ng period (S IO')
!li
Sec. 5.2 Forecasting the Marker
87
Figure 5 l shows the plotted data and the calculated regression tine (as weil as
confidence li mits, to be discussed later) . As an ex ample of how the regression equa
tion is used, suppose that disposable income for the previous quarter is $3 10 x 10 6.
Then our forecast or estima te of sales for the current quarter, y, is
J = 218.5 + 1. 081(310) = 553 .6 or $553.6 x JOl
To calcula te a confi dence interval for our forecast, we must obtai n the standard
deviation of a single esti mated value, s(y). This quantity is a measure of the tight
ness of the individual data points about the regression tine,
[
" ] l/2 [  2 ] 112 2: (y.;_ y,)2
1
+ .!_ + (xo  x)
,,. 1 n "
s( y ) = n _ 2 L (x;  iV
i l
(54)
Here x
0
is a value of the independent variable being used to obtai n t he quarter! y sales
forecast, y, and it lies in the range of observed x
1
(i .e., Xmin :5 x
0
:5 xm. , ) :
y = a + bxo
.....
700 1
..
., ...
.!! 600
,,<f:'
0
8
500
oP
c;
,.,_, s
0
vo ::
'
>
,/. . .
:x
>
,
i
/
t:
..
" 0
200
llii!HIJII ''"'ln (lf'lfoUI"u l llorl '< luil ll httiO II I rtollrtl
11'"" l 'htl l lll lllliltl l  ll1111 f111 1' \"' "1'111 \ 1,
1
1
xx EstimatinK Markets, Revenues, and Costs Chap . 5
The 1 00( 1  a) percent confidence interval for the forecast of actual quarterly sales is
defined to be between
Y + (la/l,n l)s(y) (upper limit) (55)
and
Y  (tan,n l)s( y) (lower limit) (56)
The term tan,nl is an appropriate value of the t statistic. A table of t values and
instructions for using this table are given in Appendix C of Ref. [14] as weil as in most
statistics texts.
For a 95% confidence interval (a = 0.05, a/2 = 0.025) and for n data points,
there are n  2 degrees of freedom. Thi s means that with repeated forecasts involv
ing n data points, approximately 95% of the confidence intervals calculated with the
equation above will con tain the true value of the response variable if the underlying
causal rel ationship does not change .
The information necessary to determine the confidence interval for this fore
cast is summarized in Table 52 and is used with Eq. 54 to perform the following
calculations.
 [52,482.8] l/l [ 1 (31 0  250.6)
2
]
112
s(y) 18
1
+ 20 + 160,919.6 .
= 55.88
For a 95% confidence interval we have
to.o2S.I8 = 2. 1009
and by utilizing Eqs. 55 and 56, these limits are obtained:
Y+ Uan,n 2)s(y) = 553.6 + 2.1009(55.88) = 671
.v  Uan,n2)s(y) = 553.6 2. 1009(55.88) = 436
(upper limit)
(lower limit)
To summarize, our forecast of next quarter's sales is $553,600, and we are 95%
confident that the true value (outcome) of quarterly sales will occur between
$436,000 and $671,000.
The correlation coefficient is a measure of the strength of the relation
ship between two variables only if the variables are linearly related. In
Example 51 the correlation coefficient, r, can be determined as follows :
r =
where  1 < r < 1.
n
2: (Yai  Y;)
2
1.0 :_i::_l 
2: (Yai  Ya)
2
il
If there is no relationship at ali (a shotgun etfect) between the depen
dent and independent variables , r will be zero, or nearly zero. A negative
Sec. 5 .2
Foreca.ftIII( the Market
Table 52
CALCULATIONS TO DETERMINE CONFIDENCE INTERVAL
Data Point , i
(period)
y., x, y, (y., y,)' (x1  i)
1
(y.,  .Y.J'
360 121 349.3
114.5 16,796.2
16,744
2 260 118 346. 1 7, 413 ,2
17,582.8 52,624
3 440 271 511.4
5,098.0
416.2 2,440
4
400 190 423.9
571 .2 3,672.4
7,992
5 360 75 299.6 3,648.2
30,835.4 16,744
6 500 263 502.8
7.8 153.8
11 2
7 580 334 580.0
0.0 6,955.6
8,208
8 560 368 616.3 3,169.7
13,782.8
4,984
9 505 305 541! .2 1,866.2
2,959.4
243
10
480 210 445 .5 1, 190.2
1,648.4
88
Il 602 387 636.8 1,211.0
18,605.0
12,679
12 540 270 510.4
1!76.2
376.4 2,560
13 4 15 218 454.2 1,536.6
1,1162.8
5,535
14
590 342 588.2
3.2 8,354.0 10, 120
15 492 173 405.5 7.482.2
6,021.8
7
16
660 370 618.5 1,722.2 14,256.4
29,104
17 360 170 402.3 1,789.3
6,496.4 16,744
18 410 205 440.1
906.0 2,079.4
6,304
19 680 339 585.0 9,025.0 7,814.6
36,328
20 594 283 524.4 4,844.2
1,049.8 10,941
Totals
52,482.8
160,919.6 240,501
Key:
y. 1 = actual quarter! y sales ($ 101) for period i
y1 = estimated quarterly sales ($101) for period i
x1 = disposable income in preceding period ($10)
From calculations associa led with Table 51 :
y = 218.5 + 1.081x; and x = 250.6 and .Y. = 489.4
value of r indicates that one variable decreases as the other increases, whil e
the variables both increase (or decrease) at the same time when r is posi
tive. The closer ris to 1 or +1, the more "perfect" is the correl ation .
Using Eq, 57, the correlation coefficient for data in Table 52 is
r=
n
2: (Ya; Y;Jl
1.0  .:. ':..:.. 1 
2: (Yai  Ya)
2
i=l
= V0.78I8 = 0.88
j
52,483
1.0  240,501
This value of r indicates a fairly good relationship between the independent
and dependent variables.
1!9

Vi
.1:
' 1
1
1
.
90 Estimating Markets, Revenues, and Costs Chap. 5
5.2.3.2 Exponential Smoothing
A major ad van tage of the exponential smoothing method compared to
simple linear regression is that it permits the forecaster to place relatively
more weight on current data rather than treating ali data points with equal
importance. Forecasting equations can quickly be revised with a relatively
small number of calculations as each new data point is collected. Also,
exponential smoothing does not assume linearity.
The main disadvantage of exponential smoothing concerns the basic
assomption that trends and patterns of the past will continue into the fu
ture. However, it is more sensitive to changes than is linear regression.
Because timeseries analysis cannot predict turning points in the future,
expert judgment and/or analysis of suspected causal factors should be used
in interpreting results of forecasting methods discussed in this chapter.
The basic exponential smoothing model that we shall discuss and illus
trate is shown as follows:
S, = ot'x, + (1  a')S
1

1
(58)
or
new = a' ( new ) + (1 _ a') (previous)
estimate datum estimate
This term, a', the smoothing constant, merely provides a relative weighting
for the new datum point compared to previous estimates. In general, a'
should lie between 0.01 and 0.30, but the analyst should not hesitate to use a
value outside this range if it gives better results with representative historical
data.
Single exponential smoothing is Hlustrated with data listed in Table 53
that are graphed in Fig. 52. Based on Eq. 58, the following are example
calculations forS,, which is termed a "smoothed statistic" for period t.
S, = a'x, + (1  a')S,_
1
s 1 = 0.3(50) + o. 7(50) = 50
s2 = 0.3(52) + o. 7(50) = 50.6
s] = o.3(47) + o.7(50;6) = 49.52
s4 = o.J(5I) + 0.7(49.52) = 49.96
ln determining S
1
above, a value of S
0
must be estimated. Ifthere is no trend
in the initi al data, an estimate based on the average of the flrst few data
roi nts is adequotc. Here S0 wos cstimutcd lO be 50.
To bctt cr undetstnnd the mc:n nlog of Nlll oOthi n!(, the fol
o u how d!111 11 11 ll l lh'tll \ 1 , 1 ,
1
, 11 1, '" 11 lnd111l cd
\.
80
 70
:J
'
"
0
ill 60
5
2
g
j
1
Table 53
EXPONENTI AL SMOOTHING EXAMPLE
Period
Number, 1
0
l
2
3
4
5
6
7
8
9
10
Il
12
13
14
15
16
17
18
19
20
Demand. x,
(1,000 units)

50
52
47
51
49
48
51
40
48
52
51
59
57
64
68
67
69
76
75
80
S,
(a' = 0.30)
50.00
50.00
50.60
49.52
49.96
49.67
49. 17
49.72
46.80
47. 16
48.61
49.33
52. 23
53.66
56.76
60. 13
62. 19
64.23
67.76
69.93
72.95
lt!
,.,
,.,
/
.....
,...,
..
/
.....
.1''
' l'/

. Demand. x,
l'ori;tl l lllllltlit
 JJ. Exponont lally
lffiOOthod
esti mete, $,
1 1
1
14 1 10 1tl
= $14.38
p = 16,000
A summary of the unit priees is:
(a) Based on investment
(b) Based on total costs
(c) Based on net sales
(d) Based on incrementai costs
(e) Based on value added
Product
M
$15. 18
14.30
16.80
14.60
15.00
Product
N
$13.60
13.00
17.15
12.00
14.38
Chap . 5
Since there are numerous variations in the pricing procedures illustrated, any
policy requires clear statement to avoid misunderstanding of exactly how a final
priee was established. The results in parts (a) and (b) are perhaps the most mean
ingful.
5.4 COST ESTIMATING: PERSPECTIVE AND APPROACHES
In Chapter 2 we described various cost concepts that are important to eco
nomie analyses. A key point was that the costs which are important to
economie comparisons are the marginal (variable, or incrementai) costs for
the future, and that these costs may weil include opportunities foregone as
weil as actual cash flows .
The basic difficulty in estimating costs for purposes of making eco
nomie analyses is that most prospective projects are unique; that is, substan
tially similar projects have not been undertaken in the past under conditions
that are the same as expected for the future . Hence, data that can be used in
estimating costs directly and without modification often do not exist. lt may
be possible, however, to gather data on certain past outcomes th at are re
lated to the outcomes being estimated and then to make forecasts regarding
anticipated future conditions .
Whenever an economie analysis is for a major new product or process,
estimating for that analysis should be an integral part of a comprehensive
planning procedure. Such comprehensive planning would require the active
participation of at )east the marketing, design engineering, manufacturing,
finance, and topmanagement functions. It would generally include the foi
Iowing ingredients :
1. A realistic master plan for product development, testing, phase into
production, and operation
2. Provision for working capital and facilities requirements
Sec. 5.4 Cost Estimating: Perspective and Approaches
10.5
3. Integration with other company plans
4. Evaluation against company objectives for market position, sales vol
ume, profit, and investment
5. Provision of operating controls, assuming that the project is adopted
Obviously, such comprehensive planning is costly in time and effort ,
but wh en a new product or process has major implications for the future of a
finn, it is generally a sound rule to devote a greater rather than a fesser
amou nt of effort to complete planning, including estimates for the economie
analyses that are a part of the planning. The application of this rule, of
course, is bounded by constraints of limited time and talent. Following the
rule will tend to minimize the chance of poor decisions or Jack of prepared
ness to implement projects once the decision to invest has been made.
As technology continues to accelerate, quick and reliable methods are
needed for determining the costs associated with adding new and/or improv
ing existing manufacturing processes. The methods discussed below are
excellent for estimating costs oflabor, equipment, materials , operations, and
engineering.
5.4.1 Sources of Data
The variety of sources from which information can be obtained is too
great for complete enumeration. The following four major sources, which
are ordered roughly according to decreasing importance, are described in
subsequent sections: (1) accounting records, (2) other sources within the
finn, (3) sources outside the firm, and (4) research and development.
5.4.1.1 Accounting Records
It should be emphasized that although data available from the records
of the accounting function are a prime source of information for economie
analyses, these data are very often not sui table for direct, unadjusted use. In
its most basic sense, accounting consists of a series of procedures for keep
. ing a detailed record of monetary transactions between established catego
ries of accounts, each of which has an accepted interpretation useful for its
own purposes . The data generated by the accounting function are often
inherently misleading .for economie analyses, not only because they are
based on past results, but also because of the following limitations.
First, the accounting system is rigidly categorized. These categories
for a given firm may be perfectly appropriate for operating decisions and
financial summaries, but rarely are they fully appropriate to the needs of
economie analyses for longerterm decisions .
Another limitation of accounting data for obtaining estimates is the
misstatements embedded by convention into accounting practice. These are
\
\..)...>
1
l
:!
i
Il
1
r.l
'[
ll!o
106 Estimating Markets , Revenues, and Costs Chap. 5
based philosophically on the idea that management should avoid overstating
the value of its assets and should therefore assess them very conserva
tively. This leads to such practices as (1) not changing the stated value of
one's resources as they appreciate due to rising market priees, and (2) depre
ciating assets over a much shorter fife than actually expected. As a result of
such accounting customs, the analyst should always be careful about treat
ing such resources as cheaply (or, sometimes, as expensively!) as they might
be represented.
The final limitations of accounting data are its illusory precision and
implied authoritativeness. Even though it is usual to present data to the
nearest dollar or cent, the records are not nearly that accurate in general.
Further, the results are often affected by arbitrary allocations and optional
charges and/or credits.
In summary, accounting records are a good source of historical data,
but they have severe limitations when used in making estimates for eco
nomie analyses. Further, accounting records seldom directly contain the
marginal costs, especially marginal opportunity costs, appropriate for eco
nomie analyses.
5.4.1.2 Other Sources within the Finn
The usual firm has a large number of people and records which may be
excellent sources of estimates or information from which estimates can be
made. Colleagues, supervisors, and workers can pro vide insights or suggest
sources that can readily be obtained.
Examples of records that exist in most firms are sales, production,
inventory, quality, purchasing, industrial engineering, and personnel. Table
54 provides a list of the types of data that might be needed for cost esti
mating purposes, together with typical sources (mostly intrafirm) for the
data.
5.4.1.3 Sources Outside the Finn
There are innumerable sources outside the firm that can provide infor
mation helpful for estimating. The main problem is to determine which
sources potentially are most fruitful for particular needs.
Published information such as technical directories, trade joumals,
U .S. government publi cations, and comprehensive reference books offer a
wealth of informati on to the knowl edgeable or persistent searcher.
T'crsonal contacts are excell ent potenti al sources . Vendors, snlespco
plc, profcss ional ncquaintnnccs, customcrs, bankll, govcn, mcnt fi KCncics ,
of commci'Cc, nnct vvcn compctil ors nrc ofl lln wlll lnA to
dt1d lui<H 11 11 ' 1ion 1r til t\ ICiiHi.lNt 111 t<crlnoN 11 11 d lul tf\11,
Sec. 5.4 Cost Estimating: Perspective and Approaches 107
Description of Data
General design specifications
Quantity and rate of production
Assembly or Jayout drawings
General tooling plans and list of pro
posed subassemblies of product
Detail drawings and bill of material
Test and inspection procedures and
equipment
Machine tool and equipment require
ments
Packaging and/or transportation require
ments
Manufacturing routings and operation
sheets
Detail tool, gage, machine, and equip
ment requirements
Operation analysis and workplace studies
Standard time data
Material release data
Subcontractor cost and de li very data
Area and building requirements
Historical records of previous cost es ti 
mates (for comparison purposes, etc.)
Current costs of items presently in pro
duction
Table 54
TYPES AND SOURCES OF COST ESTIMA TING DATA
Sources
Product engineering and/or sales depart
ment
Request for estimate or sales department
Product engineering or sales department
or customer' s contact man
Product engineering or manufaciuring
engineering
Product engineering or sales department
Quality control or product engineering or
sales department
Manufacturing engineering or vendors of
materials
Sales department or shipping department
or product engineering (government
specifications)
Manufacturing engineering or methods
engineering
Manufacturing engineering or material
vendors
Methods engineering
Special charts , tables, time studies, and
technical books and magazines
Manufacturing engineering and/or pur
chasing department or materials ven
dors
Manufacturing engineering and/or pur
chasing department or customer
Manufacturing engineering or plant
layout or plant engineer
Manufacturing engineering or cost de
partment or sales department
Cost department or treasurer or comp
troller
Source: Reproduced by permission of ASTME, Manufacturing Planning and Esrimaring
Handbook, McGrawHill Book Company, New York, 1%3, pp. 320.
5.4.1.4 Research and Development
.._
t
,1
i
!
"i
10!1 EstimatIII( Markets , Reue11ues, and Costs Chap. 5
are very important estimates to be made and when the sources mentioned
above are known to be inadequate.
5.4.2 Estimating Methods
Severa! methods of estimating exist. However, we limit our discussion
to the (1) conference, (2) comparison, (3) unit, and (4) detailed anal y sis
methods. The first three methods are preliminary techniques for projects
that have not been designed yet or that are in the design phase. The detailed
analysis method is used for preparing final estimates .
5.4.2.1 Conference Method
In the conference method, representatives from different functions in a
firm work together to estimate the various manufacturing costs (and reve
nues) that are needed. Primary advantages of this method are (1) its speed
and (2) the fact that it causes people with a diversity of backgrounds to focus
on the estimate(s). It is also useful when a company does not havea formai
estimating depart ment. The main disadvantage of the conference method is
th at it may lack accuracy.
5.4.2.2 Comparison Method
In the comparison method an accumulation of historical data and expe
rience is utilized. According to a publication by the Society of Manufactur
ing Engineers (SME): "The estimator applies uptodate costs derived from
si milar parts to the project, adjusting these costs to suit material, labor, and
processing variations. Care should be taken when using data from larger or
smaller manufacturing quantities. For instance, smaller production lot sizes
will usually increase setup costs per item in the lot. Converse! y, Iabor costs
normally decline with increased efficiency developed over longer production
runs" [15] . The comparison method , like the conference method, is used for
quick estimates.
5.4.2.3 Unit Method
The most popular of the preliminary estimating methods is the unit
method, which iovolves using an assumed or estimated per unit factor.
Sorne examples are:
Capital cost of plant per kilowatt of capacity
Fuel cost per kilowatthour generated
Capital cost per installed phone
Revenue per longdistance cali
ost Estimating: Perspective and Approaches 109
Operating cost per mile
Maintenance cost per day of use
These factors may be multiplied by the appropriate unit to provide the
total estimate. The following examples may be used for breaking quantities
to be estimated into units that can be estimated readily:
1. In different units (Example: dollars per week, to convert to dollars per
year)
2. A proportion instead of a number ( Example: percent defective , to
convert to number of defects)
3. A number instead of a proportion (Examp/e : number defective and
number produced, to convert to percent defective)
4. A rate instead of a number (Example. : miles per gallon, to convert to
gallons consumed)
S. A number instead of a rate (Example : miles and hours traveled, to
convert to average speed)
6. Using an adjustment factor to increase or decrease a known or esti
mated number (Example: defectives reported, to convert to total de
fectives)
Although the unit method is useful for preliminary estimating, the val
ues can be very misleading because there is no consideration of the principle
of economies of sc ale or economies of scope .
5.4.2.4 Factor Technique
The factor technique is an extension of the unit method in which one
sums the product of one or more quantities or components involving unit
factors and adds these to any components estimated directly. That is ,
c = 2: cd + 2: Ji x V;
whre C = value (cost, priee, etc.) being estimated
Cd = cost of selected components estimated directly
fi = cost per unit of component i
U; = number of units of component i
(519)
Suppose that we need a slightly refined estimate of the cost of a hou se consisting of
1,500 ft
2
, two porches, and a garage. Using unit factors of$40/ft
2
, $2,000/porch, and
$3,000/garage, we can calculate the estimate as
$40 x 1,500 + $2,000 x 2 + $3,000 = $67,000
1 1
1
' V'
110 Estimating Markets, Revenues, and Costs Chap.5
5.4.2.5 Detailed Analysis Method
The d.etailed analysis method is the most reliable form of estimating
and is normally performed by one estimator (with input from other knowl
edgeable people as needed) . The following steps should be followed in pre
paring a de tai led analysis estimate [ 15]:
1. Calculate raw material usage, including scrap allowances and salvage
material (direct material).
2. Process each individual component (write the operation sheet).
3. Compute the production time (direct labor) for each operation.
4. Determine the equipment required (new, rework, or on hand) .
5. Determine the required tools, gages, and special fixtures .
6. Determine any additional equipment needed for inspection and tes ting.
Even though the conference and comparison methods are quicker,
most manufacturers use a detailed analysis for a preliminary estimate. It
provides the most accurate results and may be used again as the final esti
mate aftei revisions are made.
5.4.3 Labor Analysis
Labor is one of the most important expenses to be considered in cost
estimating. To provide an estimate of tabor costs, the amount of lime in
volved must first be measured. Four methods of measuring labor time in
elude standard time studies, Iaborhour reports, work sampling, and prede
termined time systems.
5.4.3.1 Standard Time Studies
' 'A standard ti me may be defined as the time required by an average
worker to complete a given task while working at a normal pace" [10) .
When a job is considered to be repetitive, the stopwatch procedure may be
used to obtain this standard time . The following seven steps are involved:
(1) divide the task into elements; (2) observe and record time to complete
each element each time it is performed; (3) determine the pace (rating) at
which the work was performed ; (4) determine allowances for personal
needs, fatigue, and unavoidable delays; (5) deJete observed times that are
not performed under normal working situat ions or conditi ons; (6) determine
the normal ti me; and (7) calcul ate the standard time by adding all owances to
the nornu11 timc. Thus
Tl'\ NT, 1 (N'l' ' docl11 ml) (520)
Sec. 5.4 Cos/ Estimating: Perspective and Approaches
where TS
1
= standard ti me for work element j, j = 1, 2, ... , rn
NT; = normal time for work element
rn = number of elements
Ill
The standard times for ali elements are then summed to provide the
total standard time, TS, per unit. lfTS is in minutes, picccs pcr hour may be
calculated as 60/TS.
5.4.3.2 LaborHour Reports
Laborhour reports may be used as a basis to determine tabor hours for
nonrepetitive work or when time studies are not available. These are merely
summaries of hours worked laken from job tickets, with whatever adjust
ments are appropriate for estimating the times desired.
5.4.3.3 Work Sampling
Work sampling is another method of obtaining tabor hours for nonre
petitive work. This method does not require a stopwatch or historical re
ports and is convenient and inexpensivc . According to Ostwald:
A work sampling study consists of a number of observations taken pertaining
to the specifie activities of the person(s) or machine(s) at random intervals.
These observations are classitied into predetined categories directly related to
the work situation. Du ring the course of the work sampling study, tally marks
are made by the technician, such as "work," "idle," or "absent." The key to
the accuracy is the number of observations, which may vary according to the
requirements. [Il, p. 42]
Work sampling is a statistical technique ; therefore, probability theory
must be utilized to plan for or describe accuracy of results. Work sampling
statistics generally are based on the binomial distribution in which the mean
is equal to Np; and the variance is equal to Np;(l  p;), where N = number
of observations and p; = probability thal event i will occur. Since N is
usually large, this bino:nial distribution can be approximated by a normal
distribution.
To begin the process, N ' observations must be made so that p; may be
estimated:
Ni
Pi = N'
(521)
where pi = obscrved proportion of occurrence of an event i expressed as a
decimal
N; initinl 11111111
N' luitl nl 11111111
b.s orvutionR of event 1
l llfldlllil llhNCI VIl {
1 
'\
::
...
,,
:.
'
:r.
.f
.,
:
. '
k
: j
il
l '
\!
1;
11
L
'
11
r
1
1
" j
11 2 Estimating Markets, Revenues, and Costs . Chap. 5
Once p; has be en calculated, the estimator must decide on the maximum
desirable interval, /, as a measure of the accuracy of the estimate of p
1
Suppose that p' = 50%; if a maximum interval of 2% is desired, the estimate
of p ranges from p' 1/2 top'+ 112 (49 to 51%). The estimator must also
decide on the confidence leve!. A confidence levet of 90% is common for
worksampling studies. From a normal distribution table (see Appendix D),
for a confidence leve! of 90%, Z = 1.645. Other values of Z (number of
standard deviations for a standard normal distribution) for different confi
dence levels are listed in Table 55 .
The number of observations of event i necessary to achieve the desired
confidence lev el and interval may be found from the following equation [ 11]:
N = 4Z2p;(J .,... p;)
1 [2
(522)
Once these observations have been taken, the labor hours may be calculated
as
H = (N;IN)HR(l + PF&D)
' Np
where H, = standard labor hours per job element i
N
1
= total number of event i observations
N = total number of observations
H = total tabor hours worked during study
R = rating factor
PF&D = allowance decimal
(523)
NP = work units accomplished during period of observing this event
Example 54
Suppose that it is desired to know H, for machining a part using work sampling. A
confidence of 90% and an interval of 4% have been chosen, and 125 times out of 150
Table 55
Z VALUES FROM STANDARD NORMAL DISTRIBUTION TABLE
Area Between Number of
Limits Standard
(Confidence Level) Deviations Area Outside
(%) z to +Z Limits (%)
68 :!:1.000 32
90 :!: 1.645 10
95 :!: 1.960 5
99 :!:2.576
Source: Phillip F. Ostwald, Cos/ Estimating, 2nd ed.,
PrenticeHall, lnc., Englewood Cliffs, N.J., 1984, pp. 43
44. Reproduced by permission of the publisher.
... ,,..,.  ....  .. ,_.
Cost Estimating: Perspective and Approaches
113
initial observations the operator has been busy. From these values we canuse Eqs .
521 and 522 to find the number of observations necessary to achieve the desired
confidence levet and interval :
p; = = 0.8333
N = 4(1.645)2(0.83333)(0. 1667) =
940
1 (0.04)2
It turned out th at a total nu rn ber of observations, N = 1 ,262 were needed to
observe 940 in which the operator was busy. The opera tor worked at a rate of about
115%; therefore, R = 1.15. The PF&D allowance factor is 0. 15. The study was done
over a period of H = 6 hr, in which 20 parts were machined. From Eq. 523, the
standard time in hours per part is
H = (940/ 1 ,262)(6 hr)( 1.15)( 1 + 0. 15)
s 20
= 0.295 hr per part
5.4.3.4 Predetennined Time Systems
A predetermined time system is defined to be "an organized body of
information, procedures, and techniques employed in the study and evalua
tion of work elements ... " [10, p. 34] . Such systems are often preferred to
time studies because they are considered more accurate and because they
are generally more acceptable to workers . Numerous systems exist, but
MTM (methods time measurement) appears to be the most widely used.
Malstrom describes MTM as follows:
It assigns times for the performance of elemental motions that a re completed
by the hands, arms, and other parts of the human body. Times required to
complete different types of motions are assigned specifie values. Ti me values
are expressed intime measurement units (tmu). One tmu is assigned the value
of hr. or 0.036 sec. [10, pp. 3435]
lt should be noted that extensive training is required to use MTM with
competence.
5.4.3.5 Leaming Curves
For manufacturing in which a large amount of direct tabor is involved,
the average tabor time required to produce a unit is typically found to de
crease over time . As a general rule, each time the cumulative production
doubles, the total time required per unit is reduced by 4%. In this case when
unit time is plotted against cumulative production, the plot is known as a
96% learning curve. An 80% learning curve would show a 20% reduction
' 1
 ,"{j
1
114
Estimating Markets, Revenues, and Costs
Chap. 5
(i.e . . rate of improvement) in cumulative average labor ti me wh en produc
tion doubles. This simple mode! is an exponential function [8]:
Y;= y,;h
where Y1 = cumulative average direct labor hours or cost through the ith
production unit (thus Y10 would represent the cumulative aver
age ti me required for the first ten units, while Y
1
would represent
that for the first unit produced)
i = cumulative production count
b = the learning curve exponent
To solve for b, suppose that as cumulative production doubles from 20 to 40,
there is a 20% reduction in cumulative average Jabor hours per unit . Then
0.80 = Y4o = Y,(40)  h = 2h
Y2o Y,(20)  h
Thus. h log 2 = log(0.8), or h = log(0.8)/log 2. This can be solved to find
th at h = 0.3219. Fin ally, the equation Y; = Y
1
;
0
3219
represents an 80%
learning curve.
In general, the learning curve exponent, b, can be easily computed:
b = log of percent learning/log 2. For example, b = 0. 1520 for 90% learning
and b = 0.2345 for 85% learning. The following example shows the use of
the learning curve.
Example 55
One thousand units of a new product are to be produced. Direct labor has been
cstimated to average 8 hours per unit. Six months after production has begun, man
agement has requested the following information:
1. Average labor hours per unit, to date
2. Labor hours per unit, for the latest month' s units
3. The % learning curve thal production has followed . The cumulative labor
hours and units produced are listed in Table 56. Dividing the cumulative labor
hours by the cumulative production shows that the average labor hours per
unit to date is equal to 15.75 labor hours . For the latest month's production, an
average of 410/42 = 9.76 labor hours per unit was reached.
In examining the data, it can be seen that the cumulative production at the end
of mon th 6 ( 132 units) is double the cumulative production at the end of mon th 4 (66
units) . The learning curve can be approximated by dividing the cumul ative average
labor hours pcr unit during month 6 by the cumulati ve average labor hours pcr
unit during month 4. The lcarning curvc in effect is approxi motely 75%, as shown
low.
1' .. 1 t hli )\)1' li t HII
ll h,hn, 1111111
............... ...
a <
' f
.,.., <
"Cl
:0 z
"' 0
....
:>
0
0..
" ...
. _g '
j ;,.;.
E ..;, 5 Il
::l > 0
u..::::r:
... 
0 
.J::J c
..c;ro::l
;:..J;;,
0 . ...
.... Ol) ::l
""' > 0
" >
<::r:
"' u
"' 1o V'l
 0 ...
c .0 ::l
0 "' 0
::E..J::r:;
" c
,g
'(;
c
"' 0
..Ov.\Or"''f'V"'
,..... r ,..... 0\ \1"'1 r
NNNN
.oc.oor\GI ,......,..... __ oor
N t""l
 N
or..oVo
ONOV)OO
rl"''f'rlf"">N"''f'
=s r; 1  r ..,a f"'l '1:7" N
c: ::::s  t'l ,...,....,.
0'1j
::E 0
..c:
 1  N ..0
::E
, 1
i
:i
,.,
i
.. !
'
J
h(
j i
' i:
:(;
r;
1!
!
,,
,,
i1
ji
lj
1
.j
1
i
j ,
i'
1
1:
1
:
j:
1
1
l
i:
t:
1
11 6 Estimatin11 Markets , Revenues, and C:os ts Clwp. 5
To determine the equation for this leaming curve, the parameter b must be
calculated. This can be accomplished by applying equation 525 and solving for b, as
shown below.
Reducing and solving, we find :
Y,( 132)  h
Y
1
(66)  b = 0.75
2 b = 0.75
b log 2 = log 0.75
b = 0.415
Th us the equation is Y; = Y, ; o.l5
5.4.4 Capital Cost Analysis
Frequently, the larges! cost of a project is attributable to new equip
ment, buildings, and in sorne cases, entire production plants. Severa! meth
ods for estimating capital costs are discussed below.
5.4.4.1 Analogy Costing
Normally, the information necessary for a detailed estima te of capital
costs is not readily available du ring the initial stages of a project. Therefore,
the first estimate is usually fairly crude and uses analogy costing. Analogy
costing is based on similarity among production plants with known capital
costs. Equipment for which a cost estimate is desired is located at an exist
ing installation, and by analogy an estimate is prepared for the new equip
ment. ln this situation differences in cost from the existing to the new equip
ment would typically arise because of cost inflation and/or productivity
improvements . However, these two influences on cost have a tendency to
counteract each other in such a way that two projects, whose installations
are separated by only a few years, may have estirrated costs that are approx
imately the same. given that the existing plant was not grossly over or
underdesigned. Other factors must be considered when the same plant is to
be built in a different location, such as different delivery charges, different
site preparations, and different costs of supplementary equipment and facili
ties to support operations [4] .
5.4.4.2 Exponential Costing
Exponential cos ting may be used when the proposed plant has a differ
ent production capacity than the existing plant. According to de la Mare,
" the principle of exponential cos ting states that for many realworld produc
tion systems proportionate increases in production capacity can be achieved
by Jess than proportionate increases in capital cost. This principle is a spe
.>tC. 5.4 Cost Estimati1111 : Perspective a11 d Approaclr es 11 ,
cial manifestation of the law of increasing returns to scale, and is known as
the law of increasing technical returns to scale" [4, p. 151].
The following general equation represents most types of equipment:
c. = (Q)Il (526)
cb Qb
where Ca = capital cost of the proposed facility
Cb = known capital cost of an existing facility
Q. = production capacity of the proposed facility
Qb = production capacity of the existing plant
f3 = cost exponent factor, which can range from 0.4 to greater than
1.00, but is usually in the range 0.5 to 0.8
Table 57 provides typical cost exponent factors for selected types of indus
trial equipment. Equation 527 permits one to obtain an estimate of the
capital cost of a proposed project by including factors to adjust for the
effects of inflation and productivity differences as follows :
(
Qu)(J
Ca = Cb Qb 1,1,.
(527)
where 1, = index of cost inflation
lp = index of productivity improvement
The accuracy of the exponential costing method depends largely on the
similarity between the two projects, and the accuracy of the cost exponent
factor {3. Generally, error ranges from 10 to 30% of the actual final cost .
A certain steamgenerating boiler in the utility plant of a manufacturing complex
produces 50,000 lb/hr of saturated steam. This boiler was purchased for $250,000
eight years ago. If the priee index for this type of boil er has increased at an average
rate of 12% per year for the past 8 years, and productivity improvements have
averaged 4% per year du ring this lime, how much would a 150,000lb/hr boiler cost
now? The cost exponent factor for this boiler is 0.50.
Utilizing Eq. 527,
(
Q.)Il
c. = cb Qb 1,1p
with /
1
= (1 + 0. 12)
8
and lp = (1  0.04)
8
, the cost of the proposed boiter can be
estimated as follows:
c. == $250,000 [ ( ~ ; ~ , ; : ) 0.
5
( 1. 12)
8
(0. 96)8]
= $250,000[(3.00)
0
1
(2.476)(0.7214)]
= $773,441
~ . . . . . ! : . : )
~ ~
. ~
~
t
~
4
$
IIH
rcxt
"' "'
1 <
t.L.
....
z
w
!;;
0
u
...!
<
u
0::
>
1
"' E
Q.
.;:;
cr
w
c;;
::>
""' .5
u
c
"' Q
Q.
:;
3"
..
= :;;;
c
"'
:J;=
c;;
c
"'
'"
:::2
"' E
Q.
:;
cr
w
Ol
'
::>
""' .5
"'
"'
"'
2
0..
"' Q()
cxtl ": "'"' "f ""!"' ""! "f ooo..,oo o,.....
E
CXl.
E
0 d
 V)
0 u ...
Vl 0 V) t'
gtl ot
ct.E:S
Eu u..ocouV)
Vl 0
u c t":S:!:! .s
.... = e c "' u :; Cil 1
<<uCwciia.l
00 (7.
oo,....Oo..:t"O
1 1
00\C)OO,...._
0 d
t;
c
:.2 _g
_.,:.
E "'
00
0UU'"'
gg i: i; t 2:
t":S0000
a.lUUUW:I:
V)r<"''r0\0\0\
o=ooooV)
000000
E
:;
"' 01)
01) c
"' c "' (1) .... ....
u 0 ..= u
.:: c "' "'  E c
N"
c
0
""'
c
j
ci
1
....l
Ol
c
.s
'"
u
::>
o
w
::::1
"' u
"'
"
.,
!; ..:
" "'  .=
., "'
"'"" "'>J:J
""::>
"' Q.
.s
"'
::s c
"' 0 t:l v;
. E
"' ...
... "'
"' Q.
:::2"'
.D
"' "' u
""' ::>
fr
.
""' '"'
::s
" .
"" Q.
Sec. 5.4 Cost Estimating: Perspective and Approaches 119
5.4.5 Materials Analysis
Depending on the product, materials can be a very large, and therefore
important, part of a cost estimate. ln many industries it is common for
materials costs to comprise 50% or more of total product cost. Ma teri al cost
analysis involves making and using a bill of materials, deciding whether to
make or purchase parts, measuring the amou nt of raw material that goes into
each unit , and finding the costs of parts and raw materials.
5.4.5.1 Bill of Materials
The bill of mate rials is a complete listing of ali parts (mate rials, su bas
semblies, purchased items, etc.) going into the final prodpct. lt also lists the
part number, the quantity required to produce one unit, the material specifi
cations , whether the part is purchased or made, and perhaps other products
in which the parts are used as weil as a crossreference to similar parts. This
information is vital to the cost estimator so that an accurate estimate may be
made.
. 5.4.5.2 Decision to Purchase or Make
In most instances the decision to purchase or make parts is determined
by considering the incrementai costs of each option. That is, parts will be
purchased rather than made when the purchase priee is Jess than the incre
mentalcost to manufacture. There are situations, though, where it is desir
able to select the more expensive option . Suppose that a part can be pur
chased for Jess than the cost of inhouse manufacture, but the vendor cannot
guarantee delivery when the parts are needed. ln this case the part probably
should be made inhouse to avoid the possibility of missing the due date.
Furthermore, sometimes a part may not be readily available for purchase
without risking the Joss of competitive information .
5.4.5.3 Estlmates of Raw Materials and Purchased
Parts Costs
Once the estimator has determined which parts are to be purchased
and which are to be manufactured, the next step is to obtain cost estimates
for both raw materials and purchased parts . According to Malstrom:
In estimating material costs, severa! factors about the part to be purchased
should interest the cost esti mator. Exampl es of factors to be considered in
elude the quantit y to be purchased, whether or not the part has been purchascd
beforc, and the lcnsth of tin1C requircd to obt11ln 11n occumtc priee quOllt lon.
(10, p. SS I
l lll) (l fC IH ll y 11111 111( II,U,' 'IIIL
11 n oll on ktpltlll ll l OIIIJ'IIIt' l ill (l litiiiiiii' Y
'J , .
:;
i
1
120
Dale
3/84
9/85
8/86
12/87
Estimatinf( Markets. Revenues, and Costs
Table 58
HISTORICAL DATA
Order Quantity Unit Cost
250 1.65
1.000 1.25
160 1.!!7
730 1.72
are easily accessible. Table 58 provides a list of historical data. First, the
estimator picks two order quantities that bracket the order quantity de
si red . Suppose that the desired order quantity is 424; he or she must then
choose the quantities of 250 and 730. If today's date is 6/88 and the average
inflation for this material has been 1% per mon th du ring the last 4 years, the
unit cost today of an order size of 250 would be*
V25o = $1.65(1 + o.ol)
11
= $2.74
For a lot size of 730, the unit cost today would be
u730 = $1.72(1 + o.Ol)
6
= $1.83
Interpolation can then be used to obtain a unit cost for an order quantity of
424 unit s. If we assume thal the relationship is linear, it can be solved as
(
730  424)
u424 = $1.83 +
730
_
250
($2.74 $1.83) = $2.41
There are severa! problems with this method. For example, inflation
rates do not usually remain constant. As a result, this method is recom
mended only for parts with a low unit cost.
When the unit cost of a part is high, it is normal to obtain at least one
priee quotation even if the part has been purchased before. A priee quote
from a manufacturer's catalog is acceptable if it is up to date. "Substitution
is acceptable as long as the priee of the part may be obtained from a catalog
in a period of time equal to or Jess than that required to obtain a quote" [10,
p. 56]. If the dollar value of the order is Jess than the cost to ob tain the
estimate, the estimator may make an educated guess, use a manufacturer's
catalog, or use historical data to obtain a reasonable estimate. In instances
where a priee is not readily available, a part number match may be used as a
basis for estimating. The estimator should also be aware that priee breaks
may exist in the purchase of parts. If the company is willing to purchase
sufficient quantity of a part at one ti me (perhaps by combining requirements
See Section 5.5, Eq. 537: (1 + 0.01)" = (F/ P, 1%, 51), which merely adjusts "RS"
into "A$" by compounding inflation at 1% per mon th for the 51 months from 3/84 to 6/88.
.': '...,     __ _ .._
Sec. 5.4 Cost Estimatinf( : Perspective and Approaclzes 121
for numerous products), there could be a significant difference in the per unit
cost.
5.4.6 Other Costa
The estimator should recognize that costs for working capital, setups,
run cycles, tooling, and factory burden need to be considered when relevant.
5.4.6.1 Working Capital
Working capital is the moncy required to transform a new capital
project into an operating process, excluding fixed assets (such as equipment,
machinery, or buildings). A common accounting definition is:
working capital = current assets  current liabilities (528)
Table 59 shows an example of current as sets and liabilities for a company.
Using equation 528, working capital= $4.35 x 10
6
 $1 .2 x 10
6
= $3 . 15 x
W. Although working capital is constantly being liquidated and regen
erated, it is not available for other uses, so it must be regarded as an in
vestment.
5.4.6.2 Setup Costs
Setup is the work required to prepare the machine , process, or bench
for operation. Normally, the setup ti me can be fou nd by using standard
data, but a study of the elements necessary to perform the setup may be
required. These elements may include operator punching in or out, obtain
Table 59
A COMPANY's ASSETS AND LIABILITIES
Fixed assets
Current assets
Cas h
ln ven tory
Accounts receivable
Current liabilities
Accounts payable
Shortlerm debt
Working capi tal
Total assets employed
$ 50,000
3,300,000
1,000,000
$4,350,000
$ 700,000
500,000
$1,200,000
$ 8,000, 000
$ 3,150,000
$11 ' 150,000
Source: R. F. de la Mare , Manufacturing Sy.rtem.r Eco
nomie.<, Cassell Educational LTD.. London. 1982,
p. 167. Reproduced by permission of the publisher.
1
!(
1;
1
''
,1
:f
t
\
1
;
,.
! ' .
1'
122 Estimating Markets, Revenues, and Costs
ing tools, doing paperwork, making machine adjustments, and positioning
necessary materials. Also included in the setup are tasks performed after a
run cycle is completedsuch as teardown, returning tooling, and cleaning
workstation. Normally, the setup cost is handled as an overhead charge for
mass production [Il].
' 5.4.6.3 Cycle Time Costs
Cycle time is the time required to produce 1 unit of a product after
setup has been completed. Depending on the process, there are many dif
ferent methods of calculating the cycle time of a machine. As an example,
the steps in a metal cutting operation are ( 1) load work, (2) ad vance tool, (3)
retract tool, and (4) unload work.
The time to load and unload the workpiece is called the handling lime
and may include the time to ad vance and retract the tool from the piece. The
handling cos/ may be found by
handling cost = Coth
where C
0
= direct labor cost, dollars/min
th = time for handling, min
See Fig. 57a as an example of handling cost plotted against cutting speed.
To find the time to machine 1 unit, the following equation is used [10,
p. 270) :
L LrrD
tm = JN = 12 Vf
where tm = machining time, min
L = length of eut for metal cutting, in
f = feed rate, in/rev
N
. d 12 v 1 .
= rotary cuttmg spee = rrD , rev mm
D = diameter, in
V = cutting speed, ft/min
machining cost = Colm
The unit cost decreases as cutting velocity increases, as shown in Fig.
57b. Note that the diameter, D, may be for either the tool or the workpiece,
depending on the type of machine. The diameter of the tool is used for
milling and drilling; in lathe turning, the diameter of the workpi ece is used.
5. 4.6.4 Costs
' utti l1fl, 11111y hilC(lll l uull, 1\ nd thc r c f () lt' 11\IINI h ll 0
l i' PhH' tl d ll \i' t ol of' Jlplll l.' l ll llll l n 111 11(1 luiJHHiillll l11 Il
tlt' l ll' ld l nll 11 11 lll t'll ll llll llll il l llthlll 1111d l l ti ii' I I'!JII IInl Atillld
Sec. 5.4 Cost Estimating: Perspective and Approaches
c:
::1
..
o.
Tool changing
Cutt ing speed
le)
:
::1
8.
il
u
Cutting speed
(d)
Figure 57. Graphie costs for four parts of machine turning economies . Source:
Ostwald, Phillip F. , Cosr Esrimaring, Second Edition . Eng)ewood Cliffs. NJ :
PrenticeHall , lnc ., 1984, p. 271. Reproduced by permission of the publisher.
123
ing to Ostwald [Il], the aue rage /ife of a cutting tool may be fou nd by the
T = K""
v
where T = average tool !ife, minutes per cutting edge
(532)
n, K = empirical constants resulting from regression analysis and field
studies 0 s n ::5 1, K ;:::: 0
When V = 200ft/min and the tool life equation is VT
0
16
= 400, then 76
min of machining may be done before the tool must be indexed to a new
The toolchanging cost per operation may be found by
toolchanging cost = Co te
(5 33)
.
Il
1
,.
i
,\
124 Esrimaring Markets , Revenues , and Cosrs Chap. 5
Figure 57c shows the relationship of toolchanging cost to cutting speed.
The cast of cutting tools must also be calculated:
tool cost per operation = C
1
(t;) (534)
where C, denotes the tool cost in dollars and T is the average tool !ife in
minutes . Ostwald observes that "for insert tooling, tool cost is a function of
the insert priee, and the number of cutting edges per insert. For regrindable
tooling, the tool cost is a function of original priee, total number of cutting
edges in the li fe of the tool , and the cost to grind per edge. " As shown in
Figure 57d, the tool cos! increases as cutting speed increases [Il , p. 272].
5.4.6.5 Factory Burden
Ali costs that cannot, conveniently and economically, be allocated to a
particular product are called factory burden or ouerhead. Burden is typi
cally divided into two categories: fixed and variable. Fixed burden consists
of costs that do not change with production volume, such as salaries , mort
gage payments, rent, and insurance. Variable burden increases or decreases
as production volume rises or falls . Examples include electricity to operate
equipment. indirect materials and labor , and certain tooling. Often, compa
nies allocate burden to different cost centers and/or products in proportion
to sorne activit y basis such as direct labor hours or prime (tabor + materials)
cost.
5.4.7 Engineering Costa
When the product to be produced is similar to an existing product,
engineering costs can be estimated by using historical records. The estimate
must include costs of basic engineering services such as product design,
specifications, and bill of materi als. as weil as maintenance engineering for
the li fe of the product. The costs are made up of wages, housing, and utili
ties. In addition, costs for projects that did not materialize into final prod
uct s must be considered in the wage costs [Il) . Spreadsheets su ch as th at
shown in Fig. 58 may be used in figuring the cost of tabor for engineers.
When representing engineering rate per unit , C,, the following equation may
be used:
C = total engineering cost
projected production quantity
(535)
where the " production quantity" can be in any relevant units , which are
often " hours worked ."
Sometimes factors are used to multiply by expected salaries to provide
an estimate. They can range from 1.8 to 3.0 and depend on the complexity,
Sec. 5.4 Cosr Esrimaring: Perspective and Apprnaches 125
Design engineering estimate for Speed 1t educe1t producn
Description
6olt te6.t w eed
Customer
No.t
lnquiry or quota no. Date
Based on quantity of
480 u>t.U&
Ouri ng period of
Type of labor Hours
Rate Extended
per hour la bor
Suei'I..I>.t ,
de.g n 40 $33 . 00 $1. 32 0
Engwee.lt , 800 $30 . 00 $24,000
e.lec..tM.ca.t
a.i.M 160 $IL 75 $3. 000
Tech. Wlt.i..telt
1
80 $1 7 . 00 $1,360
Mode.l 6hop
To.ta.t labolt 1080 $29 ,680
    
Figure S8. Estimating for engineering cos ts. Source: Ostwald , Phill ip F., Cmt
Esrimating, Second Editi on. Englewood Cliffs. NJ : PrenticeHall , lnc. , 1984,
p. 318. Reproduced by permission of the publi sher .
1
novelty, and sec recy of the work. Other factors, su ch as drawing size, are
a!so being used to estimate engineering design costs .
5.4.8 Comprehensive Cost Estimating Example
The XYZ company has decided to btlild a new plant that is very similar
to an existing plant. The plant will be used to produce 5,000 units of the part
shown in Fig. 59. The loss due to scrap is expected to be 2.5%. The com
pany that sells the raw materials to the XYZ company will buy back any
scrap or waste for $0.95 per cubic inch . The recovery cos! is $0.20 per cubic
inch. Approximately 15% of the waste is not recovered for resale. Histori
l.
1
0:
126
Estimating Markets, Revenues, and Costs
0.25in. dia.
T
"" I
0.66in. dia.  +.,,'.
0.25in.
Figure 59. Sample part .
cal data show that on January 1, 1987, the cost of this material was $1.45 per
cu bic inch, in a quantity of 8,000 inches
3
On March 1, 1988, the unit cost of
3,000 inches
3
was $1.89 per cubic inch . For the past 4 years we will assume
that inflation rates have remained stable at an annual rate of 12% and that the
date wh en the material is to be purchased is February 1, 1990.
The approximate capital cost of the existing facility which was pur
chased recently was $12,000, and its production capacity is 20 units per
day. It is desirable for the future facility to have a production capacity of 30
units per day. Capital cost can be estimated using Eq. 527 with /, = 1.12,
lp = 1. 10, and (3 = 2/3. Studies were do ne at the existing facility showing thal
approximately 0.35 hr is necessary to produce each unit. This time was
achieved over a production of 3,500 units and production has followed an
85% Iearning curve. Labor costs $15 per hour, including benefits. The parts
are produced in lot quantities of 25 and the machine setup time is 1 hr. Part
of the setup involves changing the cutting tool. The cost of the tool is $20.
Factory burden is 40% of direct material. Estimate the manufacturing cost
from the information given .
Solution
The actual volume of material required per part is calculated to be
1.5375 inches
1
. The total material necessary to produce 5,000 units is there
fore (1.5375 inches
1
)(5,000) = 7,688 inches
3
. The amount of m'aterial avail
able to sell back consists of both scrap and waste. The amount of scrap per
parti s ( 1.5375 inches
3
)(0.025) = 0.0384 inches
3
The amount ofwaste (from
dri ll ing holes) is
( 1 
n  k)
(537)
where fis the average inflation rate per period over the n periods.
By ''inflation rate" is meant the average rate of inflation for the partic
ular items over the period in question . Sorne of the most valuable sources of
lnterest factors and symbols are explained in chapter 6. Factors are tabled in
Appendix A.
Sec. 5.5 Cmt Estimating Relationships and Inflat ion 129
data for estima ting inflation rates (either general, such as cost of living, or for
specifie commodities and services) are cost indexes. Cost indexes are di
mensionless nu rn bers th at provide a means for converting past costs to costs
at any later time, or vice versa.
There are many cost indexes and they cover almost every area of
interest. Sorne are based on national averages; others are very specialized
by type of item and perhaps locality. Indicative values for se veral frequent! y
used indexes are shown in Table 510. Probably the most commonlyused
measure of the general inflation rate is the Consumer Priee Index of the U .S.
Department of Labor . Note that "1year change" (i.e. , percent increase or
decrease for each year) in Table 510 is also shown for two of the indexes .
Additionally, the compounded an nuai rate of inflation for the 8year period is
shown at the bottom of the column for each index.
If we want to fi nd the cost of something at ti me n and know its cost as
of sorne (base) time k, then the relationship that can be used is
Cn ln
ck = "h
where Cn and Ck are costs and ln and h are the appropriate index values at
times n and k, respective! y. Solving gives
c. = ck (i)
(538)
Recognizing that Cn could be A$n, and Ck cou id be we can substitute
into Eq. 538 to find
A$n = i
Combining Eqs . 537 and 539, we find that
a;, ln
(F/P,f7o, n  k) = h
Example 57
(539)
(540)
Use the Consumer Priee Index in Table 510 to estimate the 1983 cost of products
that cost $1 million in 1980. Then find the annual compounded rate of inflation.
Solution
Using Eq. 538 yields
(
/83)
CsJ = Cso lw
$
. . (298.4) $ 209 . .
= ( 1 mtlhon)
246
.
8
= 1. mtlhon
\
'(..>
l '
.
.
1
'
i
iJ
; r
v
'
i
,,
i
1
li
q
' ' : i
!
;
1
i
!
i
'
il
1;
/;:
1
0"'
 w
' )(
.,., w
..!! 0
.Cl z
.. 
1f
"' 0
u
w
..J
...
<
)(
[JJ
0
'"
0::
8
"
..
.
0..
z
0
><
"
" .::
"
<.)
c
0..
" E
c
0
u
"
00
c
"'
>;
8
 "'
Il
N C
,_
a
c
"'
"'
"
>:'
8
 ><
Il
u 0.. "'
E u
ro:Orc
"'"'r
ou r.n Cl\
> 0 0
<u
"' u
c .. :
v;
0 ::J
6 8
0 a.."'
..... :::J "
0 ..  r c
V) r
r.l'l .... :l 0\
00
>o 0 "'
z
5
>
..........,..
O..:v\v\ro:oOO..:O..:..Or\
0\
v\NOOMov\c:v\
       C'l N
ooV"lrf"f""Vl<ot:tN
a\vl...or..:....:t'f"\o..or"")
NV""'V)"'=tVOO"'=t"':f' r
....:c:::)_:v)r:...0No\O oci
.crooQ\""':frooa..
NNNNN
00\0"'=t'V'lrV""'<ot:t...,.
.
NNNN
o8ooNOI""'V"'\O o
o\
g ti
'.:J
"' y
"''?>J:
?.
"'
1:!
"
.5
"
c
:::J
&.
E
0
<.)
E
e
(J.
....;
0
....;
0
c:i
00
0.0
c(
s
"' .; :::1
 0
0
.9 ]
<
Vi (,) 0
  (J.
.. 5 . E
E g .:.
"'0 8
;: < ;: UJ
o>.o""
=' 0.0 ::l L.o
1 .... 1 0
:::J [JJ :::J
Q_Q<
>o
..... OIJ ' 0
0 ,_ 0 
.Cl " .c> E
cu c "" 0
..J [JJ ..J c
'''0
0 0 0 <.)
   [JJ
c c c '
4.) u u 0
E E E :::
la la la g
a. o. o. ::J
0 u u 0
0 0 Cl u
vivi vivi
::i ::i ::i ::i
oo.o
....
.... "
';j'g.
.,,u
=
I'O
"'..,
00 .Cl
a 't:
c(<
 .,
0 c
,. K.
V'\ <+
$<
Cost Estimating Relationships and Inflation 131
Using Eq. 539, we obtain
298.4
(FIP,f%, 83  80) =
246
_
8
= 1.209
From compound iriterest tables, (FIP, 8%, 3) = 1.2.597 and (FIP, 6%, 3) = 1.1910 .
lnterpolating, the inflation rate for that 3year periO<I is approximately 6 . .5% com
pounded annually. It might be noted that the 6.5% is close to, but is not the same as,
the average of the 1year changes shown in Table 510 .
5.5.2 Real lnterest Rate, Combined lnterest Rate,
and Inflation Rate
Let us define severa! types of rates and show how they are used:
1. Real interest rate: Increase in real purchasing power expressed as a
percent per period, or the interest rate at which R$ outflow is equiva
lent to R$ inflow. lt is sometimes known as real monetary rate or
uninfiated rate and denoted as i, when it needs to be distinguished from
ic (below) .
2. Combined interest rate: Increase in dollar amount to cover real inter
est and inflation expressed as a percent per period, and is the interest
rate at which A$ outflow is equivalent to A$ inflow. It is sometimes
known as actual rate or infiated rate and is denoted as ic whenever it
needs to be distinguished from i, (above).
3. Inflation rate: As defined previously, the increase in priee of given
goods or services as a percent per period. It is denoted asf.
Because the real interest rate and the inflation rate have a multiplica
tive or compounding efTect,
ic = (1 + i,)(l + f)  1
ic = i, + J + (i, X f)
. ic f
1
' = 1 + f
Where fis not large relative to the accuracy desired, th en
ic !! i, + f
nd
(541)
(542)
(543)
(544)
1
,,
!
,
1
132 Estimating Markets , Revenues, and Costs
5.5.3 What lnterest Rate to Use in
Economy Studies
In general, the interest rate that is appropriate for timevalue calcula
lions in economy studies depends on the type of cash flow estimates as
follows :
Method
A
B
If Cash Flows Are
Estimated in Terms of:
Actual $, AS
Real$, RS
Then the lnterest Rate to Use ls:
Combined interest rate, i,
Real interest rate, i,
The above is made intuitively consistent if one thinks in terms of
method A as working with inflated (actual) dollars and interest , and method
8 being applicable to uninflated (real) dollars and interest. Method A is the
most natural to use because we usually think in terms of A$. Since interest
paid or earned is based on A$, it is a combined interest rate, ic.
method 8 is sometimes easier to use.
5.5.4 Summary of Formulas for Relating Single
Sum A$ and R$ over Time
A$n N S . .
(= A$) ote: needed only to clarify
base lime.
k Time scale n
Base time for ex pressing R$
Type of Dollars
Method or Conversion
A A$ (inflated $)
B R$ (uninflated $)
Moving Forward in Time
A$. = AS,(F/P, i " n  k)
= RS\"(F/ P, i,, n  k)
lnflating (at given time)
Moving Backward in Time
A$,= A$.(PIF, i " n  k)
RSI" =
i., n  k)
Detlating (at given time)
From R$ to A$, or AS. = n  k) = AS.(PIF.J, n  k)
from AS to R$ ,
at a given time
Example 58
A certain expense at the end of 1991 is estimated to be $10,000. The end of 1991 is
the base point for considering inflation. (Thus $10,000 = =
Find its
equivalent worth in 1996 for the following circumstances (paralleling the formulas
above) . and using the following rates .
Co.ft Estimating Relationship.f and Inflation
real rate, i, = 4%
inflation rate, f = 6%
combined rate i, = (1 + 0.04)(1 + 0.06) 
= 0.1024 = 10.24%
"" JO%
A$96 = ic%, 96  91)
= $10,000(FIP, 10. 24%, 5) = $16,282
(Note: If 10% approximation were used, answer would be $16,105. )
=
i,%, 96  91)
= $10,000(FIP, 4%, 5) = $12,167
In 1996 dollars, beginning with 1996 equivalent in 1991 dollars :
A$96 = 96 91)
= $12,167(F/P, 6%, 5) = $16,282
5.5.5 Manipulating Series Which Are Uniform
in R$
133
If cash flows expressed in R$ are uniform each year (and th us the A$' s
increase each year at the average rate of inflation, J), they can be conve
niently converted to equivalent worth(s) at other point(s) in time using uni
form series formulas at the firm's i, = real MARR.
Thus, if there is a uniform series in which each endofperiod pa y ment,
A, for n periods, is expressed in R$<kl (and th us is denoted A <k
1
), then
Phk> = (A<ki)(PIA. i" n) (546)
= (A<>)(FIA, i" n) (547)
The superscripts make clear the base point in time at which the dollars (like
present or future equivalents) are expressed. Normal! y k = 0, but estimates
can be converted to any base point in ti me at the inflation rate, f, using Eq.
536 or 537.
An equivalent formula for finding the present worth of a uniform series
that is escalating at the inflation rate f for n years is
(
1 ,
,
j
i
'
'
\!
1
i
1
'
1
1 '
1
Il\
134 Estimating Markets, Revenues, and Costs Chap. 5
p <o> _ A<O>(FIP,f, 1)[1 (PIF, ic%, n)(FIP,f%, n)]
0
 ic  J
Note that A<
0
>( FIP,f, 1) = A(l). Thus
p<O> _ Am[l (PIF, ic%, n)(FIP,f%, n)]
0
 ic  J
For the special case in which ic = f, so thal the real monetary rate = 0, Eqs.
548 and 549 become
= A<Ol(n)
and
P&
0
> = Am(PIF,f, l)(n)
Example 59
A person who is earning $21,600 salary for {assumed end of) year 1 expects that
saJary to inflate (escalate) at 10%/year, which is the same as the general rate of
inflation. If her real monetary rate, i,. is 5%, then i, = 0.05 + 0. 10 + 0.05{0. 10)"'
0. 155 , or 15.5%. Find the present worth, P&
01
for 3 years of salary.
(a) Use the approach of Eq. 546 to find P&
11
and then convert to Pb
0
>.
(b) Use Eq. 549.
Solution
(a) P&
11
= Alll(PIA, i,. 3)
= $21,600(PIA, 5%, 3) = $58,821
Pb
01
= Fb
1
J(PIF,J, 1)
= $58,821 (Pl F, 10%, 1) = $53,474
(b) p<o _ $21,600[1 (PIF, ic, 3)(FIP,f, 3)]
0
 i, f
$21,600[1 (PIF, 15.5%, 3)(FIP, 10%, 3)]
0. 155 0.10
= $53,496 [same as for part {a) except for roundoff error)
5.5.6 Manipulating Series That lnflate (Escalate)
at Rate Different from General Inflation
Whcn a cas h flow series expressed in A$ infl ates or escalates at a rate
different from general infl ati on. it wi ll not be a uniform seri es when cx
nrcsllcd in R$. The use of whot might be cnll cd "differenti ai oscolation
ITI ICII" Cll n he Il hnmly C(l!)lPIII OI () tllli COII VCfl ic nCC in ll ti Ch Il Cll llC, hui WC wil l
nol ll ln Ntt ltll' llt t nl h
Sec. 5.6 Summary and Prologue 135
eral ways to convert a series of cash ftows subject to whatever rate of
inflation (which might vary from year to year) into either R$ or A$, and then
to manipulate them (in this case, into present worths) at the general inflation
rate.
Given the same individual salary situation as in Example 59, the only difference
being that the salary will inflate (escalate) at 8%1year, which differs from the general
inflation rate of 10%. To repeat, the endoftirstyear salary A$
1
= R$\n = $21,600,
i, = 5%./ = 10%, and ic = 15.5%.
(a) Show her salary for 3 years expressed in R$ and in A$.
(b) Show the present worth (as of beginning of the tirst year) of both ways
of expressing this salary.
ln A$. :
Year n
2
3
ln
Year 11
1
2
3
( 1)
A$. = A$1(FI P, 8%, 11  Il
S21,600(FIP, 8%, 0) = 21,600
21,600(FI P, 8%, 1) = 23,328
21,600(FIP, 8%, 2) = 25,194
= AS.(PI F, 10%, 11)
S21,600(PIF, 10%, IJ = 19,637
23 ,328(PIF, 10%, 2) = 19,278
25 ,194(PIF, 10%, 3) = 18,928
(2)
(PI F, 15.5%, n )
0.8658
0.7495
0.6489
(PIF, 5%, 11)
0.9524
0.9070
0.8638
(3) = (1) x (2)
$18,701
17,485
16,348
$52,534
$18,702
17,485
16,350
$52,537
Note in the lower column {1) that even though her A$ salary is going up (at 8%
per year) , the R$ (purchasing power) of that salary is going down {approximately
10% 8% = 2% per year). The present worths of both ways of ex pressing the sa lary
are the same, except for minor roundoiT error.
5.6 SUMMARY AND PROLOGUE
ln 1hls chnpt cr w
dMIIJWCCiil'lllr y f'or
usscd toni es el u l eu lo ocvelonmcnt of
lhc lllld C(lSIS \l fii (IV!IO{' I: d OHIIHtfhl' lU
1
d
'.
'
l
l
136 Estimating Markets , Revenues, and Costs Chap. 5
ing systems . Specifically, we addressed forecasting the market for new/
modified products; estimating revenues and costs, including consideration of
learning curves; and dealing with inflation in the estimating process .
Once monetary benefits and costs of investing in new and innovative
technology are quantified , they need to be combined into economie
measure(s) of merit. This is the important subject of Chapter 6 (beforetax
analyses) and Chapter 7 (aftertax analyses, including replacement
studies) .
REFERENCES
1. Chambers. J . C., S. K. Mullick, and D. D. Smith, "How to Choose the Right
Forecasting Technique," Harvard Business Review, JulyAugust 1971, pp.
4574.
2. Corr, A. V., "The Role of Cost in Pricing," Management Accounting, Novem
ber 1974.
3. DeGarmo, E. P. , W. G. Sullivan, and J. R. Canada, Engineering Economy,
7 ed., Macmillan Publishing Company, New York, 1984.
4. de la Mare, R. F., Manufacturing Systems Economies, Holt, Rinehart and Win
ston, New York, 1982.
5. Draper. N. R .. and H. Smith, App/ied Regression Analysis, John Wiley & Sons,
!ne .. New York, 1978.
6. Gilchrist, W., Statistical Forecasting, John Wiley & Sons, lnc., New York,
1976.
7. Jones, B. W. , Inflation in Engineering Economie Analysis, John Wiley & Sons,
!ne. , New York, 1982.
8. Kleinfeld, l. H., Engineering and Managerial Economies , Holt, Rinehart and
Winston, New York, 1986.
9. R. C. Lenz, Jr., Technologica/ Forecasting, ASDTDR63414, Aeronautical
Systems Division, Air Force Systems Command, June 1962.
10. Malstrom, E. M. , What Every Engineer Shou/d Know about Manufacturing Cos/
Estimating, Marcel Dekker, !ne., New York, 1981.
Il. Ostwald, P. F., Cost Estimating, 2nd ed. , PrenticeHall, Inc., Englewood Cliffs,
N.J., 1984.
12. Porter, M. E. , Competitive Strategy, Techniques for Analyzing Industries and
Competitors, The Free Press, New York, 1980.
13. Schweyer, H. E., Analytical Models for Managerial and Engineering Eco
nomies, Reinhold Publishing Corporation, New York, 1964.
14. Sullivan, W. G . and W. W. Claycombe, Fundamentals of Forecasting, Reston
Publishing Co. , !ne., Reston, Va. , 1977.
15. Vernon, 1. R. (ed.), Realistic Cost Estimating for Manufacturing; Society of
Manufacturing Engineers. Dearborn, Mich., 1968.
Chap. 5 Exercises 137
16. Wheelwright, S. C., and S. Makridakis, Forecastinf( Methodsfor Management ,
2nd ed., John Wiley & Sons , lnc .. New York, 1977.
EXERCISES
51. List as many considerations as you can th at might affect the choice of forecast 
ing techniques by a local company familiar to you.
52. If a company manufactures a "faddish" li ne of sporting goods equipment, how
would its forecasting strategy differ from that of another company producing
aluminum beverage containers?
53. Explain how .the different stages of a product's !ife cycle will influence the
choice of a forecasting strategy.
S4. (Section 5.2.3) Suppose that you own a small company that manufactures
metal castings for severa! large automotive companies. Over the past severa!
years you have found that quarter! y newcar sales tend to lag the prime interest
rate by 3 months . Y ou would like to make a forecast ofnext quarter' s car sales
so that the size of your work force can be anticipated. There is a direct rela
tionship between car sales and demand for castings that your company pro
duces. The following data are gathered:
lnterest Next
Rate Quarter
Year Quarter (%) Sales ($M)
1 8.00 $23
2 8.25 17
3 8.50 18
4 8.25 20
2 1 7.75 21
2 7.25 25
3 7.70 24
4 7.25 29
3 1 7.50 24
2 7.75 23
3 7.25 26
4 7.00 30
ln te rest
Rate
Year Quarter (%)
4 1 7.00
2 7.50
3 7.50
4 8.25
5 1 8.75
2 8.50
3 7.50
4 7.00
6 7.50
Ne xl
Quarter
Sales ($M)
$25
26
17
20
15
18
22
23
(a) Calculate a linear regression equation for these data, assuming that the
interest rate is the independent variable.
(b) Calculate the correlation coefficient.
(c) Make a forecast of sales for the next quarter based on this quarter's prime
interest rate of 7.50%.
5S. (Section 5. 1.3) Consider the following ti meseries data of demand for a certain
company's product.
\
~
:.:
\'
L
1
;.j
:tl :t
j
1 \Jj
ft
11
~ ' '"
138 Estimating Markets, Revenues, and Costs Chap. 5
Mon th Demand (Booked Orders) Mon th Demand (Booked Orders)
4
5
6
7
8
9
10
3,1109
2,641
2.934
3.239
3,490
2,569
3,205
2,561
3,047
2,607
Il
12
13
14
15
16
17
18
19
20
(a) Plot these data on a piece of graph paper.
3,387
3,138
2,908
3,512
3,291
2,804
3,096
3,106
3,195
3,605
(b) Apply single exponential smoothing to the data when a'= 0.20 and make a
forecast for T = 1 month into the future .
(c) Repeat part (b) when a' = 0.05 and T = 3 months.
56. (Section 5.2.4) Discuss the principal advantages and disadvantages of the
Delphi method of forecasting.
57. (Section 5.2.4) In your class, attempt torun a Delphi study to determine the
priee of a compact disk player 3 years from now. Was group consensus af.
fected by conducting two or three rounds of the procedure?
58. (Section 5.2.4) Try to think of sorne products that are presently in the carly
stages of a substitution curve effect. List them and try to estimate when the
newer product will take more than half the market.
59. (Section 5.2.4) How could trend extrapolation be used to forecast future inno
vations in the aerospace industry? What performance characteristics do you
believe are important here?
510. (Section 5.3)
(a) Determine the priees for two products A and B, basing priee on a return on
investment of 15%.
(b) At the same priee as in part (a) , what is the markup based on (1) total
costs. (2) incrementai costs, and (3) value added in production?
Available data for a year's operation. with an effective income tax rate of 52%
and 10,000 annual units of each product , are as follows:
Initi al investment . 1
Mnterinls,
ircct, Cn
ther tl xed
1\!HII ,
Product
A
s 90,000
20, 000
30.000
50.000
1()(),00()
B
$120,000
40,000
I.S.OOO
15.000
10,000
Chap . 5 Exercises
139
511. (Section 5.4.2) Use the factor technique to estimate the cost of installing a local
area network in a factory environment having the following characteristics.
One large building on a single leve! will require a total of 3,000 ft of coaxial
(broadband) cable to network its six departments . Six network interface units
(NIUs) will be required and a total of 50 taps will have to be made to connect
ali the anticipated workstations and programmable deviees . Two modems are
needed in addition to one network manager/analyzer that costs $30,000. The
information necessary to make the estimate may be obtained from the work
sheet shown. How accurate do you think such an estimate would be?
Component Cost Estimating Relationship
1. Interbuilding connections $100 150 per foot x
2. lntrabuilding connections $2050 per foot x
3. Cable installation $20 per foot x
4. Equipment
a. Broadband
CA TV amplifier $500 1500 x
Taps $17 20 each x
Splitters $5 15
x __ = __
NI Us $500$1 .000 per port
x __ = _ _
Modems $1,000 each
x __ = __
b. Baseband
NI Us $600 per port x
Repeaters SI ,2001 ,500 each x
Taps/transceivers $200 300 each
x __ = __
c. Network manager $10,00030,000
Network analyzer $30,000
5U. (Section 5.4.2) A residential builder just finished constructing a 3,000ft 2 home
for $96,000. This cost did not include the lot or utility access fees . A detailed
breakout of costs for this job is as follows :
Item Fraction of Finished Cost
Lumber and carpentry
Electrical wiring
Plumbing
Concrete and masonry
Wall board
Flooring
Foundation preparation
and npplinnces
Heu ti ng and air contlilionlng
RoonnR
l'ul nt ln
M nllllriOUII N
0.20
0. 10
0. 14
0.09
0.04
0.06
0.02
0.0
O.llY
() , 1()
1) ,()/
tl 01
1 IMI
i 1
''
' (
1
'
1
j :
,.
\.

