Académique Documents
Professionnel Documents
Culture Documents
1. INTRODUCTION
This paper is a presentation of selected materials from the works of Dr. Thomas L. Saaty to
highlight the similarities and differences of the Analytic Hierarchy Process (AHP) and the
Analytic Network Process (ANP). Some examples in prediction and estimation are included to
illustrate ANP. It is assumed that the readers are familiar with the basics of AHP.
2. THE SEVEN PILLARS OF THE AHP
Dr. Saaty presented his paper titled "The Seven Pillars of the Analytic Hierarchy Process" at the
Second International Symposium on the Analytic Hierarchy Process (1999) in Pittsburgh. He
described the seven pillars as follows:
a. Ratio scales.
Ratio scales (normalized) are central to the generation and synthesis of priorities in any
multicriteria method. In addition, ratio scales are the only way to generalize a decision theory to
the case of dependence and feedback. Ratio scales can also be used to make decisions involving
several hierarchies such as in selecting strategies based on benefits, costs, opportunities, and
risks.
b. Reciprocal paired comparisons.
Priorities of the alternatives are derived from the set of pairwise comparison judgments. There
are at least three modes for arriving at a ranking of the alternatives:
Relative, which ranks a few alternatives by comparing them in pairs with respect to a
common property.
Absolute, which rates an unlimited number of alternatives one at a time on intensity scale
constructed separately for each covering criterion.
Benchmarking, which ranks alternatives by including a known alternative in the group and
comparing the other against it.
c. Conditions for sensitivity of the principal right eigenvector to changes in judgments.
Sensitivity of the eigenvector to perturbation in judgments limits the number of elements in each
set of comparisons to a few and requires that they be homogeneous. The comparisons must be by
selecting the small element as a unit and ask how much more the larger element is.
d. Homogeneity and clustering.
Clustering is used, when the difference of the elements are more than one order of magnitude, to
extend the fundamental scale gradually, eventually enlarging the scale from 1-9 to 1-.
Keynote - 1
Kirti Peniwati
Linear Hierarchy
Goal
Criteria
component,
cluster
(Level)
Subcriteria
element
Alternatives
A loop indicates that each
element depends only on itself.
The elements of a decision problem are assigned to levels depending on their concreteness,
controllability and certainty. Ordinarily the narrower and more concrete are the properties the
lower down is the level to which they belong. The less concrete and more general, less
controllable, more uncertain and risky are the properties or elements, the higher up is the level to
which they belong. In a hierarchy the elements in each level are influenced or controlled by the
elements in the level immediately above. Influence is distributed downwards from the top, with
the goal having the greatest influence or importance. It has a value of one. This value of one is
divided among the elements of the second level, and the values of each of these in turn is divided
among those of the third level, and so on down to the level of alternatives of the decision at the
bottom.
Keynote - 2
There is a more general way to structure a decision problem involving functional dependence. It
consists of components which in turn consist of elements and feedback is allowed between
components. A hierarchy is a special case of such a network. In both hierarchies and networks the
elements in a component may be dependent on each other. Figure 2 below shows the structure of
a network. In this figure, a loop means that there is inner dependence of elements within a
component.
C4
C1
C2
Feedback
A cycle between C
2 and C3
C3
A loop in a component indicates inner dependence of the elements in that
component with respect to a common property.
Keynote - 3
Kirti Peniwati
In te rm e diate
c om p o ne nt
(Tra ns ien t S ta te)
S ou rc e C om po ne nt
C1
C2
O ute rd e pe nd e nc e
Inte rm ed iate
C o m p on en t
(R e cu rre nt S ta te )
S in k C o m p on e nt
(Ab so rbing S ta te )
C3
C5
In te rm e diate
C om p o ne nt
(R e cu rre n t S ta te)
In ne r d e pe nd e nc e lo op
C4
Figure 3 Connections in a Network
If the criteria or attributes in a hierarchy cannot be compared with respect to an overall objective
because of lack of information about them, they can instead be compared in terms of the
alternatives in the level below them by examining each alternative and asking which criterion or
attribute is perceived to be more important for the overall integrity well being or behavior of that
alternative. The alternatives would naturally be also compared in terms of each attribute as is
normally done. The result is a system of two interdependent components.
A holarchy, illustrated in Figure 4 is a hierarchy of two or more levels in which the goal is
eliminated and what was the second level that used to depend on the goal, now depends on the
bottom level of alternatives, thus as a whole the hierarchy is a cycle of successively dependent
levels. We have encountered such a form in the analysis of the turn around of the US economy in
which the importance of the primary factors is determined in terms of the time periods.
Figure 4 The U.S. Holarchy of Factors for Forecasting Turnaround in Economic Stagnation
The structure of a multiple networks is shown in Figure 5 below.
Keynote - 4
Goal
(Control
Hierarchy)
Strategic Criteria
BOCR Rating
ANP model
Of Costs
ANP model
Of Opportunities
ANP model
Of Benefits
ANP model
Of Risks
5. SUPERMATRIX OF ANP
With AHP judgments are presented in a set of matrices, from each an eigenvector is derived to
obtain a set of local priority scales. The outcome (priority of the alternatives) is obtained by
systematic synthesizing the set of eigenvectors from the top down the hierarchy. ANP uses a
supermatrix of eigenvectors as shown in Figure 6, of which AHP has its special form as shown in
Figure 7.
C1
e11e12
C1
e11
e12
e1n
W=
C2
CN
e1n
e21e22
CN
e2n
eN1eN2
eNn
W11
W12
W1N
W21
W22
W2N
WN1
WN2
WNN
e21
e22
e2n
C2
eN1
eN2
eNn
0 0 K
0
0 0 K
0
W 21
0 W 32 0 K
0
W=
M M M
M
M
M
K W n-1, n-2
0 0 K
W n,n-1
0
Figure 7 The Structure and Supermatrix of Hierarchy
Keynote - 5
0
0
0
Kirti Peniwati
Marketing Mix
Competitors
Customer Group
Contemporary
Issues
Time
Horizon
Indirect
Competitors
Traits
Public Health
Figure 9 shows the outcomes of the same market share estimation using three different structures.
ANP gives the most accurate outcome.
Market
Share
McDonald
Burger King
Wendys
Simple
Hierarchy
Complex
Hierarchy
Network
Actual
0.4640
0.2305
0.3055
0.5427
0.2689
0.1884
0.5603
0.2778
0.1621
0.5823
0.2857
0.1320
Keynote - 6
Figure 11 below is an ANP model of predicting market share of airlines, which also produced a
close estimation of the actual market share as shown in Figure 12. Notice that the model uses
internal elements which can be observed and experienced directly by their customers. This is
usually true for such models with ANP.
Keynote - 7
Kirti Peniwati
Airlines
American
United
Delta
Northwest
Continental
US Airways
Southwest
American West
Model Estimate
23.9
18.7
18.0
11.4
9.3
7.5
5.9
4.4
24.0
19.7
18.0
12.4
10.0
7.1
6.4
2.9
[1]
[2]
[3]
[4]
Keynote - 8