DecTree Supplement
DecTree Supplement
CONTENTS
1.1
1.2
1.3
1.4
1.5
1.6
Case Modules
Kendall Crab and Lobster, Inc .
Buying a House
The Acquisition of DSOFT
National Realty Investment Corporation
1.7
Exercises
PERHAPS THE MOST FUNDAMENTAL AND IMPORTANT TASK THAT A MANAGER FACES
CHAPTER 1
Decision Analysis
Bill's DecisionCriterion
Let us suppose, for the sake of simplicity, that Bill feels that all summer job opportunities (working for John, working for Vanessa's firm, or obtaining a summer job
through corporate recruiting at school) would offer Bill similar learning, networking,
and resume-building experiences. Therefore, we assume that Bill's only criterion on
which to differentiate between summer jobs is the summer salary, and that Bill obviously prefers a higher salary to a lower salary.
1.1
job offer. Second, if he were to reject John's offer, and Vanessa's firm were to offer him
a job in mid-November, he must then decide whether to accept Vanessa's offer or to
instead participate in the school's corporate summer recruiting program in January
and February.
These decisions are represented chronologically and in a systematic fashion in a
drawing called a decision tree. Bill's first decision concerns whether to accept or reject John's offer. A decision is represented with a small box that is called a decision
node, and each possible choice is represented as a line called a branch that emanates
from the decision node. Therefore, Bill's first decision is represented as shown in Figure 1.1. It is customary to write a brief description of the decision choice on the top
of each branch emanating from the decision node. Also, for futun rl'ftrenct', we have
given the node a label (in this case, the letter "A").
If Bill were to accept John's job offer, then there are no other dtcisi1ins or uncertainties Bill would need to consider. However, if he were to reject John's job offer, then Bill
would face the uncertainty of whether or not Vanessa's firm would subsl'lJlll'ntly offer
Bill a summer job. In a decision tree, an uncertain event is represl'lltl'd with a small circle called an event node, and each possible outcome of the event is rl'pn stnll'd ,is a line
(or branch) that emanates from the event node. Such an event nodt with its outcome
branches is shown in Figure 1.2, and is given the label "B." Again, it is nistom,iry to write
a brief description of the possible outcomes of the event above each outrnnw branch.
Unlike a decision node, where the decision-maker gets to sl'ltct which branch to
opt for, at an event node the decision-maker has no such choice. Ratlwr, onl' can think
that at an event node, "nature" or "fate" decides which outcome will t.1ktplace .
The outcome branches that emanate from an event node must rl'pnslnt a mutually
exclusive and collectively exhaustive set of possible events. By mutually l'xdusive , we
mean that no two outcomes could ever transpire at the same time. By rnlltctivcly exhaustive, we mean that the set of possible outcomes represents the entin r,mi-;1of possible outcomes. In other words, there is no probability that anothl'r no1H1prtstnted
outcome might occur. In our example, at this event node there arc two, and on Iy two, distinct outcomes that could occur: one outcome is that Vanessa's firm will offl'r Billa summer job, and the other outcome is that Vanessa's firm will not offer Bill ,1 summer job.
If Vanessa's firm were to make Bill a job offer, then Bill would subsl'qul'lltly have
to decide to accept or to reject the firm's job offer. In this case, and if Bill Wl'fl' to accept
the firm's job offer, then his summer job problem would be resolvl'd. If Bill w1rtto instead reject their offer, then Bill would then have to search for sumnwr l'mployrnent
through the school's corporate summer recruiting program. Tht d1cision lrt'l' shown in
Figure 1.3 represents these further possible eventualities, where thl' .idditional decision
FIGURE
1.1
Representation of a
decision node.
CHAPTER 1
Decision Analysis
FIGURE
1.2
Representation of an
event node.
~c<-
FIGURE
1.3
~c<.
Further
representation of the
decision tree.
,c,O
,c,o
~'I>
c"'
.._\:#"
.::,
.'I>"'
e,C~
e,C
'?"c
'?"c
node C represents the decision that Bill would face if he were to receive a summer job
offer from Vanessa 's firm.
Assigning Probabilities
Another aspect of constructing a decision tree is the assignment or determination of the
probability, i.e., the likelihood, that each of the various uncertain outcomes will transpire.
