Decision Making and Problem Solving PDF

Decision making and
problem solving –
Lecture 1
• Review of basic probability
• Monte Carlo simulation
Liesiö, Punkka, Salo, Vilkkumaa

Why probabilities?
q Decisions are often made under uncertainty
q “How many metro drivers should be recruited = trained, when future traffic is
uncertain?”
q Probability theory dominates the modeling of uncertainty in
decision analysis
– Well established rules for computations
– Understandable
– Other models (e.g., evidence theory, fuzzy sets) are not covered here
q Learning objective: refresh memory about probability theory and
calculations
28.3.2018
2
The sample space
q Sample space S = set of all possible outcomes
q Examples:
– A coin toss: S = {H,T}
– Two coin tosses: S = {HH, TT, TH, HT}
– Number of rainy days in Helsinki in 2018: S={1,…,366}
– Grades from four courses: S=G × G × G × G=G4, where G={0,…,5}
– Average m2-price for apartments in Helsinki area next year S = [0,∞) euros
28.3.2018
3
Simple events and events
q Simple event: an individual outcome from S
– A coin toss: T
– Two coin tosses: TT
– Number of rainy days in Helsinki in 2018: 180
– Grades from four courses: (4, 5, 3, 4)
– Average m2-price for apartments in Helsinki in 2019: 4000 €
q Event: a collection of one or more outcomes (i.e., a subset of the
sample space: E⊆S)
– Two coin tosses: First toss tails, E={TT, TH}
– Number of rainy days in Helsinki in 2018: Less than 100, E={0,…,99}
– Grades from four courses: Average at least 4.0, = ∈ ∑ ≥ 4.0
– Average m2-price for apartments in Helsinki in 2019: Above 4000€, E=(4000, ∞)
28.3.2018
4
Events derived from events:
Complement, union, and intersection
q Complement Ac of A = all outcomes in S that are
not in A
q Union ∪ of two events A and B = all
outcomes that are in A or B (or both) S
q Intersection ∩ = all outcomes that are in both
A B
events
q A and B with no common outcomes are mutually
exclusive
q A and B are collectively exhaustive if ∪ =
28.3.2018
5
Events derived from events: Laws of
set algebra
Commutative laws: ∪ = ∪ , ∩ = ∩
Associative laws: ∪ ∪ = ∪ ∪ , ∩ ∩ = ∩ ∩ ,
Distributive laws: ∪ ∩ = ∩ ∪ ∩ , ∩ ∪ = ∪ ∩ ∪
DeMorgan’s laws: ∪ = ∪ , ∩ = ∩
S
A B
C
28.3.2018
6
Probability measure
q Definition: Probability P is a function that maps all events A onto
real numbers and satisfies the following three axioms:
1. P(S)=1
2. 0 ≤ P(A) ≤ 1
3. If A and B are mutually exclusive (i.e., ∩ = ∅) then ∪ =
+ ( )
28.3.2018
7
Properties of probability (measures)
q From the three axioms it follows that
S
I. (∅)=0
II. If ⊆ , then ≤ ( ) Young Young
and Restless
restless
III. ( )=1− ( )
Neither
IV. ∪ = + − ( ∩ ) young nor
restless
q In a given population, 30% of people are young, 15% are restless, and 7%
are both young and restless. A person is randomly selected from this
population. What is the chance that this person is
– Not young? 1. 30% 2. 55% 3. 70%
– Young but not restless? 1. 7% 2. 15% 3. 23%
– Young, restless or both? 1. 38% 2. 45% 3. 62%
http://presemo.aalto.fi/antti1/ 28.3.2018
8
Independence
Definition: Two events A and B are independent if
∩ = ( )
q A person is randomly selected

from the population on the right. S
Young
Young and Restless
(30%) Restless (15%)
q Are events ”the person is Neither
(7%)
young” and ”the person is young nor

restless
restless” independent?
q No: 0.07 ≠ 0.3 × 0.15
28.3.2018
9
Conditional probability
Definition: Conditional probability P(A|B) of A
given that B has occurred is
∩
≜ .
Note: If A and B are independent, the probability

of A does not depend on whether B has
occurred or not: Source: Wikipedia
( ∩ ) ( )
= = = .
( ) ( )
28.3.2018
10
Joint probability vs. conditional
probability Joint probability
Example: Wheat demand
Rye demand Low High Sum

q A farmer is trying to decide on a farming strategy for
Low 0.05 0.4 0.45
next year. Experts have made the following forecasts
about the demand for the farmer’s products. High 0.3 0.25 0.55
q Questions: Sum 0.35 0.65 1
– What is the probability of high wheat demand?

1. 40% 2. 65% 3. 134% Conditional probability
– What is the probability of low rye demand?
Wheat demand
1. 11% 2. 35% 3. 45%
– What is the (conditional) probability of high wheat demand, if rye Rye demand Low High Sum
demand is low? Low 0.11 0.89 1
1. 40% 2. 55% 3. 89%
High 0.55 0.45 1
– Are the demands independent?
1. Yes 2. No Sum 0.66 1.34
http://presemo.aalto.fi/2134l0102 28.3.2018
11
Law of total probability
q If E1,…,En are mutually exclusive and A = ⋃ , then
P(A)=P(A|E1)P(E1)+…+P(A|En)P(En)
q Most frequent use of this law:

– Probabilities P(A|B), P(A|Bc), and P(B) are known
– These can be used to compute P(A)=P(A|B)P(B)+P(A|Bc)P(Bc)
28.3.2018
12
Bayes’ rule
( | ) ( )
q Bayes’ rule: =
( )
q Follows from
( ∩ ) ( ∩ )
1. The definition of conditional probabilty: =
( )
, =
( )
,
2. Commutative laws: ∩ = ∩ .
28.3.2018
13
Bayes’ rule
Example:
q The probability of a fire in a certain building is 1/10000 any given day.
q An alarm goes off whenever there is an actual fire, but also once in every 200 days for
no reason.
q Suppose the alarm goes off. What is the probability that there is a fire?
Solution:
q F=Fire, Fc=No fire, A=Alarm, Ac=No alarm
q P(F)=0.0001 P(Fc)=0.9999, P(A|F)=1, P(A|Fc)=0.005
Law of total probability: P(A)=P(A|F)P(F)+P(A|Fc) P(Fc)=0.0051

( | ) ( ) .
Bayes: = = ≈ 2%
( ) .
28.3.2018
14
Random variables
q A random variable is a mapping from sample space S to real
numbers (discrete or continuous scale)
q The probability measure P on the sample space defines a

probability distribution for these real numbers
q Probability distribution can be represented by

– Probability mass (discrete) / density (continuous) function
– Cumulative distribution function
28.3.2018
15
Probability mass/density function (PMF
& PDF) 0.4
0.35
q PMF of a discrete random variable is fX(t)

0.3 X~Bin(5,0.3)
0.25
such that
f X(t)
0.2
– fX(t)=P({s ∈ S|X(s)=t}) = probability

0.15
0.1
– ∑ ∈( , ] = ( ∈ ( , ]) = probability 0.05
0
0 1 2 3 4 5
t
q PDF of a continuous random variable is fX(t) 1.8
1.6
such that 1.4
1.2 X~N(0.0.5 2)
– fX(t) is NOT a probability 1
f X(t)
0.8
– ∫ = ∈ , is a probability 0.6
0.4
0.2
0
-1 -0.5 a b 0 0.5 1
t
28.3.2018
16
Cumulative distribution function (CDF)
1
q The CDF of random variable X is 0.8
= ∈ | ( )≤
0.6
FX(t)
0.4
(often = ( ≤ )) 0.2
q Properties 0
-1 0 1 2
t
3 4 5 6
– FX is non-decreasing 1
– FX(t) approaches 0 (1) when t decreases

0.9
0.8
(increases) 0.7
0.6
– P(X>t)=1-FX(t)
FX(t)
0.5
0.4
– P(a<X≤b)= FX(b)- FX(a) 0.3
0.2
0.1
0
-1 -0.5 a b 0 0.5 1
t
28.3.2018
17
Expected value
• The expected value of a random variable is the weighted average of all possible
values, where the weights represent probability mass / density at these values
Discrete X Continuous X
= ( ) =
• A function g(X) of random varibale X is itself a random variable, whereby
( ) = ( ) ( ) ( ) = ( )
28.3.2018
18
Expected value: Properties
q If ,…, and =∑ are random variables, then
= [ ]
q If random variable Y=aX+b where a and b are constants, then

= +
q NB! In general, E[g(X)]=g(E[X]) does NOT hold:

– Let ∈ {0,1} with P(X=1)=0.7. Then,
= 0.3 0 + 0.7 1 = 0.7,
= 0.3 0 + 0.7 1 = 0.7 ≠ 0.49 = ( ) .
28.3.2018
19
Random variables vs. sample space
q Models are often built by directly defining distributions (PDF/PMF or CDF)
rather than starting with the sample space
– Cf. alternative models for coin toss:
1. Sample space is S={H,T} and its probability measure P(s)=0.5 for all ∈
2. PMF is given by fX(t)=0.5, t ∈{0,1} and fX(t)=0 elsewhere
q Computational rules that apply to event probabilities also apply when these
probabilities are represented by distributions
q Detailed descriptions about the properties and common uses of different
kinds of discrete and continuous distributions are widely documented
– Elementary statistics books
– Wikipedia
28.3.2018
20
Binomial distribution
q n independent binary (0/1, no/yes) trials,
each with success probability p=P(X=1)
q The number X ~ Bin(n,p) of successful
trials is a random variable that follows the
binomial distribution with parameters n
and p
q PMF: = = = (1 − )
q Expected value E[X]=np
q Variance Var[X]=np(1-p)
Source: Wikipedia
28.3.2018
21
Other common discrete distributions
q Bernoulli distribution
– If X ∈{0,1} is the result of a single binary trial with
success probability p, then X~Bernoulli(p).
– = (1 − )
q Geometric distribution
– If X ∈{1,2,3,…} is the number of Bernoulli trials needed to
get the first success, then X~Geom(p).
– = (1 − )
q Poisson distribution
– Let X ∈{1,2,3,…} be the number of times that an event
occurs during a fixed time interval such that (i) the
average occurence rate is known and (ii) events occur
independently of the last event time. Then, X~Poisson( ).
– =
!
28.3.2018
22
Source: Wikipedia
Uniform distribution
q Let X ∈[a,b] such that each real value
within the interval has equal probability.
Then, X~Uni(a,b)
, for ≤ ≤
q =
0, otherwise
q =
q = ( − )
Source: Wikipedia
28.3.2018
23
Normal distribution N( , )
( )
q =
q = , =
q The most common distribution for
continuous random variables
q Central limit theorem: Let X1,…,Xn be

independent and identically distributed
random variables with E[Xi]= and
Var[Xi]= . Then,
∑
~ , .
Source: Wikipedia
28.3.2018
24
Other common continuous
distributions
q Log-normal distribution: if X~N( , ), then
eX~LogN( , )
q Exponential distribution Exp( ): describes

the time between events in a Poisson
process with event occurrence rate
q Beta distribution Beta(α,β): distribution for

X∈[0,1] that can take various forms
28.3.2018
25
Why Monte Carlo simulation?
q When probabilitistic models are used to support decision making, alternative
decisions often need to be described by ̒performance
̒ indices̕ such as
– Expected values – e.g., expected revenue from launching a new product to the
market
– Probabilities of events – e.g., the probability that the revenue is below 100k€
q It may be difficult, time-consuming or impossible to calculate such measures

analytically
q Monte Carlo simulation:

– Use of a computer program to generate samples from the probability model
– Estimation of expected values and event probabilites from these samples
28.3.2018
26
Monte Carlo simulation of a probability
model
Probability model Monte Carlo simulation
• Random variable X~fX • Sample (x1,…,xn) from fX
∑ ( )
∈ {1, … , }| ∈( , )
( < ≤ )
28.3.2018
27
Uni(0,1) distribution in MC – discrete
random variables
q Some softwares only generate random numbers from Uni(0,1)-distribution
q Samples from Uni(0,1) can, however, be transformed into samples from many
other distributions
U~Uni(0,1) X~fX
1
q Discrete distribution: = 0.4565
X=3 =1
0.9 Demand x Prob. fX of
• Let ∈ ,…, such that = = 0.8910 =2 / week demand
X=2
= = . = 0.3254 0.7 =1 0 0.3
• Divide interval [0,1] into n segments ⋮ ⋮
1 0.4
of lengths , … , . X=1
• Sample values from Uni(0,1). 2 0.2
• Transform the sample: If ∈ 0.3
3 0.1
[∑ ,∑ ) where = 0, X=0
then = .
28.3.2018
28
Uni(0,1) distribution in MC – continuous
random variables
q Assume that the CDF of random variable X has an inverse function . Then,
the random variable = ( ) where U~Uni(0,1) follows the same distribution
as X:
= ≤ = ( )≤ = ≤ =
q Continuous distribution: U~Uni(0,1) X~fX 2500
• Let ~ (CDF) = 945.4

2000 f =N(1000,500 2)
= 0.4565 X
• Sample values from Uni(0,1). = 0.8910 = 1615.9 1500
• Transform the sample: = ( ) = 0.3254 = 773.7
F-1(u)
1000
⋮ ⋮
X
500
-500
0 0.2 0.4 0.6 0.8 1
u
28.3.2018
29
Monte Carlo simulation in Excel
VLOOKUP looks for the cell

value in the 1st column of the STDEV.S(E8:E207)
table. The value in the 3rd
column of the table is
returned to the current cell.
AVERAGE(H7:H206)
RAND() generates a random

number from Uni(0,1)
28.3.2018
30
Monte Carlo simulation in Matlab
28.3.2018
31
Monte Carlo simulation in Matlab
q Statistics and Machine Learning Toolbox makes it easy to
generate numbers from various distributions
q E.g.,
– Y=normrnd(mu,sigma,m,n): m×n-array of X~N(mu,sigma)
– Y=betarnd(A,B,m,n) : m×n-array of X~Beta(A,B)
– Y=lognrnd(mu,sigma,m,n) : m×n-array of X~LogN(mu,sigma)
– Y=binornd(N,P,m,n) : m×n-array of X~Bin(N,P)
– …
28.3.2018
32
Summary
q Probability is the dominant way of capturing uncertainty in decision models
q Well-established computational rules provide means to derive probabilities of

events from those of other events
– Conditional probability, law of total probability, Bayes’ rule
q To support decision making, probabilistic models are often used to compute

performance indices (expected values, probabilities of events, etc.)
q Such indices can easily be computed through Monte Carlo simulation
28.3.2018
33
Decision making and
problem solving –
Lecture 2
• Decision trees
• Elicitation of probabilities
• Common biases in probability assessment

Motivation
q Last time:
– Probability is the dominant way of capturing uncertainty in decision models
– Well-established rules to derive probabilities of events from those of other events
o Conditional probability
o Law of total probability
o Bayes’ rule
q This time:
– How to build a probability-based model to support decision-making under
uncertainty?
– How to elicitate the probabilities needed for these models?
– What kinds of biases are often associated with probability elicitation?
28.3.2018
2
Decision-making under uncertainty
A decision problem under uncertainty is Corresponding components in a
characterized by: probability model:
q States of nature:
– Mutually exclusive and collectively exhaustive future Sample space and simple events
events
– Uncontrollable by the decision-maker (DM)
q Decision alternatives:
– Possible strategies that the DM can employ Random variables
q Resulting consequences:
– Outcomes associated with each decision alternative
in each state of nature Realizations of random variables
28.3.2018
3
Decision trees
q Decision-making under uncertainty can be It will rain, p=0.4 4
modeled by a decision tree
q Decision trees consist of
– Decision nodes (squares) – DM can choose which arc to follow Take an umbrella 5
– Chance nodes (circles; cf. states of nature) – chance represented by It will not rain, p=0.6
probabilities dictates which arc will be followed (states of nature). The
probabilities following a chance node must sum up to 1
It will rain, p=0.4
– Consequence nodes (triangles; resulting consequences) – at the end of Do not take an 0
the tree; describe the consequence (e.g., profit, cost, revenue, utility) of umbrella
following the path leading to this node
q Decisions and chance events are displayed in a 10

It will not rain, p=0.6
logical temporal sequence from left to right
q Only chance nodes whose results are known can precede a decision
node
q Each chain of decisions and chance events

represents a possible outcome
28.3.2018
4
Solving a decision tree
q A decision tree is solved by starting EV=4.6
It will rain, p=0.4
4
from the leaves (consequence
Take an umbrella
nodes) and going backward toward 5
It will not rain, p=0.6
the root:
– At each chance node: compute the expected
It will rain, p=0.4
value at the node Do not take an 0
– At each decision node: select the arc with umbrella
the highest expected value
EV=6
It will not rain, p=0.6 10
q The optimal strategy consists of the

arcs selected at decision nodes
28.3.2018
5
Example: investing in the stock market
q You are considering between three Decision tree
investment alternatives: high-risk stock, low-
risk stock, and savings account
q Savings account: certain payoff of 500€
q Stocks:
– 200€ brokerage fee
– Payoffs depend on market conditions
Up Same Down
High-risk 1700 300 -800
Low-risk 1200 400 100
Probability 0.5 0.3 0.1
Source: Clemen, R.T. (1996): Making Hard Decisions: An Introduction to
Decision Analysis, 2nd edition, Duxbury Press, Belmont.
28.3.2018
6
Example: investing in the stock market
q The expected monetary values Decision tree
(EMVs) for the different alternatives
are EMV=580€
– HRS: 0.5∙1500+0.3∙100-0.2∙1000=580
– LRS: 0.5∙1000+0.3∙200-0.2∙100=540
EMV=540€
– SA: 500
→ It is optimal* to invest in high-risk

stock EMV=500€
Source: Clemen, R.T. (1996): Making Hard Decisions: An Introduction to

Decision Analysis, 2nd edition, Duxbury Press, Belmont.
28.3.2018
7
* Assuming that you are risk-neutral
Expected value of information
q How much could the expected value be expected to increase, if
– Additional information about the uncertainties was received before the decision, and
– The decision would be made according to this information?
– Note: this analysis is done before any information is obtained
q Perfect information: certain information about how the
uncertainties are resolved – ”if we could choose after we know the
state of the world”
q Expected value of perfect information: EVPI=EVwPI-EVwoPI
– EVwoPI=Expected value without perfect information
– EVwPI=Expected value with perfect information
– EVwPI is computed through a reversed decision tree in which all chance nodes
precede all decision nodes
28.3.2018
8
Expected value of perfect information
Decision tree Reversed decision tree
EMV=580€
EMV=540€
EVwoPI=580€ EVwPI=
0.5∙1500+0.3∙500
+0.2∙500=1000€
EMV=500€
EVPI=1000€-580€=420€
28.3.2018
9
Sample information
q Perfect information can rarely be obtained,
but…
q What if you could obtain some sample
information by, e.g., consulting an
economist about the market conditions?
True state
Up Same Down
Up 0.8 0.15 0.20
Economist’s
Same 0.1 0.70 0.20
forecast
Down 0.1 0.15 0.60
Sum 1 1 1
P(forecast ‘up’ | true state = ‘up’) 28.3.2018

10
Computing probabilities
q Missing probabilities are computed True state
using Up Same Down
Up 0.8 0.15 0.20

– The law of total probability Economist’s
Same 0.1 0.70 0.20
– Bayes’ theorem forecast
Down 0.1 0.15 0.60
q E.g., Sum 1 1 1
– P(Forecast Up) =
P(FU|U)P(U)+P(FU|S)P(S)+P(FU|D)P(D) P(Up)=0.5, P(Same)=0.3, P(Down)=0.2.
=0.8∙0.5+0.15∙0.3+0.2∙0.2=0.485
( | ) ( ) . ∙ .
– P(Up|Forecast Up)= = =
( ) .
.
28.3.2018
11
Expected value of sample information
q EVSI=EVwSI - EVwoSI
– EVwSI=Expected value with sample
information
– EVwoSI=Expected value without
sample information
q EVSI = 822€ - 580€ = 242€

– The benefits of consulting the
economist can be expected to offset the
costs only if her fee is less than 242€
– NOTE! Knowing the fee (c) at the outset would
simply have reduced the monetary outcomes of
all consequence nodes that follow the decision
to consult the economist by c 28.3.2018
12
Decision trees
Example: Decision tree (1/12)
§ Your uncle is going to buy a tractor. He has two alternatives:

1. A new tractor (17 000 €)
2. A used tractor (14 000 €)
§ The engine of the old tractor may be defect, which is hard to
ascertain. Your uncle estimates a 15 % probability for the defect.
§ If the engine is defect, he has to buy a new tractor and gets 2000
€ for the old one.
§ Before buying the tractor, your uncle can take the old tractor to a
garage for an evaluation, which costs 1 500 €.
§ If the engine is OK, the garage can confirm it without
exception.
§ If the engine is defect, there is a 20 % chance that the
garage does not notice it.
Systems Analysis Laboratory
§ Your uncle maximizes expected monetary value
eLearning
Helsinki University of Technology 13/56
Decision trees
§ Before making the buying decision and before you get to know
the result of any uncertain event, you must decide upon taking
the old tractor to a garage for an evaluation.
§ The decision node ‘evaluation’ is placed leftmost in the tree
Evaluation
No evaluation

Helsinki University of Technology eLearning 14/56
§ If the old tractor is evaluated, your uncle receives the results of the evaluation
Result: “OK”
Evaluation
Result: “Defect”
No evaluation

§ The next step is to decide which tractor to buy
Result: “OK” New
Old
Evaluation
New
Old
New
No evaluation
Old

§ …But the engine of the old tractor can be defect

Result: “OK” New
Defect
Old
Evaluation No defect
New
Defect
Old
No defect
New
No evaluation Defect
Old
No defect
§ Now all chance nodes and decisions are in chronological order such
that in each node, we can follow the path to the left to find out what we
know
§ We next need the probabilities for all outcomes of the chance nodes
P(result ”OK”) Result: “OK” New P(”Defect” | result ”OK”)

