15.

053 Tuesday, May 8

z Decision Trees

The example in this lecture is a modification of the example in

Bertsimas and Freund: Data, Models, and Decisions.
1
Quote of the Day
Truly successful decision making relies on a balance

between deliberate and instinctive thinking.

Malcolm Gladwell
Blink: The Power of Thinking Without Thinking

Every decision has a cost. Do I have to make this

decision at all or can I move on to the next thing?
What we decided to leave out is almost as important
as what we put in.
Joshua Schachter

2
 Newsvendor Problem
z News vendor Phyllis orders newspapers for sale
ƒ ordered newspapers cost 20¢ each
ƒ she sells newspapers for 25¢ each
ƒ unsold newspapers are worthless
ƒ Prob(demand = j) = 1/5 for j = 6 to 10.

rij is the “net revenue of ordering i papers if the

demand is j.

rij = 25 i – 20 i = 5 i if i ≤ j

rij = 25 j – 20 i if i > j
3

The newsvendor problem is a classic model in operations management. It models a

situation in which one purchases goods in advance of sales, and unsold items are
valueless. But it could also be used for times in which unsold items are disposed of
for a small revenue.
 Papers Demanded
6 7 8 9 10

Papers Ordered
6 30 30 30 30 30
7 10 35 35 35 35
8 -10 15 40 40 40
9 -30 -5 20 45 45
10 -50 -25 0 25 50

rij is the value of ordering i papers if the demand is j.

We consider a simple case in which we order 6 to 10 newspapers and we sell 6 to 10

newspapers. Of course, one could order 5 or fewer or 11 or greater, but that would
clearly not be optimal.
 Dominated Actions
Papers Demanded
6 7 8 9 10
“Ordering i”
Papers Ordered
6 30 30 30 30 30
dominates
7 10 35 35 35 35
“ordering k” if
8 -10 15 40 40 40 rij ≥ rkj for all j.
9 -30 -5 20 45 45
10 -50 -25 0 25 50

5 25 25 25 25 25
11 -70 -45 -20 5 30

“Ordering 5” is dominated by “ordering 6”

“Ordering 11” is dominated by “ordering 10” 5

Ordering 5 is not as profitable as ordering 6. So, “ordering 5” is dominated by

“ordering 6”, regardless of what the demand is. Similarly, ordering “ordering 11” is
not as profitable as ordering 10.
 The Maximin Criterion
Papers Demanded
6 7 8 9 10 min (rij)
30
Papers Ordered
6 30 30 30 30 30
7 10 35 35 35 35 10
8 -10 15 40 40 40 -10
9 -30 -5 20 45 45 -30
10 -50 -25 0 25 50 -50

maximin criterion: choose the action with the

largest guaranteed return
maximize i min {rij : j = 6 to 10} = 30.
6

We now consider various metrics for measuring the quality of a solution. The
maximin criteria is very conservative, and is arguably very pessimistic. For each
number that you order, we consider the worst possible profit. This is stored in the
column min (rij). We then maximize over all possible orders. In the newsvendor
problem, this will always lead to ordering the minimum possible demand for
newspapers.
 The Maximax Criterion
Papers Demanded
6 7 8 9 10 max (rij)
30
Papers Ordered
6 30 30 30 30 30
7 10 35 35 35 35 35
8 -10 15 40 40 40 40
9 -30 -5 20 45 45 45
10 -50 -25 0 25 50 50

maximax criterion: choose the action with the

largest possible return
maximize i max {rij : j = 6 to 10} = 50.
7

The maximax criteria is really dumb. For each number that you order, we consider
the best possible profit. This is stored in the column max (rij). Then you would
order the number that maximizes the max. For the newsvendor problem, you would
always order the maximum number of newspapers that can be sold.
 The Minimax Regret Criterion
Papers Demanded
Papers Ordered 6 7 8 9 10

