Prof. Prof. Ganapati Ganapati Panda, Panda, FNAE, FNAE, FNASc FNASc..

Dean Academic Affairs Dean Academic Affairs
Professor, School of Electrical Sciences Professor, School of Electrical Sciences
IIT Bhubaneswar IIT Bhubaneswar
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
³Multiobjective optimization is the process of
simultaneously optimizing two or more conflicting
objectives subject to certain constraints.´
Multiobjective Multiobjective Optimization Optimization
2/13/2012 2
Examples of multi-objective optimization problems:-
Maximizing profit and minimizing the cost of a product.
Maximizing performance and minimizing fuel consumption of a vehicle.
Minimizing weight while maximizing the strength of a particular
component.
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Difference Difference
2/13/2012 3
Single Objective Optimization
Optimize only one objective
function
Single optimal solution
Maximum/Minimum fitness value
is selected as the best solution.
Multiobjective Optimization
Optimize two or more than two
objective functions
Set of optimal solutions
Comparison of solutions by
‡ Domination
‡ Non-domination
Minimize
where -10 < x < 20
Optimal solution:-
0 10 20 30 40 50 60 70 80 90 100
0
10
20
30
40
50
60
70
80
90
100
f1
f
2
-10 -5 0 5 10 15 20
0
100
200
300
400
x
f
(
x
)
f1(x)
f2(x)
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
2/13/2012 4
Standard Approach :Weighted Sum Standard Approach :Weighted Sum
of Objective Functions of Objective Functions
Limitations:
× Result depends on weights.
× Some solutions may be missed.
× Multiple runs of the algorithm are required in order to get the whole
range of solutions.
× Difficult to select proper combination of weights.
× Combining objectives loses information and predetermines trade-offs
between objectives.
) ( ) ( ) ( ) (
2 2 1 1
x f x f x f x g
m m
T
- -
T T T
P P P !
_ a ) ( , ), ( ), (
2 ` 1
x f x f x f Minimize
m
T
- -
T T

1 &
2 1
!
m
P P P - -
m
P P P , ,. ,
2 1
- - where are weights values
and m represents the number of objective functions.
Formulate as a single objective with weighted sum of all objective functions -
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Definitions Definitions
Domination :
One solution is said to dominate
another if it is better in all objectives.
Non-Domination [Pareto points] :
A solution is said to be non-
dominated if it is better than other
solutions in at least one objective.
2/13/2012 5
Minimize function
M
i
n
i
m
i
z
e

f
u
n
c
t
i
o
n
A
B
C
D
1
f
2
f
A dominates B (better in both and )
A dominates C (same in but better in )
A does not dominate D (non-dominated points)
A and D are in the ³Pareto optimal front´
These non-dominated solutions are called Pareto optimal solutions.
This non-dominated curve is said to be Pareto front.
2
f
1
f
2
f
1
f
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
2/13/2012 6
Definitions Definitions
Pareto Optimal
A vector variable is Pareto optimal if for every
and either
or, there is at least one such that
where is the vector of decision variables,
is the vector of objective
functions, is the feasible region ,where represents
the whole search space.
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Definitions Cont«. Definitions Cont«.
Pareto Optimal Set
For a given MOP the Pareto optimal set is
defined as
Pareto Front
For a given MOP and Pareto optimal set , the
Pareto front is defined as
2/13/2012 7
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
2/13/2012 8
Desirable MOEA features Desirable MOEA features
Best Solutions
Lie on true Pareto front
They are uniformly
distributed on the front
Aim:
To achieve convergence to Pareto optimal front
To achieve diversity (representation of the entire Pareto
optimal front)
Minimize function
M
i
n
i
m
i
z
e

f
u
n
c
t
i
o
n
1
f
2
f
True Pareto
front
Possible
solutions
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
2/13/2012 9
Non Dominated Sorting based Non Dominated Sorting based
Genetic Algorithm II (NSGA Genetic Algorithm II (NSGA- - II) II)
Developed by Prof. K. Deb at Kanpur Genetic Algorithms
Laboratory (2002)
Famous for Fast non-dominated search
Fitness assignment - Ranking based on non-domination
sorting
Diversity mechanismis based on Crowding distance
Uses Elitism
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
2/13/2012 10
Minimize
where
,
2
1
x x f !

