Chapter 15

Decision analysis and Games

Real-Life Application—Layout Planning of a Computer Integrated

Manufacturing (CIM) Facility
The engineering college in an academic institution wants to establish a computer
integrated manufacturing (CIM) laboratory in a vacated building. The new lab will
serve as a teaching and research facility and as industry center of technical excellence.
Recommendations regarding the ideal and absolute minimum square footage for each
unit are solicited and compiled from the faculty. The study uses both AHP (analytic
hierarchy process) and goal programming to reach a satisfactory compromise solution
that meets the goals for teaching, research, and service to industry. The details of the
study are given in Case 10 in Chapter 26 on the website.

15.1 DeCIsIon MAkIng UnDeR CeRtAInty—AnALytIC HIeRARCHy

PRoCess (AHP)
The LP models presented in Chapters 2 through 9 are examples of decision making
under certainty (all the data are known with certainty). AHP is designed for situations
in which ideas, feelings, and emotions affecting the decision process are quantified to
provide a numeric scale for prioritizing the alternatives.

example 15.1-1 (overall Idea of AHP)

Martin Hans, a bright high school senior, has received full academic scholarships from three
institutions: U of A, U of B, and U of C. Martin bases his choice on two criteria: location and
academic reputation. To him, academic reputation is five times as important as location, and
he assigns a weight of approximately 83% to reputation and 17% to location. He then uses a

567
568 Chapter 15 Decision Analysis and Games

Select a
Decision:
university

Hierarchy 1 Location Reputation

criteria: (.17) (.83)

Alternatives: U of A U of B U of C U of A U of B U of C
(.129) (.277) (.594) (.545) (.273) (.182)

.17 3 .129 1 .83 3 .545 5 .4743 .17 3 .277 1 .83 3 .273 5 .2737 .17 3 .594 1 .83 3 .182 5 .2520
U of A U of B U of C

FiGure 15.1
Summary of AHP calculations for Example 15.1-1

systematic process (which will be detailed later) to rank the three universities from the stand-
point of location and reputation, as the following table shows:

Percent weight estimates for

Criterion U of A U of B U of C

Location 12.9 27.7 59.4

Reputation 54.5 27.3 18.2

The structure of the decision problem is summarized in Figure 15.1. The problem involves
a single hierarchy (level) with two criteria (location and reputation) and three decision alterna-
tives (U of A, U of B, and U of C).
The ranking of each university is based on the following composite weights:
U of A = .17 * .129 + .83 * .545 = .4743
U of B = .17 * .277 + .83 * .273 = .2737
U of C = .17 * .594 + .83 * .182 = .2520
Based on these calculations, Martin chooses U of A because it has the highest composite weight.

remarks. The general structure of AHP may include several hierarchies of crite-
ria. Suppose in Example 15.1-1 that Martin’s twin sister, Jane, was also accepted with
full scholarship to the three universities. The parents insist that the two siblings at-
tend the same university. Figure 15.2 summarizes the decision problem, which now
involves two hierarchies. The values p and q at the first hierarchy are the relative
weights representing Martin’s and Jane’s opinions (presumably equal). The weights
(p1, p2) and (q1, q2) at the second hierarchy, respectively, represent Martin’s and
 15.1 Decision Making under Certainty—Analytic Hierarchy Process (AHP) 569

Select a
Decision:
university

Hierarchy 1
Martin (p) Jane (q)
criteria:

Hierarchy 2 Location (p1) Reputation (p2) Location (q1) Reputation (q2)

criteria:

Alternatives:

U of A U of C U of A U of C U of A U of C U of A U of C
(p11) (p13) (p21) (p23) (q11) (q13) (q21) (q23)

U of B U of B U of B U of B
(p12) (p22) (q12) (q22)

U of A 5 p(p1 3 p11 1 p2 3 p21) 1 q(q1 3 q11 1 q2 3 q21)

FiGure 15.2
Embellishment of the decision problem of Example 15.1-1

Jane’s preferences regarding location and reputation of each university. The remain-
der of the decision-making chart can be interpreted similarly. Note that p + q = 1,
p1 + p2 = 1, q1 + q2 = 1, p11 + p12 + p13 = 1, p21 + p22 + p23 = 1, q11 + q12 + q13 = 1, and
q21 + q22 + q23 = 1. The bottom of Figure 15.2 demonstrates how the U of A composite
weight is computed.

