THE HONG KONG POLYTECHNIC UNIVERSITY

DEPARTMENT OF APPLIED MATHEMATICS My background

Decision Analysis - AMA4840 • I got my Master from University of London and my
Lecturer: Cedric Yiu Doctorate from University of Oxford
Office: TU725 Telephone: 34008981
Email: [email protected]
• Since my graduation, I have been working on
Lecture time: research projects on various engineering dept.
Wed 16:30-18:20 (Y302) including aeronautics, mechanics, electrical and
Tutorial: Tue 10:30 -11:20 (Y302) electronics, civil, industrial and logistics.

Tutor: Yuhan Zhang

• I am working extensively on research projects on
Recording: Blackboard -> zoom
economics, finance, operations research and
Home page for course materials: Blackboard of PolyU.
1 optimization. 2

Pass requirement Assessment

Continuous Assessment 40%:
To pass this subject, you are
required to obtain Grade D or above 1. Assignment (15%: 3 assignments and 5% each)
2. One midterm test (25%)
in the Overall mark. (19 March 2025, room: ST111, lecture hours)

Final Examination 60%:

There will be a two hours examination

3 4
 References Teaching schedule
1. B. Render, R.M. Stair and M.E. Hanna, Quantitative
analysis for management, 9ed., 2006 Pearson 1. Introduction and probability review (4)
2. Basic Decision Analysis (4)
2. W.L. Winston, Introduction to probability models, 4ed.,
2004 Thomson 3. Utility theory (4)
4. Multi-attributes (4)
3. Pratt, J.W., Raiffa H. and Schlaifer R.,
5. Game theory (4)
Introduction to Statistical Decision Theory,
1995 MIT Press 6. Decision analysis with sampling (4)

4. Clemen, R.T. and Reilly, T. Making hard decisions with

DecisionTools, 2014, Cengage.
5 6

Calculus Optimization is Everywhere

Maximum • It is embedded in our daily language, and
part of the way we think.
– firms want to maximize profits, or to minimize
costs
Minimum – people want to make the best choices
– We want the highest quality at the lowest price
– When playing games, we want the best strategy
– When we have too much to do, we want to
Fermat, Newton, Euler, Lagrange, Gauss, optimize the use of our time
and more 7 8
 Overview of Quantitative Analysis
Quantitative Analysis Quantitative Analysis:
A scientific approach to managerial decision
§ Mathematical tools have been used for making whereby raw data are processed and
thousands of years. manipulated resulting in meaningful information.
§ QA can be applied to a wide variety of
problems. Raw Data
Quantitative Meaningful
Analysis Information
§ One must understand the specific
applicability of the technique, its
limitations, and its assumptions. Qualitative Factors:
Information that may be difficult to quantify but can affect
the decision-making process such as the weather, outcome
9 of nature. 10

Broadband Example Broadband Example

• Company A: 3M -> $120 per month Factors of your decision:
• Broadband usage
• Company B: 10M-> $180 per month
• The subjective preference.
• Company C: 100M-> $300 per month • Contract terms and conditions
What is your decision? • Broadband stability
Company A: 3/120=0.025 (M/$) • The relations between the actual download speed
and the bandwidth.
Company B: 10/180=0.056 (M/$)
Company C: 100/300=0.333 (M/$) Difficulty: How to formulate all the factors in
Is Company C the best choice? 11
mathematical way? 12
 Travelling Problem Travelling Problem
If the actual displacement is:
Three locations: A, B and C.
Location A (C) is closest (farthest) to your C  A Apartment B
4 2 3
apartment,
Starting from your apartment, the path is: Clearly, the optimal path may be
Apartment-> A->B->C Apartment-> B-> Apartment->A->C
Is it the optimal path for travelling to all the
locations?
13 14

Travelling Problem • Information affect your decision!!