2
2
2 ! x x f
5 5 e e x
Initialize Population Initialize Population
‡ Search space is of single dimension (given).
‡ Objective space is of two dimension (given).
‡ Let population size = 10
‡ Initialize population with 10 chromosomes having single
dimensioned real value.
‡ These values are randomly distributed in between [-5,5].
0.4678
1.7355
0.8183
-0.414
3.2105
-1.272
-1.508
-1.832
-2.161
-4.105
x
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
-0.414 0.171 5.829
0.467 0.218 2.347
0.818 0.669 1.396
1.735 3.011 0.07
3.210 10.308 1.465
-1.272 1.618 10.708
-1.508 2.275 12.308
-1.832 3.355 14.682
-2.161 4.671 17.317
-4.105 16.854 37.275
x
x f
1
x f
2
Evaluate Fitness values Evaluate Fitness values
‡ Find out all objective functions values (fitness values) for all
chromosomes.
0 2 4 6 8 10 12 14 16 18
0
5
10
15
20
25
30
35
40
f
1
(x)
f
2
(
x
)
2/13/2012 11
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
1
1
1
1
2
2
3
4
5
6
-0.414 0.171 5.829
0.467 0.218 2.347
0.818 0.669 1.396
1.735 3.011 0.07
3.210 10.308 1.465
-1.272 1.618 10.708
-1.508 2.275 12.308
-1.832 3.355 14.682
-2.161 4.671 17.317
-4.105 16.854 37.275
x x f
1
x f
2
Rank
Fast Non Fast Non- -domination Sorting domination Sorting
‡ Assigning the rank to each individual of the population.
‡ Rank based on the non-domination sorting (front wise).
‡ It helps in selection and sorting.
2 } , , , {
1 } , , {
1 } , {
0 } {
0 } {
0 } {
0 } {
7 6 3 2 1 7
6 3 2 1 6
5 4 3 5
4 4
3 3
2 2
1 1
! !
! !
! !
! !
! !
! !
! !
n x x x x x
n x x x x
n x x x
n x
n x
n x
n x
-
-
-
-
Reference
chromosome
Dominated
chromosomes
Counter
2/13/2012 12
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
2/13/2012 13
-0.414 0.171 5.829 1
0.467 0.218 2.347 1
0.818 0.669 1.396 1
1.735 3.011 0.07 1
3.210 10.308 1.465 2
-1.272 1.618 10.708 2
-1.508 2.275 12.308 3
-1.832 3.355 14.682 4
-2.161 4.671 17.317 5
-4.105 16.854 37.275 6
x x f
1
x f
2
Rank
0 2 4 6 8 10 12 14 16 18
0
5
10
15
20
25
30
35
40
f
1
(x)
f
2
(
x
)
4
3
5
6
2
1
Fast Non Fast Non- -domination Sorting domination Sorting
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Crowding Distance Assignment Crowding Distance Assignment
2/13/2012 14
‡ To get an estimate of density of
solutions surrounding a particular
solution in population.
‡ Choose individuals having large
crowding distance.
‡ Help for obtaining uniformly
distribution.
where represent objective function value of solution.
and is the maximum value of function in the Pareto front.
max
m
f
m
f
T
) 1 ( , , 3 , 2 ! l i
where
-
, 1
. . . .
g ! !
D C D C
l
§
¹
¹
º
¸
©
©
ª
¨

!
m
m m
m m
D C
f f
i f i f
i
min max
. .
] 1 [ ] 1 [
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
0.5 1 1.5 2 2.5 3
0
1
2
3
4
5
6
f
1
(x)
f
2
(
x
)
i-1
i
i+1
1
2
4
3
0 2 4 6 8 10 12 14 16 18
0
5
10
15
20
25
30
35
40
f
1
(x)
f
2
(
x
)
4
3
5
6
2
1
-0.414 0.171 5.829 1
0.467 0.218 2.347 1 0.945
0.818 0.669 1.396 1 1.378
1.735 3.011 0.07 1
3.210 10.308 1.465 2
-1.272 1.618 10.708 2
-1.508 2.275 12.308 3
-1.832 3.355 14.682 4
-2.161 4.671 17.317 5
-4.105 16.854 37.275 6
x
x f
1
x f
2
Rank
g
. .D C
g
g
g
g
g
g
g
Crowding Distance Assignment Crowding Distance Assignment
‡ Crowning distance can be calculated
for all chromosomes of same Pareto
front.
2/13/2012 15
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Selection Selection
2/13/2012 16
Selection is the stage of a genetic algorithm in which individual are chosen from a
population for later breeding (recombination or crossover).
The crowding operator guides the selection process at the various stages
of the algorithm toward a uniformly spread-out Pareto optimal front.
where shows non-domination rank & is crowding
distance of individual.
Crowding operator based sorting
j i
n
e

rank rank
j i
n
e
. .D C
i
rank
i
th
i

. . . . D C D C rank rank
j i and j i
or
" !
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
2/13/2012 17
Runs a µtournament¶ among a few individuals chosen at random from the
population and selects the winner (the one with the best fitness) for crossover.
‡ In tournament selection, a number Tour size of individuals is chosen
randomly from the population and the best individual from this group is
selected as parent. (Based on the crowding operator)
Tournament Selection Tournament Selection
0.818 0.669 1.396 1 1.378
-1.508 2.275 12.30 3
x x f
1
x f
2
Rank . .D C
g
0.818 0.669 1.396 1 1.378
rank rank
2 1
0.467 0.218 2.347 1 0.945
0.818 0.669 1.396 1 1.378
x x f
1
x f
2
Rank . .D C
0.818 0.669 1.396 1 1.378
rank rank
2 1 !
. . . .
2 1
D C D C

INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
where
‡ is a random number {0,1}
‡ is a crossover operator
‡ represent dimension of
individual.
r
Q
j
Crossover Crossover
2/13/2012 18
Crossover is a genetic operator that combines (mates) two individuals (parents)
to produce two new individuals (Childs).
‡ The idea behind crossover is that the new chromosome may be better
than both of the parents if it takes the best characteristics from each of the
parents.

