Methods and Models for Decision Making
Alberto Colorni – Dipartimento INDACO, Politecnico di Milano Alessandro Lué – Consorzio Poliedra, Politecnico di Milano
5°
© Alberto Colorni
MMDM – Lesson 5
• (1) Introduction (2) Tools & frame
• (3) Mental models (4) Design & decision
• (5) Classification (6) Ranking-1, risk analysis
• (7) Ranking-2, multicriteria (8) A tentative case
• (9) Rating problems (10) Seminar M. Henig
• (11) Group decision (12) Genetic alg. + …
• (13) Research topics (14) Case results (if any …)
• (15) Conclusions
Index:
• Variable space (decision) & objective space (results)
• Three phases in MCDM (or MODM) problems
• Phase 1 from indicators to utilities
• Phase 2 elimination of dominated solutions
• Phase 3 DM preferences & final choice
• The second is the only “objective” phase
• Utility functions & preferences of this DM
• Sensitivity analysis with respect to the weight vector
2
God in 7 steps:
© Alberto Colorni 3
A tentative case:
the “Colorni award”
(the best italian newspaper on the web)
© Alberto Colorni
The best italian newspaper on the web
(you are in the jury …)
• What are the alternatives ?
• What are the attributes ?
• What are the utility functions ?
• What are the weights ?
• What is the ranking ? (the personal one and the collective one)
© Alberto Colorni
Alternatives
• Repubblica
• Corriere Sera
• Sole24Ore
• Ansa online
• RaiNews24
• Foglio
• Gazzetta d. Sport ? NO, because it is too specific (only sports)
• Novella 2000 ? NO, because it is a magazine (weekly)
What are the rules ? (if there are rules …)
The alternatives must be “similar” (but the concept of similarity is subjective)
© Alberto Colorni
Attributes
• A lot of (tentative) attributes; note that the attributes must generate a set of indicators that could be measurables
• Four main indicators:
items in the homepage
upgrades during the day
daily visitors (declaration)
quality, in a scale [1, 10]
• A search (made by the students) to obtain the data set
© Alberto Colorni
Performance matrix
7
A1 A2 A3 A4 A5 A6
I1 29 24 23 16 6 8 items
I2 120 240 288 288 288 1 upgrades
I3 3.6 3.3 0.03 0.02 0.23 0.01 visitors
I4 8 8 7 8 8 2 quality
© Alberto Colorni
Utility functions
• One utility function for each attribute/indicator
• Three steps for each utility function (easy version)
• The results
• The evaluation matrix (in a common [1, 10] scale)
• Phase 2 dominated alternatives ?
© Alberto Colorni
Utility functions
9
5
ind.1
10
0
10 20 30
ind.2
10
0
100 300 .5
ind.3
10
0
… 3.3 5
0 0
© Alberto Colorni
Evaluation matrix
10
A1 A2 A3 A4 A5 A6
C1 1 6 7 10 2 6
C2 5.5 8.5 9.75 9.75 9.75 0
C3 10 10 0.5 0.4 4.6 0
A6 (Il Foglio) is dominated, but …
C4 8 8 7 8 8 2
© Alberto Colorni
Pairwise comparison (to obtain the weights)
11
C1 C2 C3 C4
C1 1 *
C2 * 1 *
C3 1
C4 * * * 1
C4 > C2 > C1 > C3
weights?
(numerical values)
© Alberto Colorni
Determination of weights / 1
C1 C2 C3 C4
C1 1 1/3 2 1/4
C2 3 1 3 1/2
C3 1/2 1/3 1 1/5
C4 4 2 5 1
(reciprocal matrix) coherent?
C1 C2
C4 C3
3
3
5 4 2
2
© Alberto Colorni
Determination of weights / 2
1 1/2 2 1/4
2 1 4 1/2
1/2 1/4 1 1/8
4 2 8 1
(consistent matrix) coherent ?
C1 C2
C4 C3
2
4
8 4 2
2
2/15 4/15 4/15 8/15 collective weights
w1
w2
w3
w4
© Alberto Colorni
Weights (the preference structure)
• We are 30 (approximately) DMs
• What is the way to obtain (shared) weights ?
• Discussion …
• A common vector of weights ?
• Total utility of each candidate final ranking
© Alberto Colorni
General utilities
1 6 7 10 2 6
5.5 8.5 9.75 9.75 9.75 0
10 10 0.5 0.4 4.6 0
8 8 7 8 8 2
98 120 109.5 123.4 111.6 28 … /15
2
4 1 8
… /15
Ranking: A4 – A2 – A5 – A3 – A1 – A6
1. RAI news 2. Corriere 3. ANSA 4. Sole24 5. Repubblica 6. Foglio
weighted sum
© Alberto Colorni
A lot of rankings (the personal paths)
• A set of (∼ 30) individual rankings
• What can we do ?
