Decision Trees

Decision trees represent a systematic analysis of a sequence of decisions. The tree is similar in
form to an event tree, however the head node, and possibly some other nodes, are decision nodes.
At these nodes the decision-maker chooses which option to take rather than having probabilities
assigned. Thus the probability of the preferred option is one and all others are zero. The other
nodes are referred to as chance nodes.

D
0.3
B

A 0.7 E
C
First Subsequent
Decision possible outcomes
Choices
or decisions

Node A in the diagram is the head node for the tree. It is also a decision node where the decision
maker needs to choose between B and C. After the decision maker makes the choice probability
of the preferred option is one and all others are zero

Node B in the diagram is a chance node. In this case the decision maker does not know which of
the child nodes (D and E in the example) will occur. The numbers on the diagram represent the
probabilities for the different outcomes.

Nodes C and E in the diagram are leaf nodes. These represent known outcomes.

Node D in the diagram is another decision node. It will have its own child nodes and the tree will
continue. This represents a multi-stage decision where the outcome of Decision A is known
before Decision D needs to be made.

Analysis

While construction of the tree proceeds from the head node to the leaf nodes, analysis of the tree
commences at the tips of the branches.

Commonly decisions will be made based on highest value. Value is usually measured as
Expected Monetary Value.

Management of Risk Decision Trees Page 1

 Strong market For a leaf node the value is known (or can be
calculated). The position of the node in the tree gives
details about its circumstances. From this the profit can
D $5M be calculated. In the diagram Node D will be profit
based on an assumption of a strong market and node E
0.3 will be profit based on the assumption of a weak market
B

0.7 E $1M For a chance node the EMV is the sum of the products
of the probabilities and the EMV’s of the respective
branch nodes. In the diagram B will have an EMV of
Weak market 0.3 * $5M + 0.7 * $1M = $2.2M

Start a Business
B For a decision node the decision maker simply picks the
$2.2M
best option. In the diagram this would be to start a
A business. The EMV for node A thus becomes $2.2M

C $2M

Get a Job

Utility

The EMV approach is good when applied to a lot of small decisions. However it has problems
when applied to one off decisions involving a loss of a significant proportion of a person's or
company's resources. For example not many people are willing to bet their home, even if they
have a better than 50% chance of getting more than twice as much.

To deal with this we can use a utility curve. We use this to convert the dollar amounts on a
decision tree into utility values before we calculate the EMV. We then proceed as normal except
that we are calculating an expected mean utility. Some examples of utility diagrams are shown in
the following figure.
Utility
Risk Averter
100

Risk Neutral

Risk Accepter

Monetary Value

Management of Risk Decision Trees Page 2

The risk averter is someone who is concerned about losing a significant amount of money and so
wants less exposure to loss than the EMV would suggest. The risk accepter is a gambler who has
their eyes on the gain to be made and is unconcerned about the loss. The straight line between is
risk neutral and uses the EMV criterion.

Example

A company is deciding whether to make a large or small resort on an island that it owns. The
payoff matrix is given below and the utility curve for the company is given by 51 + 1.4x - 0.01x2.

High demand (70% likely) Low demand (30% likely)

Large resort $70 million profit $30 million loss
Small resort $20 million profit $10 million profit

EMV = $40 million High demand 0.7 $70 M U = 100

EMU = 70
Large Resort
-$30 M U = 0
Low demand 0.3

High demand 0.7 $20 M U = 75

Small resort
EMV = $17 million $10 M U = 64
Low demand 0.3
EMU = 71.7

If we calculate the EMV for the decision tree we find that the Large resort is the preferred option.
However, if we calculate the expected mean utility we find that the small resort is the better
choice because the loss that a large resort would make under low demand is too big a problem for
the company.

The utility curve is determined by using the following decision tree.

Probability = p U($100,000) = 100

Gamble
Probability = 1- p U(-$50,000) = 0

Gift

U(gift) = 100 * p

An individual is asked whether he or she would prefer a gamble between two options (for
example winning $100,000 or losing $50,000) at a given probability, or a gift (for example
$30,000). The probability is adjusted until the individual finds both options equally apealling.
The utility of the gift is then calculated as 100 times the probability. For example if the
individual chooses a probability of p=0.6 for a gift of $30,000 then the utility of $30,000 is 60.
The experiment is repeated with gifts of different values. When enough data points have been

Management of Risk Decision Trees Page 3

collected a utility curve can be drawn up for that particular individual in those particular
circumstances.