±
±
°
±
±
¯
®
"
¹
¹
º
¸
©
©
ª
¨

e
!
¹
¹
º
¸
©
©
ª
¨

¹
¹
º
¸
©
©
ª
¨

5 . 0
1 * 2
1
5 . 0 * 2
1
1
1
1
r if
r
r if r
b
Q
Q
Simulated Binary Crossover

) ( * 1 ) ( * 1
2
1
) (
) ( * 1 ) ( * 1
2
1
) (
2 1 2
2 1 1
j parent b j parent b j child
j parent b j parent b j child
!
!
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
2/13/2012 19
Mutation Mutation
Mutation is a genetic operator that alters one ore more gene values in a
chromosome from its initial state.
‡ Mutation is an important part of the genetic search as helps to prevent the
population from stagnating at any local optima.
Polynomial Mutation


±
°
±
¯
®
"
e
!
¹
¹
º
¸
©
©
ª
¨

¹
¹
º
¸
©
©
ª
¨

5 . 0 1 * 2 1
5 . 0 1 * 2
1
1
1
1
r if r
r if r
d
L
L
d j parent j child ! ) ( ) (
where
‡ is a random number {0,1}
‡ is a mutation operator
‡ represent dimension of
individual.
r
L
j
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Pt
Qt
Rt ={Pt , Qt}
F1
Non-dominated
sorting (Rank)
Crowding distance
sorting
F2
F3
Rejected
Selection for next generation Selection for next generation
Pt+1
Elitist Replacement
2/13/2012 20
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
2/13/2012 21
0 0.5 1 1.5 2 2.5 3 3.5
0
0.5
1
1.5
2
2.5
3
3.5
f
1
(x)
f
2
(
x
)
0 0.5 1 1.5 2 2.5 3 3.5 4
0
0.5
1
1.5
2
2.5
3
3.5
4
f
1
(x)
f
2
(
x
)
0 2 4 6 8 10 12 14 16 18
0
5
10
15
20
25
30
35
40
f
1
(x)
f
2
(
x
)
0 0. 5 1 1. 5 2 2. 5 3 3. 5 4
0
0. 5
1
1. 5
2
2. 5
3
3. 5
4
f
1
(x)
f
2
(
x
)
Initial State After 20 generation
After 10 generation After 40 generation
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Report Final
Population and Stop
Begin : Initialize
Population (N)
Evaluate objective
functions
Non-dominated
Sorting
Tournament Selection
Combine parent and
child populations ,
Non-dominating
Sorting
Crossover &
Mutation
Evaluate objective
functions
Flowchart of NSGA Flowchart of NSGA- -II II
Yes
No
Select N
individuals
2/13/2012 22
Stopping
criteria
met ?
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Performance Measures Performance Measures
2/13/2012 23
There are two main goals in a multi-objective optimization:
1) Convergence to the Pareto-optimal set
2) Maintenance of diversity in solutions of the Pareto-
optimal set.
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0
2
4
6
8
10
12
function - 1
f
u
n
c
t
i
o
n

-

2
DISTANCE MEASURE BETWEEN PARETO FRONTS


T R U E P AR E T O F R ON T
P AR E T O F R ON T
min distance
min distance
Convergence metric
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Cont«. Cont«.
2/13/2012 24
Non uniformity in the distribution,
If distance between the solutions
is equal to average distance , that
gives uniformly distribution.
The parameters and are the Euclidean distances between the extreme solutions of
true Pareto front and the boundary solutions of the obtained non-dominated set.
The parameter is the average of all distances , , assuming that
there are solutions on the best non- dominated front.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
0
0.5
1
1.5
2
2.5
3
3.5
4
function -1
f
u
n
c
t
i
o
n

-
2
DIVERSITY PLOT
Pareto front
Extreme
solution
Extreme
solution
1
d
2
d
3
d
4
d
1 n
d
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
2/13/2012 25
Simulation and Results Simulation and Results
NSGA II parameters
Population (N) = 100
Crossover Probability (Pc)= 0.9
Mutation Probability (Pm) = 0.1
Distribution index for crossover (µ)= 20
Distribution index for mutation (Ș) = 20
Tour size (selection) = 2
Implementation use real numbers representation.
* These parameters were kept in all test functions optimization.
* Only changed the total number of fitness function evaluations.
MOPSO parameters
Population = 100 particles
Repository (Archive) size = 100 particles
Mutation rate = 0.5
Divisions for Archive Grid = 30
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
0 0.5 1 1.5 2 2.5 3 3.5 4
0
1
2
3
4
5
F1
F
2
Nondominated solutions w ith NSGA-II on SCH


Pareto-optimal Front
NSGA-II
0 0.5 1 1.5 2 2.5 3 3.5 4
0
1
2
3
4
5
F1
F
2
Nondominated solutions w ith MOPSO on SCH


Pareto-optimal Front
MOPSO
Test Problem : SCH Test Problem : SCH
2/13/2012 26
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Convergence Metric NSGA-II MOPSO
Best 0.0148 0.0093
Worst 0.9578 0.1569
Mean 0.2096 0.0259
Diversity Metric NSGA-II MOPSO
Best 0.5104 0.6947
Worst 0.7904 1.3575
Mean 0.6425 0.8582
Tab. 1: Results of the Convergence Metric for the
SCH Test Function
Tab. 2: Results of the Diversity Metric for the
SCH Test Function
Comparison for SCH Test Function Comparison for SCH Test Function
2/13/2012 27
ȗ Total numbei of fitness evaluations was set to ͳͲǡͲͲͲǤ
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
1.2
1.4
F1
F
2
Nondominated solutions w ith NSGA-II on DEB-1


Pareto-optimal Front
NSGA-II
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
1.2
1.4
F1
F
2
Nondominated solutions w ith MOPSO on DEB-1


