Application of the Analytic Hierarchy Process (AHP) for Selection of Forecasting Software
Description of Analytic Hierarchy Process methodologymocenni/AHPpresentazione.pdf · AHP The...
Transcript of Description of Analytic Hierarchy Process methodologymocenni/AHPpresentazione.pdf · AHP The...
Description ofAnalytic Hierarchy Process
methodologyD.I.I., Universita di Siena
DITTY project
Ferrara, 13 December 2005
DITTY – p. 1
OutlineAnalytic Hierarchy Process
Implementation
Example
Conclusions
DITTY – p. 2
OutlineAnalytic Hierarchy Process
Implementation
Example
Conclusions
DITTY – p. 3
AHPThe Analytic Hierarchy Process(AHP), [T.Saaty (1980)], is a powerful tool thatmay be used to make decisions when
multiple and conflictingobjectives/criteria are present,
and both qualitative and quantitativeaspects of a decision need to beconsidered.
DITTY – p. 4
Literature review
Thousands of AHP applications havebeen proposed in the literature.
However, there exist relatively fewapplications of AHP to environmental ornatural resource problems.
DITTY – p. 5
How the AHP works
The AHP considers a set of evaluationcriteria, and a set of alternativescenarios among which the bestdecision is to be made.
It generates a weight for each evaluationcriterion and scenario according to theinformation provided by the DM.
The AHP combines the objective andscenario evaluations determining aranking of the scenarios.
DITTY – p. 6
Features
AHP is effective in dealing with complexdecision making because it reducescomplex decisions to a series of pairwisecomparisons.
DITTY – p. 7
Features
AHP is effective in dealing with complexdecision making because it reducescomplex decisions to a series of pairwisecomparisons.
AHP reduces the bias in the decisionmaking process because it also checksthe consistency of the DM’s evaluations.
DITTY – p. 7
FeaturesAHP may be considered as a tool that isable to translate the pairwise relativeevaluations (both qualitative andquantitative) made by the DM into amulticriteria ranking.
DITTY – p. 8
FeaturesAHP may be considered as a tool that isable to translate the pairwise relativeevaluations (both qualitative andquantitative) made by the DM into amulticriteria ranking.
The AHP is simple because there is noneed of building a complex expertsystem with the DM’s knowledgeembedded in it.
DITTY – p. 8
Concepts
Three major concepts behind the AHP:
The AHP is analytic
The AHP structures the problem as ahierarchy
The AHP helps in thedecision-making process
DITTY – p. 9
Concepts
Three major concepts behind the AHP:
The AHP is analytic
The AHP structures the problem as ahierarchy
The AHP helps in thedecision-making process
The AHP is analytic because it convertsDM’s inputs into numbers.
DITTY – p. 9
Concepts
Three major concepts behind the AHP:
The AHP is analytic
The AHP structures the problem as ahierarchy
The AHP helps in thedecision-making process
The AHP is hierarchical becausereduces complex decision-makingproblems into pairwise comparisions.
DITTY – p. 9
Concepts
Three major concepts behind the AHP:
The AHP is analytic
The AHP structures the problem as ahierarchy
The AHP helps in thedecision-making process
The AHP incorporates DM’s inputs anddefines a process for decision-making.
DITTY – p. 9
OutlineAnalytic Hierarchy Process
Implementation
Example
Conclusions
DITTY – p. 10
Implementation
The AHP consists of these steps:
1. Computing the vector of objectiveweights
2. Computing the matrix of scenarioscores
3. Ranking the scenarios
4. Checking the consistency
We consider m evaluation criteria and nscenarios.
DITTY – p. 11
Implementation
1. Computing the vector of objectiveweights
2. Computing the matrix of scenarioscores
3. Ranking the scenarios
4. Checking the consistency
DITTY – p. 12
Step 1
Pairwise comparison matrix A [m × m].
Each entry ajk of A represents theimportance of criterion j relative tocriterion k:
If ajk > 1, j is more important than k
if ajk < 1, j is less important than k
if ajk = 1, same importance
ajk and akj must satisfy ajkakj = 1.DITTY – p. 13
Step 1
The relative importance between twocriteria is measured according to anumerical scale from 1 to 9.
Value Interpretation
1 j and k are equally important
3 j is slightly more important than k
5 j is strongly more important than k
7 j is very strongly more important than k
9 j is absolutely more important than k
DITTY – p. 14
Step 1
The normalized pairwise comparisonmatrix Anorm is derived by A by makingequal to 1 the sum of the entries on eachcolumn.
DITTY – p. 15
Step 1
The normalized pairwise comparisonmatrix Anorm is derived by A by makingequal to 1 the sum of the entries on eachcolumn.
Finally, the objective weight vector w isbuilt by averaging the entries on eachrow of Anorm.
DITTY – p. 15
Implementation
1. Computing the vector of objectiveweights
2. Computing the matrix of scenarioscores
3. Ranking the scenarios
4. Checking the consistency
DITTY – p. 16
Step 2
The matrix of scenario scores S [n × m].
Each entry sij of S represents the scoreof the scenario i with respect to thecriterion j.
The score matrix S is obtained by thecolumns sj calculated as follows:
DITTY – p. 17
Step 2
A pairwise comparison matrix Bj is builtfor each criterion j.
Each entry bjih represents the evaluation
of the scenario i compared to thescenario h with respect to the criterion jaccording to the DM’s evaluations.
From each matrix Bj a score vectors sj
is obtained (as in Step 1).
