SUMMER SCHOOL

19-21 September 2007

Gaeta (LT)

The Analytic Hierarchy Process (AHP)

for Decision Making

PROF. THOMAS L. SAATY

(University Of Pittsburgh – Pennsylvania – USA)
The Analytic Hierarchy
Process (AHP)
for
Decision Making

Decision Making involves

setting priorities and the AHP
is the methodology for doing
that.
 Decision Making
Creative Thinking
1) Precedes Creative Thinking to
Precedes Decision Making to
determine which area to brainstorm first
brainstorm, connect and organize
2) It also follows Creative Thinking to
the criteria needed to
determine which avenues to specialize
make a decision
in first for implementation needs

Problem Solving
Creative Thinking and Decision
Making are the corner stones of
Problem Solving
 Decision Making
Present and likely future?
Creative Thinking
Problem to solve? What to brainstorm?
Opportunity to seize? How to structure?

Best action?
Resource allocation?

Solution Strategy

Problem Solving Opportunity Finding

Remove obstacles Help to add value

Implementation
• Design, Planning and Action
• Physical (engineering) and Behavioral
(management)
 A Simple Example of Ranking
You have five people ranked on three
criteria: age (in years), wealth (in dollars)
and health (in relative priorities), how do
you determine their overall ranks. Assume,
in addition for this exercise, that the
younger a person is the more favorable it is
for his/her rank with respect to age.
 Part 1: Decision Making Challenges

• Current State of Corporate Decision Making

• Inefficient Meetings
• Consequences
• Benefits From a Structured Process for Decision Making
• Benefits of SuperDecisions (free!) software
 Current State of Corporate Decision Making
• At least 50% of all decisions end in failure.
• 33% of all decisions made are never implemented.
• 50% of decisions implemented are discontinued after 2
years.
• 66% of decisions are based on failure prone methods.
• Decisions using high participation succeed 80% of the
time but occurs only 20% of the time.
• Practically every decision failure is preventable

Source: Why Decisions Fail - Author Paul Nut - Publisher; Berret & Koehler 2002
20 year study of over 400 business decisions from Public, Private and Not-
For Profit organizations in the United States, Canada & Europe.
 Current State of Corporate Decision Making
• Only 37% of decision makers said they have a
clear understanding of what their organization
is trying to achieve.
• Only 1 in 5 was enthusiastic about their team’s
and organization’s goals.
• Only 1 in 5 said they have a clear line of site
between their tasks and their organization’s
goals.
SOURCE: Harris Poll of 23,000 respondents
 Inefficient Meetings
• 11 Million meetings in the U.S. per day
• Most professionals attend a total of 61.8 meetings per
month
• Research indicates that over 50 percent of this
meeting time is wasted
• Professionals lose 31 hours per month in
unproductive meetings, or approximately four work
days
• A network MCI Conferencing White Paper. Meetings in America: A study of trends, costs and attitudes toward business
travel, teleconferencing, and their impact on productivity (Greenwich, CT: INFOCOMM, 1998), 3.
• Robert B. Nelson and Peter Economy, Better Business Meetings (Burr Ridge, IL: Irwin Inc, 1995), 5.
 Benefits from Structured Decision Making
• Increase the success of decisions made by at least 50%
• Increase benefits by blending diverse perspectives
• Achieve and sustain more excellent operational performance
• Maximize return on investments in business change
• Increase business leadership and stakeholder confidence and
commitment
• Achieve more consistent business performance
• Sustain increasing business prosperity

Source: Why Decisions Fail - Author Paul Nut - Publisher; Berret & Koehler 2002
20 year study of over 400 business decisions from Public, Private and Not-For Profit
organizations in the United States, Canada & Europe.
 Benefits of Systematic Decision Making

• Rapidly build consensus on goals, objectives and

priorities
• Align work activities to what is important for
organizational success
• Objectives based budgeting versus resource based
budgeting (save the salami and peanut butter for
lunch)
• Get past corporate stovepipes (tear down the walls)
Benefits of Systematic Decision Making (cont’d)

• Improve speed and effectiveness of decision

making
• Synthesize existing corporate data with
management priorities
• Achieve buy-in on major corporate decisions
• Improve accountability and outcomes over time
• Achieve cost savings through efficient allocation
of resources
Real Life Problems Exhibit:

Strong Pressures
and Weakened Resources

Complex Issues - Sometimes

There are No “Right” Answers

Vested Interests

Conflicting Values
Most Decision Problems are Multicriteria

• Maximize profits
• Satisfy customer demands
• Maximize employee satisfaction
• Satisfy shareholders
• Minimize costs of production
• Satisfy government regulations
• Minimize taxes
• Maximize bonuses
 Knowledge is Not in the Numbers
Isabel Garuti is an environmental researcher whose father-in-law is a master chef
in Santiago, Chile. He owns a well known Italian restaurant called Valerio. He
is recognized as the best cook in Santiago. Isabel had eaten a favorite dish
risotto ai funghi, rice with mushrooms, many times and loved it so much that she
wanted to learn to cook it herself for her husband, Valerio’s son, Claudio. So
she armed herself with a pencil and paper, went to the restaurant and begged
Valerio to spell out the details of the recipe in an easy way for her. He said it
was very easy. When he revealed how much was needed for each ingredient, he
said you use a little of this and a handful of that. When it is O.K. it is O.K. and it
smells good. No matter how hard she tried to translate his comments to
numbers, neither she nor he could do it. She could not replicate his dish.
Valerio knew what he knew well. It was registered in his mind, this could not
be written down and communicated to someone else. An unintelligent
observer would claim that he did not know how to cook, because if he did, he
should be able to communicate it to others. But he could and is one of the best.
 Knowing Less, Understanding More

You don’t need to know everything to get to

the answer.

Expert after expert missed the revolutionary

significance of what Darwin had collected.
Darwin, who knew less, somehow understood
more.
 Aren’t Numbers Numbers?
We have the habit to crunch numbers
whatever they are
An elderly couple looking for a town to which they
might retire found Summerland, in Santa Barbara
County, California, where a sign post read:

Summerland
Population 3001
Feet Above Sea Level 208
Year Established 1870
Total 5079

“Let’s settle here where there is a sense of humor,” said

the wife; and they did.
 Do Numbers Have an Objective Meaning?
Butter: 1, 2,…, 10 lbs.; 1,2,…, 100 tons

Sheep: 2 sheep (1 big, 1 little)

Temperature: 30 degrees Fahrenheit to New Yorker, Kenyan, Eskimo

Since we deal with Non-Unique Scales such as [lbs., kgs], [yds,

meters], [Fahr., Celsius] and such scales cannot be combined, we need
the idea of PRIORITY.