Pareto-optimal Front
MOPSO
Test Problem : DEB Test Problem : DEB- -11
2/13/2012 28
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Convergence Metric NSGA-II MOPSO
Best 0.0066 0.0070
Worst 0.5140 0.1664
Mean 0.0078 0.0079
Diversity Metric NSGA-II MOPSO
Best 0.3467 0.5112
Worst 0.5140 0.7168
Mean 0.4243 0.5938
Tab. 1: Results of the Convergence Metric for the
DEB-1 Test Function
Tab. 2: Results of the Diversity Metric for the
DEB-1 Test Function
Comparison for DEB Comparison for DEB- -1 Test Function 1 Test Function
2/13/2012 29
ȗ Total numbei of fitness evaluations was set to ͳͷǡͲͲͲǤ
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
1
2
3
4
5
6
7
8
F1
F
2
Nondominated solutions w ith MOPSO using DEB-2


Pareto-optimal Front
MOPSO
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
1
2
3
4
5
6
7
8
F1
F
2
Nondominated solutions w ith NSGA-II on DEB-2


Pareto-optimal Front
NAGA-II
Test Problem : DEB Test Problem : DEB- -22
2/13/2012 30
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Convergence Metric NSGA-II MOPSO
Best 0.0449 0.0515
Worst 0.0559 0.0725
Mean 0.0516 0.0608
Diversity Metric NSGA-II MOPSO
Best 0.7248 0.6800
Worst 0.7939 0.7582
Mean 0.7597 0.7193
Tab. 1: Results of the Convergence Metric for the
DEB-2 Test Function
Tab. 2: Results of the Diversity Metric for the
DEB-2 Test Function
Comparison for DEB Comparison for DEB- -2 Test Function 2 Test Function
2/13/2012 31
ȗ Total numbei of fitness evaluations was set to ʹͷǡͲͲͲǤ
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
-20 -19 -18 -17 -16 -15 -14
-12
-10
-8
-6
-4
-2
0
2
F1
F
2
Nondominated solutions w ith MOPSO on KUR


Pareto-optimal Front
MOPSO
Test Problem : KUR Test Problem : KUR
2/13/2012 32
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Convergence Metric NSGA-II MOPSO
Best 0.0021 0.0021
Worst 0.0041 0.0034
Mean 0.0028 0.0026
Diversity Metric NSGA-II MOPSO
Best 0.3344 0.4803
Worst 0.7825 0.6413
Mean 0.4399 0.5602
Tab.1: Results of the Convergence Metric for the
KUR Test Function
Tab.2: Results of the Diversity Metric for the
KURTest Function
Comparison for KUR Test Function Comparison for KUR Test Function
2/13/2012 33
ȗ Total numbei of fitness evaluations was set to ʹͲǡͲͲͲǤ
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Layout Optimization for a Wireless Layout Optimization for a Wireless
Sensor Network using NSGA Sensor Network using NSGA - - II II
a) Coverage
b) Lifetime
2/13/2012 34
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Wireless Sensor Network (WSN) Wireless Sensor Network (WSN)
Example of a WSN where sensor nodes are communicating with
the DPU through HECN
Data Processing Unit
(DPU)
High Energy Communication Node
(HECN)
Node 1
Node 2
Node 3
Node 6
Node 5
Node 8
Node 4
Node 9
Node 7
2/13/2012 35
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Optimization of Coverage Optimization of Coverage
Coverage is defined as the ratio of the union of
areas covered by each node and the area of the
entire ROI.
A
A
C
i N i ,..., 1 !
!
7
A
i
- Area covered by the i
th
node
N - Total number of nodes
A - Area of the ROI
2/13/2012 36
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Optimization of Lifetime Optimization of Lifetime
The lifetime of the whole network is the time until one
of the participating nodes run out of energy.
In every sensing cycle, the data from every node is
routed to HECN through a route of minimum weight
max
failure
T
T
Lifetime !
T
failure
= maximum number of sensing cycles before failure of any node
T
max
= maximum number of possible sensing cycles
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Competing Objectives Competing Objectives
Lifetime Coverage
‡ try to spread out the nodes for
maximizing coverage
‡ try to arrange the nodes as
close as possible to the
HECN for maximizing lifetime
HECN
2/13/2012 38
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Simulation Parameters Simulation Parameters
Number of chromosomes 100
Number of generations 50
Crossover Probability 0.9
Mutation Probability 0.5
Distribution index for crossover 20
Distribution index for mutation 20
Tour size 2
Parameters of NSGA-II
2/13/2012 39
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
NSGA NSGA- -II Results II Results
Pareto Front obtained for a WSN with 10 sensors, 100 chromosomes
and 50 generations
0 0. 1 0. 2 0. 3 0. 4 0. 5 0. 6 0. 7 0. 8 0. 9 1
0
0. 1
0. 2
0. 3
0. 4
0. 5
0. 6
0. 7
0. 8
0. 9
1
Pareto optimal front
Coverage
L
i
f
e
t
i
m
e
2/13/2012 40
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
NSGA NSGA- -II Results (Cont¶d) II Results (Cont¶d)
-6 -4 -2 0 2 4 6
-6
-4
-2
0
2
4
6
Coverage = 0.3335 Lifetime = 0.999
HECN
Best Lifetime
-6 -4 -2 0 2 4 6
-6
-4
-2
0
2
4
6
Coverage = 0.63709
HECN
Initial Disconnect
Network
-6 -4 -2 0 2 4 6
-6
-4
-2
0
2
4
6
Coverage = 0.5353 Lifetime = 0.249
Best Coverage
HEC
N
2/13/2012 41
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
ё
ё
ё
ё
ё
є
Input Layer Hidden Layer Output Layer
Radial Basis Function Network Radial Basis Function Network
Accuracy
Complexity of the model
i
d
Ö
1
x
2
x
3
x
4
x
1
w
2
w
3
w
4
w
5
w
0
w
1
42
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Structure determination of RBF network can be considered as the multiobjective
optimization problem concerning with accuracy and complexity of the model.
Multiobjective Multiobjective Problem Formulation Problem Formulation
M f !
1