\
140 Estimating Markets, Revenues, and Costs
(a) What is the unit cost for the justfinished home?
(b) If a 4,000ft
2
home is to be built, es ti mate the total cost from the answer to
part (a) and compare it with the total of estimated item costs based on the
breakout given above.
513. (Section 5.4.3) Develop the standard time to perform a manual assembly task
having the following elements and compute the standard number of assemblies
that can be completed each hour.
Normal Time Allowance
Element (min) (%)
1 4.2 15
2 1.7 15
3 2.9 15
514. (Section 5.4.3) For a particular activity that is being work sampled, determine
the number of observations necessary to ens ure a confidence interval of 95%
and a maximum interval of 5%. The observed proportion of occurrence of this
activity is 0.4 (i.e. , 40% of the random observations have identified this activity
in progress) . Next, repeat this exercise when the confidence interval is low
ered to 90%.
515. (Section 5.4.3) In a production run of R130jet engines, the cumulative time to
produce units 1 through 100 is 80,000 hr.
(a) Calculate the unit time for the fiftieth engine with a 71 % leaming curve in
effect.
(b) Calculate the cumulative time to produce units 1 through 40 with a 71%
leaming curve.
516. (Section 5.4.4) A 100kW diesel generator cost $140,000 seven years ago when
a certain equipment cost index was arbitrarily set at 100. A similarly designed
generator rated at 150 kW is now being proposed and the cost index is 140.
The cos! exponent factor. {3. is O. 7 for this type of equipment.
(a) Determine the estimated cost of the proposed generator by using the ap
proptiate cost estimating relationship.
(b) Repeat part (a) when {3 = 0.4.
517. (Section 5.4.4) The Neptune Manufactuting Company is consideting abandon
ing their old plant , built 23 years ago, and constructing a new plant that has
50% more square footage. The original cost of the old plant was $300,000 and
its capacity, in terms of standardized production units, was 250,000 units per
year. Capacity of the proposed plant is to be 500,000 units per year. During
the past 23 years, costs of plant construction have risen by an average of 5%
per year. If the cost exponent factor is 0.8, what is the estimated cost of the
new plant ?
518. (Section 5 .5) Answer the following questions based on the Consumer Priee
Index in Table 510.
(a) What is the annual compounded rate of inflation in living costs from 1978
through 1981 ?
Exercises 141
(b) If a person had takehome pa y of $20,000 in 1978, how much should it have
been in 1981 in order for that person to be living comparably to 1978?
(c) If, during 19781981 a person was earning 8% through investments (such
as moncymarket accounts}, what would have been his real rate of return?
S19. (Section 5.4)
(a) It is desired to estimate the 1990 construction cost of a new plant for which
1985 component costs, and applicable rates of inflation, are as follows :
La bor
Building materials
Equipment
Total
19!!5 Cost
(Millions)
$1
5
3
$9
Projected Annual Inflation Rate
( 1985 through 1990)
(%)
10
0
15
(b) Using the answer to part (a) and using a (combined) MARR of 25%, find
the equivalent worth of the 1990 construction cost as of 1985.
S20. (Section 5.5) Estimate the salary you hope to be making 5 years from now,
stated in today's purchasing power (i.e. , RS\
0
l) . If the average inflation rate for
goods and services you expect to bu y is 10% compounded annually, wh at will
your hopedfor salary need to be in actual dollars (i .e., A$
5
)?
S21. (Section 5.5)
(a) Y ou hope to prepare for someone's college education costs by depositing
enough in a bank savings account today to provide, as of 15 years from
now, an amou nt equivalent to $20,000 of today's purchasing power. If the
bank will pa y 8% and the estimated inflation rate is 5% (both compounded
annually), what lump sum should you deposit today?
(b) Repeat part (a) except that the estimated inflation rate is 10% per year .
S22. (Section 5.5) Labor costs for alternatives X and Y are estimated on different
bases as follows:
End of Year n Ait. X (AS.) Ait. B ( R S ~ " J
1 S200M S190M
2 210M 190M
3 220M 190M
4 230M 190M
If the combined (real and inflation) MARR = 25% and the real MARR
15%/yr, show which alternative has the lower equivalent :
(a) Present worth of costs at year O.
(b) Future worth of costs at year 4.
l
vv

i
L \.
\, t
l
1
(
,,
~
f
j
1
~ .
!
Chapitre 2
INTRODUCTION LA NOTION DE RISQUE ET D'INCERTITUDE
2.1 INTRODUCTION:
Dans le monde de l'industrie et du commerce les dcideurs sont conscients qu'ils n'auront pas
toujours raison quand il s'agit d'valuer un projet ou de choisir entre plusieurs possibilits. En
effet, le profil d'un Cash flow est influenc par les vnements futurs (international , national ,
local .. . ) qui sont incertains et qui ne peuvent tre contrls par le dcideur.
Il n'est pas possible de prvoir l'avance l'arrive de tels vnements et la rentabilit d'un
certain projet ne peut tre value avec certitude, qu' la fin de celuici. Donc, la rentabilit
d'un projet donn dpend d'un facteur risque qui peut entraner qu'elle sera diffrente de celle
prvue au moment de la prise de dcision.
Nous avons considr, jusqu' maintenant dans toutes les tudes conomiques effectues,
connatre tous les lments avec certitude; nous pouvions donc dfinir la valeur d'un paramtre
conomique par un seul nombre. Mais en ralit dans une situation concrte ces paramtres:
vie utile, valeur de rcupration, revenus et cots ne peuvent tre dtermins par une valeur
unique. Pour plusieurs tudes conomiques, il devient ncessaire ou prfrable de considrer
l'aspect reli aux risques et l'incertitude introduit par la variation des diffrents paramtres.
Dans les chapitres 3,4 et 5, nous allons voir diffrentes techniques pour tenir compte du risque
et de l'incertitude dans une analyse conomique.
2.2 QU'ESTCEQUE LE RISQUE ET L'INCERTITUDE
Avant de passer aux diffrentes techniques d'analyse conomique avec risque et incertitude,
nous tenterons d'claircir cette notion dans les sections suivantes:
2. 2.1 Terminologie
Les situations de dcision pouvent tre classes en trois catgories: certitude, risque et
incertitude.
Toutes les tudes conomiques qui ont t conduites dans les chapitres prcdents ont
t ralises dans des conditions de certitude, c'estdire que la probabilit de succs
a t considre gale 1.0 pour chaque projet valu (fig. 2.1 a) . Cependant en
ralit, les exemples o les paramtres sont estims avec certitude sont rares. On parle
alors de risque ou d'incertitude.
22
Les analystes prsentent une distinction classique entre le risque et l'incertitude pour un
paramtre ou une analyse conomique; par le risque, l'analyste connat les probabilits
que les tats de la nature, les revenus et les cots se produisent (figure. 2.1 b) alors
qu'en prsence d' incertitude il ne possde pas la connaissance de la frquence de leur
distribution (figure. 2. l.c). La figure 2.1 montre graphiquement ces diffrences.
Jrnl\.
ROCl.E
intitu:E
j_
a=3
m
?
~ ;
\Ae\Jile \Ae Uile
\Ae\Jile
7friS V" 7friS
(a) (b) (c)
figure 2.1
D'autres auteurs diffrencient le risque de l'incertitude en affirmant que pour le risque ils
peuvent utiliser la distribution de la probabilit d'un paramtre alors que pour l'incertitude ils
n'ont .pas confiance que l' estim de cette probabilit soit correcte.
2.2.2. Causes du risque et de l'incertitude
Nous allons maintenant considrer quelques unes des principales causes du risquee d
l'incertitude:
1. Un nombre insuffisant d'investissement similaire
Chaque firme a un faible nombre d'investissement d'un type particulier. L' e 
des revenus dfavorables peut difficilement tre statistiquement compens par
les revenus favorables; ce phnomne devient critique si une compagnie in es .
un montant considrable par rapport ses ressources totales.
23
2. Erreurs subjectives dans les donnes
L'analyse conomique dpend de l'optimisme ou du pessimisme de l'analyse,
ou des facteurs intangibles qui ne devraient pas influencer l'analyse conomi
que.
3. Changement dans l'environnement conomique invalidant l'exprience
prcdente
Il arrive souvent que l'analyste fait ses prdictions futures partir des donnes
accumules prcdemment mais sans ajustement pour les conditions futures
prvues. Ceci implique qu'il assume implicitement que le futur se droulera
comme le pass; parfois des circonstances peuvent invalider cette hypothse, par
exemple la crise d'nergie.
4. Mauvaise interprtation des donnes
Diffrents facteurs influenant les donnes peuvent conduire une mauvaise
interprtation.
5. Erreurs d'analyse
L'analyse technique et conomique peut comporter des erreurs.
6. Disponibilit d'administrateurs comptents
La limitation des ressources humaines et physiques l'intrieur d'une
organisation peut handicaper certains projets relativement d'autres.
7. Rcupration de l'investissement
Une industrie qui va se procurer de l'quipement trs spcialis doit se rendre
compte qu'il pourra tre reveridu avec beaucoup plus de risque que de
l'quipement d'application gnrale.
8. Dsuet
De nouveaux procds peuvent rendre l'quipement actuel dsuet; ce dernier
facteur constitue aussi une cause plus ou moins grande de risque.
De Garmo et Canada C) dans leur livre, regroupent en quatre parties les cau
risque et de l'incertitude:
a) L'imprcision possible des valeurs utilises dans l'tude conom1
l'estimation des revenus entrane gnralement le plus d'erreur.
b) Le genre d'industrie considr; par exemple une entreprise minire co
gnralement plus de risque qu'une picerie.
c) Le type d'usine et d'quipement; de l'quipement spcialis se re end
difficilement que de l'quipement rgulier.
2
d) Le temps qui doit s'couler avant que toutes les conditions de l'tude .so =:
ralises; un temps plus long diminue la probabilit que tous les a<
demeurent tels qu'estims, donc le risque augmente puisqu'on ne
prophtiser exactement ce que l 'avenir apportera. La mthode de la pri
recouvrement fournit une indication de ce niveau de risque.
2.2.3 Faiblesse du traitement probabiliste
Les probabilits utilises dans une analyse du risque constituent gnralem
lments non vrifiables et trs subjectifs. De plus, les diffrentes probabilits u '
peuvent dans une mme tude conomique diffrer en qualit et en quantit entree es.
L'anal y ste se doit de raliser que l'utilisation de probabilits ne fait pas dispara
risque et l'incertitude mais fournit des informations supplmentaires qui aidero
prise de dcision. L'utilisation des limites de confiance des probabilits aidera
gestionnaire prendre une dcision. Pour les problmes industriels les ex
recommandent d'utiliser une limite de confiance de 95%; c'estdire d'ajoute
valeur moyenne plus ou moins deux fois la dviation standard (2a).
2.2.4 Moyens de changer le degr d'incertitude
Une firme peut prendre diffrents moyens pour changer son degr d' ince
mentionnons:
1) Obtenir des informations techniques et conomiques supplmentaires.
2) Augmenter le nombre d'activits pour avoir suffisamment de pro
d'investissement pour que les valeurs moyennes utilises soient statistique
valables.
3) Diversifier les oprations; choisir des produits que les cycles conomiq
influencent diffremment.
25
2.2.5 Choix entre diffrents niveaux de risques et retours sur l'investissement
Les conomistes acceptent gnralement qu'un projet plus risqu doit rapporter un
retour sur l'investissement plus lev. La connaissance d'une relation entre le risque
et le retour permettrait de comparer tous les projets sur une mme base. Malheureu
sement cette relation ne peut tre dtermine d'une faon gnrale et prcise; la
politique d'une firme par rapport au risque doit donc tre connue avant qu'une dcision
d'investir puisse tre prise. Cette politique dpend grandement des prfrences du
responsable dcisionnel et du risque total auquel la firme doit dj faire face.
2.2.6 Guide pour considrer le risque et l'incertitude
Le risque et l'incertitude existent sous une varit de forme et de degr dans toutes les
analyses conomiques. Limportance de considrer le risque et l'incertitude pour un
investissement augmente, avec l'augmentation de la complexit du projet et de son
importance conomique relativement la valeur totale de la firme. Une analyse
conomique supplmentaire tenant compte du risque et de l'incertitude se justifie si elle
cote moins que les conomies potentielles qui en rsulteraient; en pratique leurs
estimations deviennent difficiles.
En gnral, lorsqu'une firme dsire effectuer une tude conomique, soit pour un
investissement ou un remplacement, elle considrera les diffrentes tapes telles que le
montre la figure 2.2. Ces tapes s'appliquent aussi bien des projets mutuellement
exclusifs ou non. Elle fait voir l'existence de quatre points o les projets peuvent tre
accepts ou rejets.
Le point 1: comprend l'ensemble de l'argent que la compagnie devra payer si elle
accepte le projet. Elle effectuera un choix immdiatement si les montants pour chaque
projet demeurent faibles compars ceux pour une tude supplmentaire. Ce critre
subjectif dpend de l'tat financier de l'entreprise, de la grosseur des projets
gnralement considrs et de la disponibilit des ressources pour une analyse plus
pousse.
Le point 2: considre le rapprochement des choix; si l'un se dtache nettement la
compagnie arrtera sa dcision.
Le point 3: exan1ine les analyses dj effectues en fonction de l'importance relative du
projet pour la compagnie, elle peut alors dcider de faire une tude plus prcise de
certains lments.
Le point 4; la compagnie rpte les considrations du point 3.
26
2.3 DCISION SOUS INCERTITUDE COMPLTE
Les diffrentes techniques d'analyse sont utilises selon qu'on soit dans le cas de risque ou celui
d'incertitude. Dans le prsent chapitre nous allons dvelopper les techniques d'anal se
conomique dans le cas d'un investissement sous incertitude COMPLTE c'estdire dans une
situation ou l'analyste ne peut pas assumer des probabilits aux diffrents tats de la nature.
Supposons qu'une dcision doit tre prise entre plusieurs possibilits A
1
, A
2
, ~ ; l e dcideu
dsire choisir la meilleure. n y a rn conditions futures, sous appeles tat de la nature SI s1
... Sm. Ces tats peuvent survenir indpendamment du contrle de celui qui doit prendre la
dcision.
Prenons P;i = Valeur pour celui qui prend la dcision s'il choisit la possibilit A; et qu ' en
mme temps, l'tat de la nature Si se produit. Supposons que les valeurs positives indiquent
des situations favorables alors que celles ngatives en montrent des nonfavorables. Ces
valeurs peuvent tre des revenus, des valeurs prsentes, des utilits ou tout autre indice
caractrisant un investissement. Les diffrents choix et tats de la nature peuvent tre prsents
sous la forme d'une matrice, tableau 2.1 illustre une matrice de dcision.
Possibilits
(contrl par
le dcideur)
A
c
T
I
0
N
s
A\S.
A
A?
A ~
1
A,
1
AN
sl
p
p?l
p ~ l
1
Pn
1
PN1
Tableau 2.1: matrice de dcisions
TATS DE LA NATURE
'aucun contrle du dcideur
s?