Let us suppose that Bill has visited the career services center at Sloan and has
gathered some summary data on summer salaries received by the previous class of
1.1
MBA students. Based on salaries paid to Sloan students who worked in the Sales and
Trading Departments at Vanessa's firm the previous summer, Bill has estimated that
Vanessa's firm would make offers of $14,000 for twelve weeks' work to summer
MBA students this coming summer.
Let us also suppose that we have gathered some data on the salary range for all
summer jobs that went to Sloan students last year, and that this data is conveniently
summarized in Table 1.1. The table shows five different summer salaries (based on
weekly salary) and the associated percentages of students who received this salary .
(The school did not have salary information for 5% of the students. In order to be
conservative, we assign these students a summer salary of $0.)
Suppose further that our own intuition has suggested that Table 1.1 is a good approximation of the likelihood that Bill would receive the indicated salaries if he were
to participate in the school's corporate summer recruiting. That is, we estimate that
there is roughly a 5% likelihood that Bill would be able to procure a summer job with
a salary of $21,600, and that there is roughly a 25% likelihood that Bill would be able
TABLE1.1
Weekly Salary
Distribution of
summer salaries.
$1,800
$1,400
$1,000
$500
$0
40%
25%
5%
FIGURE
1.4
Further
representation of the
decision tree.
$21,600
0.05
$16,800
0.25
0.40
$12,000
$6,000
0.25
$0
0.05
$21,600
0.05
$16,800
0.25
$12,000
0.40
0.25
$6,000
$0
0.05
---
--
---
-----
---------
--
CHAPTER 1
Decision Analysis
to procure a summer job with a salary of $16,800, etc. The now-expanded decision
tree for the problem is shown in Figure 1.4, which includes event nodes D and E for
the eventuality that Bill would participate in corporate summer recruiting if he were
not to receive a job offer from Vanessa's firm, or if he were to reject an offer from
Vanessa's firm. It is customary to write the probabilities of the various outcomes underneath their respective outcome branches, as is done in the figure.
Finally, let us estimate the likelihood that Vanessa's firm will offer Bill a job.
Without much thought, we might assign this outcome a probability of 0.50, that is,
there is a 50% likelihood that Vanessa's firm would offer Bill a summer job. On further reflection, we know that Vanessa was very impressed with Bill, and she sounded
certain that she wanted to hire him. However, very many of Bill's classmates are also
very talented (like him), and Bill has heard that competition for investment banking
jobs is in fact very intense. Based on these musings, let us assign the probability that
Bill would receive a summer job offer from Vanessa's firm to be 0.60. Therefore, the
likelihood that Bill would not receive a job offer from Vanessa's firm would then be
0.40. These two numbers are shown in the decision tree in Figure 1.5.
'!<.."<.
FIGURE
1.5
o"
~c<.
,;1;
'""
,.,o
Further
representation of the
decision tree.
c"'
.::.'/;"'
.,_,'-,~
c-<:i."'
c.,C'<
c.,C
'?-c
'?"
-"'er,, 4b.
~
~(>0
-f:?~
c'"'
Jc,,,,
),..,~
,,o
~-(}.
IS.::.'/;"'
~~
~&
O
~ ~s
I'~
0 .6
(>s
,;,-
$21,600
~-'(,,,_
0.05
$16,800
0.25
$12,000
-t,:,
0-7,:
f',,_A
"o
$6,000
0.25
$0
"Q;!-,
0.05
<)I').
(>&
:s;)
$21,600
0.05
$16,800
$12,000
0.40
$6,000
0.25
0.05
$0
1.1
FIGURE
1.6
The completed
decision tree.
$21,600
0.05
$16,800
0.25
$12,000
0.40
$6,000
0.25
$0
0.05
$21,600
0.05
$16,800
0.25
$12,000
0.40
0.25
$6,000
$0
0.05
FundamentalAspectsof DecisionTrees
Let us pause and look again at the decision tree as shown in Figure 1.6. Notice that
time in the decision tree flows from left to right, and the placement of the decision
nodes and the event nodes is logically consistent with the way events will play out
in reality. Any event or decision that must logically precede certain other events and
decisions is appropriately placed in the tree to reflect this logical dependence.