6 30 30 30 30 30
7 10 35 35 35 35
8 -10 15 40 40 40
9 -30 -5 20 45 45
10 -50 -25 0 25 50

max i rij 30 35 40 45 50

regret (i, j) = max k rkj - rij

regret (6, 6) = 0 regret (6, 7) = 5

8

The minimax regret criterion is a somewhat weird criterion. It assumes that you
will always look back with regret on what you do. So, if you order 6 papers and the
number of customers is 8, you have a regret of $.10 because had you ordered
exactly 8 you would have made $.10 more. If you order 10 papers, and the number
of customers is 8, then you have a regret of $.40 since had you ordered only 8
newspapers, you would have made $.40 more.

The regret is 0 if you happen to buy the number that is demanded.

 The Minimax Regret Criterion
Papers Demanded max
6 7 8 9 10 regret
20
Papers Ordered
6 30
0 30
5 30
10 30
15 30
20
7 20
10 35
0 35
5 35
10 35
15 20
8 -10
40 15
20 40
0 40
5 40
10 40
9 -30
60 40
-5 20 45
0 45
5 60
10 -50
80 -25
60 40
0 20 50
0 80

max i rij 30 35 40 45 50

minimax regret : choose the action that

minimizes the maximum possible regret
minimax regret = min i max j regret (i, j) = 20 9

Here we list all of the regrets. Then we create a column of maximum regrets. The
minimax regret criterion is to order the number that minimizes the max regret. In
this case, we could order 6 or 7 papers. In both cases, the maximum possible regret
is $.20.
 The amount of water is
just right. But the glass
is twice as big as it should
be.

too large a Just right

Wally glass 10

Wally doesn’t think in terms of the amount of water. As an engineer he focuses on

the glass itself, and wonders whether it is the right object for the task at hand.
 Expected Value
Papers Demanded
Papers Ordered 6 7 8 9 10

6 30 30 30 30 30
7 10 35 35 35 35
8 -10 15 40 40 40
9 -30 -5 20 45 45
10 -50 -25 0 25 50

Let p(j) = prob. that exactly j papers are demanded.

The expected value of ordering i papers is

E(i) = Σj p(j) rij

11

The expected value ( a horribly named word) is the average value one would obtain
if one could repeat this event an arbitrarily large number of times. It is also a
standard term used in probability, and is also called the “mean”.

It’s also a very useful way of assessing the return of an investment, and is arguably
the most common measure of performance used.
 Expected Value
Papers Demanded Expected
6 7 8 9 10 Value
30
Papers Ordered
6 30 30 30 30 30
7 10 35 35 35 35 30
8 -10 15 40 40 40 25
9 -30 -5 20 45 45 15
10 -50 -25 0 25 50 0

Suppose that p(j) = 1/5 for j = 6 to 10.

The expected value of ordering i papers is E(i) = Σj 1/5 rij

The maximum expected value is 30.

12

The expected value of ordering 6 or 7 papers is $.30, which is maximum.

 Another Example of Expected Value
Papers Demanded Expected
6 7 8 9 10 Value
60
Papers Ordered
6 60 60 60 60 60
7 40 70 70 70 70 64
8 20 50 80 80 80 62
9 0 30 60 90 90 54
10 -20 10 40 70 100 40

Suppose that p(j) = 1/5 for j = 6 to 10.

Each ordered newspaper is $.20
Each sold newspaper is $.30

13

If we changed the selling price from $.25 to $.30, the values rij would all change as
would the expected value of each number ordered.

Note that the expected value increases with the number of papers ordered and then
decreases. This always happens with newsvendor problems, and it is the key to the
solution approach.
 Back to the newsvendor problem
z Assumptions:
ƒ D = demand (random variable)
ƒ known probability p(d) that demand is d;
ƒ q is the amount ordered (decision variable);
ƒ c(q, d) = “profit” if q items are ordered and d items
are demanded .
ƒ co = Overstocking cost = ordering cost.
It is the unit cost of ordering too many.
In previous example this was 20¢.
ƒ cu = Understocking cost = price - ordering cost
It is the unit “cost” of ordering too few.
In previous example this was 25¢ - 20¢ = 5¢.
It is an opportunity cost.
14

The standard assumptions for a newsvendor problem is that you know the demand
distribution, you pay a cost for each item ordered, and you receive an amount for
each sold.