Defect
Old
New P(”No defect” | result ”OK”)
Result: “Defect” P(”Defect” | result ”Defect”)
P(result ”Defect”) Defect
Old
No defect
P(”No defect” | result ”Defect”)
New
No evaluation Defect P(”Defect”)
Old
No defect
P(”No defect”)
§ Solve all probabilities. You know that
§ ” Your uncle estimates a 15 % probability for the defect.” => P(Defect)=0.15
§ “If the engine is OK, the garage can confirm it without exception.” => P(result
“OK” | No defect)=1
§ “If the engine is defect, there is a 20 % chance that the garage does not
notice it.” => P(result “OK” | Defect)=0.20
P(result "OK") = P(result "OK"| No defect) × P (No defect) + P(result "OK" | Defect) × P(Defect)
= 1.0 × 0.85 + 0.20 × 0.15 = 0.88
P(result "defect")=1- P(result "OK") = 0.12
P(result "OK" | Defect) × P (Defect) 0.20 × 0.15
P(Defect | result "OK") = = » 0.034
P (result "OK") 0.88
P(No defect | result "OK") = 1 - 0.034 = 0.966
P (result "defect" | Defect) × P(Defect) 0.80 × 0.15
P(Defect | result "defect") = = = 1.00
P(result "defect") 0.12
Systems Analysis
P(No | result "defect") = 1 -1 = 0
Laboratory
Defect
§ Compute monetary values for each end node
§ Evaluation + new = 1500 + 17000 = 18500
§ Evaluation + old with defect = 1500 + 14000 – 2000 + 17000 = 30500
§ Evaluation + old without defect = 1500 + 14000 = 15500
§ No evaluation + new = 17000
§ No evaluation + old with defect = 14000 – 2000 + 17000 = 29000
§ No evaluation + old without defect = 14000 New
Result: “OK”
Defect
Old
New
Defect
Old
No defect
New
No evaluation Defect
Helsinki University of Technology eLearning Old 20/56
No defect
Decision trees
§ We now have a decision tree presentation of the problem
New -18 500

Result: “OK”
0.034 Defect
0.88 Old -30 500
Evaluation No defect -15 500
0.966
New -18 500
1 Defect
0.12 Old -30 500
Result: “Defect” No defect -15 500
0
New -17 000
0.15 Defect
Old -29 000
No evaluation No defect -14 000
0.85

Decision trees

§ Starting from the right, compute expected monetary values for each
decision
§ Place the value of the better decision to the decision node
EMV(Old | result “ok”)= 0.034 x -30500 +
0.966 x -15500= -16010
EMV(New | result “ok”)= -18500
Result: “OK” -18 500
0.034 Defect
0.88 Old -30 500
Evaluation No defect -15 500
-16010 0.966
New -18 500
1 Defect
0.12 Old -30 500
Result: “Defect” No defect -15 500
0
New -17 000
0.15 Defect
Old -29 000
No evaluation
Systems Analysis Laboratory No defect -14 000
Helsinki University of Technology eLearning 0.85 22/56
Decision trees

§ Starting from the right, compute expected monetary values for each
decision
§ Place the value of the better decision to the decision node
0.88 x -16010 + 0.12 x -18500 =

-16309
New -18500
Defect
0.88 -16010 Old-16010 -30 500
-16309 No defect -15 500
Evaluation
New -18500 -18 500
-18500
1 Defect
0.12 Old -30 500
-16250 Result: “Defect” -30500 No defect -15 500
0
-16250 New -17000 -17 000
No evaluation
0.15 Defect
Old -29 000
Systems Analysis Laboratory No defect -14 000
Helsinki University of Technology eLearning -16250 0.85 23/56
Decision trees

§ The optimal solution is to by the old tractor without evaluating it
New -18500
Defect
0.88 -16010 Old-16010 -30 500
-16309 No defect -15 500
Evaluation
New -18500
-18500 -18 500
1 Defect
0.12 Old -30 500
-16250 Result: “Defect” -30500 No defect -15 500
0
-16250 New -17000 -17 000
No evaluation
0.15 Defect
Old -29 000
No defect -14 000
-16250 0.85

Probability assessment
q Please use a few minutes to answer ten probability assessment
questions
– Two quesionnaires (A and B), answer to just one of them
q While answering,
– Do not communicate with others
– Do not look up the answers on the internet
q After having answered, hand in your answer sheet to the lecturer
28.3.2018
25
Estimation of probabilities
q How to obtain the probabilities needed in decision models?
1. If possible, use objective data
2. If objective data is not available, obtain subjective probability
estimates from experts through
o Betting approach
o Reference lottery
o Direct judgement
28.3.2018
26
Estimation of probabilities: Betting
approach
q Goal: to estimate the probability of event A
– E.g., A=”GDP growth is above 3% next year” or A=”Sweden
A X
will join NATO within the next five years”
Bet for A
-Y
Not A
q Betting approach: A -X
Bet
– Bet for A: win X € if A happens, lose Y € if not against A
– Expected monetary value − 1− Not A Y
– Bet against A: lose X € if A happens, win Y € if not
– Expected monetary value − + 1−
– Adjust X and Y until the respondent is indifferent between
betting for or against A
– Assuming risk-neutrality(*, the expected monetary values of
betting for or against A must be equal:
− 1− =− + 1− ⇒ =
+
*)A strong assumption 28.3.2018

27
Estimation of probabilities: Reference
lottery A X
q Lottery: Lottery
– Win X if A happens Not A
Y
– Win Y if A does not happen p
Ref. X
– X is preferred to Y lottery
1-p Y
q Reference lottery:
– Win X with (known) probability p
– Win Y with (known) probability (1-p)
– Probability p can be visualized with, e.g., a wheel of fortune
q Adjust p until the respondent is indifferent between the two lotteries:
+ 1− = + 1− ⇒ =
q Here, the respondent’s risk attitude does not affect the results (shown later)
28.3.2018
28
Reference lottery: example
q Event A: ”HIFK wins Jokerit”
A 10 € The respondent A 10 € The respondent

Lottery
Lottery chooses the 0€ chooses the
Not A
Not A 0€
reference lottery: lottery:
⚀⚁⚂⚃⚄ ⚀
10 € 5 Ref. 10 €
1
Ref.
10 < 10 lottery 10 > 10
lottery
⚅
0€
6 ⚁⚂⚃⚄⚅ 0€ 6
A 10 € A 10 € Chooses the
Lottery Chooses Lottery
reference
Not A 0€
the lottery: 0€
⚀⚁⚂
1
Not A
⚀⚁⚂⚃
lottery:
Ref. 10 €
> Ref. 10 € 2
lottery
⚃⚄⚅ 2 lottery
⚄⚅
<
0€ 0€ 3
29
These four answers revealed to probability estimate of A to

be in (0.5, 0.67). Further questions should reveal the 28.3.2018
respondent’s estimate for P(A)

Estimation of continuous probability
distributions
q A continuous distribution can be approximated by estimating several event
probabilities (X is preferred to Y)
q Example:
– Goal: to assess the distribution of the change in GDP (ΔGDP) in Finland next year
– Means: elicitation of probability p for five different reference lotteries
1
ΔGDP£-3% ΔGDP£-1% ΔGDP£0%
X X X
0.9
0.8
Y Y Y 0.7
p=0.05 0.25
X X 0.5 0.6
X
CDF
0.5
Y Y Y 0.4
ΔGDP£1.5% ΔGDP£4%
X X 0.3
0.2
Y Y 0.1
0.75 0.95
X X 0
-3 -2 -1 0 1 2 3 4
D GDP
Y Y
28.3.2018
30
Estimation of continuous probability
distributions
q Often experts assess the descriptive statistics of the distribution
directly, e.g.,
– The feasible range (min, max)
– Median f50 (i.e., P(X<f50)=0.5)
– Other quantiles (e.g., 5%, 25%, 75%, 95%)
q In the previous example:
– ”The 5% and 95% quantiles are f5 =-3% and f95 = 4%”
– ”The change in GDP is just as likely to be positive as it is to be negative”
– ”There is a 25% chance that the change in GDP is below -1%”
– ”There is a 25% chance that the change in GDP is above 1.5%”
28.3.2018
31
Biases in probability assessment
q Subjective judgements by both ordinary people and experts are
prone to numerous biases
– Cognitive bias: Systematic discrepancy between the ‘correct’ answer and
the respondent’s actual answer
o E.g., assessment of conditional probability differs from the correct value given by Bayes’ rule
– Motivational biases: judgements are influenced by the desriability or
undesirability of events
o E.g., overoptimism about success probabilities
o Strategic underestimation of failure probabilities
q Some biases can be easy, some difficult to correct
28.3.2018
32
Representativeness bias (cognitive)
q If x fits the description of A well, then P(x∈A) is assumed to be large
q The ‘base rate’ of A in the population (i.e., the probability of A) is not

taken into account
q Example: You see a very tall man in a bar. Is he more likely to be a

professional basketball player or a teacher?
– There are about 29,000 male teachers in Finland, but about 90 professional basketball players.
– If very tall refers to > 200cm, the share of very tall men should be 40 times as high among pro
basketball players as it is among teachers for it to be more likely that the man you saw is a pro
basketball player.
– Your responses: 80% teacher, 20% basketball player
28.3.2018
33
Representativeness bias
q Another example: Linda is 31 years old, single, outspoken, and very
bright. She majored in philosophy. As a student, she was deeply
concerned with issues of discrimination and social justice, and also
participated in antinuclear demonstrations. Please check the most
likely alternative:
a. Linda is a bank teller.
b. Linda is a bank teller and active in the feminist movement.
q Many people choose b, although b⊂a whereby P(b)<P(a)

– Your responses: 67% a, 33% b.
28.3.2018
34
Conservativism bias (cognitive)
q When information about some uncertain event is obtained, people typically do not
adjust their initial probability estimate about this event as much as they should based
on Bayes’ theorem.
q Example: Consider two bags X and Y. Bag X contains 30 white balls and 10 black balls,
whereas bag Y contains 30 black balls and 10 white balls. Suppose that you select one
of these bags at random, and randomly draw five balls one-by-one by replacing them in
the bag after each draw. Suppose you get four white balls and one black. What is the
probability that you selected bag X with mainly white balls?
q Typically people answer something between 70-80%. Yet, the correct probability is
27/28 ≈ 96%.
q Your responses: mean response 59%. The majority (57%) answered 50%.
28.3.2018
35
Representativeness and
conservativism bias - debiasing
q Demonstrate the logic of joint and conditional probabilities and
Bayes’ rule
q Split the task into an assessment of
– The base rates for the event (i.e., prior probability)
– E.g., what is the relative share of bank tellers in the population? What are the relative shares
of teachers and pro basketball players?
– The likelihood of the data, given the event (i.e., conditional probabilities)
– E.g., what is the relative share of people active in the feminist movement? Is this share
roughly the same among bank tellers as it is among the general population or higher/lower?
– What is the likelihood that a male teacher is taller than 200cm? How about a pro basketball
player?
28.3.2018
36
Availability bias (cognitive)
q People assess the probability of an event by the ease with which instances or
occurences of this event can be brought to mind.
q Example: In a typical sample of English text, is it more likely that a word starts
with the letter K or that K is the third letter?
– Most people think that words beginning with K are more likely, because it is easier to think of
words that begin with "K” than words with "K" as the third letter
– Yet, there are twice as many words with K as the third letter
– Your responses: 13% first letter, 87% third letter.
q Other examples:
– Due to media coverage, the number of violent crimes such as child murders seems to have
increased
– Yet, compared to 2000’s, 18 times as many children were killed per capita in 1950’s and twice as
many in 1990’s
28.3.2018
37
Availability bias - debiasing
q Conduct probability training
q Provide counterexamples
q Provide statistics
q Based on empirical evidence, availability bias is difficult to

correct
28.3.2018
38
Anchoring bias (cognitive)
q When assessing probabilities, respondents sometimes consider
some reference assessment
q Often, the respondent is anchored to the reference assessment
q Example: Is the percentage of African countries in the UN
A. Greater or less than 65? What is the exact percentage?
o Average answer: Less, 45%.
o Your responses: Less, median 22%, mean 34%.
B. Greater or less than 10? What is the exact percentage?
o Average answer: Greater, 25%.
o Your responses: Greater, median 23%, mean 27%.
28.3.2018
39
Anchoring bias - debiasing
q Avoid providing anchors
q Provide multiple and counteranchors
q = if you have to provide an anchor, provide several which differ significantly
from each other
q Use different experts who use different anchors
q Based on empirical evidence, anchoring bias is difficult to

correct
28.3.2018
40
Overconfidence (cognitive)
q People tend to assign overly narrow confidence intervals to their probability estimates
Your responses:
1. Martin Luther King’s age at death 39 years
2. Length of the Nile River 6738 km
3. Number of Countries that are members of OPEC 13
4. Number of Books in the Old Testament 39
5. Diameter of the moon 3476 km
6. Weight of an empty Boeing 747 176900 kg
7. Year of Wolfgang Amadeus Mozart’s birth 1756
8. Gestation period of an Asian elephant 645 days
9. Air distance from London to Tokyo 9590 km
10. Depth of the deepest known point in the oceans 11033 m
q If three or more of your intervals missed the actual value, you have demonstrated
overconfidence
28.3.2018
41
Overconfidence - debiasing
q Provide probability training
q Start with extreme estimates (low and high)
q Use fixed values instead of fixed probability elicitations:
– Do not say: ”Give a value x such that the probability for a change in GDP lower than
x is 0.05”
– Do say: ”What is the probability that the change in GDP is lower than -3%?”
q Based on empirical evidence, overconfidence is difficult to
correct
28.3.2018
42
Desirability / undesirability of events
(motivational)
q People tend to believe that there is a less than 50 % probability that negative
outcomes will occur compared with peers
– I am less likely to develop a drinking problem
– Your responses: 25% more likely, 31% less likely, 44% equally likely
q People tend to believe that there is a greater than 50 % probability that

positive outcomes will occur compared with peers
– I am more likely to become a homeowner and have a starting salary of more than 3,500€
– Your responses on owning a home: 44% more likely, 13% less likely, 44% equally likely
– Your responses on salary: 38% more likely, 19% less likely, 44% equally likely
q People tend to underestimate the probability of negative outcomes and

overestimate the probability of positive outcomes
28.3.2018
43
Desirability / undesirability of events -
debiasing
q Use multiple experts with alternative points of view
q Place hypothetical bets against the desired event
q “Make the respondent’s money involved”
q Use decomposition and realistic assessment of partial probabilities
q “Split the events”
q Yet, empirical evidence suggests that all motivational biases are
difficult to correct
Further reading: Montibeller, G., and D. von Winterfeldt, 2015. Cognitive and
Motivational Biases in Decision and Risk Analysis, Risk Analysis
28.3.2018
44
Summary
q Decision trees are probability-based models to support decision-
making under uncertainty
– Which decision alternative(s) should I choose?
– How much would I be willing to pay for perfect (EVPI) or imperfect sample (EVSI)
information about how the uncertainties are resolved?
q Probabilities for decision models often need to be elicited from
experts
q Probability elicitation is prone to cognitive and motivational biases
– Some cognitive biases can be easy to correct, but…
– Some other cognitive biases and all motivational biases can be difficult to overcome
28.3.2018
45
Decision making and
problem solving –
Lecture 3
• Expected Utility Theory (EUT)
• Assessment of utility functions
• Behavioral aspects

Motivation
q Last time:
– We learned how to support decision-making under uncertainty,
when the DM’s objective is to maximize the expected value
– Maximizing expected value is rational only if the DM is risk
neutral, i.e., indifferent between A 10000
o obtaining x for sure and
o a gamble with uncertain payoff Y such that x=E[Y] 0.6 20000
0.3
– Usually, DMs are risk averse = they prefer obtaining x for B 10000
sure to a gamble with payoff Y such that x=E[Y] 0.1
-5000
q This time: EV=14500

– We learn how to accommodate the DM’s risk attitude
(=preference over alternatives with uncertain outcomes) in
decision models
28.3.2018
2
Expected utility theory (EUT)
q John von Neumann and Oscar Morgenstern (1944) in Theory of
Games and Economic Behavior:
– Axioms of rationality for preferences over alternatives with uncertain outcomes
– If the DM follows these axioms, she should prefer the alternative with the highest
expected utility
q Elements of EUT
– Set of outcomes and lotteries
– Preference relation over the lotteries satisfying four axioms
– Representation of preference relation with expected utility
28.3.2018
3
EUT: Sets of outcomes and lotteries
q Set of possible outcomes T: Lottery
– E.g., revenue euros, demand Decision tree PMF
q Set of all possible lotteries L: 0.6 20000 0.6, = 20000
0.3 0.3, = 10000
– A lottery ∈ associates a probability 10000 =
0.1, = −5000
∈ [0,1] with each possible outcome 0.1
-5000 0, ℎ
∈
o Finite number of outcomes with a positive
probability >0
Degenerate lottery
o Probabilities sum up to one ∑ =1
Decision tree PDF
o Lotteries are thus discrete PMFs / decision trees
with a single chance node 1 1, = 10000
10000 =
0, ℎ
q Deterministic outcomes are modeled as
degenerate lotteries
28.3.2018
4
EUT: Compound lotteries
q Compound lottery:
– Lottery ∈ with probability
– Lottery ∈ with probability 1 −
q Compound lottery can be modeled as lottery ∈ :
= + 1− ∀ ∈ ≃ = + (1 − )
q Example:
– You have a 50-50 chance of getting a ticket to lottery ∈ or to lottery ∈
0.2 2 0.25 −1
= 0.5 0 0.15
0.3 0
0.5 −1
0.6 2 2
0.4
1− = 0.5
0.4 0.2 3
3
28.3.2018
5
Preference relation
q Let ≽ be preference relation among lotteries in L
– Preference ≽ : at least as preferable as
– Strict preference ≻ defined as ¬( ≽ )
– Indifference ~ defined as ≽ ∧ ≽
28.3.2018
6
EUT axioms A1-A4 for preference
relation
q A1: ≽ is complete
– For any , ∈ , either ≽ or ≽ or both
q A2: ≽ is transitive
– If ≽ and ≽ , then ≽
q A3: Archimedean axiom
– If ≻ ≻ , then ∃ , ∈ (0,1) such that
+ (1 − ) ≻ and ≻ + (1 − )
q A4: Independence axiom
– Let ∈ (0,1). Then,
≻ ⇔ + (1 − ) ≻ + (1 − )
28.3.2018
7
If the EUT axioms hold for the DM’s
preferences
q A3: Archimedean axiom
– Let ≻ ≻ . Then exists ∈ (0,1) so that ~ + 1−
q A4: Independence axiom
– ~ ⇔ + (1 − ) ~ + (1 − )
– Any lottery (or outcome = a degenerate lottery) can be replaced by
an equally preferred lottery; According to A3, such lotteries /
0.5 250
outcomes exist = 0.5 100 = 0.5
0.5 250
1
100 ~ ⇒ ~ 0.5 0
0.5 0 0.5 0.5
28.3.2018
8
Main result: Preference representation
with EU
q ≽ satisfies axioms A1-A4 if and only if there exists a real-valued utility
function u(t) over the set of outcomes T such that
≽ ⇔ ≥ ( )
∈ ∈
q Implication: a rational DM following axioms A1-A4 selects the
alternative with the highest expected utility
= ( )
∈
– A similar result can be obtained for continuous distributions:
o ≽ ⇔ ≥ , where =∫
28.3.2018
9
Computing expected utility
q Example: Joe’s utility function for the number of
= 2 =5
apples is u(1)=2, u(2)=5, u(3)=7. Would he prefer
– Two apples for certain (X), or = 0.5 1 + 0.5 3
– A 50-50 gamble between 1 and 3 apples (Y)? = 0.5 2 + 0.5 7 = 4.5
q Example: Jane’s utility function for money is =

. Which alternative would she prefer? = 0.5 3 + 0.5 5
= 0.5 9 + 0.5 25 = 17
– X: 50-50 gamble between 3 and 5M€
– Y: A random amount of money from Uni(3,5) distribution 1
= =
– What if her utility function was = ? 2
1 1
= 5 − 3 = 16.33333
6 6
28.3.2018
10
Uniqueness up to positive affine
transformations 5
u2(t)=1+0.5u1(t)
q Two utility functions ( ) and = +

4
u1(t)=t3-2.5t2
3
, ( > 0) establish the same preference order 2
among any lotteries: 0
= + = + .
-1
-2
q Implications:
-3
0 0.5 1 1.5 2 2.5 3
– Any linear utility function = + , ( > 0) is a positive 2.5

u1(t)=t
u2(t)=1.5+1/3u1(t)
affine transformation of the identity function = ⇒ 2
establishes the same preference order as expected value

1.5
– Utilities for two outcomes can be freely chosen: 1
o E.g., scale such that and ∗ = 1 and = 0:

− 1 0.5
= ∗ = ∗ − ∗
− − − 0
0 0.5 1 1.5 2 2.5 3
=α>0 =β
28.3.2018
11
Reference lottery revisited
q Assume that an expected utility maximizer with utility
function u uses a reference lottery to assess the
A t+
probability of event A Lottery X
q She thus adjusts p such that she is indifferent Not A
t-
between lottery X and reference lottery Y: p t+
= Ref.
Lottery Y
1-p t-
⇔ + 1− = + 1−
⇔ − = −
⇔ =
q Utility function u does not affect the result
28.3.2018
12
Expected utility in decision trees
q Do everything in the usual way,
Profit Utility
but EU=1.07 1.78
– Chance node: compute the 1.10
expected utility -0.71
– Decision node: select the 1.63
alternative corresponding to 1.18
maximum expected utility EU=1.35 0.89
– Cf. the umbrella example, in which
‘some numbers’ represented 1.39
EU=1.39
preferences
=2−
28.3.2018
13
Expected utility in Monte Carlo
q For each sample , … , of
random variable X,
compute utility ( )
q Mean of sample
utilities ( ), … , ( )
provides an estimate for
[ ]
28.3.2018
14
Assessment of utility functions
q Utility functions are assessed by asking the DM to choose between a simple
lottery and a certain outcome (i.e., a degenerate lottery) X t
– X: Certain payoff t
– Y: Payoff with probability p (1-p) p t+
Y
q General idea: 1-p t-
– Vary the parameters (p,t, , ) until the DM is indifferent between X and Y:
= ⇔ = + (1 − )
– Repeat until sufficiently many points for the utility function have been obtained
q Because u is unique up to positive affine transformations, u can be fixed at
two points
q Usually, u is set at 1 at the most preferred level, and at 0 at the least preferred
28.3.2018
15
Assessment: The certainty equivalence
approach
q The DM assesses t
q Example: Assess utility function for the interval [-10,50] euros
– Normalization: we can fix u(-10)=0 and u(50)=1 1
30€
0.8
20€ 40€
0.6
0.5 0.5 0.5
u(t)
50€ 30€ 50€ 0.4
-10€ 0.5 -10€ 0.5 30€ 0.2