PRIORITY becomes an abstract unit valid across all scales.

A priority scale based on preference is the AHP way to standardize

non-unique scales in order to combine multiple criteria.
 Stability, Objectivity, and Measurement
We deal with many things whose stimuli are qualitative and change
according to our state of mind and our judgment about them is variable
and unstable and subjective. For some things we have created scales of
measurement that enable us to get the same readings at any time no
matter who we are. But the readings have to be interpreted as to their
significance, importance or priority. We call such things tangibles.

How can we deal with intangibles as we experience them directly like

the color red from sense data, beauty as we see a work of art and
respond to it using experience, or goodness and love that are abstract
modes of behavior? We can compare one shade of red with another,
one object of beauty with another, and one act of goodness or love with
another according to our preference or sense of importance or likelihood
of occurrence. The comparisons yield relative magnitudes of dominance
that are stable at the time we do them and we can also compare them
with what other people get in their evaluation.
Nonmonotonic Relative Nature of Absolute Scales

Good for 100 Bad for

preserving food comfort

Bad for Good for

preserving food comfort

Good for 0 Bad for

preserving food comfort

Temperature
 OBJECTIVITY!?

Bias in upbring: objectivity is agreed upon subjectivity.

We interpret and shape the world in our own image. We
pass it along as fact. In the end it is all obsoleted by the
next generation.

Logic breaks down: Russell-Whitehead Principia; Gödel’s

Undecidability Proof.

Intuition breaks down: circle around earth; milk and coffee.

How do we manage?
 Organizing the Ways to Access,
Connect, and Prioritize Information
• We cannot use randomly gathered information very well without
thinking in organized ways. The best way to use the plethora of
sporadic, randomly acquired information in our memories is to create
well-organized complex structures and represent the basic concepts
with criteria and alternatives with which such information deals and
then do prioritization to determine what is important and what is not.
We quickly learn that we know much to use in predicting what
happens in the real world than we can do without such systematic ways
to represent understanding. We have no better ways to do it. The
modern computer is a gift that we can use to deal with complexity to
enhance the power of our minds. Without them it is extremely difficult
to mange complexity in the integrated ways that now we can. It is a
new revolution for better decision making.
 Making a Decision

Widget B is cheaper than Widget A

Widget A is better than Widget B

Which Widget would you choose?

 Basic Decision Problem

Criteria: Low Cost > Operating Cost > Style

Car: A B B
V V V
Alternatives: B A A

Suppose the criteria are preferred in the order shown and the
cars are preferred as shown for each criterion. Which car
should be chosen? It is desirable to know the strengths of
preferences for tradeoffs.
To understand the world we assume that:

We can describe it

We can define relations between

its parts and

We can apply judgment to relate the

parts according to

a goal or purpose that we

have in mind.
 Hierarchic
GOAL
Thinking

CRITERIA

ALTERNATIVES
 Relative Measurement
The Process of Prioritization
In relative measurement a preference, judgment
is expressed on each pair of elements with respect
to a common property they share.

In practice this means that a pair of elements

in a level of the hierarchy are compared with
respect to parent elements to which they relate
in the level above.
 Relative Measurement (cont.)

If, for example, we are comparing two apples

according to size we ask:

• Which apple is bigger?

• How much bigger is the larger than the smaller apple?

Use the smaller as the unit and estimate how
many more times bigger is the larger one.

• The apples must be relatively close (homogeneous)

if we hope to make an accurate estimate.
 Relative Measurement (cont.)

•The Smaller apple then has the reciprocal value when

compared with the larger one. There is no way to escape this sort
of reciprocal comparison when developing judgments
•If the elements being compared are not all homogeneous, they are
placed into homogeneous groups of gradually increasing relative
sizes (homogeneous clusters of homogeneous elements).
• Judgments are made on the elements in one group of small
elements, and a “pivot” element is borrowed and placed in the next
larger group and its elements are compared. This use of pivot
elements enables one to successively merge the measurements of
all the elements. Thus homogeneity serves to enhance the accuracy
of measurement.
 Pairwise Comparisons
Size

Apple A Apple B Apple C

Size Apple A Apple B Apple C
Comparison

Apple A S1/S1 S1/S2 S1/S3

Apple B S2/S1 S2/S2 S2/S3

Apple C S3/S1 S3/S2 S3/S3

We Assess The Relative Sizes of the

Apples By Forming Ratios
 Pairwise Comparisons
Size

Apple A Apple B Apple C

Size Apple A Apple B Apple C
Comparison
Resulting Relative Size
Priority of Apple
Eigenvector

Apple A 1 2 6 6/10 A

Apple B 1/2 1 3 3/10 B

Apple C 1/6 1/3 1 1/10 C

When the judgments are consistent, as they are here, any
normalized column gives the priorities.
 Consistency (cont.)
• Consistency itself is a necessary condition for a better
understanding of relations in the world but it is not
sufficient. For example we could judge all three of
the apples to be the same size and we would be perfectly
consistent, but very wrong.

• We also need to improve our validity by using redundant

information.

• It is fortunate that the mind is not programmed to be always

consistent. Otherwise, it could not integrate new information
by changing old relations.
 Consistency
In this example Apple B is 3 times larger than Apple C.
We can obtain this value directly from the comparisons
of Apple A with Apples B & C as 6/2 = 3. But if we
were to use judgment we may have guessed it as 4. In
that case we would have been inconsistent.

Now guessing it as 4 is not as bad as guessing it as 5 or

more. The farther we are from the true value the more
inconsistent we are. The AHP provides a theory for
checking the inconsistency throughout the matrix and
allowing a certain level of overall inconsistency but not
more.
 Consistency (cont.)
Because the world of experience is vast and we deal with it in pieces according to
whatever goals concern us at the time, our judgments can never be perfectly
precise.

It may be impossible to make a consistent set of judgments on some pieces that

make them fit exactly with another consistent set of judgments on other related
pieces. So we may neither be able to be perfectly consistent nor want to be.