The tree has two decision nodes, namely node A and node C. Node A represents
the decision Bill must make soon: whether to accept or reject John's offer. Node C
represents the decision Bill might have to make in late November: whether to accept
or reject Vanessa's offer. The branches emanating from each decision node represent
all of the possible decisions under consideration at that point in time under the appropriate circumstances.
There are three event nodes in the tree, namely nodes B, D, and E. Node B represents the uncertain event of whether or not Bill will receive a job offer from
Vanessa's firm. Node D (and also Node E) represents the uncertain events governing
the school's corporate summer recruiting salaries. The branches emanating from
each event node represent a set of mutually exclusive and collectively exhaustive
outcomes from the event node . Furthermore, the sum of the probabilities of each outcome branch emanating from a given event node must sum to one . (This is because
the set of possible outcomes is collectively exhaustive .)
These important characteristics of a decision tree are summarized as follows:
CHAPTER 1
Decision Analysis
cision nodes and the event nodes is logically consistent with the way events
will play out in reality. Any event or decision that must logically precede certain other events and decisions is appropriately placed in the tree to reflect
this logical dependence.
2. The branches emanating from each decision node represent all of the possible decisions under consideration at that point in time under the appropriate circumstances.
3. The branches emanating from each event node represent a set of mutually
exclusive and collectively exhaustive outcomes of the event node.
4. The sum of the probabilities of each outcome branch emanating from a given
event node must sum to one.
5. Each and every "final" branch of the decision tree has a numerical value associated with it. This numerical value usually represents some measure of
monetary value, such as salary, revenue, cost, etc.
Notice that in the case of Bill's summer job decision, all of the numerical values
associated with the final branches in the decision tree are dollar figures of salaries,
which are inherently objective measures to work with. However, Bill might also wish
to consider subjective measures in making his decision. We have conveniently assumed for simplicity that the intangible benefits of his summer job options, such as
opportunities to learn, networking, resume-building, etc., would be the same at either his former employer, Vanessa's firm, or in any job offer he might receive through
the school's corporate summer recruiting. In reality, these subjective measures would
not be the same for all of Bill's possible options. Of course, another important subjective factor, which Bill might also consider, is the value of the time he would have
to spend in corporate summer recruiting. Although we will analyze the decision tree
ignoring all of these subjective measures, the value of Bill's time should at least be
considered when reviewing the conclusions afterward.
ferent job offer of $14,000, then his decision would be easy: he would take the higher
salary offer. However, in the presence of uncertainty, it is not necessarily obvious
how Bill might proceed.
Suppose, for example, that Bill were to reject John's offer, and that in midNovember he were to receive an offer of $14,000 from Vanessa's firm. He would then
be at node C of the decision tree. How would he go about deciding between obtaining a summer salary of $14,000 with certainty, and the distribution of possible
salaries he might obtain (with varying degrees of uncertainty) from participating in
the school's corporate summer recruiting? The criterion that most decision-makers
feel is most appropriate to use in this setting is to convert the distribution of possible
salaries to a single numerical value using the expected monetary value (EMV) of the
possible outcomes:
1.1
+ 0.25 x
X $12,000
$6,00() ~ 0.05
x $0
= $11,580 .
The EMV of a certain event is defined to be the monetary valul' of thl' l'Vcnt. For
example, suppose that Bill were to receive a job offer from Varwssa's firm, and that
he were to accept the job offer. Then the EMV of this choice would simply bl' $14,000.
Notice that the EMV of the choice to participate in corporatl rl'nuit ing is $11,580,
which is less than $14,000 (the EMV of accepting the offer from Va1wssa's firm), and
so under the EMV criterion, Bill would prefer the job offer from Vamss,1's firm to the
option of participating in corporate summer recruiting .
The EMV is one way to convert a group of possible outcoml's with mo1wtary values and probabilities to a single number that weighs each possibll' outniml' by its
probability. The EMV represents an "averaging" approach to uncl'rlainty . It is quite
intuitive, and is quite appropriate for a wide variety of decision pmbllms undl'r uncertainty. (However, there are cases where it is not necessarily thl' blst nwthod for
converting a group of possible outcomes to a single number. In Sl'ction 1.5, we discuss several aspects of the EMV criterion further.)