2
1 1
0
1
2
2
1
Ö
1
§ §
§
! !
!
¹
¹
º
¸
©
©
ª
¨
¹
¹
º
¸
©
©
ª
¨
!
!
!
n
i
M
j
j i j i
n
i
i i
w c x w d
n
d d
n
mse f
J
here is total number of basis functions (centers) in RBF network,
: Desired output
: Estimated output during the training of RBF network.
: Weight vector of the RBF network
: Center vector of the RBF network
: Gaussian Function
M
i
d
Ö
i
d
w
j
c
. J
43
¹
º
¸
©
ª
¨
!
2
2
2
1
exp
j j
c x c x
W
J
where is the spread of the Gaussian function.
W
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Desired Output
0 0 0 +1 0 0 0
In the chromosome, the position of gene value ³1´ indicate the center position
of the basis function (selected center) and number of ³1´ genes in chromosome
indicates the number of basis functions (number of centers).
Chromosome 1 0 1 1 0 0 1
Input Data Points
G
a
u
s
s
i
a
n

D
i
s
t
r
i
b
u
t
i
o
n

Selected centers
Input
Data
Points
+1 +1 -1 +1 0 0 0
0 +1 +1 -1 +1 0 0
0 0 +1 +1 -1 +1 0
0 0 0 +1 +1 -1 +1
Structure selection of RBF network Structure selection of RBF network
44
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
Pareto Fronts Pareto Fronts
‡ This Pareto Front shows that for the different number of centers, MSE changes.
‡ The performance of an RBF network critically depends upon the chosen centers.
45
2 4 6 8 10 12 14 16 18 20 22 24 25
10
-40
10
-30
10
-20
10
-10
10
0
Numbers of Centers
M
e
a
n

E
q
u
a
r
e

E
r
r
o
r

(
l
o
g

s
c
a
l
e
)
Pareto Front for 13-element Barker Code
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
References References
1. K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan,´ A fast and elitist multi-objective genetic
algorithm: NSGA-II´, IEEE Transaction on Evolutionary Computation, 6(2), 181-197,2002.
2. K. Deb and R. B. Agrawal, ³Simulated binary crossover for continuous search space,´ in
Complex Syst., vol. 9, pp. 115±148., Apr. 1995.
3. N. Srinivas and K. Deb, ³Multiobjective function optimization using nondominated sorting
genetic algorithms,´ Evol. Comput., vol. 2, no. 3, pp. 221±248, Fall 1995.
4. J. Horn, N. Nafploitis, and D. E. Goldberg, ³A niched Pareto genetic algorithm for
multiobjective optimization,´ in Proceedings of the First IEEE Conference on Evolutionary
Computation, Z. Michalewicz, Ed. Piscataway, NJ: IEEE Press, pp. 82±87 , 1994.
5. J. D. Knowles and D.W. Corne, ³Approximating the nondominated front using the Pareto
archived evolution strategy,´ Evol. Comput., vol. 8, pp. 149±172, 2000.
6. Carlos A. Coello Coello, Member, IEEE, Gregorio Toscano Pulido, and Maximino Salazar
Lechuga, ³Handling multiple objectives with particle swarm optimization´, Evol. Comput.,
vol. 8, pp. 256±279, No. 3, June 2004
2/13/2012 46
INBIAN INSTIT0TE 0F TECBN0L0uY BB0BANESWAR
THANK YOU
2/13/2012 47

Multiobjective Optimization

³Multiobjective optimization is the process of simultaneously optimizing two or more conflicting objectives subject to certain constraints.´
Examples of multi-objective optimization problems: Maximizing profit and minimizing the cost of a product.  Maximizing performance and minimizing fuel consumption of a vehicle.  Minimizing weight while maximizing the strength of a particular

component.

2/13/2012





.

     2 .

Difference Single Objective Optimization  Optimize Multiobjective Optimization  Optimize two or more than two only one objective function  Single optimal solution  Maximum/Minimum fitness value objective functions  Set of optimal solutions  Comparison of solutions by ‡ ‡ is selected as the best solution. Domination Non-domination 100 90 80 Minimize 400 f1(x) f2(x) 300 70 60 f(x) where -10 < x < 20 200 50 f2 40 Optimal solution:- 100 30 20 10 0 -10 -5 0 5 10 15 20 0 0 10 20 30 40 50 60 70 80 90 100 x f1 3 2/13/2012 .

.

.

.

     .

. Pm ! 1 and m represents the number of objective functions. f m ( x )a Formulate as a single objective with weighted sum of all objective functions - T T T T g ( x ) ! P1 f1 ( x )  P2 f 2 ( x )  .. Difficult to select proper combination of weights. Limitations: × × × × × Result depends on weights. P2 . Multiple runs of the algorithm are required in order to get the whole range of solutions.... Pm are weights values & P1  P2  ... f 2 ( x ).Standard Approach :Weighted Sum of Objective Functions T T T Minimize  _f1` ( x ). Some solutions may be missed. Pm f m ( x ) where P1 .. Combining objectives loses information and predetermines trade-offs between objectives. 4 ..