In order to determine the optimal solution, we focus on two values. The first value
is the overstocking cost co. This is the cost incurred for every newspaper purchased
that is not sold. In this case, it is 20¢.

We also keep track of an understocking cost, which is not a real cost. It is an

opportunity cost. It is how much additional revenue would have been attained by
another newspaper in the case that you order fewer papers than is demanded. In this
case, it is 25¢ - 20¢ = 5¢.
 Papers Demanded
Papers Ordered 6 7 8 9 10

6 30 30 30 30 30
7 10 35 35 35 35
8 -10 15 40 40 40
9 -30 -5 20 45 45
10 -50 -25 0 25 50

If you order at least as many as the demand, then

every extra paper costs you co = 20¢.

If you order at most as many as the demand, then

every fewer paper ordered costs you cu = 5¢. 15
 Applications of Newsvendor Problem
z Relevant to problems with perishable inventories
ƒ Newspapers, magazines,
ƒ restaurants with food orders
ƒ airline overbooking
• underbooking wastes a seat
• Overbooking results in having to make an
offer to a passenger to give up a seat
ƒ Flexible reimbursement plans: MIT staff can
put in money at the beginning of the year to
pay for medical expenses pretax
• any money left over at the end of the year in
the plan is lost.
16

The newsvendor problem arises in many different applications.

 Solving using marginal analysis
We will find the value of q that minimizes the
expected cost E(q).
Fact: for the
MV(q) = E(q) – E(q+1)
newsvendor
Expected Marginal problem, MV(q) is
Value Value nonincreasing.
5 50
Papers Ordered

6 60 10
Find the largest
7 64 4
value of q such
8 62 -2 MV(q) is positive.
9 54 -8
10 40 -14
17

Here are the expected values when the selling price of the newspaper was 30¢. We
then list the marginal values. The marginal value for purchasing q newspapers is
E(q) – E(q-1), is the value of the q-th newspaper to the expected return. We note
that MV(q) is a decreasing function. So, if we choose the largest value q for which
MV(q) is positive, we will also be maximizing E(q). In subsequent slides we focus
on MV(q).
 Marginal Analysis
We will find the largest value of q such that MV(q) is
positive
Case 1: D ≤ q-1 Value
-co decreases
Prob(D ≤ q-1)
by co.
MV(q) =
E(q) – E(q-1) A

Case 2: D ≥ q Value
cu increases
1 - Prob(D ≤ q-1) by cu.

If D ≤ q-1, then MV(q) = -co.

If D ≥ q, then MV(q) = cu.

18

To compute MV(q), we consider two cases. In the first case, the number demanded
is at most q-1, in which case the q-th newspaper is unsold. This leads to MV(q)
being –co. It occurs with probability Prob(D ≤ q-1).

In the second case, the number demanded is at least D. In this case, the q-th
newspaper is sold, and the revenue increases by cu. This occurs with probability
Prob(D ≥ q) = 1 - Prob(D ≤ q-1).
 Solving using marginal analysis

If D ≤ q-1, then MV(q) = -co.

If D ≥ q, then MV(q) = cu.

MV(q) = - c0 Prob(D ≤ q -1) + cu (1 – Prob(D ≤ q - 1))

MV(q) > 0 iff - c0 Prob(D ≤ q -1) + cu (1 – Prob(D ≤ q - 1)) > 0

iff Prob(D ≤ q - 1) (c0 + cu) - cu < 0

iff Prob(D ≤ q - 1) < cu / (c0 + cu)

19

We can now compute the expected value of MV(q) using the probabilities.

