Sei sulla pagina 1di 40

ANALYTIC HIERARCHY PROCESS

Yan Liu

Department of Biomedical, Industrial & Human Factors Engineering


Wright State University
Yan Liu, Department of BIE, Wright State University

Introduction
Many decisions have conflicting objectives and tradeoffs
e.g. higher returns vs. lower risks, better performance vs. lower price
Some objectives have incomparable attribute scales
Multi-Attribute Decision Making (MADM)
A study of methods and procedures that handle multiple attributes
Various usages
Identify a single most preferred option
Rank options
Shortlist a limited number of options for subsequent detailed appraisal
Distinguish acceptable from unacceptable possibilities

2
Yan Liu, Department of BIE, Wright State University

Introduction
Types of MADM Techniques
Multiattribute scoring model (learned in Chapter 4)
Covert attributes to comparable measures if they are not
Assign weights to these attributes and then calculate the weighted average of each
consequence set as an overall score
Compare alternatives using the overall score
Multi-objective mathematical programming
Tackle complex problems involving a large number of decision variables that are subject to
constraints
Analytic hierarchy process (AHP)
Multi-attribute utility theory (MAUT)

3
Yan Liu, Department of BIE, Wright State University

What is AHP?
A process that leads one to (Saaty, 1980)
Structure a problem as a hierarchy or as a system with dependence loops
Elicit judgments that reflect ideas, feelings, and emotions
Represent those judgments with meaningful numbers
Synthesize results
Analyze sensitivity to changes in judgment
AHP also uses a weighted average approach idea, but it uses a
method for assigning ratings and weights that is considered more
reliable and consistent

4
Yan Liu, Department of BIE, Wright State University

Purposes of AHP
To structure complexity in gradual steps from the large to the
small, or from the general to the particular, so we can relate them
with greater accuracy according to our understanding
To improve our awareness by richer synthesis of our knowledge
and intuition; AHP is a learning tool rather than a means to
discover the TRUTH because truth is relative and changing

5
Yan Liu, Department of BIE, Wright State University

Phases in AHP
Phase 1: Decompose the problem into a hierarchy
AHP starts with an identification of the criteria to be used in evaluating
different alternatives which are organized in a tree-like hierarchy
Phase 2: Collect input data by pairwise comparisons of criteria at each
level of the hierarchy and alternatives
Phase 3: Estimate the relative importance (weights) of criteria and
alternatives and check the consistency in the pairwise comparisons
Phase 4: Aggregate the relative weights of criteria and alternatives to
obtain a global ranking of each alternative with regards to the goal
6
Yan Liu, Department of BIE, Wright State University

Hierarchy

GOAL

CRITERIA

ALTERNATIVES

7
Yan Liu, Department of BIE, Wright State University

Hierarchy
Overall Job Satisfaction GOAL

Research Growth Benefits Colleagues Location Reputation CRITERIA

Job A Job B Job C ALTERNATIVES

8
Yan Liu, Department of BIE, Wright State University

Hierarchy
How to Structure a Hierarchy
Identify the overall objective or goal
Identify criteria to satisfy the goal
Identify, where appropriate, sub-criteria under each criterion
Identify alternatives to be evaluated in terms of the sub-criteria
If the relative importance of the sub-criteria can be assessed and the
alternatives can be evaluated in terms of the sub-criteria, the
hierarchy is finished
Otherwise, continue inserting levels until it is possible to link levels
and set priorities on the elements at each level in terms of the
elements at the level above it
9
Yan Liu, Department of BIE, Wright State University

Hierarchy

Goal More General

C1 C2 C3

C11 C12 C13 C21 C22 C31 C32 C33

Sub-criteria at More Specific


the lowest level

Alternatives
10
Yan Liu, Department of BIE, Wright State University

Hierarchy
Societys Overall Benefit

National Health, Safety Political


Economy Environment Factors

Accidents Political
Cheap Foreign Capital Natural Unavoidable Independ- Centra-
Long-Term Cooperative-
Electricity Trade Resources Resources Pollution ence lization
Risks ness

No big Coal-fired Nuclear


power plants power plant power plant
11
Energy Decision in the Parliament of Finland
Yan Liu, Department of BIE, Wright State University

Hierarchy
How large should a hierarchy be?
Large enough to capture your major concerns
Small enough to remain sensitive to change in what is important