SI

SM
Pn

p.
 P_rn_
Pn

1

1
p ~ ? 
1

1
1 1 1 1 1
Pn  p..l.l_
l J
1 1 1 1 1
Pm
  
PNM
_ __JI
QEis llmrtaui
FI>ft3 ....... l1bnct <Dnrn tt
...... crwdt
IWillictl
A
27
R:irt4
rrn
OlrSrer le mrq.Je
c:on:re aralyse sertMt,
10.2: ders reca m OOes p:u ue amy1yoo
28
(:\
29
Maintenant considrons l'exemple prsent dans le tableau 2.2
Tableau 2.2 Exemple d'une matrice de dcision
TATS DE LA NATURE
A\S.
1 J
rcession prosprit grande
normale prosprit
Possibilits (actions) Sl S2 S3
Aucune construction Al +25 15 50
Construire une petite usine A2 50 +40 +50
Construire une grosse usine A3 150 +10 +200
Une analyse conomique sous certitude comporte un seul tat de la nature c'estdire une seule
colonne. La meilleure possibilit consiste prendre la valeur la plus leve dans la colonne
unique (ex: en cas de rcession (S
1
) aucune construction (A
1
)). Mais, maintenant dans une
analyse sous incertitude avec plusieurs tats de la nature donc plusieurs colonnes, quelles
possibilits choisir?
En premier, liminons toutes les possibilits qui mdpendamment du principe de choix que le
dcideur utilisera montrent des infriorits vidntes, et ensuite, on procdera au choix entre
les autres possibilits en tenant compte de l'incertitude.
Exemple 2.1
Considrons la matrice du profit du tableau 2.3 qui reprsente la valeur prsente des profits de
projets (A
1
, A
2
, A
3
, A
4
et A
5
) selon quatre tats de la nature (S
1
, S
2
, S
3
et S
4
).
2 10
Tableau 2.3: matrice de profits
TATS DE LA NATURE
Possibilits s s? S, s ~ ~ minimum maximum
A 4 4 0 2 0 4
A? 2 2 2 2 2 2
A, 0 8 0 0 0 8
A4 2 6 0 0 0 6
As 0 1 1 0 0 1
L'analyste se rend immdiatement compte par inspection de la matrice qu'indpendammen e
l'tat de la nature qui surviendra, il aura toujours avantage choisir la possibilit A
2
au li
de As. Il peut donc liminer immdiatement la possibilit As de toutes considrations futures.
On se rfre cette situation en mentionnant que la possibilit A
2
domine la possibilit A
5
D'une faon gnrale s'il y a deux possibilits Ag et Ah tel P gj ~ p h i pour toutes valeurs de
alors la possibilit Ah domine la possibilit Ag qui peuttre limin de cette faon. ll restera.
plusieurs possibilits et l'analyste aura besoin de procdures supplmentaires pour faire
choix.
2.3.1 tude de quelques principes sous incertitude>>
2.3.1.1
Nous allons maintenant considrer diffrents principes de choix que les experts no
suggrent. Nous discuterons aussi des points forts et des faiblesses qu 'on
principes.
Principes de l'opportunit gale
Les analystes se rfrent souvent ce principe sous les noms de <<principe de
insuffisante ou principe de Laplace. En absence d'vidence contraire l 'anal
assume que tous les tats de nature Si ont une probabilit gale de se produire.
choisit donc l'action qui permettra d'optimiser la valeur espre.
Appliquons ce principe l'exemple 2.1
pi: probabilit que l'tat de la nature si se produise
PI =p2 =p3 =p4 = !,4
2.3.1.2
2.3.1.3
211
La possibilit A
1
a donc une valeur espre gale :
E(A
1
)= IAx4+
1
Ax4+ IAxO+
1
Ax2=2.5
De la mme faon les autres possibilits ont comme v ~ e u r s espres:
E(A0=E(A
3
)=E(A
4
)=2.0
Le choix se porte donc sur la possibilit A
1
qui permet d'avoir un profit espr maximal
de 2.5.
Principe maximin ou minimax
Ce principe consiste choisir la possibilit qui maximise la valeur minimale qu'on
pourrait recevoir ou minimiser la valeur maximale qu'on pourrait dbourser. En
d'autres mots, l'analyste choisit la possibilit Ai de la faon suivante dans le cas d'une
matrice de profits.
maximum minimum Pij
J
Ce principe reflte une attitude trs pessimisme et conservatrice; en effet, il suppose
qu'indpendamment de la possibilit choisie, l'tat de la nature le plus dfavorable va
se produire. Dans le cas d'une matrice contenant des cots au lieu des profits, le
principe consistera choisir l'action qui minimisera le cot maximal. Le principe
s'appellera alors minimax:
minimum maximum p ..
IJ
j
Dans l'exemple 2.1, le principe maximin choisit la possibilit A
2
laquelle correspond
le maximum (2) des minimums de (0,2,0 et 0) respectivement pour les possibilits (A
1
,
A
2
, A
3
et A
4
), (tableau 2.3).
Principe maximax ou minimin
Ce principe consiste choisir la possibilit Ai, qui maximise la valeur maximum des
profits, c'estdire l'analyste choisit la possibilit Ai associe avec le maximum Pij dans
le cas d'une matrice de profits. Dans le cas d'une matrice de cots, ce principe permet
de slectionner la possibilit qui minimise le cot minimum.
minimum minimum Pij
i j
Cette rgle optimisme et aventureuse suppose que, indpendamment de la possibilit
choisie, le meilleur tat de la nature possible associ cette possibilit va arriver.
2.3.1.4
2 12
Pour l'exemple 2. 1, le principe maximax choisit la possibilit A
3
laquelle correspond
le maximum (8) des maximums de (2,8,6,et 1) respectivement pour les possibilits (A
1
,
A
2
, A
3
et A
4
) , (Tableau 2.3).
Principe d'Hurwicz
Hurwicz a gnralis les deux principes cidessus pour tous les niveaux d' optimisme
entre le pessimisme extrme de MAXIMIN et l 'optimisme exagr de MAXIMAX.
L'analyste choisit une indice d'optimisme o< a< 1. Alors pour chaque possibilit
il calcule la relation:
Hi = a max Pii +(1a) min Pii
1 1
Dans le cas de matrice de profits, on choisit la possibilit qui maximise Hi et dans le
cas de matrice de cots, on choisit la possibilit qui minimise Hi, cette relation se
dfinie comme le principe d 'Hurwicz. Aux deux limites lorsque le degr d' op ti mis o
gale zro et un, on retrouve respectivement les principes maximin et maximax.
a=O=>Hi= min Pii
j
= max min P .. =>maximin
IJ
i j
a=1 =>Hi= max Pii
j
Max Hi => max max Pi(>maximax
i j
La figure 2.3 nous fait voir le principe d'Hurwicz appliqu l'exemple 2. 1.
2.3.1.5
Valeur du
critre
d'Hurwicz
8 ~ A 3
6 A4
4 A1
~ ~ A 2
0.25
Indice d'optimiste: a
Supposons un degr d'optimisme a=0.25 dans l'exemple 2.1
H
1
pour A
1
= 0.25 x 4 + O. 75 x 0 = 1
H
2
pour A
2
= 0.25 x 2 + 0.75 x 2 = 2 = > choisir A
2
ou A
3
H
3
pour A
3
= 0.25 x 8 + O. 75 x 0 = 2
H
4
pour A
4
= 0.25 x 6 + O. 75 x 0 = 1.5
213
Dans l'exemple 2.1, le choix se portera sur la possibilit A
2
pour des degrs
d'optimisme (a) infrieur 0.25. Alors qu'il choisira A
3
pour des degrs suprieurs
0.25. Pour un degr d'optimisme gale la valeur 0.25, le principe n'indique aucune
prfrence entre les possibilits A
2
et A
3
Dans la relation d'Hurwicz, le degr d'optimisme a influencera les profits maximaux
et les cots minimaux moins de l'utilisation d'une autre convention.
Principe du regret minimax (Savage)
Considrons de nouveau l'exemple 2.1, si aprs qu'une possibilit ait t choisie et les
mesures prises pour son application, le dcideur constate que l'tat de la nature S
2
se
produit, il aimerait certainement avoir choisi la possibilit A
3
S'il avait dans les faits
choisi une autre possibilit, disons A
2
il regretterait son choix actuel. Pour un tat de
la nature donn la mesure du regret qu'il peut avoir; s'appelle REGRET. Pour une
matrice content ~ possibilit et Sm tats de la nature les regrets pourront tre calculs
1
. max {Pii}Pi
1
=Ri
1
..
par la re atwn '
L'application de la dfinition du regret une possibilit A
2
et l'tat de la nature S
2
de
l'exemple 2.1, donne un regret de R
22
=82=6. Considrant d'une faon similaire les
214
15 autres combinaisons possibles (Ai, Sj) de la matrice de cet exemple l'analyste pourra
construire la matrice du regret, (tableau 2.3).
Tableau 2.4: matrice de regrets
TAT DE LA NATURE
Possibilits
sl s2 s3 s4
Regret maximum pour
pour la possibilits A
A1 0 4 2 0 4
A? 2 6 0 0 6
A3 4 0 2 2 4
A4 2 2 2 2 2
Dans cet exemple, le choix de la possibilit A
4
minimisera le regret maximal .
. 2.3.2 Problme de choix de principe
2.3.2.1
En rsum les diffrents principes considrs jusqu' maintenant choisissent:
L'opportunit gale  > A
1
Hurwicz pour a< 0.25 > A
2
Hurwicz pour a> 0.25 > A
3
Regret minimax  > A
4
Nous avons donc converti le problme de choisir une possibilit en celui de choisir
principe. Pour faciliter ce choix, considrons maintenant quelques unes des proprits
de ces principes.
Problmes dans l'utilisation du principe d'opportunit gale
Si l'on utilise le principe d'opportunit gale, le choix d'une possibilit est sensibl 2
la faon dont les diffrents tats futurs sont prsents.
Considrons l'exemple suivant:
Exemple 2.2
On dsire choisir entre deux pices d'quipement; l'une tant d'application
gnrale, l'autre ayant un emploi bien spcifique. Or, on peut obtenir trois co
possibles pour utiliser la machine. Il est aussi possible que l'on ait aucun contra
215
ce cas la machine d'application gnrale pourra tre utilise pour une autre application,
mais la machine d'application spcialise demeurera inutilise.
Donc la liste des tats futurs possibles:
S
1
aucun contrat
S
2
un contrat
A
1
on achte la machine spciale
A
2
on achte la machine gnrale
La matrice, prsente dans le tableau 2.5, montre les profits (pertes) correspondant
chaque possibilit et tat de la nature; avec le principe d'opportunit gale p
1
=p
2
=lf2
Tableau 2.5: matrice de profits
(2 tats de la nature)
ETAT DE LA NATURE
Possibilits s s ?
A 1 6
A? 1 5
Le profit espr rsultant du choix de chacune des actions gales: (p
1
=p
2
=
1
/z)
E{A
1
) =
1
/z(1)+
1
/z{6) =2.5
E(A
2
)=
1
/z ( 1)+
1
/z(5) =3.0
Ce principe permet de slectionner la possibilit A
2
, car elle a un profit espr (3.0), suprieur
celui de la possibilit A
1
(2.5) .
Maintenant, supposons que 1 'analyste veut solutionner le mme problme en prsentant les tats
de la nature (obtention d'un contrat) de la faon suivante:
S
1
aucun contrat
S
2
contrat A
S
3
contrat B
s4 contrat c
La matrice des profits devient celle prsente au tableau 2.6
2.3.2.2
216
Tableau 2.6: matrice de profits
( 4 tats de la nature)
TAT DELA NATURE
Possibilits
SL
S, s3_ sd.
A1 1 6 6 6
A, 1 5 5 5
Le principe d'opportunit gale nous donne donc maintenant:
Les profits esprs des deux possibilits deviennent:
E(AI) =
1
A(1) +
1
A (6) +
1
A (6) +
1
A (6) = 4.25
E(A2) = lA( 1) +
1
A (5) +
1
A (5) + 1,4 (5) = 4.00
Maintenant avec le principe d'opportunit gale l'analyte choisi la possibilit A
1
, car
elle gnre un profit espr de 4.25 suprieur celui de la possibilit A
2
Pour viter cet inconvnient, l'analyste devrait tablir la liste des tats de la na
de faon que chacun ait la mme probabilit d'arriver. Ceci implique qu'il conruu
d'avance les probabilits, alors que sous incertitude complte>> il assume ne pas les
connatre.
Problme dans l'utilisation du principe d'Hurwicz
Considrons maintenant des exemples pour illustrer deux problmes que le principe
d'Hurwicz engendre le tableau 2. 7 illustre une matrice des profits.
Exemple 2.3
Soit la matrice des profits (tableau 2. 7).
Tableau 2.7: matrice de profits
TATS DELA NATURE
Possibilits
SI
A1 1
A? 3
H
1
= a(4) +(la) 1= 3 a+l
H
2
= a(3) + (la) 3= 3
s2 max Pu
.i
4 4
3 3
217
min p ..
IJ
j
1
3
Le choix de la possibilit A
1
donne selon le critre d 'Hurwicz les profits de 3a + 1,
alors que pour celui de la possibilit A
2
, il donne un profit constant de 3, qui ne dpend
pas du degr d'optimisme a.
galant ces deux quations, 3a+ 1 =3l'analyste trouve une valeur de 2/3 pour le degr
d'optimisme a au point d'intersection. Le choix se porte donc sur la possibilit A
2
pour
un degr d'optimisme infrieur 2/3, car ces profits galent 3, alors que ceux de la
possibilit A
1
demeurent infrieurs, car il galent 3a + 1.
Maintenant, ajoutons une constante 5 chaque lment de la premire colonne de la
matrice du tableau 2.7; le tableau 2.8 fait voir la nouvelle matrice.
1
Tableau 2.8: matrice de _Qrofits
TATS DE LA NATURE
Possibillts
SI s2
max min Pu
p ..
IJ
j
.i
A 6 4 6 4
A? 8 3 8 3
Le critre d 'Hurwicz permet de calculer les profits pour
la possibilit A
1
: H
1
de a(6) + (la) 4= 2 a+4
la possibilit A
2
: H
2
de a(8) + (la) 3= 5 a+3
1
218
Le point d'intersection devient:
5a+3 = 2a+4
La solution de cette quation donne une degr d'optimisme a, gale 113, donc po
toute valeur de a infrieur 113, le choix se dplace de la possibilit A
2
la possibilit
A
1
par rapport la solution prcdente, car 2a+4 demeure suprieur 5a+3.
Les experts rfrent ce problme comme un manque de linarit de la colonne .
Si le futur S
1
se produit, l'avnement de l'tat de la nature S
1
procurera cinq units
supplmentaires de profit indpendamment de la possibilit choisie. Le changemen de
la possibilit A
2
A
1
dpend donc, seulement de l'tat de la nature et non pas de la
possibilit choisie; idalement la dcision prise ne devrait pas tre influence.
Le principe du regret minimum et de 1 'opportunit gale possdent la linarit d la
colonne.
Exemple 2.4
Considrons la matrice des profits prsente dans le tableau 2.9, pour illustrer
deuxime problme associ au principe d'Hurwicz:
Tableau 2.9: matrice de profits
TATS DE LA NATURE
Possibilits sl s? s ~ s<l s ~ S,., s7 SR
SQ slO
A1
A?
1 0 0 0
0 .
0 0 0 0 0
0 1 1 1 1 1 1 1 1 1
L'analyste qui utilise le critre d'Hurwicz considrera les possibilits A
1
et A
2
co e
quivalentes, car pour:
H
1
=a(1) +(1a) O=a
H
2
=a(1) +(1a) O=a
et pour
Mais la plupart des gens auront une prfrence intuitive pour la possibilit _
impliquant une notion prconue au sujet des probabilits, ce qui ne peut tre le cas
lorsqu'on traite d'un problme sous incertitude complte. Pour cet exemple l'anal
qui utilise le principe d'opportunit gale choisira la possibilit A
2
2.3.2.3
219
Problmes associs au principe du regret minimax
Savage a tabli ce principe pour contourner le trs grand conservatisme du principe
MAXIMIN (ou MINIMAX). Le tableau 2.10, illustre une matrice de profit o
l'utilisation du principe de Savage a un avantage sur celui du principe maximin.
Tableau 2.10: matrice de profits .
TATS DE LA NATURE
Possibilits
SI s2
min p ..
IJ
j
A, 0 1000000 0
A, 1 1 1
L'analyste qui utilise le principe maximin choisit la possibilit A
2
, mais intuitivement
la possibilit A
1
semble plus dsirable, mais comme on peut le voir partir de la
matrice du regret montre au tableau 2.11, le principe du regret minimax choisit la
possibilit A
1
Tableau 2.11: matrice de regrets
TATS DE LA NATURE
Possibilits s_j
s2
Regret maximum
A, 1 0 1
A"
0 999999 999999
Par ailleurs, le principe du regret minimax a aussi des problmes propres;
Exemple 2.6
Les tableaux 2.12 et 2.13 illustrent respectivement une matrice des profits etla matrice
de regrets correspondants.
220
Tableau 2.12: matrice de profits
TATS DE LA NATURE
Possibilits s,
s2 s3
A, 1 6 4
A? 5 3 6
Tableau 2.13: matrice de regrets
TATS DE LA NATURE
Possibilits s,
s2 s3
regret
maximum
A, 4 0 2 4
A? 0 3 0 3
L'utilisation du principe du regret minimax permet de choisir la possibilit ~ .
Maintenant, supposons qu'une 3e possibilit A
3
devient possible. La matrice des pro :.s
est prsente dans le tableau 2.14 et celle des regrets par le tableau 2.15.
Tableau 2.14: matrice de profits
TATS DELA NATURE
Possibilits S, s? S,
A., 1 6 4
A? 5 3 6
A, 4 8 1
221
Tableau 2.15: matrice de regrets
TATS DE LA NATURE
Possibilits
SI s2 s3
regrets
maximum
A, 4 2 2 4
A? 0 5 0 5
A_3 1 0 5 5
L'ajout d'une troisime possibilit A
3
, entrane que l'analyste choisit maintenant en
utilisant le principe de regret minimax la possibilit A
1
Ce rsultat dsappointe, car
le fait de considrer une nouvelle possibilit ne devrait pas changer le choix fait parmi
les possibilits prcdemment considres.
L'exemple suivant illustre ce point. Une personne dsire acheter une nouvelle auto
aprs avoir considr les modles Ford et Chevrolet, elle choisit un modle Ford.
Avant d'avoir complt la transaction, elle apprend qu'un nouveau dtaillant des
modles Dodge vient de s'installer prs de chez elle. Elle va donc voir ces autres
modles pour dterminer s'ils prsentent des avantages par rapport au modle Ford
choisi. Mais voil qu' aprs avoir vu les modles Dodge, cette personne se ravise et
dcide maintenant de choisir une auto de modle Chevrolet.
L'utilisation du principe du regret minimax dpend donc de l'addition d'une nouvelle
possibilit qui peuttre ou non pertinente. Les principes d'opportunit gale et
d 'Hurwicz ne possdent pas ce problme.
2.3.3 Choix d'un principe
Plusieurs auteurs ont list les proprits dsirables qu'un principe devrait avoir. L'une
des plus connue a t faite par Milnor (2). Il liste les dix conditions auxquelles on
aimerait qu'une bonne rgle de dcision respecte:
1. Classification: Classifier les diverses actions.
2. Symtrie: La classification devrait tre indpendante de l'ordre des colonnes et
des ranges de la matrice.
222
3. Forte domination: L'analyste prfre la possibilit A; Ak, s'il prfre la valeur
P;i Pki pour tous lesj.
4. Continuit: Dans une matrice convergente, la classification ne doit pas tre
inverse la limite.
5. Linarit: La classification ne devrait pas changer si l'on remplace les valeurs
de la matrice (P;i) par a Pij + ~ ou a> 0
6. Addition d'une range: L'addition de nouvelles possibilits, i.e. de nou e es
ranges la matrice, ne devrait pas changer la classification des possibilits.
7. Linarit de la colonne: L'addition d'une constante tous les lments d'
colonne ne devrait pas changer la classification des possibilits.
8. La duplication d'une colonne: L'addition la matrice d'une nouvelle colonn
identique une existante ne devrait changer la classification des possibilits.
9. Convexit: Si les possibilits A; et Ak s'quivaillent aucune n'est prfre.
2. Addition d'une range: L"addition d'une nouvelle range o aucun lmen de
cette range n'est prfr aux lments correspondants de toutes ranges d ,.
existantes ne devrait pas changer la classification des possibilits.
Milnor (2) a prouv qu'aucun des principes discuts prcdemment ne satisfait les dix
conditions. On peut voir facilement que la rgle de Laplace viole la condition
Minimax la condition 7, Hurwicz la condition 7 et 9 et Savage la condition 6.
Milnor a aussi prouv mathmatiquement que le principe:
de Laplace satisfait les conditions 1, 2, 3, 6 et 7
Minimax satisfait les conditions 1, 2, 3, 4, 6, 8 et 9
d'Hurwicz satisfait les conditions 1, 2, 3, 4, 5, 6 et 8
de Savage satisfait les conditions 1, 2, 3, 4, 7, 8, 9 et 10
Pour choisir un principe, l'analyste doit donc d ~ i d e r des proprits les plus importan
pour un cas bien particulier et choisir celui qui en satisfait le plus. La liste d
proprits ne peut tre trs longue.
223
Problme no 2.1
tablir un principe personnel qui vous permet de prendre une dcision sous incertitude
complte. Monter un exemple comment ce principe peuttre utilis. Quels sont les
avantages et inconvnients de votre principe?
Problme no 2.2
Appliquer les diffrents principes de choix que nous avons vu aux cours la matrice
de cots suivante:
TATS DE LA NATURE
Possibilits S, s? s1
s"
A1 18 18 10 14
A? 14 14 14 14
A1 5 26 10 10
A"
14 22 10 10
A_5_ 10 12 12 10
Problme no 2.3
Former une nouvelle matrice en prenant chaque cot de la matrice du problme 2.2 et
en lui additionnant 2 et en multipliant le rsultat par 3. Utiliser les mmes principes
de choix qu'au problme 2.2. Qu'estceque ces rsultats vous suggrent sur les
proprits de diffrents principes?
Problme no 2.4
Dans la matrice de profits cidessous, quelle possibilit choisiriezvous en utilisant le
principe du regret minimal?
22
TAT DE LA NATURE
Possibilits s,
s2 s3
A 8 7 4
A? 10 0 4
A, 1 9 5
At
5 6 7
1 . De Garmo et Canada, Engineering Economy
Chapitre 3
ANALYSE CONOMIQUE TRADITIONNELLE ET A V ANCE
SOUS RISQUE ET INCERTITUDE
3.1 INTRODUCTION:
Dans ce chapitre nous allons dvelopper d'autres techniques d'analyse conomique avec risque
et incertitude qu'on classera en deux grandes catgories soit, celles d'analyse traditionnelle et
celles d'analyse avance. L'accent sera mis sur l'analyse conomique avec risque, c'estdire
le cas o l'analyste connat les probabilits des divers paramtres conomiques, notamment,
que les tats de la nature, les revenus et les cots se produisent.
3.2 BREF RAPPEL DES NOTIONS DE BASE DE PROBABILITS
La probabilit peuttre dfinie comme le rapport du nombre de cas favorables l'vnement
sur le nombre de cas possibles.
Si un vnement A peut survenir S>> fois partir d'un nombre total de n>> cas possibles et
galement identiques, la probabilit que l'vnement survienne sera de:
s
P(A) =
n
o P(A) = probabilit que l'vnement survienne
s = nombre de cas o l'vnement peut survenir
n = nombre total de cas possibles.
Par exemple, en jouant pile ou face, la probabilit d'obtenir pile est de
1
/2 et la probabilit
d'obtenir face est de
1
h. Quand un d est lanc sur la table, la probabilit d'obtenir 1 est de
1/6, la probabilit d'obtenir 2 est de 1/6, ... , la probabilit d'obtenir un 6 est de 116. Aux
cartes, la probabilit d'obtenir du carreau est de 13/52. ll en est de mme pour le pique, le
coeur et le trfle.
La dfinition prcdente est utile dans les cas de situations finies o S>> et n>> sont connus.
Pour une situation infinie, o n = "", la dfinition conduira toujours une probabilit de O.
Ainsi, dans un cas de situation infinie, la probabilit qu'un vnement survienne est
proportionnelle la distribution de l'univers.
3.2.1 Thormes de probabilit
Thorme 1. La probabilit est exprime par un nombre variant entre 1.00 et 0, o la
valeur 1.00 reprsente la certitude que l'vnement arrivera et 0, la certitude que
l'vnement n'arrivera pas.
32
soit un vnement E; 0:5; P(E;) :5; 1 (quation 3.1)
Thorme 2. Si P(A) est la probabilit que l'vnement A survienne, la probabilit que
A n'arrive pas est donne par P(A)=l.OO P(A)
P(A) + P(A) = 1 (quation 3.2)
Exemple: Si la probabilit d'obtenir une pice d'quipement lectrique dfectueuse es
0.04, la probabilit d'obtenir une bonne pice d'quipement est:
P(B) = 1.00  P(A)
= 1.00 0.04
= 0.96
Ainsi la probabilit d'obtenir une pice d'quipement acceptable est de 0.96.
Thorme 3. Si A et B sont deux vnements mutuellement exclusifs, alor la
probabilit que soit l'vnement A ou soit l'vnement B survienne est la somme de
leurs probabilits respectives.
P(A ou B) = P(A) + P (B) (quation .3)
L'expression mutuellement exclusive signifie que le fait qu'un vnement se produise
rend l'autre vnement impossible. Ainsi, si en jetant un d, on obtient un
(vnement A) , il n'est pas possible d'obtenir un 5.
Thorme 4. Si les vnements A et B sont indpendants, la probabilit q e
l'vnement A ou l'vnement B ou que A et B se produisent est donne par
P(A ou B ou les deux) = P(A) + P(B)  P(A et B) (quation 3. )
Un vnement indpendant est celui dont la ralisation ne prsente pas d'influences
la probabilit de ralisation de l'autre ou des autres vnements.
Exemple: Dterminer la probabilit d'obtenir un valet ou un carreau en tirant u
carte d'un paquet.
P(A ou B ou les deux) = P(A) + P(B)  P(A et B)
4 13 1
P(valet ou carreau) = +
52 52 52
16
==0.308
52
33
Thorme 5. La somme des probabilits des vnements relatifs une situation est
gale 1.
P(A) + P(B) + ... P(N) = 1.00 (quation 3.5)
Exemple: Un inspecteur examine 3 produits d'un sousgroupe pour dterminer s'ils
sont dfectueux. L'exprience nous dit que la probabilit de ne trouver aucun produit
dfectueux d'un chantillon de 3 est 0.89, que la probabilit de trouver un produit
dfectueux d'un chantillon de 3 est 0.06, et la probabilit de trouver deux produits
dfectueux d'un chantillon de 3 est 0.03. Quelle est la probabilit de trouver 3
produits dfectueux d'un chantillon de 3?
Cette situation ne comprends que 4 vnements: 0, 1, 2 et 3 produits dfectueux.
P(O) + P(l) + P(2) + P(3) = 1.00
0.89 + 0.06 + 0.03 + P(3) = 1.00
P(3) = 0.02
Thorme 6. Si A et B sont deux vnements indpendants, la probabilit que les deux
vnements arrivent est le produit de leurs probabilits respectives.
P(A et B) = P(A) x P(B) (quation 3.6)
Exemple: La probabilit que le projet A soit rentable est de 0.30 et la probabilit
que le projet B soit rentable est de 0.10. Quelle est la probabilit que les deux projets
soient rentables.
P(A et B) = P(A) x P(B)
= (0. 30) (0.10)
= 0.03
Thorme 7. Si A et B sont deux vnements dpendants, la probabilit que les deux
vnements se produisent est le produit de la probabilit que A arrive et que si A arrive,
B arrivera aussi.
P(A et B) = P(A) P(B/ A)
P(B/A)
= P(A et B) si P(A) *O
P(A)
(quation 3. 7)
(quation 3.8)
Le symbole P(B/ A) est dfini comme tant la probabilit que 1 'vnement B se
produise, si 1 'vnement A est arriv. Un vnement dpendant est celui dont la
ralisation influence la probabilit de l'autre ou des autres vnements.
3
Exemple: Une bote contient 50 engrenages et 3 d'entre eux sont dfectueux. i
un chantillon de 2 est enlev, quelle est la probabilit que les deux soient dfectueux?
La probabilit que le premier soit dfectueux (vnement A) est P(A) = 3/50. La
probabilit que le second soit dfectueux est P(B/A) = 2/49, tant donn que la
premire pice dfectueuse a dj t enleve.
P(A et B) = P(A) x P(B/ A)
P(A et B) = (3/50) x (2/49) = 0.002
3.2.2 Valeur espre et variance
Soit X
1
, X
2
.. . X, des vnements mutuels exclusifs qui reprsentent des re en
possibles et P
1
, P
2
, ... Pn, les probabilits de ces vnements tel que:
La valeur espre des revenus (moyenne) est de:
m
E(x) = P.. = L xi Pi
i ~ I
(quation . 10)
La mesure de dispersion de la distribution de probabilit appel variante V(x) ou ci es
donne que d'expression.
m
o
2
= V(n) = L ((xiE(x))
2
Pi (quation . 10)
i ~ I
o
2
= V(x) = E(x
2
) [E(x)f (quation . 11
Exemple 3.1: Une compagnie considre l'investissement dans un projet don 1
revenus probabilistes sont donns au tableau 3 .1.
Tableau 3.1:
donne pour l'exemple 3.1
Revenu Probabilit des revenus
X.$ P(xi)
llO 000 0.8
530 000 0.2
Dterminer les quantits suivantes:
a) La valeur espre du projet
b) La variante du projet
Solution:
a) de l'quation 3.9
E(X) = X1 P(Xl) + X2 P(X2)
= 110 000 x (0.8) + 530 000 x (0.2) = 18 000$
b) de l'quation 3.10
a
2
= V(X) = (X
1
 E(X
1
))
2
x P(X
1
) + (X
2
 E(X
2
))
2
x P(X
2
)
= (110 000 18 000)
2
x (0.8) + (530 00018 000)
2
x (0.2)
a
2
= V(X) = 6,553.10
10
$
3.3 ANALYSES TRADITIONNELLES
35
Nous allons en premier voir quelques unes des analyses traditionnelles le plus importantes
utilises pour considrer le risque.
1. Jugement intuitif
2. Ajustement conservateur
3. Optimistepessimiste
4. Sensibilit
5. Point mort
6. Escompte du risque
3.3.1 Jugement intuitif
Cette mthode consiste effectuer l'analyse conomique en supposant que l'analyste
connat tous les paramtres conomiques avec certitude et applique un jugement
subjectif sur le risque de chaque projet; il peut renverser la dcision base sur les
rsultats quantitatifs s'il considre le risque suffisamment grand.
3.3.2 Ajustement conservateur
Cette procdure consiste changer les estims d'un ou plusieurs paramtres d'une
analyse dans une direction conservatrice, de faon diminuer le risque qu'une situation
survienne qui ne soit pas aussi favorable que celle prdite par l'analyse conomique.
L'analyste peut faire l'tude diffrents degrs de conservatisme. Le choix de valeurs
conservatrices pour les paramtres conomiques augmente la certitude d'un rsultat
favorable.
36
Exemple 3. 2:
Considrons l'achat d'une nouvelle machine pour imprimer les circuits lectriques e
remplacer celle existante, ceci devrait permettre une compagnie de raliser une
conomie. L' achat et l'installation cotera 10 000$, avec certitude. Elle esti me
produire 5 000 circuits de plus par anne durant les six prochaines annes et obtenir une
conomie annuelle de 5 200$. La compagnie estime que la nouvelle machine aura une
vie utile de six ans, possdera alors une valeur de rcupration de zro et cotera
000$ par an de plus pour 1 'opration et la maintenance. La banque financera cet achat
un taux annuel effectif de 5%.
Considrant la valeur des paramtres estims comme exacte, on peut calculer le profi
annuel ralis durant les six prochaines annes.
Rcupration du capital avec un retour 10 000 (A/P, 5, 6)
Cot d'opration et de maintenance supplmentaire
Cot total annuel
Revenu total annuel
Profit annuel
(3 9 )
5 2
1 2
a) Pour tenir compte de la variation possible des valeurs des estims, la compagnie
appliquera un jugement conservateur en utilisant un taux d'intrt de 25 %. Les
cots et les revenus annuels deviennent:
Rcupration du capital avec une retour 10 000 (AfP, 25, 6)
Cot d'opration et de maintenance supplmentaire
Cot total annuel
Revenu total annuel
Profit (perte) annuel
(3 390 )
(2 )
(5 39
52
(190 )
Un taux d'intrt de 25% entranera une perte de 190$ par an et la compagnie pourra
dcider de ne pas effectuer le remplacement.
b) D'une faon similaire la compagnie pourra conserver le taux d' intrt de 5 o
appliquer un jugement conservateur en rduisant la priode de rcupration de
6 3 ans. Les cots et les revenus annuels deviennent:
Rcupration du capital avec une retour 10 000 (AfP, 5, 3)
Cot d'opration et de maintenance supplmentaire
Cot total annuel
Revenu total annuel
Profit (perte) annuel
De nouveau le jugement conservateur rend le remplacement non favorable.
3. 3. 3 Optimistepessimiste
37
(3 670$)
(2 000$)
(5 670$)
5 200$
(470$)
Cette procdure consiste changer la valeur estime pour un ou plusieurs paramtres
d'une faon favorable (optimiste) et nonfavorable (pessimiste), pour dterminer l'effet
de ces variations sur les rsultats de l'analyse conomique. L'utilisation de mthodes
statistiques permet de mesurer le degr d'optimisme. La valeur pessimiste correspond
l'estim des rsultats lorsque l'analyste attribue chaque paramtre conomique la
valeur la moins favorable qui a une chance raisonnable de se produire. Cette valeur ne
correspond pas la plus mauvaise valeur qui peut se produire. L'estim doit tre fait
pour chaque paramtre individuel.
La valeur optimiste correspond l'estim des rsultats lorsque l'analyste attribue
chaque paramtre la valeur la plus favorable qui a une chance raisonnable de se
produire. Canada dans son livre dfmit un estim optimiste comme la valeur d'un
paramtre dont l'analyste prvoit qu'elle sera atteinte pas plus que 5% du temps, alors
qu'un estim pessimiste correspond la valeur d'un paramtre dont il prvoit qu'elle
sera plus favorable que la valeur finale pas plus que 5% du temps.
La figure 3.1 illustre une reprsentation possible des valeurs pessimistes et optimistes
des revenus.
valeur prvue
estim pessimist; .. optimiste
5% 2 a 2 a 5%
distribution des revenus
Fig. 3.1
o: dviation standard
Cette mthode optimismepessimisme apporte des informations additionnelles dans une
tude conomique; elle montre les consquences rsultantes des dviations de la valeur
prvue. Les valeurs pessimistes utilises dans cette analyse peuvent correspondre aux
valeurs conservatrices de l'anal y se prcdente.
Exemple 3.3
Appliquons cette analyse la situation dcrite dans le tableau 3. 1
Tableau 3.1: donne de l'exemple 3. 3
Optimiste Prvue Pessimiste
(estim)
Investissement $ 10 000 10 000 10 000
Vie utile, ans 7 5 4
Valeur de rcupration $ 2 000 2 000 2 000
Profit annuel, $ 6 000 5 000 4 500
Cot annuel $ 2 200 2 200 2 400
Taux d'intrts % 8 8 8
39
Calculons les profits (cots) quivalents annuels en considrant successivement les
valeurs pessimiste, prvue et optimiste des diffrents paramtres.
AE (pessimiste) = [(10 0002 000) (A/P, 8%, 4) +2 000 (.08)]
 2 400 + 4 500 = (470$)
AE (prvue) = [(10 000 2 000) (A/P, 8%, 5) +2 000 (.08)]
 2 200 + 5 000 = 630$
AE (optimiste) = [(10 000 2 000) (A/P, 8%, 7) +2 000 (.08)]
 2 200 + 6 000 = 2 100$
Les informations additionnelles dans des conditions d'optimiste et de pessimiste aideront
l'investisseur prendre une dcision en considrant les profts et les pertes possibles
dans ces conditions extrmes.
Note: Dans cet exemple, l'analyste a considr que tous les lments pessimistes ou
optimistes se produisent en mme temps, mais il pourrait considrer diffrentes
combinaisons de ces valeurs.
3. 3. 4 Analyse de sensibilit
Une analyse de sensibilit apporte des informations supplmentaires une simple
analyse sous certitude. Elle permet de dterminer le degr de sensibilit des rsultats
d'une analyse conomique aux variations des diffrents paramtres utiliss. Les
experts, parfois, dfiniront le concept de sensibilit comme le changement relatif de
grandeur d'un ou plusieurs paramtres, dans une analyse conomique, qui va renverser
le choix des actions possibles: cette analyse s'effectue d'une faon similaire celle
optimistepessimiste.
L'une des faons les plus intressantes pour tudier la sensibilit consiste mettre en
graphique les rsultats de tous les paramtres indpendants considrs; l'analyste
montre la variation de chacun sur une abscisse commune, comme pourcentage de la
valeur estime; l'ordonne reprsentant le critre conomique utilis pour valuer le
projet.
La figure 3.2 montre un exemple de ce type:
Q)
::::J
c:
c:
m
laux d'int,
ert
100 50 0 50 100
Dviation de la valeur estime, %
Fig.3.2 Analyse de sensibilit
10
Dans ce cas particulier la variation des cots annuels et de la vie utile ont une influence beaucoup pl
importante sur la variation des profits annuels, que celle du taux d'intrt; l'analyste devrait leu
accorder une attention particulire.
Exemple 3.4
Une socit 0 dsirant construire un nouvel atelier de fabrication de pte doi
dterminer le rendement en pte le plus conomique. Une analyse prliminaire a
permis d'identifier les cots du bois et du raffinage comme les paramtres les plus
importants sur le cot de la fabrication de la pte. La socit a conduit une tude d
sensibilit sur ces deux paramtres. La figure 3.3 montre l'influence des variations de
60% respectivement dans les cots du bois et de l'nergie sur le cot de fabrication
de la pte, pour trois niveaux de rendement.
titre d'illustration cette analyse permet de constater que si les cots d'ne e
augmentaient par 50%, il n'y aurait pratiquement pas d'avantage conomique d
produire une pte 90% de rendement au lieu de 70%. De mme, si une compagni
pouvait rduire ses cots en bois d'environ 30%, elle pourrait utiliser aussi cono
quement des ptes bas (50%) et haut rendements (90%). On peut facilem
comprendre que ce genre d'information a grandement aider la compagnie prendre
dcision concernant son investissement.
8
6
du cot d'n
60/o 40 20
Variation du cot individuel
Rendement en pte
50
7
cot dubois
0 +40 +600!o'
Fig.3.3: cot de production de pte diffrents niveaux de rendement
311
3.3.4.1
8
6
4
2
0
3 1
Autre type d'analyse de sensibilit
Les experts suggrent aussi une autre faon d'tudier la sensibilit; elle consiste
mettre en graphique deux des paramtres importants, par exemple le taux d'intrt
et la vie utile, et de tracer une ou des isocourbes, pour le critre conomique
choisi, tels les profits annuels. Une ligne d'indiffrence limitera deux zones; l 'une
o l'analyste suggre l'investissement et l'autre o ille dconseille. La figure 3.
illustre cette mthode.
Investir
pas investir
0 2 4 6 8 1 0 12 14 16 18 20
Taux d'intrt, %
Fig . 3.4 : sensibilit d'un projet deux paramtres importants
Exemple 3.5
Appliquons cette analyse de sensibilit aux valeurs prvues de l'exemple
considr pour illustrer l'analyse optimistepessimiste. L'analyste dsire conna
l'effet d'une diminution de la vie utile de 40% et d'une augmentation du t2.
d'intrt de 50%, car il considre ces deux lments comme critiques. La m od
des cots annuels quivalents donne:
CA= + 5 0002 200 [(10 0002 000) (A/P, 12%, 3) + 2 000 (. 12)]
= 773$
Prcdemment pour un taux d'intrt de 8% et une dure de vie utile de 5 ans.
cette mthode indiquait un profit annuel de 630$. Dans le mme exemp
l'analyste dsire trouver pour quelle valeur du taux d'intrt la dcision d'in
changera. Il suffit donc de trouver la valeur du taux <;l'intrt pour que les co
annuels galent zro.
CA= 5 0002 200 [(10 0002 000) (A/P, i%, 5) + 2 000 (i %)] =0
pour i = 12%
i = 20%
CA= +330
CA= 300
3 13
Donc par interpolation, ou graphiquement l'on trouve que:
1=12% . (20%12%)=16.2%
. ( 330 l
330+300
3.3.5 Analyse du point mort
L' analyste utilise cette analyse dans des tudes conomiques o il existe une
incertitude particulire, concernant un des paramtres. Il dfinit le point mort
comme la valeur du paramtre auquel la rentabilit du projet devient marginale.
Dans le cas de plusieurs choix, il dfinit le point mort comme la valeur du
paramtre auquel les choix deviennent galement dsirables.
Exemple 3.6
Supposons que dans l'exemple 3.3, l'estimation de la vie utile soit trs incertaine,
l'analyste dsire dterminer la vie utile minimal permettant de justifier le projet.
galant les cots annuels quivalents zro l'anal y ste dtermine la dure de vie
utile (N) minimal.
5 0002 200((10 0002 000) (A/P,8%,N)2 000(.08)]=0
(AIP, 8%,N)=
2640
=0.330 Dans les tables on trouve un
8000
temps de 3. 7 ans qui correspond la valeur de 0.330.
Exemple 3.7
Une compagnie considre la fabrication d'un nouveau produit, elle dispose des
donnes suivantes:
Prix de vente:$/unit
Cot d'quipement:$
Frais fixe par anne:$
Cot d'opration et de maintenance:$/heure
Temps de production /1000 unit, heure
Dure de vie du projet, ans
TRAM,%
12,50
200 000
50 000
25
100
5
15
La compagnie assume une valeur de rcupration des quipements nulle aprs 5 ans
et prend X= volume de vente /anne. La valeur annuelle de l'investissement est:
VA = 200 000 (A/P, 15, 5) 50 000 0.1 * (25) X + 12.5.X
VA = 109 660 + 10.X
Au point mort V A = 0 = > X = 10 966 units/ anne
31
La figure 3.5 montre une reprsentation graphique de ce problme.
1 1 1 1 1 1 1 1 1 1
J
1
Cots
annuel
1
l
1
20 00
1 1 1 1 1 1 1 1
1
1
0
5000
11000 150 000
X (Ventes annuelles)
Fig. 3.5 : graphique du point mort
3.3.6 L'escompte du risque
Cette analyse utilise un taux d'intrt proportionnel au degr de risque; lorsque ce
dernier crot le retour sur l'investissement doit tre plus lev. L'analyste trou e
gnralement difficile et subjectif d'tablir un taux d'intrt associ un degr de
risque. Il doit connatre les revenus et les cots pour chaque proje .
L'interprtation de cette analyse doit tre change lorsqu'un projet comporte u
revenu net ngatif, car un taux d'intrt plus lev pour un risque supri
entranera une valeur prsente du cot plus faible pour le projet. Cette analyse ne
s'applique pas aux projets ayant une dure de vie utile trs faible, car dans ce cas
l'analyste n'escompte pas les revenus et les cots.
Pour viter d'avoir attribuer un taux d'intrt diffrent chaque projet selon le
risque qu'il comporte, l'analyste peut calculer les valeurs prsentes pour cha e
projet correspondant divers taux d'intrt. Mettant en graphique la val
prsente en fonction du taux d'intrt, pour les projets les plus risqus il dtermin
les taux d'intrt ou de retour correspondant une mme valeur prsente.
315
Taux d'intrt,
0
/o
Fig. 3.5 : exemple d'analyse avec escompte de risque
Les projets A et B ont une mme valeur prsente de 10 lorsque l'analyste escompte
les revenus du projet B par un taux de 5% et celui du projet A par 20%. On peut
voir aussi, que s'il escompte les revenus du projet A un taux suffisamment lev,
si le projet comporte un haut niveau de risque, la valeur prsente du projet B peut
devenir suprieure celle du projet A. Cette analyse fournira un lment de plus
pour la prise de dcision.
3.4 ANALYSE AVANCE
Aprs avoir considr quelques analyses traditionnelles, nous allons maintenant voir quatre
(4) mthodes d'analyse conomique plus avances:
1. Montaire probabilistique
2. Estim  variance
3. Escompte des variables
4. Simulation de MonteCarlo
3.4.1 Analyse montaire probabilistisque
L'anal y se montaire probabilistique comprend l'estimation de certains paramtres
dont l'analyste considre la variation comme importante en terme de probabilit.
Ces valeurs estimes peuvent tre modifies pour obtenir les caractristiques
3.4.1.1
3.4.1.1.1
316
dsires de la mesure du mrite (ex. la valeur prsente). Les rsultats d'une telle
tude peuvent se prsenter sous l'une des formes suivantes:
a) L'estim des bnfices:
i) valeur prvue
) tat futur le plus probable
i) niveau d'aspiration
b) L'estim des bnfices et de leurs variations possibles.
L'estim des bnfices
Considrons en premier l'analyse relie l'estim des bnfices.
Posons: Pi = probabilit que l'tat futur possible Si se produise
Pi ~ 0 pour tous les j et
j = rn
:E
P.
J
1
j = 1
Dans une situation sous risque l'analyste connat ces probabilits ou peut les
estimer alors que dans une situations sous incertitude complte, il n'a pas confiance
que l'estim des probabilits soit correct.
V ale urs prvues
La rgle de la valeur prvue sert souvent pour faire un choix dans une situatio
comportant des risques. L'analyste choisit la possibilit qui maximise le gain pr
ou minimise les pertes prvues.
Pour la possibilit A;, on calcule le gain ou la perte prvue partir, de la
sommation du produit de chacun des gains ou pertes par la probabilit qu'il se
produise:
j = rn
L
P .. P.
IJ J
j = 1
La dtermination des probabilits constitue l'un des principaux problmes lors
l'analyste dsire utiliser la mthode de la valeur prvue. Elles peuvent tre bases
sur des donnes historiques ou sur un traitement statistique rigoureux, mais dans
la plupart des cas l'analyste doit se servir de son intuition et de son j ugemen .
ne possde aucune assurance que le futur se droulera d'une faon identique a
pass mme si les probabilits peuvent avoir pour base des donnes historiques.
317
En revanche, l'obligation de penser et d'estimer la valeur des diffrents facteurs
dans le futur procurera des rsultats suprieurs ceux provenant d'aucune
rflexion. Certains analystes trouvent utile d'utiliser des distributions statistiques
(ex: Poisson, Normale) pour dterminer les valeurs prvues.
Parmi les distributions de probabilit thoriques , qui sont le plus utilises dans
l'analyse conomique avec risque, on cite les distributions Normale et Beta. La
figure 3.6 montre un exemple de distribution Normale. Les probabilits qu'une
variable alatoire X, normalement distribue ait une valeur dans un intervalle de
a, 2a ou 3a de chaque ct de la moyenne sont respectivement 68.26, 95.45 et
99.73%.
99.73%
Fig. 3.6 : exemple d'une distribution normale
Dans certains cas, on ne peut reprsenter la distribution de probabilit par une
distribution thorique connue. Une approche qui peuttre utilise pour estimer
une distribution de probabilit subjective consiste dterminer un estim
pessimiste, optimiste et la valeur la plus probable d'une variable alatoire.
La rgle de la valeur prvue constitue une mthode trs rpandue; pourquoi estelle
bonne? Considrons une squence de dcisions ayant des rsultats indpendants:
j = rn
E
Pj ou j = 1 ,2,3, .. . m
j = 1
z'r.. gale la somme des valeurs des rsultats des N premires dcisions lorsque
l'analyste applique le principe de la valeur prvue. Ici , par hypothse pour chaque
dcision l'analyste choisit la possibilit qui maximise VP; zN symbolise cette
somme partielle.
Alors que: ~ gale la somme des valeurs des rsultats des N premires dcisions
lorsque l'analyste applique tout autre principe de choix. Lorsque le nombre de
3.4.1.1.2
3.4.1.1.3
31
dcision N croit et tend vers l'infini la probabilit que zN soit plus grand q u e ~
tend vers un. Ceci constitue la Loi du succs long terme. Plusieurs volumes
de statistiques prsentent la preuve de cette loi.
L'tat futur le plus probable
Cette rgle consiste trouver l'tat de la nature possible, Sj, qui possde la
probabilit maximal de se produire; pour cet tat de la nature choisir la possibilit
~ qui maximisera Pij Cette rgle considre que l'tat de la nature le plus probable
va se produire avec certitude. En pratique, beaucoup d'investisseurs prennent des
dcisions de cette faon.
Niveau d'aspiration
Pour un iveau d'aspiration K, slectionnez la possibilit Ai qui maximisera la
probabilit que la valeur du rsultat soit plus grande ou gale K; K tant une
mesure d'un niveau de rentabilit que le propritaire d' une firme dsire obten
pour un projet. Cette rgle reflte le fait que la plupart des humains essaient de
dcouvrir et de slectionner des possibilits satisfaisantes et non pas ncessairement
optimales.
Exemple 3.7
Pour mieux comprendre l'application de ces rgles, considrons la matrice de
profits (tableau 3.2); elle montre aussi les probabilits qu'ont chaque tat de la
nature de se produire.
Tableau 3.2: matrice de profits
TATS DE LA NATURE
Possibilits
SI
S2 S3
actions
p_l = .2 p2 = .7 P
1
= . 1
Al 50 120 2 000
A, 130 130 130
AJ 500 110 100
Calculons en premier la valeur prvue pour chaque action:
VA(A
1
) = 50* .2 + 120 * .7 + 2 000 * .1 = 294
VA(A
2
) = 130 * .2 + 130 * .7 + 130 * . 1 = 130
VA(A
3
) = 500 * .2 + 110 * .7 + 100 * .1 = 187
319
Donc ici la rgle de la valeur prvue>> va permettre de choisir la possibilit A
1
, car
294> 130 et 187.
L'tat de la nature S
2
possde la probabilit la plus leve de se produire donc la
rgle de l'tat futur le plus probable va slectionner pour cet tat de la nature
l'action qui rapportera le profit maximal; ici, la possibilit A
2
qui donnera un profit
de 130 compar 120 et 110 respectivement pour A
1
et A
3
Maintenant, supposons que l'investisseur dsire un profit suprieur ou gal 300
(K = 300), il considrera pour chaque possibilit la ou les probabilits d'obtenir un
tel profit.
Pour la possibilit A
1
seul l'tat de la nature S
3
qui a une probabilit de 0.1 de se
produire pourra lui procurer un profit de 2 000. Pour la possibilit A
2
aucun tat
de la nature ne pourra lui procurer un tel profit. Pour la possibilit A
3
, l'tat de
nature S
1
qui a une probabilit de 0.2 de se produire lui procurera un profit de 500.
Donc, la rgle de niveau d'aspiration, pour un niveau de profit de 300, slectionne
ra la possibilit A
3
qui a une probabilit de 0.2 de lui procurer ce profit alors que
les possibilit A
1
et A
2
ont respectivement des probabilits de 0.1 et 0.0.
Plusieurs autres principes de choix peuvent tre prsents, mais lequel devonsnous
utiliser? Beaucoup d'tudes utilisent le principe de la valeur prvue principale
ment cause de la loi du succs long terme. En revanche, cette loi s'applique
une longue squence de dcisions. Or, la plupart des investisseurs ont effectuer
un seul investissement ou un nombre restreint ce qui n'assure pas la validit de la
loi du succs long terme.
Exemple 3.8
Considrons un exemple familier chacun d'entre nous, soit celui d'acheter ou non
une assurance automobile. Les statistiques publies montrent qu'un individu a une
chance sur mille d'avoir un accident (P
1
= 0.001); donc, la probabilit de ne pas
avoir d'accident gale 0.999. Si une police d'assurance cot 600$ et que le cot
moyen d'un accident gale 40 000$, l'individu doitil ou non acheter une
assurance? La matrice du tableau 3.3 montre les diffrents cots:
Les rgles de la valeur prvue (VP(A
1
) =600 et E(A
2
) =40) de l'tat futur le plus
probable ( S ~ et du niveau d'aspiration pour K suprieur 600 slectionnent toutes
la possibilit A
2
c'estdire de ne pas acheter d'assurance.
Mais beaucoup de personnes dans ce cas vont prfrer payer 600$ pour liminer
mme une chance trs faible de perdre une somme considrable de 40 000$.
1
3.4.1.2
320
Tableau 3.3: donne de l'exemQle 3.8
1
TATS DELA NATURE
Avoir un accident ne pas avoir
Possibilits
SI
d'accident
s2
P=O.OOl P=0.999
A
1
: acheter de l'assurance
600 600
A
2
: ne pas acheter d'assurance 40 000 0
Estim des bnfices et de leurs variations possibles
La rgle de la valeur prvue peuttre considrablement amliore lorsque
1' analyste ajoute aux valeurs prvues une mesure de la variation possible des
bnfices et si possible une description complte de variation de la probabilit des
bnfices.
Pour mieux comprendre cette notion, considrons l'exemple suivant:
Exemple 3.9
Un projet A a un bnfice annuel prvu de 10 000$ avec une dviation standard de
25 000$. Un projet Ba un bnfice annuel prvu de 8 000$ avec une dviatio
standard de 4 000$. Un investisseur qui dsire prendre le minimum de risque
pourra choisir le projet B qui a un bnfice annuel prvu infrieur au projet A.
Supposons que les bnfices prvus possdent une distribution normale; la forme
de la courbe de distribution des bnfices procurera l'investisseur des informa
tions supplmentaires et lui permettra de mieux j u g ~ r des avantages respectifs des
diffrents choix.
La figure 3.7 montre les bnfices de l'exemple distribus normalement.
3.4.1.3
3.4.1.3.1
probabilit de perte
pour le projet A
20 10
321
projet B
~ = 8 000 0 = 4 000
projet A ~ = 1 0 000
~ = 25 000
0
10 20 30
40
Fig. 3.7 : distribution de bnfices
On peut voir, que le projet A a une probabilit plus leve d'encourir une perte que
le projet B; le projet B pourrait tre alors choisi. Une distribution statistique des
bnfices autre que normale pourrait changer la dcision.
Les sries de Taylor peuvent servir pour approximer les variations simples des
diffrents paramtres conomiques, mais 1 'analyste doit recourir aux techniques de
simulation pour l'tude des variations plus complexes.
Exemple sur l'analyse avance
Exemple sur la mthode montaire probabilistique
Le dpartement de gnie industriel a fourni aux cadres d'une usine les informations
suivantes concernant un projet d'investissement.
Investissement,$
dure de vie, ans
3
Valeur de rcupration, $
Revenu annuel, $
Cot annuel , $
5
7
Financement (taux d'intrt annuel effectif), %
10 000
Probabilit
0.3
0.4
0.3
2 000
5000
2 200
8
3.4.1.3.2
322
Les cadres peuvent donc calculer l'aide de la mthode des revenus (cots) annuels
quivalents les bnfices prvus et leur dviation standard.
Pour une vie de 3 ans
VA=5 0002 200 [(10 0002 000) (A/P, 8%, 3) + 2 000 (8%)] = 460$
Pour une vie de 5 ans
VA=5 000  2 200 [(10 0002 000) (A/P, 8%, 5) + 2 000 (8%)] = 630$
Pour une vie de 7 ans
VA=5 0002 200 [(10 0002 000) (A/P, 8%, 7) + 2 000 (8%)] = 1 110
L'investissement aura donc un revenu annuel prvu de:
VAprvu= 460$ (0.3) + 630$ (0.4) +1110 (0.3) = 446$
Les bnfices auront comme variance:
2
Variance(a =L [(AE)
2
* probabilit(AE)]  (AEprvu)
2
AE
= ( 460? * 0.3 + (630? * 0.4 + (1110)
2
* 0.3 (446)
2
= 401000
a= dviation standard= V401000 = 631$
Pour une limite de confiance de 95% les bnfices pourront varier entre  81 6S e
1 708$ car;
AE = JL 2a
AE = 446$ + 2 * 631$
AE =  816$ 1 708$
Les cadres savent maintenant que l'investissement a un bnfice prvu de 44
mais qu'il pourra encourir aussi une perte de 816$ et un bnfice maximum de
1 708$ pour les limites de confiance fixes.
Exemple sur la valeur et l'utilit prvues
Un investisseur qui a la courbe utilitaire cidessous dsire tudier la possibili
d'investir dans un projet qui a des revenus prvus en fonction de diffren
probabilits.
30
20
10
1 / 1 ~                           
/
2 000 i
1
1
1
Revenu
$
5 000
20 000 000
10 000 000
0
2 000 000
10 000 15 000 20 000
Dollars
Probabilit
0.05
0.15
0.30
0.50
323
L'investisseur dsire effectuer l'tude en utilisant les mthodes de l'utilit et de la
valeur prvues. Le tableau 3.4 cidessous montre ces diffrents calculs.
Tableau 3.4
Revenu Probabilit A *B Utilit B*D
(1 000 000$) du revenu (1 000 000$) du revenu
(A) (B) (C) (D)
20 0.05 1.0 26.3 1.31
10 0. 15 1.5 20.0 3.00
0 0.30 0 1.0 0.30
2 0.50 1.0 10.0 5.00
Revenu prvu 1. 5 utilit _Qrvue 0.39
324
La mthode de l'utilit prvue indique que le projet n'est pas favorable car il a une
utilit prvue ngative alors que l'utilit d'un revenu de 0 gale un. Par contre,
. la mthode de la valeur prvue indique un projet favorable, car elle montre un
revenu prvu suprieur zro.
3.4.2 Estim variance:
Cette mthode d'analyse conomique, parfois appele mthode quivalente la
certitude, consiste rduire les considrations conomiques en une mesure simple
des bnfices prvus ainsi que de leur variation. Mathmatique, elles 'exprime par
la relation:
V= f.1. Aa
o V= estim  variance
f.1. = bnfices prvus
a= dviation standard des bnfices
A= coefficient d'aversion du risque
Les experts suggrent de prendre
u"
A= ___ f.l. ou u"
2
correspond
la drive seconde de la fonction utilit versus bnfices. Cette valeur de A
s'applique lorsque l'utilit marginale de l'argent diminue; ce cas se rencontre
souvent en pratique.
3.4.3 L'escompte des variables
Une faon de considrer une augmentation du risque avec des temps de plus en
plus loigns consiste escompter les diffrents revenus possibles par des taux
diffrents; des taux plus levs accorderont une pondration plus faible aux
bnfices les plus loigns. Les experts suggrent des modles mathmatiques pour
tablir les diffrents taux d'intrt, mais ici nous nous limiterons au principe.
3.4.4 Simulation de MonteCarlo
Mthode dveloppe durant les annes 40 par Von Neuman, Vlan et Fermi pour
rsoudre certains problmes de design d'crans antirayonnement. Elle es
coteuse exprimentalement et difficile rsoudre analytiquement.
Elle consiste reprsenter un problme dterministe par un processus stochastique
dont les distributions de probabilit satisfont les relations mathmatiques du
problme dterministe complexe. La premire tape consiste construire un
modle analytique qui reprsente l'investissement considr (exemple une quation
pour la valeur prsente) .
325
Dans la deuxime tape, l'analyste dveloppe une distribution de probabilit pour
chaque facteur qui est sujet incertitude dans le modle. Il peut utiliser des
donnes historiques ou une approche subjective.
partir des distributions de probabilits pour chaque facteur dans le modle, Il
gnre au hasard une rponse tentative. La rptition de ce mode d'chantillon
nage un nombre important de fois, permet de gnrer une distribution de frquence
pour la rponse (ex. Valeur Prsente). On peut utiliser n'importe laquelle des
mthodes de mesure de rentabilit. La distribution de frquence rsultant peut tre
utilise pour obtenir des donnes probabilistiques sur le problme original.
La simulation MonteCarlo est une technique efficace pour l'analyse conomique
sous risque . Elle permet de varier les facteurs simultanment, en tenant compte,
aussi bien de leur tendance centrale que de leur dispersion, et cela, quelque soit la
distribution de probabilit qui les reprsente. Le chapitre 6 permet d' expliquer plus
en dtail cette technique.
326
Problme no 3.1
Des amateurs de tennis dsirent construire deux cours intrieurs au cot de 125
000$. Ce tennis intrieur aura une dure de vie utile de 10 ans et une valeur de
rcupration de 20 000$. Les cots d'entretien sont estims 23 000$ par anne.
S'ils peuvent louer au taux de 7$ par heure par cours, combien d'heures devrontils
louer chaque anne pour oprer au point mort? Le taux d'intrt bancaire est
actuellement de 10%:
Problme no 3.2
Une ligne de production sera requise encore pour trois ans. Elle a actuellement des
cots annuels de 310 000$; l'on prvoit qu'ils vont demeurer constants. Un
ingnieur industriel a soumis un plan pour une nouvelle conception de la ligne de
production. Cette nouvelle conception coterait 150 000$ installer et il y a 50%
des chances qu'elle permettrait de rduire les cots annuels de fonctionnement
210 000$. Cependant, il y a une probabilit de 25% que les cots annuels de la
nouvelle ligne augmentent de 20 000$ dans chacune des deux dernires annes.
Il y a aussi 25% que l'augmentation soit de 75 000$ dans chacune des deux
dernires annes.
Doiton conserver la ligne actuelle ou installer la nouvelle?
Problme no 3.3
Considrant la matrice de profits annuels en milliers de dollars cidessous:
TATS DE LA NATURE
Actions
SI s2 s3 s4 s5 s6
P=.l . 1 .4 .2 .1 .1
Al 12 5 8 3 6 9
A? 7 0 1 5 20 7
Al 3 3 7 9 5 5
A4 0 12 15 2 8 200
A ~ 10 22 9 0 4 12
a) Quelle(s) action(s) devriezvous choisir pour minimiser votre probabilit de
faire des pertes?
327
b) Quelle(s) action(s) choisiriezvous si vous aviez un niveau d'aspiration d' au
moins 9 000$?
Problme no 3.4
Deux (2) projets A et B ont les profits prvus selon les probabilits montres dans
la matrice cidessous. Lequel choisiriezvous et pourquoi?
A Probabilit
B Probabilit
Profits
0.2
0
11
0.2
0.4
3
Problme 3.5
0.2
0.3
5
0.2
0.2
13
0.2
0.1
21
Il existe plusieurs mthodes pour dtecter des soudures imparfaites. Une
compagnie considre actuellement deux mthodes possibles. La mthode 1 cote
0.50$ par inspection et dtecte les dfauts 80% par temps. La mthode 2 cote
2.00$ par test et dtecte toujours une mauvaise soudure. Lorsqu'une soudure
dfectueuse est non dtecte, la compagnie estime 30.00$ les pertes qu'elle doit
encourir pour le remplacement et la mauvaise publicit qui en rsulte. La
probabilit d'avoir une soudure dfectueuse est de 0.05. Utilisant la mthode de
la "valeur prvue", dtermine quelle mthode devrait tre utilise ou estil
prfrable pour la compagnie de n'en utiliser aucune.
Problme 3. 6
Appliquer les diffrents principes de choix (valeur prvue, l'tat futur le plus
favorable>>, niveau d'aspiration) la matrice suivante; les valeurs de cette matrice
sont des profits.
TATS DE LA NATURE
P
1
=0.1 P
2
=0.7 P
3
=0.2
Actions SI s 2 s3
At
200 10 60
A2
150 30 0
A3
35 35 35
4.1 INTRODUCTION:
Chapitre 4
TECHNIQUES DE DCISION STATISTIQUE
AVEC INFORMA TI ONS ADDITIONNELLES
Plusieurs techniques statistiques ont t dveloppes pour aider l'investisseur prendre une
dcision. Mme en prsence d'incertitude des informations additionnelles peuvent tre obtenues
par exprience. Dans ce chapitre on verra l'importance d'informations additionnelles sur la prise
de dcision dans le cas de risque.
4.2 Statistique Bayesian
Pour tenir compte des informations additionnelles dans le cas de calcul des probabilits, on fait
appel aux statistiques Bayesian. Les statistiques Bayesian sont caractrises par l'ajustement des
probabilits connues priori pour un paramtre inconnu a des probabilits posteriori plus
certaines. Ces probabilits peuvent tre obtenues de simples vidences ou d'tudes supplmentaires.
4.2.1 Thorme de Bayes
La probabilit que 2 vnements se produisent est donne par la thorie des probabilits.
P(A et B) = P(A/B)*P(B) [quation 4. 1]
ou P(A et B) = probabilit que deux vnements A et B se produisent ensemble
P(A/B) = probabilit que A se produise lorsque B s'est dj produit
P(B) = probabilit que B se produise
On a aussi que: P(A et B) = P(B et A)
Donc P(B et A) = P(B/A)*P(A) et il rsulte que:
P(A/B)*P(B) = P(B/A)*P(A)
Le thorme de Bayes peut tre dduit si l'on assume que P(B) et P(A) * 0
P(AIB)
= P(BIA) P(A)
P(B)
[quation 4.2]
[quation 4.3]
P(BIA)
Exemple 4.1
= P (Al B) P(B)
P(A)
[quation . ]
La probabilit d'obtenir l'as de pique d'un jeu de 52 cartes = 1/52 ou peut tre calcul:
P(as et pique)= P(as/pique) P(pique)
= (1 as/13 piques) (13 piques/52 cartes)
= (1/13) (13/52) = 1152
P(as/pique) = P(pique/as
P(piqu
= (1 pique/4 as) (4 as/52 cartes)
(13 piques/52 cartes)
= (114 (1/13) = 1/13
1/4
Donc, en gnral s'il yan tats de la nature S
1
, S
2
, S
3
, Sn et le rsultat d'une tud
additionnelle tel que chantillonner, est X tel que X soit discrte et P(X) y; 0 e 1
probabilits priori P(S;) onft tablies, alors le thorme de Bayes peut s'crire:
P(XIS;) P(S)
P(SJX) = P(X)
[quation .
La probabilit postrieure P(S/X) est la probabilit d'avoir Si si le rsultat d'une d
additionnelle donne X. La probabilit de X et Si de se produire P(X/S;) P(S;) est la
probabilit conjointe de X et Si . La somme de toutes les probabilits conjointes gale la
probabilit de X. Donc l'quation [4.5] peut s'crire:
P(S/X) =
L P(X/S;) P(S;) [quation .6]
43
Le tableau 4.1 reprsente un modle pour 1 'application du thorme de Bayes.
Tableau 4.1: application du thorme de Bayes
1 2 3 4 5
tat de Probabilit Probabilit Probabilit Probabilit
la priori d'obtenir conjointe postrieur
nature 1' chantillon X P(S./X)
SI
s2
S;
sn
P(S
1
) P(X/S
1
) P(X/S
1
) P(S
1
) P(X/S
1
lX(S.
1
}
P(X)
P(S
2
) P(X/S
2
) P(X/S
2
) P(S
2
)
P(X)
P(S;) P(X/S;) P(X/S;) P(S;) P(X/SJ__ffil
P(X)
P(SJ P(X/SJ P(X/Sj P(SJ
P_{X)
n
n n
P(S) =1.
L P(XIS)P(S)=P(X L P(SJX)=
1
i =l i =l
Colonne:
1 S; : les tats de la nature possibles
2 P(S;): la probabilit priori S; estime (la somme gale 1)
3 P(X/S;): la probabilit conditionnelle d'obtenir un chantillon ou des rsultats d'une tude
supplmentaire X connaissant l'tat de natureS;
4 P(X/S;) P(S;): la probabilit commune d'obtenir S; et X; la somme de cette colonne est P(X), la
probabilit que 1 'chantillon ou les rsultats d'une tude supplmentaire soit X.
5 P(S/X): la robabilit postrieure deS;, tant donn qu'une tude additionnelle a produit le
rsultat X; la iime valeur gale la iime entre dans la colonne 4 divise par la somme de la colonne
4.
s.
1
SI
s?
Exemple 4.2
Une production requiert que le procd soit arrang pour fabriquer 200 units. Si les ajustements
sont bien faits, des dfauts dans les units produites se produiront avec une probabilit de o.o.
Un mauvais ajustement avec une probabilit de 0.2 alors le taux des dfauts dans les units
produites est de 0.25. Posant S
1
pour un bon ajustement, S
2
pour un mauvais et X l'vnemen
qu'un chantillon contienne un dfaut. On dsire calculer la probabilit postrieure que les
ajustements soit bien fait ou mal fait tant donn que X s'est produit.
Solution:
Dans ce cas P(S
1
) = O. 8 et P(S
2
) = 0.2
Le tableau 4.2 rsume les calculs:
P(S) P(XIS) P(X/S)P(S) P(S/X)
0.80 0.05 0.04 4/9 = 0.44
0.20 0.25 0.05 5/9 = 0.56
L
P(Si) = 1.0 L =P(X)= 0.09
L
P(S/X) = 1.00
i i i
Tableau 4.2: rsultat de l'exemple 4.8
Donc la probabilit a priori de 0.8 que l'ajustement soit bien fait est rvise, par le fait qu 'on a
trouv une pice dfectueuse, en une probabilit postrieure de 0.44.
Exemple 4.3
Considrons un investissement ayant un retour (valeur prsente nette) de 6 000$ si l'vnemen
1
se produit et 4 000$ si c'est l'vnement S
2
qui survient. Les probabilits priori sont de 0.4 pour
S
1
et 0.6 pour S
2
Donc le retour prvu E(R):
E(R) =0.4 (6 000$) + 0.6 ( 4 000$) = 0
Supposant que la possibilit de ne pas investir un retour prvu nul; donc il y aura alors ni perte
ni gain.
45
Un tude additionnelle permettra d'obtenir soit X
1
montrant un retour de 6 000$ ou X
2
pour une
perte de 4 000$. Si S
1
se produit, X
1
sera indiqu avec une probabilit de 0.8; de mme pour S
2
avec une probabilit 0.6.
Le problme se rsume ainsi:
E(S
1
) = 6 000$
E(S
2
) =  4 000$ P ( X / S ~ = 1.0  P ( X
2
/ S ~ = 0.4
Les probabilits postrieures rsultantes de l'tude additionnelle peuvent tre calculs et la dcision
d'investir ou non peut maintenant tre dtermine en fonction du rsultat X de l'tude additionnelle.
Probabilits postrieures tant donnes que X
1
se produit.
s.
1
P(SJ P(X/ SJ P(X/SJP(SJ P(S/X
1
)
SI 0.4 0.8 0.32 0.32/0.56=0.57 _,
s2
0.6 \ 0.4 0.24 0.24/0.56=0.43
I:=
1.0 L =P(X) = 0.56
I:
1.00
Quand xl se produit la probabilit d'avoir SI change de 0.4 (priori) 0.57 (posteriori). Le retour
prvu devient alors:
E(R/X
1
)=0.57 (6,000$) + 0.43 (4,000$) = 1 714$
Donc si X
1
se produit, le projet devra tre entrepris pour obtenir un retour de 1 714$.
Probabilit postrieures tant donn que X
2
se produit:
si P(SJ P(XiSJ
s, 0.4 0.2
s2 0.6 0.6
L: = Lo
P(X
2
1SJP(SJ
0.08
0.36
L: = o.44
P(S/X
2
)
0.18
0.82
L: Loo
Il rsulte une probabilit de 0.18 d'obtenir S
1
du fait que X
2
se produit; le retour prvu de vien
alors;
E(R/X
2
)=0.18 (6 000$) + 0.82 (4 000$) = 2 182$
Comme E(R/X
2
) est ngatif on prfrera ne pas investir si X
2
se produit prfrant un retour de zro.
Donc on aura:
E(R)=
[
0 si X=X
2
1 714$ si X=X
1
Comme P(X
1
) =0.56 et P(X
2
) = 0.44l'tude additionnelle donnera le retour suivant.
E(R) = E(R/X
1
) P(X
1
) + E(R/X
2
) P(X
2
)
= (1 714$) (0.56) + (0) 0.44 = 960$
Donc en gnral le .retour prvu suite une tude additionnelle (ou chantillonnage) Xj est:
E ( R 1 s 1) = ~ max [E (actions) 1 xi] P(Xj)
J
ou E(R/SI) est le retour prvu tant donn les informations de l'chantillon.
Le changement de la valeur prvu de 0 960$ s'appelle souvent la valeur prvue provenant de
l'chantillon (VPPE) qui se reprsente:
VPPE = E(R/SI)  E(R)
47
Exemple 4.4
Un lot de cinq articles doit tre accept (probabilit a
1
) ou rejet L'tat de la nature
est caractrise par le nombre de probabilits. M d'articles dfectueux parmi les 5; donc M=O,
1, 2, 3, 4 ou 5. Pour simplifier, supposons que les pertes galent deux fois le nombre d'articles
dfectueux (2M) dans un lot qui est vendu; et gal le nombre de bons articles dans un lot qui est
rejet (5M). Dterminez la possibilit Bayes si les probabilits priori sont P(O) = 0.6, P(1) =
0.2, P(2) = P(3) = 0.1, P(4) = P(5) =O.
Solution
Le tableau 4.3, cidessous, rsume les donnes
Tableau 4.3
Donnes pour l'exemple 4.4
Nombre d'article
dfectueux S; = 0 1 2 3 4 5
Action: a
1
(accept)= 0 2 4 6 8 10
Action: (rejet)= 5 4 3 2 1 0
P(S) ( priori)= .6 .2 .1 .1 0 0
Calculons les pertes si l'on accepte et rejete respectivement les lots.
B(a
1
) = 0 * 0.6 + 2 * 0.2+ 0.1 * 4 + 0.1 * 6 + 0 * 8
=1.4
B(a
2
)  5 * 0.6 + 4 * 0.2 + 3 * 0.1 + 2 * 0.1 + 1 * 0 + 0 * 0
= 4.3
Donc, accepter les lots entranera une perte moindre
Maintenant, partir des donnes prcdentes et des probabilits conditionnelles pour les donnes
X tant donnes et l'tat de la nature M cidessous
P(X/M) =
M
5
si X = 1 (mauvais)
1  M
5
si X = 0 (bon)
Calculons successivement les probabilits posteriori et la nouvelle action Bayes.
Sachant que P(X et M) = P(X/M) P(M)
X\M 0 1 2
(bon) 0 1 (.6) = .6 4/5 (.2) = .16 3/5 (.1) = .06
!(mauvais) 1 0 (.6) = 0 1/5 (.2) = .04 2/5 (.1) = .04
X\M 4 5
(bon) 0 1/5 (0) = 0 0 (0) = 0
!(mauvais) 1 l4!5 (0) = 0 1 (0) = 0
Donc, P(X = 0) = 0.6 + 0.16 + 0.04 + 0 = 0.86
P(X = 1) = 0.04 + 0.04 + 0.06 + 0 = 0.14
P(X et M)
Calculant les probabilits posteriori : P(M/X)
P(X)
X\M 0 1 2 3 4
3
2/5 (.1) = .04
3/5 (.1) = .06
5
0 0.6 30 0.16 8 0.06 3 0.04 2 0 0
 =   =0 =0
.86 43 0.86 43 0.86 43 0.86 43 0.86 0.86
1
0 0.04 2 0.04 2 0.06 3 0 0
=0      =0 =0
0.14 0.14 7 0.14 7 0.14 7 0.14 0.14
48
On dsire avoir le minimum de perte; les pertes peuvent tre calcules pour X = 0 c'estdire
lorsque l'chantillon test est bon. Donc, les pertes totales galent la somme des pertes pour les
diffrents tats de la nature.
30 . 8 3 2 40
= (0)+ (2)+ (4)+ (6)+ 0 (8)=
43 43 43 43 43
Donc, si on prend la possibilit a
2
c'estdire si on rejette le lot lorsque l'chantillon test est bo
(X =0) les pertes totales seront:
30 8 3
=5 *  + 4 * + 3 *  + 2
2 195
* 
43 43 43 43 43
40 195
Donc comme  <
43 43
pour X = 0, la possibilit Bayes est a
1
En d'autres mots, lors ue
l'chantillon test est bon, on accepte le lot. De la mme faon si l'chantillon test est mau ais.
on peut calculer les pertes pour les 2 possibilits a
1
et a
2
Pertes pour action a
1
:
49
2 2
+ 6 *
3 30
=0 * 0 + 2 *  + 4
*
 
7 7 7 7
Pertes pour action a
2
:
2
+ 3 *
2
+ 2 *
3 20
=5 * (0) + 4 * 

7 7 7 7
20 30
Donc, comme  <
7 7
, la possibilit Bayes est az pour X = 1, on rejette le lot si
l'chantillon test est mauvais de faon minimiser les pertes.
4.2.2 Mthode de Bayes applique un chantillonnage
Nous avons vu prcdemment qu'une compagnie pouvait raliser une conomie, en testant
un chantillon dans un lot de cinq. Maintenant le management dsire savoir s'il peut
raliser une conomie supplmentaire en testant un deuxime chantillon.
Dans ce cas les probabilits posteriori trouves prcdemment aprs avoir test un
chantillon deviendront les probabilits priori avant de tester le deuxime chantillon. Le
tableau cidessous montre ces probabilits.
X\M 0 1 2 3 4 5
0 30 8 3 2
0 0    
43 43 43 43
1
2 2 3
0 0
0
  
7 7 7
Considrant maintenant qu'un chantillon a dj t test et qu'il reste seulement quatre
chantillons dans le lot. Les relations cidessous caractriseront 1 'tat de la nature lorsque
le testeur aura trouv un premier chantillon bon.
P(X/M)=
M
4
si x2 = 1 (mauvais)
M
P(X/M)= 1   si X
2
= 0 (bon)
4
Par contre lorsqu'il aura trouv un premier chantillon mauvais les relations suivantes
caractriseront 1' tat de la nature.
10
P(X/M)=
M 1
si X
2
= 1 (mauvais)
4
P(X/M)= 1
M1
si X
2
= 0 (bon)

4
M a t dfinit comme le nombre de mauvais chantillon dans le lot initial.
partir de ces relations calculons en premier les probabilits conjointes en utilisant la
relation P (X et M) = P(X/M) P (M) lorsque le testeur a trouv un premier chantillon bon.
X\M 0 1 2 3 4
0 (bon)
1
30 3 8 1 3 1 2 0 * 0
*

*   *

*
43 4 43 2 43 4 43
0.70 0.14 0.035 0.011
0
1 (mauvais)
0
30 1 8 1 3 3 2
0 * 1
*  *  *  *
43 4 43 2 43 4 43
0 0.046 0.035 0.035
0
tant donn que le testeur a trouv un premier chantillon bon:
La probabilit que le deuxime chantillon soit bon gale:
P(X
2
=0) = 0.70 + 0.14 + 0.035 + 0.011 = 0.885
De mme, la probabilit que le deuxime chantillon soit mauvais gale:
P(X
2
= 1) = 0.046 + 0.035 + 0.035 = 0.115
' 1
ons1 erons mamtenant e cas ou
, h 11
e testeur a trouve un premier ec antl on mauv ais.
X\M 0 1 2 3 4
0 (bon) 0 * 0
1
2 3 2 1 3 1
* 0 *  *  *
 
0
7 4 7 2 7 4
0.286 0.214 0.214
0
1 (mauvais) 0 * 0 2 1 2 1 3 3
* 0 0*  
*

* 

0
7 4 7 2 7 4
0 0.071 0.214
0
411
tant donn que le testeur a trouv un premier chantillon mauvais:
La probabilit que le deuxime chantillon soit bon gale:
P(X
2
=0) = 0.286 + 0.214 + 0.214 = 0.714
De mme, la probabilit que le deuxime chantillon soit mauvais gale:
P(X
2
= 1) = 0.071 + 0.214 = 0.286
On peut maintenant calculer les probabilits posteriori que le deuxime chantillon soit
bon ou mauvais tant donn que le testeur a trouv un premier chantillon bon.
X\M 0 1 2 3 4
0 (bon) .70 . 14 .035 .011 0
  . 
 
.885 .885 .885 .885
0.79 0.158 0.0395 0.0124
1 (mauvais) 0 .046 .035 .035 0
  
.115 .115 . 115
0
0.40 0.30 0.30
0
Calculons de la mme faon pour le cas o le testeur a trouv un premier chantillon
mauvais.
X\M 0 1 2 3 4
0 (bon) 0 .286 .214 .214 0
  
.714 .714 .714
0.400 0.300 0.300
1 (mauvais) 0 0 .071 .214 0
   
.286 .286
0 0
0.249 0.748
Considrons maintenant la logique des rsultats. Lorsque le testeur a test un seul
chantillon et qu'ill'a trouv bon, il rsultait une probabilit posteriori de 3/43 (0.0698)
d'avoir deux chantillons de mauvais dans le lot. Si maintenant il teste un deuxime
chantillon et qu'ille trouve encore bon, la probabilit posteriori deviendra (.0395). Il
est donc normal que la probabilit de trouver deux chantillons mauvais dans le lot diminue
du fait qu'on en a trouv un deuxime de bon. On pourrait dmontrer un raisonnement
similaire pour les autres probabilits.
1
Connaissant les probabilits posteriori, calculons les pertes lorsque le testeur trou e u
premier chantillon bon et un deuxime:
a) bon
accepter le lot: 0.79 * 0 + 0.158 * 2 + 0.0395 * 4 + 0.0124 * 6
: 0 + 0.316 + 0.158 + 0.0744 = 0.548
rejeter le lot: 0.79 * 5 + 0.158 * 4 + 0.0395 * 3 + 0.0124 * 2 +0
: 3.95 + 0.632 + 0.118 + 0.024 = 4.72
Le management minimise ses pertes en acceptant le lot.
b) mauvais
accepter le lot: 0.40 * 2 + 0.30 * 4 + 0.30 * 6
: 0.80 + 1.20 + 1.80 = 3.80
rejeter le lot : 0.40 * 4 + 0.30 * 3 + 0.30 * 2
: 1.60 + 0.90 + 0.60 = 3.10
Le management minimise ses pertes en rejetant le lot.
Maintenant calculons les pertes lorsque le testeur trouve un premier chantillon mau ais et
un deuxime:
a) bon
accepter le lot: 0.40 * 2 + 0.30 * 4 + 0.30 * 6
: 0.80 + 1.20 + 1.80 = 3.80
rejeter le lot : 0.40 * 4 + 0.30 * 3 + 0.30 * 2
: 1.60 + 0.90 + 0.60 = 3.10
Le management minimise ses pertes en rejetant le lot.
b) mauvais
On peut immdiatement prvoir que le management rejettera le lot puisqu'ille fai
lorsque le testeur trouve un deuxime chantillon bon.
accepter le lot: 0.249 * 4 + O. 748 * 6
: 0.996 + 4.48 = 5.48
rejeter le lot : 0.249 * 3 + O. 748 ~ 2
413
: 0.747 + 1.49 = 2.24
Le management minimisera ses pertes en rejetant le lot.
Donc, pour minimiser l'ensemble de ses pertes, le management adoptera comme
stratgie de contrle:
Rejeter le lot lorsque le testeur trouve un premier chantillon mauvais et ne pas en
tester un deuxime. S'il trouve un premier chantillon bon il en teste un deuxime;
un mauvais chantillon entranera, le rejet du lot alors qu'un bon chantillon son
acceptation.
Prcdemment l'acceptation du lot sans aucun test entranait des pertes de 1.4. Le
test d'un chantillon permettait de rduire les pertes 1.2. Le test d'un deuxime
chantillon et l'application de la stratgie de contrle cidessus entranera des pertes
de:
0.86 (0.885 * 0.548 + 0.115 * 3.10) + 0.14 * 2.85
0.86 (0.485 + 0.359) + 0.40 = 1.13
Donc le test d'un deuxime chantillon permettra de rduire les pertes de 1.20 1. 13
= 0.07
Dpendamment des cots pour effectuer ces tests le management dcidera
d'effectuer, aucun test un ou deux chantillons par lot.
4.2.3 L'utilisation de donnes dans la prise de dcision
Par donnes, on entend les rsultats d'une ou plusieurs observations que l'on crot
intimement relis l'tat de la nature. La disponibilit de telles donnes va fournir une
certaine illumination sur l'tat de la nature facilitant ainsi le choix d'une action. Une
utilisation intelligente de ces donnes va permettre d'obtenir des pertes infrieures celles
qu'on auraient eues sans leur utilisation. Pour obtenir les donnes, on doit effectuer une
exprience (tude) dont les rsultats doivent tre diffrents selon les divers tats de la nature.
Exemple 4.5
Supposons qu'une personne entrant dans un difice peut voir les lumires sur les planchers
infrieurs et suprieurs d'un ascenseur. Lorsque quelqu'un demande l'ascenseur la lumire,
s'allume et s'teint ds que la porte de l'ascenseur s'ouvre. Les donnes peuvent tre
exprimes ainsi: x = nombre de lumires allumes.
x= 0 :
x= 1:
x= 2:
personne n'a demand l'ascenseur.
l'ascenseur a t demand sur un des planchers et peut ou non tre e
mouvement.
l'ascenseur a t demand sur les deux tages (soussol et premier tage)
peut alors penser que l'ascenseur ne fonctionne pas puisqu'elle a t demand
depuis un certain temps.
Supposons que l'on connaisse les probabilits pour x = 0, 1, 2 pour chaque tat de la na
et le tableau des pertes. Dans ce cas particulier, on pourrait assimiler les pertes l 'ner e
ncessaire pour monter un escalier. Plus on devra monter de niveau plus on perdra de
l'nergie.
Actions
Tableau des pertes
tats de la nature
Fonctionne Ne fonctionne
pas
A
1
(descendre au soussol)
A
2
(monter au premier)
0 6
1 5
Probabilit pour chaque tat de la nature
SI s2
x Fonctionne Ne fonctionne pas
0 .6 .1
1 .3 .4
2 .1 .5
On peut se demander d'o peuvent provenir de telles informations?
Il y a deux sources possibles:
. a) par observation sur une longue priode de temps.
b) de considrations thoriques; exemple; un modle mathmatique qui tiendrait compte
du dbit de personnes entrant dans l'difice et celles utilisant l'ascenseur. On note
qu'on doit aussi faire des dterminations pour tablir les paramtres du modle.
415
Les probabilits postrieures utilisant la relation de Bayes, sachant par 1 'exprience passe
que l'ascenseur fonctionne 7 fois sur 10, c'estdire P(S
1
) = 0.7, P(S:z) = 0.3
P(SJX)= P(X/S) P(Si)
P(X)
tat de la nature
x=O
fonctionne S
1
:
(.6)(.7) 14

