Vous êtes sur la page 1sur 14

Analytic Hierarchy Process (What is AHP)

Analytic Hierarchy Process (AHP) is one of Multi Criteria decision making method
that was originally developed by Prof. homas !. "aaty. #n short$ it is a method to
derive ratio scales from paired comparisons. he input can be obtained from actual
measurement such as price$ weight etc.$ or from sub%ective opinion such as
satisfaction feelings and preference. AHP allow some small inconsistency in %udgment
because human is not always consistent. he ratio scales are derived from the
principal &igen vectors and the consistency inde' is derived from the principal &igen
value.
(on)t worry if you don)t understand yet about all of those terminologies above
because the purpose of this tutorial is to e'plain that in a very simple way. *ou %ust
need to read on and at the end you will understand.
Pair+wise Comparison (What is pair-wise comparison?)
,ow let me e'plain what paired comparison is. #t is always easier to e'plain by an
e'ample. "uppose we have two fruits Apple and -anana. # would like to ask you$
which fruit you like better than the other and how much you like it in comparison with
the other. !et us make a relative scale to measure how much you like the fruit on the
left (Apple) compared to the fruit on the right (-anana).
#f you like the apple better than banana$ you thick a mark between number . and / on
left side$ while if you favor banana more than apple$ then you mark on the right side.
0or instance # strongly favor banana to apple then # give mark like this
,ow suppose you have three choices of fruits. hen the pair wise comparison goes as
the following


*ou may observe that the number of comparisons is a combination of the number of
things to be compared. "ince we have 1 ob%ects (Apple$ -anana and Cheery)$ we have
1 comparisons. able below shows the number of comparisons.
able 23 ,umber of comparisons
,umber of things . 4 1 5 6 7 2
number of comparisons 8 . 1 7 .8 .6 4.
he scaling is not necessary . to / but for 9ualitative data such as preference$ ranking
and sub%ective opinions$ it is suggested to use scale . to /.
#n the ne't section you will learn how to analy:e this paired comparisons
Making Comparison Matri' (How to make reciprocal matrix?)
-y now you know how to make paired comparisons. #n this section you will learn
how to make a reciprocal matri' from pair wise comparisons.
0or e'ample ;ohn has 1 kinds of fruits to be compared and he made sub%ective
%udgment on which fruit he likes best$ like the following



<e can make a matri' from the 1 comparisons above. -ecause we have three
comparisons$ thus we have 1 by 1 matri'. he diagonal elements of the matri' are
always . and we only need to fill up the upper triangular matri'. How to fill up the
upper triangular matri' is using the following rules3
.. If the judgment value is on the left side of 1, we put the actual judgment
value.
4. If the judgment value is on the right side of 1, we put the reciprocal value .
Comparing apple and banana$ ;ohn slightly favor banana$ thus we put in the row .
column 4 of the matri'. Comparing Apple and Cherry$ ;ohn strongly likes apple$ thus
we put actual %udgment 6 on the first row$ last column of the matri'. Comparing
banana and cherry$ banana is dominant. hus we put his actual %udgment on the
second row$ last column of the matri'. hen based on his preference values above$ we
have a reciprocal matri' like this
o fill the lower triangular matri'$ we use the reciprocal values of the upper diagonal.
#f is the element of row column of the matri'$ then the lower diagonal is filled
using this formula
hus now we have complete comparison matri'
,otice that all the element in the comparison matri' are positive$ or .
,e't section will discuss about how you will use this matri'.
Priority =ectors (How to compute Eigen Value and Eigen ector?)
Having a comparison matri'$ now we would like to compute priority vector$ which is
the normali:ed &igen vector of the matri'. #f you would like to know what the
meaning of &igen vector and &igen value is and how to compute them manually$ go to
my other tutorial and then return back here. he method that # am going to e'plain in
this section is only an appro'imation of &igen vector (and &igen value) of a reciprocal
matri'. his appro'imation is actually worked well for small matri' and there is
no guarantee that the rank will not reverse because of the appro'imation error.
,evertheless it is easy to compute because all we need to do is %ust to normali:e each
column of the matri'. At the end # will show the error of this appro'imation.
"uppose we have 1 by 1 reciprocal matri' from paired comparison
<e sum each column of the reciprocal matri' to get
hen we divide each element of the matri' with the sum of its column$ we have
normali:ed relative weight. he sum of each column is ..
he normali:ed principal &igen vector can be obtained by averaging across the rows
he normali:ed principal &igen vector is also called priority vector. "ince it is
normali:ed$ the sum of all elements in priority vector is .. he priority vector shows
relative weights among the things that we compare. #n our e'ample above$ Apple is
4>.4>?$ -anana is 75.15? and Cherry is 2.1>?. ;ohn most preferable fruit is -anana$
followed by Apple and Cheery. #n this case$ we know more than their ranking. #n fact$
the relative weight is a ratio scale that we can divide among them. 0or e'ample$ we
can say that ;ohn likes banana 4.42 (@75.15A4>.4>) times more than apple and he also
like banana so much >.24 (@75.15A2.1>) times more than cheery.
Aside from the relative weight$ we can also check the consistency of ;ohn)s answer.
o do that$ we need what is called Principal &igen value. Principal &igen value is
obtained from the summation of products between each element of &igen vector and
the sum of columns of the reciprocal matri'.
Computation and the meaning of consistency are e'plained in the ne't section.
As a note$ # put the comparison matri' into Matlab to see how different is the result of
numerical computation of &igen value and &igen vector compared to the
appro'imation above.
<e get three &igen vectors concatenated into 1 columns of matri'
he corresponding &igen values are the diagonal of matri'
he largest &igen value is called the Principal &igen value$ that is
which is very close to our appro'imation (about .? error). he
principal &igen vector is the &igen vector that corresponds to the highest &igen value.
he sum is ..58>. and the normali:ed principal &igen vector is
his result is also very close to our appro'imation
hus the appro'imation is 9uite good.
hus the sum of &igen vector is not one. <hen you normali:ed an &igen vector$ then
you get a priority vector. he sum of priority vector is one.
In next section you will learn how to make use of information of principal eige
value to measure whether the opinion is consistent.
Consistency #nde' and Consistency Batio (What is the meaning o!
consistent?)
<hat is the meaning that our opinion is consistentC How do we measure the
consistency of sub%ective %udgmentC At the end of this section will be able to answer
those 9uestions.
!et us look again on ;ohn)s %udgment that we discussed in the previous section. #s
;ohn %udgment consistent or notC