12
Yan Liu, Department of BIE, Wright State University

Judgment and Preference

In AHP, we use subjective judgment to express preference and its


intensity
e.g. Which of two apples is more red and how strongly more red we
perceive it to be
From this preference we derive one scale of relative strength of
preference

13
Yan Liu, Department of BIE, Wright State University

Pairwise Comparison
Define the relative importance of criteria at each level of the hierarchy and
relative importance of alternatives by means of pairwise comparisons
Criterion Criterion Criterion

(Alternative) 1 (Alternative) 2 (Alternative) n
Criterion
W1/W1 W1/W2 W1/Wn
(Alternative) 1
Criterion
W2/W1 W2/W2 W2/Wn
(Alternative) 2

Criterion
Wn/W1 Wn/W2 Wn/Wn
(Alternative) n
Pairwise Comparison Matrix 14
(Wi is the relative weight of ith criterion (alternative))
Yan Liu, Department of BIE, Wright State University

Pairwise Comparison

Scale for Pairwise Comparisons


1. Equally preferred
3. One is moderately preferred over another
5. One is strongly preferred over another
7. One is very strongly preferred over another
9. One is extremely preferred over another
2,4,6,8 intermediate values
Reciprocals for inverse comparison

15
Yan Liu, Department of BIE, Wright State University

Relative Weights
Denote the pairwise comparison matrix and the weight matrix as
WW
1 W1
W2 ...
W1
Wn w1
W12 W2
W2
W1 ... W = w 2
A nn = W2 Wn
n1 ...
... ... ... ...
w n
W n Wn Wn
... Wn
W1 W2

Then W1 W1
... w 1
W1
w1
W W2 Wn
W21
W2
W2 w
W1 ... 2
w 2

W2 Wn
= Or AW = W
... ... ... ... ... ...
W n Wn
... W n
w
Wn n
w n
1
W W2

is the eigenvalue of A and W is its corresponding right eigenvector


16
There are n eigenvalues and corresponding eigenvectors for Anxn
Yan Liu, Department of BIE, Wright State University

Relative Weights
Consistency
Transitivity
a1>a2 and a2>a3, then a1>a3
Measurement consistency
aij ajk = aik (aij the cell at the ith row and jth column of the comparison matrix)
Since achieving consistency is hard in pairwise comparisons, AHP introduces
notions of deviations of consistency
Consistency Index (CI)
max n
CI = n 1 max is the maximum eigenvalue of the pairwise comparison matrix
CI=0 or max = n implies perfect consistency
CI=0.1 is the generally accepted threshold value
17
Yan Liu, Department of BIE, Wright State University

Relative Weights
AW = W (A I) W = 0 (Equation 1) 1
1
, where I is the identity matrix I=
...

1
W 1
W1
W1
W2
... x 1 0
W1
Wn
W2

W2
... W2
x 2 0
Or W1 =
W2 Wn

... ... ... ... ... ...


W n Wn
... Wn
x n 0
1 W W2 Wn

The solution to Equation (1) are given by det(A I) = 0 | A I |= 0


You can solve eigenvalue and eigenvector problems in Matlab using the command
[V,E]=eig(A), where E is the eigenvalue and V is the corresponding eigenvector

18
Yan Liu, Department of BIE, Wright State University

Relative Weights
Suppose the comparison matrix
of three criteria is
1 3 5

A33 = 1 / 3 1 3
1 / 5 1 / 3 1

There is only one non-zero real
eigenvalue for the pairwise eigenvectors
comparison matrix in AHP
In this example, max =3.0385
Its corresponding eigenvector is eigenvalue
(0.9161, 0.3715, 0.1506) T; the
three numbers in the eigenvector
are proportional to the relative
Finding its eigenvalue and eigenvector Using Matlab 19
weights of the three criteria
Yan Liu, Department of BIE, Wright State University

Relative Weights
Because relative weights must sum up to 1, we have to normalize the eigenvector by
dividing each number in it by the sum of all numbers
e.g. In the previous example, the eigenvector is (0.9161, 0.3715, 0.1506) T
The normalized eigenvector is
( 0.9161+ 00..3715
9161
, 0.3715
, 0.1506
+ 0.1506 0.9161+ 0.3715 + 0.1506 0.9161+ 0.3715+ 0.1506 )