Determination of the weights. The crux of AHP is the determination of the

relative weights (such as those used in Example 15.1-1) to rank the alternatives.
Assuming that we are dealing with n criteria at a given hierarchy, AHP establishes
a pairwise n * n comparison matrix, a, that quantifies the decision maker’s judg-
ment of the relative importance of the criteria. The pairwise comparison is made
such that the criterion in row i 1i = 1, 2, c, n2 is ranked relative to every other
criterion. Letting aij define the element (i, j) of a, AHP uses a numeric scale from
1 to 9 in which aij = 1 signifies that i and j are of equal importance, aij = 5 indi-
cates that i is strongly more important than j, and aij = 9 indicates that i is extremely
more important than j. Other intermediate values between 1 and 9 are interpreted
correspondingly. Consistency in judgment means that if aij = k, then aji = k1 . Also,
all the diagonal elements aii of a equal 1, because these elements rank each crite-
rion against itself.
570 Chapter 15 Decision Analysis and Games

example 15.1-2
To show how the comparison matrix a is determined for Martin’s decision problem of
Example 15.1-1, we start with the top hierarchy dealing with the criteria of location (L) and
reputation (R). In Martin’s judgment, R is strongly more important than L, and hence a21 = 5
and, automatically, a12 = 15, thus yielding the following comparison matrix:

L R
1
L 1 5
a = a b
R 5 1

The relative weights of R and L can be determined by normalizing a to create a new matrix
N. The process requires dividing the individual elements of each column by the column sum.
Thus, we divide the elements of columns 1 by 6 11 + 52 and those of column 2 by 1.2 1 = 15 + 12.
The desired relative weights, wR and wL, are then computed as row averages:

L R Row averages
L .17 .17 wL = .17 +2 .17 = .17
N = a b
R .83 .83 wR = .83 +2 .83 = .83

The computations yield wL = .17 and wR = .83, the weights we used in Figure 15.1. The
columns of N are equal, an indication that the decision maker is exhibiting consistent judgment
in specifying the entries of the comparison matrix a. Consistency is always guaranteed in 2 * 2
comparison matrices but not in higher-order matrices (as we explain shortly).
Martin’s preferences regarding the relative importance of the three universities from the
standpoint of the two criteria L and R are summarized in the following comparison matrices:

A B C A B C
1 1
A 1 2 5 A 1 2 3
1 1 3
aL = B ° 2 1 2 ¢, aR = B ° 2 1 2¢
1 2
C 5 2 1 C 3 3 1

Next, we have

aL@column sum = 18, 3.5, 1.72

aR@column sum = 11.83, 3.67, 5.52

The normalized matrices are determined by dividing each column-entry by its respective column-
sum—namely,

A B C Row averages
A .125 .143 .118 wLA = .125 + .143
3
+ .118
= .129
.250 + .286 + .294
NL = B ° .250 .286 .294 ¢ wLB = 3 = .277
C .625 .571 .588 wLC = .625 + .571
3
+ .588
= .594
A B C Row averages
A .545 .545 .545 wLA = .545 + .545
3
+ .545
= .545
.273 + .273 + .273
NR = B ° .273 .273 .273 ¢ wLB = 3 = .273
.182 + .182 + .182
C .182 .182 .182 wLC = 3 = .182
 15.1 Decision Making under Certainty—Analytic Hierarchy Process (AHP) 571

The values of (wLA, wLB, and wLC) ( = .129, .277, and .594) provide the respective location
weights for U of A, U of B, and U of C, respectively. Similarly, the values of (wRA, wRB, and
wRC) ( = .545, .273, .182) give the relative weights regarding academic reputation of the three
universities. These are the values used in Figure 15.1.

Consistency of the comparison matrix. In Example 15.1-2, all the columns of the nor-
malized matrices N and NR are identical, and those of NL are not. This means that a
and aR are consistent and aL is not.
Consistency implies rational judgment on the part of the decision maker.
Mathematically, we say that a comparison matrix a is consistent if

aijajk = aik, for all i, j, and k

For example, in matrix aR of Example 15.1-2, a13 = 3 and a12a23 = 2 * 32 = 3. This

property requires all the columns (and rows) of aR to be linearly dependent. In par-
ticular, the columns of any 2 * 2 comparison matrix, such as a, are by definition
dependent, and hence a 2 * 2 matrix is always consistent.
It is unusual for higher-order comparison matrices to be always consistent, and
a degree of inconsistency is expected. To decide what level of inconsistency is “toler-
able,” we need to develop a quantifiable measure of consistency for the comparison
matrix a. We have seen in Example 15.1-2 that a consistent a produces a normalized
matrix N in which all the columns are identical—that is,