• More Information, better decision.
C  A Apartment B If all the available information is known in
4 2 3
advance, does it mean that you can always give
If C is a museum (close at 5:00pm) and it is the best decision?
required to visit it in the first two places, then
No, owing to uncertainty (hidden information),
Apartment->A->C -> Apartment ->B (such as traffic)

If the travelling time don’t depend on • If we have various relevant information, how to
displacement, but depend on the actual distance use them and solve the problem in mathematical
and the traffic. What is your decision? way?
15 16
Why Study Decision Analysis (DA)? Why Study Decision Analysis (DA)?
• Based on the limited information, DA may helps us to • Instead of providing solutions, DA is perhaps best
obtain the best decision. thought of as simply an information source,
• Although decision analysis cannot improve your luck, it providing insight about the situation, uncertainty,
helps you to understand better the problems you face and objectives, and trade-offs, and possibly yielding a
hence make better decisions.
recommended course of action.
• It includes the structure of the problem as well as the
uncertainty and trade-off. • DA can be used to justify why a previously chosen
• Furthermore, it helps people deal with difficult decisions. action was appropriate.
Even we are not perfect decision marker, we can do better • Using DA in finance aspect, you may enjoy the
through more structure and guidance. same expected return (outcome) with the lowest
• DA offers guidance to normal people working on hard risk (unpleasant surprises).
decisions.
What is the best decision? Give the best outcome.
17 18

Define
the problem

Where is Decision Analysis used? Develop

a model

Almost everywhere!!
Acquire
• Managing research and development programs input data
• Negotiating for oil and gas leases
• Forecasting sales for new product
• Even in game strategy. Develop
a solution
DA can be used in any decision, but depend on whether it is worth
to use DA on a particular problem. Test
the solution
Example
It is cloudy outside. Should we collect the weather report to perform Analyze
the results
DA for bringing an umbrella or not?

19 Implement 20
the results
 Decision Analysis - Introduction
1.1 Decision Making Problems
§ We need to act in a world that is full of uncertainties.
§ ¨ Action
under Uncertainty
Þ Consequence
¨ State of World
An inventory problem
§ ¨ Decision maker’s preference on consequences
Þ Action A retailer is about to place an order for a number of units of
¨ judgments concerning the uncertain states
a perishable commodity which spoils if it is not sold by the end of
§ We can only act on the basis of what we know. The more the day on which it is stocked. The retailer does not know what
information, the better the decision. the demand for the item will be, but nevertheless decide on
a definite number of units to stock.
§ Decision Analysis : We want to minimize the loss caused by either overage or
¾ Mathematical analysis of the decision making shortage cost. If overage happen, should we give a discount to the
customer in the last hour(s) before closing?
problems under uncertainty.
21 22

An investment problem

An oil wildcatter who holds an option on a plot of land in

1.2 Decision Trees
an oil-producing region must decide whether to drill on the site Two essential characteristics:
before the option expires or to abandon his rights.
1. A choice, or in some cases a sequence of choices, must be made
The profitability of drilling will depend on a large number of unknowns: among various possible courses of action.

• Cost of drilling 2. This choice (or choices) will ultimately lead to some consequence,
• Amount of oil or gas discovered but the decision maker cannot be sure in advance what this
• Price at which oil or gas can be sold consequence will be because it depends not only on his or her
• and others choice(s) but on an unpredictable event or sequence of events.

His problem is further complicated by the fact that it is possible Such a kind of problem can be presented by a diagram known as a
to perform various tests or experiments that will yield a certain amount decision tree.
of information on the geophysical structure below the land.
23 24
 1.3 The Problem of Analysis A simplest tree
The problem can be represented by a decision tree:
Analysis of the Simplest Problems:
Mr Lee, a manufacturer, has experienced a serious decline in demand
for his product and will be forced to lay off a substantial portion of
his work force, etc, unless he can obtain a large order which
the XYZ Company is about to place with some suppliers.

To have a chance at obtaining this order, Lee has to incur

large expense for making samples and sending a team of engineers
to visit the Company.
Two separate considerations are required:
Three consequences: 1. His judgment concerning his chances of getting the order
C1 = Lay off work force = Do not try 2. A comparison of the cost of trying with the advantages
C2 = Cost of attempt and Lay off work force = Try and Do not obtain
C3 = Cost of attempt and Keep work force = Try and Obtain Furthermore, Mr Lee may quantify some of his reasoning.
25 26

Decision tree
The Oil Wildcatter’s Decision Tree

♥ Holding an option on a plot of land in an oil-producing region

(having the seismic sounding test made or not);

♥ Must decide whether to drill on the site before the

option expires or to abandon the right.

Nodes under decision maker’s control will be called choice nodes

and designated by squares; nodes not under decision maker’s
control will be called event nodes and designated by circles.