(.6)(. 7) + (.1)(.3) 15
ne fonctionne pas S
2
:
(;1)(.3) 1

(.6)(.7) + (.1)(.3) 15
x = 1 x=2
(.3)(.7) 7 (.1) (. 7) 7
 
(.3)(.7) + (.4)(.3) 11 (.1)(.7) + (.5)(.3) 22
(.4)(.3) 4 (.5) (.3) 15
 
(.3)(.7) + (.4)(.3) 11 (.1)(.7) + (.5)(.3) 22
Il nous faut maintenant une rgle pour prendre une dcision. Si la personne entrant dans
1' difice trouve deux lumires allumes
X=2, il sait alors que P(S/2) = 7/22 et P(S
2
/2) = 15/22.
Se servant de ces valeurs les pertes postrieurs peuvent alors tre calcules:
La possibilit a
1
: descendre =
7 15 90
0*+6*=
22 22 22
7 15 82
1*+5*=
22 22 22
La monter =
Donc, pour X = 2 les pertes sont infrieures en prenant la possibilit a
2
c'estdire de
monter.
De mme pour X = 1
La possibilit a
1
: descendre =
7 4
0*+6*
11 11
24
11
La possibilit a
2
: monter =
7 4
1*+5*
11 11
27
11
Donc, pour minimiser les retards on choisit a
1
, descendre
Pour X= 0
ai : 0 * 14 + 6
15
1
*
15
6
15
14 19
~ : 1 *  + 5 *
15 15 15
1
Donc la rgle sera de descendre au soussol s'il n'y a pas de lumire allume ou une seule
lumire et de monter au premier tage si les deux lumires sont allumes.
Donc en utilisant les probabilits priori que 1' ascenseur fonction c'estdire O. 7, on aura
comme perte prvue:
Si on descend: E = 0.7 * 0 + 0.3 * 6 = 1.8
Si on monte: E = 0.7 * 1 + 0.3 * 5 = 2.2
Donc, si on descend toujours, on aura une perte de 1. 8 qui est une possibilit Bayes qui
minimise les pertes.
Maintenant, si on utilise les rsultats de 1 'tude, quelle sera alors la perte?
Comme P(X) = E P(XIS) P(S)
i
P(X = 0) = 0.6 * 0.7 + 0.1 * 0.3 = 0.45
P(X = 1) = 0.3 * 0.7 + 0.4 * 0.3 = 0.33
P(X = 2) = 0.1 * 0.7 + 0.5 * 0.3 = 0.22
Total = 1.0
 (.45) +
24
(.33) +
82
(.22) = 1
11 22
Donc, dans ce cas, le prix que rapporte l'chantillonnage est de:
1.80 1. 72 = 0.08
417
On peut noter que ce problme pourrait tre solutionner en considrant l'ensemble des rgles
possibles. Dans ce problme, il y a deux possibilits: a
1
: descendre et a
2
monter. De plus,
il y a trois valeurs possibles observes: X = 0, 1, 2; donc, il y a 2
3
c'estdire 8
combinaisons ou rgles possibles (d) de dcisions.
x
0
1
2
En calculant les pertes pour chaque rgle possible et en choisissant celle qui donne la plus
faible perte, on obtiendrait le mme rsultat que prcdemment. Mais, on peut immdiate
ment voir que la rgle d
3
signifie: que si on observe 0 ou 2 lumires allumes on choisit la
possibilit a
2
, c'estdire monter et la possibilit a
1
descendre si une seule lumire est
allume. Les rgles d
1
et d
8
ignorent les donnes. On remarque aussi que les rgles d
4
et
d
5
sont opposes.
417
On peut noter que ce problme pourrait tre solutionner en considrant 1' ensemble des rgles
possibles. Dans ce problme, il y a deux possibilits: a
1
: descendre et a
2
monter. De plus,
il y a trois valeurs possibles observes: X = 0, 1, 2; donc, il y a 2
3
c'estdire 8
combinaisons ou rgles possibles (d) de dcisions.
x
0
1
2
En calculant les pertes pour chaque rgle possible et en choisissant celle qui donne la plus
faible perte, on obtiendrait le mme rsultat que prcdemment. Mais, on peut immdiate
ment voir que la rgle d
3
signifie: que si on observe 0 ou 2 lumires allumes on choisit la
possibilit a
2
, c'estdire monter et la possibilit a
1
descendre si une seule lumire est
allume. Les rgles d
1
et d
8
ignorent les donnes. On remarque aussi que les rgles d
4
et
d
5
sont opposes.
418
Problme 4.1
La compagnie Papier dsire augmenter sa production; elle considre actuellement les trois choix
suivants:
cl: faire un grand agrandissement
C
2
: faire un petit agrandissement
C
3
: louer les facilits additionnelles ncessaires
La compagnie demeure incertaine sur son choix, car la demande pour ses produits pourra tre:
leve (S
1
), bonne (SJ, normale (S
3
) ou faible (S
4
). Son dpartement de marketing value les
revenus et probabilits relis chaque niveau de demande de la faon suivante:
Revenus et probabilits de la demande
(Revenus en million de dollars)
Choix leve(S
1
) bonne(SJ normale (S
3
) faible (S
4
)
P(Sl) 0.2 P(S
1
) 0.45 P(S
1
) 0.25 P(SJ_ 0.1
Grand agrandissement (C
1
) 18 12 6 12
Petit agrandissement (C
2
) 12 12 9 6
Location (C
1
) 13 10 6 1
Considrant sa difficult arrter son choix, la compagnie dcide de faire appel un consultant.
Ce dernier s'engage prdire trois niveaux .de la demande avec certaines probabilits; le tableau
suivant montre ses prdictions.
Les probabilits de la prdiction du consultant
Demande future levebonne bonnenormale normalefaible
xl x2 x3
~ l e v e S
1
0.6 0.3 0.1
bonne s2 0.3 0.5 0.2
normale s3 0.1 0.3 0.6
faible S
4
0.1 0.1 0.8
titre d'exemple si le consultant prdit une demande levebonne dans 60% des cas cette
demande se rvlera dans les faits leve P (X
1
1 S
1
) = 0.6
a) Dterminer les revenus prvus pour chacun des choix avant de faire appel au consultant?
419
b) Dterminer les probabilits postrieures l'tude du consultant?
c) Dterminer les revenus prvus aprs l'tude du consultant?
d) Quel est le montant maximal que la compagnie Papier voudra payer le consultant?
Problme no 4.2
Un tudiant possde trois (3) ds; un vert, un rouge et un blanc. Le d vert a cinq (5) faces
marques H et une face marque T, de mme, le d rouge a deux (2) faces marques H et quatre
(4) faces marques Tet le blanc trois (3) faces marques H et trois (3) faces marques T.
L'tudiant tire les trois (3) ds sans que vous puissiez les voir. Il en cache deux (2) et vous informe
que celui non cach montre la face H. Dmontrez l'aide du thorme de Bayes le choix de la
couleur du d que vous feriez pour maximiser vos chances.
Problme no 4.3
Lorsqu'une machine est ajuste correctement 9 sur 10, des articles proluits sont acceptables. Lors
d'un mauvais ajustement, il y a 0.4 chance de produire des articles acceptables. La probabilit que
la machine soit bien ajuste est de 0.95.
a) Si le premier article test aprs un ajustement est non acceptable, quelle est la probabilit
que la machine soit bien ajuste?
b) Si les deux premiers articles tests sont acceptables, quelle est la probabilit que la machine
soit bien ajuste?
Problme 4.4
Nous avons vu, aux cours, l'application de la mthode de Bayes l'chantillonnage (voir notes).
Avant de prendre une dcision finale sur le nombre d'chantillons qu'il faudra tester, votre patron
dsire connatre quelques informations supplmentaires; veuillez lui fournir ces informations.
tant donn que le testeur a trouv un premier chantillon bon et un deuxime mauvais:
a) Quelles seront les relations qui caractriseront 1 'tat de la nature?
b) Quelles seront les probabilits a priori avant de tester le troisime chantillon?
c) Quelle est la probabilit que le troisime chantillon soit bon?
420
d) Dterminer les probabilits a posteriori que le troisime chantillon soit bon.
e) Sans rien calculer, si les cots pour tester ce troisime chantillon sont nuls, quelle est
l'conomie maximale qu'on pourra raliser en le testant? (maximum 5 lignes)
Problme 4.5
La socit lectronique doit dcider si elle va produire un nouveau capteur; elle devra alors investir
5 $ millions. La demande pour ce capteur n'est pas connue. Si la demande tait leve, la socit
ferait un profit de 2 $ millions par anne durant cinq (5) ans. Si la demande est modre, le profit
serait de 1,6$ million par anne durant quatre (4) ans. Pour une demande faible, le profit ne serait
que de 0,8 $ million par anne durant quatre (4) ans.
On estime 10% les chances d'une faible demande et 50%, celles pour une demande leve. Le
taux d'intrt est actuellement de 10%.
a) Utilisant la mthode de la valeur prsente, la socit doitelle faire l'investissement?
b) L'un des cadres dsire que l'on fasse une tude de march; quel est le montant maximal que
l'on pourrait tre prt dpenser pour cette tude?
c) Un consultant propose une tude au cot de 75 000$ qui a les caractristiques suivantes?
Si la demande est
leve
modre
faible
L'tude indiquera favorable ou nonfavorable
avec les probabilits
favorable '&.
0.9
0.5
0
nonfavorable
0.1
0.5
1.0
Estce que la socit doit demander l'tude?
d) Si l'tude indique nonfavorable, la socit doitelle produire le capteur? Quel est le profit
prvu dans ce cas?
421
Problme 4.6
Un diteur d'une maison d'dition doit dterminer s'il va accepter ou non un manuscrit pour
publication. Il a dj dpens 1 000$ pour le dveloppement de ce manuscrit. Il a le choix entre:
 A 1: rejeter le manuscrit et perdre 1 000$
A 2: accepter le manuscrit sans consulter un expert extrieur
 A 3: obtenir l'avis d'un expert extrieur au cot de 800$
Les donnes sont rsumes dans le tableau cidessous:
Dcision et rsultats sivaluation de l'diteur valuation de l'expert Retour si
2ublis {10 c a s ~ ~ 2 0 c a s ~ _Qubli
Mauvais Acceptable Trs bon $
MO : pas publier 4 9 0 0 0
Ml : demande faible 3 3 2 0 10 000
M2 : bonne vente 2 1 1 1 50 000
M3 : grand succs 1 0 2 1 200 000
Il y a trois niveaux de succs possibles si le livre est publi: Ml, M2 et M3. La dernire colonne
donne la valeur prsente des revenus moins les cots de publication pour chaque condition de
march. Le tableau indique aussi les dcisions antrieures prises aprs l'avis de l'expert. L' diteur
a accept 6 des 10 derniers manuscrits soumis; un a eu un trs grand succs. L'expert en a class
20 comme mauvais, acceptable et trs bon et charge 800$ par tude.
a) Quel est le profit prvu si l'diteur dcide sans consulter l'expert?
b) Quel est le profit prvu aprs consultation de l'expert?
c) Quel est le montant maximal que l'diteur sera prt payer pour l'expert?
Problme 4.7
Un inventeur a dvelopp un instrument pour mesurer le taux de bchettes dans le papier. Une
compagnie considre actuellement la possibilit d'acheter les brevets relis cet instrument pour
un montant de 40 000$. Elle pense qu'il y actuellement une chance sur cinq que l'instrument puisse
tre commercialis avec succs; dans un tel cas, elle obtiendrait des revenus nets de 150 000$ par
anne durant les cinq prochaines annes. Dans les autres cas l'instrument ne rapportera aucun
revenu. Cette compagnie utilise un taux d'intrt annuel effectif de 15%.
422
a) La compagnie doitelle l'acheter?
b) Si un dpartement de marketing peut fournir une information avec certitude du march
futur, combien la compagnie pourrait tre prte payer pour ce service?
c) Ce mme dpartement de marketing peut aussi faire une tude de march avec une
probabilit de 0.7 que si le rsultat de l'tude s'avre positif la commercialisation sera un
succs et une probabilit de 0.9 que si l'tude montre un rsultat nonfavorable l'instrument
ne pourra tre commercialis. Dterminer combien la compagnie voudra payer pour cette
dernire tude?
Chapitre 5
V ARBRE DE DCISION
Dans les problmes dcisionnels l'analyste doit reconnatre que des dcisions prises immdiatement
influenceront directement celles qui devront tre prises dans le futur. Trs souvent les industriels
prennent des dcisions sans considrer les effets long terme; une dcision qui semble bonne
actuellement pourrait placer la compagnie dans une position dfavorable et parfois non comptitive par
rapport d'autres prendre dans l'avenir. L'utilisation des arbres de dcision s'avre un moyen trs
efficace pour reprsenter l'interaction entre une squence de dcision. Ce diagramme des vnements
futurs qui pourraient influer sur la dcision prsente permet d'clairer le dcideur sur les choix faire
en prsence de risques pour atteindre les objectifs fixs.
L'arbre de dcision explicite graphiquement la squence des dcisions prendre et les divers
vnements qui peuvent arriver. Il montre les mmes informations qu'un tableau de cots mais d'une
faon beaucoup plus clairement, particulirement dans le cas de situations complexes.
Pour Belzile et al (1) la procdure consiste :
bien se situer dans le temps;
dfinir les vnements possibles qui peuvent arriver (tenir compte de l'chelle de temps);
dterminer les actions qui peuvent tre entreprises;
dterminer la valeur en dollars ou en utilit de chaque action combine avec l'vnement;
associer chaque vnement une probabilit d'occurrence;
trouver la valeur espre de chaque solution possible;
choisir la possibilit qui offre le rsultat espr le plus lev (en dollars ou en utilit).
Graphiquement l'arbre se compose d'une srie de noeuds et de branches. Chaque branche reprsente
une action possible. Le noeud localis l'extrmit gauche de chaque branche reprsente une chance
que l'vnement ou l'tat de la nature se produise. Toute autre branche la droite du noeud reprsente
d'autres actions possibles associes aussi un tat de la nature reprsent par le noeud. Un cot, un
profit ou une utilit correspond chacune des actions reprsentes; il est indiqu l'extrmit droite
de chacune des brariches terminales.
Dans le prsent texte les carrs reprsentent les points de dcision et les cercles les points de chance.
Un arbre de dcision comprendra donc toujours une combinaison d'actions possible entirement sous
le contrle du dcideur relies par des points de chance, lesquels sont soumis aux des lois des
probabilits dont la dcision ne possde aucun contrle sur leur occurrence.
L'arbre de dcision contiendra seulement les actions et les tats de la nature qui ont une importance
pour l'investisseur et dont les consquences dsirent tre compares. De plus, cette approche assume
comme hypothse qu'aucun changement important n'influencera les diffrentes situations dans le
temps.
52
Un arbre de dcision ne donne pas au dcideur la rponse un problme dcisionnel mais l ' aide plut
dterminer quelle action en un point particulier du temps va produire l'utilit ou le profit le plus
lev. Il aide ainsi visualiser l'information pour prendre une dcision. Le dcideur devra toujours
considrer les gains prvus en fonction des risques encourir. La nature de ces risques, dpendant
de la faon qu'illes percevra, influencera non seulement les hypothses qu'il a faites mais aussi la
stratgie suivre pour en tenir compte.
La mthode de construction d'un arbre de dcision peut ~ e rsume ainsi :
1. identifier les points de dcision et les actions prsentes chacun des points;
2. identifier les points de risque et la valeur des revenus ou utilits qu'engendreront les actions
chaque point;
3. estimer les valeurs ncessaires pour faire l'analyse; plus spcifiquement dterminer les
probabilits des diffrents tats de la nature, les cots, les gains ou les utilits associs
chacune des actions;
4. analyser les valeurs des choix pour prendre une dcision.
Solution d'un arbre de dcision
En pratique, l'analyste utilise une procdure contre courant pour analyser le problme reprsent par
un arbre de dcision. Il effectue l'analyse en partant aux bouts des branches et en se dplaant ers
le tronc en appliquant les deux rgles suivantes :
1. s'il rencontre un point de chance, c'estdire associ un tat de nature, il calcule la valeur
prvue du cot, du gain ou de l'utilit en se basant sur les valeurs places la droite du point;
2. s'il rencontre un point de dcision, c'estdire une action prendre, il choisit le profit ou
l'utilit maximale ou les cots minimaux en considrant les branches de droite adjacentes au
point.
En procdant de cette faon, il peut liminer certaines actions de considration future; ceci est
particulirement important dans le cas de situation complexe qui exige la construction d'arbre trs
grand et l'utilisation d'ordinateurs pour leur solution.
Rfrence
1. Belzile R., Mercier G., Rassif F., <<Analyse et gestion financire Presse universit du Qubec,
Qubec 1989.
53
Arbre d dcision
Exemple
La socit PAPIER VERT dsire fabriquer un nouveau type de papier sanitaire, la fois
plus conomique et plus cologique, partir d'un nouveau procd de fabrication de pte
mcanique. Malgr les tudes techniques et l'analyse de march effectues, la socit
n'est pas certaine si la qualit de son produit permettra un volume de vente important.
La socit considre actuellement deux possibilits; la premire consiste construire un
atelier de fabrication de pte mcanique de 500 t/d et de tenter de prendre une part
importante du march. La deuxime possibilit serait de construire une usine de
seulement 250 t/d et de faire une nouvelle tude aprs deux ans. Si la demande est
suffisamment leve, la socit construirait une autre unit de 250 t/d pour minimiser
l'influence de ses cots indirects de fabrication et maximiser ses profits.
Par contre, si la demande pour ce nouveau produit tait faible, la socit se retirerait du
march et rorienterait la fabrication de ses 250 t/d vers la fabrication de papiers ayant
une valeur ajoute infrieure.
Les rsultats de l'analyse de march font voir qu'il y a 60 % de chance que la demande
initiale soit leve et demeure leve durant vingt ans et 10 % de chance qu'elle soit
initialement leve et qu'elle devienne faible aprs deux ans; donc, il y a 70 %de chance
que la demande initiale soit leve pour ce type de papier et 30 % qu' elle soit faible dans
deux ans. Il y a 86 % (60/70) des chances qu'une demande leve resterait leve et
seulement 14% ( 1 0/70) qu'elle devienne faible. Si, avec toute la publicit qui entourerait
le lancement de ce nouveau produit, la demande initiale tait faible, elle demeurerait
certainement faible.
Pour sa part, le dpartement de gnie industriel de la socit estime que les cots de
construction seraient respectivement de trente et soixante millions de dollars pour un
atelier de 500 et 250 t/d. L'addition de 250 t/d, aprs deux ans, coterait vingt millions
de dollars.
Le tableau 1 fait voir l'estim des profits selon que l'on aurait une capacit de production
de 250 et 500 t/D et que la demande soit leve ou faible.
La figure 1 illustre, sous la forme d'un arbre de dcision, l'ensemble des informations.
Cette reprsentation n'apporte par de solution particulire au problme de la socit
PAPIER VERT, mais elle aide ses dirigeants visualiser la suite des dcisions prendre.
Cette socit exige un taux de retour minimum de 10 5/6 sur cet investissement qui
permettrait de consolider sa position sur les marchs internationaux.
54
L'analyse utilise la procdure contrecourant pour solutionner cet arbre, en commenant
au point de dcision 2 et en attribuant une valeur montaire qui lui permettra de prendre
une dcision au point 1. Le tableau 2 illustre l'analyse de la dcision au point 2 et fait
voir les valeurs prsentes prvues.
Donc, si la socit dcide maintenant, c'estdire au point 1, de construire l'atelier de
250 t/d, elle tiendra compte que l'addition de 250 t/d aprs deux ans lui procurerait un
profit net de quarantequatre millions de dollars.
La figure 2 illustre l'arbre de dcision simplifi et ne contenant plus que le point de
dcision 1. Multipliant les valeurs prsentes des revenues prvues par les probabilits.
L'analyste obtient les profits pour la construction d'une usine de 500 t/d et d'une usine
de 250 t/d et cela, aprs que la dcision au point 2 a t prise de procder une addition
de 250 t/d dans les deux ans. Le tableau 3 illustre les calculs des valeurs prsentes
prvues au point de dcision 1.
La construction immdiate d'un atelier de 500 t/d entranerait:
0.6 x 85.1 + 0.10 x (17.3 + 6.78) + 0.30 x 8.5130 = 26.1
Celle d'un atelier de 250 t/d:
o. 7 x {7.81 + 36.4) + 0.3 x 29.8 16 = 23.9
Donc, la maximisation du revenu proviendrait de la construction d'un atelier de 500 t/d
immdiatement. Avec ces informations et d'autres pertinentes, le management de cette
socit pourra prendre un dcision finale.
Tableau 5.1
Estimation des profits annuels en fonction de la
grosseur de l'atelier et du volume de la demande
Atelier tat de la nature Revenu 1 an
capacit, t/d de la demande (millions $)
leve 10
500 faible 1
leve 4.5
250 faible 3.5
leve 9
250 + 250 faible 0.5
55
Si la demande reste faible, la socit rorientera la production et le profit annuel
ne sera que de 3.5 millions de dollars.
56
Tableau 5.2
Les valeurs prsentes prvues au point de dcision 2
Revenu
tat de la an Valeur Valeur
Choix nature de Probabilit (millions $) prsente prvue
la demande durant (millions $) escompte
18 ans (millions$}
ajouter 250 Ud leve 0.86 9 73.8 63.5
ajouter 250 Ud faible 0.14 0.5 4.10 0.57
Total 64.1
 l'investis.
.2.L
NETTE 44.1
pas de leve 0.86 4.5 36.9 31.7
changement
pas de faible 0.14 3.5 28.7 4.02
changement
Total 35.7
 l'investis. 0
NETTE 35.7
Tableau 5.3
Les valeurs prsentes prvues au point de dcision 1
tat de Revenu Revenu Valeur
Choix la nature Probabilit an Nombre escompt prvue
de la (millions $) d'annes (millions $) escompte
demande (millions $)
leve 0.60 10 20 85.1 51 .1
initiale 0.10 10 2 17.3
leve et 2.41
aprs faible 1 18 6.78
faible 0.30 1 20 8.51 2.55
Total 56.1
 l' investis.
~
'
NETTE 26.1
leve 0.70 4.5 2 7.81
et
valeur de la
dcision au
point 2 31 .0
44.1 la 36.4
fin de la 2
9
anne
faible 0.30 3.5 20 29.8 8.94
Total 39.9
 l'investis. 16
NETTE 23.9
   
58
Demande Probabilit Revenu/an Nombre
d'annes
leve
0.6 10
20
leve
0.1
10 2
Faible 1
18
Faible 0.3 1
)
20
c? 0"'
t':l"''lle
0.86 9
18
<:!
'.__.../
0.14 0.5
18
"
)
19
0.86 4.5
i:\e'lle
18
'7.
6'
' / i:I/Ofe
/:'l/6
0.14 3.5
18
v&o
.a
"
Faible 0.3 3.5
20
0 2 20
ans
Figure 5.1 L'arbre de dcision pour la construction d'un atelier de fabrication de pte mcanique
Demande Probabilit Revenu/an Nombre d'annes
0.6 10 20
/
leve initiale 0.1 10
2
)
Faible aprs
1 18
:.<..0
0
"6;$
cP
0.3 1 20
Provenant de
l'addition de
.t.:
250 tld aprs 2 ans
'?k
(\)d'
(9
0.7 2 \$'(9. +4.5
., ....
'7..
6'
0.3 3.5 20
Figure 5.2 L'arbre de dcision simplifi: aprs que la dcision a t prise de procder l'addition de 250 t/d aprs 2 ans
510
Problme 5. 1
Un jeune dbrouillard a dcid de tirer avantage de ses connaissances en statistiques
pour faire quelques dollars durant ses vacances. Il s'est achet deux chapeaux
identiques; dans le fond intrieur il a crit X dans l'un et Y dans l'autre. Dans le chapeau
marqu d'un X, il dpose 5 balles de golf; 3 blanches et 2 jaunes, alors que dans celui
identifi par un Y, il y dpose aussi 5 balles, mais cette fois, 4 jaunes et 1 blanche.
Il propose des amis les paris suivants:
Payer 10 $ et identifier le chapeau marqu d'un X. S'il l'identifie correctement, il
gagnera 8 $,s'il se trompe, il perdra son 10 $;
Payer 2 $ et tirer une balle de l'un des chapeaux, aprs qu'il ait vu si elle est
blanche ou jaune il peut ou non payer 10 $ et tenter d'identifier le chapeau marqu
d'un X. Il pourrait dcider de tirer 2 ou 3 balles et payer 2 $ pour chacune d'elles
avant de payer 10 $ pour identifier le chapeau . L'ami a toujours la possibilit de
refuser le pari.
a) Tracer un arbre de dcision pour montrer toutes les possibilits du jeux.
b) Solutionner l'arbre et dterminer la possibilit la plus intressante pour l'ami.
Problme 5.2
La compagnie Boisbriand dsire investir 100 millions de dollars dans la construction d'un
nouveau complexe rsidentiel. Elle prvoit des revenus de la location de ces rsidences
qui varieront avec la situation conomique de la province. Le tableau 1 dcrit ces
prdictions.
tat de la nature
de l'conomie
Prospre 0 0. A
Normal . .... . ... ..... B
Rcession 0 c
Tableau 1
Variation des revenus nets
Valeur actuelle des revenus
nets  Mill ions $
140
120
50
Probabilit
d'occurrence
0.2
0.5
0.3
5 11
Comme la compagnie n'a pas encore pris la dcision finale d'investir, elle cons idre la
possibilit d'engager un consultant . Ce dernier exige un montant de 100 000 $ et
s'engage prdire l'tat de la nature avec certaines prcisions. Ces dernires
proviennent des rsultats de ses tudes prcdentes. Le tableau 2 rsume ces
prdictions.
Prdictions
du consultant
An
Bn
en
Tableau 2
Les probabilits que le consultant
a dj obtenues d'tudes similaires
Les probabilits d'obt enir
les tats de la nature
A 8
0.8 0.1
0 . 1 0.9
0 . 1 0.0
c
0.1
0.2
0.7
a) Dmontrer, l'aide d'un arbre de dcision, si a compagnie doit faire
l'investissement en se basant sur ses propres prdictions .
b) Estil avantageux pour la compagnie d'engager le consultant?
c) Quel est le montant maximal que la compagni 'e pourrait payer un consultant?
d) Si une information parfaite pouvait tre obtenue, combien la compagnie pourrai t
elle payer?
 
13.3 SEQUENTIAL DECISION MAKING: USE
OF DECISION TREES
Decision trees, also commonly called decision flow networks and decision
diagrams, are powerful means for depicting and facilitating the analysis of
important problems, especially those that involve sequential decisions and
variable outcomes over time. Decision trees have great usefulness in prac
tice because they make it possible to look at a large complicated problem in
terms of a series of smaller simple problems and they enable objective anal y
sis and decision ma king which includes explicit consideration of the risk and
effect of the future.
The name decision tree is descriptive of the appearance of a graphical
portrayal, for it shows branches for each possible alternative for a given
decision and branches for each possible outcome (event) which can result
from each alternative. Such networks reduce abstract thinking to a logical
visual pattern of cause and effect. When costs and benefits are associ ated
with each branch and probabilities are estimated for each possible outcome,
then analysis of the How network can clarify choices and ri sks.
13.3.1 Deterministic Exemple
The most basic form of decision tree occurs when each alternative can
be assumed to result in a single outcomethat is, when certainty is as
sumed. The replacement problem in Fig. 134 illustrates this. The problem
as shown retlects that the decision on whether to replace the old machine
with the new machine is not just a onelime decision, but rather one which
recurs periodically. That is, if the decision is made to keep the old machine
at decision point 1, then la ter, at decision point 2, a choice again has to be
made. Similarly, if the old machine is chosen at decision point 2, then a
choice again has to be made at decision point 3. For each alternative, the
cash inflow and duration of the project is shown above the arrow and the
cash investment opportunity cost is shown below the arrow.
For this problem, one is concerned initially with which alternative to
choose at decision point 1. But an intelligent choice at decision point 1
should take into account the later alternatives and decisions which stem
from it. Hence, the correct procedure in analyzing this type of problem is to
start at the most distant decision point, determine the best alternative and
quantitative result of that alternative, and then " roll back" to each sucees
Vl
N
342
1.....
x
".s:
.9J,..
Risk and Uncertainty
% ...
.s:
'.r
': 6'
J,..
Old + $3M/yr, 3 yr
3yr
1< $2M
J:
Figure 134. Deterministic replacement example.
Chap. 13
sive decision point, repeating the procedure until finally the choice at the
initial or present decision point is determined. By this procedure, one can
make a present decision which directly takes into account the alternatives
and expected decisions of the future.
For simplicity in this example, timing of the monetary outcomes will
first be neglected. which means that a dollar has the same value regardless of
the year in which it occurs. Table 135 shows the necessary computations
and decisions . Note that the monetary outcome of the best alternative at
decision point 3 ($7.0M for the "old") becomes part of the outcome for the
"old" alternative at decision point 2. Similarly, the best alternative at deci
sion point 2 ($22.0M for the "new") becomes part of the outcome for the
"old " alternative at decision point 1.
Table 135
MONETARY UTCOMES AND D ECISIONS AT EAOI POI NTDETERMINISTIC
REPLACEMENT EXAMPLE OF FIG. 134'
Decision
Point Alternative Monetary Out come Choice
3
{
Old S3M(3)  $2M = S 7.0M Old
New S6.5M(3) SISM = S 1.5M
{
Old $7M + $3.5M(3)  SIM = Sl6.5M
2
New $6, 5M(6)  Sl7M = S22.0M New
{
Old S22.0M + S4M(3)  SO.SM = $33.2M Old
New S5M(9) $15M = SJO.OM
lnterest = 0%; that is. ignoring timing.
Sec. 13.3
Sequential Decision Making: Use of Decision Trees
343
By following the computations in Table 135, one can see that the
answer is to keep the "old" now and plan to replace it with the "new" at the
end of 3 years (at decision point 2). But this does not mean that the old
machine should necessarily be kept for a full 3 years and then a new machine
bought without question at the end of 3 years. Conditions may change at any
time, thus necessitating a fresh analysisprobably a decision tree analy
sisbased on estimates which are reasonable in light of conditions at that
later time.
13.3.1.1 Dctenninistic Example Considcring Timing
For decision tree analyses, which involve working from the most dis
tant decision point to the nearest decision point, the easiest way to take into
account the timing of moncy is to use the present worth approach and th us
discount ali monetary outcomes to the decision points in question. To dem
onstrate, Table 136 shows computations for the same repl acement problem
of Fig. 134 using an interest rate of 25% per year.
Note from Table 136 that when taking into account the effect of timing
by calculating present worths at each decision point, the indicated choice is
not only to keep the "old" at decision point 1, but also to keep the " old'' at
decision points 2 and 3 as weil. This result is not surprising sincc the high
interest rate tends to favor the alternatives with lower initial investments,
and it also tends to place Jess weight on longterm returns .
3.3.2 Typical Example with Random Outcomes
The following is a brief description of a typical automation problem for
which the analytical capabilities of decision trees are very useful . A decision
maker must decide whether to automate or not to automate a given pro
cess. He is uncertain whether the results will turn out to be poor. fair, or
excellentdepending on the technological success of the automation proj
ect. The net payoffs for possible outcomes (expressed in net present
worths) are $90M, $40M, and $300M, respectively. The initially estimatcd
probabilities thal each outcome will occur are 0.5, 0.3. and 0.2. respec
tively. Figure 135 is a decision tree depicting this simple situation. Table
137 shows calculations to determine that the best choice for the firm is to
Figure 1:35. Automation problem di a
gram before consideration of techno(ogy
study. ($ are in net PW)
Poor S  90M (0.5)
Fair S40M 10.3)
Exc. SJOOM (0.2)
Not automate
10
Y'
......
.......
Decision
Point
3
{
{
{
Risk and Uncertainty Chap. 13
Table 136
DECISIONS I'IT EI'ICH POINT WITH INTEREST = 25% PER YEAR FOR
DETERMINISTIC REPLACEMENT EXAMPLE OF fiG. 134
Alternative PW of Monetary Outcome
Old $3M(P/A,3)  S2M
S3M( 1. 95)  $2M = S3.85M

New $6.5M(PIA.3) $18M
$6.5M(I.95) SI8M = $5.33M
Old $3.85(PIF.3) + S3.5M(P/A.3) SIM
$3.85(0.512) + SIM = $7.79M
New $6.5M(PIA,6) $17M
S6.5M(2.95)  SI7M = $2.18M
Old S7.79M(P/F.3) + $4M(P/A.3) $0.8M
$7.79M(0.512) + $4M(I.95)  $0.8M SI0.99M
New $15M
$5.0M(3.46)  SI5M = S2.30M
Table 137
EXPECTED NET PW CALCULATION FOR THE AUTOMATION
PROBLEM BEFORE CONSIDERATION OF TECHNOLOGY
FEASIBILITY STUDY
Automate: S90M(0.5) + $40M(0.3) + $300M(0.2) = $27M
Not automate : = $0
Choice
Old
Old
Old
automate based on an expected* net PW of $27M versus $0 if it does not
automate. Nevertheless. this may not be a clearcut decision because of the
risk of a $90M Joss and because the decision maker might reduce the risk by
obtaining further information.
Suppose that it is possible for the decision maker to make a further
technology study at the PW cost of $10M. The study will disclose that the
enabling technology is either "shaky," "promising," or "solid."
lnstead of using Bayesian statistics for revision of probabilities at this
point (this topic will be covered in Section 13.3.5 and Appendix 13A), let us
assume that the probabilities of the various possible automation outcomes,
given the technology study outcomes, are as shown in Fig. 136, which is a
decision tree diagram for the entire problem.
Expected (longrun average. or mean) monetary value will be used as a choice criterion
for problems involving probabilistic monetary outcomes al most throughout this book. [Excep
tion: Where certain monetary equivalents are used, as in Section 13.4].
Sec. 13.3 Sequential Decision Making: Use of Decision Trees
345
Poor $90M (0.6)
Poor 100M (0.73)
Fair $40M 10.3) Fair $40M 10.22)
Exc. $300M 10.05)
$0 Not automate
"\
Poor $90M (0.43)
.l'of, "o/. Fair $40M (0.34)
Q,
*
.
") "l q"' .J\tt'e Exc. $300M (0.23)
o, "o:
"), 'J.
""'
'J. '_,

Not automate
Poor $  90M 10.21)
Fair $40M (0.37)
Exc. $300M (0.42)
Figure 136. Automation example with technology study laken in consideration.
($are in net PW be fore consideration of any cost for "Technology Study").
10
10
10
The tree diagram shows expected future events (outcomes), along with
their respective cash flows and probabilities of occurrence. The rectangular
blocks represent (decision) points intime at which the decision maker must
elect to take one and only one of the paths (alternatives) considered. These
decisions are normally based on a quantifiable measure, su ch as money,
which has been determined to be the predominant "cost" or "reward" for
comparing alternatives. The general approach is to find the action or alter
native that will maximize the expected net PW equivalent of future cash
flows at each decision point, starting with the furthermost decision point(s)
and then "rolling back" until the initial decision pointis reached .
The circular (chance) nodes represent points at which uncertain events
(outcomes) occur. At a chance node the expected value of ali paths leading
(from the right) into the node can be calculated as the sum of the anticipated
value of each path multiplied by its respective probability. (The probabilities
of ali paths leading into a node must sum to 1.) As the project progresses
through time, the chance nodes are automatically reduced to a single out
come based on the "state of nature" which occurs at that time.
The solution to the problem in Fig. 136 is shown in Table 138. It can
be noted that the alternative "technology study" is shown to be best with an
expected net PW of $34.62M.
346
Poinl
2A
28
2C
Risk and Uncertainty Chap. 13
Table 138
ExrF.\THl NFT PW CALCUl.ATIONs FOR THE AuTOMATION PROBLEM WITH
CoNSIDERATION oF TECHNOLOGY STUDY
Alternai ive
{Automate
Not automate
{Automate
Not automate
{Automate
Not automate
{ '"'".'" Not automate
Technology
study
Expecled Net PW
$90M(0.73) + $40M(0. 22)
$ + JOOM(0.05) = $4 1.9M
$0
S 90M(0.43) + S40M(0.34)
$+ 300M(0. 2Jl = $43.9M
so
$ 90M(0.21) + $40M(0.37)
$ + JOOM(0.42) = $12l.9M
$0
From labie 1 J7 $27M
From table 1 37 $0
$0(0.41) + $43. 9M(0.35)
$+ 121.9M(0.24)  $10M = $34.62M
Choice
Not automate
Automate
Automate
Technology
study
13.3.3 Decision Tree Steps
Now thal decision trees (diagrams) have been introduced and the me
chanics of using the diagrams to arrive at an initial decision have been
illustrated, the steps involved can be summarized as follows :
1. Identify the points of decision and alternatives available at each point.
2. Identify the points of uncertainty and the type or range of possible
outcomes at each point (layout of decision flow network) .
3. Estimate the values needed to make the analysis, especially the proba
bilities of different outcomes and the costs/returns for various out
cornes and alternative actions.
4. Analyze the alternatives, starting with the most distant decision
point(s) and working back, to choose the best initial decision.
The example above used the expected net PW as the decision crite
rion. However. if outcomes can be expressed in terms of utility units, the
decision ma ker can use the expected utility as a decision criterion. Alterna
tive) y, the decision maker may be willing to express his or her certain mone
tary equivalent for each chance outcome node and use thal as his or her
decision criterion as explained in Sections 13.4 and 13.5.
Because a decision diagram can quickly become discouragingly, if not
unmanageably, large, it is generally best to start out by structuring a problem
Sec. 13.3 Sequential Decision Making: Use of De_cision 347
simply by considering only major alternatives and outcomes in order to gel
an initial understanding or "feel" for the problem. Then one can develop
more information on alternatives and outcomes which seem sufficiently im
portant to affect the final decision until one is satisfied thal the study is
sufficiently complete in view of the nature and importance of the problem
and the lime and stndy resources available .
13.3.4 General Principles of Diagramming
The proper diagramming of a decision problem is, in itself, general! y
very useful to the understanding of the problem, and it is essential to correct
subsequent analysis.
The placement of decision points and chance outcome nodes from the
initial decision point to the base of any later decision point should give a
correct representation of the information thal will and will not be available
when the decision maker actually has to make the choice represented by the
decision point in question. The decision tree diagram should show the fol
lowing:
1. Ail initial or immediate alternatives among which the decision maker
wishes to choose
2. Ali uncertain outcomes and future alternatives thal the decision maker
wishes to consider because they may directly affect the consequences
of initial alternatives
3. Ali uncertain outcomes thal the decision maker wishes to consider
because they may provide information thal can affect his or her future
choices among alternatives and hence indirectly affect the conse
quences of initial alternatives
It should also be noted !hat the alternatives at any decision point and
the outcomes at any outcome node must be:
1. Mutually excl usive; thal is, no more than one can possibly be chosen .
2. Collectively exhaustive; that is, sorne one must be chosen or some
thing must occur if the decision point or outcome node is reached.
Figure 136 retlects these points. For example, decision points 2A, 28 ,
and 2C are each reached only after one of the mutually exclusive results of
the technology study are known; and each decision point retlects ali alterna
tives to be considered at thal point. Further, ali possible outcomes to be
considered are shown as evidenced by the fact thal the probabilities sum to
1.0 for each chance node.
'h
1
..._
'\
48
Risk and Uncertainty
13.3.5 Use of Statistics to Evaluate the Worth
of Further lnvestigative Study
Chap. 13
As was included in the automation example above, one alternative that
frequcntly exists in an investment decision problem is further research or
investigation before deciding on the investment. This means making an in
tensive objective study. hopefully by a fresh group of people. It may involve
su ch aspects as undertaking additional research and development study,
making a new analysis of market demand, or possibly studying anew future
operating costs for particular alternatives.
The concepts of Bayesian statistics pro vide a means for utilizing subse
quent information to modify estimates of probabilities and also a means for
cstimating the economie value of further investigation study.
To illustrate, consider the onestage decision situation shown in Fig.
137 in which each alternative has two possible chance outcomes: "high" or
"low" demand. lt is estimated that each outcome is equally likely to occur,
and the monetary result expressed as PW is shown above the arrow for each
outcome. Again. the amount of investment for each alternative is shown
below the respective !ines. Based on these amounts, the calculation of the
expected monetary values (in net PW) is shown in Table 139, which indi
cates that the "new FMS" should be selected.
To demonstrate the use of Bayesian statistics, suppose that one is
considering the advisability of undertaking a fresh intensive investigation
before deciding upon the "old system" versus the "new FMS." Suppose
also that this further study would cost $2.0M. To use the Bayesian ap
proach. it is necessary to assess the conditional probabilities that the investi
gation or "technology study" will yield certain results. These probabilities
reflect explicit measures of management's confidence in .the ability of the
investigation to predict the outcome. Sample assessments are shown in Ta
ble 131 O. As an explanation, P(ltill) means the probability that the pre
dicted demand is "high" (Ir), given that the actual demand will turn out to be
"high" (Il) .
PW of returns
H $45M (0.5)
$27.5M (0.5)
H $SOM (0.5)
$48M (0.5) Figure 137. Onestage FMS replace
ment problem.
Sec. /3.3 Sequential Decision Making: Use of Decision Tree.t 349
Appendix 13A contains a formai statement of Bayes' theorem as weil
as a tabular format for ease of calculations in the discrete outcome case.
Tables 131 1 and 1312 use this format for revision of probabilities based on
the data in Table 1310 and the prior probabilitics of0.5 that the dcmand will
be high and 0.5 that the demand will be low.
Table 139
EXPECTF.D NF.T PW FOR PRORLHI IN FIG. 13 7
Old system: S45M(0.5) + $27.5M(0.5)  SJOM S21i.25M
New FMS: S80M(0.5) + $48M(0.5)  $35M S29.0M
(1)
State
(Actual
Demand)
H
L
( 1)
State
(Actual
Demand)
H
L
Table 1310
MANAGEMF.NT's AssESSMENT OF CoNFIDENCE
IN INVESTIGATION RESULTS'
P(hj HJ O. 70
P(/1jL) = 0.20
= 0.30
= 0.80
lnvestigationpredictcd dcmand: h , high;
/, low. Actual demand: Il, high; L, Jow.
Table 1311
CoMPUTATION OF PoSTERIOR PRORABIUTIF.S THAT
(NVESTIGATIONPREDICTED DEMAND ls HIGII Ch)
(2) (3) (4) (5)
= (2)(3) = (4)(4)
Prior Confidence Posterior
Probability, Assessment. Joint Probability,
P (State) P (lrjState) Probability P (Statejlr)
0.5 0.70 0.35 0.7R
0.5 0.20 0. 10 0.22
:E = 0.45
Table 1312
CoMPUTATION oF PosTERIOR PROBABILITIES GIVEN THAT
INVESTIGATIONPREDICTED DEMAND (S LOW (/)
(2) 0) (4) (5)
(2)(3) = (4)/L(4)
Prior Confidence Posterior
Probability, Assessment, Joint Probability,
P(State) P (ljState) Probability P (Statej/)
0.30 0.15 0.27
0.5/ 0.80 0.40 0.73
:E = 0.55
V\
\
350 Risk and Uncertainty Chap. /3
The probabilities calculated in Tables 1311 and 1312 can now be used
to assess the "technology study" alternative . Figure 138 shows a decision
tree diagram for this alternative as weil as the two original alternatives .
Note the demand probabilities are entered on the branches according to
whether the investigation indicates "high" or "low" demand.
PW of benefits
H
$45M (0.5)
$27 .5M (0.5)
H $45M (0.78)
L $27.5M (0.22)
H $80M (0.78)
L $48M (0.22)
T eehnology study
H $45M (0.27)
$27.6M (0.73)
H $80M (0.27)
$48M (0.73)
H $80M (0.5)
$48M (0.5)
figure 138. Replacement problem with alternative of further investigation.
Sec. 13.3
Sequential Decision Making: Use of Decision
351
The expected outcome for the "technology study" alternative can now
be calculated. This is done by the standard decision tree "rollback" princi 
ple. This is shown in Table 1313. His worthy of note thal the 0.45 and 0.55
probabilities that investigationpredicted demand will be "high .. and "low:
respectively, are obtained from the totals in column (4) of the Bayesian
revision calculations shown in Tables 1311 and 1312.
Thus, from Table 1313, it can be seen that the "new FMS .. alternative
with an expected net PW of $29.0M is the best initial course of action by a
slight margin. Wh ile the figures used here do not reftect any ad van tage to
this technology study, the ad v an tage potentially can be great.
A 3.3.6 Expected Value of Perfect Information
The expected value of perfect information (EVPI) is the maximum
expected loss due to imperfect information or foreknowledge as to wh at will
be the outcome(s) in a situation involving probabilities associated with the
possible returns or costs. It is described more complete( y in Appendix 13B.
Calculations of the EVPI for the cxample replacement problem of Fig
ure 137 are shown in Table 1314. It can be seen in the table thal the
expected net PW with perfect foreknowledge, $31.25M, is greater than the
expected net PW of the preferred alternative before the further investigation.
$29.0M (from Table 139), by $2.25M, which is a measure of the maximum
expected value of further investigation.
For the same example problem except with the different probabi lities
resulting from "technology study" as given in Fig. 138 and Table 1313. the
technology study results in a decrease in the expected net PW of $1.70M.
(i.e., $29.0M for the "new FMS" minus $27.3M for "technology study .. ). If
the "technology study" had cost $0 instead of $2.0M assumed in the prob
Decision
Point Alternative
2A
{
Old system
New FMS
2B {
Old system
New FMS
1
{
Technology
study
Old system
New FMS
Table 1313
EXPECTED NET PW FOR REPLACEMENT PRORLEM OF FIG. 138
Expected Net PW Choice
$45M(0.78) + $27.5M(0.22) SIOM = S31.15M
$80M(0.78) + $48M(0.22)  $35M = S37.%M New FMS
$45M(0.27) + $27.5M(0.73) SIOM = $22.23M Old system
$80M(0.27) + S48M(0.73)  $35M = S21.64M
$37.96M(0.45) + $22.23M(0.55) S2M = S27.31M
(from Table 139)
(from Table 139)
S26.25M
S29.0M New FMS
0
 '\)
.,
3.54 Risk and Uncertainty
13.5 ALTERNATIVE METHOD OF DECISION TREE
ANALYSIS; NOTATION/ANALYSIS CONVENTIONS
Chap. 13
Sorne analysts and decision makers prefer to show the criterion values at the
end of each possible pa th through the tree. (Note: If the monies along the
path are at significantly different points in time, the criterion values should
be expressed in terms of equivalent worths, such as PW or FW.) Then the
"rollback" technique can be used to determine the optimal choice at each
decision point and obtain the same initial decision as when using the pre
vious method.
Suppose, for the replacement problem in Fig. 138, that it is desired to
evaluate the alternatives using the certainty equivalent (CE) criterion while
formally considering th at the decision maker's initial asset position is, say,
$10M. Figure 1310 depicts this as weil as a hypothetical solution to the
problem. Note !hat the initial asse! position is shown in a dashed oval to the
left of the initial decision point. and that the solid ovals at the extreme right
indicate the outcomc's criterion values, (including the initial asset position)
for each path of branches. The conventions which produce the results in
Fig. 1310 are as follows:
1. At each chance node (circle), the CE is shown in a rectangular box and
represents the value of being at that node, and th us takes into account
everything subsequent to (i.e., to the right of) the node.
2. At each decision point node (square) the CE of the best alternative is
shown in a rectangular box. Further, ali alternatives except the best at
each decision node are marked with double slashes across their respec
tive paths.
To complete the solution, the decision maker must specify his CE for
ali chance outcome nodes. Note that the uppermost node is the same as in
Fig. 139a. As in thal example, let's assume thal the decision maker decides
that a certain $35.5M is just as desirable as 0.5 chance at $45M and 0.5
chance at $27.5M. This is shown in the small block above the node. Subjec
tively determined CEs are shown at ali other nodes. The "rollback" analy
sis procedure is applied but this time taking into account only the probabili
ties . and not the cash flow s. along each path (sin ce the effect of the cash
flows along each path has already been laken into account by the equivalent
worth outcomes (in ovals) at the end of each path).
The final "rollback" to decision point 1 shows thal the alternative of
"technology study" (with a CE of $35.7M) is better than either "old sys
tem (with a CE of $35.5M) .or "new FMS" (with a CE of $34M). This
differs from the expected monetary (net PW) results in Table 1313 as typi
cally would be anticipated.
Sec. 13.5
Initial asset
position
Alternative Method of Decision Tree Analysi.t
H
355
127.5M 10.51
H $SOM
$48M
'""""!
Outcome1
incl udlng
ini tia18'SS!et
Figure 1310. Replacement problem with alternative of further investigation
(showing outcome criterion values) at the end of each branch and subsequent
analysis using CE criterion and including initial asse! position.
Example for lowest branch path: SIOM  SJ5M + S48M = SZJ M.
V\
'
If
demand
Risk and Uncerrainty Chap. 13
Table 1314
EXPECTED VALUE OF PF.RFECT INFORMATION FOR NESTAGE REPLACEMENT
PROBLEM OF FIG. 137 (ALL $ARE IN PW)
Oulcomes
for alternatives
would he:
Then the And the The prior Thus the
preferred monetary probability expected
alternative is: oulcome is: estimate is: value is:
High Old:  $10M = $J5M New FMS $45M 0.5 S22.50M
New:  35M = 45M
Low Old: 27.5M  IOM = 17.5M Old system 17.5M 0.5
New: 48M  35M =
Expected net PW with perfect foreknowledge:
Expected net PW (from Table 139) = S29.0M
:. EVPI = SJ1.25M  S29.0M = $2.25M (in net PW)
8.75M
r = S3J.25M
lem. it would have resulted in an increase in the expected net PW of $1.7M
+ $2.0M = $+0.3M. Thal $0.3M represents the "expected value of sample
information:: i.e. , how much out of the possible maximum EVPI of $2.25M
can be achieved by this particular "technology study."
13.4 USE OF CERTAIN MONETARY EQUIVALENTS
Any problem involving variable outcomes, such as the replacement problem
in Fig. 138. could just as weil be analyzed using the socalled certain mone
tary equivalent (CME) or certainty equivalent (CE) criterion rather than an
expccted monetary value criterion.
A certainty equivalent (CE) for any situation involving variable out
cornes is a monetary amou nt (lump sum) which the decision maker considers
to be equally as valuable (to pay or to receive) as the variable possible
outcomes for thal situation. Socalled "riskaverse" (riskavoiding) deci
sion. makers typically have (or will choose) a CE which is Jess than the
expected monetary value (EMV) for a given situation. The amount by which
the EMV is greater than the CE, called a "risk premium" (RP), is a measure
of how much the decision maker is willing to give up, on the average, to
eliminate the variability (riskiness) in a decision problem. Thus
CE = EMV  RP and RP = EMV  CE
While RP is positive for a riskaverse decision maker (i .e., he or she would
opt to "bu y out" of a ris ky situation), it is negative in the case of a socalled
"riskseeking" (risktaking) decision maker, and it is 0 in the case of a "risk
neutral" decision maker.
j.
Sec. 13.4
Use of Certain Monetary Equivalents
53
Example 13l(a)
Suppose thal it is desired for a decision maker to specify his CE for the situation in
the left part of Fig. 139. The question to be answered can be phrased as: What
amount for certain (to be received or paid) would be just as good Cor bad) as a 50%
chance of $45M and a 50% chance of $27.5M? Another way of phrasing this is :
"How much would you be wiUing to pay (or get paid) in place of [sorne probabilistic
payoff situation]?" The decision maker might consider th at situation not very ris ky
and answer, say $35.5M. The EMV for the situation is $36.25M. which mcans that
RP = $36.25M  $35.5M = $0.75M.
Example 13l(b)
Suppose thal the outcomes in Example 13l(a) are made more risky (highly variable)
by increasing the $45M by $40M to $85M and decreasing the $27.5M by $40M to
$12.5M. The situation is shown in the righi part of Fig. 139. The EMV is stiJl
$36.25M, but the CE answer by a given decision maker might vary considerably
according to his perception of the impact of the new possible outcomes . If the
decision maker is very slightly risk averse for the amounts involved, he might stiJl
have a CE close to the $35.5M in Example 13l(a) . However, if he is significantly
concerned about the impact of a possible low outcome su ch as the  $12.5M. his CE
might be significantly lower , such as , say, $28.0M. If so. his RP = $3t>.25M 
$28.0M = $8.25M.
A decision maker's certainty equivalent for a set of uncertain cash flow
outcomes should depend on the complete context in which the cash flow
occurs . Thal is, the decision maker should be cognizant of the firm's finan
cial position when deciding on a CE. One way this could be taken into
account is to specify the decision maker' s financial position before the
initial decision is to be made. For example, perhaps net liquid assets (such
as net assets thal can be converted into cash within a month or within a
year) might be laken as the most appropriate mcasurc of the firm ' s initial
asse! position.
(0.5) S45.0M
10.5) S85M
(0.51 S27.5M
10.5) S 12.5M
Choice 2 CE a 7 + (to equal choie 11
(a)
(b)
Figure 139.
determined.
Example of outcomes for which certainty equivalents are lo be
3Sb Risk and Uncertainty Chap. 13
13.6 STOCHASTIC DECISION TREES
lt is generally true that the number of possible outcomes that could result
from the choice of an alternative is greater than the few outcomes typically
assumed in decision tree problems so that the number of branches is kept to
manageable size. Hence, information (and sometimes, objectivity) is sacri
ficed by the assumed reduction of branches. It is corn mon that the possible
out cornes emana ting from an alternative could be described most adequately
as a continuous distribution. These are sometimes called stochastic decision
trees .
As an example of a problem for which sorne kind of stochastic decision
tree analysis is probably applicable, consider the automation problem in Fig.
136. Only thrce possible outcomes were considered in case the process
were automated. However. the problem would have been more adequately
mode led if, say, ten or more possible monetary outcomes were considered if
the process were automated, taking into account the variable costs and
benefits of automation. lndeed, the problem probably can be represented
best by continuous probability distributions. Monte Carlo simulation can be
used to analyze such decision situations efficiently to arrive at probability
distributions of the decision criterion for ali initial alternatives. Section 13.7
will explain Monte Carlo simulation in general.
Often, variable outcomes are thought to have a continuous distribution
. but the decision maker or analyst feels that it is not.feasible or reasonable to
specify the distribution (i.e., the decision problem fits the classical definition
of uncertainty). A useful way to diagram an uncertain outcomefor a deci
sion tree problem is shown in Fig. 1311 . Note that this outcome "fan" can
show maximum and minimum values if these can be estimated.
As a further example, suppose the above automation problem involves
corltinuous outcomes for which probabilities cannot be estimated but for
which the minimum and maximum extremes for each outcome can be esti
mated. In this case, a diagram representing the problem (with assumed max
Figure 1311. Representation of a con
tinuous outcome for which probabilities
are not known.
Sec. 13.6 Stochastic Decision Trees
' /
Not autom. $0
Figure 131Z. Automation problem assuming continuous outcomes for which
probabilities may not be known, but extremes can be estimated.
j 57
imums and minimums) could be as shown in Fig. 1312. Note that even the
outcomes of the technology study are represented to have continuous, but
unknown, probability distributions. White information in this form does not
le nd itself to mathematical calculation of decision criteria such as EMV, the
mere diagramming of the problem can be very helpful in reaching a deci
sion. Indeed, a decision maker can be asked to determine his CE at any
outcome node even though probabilities of the various outcomes are not
known. Any such CE would be a gross guesstimate (certainly, it would be
more subjective than if probabilities of various possible outcomes were
known), but at !east it represents the decision maker's intuitive judgment
under the circumstances and can be used for comparing alternatives .
r
, ..
J74
Risk and Uncertainty Chap. 13
/)rcinn Trers (Srctions 13.3 tllrou!!ll 13.6) :
B9. Given : There is a 60% chance !hat Dynamics General Corporation will require
a conversion from haselmnd Ethernet cabling in a certain manufacturing plant
to hroadhand Ethernet cabling at a rate of 1000 feet per year (there is a 40%
chance that the conversion will occur at a rate of 500 feet per year). The
recabling will be required if a new government contract is negotiated with them
now. which is 10 years prior to expiration of the present con tract.
The VP of Manufacturingfeels that if he wait.r for the present contract to
expire in 10 years. there is a 25% chance that the new contract will require 5000
feet per year of conversion from baseband to broadband Ethernet cable and a
75% chance of conversion at the rate of 1000 feet per year. Each 1000 feet of
Ethernet conversion (baseband to broadband) costs $100,000.
IJ10.
Pwpert y ta x for either type of cable is the sa me and should not be
C<lnsidcred . Ail new cabling will be broadband Ethernet; and at the MARR of
21l<::f. pcr year. it can be considered lo have perpetuai life. The terms of any
new contract will cali for a continuing program of converting existing
bascband cable to broadband cable until the total anticipated cabling require
ment of 25.000 feet has been completed.
Required: Draw a decision tree for this problem, including the appropri
ate cash nows and probabilities. State any assumptions you make. Carefully
label ali alternatives.
A purchasing manager is faced with deciding whether or not to stock a large
supply of metal. The uncertain variable is the future priee of the metal. The
following are present worths of consequences and prior probabilities for the
various perceived outcomes:
Future Priee P(Furure Priee) PW If Stock
PW If Don't Stock
High
O.J $ 100,000 $0
Medium 0.5  10,000
$0
Low
0.2 100,000
$0
For $6.000 it is possible to hire a consulting firm that would be able to make a
fairly accurate forecast in terms of whether the priee will go up or down as
follows :
If the future
is going to be:
High
Medium
Low
Then the probabilities
r he consultant will
predict the priee will
go up or down are as
follows:
Up
0.9
0.4
0.8
Down
0.1
0.6
0.2
Chap. 13 Exercise.r 375
(a) Diagram the problem in the form of a decision trec .
(b) Determine what would be the best alternative using the EMY criterion.
(c) Determine the maximum expected value of the consulting firm' s
if the firm could perfectly predict the future.
(d) Determine the best alternative if you were the decision ma ker using your
own unique CE values.
1311. The Norva Company has already spent $80,000 developing a new electronic
gage and is now considering whether or not to market it. Tooling for produc
tion wou id cost $50,000. If the gage is produced and marketed, the company
estimates that there is only one chance in four thal the gage would be success
fui. If successful, the net cash infiows would be $100.000 per year for 8
years . If not successful, the net cash outnows wou id be $30,000 per year for 2
years, after which lime the venture would be terminated. The minimum at
tractive rate of return on money is 20% per year.
(a) Draw a decision tree and determine the best alternative using the ex
pected net PW criterion .
(b) If there is a market research group that can provide perfect information
about the success of this product, what would be the most the company
should be willing to pay for the group's service?
(c) Suppose the market research group can make a market survey thal with
probability 0.8 will predict a successif the gage will actually tu rn out to be
a success and with probability 0.9 will predict failure if the gage will tu rn
out to be unsuccessful. Should the survey be undertaken first? What is
the expected value of the survey to the company?
13U. Given the following twostage decision situation shown in Fig. E 1312. deter
mine which is the best initial decision. Notice thal sorne annual costs are
negatively signed, which makes them positive cash flows. Use the expected
PW of costs method and a MARR of 12%. To give the problem a physical
context, the following letter symbols have been employed for each alterna
tive:
BSW: Build small warehouse.
RLW: Rent large warehouse.
BA: Build addition.
NC: No change.
1313. A firm must decide between purchasing an automatic machine which costs
$50,000 and will last 10 years and have 0 salvage value, or purchasing a
manual machine which costs $20,000 and will last 5 years and have 0 salvage
value. If the manual machine is purchased initially, a ft er 5 years a decision
will have to be between a manual machine having the same characteristics
affecting cost as the first manual machine and a semiautomatic machine cost
ing $40,000 which would have a $20,000 salvage value after 5 years of life.
The annual operating tosts for each of the machines are as follows : auto
matie, $10,000; manual, $14,000; semiautomatic, $11.000.
(a) Graphically construct a decision tree to represent this situation.
N