=(0.64, 0.26, 0.10) Weights for the three criterion

CI=(max-n)/(n-1) = (3.0385-3) /(3-1) = 0.0193

20
Yan Liu, Department of BIE, Wright State University

School Selection Example

Satisfaction With School GOAL

School Vocational College Music


Learning Friends
Life Training Preparation Classes
CRITERIA

School A School B School C ALTERNATIVES

21
Yan Liu, Department of BIE, Wright State University

School Selection Example

L F SL VT CP MC

L 1 4 3 1 3 4

F 1/4 1 7 3 1/5 1

SL 1/3 1/7 1 1/5 1/5 1/6

VT 1 1/3 5 1 1 1/3

CP 1/3 5 5 1 1 3

MC 1/4 1 6 3 1/3 1

Pairwise Comparison Matrix of Six Criteria


(L,F,SL,VT,CP, and MC denote Learning, Friends, School Life, Vocational Training, 22
College Preparation, and Music Classes, respectively)
Yan Liu, Department of BIE, Wright State University

School Selection Example

1 4 3 1 3 4
1 / 4 1 7 3 1 / 5 1

1 / 3 1 / 7 1 1 / 5 1 / 5 1 / 6
A=
1 1/ 3 5 1 1 1 / 3
1 / 3 5 5 1 1 3

1 / 4 1 6 3 1 / 3 1

Using Matlab, we can get max=7.4199,


and its corresponding eigenvector is ( 0.6906, 0.3003, 0.0748, 0.2766, 0.5111, 0.2994) T
The normalized eigenvector is (0.32, 0.14, 0.03, 0.13, 0.24, 0.14) T
CI =( 7.4199 6) / (6 1) = 0.284 > 0.1
So there is large inconsistency in the pairwise comparison matrix of six criteria
23
Yan Liu, Department of BIE, Wright State University

School Selection Example

L F SL VT CP MC Weight

L 1 4 3 1 3 4 0.32

F 1/4 1 7 3 1/5 1 0.14

SL 1/3 1/7 1 1/5 1/5 1/6 0.03

VT 1 1/3 5 1 1 1/3 0.13

CP 1/3 5 5 1 1 3 0.24

MC 1/4 1 6 3 1/3 1 0.14


24
Yan Liu, Department of BIE, Wright State University

School Selection Example

A B C Weight

A 1 1/3 1/2 0.16

B 3 1 3 0.59

C 2 1/3 1 0.25

Pairwise Comparison Matrix of Three


Alternative Schools With Respect to Learning
max=3.0536, and the normalized eigenvector is (0.16, 0.59, 0.25) T

CI =( 3.0536 3) / (3 1) = 0.027 25
Yan Liu, Department of BIE, Wright State University

School Selection Example

A B C Weight

A 1 1 1 0.33

B 1 1 1 0.33

C 1 1 1 0.33

Pairwise Comparison Matrix of Three


Alternative Schools With Respect to Friends
max=3 , and the normalized eigenvector is (1/3, 1/3, 1/3) T
CI =( 3 3) / (3 1) = 0 26
Yan Liu, Department of BIE, Wright State University

School Selection Example

A B C Weight

A 1 5 1 0.46

B 1/5 1 1/5 0.09

C 1 5 1 0.46

Pairwise Comparison Matrix of Three


Alternative Schools With Respect to School Life
max=3 , and the normalized eigenvector is (0.46, 0.09, 0.46) T
CI =( 3 3) / (3 1) = 0 27
Yan Liu, Department of BIE, Wright State University

School Selection Example

A B C Weight

A 1 9 7 0.77

B 1/9 1 1/5 0.05

C 1/7 5 1 0.17

Pairwise Comparison Matrix of Three Alternative


Schools With Respect to Vocational Training
max=3.2085, and the normalized eigenvector is (0.77, 0.05, 0.17) T

CI =( 3.2085 3) / (3 1) = 0.104 28
Yan Liu, Department of BIE, Wright State University

School Selection Example

A B C Weight

A 1 1/2 1 0.25

B 2 1 2 0.50

C 1 1/2 1 0.25

Pairwise Comparison Matrix of Three Alternative


Schools With Respect to College Preparation
max=3 , and the normalized eigenvector is (0.25, 0.50, 0.25) T