0irst he prefers -anana to Apple. hus we say that for ;ohn$ -anana has greater value
than Apple. <e write it as .
,e't$ he prefers Apple to Cherry. 0or him$ Apple has greater value than Cherry. <e
write it as .
"ince and $ logically$ we hope that or -anana must be preferable
than Cherry. his logic of preference is called transitive property. #f ;ohn answers in
the last comparison is transitive (that he like -anana more than Cherry)$ then his
%udgment is consistent. Dn the contrary$ if ;ohn prefers Cherry to -anana then his
answer is inconsistent. hus consistency is closely related to the transitive property.
A comparison matri' is said to be consistent if for all $ and .
However$ we shall not force the consistency. 0or e'ample$ has value and
has value $ we shall not insist that must have value . his too
much consistency is undesirable because we are dealing with human %udgment. o be
called consistent $ the rank can be transitive but the values of %udgment are not
necessarily forced to multiplication formula .

Prof. "aaty proved that for consistent reciprocal matri'$ the largest &igen value is
e9ual to the number of comparisons$ or . hen he gave a measure of
consistency$ called Consistency #nde' as deviation or degree of consistency using the
following formula


hus in our previous e'ample$ we have and three comparisons$ or
$ thus the consistency inde' is
Enowing the Consistency #nde'$ the ne't 9uestion is how do we use this inde'C
Again$ Prof. "aaty proposed that we use this inde' by comparing it with the
appropriate one. he appropriate Consistency inde' is called Bandom Consistency
#nde' ( ).
He randomly generated reciprocal matri' using scale $ $ F$ $ F$ >$ / (similar to
the idea of -ootstrap) and get the random consistency inde' to see if it is about .8?
or less. he average random consistency inde' of sample si:e 688 matrices is shown
in the table below
able >3 Bandom Consistency #nde' ( )
n . 4 1 5 6 7 2 > / .8
B# 8 8 8.6>8./ ...4..45..14..5...56..5/
hen$ he proposed what is called Consistency Batio$ which is a comparison between
Consistency #nde' and Bandom Consistency #nde'$ or in formula
#f the value of Consistency Batio is smaller or e9ual to .8?$ the inconsistency is
acceptable. #f the Consistency Batio is greater than .8?$ we need to revise the
sub%ective %udgment.
0or our previous e'ample$ we have and for is 8.6>$ then we
have . hus$ ;ohn)s sub%ective evaluation about his
fruit preference is consistent.
"o far$ in AHP we are only dealing with paired comparison of criteria or alternative
but not both. #n ne't section$ # show an e'ample to use both criteria and alternative in
two levels of AHP.
#llustrative e'ample (how to compute in a !ull hierarchy?)
#n this section$ # show an e'ample of two levels AHP. he structure of hierarchy in
this e'ample can be drawn as the following
!evel 8 is the goal of the analysis. !evel . is multi criteria that consist of several
factors. *ou can also add several other levels of sub criteria and sub+sub criteria but #
did not use that here. he last level (level 4 in figure above) is the alternative choices.
*ou can see again able . for several e'amples of Goals$ factors and alternative
choices. he lines between levels indicate relationship between factors$ choices and
goal. #n level . you will have one comparison matri' corresponds to pair+wise
comparisons between 5 factors with respect to the goal. hus$ the comparison matri'
of level . has si:e of 5 by 5. -ecause each choice is connected to each factor$ and you
have 1 choices and 5 factors$ then in general you will have 5 comparison matrices at
level 4. &ach of these matrices has si:e 1 by 1. However$ in this particular e'ample$
you will see that some weight of level 4 matrices are too small to contribute to overall
decision$ thus we can ignore them.
-ased on 9uestionnaire survey or your own paired comparison$ we make several
comparison matrices. Click here if you do not remember how to make a comparison
matri' from paired comparisons. "uppose we have comparison matri' at level . as
table below. he yellow color cells in upper triangular matri' indicate the parts that
you can change in the spreadsheet. he diagonal is always . and the lower triangular
matri' is filled using formula .
able /3 Paired comparison matri' level . with respect to the goal
"riteria A # " $ Priority Vector
A ..88 1.88 2.88 /.88 %&'()*
# 8.11 ..88 6.88 2.88 +)',(*
" 8..5 8.48 ..88 1.88 )'-(*
$ 8... 8..5 8.11 ..88 .'.%*
/um ..6/ 5.15 .1.11 48.88 .88.88?
@5.47/4$ C# @ 8.8>/2$ CB @ /./2? H .8? (acceptable)