0.5
0
-10 0 10 20 30 40 50
30 20 40 t
= 0.5 −10 + 0.5 50 = 0.5 −10 + 0.5 30 = 0.5 30 + 0.5 50

= 0.5 0 + 0.5 1 = 0.5 = 0.5 0 + 0.5 0.5 = 0.25 = 0.5 0.5 + 0.5 1
= 0.75
28.3.2018
16
Other approaches to utility assessment
q Probability equivalence: X t X t
– The DM assesses p p ? ? t+
Y Y
q Gain equivalence: 1-p t- 1-? t-
– The DM assesses t+
q Loss equivalence: X t
– The DM assesses t- p t+
Y
1-p ?
q Often in applications, the analyst chooses a family of utility functions

and then asks the DM to compare lotteries to fix the parameter(s)
– E.g., the exponential utility function (parameter )
=1− , >0
28.3.2018
17
EUT for normative decision support
q EUT is a normative theory: if the DM is rational, she should select
the alternative with the highest expected utility
– Not descriptive or predictive: EUT does not describe or predict how people
actually do select among alternatives with uncertain outcomes
q The four axioms characterize properties that are required for

rational decision support
– Cf. probability axioms describe a rational model for uncertainty
– The axioms are not assumptions about the DM’s preferences
28.3.2018
18
Question 1 http://presemo.aalto.fi/2134lec2
q Which of the below alternatives would you choose?
1. A sure gain of 1 M€
2. A gamble in which there is a
o 1% probability of getting nothing,
o 89% probability of getting 1M€, and
o 10% probability of getting 5M€
28.3.2018
19
q Imagine that a rare disease is breaking out in a community and is

expected to kill 600 people. Two different programs are available
to deal with the threat.
– If Program A is adopted, 200 people will be saved
– If Program B is adopted, there’s a 33% probability that all 600 will be
saved and a 67% probability that no one will be saved.
Which program will you choose?
1. Program A
2. Program B
28.3.2018
20

o 89% probability of getting nothing and

o 90% probability of getting nothing, and
28.3.2018
21
q Imagine that a rare disease is breaking out in some community

and is expected to kill 600 people. Two different programs are
available to deal with the threat.
– If Program C is adopted, 400 of the 600 people will die,
– If Program D is adopted, there is a 33% probability that nobody will die
and a 67% probability that 600 people will die.
Which program will you choose?
1. Program C
2. Program D
28.3.2018
22
Allais paradox
q Which of the below alternatives would you choose? Most people choose A; hence
A. A sure gain of 1 M€ E[u(A)]>E[u(B)]:
B. A gamble in which there is a u(1) > 0.10u(5)+0.89u(1)+0.01u(0) ⇒
o 1% probability of getting nothing,
o 89% probability of getting 1M€, and 0.11u(1) > 0.10u(5)+0.01u(0)

C. A gamble in which there is a Most people choose D; hence
o 89% probability of getting nothing and E[u(D)]>E[u(C)]:
11% probability of getting 1M€
o
0.10u(5)+0.90u(0) > 0.11u(1)+0.89u(0) ⇒
D. A gamble in which there is a
o 90% probability of getting nothing, and
0.11u(1) < 0.10u(5)+0.01u(0)
q Actual choice behavior is not always consistent with EUT
28.3.2018
23
Framing effect
q Most people choose A and D
q People tend to be ”risk-averse” about gains and ”risk-seeking”
about losses
A 200 (=600-400) C -400

33% 33%
B
600 (=600-0) 0
D
67% 0 (=600-600) 67% -600
28.3.2018
24
Summary
q The DM’s preferences over alternatives with uncertain outcomes
can be described by a utility function
q A rational DM (according to the four axioms of rationality) should

choose the alternative with the highest expected utility
q Yet, observed choice behavior may not be consistent with EUT
28.3.2018
25
Decision making and
problem solving –
Lecture 4
• Modeling risk preferences
• Stochastic dominance
• Risk measures

Motivation
q Last time:
– We learned how to accommodate the DM’s preferences over alternatives with
uncertain outcomes in decision models
q This time:
– We learn how these preferences correspond to the DM’s attitude towards risk
28.3.2018
2
Risk and risk preferences
q Risk: possibility of loss (or some other unpreferred outcome)
– Characterized by both the probability and magnitude of loss
q Risk preferences:
– How does the riskiness of a decision alternative affect its desirability?
– E.g., risk neutrality: choose the alternative with the highest expected (monetary) value, riskiness
is not a factor
q Definition of risk preferences requires that outcomes T are quantitative and
preferences among them monotonic
– E.g., profits, costs, lives saved etc.
q Here, we assume that more is preferred to less, i.e., u(t) is increasing (and
differentiable) for all t
28.3.2018
3
Certainty equivalent in EUT
q Definition: Certainty equivalent of a random variable X, denoted by
CE[X], is an outcome in T such that CE[X]
= ⇔ X
= ( ) Allowed
because u is
monotonic
– CE[X] is the certain outcome such that the DM is indifferent between alternatives X
and CE[X]
– CE[X] depends on both the DM’s utility function u (preferences) and the distribution
of X (uncertainty)
o My CE for roulette may be different from yours
o My CE for roulette may be different from my CE for one-armed bandit
28.3.2018
4
Certainty equivalent - Example
q Consider a decision alternative X with 3 = 0.5 and 5 = 0.5 and
three DMs with the below utility functions
q Compute each DM’s certainty equivalent for X
( ) ( ) ( )
(5) (5) (5)

[ ]
(3) [ ]
[ ]
(3)
(3)
[ ]
3 5
[ ]
[ ]
3 5 3 5
q The shape of the utility function seems to determine whether CE[X] is

below, above, or equal to E[X]=4
28.3.2018
5
Convex and concave functions
q Definition: u is concave, if for any , :
+ (1 − ) ≤ + (1 − ) ∀ ∈ [0,1]
– A line drawn between any two points and is below (or
equal to)
– ′′ ≤ 0 ∀ ∈ , if the second derivative exists
q Definition: u is convex, if for any , :

+ (1 − ) ≥ + (1 − ) ∀ ∈ [0,1]
– A line drawn between any two points and is above (or
equal to)
– ′′ ≥ 0 ∀ ∈ , if the second derivative exists
28.3.2018
6
Jensen’s inequality
q For any random variable X, if function u is
I. Convex, then [ ]≥ ( )
II. Concave, then ≤
⇒
concave convex
⇒ ≤ ⇒ ≥
⇔ ( )≤ ( ) ⇔ ( )≥ ( )
Allowed ⇔ [ ]≤ ⇔ [ ]≥
because u is
increasing
28.3.2018
7
Risk attitudes in EUT
I. u is concave iff CE[X] ≤ E[X] for all X CE[X] ?
II. u is convex iff CE[X] ≥ E[X] for all X
III. u is linear iff CE[X]=E[X] for all X X
q A DM with a linear utility function is called risk neutral

– Indifferent between uncertain outcome X and a certain outcome equal to E[X]
q A DM with a concave but not linear utility function is called risk averse
– Prefers a certain outcome smaller than E[X] to uncertain outcome X
q A DM with a convex but not linear utility function is called risk seeking
– Requires a certain outcome larger than E[X] to not choose uncertain outcome X
28.3.2018
8
Risk premium in EUT
q Definition: Risk premium for random variable X is RP[X]=E[X]-CE[X]
– RP[X] depends on both the DM’s preferences (u) and the uncertainty in the decision
alternative (distribution of X)
– RP[X] is the premium that the DM requires on the expected value to change a
certain outcome of CE[X] to an uncertain outcome X
( )
I. DM is risk neutral, iff RP[X]=0 for all X (5)

[ ]
II. DM is risk averse, iff RP[X] ≥ 0 for all X (3)
III. DM is risk seeking, iff RP[X] ≤ 0 for all X
3 5
[ ]
28.3.2018
9
Computing CE and RP
Example: Jane’s = and her
1. Compute E[u(X)] and E(X)
payoff is Y~Uni(3,5)
2. Solve
1. =∫ ( ) = 16.33
3. Compute = [ ( )]
2. = = ⇔ = =
4. Compute RP[X]=E[X]-CE[X]
3. = 16.33 = 16.33 = 4.04
4. RP[X] = 4 - 4.04 = -0.04
q Step 2: if cannot be solved
analytically, solve it numerically from
=
– Trial and error
– Computer software
28.3.2018
10
Prospect theory
utility
q EUT assumes that people only care about the outcome in the
absolute sense
q Yet, empirical evidence suggests that people tend to
– think of possible outcomes relative to a certain reference point (often the
status quo),
– have different risk attitudes towards gains and losses with regard to the
reference point,
– overweight extreme, but unlikely events, but underweight "average" events.
q Prospect theory seeks to accommodate these empirical findings:
Tversky, A. and D. Kahneman. ”Advances in prospect theory: Cumulative Gains Losses
representation of uncertainty.” Journal of Risk and uncertainty 5.4 (1992): 297-
High Risk Risk
323. probability averse seeking
q NOTE:
Low Risk Risk
– EUT is a normative theory: tells what rational people should do probability seeking averse
– Prospect theory is a descriptive theory: tries to describe what people tend to
do in real life
28.3.2018
11
Stochastic dominance
q Question: Which decision alternative would you choose?
1
1. X F (t)
Y
F (t)
2. Y 0.8 X
0.6
0.4
0.2
0
-10 -5 0 5 10 15 20
t euros
28.3.2018
12
First-degree Stochastic Dominance 1
F Y(t)
0.8 F X(t)
Definition: X dominates Y in the sense of First-
degree Stochastic Dominance (denoted X ≽FSDY), if 0.6
X ≽FSD Y
0.4
≤ ∀ ∈
0.2
0
with strict inequality for some t. -10 -5 0 5
t euros
10 15 20
Theorem: X ≽FSDY if and only if 0.8

FZ(t)
FX(t)
Neither
≥ ∀ ∈ , alternative
0.6 dominates
where is the set of all strictly increasing functions the other in
0.4 the sense of
FSD
Implication: If an alternative is strictly dominated in the sense 0.2
of FSD, then any DM who prefers more to less should not

0
choose it. -10 -5 0 5 10 15 20
28.3.2018
13
FSD: Mining example
q A mining company has an opportunity
to bid on two separate parcels of land
q Decisions to be made:
q Overall commitment of some $500
million
– How much to bid?
– Bid alone or with partner?
– How to develop the site if the bid turns out
successful?
q Large decision tree model built to
obtain cumulative distribution functions
of different strategies (= decision
alternatives)
Source: Hax and Wing (1977): ”The use of decision analysis in a capital investment
probelm” In Bell, Keeney, and Raiffa (eds.): Conflicting Objectives in Decisions, Wiley. 28.3.2018
14
FSD: Example (cont’d)
q Assume that the
company prefers a
larger NPV to a
smaller one
q Which strategies
would you
recommend?
Source: Hax and Wing (1977): ”The use of decision analysis in a capital investment
probelm” In Bell, Keeney, and Raiffa (eds.): Conflicting Objectives in Decisions, Wiley. 28.3.2018
15
Second-degree Stochastic Dominance
q Theorem:
≥ ∀ ∈ ⇔ − ≤0 ∀ ∈ ,
where = ∈ | is concave .
q Definition: X dominates Y in the sense of Second-degree Stochastic Dominance

(denoted X ≽SSD Y), if
− ≤0 ∀ ∈ .
q Implication: If an alternative is strictly dominated in the sense of SSD, then any risk-
averse or risk neutral DM who prefers more to less should not choose it.
28.3.2018
16
SSD: graphical interpretation
0.08 1
f X(t) F Y(t)
0.07
f Y(t) 0.8 F X(t) B
0.06
− ≤0 ∀ ∈ 0.05 0.6
0.04
0.03 0.4
q Integral 0.02
0.2
= the area between and 0.01
A
up to point z 0
-10 0 10 20 30
0
-10 0 10 20 30
= the area between the -

0.2 0.5
and the horizontal axis up to point z F X(t)-FY(t) Integral of FX(t)-F Y(t) up to z
0
0.1
q If it is non-positive for all z, then B -0.5

0
X ≽SSD Y -1
-0.1 A
-1.5
-0.2
q Here: X ≽SSD Y, because area A is -2
bigger than area B, and A is left of B -0.3

-10 0 10 20 30
-2.5
-10 0 10 20 30
z
28.3.2018
17
SSD: Mining example revisited
q Assume that the
mining company is
either risk-averse or
risk-neutral
q Which strategies
would you
recommend?
28.3.2018
18
Properties of FSD and SSD
q Both FSD and SSD are transitive:
– If X ≽FSD Y and Y ≽FSD Z, then X ≽FSD Z
o Why? Take any t. Then, ≤ ≤ .
– If X ≽SSD Y and Y ≽SSD Z, then X ≽SSD Z
o Why? Take any ∈ . Then, − [ ]≥ − [ ] ≥ 0.
q FSD implies SSD:
– If X ≽FSD Y, then X ≽SSD Y.
o Why? Take any ∈ . Then, ∈ , and since X ≽FSD Y, we have ≥
[ ].
o Or consider the definitions of FSD and SSD: If ≤ ∀ ∈ , then
− ≤ 0 ≤0 ∀ ∈
28.3.2018
19
Risk measures
q Risk measure is a function that maps each decision alternative to a
single number describing its risk
– E.g., variance = [( − [ ]) ]
– The higher the variance, the higher the risk
q Risk measures are not based on EUT, but can be used together with
expected values to produce decision recommendations
– Risk constraint: Among alternatives whose risk is below some threshold, select the
one with the highest expected value
– Risk minimization: Among alternatives whose expected value is above some
threshold, select the one with minimum risk
– Efficient frontier: Select one of those alternative compared to which no other
alternative yields higher expected value and smaller risk
28.3.2018
20
Risk measures: Value-at-Risk (VaR)
1
F Y(t)
0.8 F X(t)
q Value-at-Risk (VaR [ ]) is the outcome

0.6
such that the probability of a worse or
equal outcome is : 0.4
VaR % = 0.6€
VaR = −1.4€
VaR [ ] 0.2
%
= VaR [ ] = . 0
-10 -5 0 5 10 15 20
t euros
q Higher VaR means smaller risk 0.08
0.07
– Unless applied to a loss distribution 0.06
q Common values for : 1%, 5%, and 10% 0.05
0.04 10%
q Problem: the length/shape of the tail is not 0.03 90%
taken into account 0.02
0.01
0
-10 -5 0 5 10 15 20
28.3.2018
21
Mining example revisited
q Assess VaR % for
strategies 1 and 25
28.3.2018
22
Risk measures: Conditional Value-at-
Risk (CVaR) 0.14
VaR = −1.85
%
0.12
VaR % = −0.97
q Conditional Value-at-Risk (CVaR [ ]) is the ( )
0.1
expected outcome given that the outcome is at most CVaR % = −3.26
CVaR = −4.23
VaR : 0.08 %
CVaR = [ | ≤ VaR ] 0.06

( )
0.04
q Higher CVaR means smaller risk (unless applied to
0.02
losses)
0
-10 0 10 20
q Computation of CVaR to discrete and continuous X:

( ) VaR ( )
≤ VaR = ∑ VaR , ≤ VaR =∫ .
– Note: = ≤ VaR ; PMF/PDF ( ) is scaled such that it sums/integrates up to 1.
28.3.2018
23
Computation of VaR and CVaR
q If the inverse CDF of X is well-defined, VaR can be obtained from
VaR = ( )
– In Excel: norm.inv, lognorm.inv, beta.inv, binom.inv etc
– In Matlab: norminv, logninv, betainv, binoinv etc
q CVaR can then be computed using the formulas on the previous slide
– Sometimes an analytic solution can be obtained; if, e.g., ~ , and VaR = , then
CVaR = − ,
where and Φ are the standard normal PDF and CDF, respectively.
– Sometimes numerical integration is needed
28.3.2018
24
Computation of VaR and CVaR
q With discrete random variables VaR and CVaR are not always well
defined for small values of
– Example:
t -10 -5 1 10 20
fX(t) 0.06 0.02 0.02 0.5 0.4
– VaR % =1
0.06(−10)+0.02(−5)+0.02(1)
– CVaR % = =-6.8
0.06+0.02+0.02
– But what are VaR % , CVaR % ?
28.3.2018
25
VaR and CVaR with Monte Carlo - Excel
=AVERAGE(D12:D211)
=PERCENTILE.INC(C12:C211;0.1)
=IF(C12<=$F$10;C12;”above”)
Note! 200 samples is very

little, because only 1/10=20
are used to estimate CVaR
28.3.2018
26
VaR and CVaR with Monte Carlo -
Matlab
28.3.2018
27
Risk measures and stochastic
dominance
q Theorem: X ≽FSD Y if and only if 1
VaR ≥ VaR ∀ ∈ 0,1 F Y(t)
0.8 F X(t)
q Theorem: X ≽SSD Y if and only if 0.6
CVaR ≥ CVaR ∀ ∈ 0,1 0.4
0.2
0
-10 -5 0 5 10 15 20
t euros
VaR
VaR
28.3.2018
28
EUT vs. Risk measures
q EUT provides a more comprehensive way to capture the DM’s
preferences over uncertain outcomes
q With risk measures, one must answer questions such as
– Which measure to use?
– Which to use in VaR and CVaR?
– How to combine EV and the value of a risk measure into an overall performance
measure?
q Yet, if answers to such questions are exogenously imposed, the use

of risk measures can be easy
– E.g., laws, regulations, industry standard etc.
28.3.2018
29
Summary
q The shape of the utility function determines the DM’s risk attitude
– Linear utility function = risk neutral
– Concave utility function = risk averse
– Convex utility function = risk seeking
q Even if the utility function is not completely specified, decision

recommendations may be implied by stochastic dominance
– If the DM prefers more to less, she should not choose an FSD dominated alternative
– If the DM is also risk averse, she should not choose an SSD dominated alternative
q Risk considerations can also be accommodated by using risk measures such

as VaR or CVaR in conjuction with expected value
28.3.2018
30
Decision making and
problem solving –
Lecture 5
• Multiattribute value theory
• Axioms for preference relations
• Elicitation of attribute-specific value functions
• Preferential and difference independence
• Aggregation of values with an additive value function
Motivation
q So far:
– We have considered decision-making situations in which the DM has one
objective (e.g., maximize the expected value/utility of a monetary payoff)
q This time:
– We consider decision-making situations in which the DM has multiple
objectives or, more precisely…
– Multiple attributes with regard to which the achievement of some
fundamental objective is measured
28.3.2018
2
Multiattribute value theory
q Ralph Keeney and Howard Raiffa (1976): Decisions with Multiple Objectives:
Preferences and Value Tradeoffs
q James Dyer and Rakesh Sarin (1979): Measurable multiattribute value functions,
Operations Research Vol. 27, pp. 810-822
q Elements of MAVT
– A value tree consisting of objectives, attributes, and alternatives
– Preference relation over the alternatives’ attribute-specific performances and differences thereof &
their representation with an attribute-specific value function
– Preference relation over the alternatives’ overall performances and differences thereof & their
representation with a multiattribute value function
28.3.2018
3
Value tree: objectives, attributes, and
alternatives
q A value tree consists of Ideal car
– A fundamental objective
– Possible lower-level objectives Economy Driving
– Attributes that measure the
achievement of the objectives
– Alternatives whose attribute-
Price Expenses Acceleration Top speed
specific performances are being
measured
Audi A4 VW Passat Citroën
28.3.2018
4
Value tree: objectives, attributes and
alternatives
q The attributes a1,…, an have
Job
measurement scales Xi, i=1,…,n; e.g.,
– X1=[1000€/month, 6000€/month]
– X2 =[2 weeks/year, 8 weeks/year]
– X3 =[0 days/year, 200 days/year] Business Fit with
Salary Vacation
– X4 ={poor, fair, good, excellent} travel interests
q Alternatives = ( , , … ) are
characterized by their performance Banker Researcher
Engineer in
Consultant
industry
w.r.t. the attributes; e.g.,
– Banker=(6000€/month, 5 weeks/year, 40
days/year, fair) ∈ × × × .
28.3.2018
5
Preference relation: attribute-specific
performance
q Let ≽ be preference relation among performance levels a and b on a given
attribute
– Preference ≽ : at least as preferable as
– Strict preference ≻ defined as ¬( ≽ )
– Indifference ~ defined as ≽ ∧ ≽
q Let ≽ be preference relation among differences in performance levels on a

given attribute
– Preference ( ← ) ≽ ( ← ): a change from to is at least as preferable as a change from
to
– Strict preference ( ← ) ≻ ( ← ) defined as ¬(( ← ) ≽ ( ← ))
– Indifference ← ~ ( ← ) defined as ( ← ) ≽ ( ← ) ∧ ( ← ) ≽ ( ← )
28.3.2018
6
Axioms for preference relation
q A1: ≽ is complete
– For any , ∈ , either ≽ or ≽ or both
q A2: ≽ is transitive
– If ≽ and ≽ , then ≽
28.3.2018
7
Ordinal value function
Theorem: Let axioms A1-A2 hold. Then, there exists an ordinal
value function : → ℝ that represents preference relation ≽ in
the sense that
≥ ⟺ ≽
q An ordinal value function does not describe strength of preference,

i.e., it does not communicate much more an object is preferred to
another
28.3.2018
8
qAssume you have two mopeds A and B with top speeds of 30 and
35km/h, respectively
qYou have two alternatives for upgrade
q Increase top speed of moped A to 40
q Increase top speed of moped B to 45
qYour prefer a higher top speed to a lower one
q 45>40>35>30
q v(45)=1, v(40)=0.8, v(35)=0.5, v(30)=0.4
q w(45)=0.9, w(40)=0.8, w(35)=0.6, w(30)=0.4
qBoth v and w are ordinal value functions representing your
preferences but they do not describe your preferences between the two
upgrade alternatives
q v(45)-v(35)=0.5 > v(40)-v(30)=0.4, but w(45)-w(35)=0.3 < w(40)-w(30) =0.4
28.3.2018
9
Theorem: Ordinal value functions and represent the same

preference relation ≽ if and only if there exists a strictly increasing
function : ℝ → ℝ such that = ∀ ∈ .
28.3.2018
10
Axioms for preference relation (cont’d)
q A3: ∀ , , ∈ : ≽ ⇔( ← )≽ ( ← )
– If a is preferred to b, then a change from b to a is preferred to no change
q A4: ∀ , , , ∈ : ( ← )≽ ( ← )⇔( ← )≽ ( ← )
– E.g., if an increase in salary from 1500€ to 2000€ is preferred to an increase from 2000€ to 2500€, then a
decrease from 2500€ to 2000€ is preferred to a decrease from 2000€ to 1500€
q A5: ∀ , , , , , ∈ : ( ← )≽ ( ← )∧( ← )≽ ( ← )⇒( ← )≽ ( ← )
– If two incremental changes are both preferred to some other two, then the overall change resulting from the
first two increments is also preferred.
q A6: ∀ , , ∈ ∃ ∈ such that ← ~ ← and ∀ , ∈ ∃ ∈ such that ( ←
)~ ←
– Equally preferred differences between attribute levels can always be constructed
– There is always an attribute level a between b and c such that a change from c to a is equally preferred to a
change from a to b.
q A7: The set (or sequence) | ≻ ℎ ← ~ ( ← ) is finite for any b in Xi
– The sequence of equally preferred differences over a fixed interval is finite
– “No b can be infinitely better than other performance levels”
As French (1988) incorrectly puts it; the idea here is that it is possible to
28.3.2018
construct equally preferred changes in order to represent preferences 11
Cardinal value function
q Theorem: Let axioms A1-A7 hold. Then, there exists a cardinal
value function : → ℝ that represents preference relations ≽
and ≽ in the sense that
≥ ⟺ ≽
− ≥ − ⟺ ← ≽ ← .
Note: A cardinal value function is unique up to positive affine

transformations, i.e., ( ) and = ( ) + , > 0 and
represent the same preferences
28.3.2018
12
Attribute-specific value functions
q A value function maps the
attribute-specific measurement
scale onto a numerical scale in
accordance with the DM’s
preferences
q Value and utility:

