Decision Trees

Decision Trees
• Also called decision flow networks and decision
diagrams
• are powerful means of depicting and facilitating the
analysis of important problems, especially those
that involve sequential decisions and variable
outcomes over time
• Are used in practice because they make it possible
to break down a large, complicated problem into a
series of smaller simple problems, and they enable
objective analysis and decision making that
includes explicit consideration of the risk and effect
of the future
 Example 12-10
Probabilistic Decision Tree Analysis
The Ayax Corporation manufactures compressors for
commercial air-condition systems. A new compressor
design is being evaluated as a potential replacement for
the most frequently used unit. The new design involves
major changes that have the expected advantage of
better operating efficiency. From the perspective of a
typical user, the new compressor (as an assembled
component in an air-condition system) would have an
increased investment of $8,600 relative to the present unit
and an annual expense saving dependent upon the
extent to which the design goal is met in actual
operations.
Estimates by the multidisciplinary design team of the new
compressor achieving four levels (percentages) of the
efficiency design goal and the probability and annual
expense saving at each level are as follows:
Level (Percentage) of Probability Annual Expense
Design Goal Met (%) p(L) Saving
90 0.25 $3,470
70 0.40 2,920
50 0.25 2,310
30 0.10 1,560

Based on a before-tax analysis (MARR = 18% per

year, analysis period = 6 years, and market value =
0) and E(PW) as the decision criterion, is the new
compressor design economically preferable to the
current unit?
Figure 12-10 Single-State Decision Tree

Probability Annual
p(L) Expense Saving PW E(PW)

90: 0.25 $3,470 $3,538 $885

$1,086 70: 0.40 2,920 1,614 646

n
e sig 50: 0.25 2,310 -520 -130
D
w
Ne
30: 0.10 1,560 -3,143 -314
Cu $1,086
rre
=

nt
De
si g
n
$0 $0
Solution:
The PWs associated with each of the efficiency design goal levels being met
are as follows:
PW(90) = -$8,600 + $3,470(P/A,18%,6) = $3,538
PW(70) = -$8,600 + $2,920(P/A,18%,6) = $1,614
PW(50) = -$8,600 + $2,310(P/A,18%,6) = -$ 520
PW(30) = -$8,600 + $1,560(P/A,18%,6) = -$3,143
(P/A;18%;6) = 3,4976
Based on these values, the E(PW) of each installed unit of the new com-
pressor is as follows:
E(PW) = 0.25x$3,538 + 0.40x$1,614 + 0.25x(-$520) + 0.10x(-$3,143)
= $885 + $646 - $130 - $314
= $1,086
The E(PW) of the current unit is zero since the cash-flow estimates for the
new design are incremental amounts relative to the present design.
Therefore, the analysis indicates that new design is economically preferable
to the present design.
 The Expected Value of Perfect Information (EVPI)
The probability estimates of achieving the efficiency design
goal levels, p(L), developed by the design team in
Example 12-10 reflect the uncertainty about the future
operational performance of the new compressor. These
probabilities are based on present information and are
prior to obtaining any additional data.
If more test data were obtained to reduce the uncertainty,
additional costs would be incurred. Therefore, these
additional costs must be balanced against the value of
reducing the uncertainty. Obviously, if perfect information
were available about the future operational efficiency of the
new compressor, all uncertainty would be removed and we
could make an optimal decision between the current unit
and the new design. Even though perfect information is
unobtainable, its expected value indicates the maximum
amount (upper limit) we should consider spending for
additional information.
 Example 12-11
Expected Value of Perfect Information
Refer to Example 12-10. What is the EVPI regarding
future operational performance of the new compressor
to the typical user of an air-conditioning system?
Solution:
We can calculate the EVPI by comparing the optimal
decision based on perfect information with the original
decision in Example 12-10 to select the new
compressor. This comparison is shown in the Table 12-
16. Based on this comparison, the EPVI to the typical
user of a unit is
EVPI = $1,530 - $1,086 = $444
TABLE 12-16 Expected Value of Perfect Information (Example 12-11)