376 Risk and Uncertainty Chap. /3
First cost and resale
value of alternatives
Annual cost (or savings, if negative) Probabillties
<.
"J.
$2,000/yr (10 yr) 0.15
11 ,000/ yr (3 yr) and 110,000/ yr (7 yr) 0.15
1 8,000/ yr (7 yr) 0.2
12.000/ yr (7 yr) 0.8
$1 ,000/ yr (7 yr) 0. 2
118,000/yr (7 yr) 0.8
118,000/ yr l10yr)
118,000/ yr (3yr)and $21,000/yr (7 yr) 0.15
$23,000/ yr (3 yr) and $21 ,000/ yr (7 yr) 0.56
123,000/yr (10 yr) 0.14
Figure E1312. Twostage decision situation for Exercise 1312.
(b) Determine which decision would be made at each point using the PW of
cost s method and an interest rate of 10%.
(c) At what interest rate would the decision between the manual and semiau
tomatic machine be reversed?
1314. Suppose one is faced with the same alternatives and dollar outcome conse
quences as in the replacement problem depicted in Fig. 138. However, the
initi al estimates of probability of demand are: high, 0.6; low, 0.4. Further
more. management's assessment of confidence in further investigation
results, using the notation in Table 13 10 are
P(h !Hl = 0.80
P(hiL) = 0.40
P(liH) = 0.20
P(I!Ll = 0.60
Calculate the choice at each decision poini to determine the best initial dcci .
sion. How close is the initial decision with these revised probabilities to the
initial decision for the original problem depicted in Fig. 138?
1315. Figure E13 15 is a decision trec portrayal of a building lease versus buy
problem with input data supplied. lnvestment requirements are shown as
negative numbers : probabilities associated with each outcome are shown in
Chap. 13 Exercises Jll
$20,000/ yr, 25 yr
1 15,000/yr, 15 yr
Keep 11 0,000/ yr, 1 0 yr, 10 salvoqe
Abandon $0/yr, 10 yr
Salvage 120,000
$0/yr, 15 vr. salvage $25,000
$20,000/yr, 15 yr
Keep $15,000/yr, 15 yr, 10
Abandon 10/yr, 15 yr
Salvage $30,000
$20,000/ yr, 25 yr
110,000/ yr, 15 yr
Keep $5,000/ yr, 10 yr
Abandon $0/yr , 10 yr
Abandon 10/yr, 15 yr
Figure E1315. Decision lree for Exercise 13 15.
parentheses. The annual cash savings and duration of those savings are
shown together at each relevant outcome. Salvage values in the cases of
abandonments are assumed to occur at the end of the 25year study period .
Determine the best decision using the expected net PW method with a mini
mum required RR of 0%.
1316. Set up a decision tree to reHect the persona! automobile alternatives which
you expect over the next severa! years. Carry the trec far enough in ti me to
show severa! decision points at which a decision must be made between
keeping an old car and buying a new car (from a possible choice of several).
Show your roughly estimated assumed certain investment costs and salvage
values and an nuai operating costs for each alternative and determine the best
initial decision using the PW method and a 10% minimum persona! opportu
nity cost of moncy.
1317. Suppose, for the automation problem for which prior probabilities are shown
in Fig. 135, the probabilities of the possible outcomes for the added study
(technology study) are as follows :
V\
\
('0
371!
Givcn lhnl lhc
mllomnlion will
IUrn OUI 10 be:
Pnor
Fair
Excellcnl
Ri.<k and Uncertainty
Thcn lhc prohahilily lhal
lhc lcchnnlogy
will indicnle   is:
Shaky Sol id
0.6 0.3 0. 1
O.J 0.4 0.3
0. 1 0.4 0.5
Clzap. 13
lise Rayes" thcorcm to cnlculnlc the posterior prohnhililies regarding whal
the automation results will lurn oui ln he. givcn whal lhe lcchnology sludy
in<l ic ntcs . Chct"k ynur rcsulls n!!ainsl lhe pmhahilitics shown in Fig. 136.
J\
1
61
CHAPITRE6
LMENTS DE BASE DE LA SIMULATION DES SYSTMES
6.1 Technique de Monte Carlo :
Mthode dveloppe durant les annes 40 par Von Neuman, Vlan et Fermi pour rsoudre
certains problmes de design d'crans antirayonnement. Elle est coteuse exprimentalement et
difficile rsoudre analytiquement.
Elle consiste reprsenter un problme dterministe par un processus stochastique dont les
distributions de probabilit satisfont les relations mathmatiques du problme dterministe complexe.
La premire tape consiste construire un modle analytique qui reprsente l'investissement considr
(exemple: une quation pour la valeur prsente).
Dans la deuxime tape, l'analyste dveloppe une distribution de probabilit pour chaque
facteur qui est sujet l'incertitude dans le modle. Il peut utiliser des donnes historiques ou une
approche subjective.
partir des distributions de probabilits pour chaque facteur dans le modle. Il gnre au
hasard une rponse tentative. La rptition de ce mode d'chantillonnage un nombre important de fois,
permet de gnrer une distribution de frquence pour la rponse. On peut utiliser n'importe laquelle
des mthodes de mesure de rentabilit. La distribution de frquence rsultant peut tre utilise pour
obtenir des donnes probabilistes sur le problme original.
Exemple 6.1 : Calcul de 1t
Si on met alatoirement un certain nombre de points (n) dans le carr.
Un certain nombre de points appartiendront au cercle et d'autres non
 Quand n est grand, la surface du carr sera couverte.
n => oo ~ = >
#cercle
# total
nb. pt. _ds c ercle
nb. pt. ds carr
surface cercle
surface du carr
(2a)
2
n
4
nb. ds cercle
n = 4 x 
nb. ds. carr
a = 1 M = [ : : ]
r
1
r
2
des nombres alatoires entre 1 et 1.
On a gnr 30 pts, 24 e cercle et 6 e cercles.
n = 4 * ~ = 3 .2
30
Sinoo, 1t3.14
Exemple 6.2 : Le pari
62
Supposons que l'on vous propose de jouer au jeu suivant: pour une mise de 20 dollars, on vous
permet de jouer pile ou face avec une pice bien quilibre et de gagner 2\ x tant le nombre de lancs
ncessaires pour avoir pile. Accepterezvous de faire la mise pour jouer? Quel est le montant maximum
(suprieur ou infrieur 20 $)que vous accepterez de risquer pour jouer? Votre rponse dpendra de
votre got pour ce type de risque. Je n'ai encore jamais rencontr quelqu'un qui dit accepter de miser
jusqu' 100 $. Pourtant. ..
Une approche analytique de ce problme de pari suggrera le recours l'esprance mathmatique.
Dans ce cas, on arrivera la conclusion qu'il faut toujours jouer ce jeu quel que soit le montant de la
mise demande, car l'esprance mathmatique du gain est infinie!
C'est ce qui ressort du calcul ciaprs.
63
E ( gai n) = " ( 2n 2n 1 2 )
LJ 2 . 2
E (gain ) =L 1
= !><)
Une approche de simulation est facile concevoir et mettre en oeuvre: essayez sans frais, avec
une pice de monnaie, de jouer pile ou face et dterminez ce qu'auraient t vos revenus si vous aviez
rellement jou.
Vous pourriez obtenir les rsultats suivants :
1
2
3
4
5
ftfp
fp
ffifp
p
p
Chaque essai est bien termin avec une sortie de pile et les montants gagns sont de 2 une
puissance gale au nombre de fois qu'il a fallu lancer la pice pour avoir pile.
la vue de ces rsultats on va tre frapp par le gros gain comme 32 ou le petit 2 selon qu'on a
le temprament de joueur ou non. En faisant la moyenne des gains on trouve 56/5, soit environ 11 . Mme
si ces rsultats ne sont pas trs significatifs (chantillon trop petit et probablement biais par la
manipulation), ils livrent nanmoins un message intressant: au del de 30 la mise est trop risque.
La simulation ne nous donne pas LA solution; elle nous donne des indications nous permettant de
prendre une dcision selon notre temprament. Cette situation est confortable pour le dcideur car on ne
lui impose pas une solution futelle optimale.
L'exemple du pari va nous permettre d'illustrer aussi un aspect essentiel de la simulation Monte
Carlo : la simulation de la ralisation d'un vnement partir de la probabilit attache la ralisation de
cet vnement. Une pice de monnaie permet normalement de bien simuler le comportement d'une autre
pice de monnaie quivalente.
Cependant les rsultats obtenus par le jet d'une pice de monnaie ne sont pas tout fait alatoires:
les faons de jeter la pice, de la saisir sont propres l'oprateur et non au hasard. Une autre limitation
de la technique de simulation par le jet de la pice de monnaie apparat lorsque l'on envisage de simuler
la ralisation non pas de deux vnements quiprobables (pile face) mais de plusieurs (pile 1 fois ou 2 fois
ou 3 fois ou 4 fois en 4 lancs).
64
La simulation Monte Carlo permet d'arriver au mme rsultat que la simulation manuelle du jet
de la pice de monnaie. En faisant la correspondance suivante :
pile= {0, 1, 2, 3, 4}
face= {5, 6, 7, 8, 9}
et en supposant que lors d'un tirage au sort parmi ces nombres quiprobables la sortie d'un chiffre traduit
le fait que la pice de monnaie a montr le ct (pile ou face ) correspondant. Par exemple, si le nombre
tir est 2, c'est pile; 8 c'est face, etc. Il ne restera plus alors qu' trouver un systme de gnration de
nombres alatoires pour simuler le jet de la pice de monnaie.
Ce procd a l'avantage d'tre plus neutre. En plus, il permet de faire face au cas o il y a plus de
deux vnements. Si nous avons trois vnements par exemple, il suffit de connatre la probabilit attache
chacun de ces vnements et de faire la correspondance des chiffres consquemment.
Exemple 6.3 : Estim des ventes annuelles
On a estim que les ventes annuelles d'un produit d'entretien mnager ont 60% de chance
d'atteindre 200,000 dollars, 30% de chance d'atteindre 300, 000 dollars et 10% de chance d'atteindre
400,000 dollars.
Pour simuler le comportement des ventes annuelles, on fera les correspondances suivantes sur la
base de 10 :
200 000$
300 000$
400 000$
60% {0, 1, 2, 3, 4, 5}
30% {6, 7, 8}
10% {9}
Il ne restera plus qu' tirer les nombres alatoires pour simuler le comportement des ventes
annuelles. Si les pourcentages n'taient pas arrondis, il aurait fallu faire la correspondance sur la base de
100 nombres. Essayez pour l'exemple prcdent : 61%, 29% et 10% respectivement. Essayez ensuite
61.5%, 28.5% et 10%.
Exemple 6.4 : Comparaison de rsultats analytiques et de simulation
[!] Une clinique mdicale dispose de 5 centres.
[!] Une ( 1) infirmire est affecte chaque centre.
[!] Il y a 3 infirmires rservistes qui peuvent tre appeles pour remplacer celles qui
s'absenteraient.
Les infirmires peuvent s' absenter en cas de maladie. La distribution de probabilit des infirmi res
65
malades est reprsente dans le tableau suivant :
Nb. de malades 0 1 . 2 3 4 5
Probabilit .2 .25 .2 .15 .1 .1
On dsire simuler 25 jours de fonctionnement de la clinique l'aide de la technique MONTE
CARLO pour obtenir des estims de :
[!] le taux d'utilisation des infirmires rservistes;
[!] la probabilit qu'au moins un centre mdical soit ferm par manque d'infirmires.
Feuilles de statistiques pour la simulation du cas des infirmires.
NIG
NRU
NJCF
NIG
NRU
NJCF
Analytiquement
Nombre moyen d'infirmires malades=
Nombre moyen d'infirmires rservistes utilises=
La probabilit qu'au moins un centre mdical soit ferm par manque d'infirmires =
Exprience de simulation
Nombre moyen d'infirmires malades =
Nombre moyen d'infirmires rservistes utilises =
La probabilit qu'au moins un centre mdical soit ferm par manque d'infirmires =
Disque d'chantillonnage pour le cas des infirmires :
D
1
s
Q
u
E
NIG U) = 0
NIG U) = 1
NIG (j) =nombre d'infinnires malades au jour j
66
Logigramme du modle de simulation
INITIALISA noN
NRU = NJFC = 0
NIG=O
UIR = NRU/75
PCF = NJCF /25
NJCF = NJCF + 1
NRU= NRU+ 3
NIG = NIG + NIG UJ
_j 1 Ill lill 1 1 1 1 Il 1 1 1 1 Il 1 Il 1 1 1 1 1 1 1 1 1 1 1 1 Il 1 1 1 1 1 1 L
GLOSSAIRE
NJCF = #jour oa au moins 1 centre ferm
NRU = # rserviste jours utiliss
NIG (j) = 11 infinmire en grve au jour j
UIR = utilisation des infinmires de rserve
PCF = probabilrt qu'au moins un centre sort fenm
1 11111111 Il l Ill 111 11 11 Ill 1 I l Ill lill 1 I l Il l I l 1 1
67
68
6.2 Considrations dans la sim.ulation de MonteCarlo
Une question importante connatre dans la simulation de MonteCarlo est le nombre d'essais (ou
d'excution) ncessaire pour avoir une rponse satisfaisante. La rponse est que le nombre d'essais doit tre
assez lev pour atteindre le rgime permanent. On peut dfinir les conditions de rgime permanent comme
une situation dans laquelle le rsultat d'essais conscutifs ne varie pas significativement.
Dans un simulateur, on utilise habituellement une priode de rchauffement avant de collecter les
statistiques, l'utilisation des simulations de MonteCarlo prsente des avantages pour:
Les situations dans lesquelles des mthodes de rsolution analytique n'ont pas encore t trouves;
Des cas o les modles analytiques sont trs complexes rsoudre;
Ne pas perturber le fonctionnement de la compagnie;
Des situations pour lesquelles il est trs difficile ou dangereux de crer les mmes conditions
d'opration pour chaque exprience;
Faciliter l'exprimentation de plusieurs possibilits qui ne peuvent tre ralise en vraie vie;
Des situations qui peuvent coter chres et/ou prennent beaucoup de temps pour gnrer le rsultat
voulu.
Les modles analytiques et la simulation de MonteCarlo ont les mmes problmes, savoir, la
validit du modle. Cependant, dans le cas de la simulation de MonteCarlo, la taille de l'chantillon doit tre
assez grande pour dcrotre la variation d'chantillonnage un niveau acceptable. Les probabilits d'es
vnements doivent tre bases sur les jugements de personnes impliques dans le projet, par la suite,
l'analyse doit tre dose sur ces estims bien qu'ils soient subjectifs tant donn que chaque modle est aussi
bien que les estims de ses paramtres d'entre, la simulation de MonteCarlo doit tre utilise avec le
jugement du dcideur.
6.3 GNRATION DE NOMBRES ALATOIRES
Les systmes habituellement simuls comportent des processus stochastiques;
ex. : demande pour un produit;
temps d'opration sur une machine;
arriv de clients une station de service;
temps de ralisation d'un projet.
etc.
Pour simuler adquatement la ralisation des vnements dont on connat les probabilits de
ralisation, il ne suffit pas de faire les correspondances que nous venons de voir. Il faut avoir aussi un
systme de gnration de nombres quiprobables (ou alatoires). Malgr l'abondance des sries de chiffre
dans notre entourage: bottin de tlphone, numro de rue, etc., on ne compte que deux sources fiables de
nombres alatoires: les tables de nombres alatoires publies (on en trouve dans les annexes de la plupart des
manuels de statistiques et de simulation) et les gnrateurs de nombres alatoires des ordinateurs.
69
Les tables se prtent l'utilisation manuelle. Les nombres contenus dans ces tables ont t soumis
divers tests statistiques pour s'assurer de leur caractre alatoire dans toutes les directions: horizontale,
verticale et oblique. Tous les chiffres sont quiprobables; il en est de mme des nombres de deux chiffres,
de trois chiffres. Lorsque l'on veut utiliser ces tables, il est conseill de fixer la direction de lecture de
nombres et la position de dpart et de ne pas revenir sur ses pas. Voir la table de nombres alatoires.
Les ordinateurs possdent des gnrateurs de nombres alatoires. Il suffit gnralement de les
voquer. On peut aussi programmer la gnration des nombres alatoires ou les calculer soimme.
On a donc besoin de gnrer des variables alatoires (dpendantes ou indpendantes) ayant des
distributions spcifies. C'estdire un chantillonnage d'une certaine population.
L'ide est : gnrer des variables alatoires uniformment distribues;
t gnrer des variables d'autres distributions.
6.3.1 Distributions uniformes
Ses caractristiques sont :
1) Chacun des nombres de la population a la mme probabilit d'tre dans la srie :
2) Chacun des nombre est compltement indpendant de tous les autres dj gnrs.
....
Variables alatoires indpendantes uniformment distribues .
6.3.1.1 Mthode manuelle
Pile ou face, ds, cartes, roulette, etc.
Avantages : Trs simples, honntes.
Dsavantages : Lentes, peu pratiques, impossible reproduire.
6.3.1.2 Tables
Pris dans n'importe quelle direction prdtermine, partir de n'importe quel point.
A vanta ge : reproductibilit
Dsavantages : Lenteur de l'emploi, utilisation continuelle des mmes squences (tables limites).
610
Table de nombres alatoires
Table A.t. RANOOM DIGITS
94737 08225 35614 24826 88319 05595 58701 57365 i4759
87259 85982 13296 89326 74863 99986 68558 06391 50248
63856 14016 18527 11634 96908 52146 53496 51730 03500
66612 54714 46783 61934 30258 61674 07471 67566 31635
30712 58582 05704 23172 86689 94834 99057 55832 21012
69607 24145 43886 86477 05317 30445 33456 34029 09603
37792 27282 94107 41967 21425 04743 42822 28111 09757
01488 56680 73847 64930 11108 44834 45390 86043 23973
66248 97697 38244 50918 55441 51217 54786 04940 50807
51453 03462 61157 65366 61130 26204 15016 85665 97714
92168 82530 19271 86999 96499 12765 20926 25282 39119
36463 07331 54590 00546 03337 41583 46439 40173 46455
47097 78780 04210 87084 44484 75377 57753 41415 09890
80400 45972 +till 99708 45935 03694. 81421 60170 58457
94554 13863 88239 91624 00022 40471 78462 96265 55360
31567 53597 08490 73544 72573 30961 12282 97033 13676
07821 24759 47266 21747 72496 77755 50391 59554 31177
09056 10709 69314 11449 40531 02917 95878 74587 60906
19922 37025 80731 26179 16039 01518 82697 73227 13160
29923 02570 80164 36108 73689 26342 35712 49137 13482
29602 29464 99219 20308 82109 03898 82072 85199 13103
94135 94661 8724 88187 62191 70607 63099 40494 49069
87926 34092 34334 55064 43152 01610 03126 47312 59578
85039 19212 59160 83537 54414 19856 90527 21756 64783
66070 38480 74636 45095 86576 79337 39578 40851 53503
78166 82521 9261 12570 10930 47564 77869 16480 43972
94672 07912 26153 10531 12715 63142 88937 94466 31388
56406 70023 27734 22254 27685 67518 63966 33203 70803
67726 57805 94264 77009 08682 18784 47554 59869 66320
07516 45979 16i35 46509 17696 67177 92600 55512 17245
43070 22671 00152 81326 89428 16368 S659 79424 57604
36917 60370 80812 87225 02850 47118 23790 55043 75117
03919 82922 02312 31106 05573 17470 25900 91080
46724 22558 64303 78804 05762 70650 56117 06707 90035
16108 61281 86823 20286 14025 24909 38391 12183 89393
74541 75808 89669 87680 72758 60851 55292 95663 883:!6
82919 31285 01850 72550 42986 57518 01159 01786 98145
31388 26809 77'158 99360 92362 21979 41319 75739
17190 75522 15687 07161 99745 48767 03121 20046 28013
00466 88068 68631 98745 97810 35886 90230 69264
611
6.3.1.3 Utilisation de formules rcursives
Utilisation d'quations mathmatiques de rcursion (ex. : partir d'un nombre, un autre nombre de
la srie est gnr). Puis qu'on peut reproduire ces nombres, ils sont appels pseudoalatoire. Cette
reproductivit est leur avantage majeur.
6.3.1.3.1 Mthode du centre du carr
Chaque nombre alatoire compos den chiffres est form des n chiffres du centre du carr du nombre
prcdent.
Exemple : nombre alatoire de 2 chiffres (0 ;!; Xj ;!; 99).
Nombre
Xo= 12
X
1
=44
X
2
=93
X
3
=64
Source= 12
Xo
2
= 144
x
2
= 1936
1
X/= 8649
(NB. On pourrait aussi alterner)
Dsavantage : Dgnrescence rapide (cycle court) notons qu'aussitt qu'un nombre apparat pour
la 2ime fois, il y a cration d'un cycle.
6.3.1.3.2 Mthode du centre du produit
Chaque nombre alatoire Xj compos de n chiffre est form des n chiffres du centre de
ou
de
o K est une constante :
Exemple : Nombre de 2 chiffres
Nombre source X
1
=8; X
2
= 86.
X
1
=8
X
2
= 82
X
3
=56
X
4
=59
X
1
X
2
= 656
X
2
X
3
= 4592
6.3.1.3.3 Mthodes congruentielles
Additives : "n+I =(a "n + C) (modulo rn)
Multiplicatives : "n+I = ( a ~ (modulo rn)
{I(modulo rn)= reste de la division deI par rn}
Exemple:
Exemple :
13 (mod 5) = 3.
rn, c, a, sont des entiers positifs.
a= 2; rn= 11; X
0
= 4; c =O.
x
1
= (2*4) mod 11 = 8
x
2
= (2*8) mod 11 = 5
x
3
= (2*5) mod 11 = 10 ...
Notons que les nombres gnrs sont entre 0 et 10. 0 :S "n :S m1.
TOUTES CES MTHODES POSSDENT 1 OU PLUSJEURS DES CARACTRISTIQUES:
les nombres peuvent tre reproduits;
possibilit de cyclage;
possibilit de dgnrescence (toujours mme nombre);
les n.a. sont uniformes.
612
Il existe des tests pour vrifier l'uniformit et l'autocorrlation des sries (pas dans le cadre du
cours) .
La mthode Congruentielle est la plus utilise en pratique :
Poser rn = 2b (b entier positif)
a = 8t 3 (t entier positif)
a= 2 bn
ko = entier impair.
Tous les nombres seront :S 2b 1.
Il y aura au maximum 2b
2
rcursions avant de former un cycle.
Exemple : Si b = 21 ; 2
19
.
Pour U(O, 1 ), diviser les nombres par 2b  1.
613
Exemple 6.5 : Prvision de la demande
La compagnie SOW Inc. cherche se faire une ide plus prcise de ce que sera la demande pour son
nouveau modle d'ordinateur. Ce type de produit informatique tant vite dsuet, la compagnie ne veut pas
se retrouver avec des invendus la fin de la priode de planification de trois mois.
Utilisons des donnes historiques pour simuler la demande trimestrielle.
.09
3 .27
2 4 .37
3 2 . 18
4 1 .09
Total= 11
Correspondances : 0 = {00, 01, 02, 03, 04, 05, 06, 07, 08} (9)
1 = {09, 10, 11, 12, 13, 14, ... '35} (27)
2 = {36, 37, 38, 39, 40, 41, ... ' 72} (37)
3 = {73, 74, 75, 76, 77, 78, 000 ' 90} (18)
4 = {91, 92, 93, 94, 95, 96, 97, 98, 99} (9)
614
Tirage de nombres alatoires:
Dans la table des nombres alatoires, prenons la colonne forme par les chiffres l'extrme droite
et associons chaque nombre trouv avec sa demande correspondante. Simulons 20 semaines.
 =, ... . f 
"' .... Je_atgire,..,,..... _. _
95 4
31 1
13 1
86 3
02 0
79 3
85 3
87 3
87 3
64 2
54 2
59 2
39 2
03 0
52 2
53 2
22 1
34 1
74 2
38 2
Total=> 39
Ce qui donne une demande de 1.95 en moyenne, arrondie 2.
On pourrait penser utiliser l'esprance mathmatique pour faire l'estim de la demande partir des
probabilits attaches la ralisation de chaque volume de demande hebdomadaire. Il faut remarquer
cependant que l'esprance mathmatique est fige. Elle indique ce vers quoi tend la moyenne des demandes 
sur le long terme sans permettre dame nature d'tre capricieuse. La simulation permet ce caprice de la
nature: par exemple, mme si la demande de 1 unit ne survient normalement que 9 fois sur 100, la
simulation peut gnrer plus ou moins que cette proportion, donnant ce volume de demande un poids plus
grand ou plus petit dans la moyenne calcule. L'esprance mathmatique figera ce poids 9 pour cent.
615
Exemple 6.6 : Gestion des stocks
Modle analytique de base
Le problme de la gestion des stocks est bien classique: minimiser le cot total de cette gestion en
faisant le tradeoff entre le cot occasionn par les lancements de commandes d'approvisionnement et le cot
de conserver des 1veaux levs d'inventaire. Un modle trs simple a t mis au point par l'approche
analytique sur la base de quelques hypothses simplificatrices.
Soient :
D la demande annuelle
A le cot de lancement d'une commande
H le cot unitaire de maintient des inventaires
L le dlais de livraison
R le niveau de stocks partir duquel on doit commander
Q la quantit commander chaque fois qu'il faut le faire
La quantit optimale commander note Q* pour minimiser le cot total de la gestion des stocks est
donne par la formule suivante :
Q* = /(2AD/H)
et
R =dL (d tant la demande durant le dlais de livraison R)
Exemple :
D = 1 000 units par an; A = 20 $ par commande; H = 4 $ par unit;
L = 2 semaines; alors :
Q* = v'(2*20* 1000/4)
= 100 units
R = (1 000 units/50 semaines) * 2 semaines = 40 units
On le voit : la formule de la quantit optimale commander est lgante et facile utiliser. Cependant
on comprend aussi qu'elle n'est approprie que pour des cas d'une simplicit rarement rencontre dans la vie
relle: demande connue, fixe et rgulire; dlais de livraison connus et constants, etc. La simulation permet
de faire face aux situations plus ralistes.
Modle de simulation
Ici on va supposer que la demande varie la semaine, que les dlais de livraison sont susceptibles de
changer; que des ruptures de stocks peuvent survenir et que cela comporte un cot estim 30 dollars par
unit manquante. Le lancement d'une commande cote comme prcdemment 20 $; on suppose que l'on
commence avec 40 units en inventaire; le cot de maintient des inventaires est de 4 $ par unit par semaine.
Par ailleurs on a obtenu les donnes suivantes sur les demandes hebdomadaires et les dlais de
livraison observs par le pass.
10
20
30
40
10
20
10
10
1
2
3
4
8
4
616
La simulation nous permettra de comparer plusieurs politiques alternatives dans la gestion des stocks,
chacune des politiques consistant dans une rponse aux 2 questions :
Combien commander (Q = ?)
Quand (R= ?)
Dans cet exemple manuel, on se contentera de faire la simulation pour la politique qui consiste
prendre Q = 40, R = 1 O.
Les semaines successives traduisant le droulement du temps; les nombres alatoires pour simuler
la demande de chaque semaine; les nombres alatoires pour simuler les dlais de livraison lorsqu'une
commande est lance; les demandes simules; les dlais simuls; les commandes lances; les commandes
reues; les units en stock; les cots de maintien; les cots de commandes et les cots de rupture.
617
!
..2
1
_,
.... '
...
..c::
u
>
<
;;
Ol
3
en
<
..c::
t:
0
:l:
" :::
o.
z
ai
:
"'
" .!.!
" en
Li
t:
g
u
::l'
c
"
0
" ' tl>
"
" " E
:;j
0
'
<
1 ""'
1 1:!1
1 Ut
'>
i
1 o
l .... .J
e
u
>
<
'
..><
]
"'
"'
"
" :;;!
iii
ILi
0
.;
 "' "' .0
'5 '
 "'
"' ...
0 .g
"" 0
0 >
> 
E "
u 
 Ol
v;
"' ...
" 0
CQ
on
....
Sec. 8.4 Graphical Techniques
Alternati""

A. CIMS tactical ai ms
B. Net present worth
C. Serviceability
O. Management/engineering
effort
E. Rlskiness, Jack of
Key: (a) For net present worth:
P 1



'it
P2
'1
a
'it
e For $500M, with shaded area decrening proportionately to
0 For$0M
P 3
0
P4
"
219
(b) For ali other ortributes:
e ExceptioN!
'it Abo"" average
Balowaverage
0 Poor
Figure 81. Graphical portrayal (using shaded circles) for 4 alternatives and
5 auributes. j
to 10) correspond1tg to possible outcomes (from worst to best, respectively)
for each of four attributes. The evaluation ratings could be "utility" values
(usually on a scale of 0 to 1), as shawn in Chapt er 9. Although such graphs
are not recommended as sufficient by themselves for comparing alternatives,
they do help by portraying how ratings or scoring values vary with outcome
changes for whatever attributes to facilitate the use of other decisionmaking
techniques.
8.4.4 Polar Graphs
Figure 83 is an example polar graph showing evaluation ratings for
alternatives P1 and P2. Note that four of the rays drawn from the center
.....]
w
N
220
10
"'
8
~
c 6
0
~
:>
~ 4
w
0
8
C1.
10
8
8'
 ~
c 6
0
~
:>
;;;
4
>
w
2
0
.J:
M
6
5
0
C1.
t
~
'!;
0
.;
al
CIMS tactical ai ms
{how weil met)
8.
i ~
<1:
..
~
.8
<1:
Serviceability
{expected repair ti me,
.J: E E
M N
~
' '
!::! ~
8, '
:::
!!
"
~
8,
;;
!!
'!;
" ~ 0 >
<1: .;
"
al >
.8
<t
~
.2
.
8
)(
w
.c

~
n;
c
.g
a.
1l
x
w
Multiallribute Decision Ana/ysis
8
"'
 ~
~ 6
.g
..
:>
~ 4
w
2
10
8
. ~
~
c 6
0
 ~
:>
~ 4
w
0
~
61
5
0
c..
8,
~
;;
3:
0
..
al
Riskiness, lack of
f
~
<1:
f
!
~
.8
<1:
Management/engineering
effort (hours required)
~ ~
~
1
8
8
::!
co

~
11.
~
!!
~
" ~ ;;
> <1:
..
~
'!;
.8 0
~
<1:
Figure 82. Sample graphs relating attribute outcomcs to evaluation ratings.
Chap. 8
1
o.
8
)(
w
8
...
~
;;;
c
.g
a.
1l
)(
w
Sec. 8.4 Graphical Tech11iques
CIMS tactical ai ms
10
10
Management
engineering
effort
" Ideal " or "be$t" for
each attribute
221
... ~ !)  i Net P. W.
$500M
Figure 83. Pof graph for 2 alternatives and S attributes .
correspond to evaluation ratings on a scale of 0 to 10 (such as might be
obtained from Fig. 82) and the rightmost ray corresponds to the attribute
"net PW" (on a scale of $0 to $500M). Any attributes thal can be measured
or given an evaluation rating can be included if the maximum and minimum
for each are defined.
Note that after the appropriate evaluations or other measures are plot
led (using linear scales) on each ray, those points are connected to form
polyhdrons for each alternative. Thi s aids one in judging differences be
tween alternatives. However, one should not be misled into thinking that
the overall desirability of each alternative is in proportion to the areas of
their respective polyhedrons . This technique as weil as the other graphical
techniques up to this point have not considered the relative weights or
importances of the attributes, which is the subject of the next section.
'J
w
w
222 Multiallribute Decision Analysis Chap. 8
8.5 RANKING OF ATIRIBUTES (OR ALTERNATIVES)
8efore attribut es are weighted (or alternatives are evaluated) it is often
desired to rank them in order of decreasing preference. This might be done
by presenting the decision maker with a list of factors (usually
they could be alternatives) and asking him or her to rank them in order of
preference. There is, however, a procedure that may make this task easier
and provide a check on the internai consistency of the value judgments
obtained. This is ca lied the method of paired comparisons and is illustrated
by example below.
Consider the five factors (attributes) in Table 82, which we now want
to rank order. They are designated as follows:
A. CIMS tactical aims
B. Net present worth
C. Serviceability
D. Management/engineering effort
E. Riskiness, lack of
The method of paired comparisons suggests that the factors be submit
ted to the decision maker two at a ti me for preference judgments. In gen
eral. ifthere areN factors. N(N 1)/2 pairs must bejudged. 'Assume that
the results of this process are indicated by the following list of preference
statements (as normal, the symbol > means "is preferred to" and = means
is "equally preferred to." etc.).
1. A < 8 6. 8 > D
2. A > C 7. B > E
3. A > D 8. C > D
4. A > E 9. C = E
5. 8 > c 10. D < E
A good way to depict the pairwise comparisons above and then deter
mine rankings is shown by the matrix in Table 83. In that matrix, "P" is
shown for each pair in which the row factor is preferred to the column
factor. Note thal the diagonal of the matrix is empty (since a given factor
cannot be preferred to itself). A good way to make sure thal ali pairs of
factors have been considered is to recognize thal there should be a P either
above or below the diagonal for ali pairs, or in case of equal preferences, an
" = " is shown both above and below the diagonal. Note thal on..the right
hand side of the mat rix is shown the number of times the factor in each row
is preferred . Thus, for this example, it is found that the rank order is 8 >
A > C = E > D.
Sec. 8.6
Weighting of Attributes
A
A 
B
p
c
D
E
Table 83
I LLUSTRATION OF PREFERENCE CoMPARISONS'
Number
or Times
8 c D E Prererred
p p p 3

p p p 4

p ; Il

0
; p

Il
"P" irrow ractor is prererred to column ractor or"; .. ir
they are equally prererred.
223
The foregoing scheme of deducing rankings assumes rransiriuiry of
preferences. Th at is, if X> Y and Y > Z, th en X must be> Z. If the number
of times a gi ven factor is preferred to another is equal fort wo or more factors
(except in the case of ties), there is evidence of lack of consistency, which
suggests the need for questioning preference judgments for the factors af
fected.
8.6 WEIGHTING OF ATIRIBUTES*
Weighting of attributes involves quantifying the relative importance of
each. Thefollowing metho ofweighting attributes (again called " factors")
is based on two assumptions:
1. It must be possible for the.Jecision maker to consider and judge the
relative weight of any combination of factors. That is, it must be possi
ble to cons id er not only the weight of F 1 , but also the weight of both F
1
and F2.
2. W_ights are assumed to be additive. That is, given the weight of F
1
and
the weight of F
2
, the weight of both F 1 and F1 is the sum of their
individual weights.
The decision maker could proceed in weighting the factors as follows :
1. Rank the factors according to decreasing preference or importance by
any method, such as ordinal scaling illustrated above. Once this is
This section is adapted rrom William T. Morris, Tire Analysis of Management Deci
sions, Richard D. Irwin, lnc. , Homewood, Ill . , !964, pp. 417424. with permission or the
au thor.
....J
w
+::>
L24
Multiauribute Decision Analysis
Chap. 8
done, ir mighr be desirable to assign or reassign subscripts to indicate
the rankings. Thus F, is ranked first, F
2
is ranked second, and so on.
. 1
2. Let the weight of F1 equal lOO [i.e., W(F
1
) = 100] and weight the(other
factors so asto retlectjudgment oftheir weight relative to the weight of
F
1
Th us the weights of F2 F N will range from a possible .high
value of lOO to a possible low value of O. Ail further steps (3 through 6)
serve ro retine these initial weightingjudgments and to ensure they are
internally consistent.
3. Compare F
1
with the combination of F
2
+ F
3
(the plus sign here means
" and"). If F1 is considered to have Jess weight than F
2
+ F
3
, the
weighrs assigned in step 2 must satisfy the relation W(F
1
) < W(F
2
) +
W( FJ). If !he assigned weights do not satisfy this relation, W(F
1
)
should be adjusred until the relation above corresponds to thejudgment
on relative weights. If, on the other hand, F
1
is considered to have
greater weight than F2 + F3 , the inequality sign of the relation above
would be reversed, and, ifnecessary, the weight of F
1
may be adjusted.
4. Compare F, with the combination F
2
+ F
3
+ F
4
, and repeat the process
of adjusting the value of W(F
1
) if necessary.
5. The process of comparison ilnd adjustment is continued according to
rhe following pattern (for a given problem, it is to be expected that
many of these comparisons will not be needed, thus shortening the
process):
Compare F, with F
2
+ F
3
+ F4 + F
5
Compare F1 with F2 + F
3
+ F4 + F
5
+ + FN
Compare F2 with F
3
+ F4
Compare F2 with F
3
+ F4 + Fs
Compare F2 with F
3
+ F4 + F
5
+
. + FN
Compare FN z wirhFN 1
+ FN
~
6. Once !he weights [ W(F1) + W(F
2
) + + W(FN)] have been ad
justed and checked for consistency to the satisfaction of the decision
! maker, it is common to "normalize" them to sum to 100 points by
multiplying each individual weight by
Sec. 8.6 Weighting of Allributes 225
100
N
2,: W(F;)
i= l
Example 85
Suppose that management desires to establish weights for the five factors or attri
butes identified above. The following are the identifications and initial weighting
assignments for the factors in rank order.
Factor (Attribute) Name Identificati on
Net present worth B
CIMS tactical aims A
Serviceability C
\
Riskiness, lack of E
Management/engineering effort D
Initial
Weighting
Asignment
1{)()
65
40
40
15
It is desired to illustrate the determination of weights in the spirit of, but not exactly
following, the consistency checking procedure in steps 3 through 5 (above).
The first consistency check will be to compare B with A + C. Suppose that the
decision maker feels that B is slightly less important than the combination of A and
C. This judgment agrees with the initial assignment of weights. That is,
Judgment: W(B) < W(A) + W(C), where W(B) is weight of factor B, etc.
Weights: 100 < 65 + 40; 100 < 105; .'. OK
The next comparison could be bJtween A and C + E. Suppose the decision
maker feels that A is more important fuan the combination of C and E. A check of
this compared to the assigned weights shows
Judgment : W(A) > W(C) + W(E)
Weights: 65 < 40 + 40; 65 < 80; .' . NOT OK
Thus there is an inconsistency to be resolved. This might be done by increas
ing W(A) alone by m o ~ than 15 points (i.e., 80  65), by decreasing W(C) and/or
W() by a sum of more than 15 points, or by sorne combination ofthese adjustments
for consistency. Suppose it is judged that W(A) should be increased by 5 points and
that W(C) and W(E) each should be decreased by 10 points . With these changes, the
weights are fou nd to be consistent with the judgment.
Judgment : W(A) > W(C) + W(E)
Weights : 70 > 30 + 30; 70 > 60; . . . OK
Ifadjustments in point assignments are significant, it may be necessary to recto prior
consistency checks. This process may be continued until one is satisfied with tht

"'J
w
V1
i
i
1':
( .
i '
l
'
1
1
' {;
:I:
.
t'
1
t
:/,
1
!
226
Mu/tialtribute Decision Ana/ysis
Chap. 8
84
C. .. LCULATION OF NORMALIZED FACTOR WEIGHTS
Factor (Attribute)
B. Net present worth
A. CIMS tactical aims
C. Serviceability
Weight
1= W()J
100
70
D. Management/engineering eiTort
E. Riskiness , Jack of
30
20
30
LW( ) = 250
Nonnalized Factor Weight
[
= W() x 100 w,]
LW()
40
28
12
8
12
L = 100
w_eights . Table 84 shows the normalization (to sum to 100) of the factor weights,
which presumably have been checked sufficiently for consistency.
8.6. 1 Formula Methods for Weighting
Many numerical formula methods for assigning weights exist that are
easier but generally much less defensible than the procedure advocated
above (even if checking for consistency is not done completely). Severa! are
described below, and example calculations are shown in Table 85 for the
same five factors (altributes) as shown in Table 84. As in Table 84, the
weighls are expressed as percentages.
1. Uniform or equa/ weights. Given N altributes, the weight for each,
W1 == liN x 100%
(81)
2. Rank sum weights. If R1 is the rank position of attribute i (with 1 the
highest rank, etc.) and there are N attributes, rank sum weights for
each altribute, W;, may be calculated as
W == N  R; + 1 x 100%
' N .
L (N  R
1
+ 1)
(82)
3. Rank reciproca/ weights. Rank reciprocal weights using the same no
tation as above may be calculated as
IIR
1
W; == N x 100%
(83)
L (t i R;)
j ::z l
"

\
V) "'
oj
"::>
:;;l
1:< 0
l
...
2
;:!
"'
Il
1 VI
"'
;;;
>
"
"'
"'
en
.:.:
)o c:
.. ..
3
"' ::c
1
..,
11:::0 0000 N
;.:;
6 Il
...
""'
0
z
0
1=
j
::>
x;; 1 .., r 0.., ,_,8
u
...
.. 1"'"\NN
< 
u
1.
"'
1 H
...
o.
E
::>
< Vl
)(
.:.:
+
1 Il
:C:, H
3:
jj
1 oooool8  NNNMN
Il Il
H 1
v
t:
li
@
" 00
c:
c:
"' g
5
"' .z: E ...., '60
'5 5 ...
c:
... u 
li
<  (,) 0
C ::::  ,.. U .z;;
'J
"'
"' g bO .:.:
w
"'"' :.;;! ..
c:
:::n:
..
0'\
"'
227