Using the EMV criterion, we can now "solve" the decision tree. Wl' do so by evaluating every event node using the EMV of the event node, and evaluating every de cision node by choosing that decision which has the best EMV. This is accomplished
by starting at the final branches of the tree, and then working "backwards" to the
starting node of the decision tree . For this reason, the process of solving the decision
tree is called folding back the decision tree. It is also occasionally referred to as
backwards induction. This process is illustrated in the following discussion.
Starting from any one of the "final" nodes of the decision tree, we proceed backwards. As we have already seen, the EMV of node Eis $11,580. It is customary to
write the EMV of an event node above the node, as is shown in Figure 1.7. Similarly,
the EMV of node Dis also $11,580, which we write above node D. This is also displayed in Figure 1.7.
We next examine decision node C, which corresponds to the event that Bill receives a job offer from Vanessa's firm. At this decision node, there are two choices.
The first choice is for Bill to accept the offer from Vanessa's firm, which has an EMV
of $14,000. The second choice is to reject the offer, and instead to participate in corporate summer recruiting, which has an EMV of $11,580. As the EMV of $11,580 is
less than the EMV of $14,000, it is better to choose the branch corresponding to accepting Vanessa 's offer. Pictorially, we show this by crossing off the inferior choice by
drawing two lines through the branch, and by writing the monetary value of the best
choice above the decision node . This is shown in Figure 1.7 as well.
We continue by evaluating event node B, which is the event node corresponding
to the event where Vanessa's firm either will or will not offer Bill a summer job. The
methodology we use is the same as evaluating the salary distributions from participating in corporate summer recruiting. We compute the EMV of the node by computing the
-----
--
~ -
- ---
--
10
i
$14,ooo
0~iz,<-
FIGURE
1.7
'"
Solution of the
decision tree.
ll,c,4'
.::,_'Ir~
C,i:
'l"'c
$21,600
$16,800
0.25
$12,000
$6,000
$0
$21,600
$16,800
0.25
$12,000
$6,000
$0
weighted average of the EMVs of each of the outcomes, weighted by the probabilities
corresponding to each of the outcomes. In this case, this means multiplying the probability of an offer (0.60)by the $14,000value of decision node C, then multiplying the probability of not receiving an offer from Vanessa's firm (0.40) times the EMV of node D,
which is $11,580,and then adding the two quantities. The calculations are:
EMV = 0.60 x $14,000 + 0.40 x $11,580 = $13,032.
This number is then placed above the node, as shown in Figure 1.7.
The last step in solving the decision tree is to evaluate the remaining node, which
is the first node of the tree. This is a decision node, and its evaluation is accomplished
by comparing the better of the two EMV values of the branches that emanate from it.
The upper branch, which corresponds to accepting John's offer, has an EMV of $12,000.
The lower branch, which corresponds to rejecting John's offer and proceeding onward,
has an EMV of $13,032. As this latter value is the highest, we cross off the branch corresponding to accepting John's offer, and place the EMV value of $13,032 above the initial node. The completed solution of the decision tree is shown in Figure 1.7.
Let us now look again at the solved decision tree and examine the "optimal decision strategy" under uncertainty. According to the solved tree, Bill should not accept John's job offer, i.e., he should reject John's job offer. This is shown at the first
decision node. Then, if Bill receives a job offer from Vanessa's firm, he should accept
this offer. This is shown at the second decision node. Of course, if he does not receive
a job offer from Vanessa's firm, he would then participate in the school's corporate
summer recruiting program. The EMV of John's optimal decision strategy is $13,032.
Summarizing, Bill's optimal decision strategy can be stated as follows:
1.1
11
If Vanessa's firm offers him a job, he should accept it. If Vanessa's firm does
not offer him a summer job, he should participate in the school's corporate
summer recruiting.
For an event node, compute the EMV of the node by computing the
weighted average of the EMV of each branch weighted by its probability.
Write this EMV number above the event node.
For a decision node, compute the EMV of the node by choosing that branch
emanating from the node with the best EMV value. Write this EMV number
above the decision node, and cross off those branches emanating from the
node with inferior EMV values by drawing a double line through them.