CI =( 3 3) / (3 1) = 0 29
Yan Liu, Department of BIE, Wright State University

School Selection Example

A B C Weight

A 1 6 4 0.69

B 1/6 1 1/3 0.09

C 1/4 3 1 0.22

Pairwise Comparison Matrix of Three Alternative


Schools With Respect to Music Classes
max=3.0536, and the normalized eigenvector is (0.69, 0.09, 0.22) T

CI =( 3.0536 3) / (3 1) = 0.027 30
Yan Liu, Department of BIE, Wright State University

Relative Weights
Other than computing the eigenvector of a pairwise comparison
matrix to find the weights of compared criteria or alternatives, we
can also approximate the weight by:
First, normalizing each column in the comparison matrix
Then, calculating the average of each row in the normalized matrix as the
estimate of the relative weight for its corresponding criterion or
alternative

31
Yan Liu, Department of BIE, Wright State University

Relative Weights
Criterion Criterion Criterion
(Alternative) (Alternative) (Alternative) Approximated Consistency
1 2 n Weight Measure
Criterion n W /W n ~
1Ti i (W1/Wi ) Wi
(Alternative) W1/W1 W1/W2 W1/Wn ~ CM1= i=1 ~
1 T1 T2 Tn W1= i=1
n W1

Criterion n ~
n W /W
(Alternative) W2/W1 W2/W2 W2/Wn 2Ti i (W2 /Wi ) Wi
~ CM2=
W2 =
i =1
i=1
2 T1 T2 Tn n
~
W2


Criterion n W /W n ~
(Alternative) Wn/W1 Wn/W2 Wn/Wn n i (Wn /Wi ) Wi
~ Ti

n T1 T2 Tn Wn = i=1
n
CMn= i =1
~
Wn
n n n

n
Wi Wi Wi CMi
Total T=
1 i =1
W1
T=2 i =1 T=n i =1 CI ( i=1 n n) /(n 1) 32
W2 Wn
Yan Liu, Department of BIE, Wright State University

School Selection Example

L F SL VT CP MC
L 1 4 3 1 3 4
F 1/4 1 7 3 1/5 1
SL 1/3 1/7 1 1/5 1/5 1/6
VT 1 1/3 5 1 1 1/3
CP 1/3 5 5 1 1 3
MC 1/4 1 6 3 1/3 1
Sum 3.17 11.48 27.00 9.20 5.73 9.50

Pairwise Comparison Matrix of Six Criterion


33
Yan Liu, Department of BIE, Wright State University

School Selection Example


Approximated Eigenvector
L F SL VT CP MC Weight Weight
1/3.17 4/11.48 3/27.00 1/9.20 3/5.73 4/9.50
L
=0.32 =0.35 =0.11 =0.11 =0.52 =0.42 0.30 0.32
(1/4)/3.17 1/11.48 7/27.00 3/9.20 (1/5)/5.73 1/9.50
F
=0.08 =0.09 =0.26 =0.33 =0.03 =0.11 0.15 0.14
(1/3)/3.17 (1/7)/11.48 1/27.00 (1/5)/9.20 (1/5)/5.73 (1/6)/9.50
SL
=0.11 =0.01 =0.04 =0.02 =0.03 =0.02 0.04 0.03
1/3.17 (1/3)/11.48 5/27.00 1/9.20 1/5.73 (1/3)/9.50
VT
=0.32 =0.03 =0.19 =0.11 =0.17 =0.04 0.14 0.13
(1/3)/3.17 5/11.48 5/27.00 1/9.20 1/5.73 3/9.50
CP
=0.11 =0.44 =0.19 =0.11 =0.17 =0.32 0.22 0.24
(1/4)/3.17 1/11.48 6/27.00 3/9.20 (1/3)/5.73 1/9.50
MC
=0.08 =0.09 =0.22 =0.33 =0.06 =0.11 0.15 0.14
34
Normalized Pairwise Comparison Matrix of Six Criterion
Yan Liu, Department of BIE, Wright State University

School Selection Example


Approximated Consistency
L F SL VT CP MC Weight Measure
L 1 4 3 1 3 4 0.30 8.07
F 1/4 1 7 3 1/5 1 0.15 7.46
SL 1/3 1/7 1 1/5 1/5 1/6 0.04 6.46
VT 1 1/3 5 1 1 1/3 0.14 6.86
CP 1/3 5 5 1 1 3 0.22 8.45
MC 1/4 1 6 3 1/3 1 0.15 7.39
(10.30+40.15+30.04+10.14+30.22+40.15)
CM1= 0.30 = 8.07
(1/ 40.30+10.15+70.04+30.14+1/ 50.22+10.15)
CM2=
0.15 = 7.46
Likewise, we can calculate CM3, CM4, CM5, and CM6.
( 8.07 + 7.46 + 6.46 + 6.86 + 8.45 + 7.39 )
CI [ 6 6] /(6 1) = 0.290 CI =0.284 from eigenvalue35
Yan Liu, Department of BIE, Wright State University