• It is a group decision see lesson 6
• The final result is the “less conflicting” (or more shared) ranking
• Two main procedures Distillation or Maastricht
© Alberto Colorni 17
DI – Deeper Investigations
© Alberto Colorni
DI-1 Utility & dominance
cost
alt 1 alt 2 alt 3 0.3 0.8 0.7
25 47 12
1 6 5
250 710 940
7 6 4
53 71 98
409 524 912 air quality
water quality flora
fauna
landscape employment
© Alberto Colorni
Utility functions
alt 1 alt 2 alt 3
80 60 70
70 50 80
40 80 70
50 70 90
70 60 40
50 70 80
90 70 40
air quality water quality flora
fauna landscape employment
cost cost
u7 100
0
u2
Water quality
100
0
u4
fauna
100
0 alt 1 alt 2 alt 3
0.3 0.8 0.7
25 47 12
1 6 5
250 710 940
7 6 4
53 71 98
409 524 912 air quality
flora
landscape employment
cost
water quality fauna
© Alberto Colorni
Pareto dominance
The Pareto criteria has to be verified only after the application of the utility
functions
alt 1 alt 2 alt 3
80 60 70
70 50 80
40 70 70
50 70 90
70 30 40
50 70 90
90 40 40
air quality water quality flora
fauna landscape employment cost
cost
alt 1 alt 2 alt 3 0.3 0.7 0.5
25 47 12
1 5 5
250 940 710
7 3 4
53 71 83
409 912 912
air quality water quality flora
fauna landscape employment
(attributi)
(criteria)
utility
alt2 dominated by alt3 Performance
matrix
© Alberto Colorni
The preferential rank (and the structure) between 2 values of an ttribute doesn’t depend on the value of the other attribute
Example the purchase of a radio
2 attributes • price ↓
• ratio signal/noise ↑
Counterexample1 choice in a menu
2 attributes • food (fish, meat)
• beverage (white wine, red wine)
Counterexample2 the chemical reaction
2 attributi • reagent A
• reagent B (4,5) > (3,5) but (4,2) < (3,2)
complementary goods, synergic effects keep the ratio 1:1
DI-2 Mutual preferential independence
© Alberto Colorni
separabilità U(x,y)= f(ux(x) , uy(y))
additività U(x,y)= f1 [ux(x)] + f2 [uy(y)]
Mutual preferential indipendence
2 steps:
• determine ux(x) e uy(y)
• determine the functions f1 e f2
More formally
The mutual preferential indipendence is a necessary condition (but not sufficient) for the additivity
© Alberto Colorni
DI-3 Tasso di sostituzione
Il decisore è sempre disposto a compensare il peggioramento della prestazione di un attributo con un opportuno miglioramento della prestazione di uno qualsiasi degli altri.
Tasso marginale di sostituzione
© Alberto Colorni
Tasso marginale
1 1 1
λ
Aλ
Bλ
Cλ
Eλ
DA B
C
D E
λ
B<λ
A< λ
Cλ
E<λ
A< λ
Dx
y
A’
© Alberto Colorni
DI-4 Weight assignment
y
x0 x y0 y1
x1
2 domande:
1. è preferibile (x0, y1) o (x1,y0)?
se la risposta è (x1,y0)
⇒
w1>w22. supponiamo di essere in (x0, y1) , qual è il valore di x per cui (x0, y1) è indifferente a (x, y0)?
w1 ux(x0) + w2 uy(y1) = w1 ux(x) + w2 uy(y0) w1 + w2 = 1
x
© Alberto Colorni
Weight assignment: pair comparison
verde pubblico (m2 /ab) x1
x2 % occupazione
•A
•B principio di compensazione
tasso di sostituzione
1
w2
w
n pesi ⇒ n-1 confronti a coppie
⇒ n-1 rapporti
j
wi
w
∑
=i
i 1
& w
80 ? 82
7.0 6.7
è il rapporto tra le differenze di utilità
© Alberto Colorni
DI-5 The quality of life
• La qualità della vita nelle città italiane
• Indagine del Sole24ore (29 dicembre 2008)
• Analizzati 6 settori, con 36 indicatori
• Funzioni di utilità, pesi, classifiche
• Confronto tra classifiche basate sugli indicatori e
percezione di un campione (circa 700 ab. per provincia)
© Alberto Colorni
Test examples
© Alberto Colorni
Describe how the Sudoku can be seen as a decision (or operational research) problem
or
What is the Pareto frontier ? Describe it and show an example or
Define the concept of lottery. Elaborate with two numerical examples or
Describe the main ideas behind the “C-K theory”
or
… ( a concept to be explained) …
Example-1: free text concerning an argument
(no more than 300 words or 2000 characters)
© Alberto Colorni
In a problem characterized by two attributes (x and y), you know the utility functions of such attributes:
Moreover, you know that the decision-maker is indifferent to two following situation (A and B)
A B
Please discuss if the situation K is preferable to H Example-2
x x
ux( ) =1− ) 2
(y y uy =
=
=
2 1 1 y x
=
= 0 2 1 y x
=
=
4 1 2 1
y x
=
=
2 1 4 3
y x
© Alberto Colorni
A multicriteria decision problem (6 alternatives, 3 criteria = utilities) is showed in this matrix, with its weight vector.
1. Are there dominated alternatives ?
2. What is the ranking and the final choice ?
3. Is the result changing if w2 increase ? Is there a rank reversal ? Explain (briefly) all the answers.
Example-3
a1 a2 a3 a4 a5 a6
c1 60 40 20 70 100 80 w1 0.20
c2 40 40 35 35 35 40 w2 0.40
c3 20 30 60 40 50 50 w3 0.40