;it
{
''l
.,_(
) ,
L L /l
Multiatlribute Decision Analysis Chap. 8
0
Comparing the methods in Table 85, note that the rank recipmcal
method gives the highest weight for the firstranked attribute. One might
arbitrarily choose among the weighting methods above according to which
provides the closest approximation to the independently judged weight for
the highest ranked attribute. If we take that independently judged weight to
be the 40% in Table 84, it is seen to be closest to the 44% for the rank
reciprocal method.
8.7 WEIGHTEO EVALUATION OF ALTERNATIVES
Once weights have been assigned to attributes, the next step is to assign
numerical values regarding the degree to which each alternative satisfies
each attribute. This is generally a difficult judgment task using an arbitrary
scale of, say, between 0 and 10 or between 0 and 1,000 to refl,ect relative
evaluations for eac!h alternative and each attribute.
Example 86
Suppose thal we are comparing two alternatives on the basis of how.wellthey satisfy
the live attributes with the weights developed in Table 84. The attributes, together
with the subjective evaluation of how weil a particular alternative meets each on the
ba sis of a sc ale of 0 to 10, are shown in Table 86. [Note: The se evaluation ratings
roughly correspond to the " scorecard" (Table 82), and sample graphs in Fig. 82.
Additionally, the " net present worth" amounts have been converted to a scale of 0 to
10. ].
Once the evaluations have been made, the results can be calculated as in Table
87 to arrive at weighted evaluations of attributes for each alternative. Thus the
summed weighted evaluation is 70.0 for alternative P1 and 74.2.for. alternative..P2,
(using Equation 8.4 at the bottom of Table 87), which indicates that alternative P2 is
the better even though it happened to have the lower evaluation rating for three out
of the live attribut es.
Table 86
ExAMPLE EVALUATION RATtNG OF How WELL
Factor
Net present worth
CIMS tactical aims
Serviceability
Riskiness, Jack of
EACH ALTERNATIVE SATISFIES
EACH FACTOR (ATTRIBUTE)
Allernative
P1 P2
6 7
7.5 9
10 7.5
8 6
Management/engineering effort
8 6
\
,_ "'
. "'
CIO >
..0 <
"'z
.... ..:
"' !::;
<(
...
0
"' z
0
::>
...
<
>
Ill
c
"'
....
:.:
0
j;j
...
0
z
0
...
::>
0
N
tl..
.,
.::
E
<
'il il
c
.g CIJ
"' c ,;;
<3 "'
>QG
Ill
.,
.2
1il,'ii
 :l
.. 
;,::
<
c
.g
=
Ul
"E
:ifo
z:ge
 0
;:s
u
:;
..0
s
<
C? oql "!
OOV"\0\f'"11' "'f'
N N t
"' r0\ r: \CI...o
Il
vI
M
N N r
"' ..or=:2oo tX>
OOO NN OO
"'t N 
t:
.g
0
QI)
c
c:
0
u
"' c
' 'o
0 .>< u
B]:;
s
Q,Vl .\!] "'
 :,z t "';
tl
vI
...
.5
.SI!=
'iii
:l
il
>
u
x
.E
00
'
u
:;
..0
s
'iii
.,
0
c
Il
c
.!2
'iii
:l
il
>
.,
.,
.,
.E
co
'
229
"'J
w
"'J
Ait. P 1
figure 84. Sample rormat to racilltat e calcul ation or weighted evaluation.
Format adapted substantially rrom "objectives matrix" ror measuring productivity
performance by Oregon Productivity Center, Corvalli s, OR; adapted with permission or ils
highl y productive direct or, the tate Dr. James E. Riggs.
230
.
t 
Sec. 8.7
Weighted Evaluation of Alternatives
Project P 1
10
"'
8
c
. .,
:!
c
s 0
::>
;;;
>
4 w
\
1
r_l _____ ;.
1 1
L __
r
1 1
1 1
1
1 
1
1 1 1 1 t:l
1
1
1
1 1;:; 2 1
1
V) i 1 :0 1 1 1
1
h 1 asn 1
1 lE!!' 1
!5
1
u:l 1
t&;: 1
1
1 ij
\. 1
l!l 1 cl!
1 1
z
1 1
1 a:
1
1
1
1 1 L
1 1 ii 1
li
I O 10 20 30 40 50 60 70 80 90 100
Normalized weight , cumulative
Figure 8S. Graphical portrayal or weighted evaluations ror two allemalives.
4_
8.7.1 Tabular and Graphical Ways to Show
Weighted Evaluation of Alternatives
231
Figure 82 showed sorne typical graphs relating attribute outcomes to
evaluaijpn ratings . Now we illustrate in Fig. 84 a tabular means which
shows the same information and in addition provides a format* for easily
calculating weighted evaluations. The illustration is for project P1 only,
and we see once again that it has a weighted evaluation of 700 (out of 1,000) ,
which is the same as 70 out of 100. The same form can be used for compar
ing alternatives, together if desired.
Figure 85 is a graphical way of depicting the weighted evaluation of
alternatives Pl and P2 as calculated in Table 87. The attribute weights are
shown to scale on the horizontal axis , and the evaluation ratings are shown
to scale on the vertical axis. Hence the areas for each alternative are in
proportion to their respective weighted evaluations. ln this case the differ
ence does not stand out , but project P2 does have a somewhat greater
area. Figure 86 shows the differences in weighted evaluations for the two
lfernatives, which may facilitate graphical interpretation better than
Fig. 85.
..
'1
w
co
BL Multiattribute Decision Ana/ysis Chap. 8
"'
c:
2
d. d. 1 
0
K 1
<] 2
P2 > P1
40
P2 < P1
Normalized
weight,
cumulative
Figure 86. Graphical portrayal of difference in weighted evaluations for Iwo
alternatives.
8.8 VARIATION OF WEIGHTED EVALUATION METHOD
THAT CONSIDERS OBJECTIVE (ECONOMIC)
MEASURES AND SUBJECTIVE MEASURES*
The following variation of the weighted evaluation mode! involves first
transforming any objective (economie) measure of merit for each alternative
into a score between 0 and 1 so that the sum of the scores totals
1. For subjective attributes, such as lack of riskiness, nominal ratings are
th en transformed into nu me rica! scores, and the relative importances .of the
attribi.Jtes are weighted so thal the scores of each attribute over ali alterna
tives and the weighted (subjective) scores over ali attributes each sum to 1.
Next, weights are assigned to the objective and the subjective score.s such
that those two weights sum to 1. The combined weighted scores (which also
sum to 1) show the relative desirability for each alternative.
Expressed in formula form, the combined measure (weighted evalua
tion) for each alternative k, WEt. is
WEk = (a)(OMk) + (1  a)(SMk)
where OMk = objective measure for alternative k
SMk = subjective measure for alternative k
a = relative importance weighting for OMk
and where OMk, SMk, and a are each 0 and :s 1.
(85)
We will demon strate determination of OMk, SMk, and WEk by exam
pies below. The examples also include the use of certain formulas or ap
proaches for determining weights and evaluation ratings which are merely
different than the more general methods for determining weighted evalua
tions shown previously.
Method tirst developed in 1972 as the BrownGibson mode!. One signiticant version
with generalized modifications was b.y . .E' . Y. Huangand P. Ghandfo.{oush,
for Evaluating, Robots ," Indus trial Engineering, M48. Adapted
with permission of publisher. . . . ..... .. . .
"'
Sec. 8.8
Variation of Weight ed Evaluation Method
J3
Objective Measure
The objective measure for each alterative k, OMk, can be calculated by
one of the following:
(a)
If measure is returns or profits for each alternative k, cali it OFPt
(such as net PW), then
\ OM
_ OFPk
k K
L OFPk
where K = total number
of alternatives
(86)
(b) Jf measure is costs, for each alternative k, cali it OFCk (such as PWC
or AC), then
OMk = OFC
1
x S
(87)
where
s=L.
k=t OFCk
Example 87
Alternative CIM systems are expected to have the following total equivalent a nnual
costs.
....
Alternative k
Il
Ill
OFC,
$130,000
150,000
165,000
Determine the OM for each alternative.
Solution
and
/
Similarly,
1 1 1
s = 130,000 + 150,000 + 165,000
= 0.000020418
1
OMt = 130,000 x 0.000020418 =
0
377
OMu=
OMw =
= 0.326
= 0.297
L = 1.00
'.J
w
1.!)
234
00 z
i=
.0 <
z
0
=
<
:::>
....
<
>
t1l
0
z
<
"' ...
"'
0
;;;
w
0!
:::>
Cil
il:
>
E
<
...
.8
.
.. 00
3 .5
..
> ..
tl.JO::
H
"'
"" .5
..
a::
c
.Q
..
::s
...
>
t1l
:::
" "
>
::: ::s
u.o
" s 'E_
,;j<
il:,
.. 
E::
... " <
0
U.tllO
tllU.O
c
NON
8
O>O
"'
.:;
gg
0000
>>
5
Soo
... >.
8 t;
o..>
:!.
U.tll
"' >.
e o o " >
c....; .iJ:o
.... {l
.:: 0 c. > ti.
:::::1 0 ")(
Cl u <"l
0 V\0
N
H 
Sec. 8.8 Variation of Weighted Evaluation Method
235
Subjective Measure
The subjective measure for each alternative k, SM*, can be calculated
for N subjective attributes, as we demonstrated earlier for weighted evalua
. tions in Section 8. 7\
N
SMk = 2: (subjective attribute weight1) x (subjective evaluation rating;t)
1= 1
1
(88)
For the example below, let us assume thal subjective attribute weights are
assigned with a 20 being the highest, and that the subjective evaluation
ratings are assigned as follows:
Example 88
excellent = 4
very good = 3
good = 2
fair= 1 ..J_
poor = 0
For the same alternative CIM systems for which objective measures were shown
above, Table 88 gi'les mw subjective weights and evaluation ratings for each. Table
89 shows those numbers each divided by their respective totals, so they will add to
1. These we might cali "unitized" weights and ratings . Calculate the SM for each.
Solution
20 (2) 15 (3) 10(2) 10 (1) 5 ( 4 )
. = 60 7 + 60 8 + 60 6 + 60 4 + 60 10 = 0.3
20
Similarly (as shown on bottom of Table 89),
SMn =
SM111=
= 0.302
= 0.378
L = 1.000
Table 89
UNITIZED SUBJECTIVE WEIGIITS AND EVALUATION RATINGS'
Unitized /
Weight
20160 \
' 15/60 \
10/60 1
10/60 /
5160/
L a J.OQ
Subjective
Attribute
Quality of results
Ease of use
Competitive
Adaptability
Expandability
SM1
U nitized Evaluation
Ratings for Alternatives
1 Il Ill
2n 417 117
3/8 1/8 4/8
2/6 2/6 2/6
114 0 3/4
4/10 3/10 3/10
0.320 0.302 0.378
Final SMs are shown at bottom for each alternative .
l: Evaluation
Ratings
1.0
1.0
1.0
1.0
1.0
L SM, = 1.0
.....
'.!
0
l36 Multiattribute Decision Analysis Chap. 8
Weighted Evaluation
The final weighted evaluation results for each alternative k (WE*)
for any a ( 2!: 0 and :5 1) can now be calculated as WE* = (a)(OM*) +
( 1  a)( SM*) .
Example 89
If a = 0.3, calculate the WEs.
Solution
WE
1
: (0.3)(0.377) + (1  0.3)(0.320) = 0.337
WE
11
: (0.3)(0.326) + (1  0.3)(0.302) = 0.309
WE
111
: (0.3)(0.297) + (1  0.3)(0.378) = 0.354
L = 1.000
Thus. for a = 0.3, alternative III is slightly better than alternative 1, which is some
what better than alternative IL
Very often it is useful to depict the effects of differing weights on the
weighted evaluation as illustrated in Fig. 87. The vertical dashed line at
0.38
0.377
0.37
0.36 
c 0.35
0.34
'0
Il 0.33
.c
0> 0.326
0.32
0.31
0.30 0.42 = " breakeven" for
alts. 1 and Ill "'1 0.297
0.29
0 0.2 0.4 0.6 0.8 1.0
a '"' weight of objective measure
1.0 0.8 0.6 0.4 0.2 0
( 1  a) a weight of subjective measure
Figurr 87. Weighted evaluation for ali possible weights ofOM and SM in Exam
ple 89.
t
Chap. 8 References 237
a = 0.3 helps 'onfirm the results above for the three alternatives. The graph
shows, among other things, that alternative 1 is preferred to alternative li for
ali values of a , and that alternative 1 becomes preferred to alternative Ill for
ali a slightly greater than 0.4. lndeed, the " breakeven" value of a can be
calculated by equating the WEs for alternatives 1 and Ill as follows:
{a)(0.377) + (l  a)(0.320) = (a)(0.297) + (1  a)(0.378)
a = 0.42
8.9 SUMMARY AND PROLOGUE
This chapter has concentrated on basiq{raphical, tabular, and additive
weighting techniques for aiding in the analysis and resolution of multipl e
attribu._te deeisin problems. In the following four chapters we explain ot her
more detailed, quantitative methodologies; namely, use of the multiattribut e
utility method, analytic hierarchy process, goal programming, and expert
system technology.
It should be recognized th at presentation of most of the methodologies
in this and other chapters is based on "assumed certainty. " Thal is, we
normally assume that required input estimates (such as weights, ratings,
tradeoffs, etc.) are known quantities . Of course, most real
decisionmaking situations involve uncertainties. The main recommended
approach to dealing with uncertainties is to perform sensitivity studies to
determine the relative effect on the indicated decision(s) due to changes in
important input estimates for whatever methodology and to display the
results in easytounderstand ways .
The determination of what methodology(ies) to use depends on the
nature of the decision problem and the preferences of the decision
maker(s). Although this determination is a very inexact science, it can be
shown J)lat the use of widely differing models will often have Jess effect on
the probable quality of the solution than does the unintended omission of
possible outcomes, alternatives, or important criteria. Whatever methodol 
ogy is chosen should make the decision maker more confident and ultimately
make his or her job easier.
REFERENCES
1. MacCrimmon, Ken R., Decision Making among Multiple Attribute Alternatives :
A Survey and Consolidated Approach, Memo RM4823ARPA, Rand Corpora
tion, December 1968.
2. Canada, John R. and Robert L. Edwards, Should We Alllomate Now? Evaluation
of Computer lntegrated Manufacturing Systems, Industrial Extensiotl Service,
North Carolina State University, 1986.
'.1
t:>.
f'
1
:1
. 1
,,
1 i
238
Multiallribwe Decision Analysis Chap. 8
EXERCISES
81. !Section 1!.4) Select a class or type of significant decision problem in your work
or personallife. Name the three or more significant auri butes to be considered
for thal class of problem. Assume that the weighting of these attributes is
nebulous, so you think il desirable to develop only a matrix of alternatives
versus attributes, with entries to show how weil each alternative meets each
attribute in whatever measures are appropriate, such as in dollars, lime, rank,
and so on. Th en use co lors or other symbol codings to facilitate ease of under
standing of the differences by the persons making the final selection among the
alternatives .
82. (Sections 8.5 through 8.7) Select a class or type of decision problem in your
work or personallife involving multiple objectives (not necessarily same as for
Problem 8 1 ). Na me three to live most important objectives for that class of
problem. as you see them, and do the following:
(a) Using the method of paired comparisons, rank the objectives in order of
decreasing importance.
(b) Weight them and check for consistency and then normalize the weight to
sum to 100.
(c) ldentify two to four alternatives for the decision problem and evaluate how
weil each alternative meets each objective on a scale of 0 to 10. Then
multiply the evaluation ratings by the objective weights and sum these for
each objective to compute a weighted evaluation for each alternative.
83. (Sections 8.5 through 8.7)
(a) Weight the relative importance of the three or four most important aUri
butes you would consider in selecting a job. Assume that these are the
only attributes you will quantitatively consider. Show how you make
comparisons for internai consistency and normalize the factor weights to
sum to 100.
(b) Using the attribute weights developed above, obtain a weighted evaluation
of two alternative jobs in which you might be interested by evaluating how
weil each job satisfies each attribute using a scale from 0 to 10.
84. (Sections 8.5 through 8.7)
(a) Weight the relative importance of the three to six most important auri
butes you would consider in selecting a persona! car. Assume that these
are the only attributes you will quantitatively consider. Show how yo6.
make comparisons for internai consistency and normalize the attribute
weights to sum to 1,000.
(b) Using the attribute weights developed above, obtain a weighted evaluation
of three alternative makes of cars you might consider for purchase by
evaluating how weil each make of car satisfies each attribute. Use a scale
from 0 to 20.
85. (Sections 8.4 through 8.7) Describe a significant multiattribute decision prob
lem of existing or potential meaningful interest to you. Examples are: housing,
car, job, spouse, !ravel, and equipment. ldentify at !east four most .relevant
attributes which are as independent of each other as possible and at !east two
Chap. 8 Exercises
i 0
, , L 239
'
.. .
or three mutually exclusive alternatives. (Note: ln the process of further anal
ysis, you might weil add or delete alternatives and/or attributes .)
(a) Describe the problem as meaningfully as possible using two or more
graphical techniques . What conclusion (choices), if any, can you make
from these?
1
(b) Rank order the attributes. Can you combine this with one or more of the
graphical techniques in part (a) to make choices? If so. describe .
(c) Weight the aUri butes by the method or formula(s) of your choice and show
at !east two checks and/or revisions of weights for consistency between
initial judgments and preferences.
(d) For each attribute show a graph of attribute out come (x axis) versus evalu
ation rating (y axis) .
(e) Using parts (c) and (d), show calculations weighted evaluations for
each alternative and display the results for ease of comparison.
86. (Section 8. 1) For' multiattribute decision problem of existing or
potential meaningful interest to you (preferably an entirely different problem
from Exercises 81 through 85), apply "Ben Franklin's method" of listing
pros and cons and canceling out combinations thereof. His description as
sumed just one alternative or course of action is being considered. If two or
more altematives are considered, his process can be used for each alternative
by il self, and further cancellations can be do ne if one can fi nd one or more pros
which areequal for ali alternatives, or one or more cons which are equal for ali
alternatives .
87. (Sections through 8.7) De scribe a significant multiattribute decision prob
lem of exasting meaningful interest to a "client" (friend, business person, or
relative). Describe the problem as clearly as possible using words, tables,
graphs, and so on, in sufficient detail to facilitate analysis by two or more
methods described in this chapter. [Note: Problem should consist of at !east
three (independent as possible) attributes and at !east two alternatives . ln the
process of analysis you might weil add or del ete alternatives and/or attributes ,
1
\ but do not go below the minimums above.)
\ 88,1 (Section !r.'3) Given the following matrix of outcomes for alternatives and
, attributes (with higher numbers being better), show what you can conclude
using the following sequential elimination methods:
(a) Conjunctive constraint.
(b) Dominance.
(c) Lexicography, with rank order of attributes D > C > B > A.
Alternative
Standard
Attribute t 2 3 "Ideal" (Minimum Acceptable)
A 60 75 90 100 70
B 7 7 8 JO 6
c Poor Excellent Fair Excellent Good
D 7 8 8 10 6
'.J
N
240 Multiallribute Decision Analysis Chap. 8
89. (Sections 8. 3, 8.6, and 8. 7) Management has asked you to look into moving the
"cutting" operation from where it is now to a new location. Management has
given you three alternative areas, which are mutually exclusive, to consider.
After inspecting the three areas and considering which factors reHect signifi
cant differences among the alternatives, you have decided on five independent (
attributes by which to evaluate the alternatives, listed in descending order of
importance:
A. Distance traveled from cutting operation to next operation [more distance
is worse]
B. Stability of foundation [strong (excellent) to weak (poor)]
C. Access to loading and unloading [close (excellent) to far (poor)]
D. Cost of the moving operation
E. Storage capacity
(Note: The cost for operating the cutter once the operation has been moved is
independent of the area chosen and hence is the same for each area.) The data
of these factors for the three alternatives are as follows:
Alternative
Anribute Area 1 Area Il Area Ill Ideal Standard Worst
A 5000 ft 3000 ft 750ft Oft 3000 ft 10,000 ft
ll Good Very good Good Excellent Good Poor
c Excellent Very good Good Excellent Good Poor
D $7500 $3000 $8500 so $5000 $10,000
E 60,000 ft' 85,000 ftl 25,000 ft' 10,000 ftl 25,000 ft' 150,000 ftl
(
Determine which alternative is to be chosen using:
(a) Alternative versus aliernative: comparison across alternatives.
(l) Lexicography.
(2) Elimination by aspects .
(b) Weighted evaluation of alternatives using the rank reciprocal method for
weighting attributes.
(c) Weighted evaluation of alternatives using uniform or equal weights for
attributes .
For parts (b) and (c) the following are evaluation ra ting functions on a scale of 0
to 10.
r '
"
241
Chap. 8 Exercises
Range for
A.uribute Attribute Outcome Evaluation Rating
A 0 s x s ft x/600 + 10
3,000 s x s 10,000 ft xli ,400 + 5017
Band Excellent 10.0
Very good 7.5
Good 5.0
Fair 2.5
Poor 0
D $0 s x s $10,000 x/1,000 + 10
El  15,000 s x s 25,000 ftl 35/2  x/2.000
25,000 s x s 100,000 20/3  x/15 ,000
:r Two highway alignments have been proposed for access to a new manufactur
\_./ ing plant. Based on the data and information given below, comparison can be
made of the two alignments. The better one must be chosen as the proposed
highway alignment that will be the connector between an interstate highway
and tho proposed site. The route length for Alignment A is 4.7 miles and for
Alignment B, 5.1 miles. The monetary costs are as follows:
Item Alignment A Alignment B
Land s 4,044,662 s 4,390,000
Bridges 10,134,000 8,701,000
Pavement 4,112,500 4,;2,.500
Grade and drainage 7,050,000 7,650,000
Erosion control 470,000 510,000
Clearing and grading 188,000 204,000
/
Total $25,999,162 $25,917,500
Nonmonetary attributes are:
Alignment A Alignment B
Maintenance High
Noise pollution Very good Good
Cost savings (on gas) Excellent Poor
Accessibility to another U.S. Highway 41 None
roadway
Impact on wildlife Little Little
Relocation of residences 2 3
Road condition Flat Hill y
Which highway alignment would you recommend? Show ali work.
""'J
w
242 Multiattribute Decision Analysis Chap. 8
811. (a) Use the weighted evaluation mode! to make a selection of one of the three
automobiles for which sorne data are given below. State your assumptions
regarding miles driven each year; !ife of the automobile (how long you
would keep it); market (resale) value at end of life; interest cost; priee of
fuel; cost of annual maintenance; and attribute weights and other subjec
tively based determinations.
(b) Use the data developed in part (a) and the weighted evaluation method in
which objective and subjective measures are combined (OMk and SMk, as
in Section 8.8) to make a. selection. Do your answers in parts (a) and (b)
agree? Explain why they should (or should not) agree.
Alternative
Attribute Domestic 1 Domestic 2 Foreign
Priee $8,400 $10,000 $9,300
Gas mileage 25 mpg 30mpg 35 mpg
Type of fuel Gasoline Gasoline Diesel
Corn fort Very good Excellent Excellent
Aesthetic appeal 5 o/11 of 10 7 OUI of JO 9 OUI of JO
Passengers 4 6 4
Ease of servicing Excellent Very good Good
Performance on road Fair Very good Very good
Stereo system Po or Good Excellent
Ease of cleaning Excellent Very good Poor
upholstery
Storage space Very good Excellent Po or
812. Utilize any three multiattribute techniques described in this chapter to analyze
the following important si tuation; Y ou need to determine which job to take,
given acceptable offers from firms A, B, and C. Y ou first determine a set of
attributes thal are most important in evaluating and then construct a decision
matrix with alternatives as columns and attributes as rows, and fil! in. (Results
are given in the matrix below.) Analyze the alternatives with the different
techniques according to your perceptions (be completeil look years to gel to ~
this point!) .
Alternative (Finn)
All ri bute A B c
Starting salary $30,000 $28,500 $33,000
Opportunity for advancement Excellent Excellent Fair
Management attitude Very good Excellent Good
Location Good Fair Poor
Type of work Excellent Poor Very good
Opportunity for continuing Very good Good Fair
education
1
"
243
Chap. 8 Exercises
813. (Section 8.3) The follow)ng are evaluation ratings (on a scale of 0 to 20) of how
weil each of two alternatives satisfies each of four attributes.
Alternative
Attribute 1 2
A 15 12
B 8 20
,c
14 14
ID
13 12
If the attributes are ranked A > B > C > D in order of importance, the
minimum standard evaluation rating is 10 for each. Show which alternative
would be best by;
(a) Lexicography. ('
(b) Elimination by aspects .
(c) Dominance cheekt (assuming attribute B is eliminated from consideration).
814. (Sections 8.6 and 8.7) Given the same attributes and alternatives with evalua
tion ratings as in Exescise 813, show which alternative would be best using the
weighted evaluation methodology and each of the following attribute weighting
formulas . (Note: Normalize the summed weighted evaluations to equal 100.)
(a) Uniform.
(b) Rank sum, with A > B > C > D.
(c) Rank reciprocal, with A > B > C > D.
815. (Section 8.4) Construct a polar chart depicting the evaluation ratings for the
.two alternatives in E ~ r c i s e 813.
816. (Section 8.7) Construct a graphical portrayal of Exercise 8 14(b) with cumula
tive weights on the x axis and evaluation ratings on the y axis .
817. (Section 8.8) Volunteer alumni A and B have contributed $100M and $150M,
respectively (higher is better!). However, evaluations of their leadership suc
cess (on a 1 to 5 maximum scale) are 4 and 3, respectively . Contributions
(objective) are considered to be half as important as leadership success (sub
jective). Using the weighted evaluation method combining these measures,
which should be the" Alumnus of the Year"? What part of 1.0 total "weighted
818.
brown ies" does he score?
(Section 8.8) For Exercise 817, assume the relative weight, a, for the objec ,
tive measure (contributions) is not known. Plot the weighted evaluations for
each of the alumni for a ranging from 0 to 1.0 on the x axis .
'.1
~
~
CHAPITRE 8
Dcider face la complexit.
In traduction
Our present complex environment calls for a new logica new way to
cope with the myriad factors that affect the achievement of goals and
the consistency. of the judgments we use to draw valid conclusions. This
approach should be justifiable and appeal to our wisdom and good sense.
lt should not be so complex that only the educated can use it, but should
serve as a unifying tool for thought in general.
That we are ali very much creatures of the moment needs little debate.
Try to reconstruct what was said two minutes before in a conversation and
you will soon discover that, even for a short duration, your recollection is
fuzzy. Even when you remember, the precision of your recollection is
usually less than exacting. Our understanding of the world not only needs
repetition to improve our recoll.ection and precision, but it also depends
very much on the in/ensi/y of our participation. lt is an exaggeration to
conclude that humans are logical creatures. It is more accurate to say that
our judgment depends on the totality of our impressions, even if they
cannot ,be logically and rigorously justified. For better understanding we
need to deal with experience as an ongoing process . We need concentra
tion, repetition, diversity, debate, and, when necessary, consensus.
Severa! good !essons have been learned in recent years from political
leadership. Having a clearly defined goal has again been demonstrated to
be the core of successful statecraft. Being consistent is another vital ingre
dient . Persuasion and support for one's view are still considered two of the
main attributes of a great leader. What ensues is largely designed to help
leaders get their points across .
More and more people are finding it hard to put ali their trust in the
unspoken, unjustified, and intuitive thinking of their leaders' decisions on
complex matters. Whatever internai mechanism leaders have needs to be
1
00
1'
2 INTRODUCTION
articulated and understood. Just as language itself and the rules of thought
had to be organized in a formai manner long ago, so must we now organize
our thought processes so they lead us to good decisions. We_should be able
to say th at, given the information, we agree on the methq_g._ o(maklng the
decision (but not necessarily on the quality that Th! matter
will then become a common concern rather than a mystical phenomenon.
The following pages offer the reader a method for organizing the in
formation an.<J:judgments metfiofeflts
of feelings and logic bearing on the issues and then synthe
sizes these diverse judgments into an outcome that agrees with our intui
tive expectations as represented in ali the judgments we give.
The process contributes to solving complex problems by structuring a
hierarchy of criteria, stakeholders, and outcomes and by eliciting judg
ments to develop priorities. lt also leads to prediction of likely outcomes
according to these judgments.
The outcome can be used to rank alternatives, allocate resources, con
duct benefitlcost comparisons, exerciseontrol in the systemh.y;;;filating
the se!'lsitivity of the outcome to changes in judgment, __()ut
planning of projected and desired futures . A useful byproduct is the
measurement of how weil the leader understands the relations among
factor': . Although people generally are not consistent, the main concern
here is the strength of their inconsistency. Is their understanding close to
capturing the interactions observed? Oris it a random understanding that
only hits the target now and then?
The process described in this book has had many applications so
farto energy rationing, the Middle East conflict (1972), Sudan transport
planning (19731975), mineral exploration in Mauritania (1976), planning
for higher education in the United States (1976), the presidential election
(1976), the conflict in Northern Ireland (1977), planning for a research insti
tute (1977), terrorism (1978), predicting the outcome of a world chess
championship match (1978), product portfolio selection (1979), the stock
market (1980), the presidential election (1980), oil priees in 1990 (1980), the
conflict in South Africa (1981), and severa! specifie corporate applications.
AU the applications have contributed to refinement of the principles set
forth in tlle following pags:They nave a!so
us how to deal with larger and more complex decisig!l pr_q.blems. The
theory has been enriched and immensiyolfered by the scope and variety
6f these applications.
...
1
Making
Decisions in a
Complex World
This chapter deals with the following questions:
How can we understand com.plex problems
involvi_ng a great many factors?
What are the basic processes of thought and
behavior involved in making decisions?
What role should ethics play in the
decisionmaking process?
Why do we need a new way of thinking about
complex problems?
How might a group of people holding
varying opinions debate an issue in search of an
acceptable compromise?
What can they do about their differences?
3
<Xl
N
MAKING DECISIONS IN A COMI'I.EX WlliU.D
COPING WITH COMPLEXITY
To the best of our understanding, the world is a complex system of
interacting elements. The economy, for example, depends on energy and
other resources; the availability of energy depends on geography and poli
tics; poli tics depends on military strength; military strength depends on
technology; technology depends on ideas and resources; ideas depend on
politics for their acceptance and support; and so on. In such an intricate
network of factors, first causes and final effects cannot be identified easily.
Our minds have not yet evolved to the point where we can clearly see these
ultimate relationships and readily resolve important issues like nuclear
energy, world trade, and environmental regulations.
ln our complex world system, we are forced to cope with more prob
lems than we have the resources to handle. To deal with unstructured
social, economie, and political issues, we need to order our priorities, to
agree that one objective outweighs another in the short term, and to make
tradeoffs to serve the greatest common interest.
But it is often difficult to agree on which objective outweighs
another'particularly in complex issues where a wide margin of error is
possible in making tradeoffs. Leaders may be confused by the diverse
information provided by their assistants; they may need help in identify
ing differences of opinion and seeing positions where compromise can be
reached . They may need to know which important issues must be re
searched in depth to obtain better information and how sensitive the out 
come is to slight or drastic changes in opinion and judgments. Intuitive
thought processes that serve us weil in the familiar routine of daily )ife can
. mislead us on complicated matters where sources of information and opin
ions are varied. lncreasingly we need to articulate and map out the issues
to see whether what we think and what we feellead us to the same kind of
answers .
Most of us be lieve !ife is so complicated thal in order to solve problems
we need more complicated ways of thinking. Yet thinking even in simple
ways can be taxing. If we struggle to examine collections of only a few
ideas at a lime, how can we understand complex problems involving a
great many factors? Simple thinking about such problems leads to combi 
nations of ideas whose structure is not unlike a dish of spaghetti in which
ali strands are separate but tangled.
RATIONALE FOR A NEW FRAMEWORK
What we need is not a more complicated way of thi nking, since it is
d ifficult enough to do simple thinking. Rather, we need to view our prob
lems in an organized but complex framework thal allows for interaction
and int erdependence among factors and still enables us to think about
ORGANIZING KNOWLEDGE FOR DF.CISIONS 5
them in a simple way. This new way of thinking should be accessible to ali
without straining our innate capabilities.
The analytic hierarchy process (AHP) described in this book provides
such a framework. lt enables us to make effective decisions on complex
issues by simplifying and expediting our natural decisionmaking pro
cesses. Basically the AHP is a method of breaking down a complex, un 
structured situation into its component parts; arranging these part s, or
variables, into a hierarchie order; assigning numerical values to subj ect ive
judgments on the relative importance of each variable; and synthesizing
the judgments to determine which variables have the highest priority and
should be acted upon to influence the outcome of the situat ion.
The AHP also provides an effective structure for group decision mak
ing by imposing a discipline on the group's thought processes. The neces
sity of assigning a numerical value to each variable of the problem helps
decision makers to maintain cohesive thought patterns and to reach a
conclusion. ln addition, the consensual nature of group decision making
improves the consistency of the judgments and enhances the reliability of
the AHP as a decisionmaking tool.
In this chapter we consider the human behavioral and thought pro
cesses involved in decision making. These natural processes form the basis
for the analytic hierarchy process, which is described in Chapter 2 and
explained in greater detail in Chapters 3, 4, and 5. The rest of the book
shows how the AHP can be applied to a variety of decisionmaking situa
tions and problems .
ORGANIZING KNOWLEDGE FOR DECISIONS
The AHP can best be appreciated by first looking into how the human
mind organizes knowledge for decisions . The Iwo fundamental ap
proaches humans have developed so far for analysis are the deductive
approach and the inductive or systems approach.
The Deductive Approach and the Systems Approach
We can analyze a system logically by representing it as a network and
structuring it into chains and cycles. ln analyzing natural systems, for
example, biologists break down networks into food chains, hydrologie
cycles, and so on. After structuring the network, we look for explanations
for the functioning of ils parts . Then, by an act of imagination, as no rules
of logic exist for combining these piecemeal explanations, we synthesize
an explanation for the whole network . But thi s sci entific, deductive ap
proach ignores the feedback mechanisms among the parts and betwee n !he
parts and the environment thal affect the whole system.
00
UJ
     ~   ~             .           ~   ~ 
MAKING DECI SIONS IN A COMPLEX WORLD
Systems theorists have pointed out that we can better understand an
en tire system by examining it from a general, holistic perspective thal does
not give as much attention to the function of the parts. For example, a caris
better understood by observing how il functions in the environment than
by studying the operation of its mechanical parts . In this way we see it as a
whole; we can simultaneously perceive how the car runs and how it inter
acis with other cars, road conditions, traffic signais, and so on.
Clearly both the deductive and the systems approach contribute to our
understanding of complex systems. We can benefit by combining the two
within an integrated, logical framework: the anaJytic hierarchy process.
The AHP enables us to structure a system and its environment into mutu
ally interacting parts and then to synthesize them by measuring and rank
ing the impact of these parts on the entire system. We will see how this
process works with systems as diverse as a transport network in the Sudan
and the steel industry in the United States. By providing a new logic for
synthesis, this structured approach to decision making eliminates much of
the guesswork and confusion of our ordinary method of synthesizing an
overall explanation for a system from piecemeal explanations arrived at
through deduction.
The Role of Logic, Intuition, and Experience
The everyday way we proceed to understand and 5olve problems is to
use logical deduction to argue through familiar matters. For example, it is
easy to reason that to increase capital one must obtain a good retum on it
by investing it or by obtaining a good interest rate. In this case we know
that in order for capital to increase from size A to size B, money has to be
added to A somehow. That much we can say with bold certainty. But we
tend to treat larger, imprecise perceptions of a problem by relying on
feeling, experience, emotion, appeal to other people's understanding, and
sometimes even force . Many political problemsin advanced as weil as Jess
developed countries are handled in such a fashion . In unstructured situa
tions people often act on the ir" gut feelings" rather th an strict! y on rational
grounds . Logic plays a role mostly in arranging words and ideas after the
conclusions have been reached.
People in the public and private sectors tend to cooperate in defining
and structuring their problems broadly and richly so that ali their ideas can
be included. But when they need to explain which factors have the greatest
impact on the outcome of a decision, not even experts with the clearest
logic can hold fast to their positions in the face of objections. As a result
they are willing to compromise. Thus decisions are based not so much on
the clarity of ideas or amount of information exchanged as on the persis
tence of sorne participant in the decisionmaking process and on that per
ORGANIZING KNOWLEDGE FOR DECI SIONS 7
son's ability to persuade others to accept his or her ideas, like a politician
selling himself in a campaign.
People, then, not only have different feelings about the same situa
tion, but their feelings change or can be changed by discussion, new
evidence, and interaction with other experienced people. Usually the out
come is a compromise of many viewpoints involving substantial change in
individual attitudes . The fact is that when we make decisions, persona!
preference and persuasion usually prevail over clear and straight logic.
Our actual d.ecisionmaking processes have been illuminated by recent
studies conducted by behaviorists, other psychologists, and brain re
searchers. Let us examine their findings .
Behaviorist Theories
Those who study and explain ways, reasons, and consequences of
human and animal behavior are finding it difficult to uphold the idea that
humans are rational animais. Their theories are helping to create an atmo
sphere in which people are accepted as they really are rather than as they
were idealistically portrayed during the Renaissance and the Age of Ra
tionalism. Human behavior is enormously complex; the many theories
explaining human action are deep and multilayered, and probably ali con
tribute to our understanding of human behavior .
InstinctDrive Theory. Sorne theorists consider rational thinking to be
but a thin veneer over human behavior. Much of our action is driven by
instinctpatterns woven into the mind, bone, and muscle. Just as wasps
have an instinct for nest building and birds have their characteristic songs,
humans also follow certain unleamed patterns of behavior, such as seeking
food, mating, avoiding pain, caring for young, and so on. Although
instinctdrive theory describes such patterns, it does not explain them. lt is
inadequate to account for most adult behavior, including sentiment, value,
ambition, attitude, taste, and inclination.
ReasonImpulse Theory. We tend to regard ourselves as rational animais
capable of making choices based on objective, or real , criteria. We fee! thal
most of our decisions flow from logical necessity, not from whim and
caprice. Although we may acknowledge that needs and persona! motives
are the driving forces behind human behavior, we contend that we use
reason to attain our goals efficiently and without harm or injury. Through
reason we get what we want within the limits of available resources . And
many of us ultimately do learn to apply rational techn iques to deci sion
making, regardless of what our persona! wishes may dict ate.
But critics of this view say that our socalled reason is an abyss of
(X)
_p.
8 MAKING DECISIONS IN A COMPLEX WORLD
unconscious or barely conscious urges and habits that overwhelm the intel
lect . They argue that human relationships are essentially governed by irra
tional, emotional forces; rational appeal in most cases plays only a minor
role. These reasonimpulse theorists hold that our actions ar bas.ed on
imitation, habit, suggestion, or other subrational forms of thinking and
are rarely due to pure logic. Planned actions are the result of analysis based
on preferences as to which objectives are served bestand preferences are
strongly influenced by habit and training rather than by rational thinking.
Dynamic Field Theory. Other behaviorists point to the influence of envi
ronmental factors on human behavior. We act in response to a "dynamic
field" of stresses and tensions when we perceive the environment to deny
or fulfill the satisfaction of our wants and needs. The hierarchy of human
needs that motivate behavior has been examined by Abraham H. Maslow
and others; these needs range from the most basic physiological and safety
and security needs to sophisticated selfactualization and esthetic needs.
Leaming Theories
Most people tend to assume that the way we think and
use to develop our thinking are innately human _and that the basics of
human knowledge have come down tous as a packa.ge from heaven: But
recent learning theories argue that we learn mainly by triaLmd error and
through feeling rather than through logic. Stimulusresponse theory, .. for
example, maintains that we gravitate somewhat randomly in certain direc
tions to satisfy our needs and desires; acts that bring satisfaction are rein
forced. Gestalt theory tells us that even what we consider to be
" insight" perceiving the necessary relationships in a situationresults
from feeling.
Generally we have vague feelings and inklings about what we think
we have experienced, but we are not attentive enough to register ideas and
feelings sharply. We have no systematic way of reconstructing from mem
ory that which we did not learn, understand, or memorize consciously.
Most of our daily experience goes before our senses and passes through our
feelings like a hazy cloud that slightly moistens the environment but
makes little difference to the growth of our understanding.
Our understanding does grow when a particular experience connects
weil with our earlier experiences, not mere! y with our knowledg;or when
it shocks us, capturing our attention involuntarily as it intrudes into our
being (sometimes unconsciously) either pleasantly or forcefully.
can be defined as the ability to recognize a specifie act in the light _()f
1
,.
,.
1
r
1
(:
:
1
1
ORGANIZING KNOWLEDGE FOR DECISIONS 9
previous experience. lt is an iterative, or repeated, process of adding
knowledge thatelaborates on or expands existing knowledge. Leaming
can" be conscious and intentional, as in memorizing facts, or it can be
unconscious and unintentional, as in discovering physically from experi
ence that eating green apples results in a stomachache. We generally agree
that people who experience a phenomenon firsthand are the ones who can
best shed light on our understanding of it; indeed, knowledge derived
from experience is basic to ali understanding.
Brain Research Findings
The importance of intuition, feeling, and experience in human behav
ior and decision making is further underscored by the findings of brain
researchers. In particular, they have discovered a distinction in the func
tiors of the two halves of the neocortex of the brainthe left is the logical,
rational, and calciiiating member; the right is the intuitive, creative, and
inarculate half. The verbal half's job is to interpret for ourselves
and the world the dcisions of its mute brother. Note that the decisions are
actually made by the intuitive, not the logical, half. The left hemisphere
mere! y arranges and puts into words the insights of the right.
Studies of human perception show that our senses both condition and
limit whatever enters our consciousness. Older people who have experi
enced !ife longer often recognize that illusion is primary experience. (Per
haps this awareness results from the aging of the senses, but more likely it
is due to the wisdom of age.) Our senses shape our world; thus we can
never interpret the universe with absolute accuracy. The qualities we
study are simply those we can perceive, and the laws we develop are
concoctions of our senselimited brains.
Another !ife form could have many more senses than we have; for
example, it could have a magnetic eye, could perceive the colors of the
spectrum in white light, or could see through objects. As a result, its
consciousness would be different from ours. Of course, our senses have
expanded through human inventions such as the microscope, telescope,
and X ray. We can perceive many qualities not directly accessible to our
senses, and these are only a small portion of the potential total.
If we had more senses acting at once, we would probably have diffi
culty sorting out our perceptions and understanding their relations, given
the present stage of the evolution of our brains. Although sense data can be
organized in chronological sequence, brain events are not rigidly bound
by time. Ideas can occur before or after other ideas . Because the way we
think is fluid, we are freer to arrange our ideas in the manner we desire
for good or evil purposes.
co
lJl
!
i
t
l
~
 ..    .................           ~        ..
MAKING DECISIONS IN A COMPLEX WORLD
ETHICAL CONSIDERATIONS
When complexity makes normal !ife appear difficult and even hopeless
to many, a strong individual may emerge to take over leadership charis
matically by declaring that ali the pain and confusion are simply due to a
single problem that must be solved; Hitler is an infamous example. Such
leaders draw attention to one problem and persuade or force others to
believe in them. By oversimplifying the situation, they banish ali other
problems to narrow the perspective and make themselves appear logical in
their explanations; they linearize the workings of a system in a deductive
fashion .
We accept new ideas as truths on the basis of the manner in which
they cohere with knowledge we already possess. A system of beliefs could
be perfectly consistent, and yet each belief could be false . Severa! examples
come to mind. The Gregorian calendar (1582) was based on a set of consis
tent assumptions that the earth is the center of the universe. Yet it was so
accurate that the accumulated error was one day every 3323 years . Our
calendar is in error every 20,000 years . This is an example of a consistent
prevailing view that was in error according to modern astronomy but
produced good results. An example of a consistent theory of our ti me that
produced bad results until recently involves the humble golf bali. People
used to believe that it is the perfectly spherical golf bali thal travels farthest
when hituntil it was discovered that the more a bali is used, the more
dents and dimples it develops, which serve as wings countering drag and
sustaining it longer in the air. Now golf balls are made according to pre
cisely dimpled patterns.
Consistent thinking with no real validity is frequently espoused by
lunatics and other mental cases (sometimes of a seemingly respectable
genre) . Beliefs can be deduced from one another in perfect consistency
depending solely on the observance of formai, linear relationships be
tween them. Thus one may develop a consistent deductive system with no
real validity in this world.
Because it is possible in complex, unstructured situations to present
convincing arguments that have little correspondence to reality and may
harm society, one must apply certain ethical standards to the decision
making process. The philosopher Alasdair Maclntyre of Boston University
has identified four qualities thal should characterize a decision maker's
approach to dealing with social issues :
Truthfulness by not oversimplifying complexity . Our political and legisla
tive processes demonstrate thal it is easier to consider issues such as
environmental protection or health care in a narrow, piecemeal way
than to look at ali the critical variables, fit them together, and determine
their priorities and implications. ln the short run a simplistic approach
PERSPECTIVE 11
may satisfy local contending parties, but it is no way to get at the
answers to complex problems.
Justice by evaluating costs and benefits and assig11ing costs to those who get
the benefits. Everyone involved in a decisionmaking situation family
members deciding whether to purchase a home computer or corporate
executives deciding which companies to invest inshould have a
chance to weigh costs and benefits. Those who receive the benefits
should be the ones who pay the costs, and vice versa. Justice demands
not only thal everyone have a voice and a vote but also thal those who
will bear the risks and dangers have more of a voice and vote than
others. (People should also be informed about where and how they are
paying for benefits, particularly for commodities whose priees are con
trolled or subsidized.)
Ability to plan for the unknown by calculating changes, determi11il18 where
they are likely to occur, and deciding which przorities should di ctate actio11 .
Leaders must be able tb plan and deal both with projected futures, such
as higher energy priees in 1985, and with desired and less predictable
futures, such as energy independence by the year 2000.
Flexibility in adapting to cha11ge by plallni11g , implementi11g , a11d, i11 re
spOIISe to new conditions , repla1111ing and reimplementing . This iterative
approach is essentially a learning process; it tempers our tendency to let
immediate needs dictate short lerm solutions . For example, fle xibility is
necessary in planning strategies for the use of alternative resources in
dealing with the energy situation.
As we will see, the analytic hierarchy process encourages socially re
sponsible decision making by helping leaders to avoid oversimplification,
to identify and evaluate costs and benefits, to plan for the future, and to
adapt to change.
PERSPECTIVE
Most of us have trouble coping with ordinary problems of society that
cannot be understood bya deductive, linear, causeandeffect explanation.
Our respect for the scientific method, which relies on deduction, has led us
to try to solve ali our problems through logical debate. As a result of our
scientific education and because science usually deals with things we can
observe through our physical senses, we are made to fee! that there is
precis_ion in what we do. Our senses are trained to be consistent in focus 
ing on their abjects; thus our minds have a sense of consistency in syn
thesizing and interpreting sense data. But when we deal direct! y with ideas
rather than with sense perceptions, things tose their precision. The reason
(X)
0"'1
...   
MAKING DECISIONS IN A COMPLEX WOHLD
is that we use words whose meanings are imprecise. Philosophers have
long recognized that primary language does not express thoughts or ideas
but feelings and affections. Moreover, we cannot be exact in describing
abstract relationships, and our understanding is conditioned by our states.
of mind, feeling, and imagination at the moment we are thinking.
Thought without language is impossible. Abelard says that "language is
generated by intellect and generales intellect." According to the Chan
dogya Upanishad, "The essence of man is speech."
We are very much creatures of the moment. At any given time our
attention is captured by whatever our senses perceive. We cannot re
member the past de arly, not even what the tomorrow of our persona! lives
will be like, despite the fact that sorne of us may have lived more than ten
thousand days on this earth, repeating the same pattern that many times .
With ali this experience we are still unable to see the immediate future
with adequate clarity. But sorne would try to venture predictions on seri
ous issues of politics or economies in which they have had little expe
rience.
To understand and deal with what is going on in the world, we need
to improve our recollection of events and the precision of our knowledge
by reviewing the facts and organizing them in a logical framework. If we
are to make decisions thal are rational and effective, we must participate
intensely in the act of understanding the world around us. It is an exagger
ation to say thal humans are logical creatures. More accurately, our under
standing is filtered through our senses, and our judgment relies on often
hazy impressions of reality. With experience and through the perceptions
and opinions of other people, our views of reality may change and become
more precise. For a better understanding of the world, we need to perse
vere in thinking matters through carefully and to debate with others who
hold different views.
But the complexity of social systems cannot await a full, logical analy
sis of situations on which our health, safety, and even survival depend. We
need to rethink the traditional use of logic to derive knowledge. We also
need to expand our analytic procedures to improve our understanding of
situations in which not only time and space but also human behavior pla ys
a fundamental role in determining the outcome.
The analytic hierarchy process enables decision makers to represent
the simultaneous interaction of many factors in complex, unstructured
situations. lt helps them to identify and set priorities on the basis of their
objectives and their knowledge and experience of each problem. As we
have seen, our feelings and intuitive judgments are probably more rep
resentative of our thinking and behavior than are our verbalizations of
them. The new framework organizes feelings and intuitive judgments as
weil as logic so that we can map out complex situations as we perceive
them. lt reflects the simple, intuitive way we actually deal with problems,
KEY CONCEPTS 13
but it improves and streamlines the process by providing a structured
approach to decision making.
KEY CONCEPTS
0 In our complex world system, we are forced to cope with more prob
lems than we have the resources to handle.
0 What we need is not a more complicated way of thinking but a frame
work thal will enable us to think of complex problems in a simple way.
0 There are two fundamental approaches to solving problems: the de
ductive approach and the systems approach. Basically, the deductive
approach focuses on the parts whereas the systems approach concen
trates on the workings of the whole. The analytic hierarchy process,
the approach proposed in this book, combines these two approaches
into one integrated, logical framework.
0 Humans are not often logical creatures. Most of the time we base our
judgments on hazy impressions of reality and then use logic to defend
our conclusions.
0 The analytic hierarchy process organizes feelings and intuition and
logic in a structured approach to decision making.
co
'1
i
!
l
2
The Analytic
Hierarchy
Process
.       
This chapter deals with the following questions:
What are the three basic principles of logical
analysis?
How do these three principles relate to a new
approach to decision makingthe analytic .
hierarchy process?
What can we do when the usual scales of
measurementdollars, time, tons, and so
forthfail to measure intangible qualities?
Why is the analytic hierarchy process such a
powerful method for tackling complex political
and socioeconomic problems?
What can you expect to gain by using the
analytic hierarchy process?
14
WHAT IT IS AND HOW IT WORKS 15
WHAT IT IS AND HOW 1T WORKS
To introduce the analytic hierarchy process, consider the following
example of a decision problem. The Brandywine River Region in Pennsyl
vania faces possible urbanization and its environmental effects. What
actions should the people of the region take to maintain environmental
quality? Should they allow development and invest money to prevent
environmental deterioration or should they limit development?
Planners who used the AHP to study this problem first defined the
situation carefully, including as many relevant details as possible. Then
they structured it into a hierarchy of levels of deil (Figure 21) . The
highest leve! was the overall objective of protecting environmental quality.
The lowest included the final actions, or alternative plans, thal would con 
tribute positively or negatively to the main objective through their impact
on the intermediate criteria. The alternatives were (A) to leave the area
nonurbanized, (B) to allow partial urbanization, and (C) to allow total
urbanization. The intermediate levels of the hierarchy comprised the Iwo
basic criteria for evaluating environmental quality: (1) esthetic criteria,
which were further structured into properties of vividness, intactness, and
Environmental Quality
Esthetics Hydrology
~ ~
Vividness
~
No Noise ,
No Disturbances
lntactness
No
Flooding
Water
Quality
~
~ ~ ~
Nonurbanized
(A)
Partly
Urbanized
(B)
Totally
Urbanized
(C)
Channel
Naturalness
~
Figure 21 Hierarchy for Brandywine River Regiof1
(X)
(X)
THE ANAL YTIC lliERARCHY PROCESS
no noise or disturbances; and (2) hydrologie criteria, subdivided into no
flooding, water quality, and channel naturalness. This hierarchy graphi
cally depicts the interdependence of elements in the problem; it both iso
laies the relevant factors and displays them in the larger context of their
relationship to each other and to the system as a whole.
After developing the hierarchy, the planners judged the relative im
portance of ali the elements. They quantified these judgments by assigning
them numbers from 1 to 9and sometimes they disagreed. On many
major issues where an impasse in the judgment of different people occurs,
careful assessment of the differences in the intensity with which these
people defend their preferences and opinions is necessary. Often words
alone or logical argument cannot express the subtleties of deeply felt differ
ences. But these differences can be measured by numbers, as we will see
later on. After debate and compromise, the planners determined priorities
for the elements of the hierarchy. Through a sequential process the judg
ments were synthesized and the desirability of each of the three alternative
plans was estimated mathematically. The plan with the highest numerical
value, and therefore priority (in this case, plan B), was the obvious best
choice.
Judgments on the relative importance of each element in the hierarchy
were made by people who were knowledgeable about the Brandywine
River Region and about problems of urbanization and environmental
quality. Yet even experts can make mistakes in setting up a hierarchy or
discriminating between pairs of elements to judge priorities. The AHP also
tests the consistency of judgments; too gre at a departure from the perfectly
consistent value indica tes a need to improve the judgments or to restruc
ture the hi erarchy. Suppose we take a doser look at the question of consis
tency.
The consistency is perfect if ali the judgments relate to each other in a
perfect way . lf y ou say thal you pre fer spring three times more to summer
and thal you prefer summer twice more to winter, then when you give the
judgment comparing your preference of spring to winter it should be 6 and
not anythi ng else. The greater your deviation from 6, the greater your
inconsistency. This observation applies to relations among ali the judg
ments given . We would have perfect consistency, then, if ali the relations
checked out As we will see, there is a rather simple way of
verifyi ng inconsistency and how much it deviates from perfect consis
tency. There is also a good way for interpreting what inconsistency means
in practical terms . When we are revising judgments, this method is useful
and necessa ry .
;; This a pproach to the Brandywine River problem illustrates the basic
(1 principles of the analytic hi erarchy process . Now let us take a doser look at
\i these principles .
!
PRINCIPLES OF ANAL YTIC THIN KING 17
PRINCIPLES OF ANALYTIC THINKING
In solving problems by explicit logical analysis, three principles can be
distinguished: the principle of constructing hierarchies, the principle of
establishing priorities, and the principle of logical consistency. As sug
gested in the Brandywine River example, these natural principles of analy
tic thought underlie the AHP.
Structuring Hierarchies
Humans have the ability to perceive things and ideas, to identify
them, and to communicate what they observe. For detailed knowledge our
minds structure complex reality into its constituent parts, and these in tum
into their parts, and so on hierarchically. The number of parts usually
ranges between five and nine. In the Brandywine River Region study, the
idea of environmental quality was structured into six elements: vividness,
intactness, no noise or disturbance, no flooding, water quality, and chan
nel naturalness. By breaking down reality into homogeneous dusters and
subdividing these clusters into smaller ones, we can integrale large
amounts of information into the structure of a problem and form a more
complete picture of the whole system. (This process is explored further in
the next chapter.)
Setting Priorities
Humans also have the ability to perceive relationships among the
things they observe, to compare pairs of similar things against certain
criteria, and to discriminate between both members of a pair by judging
the intensity of their preference for one over the other. Then they synthe
size their judgments through imagination or, with the AHP, through a
new logical process and gain a better understanding of the whole system.
In the Brandywine River Region study, the planners established rela
tionships between the elements of each leve! of the hierarchy by compar
ing the elements in pairs. These relationships represent the relative impact
of the elements of a given leve! on each element of the next higher leve!. ln
this context the latter element serves as a criterion and is called a pro pert y .
The result of this discrimination process is a uector of priority , or of relative
importance, of the elements with respect to each property. This pairwise
comparisoh is repeated for ail the elements in each leve!. The final step is to
come down the hierarchy by weighing each vector by the priority of its co
property. This synthesis results in a set of net priority weights for the
bottom leve!. The element with the h ighest weightplan B (part ial ur
[,
 . 
THE ANAL YTI C HI ERARCHY PROCESS
banization) in our example is the one that merits the most serious con
sideration for action, although the others are not ruled out entirely. This
principle and the next are fully explained in Chapter S.
Logical Consistency
The third principle of analytic thought is logical consistency. Humans
have the ability to establish relat ionships among objects or ideas in such a
way that they are coherentthal is, they relate well to each other and their
relations exhibit consistency. Consistency means two things. The first is
that similar ideas or objects are grouped according to homogeneity and
relevance. For example, a grape and amarble can be grouped into a homo
geneous set if roundness is the relevant criterion but not if flavor _is the
criterion. The second meaning of consistency is that the intensities of rela
tions among ideas or objects based on a particular criterion justify each
other in sorne logical way. Thus if sweetness is the criterion and honey is
judged to be five times sweeter than sugar, and sugar twice as sweet as
molasses, then honey should be taken to be ten times sweeter than molas
ses. If honey is judged to be only four times sweeter than molasses, then
the judgments are inconsistent and the process may have to be repeated if
more accurate judgments could be obtained .
ln utilizing these principles, the anal ytic hierarchy process incorpo
rates both the qualitative and the quantitative aspects of human thought :
the qualitative to define the problem and its hierarchy and the quantitative
to express judgments and preferences concisely. The process itself is de
signed to integrale these dual properties. lt clearly shows thal for better
decision making the quantitative is basic to making decisions in
complex situations where it is necessary to determine priorities and make
tradeoffs. To calculate priorities, we need a practical method of generating
scales for measurement.
MEASUREMENT
People are generally wary, if not distrustful, when numbers are intro
duced into the traditional process of decision making. But appropriately
chosen numbers can represent variat ions in feelings more faithfully than
can words or rhetoric. In the face of complexity, we run out of words to
express adequately our full awareness of what we sense to be taking place.
Words limit the perspectives of our feelings.
Numbers are used in many different ways in our civilization to mea
sure ali kinds of physi cal experience. We find this application acceptable.
The question is whether we can extend and justify the use of numbers in
sorne reasonable , easily underst ood way to reflect our feelings on various
social, economie, and political matters . We need to look into whether
MEASUREMENT 19
numbers are si mply artifacts that gi ve us the illusion of greater precision
than we are capable of feeling, or whether we are missing a great deal by
not realizing that numbers are a creation of our minds to reflect feelings
and distinctions. Perhaps we have not yet recognized and appreciated
their value in solving complex, unstructured problems.
In a moment we will briefly consider how numbers have come to be
used in our lives to measure our perceptions of physical stimul i. Later we
will see that numbers can also be used to reflect accurately our subj ective
judgments and their intensity; they can be used to distinguish among
intangible as well as physical stimuli. Chapter 5 describes a simple way to
use numbers to synthesize outcomes that faithfully represent our intuitive
feeling and understandi ng of what we perceive the outcomes to be. The
advantage is that finer shades of differences in judgment can be identified
for their effect on the outcome and that we can accommodate different
opinions in the decisionmaking framework .
Evolution of Scales
Our highly organized civili zation depends on scales to measure such
qualities as time, length. temperature, and money. Such measures were
not handed down through the burning bush but evolved hi storically.
Time. Time is a fundamental quality of nature; ils measurement is at the
foundation of science. The Sumerians were the first to divide the year and
the day into units . Their year contained twelve months; and each montlt,
thirty days. Egyptian priests divided the year into 365 days . The full sun
light period of a day was divided into ten hours, corresponding to ten
fi,ngers; dawn and dusk were each allotted one hour, bringing the total to
twelve. The night was also allotted twelve hours, making twentyfour
hours in a day. The daylight hours were marked by shadow docks, precur
sors of sundials, and the night hours were marked by the appearance of
stars. Thus neither daylight nor night hours were uniform in length; they
depended on the seasonal transit of the sun and the rising of the stars.
The real beginning of time measurement was the depsydra, a water
dock that measured time by the emptying and filling of a vessel. As the
depsydra became more corn mon, so too did the notion of lime as a thing in
itself, a flowing reality measured independently of the heavens . ln the
tenth century Arab scientists developed an improved sundial thal marked
off the hours accu ratel y year round and was the first use of fixed lime units.
In the thirteenth century mechanical docks measured lime by uniform 00
periodic motions . Europeans developed precision in meas uring lime dur ;......
ing the sev{mteenth and eighteenth centuries. In October 1960 the 0
Eleventh General Conference on Weights and Measures, in Paris,:
THE ANAL YTI C HIERARCIIY PROCESS
refined chronological precision by defining the second as the duration of
9,192,631,770 cycles of the radiation associa led wi th a specified transi lion,
or change in energy leve!, of the cesium atom.
Length. The Babylonians, Egyptians, Greeks, and Chinese ali had their
own un ils and subunits of length and other physical qualities such as area,
weight, and liquid volume. In the thirteenth century the English i n ~ r o
duced a standard yard and divided it into three feet of twelve inches each.
They also defined a rodas five and a half yards and a furlong as oneeighth
of a mile. In 1878 the yard was redefined as "the straight line or distance
betwe:en the centers of two gold plugs or pins in the bronze bar ...
measured when the bar is at the temperature of sixtytwo degrees of
Fahrenheit's thermometer, and when it is supported by bronze rollers
placed und er it in such a manner as best to a void flexure of the bar."
The French Revolution contributed the metric system. In 1792 Louis
XVI issued a proclamation directing two engineers to determine the length
of the meler. They set out to measure the distance on the meridian from
Barcelona, Spain, to Dunkirk, France, but civil war intervened and the task
took years to complete. A provisional meler was established in 1795 and
was adopted by the French Assembly four years later: The meler was "one
tenmillionth part of a meridional quadrant of the earth." Again in Oc
lober 1960 the meler was redefined as 1,650,753.73 wavelengths in a vac
uum of the orangered line of the spectrum of krypton86.
Temperature. The mercury thermometer was invented by the German
physicist Gabriel Daniel Fahrenheit in the eighteenth century. He consid
ered body temperature to be 100 on his scale (later found to be 98.6, as he
erred in determining his fixed points) and the temperature of the coldest
thing he could produce in his laboratory a mixture of salt and iceto be
0. The freezing and boiling points of pure water at sea leve! had to be 32
and 212, respectively. Although the thermometer has been improved,
Fahrenheit's scale survives unchanged. It is sometimes preferred to the
Celsius scale, devised by the Swedish astronomer Anders Celsius in 1742,
because it has smaller subdivisions. For highly accurate temperature read
ings, Lord Kelvin (William Thompson) first suggested the use of the gas
thermometer in the nineteenth century. AbS()lute zero is equal to 273
Celsius, the temperature at which ali atomic activity stops.
Money. Monetary standards differ from the other units of measure in
thal they are neither uniform nor consistent. The monetary value we attrib
u te to goods and services fluctuates according to supply and demand and
perce ived desirability or utilit y. The evolution of money is closely inter
woven with thal of ci vili zation . In sorne cases, as in ancient Egypt, the
preva iling monetary sys tem evolved from the political, social, and eco
MEASUREMENT 21
nomic systems, whereas in ancien! Lydia these institutions followed the
progress of monetary evolution. Although goods and services can be ex
changed directly, the adoption of a monetary system simplifies trade and
promotes higher productivity.
The Need for a New Scaie
Just as we can distinguish and measure physical relationships
meters for length, for example, and seconds for timewe are capabl e of
doing the same with abstract relationships. We have the capacity for a
range of feeling and discrimination thal permits us to develop relat ion
ships among the elements of a problem and to determine which element s
have the greatest impact on the desired solution. In dealing with concret e
matters, such as repairing a car, we perceive the varying intensity of im
pact through our senses, by hearing a faulty motor or seeing a leak, or
through their refinement by scientific instruments such as a voltmeter or
pressure gauge. We carry out this process of measuring the priorities of
impacts in order to solve problems.
Soto determine the intensity of impact of the various components of a
system, we must perform sorne type of measurement on a scale with units
such as pounds, seconds, miles, . and dollars . But these sc ales li mit the
nature of ideas we can deal with. Social, political , and other qualitative
factors can in no reasonable way be assessed in terms of physical or eco
nomie measurement . What then can we do?
We can devise a scale thal enables us to measure intangible qualities,
just as scales evolved for measuring physical qualities . Chapter 5 presents
such a scale to measure priority impacts in unstructured systems . This new
way of assessing intangible qualities should hold up in areas where we
already know the unit of measurement, which can then be used to validate
the method. And in fact examples show that this approach to measuring
priorities can be used to generale results conforming to classic ratio scale
measurement in physics, economies, and other fields where standard mea
sures already exist.
To measure priorities, we compare one element with another . The old
adage thal one cannot compare apples and oranges is false . Apples and
oranges have many properties in common: size, shape, !aste, aroma, color,
seediness, juiciness, and so on. We may prefer an orange for sorne proper
ties and an apple for others; moreover, the strength of our preference may
vary. We may be indifferent to size and color, but have a strong preference
for !aste,. which again may change with the time of day. It is my thesis thal
this sort of complicated comparison occurs in real !ife over and over again,
and sorne kind of mathematical approach is required to help us determine
priorities and make tradeoffs. This approach is the analytic hierarchy
process:
(X)
........
........
j
j
l
,;
l'
.
THE ANi\l.YTIC IIIFRARCIIY PROCESS
AHP: A FLEXIBLE MODEL FOR DECISION MAKING
These basic observations on human nature, analytic thinking, and
measurement have led to the development of a useful mode! for solving
problems quantitatively. The analytic hierarchy process is a flexible mode!
that allows individuals or groups to shape ideas and define problems by
making their own assumptions and deriving the desired solution from
them. lt also enables people to test the sensitivity of the solution, or out
come, to changes in information. Designed to accommodate our human
nature rather than force us into a mode of thinking that may violate our
better judgment, the AHP is a powerful process for tackling complex poli ti
cal and socioeconomic problems.
The AHP incorpora tes judgments and persona! values in a logical way.
lt depends on imagination, experience, and knowledge to structure the
hierarchy of a problem and on logic, intuition., and experience to provide
judgments. Once accepted and followed, the AHP shows us how to con
nect elements of one part of the problem with those of another to obtain the
combined outcome. lt is a process for identifying, understanding, and
assessing the interactions of a system as a whole.
To define a complex problem and to develop sound judgments, the
AHP must be progressively repeated, or iterated, over time; one can hardly
expect instant solutions to complicated problems with which one has wres
tled for a long time. The AHP is flexible enough to allow revision
decision makers can both expand the elements of a problem hierarchy and
change their judgments. lt also permits them to investigate the sensitivity
of the outcome to whatever kinds of change may be anticipated. Each
iteration of the AHP is like hypothesis making and testing; the progressive
refinement of hypotheses leads to a better understanding of the system.
The many practical applications of the AHP have generated sample hierar
chies, sorne of which are presented in Chapter 4. With mi nor modification,
sorne of thse paradigms can be used to structure new problems.
Another feature of the AHP is that it provides a framework for group
participation in decision making or problem solving. We have seen that
ideas and judgments can be questioned and strengthened or weakened by
evidence that other people present. The way to shape unstructured reality
is through participation, bargaining, and compromise. Indeed, the con
ceptualization of any problem by the analytic hierarchy process requires
one to consider ideas, judgments, and facts accepted by others as essential
aspects of the problem. Group participation can contribute to the overall
validity of the outcome, although perhaps not t the ease of implementa
tion if the views diverge widely. Th us one could include in the process any
information derived scientifically or intuitively.
The process can be applied to real problems and is particularly useful
for  " ""Ca ting resources, planning, analyzing the impact of policy, and
AHP: A FLEXIBLE MODEL FOR DECISION MAKING
Unlty:
The AHP provides a single.
easily understood, flexible mode!
for a wide range of unstructured
problems
Process Repetition:
The AHP enables people to re fine
their definition of a problem
and to improve judgment
and understanding through
repetition
Judgment and Consensus:
Complexlty:
The AHP integrales deductive
and systems approaches
in solving complex
problems
23
on consensus but synthe The AHP can deal w1th the
The AHP does not insist \ 1 Interdependance:
sizes a representative 1nterdependence of elements
outcome from d1verse ......______ 1n a system and does not
judgments e  tnSISt on hnear th1nk1ng
AHP
Tradeoffs: \....___________ Hierarchie Structurlng:
The AHP takes into The AHP rellects the
consideration the natural tendency of the
relative priorities of mind to sort elements of
factors in a system and a system into different
enables people to select levels and to group like
the best alternative elements in each ievel
based on the ir goals
Synthesls: Measurement:
The AHP leads to
an overall estima te of the
desirability of each
alternative
The AHP provides a scale
for measuring intangibles
and a method lor
establishing priorities
Conslstency:
The AHP tracks the logical
consistency of judgments used
in determining priorities
Figure 22 Advantages of the Analytic Hierarchy Process
j
1
1
'
j
!
.j
:1
1
i
:0
J
.; ,
co i:
. :
t" ,.
N
1
1
 .. ....         .. .    .
1 ANALYTIC lllERARCHY PROCESS
resolving conflicts. Social and physical scientists, engineers, policymakers,
and eyen laypersons can use the method without intervention by socalled
experts; those who have a problem are ordinarily best informed about that
particular problem. Currently the AHP is being widely used in corporate
planning, portfolio selection, and benefit/cost analysis by government
agencies for resource allocation purposes . And it is being used more
widely on an international scale for planning infrastructure in developing
countries and for evalating natural resources for investment. Figure 22
summarizes the advantages of using the AHP as a new approach to prob
lem solving and decision making.
PERSPECTIVE
The analytic hierarchy process reflects the way we naturally behave
and think. But it improves upon nature by accelerating our thought pro
cesses and broadening our consciousness to include more factors than we
would ordinarily consider. The AHP is a process of "systemic rationality":
lt enables us to consider a problem as a whole and to study the simulta
neous interaction of its components within a hierarchy. This process is
rather sophisticated. If we observed an individual, such as a child, who
does not have to deal with complex issues or do severa! tasks at the same
time, we most likely would not note a talent for hierarchie organization.
Over long periods of time, however, the mind learns shortcuts and groups
activities into clusters. This mental process gradually develops into a style
for looking at the world and organizing it in such a way that we can deal
with it efficiently.
Sorne people have gone so far as to point out that nature itself or
ganizes matter and !ife hierarchically. But it is difficult to separate this
observation from the fact that to understand the complexity of nature, it is
we who have to sort and arrange what we perceive and see hierarchically.
Moreover, the interactions among elements of a hierarchy do not always
have an inherent structure apart from what we can observe. We must
identify and synthesize these interactions based on our objectives aild the
knowledge and experience we have of each problem.
The AHP addresses complex problems on their own terms of interac
tion . It allows people to lay out a problem as they see it in ils complexity
and to refine ils definition and structure through iteration. To identify
criti cal problems, to define their structure, and to locale and resolve con
fli cts, the AHP calls for informati on and judgments from severa! partici 
pants in the process. Through a mathematical sequence it synthesizes their
judgments into an overall est imate of the relative priorit ies of alternative
courses of acti on. The pri orities yielded by the AHP are the basic units
used i n ali types of analysis; for example, they can serve as guidelines for
all ocati ng resources or as probabilities in making predictions.
KEY CONCEPTS 25
The AHP can be used to stimulate ideas for creative courses of action
and to evaluate their effectiveness. lt helps leaders determine what infor
mation is worth acquiring to evaluate the impact of relevant factors in
complex situations. And it tracks inconsistencies in the participants'
judgments and preferences, thereby enabling leaders to assess the quality
of their assistants' knowledge and the stability of the solution .
The next three chapters on this new approach to decision
making. Here is what one can expect to gain by using it:
1. A practical way to deal quantitatively with different kinds of functi onal
relations in a complex network.
2. A powerful tool for integrating forward (projected) and backward (de
sired) planning in an interactive manner that reflects the judgment s of
ali relevant managerial personnel. The output of this process is explicit
rules for allocating resources among current and new st rategy
offeringsor to satisfy a specifie set of corporate objectivesor under
alternative environmental scenarios .
3. A new way to:
Integrale hard data with subjective judgments about intangible fac
tors.
lncorporate judgments of severa! people and resolve conflicts among
them.
Perform sensitivity analysis and revision at low cost .
Use marginal as weil as average pri orities to guide allocation.
Enhance the capacity of management to make tradeoffs explicitly.
4. A technique complementing other on es (benefitlcost, priority, risk
minimization) for selecting projects or activities.
5. A single replacement for a variety of schemes for projecting the future
and protecting against risk and uncertainty.
6. A vehicle for monitoring and guiding organizational performance ta 
ward a dynamic set of goals.
KEY CONCEPTS
0 There are three basic prin ci pies of the analytic hierarchy process:
1. Hierarchie representation and decomposition, whi ch we cali h ierar
chie structuringthat is, breaking clown the problem int o separa te
elements.
2. Priority discrimination and synthesis, which we cali pri orit y
settingthat is, ranking the element s by relative importance.
(X)
1'"
w
,;
~ !
 .  ....
TIIE ANAL YTIC HIERARCHY PROCESS
3. Logical consistencythat is, ensuring that elements are grouped
logically and ranked consistently according to a logical criterion.
0 We cannot measure without a scale, but traditional scales such as ti me
and money limit the nature of ideas we can deal with. Thus we need a
new scale for measuring intangible qualities.
0 The analytic hierarchy process is a flexible mode! that allows us to
make decisions by combining judgment and persona! values in a logi
ca! way.
3
Analyzing and
Structuring
Hierarchies
This chapter deals with the following questions :
Why are hierarchies fundamental to human
thinking?
How can they be used to understand complex
systems?
What is the difference between a structural
hierarchy and a functional hierarchy?
How do we go about constructing a
hierarchy?
77
_j
co
!"
.p..
ANALYZING AND STRUCfURING IIIERARCHIES
HIERARCHIES: A TOOL OF THE MIND
Complex systems can best be understood by breaking them down into
the ir constituent elements, structuring the elements hierarchically, and
then composing, or synthesizing, judgments on the relative importance of
the elements at each leve! of the hierarchy into a set of overall priorities.
This chapter explains how to structure problems hierarchically.
Hierarchies are a fundamental tool of the human mind. They involve
identifying the elements of a problem, grouping the elements into homo
geneous sets, and arranging these sets in different levels . The simples!
hierarchies are linear, rising or descending from one levet to another, such
as physics hierarchies that rise from atoms to molecules to compounds and
so on; the most complex are networks with interacting elements, such as
systems representing the learning process of a child.
CLASSIFYING HIERARCHIES
Hierarchies can be divided into two kinds: structural and functional.
ln structural hierarchies, complex systems are structured into their con
stituent parts in descending order according to structural properties such
as size, shape, color, or age. A structural hierarchy of"the universe would
descend from galaxies to constellations to solar systems to planets, and so
on, down to atoms, nucleii, protons, and neutrons . Structural hierarchies
relate closely to the way our brains analyze complexity by breaking down
the abject s perceived by our senses into clusters, subclusters, and still
smaller clusters.
ln contras!, functional hierarchies decompose complex systems int o
their constituent parts according to their essential relationships. A conf!ict
over school busing to achieve integration can be structured into a cluster of
major st akeholders (majority and minority communities, city officiais,
board of education , federal government) ; a cluster of stakeholders' objec
tives (education for children, retention of power, and the like) ; and alter
native outcomes (complete, partial, or no busing) . Such functional hierar
chies help people to steer a system toward a desired goal like conflict
resolution, efficient performance, or overall happiness. For the purposes of
this book, functional hierarchies are the only ki nd thal need be considered.
Each set of elements in a functional hierarchy occupies a leve! of the
hierarchy. The top leve! , called the foCIIs , consists of only one element : the
broad, overall objective . Subsequent levels may each have severa! ele
ments, although thei r number is usually small between five and nine.
Beca use the element s in one leve! are to be compared with one another
against a crit er ion in the next higher leve!. the elements in each leve! must
be of the same order of magnitude. If the d isparity between them is great,
they should belong to different levels. For example, we cannat make a
~
CONSTRUCflNG HIERARCHIES 29
precise comparison between two jobs whose performances differ in diffi
culty by a factor of 100 because our judgment would be subject to signifi
cant error. lnstead, we first group simple jobs into a cluster and compare the
cluster with a job one arder of magnitude more difficult to perform than a
simple job. We then compare the jobs in the cluster among themselves
according to difficulty of performance. When we compare the results of the
two comparison processes, we obtain a net comparison of a simple job
with the more difficult one. To avoid making large errors, we must carry
out this process of clustering. By forming hierarchically arranged cl usters
of like elementssimple jobs in this casewe can efficiently compl e te the
process of comparing the simple with the very complex. Similarl y, to com
pare small stones with boulders or atoms with stars, we must int etvene
between them severa! levels of abjects of slightly differe(1t magnitude to
make the transition and comparison possible.
Because a hierarchy represents a mode! of how the brain analyzes
complexity, the hierarchy must be flexible enough to deal with that com
plexity. The levels of a hierarchy interconnect like layers of cell tissue to
form an organic whole that setves a certain function . A spiraling effect is
noticeable when we move from the focus expanding the hierarchy to the
leve! of simple elements. The expansion may be continued to the leve! of
elements of minutest concern. The next chapter illust rates the flexibility of
hierarchies with a variety of examples. Ali are functional hierarchies, but
sorne are completethat is, ali the elements in one leve! share every prop
erty in the next higher level and SOrne are i11 C0 111 p /ete in thal Sorne ele
ments in a leve! do not share properties.
CONSTRUCTING HIERARCHIES
No inviolable rule exists for constructing hi erarchies . The sample
hierarchies offered throughout the book are presented not to prescribe
certain frameworks but to stimulate thinking about what types of hierar
chicallevels to choose and what kinds of elements to include in the levels.
The numbers of levels and elements may be more or less than those in the
examples.
The great variety of examples thal !end themselves to hierarchies sug
gests thal the Sl.\,bjects thal can be approached with the an.,Jytic hierarchy
process are infinlt&. ln ali these areas we are limited only by our experi 
ences and feelings as represented by words from the di cti onary. Lan
guages that are limited in vocabulary may pose problems of ambiguit y or
may no.t represent human experience adequately. Recogni zing these lim
itations, we may be encouraged to innovate by creating the needed vocab
ulary .and other symbols, such as computer languages, to represe nt feelings
and ideas that we may become aware of i n the process of identificati on and
structuring.
co
~
l1l
ANALYZING AND STRUCTURING HI ERARCHIES
One' s approach to constructing a hierarchy depends on the kind of
decision to be made. If it is a matter of choosing among alternatives, we
could start from the bottom leve! by listing the alternatives. The next leve!
would consist of the criteria for judging the alternatives. And the top leve!
would be a single element, the focus or overall purpose, in terms of which
the criteria can be compared according to the importance of their contri
bution.
Suppose we wanted to decide whether to buy one of five sports cars
(Figure 31) . These alternatives form the bottom leve! of the hierarchy. The.
criteria in terms of which the alternatives will be judged form another leve!
and might include adequacy of salary, prestige, basic necessities, comfort,
satisfaction of other needs, large savings account, and freedom from
worry. The priorities of these criteria will be judged in terms of their
contribution to the focus of the hierarchy: our overall happiness.
Note that once we construct this hierarchy, it is not necessarily cast in
bronze. We can always alter parts of it later to accommodate new criteria
that we may think of or that we did not consider to be important when we
first designed it . The computer programs that assis! us in the task are
constructed with this flexibility in mind. After we rank the criteria and
arrive at overall priorities for the alternatives, we might still have sorne
doubts about the final decision. In that case we would simpl y go through
the process again and perhaps change sorne of our judgments on the rela
tive importance of the criteria. If the same alternative is still significantly
ahead of the others in overall priority, we know that it is the right choice
for us .
Sometimes the criteria the mselves must be examined in detail, so a
leve! of subcriteria should be inserted between those of the criteria and the
alternatives. To select a school from three possible choices, for example,
severa! criteria might be usedsuch as educational , cultural, and social
Sa lary
Prestige
Good Choice for a Sports Car
Basic
Needs
Comfort
Satisfaction
ofOther
Needs
Large
Savings
Account
Freedom
from
Worry
Porsche Mercedes Triumph Dats un Corvette
Figure 31 Hierarchy for Choosing a Sports Car
CONSTRUCTING HIERARCHIES
Educational
Quality
T eachers \ \ "
General
of Students \ '\. ...__
Preparation \ ",
College \,
Environment '.,_
School
A
Selecting a Best School
Cultural
!\
School
B
School
c
Figure 32 Hierarchy for Selecting a School
31
Social
J
(Figure 32) . The educational criterion might be broken down into subcri 
teria of (1) quality of teachers, (2) general standard of students, (3) disci 
pline, (4) preparation for college, and (5) learning environment ; the other
criteria could similarly be broken down as weil. In this case the subcri teria
would be .compared in terms of the criterion to which they belonged and
not to the other criteria. Such a hierarchy would then be called incomplete
because the subcriteria are not ali compared in terms of ali the criteria of
the next higher leve!.
There is no limit to the number of levels in a hierarchy. If one; ., unable
<Xl
t<
0'1
ANAL YZING AND STRUCTURING HIERARCHIES
to compare the elements of a leve! in terms of the elements of the next
higher leve!, one must ask in what terms they can be compared and then
seek an intermediate leve! that should amount to a breakdown of the
elements of the next higher leve!. Th us a new leve! has been introduced to
facilitate the analysis for comparisons and to increase the precision of the
judgments. Now we can answer the main question: How much more does
one element contribute than another to satisfying a criterion in the next
higher leve! of the hierarchy?
Hierarchies may be more complicated than the ones described above.
Those related to projected planning and repeated later, for example, in
elude the following levels:
Uncontrollable environmental constraints
Risk scenarios
Controllable systemic constraints
Overall objectives of the system
Stakeholders
Stakeholders' objectives (separate ones for each stakeholder)
Stakeholders' policies (separate ones for each stakeholder)
Exploratory scenarios (outcomes)
Composite or logical scenario (outcome)
But decision makers do not have to pursue a problem to the full leve) of
detail indicated here. The depth of detail depends on how much knowl
edge one has about the problem and how much can be gained by using
that knowledge without unnecessarily tiring the mind.
PERSPECTIVE
Although hierarchies have been known for a very long ti me, the anal y
tic hierarchy process makes it possible to generale new levels and arrange
them in a logical fashion so they relate to each other naturally. By making
paired comparisons of the elements in a level in terms of the elements of
the next higher level, it is possible to decide on an appropriate choice of
thal upper level. Moreover, when the elements of a level cannot be com
pared except in terms of finer criteria than identified so far, a new leve!
must be created for this purpose. Thus the analytic aspects of the AHP
serve as a stimulus to create new dimensions for the hierarchy. It is a
process for inducing cognitive awareness . A logically constructed hierar
chy is a by product of the entire AHP approach. lndeed, experience indi
cates that there are a few patterns that ali decision hierarchies seem to
follow. ln the next chapter we will see how this concept of hierarchies can
be appli ed to a broad range of real situations.
KEY CONCEPTS
33
KEY CONCEPTS
0 In a functional hierarchy, complex systems are broken down into their
constituent parts according to their essential relationships.
0 The top level of the hierarchythe focusconsists of only one ele
ment: the overall objective. The other levels contain severa! elements
(usually between five and nine) .
0 There is no limit to the number of levels in a hierarchy.
0 When the elements of a level cannot be compared readily, a new level
with finer distinctions must be created.
0 Hierarchies are flexible. We can always alter them to accommoda te
new criteria.
HOW TO STRUCTURE A HIERARCHY
The basic principle to follow in structuring a
hierarchy is to see if one can answer the question: "Can
you compare the elements in a lower leve! in terms of sorne
or all the elements in the next higher level?"
A useful way to proceed is to come down from the goal
as far as one can and then go up from the alternatives
until the levels of the two processes are linked in a way
to make comparison possible. Here are sorne suggestions.
1. Identify overall goal. What are you trying to
accomplish? What is the main question?
2. Identify subgoals of overall goal. If relevant,
i d e n ~ i f y time horizons that affect the decision.
3. Identify criteria that must be satisfied to fulfill
subgoals of the overall goal.
4. Identify subcriteria under each criterion. Note t hat
subcriteria may be intervals of numerical values; or
intensities such as high, medium, low; or excellent,
very good, good, average, poor, and very poor.
5. Identify in descending levels, as needed, actors, actor
objectives, and actor policies in this order.
6. Identify alternatives or outcomes.
7. For yesno decisions include for example doing and not
doing the alternative.
8. It is often useful to construct two hierarchies, one
for benefits and one for costs to decide on the best
alternative, particularly in the case of yesno
decisions. Ratios of benefits to costs, or better,
marginal benefits to costs, are formed and the
alternative with the largest ratio is chosen. Answer
the question: Which alternative yields the greatest ~
benefit (for benefits) or costs the most (for costs)7 ~
'J
1
~
r
t
!
1
l__
.  .  ..   .....  
4
Practical
Examples of
Hierarchies
This chapter deals with the following questions :
What is the main purpose of arranging goals,
attributes, issues, and stakeholders in a
hierarchy?
How can hierarchies be applied to business
decisionschoosing equipment, deciding
whether to buy or lease, making financial
decisions, and so on?
How can hierarchies be applied to persona!
and domestic decisionschoosing a car,
choosing a career, purchasing a house, and so
on?
How can hierarchies be applied to public
policy decisionschoosing areas for R&D
programs, allocating resources in a juvenile
34
J
AN APPROACH TO HIERARCHIES
correction program, analyzing a school busing
conflict, and so on?
How can hierarchies be applied to planning
economie policiesfor example, planning
economie strategy for an underdeveloped
country?
How can hierarchies be used for estimating
and predictingpredicting the presidential
election, estimating the popularity of a rock
group, and so on?
How can hierarchies be used for measuring
influences for example, parental influence on a
child's psychological we!lbeing?
How can hierarchies be used to represent
systems networksthe network of a child' s
leaming system, the network of a volleyball
team, and so on?
AN APPROACH TO HIERARCHIES
35
ln later chapters we will discuss in detail specifie applications of the
AHP to planning, conflict resolution, benefit/cost analysis, and resource
allocation. This chapter offers examples of different kinds of hierarchies to
suggest ways of approaching problems and structuring them into their
constituent parts with enough detail to make reasonable decisions .
Most prob1ems arise because we do not know the internai dynamics of
a system in enough detail to identify causeandeffect relationships. If we
were able to do so, the problem could be reduced to OJ1e of soci al engineer 
ing, as we would know at what points in the system interventio n is neces
sary to bring about the desired objective. The crucial contribution of the
analytic hierarchy process is that it enables us to make practical decisions
based on a "precausal" understandingnamely, on our feelings and
judgments about the relative impact of one variable on another.
ln sum, when constructing hierarchies one must include enough rele
vant detail to depict the problem as thoroughly as possible . Consider thE
environment surrounding the problem. Jdentify the issues <' ribut e!
i .
~
i'i
l i ~
; t
1
1
[
f
1
i
~
' 1
!
l
\
~
l
;
il
:1
i ~
"
1
1
!j
" :1
i\
\
l
..
!,
.,,
l'
~
,
1
)
l
t
1
(X)
,_...
(X)
1
r
H
::
i=
li
; ,
; ,
.,
,,
,,
l '
i
. ... . ......                   ~
PRACTICAL EXAMI'LES OF IIIERARCIIIES
that you feel contribute to the solution. ldentify the participants associated
with the problem. Arranging the goals, attributes, issues, and stakeholders
in a hierarchy serves two purposes: lt provides an overall view of the
complex relationships inherent in the situation, and it permits the decision
maker to assess whether he or she is comparing issues of the same arder of
magnitude in weight or impact on the solution.
The elements should be clustered into homogeneous groups of five to
nine so they can be meaningfully compared to elements in the next higher
levet. The only restriction on the hierarchie arrangement of elements is that
any element in one level must be capable of being related to sorne elements
in the next higher leve!, which serves as a criterion for assessing the rela
tive impact of elements in the level below.
The hierarchy does not need to be complete; thal is, an element in a
given level does not have to function as a criterion for al/ the elements in
the leve! below. Thus a hierarchy can be divided into subhierarchies shar
ing only a corn mon topmost element . Further, a decision maker can insert
or eliminate levels and elements as necessary to clarify the task of setting
priorities or to sharpen the focus on one or more parts of the system.
Elements that are of Jess immediate interest can he represented in general
terms at the higher levels of the hierarchy and elements critical to the
problem at hand can be developed in grea ter depth and specificity.
ln addition to identifying within a hierarchie structure the major fac
tors that influence the outcome of a decision, we need a way to decide
whether these factors have equal effects on the outcome or whether sorne of
them are dominant and others so insignificant they can be ignored. This .is
accomplished through the process of priority setting. The task of setting
priorities requires that the criteria, the subcriteria, the properties or fea
tures of the alternatives being compared, and the alternatives themselves
are gradually layered in the hierarchy so. thal the elements in each level are
comparable among themselves in relation to the elements of the next
higher leve!. Now the priori lies are set for the elements in each level
severa! times<Jnce with respect to each criterion of the upper level. These
in tum are prioritized with respect to the elements of the next higher level
and so on. Finally a weighting process is used to obtain overall priorities.
This is done by coming down the hierarchy and weighting the priorities
measured in a level with respect to a criterion in the next higher Ievel with
the weight of that criterion. The weighted priorities can then be added for
each element in the leve! to obtain ils overall priority.
Finally, a ft er judgments have been made on the impact of ali the
elements, and priorities have been computed for the hierarchy as a whole,
the Jess important elements can be dropped from further consideration
because of the ir relatively small impact on the overall objective. Now let us
exa mine sorne of the realworld applications of the hierarchy.
:"!
BUSINESS DECISIONS 37
BUSINESS DECISIONS
Hierarchy for Choosing Urethane Equipment
To decide on the purchase of urethane manufacturing equipment
(Figure 41), three principal considerations are priee, technical features,
and service. "Technical features" is really a cluster of severa! specifie de
sired qualities to which it is further decomposed. The overall priorities
obtained through successive priority setting represent the relative impor
tance of the considerations bearing upon the purchase decision.
The three products under consideration are then ranked with respect
to each criterion or subcriterion. The overall priority gives the relative
superiority of the brands.
Hierarchy for Choosing Word Processing Equipment
For the selection of word processing equipment for an office (Fi gure
42), the benefit and cost hierarchies are considered separately and
benefit/cost ratios are obtained. The qualities desired of a word processing
machine form the second leve! in the benefit hierarchy, and thei r pri or it ics
represent the relative weights assigned by the user. The eq ui pment chara
Level1 :
Focus
Level2:
Criteria
Priee
Level3:
Subcriteria
Leve14:
Alternatives
Purchase of Urethane Machinery
Technic_al Features
~
Ratio
Control
Recirculat ion
Control
Temperature
Control
~ ~ ~
Brand X Brand Y Brand Z
Figure 41 Hierarchy for Choosing Urethane Equipment
Servce
1
~
rt.;
}
" )
. I
1
;1
t
{
t
' l
1
!
1
' 1
'
li
! i
,.
1
1
(
1
1
1
1
1
~ i l
~ I l
1
!
. ..  ....... ...
Level1 :
Foc us
Level2:
Criteria
Level3:
Features
Level4:
Alternatives
Level1 :
Focus
Level2:
Criteria
Level3:
Alternatives
PRACTICAL EXAMPLES OF HIERARCHIES
Ti me
Saving
Filing
Benefits
Ouality of
Document
Accuracy
Training Screen Service Space
Required Capability Ouality Required
Printer
Speed
l'
Lanier Syntrex Qyx
Costs
Capital Supplies Service Training
Lanier Syntrex Oyx
Figure 42 Hierarchies for Choosing Word Processing Equipment
teristics that contribute to these qualities fall in the third level. They are
ranked with respect to each de si red quality, and the overall priority repre
sents the relative importance ofeach characteristic. The overall priorities of
the brands under consideration represent the relative superiority regard
ing benefits expected from each.
In the cost hierarchy, likewise, the overall priorities represent the rela
tive weights of the costs . The benefit/cost ratio for each indicates the rela
tive superiority of one machine over the others.
"
BUSINESS DECISIONS 39
Hierarchy for Deciding on Buying or Leasing
The decision regarding company ownership or leasing of a piece of
capital equipment (Figure 43) depends on its contribution to the com
pany's profitability. This profitability has two dimensions : economie and
intangible. The benefits depend on a number of factors that, in turn,
depend on certain characteristics of the company. Buying or leasing would
promote characteristics to a varying extent .
By setting priorities for the factors at a certain leve! with respect to the
relevant factors at the previous leve! and finding the composite priorities,
we can find to what extent, relatively speaking, the factors in the same
leve! contribute to the firm's overall profitability. Extending this logic to
the question of company ownership or leasing, we can say, in the judg
ment of the decision maker, which alternative is preferable.
In this example we take the intangible benefits explicitly into consid
eration for a decision , so the subjective judgments of the decision maker
Level1 :
Focus
Level2:
Bene lits
Level3:
Factors
Level4:
Company
Status
Level5:
Alternatives
Profitabrlity
EHicient
Allocation
ofFunds
Economie Benefits
lnerease lnerease
Current Total
and Quick Re sources
Ratios ol Company
Intangible Benefits
EHicient
Allocation
of Personnel
and Time
Economie
Security
lmprove
Company
Image
Financial
Stalus of
Company
Availability
of Working
Capital
r
Company Ownership
Centralizalion
or Tax Advantagas
Decentralrzatron
Leasing
Figure 43 Hierarchy for Deciding on Buying or Leasing
co
N
0
Level1:
Focus
Level2:
Education
and Skill
Level3:
Abilities
Leve14:
Alternatives
PRACTICAL EXAMI'LES OF IIIERARCIIIES
Education
Select Manager
Management
Skills
Technical
Skills
Persona!
Skills
L . Probl em Job Organize Respon Decision J d
1
eadershp Solving Knowledge and Plan sibilily Making u gmen
Candidate 1 Candidate 2 Candodale 3
Figure 44 Hierarchy for Choosing a Management Candidate
are also considered. This is unlike a conventional exercise where only the
hard economie data are considered and then managerial judgment is used
in a qualifying manner at the end.
Hierarchy for Choosing a Candidate for Management
To select the proper person to fill a management position, we first
identify the four areas of evaluation and then the specifie traits to which
these areas contribute (Figure 44). Successive stages of priority setting
give the overall priorities for the relevant traits in leve) 3, which represent
the relative weights believed to be associated with them for a person to
function effectively in that position.
Once ali the candidates have been reviewed and the choice has been
narrowed clown to a few seemingly equal candidates, they may be ranked
with respect to each trait . The composite priority for each candidate then
represents his or her relative superiority on the basis of overall judgment
and is useful in comparing candidates and ranking them in order of pref
erence .
Hierarchy for Choosing a Beverage Container
To evaluate the desirability of different containers to be used by the
soft drink beverage industry (Figure 45), we first consider the criteria for
0) .
;:
BUSINESS DECISIONS 41
evaluation and then rank them according to their relative importance on
the final outcome. Next we judge the alternative containers with respect to
each criterion per unit of beverage delivered. This prioritization shows the
desirability from the point of view of each criterion, and the composite
priorities show the overall superiority of the containers in relative terms.
Hierarchy for Staggering lndustry Hours for Energy
Conservation
To decide how to stagger industry work hours, we firs t examine the
relevant consequences of staggered work hours and judge ho w importan t
they are with respect to one another. This is clone by ranking the re levan t
criteria with respect to the focus (Figure 46).
Next the various shift patterns under consideration are ranked with
respect to each of the criteria above to see how much they wou id affect it in
relative terms. The composite priorit ies indicate the overall desirability of
the considerations in relative terms. The highestpriority shift is the most
desirable decision .
Hierarchy for Choosing a Site for Combustion Turbines
The problem of site selection for an electric utility company' was nar
rowed to four alternatives after preliminary screening (Figure 47). The
company identified nine relevant factors they would have to consider in
the selection process. Since many of the considerations were in conflict,
they were first ranked to find their relative importance bearing upon the
Level1 :
Focus
Level2:
Criteria
Level3:
Alternatives
Energy to
Produce
Best Bever age Container
Cosl
Environmental
Wasl e
Customer
Service
Glass Bimelall ic Aluminum Steel
Figure 4 5 Hierarchy for Choosing a Beverage Container
(X)
N
1
1
1
1
  
