carlos a. bana e costa - université paris dauphinetsoukias/download/tutorials/shortvaria/... ·...
TRANSCRIPT
International Doctoral School onMultiple Criteria Decision Analysis (MCDA), Data Mining and Rough SetsMultiple Criteria Decision Analysis (MCDA), Data Mining and Rough Sets
Carlos A. Bana e CostaFull Professor of Decision & Information
School of Engineering (IST), Technical University of LisbonMember of the Centre of Management Studies of IST (CEG-IST)g ( )
Visiting Professor of Decision SciencesDepartment of Management
L d S h l f E i d P liti l S iLondon School of Economics and Political Science
http://web.ist.utl.pt/carlosbana
Troina, Sicily, Italy, 12 April 2008
Development and use of logical methodsfor the improvement of decision makingfor the improvement of decision-making
in public and private enterprise.
Such methods include:
• models for decision-making under conditions of
uncertainty or multiple objectives
• techniques of risk analysis and risk assessment;
i l d d i i di f d i i ki b h i• experimental and descriptive studies of decision-making behavior
• economic analysis of competitive and strategic decisions
• techniques for facilitating decision-making by groupstechniques for facilitating decision making by groups
• computer modeling software and expert systems
for decision support
http://decision-analysis.society.informs.org/© 2008 Carlos A. Bana e Costa
© 2008 Carlos A. Bana e Costa
Problem dominated by
i i
Uncertainty Multiple Objectives
EXTEND conversation•Event tree•Fault tree
EVALUATE options•Multi-criteria decision analysisCHOOSE option
REVISE opinion
Fault tree•Influence diagram
decision analysis
ALLOCATE resources•Multi criteria
p•Payoff matrix•Decision tree
•Bayesian nets•Bayesian statistics
SEPARATE into components
•Multi-criteria commons dilemma
NEGOTIATESEPARATE into components•Credence decomposition•Risk analysis
•Multi-criteria bargaining analysis
y
© 2005 Larry PhillipsReference: L.D.Phillips, Decision Analysis in 2005
MULTIPLE CRITERIA DECISION ANALYSIS
Several issuesFamíly of
Multiple(consequences, attributes, characteristics, objectives,
dimensions, etc.)
fundamental
Points of view
Multiple
criteria
Descriptorsof performance
Intracriterion judgments:Local evaluation
models
Coherent family ofEvaluation criteria
Intercriteria Multicriteriaaggregation Overall
judgments:Weighting criteria
aggregationof local preferences evaluation
model
© 2008 Carlos A. Bana e Costa
MULTIMULTI--CRITERIA VALUE MEASUREMENTCRITERIA VALUE MEASUREMENT
• Measuring the relative value of options in each criterion:
Numerical (e.g. direct rating) andNon-numerical approaches (e.g. MACBETH)pp ( g )
• Criteria weighting procedures• Criteria weighting proceduresNumerical techniques (e.g. swing weighting)N i l t h i ( MACBETH)Non-numerical techniques (e.g. MACBETH)
© 2008 Carlos A. Bana e Costa
MULTI-CRITERIA VALUE MEASUREMENT
n
Evaluation framework: Additive value model
)a(v.k)a(V j
n
1jj∑
=
=
⎪⎪⎧
∀
∀=
j0)hl(j ,100)anchorupper(v jjWith:
V(a) overall value of option a⎪⎪
⎪⎪
⎨=
∀=
)(100)anchorsupperall(Vj,0)anchorlower(v jj
vj(a) local value (score)f ti
⎪⎩ = 0)anchorslowerall(V
of option aagainst criterion j
∑n
kj scaling constant(relative weight)
∑=
=1j
j 1k and kj > 0 ( j = 1,…,n )
of criterion j
© 2008 Carlos A. Bana e Costa
Non-numerical approach: MACBETH
Measuring Attractiveness by a Categorical Based Evaluation Technique
An interactive pairwise comparison approachto guide the construction of a quantitative value modelg q
from qualitative value judgments
When you take all non-verbal judgment out of a decision it becomes a calculation and not a decisionit becomes a calculation and not a decision.