– Value is a measure of preference
under certainty
– Utility is a measure of preference
under uncertainty
28.3.2018
13
Elicitation of value functions
q Phases:
∗ ∗
– Define the measurement scale =[ , ] (or , )
– Ask a series of eliciation questions
– Check that the value function gives realistic results
28.3.2018
14
Elicitation of value functions:
Indifference methods
q Bisection method:
– Ask the DM to assess level . ∈ [ , ∗ ] such that she is indifferent
between change . ← and change ∗ ← . .
– Then, ask her to assess levels . and . such that she is indifferent
between
o changes . ← and . ← . , and
∗
o changes . ← . and ← . .
– Continue until sufficiently many points have been obtained
o Use, e.g, linear interpolation between elicited points if needed
– The value function can be obtained by fixing ( ) and ( ∗ ) at, e.g., 0
and 1
28.3.2018
15
Indifference methods 1
0.8
q Example of the bisection method 0.6
v (x)
3
0.4
– Attribute : Traveling days per year
Measurement scale ∗ , , where ∗ = 0 and
0.2
–
= 200; fix =0 and ( ∗ ) =1
0
0 50 100 150 200
x
o ”What would be the number . of traveling days such that
130 − 200 = 0 − 130 ⇒
you would be indifferent between a decrease from 200 to .
0 + 200
days a year and a decrease from . to zero days a year?” 130 = = 0.5
(Answer e.g., ”130”) 2
170 − 200 = 130 − 170 ⇒
you would be indifferent between a decrease from 200 to . 130 + 200
days a year and a decrease from . to 130 days a year?” 170 = = 0.25
2
(Answer e.g., ”170”)
80 − 130 = 0 − 80 ⇒
you would be indifferent between a decrease from 130 to . 0 + 130
days a year and a decrease from . to zero days a year?” 80 = = 0.75
2
(Answer e.g., ”80”)
28.3.2018
16
Indifference methods 1
q Sequence of equally preferred 0.8

differences:
– Set ∈ ( , ∗ ) 0.6
vi(x)
– Ask the DM to assess level ∈ ( , ∗ ] such that he is
indifferent between changes ← and ← 0.4
o − = − ⇒ =2
– Then, ask him to assess level ∈ ( , ∗ ] such that he is 0.2
indifferent between change ← and ←

o − = − ⇒ =3 0
1000 2000 3000 4000 5000 6000
∗
– Continue until = and solve the system of linear equations Salary x€/month
o = = ⇒ = etc. Example:
– If > ∗
(see the exercises!) [ , ∗] = 1000, 6000 , = 1500
o Change ∗
to and interpolate, or = 2500, = 4000, = 6000 = ∗ ⇒
∗
Interpolate to get −
o
1500 = , 2500 = , 4000 = .
28.3.2018
17
Indifference methods
q Indifference methods are likely to result in a cardinal value function
that captures the DM’s preferences
q Therefore, they should be used whenever possible
q Yet: indifference methods cannot be used when the measurement

scale is discrete
– E.g., Fit with interest: X4 ={poor, fair, good, excellent}
– Cf. Axiom A6
28.3.2018
18
Elicitation of value functions: direct
methods
q Direct rating
– Ask the DM to directly attach a value to each attribute level
– E.g. ”Assume that the value of poor fit with interests is 0 and the value of excellent fit with
interests is 1. What is the value of fair fit with interests? How about good fit?”
q Class rating
– Divide the measurement scale into classes and ask the DM to attach a value to these classes
q Ratio evaluation
– Take one attribute level as a reference point and ask the DM to compare the other levels to this
– E.g., ”How many times more valuable is 1000€ than 900€?”
q Direct methods should be avoided whenever possible

– Usually do not result in a cardinal value function
28.3.2018
19
Aggregation of values
q Problem: How to measure the overall value of alternative =
, ,… ?
, ,… =?
q Question: Could the overall value be obtained by aggregating
attribute-specific values?
, ,… = ,…, ?
q Answer: Yes, if the attributes are

– Mutually preferentially independent and
– Difference independent
28.3.2018
20
Preferential independence
q Definition: Attribute X is preferentially independent of the other
attributes Y, if for all x,x’ ∈ X
( , ) ≽ ( , ′) ⇒ , ≽ , for all y ∈ Y
q Interpretation: Preference over the level of attribute X does not

depend on the levels of the other attributes, as long as they stay
the same
q “All other things Y being equal (no matter what they are), an
alternative with performance level x w.r.t. X is preferred to an
alternative with level x’ ∈ X”
28.3.2018
21
Preferential independence: example
q Consider choosing a meal using two attributes:
1. Food ∈ {beef, fish}
2. Wine ∈ {red, white}
q Preferences:
1. Beef is preferred to fish (no matter what the wine is):
o (beef, red) ≽ (fish, red)
o (beef, white) ≽ (fish, white)
2. White wine is preferred with fish and red wine with beef
o (fish, white) ≽ (fish, red)
o (beef, red) ≽ (beef, white)
q Food is preferentially independent of wine

q Beef is preferred to fish, no matter what the wine is: ( , ) ≽ ( , ′) ⇒ , ≽ , for all y ∈ Y
q Wine is not preferentially independent of food
q Attribute-specific valuation of wine is not meaningful from the meal’s perspective
28.3.2018
22
Mutual preferential independence
q Definition: Attributes A are mutually perferentially independent, if
any subset of attributes X⊂A is preferentially independent of the
other attributes Y=A\X. I.e., for any X⊂A, Y=A\X:
( , ) ≽ ( , ′) ⇒ , ≽ , for all y ∈ Y
q Interpretation: Preference over the levels of attributes X does not

depend on the levels of the other attributes, as long as they stay
the same
28.3.2018
23
Mutual preferential independence:
example
q Consider choosing a meal using three attributes:
1. Food ∈ {beef, fish}
2. Side dish ∈ {potato, rice}
3. Wine ∈ {red, white}
q Preferences:
1. All other things being equal, red ≽ white, beef ≽ fish, potato ≽ rice
2. Full meals:
o (beef, rice, red)≽(beef, potato, white)
o (fish, potato, white) ≽ (fish, rice, red)
Each attribute is preferentially independent of the other two, but the

attributes are not mutually preferentially independent:
( , , )≽( , , )⇏ , , ≽ , ,
28.3.2018
24
Mutual pref. independence: example 2
q Choosing a car w.r.t. attributes A={top speed, price, CO2
emissions}
q Attributes defined on continuous scales
q Are all A’s subsets (X) preferentially independent of the other
attributes (Y)?
q Each single attribute is preferentially independent of the other
attributes, because
q Lower price is preferred to higher price independent of other attributes (if other
attributes are equal)
q Higher top speed is preferred to lower
q Smaller emissions are preferred to bigger ones
28.3.2018
25
Mutual pref. independence: example 2
q Is X={price, CO2 emissions} pref. independent of Y={top speed}?
q Consider two cars which differ in price (e.g., 30000 e, 25000 e) and emissions
(150 g/km, 200 g/km) so that one of the alternatives is better in emissions and the
other in price. Set the same top speed for the alternatives (e.g. 230 km/h). Which
one is better?
q DM says (230 km/h, 30000 e, 150 g/km) ≻ (230 km/h, 25000 e, 200 g/km)
q Change the top speed. Is the first car still preferred to the second? e.g. does (150
km/h, 30000 e, 150 g/km) ≻ (150 km/h, 25000 e, 200 g/km) hold?
q “No matter what the top speed is, (30000 e, 150 g/km) ≻ (25000 e, 200 g/km)”
q Consider other prices and emission levels; does your preference hold for all top
speeds?
q If varying the top speed does not influence preference between alternatives, then
{price, CO2 emissions} is preference independent of {top speed}
28.3.2018
26
Difference independence
q Definition: Attribute X is difference independent of the other
attributes Y if for all x,x’ ∈ X
( , ) ← ( , ′)~ ( , ) ← ( , ) for all ∈
q Interpretation: The preference over a change in attribute X does

not depend on the levels of the other attributes Y, as long as they
stay the same
28.3.2018
27
Difference independence: example
q Is {top speed} difference independent of the other attributes {price,
CO2 emissions}?
q Construct y and y’ from any two levels of price and CO2 emissions; y=(25000 e,
150 g/km) and y’=(30000 e, 200 g/km)
q Consider any two levels of top speed; x’=200 km/h, x=250 km/h
q Does (250 km/h, 30000 e, 200 g/km) ← (200 km/h, 30000 e, 200 g/km) ~d (250
km/h, 25000 e, 150 g/km) ← (200 km/h, 25000 e, 150 g/km) hold?
q If yes (for all x,x’,y,y’), then difference independence holds
q That is, does the value of increased top speed depend on the levels of other attributes or not?
q Is the ”amount of” value added by a fixed change in top speed independent of the other
attributes?
28.3.2018
28
Additive value function
Theorem: If all attributes are mutually preferentially independent and each
attribute is difference independent of the others, then there exists an additive
value function
= ,…, = ( )
which represents preference relations ≽, ≽ in the sense that

≥ ⇔ ≽
− ≥ − ( )⇔( ← )≽ ( ← )
Note: The additive value function is unique up to positive affine transformations,

i.e., V(x) and V’(x)=αV(x)+β, α>0 represent the same preferences
28.3.2018
29
Conditions
q What if the conditions (mutual preferential independence and
difference independence) do not hold?
– Reconsider the attribute ranges [ , ∗ ]; conditions are more likely fulfilled
when the ranges are small
– Reconsider the attributes; are you using the right measures?
q Even if the conditions do not hold, additive value function is often

used to obtain approximate results
28.3.2018
30
Example (Ewing et al. 2006*): military
value of an installation
• “How to realign US Army units and which bases to close in order to
operate more cost-efficiently?”
• Many attributes, including ”total heavy maneuver area” (x1) and

”largest contiguous area” (x2; a measure of heavy maneuver area
quality)
- ”Total heavy maneuver area” is not difference independent of the other attributes x2
∪ ′′ because (1000 ha, 100 ha, y’’) ← (100 ha, 100 ha, y’’) ~d (1000 ha, 10 ha, y’’)
← (100 ha, 10 ha, y’’) as the ncrease from 100 to 1000 ha in total area is found quite
useless, if total area consists of over 100 small isolated pieces of land
28.3.2018
* Ewing, Tarantino, Parnell (2006): Use of Decision Analysis in the Army Base 31
Realignment and Closure (BRAC) 2005 Military Value Analysis. Decision Analysis 3, 33-49
Example (Ewing et al. 2006*): military
value of an installation
q Solution: unite the two attributes x1 and x2 into one attribute ”heavy
maneuver area”
q Then (1000 ha, 100 ha, Y) ← (100 ha, 100 ha, Y) ≻d (1000 ha, 10 ha, Y) ← (100
ha, 10 ha, Y) does not violate required difference independence conditions
( , ) ← ( , ′)~ ( , ) ← ( , ) for all ∈ , because x2 is no longer an element
of y or y’
q BUT we need to elicit preferences between different ’pairs’ (x1, x2)
28.3.2018
32
Normalized form of the additive value
function = ,…, = ( )
q Denote
– = Least preferred level w.r.t to attribute i
∗
– = Most preferred level w.r.t to attribute i
q Then,
= − + ( )
=∑ −∑ + =∑ [ − ]+
∗
=∑ [ − ] ∗ +
=∑ ∗ + ∗ + …
28.3.2018
33
Normalized form of the additive value
function (cont’d)
…=∑ + + Normalized attribute-
∈[ , ]
specific value
function ( ) ∈
=∑ ∑ ( ) + [0,1]
∑
,∑
= ∑ ∑ ( )+
( ) Normalized additive value function
=χ + ( )=∑ ( ) ∈ [0,1]
=χ + is a positive affine transformation of

; they represent the same preferences!
28.3.2018
34
Summary: Using the additive value
function
q Elicit the attribute-specific value functions
– Use indifference methods if possible
– If not, use direct elicitation methods
q Normalize these value functions to get ( )
q Elicit attribute weights , = 1, … , (next lecture)
q Compute the alternatives’ overall values ( )=∑ ( )
q Select the alternative with the highest overall value
q To be continued…
28.3.2018
35
Decision making and
problem solving –
Lecture 6
• Interpretation of attribute weights
• Elicitation of attribute weights
• Trade-off methods
• SWING and SMART(S) methods
• Ordinal weighting methods
Recap from previous lecture (1/3)
q Given certain axioms, a DM’s preferences about a single attribute can be represented
by a cardinal value function such that
≥ ⟺ ≽
− ≥ − ⟺ ← ≽ ← .
q Attribute-specific value functions are obtained by

– Defining measurement scales [ , ∗ ]
– Asking a series of elicitation questions through, e.g.,
1. Bisection method
2. Equally preferred differences
3. Giving a functional form; e.g., is linear and increasing
∗
q Value functions can be normalized such that = 0 and = 1.
28.3.2018
2
q Attribute X is preferentially independent of the other attributes Y, if
preference between different performance levels on X does not
depend on the levels of the other attributes.
q Attributes A are mutually preferentially independent, if any subset of

attributes X is preferentially independent of the other attributes Y=A\X
q Attribute X is difference independent of the other attributes Y, if

preference over a change in attribute X does not depend on the levels
of the other attributes.
28.3.2018
3
q If the attributes are mutually preferentially independent and each
attribute is difference independent of the others, then there exists
an additive value function
( )=∑ ( )
such that
≥ ⟺ ≽
− ′ ≥ − ′ ⟺ ← ′ ≽ ← ′ .
q Decision recommendation: choose the alternative with the highest

overall value
28.3.2018
4
This lecture
q How to interpret the attribute weights ?
q How to elicit these weights?
28.3.2018
5
Interpretation of attribute weights
∗
∗
q By definition, =∑ =∑ ∗ ∝ −
( )
q Attribute weight reflects the increase in overall value when the performance level on
attribute ai is changed from the worst level to the best – relative to similar changes in
other attributes
q Weights thus reflect trade-offs between attributes; not their absolute ”importance”
q Elicitation of attribute weights without this interpretation is not meaningful

– Do not ask: ”What is more important: environment or economy?”
– Do ask: ”How much is society willing to pay to save an insect species?”
28.3.2018
6
Interpretation of attribute weights
q Correct interpretation and hence application of the weights may lead to
‘resistance’
q Let the least preferred and the most preferred levels in
q cost savings be 0 € and 1 B€ (“money”)
q the number of insect species saved from extinction in Finland be 0 and 1 (“environmental
aspects”)
q Environmental aspects are likely to receive a small weight, as for example weighting (0.5, 0.5)
would mean that we equally prefer saving 1 B€ and saving 1 species
q Cf. …. Let the least preferred and the most preferred levels in
q cost savings be 0 € and 1 B€
q the number of insect species saved from extinction in Finland be 0 and 100
28.3.2018
7
Elicitation of attribute weights
q Attribute weights are derived from the DM’s preference

statements
q Approaches to eliciting attribute weights:

– Trade-off weighting
– ”Lighter” techniques: SWING, SMART(S), and ordinal methods
28.3.2018
8
Trade-off weighting
q The DM is asked to
1. Set the performance levels of two imaginary alternatives x and y such that they are equally
preferred (x~y):
+ ⋯+ = + ⋯+ , or
2. Set the performance levels of four imaginary alternatives x, x’, y, and y’ such that changes
x ← x’ and y ← y’ are equally preferred ( ← ′~ ← ′):
( − )+ ⋯+ ( − )= ( − ) + ⋯+ ( − )
28.3.2018
9
Trade-off weighting
q n-1 pairs of equally preferred alternatives/changes → n-1 linear
constraints + 1 normalization constraint
q If the pairs are suitably selected (no linear dependencies), the system
of n linear constraints has a unique solution
– E.g., select a reference attribute and compare the other attributes against it
– E.g., compare the ”most important” attribute to the second most important, the
second most important to the third most important etc
28.3.2018
10
Trade-off weighting: example (1/7)
q Consider two magazines A and B reporting a comparison of cars
, , and , based on the same expert appraisal, using the
same attributes:
: Top speed : Acceleration : CO2 : Maintenance

km/h 0-100 km/h emissions g/km costs €/year
192 km/h 12.0 s 120 g/km 400 €/year
200 km/h 10.4 s 140 g/km 500 €/year
220 km/h 8.2 s 150 g/km 600 €/year
28.3.2018
11
q Consider changing top speed (reference attribute) from 150 to
250 km/h. All other things being equal, what would be an equally
preferred change in
– Acceleration time? Expert’s answer: from 14 to 7 s ⇒
250 − 150 = 7 − 14 ⇒ =
– CO2 emissions? Expert’s answer: from 100 to 0 g/km ⇒
250 − 150 = 0 − 100 ⇒ =

– Maintenance costs? Expert’s answer: from 800 to o €/year ⇒
250 − 150 = 0 − 800 ⇒ =
28.3.2018
12
q Attribute-specific value functions according to the expert:
1 1
0.8 0.8 (8.2,0.95)

(192,0.7) (220,0.87)
(200,0.75) 0.6 (10.4,0.75)
0.6
v2(x2)
v1(x1)
0.4 (180,0.5) 0.4 (12,0.5)
0.2 0.2
0 0
150 160 170 180 190 200 210 220 230 240 250 7 8 9 10 11 12 13 14
x : Top speed x2: Acceleration
1
1 1
0.8 0.8
0.6 0.6
v (x )
v4(x4)
3 3
0.4 0.4
0.2
0.2
0
0 50 100 150 200 250 0
0 100 200 300 400 500 600 700 800 900 28.3.2018
1000
x3: CO2 emissions 13
x4: Maintenance costs
q Magazine A uses the following measurement scales:
Attribute Measurement scale
: Top speed (km/h) [150, 250] 180 = 0.5, 192 = 0. 7, 200 = 0.75, 220 = 0.87
: Acceleration time (s) [7, 14] 12 = 0.5, 10.4 = 0.75, 8.2 = 0.95
: CO2 emissions (g/km) [120, 150] 5− /30
: Maintenance costs (€/year) [400,600] 3− /200
– = = 1
100
30 ( 120 − 150 ) 10
– = = 1
=
3
800
– = = 200 ( 400 − 600 )
=4
1
q The three equalities and ∑ = 1 give = = 0.39, = 0.12, = 0.10.
28.3.2018
14
q Magazine A reports the alternatives’ attribute-specific values multiplied by 10
(i.e., scaled to interval [0,10]) and the attribute weights:
: Top speed : Acceleration : CO2 : Maintenance Overall value
7 5 10 10 6.86
7.5 7.5 3.3 5 6.76
8.7 9.5 0 0 7.14
Weights 39% 39% 12% 10%
q Possible (mis)interpretations / ”headlines”:

– ”Only power matters – minor emphasis on costs and environment”
– ”Car terrible w.r.t. CO2 emissions and maintenance costs – yet, it’s the expert’s choice!”
– ”No significant differences in top speed – differences are in CO2 emissions and maintenance
costs”
28.3.2018
15
q Magazine B uses the following measurement scales:
Attribute M. scale
: Top speed [192, 220] 150 = −4.12, 180 = −1.18, 192 = 0, 200 = 0.29, 220 = 1, 250 = 1.76
: Acceleration [8.2, 12] 14 = −1.11, 12 = 0, 10.4 = 0.56, 8.2 = 1, 7 = 1.11
: CO2 emissions [0, 250] 1− /250
: Maintenance [0,1000] 1− /1000
. .
– = = = 0.378
. .
– = = = 0.068
. .
– = = = 0.136
. .
q The three equalities and ∑ = 1 give = 0.039, = 0.103, = 0.572, = 0.286.
28.3.2018
16
q Magazine B reports the alternatives’ attribute-specific values multiplied by 10
(i.e., scaled to interval [0,10]) and the attribute weights:
: Top speed : Acceleration : CO2 : Maintenance Overall value
0 0 5.2 6 4.7
2.9 5.6 4.4 5 4.6
10 10 4 4 4.9
Weights 3.9% 10.3% 57.2% 28.6%
q Possible (mis)interpretations:
– ”Emphasis on costs and environmental issues”
– ” wins only on the least important attributes – yet, it’s the expert’s choice!”
– ”Car terrible w.r.t. top speed and acceleration time”
28.3.2018
17
Trade-off weighting
q Weights reflect value differences over the measurement scales →
changing the measurement scales changes the weights
q The attribute-specific values used in trade-off weighting take the

measurement scales explicitly into account → weights represent
the DM’s preferences regardless of the measurement scales
q Trade-off weighting has a solid theoretical foundation and requires

thinking; use whenever possible
28.3.2018
18
SWING
q Swing-weighting process:
1. Consider alternative = ( , … , ) (each attribute on the worst level).
2. Choose the attribute that you would first like to change to its most
preferred level ∗ (i.e., the attribute for which such a change is the most
valuable). Give that attribute a (non-normalized) weight = 100.
3. Consider again. Choose the next attribute that you would like to
change to its most preferred level. Give it weight ∈ 0,100 that reflects
this improvement relative to the first one.
4. Repeat step 3 until all attributes have been weighted.
5. Obtain weights by normalizing .
28.3.2018
19
SWING: example
q Magazine A’s measurement scales
€
– Alternative = 150 , 14 , 150 , 600
Attribute Measurement
– The first attribute to be changed from the worst to scale
the best level: → = 100 : Top speed [150, 250]
– The second attribute: → = 100 : Acceleration [7, 14]
– The third attribute: → = 30 : CO2 emissions [120, 150]
: Maintenance [400,600]
– The fourth attribute: → = 20
– Normalized weights: = = 40% =
12%, = 8%.
28.3.2018
20
About SWING weighting
q The mode of questioning explicitly (but only) considers the
least and most preferred levels of the attributes
q Assumes that the DM can directly numerically assess the strength of preference of
changes between these levels
q NOTE that we only have two preference relations: ≽ and ≽

∗
q For example preference statement = 100, = 20 is equal to −
∗ . . .
= 5[ − ], which assumes that there exist levels , , ,
. . .
so that ⟵ ∼ ⟵ . ∼ …∼ ( ∗⟵ . )
∗ . ∗ .
q Then − =5 − = 5[ − ] if ( , , , )⟵
∗
( , , , )∼ ( , , , )⟵( , , , )
28.3.2018
21
SMART
q Simple Multi-Attribute Rating Technique process:
1. Select the least important attribute and give it a weight of 10 points.
2. Select the second least important attribute and give it a weight (≥10 points) that
reflects its importance compared to the least important attribute.
3. Go through the remaining attributes in ascending order of importance and give
them weights that reflect their importance compared to the less important
attributes.
4. Normalize the weights.
q This process does not consider the measurement scales at all →

interpretation of weights is questionable
28.3.2018
22
SMARTS
SMARTS = SMART using Swings
1. Select the attribute corresponding to the least preferred change from
worst to best level and give it a weight of 10 points.
2. Go through the remaining attributes in ascending order of preference over
changing the attribute from the worst to the best level, and give them
weights that reflect their importance compared to the less preferred
changes.
3. Normalize the weights.
28.3.2018
23
SMARTS: example
q Magazine A’s measurement scales
€
– Alternative = 150 , 14 , 150 , 600
– Least preferred change from the worst to the best Attribute Measurement
scale
level: → = 10 : Top speed [150, 250]
– The second least preferred change: → = 20 : Acceleration [7, 14]
– The third least preferred change : → = 40 : CO2 emissions [120, 150]
– The fourth least preferred change: → = 40 : Maintenance [400,600]
– Normalized weights: = = 36%, =

18%, = 9%.
28.3.2018
24
Empirical problems related to SWING &
SMARTS
q People tend to use only multiples of 10 when assessing the
weights, e.g.,
– SWING: = = 100, = 30, = 20 → = = 0.40, = 0.12, = 0.08
– SMARTS: = = 40, = 20, = 10 → = = 0.36, = 0.18, = 0.09
® SWING and SMARTS typically produce different weights
q Assessments may reflect only ordinal, not cardinal information

about the weights
– E.g., SMARTS weights = 10 and = 20 only imply that < , not that
/ =2
28.3.2018
25
Ordinal weighting methods
q The DM is only asked to rank the attributes in terms of their
importance (i.e., preferences over changing the attributes from the
worst to the best level)
– = 1 for the most important attribute
– = for the least important attribute
q This ranking is then converted into numerical weights such that

these weights are compatible with the ranking
– > ⇔ <
28.3.2018
26
q Rank sum weights are proportional to the opposite number of the
ranks
∝( − + 1)
q Rank exponent weights are relative to some power of ( − +

1)
∝( − + 1)
– If z > 1 (z < 1), the power increases (decreases) the weights of the most
important attributes compared to Rank sum weights.
28.3.2018
27
q Rank reciprocal weights are proportional to the inverse of the ranks
1
∝
q Centroid weights are in the center of the set of weights that are
compatible with the rank ordering
– Order the attributes such that ≥ ≥⋯≥ .
– Then, the extreme points of the compatible weight set are (1,0,0,0…), (½, ½,0,0,…),
(1/3, 1/3, 1/3,0,…),… (1/n,…,1/n).
– The average of these extreme points is
1 1
=
28.3.2018
28
Example: centroid weights
w3
q Rank ordering ≥ ≥ : (0,0,1)
1 1 1
, ,
3 3 3
1 1 1 11
= 1+ + = ≈ 0.61 =
3 2 3 18
1 1 1 5 w1
= + = ≈ 0.28 =
3 2 3 18
1 1 1 (1,0,0)
= = ≈ 0.11 1 1
3 3 9 w2 , ,0
2 2
(0,1,0) ≥ ≥
28.3.2018
29
Ordinal weighting methods: example
q Four attributes , , , in descending order of importance → =
1, = 2, = 3, = 4.
a1 a2 a3 a4 ∑
Rank sum 4 3 2 1 10
weights 0.4 0.3 0.2 0.1 1
Rank exp(z=2) 16 9 4 1 30
weights 0.53 0.30 0.13 0.03 1
Rank rec. 1 1/2 1/3 1/4 25/12
weights 0.48 0.24 0.16 0.12 1
Centroid 25/48 13/48 7/48 3/48 1
weights 0.52 0.27 0.15 0.06 1
q Different methods produce different weights!
28.3.2018
30
Ordinal weighting methods: example
(cont’d)
q Assume that the measurement scale of the most important
attribute is changed from [0€,1000€] to [0€,2000€].
∗
q Because ∝ − ( ), the weight of attribute should be
even larger.
q Yet,
– Ranking among the attributes remains the same → rank-based weights
remain the same
– The alternatives’ normalized scores on attribute become smaller →
attribute has a smaller impact on the decision recommendation
q Avoid using ordinal methods
28.3.2018
31
Weighting in value trees Ideal car
q Two modes of weighting = 0.22 = 0.78

Economy Driving
– Hierarchical: all weights are elicited
and then multiplied vertically
o Problem: elicitation questions for the higher- Price Expenses Acceleration Top speed
level attributes are difficult to interpret: 0.45 0.55 0.50 0.50
= 0.22 0.45 = 0.10 = 0.12 = 0.39 = 0.39
= + ∝ ( ( ∗ )- ( )) +( ( ∗ )- ( ))
→ Avoid!
Ideal car
– Non-hierarchical: weights are only

elicited for the twig-level attributes Economy Driving
Price Expenses Acceleration Top speed
= 0.10 = 0.12 = 0.39 = 0.39
0.22 0.78
28.3.2018
32
Summary
q The only meaningful interpretation for attribute weight :
The improvement in overall value when attribute is changed from its worst
level to its best relative to similar changes in other attributes
q Prefer weighting methods that apply indifference relations between

alternatives / changes in alternatives
– Trade-off methods
– SWING and SMARTS – but be aware of associated problems
q Avoid weighting methods based on rank-orderding among the attributes
q Weighting other than twig-level attributes is difficult
28.3.2018
33
Decision making and
problem solving –
Lecture 7
• Incomplete preference statements
• Modeling incomplete information
• Dominance and non-dominated alternatives
• Computing dominance relations
Recap: elements of MAVT
q Elements of MAVT:
– Alternatives = ,…,
– Attributes = ,…,
– Attribute weights = [ , … , ]∈ℝ
– Attribute-specific (normalized) values ∈ℝ × , = ∈ [0,1]
– Overall values of alternatives , , =∑ , = 1, … ,
28.3.2018
2
Recap: Elicitation of attribute weights
q Attribute weights are elicited by defining equally preferred
alternatives / equally preferred changes between alternatives
– E.g., ”All else being equal, a change 150 → 250 km/h in top speed is
equally preferred to a change 14 → 7 s in acceleration time” ⇒
250 − 150 = 7 − 14 ⟺
7 − 14
=
250 − 150
q Question: What if the DM finds it difficult or even impossible to
define such alternatives / changes?
– E.g., she can only state that a change 150 → 250 km/h in top speed is
preferred to a change 14 → 7 s in acceleration time?
28.3.2018
3
This lecture
q We learn how to
– Accommodate incomplete preference statements in the decision
model
– Generate robust decision recommendations that are compatible
with such statements; i.e., recommendations that are not sensitive
to (small) changes in model parameters
28.3.2018
4
Incomplete preference statements
q Set the performance levels of two imaginary alternatives and such that
≽ ⇒
+ ⋯+ ≥ + ⋯+ .
q For instance, a change 150 → 250 km/h in top speed is preferred to a change
14 → 7 s in acceleration time ⇒
7 − 14
≥
250 − 150
q Incomplete preference statements result in linear inequalities between the

weights
28.3.2018
5
Incomplete preference statements:
example
q Consider attributes
– CO2 emissions ∈ [120 , 150 ]
– Maintenance costs ∈ [400€ , 600€]
q Preferences are elicited with SMARTS:

– Q: ”If the change 600€ → 400€ in maintenance costs is worth 10 points,
how valuable is change 150g → 120g in CO2 emissions?”
– A: ”Between 15 and 20 points”
1.5 [ 400 − 600 ] ≤ [ 120 − 150 ] ≤ 2 [ 400 − 600 ]
⇒ 1.5 ≤ ≤2
28.3.2018
6
Incomplete preference statements:
example
q Consider again attributes
– CO2 emissions ∈ [120 , 150 ]
– Maintenance costs ∈ [400€ , 600€]
q Preferences are elicited with trade-off methods:

– Q: ”Define an interval for x such that 600€ → 400€ in maintenance costs is
as valuable as 150 g → x g in CO2 emissions.”
– A: ”x is between 130 and 140 g”
[ 140 − 150 ] ≤ [ 400 − 600 ] ≤ [ 130 − 150 ]
⇒ 140 ≤ ≤ 130
⇒ ≤ ≤ , if is linear and decreasing.
28.3.2018
7
Modeling incomplete informaation
q Incomplete information about attribute weights is modeled as set
of feasible weights that are consistent with the DM’s preference
statements:
⊆ = ∈ℝ | = 1, ≥0 ∀
28.3.2018
8
Modeling incomplete information
w3 S
q Linear inequalities on weights can (0,0,1)
correspond to
1. Weak ranking ≥ 3 =
2. Strict ranking − ≥ 4 =
3. Ranking with multiples ≥ w1
(equivalent to incompletely defined (1,0,0)
weight ratios / ≥ ) =
w2
4. Interval form ≤ ≤ + (0,1,0) 2 =
5. Ranking of differences − ≥ −
≤ ≤3 ,
2 ≤ ≤4
28.3.2018
9
Overall value intervals
q Due to incompletely specified weights,
the alternatives’ overall values are
intervals: ( )
, , ∈ min ( , , ) , max ( , , ) ( )
∈ ∈ ( )
Value
intervals
q Note: linear functions obtain their minima
and maxima at an extreme point of 0.4 0.7
– E.g., = ∈ ⊆ ℝ |0.4 ≤ ≤ 0.7 ⇒ 0.6 0.3

= 0.4, 0.6 , (0.7, 0.3) =1−
28.3.2018
10
Dominance
q Preference over interval-valued alternatives can be established through a
dominance relation
q Definition: dominates in , denoted ≻ , iff

, , ≥ , , for all ∈
, , > , , for some ∈
i.e., iff the overall value of is greater than or equal to that of for all
feasible weights and strictly greater for some.
28.3.2018
11
Non-dominated alternatives
q An alternative is non-dominated if no other alternative dominates it
q The set of non-dominated alternatives is
= ∈ |∄ such that ≻
q contains all good decision recommendations

– I.e., alternatives compared to which no other alternative has at least as high
value for all feasible weights and strictly higher for some
28.3.2018
12
Non-dominated alternatives
is non-dominated if no other alternative has
higher value than for all feasible weights
( )
• Alternative dominates
( )
( )
• Alternatives and are non-dominated
Value
intervals
0.4 0.7
0.6 0.3
=1−
28.3.2018
13
Non-dominated vs. potentially optimal
alternatives
q A non-dominated alternative is not necessarily
optimal for any ∈
( ) ( )
– , and are all non-dominated

– Only and are potentially optimal in that they maximize V ( )
for some ∈
– Still, neither of them can be guaranteed to be better than
0.4 0.7
0.6 0.3
28.3.2018
14
Properties of dominance relation
Non-dominated
A J L
q Transitive alternatives
– If A dominates B and B K
dominates C, then A dominates
C I
q Asymmetric B G C E
– If A dominates B, then B does
not dominate A
q Irreflexive
D F H
– A does not dominate itself Dominance relations
expressed with a directed
arc: B dominates D
28.3.2018
15
Computing dominance relations
q If dominates :
1. , , ≥ , , for all ∈
⇔ min , , − , , ≥ 0 ⇔ min ∑ ( − ) ≥0
∈ ∈
2. , , > , , for some ∈

⇔ max , , − , , > 0 ⇔ max ∑ ( − ) >0
∈ ∈
q Dominance relations between two alternatives can thus be

established by comparing their minimum and maximum value
differences
28.3.2018
16
Computing dominance relations:
example
q Consider three cars with normalized attribute-specific values:
Car : Top speed : Acceleration : CO2 emissions : Maintenance
0.7 0.5 1 1
0.75 0.75 0.33 0.5
0.87 0.95 0 0
q Incomplete preference statements have resulted in feasible set of

weights S:
= ∈ ⊆ℝ | = ≥3 , ≥ ≥ 0.1
28.3.2018
17
example
Matlab function
linprog(f,A,b,Aeq,beq)
solves the optimization
problem:
min such that

≤
=
28.3.2018
18
example
q Minimum and maximum value differences
min , , − , , = −0.003 < 0 → Neither nor
∈
max , , − , , = 0.0338 > 0 dominate the other
∈
min , , − , , = −0.045 < 0

∈ → dominates
max , , − , , = −0.0163 < 0
∈
min
∈
, , − , , = −0.048 < 0 → Neither nor
max , , − , , = 0.0175 > 0 dominate the other
∈
q ={ , }
28.3.2018
19
example
q Note: because value differences are linear
in w, minimum and maximum value
differences are obtained at the extreme
points of set S:
= 0.4 0.4 0.1 0.1
27 27 9 1
= , , , ≈ (0.386, 0.386, 0.129, 0.10)
70 70 70 10 - -0.003 0.0204 0.0338
3 3 1 1
= , , , = (0.375, 0.375, 0.125, 0.125) − -0.045 -0.031 -0.0163
8 8 8 8
− -0.048 -0.0106 0.0175
28.3.2018
20
Additional information
q If information set S results in too many non-dominated alternatives, additional
preference statements (i.e., linear constraints) can be elicited
q New information set ′ ⊂ preserves all dominance relations and usually

yields new ones → stays the same or becomes smaller
≻ ⇒ ≻
⊂ , ∩ ≠ ∅: ,
( )⊇ ( ′)
where is the relative interior of S.

– ∩ ≠ ∅: S’ is not entirely on the “border” of S
28.3.2018
21
Additional information: example
q No weight information
= = ∈ℝ | = 1, ≥0
E
q Dominance relations A
C
1. B dominates D B
B
2. C dominates D
C D
q Non-dominated alternatives
D A
– A,B,C,E
E
=0 =1
=1 =0
28.3.2018
22
Additional information: example (2/3)
q Ordinal weight information
= ∈ | ≥
q Dominance relations A
E
1. B dominates D C
B
2. C dominates D B
3. E dominates D C D
4. B dominates A
5. C dominates A
D A
E
q Non-dominated alternatives =1
=0 = 0.5
– B,C,E =1 = 0.5 =0
28.3.2018
23
Additional information: example (3/3)
q More information
= ∈ | ≤ ≤2
q Dominance relations
E
1. B dominates D A
2. C dominates D B C
3. E dominates D B
4. B dominates A C D
5. C dominates A
6. B dominates C D A
7. B dominates E E
=0 = 0.5 = 0.67 =1
q Non-dominated alternatives: B =1 = 0.5 = 0.33 =0
28.3.2018
24
Value intervals Central values
Maximax
Can value intervals be used in deriving Maximin
decision recommendations?
Some suggestions for “decision rules” from
literature: ( )
• Maximax: choose the alternative with the

highest maximum overall value over the ( )
feasible weights ( )
• Maximin: choose the alternative with the Value

highest lowest overall value over the feasible intervals
weights
0.4 0.7
• Central values: choose the alternative with the
highest sum of the maximum and minimum 0.6 0.3
=1−
values
28.3.2018
25
…more decision rules Domain criterion
Minimax regret
• Minimax regret: choose the

alternative with the smallest maximum ( )
regret (= value difference compared to
any other alternative)
( )
• Domain criterion: choose the ( )
alternative which is favored by the
largest set of weights
0.4 0.55 0.7
• 0.6 0.3
=1−
28.3.2018
26
Example
q DM asks 2 experts to compare fruit baskets (x1,x2) containing
apples x1 and oranges x2
q Linear attribute-specific value functions v1 and v2
q DM: (2,0) >~ (0,1) and (0,2)>~(1,0)

q One orange is not preferred to 2 apples, one apple is not preferred to 2 oranges
q Fruit baskets (1,2) and (2,1) do not dominate each other
q What do the decision rules recommend?
28.3.2018
27
Expert 1: Expert 2:
x0=(0,0), x*=(2,4) x0=(0,0), x*=(4,2)
x1 N x x1 N x
v1N ( x1 ) = , v2 ( x2 ) = 2 v1N ( x1 ) = , v2 ( x2 ) = 2
2 4 4 2
V (2, 0) ³ V (0,1) Û V ( 2 , 0 ) ³ V ( 0 ,1) Û

2 1 1 1 2 1 1 1
w1 + 0w2 ³ 0w1 + w2 = (1 - w1 ) Û w1 ³ w1 ³ w 2 = (1 - w 1 ) Û w1 ³
2 4 4 5 4 2 2 2
V (0, 2) ³ V (1, 0) Û V ( 0, 2 ) ³ V (1, 0 ) Û

2 1 1 1 1 4
w2 = (1 - w1 ) ³ w1 Û w1 £ w 2 = 1 - w1 ³ w 1 Û w1 £
4 2 2 2 4 5
x1 x æx x ö x æ x x ö x
V ( x) = w1 + w2 2 = w1 ç 1 - 2 ÷ + 2 V ( x ) = w1 ç 1 - 2 ÷ + 2
28.3.2018
28
2 4 è2 4ø 4 è 4 2 ø 2
æx x ö x
V ( x) = w1 ç 1 - 2 ÷ + 2
è2 4ø 4
æ1 2ö 2 1 æx x ö x
V (1, 2) = w1 ç - ÷ + º V ( x) = w1 ç 1 - 2 ÷ + 2
è2 4ø 4 2 è4 2ø 2
æ2 1ö 1 3 1 3
V (2,1) = w1 ç - ÷ + = w1 + V (1, 2) = - w1 + 1
è2 4ø 4 4 4 4
1
(1,2) V (2,1) º (1,2)
(2,1)
0.6
2 (2,1)
0.6
2/15<1/6 2/15<1/6
0.125 0.125
0.5 0.5
0.1 0.1
0.4
1/6 0.4 1/6
V
V
(2,1) is the maximax solution (1,2) is the maximax solution