2. The decision tree is solved when all nodes have been evaluated.
3. The EMV of the optimal decision strategy is the EMV computed for the start-
sion strategy derived above, without a critical evaluation of the impact of the key
data assumptions that were made in the development of the decision tree model. For
example, consider the following data-related issues that we might want to address:
Issue 1: The probability that Vanessa's firm would offer Bill a summer job. We
have subjectively assumed that the probability that Vanessa's firm would offer
Bill a summer job to be 0.60. It would be wise to test how changes in this probability might affect the optimal decision strategy.
Issue 2: The cost of Bill's time and effort in participating in the school's corporate summer recruiting. We have implicitly assumed that the cost of Bill's
time and effort in participating in the school's corporate summer recruiting
would be zero. It would be wise to test how high the implicit cost of participating
~- - ---
. -
- --
---
--
------
------
~~
12
CHAPTER 1
Decision Analysis
FIGURE
1.8
Spreadsheet
representation of Bill
Sampras' summer
job problem.
Data
Value of John'soffer
$12,000
$14,000
0.60
$0
Distribution of Salaries from Recruiting
Weekly Salary
$1,800
$1.400
$1,000
$12,000
$6,000
$500
$0
$0
EMVofNodes
Nodes
EMV
A
B
$13,032
c
D
$14,000
$11,580
$11580
$13,032
25%
40%
25%
5%
1.1
13
In the spreadsheet representation of Figure 1.8, the data for the decision tree is
given in the upper part of the spreadsheet, and the "solution" of the spreadsheet is
computed in the lower part in the "EMV of Nodes" table. The computation of the
EMV of each node is performed automatically as a function of the data. For example,
we know that node E of the spreadsheet has its EMV computed as follows :
+ 0.25
0.05 X $21,600
EMV of node E =
+ 0.40 x $12,000 + 0.25 x $6,000 + 0.05 x $0
$16,800
= $11,580.
(EMV of node D)
and
EMV of node A = MAX{EMV of node B, $12,000}.
All of these formulas can be conveniently represented in a spreadsheet, and such a
spreadsheet is shown in Figure 1.8. Note that the EMV numbers for all of the nodes
in the spreadsheet correspond exactly to those computed "by hand" in the solution
of the decision tree shown in Figure 1.7.
We now show how the spreadsheet representation of the decision tree can be used
to study how the optimal decision strategy changes relative to the three key data issues
discussed above at the start of this subsection . To begin, consider the first issue , which
concerns the sensitivity of the optimal decision strategy to the value of the probability
that Vanessa's firm will offer Bill a summer job. Denote this probability by p, i.e.,
p
If we test a variety of values of pin the spreadsheet representation of the decision tree,
we will find that the optimal decision strategy (which is to reject John's job offer, and
to accept a job offer from Vanessa's firm if it is offered) remains the same for all values of p greater than or equal top = 0.174. Figure 1.9 shows the output of the spreadsheet when p = 0.18, for example, and notice that the EMV of node B is $12,016,
which is just barely above the threshold value of $12,000. For values of p at or below
p = 0.17, the EMV of node B becomes less than $12,000, which results in a new opti mal decision strategy of accepting John's job offer. We can conclude the following:
As long as the probability of Vanessa's firm offering Bill a job is 0.18 or larger,
then the optimal decision strategy will still be to reject John's offer and to accept
a summer job with Vanessa's firm if they offer it to him.
This is reassuring, as it is reasonable for Bill to be very confident that the probability
of Vanessa's firm offering him a summer job is surely greater than 0.18.
We next use the spreadsheet representation of the decision tree to study the second data assumption issue, which concerns the sensitivity of the optimal decision
strategy to the implicit cost to Bill (in terms of his time) of participating in the school's
corporate summer recruiting program. Denote this cost by c, i.e.,
c
14
CHAPTER 1
Decision Analysis
FIGURE
1.9
Output of the
spreadsheet of Bill
San1pras'sun1n1er
job problen1 when
the probability that
Vanessa's firn1will
n1ake Bill an offer is
0.18.
Data
Value of John's offer
$12,000
$14,000
0.18
$0
Distribution of Salaries from Recruiting
Weekly Salary
Total Summer Pay Percentage of Students
(based on 12 weeks) who Received this Salary
$1,800
$21,600
5%
$1,400
$16,800
25%
$1,000
$12,000
40%
$6,000
$500
25%
$0
$0
5%
EMVofNodes
Nodes
EMV
A
c
D
$12,016
$12,016
$14,000
$11,580
$11 580
1.1
15
FIGURE
1.10
Output of the
spreadsheet of Bill
Salllpras'sununer
job problelll if the
cost of Bill's time
spent participating
in corporate sullllller
recruiting is $2,578.