We must allow for a modicum of inconsistency.

 Consistency (cont.)
How Much Inconsistency to Tolerate?
• Inconsistency arises from the need for redundancy.
• Redundancy improves the validity of the information about the real world.
• Inconsistency is important for modifying our consistent understanding, but it must not be too large
to make information seem chaotic.
• Yet inconsistency cannot be negligible; otherwise, we would be like robots unable to change our
minds.
• Mathematically the measurement of consistency should allow for inconsistency of no more than
one order of magnitude smaller than consistency. Thus, an inconsistency of no more than 10%
can be tolerated.
• This would allow variations in the measurement of the elements being compared without
destroying their identity.
• As a result the number of elements compared must be small, i.e. seven plus or minus two. Being
homogeneous they would then each receive about ten to 15 percent of the total relative value in the
vector of priorities.
• A small inconsistency would change that value by a small amount and their true relative value
would still be sufficiently large to preserve that value.
• Note that if the number of elements in a comparison is large, for example 100, each would receive
a 1% relative value and the small inconsistency of 1% in its measurement would change its value
to 2% which is far from its true value of 1%.
 Pairwise Comparisons using Inconsistent Judgments

Paris London New York

Nicer ambience
comparisons
Normalized Ideal

Paris 1 2 5 0.5815 1

London 1/2 1 3 0.3090 0.5328

New York 1/5 1/3 1 0.1095 0.1888

Pairwise Comparisons using Judgments and the Derived Priorities

B. Clinton M. Tatcher G. Bush

Politician
Normalized Total
comparisons

B. Clinton
1 3 7 0.6220 1

M. Tatcher
1/3 1 5 0.2673 0.4297

G. Bush 1/7 1/5 1 0.1107 0.1780

When the judgments are consistent, we
have two ways to get the answer:
1. By adding any column and dividing each entry by the
total, that is by normalizing the column, any column
gives the same result. A quick test of consistency if all
the columns give the same answer.
2. By adding the rows and normalizing the result.

When the judgments are inconsistent we

have two ways to get the answer:
1. An approximate way: By normalizing each column,
forming the row sums and then normalizing the result.
2. The exact way: By raising the matrix to powers and
normalizing its row sums
 Comparison of Intangibles

The same procedure as we use for size can be used to

compare things with intangible properties. For example,
we could also compare the apples for:

• TASTE
• AROMA
• RIPENESS
 The Analytic Hierarchy Process (AHP)
is the Method of Prioritization
• AHP captures priorities from paired comparison judgments of the
• elements of the decision with respect to each of their parent criteria.

• Paired comparison judgments can be arranged in a matrix.

• Priorities are derived from the matrix as its principal eigenvector,

• which defines an absolute scale. Thus, the eigenvector is an intrinsic
• concept of a correct prioritization process. It also allows for the
• measurement of inconsistency in judgment.

• Priorities derived this way satisfy the property of an absolute scale.

 Decision Making
We need to prioritize both tangible and intangible criteria:

♦ In most decisions, intangibles such as

• political factors and
• social factors
take precedence over tangibles such as
• economic factors and
• technical factors
♦ It is not the precision of measurement on a particular factor
that determines the validity of a decision, but the importance
we attach to the factors involved.

♦ How do we assign importance to all the factors and synthesize

this diverse information to make the best decision?
 There are Primitive People Living in
our Time Who can’t Count
• I have to tell you something. I did a small research in Papua with the local tribe. These
people still live literally in a stone age environment. They do not wear clothes, and
many live in the trees and practice cannibalism. They don't speak the Indonesian
language. It was so funny that people from the central bank who accompanied me to
this region noticed and took a picture where I am holding my laptop interviewing these
people using the AHP approach....a combination of stone age respondents with modern
technology!!!!. I will send you the photos when I return to Indonesia next month. For
this research the AHP questions had to be designed is such that they are very simple yet
straight to the most critical points that we want to know about the livelihood of this
society. It was a fascinating experience. Those who accompanied me were so impressed
with the power of the AHP in conducting research on such a unique society. According
to the central bank people who paid attention to my interviews: "....the outcome is clear
and quantified, and yet the respondents do not need to have any knowledge about any
numbers and their corresponding meaning. What it needs are just their honest
perceptions and expressions.” They hardly know how to count. In fact they prefer to have 100
rupiah than 1000 rupiah simply because the color of the money (red) is what they are used to hold.
The extra zero on the bill does not mean much to them.
• Love,
• Iwan (Iwan Y. Azis is Indonesia’s greatest economist and is Professor at Cornell University)
Verbal Expressions for Making
Pairwise Comparison Judgments

Equal importance
Moderate importance of one over another
Strong or essential importance
Very strong or demonstrated importance
Extreme importance
Fundamental Scale of Absolute Numbers
Corresponding to Verbal Comparisons
1 Equal importance
3 Moderate importance of one over another
5 Strong or essential importance
7 Very strong or demonstrated importance
9 Extreme importance
2,4,6,8 Intermediate values
Use Reciprocals for Inverse Comparisons
0.05
0.47

0.10

0.15 0.24
 Which Drink is Consumed More in the U.S.?
An Example of Estimation Using Judgments
Drink
Consumption
in the U.S. Coffee Wine Tea Beer Sodas Milk Water

Coffee 1 9 5 2 1 1 1/2
Wine 1/9 1 1/3 1/9 1/9 1/9 1/9
Tea 1/5 2 1 1/3 1/4 1/3 1/9
Beer 1/2 9 3 1 1/2 1 1/3
Sodas 1 9 4 2 1 2 1/2
Milk 1 9 3 1 1/2 1 1/3
Water 2 9 9 3 2 3 1

The derived scale based on the judgments in the matrix is:

Coffee Wine Tea Beer Sodas Milk Water
.177 .019 .042 .116 .190 .129 .327
with a consistency ratio of .022.
The actual consumption (from statistical sources) is:
.180 .010 .040 .120 .180 .140 .330
 Estimating which Food has more Protein

Food Consumption
in the U.S. A B C D E F G

A: Steak 1 9 9 6 4 5 1
B: Potatoes 1 1 1/2 1/4 1/3 1/4
C: Apples 1 1/3 1/3 1/5 1/9
D: Soybean 1 1/2 1 1/6
E: Whole Wheat Bread
(Reciprocals) 1 3 1/3
F: Tasty Cake 1 1/5
G: Fish 1