he priority vector is obtained from normali:ed &igen vector of the matri'. Click here
if you do not remember how to compute priority vector and largest &igen value
from a comparison matri'. C# and CB are consistency #nde' and Consistency ratio
respectively$ as # have e'plained in previous section. 0or your clarity$ # include again
here some part of the computation3
(hus$ DE because 9uite consistent)

Bandom Consistency #nde' (B#) is obtained from able >.
"uppose you also have several comparison matrices at level 4. hese comparison
matrices are made for each choice$ with respect to each factor.
able .83 Paired comparison matri' level 4 with respect to 0actor A
Choice 0 1 2 Priority Vector
I ..88 ..88 2.88 %,'-%*
* ..88 ..88 1.88 (3')(*
J 8..5 8.11 ..88 ,-'-,*
"um 4..5 4.11 ...88 .88.88?
@1..85$ C# @ 8.86$ CB @ >./2? H .8? (acceptable)

able ..3 Paired comparison matri' level 4 with respect to 0actor -
"hoice 0 1 2 Priority Vector
0 ..88 8.48 8.68 ,,'.)*
1 6.88 ..88 6.88 &-'+3*
2 4.88 8.48 ..88 ,3'++*
/um >.88 ..58 7.68 .88.88?
@1.8>>$ C# @ 8.85$ CB @ 2.6>? H .8? (acceptable)

<e can do the same for paired comparison with respect to 0actor C and (. However$
the weight of factor C and ( are very small (look at able / again$ they are only about
/? and 6? respectively)$ therefore we can assume the effect of leaving them out from
further consideration is negligible. <e ignore these two weights as set them as :ero.
"o we do not use the paired comparison matri' level 4 with respect to 0actor C and (.
#n that case$ the weight of factor A and - in able / must be ad%usted so that the sum
still .88?
Ad%usted weight for factor A @

Ad%usted weight for factor - @

hen we compute the overall composite weight of each alternative choice based on
the weight of level . and level 4. he overall weight is %ust normali:ation of linear
combination of multiplication between weight and priority vector.

able .43 Dverall composite weight of the alternatives
4actor A 4actor # "omposite Weight
(Ad%usted) Weight 0.663 0.337
"hoice 0 6..86? ...5/? (&'&+*
"hoice 1 1>./1? 28.4>? .)'.)*
"hoice 2 .8.8.? .>.44? ,+'&3*
0or this e'ample$ we get the results that choice * is the best choice$ followed by I as
the second choice and the worst choice is J. he composite weights are ratio scale.
<e can say that choice * is 1.>2 times more preferable than choice J$ and choice * is
..1 times more preferable than choice I.
<e can also check the overall consistency of hierarchy by summing for all levels$
with weighted consistency inde' (C#) in the nominator and weighted random
consistency inde' (B#) in the denominator. Dverall consistency of the hierarchy in our
e'ample above is given by

(Acceptable)

4inal 5emark
-y now you have learned several introductory methods on multi criteria decision
making (MC(M) from simple cross tabulation$ using rank$ and weighted score until
AHP. Ksing Analytic Hierarchy Process (AHP)$ you can convert ordinal scale to ratio
scale and even check its consistency.