Economists also talk about a concept of utility. This is the same thing as discussed here.
Namely, the utility is the value of something (in this case money) to a particular person under
particular circumstances, as opposed to the price, which is determined by supply and demand.
The underlying concept is that for someone (who is a risk averter) that has a large sum of money
a little extra is not worth as much as it would be if the extra amount was all that they had.

Bayes Theorem

If we have many events X1...Xj...Xn which are possible outcomes of a decision with initial
estimates of probabilities P(X1)...P(Xj)....P(Xn), and we perform an experiment that leads to
event E occurring
n
P(E) = j=1
P(E Xj ) P(Xj )

P(E X j ).P(Xj )
 (Xj E) =
P
n


i=1
P(E X i ).P(X i)

Example (Dandy and Warner, 1989)

The construction of a pre-stressed concrete bridge has been delayed. The pre-stressing strand has
rusted. We need to make a decision to replace the strand or not due to a possible reduction in
properties due to the rusting.

Replacement cost for strand is $20,000.

Failure of the bridge would result in replacement cost $750,000

States of Nature
Decision Alternatives Failure No Failure
___________________________________________________________________________
Replace Strand (a1) 0.00001 0.99999
Do not Replace Strand (a2) 0.10 0.9

If we carry out a decision analysis based on these figures and the EMV criterion we would
choose to replace the strand.

We have a third option a3 which involves postponing the decision to allow tests to be carried out.
(The tests cost $2000) Action a3 has two possible outcomes E1 - tests indicate no significant
reduction in strength E2 - tests indicate significant reduction in strength.

Management of Risk Decision Trees Page 4

The probability of the tests indicating no significant strength reduction in the two cases where the
bridge will or will not fail are:

P(E1|X3) = 0.95
P(E1|X4) = 0.09

The problems that we have in this example are that we do not know the probabilities of the two
different test outcomes )P(E1) and P(E2) or the probabilities that the bridge will or will not fail in
the case of either test result. That is we do not know P(X3|E1), etc. (notice that the symbols in the
conditional probability are reversed compared to what we are given).

To solve the first problem we use P(X3) = 0.9 and P(X4) = 0.1 with the first equation from Bayes'
Theorem above to obtain:

P(E1) = P(E1|X3).P(X3) + P(E1|X4).P(X4)

= 0.95 x 0.9 + 0.09 x 0.1
= 0.864

And the second equation from Bayes’ Theorem above to obtain:

0.95 0.9
| 0.99
0.864

0.09 0.1
| 0.01
0.864

Similarly, if E2 occurs,

P(E2 |X3) = 0.05

P(E2 | X4) = 0.91

P(E2) = 0.136

and
P(X3|E2) = 0.33

P(X4|E2) = 0.67

Management of Risk Decision Trees Page 5

 x1(0.99999) 20000
EMV = 20008 no failure

failure 770000
x2(0.00001)
a1
replace x3(0.90)
no failure 0
EMV= 75000

failure 750000
a2 x4(0.10)
don't replace
EMV = x1(0.99999) 22000
22008

772000
a3 a1
EMV=9500
tests x2(0.00001)
EMV =
9500 2000
(0.864) x3(0.99)
E1 a2 752000

x4(0.01)
EMV = $11 201
EMV =
22008 22000
x1(0.99999)
(0.136)
E2 a1 772000
x2(0.00001)

EMV =
504500
EMV=22008 x3(0.33) 2000
a2

x4(0.67) 752000

In summary
EMV(replace) = $20,008
EMV(don't replace) = $75,000

If E1 is the outcome of the test option a2 is the preferred option. If E2 is the outcome of the test
then it is better to replace the strand (option a1). This gives an EMV (a3) = (0.136*22,008) +
(0.864 x 9500) = $11,201.

This indicates that carrying out the tests is the best option.