School Selection Example

A B C Approximated
A B C Weight
1/6 (1/3)/1.67 (1/2)/4.5
A 1 1/3 1/2 A 0.16
=0.17 =0.20 =0.11
B 3 1 3 3/6 1/1.67 3/4.5
B
=0.50 =0.60 =0.67 0.59
C 2 1/3 1
Sum 6 1.67 4.5 2/6 (1/3)/1.67 1/4.5
C 0.25
=0.33 =0.20 =0.22
Pairwise Comparison Matrix of
Normalized Pairwise Comparison
Three Alternative Schools With
Matrix of Three Alternative
Respect to Learning
Schools With Respect to Learning

36
Yan Liu, Department of BIE, Wright State University

School Selection Example


Approximated Consistency
A B C Weight Measure
A 1 1/3 1/2 0.16 3.01
B 3 1 3 0.59 3.08

C 2 1/3 1 0.25 3.07

(10.16 +1 / 30.59 +1 / 20.25 )


CM1= 0.16 = 3.01
( 30.16 +10.59 + 30.25 )
CM2= 0.59 = 3.08
( 20.16 +1/ 30.59 +10.25 )
CM3= 0.25 = 3.07
(3.01+3.08+3.07)
CI [ 3 3] /(3 1) = 0.027
37
Yan Liu, Department of BIE, Wright State University

Composition and Synthesis


Combine the relative importance of criteria and alternatives to obtain a global
ranking of each alternative with regards to the goal
Criteria Cj (j=1,2,,n) and
their corresponding weights
wC1 wC2 wCn
Composite impact
C1 C2 Cn
A1 wA1C1 wA1C2 wA1Cn OA1 = wC1wA1C1+wC2 wA1C2++wCnwA1Cn
A2 wA2C1 wA2C2 wA2Cn OA2 = wC1wA2C1+wC2 wA2C2++wCnwA2Cn

Am wAmC1 wAmC2 wAmCn OAm = wC1wAmC1+wC2 wAmC2++wCnwCnAm
Weights of alternatives Ai (i=1,2,,m)
38
regarding criteria Cj (j=1,2,,n)
Yan Liu, Department of BIE, Wright State University

School Selection Example

0.32 0.14 0.03 0.13 0.24 0.14 Composite


Impact of
L F SL VT CP MC Schools
A 0.16 0.33 0.45 0.77 0.25 0.69 0.37
B 0.59 0.33 0.09 0.05 0.5 0.09 0.38
C 0.25 0.33 0.46 0.17 0.25 0.22 0.25
OA=0.32 0.16 + 0.14 0.33 + 0.03 0.45 + 0.13 0.77 + 0.24 0.25 + 0.14 0.69 = 0.37
OB= 0.32 0.59 + 0.14 0.33 + 0.03 0.09 + 0.13 0.05 + 0.24 0.5 + 0.14 0.09 = 0.38
OC= 0.32 0.25 + 0.14 0.33 + 0.03 0.46 + 0.13 0.17 + 0.24 0.25 + 0.14 0.22 = 0.25
In conclusion, schools A and B appear better than school C, and schools A
and B seem to be similar 39
Yan Liu, Department of BIE, Wright State University

Summary of AHP
AHP has been widely in varieties of applications
e.g. resource allocation , conflict resolution, prediction, planning, etc.
Advantages
Decision hierarchy and pairwise comparisons make the AHP process easy to comprehend
The use of a subjective scale, such as strongly preferred, rather than a quantitative scale
is particularly useful when it is difficult to formalize some criteria (attributes)
quantitatively
It is usually much easier to compare two items at a time than to compare many items all at
once
Disadvantages
The decision hierarchy in AHP assumes independence among criteria, which is not always
appropriate
The subjective scale is subject to human errors and biases
The number of pairwise comparisons becomes quite extensive when the number of
40
attributes and alternatives is large

Potrebbero piacerti anche