w1 w1 c w1
w w2 c w2
N = § 2 ¥
c c c c
wn wn c wn

The original comparison matrix a can be determined from N by a reverse process that
divides the elements of column i by wi —that is,

c
w1 w1
1 w2 wn
c
w2 w2
w1 1 wn
a = § ¥
c c c c
c
wn wn
w1 w2 1

Post-multiplying a by w = 1w1, w2, c, wn 2 T , we get

c
w1 w1
1 w2 wn w1 nw1 w1
c
w2 w2
1 w2 nw2 w
¥ = n§ 2 ¥
w1 wn
§ ¥ § ¥ = §
c c c c c c c
c
wn wn
w1 w2 1 wn nwn wn

Hence, a is consistent if,

aw = nw
572 Chapter 15 Decision Analysis and Games

For the case where a is not consistent, the relative weight, wi, is approximated by
the average of the n elements of row i in the normalized matrix N (see Example 15.1-2).
Letting w be the vector of computed averages, it can be shown that
aw = nmaxw, nmax Ú n
In this case, the closer nmax is to n, the more consistent is the comparison matrix a.
Based on this observation, AHP computes the consistency ratio as
CI
CR =
RI
where
CI = Consistency index of a
nmax - n
=
n - 1
RI = Random consistency of a
1.981n - 22
=
n
The random consistency index, RI, is determined empirically as the average CI of a
large sample of randomly generated comparison matrices, a.
If CR … .1, the level of inconsistency is acceptable. Otherwise, the inconsistency
is high, and the decision maker may need to revise the estimates of the elements aij to
realize better consistency.
The value of nmax is computed from aw = nmaxw by noting that the ith equation is

a aijwj = nmaxwi, i = 1, 2, c, n
n

j=1

Given a i = 1wi = 1, we get

n

a a a aijwj b = nmax a wi = nmax

n n n

i=1 j=1 i=1

This means that the value of nmax equals the sum of the elements of the column vec-
tor aw.

example 15.1-3
In Example 15.1-2, the matrix aL is inconsistent because the columns of its NL are not identical.
To test the consistency of NL, we start by computing nmax. From Example 15.1-2, we have
w1 = .129, w2 = .277, w3 = .594
Thus,
1 1
1 2 5 .129 0.3863
1
aLw = £ 2 1 2 ≥ £ .277 ≥ = £ 0.8320 ≥
5 2 1 .594 1.7930
nmax = .3863 + .8320 + 1.7930 = 3.0113
 15.1 Decision Making under Certainty—Analytic Hierarchy Process (AHP) 573

Now, for n = 3,
nmax - n 3.0113 - 3
CI = = = .00565
n - 1 3 - 1
1.981n - 22 1.98 * 1
RI = = = .66
n 3
CI .00565
CR = = = .00856
RI .66

Because CR 6 .1, the level of inconsistency in aL is acceptable.

excel Moment
Template excelAHP.xls is driven by user input and can handle comparison matrices of size 8 * 8
or less. Figure 15.3 demonstrates the application of the model to Example 15.1-2 (columns F:I and
rows 10:13 are hidden to conserve space). The comparison matrices of the problem are entered
one at a time in the (top) input data section of the spreadsheet. The order in which the compari-
son matrices are entered is unimportant, though it makes more sense to consider them in their
natural hierarchal order.
The output (bottom) section of the spreadsheet provides the associated normalized matrix
and its consistency ratio, CR.1 The weights, w, are copied from column J and pasted into the
solution summary area (the right section of the spreadsheet). Remember to use Paste Special
1 Values when performing this step to guarantee a permanent record. The process is repeated
until all the weights for all the comparison matrices have been stored in the solution summary
area starting at column K.
In Figure 15.3, the final ranking is given in cells (K18:K20). The formula in cell K18 is
= $L$4*$L7+ $L$5*$N7

FiGure 15.3
Excel solution of Example 15.1-2 (file excelAHP.xls)

1
The more accurate results of the spreadsheet differ from those in Examples 15.1-2 and 15.1-3 because of
manual roundoff approximation.
574 Chapter 15 Decision Analysis and Games