.

.

.

     2/13/2012 .

f1 2/13/2012 . Non-Domination [Pareto points] : A solution is said to be nondominated if it is better than other solutions in at least one objective. This non-dominated curve is said to be Pareto front.       f2 Minimize function B A C D Minimize function A dominates B (better in both f1 and f 2 ) A dominates C (same in f1 but better in f 2 ) A does not dominate D (non-dominated points) A and D are in the ³Pareto optimal front´ These non-dominated solutions are called Pareto optimal solutions.Definitions Domination : One solution is said to dominate another if it is better in all objectives.

.

.

.

     5 .

2/13/2012 . there is at least one where is Pareto optimal if for every either such that is the vector of decision variables.Definitions Pareto Optimal A vector variable and or. is the vector of objective functions.where represents the whole search space. is the feasible region .

.

.

.

     6 .

Pareto Optimal Set For a given MOP defined as the Pareto optimal set is Pareto Front For a given MOP and Pareto optimal set Pareto front is defined as .Definitions Cont«. the 2/13/2012 .

.

.

.

     7 .

Desirable MOEA features Best Solutions Lie on true Pareto front They are uniformly distributed on the front f2 Minimize function Possible solutions True Pareto front Minimize function Aim:  To achieve convergence to Pareto optimal front f1  To achieve diversity (representation of the entire Pareto optimal front) 2/13/2012 .

.

.

.

     8 .

II) (NSGADeveloped by Prof.Ranking based on non-domination sorting Diversity mechanism is based on Crowding distance Uses Elitism 2/13/2012 . Deb at Kanpur Genetic Algorithms Laboratory (2002) Famous for Fast non-dominated search Fitness assignment . K.Non Dominated Sorting based Genetic Algorithm II (NSGA.

.

.

.

     9 .

Initialize Population Minimize f1 .

f 2 .x ! x .

x ! .

414 3. ‡ These values are randomly distributed in between [-5. ‡ Let population size = 10 ‡ Initialize population with 10 chromosomes having single dimensioned real value.2105 -1. ‡ Objective space is of two dimension (given).x  2 2 where 2 x 0.161 -4.105 5 e x e 5 ‡ Search space is of single dimension (given).8183 -0.4678 1.272 -1. 2/13/2012 .832 -2.5].7355 0.508 -1.

.

.

.

     10 .

508 -1.210 -1.735 3.467 0.105 f1 . x -0.414 0.161 -4.272 -1.832 -2.Evaluate Fitness values ‡ Find out all objective functions values (fitness values) for all chromosomes.818 1.

308 1.275 3.171 0.011 10.355 4.854 f 2 .671 16.669 3.x 0.218 0.618 2.

465 10.396 0.317 37.708 12.275 40 35 30 25 f2 (x ) 20 15 10 5 0 0 2 4 6 8 f 1 (x) 10 12 14 16 18 2/13/2012 .829 2.x 5.682 17.347 1.07 1.308 14.

.

.

.

     11 .

‡ Rank based on the non-domination sorting (front wise).210 -1. x -0.105 f1 .818 1.735 3.832 -2.Fast Non-domination Sorting Non‡ Assigning the rank to each individual of the population.508 -1.467 0. ‡ It helps in selection and sorting.161 -4.272 -1.414 0.

355 4.308 1.618 2.x 0.671 16.275 3.854 f 2 .011 10.669 3.218 0.171 0.

}   n 4 ! 0 x ! {x . x }   n ! 1 x ! {x .829 2. x 2 .}   n ! 0 x 3 ! {. x }   n ! 1 x 7 ! {x1 .07 1.}   n3 ! 0 x 4 ! {. x . x 3 .708 12.396 0.275 Reference chromosome Dominated chromosomes Counter Rank 1 1 1 1 2 2 3 4 5 6 12 x1 ! {.465 10.}   n1 ! 0 x ! {.308 14. x 6 }   n 7 ! 2 .x 5.317 37.682 17.347 1.

.

.

.

     2 2 5 3 4 5 6 1 2 3 6 2/13/2012 .

105 f1 .414 0.467 0.832 -2.508 -1.Fast Non-domination Sorting Nonx -0.818 1.735 3.272 -1.210 -1.161 -4.

671 16.171 0.854 f 2 .011 10.308 1.618 2.218 0.x 0.275 3.669 3.355 4.

829 1 2.317 37.465 10.347 1.396 0.308 14.708 12.07 1.275 1 1 1 f 2 (x ) 40 35 30 25 6 5 20 15 10 5 0 0 2 2 3 4 5 6 4 3 2 1 2 4 6 8 f 1 (x) 10 12 14 16 18 2/13/2012 .682 17.x Rank 5.

.

.

.

     13 .

obtaining uniformly 1C . iC .Crowding Distance Assignment ‡ To get an estimate of density of solutions surrounding a particular solution in population. ‡ Help for distribution. where and f max m ¨ f [i  1]m  f [i  1]m ! §© max min © m fm  fm ª represent ¸ ¹ ¹ º where i ! 2. (l  1) T is the maximum value of function f m in the Pareto front.. ‡ Choose individuals having large crowding distance.. ! lC .3. ! g. . D. D. D.

.

.

.

14 2/13/2012 .     objective function value of solution.

6 5 20 15 4 3 x f1 .Crowding Distance Assignment 40 35 30 25 f (x) 2 ‡ Crowning distance can be calculated for all chromosomes of same Pareto front.

x f 2 .