27 28
 Bases of decision: preference and judgment
• The problem becomes complicated if § There is in general no “objectively correct” solution to
any realistic decision problem.
there is time constraint or the variables vary
with time. § A “reasonable” decision must necessarily rely on
• The cost of drill depends on the solid nature, * Decision maker’s personal preference for consequences;
the depth required. * Decision maker’s personal judgment concerning
the chances of events.
• The labor cost, the amount of oil, etc.
• All the available information are required If the decision maker behaves reasonably, she or he will choose
a solution which is consistent with his or her personal preference
for DA. and personal judgment.

§ Not a pure “subjective” decision. All available objective evidence

29 will be taken into account in arriving at considered judgments. 30

Review of Probability Union and intersection

Union: The union of two events A and B is denoted by and defined
Sample space: The set of all possible outcomes from a statistical
by
experiment is called a sample space and denoted by S.
A È B = {x : x Î A or x Î B (or both)}
For example,
which occurs if at least one of A and B occurs. For example,
(i) S = {1, 2, 3, 4, 5, 6}
S = {1, 2, 3, 4, 5, 6} - A = {1, 2} - B = {1, 4, 5}, then
(ii) S = {D, N},
where D denotes a defective and N a non-defective; AÈ B = {1, 2, 4, 5}.

Intersection: The intersection of two events A and B is denoted

Events: A subset of a sample space S is called an event, e.g.
and defined by
S = {1, 2, 3, 4, 5, 6} - a sample space
A Ç B = {x : x Î A and x Î B }
A = {1, 2} - an event.
which occurs if both A and B occur. For example,
S = {1, 2, 3, 4, 5, 6} - A = {1, 2} - B = {1, 4, 5}, then
AÇ B = {1}.
31 32
 Definition of Probability Mutually exclusive
Probability: A probability P is a rule which assigns a value If AÇ B = Æ, then A and B are called to be mutually exclusive.
to each event A, denoted by P(A). A probability must satisfy
Example 1.1. In cutting a standard pack of cards, two events
the following three axioms:
Ei: card is an ace
(i) 0 £ P( A) £ 1; Ej: card is a king
(ii) P( S ) = 1 and P(f ) = 0; and are mutually exclusive. We have
(iii) If A Ç B = f , then P( A È B) = P( A) + P( B),
P( Ei È E j ) = P( Ei ) + P( E j ) =

Example, Toss a dice, S = {1, 2, 3, 4, 5, 6}, A = {1, 3} and B = {5}.

The formula
Then
P(A) = 2/6 = 1/3. P(B) = 1/6. P ( A È B ) = P ( A) + P ( B ),

is called the addition law (first law).

33 34

Conditional probability Example 1.2

Data of a carpet firm is given below.

Given that event E has occurred, the Conditional Probability Quality required
of event R is denoted and defined by Regular Special De-luxe
Home 30 20 10
Export 5 15 20
P( R Ç E )
P( R | E ) = (if P( E ) > 0).
P( E ) Then

P( R Ç E )
P( R | E ) = =
P( E )
P( S Ç H )
P( S | H ) = =
P( H )

35 36
 Independence between two events Prior and posterior probability
Two events A and B are said to be independent if When the unconditional probability P(R) is compare with
the conditional probability P(R|E), we refer to the former as a
P( A Ç B) = P( A) P( B)
which is equivalent to prior probability and to the latter a posterior probability.
P( B | A) = P( B) and P( A | B) = P( A) (if P( A) > 0, P( B) > 0)
Rewrite the definition of conditional probability P(R|E), we obtain
Example 1.3. Toss a fair dice and let A = {1, 3, 5}, B = {1, 2, 3}, The multiplication law (second law):
C = {1, 2}. Then P( R Ç E ) = P( R | E ) P( E ).
P({1,3}) 1
P( B | A) = = ¹ P( B ) = Example 1.2 (cont.)
P( A) 2
Hence A and B are dependent (not independent). But A and C are 5 5 40
independent since P( R Ç E ) = = ´ = P( R | E ) P( E ).
100 40 100
P( A Ç C ) = P({1}) = = P( A) P(C )

37 38

Third law Third law (cont.)