DITTY – p. 18
Implementation
1. Computing the vector of objectiveweights
2. Computing the matrix of scenarioscores
3. Ranking the scenarios
4. Checking the consistency
DITTY – p. 19
Step 3
Once the weight vector w and the scorematrix S have been computed, the AHPobtains a vector v of global scores bymultiplying S and w, v = S · w.
The i-th entry vi of v represents theglobal score assigned by the AHP to thescenario i.
DITTY – p. 20
Step 3
Once the weight vector w and the scorematrix S have been computed, the AHPobtains a vector v of global scores bymultiplying S and w, v = S · w.
The i-th entry vi of v represents theglobal score assigned by the AHP to thescenario i.
The scenario ranking is accomplished byordering the global scores in decreasingorder.
DITTY – p. 20
Implementation
1. Computing the vector of objectiveweights
2. Computing the matrix of scenarioscores
3. Ranking the scenarios
4. Checking the consistency
DITTY – p. 21
Step 4
When many pairwise comparisons areperformed, inconsistencies may arise.
Consider the case
criterion 1 is slightly more importantthan criterion 2
criterion 2 is slightly more importantthan criterion 3
inconsistency arises if criterion 3 ismore important than criterion 1
DITTY – p. 22
Step 4
The Consistency Index (CI) is obtained:
x is the ratio of the j-th element of thevector A · w to the correspondingelement of the vector w
CI is the the average of the x
A perfectly consistent DM should alwaysobtain CI = 0, but inconsistenciessmaller than a given threshold aretolerated.
DITTY – p. 23
OutlineAnalytic Hierarchy Process
Implementation
Example
Conclusions
DITTY – p. 24
Example
Small example, m = 3 criteria and n = 3scenarios.
DITTY – p. 25
Example
The DM first builds the following pairwisecomparison matrix A for the 3 criteria:
A =
1 3 5
1/3 1 3
1/5 1/3 1
to which corresponds the weight vectorw = [0.633 0.261 0.106]T .
DITTY – p. 26
Example
Then, the DM builds the followingpairwise scenario comparison matricesfor the first criterion:
B1 =
1 3 7
1/3 1 5
1/7 1/5 1
to which correspond the score vectors1 = [0.643 0.283 0.074]T .
DITTY – p. 27
Example
Then, the DM builds the followingpairwise scenario comparison matricesfor the second criterion:
B2 =
1 1/5 1
5 1 5
1 1/5 1
to which correspond the score vectors2 = [0.143 0.714 0.143]T .
DITTY – p. 28
Example
Then, the DM builds the followingpairwise scenario comparison matricesfor the third criterion:
B3 =
1 5 9
1/5 1 3
1/9 1/3 1
to which correspond the score vectors3 = [0.748 0.180 0.072]T .
DITTY – p. 29
Example
Hence, the score matrix S is
S = [s1s2s3] =
0.643 0.143 0.748
0.283 0.714 0.180
0.074 0.143 0.072
and the global score vector isv = S · w = [0.523 0.385 0.092]T .
DITTY – p. 30
Example
The rank is:
scenario 1 (0.523)
scenario 2 (0.385)
scenario 3 (0.092)
Note that the first scenario results to bethe most preferable, though it is theworst of the three with respect to thesecond criterion.
DITTY – p. 31
OutlineAnalytic Hierarchy Process
Implementation
Example
Conclusions
DITTY – p. 32
Main advantages
Flexible tool
multiple and conflictingobjectives/criteria
qualitative and quantitative aspects ofa decision
DITTY – p. 33
Main advantages
Flexible tool
multiple and conflictingobjectives/criteria
qualitative and quantitative aspects ofa decision
Driven by the Decision Maker
reduces complex decisions to a seriesof pairwise comparisons
computations guided by the DM’sexperience DITTY – p. 33
Main advantages
Robust
capture both subjective and objectiveaspects of a decision
incorporates a technique for checkingthe consistency of the DM’sevaluations
DITTY – p. 34
Main advantages
Robust
capture both subjective and objectiveaspects of a decision
incorporates a technique for checkingthe consistency of the DM’sevaluations
Easy to implement
few matrix manipulations are required
DITTY – p. 34
Drawback
The AHP may require a large number ofevaluations by the DM, especially forlarge problems.
Every evaluation requires the DM toexpress how well two scenarios orcriteria compare to each other...
DITTY – p. 35
Drawback
The AHP may require a large number ofevaluations by the DM, especially forlarge problems.
Every evaluation requires the DM toexpress how well two scenarios orcriteria compare to each other......but the number of pairwisecomparisons grows quadratically withthe number of criteria and scenarios.
DITTY – p. 35
DrawbackAn example
With n = 10 scenarios and m = 4objectives.
Pairwise comparisons required:
Weight vector 4 · 3/2 = 6
Score matrix 4 · (10 · 9/2) = 180
DITTY – p. 36
DrawbackHowever, it may be completely orpartially automated for deciding somepairwise comparisons.
DITTY – p. 37
DrawbackHowever, it may be completely orpartially automated for deciding somepairwise comparisons.
Bjab = 1 + 8(
xja−x
jb
max xji−min x
ji
)
DITTY – p. 37
DrawbackHowever, it may be completely orpartially automated for deciding somepairwise comparisons.
Bjab = 1 + 8(
xja−x
jb
max xji−min x
ji
)
Different behaviours may beimplemented.
DITTY – p. 37