Elliot JaquesElliot JaquesRequisite Organization, 1988
© 2008 Carlos A. Bana e Costa
Download the Demo Version of M-MACBETH http://www m-macbeth comhttp://www.m-macbeth.com
© 2008 Carlos A. Bana e Costa
How does it work?MACBETH i l ti t lMACBETH uses a simple question-answer protocolthat involves only two options in each question:Ask the evaluator to pairwise compare optionsby given a qualitative judgementy g q j gof the difference in attractivenessbetween each two optionsp
For x and y such thatFor x and y such thatx is preferred to y,
the difference in attractivenessthe difference in attractivenessbetween x and y is:
© 2008 Carlos A. Bana e Costa
MACBETH semantic categories ofdifference of attractiveness:
CC6
C5
C4
C3C3
C2
CC1
C0
Note: the ‘weak’, ‘strong’ and ‘extreme’ were initially called the fundamental categories, but the M-MACBETH software that implements the MACBETH approach does not make this distinction and even allows for group judgments that do not
© 2008 Carlos A. Bana e Costa
does not make this distinction and even allows for group judgments that do not distinguish between several consecutive categories, such as ‘strong or very strong’.
How many judgements?
For a set X of m options, the number of pairwisecomparisons can vary from a maximum of m(m-1)/2 judgments when all pairwise comparisons are madejudgments, when all pairwise comparisons are made,to a minimum acceptable number of m–1 judgments, as when comparing only each two consecutive options in the ranking or one options with all of the other m 1in the ranking or one options with all of the other m–1 (however, it is recommended to ask for some additional judgments to perform several consistency checks)
© 2008 Carlos A. Bana e Costa
checks).
Larry’s summerholidays in Hawai
Kaua’iNi’h holidays in HawaiO’ahu
M l k ’iNi’hau
Moloka’iLãna’i M iLãna i Maui
Hawai’i
In which island?
© 2008 Carlos A. Bana e Costa
© 2008 Carlos A. Bana e Costa
Assessing MACBETH intracriterion preferenceintracriterion preference information
As each judgement is entered in the matrix, its consistency with the judgments already inserted is checked and possible
if an inconsistency is detected
already inserted is checked and possible inconsistencies are detected.
is detected, suggestions to overcome it are presentedpresented. Technically, this is done by a
th ti lmathematical programming algorithm (see Bana e Costa et al. 2005 for details).
© 2008 Carlos A. Bana e Costa
Molo
Very weak
Lana
Strong
ModerateMaui
Moderate
Weak
Oahu© 2008 Carlos A. Bana e Costa
For a set of consistent judgements, MACBETH suggestsj g , gga numerical scale v on X that satisfies the followingmeasurement rules:Rule 1Rule 1∀ x, y ∈ X : v(x) = v(y) iff x and y are equally attractive∀ x, y ∈ X : v(x) > v(y) iff x is more attractive than y;Rule 2Rule 2∀ k, k’ ∈ { 1, 2, 3, 4, 5, 6 }, ∀ x, y, w, z ∈ X,with (x, y) ∈ Ck and (w, z) ∈ Ck’ :
k ≥ k’ + 1 ⇒ v(x) -v(y) > v(w) -v(z)
© 2008 Carlos A. Bana e Costa
The softwareThe software determines the interval within which each scorewhich each score of each option can vary when the other m-1 scores are fixed and still remain compatible with the matrix of judgments.
This allows the adjustment of the scale by comparing differences of scores, to arrive to a cardinal scale.