308 -1.011 0.735 3.818 0.669 1.682 -2.618 10.308 1.508 2.467 0.105 16.x Rank 1 1 1 1 2 2 3 4 5 6 10 C.161 4.210 10.465 -1. -0.832 3.275 12.5 3 2 3 15 .07 3.829 0.414 0.708 -1.272 1.396 1.945 1.854 37.378 6 5 4 f (x) 3 2 1 0 1 g g g g g g g 2 i-1 i i+ 1 4 0 .347 0.218 2.171 5.671 17.D.5 1 1 .5 f 1 (x ) 2 2 .355 14.275 2/13/2012 g 5 0 0 2 1 2 4 6 8 f 1 (x ) 10 12 14 16 18 0.317 -4.

.

.

.

     .

Selection Selection is the stage of a genetic algorithm in which individual are chosen from a population for later breeding (recombination or crossover). i en j . Crowding operator based sorting The crowding operator e n guides the selection process at the various stages of the algorithm toward a uniformly spread-out Pareto optimal front.

irank jrank or .

irank ! jrank and .

2/13/2012 . is crowding distance of i th individual.iC . where irank shows non-domination rank & iC .D.D. " jC .D .

.

.

.

     16 .

‡ In tournament selection.Tournament Selection Runs a µtournament¶ among a few individuals chosen at random from the population and selects the winner (the one with the best fitness) for crossover. (Based on the crowding operator) x 0. a number Tour size of individuals is chosen randomly from the population and the best individual from this group is selected as parent.508 f1 .818 -1.

x f 2 .

396 12.378 x 0.818 f1 .275 1.30 1 3 C.x Rank 0.669 2.467 0. 1.D.

x f 2 .

669 1C .669 1.378 17 .945 1.396 1 1 C. 0. 1. 1.378 1rank ! 2rank 0. D.669 2.818 0.378 g 1rank 0.x Rank 0.818 2/13/2012 2rank 0.D.396 1 2C .396 1 1.218 0. D .347 1.

.

.

.

     .

Crossover Crossover is a genetic operator that combines (mates) two individuals (parents) to produce two new individuals (Childs). Simulated Binary Crossover ¨ 1 ¸ ® where . ‡ The idea behind crossover is that the new chromosome may be better than both of the parents if it takes the best characteristics from each of the parents.

5 ± ‡ r is a random number {0.5 ‡ j represent dimension of © 2 * .2 * r ©©ª Q 1 ¹¹º if r e 0.1} ± ¨ 1 ¸ © ¹ b !¯ ‡ Q is a crossover operator ª º ¨ ¸© Q 1 ¹ 1 ± © if r " 0.

 r ¹ ¹ ± 1 ª º ° individual. 1 child1 ( j ) ! .

.

 b * parent1 ( j )  .

 b * parent2 ( j ) 1 1 2 1 child 2 ( j ) ! .

.

 b * parent1 ( j )  .

 b * parent2 ( j ) 1 1 2 2/13/2012 .

.

.

.

     18 .

Mutation Mutation is a genetic operator that alters one ore more gene values in a chromosome from its initial state. Polynomial Mutation ¨ 1 ¸ ® © ¹ ± . ‡ Mutation is an important part of the genetic search as helps to prevent the population from stagnating at any local optima.

2 * r © L 1 ¹  1 ª º d !¯ ¨ 1 ¸ © ¹ ± .

2 * .

5 r " 0.5 where ‡ r is a random number {0.  r © L 1 ¹ 1 1 ª º ° if if r e 0.1} ‡ L is a mutation operator ‡ j represent dimension of individual. 19 child ( j ) ! parent ( j )  d 2/13/2012 .

.

.

.

     .

Qt} Elitist Replacement 20 .Selection for next generation Non-dominated sorting (Rank) F1 Crowding distance sorting Pt F2 F3 Pt+1 Qt Rejected 2/13/2012 Rt ={Pt .

.

.

.

     .

5 3 3 .5 3 2.5 2 2.5 3 3.5 4 f (x) 1 f 1 (x) After 10 generation 3 .5 1 1 .5 4 2/13/2012 f (x) 1 .5 2 .5 0 0 0 .5 1 1.Initial State 40 35 30 25 4 3.5 3 3 .5 1 0 .5 1 0.5 2 2 .5 1 0 .5 0 0 2 2 2 1 .5 0 0 2 4 6 8 10 12 14 16 18 0.5 2 2 .5 After 20 generation f (x) f (x) 2 2 20 15 10 5 0 0 2 1.5 3 2 .5 3 After 40 generation f (x) f (x) 2 1 .5 4 3 .5 1 1 .5 0 .

.

.

.

     f 1 (x) 21 .

Non-dominating Sorting Stopping criteria met ? Select N individuals 22 .Flowchart of NSGA-II NSGABegin : Initialize Population (N) Evaluate objective functions Non-dominated Sorting Tournament Selection No Crossover & Mutation Evaluate objective functions Combine parent and child populations .

.

.

.

     Yes Report Final Population and Stop 2/13/2012 .

1 0 .2 0 .Performance Measures There are two main goals in a multi-objective optimization: 1) Convergence to the Pareto-optimal set 2) Maintenance of diversity in solutions of the Paretooptimal set. 12 m in d is tan c e D IS T A N C E M E A S U R E BE T W E E N P A R E T O F R O N T S 10 T R U E P AR E T O F R O N T P AR E T O F R O N T Convergence metric fu n c tio n .2 8 6 m in d is tan c e 4 2 0 0 .4 2/13/2012 .3 0 .