MV(q) has positive expected value when Prob(D ≤ q - 1) < cu / (c0 + cu).

So, to maximize E(q), we find the largest value of q such that MV(q) is positive,
which is also the largest value of q such that Prob(D ≤ q - 1) < cu / (c0 + cu). This is
the solution to the newsvendor problem.
 Solution for Newsvendor Problem

The optimum value q of newspapers to order is

the largest value of q such that

Prob(D ≤ q - 1) < cu / (c0 + cu)

Example 1. cu = 5¢
ƒ ordered cost: 20¢ each
co = 20¢
ƒ selling cost 25¢ each
cu / (c0 + cu) = 1/5
ƒ Prob(demand = j) = 1/5 for
j = 6 to 10. Prob(D ≤ 6 - 1) = 0
Prob(D ≤ 7 - 1) = 1/5
20

In our original instance, 6 is the largest value of q such that Prob(D ≤ q - 1) < cu /
(c0 + cu) = 1/5. Note that Prob(D ≤ 6 - 1) = 0, and Prob(D ≤ 7 - 1) = 1/5. (It turns
out that 7 is an alternative optimum solution.)
 Example 2 for Newsvendor Problem

Suppose that p(j) = 1/5 for j = 6 to 10.

Each ordered newspaper is $.20
Each sold newspaper is $.30
Expected
cu = 10¢ co = 20¢ Value

cu / (c0 + cu) = 1/3 5 50

Papers Ordered
6 60
Prob(D ≤ 7 - 1) = 1/5
7 64
Prob(D ≤ 8 - 1) = 2/5
8 62
Optimum order is 7 9 54
newspapers. 10 40 21

If each newspaper is sold for $.30, 7 is the largest value of q such that Prob(D ≤ q
1) < cu / (c0 + cu) = 10/30=1/3. Note that Prob(D ≤ 7 - 1) = 1/5, and Prob(D ≤ 8 - 1)
= 2/5.
 Examples of Optimal Ordering
Choose the greatest value of q such that

Prob(D ≤ q – 1 ) < cu / (c0 + cu).

Suppose that D takes on values from 1 to 100, and

that Prob(D ≤ q-1) = (q-1)/100 for each q.

Suppose the cost of selling a newspaper is 25¢.

What is the optimal order amount if
1. Cost of ordering = 20¢.
2. Cost of ordering = 10¢.
3. Cost of ordering = 5¢.

22

Case 1. cu / (c0 + cu) = 5/25 = 1/5. q = 20

Prob(D ≤ 20 - 1) = 19/100

Case 2. cu / (c0 + cu) = 15/25 = 3/5. q = 60

Prob(D ≤ 60 - 1) = 59/100

Case 3. cu / (c0 + cu) = 20/25 = 4/5. q = 80

Prob(D ≤ 80 - 1) = 79/100
Mental Break

23
 Juan Lee’s Decision Problem
It is January 10th, and Juan Lee is currently a fourth year
undergraduate in Management Science at Sloan. He has
decided to seek out a job as a consultant. He already has
received an offer from ABC consulting for $72,000 per year.
He has until February 1st to decide whether to accept the
offer. An old classmate of his, Mary Kumar, has told him
that she has recommended him highly to her consulting
firm, and feels that there is an excellent chance that they
would give him an offer for $80,000. However, they are not
prepared to make any decision until February 15th. If they
made him an offer, he would need to decide by March 1. He
also has the option of taking part in the consulting job fair in
the middle of March. He is fairly certain that he could get a
consulting job at that time, but is uncertain as to what he
would be paid.
24

This is an example used for Decision trees.

Juan’s Goal
z Assume here that Juan views all of these
consulting jobs as excellent opportunities, and
that the only differentiating factor is money.

z Juan has decided to maximize his expected

salary.