Level (%) of Probability Decision with Prior Decision

Design Goal Met P(L) Perfect Information (New Design)
Decision Outcome

90 0.25 New $3,538 $3,538

70 0.40 New 1,614 1,614
50 0.25 Current 0 -520
30 0.10 Current 0 -3,143
Expected value $1,530 $1,086
 The Use of Additional Information to
Reduce Uncertainty
• The solution to Example 12-11 shows that there is some
potential value to be obtained from additional test information
about the operational performance of the new compressor.
• From the viewpoint of the typical user of a new unit versus the
current unit, the maximum estimated value of additional
information is $444.
• The members of the management team of Ajax Corporation
are strongly customer focused and want to exceed customers’
expectations about the performance of their products. Thus,
they asked the design team to estimate the value of data that
could be obtained from an additional comprehensive test of
prototypes of the new compressor.
• The information from the test will not be perfect, because the
test cannot determine exactly the long-term operational
performance of the new design under diverse customer
applications. However, the imperfect test information may
reduce uncertainty and merit the additional cost of obtaining it.
 Example 12-12
Compressor Analysis Using Revised Probability
Refer to Example 12-10 and 12-11. The members of the design
team are confident that data from the additional comprehensive
test of prototype compressors will show whether future
operational performance will be favorable (60% or more of the
design goal met) or not favorable (60% of the design goal not
met). Based on obtaining these results from the test, and by
using current engineering data within Ajax Corporation, the
design team developed the following conditional probability
estimates:
Comprehensive Conditional Probabilities of Test Outcome
Test Outcome Given the Level (%) of the Design Goal Met
90: 0.25 70: 0.40 50: 0.25 30: 0.10

Favorable (F) 0.95 0.85 0.30 0.05

Not Favorable (NF) 0.05 0.15 0.70 0.95
Sum: 1.00 1.00 1.00 1.00
For example, given that the future operational
performance of the new compressor is 90% of the design
goal, the conditional probability of the comprehensive test
result being favorable is estimated to be 0.95, and the
conditional probability of the test results being not
favorable is estimated 0.05. That is to say, p(F|90) = 0.95,
and p(NF|90) = 0.05, where the | Means “given”.

Based on these conditional probabilities (reliability

estimates of the test results) and the choice of either
doing the test or not doing it, (a) calculate the revised
probabilities of four efficiency design goal levels being
met, and (b) estimate the value of performing the
comprehensive test to the typical user of a new
compressor unit.
Solution:
The two-stage decision tree diagram, which includes an initial
decision on whether to do the test, is shown in Figure 12-11. In
order to calculate the revised probabilities, we need to
determine the joint probabilities of each level of the efficiency
design goal being met and each test outcome occurring, as well
as the marginal probability of each test outcome. As an
example, the joint probabilities of the efficiency design goal
being met at the 90% level and each test outcome occurring
are calculated as follows:
p(90, F) = p(F|90).p(90) = (0.95)(0.25) = 0.2375;
p(90, NF) = p(NFǀ90).p(90) = (0.05)(0.25) = 0.0125.
The revised probabilities of each level of the design goal being
met [e.g., p(90) = p(90,F)/p(F) = 0.2375/0.6575 = 0.3612] are
also shown in Figure 12-11.
 Figure 12-11 Two-Stage Decision Tree (Example 12-12)
$1,086
PW(18%) E(PW)

90: 0.3612 $3,538 $1,278

est
T
No $2,029 70: 0.5171 1,614 835
n
D e s ig 50: 0.1141 -520 -59
$1,334 $2,029 New
=

Cur 30: 0.0076 -3,143 -24

75 rent
=
5 Des
0 .6 ign $0 $2,029
Do

l e: $0
ab
Te

v or
s

Fa
t

90: 0.0365 $3,538 $129

No 70: 0.1752 1,614 283
tF -$726
$1,334
0. avo
34 ra 50: 0.5109 -520 -266
25 bl e ign
Des
: $0 New -872
30: 0.2774 -3,143
Curr
=

e nt D -$726
esig $0
n $0
Joint and Marginal Probability (Example 12-12)

Level (%) of Joint Probabilities Marginal

Design Goal Met Favorable Not Favorable
Probabilities,
(F) (NF) p(L)

90 0.2375 0.0125 0.25

70 0.3400 0.0600 0.40
50 0.0750 0.1750 0.25
30 0.0050 0.0950 0.10
Marginal probabilities 0.6575 0.3425 1.00
of test outcome: (Sum)
The Revised Probabilities
Level (%) of Joint Probabilities
Design Goal Met Favorable Not Favorable
(F) (NF)

90 0.3612 0.0365
70 0.5171 0.1752
50 0.1141 0.5109
30 0.0076 0.2774
Sum 1.0000 1.0000