.. .
;
' 1
;
'
i
i
;
Level1 :
Focus
Level2:
Criteria
Level3:
Alternatives
PRACTICAL EXAMPLES OF HIERARCHIES
Work Staggering Goats
Energy Economie Employment Transportation Social
Conservation Consequences Consequences Consequences Consequences
Shiftof Shiftot
Shoft of Shiftof 80% 20%
Historical
2 Hours 8Hours Production Production
Pattern
Earlier Earlier Shifted Shifted
Earlier Earlier
Figure 46 Hierarchy for Staggering Work Hours
!:: decision. Next the sites were ranked with respect to each factor and com
posite priorities were found . These overall priorities indicate the relative
merits of selecting the sites in question when ali the factors are taken into
consideration.
i
!.
:J:,
':\;1
{
'i
i
;l "
: ;:1
. ';
Hierarchy for Allocating Resources Among R&D Projects
in a Bank
This decision regarding resource allocation is done as a benefit/cost
exercise involving the benefits expected to accrue from the projects and the
costs expected to be increased thereby (Figure 48) .
In the benefits hierarchy, the benefits are ranked according to their
impact on the bank' s performance. The projects are ranked according to
how far they can generale that benefit. The composite priorities represent
their overall benefit contributions on a ratio scale. In the costs hierarchy,
Iikewise, the various costs are ranked by their severity and the projects are
ranked with respect to their contribution to that cos!. The resulting com
posite priorities represent their overall costs. The benefit/cost ratios mea
sure the superiority of benefit for cost incurred and also each R&D project's
expected attractiveness . Ratios comparing the greatest marginal benefits to
costs are often more useful than simple benefitlcost ratios. Sometimes dis
counting of benefits and costs is more realistically done before forming
such
r:
BUSINESS DECISIONS
43
ii
ill VI
E a
c
e
su..
c
w
<ii
C VI
0
.;:; 0
iiiu..
...
E !'!
i
u
(
0 0
c 13
8 ..
w"
ill
t
u ill
Vi
E
g
<Il
M
c::
ill E :
(( 0 !'!
u i
..
ill .
i'i
'"VI
...
J!l
. 0 tiC
.s
0 ..
(ii"
1/)
0
0
.c:
c
u
0 VI VI
g
..
"'
0
.....
VI VI !'!
c co
>.
i .c:
...
8
..
<Il
:
:>.0
a. c t
0 iii
O. a
'
c
0
t

::1
.;: r./) !'!
tiC
'l!
i iZ
r=
t=
> ,
e
o
>
I
,
VI
Cl>
co
VI N
. .
<Il;:) o
(")<tl
N
g
<Il Q;E
> 0
N
...JU.
<Il <tl
><Il
...JU.
Q)=
...J<(

:!
1
l
d
H
!1
:1
li
,.
1
1
I
l '
,,
,, :
i
. ' 1
t' ;
.
l l
.  ....
Level1 :
Focus
Level2:
Benefits
Level3:
Alternatives
Level1 :
Focus
Level2 :
Costs
Level3:
Alternatives
PRACTICAL EXAMI'LES OF IIIERARCIIIES
Cost
Savings
lncome
Generation
Benefils
Customer
Satisfaction
Legal
Aequiremenls
Corp.
DDA/Aecon.
Personnel
Costs
Cuslomer
Analysis
Hardware
Cosls
Costs
EFTS
Training
Cosls
"......_
Data
Entry
lnquiry
Operating
Costs
Cor p.
DDA/Aecon.
Customer
Analysis
EFTS
Data
Entry
lnquiry
Figure 48 Hierarchies for Alloc.ating R&D Resources
Hierarchy for Making Financial Decisions
To select one from a number of financial projects, we may consider the
benefits and cos ts separately (Figure 49). In the benefits hierarchy, three
possible scenari os are considered for the future . The company would like
to base its decision on a number of considerations whose impact depends
on the scenario. Hence the overall priorities will reflect the relative impor
Level1 :
Focus
Level2:
Scenarios
Level3:
Factors
Level4:
Alternatives
Level1 :
Focus
Level2:
Scenarios
... ,
Level3:
Factors
Level4:
Alternatives
BUSINESS DECISIONS 45
Financial Decision Making
Pessimislic StatusQuo Oplimislic
lncreased
Diversification
'V
Growth
Potenlial
Strong
Mar1<el
Position
(Early)
Currenl
Demand
Independance
lrom Internai
Economie
Slrenglh
Projecl A Projecl B Projecl C Projecl 0
Cos! Hierarchy
Pessimislic Slalus Quo Opl imislic
lncreased
Competition
induslrial
Accidents
Regulation
by
Governmenl
Unmanageabie
Cyciicaiity
insuHicienl
Knowledge
?/
Projecl A Projecl B Projecl C Pro je cl 0
Figure 49 Hierarchy for Financial Decisions
:l
C/J
N
w
i
1
!
1
l:
.;
1
'!3
l
i
'
' :
1
l' i
.   
PRACfiCAL EXAMPLES OF IIIF.RARCIIIES
tance of these factors . The projects are ranked according to their contribu
tion to each factor. The composite priorities give relative measures of the
benefits accruing from them.
ln the costs hierarchy, in a similar way, we find the relative impor
tance of a number of factors the company would like to avoid or minimize.
The overall priorities of the projects in this hierarchy, th en, give the rela
tive measure of the negative contribution of these projects. The benefit/cost
ratios give the superiority of the benefits over costs on a ratio scale. The
project with the highest marginal benefitfcost ratio is the best selection.
Hierarchy for Choosing a Marketing Strategy
The decision regarding a company's product/market posture depends
on a number of externat factors thal determine how far the company can
strive to maintain the status quo or expect an optimistic or a pessimistic
environment. The company' s objective regarding economie growth and
Level1 :
Mission
Level2:
Risk Factors
Level 3:
Scenari o
Leve14 :
Objectives
Level 5:
Acti on
Regulat ory
Standards
Weii Being of Company
Competitive
lntensity
Inflation
StatusOuo
Optimislic
Environment
;/"\
Pessimistic
Environment
Market
Profitability
Share
Sal es
Growth
Re duce
Vulnerability
lncrease
Promotion
l or
Producl A
Oelete
Product B
A aise
Priee of
Producl C
Enter
New Market
c;s;ure 410 Hierarchy for Choosing a Marketing Strategy
......
PERSONAL AND DOMESTIC DECISIONS 47
risk depends on the scenarios envisioned and will be achieved in varying
degrees by following different alternatives of product and market . The
decision mode! is thus represented in the form of a complete hierarchy
(Figure 410) .
By prioritizing the factors in one leve! with respect to each factor in the
preceding leve! and finding the overall priorities, we can find the relative
influence, feasibility, importance, or contribution, as the case may be, of
the factors in a leve! with respect to the focus : the company's well being.
The priority of each course of action is therefore a relative measure of how
far that product/market posture would achieve the desired wellbeing.
Hierarchy for Evaluating a Division' s Performance
There are severa! dimensions of the performance of a division in a
corporation. The principal dimensions to be considered in this evaluation
are government dealings, management, imports, and customers (Figure
411). There are severa! factors for each dimension. Leve! 3 of the hierarchy
shows those pertaining to management alone; other factors can be simi 
larly included for the other dimensions.
The overall priorities of the factors at level3 are the relative weights by
which the evaluators would view performance in that area. Composite
priorities of the various divisions with respect to ali the factors at this leve!
show the relative performance rating of the division on an overall basis.
PERSONAL AND DOMESTIC DECISIONS
Hierarchy for Choosing a Car
The problem of choosing from a number of cars, bath new and old, is
structured into a hierarchy of three levels (Figure 412) . ln the second leve!,
the various factors<osts as weil as benefitsgoing into the decision
maker's judgment are prioritized. Next the specifie alternatives are com
pared with respect to each factor in leve! 2. The overall priority of each
alternative indicates its ranking and strength of preference as far as the
buyer is concerned. (This is an example of a complete hierarchy because ali
the factors at any leve! relate to ali the factors in the next higher leve!.)
Hierarchy for Choosing a Home Computer
There are both benefits and costs in having a home comput er. Three
choices are prioritized in two separate hierarchies for benefits and costs
(Figure 413) . In each hierarchy, the priorities are determined
\ l
:1
' .,
1
' i
i
'j
CIJ
N
_p.
!
t
q
.,
;,j
;;
!1
' 1
ii ,.
j ;
1
I
l '
,,
,, :
1
1
. ;
, . !
' j j
"....
Level1 :
Focus
Level2:
Benelits
Level3:
Alternatives
Level1 :
Focus
Level2:
Costs
Level3:
Alternatives
PRACTICAL EXAMI'LES OF HIERARCHIES
Cost
Savings
lncome
Generat ion
Benelits
Cuslomer
Satisfaction
Legal
Aequiremenls
Corp.
DDA/Aecon.
Personnel
Costs
Customer
Analysis
Hardware
Costs
Costs
EFTS
Training
Costs
.............
Data
Entry
lnquiry
Operating
Costs
Corp.
DDA/Recon.
Customer
Analysis
EFTS
Data
Entry
lnquiry
Figure 48 Hierarchies for Allocating R&O Resources
Hierarchy for Making Financial Decisions
To select one from a number of financial projects, we may consider the
benefits and costs separately (Figure 49). In the benefits hierarchy, three
possible scenari os are considered for the future . The company would like
to base ils decision on a number of considerations whose impact depends
on the scenario. Hence the overall priorities will reflect the relative impor
Level1 :
Focus
Level2:
Scenarios
Level3:
Factors
Level4:
Alternatives
Level1 :
Foc us
Level2:
Scenarios
.. ;
Level3:
Factors
Level4:
Alternatives
BUSINESS DECISIONS 45
Financial Decision Making
Pessimistic StatusQuo Opl imislic
lncreased
Diversification
Growth
Potenlial
Strong
Mali<el
Position
(Eariy)
Currenl
Demand
Independance
from Internat
Economi e
Strenglh
ProjectA Projecl B Project C Projecl 0
Cos! Hierarchy
Pessimistic Slalus Quo Oplimistic
lncreased
Competition
indu striai
Accidents
Regulation
by
Government
Unmanageable
Cyciicality
lnsuHicienl
Knowledge
Project A Project B ProjectC Pro je cl 0
Figure 49 Hierarchy for Financial Decisions
:'.t
'l
i
1
1
f
(X)
N
w
' t
5
.i
: :
1
!
\
!
.!
1
1
1
JI
1 
1
H
!
1'
1 .
j:
'
1:
j!
l
1
,
t.
'1
'
:1'
1
i
, ,1
&
c
0
'iii
;;
i5
"'iii
a;
0
 V>
_l U.
PRACTICAL EXAMPLES OF HIERARCHIES
Jt
"' t
8.
E
.,
E
.,
:?
1ij
:::1!
N (1)
> 'E
.3u
g;
.,
'ii
(/)
"2
"'
QI
a:
ii
::>
0
.
., .,
1
c 0
U:i
[l_
\0
r
<
z
t U!
r
wn_
(1)
c:
.. Q)
Q) :::1
u
"' s:
6
co
5
'iii
;;
6
<(
c:
.Q
"' :2:
0
.. c:
"<!' 0
Cll 
_JQ
.   
cu
v
c:
j
t
Il.
c:
.Sl
!Il
Cl
1:>1)
c:
.c
"'
= ;;
>
t5

>.
.c
v
:;
t
:E
....
....
.
"" QJ
...
Level1 :
Focus
Level2:
Factors
Level3:
Alternatives
PERSONAL AND DOMESTIC DECISIONS
Initial
Cos!
Maintenance
Cos!
Choosing an Automobile
Fuel
Economy
Comfort Status
49
Uncertainties
?/'
Alternative 1 Alternative 2 Alternative 3
Figure 412 Hierarchy for Choosing a Car
intennediate levels of factors and subfactors contributing to the benefit or
cost of home computers. Finally, the benefit/cost ratio for each alternative
is calculated to find out where the benefits stand relative to the costs. The
alternative with the highest marginal benefit/cost ratio is the preferred
choice for the decision maker.
Hierarchy for Choosing a Career
In choosing a career, a person desires satisfaction in severa! dimen
sions : intellectual, financial, social, and persona!. However, one's sources
of satisfaction are from expectations of learning, growth, leisure, friends,
and prestige. Each source of satisfaction may derive fulfillment from sev
era! dimensions, which therefore occupy a superior leve! (Figure 414) .
The priority of each career with respect to any criterion of satisfaction
reflects its desirability with respect to that criterion only. The overall prior
ity of the career shows the overall preference for that career; the highest
priority career is the one preferred. (This is an example of an incomplete
hierarchy, as factors in leve! 3 do not relate to ali the factors in leve! 2.)
Hierarchy for Choosing an Ideal lnvestment
The different criteria affecting an investment are prioritized to find out
how important they are in relative terms to the investor in question (Figure
415) . In the next step, the different investment alternatives are ranked
with respect to each criterion to examine how weil they satisfy the criterion
in question. The overall priority will show the relative superiority of an
:t
,;
1
,;
,;:
!_ 1
:,f
:,11
' '
.,
;a
t!l
!
1
co

N
Vl
!
'
..,
Level1 :
Foc us
Level2:
Factors
Level3:
Subfactors
Leve14:
Alternatives
Level1 :
Focus
Level2:
Factors
Level3:
Subfactors
Level4:
Alternatives
PRAC riCAL EXAMPLES OF IIIERARCIIIES
Home Computer Benefits
Business
Education Entertainment
Persona!
~ ~ ~ ~
Me mory
Capaclty
~
Spending
Monay
~
Software Expansion
Languages Special Avallabllity Capacity
r
~ :?\
~
Choice A
Choice B
Cholce C
Home Computer Cosls
Economie
Psychologie al
ldeological
/{\
{\
""'
Wasting Wastlng . Lack of lack of
Monay Tlme Frustration Fa!igue Entertainment Education
~ ~ ~ ~ ~
~
ChoiceA
Choice B
ChoiceC
Figure 413 Hierarchy for Choosing a Home Computer
. t
Level1 :
Focus
Level2:
Factors
Level3:
Subfactors
Level4:
Alternatives
Level1:
Focus
Level2:
Criteria
Level3:
Alternatives
PERSONAL AND DOMESTIC DECISIONS
51
Career Choice
lntelleelual Financial Social Persona!
1\ 1 \' ?!\,
Learnlng Growth leisure Friands Prestige
~ ~ ~ ~ ~
CareerA .Career B CareerC Ca reer D Career E
Figure 414 Hierarchy for Choosing a Career
Appreciation
~
Bank
Savings
ldeallnvestment
Rate
of Risk Experience
Return
~ ~ ~
Real
Estate
Stocks
Easeof
Monitoring
liquidity
lnvestment
on One's Own
7/\ ~
Troasury
Bonds
Figure 415 Hierarchy for Choosing an Ideal lnvestment
Q)
N
0'1
' :1
i
f
,.
1
1
l
1
  __,. ."!."'..... 
Level1 :
Foc us
Level2:
Criteria
Level3:
Subcriteria
Level4:
Alternatives
Educalional
Qualrty of
Teachers
General Standard
Discipline
Preparation for
Co liege
leammg
Environment
a:: . .. ...... ..,. _ _ .., ___ ...__ __
Selection of School
Vocational Spons Cultural
Facilities for Facilities Music
Training
Instruction Fine Arts
Ouality of
lnstructors Reputation Dramat1cs
1
SchooiA SchooiB
Figure 416 Hierarchy for Choosing a School
Social
Friends
Atmosphere
SchooiC
Spiritual
Religious
Instruction
Inculcation
of Desired
Values
lib. ki UlkX!kU lk!l .!LCJilX .... ( .,M4QtUJMJt.Ut LAL$2621 145
Level1:
Focus
Purchasing a House
Ul
N
"0
s;
(
i
n
>
m
x
>
3::
"0
r
m
IJl
0
"Tl
:;
m
;>:l
>
;>:l
Q
...
m
IJl
 
54 PRACTICAL EXAMI'LES OF IIIERARCHIES
alternative as viewed by the in vestor. For a large investor, the priorities
.also indicate the proportions in which the total investment could be dis
tributed among the various alternatives availablethat is, a representation
of the investor's portfolio.
Hierarchy for Choosing a School
First we establish priorities for the criteria of the student and the
parents to establish their relative importance (Figure 416). ln the next
leve!, we resolve the criteria in greater detail and prioritize the subcriteria.
Each of the schools under consideration is then prioritized with respect to
the subcriteria. The overall priorities show how strongly a school is rated
by student and parents in relation to the others.
Hierarchy for Purchasing a House
Considerations that apply to purchasing a house are the criteria,
which are prioritized to find their relative importance (Figure 417). ln the
next leve!, these criteria are decomposed into further subcriteria, which are
similarly prioritized. ln the next stage, the alternative houses under con
sideration are prioritized with respect to each criterion or subcriterion,
and their overall priorities indicate the buyer's preferences for the houses
in question. This is a comprehensive model in which the criteria pertain
ing to the neighborhood and those pertaining to the qualities of the house
are dealt with simultaneously.
PUBLIC POLICY DECISIONS
Hierarchy for Choosing a Mode for Crossing a River
To decide which mode to use for crossing a river would be beneficiai to
the community as a whole. We consider the nature of benefits envisaged
and enlist under each the details (Figure 418) . Setting priorities for the
benefits gives an idea which ones the community regards as important.
We can also establish priorities for the costs .
Hierarchy for Choosing Areas for R&D Programs
This hierarchy (Figure 419) is concerned with the selection of areas for
research and development to ensure adequate power and electricity in the
future. At the first stage, the focus is on the range of planning by prioritiz
ing the time horizon of the plan. Next we set priorities for the potential
... .
l i
5
>
i
<U
g>
2i
e
u
"5
V>
:t::
a;
c
>
.3u.
PUBLIC POLICY DECISIONS
.j
"'
8
c
w
ut
EE
g
8 Cl>
wco
.. o.l!l
C\1 1/)
H J!
.2!
.!.!
a;
\0
J:
ii
w
.51
E
t'
8
(
ea.
8
c)ll<Uij
a:
"' c
r
0
::J
E
E
8
c
<
0
L
E Ci!
8 [/
e
E o
YI
f
1=
" .l!l
Cl> Cl>
> c
.3J!
  ..
t
Cl>
u.
g>
;:;
"'
)(
w
ai
c
c
:>
t
Cl>
g'
c
co
"'
Cl>
>
..
t
>
..
llO
c:
0
u
._
0
"'
..
0
._
;...,
v
..
..
::r::
ao
0
::s
llO
iZ
55
co
N
,
J
.:i
i
1
1.
l
:!
l
...
:i
. 1
!
1'
' 1
,l
.i\
i '
!;!
.\
1
56
j
8
1
.:
j
1
ll.L
PRACTICAL EXAMPLES OF HIERARCHIES
i
1? / UOijeiUIISUO::J pue puUWOQ
i
u. IIW8,&AS
5
ii(
j'.
:g
! L f h8 / ::::::,
i l 8PGJ13 411eOH
UJ
! p
w
69!J9UV8
S9:l!A9Q
UQ!I8'JI6UOW9Q IIO::J IBn;j
IU9wd0,9A9Q pe:lUBAPVII8:) 19n;j
UO!GJ91\Uo:t A6J9U3
li
\ . e
....
U{)!IIII!WIUBJ! :JV
UQ!IIS' JW$U'IJ1 JO
UOfSS!WSU8Jl punoJ6J9pUO
.
IIIOIJP913
IOJIUO:) pue 6u!UU81d
IBWJ941081 SjUBUeJI tlMl
l:l8::1t'4l
N g'
Q) fa
l O..CI:
li
Q.
. . ij !J)
> 0 c 0
..3a..w(f)
i
..,. ra
 t:
Q)Q)
> :E
..3o
j
. . 1Q
Il) . Si!
 c :Q
Q)
Q) a; ..
li<
uatsn::1
&IUBid W881$ j80::J
UOIISnqUJ<Y.) P98 PBZ!P!"I::I
IBO:) PB!l!SWQ
P<i<>S
1
8
0:> uee1::> pue UO!Paoni:J!l
6U!t00:)
pue U!Ie&H 9:lUei&IS91;l
UOUlJ81\U00
'"lOS
.. 5
<D'
"':g
:;: Q)
Q) Q) ....
lCI:<
.,
a
e
Il.
0
a:
till
c::
ii
0
.8
u
;..,
1
QI
:E
0'1
....
.,j.
i.i:
PUBLIC POLICY DECISIONS 57
areas of energy resources to determine which one holds the most promise.
With respect to each energy source, several criteria are considered that
require attention; these criteria are prioritized according to the importance
they command for the energy resource.
For each criterion, several technical aspects are identified and then
prioritized for relative importance. The overall priority at this levet shows
the share of effort and resources that should be devoted to the technical
field concerned. As a continuation of the process, each technical field is
further subdivided into research areas and subareas that need attenti on.
The overall priorities at each levet again represent the extent to which
resources should be devoted to those areas according to the best overall
judgment of the decision makers.
Hierarchy for Choosing a Program to increase Harbor
Capacity
Deciding on means to improve harbor capacity in a small country with
three harbors is viewed as a joint political process. The three principal
stakeholders are the evaluation group of specialists, the transportation
committee of Jegislators, and the harbor bureau (Figure 420). Their priori
ties represent their influence in the matter.
The objectives and considerations to be followed by each stakeholder
are identified and prioritized. At the next stage, the three existing harbors
are prioritized with respect to each objective to find out how much they
contribute to the objectives. The overall priority at this levet shows the
extent to which resources and attention should be devoted to each harbor
development .
The specifie programs to increase harbor capacity are next identified
and prioritized with respect to each harbor . In the context of each harbor,
these priorities show the effectiveness of each action in achieving the goal.
The overall priorities, obtained by weighting the harbor priorities, reflect
the effectiveness for the whole country. The actual policy that is adopted
will consist of several individual programs pursued with varying degrees
of emphasis as represented by these priorities.
Hierarchy for Evaluating Energy Storage Systems
To evaluate four advanced energy storage systems, six feasibility crite
ria are set .(Figure 421) . They are then prioritized to see which would be
dominant over others in relative terms .
Next we prioritize the energy storage systems with respect to each
criterion regarding their suitability. The overall priorities, obtained by
weighting them with criteria priorities, show how they are expected to
<Xl
N
1.0
:
.i
...
,,
,.
t.'
' '
Level1:
Foc us
Level2:
Stakeholders
Level3:
Objectives
Level4:
Harbors
LevaiS:
Action
Programs
_..,. _ ....   .. ............ _   
PRACTICAL EXAMPLES OF HIERARCHIES
How to lncrease the Harbor's Capaclty
Evaluation
Group
(Speclalists)
Transportation
Commlttee
(Congress)
Harbor
Bureau
Economie
Equlty and
Stabilization
lncrease
Berths
Evaluation
of
Eco nom leal
Elficiency
and
Benelit
Cost
Desire of
Shipowners
Desire of
Cargo Owners
Keelung Taichung Kaushung
Harbor Harbor Harbor
(North) (Middle) (South)
'if\
lmprove Push
Promote
Conlrol Port Are a Procedure
Loading
Pattern of and of
Efficiency
Ship lnland Customers
Arrivai Transport and
System Administration
Figure 420 Hierarchy for Increasing Harbor Capacity
satisfy the criteria. The highestpriority storage system is the one pre
ferred.
Hierarchy for Allocating Resources in Juvenile Correction
Programs
A group of public officiais were interested in juvenile law enforcement
and wanted to allocate resources in five programs the staff had suggested.
To with, they considered three principal areas of correction and
.CTICAL EXAMI'LES OF HIERARCHIES
prioritized them to find out how much attention they should receive (Fig
ure 422).
Each principal area could be associated with the programs identified,
so the programs were prioritized regarding their effectiveness with respect
to each area. The overall priorities obtained after weighting by the area
priorities show how much relative importance each program commands
for the optimum juvenile correction system.
Hierarchy for Analyzing School Busing Conflict
The introduction of busing due to the 1954 Supreme Court ruling has
been a source of friction in a certain school district. The minority commu
nity wants to introduce complete busing for racial integration at school;
the majority community wants segregation to protee! the privileged posi
tion they now enjoy. To analyze the situation and judge the potential
outcome, we rank the stakeholders according to their relative influence on
the political scene (Figure 423). Then we prioritize the objectives of each
stakeholder to see which objectives weigh more and should thus be pur
sued in preference to others. The overall priorities give a picture of the
relative strengths of the forces at work on the scene.
The outcomes under consideration here are three scenarios spanning
the complete spectrum of possibilities. They are prioritized with respect to
each stakeholder objective to find out which outcome is favored by thal
objective. The overall priorities indicate the relative likelihood thal each
possible outcome will occur. This exercise shows the interaction of various
Level1 :
Focus
Level2:
Correction
Are a
Level3:
Programs
Resident iel
Prevention
Juvenile Correction
Preadjustory
Programs
Postadjustory
Programs
~ ~ ' l j \
Educational Vocational Family
lntake and
Referral
Figure 422 Hierarchy for Allocating Resources in Juvenile Cor
rection Programs
Level1 :
Focus
Level2: Majority
Stakeholders Community
Education
for
Children
Level3: Re tain
Objectives
Power
Ensure
Public
Safety
~
PUBLIC POUCY DECISIONS
School Busing Conflict
Minority City Board of
Community OHicials Education
Education
Higher
for
Compensatory
Chlldren
Reputation Wages
Equal
Win Public
Rights
Election Safety
Ensure
Retain Maintain
Public
Power Standard
of
Safety
Education
~ ~
~
61
Federal
Government
En force
Law
and
Or der
lmprove
Quality
of
Education
Enforce
Busing
lor Ail
7
Level4:
Outcomes
Complete Busing Part ial Busing No Busing
Figure 423 Hierarchy for Analyzing School Busing Conflict
factors and forces at work, so thal if one stakeholder wanted to influence
the outcome, he or she could decide accordingly on a course of action such
as coalition or persuading others to change their objectives .
Hierarchy for Analyzing Health Administration Conflict
This hierarchy for resolving conflict in health administration finds the
likelihood of various health plans being adopted as national po licy (Figure
424) . First we prioritize the principal actors regarding their relative influ
ence on the issue. Next their objectives are identified and prioritized to
indicate the relative extent to which the actors are motivated by various
considerations.
At the next stage, we establish which policies of the actors would
satisfy the objectives. The policies thal relate to the same objective are
Il>
J:
c;:
c:
8
::V)
....JU.
PRACTI CAL EXAMPLES OF HIERARCHIES
cr
.
!Xl
cr
cr
cr
I
<L.1.
.
<( ..C:+
cr
cr
!Xl
,!;
ij
8
Qi 0
>t;
3<
<
<(
<(0
"'
Q)
.. >
> .!t
38
cr
fl
!Xl.
cO
cr
V)
...,. Q)
"ijj (;)
>'=
Q) 0
....Jil..
..JO..
a; c:
t: "'
(JC::
c:o..
"'
QiE
..Je(
't
...
<:;:l
c::
0
u
c::
"
li:
= ;
= 1:10
c::
...
N
.
" c::
(
!S
......
"
"
t
:E
.:!'
N
0
...,.
1:10
iZ
PLANNING ECONOMIC POLICIES 63
prioritized to find the importance of the policies to serve the objective in
question. The composite priorities show the overall influence of the poli
des in national health administration.
There are three main health plans being debated nationally. By
prioritizing the plans with respect to each policy, we find to what extent a
plan satisfies the po licy in q uestiono Th us the overall priorities of the plans
show the relative extent of support for the plans from various interested
quarterso The overall priorities are also the likelihoods of the plans being
adopted nation ally o
The elements of that hierarchy are defined as follows :
A
0
: comprehensive health insurance
A
1
: mixed privatepublic financing
A
2
: containment of health care costs
A
3
: minimum government involvement
A
4
: expanded health coverage
A
5
: control of health care costs
A
8
: comprehensive health services
A
1
: containment of the total costs of health ca re
B
1
: employeremployee contribution
8 2 : limited public funding
8 3: tax subsidy
B
4
: deductibles/coinsurance
B
5
: planning/integration
B
8
: utilization control
B
7
: advisory federal agency
B
8
: state regulation of N HI
Bv: no provider's charge control
B
10
: increased spending of public funds
B
11
: employer only paid premium
B
12
: control of provider's charges
B
13
: administrative reorganization of Medicare and Medicaid
B
14
: wagerelated premiums
B
16
: nonwagerelated premium
B
18
: premium pa id from public funds
B
17
: competition between providers and insurers
PLANNING ECONOMIC POLICIES
Hierarchy for Planning Economie Strategy for an
Underdeveloped Country
To map the strategy of economie development in an underdeveloped
country where oil is the chief source of revenue, we draw up a plar
0
"'g
Level1 :
Foc us
Level2:
Premises
Level3:
Sectors of
Economy
Level4:
Policies
Level5:
Strategies
PRACfiCAL EXAMPLES OF HIERARCHIES
Developed
Agriculture
~
More Equal
lncome
Distribulion
~
TradeBalanced Economie OevelopmentYear 2000
Minimum Foreign
Exchange
Requirements
~
Developed
Capital
and
tntermediate
Goods
0'
Fuller
Employment
of Labor
and
Capital
StatusOuo
(lmport
Substitution
Policies)
/'\
Oras!lc
Reduction
in
Protection
Maximum Availabilily
of Foreign
Exchange
~
Oeveloped
Consumer
Goods
and
Services
Developed
Tourism
~ 'l1
More
Rational
Allocation
of
Resources
~
Timetable
Endto
Protection
More
Productive
New
lnvestment
~
Figure 425 Hierarchy for Planning Economie Strategy
hierarchy (Figure 425) . Although ol plays a very important role in the
economy now, it is expected to be depleted by the year 2000. The country's
planners must develop a strategy with that eventuality in view.
First, Iwo principal underlying premises for a future oilfree economy
are identified and then prioritized to indicate which one will have a more
domi nant influence. Next the important sectors of the future economy are
prior iti zed to indicate the degree to which they will become important
according to the premises. Specifie rational policies are then prioritized
wi th respect to each sector of the economy to determine which would be
ESTIMATING AND PREDICTING 65
more effective. Finally, the strategies in question are prioritized with re
spect to the policies regarding their beneficiai role on each.
The overall priorities for the policies as weil as strategies show the
national planners which ones will be more effective and hence should be
actively promoted in an overall context.
ESTIMATING AND PREDICTING
Hierarchy for Estimating Popularity of Rock Groups
To estimate the popularity of various rock groups, we structure the
overall performance of a music group into three main components by
which they are judged (Figure 426) . These components are then
prioritized regarding their relative influence.
Eah component is further judged in light of :.;everal characteristics of
performance, and the characteristics are prioritized regarding the ir relative
influence. Finally, we prioritize the rock groups with respect to the charac
teristics to give their relative standing. The composite priority we obtain
gives a relative mea:.;ure of overall superiority and hence an index of thei r
popularity.
Level1 :
Foc us
Overall Preference for Rock Group
Level2:
Music Lyrics Pertormers
Factors
~
(\
~
Level3: Compfexity Smoothness Emotional Signilicance Se x
Personality
Subfactors
of Style of Sound Appeal of Message Appeal
Level4:
Alternatives
Various Rock Groups
Figure 426 Hierarchy for Estimating Popularity of Rock Groups
())
w
w
,,
l
,,
1
1
;
j.;
t
'
):
f
j :
\
ji
':
!
l '
, J
i l
il
:;
c
c
 V)
(!) ::l
8
lU..
PRACTlCAL EXAMPLES OF HIERARCHIES
1(
:l
<llo::
.J.E
]l
.::
:::>
.!!!
cu
0>
..
_
j
:e
O
8
(") !:?
 0
<1l t1l
lU..
c
I
c ,g
[
.,
0 !E
E
n.
8
<
cl,
u u
5
a: n.
r j
w
'o
w u.ll.
J!l
1
l 1
.c Q)
n. n.
.,
O
w
6
!:?
0
'<t
rn
<ll
>.0
<1l ::l
.J(/)
"'
..
.c
u
c
8_
)(
w
.,
Q)
w
(
l
fi
1ii
m
a:
;
t
u
V)
<1l
.. iii
10'0
c
,g
t
" Q
;;
:D
5
,
...
"
l
..
CIO
c
s
i
"
ti
1
.,
...
"'
ESTIMA TING AND PREDICTING 67
Hierarchy for Predicting the Election
To predict a presidential election, we begin by identifying the princi
pal political strains influencing ali coalitions and alliesliberais, moder
ales, and conservativesand prioritize them regarding their influence on
the contemporary political scene (Figure 427) . Next we consider the
characteristics of the candidates and prioritize them to find to what extent
they are decisive. These factors are subdivided into more detailed elements
and then prioritized. The overall priorities we obtain assess how far these
individual elements influence popular choice.
Finally, the three principal candidates are prioritized with respect to
each subfactor to estimate their relative superiority. The overall priority we
obtain is an index of the predictability of the candidate' s winning the
election.
Hierarchy for Predicting Likelihood of Technical
Innovation
We can investigate the likelihood of technical innovation associated
wHh corporate planning with respect to three forms of corporate control :
traditional public ownership, employee ownership, and government own
ership. We study the relationship in three representative industries (Fig
ure 428) .
For each industry, the related corporate factors are prioritized with
respect to each form. the actors are prioritized with respect to each
corporate factor to obtain their relative impact on the factor in question.
The objectives of each actor are prioritized to find their relative strengths
in influencing decisions, and, finally, the three forms of corporate control
are prioritized to indicate how far each would facilitate the objectives in
question. Therefore, the overall priorities at this leve! show the relative
likelihood that technical innovation will be fostered by the different forms
of corporate control.
Hierarchy for Predicting Number of Children in an
Average Family
This exercise is carried out to anticipate the number of children an
average family is likely to have. Severa! criteria are considered to influence
the number of children in a family: the availability of birth control mea
sures, working mother, older age at motherhood, education of mother, cost
of raising children, and social pressure (Figure 429). These criteria are oo
prioritized to show the relative degrees of influence they exert . w
l
!
1
l
1.
i'
i
1
J
1
}
1
' l
1
l
i
!1
.E
c
'S
0
Gi
"
::J
 Vl
...J IL
. . ..
PRACTICAL EXAMPLES OF HIERARCHIES
...
Il(
Vl
.. 9!
 Vl
Q) ::l
_._
!!l
g c
ti
(.) Vl
 0
Ql ro
...JIL
i
Q: (3
a:
ra
e::J (/) J!
w
:l
()
g."' .;::; Q)
"'
8
.eni
(.!) 0
a..
i
...,::J
w
::
::l
a:
c o.
.2 Q) 
ti
(/) (.!)
(
f
t:."
>Cl>
:
 E
g> c
c
i!
c "'
g ... "'
;: ..., "'
g>
w (/)
Cl
a(!
a:
C1l 0
>u
.3<
"' 0 "' , 0
en
i?n
:l "' LLLL
(/)
C1l
.. >
Ill:.=
Qi al
.g
p
"'
(/)
e:.
(/)
C1l
<DE
Qi 8
>
Cil ::l
...JO
1
.s
;;
"'
.c
"'
.....
0
a
0
0
3
Cil
,_j
()[)
.s
Cil
..
..
0
......
>.
.c
"'
..
::r::
CIO
N
.;. g
QI ...
.. ..
::s l'il
oc>
0
Il.. c:
Level1 :
Focus
Level2:
Criteria
Level3:
Possibilities
Availability
ot
Birth Control
MEASURING INFLUENCES
Working
Mother
FamilySize
Older
Age at
Mothemood
Education
o!Mother
Costot
Raising
Children
69
Social
Pressure
No
Children
1 Child 2 Children 3 Children 4 Children
Figure 429 Hierarchy for Predicting Average Number of Chil
dren .
The number of children considered here runs from zero to four . (We
assume that the number of families having more children is very small.)
With respect to each criterion, the priority of the number of children is
found by ranking on the basis of best judgment . The overall priorities
obtained after weighting by the priorities of the criteria reflect the distribu
tion of the number of children in an average family . The expected value of
this distribution is the average number of children a family is likely to
have.
MEASURING INFLUENCES
Hierarchy for Measuring Parental Influence on
Psychological WeiiBeing
,,
A study of an individual's psychological wellbeing traces the extent
to which it is influenced by the mother and the father separately and by
both jointly. Psychological well being depends on selfrespect, a sense of
security, and the ability to adapt to others (Figure 430) . The factors that
contribute to these qualities in one's developing years are in the third
leve); ali are nfluenced by the parents separately and jointly.
The overall priorities of the factors at any leve) show their relative
contribution to the person' s well being. The overall priorities at the lowest
leve) represent the influences that the parents brought to bear separately
and jointly in the person' s formative years .
L
V1
1
r
1
1
f
t:
f
;:
;; .:
.!,
' ;
"
'
'
l1
J I
ti
1 .
r .
i'
l
li
i.
i
J.
l
'f
J':
.,.
i
.  
PRACTICAL EXAMPLES OF HIERARCHIES
Level1 :
REPRESENTING SYSTEMS NElWORKS
Focus
Level2:
Criteria
Level3:
Contributory
Factors
Visible
AHection
for
Subject
SelfRespect
Ove rail WeiiBelng
Sense
of
Security
Ability
toAdapt
toOthers
Idees of
Strict ness,
Ethics
Actual
Disciplining
ofChild
Emphasis
on
Persona!
Adjustment
wlth Others
A' ?/
Level4:
Influences
Influence
of Mother
Influence
of Father
Influence
ofBoth
Jointly
Figure 430 Hierarchy for Measuring Parental Influence
REPRESENTING SYSTEMS NETWORKS
Network of a Child's Learning System
A child's leaming process in the formative years is influenced by
many factors. These factors may be classified into certain groups. What is
significant, however, is that the various factors interact with and influence.
one another. Hence we represent them in the form of a system consisting of
subsystems and further divisions interacting with one another (Figure
431). .
By prioritizing the factors in a group with respect to each factor in
each group separately, weighting them by the relative importance of the
groups themselves, and finding how they influence one another, we can
come to a conclusion regarding the intrinsic importance of the various
factors bearing upon the child's upbringing.
The system depicted here consists of five subsystems, which ali impact
on the child' s leaming elements (CHL) (motivation, M; creativity, C;
u.1
()
en
tdCl.
E
"'
>.
Vl
1>0
c
'
l'li
cu
....J
Il)
5:!
:a
u
l'li
'""' 0
15
!
cu
z
....
<"')
'
cu
..
6b
ii:
co
w
(J'\
PRACTICAL EXAMI'LES OF HIERARCHIES
reflection ability, RA; retention and assimilation, R; character and person
ality, CHP)
1. The school (S)
2. The family (F)
3. The community (C)
4. The family atmosphere (FA)
5. The home (H)
The most important components associated with the school subsystem (S)
are:
1. Teachers (T)
2. Classmates (CL)
3. Installations (1)
4. Study programs (P)
Those for the family (F) are:
1. Parents (PS)
2. Brothers and sisters (B)
3. Close relatives (those who interact directly with the family) (CR)
Those for the community (C) are:
1. Family friends (FF)
2. The child's friends (CHF)
3. The environmental living characleristics (EL)
Those for the family atmosphere (FA) are:
1. The relation between the parents (RBP)
2. The relation between the parents and the child (RPS)
3. The relation between the child and brothers and sisters (BR)
4. The overprotection of people older than the child (OVP)
5. The family's economie status (ES)
And those for the home (H) are:
1. Physi cal characteristics (PCH)
2. The child's toys (TS)
3. The communication media (radio and television) (CM)
This mode! has been evaluated for a certain community, at an average
s o c i o e c o n o ~ i c leve!, in a democratie regime, and for children eight years
old with normal psychological characteristics.
PERSPECTIVE 73
Figure 432 Network of a Volleyball Team
Network of a Volleyball Team
Volleyball is agame of teamwork and simple skills, but the play.ers are
required to change positions and possess al/ the skills . Since few players
have ali the skills to the same extent, the coach has to use players with the
proper mix of skills. Th us the coach has to know the relative importance of
skills and play the proper combination of players. Since the various skills
depend upon one another, we use systems analysis and nol hierarchical
representation (Figure 432) .
By prioritizing the skills for a good game of volleyball and prioritizing
the players with respect to one another's skills, we get the relative stand
ings of the players and skills in the ultimate analysis . This information
helps the coach to select players with basic skills .
PERSPECTIVE
The hierarchies just presented are only a few of those thal have actu
ally been used to make decisions with the AHP. But they suggest the wide
range of problems to which the AHP can be appliedfrom choosing a car
to crossing a river. Specifically, the AHP can be used for the following
kinds of decision problems:
o Setting priorities
o Generating a set of alternatives
co
w
....,]
\
l
1
. ~
j
T
~ ~
:i'
i)
:j
, ..
J
i
~
~ :
8
m
1
~
1
~
~~ .
. ' ,,
' :.
~
~
i
~
m
PRACTI CAL EXAMPLES OF HIERARCHIES
Choosing the best policy alternative
Detennining requirements
Allocating resources
Predicting outcomes and assessing risks
Measuring performance
Designing a system
Ensuring system stability
Optimizing
Planning
Resolving conflicts
The sample hierarchies in this chapter should give sorne insight into
the need for various levels in different kinds of hierarchies . Different
people may have their own idea about how to deal with a problem, and
that is how they should carry it out . Only when a group must act together,
as in a corporation, do people need to agree on the structure of their
problem.
In forming a hierarchy, one should include as much detail as seems to
be needed to understand the problem; the prioritization process will
eliminate elements that are unimportant . A new leve! should be added to
the hierarchy if it facilitates the comparison and evaluation of the elements
in the leve! immediately below and contributes to improving precision in
the judgments. One contribution of a good hierarchy is thal it enables
people to make better guesses about the effects of the unknown by laying
out its components and studying each separately instead of Jumping ev
erything together and making one big guess at the consequences of deci
sions made in the face of that unknown. The hierarchy can provide an
effective buffer between reason and worry.
Clearl y the design of an analytic hierarchylike the structuring of a
problem by any other methodis more art than science. There is no pre
cise formula for identification or stratification of elements. But structur
ing a hierarchy does require substantial knowledge about the system or
problem in question. A strong aspect of the AHP is that the experienced
decision makers who specify the hierarchy also supply judgments on the
relative importance of the elementswhich brings us to the next topic:
establishing priorities.
....
5
Establishing
Priorities
This chapter deals with the following questions:
How do we establish priorities in a decision
problem?
Why is the matrix useful in setting priorities?
How do we synthesize our judgments to geta
set of overall priorities?
How can we check the consistency of our
judgments? And how important is consistency?
What do we do when the elements we are
ranking overlap?
75