25
Juan’s probabilities
z Juan does not know how to evaluate his
“excellent chances” at Mary’s firm, nor does he
know what to expect from the consulting fair.

z Juan’s best guesses

ƒ Probability of getting Job at Mary’s firm: 60%
ƒ Different possibilities for consulting fair offers.
• $90,000: 10%
• $70,000: 50%
• $60,000: 40%

26
 A comment on probabilities
z Juan’s best guesses are not probabilities in the
sense of frequencies.

z They are called Bayesian or subjective

probabilities.

z Most decision problems rely on “best guesses”

both in terms of the outcomes (events) and the
probabilities for those events.
ƒ Could be based on detailed analysis
ƒ Could be a first guess for a number that is not
known at all.

27

One of the first realizations in using probability in decision making is that it is

rarely about repeatable events. It is much more an internal assessment about how
likely an event is. Juan believes that there is a 60% chance of getting a job at
Mary’s firm. He knows what a 60% probability is. It is the probability of getting
one of the numbers 1 to 6 on a fair 10-sided die. In his mind, he views the
probability of getting a job from Mary as equally likely.

Juan’s use of probability is called “subjective probability” and it is often referred to

as Bayesian probability, although Bayesian has additional meanings.

A natural question is: can one use subjective probabilities in making calculations on
probabilities? What this question means is “will the answers be useful and
meaningful?”

In our class we will use the subjective probabilities as though they are the
probabilities we are used to. In practice, using subjective probabilities is far better
than ignoring the uncertainties altogether. And the value of using the subjective
probabilities depends a lot on whether Juan is reasonably good at expressing his
lack of certainty, and whether he is reasonably consistent in assigning probabilities
to events.
 Decision Trees
z Method of organizing decisions over time in the
face of uncertainties

Decision nodes:
Accept ABC

A •Represented as boxes
•lines coming from the nodes
Reject ABC
represent different choices.
.6
Mary’s Firm Event nodes:
makes offer
B •Represented as circles
•lines coming from the nodes
.4
Mary’s Firm represent different outcomes.
makes no offer 28

Decision trees map out a decision problem over time.

There are two types of nodes of the decision tree: decision nodes and event nodes.
Decision nodes are represented as rectangles, and event nodes are represented as
ovals.

At a decision node, the decision maker needs to make a decision.

At an event node, the decision maker finds out the outcome of an event. This type
of node is different than we are used to in probability. Usually, we think of an event
as having a probability (and it does). But in a decision tree, the event node refers to
the decision maker finding out the outcome of an event, and then being able to use
this information in subsequent decisions.
 Step 1. Map out Juan Lee’s Decision Problem
Accept ABC

Accept Mary’s Offer

Offer
from Mary
C High Salary

Reject
E Medium Salary
Mary’s
A B offer
Low Salary
Reject
ABC
High Salary

D
No offer Medium Salary
from Mary

Low Salary 29

This diagram reflects a mapping out of Juan Lee’s decision.

First Juan has to make a decision on whether or not to accept ABC. If he accepts
ABC, the decision process ends for Juan. Otherwise, he next finds out whether
there is an offer or not from Mary. If there is an offer, he can accept it or not.

If Juan does not receive an offer from Mary or if receives one and rejects it, he then
moves to the job fair and learns which of three salaries he will get.
 Step 2. Assign Probabilities to Events
Accept ABC

Accept Mary’s Offer

Offer
from Mary
C .1 High Salary
.6

Reject
E .5 Medium Salary
Mary
A B
.4 Low Salary
Reject
ABC
.1 High Salary
.4
D
No offer .5 Medium Salary
from Mary
.4 Low Salary 30

We now assign a probability to events.

The probability that he gets an offer from Mary is .6.

The probability of getting the high, medium and low salaries are taken from a

previous slide.