This formula provides the composite weight for alternative UA and is copied in cells K19 and
K20 to evaluate alternatives UB and UC. Note from the formula in K18 that cell reference to the
alternative UA must be column-fixed (namely, $L7 and $N7), whereas all other references must
be row-and-column-fixed (namely, $L$4 and $L$5). The validity of the copied formulas requires
stacking the (column-fixed) alternative weights of each matrix in a single column (no intervening
empty cells). In Figure 15.3, the AR -weights are in column L and the AL-weights are in column N.
There are no restrictions on the placement of the A-weights because they are row- and column-
fixed in the formula.
You can embellish the formula in K18 to capture the names of the alternatives by using
= $K7&"= "&TEXT1$L$4*$L7+ $L$5*$N7,"#### 0.00000"2
The procedure for evaluating alternatives can be extended to any number of hierarchy levels.
Once you develop the formula correctly for the first alternative, the same formula is copied to
the remaining cells. Remember that all cell references in the formula must be row-and-column-
fixed, except for references to the alternatives, which must be column-fixed only. Problem 15-2,
asks you to develop the formula for a 3-level problem.

15.2 DeCIsIon MAkIng UnDeR RIsk

Under conditions of risk, the payoffs associated with each decision alternative are rep-
resented by probability distributions, and decision can be based on the expected value
criterion—maximization of expected profit or the minimization of expected cost. The
expected value criterion is sometimes modified to account for other situations, as will
be described later in this section.

Real-Life Application—Booking Limits in Hotel Reservations

Hotel La Posada has a total of 300 guest rooms. Its clientele includes both business
and leisure travelers. Room prices are discounted, mainly to leisure travelers. Business
travelers, who usually are late in booking their rooms, pay full price. La Posada sets a
booking limit on the number of discount rooms to take advantage of the full price paid
by business customers. The case study at the end of this chapter utilizes decision tree
analysis to determine the booking limits.

15.2.1 Decision tree–Based expected Value Criterion

The expected value criterion seeks the maximization of expected (average) profit or
the minimization of expected cost. The data of the problem assumes that the payoff (or
cost) associated with each decision alternative is probabilistic.

Decision tree analysis. The following example considers simple decision situations
with a finite number of decision alternatives and explicit payoff matrices.

example 15.2-1
Suppose that you want to invest $10,000 in the stock market by buying shares in one of two com-
panies: A and B. Shares in Company A, though risky, could yield a 50% return during the next
year. If the stock market conditions are not favorable (i.e., a “bear” market), the stock may lose
20% of its value. Company B provides safe investments with a 15% return in a “bull” market
 15.2 Decision Making under risk 575

“Bull” market (.6) FiGure 15.4

$5000
Decision-tree representation
Invest in stock A of the stock market problem
2
“Bear” market (.4)
2$2000
1
“Bull” market (.6)
$1500
Invest in stock B
3
“Bear” market (.4)
$500

and only 5% in a “bear” market. All the publications you have consulted (and there is always a
flood of them at the end of the year!) are predicting a 60% chance for a “bull” market and 40%
chance for a “bear” market. How should you invest your money?
The decision problem is summarized in the following table:

1-year return on $10,000 investment

Decision alternative “Bull” market ($) “Bear” market ($)

Company A stock 5000 - 2000

Company B stock 1500 500
Probability of occurrence .6 .4

The problem can also be represented as a decision tree as shown in Figure 15.4. Two types of
nodes are used in the tree: A square (n) represents a decision point, and a circle (~) represents
a chance event. Thus, the two branches from decision point 1 represent the two alternatives of
investing in stock A or stock B. Next, the two branches emanating from chance events 2 and 3
represent the “bull” and the “bear” markets with their respective probabilities and payoffs.
From Figure 15.4, the expected 1-year returns are

Stock A = $5000 * .6 + 1 -20002 * .4 = $2200

Stock B = $1500 * .6 + $500 * .4 = $1100

Stock A is chosen because it yields a higher expected return.

remarks. In the terminology of decision theory, the probabilistic “bull” and the
“bear” markets are called states of nature. In general, a decision problem may include n
states of nature and m alternatives. If pj 1 7 02 is the probability of occurrence for state
j and aij is the payoff of alternative i, given state j 1i = 1, 2, c, m; j = 1, 2, c, n2,
then the expected payoff for alternative i is computed as

EVi = ai1 p1 + ai2 p2 + c + ain pn, i = 1, 2, c, n

p1 + p2 + c + pn = 1

The best alternative corresponds to maxi 5EVi 6 and mini 5EVi 6 for the cases of profit
and loss, respectively.
576 Chapter 15 Decision Analysis and Games

15.2.2 Variants of the expected Value Criterion

This section addresses two issues relating to the expected value criterion: the determi-
nation of posterior probabilities based on experimentation and the use of utility versus
actual value of money.

posterior (Bayes’) probabilities. The probabilities used in the expected value crite-
rion are usually estimated from historical data (see Section 14.5). In some cases, the
accuracy of these estimates can be enhanced by using additional experimentation. The
resulting probabilities are referred to as posterior (or Bayes’) probabilities, as opposed
to the prior probabilities determined from raw data.