CX>
w
CX>
l
f
 . ..
EST I'RIORITIES
THE NEED FOR PRIORITIES
ln this chapter we complete the analytic hierarchy process by estab
lishing priorities among the elements of the hierarchy, syn.thesizing our
judgments to yield a set of overall priorities, checking the consistency of
these judgments, and coming to a final decision based on the results of this
process.
Systems theorists point out thal complex relationships can always be
analyzed by taking pairs of elements and relating them through their
attributes . The abject is to find from many things those that have a neces
sary connection. This causal approach to understanding complexity is
complemented by the systems approach, whose abject is to find the sub
systems or dimensions in which the parts are connected.
We have seen thal the analytic hierarchy process deals with both ap
proaches simultaneously. Systems thinking is addressed by structuring
ideas hierarchically, and causal thinking, or explanation, is developed
through paired comparison of the elements in the hierarchy and through
synthesis.
The judgments we apply in making paired comparisons combine logi
cal thinking with feeling developed from informed experience. The math
ematical sequence described in this chapter is a more efficient method of
arriving at a solution than the intuitive means we usually employ, but the
end result is not necessarily more accurate. If the solution reached through
the AHP does not fee) right to an experienced, wellinformed decision
maker, then he or she would do weil to repeat the process and restructure
the hierarchy or improve the judgments. On the other hand, the AHP
provides its own check on the consistency of judgments, and experience
has shawn that the results of the AHP closely approximate decisions
reached more laboriously in the business world.
lt is important to note thal the calculations described here can be
ca rried out by computer. My aim is not to dwell on the mathematics of the
processa mathematical supplement is available for those interestedbut
to ex plain how subjective judgments can be quantified and converted into
a set of priorities on which decisions can be based.
SETTING PRIORITIES
Tlw first step in the priorities of elements in a decision
problem is to make pairwise comparisonsthat is, to compare the ele
ments in pairs against a given criterion. For pairwise comparisons, a ma
tri x is the preferred form. The matrix is a simple, wellestablished tool that
offer s a fr amework for tes ting consis.tency, obtaining additional informa
ti o n thr ough making ali possible comparisons, and analyzing the sensi
ti vi ty of overall priorities to changes in judgment. The matrix approach
uniquely reflects the dual aspects of priorities: dominating and dominated.
SEHING PRIORITIES 77
To begin the pairwise comparison process, start at the top of the
hierarchy to select the criterion C, or property, that will be used for making
the first comparison. Then, from the level immediately below, take th
elements to be compared: A v A 2, A 3, and so on. Let us say there are se ven
elements. Arrange these elements in a matrix as in Figure 51 .
c
A,
A2
A,
A,
A A,
Figure 51 Sample Matrix for Pairwise Comparison
In this matrix compare the element A, in the column on the left with
the elements A 1, A2, A
3
, and so on in the row on top with respect to
property C in the upper lefthand corner. Then repeat with colurrin ele
ment A2 and so on. To compare elements, ask: How much more strongl y
does this element (or activity) possess or contribute to, dominate, influ
ence, satisfy, or benefit the property than does the element with which it
is being compared?
The phrasing of the question is important. Il must reflect the proper
relationship between the elements in one leve! with the property in the
next higher leve!. If time or another probabilistic criterion is used, then
ask: How much more probable or likely is one element than another? If the
elements are dominated by the property rather than vice versa, ask how
much more strongly the element is possessed, dominated, affected by, and
so on, this property. In projecting an outcome, ask which element is more
likely to be decisive or to result in the outcome.
To fill in the matrix of pairwise comparisons, we use numbers to repre
sent the relative importance of one element over another with respect to the
property. Table 51 contains the scale for pairwise comparisons. lt defines
and explains the values 1 through 9 assigned to judgments in comparing
pairs of like elements in each leve! of a hierarchy against a criterion in the
next higher leve!. Experience has confirmed that a scale of nine units is
reasonable and reflects the degree to which we can discriminate the inten
sity of relationships between elements. When using the scale in a social ,
psychological, or political context, express the verbal judgments first and
then translate them to numerical values . The numerically translated judg
co
w
1.0
L
lntmsity of
Importance
3
5
7
9
2, 4, 6, 8
Reci procals
Rationals
ESTABLISHING PRIORITIES
Table 51 The Pairwise Comparison Scale
Definition
Equal importance .
Moderate importance
of one over another.
Esaential or
importance .
Very stronq impor
tance .
Extreme importance .
Intermediate values
between the two ad
iacent judgments.
Exp/an a ti on
Two elements contribute
equally to the property
Experience and judgment
slightly favor one element
over another
Experience and judgment
strongly favor one element
over another
An element is strongly
favored and its dominance is
demonstrated in practice
The evidence favoring one
element over another is of the
highest possible order of
affirmation
Compromise is needed
between two judgments
When act1v1ty 1 copared to j 1s ass1gned one
of the above nubers, then act1v1ty j compared
to 1 1s ass1gned its reciprocal.
Ratios aris1ng from forcing cons1stency of
judgents.
ments are approximations; their validity can be evaluated by a test of
consistency, which will be described later, and by reallife applications for
which the answers are already known.
When comparing one element in a matrix with itselffor example, A
1
with A
1
in Figure 51the comparison must give unity (1), so fill in the
diagonal of the matrix with 1s. Always compare the first element of a pair
(the element in the left hand column of the matrix) with the second (the
element in the row on top) and estimate the numerical value from the scale
in Table 51. The reciprocal value is then used for the comparison of the
second element with the f1rst. For example, if the two elements are stones
and the first is five times heavier than the second, then the second is
onefifth as heavy as the f1.rst .
Why not simply use arbitrary numbers for ranking the elements ac
.IIi :,
. .'
SYNTHESIS 79
cording to their impact on a criterion? If the problem concerns simple
ranking, and the degree to which the elements being ranked reflect the
criterion is obvious, then one can simply assign numbers. To discriminate
the relative strength with which each element possesses or contributes to
the criterion (property), numbers can be used directly by starting with the
smallest element and perhaps using it as a unit . This procedure may be
useful in organizing one's thinking, but the logic is not clear and, more
over, feeling is not integrated into the process. For fine distinctions, the
pairwise comparison matrix and scale provide a more satisfactory frame
work.
When tradeoffs must be made among severa! criteria, the problem of
ranking becomes complex. lt is no longer sufficient simply to assign arbi
trary numbers. We must select with care the numbers used to express the
strength with which each element possesses or contributes to the property
in question. Such care ensures that in the end we obtain the correct overall
priorities for the elements by considering ali tradeoffs . (These priorities can
also then be used to allocate resources.)
SYNTHESIS
To obtain the set of overall priorities for a decision problem, we have
to pull together or synthesize the judgments made in the pairwise
comparisonsthat is, we have to do sorne weighting and adding to give us
a single number to indicate the priority of each element. The following
example explains how to synthesize.
Suppose we want to decide which of three new cars a Chevrolet, a
Thunderbird, and a Lincolnta buyon the basis of comfort . We draw a
matrix with the criterion "comfort" listed in the upper lefthand corner and
the cars listed in the column on the left and in a row on top (Figure 52) . We
Comfort c T L
Chevro/et (C) 1 112 1/4
Thunderbird (T) 2 1 112
l.i 11 co/n ( L) 4 2
Figure 52 Simple Matrix Comparing Three Cars for Comfort
then put ls in the diagonal positions as indicated. This matrix has nine
entries to fiJI . Three are already committed to ls. Three of the remaining six
are reciprocals. This leaves three jud_gments to make. ln ,J, if the
<Xl
0
i
1
l:
l:
'
~
j .
ill.'
:
.H
',1
'l 
ii
~  .
~
... .         ~    ___________ ....,........, ________________ _
EST ABLISHJNG PRJORITIES
matrix deals with, say, seven elements, the number of judgments needed
to fiJI the entries is 7 x 7  7 + 2 = 21. We subtract the seven unit entries
down the diagonal and divide by 2 because half the judgments are recip
rocals th at are entered automatically.
We then ask: How much more comfortable is an average new Chev
rolet than an average new Thunderbird and an average new Lincoln?
Based on our experience and persona! preference, our judgment is that a
Chevrolet is onehalf as comfortable as a Thunderbird and onefourth as
comfortable as a Lincoln . To state these judgments in terms of the quan
tifiers in the scale (Table 51), a Thunderbird is slightly more comfortable
than a Chevrolet, and a Lincoln is between slightly and strongly more
comfortable than a Chevrolet. Thus we enter the values 2 for the Thunder
bird over the Chevrolet and 4 for the Lincoln over the Chevrolet. These
num bers are the reciprocals of the two judgments comparing the Chevrolet
with the other cars.
Remember that the element that appears in the lefthand column is
always compared with the element appearing in the top row, and the value
is given to the element in the column as it is compared with the element in
the row. If it is regarded less favorably, the judgment is a fraction . The
reciprocal value is entered in the position where the second element, when
it appears in the column, is compared with the first element when it ap
pears in the row.
In this example, because the Chevrolet is regarded less favorably when
compared with the other two cars, we enter 112 and 1/4 in the second and
third positions of the first row and enter 2 and 4 in what are known as the
transp ose positions in the first column. We then compare the Thunderbird
with the Lincoln and enter a value of 1/2 in the second row, third column
position, and its reciprocal 2 in the second column, third row position. We
now have the three judgments needed to complete the pirwise compari
son matrix (Figure 52).
Next we want to synthesize our judgments to get an overall es ti mate of
the relative priorities of these cars with respect to comfort. To do so, we
first add the values in each column (Figure 53) . Then we di vide each entry
SYNTHESIS 81
in each column by the total of thal column to obtain the normalized matrix,
which permits meaningful comparison among elements (Figure 5  ~ ) . Fi
Comfort 1 c T L

c 117 117 1/7
T 2/7 217 217
L 417 417 417
Figure 54 Normaliz:ed Matrix
nally, we average over the rows by adding the values in each row of the
normalized matrix and dividing the rows by the number of en tries in each:
1/7 + 117 + 1/7 = 117 = 0. 14
3
2/7 + 217 + 217 = 217 = 0.29
3
4/7 + 417 + 417 = 4/7 = 0.57
This synthesis yields the percentages of overall relative priorities, or pref
erences, for the Chevrolet, the Thunderbird, and the Lincoln: 14, 29, and 57
percent, respectively. As far as comfort is concemed, the Thunderbird and
the Lincoln are thus about twice and four times more preferable than the
Chevrolet.
The answer in this case was very simple, because ali the columns in the
normalized matrix were the same. They turned out to be the same because
the pairwise comparison matrix (Figure 52) was consistent . That is, from
the relationship of the Chevrolet to the Thunderbird in the first row of the
matrix,
Comfort 1 C T L 1 . C = (1/2)T
 " 1
c 1 1/2 1!4 . ~ . . . . .
and from 1ts relatlonsh1p w1th the Lmcoln,
T 2 1 112
L 4 2 1 :S... C = (1/4)L
Columrr
total 7 3.5 1.75 ' we can deduce thal
Figure 53 Synthesizing the Judgments . (li2)T = (1/4)L
co
~
:1
1
1
i
l
1
l
\
1
t
Il
lj
!.
li
i:
ij
l.
ii
"
1
,,
l '
:1
.,
~ 1
:;
: J
;;
d
q
i'
\;
H
li
ij
1.
(
'
'
~ .,
1
:j
'
l
!
1
!
... ..  ..... ... __ .. ___   ... 
ESTABL.ISIIING PRIORITIES
and that
T = (1/2)L
which is precisely what we have in the second row, third column entry. In
other words, if the Chevrolet is preferred half as much as the Thunderbird
and onefourth as much as the Lincoln, then the Thunderbird must be
preferred half as much as the Lincoln. The information in the first row is
used to force judgmental consistency.
CONSISTENCY
ln decision making problems it may be important to know how good
our consistency is, because we may not want the decision to be based on
judgments that have such low consistency that they appear to be random.
On the other hand, perfect consistency is hard to live up to. Our judgments
on the relative comfort of the three cars were consistent, but in real !ife
specifie circumstances often influence preferences, and circumstances
change.
If apples are preferred to oranges, for example, and oranges are pre
ferred to bananas, then in a perfectly consistent relationship apples must
be preferred to bananas. But the same individual may sometimes like
bananas better than apples, depending on the time of day, the season, and
other circumstances. In the example of the cars, we identified a couple of
relationships that showed the strength of our preference for a Thunderbird
over a Chevrolet and for a Lincoln over a Chevrolet and fNced these rela
tionships on the comparison between a Thunderbird and a Lincolnthe
Thunderbird was preferred half as much as the Lincoln . But often such a
relationship does not hold true. Violating it, which we do ali the time,
leads to inconsistency.
How damaging is inconsistency? Usually we cannot be so certain of
our judgments that we would insist on forcing consistency in the pairwise
comparison matrix. Rather, we guess our feelings or judgments in ali the
positions except the diagonal ones (which are always 1), force the recipro
cals in the transpose positions, and look for an answer. We may not be
perfectly consistent, but that is the way we tend to work. (lt is also the way
we grow. When we integrale new experiences into our consciousness,
previous relationships may change and sorne consistency is !ost. As long as
there is enough consistency to maintain coherence among the objects of
our experience, the consistency need not be perfect.) lt is useful to
remember that most new ideas that affect our lives tend to cause us to
rearrange sorne of our preferences, thus making us inconsistent with our
previous commitments. If we were to program ourselves never to change
.ttft'
CONSISTENCY 83
our minds, we would be afraid to accept new ideas. Ali knowledge has to
be admitted into our narrow corridor between tolerable inconsistency and
perfect consistency.
Of course, a certain degree of consistency in setting priorities for ele
ments or activities with respect to sorne criterion is necessary to get valid
results in the real world. The AHP measures the overall consistency of
judgments by means of a consistency ratio. The value of the consistency
ratio should be 10 percent or less. If it is more than 10 percent, the judg
ments may be somewhat random and should perhaps be revised. Let us
continue with the example of the cars and see how the AHP measures
consiste ney.
Suppose that we keep the first row of our pairwise comparison matrix
in Figure 52 but do not pay much attention to consistency with our previ
ous judgments. In comparing the Thunderbird with the Lincoln, we enter
the value 1/4 in the second row, third column, and enter its reciprocal 4 in
the third r.ow, second column (Figure 55) . Following the steps described
Comfort 1 c T L

c 1 1/2 114
T 2 1 114
L 4 4
Co/um11
lolr1/ 1 7 5.5 1.5
Figure 55 lnconsistent Matrix
earlier, we obtain the normalized matrix, its row sums, and the percent
ages of relative overall priorities (Figure 56) . The percentages, 13, 21, and
66 percent, constitute the priority vector of the three cars with respect to
comfort. The value of the priority vector is approximate. (We can find the
Avrragr
Comfort c T L Row Sums Row Sum
c 1/7 1/11 1/6 0.40 0.4013 = 0. 13
T 2/7 2/11 116 0.63 0.6313 = 0 .21
L 4/7 8/11 4/6 1.97 1.9713 = 0 .66
Figure 56 Normalized Matrix, Row Sums, and Overall Pri r ' ies
(X)
~
N
ESTABLISIIING PRIORITIES
exact value, but the solution is complicated. Besicles, when the judgments
are perfectly consistent, the two values are identical; nearly consislent, the
values are close.) Although the standing of the Chevrolet has not been
changed by much, the other two have changed by our reducing the value
for the Thunderbird and raising it for the Lincoln.
With inconsistency ali the values are changed. The question is: How
significant is this change? Presumably we want to compare our inconsis
tency with the value it would have if the judgments were random. To do
this, multiply the first column of the inconsistent matrix (Figure 55),
changed to decimal form, by the relative priority of the Chevrolet (0.13),
the second column by that of the Thunderbird (0.21), and the third column
by that of the Lincoln (0.66). Then total the entries in the rows (Figure 57) .
Row
Comfort c (0 . 13) T (0.21) L (0.66) Comfort J c T L Total
 
c 1 0.5 0.25 c 0.13 0. 11 0. 17 0.41
T 2 1 0.25 1' 0.26 0.21 0.17 0.64
L 4 4 1 L 0.52 0.84 0.66 2.02
Figure 57 Totaling the Entries
Now take the column of row totals and divide each of its entries by the
corresponding entry from the priority vector (figure 58). We can now find
the average of the three entries in the last column of Figure 58:
3.15 + + 3.06 = 9.;6 "" 3.09
By convention, the symbol for this number is max (lambda max) . The
consistency index (Cl) is
3.09  3 = = 0.045
The random value of the CI for n = 3 is 0.58. The consistency ratio (CR) is
0.045/0.58 = 0.08, which indicates good consistency.
If numerical judgments were taken at random from the scale 119, 1/8, 117, . . , 112, ... , 1,
2, . . . 9, then using a reciprocal matrix we would have the following average consistencies
for di fferent order random matrices:
Size of matrix
Rand<'m consistency
1 2 3 4 5 6 7 8 9 10
0.00 0.00 0.58 o. 90 1.12 1.24 1.32 1.41 1.45 1.49
EXTENDING THE PROCESS oS
[
0.41] .
0.64 ..,.
2.02
[
0.13] [3.15]
0.21  3.05
0.66 3.06
Figure 58 Determining max
A second approximation procedure is to compute the geometrie mean
of the elements in each rowthat is, to multiply the elements and then
take the ir n th root. This step is followed by normalizing the resulting
vector so that its components add to unity. ln general, the geometrie mean
is a good approximation, particularly when the consistency is high . The
calculation of max can proceed as before. The geometrie mean for the
inconsistent matrix of cars with respect to comfort yields 0. 16, 0.20, and
0.64. The exact solution by computer is 0. 13, 0.21, 0.66, and max = 3.05,
which nearly concides with the results of the column normalization pro
cess described earlier. Note that row averaging followed by a normali za
tion of the resulting vector yields 0. 13, 0.23, and 0.64 . There are many
for which this last process yields unsatisfactory results when the
matrix is inconsistent.
One way to improve consistency when it turns out to be unsatisfactory
is to rank the activities by a simple order based on the weights obtained in
the first run of the problem. A second pairwise comparison matrix is then
developed with this knowledge of ranking in mind. The consistency
should generally be better.
EXTENDING THE PROCESS
To see how the process just described can be extended to an entire
hierarchy, consider the problem of a woman who has recently earned her
Ph.D. and is being interviewed for three jobs. Which one should she
choose? Figure 59 shows how she structured the elements of the problem
and arranged them in a hierarchy. Leve) 1, the focus, is overall job satisfac
tion; leve! 2 comprises the criteria that contribute to job satisfaction; and
leve! 3 consists of the three job possibilities. The hierarchy is a complete
one: Each element in a leve) is evaluated in terms of ali the elements in the
next higher leve!.
The woman compared the leve) 2 criteria in pairs with respect to job
satisfaction and judged the relative importance of each criterion. She felt
that research, for example, would be equally as important as loca tion in
contributing to job satisfaction, but it would be slightly to strongly more
important than colleagues. Figure 510 shows the pairwise comparison
matrix of the criteria with respect to the focus . The last column gives the
ro
w
1
:l
' J
:!
j
j
:i
l
,.. _ ... .. .. .. ...._ .. ,.. ...... ...._ . . .. iloo ....... ... , r 1 a rtra._
. : ESTABLISHING PRI ORITIES
Level1:
Focus
Overall Job Satisfaction
Level2:
Criteria
Bene lit Location Repulallon Re se arch Grow1h Colleagues
Level3:
Jobs
B c A
Figure S9 Hierarchy for Choosing Among Three Job Qffers
priorities: Reputation tums out to be the most important criterion, fol
lowed by opportunities for growth and benefits .
Next she developed six matrices for comparing the three jobs with
respect to each criterion (Figure 511). Ali three entries in the vector of
priories obtained in each of the six matrices and listed in the las! column
of each are multiplied (weighted) by the priority of the corresponding
criterion. These values are shown in Figure 512. The results of this opera
tion are then added to yield the overall priori lies for the jobs: A= 0.40, B =
0.34, and C = 0.26. The differences made apparent by this synthesis were
sufficiently large for the woman to accept the offer for job A.
Although we will not go into the measurement of consistency here, it
is important to note thal in a hierarchy, the highestlevel elements usually
have the highest priorities. Inconsistency arising from comparison with
respect to these elements is very damaging because of their high priority.
The consistency index of a hierarchy is obtained by multiplying the consis
tency index of each matrix. by the priority of the criterion used for the
comparison and adding ali such quanti lies . To evaluate the consistency of a
hierarchy, compare the consistency index of the hierarchy with ils coun
terpart when the consistency indices of the matrices are replaced by aver
age random judgment consistency indices for matrices of the same size.
INTERDEPENDENCE
So far we have considered how to establish the priority of elements in
a hierarchy and how to obtain the set of overall priorities when the ele
ments of each leve! are independent . But often the elements are interde
pendent. How do we account for overlapping areas, or commonalities,
among such elements? There are t_wo principal kinds of interdependence
,.,
... 
0
 0
0
r;:
.8
::s
"
..
c.:::
r;:
0
.z::
3
0
....J
le
::s
bo
"'
..
u
"'
't
E
<rl
_.,
fj
::
Cl
_.,
0
...
"'
..
"'
....
c.:::
..Q r;:
0 0
 .:::
 0
t; . !(1
p 
INTERDEPENDENCE
....... 9"""'4 0 ..4
ci ci ci ci ci ci
1:::! 1:::! 1:::!
9"""'4 9"""'4,......! ..4
..,.... ,...... M s .... ,..,
.:t' "'tt' 1.1') ,...... M M
.4 N ... !!2 N
........
.........
........
,.... N
.....4 9"""'4 9"""'4 ;! ,...... N
"' "'
..s:: 1:
:: :: .9 .!!
" ... E o o "
c.::: Cl <rl u ....J c.:::
.0
0

ii
5
:e
.:!
"'
V'l
]
5
0
...
J..
"'
87
<Xl
_p.
q
h
li
i
FlJ AC . . Sj
::_. .
Vector of Vector of
8enefits 1 Research 1 A 8 c Priorities Growth A 8 c Priori ti es
A 1 1/4 112 0.14 A 1 114 115 0.10 A
8 4 1 3 0.63 B 4 1 1/2 0.33 B
c 2 1/3 1 0.24 c 5 2 1 057 c
Vector of
Location 1
Vector of
Reputa 1
Col/eagues 1 A B c Priorities A B c Priorities lion
A 1 113 5 0.28 A 1 1 7 0.47 A
B 3 1 7 0.65 B 1 1 7 0.47 B
c 115 1f7 1 0.07 c 117 117 1 0.07 c
Figure 511 Six Matrices for Comparing Three Jobs
Research Growth
8enelits Col/eagues Location
(0. 16) (0.19)
(0.19) (0.05) (0.12)
A 0. 14 {0. 16)
+ 0.10 {0. 19)
+ 0.32 (0.19)
+ 0.28 (0.05) + 0.47 {0.12)
+
8 0.63 (0.16)
+ 0.33 (0. 19)
+ 0.22 (0.19)
+ 0.65 (0.05)
+ 0.47 (0.12)
+
c 0.24 {0. 16)
+ 0.57 (0.19)
+ 0.46 {0.19)
+ 0.07 {0.05)
+ 0.07 (0.12)
+
Figure 512
Determining the Overall Priorities
SV' 8
A 8 c
1 3 113
1/3 1 1
3 1 1
A 8 c
1 7 9
117 1 5
1/9 115 1
Reputation
(0 .30)
0.77 (030) =
0.17 {0.30) =
0.05 {0.30) =
Vector of
Priorities
0.32
0.22
0.46
Vector of
Priori ti es
0.77
0.17
0.05
Vector of
Ovnall
Priori ti es
0.40
0.34
0.26
m
CJl
!
>
cc
r
Vi
::::
z
C"l
"0
;:1::1
;:1::1
::i
m
CJl
;:1::1
0
m
"'0
m
z
0
m
z
()
m
00
o
1
. . _..:_ . ___ ; .;.._ :..:..:,: ..... :...
0    _: _ _ _ _ __ _____ _ ..  . ..,... . . ii'cz=n '
,.......ol .. ....... .....
ESTAII! .ISIIINC I'RI O RITI ES
among elements of a hierarchy leve!: additive interdependence and syner
gistic interdependence.
Additive lnterdependence
ln additive interdependence, each element contributes a share that is
uniquely its own and also contributes indirectly by overlapping or in
teracting with other elements. The total impact can be estimated by exam
ining the impacts of the independent and the overlapping shares and then
combining the impacts. The effects of this simpler type of interdependence
can be computed precisely since we can usually tell how much of an ele
ment's contribution is due toits independent properties and how much to
ils effect on other elements. 6oth mechanization and farm size con tri bute to
agricultural productivity, for example, but mechanization also influences
farm size by enabling farmers to work larger areas . ln practice, most people
prefer to ignore the rather complex mathematical adjustment for additive
interdependence and simply rely on their own judgment. ln the example
just given, mechanization would be assigned a priority than farm
size. The precise value would be determined subjectively from the descrip
tions in the pairwise comparison scale. Such judgments can also replace
the more technical adjustment for synergistic interdependence.
Synergistic lnterdependence
ln synergistic interdependence, the impact of the interaction of the
elements is greater than the sum of the impacts of the elements, with due
consideration given to their overlap. This type of interdependence occurs
more frequently in practice than additive interdependence and amounts to .
crea ting a new entity for each interaction. Power coalitions and the biology
of marriage are examples of synergy.
The analytic hierarchy process provides a simple and direct means for
measuring interdependence in a hierarchy. The basic idea is thal wherever
there is interdependence, each criterion becomes an objective and ali the
criteria are compared according to their contributions to thal criterion.
This generales a set of dependence priorities indicating the relative de
rendence 0f each criterion on ali the criteria. These priorities are then
weighted by the independence priority of each related criterion obtained
from the hierarchy and the results are summed over each row, thus yield
ing the interdependence weights . By way of validation we find that this
approach is compatible with, for example, what econometricians do in
calculating inputoutput matrices . These ideas are illustrated in the math
emahcal supplement.
With synergistic interdependence, one needs to introduce for evalua
PERSPECTIVE 91
lion additional criteria thal reveal the nature of the interaction. The over
lapping element should be separated from its constituent parts. Ils impact
is added to theirs at the end to obtain their overall impact. Synergy of
interaction is also captured at the upper levels when clusters are compared
according to their importance. Interactions within and between clusters
are better seen and judged higher up in the hierarchy.
Much of the problem of synergistic interdependence arises from the
fuzziness of words and even the underlying ideas they represent . The full
potential of interaction is never fully understood until il has taken place in
practice. The qualities thal emerge cannot be captured by a mathematical
process. Thus a set theoretical approach using Venn diagrams would not
be useful, because we do not have a simple geometrie overlap of regions.
What we have instead is the overlap of elements with other elements to
produce an element with new properties thal are not discernible in ils
parent parts.
Note that if we increase the elements being compared by one more
element and attempt to preserve the consistency of the.ir earlier ranking,
we must be careful how we make the comparisons with the new element.
For example, if we have been comparing apples, oranges, and bananas
with respect to taste and then add melons to the problem, this new addi
tion may change the priorities . Once we compare one of the previous
elements with a new one, ail other relationships are automatically set ;
otherwise there would be inconsistency and the rank order might change.
PERSPECTIVE
The process of setting priorities captures the feelings and judgments
of informed individuals by asking them to make pairwise comparisons
among like objects as criteria. The judgments, which represent strength of
preference, are simultaneously converted to numerical values to represent
their intensity and are laid out in a matrix. The priorities are then derived
from ali the judgments and the consistency of the judgments is calculated
through the deviation of a single number from the order of a matrix . ln a
hierarchy we synthesize the priorities in a leve! by weighting the priorities
derived from each matrix by the weight of the criterion of comparison. To
obtain overall priorities, we add the results for each element . Note thal
prioritization from the top of the hierarchy downward includes less and
less synergy as we move from the larger more interactive clusters to the
small and more independent ones.
lndependence can be treated in two ways. Either the hierarchy is
structured in a way that identifies independent elements or dependence is
allowed for by evaluating in separa te matrices the impact of ali the el e 
ments on each of them with respect to the criterion being considered . As
before, a weighting process is then used to determine their prioriti es .
11
:!,
{
j
1
J
<
1
p
i
l
:l
:!
(X)
_p.
0"
1!
,.
" r,
' j
:
1
;
,,
.,.
1 ,,
J '
1!
:i
:;
i'i
::
;
;
ESTAIJLISHING PRIORITIES
If we already have an idea of the ranking of elements and wish to
derive their priorities, the judgments we give must indicate this domi
nance in rank. Otherwise we must assume that the individual does not
fully understand (or is inconsistent in) his or her subjective ranking.
KEY CONCEPTS
0 To establish the priorities of elements, we have to compare them in
pairs according to a criterion. A matrix is the best framework for this
corn parison.
0 To obtain the set of overall priorities for a decision, we have to synthe
size the results of the pairwise comparisons. That is, we have to com
bine our judgments to get an overall estimate of the relative ank of
priori ti es.
0 ln a hierarchy, the highestlevel elements usually have the highest .
priorities. They are the clusters thal give rise to smaller elements at the
lower levels. Their priorities are divided by the weighting process
among their descendants .
0 The consistency of a hierarchy can be measured by multiplying the
consistency of each matrix by the priority of its criterion and adding.
This result is then compared with a similar number obtained for ran
dom matrices of the same size . The ratio should be 10 percent or less.
Grea ter inconsistency ind ica tes lack of information or lack of under
standing.
6
StepbyStep
Examples of the
Process
This chapter deals with the following questions :
What are the basic steps of the analytic
hierarchy process?
How can the process be used to analyze
highlevel decisions like the one to rescue the
host ages in Iran?
How can the process be used to determine
consumer preference?
How can the process be used to estimate the
economy's impact on sales?
How can the process be used to select a stock
portfolio?
How valid is the analytic hierarchy process?
93
co
~
'.J
1
~
1
(
..
..
i
;,
,_.
' ,;
l1
y
 .... ........   . ....  '"' w .......... a ....
STEPBYSTEP EXAMPLES OF THE I'ROCESS
AN OUTLINE OF THE STEPS
We began our study of the analytic hierarchy process by laying out the
elements of a problem as a hierarchy. We then made paired comparisons
among the elements of a leve! as required by the criteria of the next higher
leve!. These comparisons gave rise to priorities and finally, through syn
thesis, to overall priorilies. We measured consistency and deal! with inter
dependence. These basic steps of the process can ali be condensed into a
brief outline. In broad terms, the process is stable, although certain steps
may be given special emphasis in particular problems and, as noted be
low, repetition is generally necessary.
1. Define the problem and specify the solution desired.
2. Structure the hierarchy from the overall managerial viewpoint (from the
top levels to the leve! at which intervention to solve the problem is
possible).
3. Construct a pairwise comparison matrix of the relevant contribution or
impact of each element on each goveming criterion in the next higher
leve!. In this matrix, pairs of elements are compared with respect to a
criterion in the superior leve!. In comparing Iwo elements most people
prefer to. give a judgment thal indicates the dominance as a whole
number . The matrix has one position to enter thal number and another
to enter ils reciprocal. Thus if one element does not contribute more
than another, the other must contribute more than it. This number is
entered in the appropriate position in the matrix and ils reciprocal is
entered in the other position. An element on the left is by convention
examined regarding ils dominance over an element at the top of the
matrix.
4. Obt21in ali judgments required to develop the set of matrices in step 3. If
there are many people participating, the task for each person can be
made simple by appropriate allocation of effort , which we describe in a
later chapter. Multiple judgments can be synthesized by using their
geometrie mean.
s.
6.
7.
Having collected ali the pairwise comparison data and entered the re
ciprocals together with unit entries down the main diagonal, the priori
lies are obtained and consistency is tested .
Perform steps 3, 4, and 5 for ali levels and clusters in the hierarchy.
Use hierarchical composition (synthesis) to weight the vectors of priori
lies by the weights of the criteria, and lake the sum over ali weighted
priority en tries corresponding to those in the next lower levet and so on.
The result is an overali priorit y vector for the lowest leve! of the hier
archy. If there are severa! outcomes, their arithmetic average may be
ta kP"' .
$ ... .. .rtr""",_,.,,,..... . . , .. ,...,, , ..,...., . , , ,. , , .. , ...
ANALYZING THE HOST AGE RESCUE OPERATION 95
8. Evaluate consistency for the entire hierarchy by multiplying each con
sistency index by the priority of the corresponding criterion and adding
the products. The result is divided by the same type of expression using
the random consistency index corresponding to the dimensions of each
matrix weighted by the priorities as before. The consistency ratio of the
hierarchy should be 10 percent or less. If it is not, the quality of informa 
tion should be improvedperhaps by revising the manner in which
questions are posed to make the pairwise comparisons. If this measure
fails to improve consistency, it is likely thal the problem has not been
accurately structuredthat is, similar elements have not been grouped
under a meaningful criterion. A return to step 2 is then required, al 
though only the problematic parts of the hierarchy may need revision .
These are the basic steps we will follow in working out the examples of
this chapter. In each case we will examine the problem, set up the hierar
chy, carry out the pairwise comparisons, determine the priorities, synthe
size the overali priorities, and examine the consistency. We begin with a
recent application of the method to the Iran rescue operation . An intangi
.ble factor played an important role in President Carter's thinking but not
necessarily the thinking of his advisors hence their different conclusions
regarding the rescue operation.
The second example challenges us to select the best product to man 
ufacture from three alternatives by using six criteria . The product that is
chosen costs but is more desirable ali around. The third exampl e
shows how the AHP may be used to estimate the impact of energy, reces
sion, and inflation on a company's sales. Here percentage ranges are
prioritized according to likelihood of occurrence. The fourth example ill us
traies how the AHP can be used to set priorities for stocks according to
severa! criteria. By way of validation, the highpriority stocks did in fa ct
appreciate in value.
ANALYZING THE HOSTAGE RESCUE OPERATION
It is customary for highlevel managers and political leaders to make
their most important decisions by depending on expert recommendations
governed by their own persona! judgment and understanding. President
Carter once said thal when he had to make a really crucial decisi on be
tween two alternatives, his experts were usually evenly spli t. ln essence ,
then, the hard decisions were still up to him .. If expert opini ons were
accurate and thorough, leaders would become superfluous. In splt e of
splitting on their judgments the experts do raise questi ons thal stimul ate
decision makers and draw their attention to issues they may have ne
glected to consider.
1
i
1
1
j
(X)
(X)
l
...  '            ~    ~           .....
STEPBYSTEP EXAMPLES OF THE PROCESS
Background of the Decision
A great shroud of secrecy surrounds the decision to rescue the hos
tages in Iran. On 28 April 1980 a decision was made to send an American
air rescue team to Iran to bring out the fiftythree American hostages from
Teheran where they had been held since early November 1979. The mission
was a complicated plan involving troops, airplanes, helicopters, a long
flight, .a landing in the desert, a journey to Teheran, taking out the hos
tages, and returning to safety.
It appears that not even the upperleve! national security staff was
informed about how the decision was made. It is said thal President Carter
had the Joint Chiefs of Staff draw up a couple of plans with different
options. He then discussed these plans with his dosest advisors. According
to Time (5 May 1980) these advisors were Brown, Brzezinski, Christopher,
Jordan, Mondale, Turner, and Vance. Carter asked for clarification of sorne
1
of the ideas, but he made the decision on his own. There is no indication
that the president resorted to sorne kind of voting in the discussion on
whether the operation should or should not be undertaken. It has been said
thal this was a typical Carterstyle decision in which he asked his advisors
for details and then made the decision himself.
There is no question that the likelihood of success of a rescue opera
tion is a compelling factor in deciding whether to go or not to go. Of
course, it is crucial how one defines success for such a mission. High
success means no deaths among the hostages or military personnel; me
dium success means a few military and no hostage deaths; low success
means a few military and a few hostage deaths. Greater !osses could mean
anything from failure to disaster. Sorne experts in the Department of De
fense who are familiar with that mission have indicated that it would be
generous to assign a medium chance of success to the military operation.
They also said thal this was known prior to 28 April and is not a question
of hindsight. Time writes: "Carter himself conceded that 'the operation
was certain to be difficult and it was certain to be dangerous.' He insisted
th at the operation had ' an excellent chance of success. "'
As we will see, the president's decision was consistent with his per
ception of the situation as it affected him; but it was not necessarily a good
decision for the nation. Suppose we do a sensitivity analysis of the subjec
tive factors and the emphasis the president placed on them.
Analysis of the Decision
The decision problem can be divided into two parts. The first partis to
identify the best military option among those available and evaluate its
likelihood of success . A military option would also have to be examined by
expert s on foreign relations and intelligence. The second partis the process
ANALYZING THE HOST AGE RESCUE OPERATION 97
of making the go/nogo decision based on the body of knowledge provided
by the experts . We now examine these two parts in detail.
The likelihood of success was determined by military experts. They
considered such factors as:
Transferability: to the desert, to Teheran, then to the embassy
Rounding up: getting inside the compound, creating diversion, locat ing
the hostages
Rescue: subduing the captors, transfer to aircraft, departure (avoiding
Iranian forces)
It is not my intention here to analyze the possible details of their delibera
tions. In this instance the importance of the military factors and the likeli
hood of the mission's success were probably much more important than
the reaction of our allies or the Russians. These factors do appear in the
hierarchy for the problem, but one could have construed the medium
likelihood of success from other sources. In fact, an analysis of the first part
did produce a r11edium likelihood of success, but a Pentagon expert, when
asked, strongly supported these findings. Thus we may assume that a
medium likelihood of success was presented to the president, who then
made his go/nogo decision. How he did this is what 1 wish to dwell upon .
The hierarchie structure of both parts of the pro cess is shown in Figure
61. We find that the main factors that could have played an important role
in President's Carter's mind are:
Hostages' lives: the president's concem for the safe return of ali fifty 
three hostages
Carter's politicallife: the president's concern about the influence of the
decision and his chances for reelection
Military costs : The president's concem about the loss of soldiers' lives in
the operation
United States prestige: the president's concern about the influence of
the decision on relations with foreign states and their subsequent image
of the United States
These factors differ in their impact on the president's decision . More
over, their relative importance changes as the likelihood of success is
changed. Now let us assess their priorities based on a medium likelihood
of success of the rescue operation. We can carry out a pairwise comparison
process by using the pairwise comparison scale (Table S1) . Figure 62
shows, for example, that "hostages' lives" has strong dominance over
"military costs" in the president's mind. Thus we assign the value 5 in the
first row and third column position and its reciprocal value in the first
column and third row position. We always compare the row factors on the
left over the column factors on the top.
STEPBYSTEP EXAMPLES OF THE PROCESS
<_u.
= "2 ll .>
111 "'
cal
.= <
::>UJct
/(_
3 (
t:.
:8 i ::; 0
1
0
z
1
I <
1
i .. (
ns
l
1
,,
f
t:!
I
t
t
:::>
UJ
8'
:g
0
'g u
1
8
f.
a:
'o
Qi
1
::;
(
E _
<!)
t: "' 3
1110
ua..
 l
\
0
<!)
< " H
i.
0 v ::io
.=t:.
v
Qi 0
:I:
II
f
c::
.9
0
Q
()
1
Q
z
0
()
u
=
....
;.,
5
..
s
...
x
...
.b
u
i "
...
Ilot
ANALYZING THE HOSTAGE RESCUE OPERATION 99
2 3 4 Priori lies
1. Hostages' lives 1 113 5 113 0. 15
2. Carter's political !ife 3 1 7 4 0.54
3. Military costs 115 117 1 116 0.05
4. U .S. prestige 3 1/4 6 1 0.26
Figure 62 Relative Priorities of the Factors
As can be seen from the column of priorities in Figure 62, the two
main factors are "Carter's political !ife" and "United States prestige."
" United States prestige" is important because the United States had to do
something to assert its power despite the fact that there was a medium
chance for success. "Carter's political !ife," a subjective factor, is the
highestpriority factor.
Under each factor we consider this question: Which alternative (go or
nogo) is more favorable considering just that factor under a medium
likelihood of success? The results are shown in Figures 63 to 66.
Under "hostages' lives" the decision was evenly split . The rationale
for this judgment is that there was no immediate danger to the lives of the
hostages. With respect to "Carter's political life," we note that the presi
dent was strongly in favor of performing the operation at that time; a few
months before the election the polis were not predicting a good chance for
his reelection. A successful operation would have resulted in a strong
influence on Carter's public image. Failure would be painful to him, but it
Host ages' Prior
Liu es G N ities
Go 1 1 0.5
Nogo 1 1 0.5
Figure 63 Priority of Hostages' Lives
Carter's Pol  1 Prior
itical Life
Go
Nogo
G
1
1/3
N
3
ities
0.75
0.25
Figure 64 Priority of Carter's Political Life
00
V1
0
1
_..,.... ..  ... .  ~ .. . "
STEPBYSTEP EXAMI'LES OF TllE PROCESS
Military Prior
Costs G N ities
Go 1 1/7 0.125
Nogo 7 1 0.895
Figure 65 Priority of Military Costs
Prior
U.S. Prestige 1 G N ities

Go
1
1 4 0.8
Nogo 1/4 1 0.2
Figure 66 Priority of U.S. Prestige
would not hurt him as much as a success would benefit him. The influ
ences of "military costs" and "United States prestige" on the decision are
apparent . Composing the factors' weights with those of the go/nogo alter
natives under each factor gives these results :
Go: 69 Nogo: 31
lt is very easy to appear smart after the events. Yet 1 am convinced that
if the AHP had been applied bef ore the decision was made, we would have
obtained similar results. Seven people established the priorities presented
here. The variance in their judgments was small. lt is possible to argue that
if those seven people had set the priorities individually, they would .have
arrived at the same decision.
lt is clear that "Carter's political !ife" dominated that decision. Sen
sitivity analysis shows that if the 75 percent in favor of go under that factor
is changed to 38 percent, the outcome for go/nogo would have been even.
I believe that 75 percent is a modes! estimate. Carter may not have gone
through with the decision to go if his desire to be reelected approximately
matched his concern about the hostages.
To pursue the analysis further, my colleagues and I examined the
outcome of the decision under a "low likelihood of success" recommenda
tion by the experts. (Presumably this would have been interpreted less
optimistically by the president . Recall that he interpreted medium success
as " excellent.") Our results are shown in Table 61.
ANALYZING THE HOST AGE RESCUE OPERATION
Table 61 Relative Priorities Given a Low
Likelihood of Success
Priori/y Given Priori/y Given
Medium Likelihood Low Like/ihood
Factor of Success of Success
Hostages' lives 0.15 0.35
Carter's political life 0.54 0.39
Military costs 0.05 0. 10
U.S. prestige 0.26 0.16
101
In this case the "hostages' lives" become a more important issue and
"Carter's politicallife" less important. The influence of these factors on the
go/nogo decision is shown in Table 62.
The outcome of the decision under a "low likelihood of success" rec
ommendation by the experts leads to the following results:
Go: 0.41 Nogo: 0.59
The change from the previous result is due to a greater emphasis on the
hostages' lives. Only if the hostages' lives were clearly in jeopardy in Iran
would one have decided in favor of go. Moreover, "Carter's political life"
would have had to be less important because a military operation with a
low chance of success would have had a low chance of helping Carter.
Now let us pull together the observations we have drawn from this
analysis of the Iran rescue operation. For President Carter, the subjective
factornamely, his concern with his careeraccounted for 54 percent of
the total. Perhaps the dominance of this subjective factor was perceived by
Table 62 Influence of Factors on the
Go/NoGo Decision
Ml'dium Likelihood Law Likelihood
Factor Go NoGo Go NoGo
Hostages' lives 0.50 0.50 0.20 0.80
Carter' s political life 0.75 0.25 0.75 0 25
Military costs 0. 125 0.875 0.10 0.90
U.S. prestige 0.80 0.20 0.25 0.75
cc
lJ1
.....
i
1
!
j
'1
il
~
il
.1
:1
l\
i'
Il
~ =
i\
.1
'1
l'
~
, l ~
~ 
;,
1
'
;
. ~
~
Il
1
STEPBYSTEP EXAMPLES OF THE PROCESS
Carter's aides; perhaps it accounts for Secretary of State Vance's resigna
tion. Certainly "Carter's political !ife" would not have figured as impor
tantly in Vance's analysis of the situation. He would have had to decide
against the operation. Of course he may have had other political reasons of
his own.
This application of the AHP illustrates its effectiveness in analyzing
highlevel decisions. There is a value to analysis even after a decision has
been made because the method is an efficient tool for deriving a lesson
from previous mistakes. ln this example the results were consistent with
reality. ln sorne situations, as we have seen, media information may be
. sufficient to indicate what went into a decision. If used by decision makers,
the process can sharpen thinking and reveal subjective factors. lt can
show, for example, that persona! interest is weighted much higher than
what is good for the business. lt can also show that there is not a clear
commitment to the important objectives.
DETERMINING CONSUMER PREFERENCE
The AHP is basically a simple, efficient technique for problem solving.
The following stepbystep example demonstrates this simplicity; il can
also serve as a mode! for using the process to solve other problems.
A firm wants to determine consumer preferences for three different
kinds of paper towel. The attributes considered most relevant from the
consumer's perspective are (1) softness, (2) absorptiveness, (3) priee, (4)
size, (5) design, and (6) integrity. The three kinds of paper towels, X, Y, and
Z, possess ali these attributes, but at different levels of intensity: high (H),
medium (M), and low (L). Given the consumer' s "bounded rationality"
that is, the fact that consumers do not act on perfect or complete informa
tion and are satisfied with less than the economically most rational
choicewe can best distinguish among the attributes by dividing them
into this small number of intensity categories. The resulting hierarchy is
shown in Figure 67. The problem of selecting the product with the
greatest overall consumer preference is solved in the following manner:
Step 1: Determine consumer preference among the attributes by devel 
oping a matrix that compares attributes in pairs with res.pect to product
desirability (Figure 68).
Step 2: Determine consumer preference among the intensities of the
attributes by developing six matrices that compare intensity levels in
\c pairs with respect to each attribute (Figure 69) .
":Ji
Now we want to synthesize these judgments to obtain the set of over
all priorities that will indicate which product consumers prefer. The re
maining steps lake us through this process:
~
15
~
<ii
<1>
0
u
::J
~
a.
DETERMINING CONSUMER PREFERENCE
"
v
c
"
..
"
....
~
Cl..
..
"
<Il
c
0
u
co
c
'
'
~
"
8
Cl
'1:: ..
a. 0
....
>.
= v
..
..
..
. ~
::r:
"'
1
\0
"
~
co
iZ
103 '
co
U1
N
1
1
i
!
l
]
~
STEPBYSTEP EXAMI'LES OF THE PROCESS
Attributes
Product
Desirability s A
p
SI D 1
s 1 1/4 1/5 1/4 5 116
A 4 1 113 3 6 1/2
p 5 3 1 4 7 3
SI 4 1/3 114 1 5 115
D 1/5 116 117 115 1 117
6 2 1/3 5 7 1
>...,,, = 6.66; Cl= 0.12.
Figure 68 Matrix Comparing Attributes
Absorp
Soft ness H M L Priority tiueness H M
H 1 5 8 0.7257 H 1 7
M 1/5 1 5 0.2122 M 1/7 1
L 118 115 1 0.0621 L 119 1/7
h"'"' = 3.15; Cl= 0.07. = 3.33; Cl = 0. 16.
l'ricr 1 1/ M L Priorily Size Il M

H 1 117 1/9 0.0480 H 1 3
M 7 1 117 0.1912 M 113 1
L 9 7 1 0.7608 L 1/5 1/4
= 3.33; CI= 0.16.
= 3.09; Cl = 0.04.
Design H M L Priority /ntegrity H M
H 1 1/5 2 0. 1786 H 1 7
M 5 1 5 0.7089 M 117 1
L 1/2 115 1 0. 1125 L 119 117
h,,., = 3.05; Cl = 0.03. h"'"' = 3.33; CI = 0. 16.
Figure 69 Matrices Comparing Intensity Levels
Priority
0.0570
0.1679
0.3837
0.1002
0.0269
0.2643
L Priority
9 0.7608
7 0.1912
1 0.0480
L Priority
5 0.6267
4 0.2797
1 0.0936
L Priority
9 0.7608
7 0.1912
1 0.0480
H
M
L
DETERMINING CONSUMER PREFERENCE 105
Table 63 Priorities of the Attributes
(0.0570) (0.1679) (0,3837) (0.1002) (0.0269) (0.2643)
Soft ness Absorp. Priee Sizt Design /ntegrity
0.7257 0.7608 0.0480 0.6267 0. 1786 0.7.608
0.2122 0.1912 0.1912 0.2797 0.7089 0.1912
0.0621 0.0480 0.7608 0.0936 0. 1125 0.0480
Step 3: Group the priorities of the intensities (H, M, L) for each of the six
attributes in columns and enter the priorities of the attributes, laken
from Figure 68, above the columns (Table 63). Then multiply each
column by the priority of the corresponding attribute to obtain the
weighted vectors of priority for the intensities (Table 64) .
Step 4: Now select from each column the element with the highest
priority to obtain the vector of desired attribute intensities:
HSoft
0.0413
HAbsorp.
0.1278
LPrice HSiu
0.2919 . 0.0628
MDesign
0.0190
Hllltegrity
0.2011
Then add this row and divide each entry by the total to gel the nor
malized vector of desired attribute intensities:
HSoft
0.0556
HAbsorp .
0.1717
LPrice
0.3924
HSize
0.0844
MDesign
0.0256
Fflntegrity
0.2703
Step 5: Determine the perceived product standings by developing ma
trices that compare the three paper towels (X, Y, and Z) in pairs with
respect to the most desired attribute intensities (Figure 610) .
Step 6: Group the priorities of the paper towels with respect t.o each
desired attribute intensity in columns and enter the normalized priori
Table 64 Vectors of Priority for the Intensities
Soft ness Absorp. Priee Size Design
H 0.0413 0. 1278 0.0184 0 0628 0.0048 0.2011
M 0.0121 0.0321 0.0734 0.0280 0.0190 0.0505
L 0.0035 0.0081 0.2919 0.0094 0.0030 0.0127
1.
1
l
,.
1
'
(X)
lJl
w
_. ... o,  ,.    .  ........ ...    . .... v . .. _,h_______  ..............,,.lr"".JO:f ..........,._/,w.oll..............,_f,, , ... . .. _, , ., .... . .. . _, .. ..  
=
0
r:::
Cl.
N
>
><
"'
..<:>
"(
l:
::
0
r:::
Cl.
N
>
><
a ('.. ....
LI") N ..
l::; ;g
0 0 0
('.. <Xl 
N ,..... 22

N !'.
>< > N
  t: "! .
0 0 0
('..LI")
LI")
LI")  ;::.
 !!? t::

:i:
STEPBYSTEP EXAMPLES OF THE PROCESS
0
Il
0
ci
0
,.;
E
<
g;
0
Il
0
ci
 ,.;
.
.
E
<
r::: ::::! li)
0 N (1"')
r::: . . .
Cl. 0 0 0
N 1,........ ,..... .,.....
> 1 N ..< ,....
><
"'
N
Vi
:i:
i5
r:::
Cl.
N
>
><
"' u
r:::
Cl.
...J
!:::!

>< > N
('.. :il ""
0
0 N !'.
0 0 0
Si ;3 
....  li) ;:::;
 ""' ('..
>< > N
I""'
0
0
Il
0
,i)
0
...,
<
0
0
ti
,.;
.
.
E
<
>.
t:
0
r:::
Cl.
N
>
><
= r:::
:r::
:::
0
r:::
Cl.
N
>
><
"'
"'
Cl
('.. "" a
"'  00 "" 00
\0 N 0
0 0 0
c .........
.... ..... ....
;:::;
.....
>< > N
('.. li) <Xl
8 a
.... N
6 0 0
..... "" 
N r' t:?
.....
!:::!

>< > N
li)
0
0
Il
0


""
'
.
E
<
00
.....
0
Il
0
r:;
""!
""
.
<
Ill
:3
Ill
Cil

Cil
= .0
c:
:t:
<
"Cl
;;;
Cil
0
..
.E
Ill
"ii
..
lU
Q..
.,
Q.,
Cil
..c:
(o
:S
bO
.:
.,
Q..
El
0
u
Ill
Cil
v
c:
:;
0
....
1
"'
bO
iZ
x
y
z
x
y
z
(0.0556)
HSoft
0.7147
0.2185
0.0668
ESTIMA TING THE ECONOMY'S IMPACT ON SALES
Table 65 Overall Product Attribute Perception
(0. 1717)
H Absorp.
0.5659
0.3727
0.0614
(0.3924)
LPrice
0.0727
0.2050
0. 7223
(0. 0844)
HSiu
0.4126
0.2599
0.3275
(0.0256 )
MDesign
0.4067
0.3695
0.2238
107
(0.2703)
Hlntegrity
0.6817
0.2363
0.0819
Table 66 Weighted Overall Product Attribute Perception
HSoft
0.0397
0.0121
0.0037
HAbsorp.
0.0972
0.0640
0.0105
LPrice
0.0285
0. 0804
0.2834
HSi u
0.0348
0.0219
0.0277
MDesign
0.0 104
0.0095
0.0057
Hlnt egrity
0. 1842
0.0639
0.0221
lies above the columns (Table 65) . Then multiply each column by the
normalized priority of the corresponding attribute intensity to obtain
the weighted vectors of priority for the desired attri bute intensities for
each paper towel (Table 66).
St ep 7: Add each of the three rows to obtain the overall priorities of the
three paper towels. This synthesis yields the following priorit ies:
x = 0.3949 y= 0.2519
z = 0.3532
From these result s we would select product X as most desirable from the
customer's perspecti ve.
Even though low priee was the desired attribute intensity with the
highest priority, product X, whose priority was very low with respect to
low priee, was the final choice. The reason for this choice is clear: X domi 
nated Y and Z on ali other desired attribute intensities . Thus the firm
decided to market a superior but highprieed product , a decisi on that is
not inconsistent with realworld situations.
ESTIMATING THE ECONOMY'S IMPACT ON SALES
The AHP can be used for certain special applications as weil as for
obtaining sets of priorities in decision problems. In the following examl"'"
l
;
.1
1
1
i
il
1\
~
tl
(
1
'f
:'!
' :
STEPBYSTEP EXAMPLES OF THE PROCESS
Future Sales 1 EC R 1 Priorily

Energy crisis 1 7 1 0.4667
Recession 117 1 117 0.0666
Ill fla lion 1 7 1 0.4667
.\"'"' = 3.0; Cl = 0.00.
Figure 611 Matrix for Comparing Criteria with Respect to Fu
ture Sales
Enl'rgy
Cr isis 0 5 510 10 15 15 20 Pri ority
0 .'i 1 1/5 117 1/5 0.0518
.'i  10 5 1 1/3 1/4 0. 1451
10 15 7 3 1 1/3 0.2904
1 . ~  2 0 5 4 3 1 0.5127
ma ' = 4.337; Cl = 0.11.
Recessio11 1 0 5 510 1015
15 20 ;' Priority

0 5 l 2 5 7 0.5232
S 10 112 1 3 5 0.2976
10 15 i /5 1/3 1 3 0.1222
I S 20 117 1/5 1/3 1 0.0570
.\ """ = 4.069; Cl = 0.02.
lnflatio1 0 5 510 10 15 1520 Priority
05 1 2 5 7 0.5232
5 !Cl 1/2 1 3 5 0.2976
10 15 115 1/3 1 3 0. 1222
IS20 117 1/5 1/3 1 0.0570
.\ .,,." = 4.069; Cl = 0.02.
Figure 612 Matrix for Comparing Sales with Respect to Criteria
SELECTING A PORTFOLIO 109
the process is extended to estimate the percentage of a company's sales
affected by the energy crisis, recession, and inflation .
First the sales of the companya manufacturer of heavy equipment
(oil drills and construction machines)are divided into intervals : 05 per
cent , 510 percent, 1015 percent, and 1520 percent. Then a matrix is
developed to compare the criteria (energy crisis, recession, inflation) in
pairs with respect to future sales (Figure 611). Next matrices are devel
oped to compare sales with respect to the criteria (Figure 612) . The overall
priorities of the sales percentage intervals are
0.3033 0.2264 0.2007 0.2697
To get the expected value of sales affected by the energy crisis, reces
sion, and inflation, multiply the mid point of each interval by the priorit y
weight of that interval. For example, the mid point of 05 percent is 2.5 and
so on. Thus
(2.5 x 0.3033) + (7.5 x 0.2264) + (12.5 x 0.2007) + (17.5 x 0.2697)
= 9.685%
As this example shows, the analytic hierarchy process can be used to
estimate numbers in this case percentages.
SELECTING A PORTFOLIO
This application is more complex than the preceding one. Here we use
the AHP to select a portfolio. Our hierarchical mode! consists of three
separate hierarchies : one based on extrinsic factors, one based on intrinsic
factors, and a third based on the investor's objectives. The firms being
considered are ranked (weighted) relative to the criteria in each hierarchy .
The weights are then combined to get an overall preference list of firms .
Figure 613 gives an overall view of the mode! we will use .
The various factors and objectives that influence the selection of firms
for the portfolio include:
Extrinsic Fact ors (A) : These are the outside factors or environmental
characteristics thal affect an industry' s (or firm' s) performance. The firm,
however, has no direct influence on them. These factors are economie,
political, social , and technical. By incorporating the analysis of the ex
trinsic variables we can determine the sensitivity of a firm to changes in
these factors .
Intri11sic Fact ors (8 ): These are the internai factors or operati on al charac
. teristics of the firm. They may be considered as a measure of the way the
firm is making its decisions or, in general, a measure of the firm' s capac
1
co
(J1
(J1
l
1
l
J
j
..
STEPBYSTEP EXAMPLES OF THE PROCESS
Risk Risk Risk
!
Level Lev el Class
1
A B c
Extdnsic lntrinsic lnvestor's
Factors Factors Objectives
1
A Criteria B Criteria
1
Firms Firms Firms
1
Portfolio
Figure 613 Hierarchical Model for Selecting a Portfolio
ity to compete successfully. These factors are profitability, size, technol
ogy, and philosophy.
Investor's Objectives (C): These are the values that define the actions the
investor undertakes in the business world. We are aware, of course, of
the great variety of objectives an investor may have, but to simplify the
mode! we consider four mutually exclusive objectives: profitability, se
curi ty, excitement, and control.
Since we are dealing with a mode! that is based on future conditions,
we must consider riskthe uncertainty of future events. The mode! incor
porates the uncertainty of the general business environment, the firm's
behavior (high, medium, low risk), and the risk class of the investor (high,
medium, low) .
As our frrst step we look at the hierarchy of extrinsic factors . At the ftrst
SELECfiNG A PORTFOLIO 111
leve! we have the primary extrinsic factors that affect a firm's behavior. At
the second leve! we have the criteria that influence each of the primary
factors:
AI: Primary Extrinsic Factors
Economie
Political
Social
Technological
Ail: Criteria
Employment conditions
Elasticity of demand
Elasticity of supply
International economy
Interest rates
Government regulations
International exposure
Employment conditions
Family disintegration
Age distribution
Educational achievement
Employment conditions
State of technology
Government involvement
The extrinsic. factors are compared pairwise for a highrisk environment, a
mediumrisk environment , and a lowrisk environment (Figure 614). ln
each case we base the comparisons on the questions: Which factor has
more impact on a firm' s behavior and by how much7 We can see that in a
highrisk environment the group constructing the matrix believed the
technology would have a generally strong influence on a firm's behavior
compared to the other factors.
We then compute the priorities for each of the factors for each risk leve!
(Table 67). These priorities point out that in a highrisk environment,
future technological factors have the greatest impact on a firm's behavior.
In a mediumrisk environment both the economie and technological fac
tors are most influential; in a lowrisk environment the social factors are
High Risk
1
Medium Risk
1
Low Risk
Firm E
p
s T E
p
s T E
p
s T
Economie 1 3 4 1/3 1 5 7 1 1 5 1/3 4
Po/itical 113 1 3 115 1/5 1 3 1/5 1/5 1 1/5 2
Social 1/4 1/3 1 118 117 1/3 1 1/6 3 5 1 5
Te ch nological 3 5 8 1 1 5 6 1 1/4 1/2 1/2 1
Figure 614 Pairwise Comparison of Extrinsic Factors
q
'
(X)
lJ1
0'1
.,
'
;j
:
,
1.
( .
1
l'
r
( )
~
Risle
High
Medium
Low
STEPBYSTEP EXAMPLES OF THE PROCESS
Table 67 Priorities for Each Risk Level
Economie Po/itica/
0.25 0. 12
0.43 0. 11
0.30 0.10
State of technology
Goternment irrvolvement
Social
0.06
0.05
0.54
5
1
1/4
G
4
1
Technologicn/
0.57
0.41
0.07
Figure 615 Comparison Matrix for Technology
most important . Economie factors have appreciable impact at ali three
levels whereas the impact of political actions is rather low:
We could carry through our analysis un der each of the three risk levels,
but for illustrative purposes we will use the average weight for each factor.
A weighted average could be used if we wished to favor one type of future
over the other. A veraging, we obtain these weights:
Economie Political
0.32 0.11
Social
0.22
Technological
0.36
Now we compare the criteria for each factor to see their order of impor
tance. In each pairwise comparison we ask: Which criterion has more
impact on the factor and by how much?
ln the case of technology and its two criteria, we have the comparison
matrix shown in Figure 615, which indicates that the state of technology
has somewhat more impact on technology in general than does the gov
ernment's involvement in it . This matrix and three others yield the follow
ing sets of weights for the various criteria:
Employment Elasticity of Elasticity of
Conditions Demand Supply
0.16
Economie (0.32): ~
0.45 0.07
International lnterest
Economy Rates
0.05 0.27
a:
~ l ' ,
SELECTING A PORTFOLIO
Political (0.10) :
Social (0.22) :
Government
Regulations
0.24
Family
Disintegration
0.30
Educational
Achievement
0.11
State of
Technology
Technological (0.36): 0.80
International
Exposure
0.06
Age
Distribution
0.11
Employment
Conditions
0.48
Government
lnvolvement
0.20
Employment
Conditions
0. 70
d3
To get the final weights for a criterion, we must multiply the criterion
weights just found by the weight of the factor associated with that criterion:
Employment
Conditions
0.05
Economie:
International
Economy
0.02
Government
Regulations
Political : 0.02
Fa mil y
Disintegration
0.07
Social :
Educational
Achievement
0.02
Elasticity of
Demand
0.14
Interest
Rates
0.09
International
Exposure
0.01
Age
Distribution
0.02
Employmenl
Conditions
0.11
Elasticity of
Supply
0.02
Employment
Conditions
0.07
CXJ
lJl
']
..  .   _ ..  .. ... .... ...... ..  ...
STEPBYSTEP EXAMPLES OF THE PROCESS
Technological:
State of
Technology
0.29
Government
Jnvolvement
0.07
The weights printed in boldface are the largest ones. We will shorten our
list to include only these. In order to have the weights total1, we divide
each weight by the total of ali the weights in the shortened list. We have
our list of extrinsic criteria:
State of Employment Elasticity of
Tecltnology Cortditions Demand
0.33 0.26 0. 16
Government
lnterest Fa mil y Involvement
Rates Disintegration in Technology
0.10 0.08 0.08
The firms being considered (eight in this illustration) are now
prioritized relative to each of the six extrinsic criteria. This process is the
same pairwise comparison and weighting procedure we have been apply
ing throughout . When comparing the firms pairwise relative to the state of
technology, for example, we ask: Which firm will respond more favorably
to the technological environment of the future? The matrix for this criterion
is sHown in Figure 616.
Technology T IR /C AC R DG 8 c
Tappan 1 1/3 115 4 115 117 6 5
Ing . Rand 3 1 1/4 4 1/4 116 7 6
/ . C. Industries 5 4 1 5 1/2 114 6 5
A/lied Ch . 1/4 1/4 115 1 1/6 117 4 3
Rockwe/1 5 4 2 6 1 1/3 7 6
Data General 7 6 4 7 3 1 9 7
Butler 1/6 117 116 1/4 117 1/9 1 113
Cltemetron 115 116 115 113 116 117 3
Matrix for Pairwise Comparison of Firms Relative to
State of Technology
.. ...  .. ..   .... ..
SELECTING A PORTFOLIO 115
Table 68 summarizes the weights that the eight firms received for
each criterion. A glanee at the table shows that in a technological environ
ment Data General, Rockwell, and I.C. Industries are heavily favored , as
one might expect . Tappan and Rockwell are beneficiaries if employment
conditions are goodTappan's consumer market is in appliances (80 per
cent of sales) and Rockwell's is in home and auto products (35 percent of
sales) .
To get an overall prioritized list of firms visvis the extrinsic fac
tors, we multiply the weights of the firms for a given criterion by the
weight of that criterion (shown in parentheses in Table 68). Then we add
up these new weights for each firm:
Tappan
0.16
lng . Rand
0.07
I. C. Industries
0. 13
A/lied Ch .
0.03
Rockwe/1 Data General Butler Chemetron
0.20 0.21 0.07 0.13
For purposes of illustration we will use only the four highest ones with
their weights adjusted as we did wh
Bien plus que des documents.
Découvrez tout ce que Scribd a à offrir, dont les livres et les livres audio des principaux éditeurs.
Annulez à tout moment.