SYSTEMS METHODS

CHAPTER 5
 Decision Analysis
• A decision process is a process requiring either a
single or sequential set of decisions for its
completion.

• Each allowable decision has a gain or loss

associated with it, which is codetermined by
external circumstances surrounding the process,
a feature which distinguishes these processes
from the processes treated in other problems.
• The set of possible circumstances, known as the
state of nature, and a probability distribution
governing the occurrence of each state are
presumed known. Both the set of allowable
decisions and the set of states of nature will be
assumed finite (an assumption not made in the
more elaborate theory).

• We denote the allowable decisions by D1,

D2,…,Dm; the states of nature by S1, S2,…,Sn; and
the return associated with decision D1 and state Sj
by gij (i=1,2,…,m); j=1,2,…,n.
 Table 1
• A process requiring the implementation of just
one decision is defined completely by Table 1.
This payoff table is known as a gain matrix
whenever the entries gij are in terms of gains to
the decision maker. Losses are then represented
as negative gains.
States of Nature
S1 S2 … Sn
D1 g11 g12 … g1n
Decision

D2 g21 g22 … g2n

… … … … …
Dm gm1 gm2 … gnm
 Example 1
• A major energy company offers a landowner
$60,000 for the exploration rights to natural
gas on a certain site and the option for future
development. The option, if exercised, is
worth an additional $600,000 to the
landowner, but this will occur only if natural
gas is discovered during the exploration
phase. The landowner, believing that the
energy company’s interest is a good indication
that gas is present, is tempted to develop the
field herself.
• To do so, she must contract with local outfits with
expertise in exploration and development. The
initial cost is $100,000 which is lost if no gas is
found. If gas is discovered, however, the
landowner estimates a net profit of 2 million
dollars.

• The decisions for the landowner are D1 (to accept

the energy company’s offer) and D2 (to explore
and develop on her own). The states of nature
are S1 (there is no gas on the land) and S2 (there is
gas on the land).
 Table 2
• The gains (in thousand of dollars) to the
landowner for each combination of events are
given in Table 2. It remains to specify or
estimate the probabilities attached to the two
states of nature, P(S1), and P(S2).
States of Nature
S1 S2
Decision

D1 60 660
D2 -100 2000
• In a decision process, only the decision maker is
capable of making rational decisions, nature is not.

• The actual state of nature in existence at any given

time is a random event, but the underlying probability
distribution cannot be considered a “mixed strategy,”
designed to inflict losses on the decision maker.

• Furthermore, we generally rule out any randomness in

the decision maker’s choice; he or she is restricted to
one or another “pure strategy” D1,……,Dm. Because of
these differences, optimal game strategies tend, for
decision process, to be too conservative.
 Naïve Decision Trees
• The minimax (or pessimistic) criterion is to select
the decision that minimizes the maximum
possible loss to the decision maker. In terms of a
gain matrix, it is the decision that maximizes the
minimum possible gain.

• The optimistic criterion is to choose the decision

that maximizes the possible gain. The middle-of-
the road criterion is to select that decision for
which the average of the maximum and minimum
gains is greatest.
 Example No. 2
• Determine recommended decisions under each
naïve criterion for the process described in
Example No. 1.

• Solution:
The gain matrix for this process is Table 2.
The minimum gain for decision D1 is 60, while
that for D2 is -100. Since max {60, -100} = 60 is
the gain associated with D1, D1 is the
recommended decision under the minimax
criterion.
• The largest entry in the matrix is 2000, the gain
associated with D2. Therefore D2 is the recommended
decision under the optimistic criterion.

• The averages of the maximum and minimum gains for

D1 and D2 are, respectively,

660 + 60 = 360 and 2000 + (-100) = 950

2 2

• Since max {360, 950} = 950 is associated with D2, D2 is

the recommended decision under the middle-of-the-
road criterion.
 Example No. 3
• Determine recommended decisions under each
naïve criterion for the following decision process.
A customer for a large construction firm must
place orders with an item manufacturer 9 months
before the items are needed. One decision is as
to the number of one boom truck to stock. The
ultimate gain to the construction firm depends
both on this decision and on the technology
prevailing 9 months later. The customer’s
estimates of the gains (in thousands of dollars)
are given in Table 3.
 Table 3
S1: Item 1 are S2: Item 1 are S3: Item 1are not
highly new acceptable acceptable
technology
D1: Order none -50 0 80
D2: Order a little -10 30 35
D3: Order 60 45 -30
moderately
D4: Order a lot 80 40 -45
• The minimum gains for decisions D1 through D4 are,
respectively, -50, -10, -30, and -45. Since the maximum
of these four amounts is -10, a gain associated with D2,
D2 is the recommended decision under the minimax
criterion.

• The maximum gain is 80, associated with both D1 and

D4. Hence, either D1 or D4 is the recommended
decision under the optimistic criterion.

• The averages of the maximum and minimum gains for

D1 through D4, respectively, are 15, 12.5, 15, and 17.5.
Since the maximum of these averages is associated
with D4, D4 is the recommended decision under the
middle-of-the-road criterion.
 A Priori Criterion
• The a priori (or Bayes’) criterion is to select
the decision that maximizes the expected
gain.
 Example No. 5
• Determine the recommended decision under the
a priori criterion for the process of Example 1, if
the landowner estimates the probability of
finding gas as 0.6.

• Solution:
With P(S2) = 0.6, it follows that P(S1) 1 – 0.6 =
0.4. Using the data in Table 2, we calculate the
expected gain from D1, as:
E(G1) = (60)(0.4) + (660)(0.6) = 420
• And the expected gain from D2 as:
E(G2) = (-100)(0.4) + (2000)(0.6) = 1160

• Since the maximum of these two amounts, 1160,

is associated with D2, D2 is the recommended
decision under the a priori criterion.