 Step 3. Evaluate the end nodes

Accept ABC
72
Accept Mary’s Offer
Offer 80
from Mary
C .1 90
.6 High Salary

E .5 Medium Salary 70
Reject
Mary
A B 60
.4 Low Salary
Reject
ABC 90
.1 High Salary
.4
D 70
No offer .5 Medium Salary
from Mary
60
.4 Low Salary 31

The essence of calculating the values of nodes is to start at the end, and work backwards to the beginning. This
is because each end node reflects a possible decision path for Juan, including outcomes of events. We refer
to all of the nodes on the far right as “endnodes.”

For example, the second endnode from the top on the right has a value of 80. This node reflects the following
history:
1. Juan rejects ABC’s offer
2. Juan gets an offer from Mary
3. Juan accepts Mary’s offer.

As such, this is worth $80,000 to Juan, which is the value of the offer.

As another example, consider the fourth endnode with a value of 70. This reflects the following history:
1. Juan rejects ABC’s offer
2. Juan gets an offer from Mary
3. Juan rejects Mary’s offer
4. Juan goes to the job fair and is given a medium salary.

As such, this is worth $70, which is the value of the medium salary.

For some problems, one needs to look carefully at the full history to figure what the endnode is worth. For this
problem, it is more straightforward to compute the value of an endnode.
 Step 4. Work Backwards and Evaluate
Accept ABC
72
Accept Mary’s Offer
Offer 80 80
from Mary
C .1 90
.6 High Salary
68
E .5 Medium Salary 70
Reject
Mary
A B 60
.4 Low Salary
Reject
ABC 90
68 .1 High Salary
.4
D 70
No offer .5 Medium Salary
from Mary
60
.4 Low Salary 32

Once the endnodes are given values, we can work backwards (from right to left) and compute the
value of all other nodes.

There are two types of calculations, one for event nodes and one for decision nodes.

You can click on the nodes during the slide show and see the calculations in more detail. Then click
on the return button (curly arrow button on the slide) to return to this slide. Or if you forget, just
type in 32 and press enter.
The value of an event node is an expected value calculation. For example, the value of node E is
.1 x 90 + .5 x 70 + .4 x 60, which is the expected value at node E. This value is $68,000 or 68 on the
decision tree. What this means is the following: If Juan rejects the ABC offer, obtains an offer from
Mary, and then rejects the offer from Mary, he has an expected annual salary of $68,000. (Recall
once again that “expected value” is a technical term from probability. Juan has no chance at all of
getting exactly $68,000 in this example.)

A natural question is whether we need to calculate the value of an event node by calculating its
expected value. The answer is that this is the most useful way of approaching decision trees, but it
can run into difficulties if the decision maker is either risk preferring (very unusual) or risk averse
(very common). We will show how one can deal with risk aversion in the next lecture. For now, we
will work with expected values.

The other type of calculation is for a decision node. For example, consider node C, which is a
decision node. At a decision node, Juan will choose the best decision, that is, the one with the
highest expected return. For node C, he has a choice of accepting Mary’s offer for $80,000, or going
to the job fair and getting an expected income of $68,000. In this case, Juan will take the $80,000.
So, at a decision node, the value is the max of the values of the different decisions.
 Evaluate Node E

.1 90
High Salary
Take the 68 70
E .5 Medium Salary
expected value.
60
.4 Low Salary

.1 × 90

+ .5 × 70
+ .4 × 60
= 9 + 35 + 24 = 68
33

The next few slides treat the evaluations of different nodes. The button with the
curly arrow is the return button.
 Evaluate Node C

Accept Mary’s Offer

80
80
C .1 90
High Salary
For a decision 68
node, take the E .5 Medium Salary 70
Reject
best outcome Mary
leading from its 60
Low Salary
branches.

Max (80, 68) = 80

34
 Evaluate Nodes A and B
72

Offer 80
from Mary Exercise, determine the
C
.6 values for nodes B and A.