Real-Life Application—Casey’s Problem: Interpreting and evaluating

a new test
A screening test of a newborn, named Casey, reveals a C14:1 enzyme deficiency. The
enzyme is required to digest a particular form of long-chain fats, and its absence could
lead to severe illness or mysterious death (broadly categorized under the sudden infant
death syndrome or SIDS). The test had been administered previously to approximately
13,000 newborns, and Casey was the first to test positive. Though the screening test does
not in itself constitute a definitive diagnosis, the extreme rarity of the condition led her
doctors to conclude that there was an 80 to 90% chance that she was suffering from
this deficiency. Given that Casey tested positive, Bayes’ posterior probability is used
to assess whether or not the child has the C14:1 deficiency. The situation is detailed in
Case 11 in Chapter 26, on the website.

example 15.2-2
This example demonstrates how the expected-value criterion is modified to take advantage of
posterior probabilities. In Example 15.2-1, the (prior) probabilities of .6 and .4 of a “bull” and
a “bear” market are determined from available financial publications. Suppose that rather than
relying solely on these publications, you have decided to conduct a more “personal” investiga-
tion by consulting a friend who has done well in the stock market. The friend quantifies a “for/
against” investment recommendation in the following manner: In a “bull” market, there is a
90% chance the recommendation is “for.” It drops to 50% in a “bear” market. How does the
additional information affect the decision?
The friend’s statement provides conditional probabilities of the recommendations “for” and
“against” given that the states of nature are “bull” and “bear” markets. Define
v1 = “For” vote
v2 = “Against” vote
m1 = “Bull” market
m2 = “Bear” market

Thus, the friend’s statement may be written in the form of probability statements as
P5v1 0 m1 6 = .9, P5v2 0 m1 6 = .1
P5v1 0 m2 6 = .5, P5v2 0 m2 6 = .5
 15.2 Decision Making under risk 577

With this representation, the decision problem is summarized as:

1. If the friend’s recommendation is “for,” would you invest in stock A or in stock B?
2. If the friend’s recommendation is “against,” would you invest in stock A or in stock B?
The decision tree in Figure 15.5 represents the problem. Node 1 is a chance event represent-
ing the “for” and “against” possibilities. Nodes 2 and 3 are decision points for choosing between
stocks A and B, given the “for” and “against” recommendations, respectively. Finally, nodes 4 to
7 are chance events representing the “bull” and “bear” markets.
To evaluate the different alternatives in Figure 15.5, it is necessary to compute the posterior
probabilities P5mi 0 vj 6 shown on the m1- and m2-branches of nodes 4, 5, 6, and 7. These pos-
terior probabilities take into account the additional information provided by the friend’s “for/
against” recommendation and are computed according to the following general steps:

Step 1. Summarize the conditional probabilities P5vj 0 mi 6 in the following tabular form:

v1 v2

m1 .9 .1

m2 .5 .5

FiGure 15.5
Decision tree for the stock market problem with posterior probabilities

“Bull” market (m1)

$5000
P m1|y 1 5 .730
Stock A
4
“Bear” market (m2)
–$2000
“For” vote (y 1) P m 2|y 1 5 .270
2
“Bull” market (m1)
$1500
P m1|y 1 5 .730
Stock B
5
“Bear” market (m2)
$500
P m 2|y 1 5 .270
1
“Bull” market (m1)
$5000
P m1|y 2 5 .231
Stock A
6
“Bear” market (m2)
–$2000
“Against” vote (y 2) P m 2|y 2 5 .769
3
“Bull” market (m1)
$1500
P m1|y 2 5 .231
Stock B
7
“Bear” market (m2)
$500
P m 2|y 2 5 .769
578 Chapter 15 Decision Analysis and Games

Step 2. Compute the joint probabilities as

P5mi, vj 6 = P5vj 0 mi 6P5mi 6, for all i and j

Given the prior probabilities P5m1 6 = .6 and P5m2 6 = .4, the joint probabilities are
determined by multiplying the first and the second rows of the table in step 1 by .6 and
.4, respectively—that is,

v1 v2

m1 .54 .06

m2 .20 .20

The sum of all the entries in the table equals 1.