0.3 0.3
(1,2) is the maximin solution (2,1) is the maximin solution
0.2 (2,1) is the minimax regret solution 0.2 (1,2) is the minimax regret solution
(2,1) is the domain criterion solution (1,2) is the domain criterion solution
0.1 0.1
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.5 0.55 0.6 0.65 0.7 0.75 0.8
w1 w1
On decision rules
q A common problem for all of the above decision rules: changing
the measurement scales [xi0,xi*] can change the recommendations
q Different attribute weightings w and w* represent value functions V

and V* – they cannot be compared
q if V represents the DM’s preferences, so does all its positive affine
transformations, too
q How to choose one of the value functions that represent the same preferences?
q Avoid using measures which compare different value

functions
28.3.2018
30
Rank (sensitivity) analysis
V ( x3 )
q For any weights, the alternatives can
be ranked based on their overall
V ( x2 )
values
V ( x1 )
q This ranking is not influenced by
normalization (i.e., positive affine w1
transformations of V) w2 0.4 0.7
0.6 0.3
q How do the rankings of alternatives

change when attribute weights vary? ranks x1 x2 x3
minimum 1 1 1
maximum 3 2 3
28.3.2018
31
Computation of rank intervals
The minimum ranking of xk is
rS- ( x k ) = 1 + min | {x j Î X | V ( x j , w, v) > V ( x k , w, v)}|
( w ,v )ÎS
which is obtained
m
as a solution to the mixed integer LP
min
( w ,v )ÎS
å y
j =1
j
y j Î{0,1}
V ( x j , w, v) £ V ( x k , w, v) + y j M j = 1,..., m
y =1
k
Maximum rankings with a similar model
28.3.2018
32
Rank analysis – example (1/5)
q Academic ranking of world universities 2007
q 508 universities
q Additive multi-attribute model

q 6 attributes
q Attribute weights (denoted by w*) and scores
q Universities ranked based on overall values
28.3.2018
33
28.3.2018
34
Rank analysis –
example (3/5)
Scores (some of them)
28.3.2018
35
Incomplete weight information
qRelative intervals: w Î {w Î S w0 | (1 - a ) wi* £ wi £ (1 + a ) wi*}

q For a=0.1, 0.2, 0.3
q e.g. a=0.2 , wi*=0.20: 0.16 £ wi £ 0.24
qIncomplete ordinal: w Î{w Î S w | wi ³ wk ³ 0.02 " i Î{2,3, 4,5}, k Î {1, 6}}

0
q Consistent with initial weights and lower bound b = 0.02
qOnly lower bound: w Î {w Î S w | wi ³ 0.02"i = 1,..., 6}

0
qNo weight information: w Î S w

0
28.3.2018
36
exact weights
Unsensitive rankings
20 % interval
University
30 % interval
”Different weighting would
likely yield a better ranking”
incompl. ordinal
no information
10th 442nd
28.3.2018
37
Ranking
Example: prioritization of innovation
ideas*
q 28 ”innovation ideas” evaluated by several people on a scale from
1 – 7 with regard to novelty, feasibility and relevance
q Innovation ideas described by the 3 averages of these evaluations
q No preference information about the relative values of the
attributes
q ”Which 10 innovation ideas should be selected for further
development?”
q Sets of ideas called portfolios
q The value of a portfolio is the sum of its constituent projects
* Könnölä et al. (Technological Forecasting & Social Change, 2007) 28.3.2018

38
Example: prioritization of innovation
ideas
q Robust Portfolio Modeling* method was used to compute non-
dominated portfolios of 10 ideas and discriminate between
q Core ideas that belong to all non-dominated portfolios
q Borderline ideas that belong to some non-dominated portfolios
q Exterior ideas that do not belong to any non-dominated portfolio
q How do ranking intervals compare with this division?

q If the ranking of an idea cannot be worse than 10, is it a core project?
q If the ranking of an idea cannot be better than 11, is it an exterior project?
28.3.2018
* Liesiö, Mild, Salo (European Journal of Operational Research, 2007) 39
Ranking intervals vs. core, borderline
and exterior ideas
28.3.2018
40
Ranking intervals divide the innovation ideas into core, borderline and exterior ideas
among potentially optimal portfolios
Rationales for using incomplete
information
q Limited time and effort can usually be devoted to preference
elicitation
q Complete preference specification may not even be needed to reach
a decision
q DM’s preferences may evolve during the analysis → iteration can be
helpful
q Experts / stakeholders may have conflicting preferences
q Take-it-or-leave-it solutions may be resented in group decision
settings → results based on incomplete information leave room for
negotiation
28.3.2018
41
Summary
q Complete specification of attribute weights is often difficult
– Trade-off methods take time and effort
– SWING and SMARTS are prone to biases
q Incomplete preference statements can be modeled by linear inequalities on

the weights → alternatives’ overall values become intervals
q Preference over interval-valued alternatives can be established through

dominance relations
q Non-dominated alternatives are good decision recommendations
q Avoid methods which compare numerical values of different value functions
28.3.2018
42
Decision making and
problem solving –
Lecture 8
• From EUT to MAUT
• Axioms for preference relations
• Assessing attribute-specific utility functions and attribute weights
• Decision recommendations
• MAVT vs. MAUT Liesiö, Punkka, Salo, Vilkkumaa
Motivation
q Multiattribute value theory helps generate decision
recommendations, when
– Alternatives are evaluated w.r.t. multiple attributes
– Alternatives’ attribute-specific values are certain
q What if the attribute-specific performances are uncertain?
– Planning a supply chain: minimize cost, minimize supply shortage,
minimize storage costs
– Building an investment portfolio: maximize return, minimize risk
→ Multiattribute utility theory
28.3.2018
2
From EUT to MAUT
EUT Lottery
q Set of possible outcomes T: Decision tree PMF
2M€ 0.6, = 2 €
– E.g., revenue = ℝ euros, demand = 0.6
0.3 0.3, = 1 €
ℕ 1M€ =
0.1, = −0.5 €
0.1 0, ℎ
q Set of all possible lotteries L: -0.5M€
– A lottery ∈ associates a probability

∈ [0,1] with each possible outcome Degenerate lottery
∈
Decision tree PDF
q Deterministic outcomes are modelled 1 1, = 1 €
1M€ =
as degenerate lotteries 0, ℎ
28.3.2018
3
From EUT to MAUT
MAUT
Lottery
q Multidimensional set of outcomes
X:
(2M€,10%) 0.6, = (2,0.1)
= × ⋯× 0.6
0.3 0.3, = (1,0.2)
(1M€,20%) =
– E.g., = revenue ($), = market 0.1, = (−0.5,0.4)
0.1 0, ℎ
share (-0.5M€,40%)
q Set of all possible lotteries L:

– A lottery ∈ associates a Degenerate lottery
probability ∈ [0,1] with each Decision tree PDF
possible outcome = ( , … , ) ∈ 1 1, = (1,0.2)
(1M€,20%) =
q Deterministic outcomes are 0, ℎ
modelled as degenerate lotteries
28.3.2018
4
Aggregation of utilities
q Problem: How to measure the overall utility of alternative =
, ,… ?
, ,… =?
q Question: Could the overall utility be obtained by a weighted sum of
the attribute-specific utilities?
, ,… = ?
q Answer: Yes, if the attributes are

– Mutually preferentially independent and
– Additive independent (new)
28.3.2018
5
Preferential independence (old)
q Definition: Attribute X is preferentially independent (PI) of the
other attributes Y, if the preference order of degenerate lotteries
that differ only in X does not depend on the levels of attributes Y
( , )≽( , )⇒ , ′ ≽ , ′ for all ′ ∈ Y
q Interpretation: Preference over the certain level of attribute X does

not depend on the certain levels of the other attributes, as long as
they stay the same
q Same as in MAVT
28.3.2018
6
Mutual preferential independence (old)
q Definition: Attributes A are mutually perferentially independent
(MPI), if any subset X of attributes A is preferentially independent
of the other attributes Y=A\X. I.e., for any degenerate lotteries:
( , ) ≽ ( , ′) ⇒ , ≽ , for all y ∈ Y.
q Interpretation: Preference over the certain levels of attributes X

does not depend on the certain levels of the other attributes, as
long as they stay the same
q Same as in MAVT
28.3.2018
7
Additive independence (new)
q Definition: Subset of attributes X⊂A is additive 0.5
( , )
independent (AI), if the DM is indifferent between I
0.5 ( ′, ′)
lotteries I and II for any , , ( ′, ) ∈ 0.5
( , ′)
II
( ′, )
q Example: 0.5
– Profit is AI if the DM is indifferent between I and II

– However, she might prefer II, because it does not include an
outcome where all attributes have very poor values. In this 0.5
case profits is not AI. (20M€, 10%)
I
0.5 (1M€, 2%)
0.5
(20M€, 2%)
II
(1M€, 10%)
0.5
28.3.2018
8
Additive independence (new)
q Example:
– A tourist is planning a downhill skiing weekend trip to the mountains
– 2 attributes: sunshine ( {sun, no sun} ) and snow conditions ( {snow, no snow} )
– Additive independence holds, if she is indifferent between I and II
0.5
(sun, no snow)
I
0.5 (no sun, snow)
0.5
(sun, snow)
II
(no sun, so snow)
0.5
28.3.2018
9
Additive multiattribute utility function
q Theorem: The attributes are mutually preferentially

independent and single attributes are additive
independent iff preference relation ≽ is represented by an
additive multi-attribute utility function
= ( ),
∗
where = 0, = 1, and ∑ = 1, ≥ 0.
28.3.2018
10
What if the MPI & AI do not hold?
q Definition: Attribute ∈ is utility independent (UI) if the preference order
between lotteries that have equal certain outcomes on attributes Y=A\X does
not depend on the level of these outcomes, i.e.,
, ≽ , ⇒ , ≽ , ∀ ′
q Example: If profit is UI, then the DM should prefer I for
Assume DM prefers I any a
(2M€,10%) (2M€, a%)
0.6 0.6
0.3 (1M€,10%) 0.3 (1M€, a%)
I I
0.1 (-0.5M€,10%) 0.1 (-0.5M€, a%)
II (1M€,10%) II (1M€, a%)
However, for a small market share (a), the

DM may be more risk averse and choose II
→ profit is not UI
28.3.2018
11
Mutual utility independence
q Definition: Attributes are mutually utility independent (MPI), if every
subset X ⊂ is the utility independent of the other attributes Y=A\X i.e.,
, ≽ , ⇒ , ≽ , ∀ ′
(2M€, a%) (b M€,15%)

0.6 0.6
0.3 (1M€, a%) 0.3 (b M€,10%)
I I
0.1 (-0.5M€, a%) 0.1 (b M€,2%)
II (1M€, a%) II (b M€,10%)
If DM prefers I for some a, she If DM prefers I for some b, she

should prefer I for all a AND should prefer I for all b
28.3.2018
12
Other multi-attribute utility functions
q If attributes are mutually utility independent, then preferences are
represented by a multiplicative utility function:
∏ [1 + ( )] 1
= −
q If each single attribute is utility independent, then preferences are
represented by a so-called multilinear utility function
q AI is the strongest of the three preference assumptions

– Let X ⊂ A. Then, X is AI ⇒ X is UI ⇒ X is PI
28.3.2018
13
Assessing attribute-specific utility
functions
q Use the same techniques as with a unidimensional utility function
∗
– Certainty equivalent, probability equivalent, etc. & scale such that = 0, = 1.
– Also direct rating often applied in practice
q What about the other attributes?
– Fixed at the same level in every outcome I ( ? , 4€)
– Do not matter! → Usually not explicitly shown to the DM
0.5
(50 apples, 4€)
II
(−10 apples, 4€)
0.5
, 4 = 0.5 50,4 + 0.5 (−10,4)
⇔ + 4 = 0.5 50 + 0.5 4 + 0.5 −10 + 0.5 4
⇔ = 0.5 50 + 0.5 −10
⇔ = 0.5 50 + 0.5 −10
28.3.2018
14
Example: Choosing a software supplier
q Three attributes: cost, delay, quality
∗
i Name Xi
1 Cost [10,40] k€ 40 10
2 Delay {1,2,…,30} days 30 1
3 Quality {fair, good, excellent} fair excellent
28.3.2018
15
∗
q Assessment of the attribute-specific utility i Name Xi
functions 1 Cost [10,40] k€ 40 10
– Quality: Direct assessment 2 Delay {1,2,…,30} days 30 1

o (fair)=0, (good)=0.4, (excellent)=1
3 Quality {fair, good, exc.} fair exc.
– Cost: Linear decreasing utility function
o =
– Delay: Assessment with CE approach

22 15 10
0.5 0.5 0.5

30 days 22 days 15 days
1 day 0.5 1 day 0.5 1 day

0.5
22 15 10
= 0.5 1 + 0.5 30 = 0.5 1 + 0.5 22 = 0.5 1 + 0.5 22
= 0.5 ∗ 1 + 0.5 ∗ 0 = 0.5 ∗ 1 + 0.5 ∗ 0.5 = 0.5 ∗ 1 + 0.5 ∗ 0.75 28.3.2018
= 0.5 = 0.75 = 0.875 16
( ) ( )
1 1 16 0.7143
Linear interpolation between specified values
2 0.9861 17 0.6786
1 3 18
0.9722 0.6429
0.9 4 19
0.9583 0.6071
0.8 5 20
0.9444 0.5714
0.7 6 0.9306 21 0.5357
0.6 7 0.9167 22 0.5
u2(x2)
0.5 8 0.9028 23 0.4375
0.4 9 0.8889 24 0.375

10 0.875 25 0.3125
0.3
11 0.85 26 0.25
0.2
12 0.825 27 0.1875
0.1
13 0.8 28 0.125
0
0 5 10 15 20 25 30 14 0.775 29 0.0625
x2 15 30
0.75 0
28.3.2018
17
Assessing attribute weights
q Attribute weights are elicited by constructing two equally preferred
degenerate lotteries
– E.g., ask the DM to establish a preference order for n hypothetical
alternatives , … , ∗, … , , = 1, … , .
– Assume that ∗ , , … , ≽ , ∗, … , ≽⋯≽ , ,…, ∗
– Then, for each i=1,…,n-1 ask the DM to define ∈ such that
… , ,… ~ … , ∗ ,…
⇒ … , ,… = … , ∗ ,…
⇒ ( ) =
– n-1 such comparisons + 1 normalization constraint ⇒ unique set of
weights
28.3.2018
18
q Assessment of the attribute weights
– Assume preferences 40k€, 1 day, fair ≽ 10k€, 30 days, fair ≽ 40k€, 30 days, exc.
– Choose delay ∈ {1, … , 30} such that 40, , ~ 10,30,
– Answer = 8 gives
40 + 8 + = 10 + 30 +
8 =
⇔ 0.9028 =
– Choose cost ∈ 10,40 such that , , fair ~ 40, , excellent
– Answer = 20 gives
20 + + fair = 40 + + excellent
20 =
2
⇔ =
3
– Attribute weights: ≈ , ,
28.3.2018
19
MAUT: Decision recommendations
q Consider m decision alternatives = ,…, , = 1, … , ,

where is a random variable with PDF ( )
q Alternatives are ranked by their expected (multiattribute) utilities
= ( ) ( )= ( ) ( )
∈ ∈
– Integral for continuous random variables
q In a decision tree, MAU is used just like unidimensional utility
28.3.2018
20
(35 €, 1, )
q Consider three suppliers:
– Supplier 1: Expensive, fair quality, can deliver with
no delay (21 €, 7, )
= (35 €, 1, )
– Supplier 2: Cheap, good quality, can deliver in 1 week
= (21 €, 7, ) 0.7 (24 €, 1, )
– Supplier 3: Moderate price, good quality, 20% chance 0.2 (24 €, 8, )
of 1-week delay and 10% chance of 2-week delay 0.1 (24 €, 15, )
= 24 €, , ,
0.7, = (24 €, 1, )
= 0.2, = (24 €, 8, )
0.1, = (24 €, 15, )
28.3.2018
21
E[ ]
0.17 1.00 0.00 0.46 1 0.46

0.63 0.92 0.40 0.69 1 0.69
( ) 0.53 1.00 0.40 0.69 0.7 0.67
( ) 0.53 0.90 0.40 0.65 0.2
( ) 0.53 0.75 0.40 0.59 0.1
0.36 0.40 0.24
28.3.2018
22
MAVT vs. MAUT
q MAVT: Preference between alternatives with certain outcomes can be
represented by an additive multiattribute value function, iff the
attributes are
– Mutually preferentially independent
– Difference independent
q MAUT: Preference between lotteries with uncertain outcomes can be

represented by additive multiattribute utility function, iff the attributes
are
– Mutually preferentially independent
– Additive independent
28.3.2018
23
MAVT vs. MAUT
q Attribute-specific value functions are elicited by asking the DM
to specify equally preferred differences in attribute levels
– E.g., ”Specify salary x such that you would be indifferent between change
1500€ → x€ and x€ → 2000€”
q Attribute-specific utility functions are elicited by asking the DM

to specify equally preferred lotteries
– E.g., ”Specify salary x such that you would be indifferent between getting
x€ for certain and a 50-50 gamble between getting 1500€ or 2000€”
q Attribute weights are elicited similarly in MAVT and MAUT
28.3.2018
24
MAVT vs. MAUT
q In principal, the natural /
measurement scale is first
mapped to value scale and
then (if needed) to utility scale
q Yet, in practice the value

function is ”hidden” in the utility
function
– E.g, if CE of 50-50 gamble between
3k€ and 5k€ salary is 3.9k€, is this a
sign of risk aversion or decreasing
marginal value of salary?
28.3.2018
25
Summary
q Multiattribute utility theory helps establish a preference relation
between alternatives with uncertain outcomes on multiple attributes
q Preference relation is represented by an additive utility function, iff the
attributes are mutually preferentially independent and additive
independent
q Attribute-specific utility functions are elicited as in the unidimensional
case
q Attribute weights are elicited as in MAVT
q Decision recommendation: the alternative with highest expected utility
q Robust methods can also be used with MAUT
28.3.2018
26
Decision making and
problem solving –
Lecture 9
• Multiple objective optimization (MOO)
• Pareto optimality (PO)
• Approaches to solving PO-solutions: weighted sum, weighted max-norm,
and value function methods
Until this lecture
=( , )
q Explicit set of alternatives =
1, … , , which are evaluated with =( , )
regard to criteria
q Evaluations : →ℝ
q Preference modeling
( )= ,
q Value functions
max = ( ,…, )
∈
28.3.2018
2
Need for other kind of approaches
q The decision alternatives cannot necessarily be listed
q Preference modeling can be time-consuming and difficult at the

early stages of the analysis
q Conditions required for the additive value function to represent

preferences do not necessarily hold or are difficult to validate
q We might want to see some results quickly to get a better

understanding of the problem at hand
28.3.2018
3
Multi-objective optimization: concepts
= {( , ) ∈ ℝ2|
q Set of feasible solutions
1 ≥ 1, 2 ≥ 1, 1 + 2 ≤ 7}
= { ∈ ℝ | ( ) ≤ 0}
q Objective functions
= ,…, : → ℝ
q Preference modeling on trade-offs
between objectives
– Value functions
max ( ( )) = ( ( ), … , ( ))
∈
– Pareto approaches
v−max ( ( )) = ( ( ), … , ( ))
∈
– Interactive approaches (not covered)
= , =( 1 + 2 2, − 2)
28.3.2018
4
= {( , ) ∈ ℝ2|
Multi-objective 1 ≥ 1, 2 ≥ 1, 1 + 2 ≤ 7}
optimization: concepts
q Objective functions map the
feasible solutions to in the 1,1
solution space:
= ,
1 = 1+2
= { ∈ ℝ |∃ ∈ ℎ 2
2 = − 2
= }
3, −1
( ) = {( , ) ∈ ℝ2|
( )
2 ≤ −1, 2 ≤ 7 − 1, 2 2 ≥ 1 − 1}
28.3.2018
5
Preferential independence
q In multi-objective optimization (MOO), each objective is assumed
preferentially independent of the others
q Definition (cf. Lecture 5): Preference between two values of
objective function i does not depend on the values of the other
objective functions
→ Without loss of generality, we can assume all objectives to be

maximized
– MIN can be transformed to MAX: min ( ) = − max[− ( )]
∈ ∈
28.3.2018
6
Which feasible solution(s) to prefer?
q Selection of 1
cannot be
supported because ( )
Better than 1 on
other solutions both objectives
have higher 1 and
2
→ Focus on
Pareto-optimal 1
solutions
28.3.2018
7
Pareto-optimality f2 f ( X PO )
Definition. x*ÎX is Pareto-optimal if there does
f (X )
not exist xÎX such that
≥ ∗ for all ∈ {1, … , }
> ∗ for some ∈ {1, … , } f1

Set of all Pareto-optimal solutions: XPO
f
Definition. Objective vector yÎf(X) is Pareto- x2
optimal, if there exists a Pareto-optimal x*ÎX s.t. X PO
f(x*)=y X
- Set of Pareto-optimal objective vectors: f(XPO)
- Notation = v−max ( )
∈
x1
28.3.2018
8
x3
Example: Markowitz model (0,0,1)
q Optimal asset portfolio selection

– How to allocate funds to m assets based on X
o Expected returns ̅ , i=1,...,m x1
(1,0,0)
o Covariances of returns σij, i,j=1,...,m
q Set of feasible solutions (0,1,0)
x2
– Decision variables x1,...,xm
o Allocate xj *100% of funds to j-th asset
f2
– Portfolio ∈ = { ∈ ℝ |∑ = 1}
q Objective functions
return
f (X )
1. Maximize expected return of portfolio =∑ ̅
2. Minimize variance (risk) of portfolio =
∑ ∑
risk f1
28.3.2018
9
Pareto-optimality in Markowitz model
x3
(0,0,1)
q Portfolio x is Pareto-optimal, if no other X PO
portfolio yields greater or equal expected
return with less risk
x1
q One possibility for computation: (1,0,0)
- Choose d = max number of solutions computed
(0,1,0)
- Solve μ1 = max f2, μd = min f2 x2 f ( X PO )
- For all k=2,…,d-1 set μk s.t. μk-1> μk> μd and solve f2
(1-dimensional) quadratic programming problem μ1
μ2
min ∑ ∑ such that ∑ ̅ = μ3
∈
return
- Termination condition (for problems, where f(XPO) μ4 f (X )
is a connected set): stop, if solution is not PO
- Not attractive when n>2
28.3.2018
risk f1 10
Algorithms for solving Pareto-optimal
solutions (1/2)
q Exact algorithms
- Guaranteed to find all PO-solutions XPO
- Only for certain problem types, e.g., Multi-Objective Mixed Integer Linear
Programming (MOMILP)
q Use of single-objective optimization algorithms