The resulting derived scale and the actual values are shown below:
Steak Potatoes Apples Soybean W. Bread T. Cake Fish
Derived .345 .031 .030 .065 .124 .078 .328
Actual .370 .040 .000 .070 .110 .090 .320
(Derived scale has a consistency ratio of .028.)
 WEIGHT COMPARISONS

Weight Radio Typewriter Large Projector Small Eigenvector Actual

Attache Attache Relative
Case Weights
Radio 1 1/5 1/3 1/4 4 0.09 0.10
Typewriter 5 1 2 2 8 0.40 0.39

Large 3 1/2 1 1/2 4 0.18 0.20

Attache
Case
Projector 4 1/2 2 1 7 0.29 0.27
Small 1/4 1/8 1/4 1/7 1 0.04 0.04
Attache
Case
 DISTANCE COMPARISONS

Comparison Cairo Tokyo Chicago San London Montreal Eigen- Distance to Relative
of Distances Francisco vector Philadelph Distance
from ia in miles
Philadelphia
Cairo 1 1/2 8 3 3 7 0.263 5,729 0.278
Tokyo 3 1 9 3 3 9 0.397 7,449 0.361
Chicago 1/8 1/9 1 1/6 1/5 2 0.033 660 0.032
San 1/3 1/3 6 1 1/3 6 0.116 2,732 0.132
Francisco
London 1/3 1/3 5 3 1 6 0.164 3,658 0.177
Montreal 1/7 1/9 1/2 1/6 1/6 1 0.027 400 0.019
 Relative Electricity Consumption (Kilowatt Hours) of Household Appliances
Annual
Electric Actual
Consumption Elec. Dish Hair
Refrig TV Iron Radio Eigen-vector Relative
Range Wash Dryer
Weights

Electric
1 2 5 8 7 9 9 .393 .392
Range

Refrig-
1/2 1 4 5 5 7 9 .261 .242
erator

TV 1/5 1/4 1 2 5 6 8 .131 .167

Dish-
1/8 1/5 1/2 1 4 9 9 .110 .120
washer

Iron 1/7 1/5 1/5 1/4 1 5 9 .061 .047

Radio 1/9 1/7 1/6 1/9 1/5 1 5 .028 .028

Hair-dryer 1/9 1/9 1/8 1/9 1/9 1/5 1 .016 .003

 Relative coin sizes
.

Dime Quarter 5Cents Priorities Size in mm2 Actual

relative
size
Cent 1.1 2 1.5 0.182 283.38 0.212
Dime 2.1 1.6 0.171 254.34 0.190

Quarter 1.6 0.382 452.16 0.338

5Cent 0.263 346.18 0.259

Percentage of Google Yahoo MSN AOL My Web Priorities Actual
Individuals that use Percentage
different search
engines

Google 1.000 1.811 7.057 8.491 9.000 0.463 0.485

Yahoo 0.552 1.000 6.566 8.073 9.000 0.367 0.225

MSN 0.142 0.152 1.000 4.264 4.076 0.101 0.107

AOL 0.118 0.124 0.235 1.000 2.943 0.048 0.066

My Web 0.111 0.111 0.245 0.340 1.000 0.030 0.027

 Relative Distances from Pittsburgh

Distance L.A. New St. Louis Washington Priorities Actual Relative

From Pittsburgh Orleans D.C. Distance Values
In Miles
L.A. 1 3 5 7 .558 2446 .556

New Orleans 1/3 1 3 6 .267 1049 .249

St. Louis 1/5 1/3 1 4 .125 610 .139

Washington 1/7 1/6 1/4 1 .049 247 .056

D.C.
 Model Mercedez - E BMW - 5 Acura - TL Lexus - ES Audi - A6 Priorities Actual Cost Relative Values

Mercedez - E 1 1 1.6 1.3 1.25 0.237128 52000 0.245283019

BMW - 5 1 1.5 1.25 1.2 0.230349 48000 0.226415094

Acura - TL 1 1.3 2 0.151647 30000 0.141509434

Lexus - ES 1 1 0.187402 40000 0.188679245

Audi - A6 1 0.193475 42000 0.198113208

 Nonlinearity of the Priorities
RELATIVE VISUAL BRIGHTNESS-I

C1 C2 C3 C4
C1 1 5 6 7

C2 1/5 1 4 6

C3 1/6 1/4 1 4

C4 1/7 1/6 1/4 1

RELATIVE VISUAL BRIGHTNESS -II

C1 C2 C3 C4
C1 1 4 6 7

C2 1/4 1 3 4

C3 1/6 1/3 1 2

C4 1/7 1/4 1/2 1

 RELATIVE BRIGHTNESS EIGENVECTOR
The Inverse Square Law of Optics
I II
C1 .62 .63

C2 .23 .22

C3 .10 .09

C4 .05 .06

Square of Reciprocal
Normalized normalized of previous Normalized
Distance distance distance column reciprocal

9 0.123 0.015 67 0.61

15 0.205 0.042 24 0.22
21 0.288 0.083 12 0.11
28 0.384 0.148 7 0.06
 There are two Kinds of
Fundamentally Different Numbers
The behavior and arithmetic of such numbers are different

Numbers Derived from Numbers Derived from

Measurements; Metric Comparisons; Order Topology
Topology-closeness (Dominance Numbers)
(Scale Numbers) The numerical values of such numbers
Numbers that have the same depend on what other things a given
value no matter how many other thing is compared with. They are often
things there are referred to as relative numbers. The
ratio of two absolute numbers or ratio
Counting Measuring scale numbers is absolute. So is the
Absolute Scales Ratio Scale ratio of interval scale differences.
Interval Scale The lesser of two elements is used as
Need a unit except for a unit in each comparison. No unique
absolute scales defined unit for the derived scale which
in terms of equivalence defines an absolute scale in relative
classes form like probabilities.
 Meaning and Usefulness
Dominance Numbers are:
Scale Numbers are: • Abstract (dependent and
• Concrete (independent changeable)
from what and how many • Perception oriented
objects are involved and • “Subjective”
therefore unchangeable)
• Object oriented
• “Objective”
Measurement in physics is concrete because its objects are
tangible and independent, Dominance numbers are essential to
make decisions relative to human values that are abstract and
intangible. It is necessary to compare criteria to obtain their
priorities.
One can create dominance numbers from scale numbers but one
cannot create scale numbers from dominance numbers because
they are changeable and depend on the objects being compared.
Creating dominance numbers from ratios of measurement is not
always meaningful.
 Extending the 1-9 Scale to 1- ∞
•The 1-9 AHP scale does not limit us if we know how
to use clustering of similar objects in each group and
use the largest element in a group as the smallest one in
the next one. It serves as a pivot to connect the two.