Assume that E1, E2, …, En are mutually exclusive (i.e., every pair of By the multiplication law,
Ei and Ej are mutually exclusive) and exhaustive
(i.e., S = E1 È E2 È  È En ) P( A Ç Ei ) = P( A | Ei ) P( Ei ) and hence
P( A) = P( A Ç E1 ) + P( A Ç E2 ) +  + P( A Ç En )
If A is any event, then
= P( A | E1 ) P( E1 ) + P( A | E2 ) P( E2 ) +
A=
 + P( A | En ) P( En ).
Note that AÇ E1, AÇ E2, …, AÇ En are mutually exclusive, by
- third law
the addition law, we have
P( A) =

39 40
 Bayes’ formula Example 1.4
(Decision analysis, G. Gregory, p. 28)
Assume that E1, E2, …, En are mutually exclusive and exhaustive and • Machines E and F make printed circuits;
that A is any other outcome. Then • E and F produces 2 and 4 per cent defective circuits, respectively;
P( Ei | A) P( A) = P( Ei Ç A) = P( A | Ei ) P( Ei ) • If combined, F making 3 printed circuits for every 2 made by E;

Thus What is the probability that a circuit randomly selected

P( A | Ei ) P( Ei ) from the mixture is defective?
P( Ei | A) =
P( A) Let D be the outcome that the circuit is defective. Then
P( A | Ei ) P( Ei ) P( E ) = P( F ) =
P( Ei | A) =
P( A | E1 ) P( E1 ) + P( A | E2 ) P( E2 ) +  + P( A | En ) P( En ) P( D | E ) = P( D | F ) =
--- Bayes’ formula. Also E and F are mutually exclusive and exhaustive. S = E È F.
Then
q It reverses the order of the conditional probability P( D ) = P( D Ç S ) = P( D Ç ( E È F )) =
q It gives a way of relating the posterior probability
to the prior probability. 41 42

(cont.) Suppose that a circuit is selected from the mixed

Example 1.5
output of the two machines, which is defective.
What is the probability that it came from machine E? (Decision making and forecasting, K.T. Marshall, p.13)
i.e., the posterior probability P(E|D). • one in every 1,000 people in the age group 55 to 65 has cancers;
• of those who do, 40% have serious cancers, 60% have light cancers;
The outcomes E and F are mutually exclusive and exhaustive. • assume that a not perfect diagnostic test has been developed;
Hence, by Bayes’ formula, we have
(1) when the test is administered to people having serious cancers,
P( E | D ) = 99% show a positive result,
(2) when administered to people with light cancers,
65% show positive;
(3) on those known not to have the disease 10% show positive;

If this test is administered widely on a regular basis to this age

The information that the circuit is defective has reduced the prior group, what fraction of tests indicate cancers is present?
probability of 0.4 that it came from machine E to the posterior P(cancers | test is positive) ?
probability of 0.25 that was defective and came from E ! 43 44
 Example 1.5 (cont.) Example 1.5 (cont.)
Since S = CÈ A3, (exhaustive) and CÇ A3 = Æ (exclusive),
Let
A1 = the event “serious cancers” P(C ) =
A2 = the event “light cancers”
Since A1, A2,, A3 are exclusive and exhaustive, we have, by third law,
A3 = the event “no cancers”
B = the event “a test result is positive” P( B) = P( B | A1 ) P( A1 ) + P( B | A2 ) P( A2 ) + P( B | A3 ) P( A3 )
C = the event “the person has cancers”
= 0.99 ´ 0.0004 + 0.65 ´ 0.0006 + 0.10 ´ 0.999
Find P(C|B) ? = 0.100686
What is P(C|B) ? By Bayes’ formula
Then
P( A1 ) = 0.0004, P( A2 ) = 0.0006, P( A3 ) = 0.9990 P( B | C ) P(C )
P(C | B) =
P ( B | A1 ) = 0.99, P ( B | A2 ) = 0.65, P( B | A3 ) = 0.10 P( B)
So we need to find P(B|C).
45 46

In fact
Example 1.5 (cont.)
P( B | C ) = P( B | A1 È A2 )
P( B Ç ( A1 È A2 ))
=
P( A1 È A2 )
P( B Ç A1 ) + P( B Ç A2 )
=
P( A1 È A2 )
P( B | A1 ) P( A1 ) + P( B | A2 ) P( A2 )
=
P(C )
0.99 ´ 0.0004 + 0.65 ´ 0.0006
= = 0.768
0.001
Thus
P( B | C ) P(C ) 0.768 ´ 0.001
P(C | B) = = = 0.007806.
P( B) 0.100686
The information that the tested person has cancers has increased
from the prior probability of 0.001 that it came from persons with
cancers to the posterior probability of 0.007806 ! 47