- Sequentially solve ordinary (i.e. 1-dimensional) optimization problems to obtain a
subset of all PO-solutions, XPOS
- Performance guarantee: XPOSÍXPO
o Solutions may not be “evenly” distributed
- Methods:
o Weighted sum approach, weighted max-norm approach, ε-constraint approach
28.3.2018
11
Algorithms for solving Pareto-optimal
solutions (2/2)
q Approximation algorithms
- Obtain an approximation XPOA of XPO in polynomial time
- Performance guarantee: For every xÎXPO exists y ÎXPOA such that ||f(x)-f(y)||< ε
- Only for very few problem types, e.g., MO knapsack problems
q Metaheuristics
- No performance guarantees
- Can handle problems with
• A large number of variables and constraints
• Non-linear or non-continuous objective functions/constraints
- Evolutionary algorithms (e.g., SPEA, NSGA)
- Stochastic search algorithms (simulated annealing)
28.3.2018
12
Example: Multiobjective integer linear
programming (MOILP)
q Ben is at an amusement park that offers 2 different rides:
q Tickets to ride 1 cost 2 €. Each ticket lets you take the ride twice
q Tickets to ride 2 are for one ride and cost 3 €
q Ben has 20 euros to spend on tickets to ride 1 (x1Îℕ) and ride 2 (x2 Î
ℕ) → constraint 2 1 + 3 2 ≤ 20
q Each time Ben takes ride 2, his grandfather cheers for him
q Ben maximizes the number of (i) rides taken and (ii) cheers →
objective functions = 1, 2 = (2 1 + 2, 2)
28.3.2018
13
Feasible solutions ”Ben has 20 euros. He is
choosing the number of tickets to
ride 1 ( 1 ∈ ℕ) and ride 2 ( 2 ∈ ℕ)
‒ Constraint 2 1 + 3 2 ≤ 20”
2 1 +3 2 = 20
28.3.2018
14
Example: MOILP (cont’d)
q Blue points are feasible solutions; the 7 PO solutions are circled
f (X )
28.3.2018
15
Weighted sum approach
q Algorithm
f2
l 1 = (0.2, 0.8)T
1. Generate ~ ({ ∈ 0,1 | ∑ = 1})
l 2 = (0.7, 0.3)T
2. Solve max ∑ ( )
∈
3. Solution is Pareto-optimal f1
Repeat 1-3 until enough PO-solutions have been found
+ Easy to implement
– Cannot find all PO solutions if the problem is non-convex (if PO
solutions are not in the border of the convex hull of f(X))
28.3.2018
16
max [2 1 1 +( 1 + 2) 2]
, ∈ℕ
max [2 1 1 +( 1 + 2) 2]
, ∈ℕ
( 1, 2) =
=
1 , 6 is
0.2, 0.8 ⇒
Pareto-
maximize
optimal
0.4 1 + 2
= 1/3, 2/3
= 0.5, 0.5
( 1, 2) =
0.4 1 + 2 constant ( 1, 2) = 10 , 0 is
4 , 4 is Pareto-
Pareto- optimal
optimal ( 1, 2) =
7 , 2 is
Pareto-
optimal
( ) and Pareto-optimal solutions
Only 4 of the 7 Pareto-optimal

solutions can be found with
the weighted sum approach
28.3.2018
19
Weighted max-norm approach
q Idea: define a utopian vector of objective function values
f2
and find a solution for which the distance from this utopian
vector is minimized
f*
q Utopian vector: ∗ = ∗ , … , ∗ , ∗ > ∀ ∈ , = 1, … ,
q Distance is measured with weighted max-norm max ,
,…,
where is the between ∗ and , and > 0 is the
weight of objective i such that ∑ = 1.
q The solutions that minimize the distance of ( ) from ∗ are
found by solving:
f1
l
∗ ∗ Contours of f * - f ( x )
min − ( ) = min max − max
∈ ∈ ,…,
= min ∆ . . ∗
− ≤ ∆ ∀ = 1, … , when l = (0.9, 0.1) T
∈ ,∆∈ℝ
28.3.2018
20
Weighted max-norm approach (2/2)
q Algorithm l 1 = (0.5, 0.5)T
f*
1. Generate ~ ({ ∈ 0,1 | ∑ = 1})
2. Solve min ∗ − f(x) f2
∈
λ2=(0.9,0.1)T
3. At least one of the solutions of Step 2 is PO
Repeat 1-3 until enough PO solutions have been
found
f1
+ Easy to implement
+ Can find all PO-solutions
– n additional constraints, one additional variable
28.3.2018
21
Example: MOILP (cont’d)
q Find a utopian vector f* λ1=0.1, λ2=0.9:
- max f1= 2x1+x2 s.t. 2x1+3x2 ≤ 20, x1,x2 ≥ 0
o x=(10,0); f1=20
- max f2 = x2 s.t. 2x1+3x2 ≤ 20, x1,x2 ≥ 0 min ∆ s.t.
∆∈ℝ
o x=(0, 20/3); f2=20/3
2.1 − 0.2 − 0.1 ≤∆
- Let f*=(21,7)
6.3 − 0.9 ≤ ∆
q Minimize the distance from the 2 + 3 ≤ 20
utopian vector: , ∈ℕ
min ∆ s.t.
∆∈ℝ
21 − 2 + ≤∆
7− ≤∆ Solution: Δ=1.3, x=(1,6) Þ
2 + 3 ≤ 20, , ∈ ℕ x=(1,6), f=(8,6) is PO
28.3.2018
22
Example: MOILP revisited
1.λ1=0.1; solution: {Δ=1.3, x=(1,6)} Þ f*
x=(1,6), f=(8,6) is PO 0.1
2.λ1=0.2; 3 solutions x=(2,5), x=(3,4),
x=(4,4). Only x=(2,5), f=(9,5) and x=(4,4),
0.9
f=(12,4) are PO
3.λ1=0.35; x=(5,3); f=(13,3) is PO
4.λ1=0.4; 2 solutions x=(6,2) and x=(7,2);
x=(7,2), f=(16,2) is PO
5.λ1=0.55; x=(8,1); f=(17,1) is PO
6.λ1=0.70; 2 solutions x=(9,0) and x=(10,0);
x=(10,0), f=(20,0) is PO
28.3.2018
23
Value function methods (1/2)
q Use value function : ℝ → ℝ to f2
transform the MOO problem into a
single-objective problem f (X )
– E.g., the additive value function

=∑ ( ( )) f1
V
q Theorem: Feasible solution x* with
the highest value V(x*) is Pareto-
optimal V
V ( f ( X ))
28.3.2018
24
Value function methods (2/2)
q Consider the additive value function =∑ ( ( ))
with incomplete weight information ∈ ⊆
q Set of Pareto-optimal solutions = set of non-dominated

solutions with no weight information ( )
q Preference statements on weights decrease the set of feasible

weights to ⊆ → focus on preferred PO-solutions ⊆
=
28.3.2018
25
Example: MOILP revisited
q Choose vi(fi(x))=fi(x)/Ci*, constants C1*=20, C2*=6
2 +
, = ( ( )) = + 1− = + (1 − )( /6)
20
1
0.9
0.8
0.7
0.6
0.5
V
0.4
0.3
0.2
0.1
28.3.2018
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 261
w1
Example: Bridge repair program (1/7)
q Total of 313 bridges calling for repair
q Which bridges should be included in the repair program under the next three
years?
q Budget of 9,000,000€
q Program can contain maximum of 90 bridges

- Proxy for limited availability of equipment and personnel etc.
q Program must repair the total sum of damages by at least 15,000 units
P. Mild, J. Liesiö and A. Salo (2015): Selecting Infrastructure Maintenance

28.3.2018
Projects with Robust Portfolio Modeling, Decision Support Systems 27
q Set of feasible solutions X defined by linear constraints and binary
decision variables:
∑ − 9000000
= ∈ 0,1 ≤0 , = ∑ − 90
15000 − ∑
- xj = a decision variable: xj =1 repair bridge j

- x=[x1,...,x313] is a repair program
- cj = repair cost of bridge j
- dj = sum of damages of bridge j
28.3.2018
28
q Six objective indexes measuring urgency for repair
1. Sum of Damages (“SumDam”)
2. Repair Index (“RepInd”)
3. Functional Deficiencies (“FunDef”)
4. Average Daily Traffic (“ADTraf”)
5. Road Salt usage (“RSalt”)
6. Outward Appearance (“OutwApp”)
q All objectives additive over bridges: =∑ ,

where is the score of bridge j with regard to objective i:
28.3.2018
29
q A multi-objective zero-one linear programming (MOZOLP) problem
v− max( ,…, )
∈
q Pareto-optimal repair programs XPO generated using the weighted

max-norm approach
min ∆
∈ ,∆∈ℝ
∗
∆≥ −∑ ∀ = 1, … , 6
28.3.2018
30
q Additive value function applied for modeling preferences between the
objectives: , =∑ =∑ ∑
q Incomplete ordinal information about objective weights: {SumDam,RepInd}
≥{FunDef, ADTraf} ≥ {RSalt,OutwApp}
= ∈ ≥ ≥ , ∀ = 1,2; = 3,4; = 5,6
q Non-dominated repair programs
, ≥ , for all ∈
= ∈ |∄ ∈ s.t.
, > , for some ∈
= ( )⊇ ( )
28.3.2018
31
q Ca. 10,000 non-dominated bridge repair programs C
q Bridge-specific decision recommendations can be
I
obtained through a concept of core index:
J
|{ ∈ ( )| = 1}|
= D
| ( )| E
q Of the 313 bridges: H
F
– 39 were included in all non-dominated repair programs A G
(CI=1)
– 112 were included in some but not all non-dominated B
programs (0<CI<1)
– 162 were included in none of the non-dominated programs
(CI=0)
28.3.2018
32
BRIDEGES' SCORES
q Bridges listed in Bridge number and name

2109 Lavusjoen silta
Core Index
1.00
DamSum
5.00
RepInd
1.65
FunDef
4
ADTraf
2.6
Rsalt
1
OutwApp
2.6
Cost
50000
2218 Joroisvirran silta 1.00 5.00 5.00 2 5 5 2.6 180000
decreasing order of core 2217 Rautatieylikulkusilta
763 Hurukselantien risteyssilta
1.00
1.00
3.49
2.27
5.00
2.33
1.5
1
5
3.4
5
5
1.8
1
130000
280000
indices 80 Suolammenojan silta

257 Villikkalan silta
1743 Huuman silta II
1.00
0.81
0.76
1.36
1.97
1.64
1.53
1.96
1.53
2
5
1
4.2
1
5
5
1
5
1.8
1.8
1.8
10000
20000
140000
- Tentative but not binding 730 Mälkiän itäinen risteyssilta

2804 Raikuun kanavan silta
0.63
0.60
1.33
3.93
1.58
1.12
1.5
2.5
5
1
5
1
1
1
120000
20000
priority list
856 Ojaraitin alikulkukäytävä I 0.54 1.46 1.46 1 5 5 1 20000
2703 Grahnin alikulkukäytävä 0.43 1.70 1.23 1 5 5 1 60000
817 Petäjäsuon risteyssilta 0.39 1.52 1.37 1 5 5 1 50000
- Costs and other 725 Mustolan silta

2189 Reitunjoen silta
0.29
0.24
1.98
1.90
1.93
1.63
2
3
1.8
1.8
1
1
4.2
1.8
190000
10000
characteristics displayed
2606 Haukivuoren pohjoinen ylikulkusilta 0.15 1.84 2.09 1.5 2.6 1 1 70000
125 Telataipaleen silta 0.14 1.38 1.12 1 5 5 1.8 40000
608 Jalkosalmen silta 0.03 1.54 1.50 3 1.8 1 2.6 10000
556 Luotolan silta 0.00 1.74 1.26 3 1 1 1.8 10000
661 Raikan silta 0.00 1.95 1.58 2 1 1 1.8 10000
2613 Pitkänpohjanlahden silta 0.00 1.27 1.16 1 4.2 5 2.6 20000
q The list was found useful

738 Hyypiälän ylikulkusilta 0.00 1.72 1.79 1 3.4 1 1.8 90000
2549 Uitonsalmen silta 0.00 1.71 1.37 3 1 1 1 30000
703 Tokkolan silta 0.00 1.82 1.70 2 1.8 1 1 10000
by the program managers 870 Tiviän alikulkukäytävä

377 Sudensalmen silta
0.00
0.00
1.10
1.88
1.07
1.66
1
1
5
2.6
5
1
1
1.8
20000
20000
953 Sydänkylän silta 0.00 1.23 1.33 3.5 1 1 1.8 10000
700 Kirjavalan ylikulkusilta 0.00 1.42 1.98 1.5 1 1 1 60000
2142 Latikkojoen silta 0.00 1.43 1.58 2.5 2.6 1 1.8 20000
464 Jokisilta 0.00 1.19 1.25 3.5 1.8 1 1 20000
1025 Hartunsalmen silta 0.00 1.18 1.09 3.5 1.8 1 2.6 20000
95 Touksuon silta 0.00 1.83 1.18 2 2.6 1 2.6 20000
418 Laukassalmen silta 0.00 1.54 1.35 1.5 2.6 1 28.3.2018
1.8 10000
420 Sillanmäenojan silta 0.00 1.20 1.07 1.5 2.6 1 1.8 3310000
Summary
q MOO differs from MAVT in that
– Alternatives are not explicit but defined implicitly through constraints
– MOO problems are computationally much harder
q MOO problems are solved by
– Computing the set of all Pareto-optimal solutions – or at least a subset or
an approximation
– Introducing preference information about trade-offs between objectives to
support the selection of one of the PO-solutions
28.3.2018
34
Decision making and
problem solving –
Lecture 10
• Analytic Hierarchy Process
• Outranking methods

Motivation
q When alternatives are evaluated w.r.t. multiple attributes / criteria,
decision-making can be supported by methods of
– Multiattribute value theory (certain attribute-specific performances)
– Multiattribute utility theory (uncertain attribute-specific performances)
q MAVT and MAUT have a strong axiomatic basis
q Yet, other popular multicriteria methods exist
28.3.2018
2
Analytic Hierarchy Process (AHP)
q Thomas L. Saaty (1977, 1980)
q Enormously popular
– Thousands of reported applications
– Dedicated conferences and scientific journals
q Several decision support tools

– Expert Choice, WebHipre etc.
q Not based on the axiomatization of preferences – therefore remains

controversial
28.3.2018
3
Problem structuring in AHP
q Objectives, sub- Goal
Satisfaction with School
objectives / criteria,
and alternatives
are represented as
a hierarchy of
Learning Friends School Vocational College Music
elements (cf. value Life Training Preparation Classes
tree)
School School School

A B C
28.3.2018
4
Local priorities
Scale
Verbal statement 1-to-9 Balanced
q For each objective / sub-objective, a local priority
vector is determined to reflect the relative Equally important 1 1.00
- 2 1.22
importance of those elements placed immediately Slightly more important 3 1.50
below the objective / sub-objective - 4 1.86
Strongly more important 5 2.33
q Pairwise comparisons: - 6 3.00
– For (sub-)objectives: ”Which sub-objective / criterion Very strongly more important 7 4.00
is more important for the attainment of the objective? - 8 5.67
Extremely more important 9 9.00
How much more important is it?”
– For alternatives: ”Which alternative contributes more
to the attainment of the criterion? How much more
does it contribute?” 1-to-9
q Responses on a verbal scale corresponding to

Balanced
weight ratios
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
w =1-w
1 2
28.3.2018
5
Pairwise comparison matrix
L F SL VT CP MC
q Weight ratios = form a pairwise Learning 1 4 3 1 3 4
comparison matrix A: Friends 1/4 1 7 3 1/5 1
⋯ School life 1/3 1/7 1 1/5 1/5 1/6
= ⋮ ⋱ ⋮ Voc. training 1 1/3 5 1 1 1/3
= 1/ ⋯ College prep. 1/3 5 5 1 1 3

Music classes 1/4 1 6 3 1/3 1
Learning Friends School life
A B C A B C A B C
A 1 1/3 ½ A 1 1 1 A 1 5 1
B 3 1 3 B 1 1 1 B 1/5 1 1/5
C 2 1/3 1 C 1 1 1 C 1 5 1
Voc. training College prep. Music classes

A B C A B C A B C
A 1 9 7 A 1 1/2 1 A 1 6 4
B 1/9 1 5 B 2 1 2 B 1/6 1 1/3 28.3.2018
C 1/7 1/5 1 C 1 1/2 1 C 1/4 3 1 6
Incosistency in pairwise comparison
matrices
q Problem: pairwise comparisons are not necessarily consistent
q E.g., if learning is slightly more importannt (3) than college

preparation, which is strongly more important (5) than school life, then
learning should be 3 × 5 times more important than school life –
impossible
→ Weights need to be estimated
28.3.2018
7
Local priority vector
q The local priority vector w (=estimated
weights) is obtained by normalizing the
Learning W
eigenvector corresponding to the largest
A B C
eigenvalue of A:
= , A 1 1/3 1/2 0.16
1 B 3 1 3 0.59
:= .
∑ C 2 1/3 1 0.25
q Matlab:
– [v,lambda]=eig(A) returns the eigenvectors
and eigenvalues of A Only one eigenvector with all real
elements: (0.237, 0.896, 0.376) ®
normalized eigenvector w=(0.16,
0.59, 0.25).
28.3.2018
8
Local priority vector
Learning W Friends W
A B C A B C
A 1 1/3 1/2 0.16 A 1 1 1 0.33
B 3 1 3 0.59 B 1 1 1 0.33
C 2 1/3 1 0.25 C 1 1 1 0.33
L F SL VT CP MC W
School life W Voc. training W Learning 1 4 3 1 3 4 0.32
A B C A B C Friends 1/4 1 7 3 1/5 1 0.14
A 1 5 1 0.45 A 1 9 7 0.77 Schoo life 1/3 1/7 1 1/5 1/5 1/6 0.03
B 1/5 1 1/5 0.09 B 1/9 1 5 0.05 Voc. Training 1 1/3 5 1 1 1/3 0.13
C 1 5 1 0.46 C 1/7 1/5 1 0.17 College prep. 1/3 5 5 1 1 3 0.24
Music classes 1/4 1 6 3 1/3 1 0.14
College prep. W Music classes W
A B C A B C
A 1 1/2 1 0.25 A 1 6 4 0.69
B 2 1 2 0.50 B 1/6 1 1/3 0.09
C 1 1/2 1 0.25 C 1/4 3 1 0.22
28.3.2018
9
Consistency checks
Three alternatives, n=3:
q The consistency of the pairwise
comparison matrix A is studied by q Learning: = 3.05, = 0.04
comparing the consistency index (CI) q Friends: = 3.00, =0
of A to the average consistency index q School life: = 3.00, =0
RI of a random pairwise comparison q Voc. training = 3.40, = 0.34
matrix: q College prep: = 3.00, =0
−
= , = q Music classes: = 3.05, = 0.04
−1
n 3 4 5 6 7 8 9 10 Six attributes, n=6:
RI 0.58 0.90 1.12 1.24 1.32 1.41 1.45 1.49
q All attributes: = 7.42, = 0.23
q Rule of thumb: if CR>0.10,
comparisons are so inconsistent that
they should probably be revised
28.3.2018
10
Total priorities
q The total (final, aggregate, overall) priorities are obtained
recursively:
= ,
where
– is the total priority of criterion i,
– is the local priority of criterion / alternative k with regard to criterion i,
– The sum is computed over all criteria i below which criterion / alternative k
is positioned in the hierarchy
28.3.2018
11
Total priorities
Learning w Friends w L F SL VT CP MC w
A B C A B C Learning 1 4 3 1 3 4 0.32
A 1 1/3 1/2 0.16 A 1 1 1 0.33 Friends 1/4 1 7 3 1/5 1 0.14
B 3 1 3 0.59 B 1 1 1 0.33 Schoo life 1/3 1/7 1 1/5 1/5 1/6 0.03
C 2 1/3 1 0.25 C 1 1 1 0.33 Voc. Training 1 1/3 5 1 1 1/3 0.13
School life w College prep. 1/3 5 5 1 1 3 0.24

Voc. training w
A B C Music classes 1/4 1 6 3 1/3 1 0.14
A B C
A 1 5 1 0.45 A 1 9 7 0.77
B 1/5 1 1/5 0.09 B 1/9 1 5 0.05
0.32 0.14 0.03 0.13 0.24 0.14
C 1 5 1 0.46 C 1/7 1/5 1 0.17
L F SL VT CP MC Total w
College prep. w Music classes w A 0.16 0.33 0.45 0.77 0.25 0.69 0.37
A B C A B C B 0.59 0.33 0.09 0.05 0.50 0.09 0.38
A 1 1/2 1 0.25 A 1 6 4 0.69 C 0.25 0.33 0.46 0.17 0.25 0.22 0.25
B 2 1 2 0.50 B 1/6 1 1/3 0.09
C 1 1/2 1 0.25 C 1/4 3 1 0.22 E.g.,
wB=0.32*0.59+0.14*0.33+0.03*0.09+
28.3.2018
0.13*0.05+0.24*0.50+0.14*0.09 12
Problems with AHP
q Rank reversals: the introduction of an additional
alternative may change the relative ranking of the other
alternatives
q Example: C1 C2
– Alternatives A and B are compared w.r.t. two ”equally important”
criteria C1 and C2 (wC1 = wC2 = 0.5) A 1 5
– A is better than B: B 4 1
1 1 1 5 1 4 1 1
= × + × ≈ 0.517, = × + × ≈ 0.483
2 5 2 6 2 5 2 6 C 1 5
– Add C which is identical to A:
1 1 1 5 1 4 1 1
= = × + × ≈ 0.311, = × + × ≈ 0.379
2 6 2 11 2 6 2 11
– Now B is better than A!
28.3.2018
13
Methods based on outranking relations
q Basic question: is there enough preference information / evidence
to state that an alternative is at least as good as some other
alternative?
q I.e., does an alternative outrank some other alternative?
28.3.2018
14
Indifference and preference thresholds
q If the difference between the criterion-specific A preferred
A and B Uncertain
performances of A and B is below a pre- equally preference to B
defined indifference threshold, then A and preferred
B are ”equally good” w.r.t. this criterion
q If the difference between the criterion-specific

performances of A and B is above a pre-
defined preference threshold, then A is
preferred to B w.r.t this criterion 0 200 400 Difference
in cost
between A
q Between indifference and preference and B
thresholds, the DM is uncertain about Indifference
Preference
preference threshold
threshold
28.3.2018
15
PROMETHEE I & II
q In PROMETHEE methods, the degree A and B Uncertain A preferred
to which alternative k is preferred to l equally preference to B
is 1 preferred
( , ) ≥ 0,
where
– is the weight of criterion i
0
– ( , ) =1, if k is preferred to l w.r.t. criterion i, 0 200 400 Difference
– ( , ) =0, if the DM is indifferent between k in cost
and l w.r.t. criterion i, or l is preferred to k between A
– ( , ) ∈ (0,1), if preference between k and l Indifference and B
w.r.t. criterion i is uncertain threshold
Preference
threshold
28.3.2018
16
PROMETHEE I & II
There is more
evidence in
q PROMETHEE I: k is preferred to k’, if favor of k than k’
( , )> ( ′, )
(, )< ( , ′)
There is less
evidence
against k than k’
q PROMETHEE II: k is preferred to k’, if
= [ ( , )− ( , )] > [ ( ′, ) − ( , ′)] = ′
The ”net evidence” for

k is larger than for k’
28.3.2018
17
PROMETHEE: Example
q 3 alternatives, 2 criteria = revenue and market share
Revenue Market share Revenue F1 M. share F2 Weighted

