Eigenvectors and Decision Making

Download Report

Transcript Eigenvectors and Decision Making

Eigenvectors and Decision Making
• Analytic Hierarchy Process (AHP)
• Positive Symmetrically Reciprocal Matrices
(PSRMs) and Transitive PSRMs
• Estimation methods -- How close? Close to what?
Close in what sense?
Analytic Hierarchy Process
“In the 1970’s the Egyptian
government asked Tom Saaty, a
pioneering mathematician with
a fistful of awards, to help
clarify the Middle East conflict.
The Egyptians needed a
coherent, analytical way to
assess the pros and cons of their
less than cozy strategic
relationship with the Soviet
Union. Saaty, a Wharton
professor with a background in
arms-control research, tackled
the question with game theory.”
Analytic Hierarchy Process
“The Egyptian government was
pleased with his work (and
eventually did ask the Russians
to leave), but Saaty himself
wasn’t satisfied with the
process. He felt his conclusion
was incomplete – that important
but intangible information was
left out of the final equation
because game theory was too
rigid. ‘I couldn’t use it to solve
a real-life problem.’”
Analytic Hierarchy Process
•
•
•
•
•
•
•
Xerox
(Software:
Boeing
Expert
IBM
Choice)
Northrop Grumman
US Steel
Corning
Governments of U.S., South Africa, Canada, and
Indonesia
Analytic Hierarchy Process
Decision:
Choose between a
Toyota, Honda, or Citation (Chevrolet)
Analytic Hierarchy Process
One set of criteria: cost, dependability, size, and aesthetics
First we want to assign numbers to these four criteria that
represent their relative importance w1, w2, w3, w4 .
Cost Depen Size Aesth
Cost
1
2
3
3
Dependability
1/2
1
3
3
Size
1/3
1/3
1
1/2
Aesthetics
1/3
1/3
2
1
This matrix has maximum eigenvalue approximately 4.08 and
eigenvector is (.44, .31, .10, .15).
Analytic Hierarchy Process
A n x n matrix
x nx1
λ scaler
Ax = λx
(λI-A)x = 0
det(λI-A) = 0
Analytic Hierarchy Process
2
 1
1 / 2 1

1 / 3 1 / 3

1 / 3 1 / 3
3 3  .44
.44





3 3  .31
.
31
 4.08 
 .1 
1 1 / 2  .1 
 
 
2 1  .15
.15
Generate reciprocal matrices for each criterion, comparing the three types
of cars pairwise on each criterion as shown below.
For Dependability:
Toy
Hon
Cit
Toy
1
1
1/3
Hon
1
1
1/3
Cit
3
3
1
Weights
.43
.43
.14
Toy
Hon
Cit
Weights
1
1
1
2
1
1
For Cost:
Toy
Hon
Cit
1
1
1/2
.40
.35
.25
For Size:
Toy
Toy
Hon
Cit
1
1
1/3
Hon
Cit
3
1
1/3
1
3
1
Hon
Cit
4
1
1/2
3
2
1
Weights
.43
.14
.43
For Aesthetics:
Toy
Toy
Hon
Cit
1
1/4
1/3
Weights
.63
.22
.15
The weight for each criterion is then multiplied by the weight for each car
within that criterion, and added across criteria as shown below:
Cost Depen Size Aesth
(.44) (.31) (.10) (.15)
Toy
Hon
Cit
.40
.35
.25
.43
.43
.14
.43
.14
.43
.63
.33
.15
Composite Priorities
.45
.33
.22
Positive Symmetrically Reciprocal
Matrices
n x n matrix A
aij > 0
aji = 1/aij
Transitive or consistent: aik = aijajk
wi/wj = wi/wk x wk/wj
(The entries aij represent the ratio wi/wj.)
Cost
Dependability
Size
Aesthetics
w1/w1
w2/w1,
w3/w1
w4/w1
Cost
Depen
w1/w1
w2/w1,
w3/w1
w4/w1
w1/w2 w1/w3
w2/w2 w2/w3
w3/w2 w3/w3
w4/w2 w4/w3
w1/w2 w1/w3
w2/w2 w2/w3
w3/w2 w3/w3
w4/w2 w4/w3
w1/w4
w2/w4
w3/w4
w4/w4
Size
w1
w2
w3
w4
Aesth
w1/w4
w2/w4
w3/w4
w4/w4
= 4
w1
w2
w3
w4
Positive Symmetrically Reciprocal
Matrices
Theorem (Perron-Frobenius):
(Let A be non-negative and irreducible.)
(1) A has a real positive simple (not multiple) eigenvalue
λmax (Perron root) which is not exceeded in modulus by
any other eigenvalue of A.
(2) The eigenvector of A corresponding to the eigenvalue
λmax has positive components and is essentially unique (to
within multiplication by a constant).
Theorem: A positive, reciprocal matrix is consistent if and
only if λmax = n. (Saaty. J of Mathematical Psychology,
1977).
Positive Symmetrically Reciprocal
Matrices
Theorem:
Let R be an n x n positive reciprocal matrix with entries
1/S < rij < S, 1 < i, j < n, for some S >1
and let λmax denote the largest eigenvalue of R in modulus,
which is known to be real and positive. Then
n < λmax < 1 + ½(n-1)(S + 1/S),
the lower and upper bound being reached if and only if R
is supertransitive or maximally intransitive, respectively.
(Aupetit and Genest – European Journal of Operational
Research, 1993):
Positive Symmetrically Reciprocal
Matrices
Theorem:A transitive matrix is SR and has rank one. (Farkas,
Rozsa, Stubnya, Linear Algebra Appl., 1999.)
Close?
2
 1
1 / 2 1

1 / 3 1 / 3

1 / 3 1 / 3
3 3  .44
.44





3 3  .31
.
31
 4.08 
 .1 
1 1 / 2  .1 
 
 
2 1  .15
.15
Close?
• Saaty: Limit of normalized row sums of powers of the
matrix.
k
Ae
• μ =(λmax-n)/(n-1)
lim T
k  e Ae
• If we consider inconsistencies in the powers of A,
generated by cycles of length 1, 2, 3, … then the right
eigenvector represents the dominance of one alternative
over all others through these cycles. (Mathematics
Magazine, 1987.)
Close?
• Logarithmic Least Square Method


w
i
 ln aij  ln



w j 
i, j 
n
2
• Least Squares

wi

a


 ij w
i, j 
j
n




2
• Another kind of eigenvector in another metric, the “max
eigenvector” minimizes relative error (Elsner & vanden
Driesscle, Linear Algebra and Its Applications, 2004).
Finally
Why do I care?
What am I doing with these matrices and
eigenvectors?