A B
Reject
ABC
68
.4
D
No offer
from Mary

35
 Key Aspects of a Decision Tree
Time flows from left to right
Branches from a decision node represent
decisions and take into account all decisions or
events leading to that node

z Example. Juan expects to get Mary’s job, but

does not know how he is viewed in the job
market. If he does get Mary’s offer, he believes
that he must be viewed highly. Accordingly, he
would adjust his probabilities of getting salaries
in the job fair as follows:
ƒ High: 60%
ƒ Medium: 30%
ƒ Low: 10%
36

This slide reflects an interesting use of subjective probability. People will often
change their probability assessments when they get new information. Here we
suppose that Juan will change his probability assessments if he gets the offer from
Mary’s firm. (If he is rejected by Mary’s firm, for some reason he does not change
his probability assessments. It may Juan’s ego, or it may reflect reality.)

So, we assume here that the probability of a high offer increases to 60%, and the
probability of medium and low offers also adjust.
 Illustration of new probabilities
Accept ABC

Accept Mary’s Offer

Offer
from Mary
C .6
.1 High Salary
.6

Reject
E .5
.3 Medium Salary
Mary
A B .1
.4 Low Salary
Reject
ABC
.1 High Salary
.4
D
No offer .5 Medium Salary
from Mary
.4 Low Salary 37

Note that the probability of .6 for a high salary occurs at event node E. The event node E reflects the
following history:
1. Juan has rejected the ABC offer
2. Juan received an offer from Mary
3. Juan rejected the offer from Mary

Under these circumstances, Juan believes the probability of a high salary is .6.
If you look at node D, you will note that the probability there of a high salary is .1, which is Juan’s
probability if he gets no offer from Mary.

In this case, the value of node E is

.6 x 90 + .3 x 70 + .1 x 60 = 54 + 21 + 6 = 81.

So, at node C, the optimal decision is to reject the offer from Mary’s firm. It’s ironic that without the
offer, Juan’s expected salary would be $68,000 and the offer from Mary’s firm would be very
attractive. But after getting the offer, it now is optimal to reject it.
Key Aspects of a Decision Tree
z Branches from an event node represent a set of
mutually exclusive and collectively exhaustive

z Final nodes in the tree have an associated value

z Values of other nodes are computed working

backwards. The value of an event node is the
expected value of its endpoints. The value of a
decision node is the highest value of its
endpoints.

38
 But the tree is only the beginning
z Typically in decision trees, there is a great deal of
uncertainty surrounding the numbers.
ƒ Decision Trees work well in such conditions

z This is an ideal time for sensitivity analysis the

“old fashioned way.”
ƒ One varies numbers and sees the effect
ƒ One can also look for changes in the data that
lead to changes in the decisions.

39

One of the great things about Decision trees is that it is easy to change the numbers
and get new results. This type of sensitivity analysis can do a variety of things:

1. It can give additional confidence in decisions

2. It can reveal how sensitive the decisions are to the probabilities or other
numbers in the tree. Usually, the probabilities are best guesses. If the decision
is highly sensitive to one of these guesses, it may be better to gather more
information. Similarly, if the decision is not changed when the probabilities
change, then you don’t need to worry about the accuracy of the guess, and there
is no need to gather more information.
3. It can give you a sense for whether to consider new options. If all decisions in
the decision tree appear to be poor, then there may be other options to consider.
 The Airfare Problem
You are trying to get the cheapest airfare that you
can. You just called up and found that the ticket
home will cost $400, and it cannot be refunded or
exchanged. You can also buy a ticket for $450,
which can be refunded for $430 (and thus costs
you $20). The price of tickets will change in one
week, and you will have one more chance to buy
a ticket. There is a 50% chance that the ticket
would cost $300, and a 50% chance that it would
cost $600. What should you do to minimize the
expected expense?

Exercise. Write the Airfare problem as a decision

tree, and solve it.
40

We leave this as an exercise.

The Air Fare Problem

41
Summary Conclusions
z Decision trees are a very useful technique for
mapping out sequential decisions under
uncertainty.

z They can be useful for mapping out other logic as

in the newsvendor problem.

42