© 2008 Carlos A. Bana e Costa
“Behind” M-MACBETHBehind M MACBETH
João Bana e Costa Carlos Bana e Costa
A B C DA B C D
A Very Weak Moderate Very StrongExtreme 6
Very St 5
B Moderate StrongStrong
Strong 4
C StrongModerate 3
Weak 2
DVery Weak 1
© 2008 João Bana e Costa & Carlos A. Bana e Costa
A B C DA B C D
A Very Weak Moderate Very StrongExtreme 6
Very St 51
B Moderate StrongStrong
Strong 4
1
3
C StrongModerate 3
Weak 24
DVery Weak 1
© 2008 Carlos A. Bana e Costa
A B C DA B C D
A Very Weak Moderate Very StrongExtreme 6
Very St 51 4
B Moderate StrongStrong
Strong 4
1 4
3
C StrongModerate 3
Weak 24
DVery Weak 1
v(A)-v(C) = v(A)-v(B) + v(B)-v(C)
© 2008 Carlos A. Bana e Costa
A B C DA B C D
A Very Weak Moderate Very StrongExtreme 6
Very St 51 4
B Moderate StrongStrong
Strong 4
1 4
3
C StrongModerate 3
Weak 24
DVery Weak 1
v(A)-v(C) < v(C)-v(D)
© 2008 Carlos A. Bana e Costa
A B C DA B C D
A Very Weak Moderate Very StrongExtreme 7
Very St 61 4
B Moderate StrongStrong
Strong 5
1 4
3
C StrongModerate 3-4
Weak 25
DVery Weak 1
Increase v(C)-v(D) of 1
© 2008 Carlos A. Bana e Costa
A B C DA B C D
A Very Weak Moderate Very StrongExtreme 10
Very St 91 4
B Moderate StrongStrong
Strong 5-8
1 4
3 8
C StrongModerate 3-4
Weak 25
DVery Weak 1
v(B)-v(D) = v(B)-v(C) + v(C)-v(D)
© 2008 Carlos A. Bana e Costa
A B C DA B C D
A no Very Weak Moderate Very Strong
1 4 9
Extreme 10
Very St 9
B no Moderate Strong
1 4
3 8
9 Strong
Strong 5-8
C no Strong
5
Moderate 3-4
Weak 2
D noVery Weak 1
v(A)-v(D) = v(A)-v(B) + v(B)-v(C)+ + v(C)-v(D)
© 2008 Carlos A. Bana e Costa
© 2008 Carlos A. Bana e Costa
MACBETH weighting procedure
strong weakt
veryt
strongto
strongto
moderatestrongto
very strong
Qualitative swing judgementsSUM=1
© 2008 Carlos A. Bana e Costa
2) to compare the difference in overallvalue between any two swings (suchthat the first is more attractive thanthe second)
© 2008 Carlos A. Bana e Costa
REFERENCES
B C C A V i k J C D C J M (2003) MACBETH W ki P LSEOR 03 6Bana e Costa, C.A., Vansnick, J.C., De Corte, J.M. (2003), MACBETH, Working Paper LSEOR 03.56,London School of Economics, London.
Bana e Costa C.A., M.P. Chagas. 2004. A career choice problem: an example of how to useMACBETH t b ild tit ti l d l b d lit ti l j d t EMACBETH to build a quantitative value model based on qualitative value judgments. EuropeanJournal of Operational Research 153(2) 323-331.
Bana e Costa C.A., J.M. De Corte, J.C. Vansnick. 2005. On the mathematical foundations ofMACBETH J Fi i S G M Eh tt d M lti l C it i D i i A l i St t f th A tMACBETH. J. Figueira, S. Greco, M. Ehrgott, eds. Multiple Criteria Decision Analysis: State of the ArtSurveys. Springer, New York, 409-442.
Bana e Costa C.A., J.M. De Corte, J.C. Vansnick. 2005. M-MACBETH Version 1.1 User’s Guide,htt // b th /d l d ht l# idhttp://www.m-macbeth.com/downloads.html#guide.
Bana e Costa J. 2007. Behind MACBETH. Presented at the Nato Advanced Research on Risk,Uncertainty and Decision Analysis for Environmental Security and Non-chemical Stressors, Lisbon,P t l htt // b th / f ht l#b iPortugal, http://www.m-macbeth.com/references.html#basic.
Bana e Costa, C.A., Lourenço, J.C., Chagas, M.P., Bana e Costa, J.C. (2008), "Development ofreusable bid evaluation models for the Portuguese Electric Transmission Company", DecisionA l i 5 1 (22 42)
© 2008 Carlos A. Bana e Costa
Analysis, 5, 1 (22-42).