•We then compare the elements in each group on the 1-

9 scale get the priorities, then divide by the weight of
the pivot in that group and multiply by its weight from
the previous group. We can then combine all the
groups measurements as in the following example
comparing a very small cherry tomato with a very large
watermelon.
 .07 .28 .65
Unripe Cherry Tomato Small Green Tomato Lime

.08 .22 .70

Lime Grapefruit Honeydew
.08 .22 .70
=1 = 2.75 = 8.75
.08 .08 .08
1 x .6 5 = .6 5 2.75 × .65 = 1.79 8.75x.65 = 5.69

.10 .30 .60

Honeydew Sugar Baby Watermelon Oblong Watermelon
.10 .30 .60
=1 =3 =6
.10 .10 .10
1× 5.69 = 5.69 3 × 5.69 = 17.07 6 × 5.69 = 34.14
This means that 34.14/.07 = 487.7 unripe cherry tomatoes are equal to the oblong watermelon
 Clustering & Comparison
Color
How intensely more green is X than Y relative to its size?

Honeydew Unripe Grapefruit Unripe Cherry Tomato

Unripe Cherry Tomato Oblong Watermelon Small Green Tomato

Small Green Tomato Sugar Baby Watermelon Large Lime

64
Antares is the 15th brightest star in the
sky.It is more than 1000 light years away.
Comparing a Dog-Catcher w/ President
Comparing a Dog-Catcher w/ President
 Goal
Satisfaction with School

Learning Friends School Vocational College Music

Life Training Prep. Classes

School School School

A B C
 School Selection

L F SL VT CP MC Weights
Learning 1 4 3 1 3 4 .32

Friends 1/4 1 7 3 1/5 1 .14

School Life 1/3 1/7 1 1/5 1/5 1/6 .03

Vocational Trng. 1 1/3 5 1 1 1/3 .13

College Prep. 1/3 5 5 1 1 3 .24

Music Classes 1/4 1 6 3 1/3 1 .14

 Comparison of Schools with Respect
to the Six Characteristics
Learning Priorities Friends Priorities School Life Priorities
A B C A B C A B C
A 1 1/3 1/2 .16 A 1 1 1 .33 A 1 5 1 .45

B 3 1 3 .59 B 1 1 1 .33 B 1/5 1 1/5 .09

C 2 1/3 1 .25 C 1 1 1 .33 C 1 5 1 .46

Vocational Trng. Priorities College Prep. Priorities Music Classes Priorities

A B C A B C A B C
A 1 9 7 .77 A 1 1/2 1 .25 A 1 6 4 .69

B 1/9 1 1/5 .05 B 2 1 2 .50 B 1/6 1 1/3 .09

C 1/7 5 1 .17 C 1 1/2 1 .25 C 1/4 3 1 .22

 Composition and Synthesis
Impacts of School on Criteria

.32 .14 .03 .13 .24 .14 Composite

Impact of
L F SL VT CP MC
Schools

A .16 .33 .45 .77 .25 .69 .37

B .59 .33 .09 .05 .50 .09 .38

C .25 .33 .46 .17 .25 .22 .25

 The School Example Revisited Composition & Synthesis:
Impacts of Schools on Criteria
Distributive Mode Ideal Mode
(Normalization: Dividing each (Dividing each entry by the
entry by the total in its column) maximum value in its column)
.32 .14 .03 .13 .24 .14 Composite .32 .14 .03 .13 .24 .14 Composite Normal-
L F SL VT CP MC Impact of L F SL VT CP MC Impact of ized
Schools Schools
A .16 .33 .45 .77 .25 .69 .37 A .27 1 .98 1 .50 1 .65 .34

B .59 .33 .09 .05 .50 .09 .38 B 1 1 .20 .07 .50 .13 .73 .39

C .25 .33 .46 .17 .25 .22 .25 C .42 1 1 .22 .50 .32 .50 .27

The Distributive mode is useful when the The Ideal mode is useful in choosing a best
uniqueness of an alternative affects its rank. alternative regardless of how many other
The number of copies of each alternative similar alternatives there are.
also affects the share each receives in
allocating a resource. In planning, the
scenarios considered must be comprehensive
and hence their priorities depend on how many
there are. This mode is essential for ranking
criteria and sub-criteria, and when there is
dependence.
 A Complete Hierarchy to Level of Objectives
Focus: At what level should the Dam be kept: Full or Half-Full

Decision Financial Political Env’t Protection Social Protection

Criteria:

Decision
Congress Dept. of Interior Courts State Lobbies
Makers:

Potential Archeo- Current

Irreversibility
Factors: Clout Legal Position Financial logical Financial
of the Env’t
Loss Problems Resources

Groups Farmers Recreationists Power Users Environmentalists

Affected:

Protect
Objectives: Irrigation Flood Control Flat Dam White Dam Cheap Power
Environment

Alternatives: Half-Full Dam Full Dam

 Evaluating Employees for Raises

GOAL

Dependability Education Experience Quality Attitude Leadership

(0.075) (0.200) (0.048) (0.360) (0.082) (0.235)

Outstanding Doctorate >15 years Excellent Enthused Outstanding

(0.48) .48/.48 = 1 (0.59) .59/.59 =1 (0.61) (0.64) (0.63) (0.54)

Very Good Masters 6-15 years Very Good Above Avg. Above Avg.
(0.28) .28/.48 = .58 (0.25).25/.59 =.43 (0.25) (0.21) (0.23) (0.23)
Bachelor
Good (0.11) etc. 3-5 years Good Average Average
(0.16) .16/.48 = .33 (0.10) (0.11) (0.10) (0.14)
High School
Below Avg. (0.05) 1-2 years Poor Negative Below Avg.
(0.05) .05/.48 = .10 (0.04) (0.04) (0.04) (0.06)