Step 3. Compute the absolute probabilities as

P5vj 6 = a P5mi, vj 6, for all j

all i

These probabilities are the column sums in the table in step 2—that is,

P{v1} P{v2}

.74 .26

Step 4. Determine the desired posterior probabilities as

P5mi, vj 6
P e mi ` vj f =
P5vj 6

These probabilities are computed by dividing each column in the table of step 2 by the
corresponding column sum in the table of step 3, which yields

v1 v2

m1 .730 .231

m2 .270 .769

These are the probabilities used in Figure 15.5 and are different from the prior prob-
abilities P5m1 6 = .6 and P5m2 6 = .4.
We are now ready to evaluate the alternatives based on the expected payoffs for
nodes 4, 5, 6, and 7—that is,

“For” Recommendation

Stock A at node 4 = 5000 * .730 + 1 - 20002 * .270 * = $3110

Stock B at node 5 = 1500 * .730 + 500 * .270 = 1230
Decision. Invest in stock A.
 15.2 Decision Making under risk 579

“Against” Recommendation
Stock A at node 6 = 5000 * .231 + 1 - 20002 * .769 = - $383
Stock B at node 7 = 1500 * .231 + 500 * .769 = $731
Decision. Invest in stock B.
The given decisions are equivalent to saying that the expected payoffs at decision nodes 2
and 3 are $3110 and $731, respectively (see Figure 15.5). Thus, given the probabilities P5v1 6 = .74
and P5v2 6 = .26 as computed in step 3, we can compute the expected payoff for the entire deci-
sion tree. (See Problem 15-30.)

excel Moment
Excel file excelBayes.xls is designed to determine the posterior probabilities for prior probability
matrices of sizes up to 10 * 10 (some rows and columns have been hidden to conserve space).
The input data include P5m6 and P5v  m6. The spreadsheet checks input data errors and dis-
plays appropriate error messages.

Aha! Moment: An eighteenth-Century Lottery that yields Infinite expected

Payoff, or Does It?
In the early eighteenth century, Swiss mathematician Nicolas Bernoulli introduced a paradoxical
theoretical lottery game with an expected payoff of infinity. The paradox arises because the game
sets no limit on the amount of money a player can win. The game was published by Nicolas’
brother, Daniel, in 1738 in the St. Petersburg Academy Proceedings and became known as the
petersburg paradox. The rules of the game are simple: Toss a fair coin. If the outcome is heads (H),
the game continues; otherwise, the game ends at the first occurrence of tails (T). Starting with $2
for the first H, the payoff doubles with the occurrence of each successive H, yielding the monetary
stream $2, $4, $8, $16, . . . The probability that an H will recur in toss n is 1 12 2 n - 1 1 12 2 = 1 21n 2. Thus,
assuming the game is played indefinitely,

Expected payoff = a 2n 1 21n 2 = 1 + 1 + 1 + c = ∞

∞

n=1

If the infinite expected payoff is the “fair” value of the game, then, theoretically, a player should
accept any price for playing the game, a paradoxical outcome particularly when a rational deci-
sion maker realizes that a low payoff is probable (e.g., there is a 50-50 chance of winning $2) and
a high payoff is unlikely [e.g., the probability of winning the (by-comparison) modest amount of
$10241 = 210 2 is less than .001].
The paradox was resolved by Daniel Bernoulli2 by introducing the concepts of utility func-
tions and risk aversion to replace monetary amounts in the expected value computations, as the
remainder of this section explains.

2
http://www.econ.ucsb.edu/~tedb/Courses/GraduateTheoryUCSB/Bernoulli.pdf, accessed 06-14-2015. Bernoulli
acknowledged that ten years earlier his Swiss colleague Gabriel Cramer independently came very close to re-
solving the paradox.
580 Chapter 15 Decision Analysis and Games

Utility functions. In the preceding presentation, the expected value criterion is ap-
plied to situations where the payoff is real money. There are cases where the utility
rather than the real value should be used in the analysis. To illustrate this point, sup-
pose there is a 50-50 chance that a $20,000 investment will produce a profit of $40,000
or be lost. The associated expected profit is 40,000 * .5 - 20,000 * .5 = $10,000.
Although there is a net expected profit, different individuals vary in interpreting the
result. An investor who is willing to accept risk may undertake the investment for a
50% chance to make a $40,000 profit. Conversely, a conservative investor may not be
willing to risk losing $20,000. The concept of utility function is devised to reflect these
differences. The utility function then takes the place of real money in the decision-
making model.
How is the subjective attitude toward risk quantified in the form of a utility
function? In the preceding investment illustration, the best payoff is $40,000, and
the worst is - $20, 000. We can establish a utility scale, U, from 0 to 100 that speci-
fies U1 - $20,0002 = 0 and U1$40,0002 = 100. The value of U for investment re-
turn between - $20, 000 and $40,000 can be determined in the following manner: If
the decision maker is neutral (indifferent) toward risk, then U can be represented
by a straight line joining 10, - $20,0002 and (100, $40,000). In this case, both real
money and its utility lead to the same decisions. More generally, the function U can
take other forms reflecting different attitudes toward risk. Figure 15.6 illustrates the
cases of individuals X, Y, and Z. Individual Y is risk neutral, individual X is risk
averse (or cautious), and individual Z, the opposite of X, is a risk seeker. The figure