Data
Value of John's offer
$12,000
$14,000
0.60
$2,578
Distribution of Salaries from Recruiting
Weekly Salary
$1,800
Percentage of Students
(based on 12 weeks) who Received this Salary
$21,600
5%
$1,400
$16,800
25%
$1,000
$500
$12,000
$6,000
40%
$0
$0
5%
25%
EMVofNodes
Nodes
$12,001
$12,001
$14,000
D
E
$9,002
$9002
FIGURE1.11
Output of the
spreadsheet of Bill
Salllpras'sullllller
job problelll if
sullllller salaries
frolll recruiting were
$2,419higher.
EMV
A
B
Data
Value of John's offer
$12,000
$14,000
0.60
$0
Distribution of Salaries from Recruiting
Weekly Salary
Total Summer Pay Percentage of Students
(based on 12 weeks) who Received this Salary
$1,800
$24,019
5%
$1,400
$1,000
$19,219
$14,419
25%
40%
$500
$0
$8,419
25%
$2,419
5%
EMVofNodes
Nodes
EMV
$14,000
$14,000
$14,000
D
E
$13 999
$13,999
In order for Bill's optimal decision strategy to change, all of the possible summer corporate recruiting salaries of Table 1.1 would have to increase by more than $2,419.
This is also reassuring, as it is reasonable to anticipate that summer salaries from corporate summer recruiting in general would not be $2,419 higher this coming summer
than they were last summer.
16
CHAPTER 1
Decision Analysis
1.2
made. List all of the uncertain events in the problem and all of their possible outcomes.
2. Construct the basic decision tree by placing the decision nodes and the event
nodes in their chronological and logically consistent order.
3. Determine the probability of each of the possible outcomes of each of the uncertain events. Write these probabilities on the decision tree.
4. Determine the numerical values of each of the final branches of the decision
tree. Write these numerical values on the decision tree.
5. Solve the decision tree using the folding-back procedure:
1.3
17
(a) Start with the final branches of the decision tree, and evaluate each event
node and each decision node, as follows:
For an event node, compute the EMV of the node by computing the
weighted average of the EMV of each branch weighted by its probability. Write this EMV number above the event node.
For a decision node, compute the EMV of the node by choosing that
branch emanating from the node with the best EMV value. Write this
EMV number above the decision node and cross off those branches emanating from the node with inferior EMV values by drawing a double
line through them .
(b) The decision tree is solved when all nodes have been evaluated.
(c) The EMV of the optimal decision strategy is the EMV computed for the
starting branch of the tree.
6. Perform sensitivity analysis on all key data values. For each data value for
which the decision-maker lacks confidence, test how the optimal decision strategy will change relative to a change in the data value, one data value at a time.
As mentioned earlier, the solution of the decision tree and the sensitivity analysis procedure typically involve a number of mechanical arithmetic calculations. Unless the decision tree is small, it might be wise to construct a spreadsheet version of
the decision tree in order to perform these calculations automatically and quickly.
(And of course, a spreadsheet version of the model will also eliminate the likelihood
of making arithmetical errors!)
1.3
BIO-IMAGINGDEVELOPMENTSTRATEGIES
In 2004, the company Bio-Imaging, Incorporated was formed by James Bates, Scott
Tillman, and Michael Ford, in order to develop, produce, and market a new and potentially extremely beneficial tool in medical diagnosis. Scott Tillman and James
Bates were each recent graduates from Massachusetts Institute of Technology (MIT),
and Michael Ford was a professor of neurology at Massachusetts General Hospital
(MGH) . As part of his graduate studies at MIT, Scott had developed a new techniqu e
and a software package to proc ess MRI (magnetic resonance imaging) scans of brains
of patients using a personal computer. The software, using state of the art computer
graphics, would construct a three-dimensional picture of a patient's brain and could
be used to find the exact location of a brain lesion or a brain tumor, estimate its volume and shape, and even locate the centers in the brain that would be affected by the
tumor. Scott's work was an extension of earlier two-dimensional image processing
------
- -
--