Unsatisfactory Unsatisfactory
(0.03) .03/.48 = .06 (0.03)
 Final Step in Absolute Measurement
Rate each employee for dependability, education, experience, quality of
work, attitude toward job, and leadership abilities.
Dependability Education Experience Quality Attitude Leadership Total Normalized
0.0746 0.2004 0.0482 0.3604 0.0816 0.2348

Esselman, T. Outstand Doctorate >15 years Excellent Enthused Outstand 1.000 0.153
Peters, T. Outstand Masters >15 years Excellent Enthused Abv. Avg. 0.752 0.115
Hayat, F. Outstand Masters >15 years V. Good Enthused Outstand 0.641 0.098
Becker, L. Outstand Bachelor 6-15 years Excellent Abv. Avg. Average 0.580 0.089
Adams, V. Good Bachelor 1-2 years Excellent Enthused Average 0.564 0.086
Kelly, S. Good Bachelor 3-5 years Excellent Average Average 0.517 0.079
Joseph, M. Blw Avg. Hi School 3-5 years Excellent Average Average 0.467 0.071
Tobias, K. Outstand Masters 3-5 years V. Good Enthused Abv. Avg. 0.466 0.071
Washington, S. V. Good Masters 3-5 years V. Good Enthused Abv. Avg. 0.435 0.066
O’Shea, K. Outstand Hi School >15 years V. Good Enthused Average 0.397 0.061
Williams, E. Outstand Masters 1-2 years V. Good Abv. Avg. Average 0.368 0.056
Golden, B. V. Good Bachelor .15 years V. Good Average Abv. Avg. 0.354 0.054

The total score is the sum of the weighted scores of the ratings. The
money for raises is allocated according to the normalized total score. In
practice different jobs need different hierarchies.
Should U.S. Sanction China? (Feb. 26, 1995)
BENEFITS

Protect rights and maintain high Incentive to Rule of Law Bring China to Help trade deficit with China
make and sell products in China (0.696) responsible free-trading 0.206) (0.098)
Yes .80 Yes .60 Yes .50
No .20 No .40 No .50
Yes 0.729 No 0.271

COSTS

$ Billion Tariffs make Chinese products Retaliation Being locked out of big infrastructure
more expensive (0.094) (0.280) buying: power stations, airports (0.626)
Yes .70 Yes .90 Yes .75
No .30 No .10 No .25
Yes 0.787 No 0.213

RISKS

Long Term negative competition Effect on human rights and Harder to justify China joining WTO
(0.683) other issues (0.200) (0.117)
Yes .70 Yes .30 Yes .50
No .30 No .70 No .50
Yes 0.597 No 0.403

Benefits .729 .271

Result: ; YES = 1.55 NO = 3.16
Costs x Risks .787 x .597 .213 x .403
 8
.
7
. .
Benefits/Costs*Risks

6 . . No

. . .. .
5
. . Yes

. . .. . . . ...
4
.. . . . ... .
3 . . . ..
2
.
1

0 6 18 30 42 54 66 78 90 102 114 126 138 150 162 174 186 198 210

Experiments
 Whom to Marry - A Compatible Spouse

Flexibility Independence Growth Challenge Commitment Humor Intelligence

Psychological Physical Socio-Cultural Philosophical Aesthetic

Communication Food Sociability World View Housekeeping

& Problem Solving
Shelter Finance
Family & Children Theology Sense of Beauty
& Intelligence
Sex Understanding
Temper

Security

Affection
CASE 1: Marry Not Marry

Loyalty

CASE 2: Campbell Graham McGuire Faucet

 Value of Yen/Dollar Exchange : Rate in 90 Days

Relative Interest Forward Exchange Official Exchange Relative Degree of Confi- Size/Direction of U.S. Past Behavior of
Rate Rate Biases Market Intervention dence in U.S. Economy Current Account Exchange Rate
.423 .023 .164 .103 Balance .252 .035

Federal Size of Bank of Forward Size of Consistent Erratic Relative Relative Relative Size of Anticipated Relevant Irrelevant
Reserve Federal Japan Rate Forward Inflation Real Political Deficit Changes
Monetary Deficit Monetary Premium/ Rate Rates Growth Stability or
Policy Policy Discount Differential Surplus
.294 .032 .097 .007 .016 .137 .027 .019 .008 .032 .032 .221 .004 .031

Tighter Contract Tighter High Premium Strong Strong Higher More More Large Decr. High High
.191 .002 .007 .002 .008 .026 .009 .013 .048 .048 .016 .090 .001 .010

Steady No Chng. Steady Medium Discount Mod. Mod. Equal Equal Equal Small No Chng. Med. Med.
.082 .009 .027 .002 .008 .100 .009 .006 .003 .022 .016 .106 .001 .010

Easier Expand Easier Low Weak Weak Lower Lower Less Incr. Low Low
.021 .021 .063 .002 .011 .009 .001 .003 .006 .025 .001 .010

Probable Impact of Each Fourth Level Factor

119.99 119.99- 134.11- 148.23- 162.35
and below 134.11 148.23 162.35 and above

Sharp Moderate No Moderate Sharp

Decline Decline Change Increase Increase
0.1330 0.2940 0.2640 0.2280 0.0820
Expected Value is 139.90 yen/$
 Best Word Processing Equipment
Focus Benefits

Criteria Time Saving Filing Quality of Document Accuracy

Features Training Service Space Printer

Required Screen Capability Quality Required Speed

Lanier Syntrex Qyx

Alternatives (.42) (.37) (.21)

Focus Costs

Criteria Capital Supplies Service Training

Lanier Syntrex Oyx

Alternatives
.54 .28 .18
 Best Word Processing Equipment Cont.

Benefit/Cost Preference Ratios

Lanier Syntrex Qyx

.42 = 0.78 .37 = 1.32 .21 = 1.17
.54 .28 .18

Best Alternative
 Group Decision Making
and the
Geometric Mean
Suppose two people compare two apples and provide the judgments for the larger
over the smaller, 4 and 3 respectively. So the judgments about the smaller relative
to the larger are 1/4 and 1/3.
Arithmetic mean
4+3=7
1/7 ≠ 1/4 + 1/3 = 7/12
Geometric mean
√ 4 x 3 = 3.46
1/ √ 4 x 3 = √ 1/4 x 1/3 = 1/ √ 4 x 3 = 1/3.46

That the Geometric Mean is the unique way to combine group judgments is a
theorem in mathematics.
0.05
0.47

0.10

0.15 0.24
 Why Is the Eigenvector Necessary
1) Consistent matrix: Aw = nw; Ak = n k −1 A.
2) Perturbed matrix: Aw = λmax w.
1 m k
3) Transitivity → lim ∑ A e / eT Ak e → w, eT = (1,...,1). By Cesaro sumability
k →∞ m
k =1

this converges to the same limit as Ak e / eT Ak e.