References
Cross, J. (1997), Risk Management, The University of New South Wales.
Dandy, G. C. and Warner, R. F. (1989), Planning and Design of Engineering Systems,
Unwin Hyman, London.

Management of Risk Decision Trees Page 6

 Tutorial on Decision Trees

1. (after Meredith et al., 1973) Consider the problem encountered by a contractor who is
performing some work with special equipment in a river flats area that has been subjected
in the past to high-water conditions and occasional destructive flooding. There is a 4-
month period during which there is no use for the equipment either on this job or on
others. The contractor can keep the equipment on the job in the river flats, or else move it
out, store it, and then move it back, at a total cost of $1800.

2. A construction company needs to erect a temporary access bridge to a construction site in

area of known seismic activity. They must decide what degree of earthquake resistance
should be incorporated in the design to minimize total costs associated with seismic
design and damages. The following information is available.

(a) 50 years of earthquake records. 40 earthquakes have occurred.

(b) The cost of making the bridge earthquake resistant.

Earthquake No. in past 50 years Cost of Resistant

Magnitude Design
0-4 24 $1000
4-7 12 $3000
>7 4 $6000

(c) the cost of damage of an earthquake occurs,

Earthquake Magnitude Earthquake Cost of

Designed for Magnitude Occurring Damage ($)
0-4 3000
None 4-7 5000
>7 8000
0-4 0
0-4 4-7 3000
>7 7000
4-7 0-7 0
>7 4000
>7 any 0

3. A bank classifies 70% of their credit card customers as having a good credit rating and the
other 30% as having a poor credit rating. Customers with good credit rating allow their
accounts to go overdue 10% of the time on average, while customers with a bad credit
rating allow their accounts to go overdue 60% of the time on average. What percentage
of overdue credit card accounts do customers with poor credit ratings hold?

Management of Risk Decision Trees Page 7

 4. A group of scientists have written a computer code that they believe will unlock the
secrets of the universe. However even on the fastest computers it runs quite slowly and
will take 4 years to finish executing. If they are the first to solve the problem they will
receive a prize of $6,000,000. They estimate that the code has a 90% chance of solving
the problem. They have heard that there is going to be a revolution in computers in the
next year that will double the speed of computers, without increasing their cost. They
think that this is 40% likely. If this does not happen then a computer bought in a year’s
time will be exactly the same as today’s computer (this happens because at this time
silicon chip technology is at saturation point). The scientists also have the option of
further developing the code. They estimate that there is a 70% chance that in one year’s
time they can halve the execution time of their code on a current computer (with
corresponding results on newer computers). A current super computer costs $2,000,000.
Development of the new code will cost $1,000,000. Estimated probabilities of other
groups possible earliest finishing times are:

Without computer With computer revolution

revolution
less than two years 0.1 0.1
between two and three years 0.2 0.2
between three and four years 0.3 0.4
between four and five years 0.2 0.25
longer than five years 0.2. 0.05

If another group finishes first the group making the decision get nothing (but bills). If they begin
executing on a current computer and then purchase a faster computer then the code will have to
start executing from scratch.

Draw a desicion tree and determine what the scientists should do.

5. The Port of Wattalil Harbour needs some more space for its container terminal. There is a
bay in the harbour that can be used for the terminal, but it will need to be deepened. The
cost of this deepening will depend on the hardness of the sand at the bottom of the
harbour and the excavation method used. If dredging is used then the cost of excavating
the sand will be $200 000 if it is soft, and $700 000 if it is hard. If a dragline is used then
the cost of excavating the sand will be $300 000 if it is soft, and $400 000 if it is hard.
The hardness of the sand is unknown, but the engineer estimates that there is a 70%
chance that it will be soft.
It is possible for the engineer to do some SPT testing to see if the sand is hard or soft. If
he does this test and the sand is hard then there is a 90% probability that high SPT results
will be obtained and 10% chance that low SPT results will be obtained while if the sand is
soft then there is a 10% probability that high SPT results will be obtained and 90% chance
that low SPT results will be obtained.
What is the value of the test to the engineer?

Management of Risk Decision Trees Page 8