• This decision process is represented by the

decision tree in Fig. 1. The expected gain of the
process, 1160 at node B, is carried back from
node D.
Figure 1
 Decision Trees
• A decision tree is an oriented tree that represents
a decision process. The nodes designate points in
time where (i) one or another decision must be
made by the decision maker, of (ii) the decision
maker is faced with one or another state of
nature, or (iii) the process terminates.

• Directed out of a node (i) is a branch for each

possible decision; directed out of a node (ii) is a
branch for each possible state of nature.
• Under each branch the probability of the
corresponding event is written, when defined.

• Decision trees are useful in determining optimal

decisions for complicated processes. The
technique is to begin with the terminal nodes and
sequentially to move backwards through the
network, calculating the expected gains at the
intermediate nodes.

• Each gain is written above its corresponding

node. A recommended decision is one that leads
to a maximum expected gain. Decisions that turn
out to be nonrecommended have their
corresponding branches crossed out.
 Example No. 6
• Determine the recommended decision under the
a priori criterion for the decision process
described in Example No. 3, if the buyer
estimates P(S1) = 0.25, P(S2) = 0.40, and P(S3) =
0.35.

• Solution:
Using the data from Table 3, we calculate the
expected gains for decisions D1 through D4,
respectively,
E(G1)=(-50)(0.25)+(0)(0.40)+(80)(0.35)=15.5
E(G2)=(-10)(0.25)+(30)(0.4)+(35)(0.35)=21.75
E(G3)=(60)(0.25)+(45)(0.40)+(-30)(0.35)=22.5
E(G4)=(80)(0.25)+(40)(0.4)+(-45)(0.35)=20.25

• Since the maximum of these expected gains,

22.5, is associated with D3, D3 is the
recommended decision under the a priori
criterion.

• The process is represented in decision tree in Fig.

2
Figure 2
 A Posteriori Criterion
• If an imperfect experiment can be conducted that
provides information of the true state of nature,
then data from this experiment may be combined
with the initial probabilities of the various states
to yield an updated probability distribution.

• Designate the outcome of the experiment by θ

and assume that the reliability of the
experiment is given by the conditional
probabilities P(θ|S1), P(θ|S2),…,P(θ|Sn).
• The updated (or a posteriori) probabilities of
the states—P(S1| θ), P(S2| θ),…,P(Sn| θ)—are
determined from Baye’s theorem.

• The a posteriori criterion is to select the

decision that maximizes the expected gain
with respect to the updated probability
distribution.
 Example No. 7
• The landowner in Example No. 1 has soundings taken
on the site where natural gas is suspected, at a cost of
$30,000. The soundings indicate that gas is not
present, but the test is not a perfect one. The
company conducting the soundings concedes that 30
percent of the time the test will indicate no gas when
gas in fact exists. When gas does not exist, the test is
accurate 90 percent of the time. Using these data,
update the landowners’ initial estimate that the
probability of finding gas is 0.6 and then determine the
recommended decision under the a posteriori
criterion.
• Initially, P(S2) = 0.6, P(S1) = 0.4. Let θ1 designate the
event that soundings indicate no gas. Then
the reliability of the test is given by the
conditional probabilities P(θ1|S1) = 0.90 and
P(θ1|S2) = 0.30. Bayes’ theorem, of Example No. 6
gives the updated probabilities as:

P(S1| θ1)= P(θ1|S1)P(S1) .

P(θ1|S1)P(S1) + P(θ1|S2)P(S2)
= (0.90)(0.4) . = 2
(0.90)(0.4) + (0.30)(0.6) 3

P(S2| θ1) = 1 – P(S1| θ1) = 1/3

• The a posteriori gain matrix is obtained from
Table 2 by subtracting 30 (thousand dollars) from
each entry, thereby reflecting the cost of the test.
The expected gains (in thousands of dollars) for
decisions D1 and D2, respectively, in terms of the
updated probabilities are:
E(G1| θ1) = (60 – 30)(2/3) + (660 -30)(1/3)
= 230
E(G2| θ1) = (-100-30)(2/3)+(2000 – 30)(1/3)
= 570
 Figure 3
• Since the maximum expected gain is associated with D2, D2 is
the recommended decision under the a posteriori criterion.
• Figure 3 is the decision tree for this process. The probability
that the soundings indicate no gas, P(θ1), is unity, since
the result of the experiment is known.
 Example No. 8
• Solve Example No. 7 if the soundings had
indicated that gas was present.

• Solution:
Designate the event that soundings indicate
gas by θ2, From the data of Example No.
7,
P(θ2|S1) = 0.10 P(θ2|S2) = 0.70
• The initial probabilities are P(S1) = 0.40, P(S2) = 0.6;
therefore, the updated probability distribution is:
P(S1|θ2) = P(θ2|S1)P(S1) .
P(θ2|S1)P(S1) + P(θ2|S2)P(S2)
= (0.10)(0.4) = 0.087
(0.10)(0.4) + (0.70)(0.6)

• Again each entry in the original gain matrix, Table 2,

must be reduced by 30 (thousand dollars) to reflect the
cost of the test.
• Then the expected gains (in thousand of dollars) for
decisions D1 and D2 with respect to the latest
probability distribution are:
E(G1|θ2)=(60–30)(0.087)+(660-30)(0.913)
=577.8
E(G2| θ2)=(-100-30)(0.087)+(2000-30)(0.913)
= 1787.3
• Since the maximum expected gain is associated with
D2, D2 is the recommend decision under the a
posteriori criterion.
• Figure 4 is the decision tree for this process. The
probability that the soundings indicate gas is present,
P(θ2), is unity, since the result of the experiment is
known.
Figure 4