4) Necessity : Priority must be invariant with respect to whatever process of
synthesis one chooses. We start with an initial priority vector (1,...,1) and then
derive a first estimate of priorities from it. We then use this vector it to weight
the alternatives or weight squares of differences and derive a new priority vector,
and so on. For uniqueness we must have A o x = cx, where o indicates the
composition principle used and c indicates proportionality of the new vector cx
and the old vector x. For additive composition we have Ax = cx. More generally,
Ak x = c k x and because initially x = e, Ak e = c k e → w, the principal right eigenvector
of A. Thus it is necessary to use the eigenvector to derive a priority vector.
 A1 ... An
A1w1 w1 ... w1 wn w1 w1
    
Aw= M  M ... M   M  = n  M =nw 

Anwn w1 ... wn wn wn wn

 Let A1, A2,…, An, be a set of stimuli. The
quantified judgments on pairs of stimuli Ai, Aj, are
represented by an n-by-n matrix A = (aij), ij = 1,
2, . . ., n. The entries aij are defined by the
following entry rules. If aij = a, then aji = 1 /a,
a
0. If Ai is judged to be of equal relative
iaijji intensity to Aj then aij = 1, aji = 1, in particular, aii
= 1 for all i.
 1 a12 ... a1n 
1/ a 1 
... a2n 
A=  12

 M M M M
 
1/ a1n 1/ a2n ... 1 
How to go from
Aw=nw

to Aw=cw
and then to Aw=λmaxw
Clearly in the first formula n is a simple eigenvalue and all other
eigenvalues are equal to zero.
A forcing perurbation of eigenvalues theorem:
If λ is a simple eigenvalue of A, then for small ε > 0, there is an
eigenvalue λ(ε) of A(ε) with power series expansion in ε:
λ(ε)= λ+ ε λ(1)+ ε2 λ(2)+…
and corresponding right and left eigenvectors w (ε) and v (ε) such
that w(ε)= w+ ε w(1)+ ε2 w(2)+…
v(ε)= v+ ε v(1)+ ε2 v(2)+…
n

∑a
j=1
ij w j = λ max wi

∑w =
i=1
i 1
 1 2 1 0 λ 0 
A=  , I =  , λI =  
3 4   0 1   0 λ 
1 − λ 2 
( A − λI ) =  
 3 4 − λ 
A − λI = (1 − λ)(4 − λ) − 6 = λ2 − 5λ − 2 = 0

5 + 33
λ1 =
2
5 − 33
λ2 =
2
 n n wj
max ∑aij ≥∑aij = λmax for max wi
j =1 j =1 wi
n n wj
min ∑aij ≤∑aij = λmax for min wi
j =1 j =1 wi
n n
Thus for a row stochastic matrix we have 1=min ∑aij ≤ λmax ≤ max ∑aij = 1, thus λmax =1.
j =1 j =1
 Sensitivity of the Eigenvector
n
∆ w1= ∑
j= 2
( v Tj ∆ A w 1 /( λ 1 - λ j ) v Tj w j )w j

The eigenvector w1 is insensitive to perturbation in A, if 1) the number of terms is small

(i.e. n is small), 2) if the principal eigenvalue λ1 is separated from the other eigenvalues ,
here assumed to be distinct (otherwise a slightly more complicated argument can also be
made and is given below) and, 3) if none of the products vjT wj of left and right
eigenvectors is small and if one of them is small, they are all small. Howerver, v1T w1,
the product of the normalized left and right principal eigenvectors of a consistent matrix
is equal to n which as an integer is never very small. If n is relatively small and the
elements being compared are homogeneous, none of the components of w1 is arbitrarily
small and correspondingly, none of the components of v1T is arbitrarily small. Their
product cannot be arbitrarily small, and thus w is insensitive to small perturbations of the
consistent matrix A. The conclusion is that n must be small, and one must compare
homogeneous elements.
When the eigenvalues have greater multiplicity than one, the corresponding left and right
eigenvectors will not be unique. In that case the cosine of the angle between them which
is given by corresponds to a particular choice of and . Even when and correspond to a
simple they are arbitrary to within a multiplicative complex constant of unit modulus,
but in that case | viT wi| is fully determined. Because both vectors are normalized, we
always have | viT wi | <1.
 Fundamentals of the AHP/ANP
•How to structure complexity as a hierarchy or as a
network;
•Why make comparisons to derive priorities;
•Why reciprocals and why homogeneous groups of
elements;
•Why the fundamental scale 1-9, what does it mean to
assign a number for a judgment;
•Why allow inconsistency;
•What is the minimum number of judgments needed
and why use redundant judgments;
•Why the principal right eigenvector;
•Why absolute numbers and absolute scales;
•Why weight and add for synthesis;
•Why the distributive and ideal modes;
•Why the supermatrix and what does raising it
to powers do;
•Why stochastic supermatrix;
•Why weight the components;
•Why BOCR – why add benefits and
opportunities and subtract costs and risks ;
•Why the ideal mode;
•How to allocate resources –the need for ratio
scales;
 Some Answers
(Only to be a little helpful)
• One structures a hierarchy from a goal download to criteria, subcriteria and goals, involving actors
and stakeholders and terminating in alternatives at the bottom. The ideas to go gradually from the
general to the particular. In a network, elements are put in clusters or components with their
connections indicating influence.
•Comparisons are more scientific in deriving scales because they use a unit and estimate multiples of
that unit rather than simply assigning numbers by guessing.
•Reciprocals are needed because if one element is five times more important than another then the
other is a forteori one fifth as important as the first. One deals with homogeneous clusters to make
the comparisons possible, closer and more accurate.
•The scale 1-9 helps us quantify our feelings and judgments in comparing elements.
•Human judgment expressed in the form of paired comparisons is naturally inconsistent. A modicum
of inconsistency enables us to improve our understanding by focusing on the most inconsistent
judgments.
• The minimum number of judgments needed to connect n elements is n-1. Redundant judgments
improve the validity of the derived priority vector.
•Ratios and ratio scales give us information on both the rank order of the elements and on their
relative values. It also makes possible proportionate resource allocation.
•Weighting and adding follows from simple operations we do all the time and is no different for
priorities. Suppose the goal has two components of values 0.6 and 0.4, and assume that one has a 0.2
share in the first component and a 0.7 share in the second. The total share with respect to the goal is
0.6 x 0.2 + 0.4 x 0.7 = 0.4.
•The supermatrix is the framework for organizing the priorities derived from paired comparisons.
Raising it to powers gives the overall influence of each element on all the other elements.
 ASSIGNING NUMBERS vs.
PAIRED COMPARISONS
• A number assigned directly to an
object is at best an ordinal and
cannot be justified.
• When we compare two objects or
ideas we use the smaller as a unit
and estimate the larger as a multiple
of that unit.
• If the objects are homogeneous and if
we have knowledge and experience,
paired comparisons actually derive
measurements that are likely to be close
and that indicate magnitude on an
absolute scale.
WHY IS AHP EASY TO USE?
• It does not take for granted the
measurements on scales, but asks that
scale values be interpreted according
to the objectives of the problem.
• It relies on elaborate hierarchic
structures to represent decision prob-
lems and is able to handle problems of
risk, conflict, and prediction.
• It can be used to make direct
resource allocation, benefit/cost
analysis, resolve conflicts, design
and optimize systems.
• It is an approach that describes
how good decisions are made rather
than prescribes how they should be
made.
WHY THE AHP IS POWERFUL
IN CORPORATE PLANNING
1. Breaks down criteria into manage-
able components.
2. Leads a group into making a specific
decision for consensus or tradeoff.
3. Provides opportunity to examine
disagreements and stimulate
discussion and opinion.
4. Offers opportunity to change
criteria, modify judgments.
5. Forces one to face the entire
problem at once.
6. Offers an actual measurement
system. It enables one to
estimate relative magnitudes and
derive ratio scale priorities
accurately.
7. It organizes, prioritizes and
synthesizes complexity within a
rational framework.
8. Interprets experience in a relevant
way without reliance on a black
box technique like a utility function.
9. Makes it possible to deal with
conflicts in perception and in
judgment.
Table 1 Comparison of Group Decision Making Methods