Fw=w1F1+w2F2
x1 1M€ 10%
x1, x2 1 0 1
x2 0.5M€ 20%
x2, x1 0 0 0
x3 0 30%
x1, x3 1 0 1
Indiff. threshold 0 10%
x3, x1 0 1 1
Pref. threshold 0.5M€ 20% x2, x3 1 0 1
Weight 1 1 x3, x2 0 0 0
28.3.2018
18
PROMETHEE I: Example
q PROMETHEE I:
– is preferred to , if
, + , > , + ,
F1 F2 Fw
x1, x2 1 0 1
, + , < , + ,
x2, x1 0 0 0
x1, x3 1 0 1 – is not preferred to due to the latter condition
x3, x1 0 1 1 – is not preferred to due to both conditions
x2, x3 1 0 1 – is preferred to
x3, x2 0 0 0 – is not preferred to and vice versa
q Note: preferences are not transitive
– ≻ ~ ⇏ ≻
28.3.2018
19
PROMETHEE I: Example (Cont’d)
q PROMETHEE I is also prone to rank
reversals: F1 F2 Fw
– Remove
x1, x2 1 0 1
– Then,
x2, x1 0 0 0
, ≯ , x1, x3 1 0 1
x3, x1 0 1 1
x2, x3 1 0 1
, ≮ ,
x3, x2 0 0 0
→ is no longer preferred to
28.3.2018
20
PROMETHEE II: Example
q The ”net flow” of alternative
= [ ( , )− ( , )] F1 F2 Fw
x1, x2 1 0 1
– = 1−0 + 1−1 =1 x2, x1 0 0 0
– = 0−1 + 1−0 =0 x1, x3 1 0 1
– = 1 − 1 + 0 − 1 = −1 x3, x1 0 1 1
x2, x3 1 0 1
x3, x2 0 0 0
→ ≻ ≻
28.3.2018
21
PROMETHEE II: Example (Cont’d)
q PROMETHEE II is also prone to rank reversals
– Add two altrenatives that are equal to x3 in both criteria. F1 F2 Fw
Then, x2 becomes the most preferred:
x1, x2 1 0 1
= 1−0 +3× 1−1 = 1
= 0−1 +3× 1−0 =2 x2, x1 0 0 0
:
= 1 − 1 + 0 − 1 = −1 x1, x3 1 0 1
– Add two alternatives that are equal to x1 in both criteria.
x3, x1 0 1 1
Then, x2 becomes the least preferred:
, , x2, x3
= 1 − 0 + 1 − 1 + 2 × (0 − 0) = 1 1 0 1
= 3 × 0 − 1 + 1 − 0 = −2 x3, x2 0 0 0
= 3 × 1 − 1 + 0 − 1 = −1
– Remove x2. Then, x1 and x3 are equally preferred.
= = 1−1 =0
28.3.2018
22
Summary
q AHP and outranking methods are commonly used for supporting
multiattribute decision-making
q Unlike MAVT (and MAUT), these methods do not build on the
axiomatization of preferences →
– Rank reversals
– Preferences are not necessarily transitive
q Elicitation of model parameters can be difficult
– Weights have no clear interpretation
– In outranking methods, statement ”I prefer 2€ to 1€” and ”I prefer 3€ to 1€” are both
modeled with the same number (1); to make a difference, indifference and
preference thresholds need to be carefully selected
28.3.2018
23
Decision making and
problem solving –
Lecture 11
• Group techniques
• Voting
• MAVT for group decisions

Motivation
q Thus far we have assumed that
– Objectives, attributes/criteria, and decision alternatives are given
– There is a single decision maker
q This time we’ll learn

– How a group of experts / DMs can be used to generate objectives,
attributes, and/or decision alternatives
– How to aggregate the views and preferences of the group members into a
single decision recommendation
28.3.2018
2
Idea generation and evaluation
techniques
q Goals:
– Generate topics / ideas / decision alternatives
– Evaluate these topics / ideas / alternatives
– Agree on a prioritization of the topics / ideas / alternatives
q Methods:
– Brainstorming
– Nominal group technique
– Delphi method
– …and variants of the above
28.3.2018
3
Brainstorming
q Goal: to generate a large number of possible solutions for a problem
q Participants: Facilitator, recorder, and max 8-12 panel members

– Step 1 Prior notification: time for individual idea generation
– Step 2 Session for idea generation: all ideas are listed, spotaneous ideas are encouraged, no
criticism is allowed
– Step 3 Review and evaluation: a list of ideas is sent to the panel members for further study
q Principles:
– Focus on quantity
– Withhold criticism
– Welcome unusual ideas
– Combine and improve ideas
28.3.2018
4
Brainstorming
+ A large number of ideas can be generated in a short period of time
+ Simple – no expertise or knowledge required from the facilitator
- Blocking: during the process, participants may forget their ideas or

not share them because they no longer find them relevant
- Collaborative fixation: Exchanging ideas in a group may decrease
the novelty and variety of ideas
28.3.2018
5
Nominal group technique
q Goal: to generate a large number of possible solutions for a problem and
decide on a solution
q Participants: Faciliator, recorder, and max 6-12 panel members

– Step 1: Silent generation of ideas – group work not allowed
– Step 2: Round-robin sharing of ideas. Facilitator lists all ideas on a flip chart, no comments at this
point.
– Step 3: Group discussion to facilitate common understanding of the presented ideas. No ideas are
eliminated, judgment and criticism are avoided.
– Step 4: Ranking of the ideas (by, e.g., voting)
28.3.2018
6
Nominal group technique
+ A large number of ideas can be generated in a short period of time
+ Silent generation of ideas decreases blocking
+ Round-robin process ensures equal participation
- Not suitable for settings where consensus is required

- Can be time-consuming
28.3.2018
7
Delphi technique
q Goal: To obtain quantitative estimates about some future events (e.g.,
estimated probabilities, impacts, and time spans of negative trends for
Finland)
q Participants: Faciliator and a panel of experts
q Principles:
– Anonymous participation
– Structured gathering of information through questionnaires: numerical estimates and arguments
to support these estimates
– Iterative process: participants comment on each other’s estimates and are encouraged to revise
their own estimates in light of such comments
– Role of the facilitator: sends out the questionnaires, organizes the information, identifies
common and conflicting viewpoints, works toward synthesis
28.3.2018
8
Example: Decision analysis based real
world conflict analysis tools
q Workshop organized by the Finnish Operations Research Society
(FORS) Monday 5.10.2015
q Goal: to practice DA-based conflict analysis tools that Crisis

Management Initiative (CMI) uses regularly in its operations:
– Trend identification,
– Data collection,
– Visualization,
– Root-cause analysis.
28.3.2018
9
Example cont’d
q Prior to the workshop,
each participant was
asked to
– List 3-5 negative trends for
Finland (title and brief
description)
– Provide time-spans for the
impacts of these trends
(<10 years, 10-20 years,
>20 years)
28.3.2018
10
Example cont’d
q Trends listed by the
participants were organized
by the workshop facilitators
– Similar trends combined
– Marginal trends eliminated
q A final list of 21 trends was

emailed to the participants
prior to the workshop
28.3.2018
. 11
.
.
Example cont’d
q At the workshop, each
participant was asked to
evaluate
– The probability of each
trend being realized (scale
0-5) .
.
– The impact that the trends

.
would have upon

realization (scale 0-5)
28.3.2018
12
Example cont’d
q The participants were
also asked to assess .
.
cross-impacts among .
trends
– Which other trends does
this trend enhance? .
.
.
.
.
.
28.3.2018
13
Example cont’d Increased political
tension in EU
q Visualizations on Russia’s Climate

the probability actions change
The retirement
and impact Brain
bomb
drain
assessments
were shown to
the participants
to facilitate
discussion
Eating and
drinking habits
28.3.2018
14
Example cont’d Socially excluded
youth
The welfare
trap
Specialization, digitalization, and
automation driving inequality
q Cross-impacts High
were visualized, unemployment
Economic
stagnation
too Fossile fuels
Climate
change
The retirement Cuts on

bomb education
Refugees and immigration
Increasing
Bifurgation of Finns and government debt
political radicalization
Russia’s
actions
Increased political
tension in EU
28.3.2018
15
Example cont’d
q Goal of such analysis:
– To create a shared understanding of the problem
– To identify possible points of disagreement
q Next steps:
– Possible revision of estimates in light of the discussion
– The determination of policy actions to help mitigate / adapt to the most important negative
trends
– Agreement on which policy actions to pursue
– The implementation of these policy actions
28.3.2018
16
Aggregation of preferences
q Consider N alternatives x1,…, xN
q Consider K decision makers DM1,… DMK with different

preferences about the alternatives
q How to aggregate the DMs’ preferences into a group choice?

– Voting
– MAVT
28.3.2018
17
Plurality voting
q Each voter casts one vote to his/her most preferred candidate
q The candidate with the most votes wins
q Plurality voting with runoff:
- The winner must get over 50% of the votes
- If this condition is not met, alternatives with the least votes are eliminated
- Voting is continued until the condition is met
- E.g., Finnish presidential election: in the second round only two candidates
remain
28.3.2018
18
Plurality voting
q Suppose, there are three alternatives A, B, C, and 9 voters
• 4 think that A > B > C
• 3 think that B > C > A
• 2 think that C > B > A
Plurality voting Run-off
4 votes for A C eliminated

3 votes for B
2 votes for C 4 votes for A
3+2 = 5 votes for B
A is the winner B is the winner
28.3.2018
19
Condorcet
q All voters rank-order the alternatives
q Each pair of alternatives is compared - the one with more votes is
the winner
q If an alternative wins all its one-to-one comparisons, it is the
Condorcet winner
q There might not be a Condorcet winner – some other rule must be
applied, e.g.,
– Copeland’s method: the winner is the alternative with the most wins in one-to-one
comparisons
– Eliminate the alternative(s) with the least votes and recompute
28.3.2018
20
Condorcet - example
q 33 voters and alternatives A, B, C
• 17 voters: A>B>C
• 1 voter: A>C>B
• 15 voters: B>C>A
• 0 voters: C>B>A, C>A>B, B>A>C
q A is the Condorcet winner, because it wins both one-on-one

comparisons
• 18 favor A over B and 18 favor A over C
28.3.2018
21
Condorcet completion
q There might not be a Condorcet winner
– Copeland’s completion method: the winner is the alternative with the most wins in
one-to-one comparisons
5 voters and alternatives A, B, C, D, E A B C D E wins

- 1 voter: A>B>C>D>E A 2 2 3 3 2
- 1 voter: A>D>E>C>B
- 2 voters: D>E>B>C>A B 3 3 2 2 2
1 voter: C>B>A>D>E
C 3 2 2 2 1
-
D wins more one-on-one comparisons D 2 3 3 5 3

than other alternatives E 2 3 3 0 2
28.3.2018
22
Condorcet completion
qAnother possibility for Condorcet completion: Eliminate the one
with least wins and recompute results
qFirst C is eliminated
A B C D E wins
• B,D,E lose one win A 2 2 3 3 2
B 3 3 2 2 2 1
qB and E with one win are elimitated C 3 2 2 2 1
• A and D remain
D 2 3 3 5 3 2
qA wins D by 3 votes to 2 E 2 3 3 0 2 1
28.3.2018
23
Borda
q Each voter gives
– n-1 points to the most preferred alternative,
– n-2 points to the second most preferred,
– …
– 0 points to the least preferred alternative
q The alternative with the highest total number of points wins
4 state that A > B > C A : 4·2 + 3·0 + 2·0 = 8 votes

3 state that B > C > A B : 4·1 + 3·2 + 2·1 = 12 votes
2 state that C > B > A C : 4·0 + 3·1 + 2·2 = 7 votes
28.3.2018
24
Approval voting
q Each voter casts one vote for each alternative he/she approves
q The alternative with the highest number of votes is the winner
28.3.2018
25
Problems with voting: The Condorcet
paradox (1/2)
q Consider the following rank-orderings of three alternatives
DM1 DM2 DM3
A 1 3 2
B 2 1 3
C 3 2 1
q Paired comparisons:
– A is preferred to B by 2 out of 3 voters
– B is preferred to C by 2 out of 3 voters
– C is preferred to A by 2 out of 3 voters
28.3.2018
26
Problems with voting: The Condorcet
paradox (2/2)
DM1 DM2 DM3
q Three voting orders:
A 1 3 2
1. (A-B) → A wins, (A-C) → C is the winner
B 2 1 3
2. (B-C) → B wins, (B-A) → A is the winner
C 3 2 1
3. (A-C) → C wins, (C-B) → B is the winner
The outcome depends on the order in which votes are cast!
q No matter what the outcome is, the majority of voters would prefer some
other alternative:
– If C wins, 2 out of 3 voters would change it to B
– …But B would be changed to A by 2 out of 3 voters
– …And then A would be changed to C by 2 out of 3 voters…
28.3.2018
27
Problems with voting: tactical voting
q DM1 knows the preferences of the other
voters and the voting order (A-B, winner-C)
q If DM2 and DM3 vote accroding to their true DM1 DM2 DM3
A 1 3 2
preferences, then the favourite of DM1 (A)
B 2 1 3
cannot win C 3 2 1
q If DM1 votes B instead of A in the first round,

then B wins and DM1 avoids her least
favourite alternative (C)
28.3.2018
28
Social choice function
q Assume that the preferences of DMi are represented by a
complete and transitive weak preference order Ri:
DMi thinks that x is at least as good as y Û x Ri y
q What is the social choice function f that determines the collective

preference R=f(R1,…,RK) of a group of K decision-makers?
– Voting procedures are examples of social choice functions
28.3.2018
29
Requirements on the social choice
function
1. Universality: For any set of Ri, the social choice function should yield a
unique and complete preference ordering R for the group
2. Independence of irrelevant alternatives: The group’s preference between

two alternatives (x and y) does not change if we remove an alternative from
the analysis or add an alternative to the analysis.
3. Pareto principle: If all group members prefer x to y, the group should prefer
x to y
4. Non-dictatorship: There is no DMi such that x Ri y ⇒ x R y
28.3.2018
30
The big problem with voting: Arrow’s
theorem
There is no complete and transitive social

choice function f such that conditions 1-4
would always be satisfied.
28.3.2018
31
Arrow’s theorem – an example
q Borda criterion: DM1 DM2 DM3 DM4 DM5 Total
x1 3 3 1 2 1 10
x2 2 2 3 1 3 11 Alternative x2
x3 1 1 2 0 0 4 is the winner!
x4 0 0 0 3 2 5
q Suppose that the DMs’ preferences do not change. A ballot between

alternatives 1 and 2 gives
DM1 DM2 DM3 DM4 DM5 Total
x1 1 1 0 1 0 3
Alternative x1
x2 0 0 1 0 1 2
is the winner!
q Independence of irrelevant alternatives is not satisfied!
28.3.2018
32
Aggregation of values
Theorem (Harsanyi 1955, Keeney 1975):
Let vk(·) be a cardinal value function describing the preferences of DMk. There
exists a K-dimensional differentiable (ordinal) function VG() with positive partial
derivatives describing group preferences ≻ in the definition space such that
a ≻ b Û VG[v1(a),…,vK(a)] ³ VG[v1(b),…,vK(b)]
and conditions 1-4 are satisfied.
Note: Voting procedures use only ordinal information (i.e., rank ordering) about
the DMs’ preferences – strength of preference should be considered, too
28.3.2018
33
MAVT in group decision support
q From MAVT, we already know how to
combine cardinal value functions into an VG(x)
overall value function:
( ) ( )
( )=∑ ( ), ≥ 0, ∑ = 1.
( ) ( ) ( ) ( )
q This can be done for multiattribute DM1 DM2
cardinal value functions as well:
( )=∑ ∑ ( )
28.3.2018
34
MAVT in group decision support
q Weights , measure the value difference
between the worst and best achievement
levels x0 , x* for DM1 and DM2, respectively VG(x)
q How to compare these value differences – ( ) ( )

i.e., how to make trade-offs between people?
q “Compared to my preference for apples over oranges,
how strong is yours?” ( ) ( ) ( ) ( )
q Group weights = = 0.5 would mean DM1 DM2
that the value differences are equally
valuable, but…
q Who gets to define x0 and x*?
28.3.2018
35
MAVT for group decision support
q Example: for both DMs, vi’s are linear, DM1 has preferences
(1,0)~(0,2) and DM2 (2,0)~(0,1)
q Let x0=(0,0), x*=(2,4) for both DMs, and W1=W2 =0.5
- Then vk1N=0.5x1, vk2N=0.25x2 for both k=1,2
DM1 DM2
o (1,0)~(0,2) ⇒ V1N (1,0)= V1N (0,2)⇒ o (2,0)~(0,1) ⇒ V2N (2,0)= V2N(0,1)⇒
0.5w11=0.5w12 ⇒ w21=0.25w22⇒
w11=w12=0.5 w21=0.2, w22=0.8
o V1N(1,0)=0.25, V1N(0,1)=0.125 o V2N(1,0)=0.1, V2N(0,1)=0.2
q VG(1,0)=0.5*0.25+0.5*0.1=0.175 > VG(0,1)=0.1625
28.3.2018
36
MAVT for group decision support
q Interpretation of the result
- For DM1 (1,0)←(0,1) is an improvement. The ”group” values this more than
the value of change (0,1)←(1,0) for DM2
q Let x0=(0,0), x*=(4,2) for both DMs, and W1=W2 =0.5

- VG(1,0)=0.1625 < VG(0,1)=0.175
q Interpretation of the result

- (0,1)←(1,0) - which is an improvement for DM2 - is now more valuable for
the group than change (1,0)←(0,1)
28.3.2018
37
Summary
q Techniques for involving a group of experts or DMs can be helpful for
– Problem identification and definition,
– Generating objectives, attributes, and alternatives,
– Defining common terminology
q Individual preferences can be easily aggregated into a group

preference through voting procedures, but…
– Arrow’s impossibility theorem states that no ”good” voting procedure exists
q MAVT provides a sound method for aggregating preferences, but…

– The determination of group weights = interpersonal comparisons can be difficult
→ Aim at a joint model e.g. by exploiting incomplete preference information
28.3.2018
38

Decision Making and Problem Solving PDF

Uploaded by

Copyright:

Available Formats

Decision Making and Problem Solving PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Decision Making and Problem Solving PDF

Uploaded by

Copyright:

Available Formats

Decision making and

Liesiö, Punkka, Salo, Vilkkumaa

q A person is randomly selected

young” and ”the person is young nor

Note: If A and B are independent, the probability

Example: Wheat demand

Rye demand Low High Sum

q Questions: Sum 0.35 0.65 1

– What is the probability of high wheat demand?

q If E1,…,En are mutually exclusive and A = ⋃ , then

q Most frequent use of this law:

Law of total probability: P(A)=P(A|F)P(F)+P(A|Fc) P(Fc)=0.0051

q The probability measure P on the sample space defines a

q Probability distribution can be represented by

q PMF of a discrete random variable is fX(t)

– fX(t)=P({s ∈ S|X(s)=t}) = probability

q PDF of a continuous random variable is fX(t) 1.8

such that 1.4

q The CDF of random variable X is 0.8

– FX(t) approaches 0 (1) when t decreases

– P(a<X≤b)= FX(b)- FX(a) 0.3

• A function g(X) of random varibale X is itself a random variable, whereby

q If random variable Y=aX+b where a and b are constants, then

q NB! In general, E[g(X)]=g(E[X]) does NOT hold:

q Central limit theorem: Let X1,…,Xn be

q Exponential distribution Exp( ): describes

q Beta distribution Beta(α,β): distribution for

q It may be difficult, time-consuming or impossible to calculate such measures

q Monte Carlo simulation:

• Let ~ (CDF) = 945.4

• Transform the sample: = ( ) = 0.3254 = 773.7

VLOOKUP looks for the cell

RAND() generates a random

q Well-established computational rules provide means to derive probabilities of

q To support decision making, probabilistic models are often used to compute

q Such indices can easily be computed through Monte Carlo simulation

Liesiö, Punkka, Salo, Vilkkumaa

q Decisions and chance events are displayed in a 10

q Each chain of decisions and chance events

q The optimal strategy consists of the

→ It is optimal* to invest in high-risk

Source: Clemen, R.T. (1996): Making Hard Decisions: An Introduction to

P(forecast ‘up’ | true state = ‘up’) 28.3.2018

using Up Same Down

Up 0.8 0.15 0.20

q EVSI = 822€ - 580€ = 242€

Example: Decision tree (1/12)

§ Your uncle is going to buy a tractor. He has two alternatives:

Example: Decision tree (2/12)

Systems Analysis Laboratory

Systems Analysis Laboratory

§ The next step is to decide which tractor to buy

Result: “OK” New

Systems Analysis Laboratory

§ …But the engine of the old tractor can be defect

P(result ”OK”) Result: “OK” New P(”Defect” | result ”OK”)

Example: Decision tree (9/12)

§ We now have a decision tree presentation of the problem

New -18 500