FiGure 15.6
Utility functions for risk averse (X), neutral (Y), and risk seeker (Z) decision makers

100
210 110 X

a Y

b
Utility

50
c Z

e
f
0
220 0 120 140
Thousands of dollars
 15.3 Decision under uncertainty 581

demonstrates that for the risk-averse X, the drop in utility bc corresponding to a loss
of $10,000 is larger than the increase ab associated with a gain of $10,000. The op-
posite is true for risk seeker Z where de 7 ef. In general, an individual can be both
risk averse and risk seeking, in which case the associated utility curve will follow an
elongated S-shape.
Utility curves similar to the ones demonstrated in Figure 15.6 are determined by
“quantifying” the decision maker’s attitude toward risk for different levels of cash money.
In our example, the desired range is ( - $20,000 to $40,000) with U1 - $20,0002 = 0
and U1$40,0002 = 100. To specify the values of U for intermediate cash values (e.g.,
- $10,000, $0, $10,000, $20,000, and $30,000), we establish a lottery for a cash amount x
whose expected utility is
U1x2 = pU1 -20,0002 + 11 - p2U1 $40,0002, 0 … p … 1
= 0p + 10011 - p2
= 100 - 100p
To determine U(x), the decision maker must state a preference between a guar-
anteed cash amount x and the chance to play a lottery for which there is a loss of
- $20,000 with probability p and a profit of $40,000 with probability 1 - p. The value
of p reflects the decision maker’s neutrality (or indifference) toward risk. For example,
for x = $20,000, the decision maker may feel that a guaranteed $20,000 cash and the
lottery with p = .8 are equally attractive. In this case, we can compute the utility of
x = $20,000 as
U1$20,0002 = 100 - 100 * .8 = 20
Note that higher values of p for the same lottery reflect risk seeking (as opposed to risk
aversion). For example, for p = .1,
U1$20,0002 = 100 - 100 * .2 = 80

15.3 DeCIsIon UnDeR UnCeRtAInty

Decision making under uncertainty, as under risk, involves alternative actions whose
payoffs depend on the (random) states of nature. Specifically, the payoff matrix of a de-
cision problem with m alternative actions and n states of nature can be represented as

s1 s2 c sn

a1 v1a1, s12 v1a1, s22 c v1a1, sn2

a2 v1a2, s12 v1a2, s22 c v1a2, sn2
f f f f f

am v1am, s12 v1am, s22 c v1am, sn2

The element ai represents action i and the element sj represents state of nature j. The
payoff or outcome associated with action ai and state sj is v(ai, sj).
582 Chapter 15 Decision Analysis and Games

In decision making under uncertainty, the probability distribution associated with

the states sj, j = 1, 2, c, n, is either unknown or cannot be determined. This lack of
information has led to the development of special decision criteria:

1. Laplace
2. Minimax
3. Savage
4. Hurwicz

These criteria differ in how conservative the decision maker is in the face of uncertainty.
The Laplace criterion is based on the principle of insufficient reason. Because the
probability distributions are not known, there is no reason to believe that the probabil-
ities associated with the states of nature are different. The alternatives are thus evalu-
ated using the liberal assumption that all states are equally likely to occur—that is,
P5s1 6 = P5s2 6 = c = P5sn 6 = n1 . Given that the payoff v(ai, sj) represents gain,
the best alternative is the one that yields

n ja
1 n
max e v1ai, sj 2 f
ai =1

The maximin (minimax) criterion is based on the conservative attitude of making

the best of the worst-possible conditions. If v(ai, sj) is loss, then we select the action
that corresponds to the following minimax criterion:

min e max v1ai, sj 2 f

ai sj

If v(ai, sj) is gain, we use the maximin criterion given by

max e min v1ai, sj 2 f

ai Sj

The Savage regret criterion aims at “moderating” the degree of conservatism in

the minimax (maximin) criterion by replacing the (gain or loss) payoff matrix v(ai, sj)
with a loss (or regret) matrix, r(ai, sj), by using the following transformation:

v1ai, sj 2 - min 5v1ak, sj 26 , if v is loss

ak
r1ai, sj 2 = µ
max 5v1ak, sj 26 - v1ai, sj 2, if v is gain
ak

To show why the Savage criterion moderates the minimax (maximin) criterion,
consider the following loss matrix:
s1 s2 Row max
a1 $11,000 $90 $11,000
v1ai, sj 2 =
a2 $10,000 $10,000 $10,000 d Minimax
 15.3 Decision under uncertainty 583

The application of the minimax criterion shows that a2, with a definite loss of $10,000,
is the preferred alternative. However, it may be better to choose a1 because there is a
chance of limiting the loss to $90 only if s2 occurs. This happens to be the case when the
regret matrix is used:
s1 s2 Row max
a1 $1,000 $0 $1,000 d Minimax
r1ai, vj 2 =
a2 $0 $9,910 $9,910

The last criterion, hurwicz, is designed to represent different decision-making

attitudes, ranging from the most liberal (optimistic) to the most conservative (pessi-
mistic). Define 0 … a … 1. The selected action must be associated with

max e a max v1ai, sj 2 + 11 - a2min v1ai, sj 2 f, if v is gain

ai sj sj

min e a min v1ai, sj 2 + 11 - a2max v1ai, sj 2 f, if v is loss

ai sj sj

The parameter a is the index of optimism. If a = 0, then the criterion reduces to con-
servative minimax criterion, seeking the best of the worst conditions. If a = 1, then the
criterion is liberal because it seeks the best of the best conditions. The degree of opti-
mism (or pessimism) can be adjusted by selecting a value of a between 0 and 1. In the
absence of strong feeling regarding extreme optimism and extreme pessimism, a = .5
may be a fair choice.

example 15.3-1
National Outdoors School (NOS) is preparing a summer campsite in the heart of Alaska to
train individuals in wilderness survival. NOS estimates that attendance can fall into one of four
categories: 200, 250, 300, and 350 persons. The cost of the campsite will be the smallest when its
size meets the demand exactly. Deviations above or below the ideal demand levels incur addi-
tional costs resulting from constructing more capacity than needed or losing income opportuni-
ties when the demand is not met. Letting a1 to a4 represent the sizes of the campsites (200, 250,
300, and 350 persons) and s1 to s4 the level of attendance, the following table summarizes the
cost matrix (in thousands of dollars) for the situation:

s1 s2 s3 s4

a1 5 10 18 25

a2 8 7 12 23

a3 21 18 12 21

a4 30 22 19 15

The problem is analyzed using all four criteria.

584 Chapter 15 Decision Analysis and Games

Laplace. Given P5sj 6 = 14, j = 1 to 4, the expected values for the different actions are computed as
1
E5a1 6 = 15 + 10 + 18 + 252 = $14,500
4
1
E5a2 6 = 18 + 7 + 12 + 232 = $12,500 d Optimum
4
1
E5a3 6 = 121 + 18 + 12 + 212 = $18,000
4
1
E5a4 6 = 130 + 22 + 19 + 152 = $21,500
4
Minimax. The minimax criterion produces the following matrix:
s1 s2 s3 s4 Row max

a1 5 10 18 25 25

a2 8 7 12 23 23

a3 21 18 12 21 21 d Minimax

a4 30 22 19 15 30

Savage. The regret matrix is determined by subtracting 5, 7, 12, and 15 from columns 1 to 4,
respectively. Thus,
s1 s2 s3 s4 Row max

a1 0 3 6 10 10

a2 3 0 0 8 8 d Minimax

a3 16 11 0 6 16

a4 25 15 7 0 25

hurwicz. The following table summarizes the computations:

Alternative Row min Row max a1Row min2 + 11 - a21Row max2

a1 5 25 25 - 20a
a2 7 23 23 - 16a
a3 12 21 21 - 9a
a4 15 30 30 - 15a

Using an appropriate a, we can determine the optimum alternative. For example, at a = .5,
either a1 or a2 is the optimum, and at a = .25, a3 is the optimum.

excel Moment
Template excelUncertainty.xls can be used to automate the computations of Laplace, maximin,
Savage, and Hurwicz criteria. The spreadsheet is based on the use of a cost matrix. To use a reward
matrix, all entries must be multiplied by -1. The maximum matrix size is 110 * 102.