Method Group Maintenance Problem Abstraction Structure Analysis

Leadership Learning Scope Development Breadth Depth Faithfulness Breadth and

Effectiveness of of Depth of
Alternatives Judgments Analysis
(What if)

Structuring Low Medium Medium Low NA NA NA NA

Analogy, Association Medium Medium High Low NA NA NA NA
Boundary Examination Low Low Low Medium NA NA NA NA
Brainstorming/Brainwriting Low Medium High Very High NA NA NA NA
Morphological Connection Medium Medium High Very High High High NA NA
Why-What's Stopping Low Low NA NA Low Low Low Low
Ordering and Ranking Medium Medium Medium High Low Low Low Low
Voting Medium Medium Medium High Low Low Low Low
Nominal Group Technique Medium High Medium Medium High Low Medium Medium
Delphi Medium Medium Medium Low High Low Medium Medium
Disjointed Incrementalism Low Low Medium Low High Low Very High Medium
Matrix Evaluation Low Low Medium Low Low Low Very High Medium
Goal Programming Medium High Medium High High Low Medium High
Conjoint Analysis
Outranking

Structuring and Measuring Medium High Medium Low Low Low Very High Medium
Baysian Analysis Medium High Medium High High Low High High
MAUT/MAVT High Very High Medium Very High High High Very High Very High
AHP

NA = Not Applicable
Table1 (cont'd)

Method Fairness Applicability, Validity, and Truthfulness

Cardinal Prioritizing Consideration Scientific Applicability Psycho- Applicability Validity of

Separation Group of Other and to Intangibles physical to the Outcome
of Members Actors & Mathematical Applicability Conflict (Prediction)
Alternatives Stakeholders Generality Resolution

Structuring NA NA NA NA NA NA NA NA
Analogy/Association NA NA NA NA NA NA NA NA
Boundary Examination NA NA NA NA NA NA NA NA
Brainstorming/Brainwriting NA NA NA NA NA NA NA NA
Morphological Connection NA NA NA NA NA NA NA NA
Why-What's Stopping Low Low NA Medium NA NA NA Low
Ordering and Ranking NA NA NA Medium NA NA NA Low
Voting NA NA NA Medium NA NA NA Low
Nominal Group Technique NA NA Medium Low Low Low NA Medium
Delphi NA NA Medium Low Low Low NA Medium
Disjointed Incrementalism High NA Low Medium Medium NA NA Low
Matrix Evaluation High NA NA Medium Medium NA NA Low
Goal Programming High High Low Medium Medium Medium NA Medium
Conjoint Analysis
Outranking

Structuring and Measuring High NA Low High Medium Low NA Medium

Baysian Analysis High High Medium High Medium Medium Medium Medium
MAUT/MAVT High Very High High High Very High Very High High High
AHP

NA = Not Applicable
03/27/2006 09:58 PM
Dear Prof. Saaty:
Recently, I am thinking my future research fields all along. It is a really difficult decision. I want to devote
myself to MCDM. I also know that Herbert A. Simon got the Nobel prize in 1978 for his contribution to
organizational decision making, and Nash for his contribution to game theory and its applications. I wonder
if there is any possibility for the research on MCDM to make the same great contribution.

regards, Sun Yonghong

Sun Yonghong <50008602@student.cityu.edu.hk>

-----Original Message-----
From: saaty@katz.pitt.edu [mailto:saaty@katz.pitt.edu]
Sent: Tuesday, March 07, 2006 8:24 PM
To: Sun Yonghong
Subject: Re:
Dear Mr. Sun,
I assume you are just learning the subject because I think these are old papers that are not written
electronically. Would it be ok to send you other papers on the subject or should I try to find them and scan
them? I have generalized the AHP to the Analytic Network Process with dependence and feedback (ANP)
you may want something on it too. I teach a graduate course on the ANP starting tomorrow.
Kind regards,
Tom Saaty