.

.

.

8 0 .6 fu n c tio n .1 0 .7 0 .9 23 .     0 .5 0 .

. d n1 0 0 0.5 function -1 3 3. 3.5 The parameters and are the Euclidean distances between the extreme solutions of true Pareto front and the boundary solutions of the obtained non-dominated set.5 3 2. assuming that there are solutions on the best non.5 2 2.5 4 4.dominated front.5 1 1. The parameter is the average of all distances .5 Extreme solution Pareto front d1 d2 d3 d4 Extreme solution  If distance between the solutions is equal to average distance . DIVERSITY PLOT 4 Non uniformity in the distribution.5 1 0.Cont«.5 function -2 2 1. 2/13/2012 . that gives uniformly distribution.

.

.

.

     24 .

2/13/2012 . * These parameters were kept in all test functions optimization.9 Mutation Probability (Pm) = 0.1 Distribution index for crossover (µ)= 20 Distribution index for mutation ( ) = 20 Tour size (selection) = 2 MOPSO parameters     Population = 100 particles Repository (Archive) size = 100 particles Mutation rate = 0. * Only changed the total number of fitness function evaluations.Simulation and Results NSGA II parameters       Population (N) = 100 Crossover Probability (Pc)= 0.5 Divisions for Archive Grid = 30 Implementation use real numbers representation.

.

.

.

     25 .

5 2 F1 2.5 26 4 .5 3 3.5 1 1.5 3 3.5 4 0.5 2/13/2012 2 F1 2.Test Problem : SCH Nondominated solutions w ith MOPSO on SCH 5 Pareto-optimal Front MOPSO 4 Nondominated solutions w ith NSGA-II on S CH 5 Pareto-optimal Front NSGA-II 4 3 F2 3 F2 2 1 0 0 2 1 0 0 0.5 1 1.

.

.

.

     .

1: Results of the Convergence Metric for the SCH Test Function Diversity Metric Best Worst Mean NSGA-II 0. 2: Results of the Diversity Metric for the SCH Test Function ‘–ƒŽ—„‡”‘ˆˆ‹–‡••‡˜ƒŽ—ƒ–‹‘•™ƒ••‡––‘ 2/13/2012 .1569 0.6947 1.3575 0.Comparison for SCH Test Function Convergence Metric Best Worst Mean NSGA-II 0.9578 0.7904 0.0148 0.0093 0.2096 MOPSO 0.6425 MOPSO 0.5104 0.0259 Tab.8582 Tab.

.

.

.

     27 .

6 0.4 1.4 0.4 0.8 F2 F2 0.2 0.4 F1 0.8 0.2 0 0 N ondom ina ted solutions w ith N S G A-II on D E B -1 P are to-optim al Front N S G A-II 0.4 F1 0.2 1 0.6 0.6 0.8 1 2/13/2012 28 .2 0 0 0.6 0.4 1.2 0.2 1 0.8 1 P areto-optim al Front MO P S O 1.Test Problem : DEB-1 DEB- N ondom inated solutions w ith MO P S O on D E B -1 1.

.

.

.

     .

5140 0.0070 0.5938 Tab.1664 0.5140 0.7168 0.4243 MOPSO 0.3467 0.Comparison for DEB-1 Test Function DEBConvergence Metric Best Worst Mean NSGA-II 0.0079 Tab.5112 0. 2: Results of the Diversity Metric for the DEB-1 Test Function ‘–ƒŽ—„‡”‘ˆˆ‹–‡••‡˜ƒŽ—ƒ–‹‘•™ƒ••‡––‘ 2/13/2012 .0078 MOPSO 0.0066 0. 1: Results of the Convergence Metric for the DEB-1 Test Function Diversity Metric Best Worst Mean NSGA-II 0.

.

.

.

     29 .

1 0.4 0.1 0.5 F1 0.2 N ondom inated solutions w ith N S G A-II on D E B -2 P areto-optim al Front N AG A-II 0.9 1 P areto-optim al Front MO P S O 8 7 6 5 4 3 2 1 0 0.7 0.7 0.4 0.3 0.9 1 2/13/2012 30 .8 0.2 0.5 F1 0.8 0.6 0.Test Problem : DEB-2 DEB- N ondom inated solutions w ith M O P S O using D E B -2 8 7 6 5 F2 F2 4 3 2 1 0 0.6 0.3 0.

.

.

.

     .

7248 0.0608 Tab.0515 0.7939 0.0725 0.7582 0.7193 Tab. 2: Results of the Diversity Metric for the DEB-2 Test Function ‘–ƒŽ—„‡”‘ˆˆ‹–‡••‡˜ƒŽ—ƒ–‹‘•™ƒ••‡––‘ 2/13/2012 .0516 MOPSO 0. 1: Results of the Convergence Metric for the DEB-2 Test Function Diversity Metric Best Worst Mean NSGA-II 0.7597 MOPSO 0.0559 0.0449 0.Comparison for DEB-2 Test Function DEBConvergence Metric Best Worst Mean NSGA-II 0.6800 0.

.

.

.

     31 .

Test Problem : KUR N ondom inated solutions w ith MO P S O on K U R 2 0 -2 -4 F2 -6 -8 -10 -12 -20 -19 -18 -17 F1 -16 -15 -14 P areto-optim al Front MO P S O 2/13/2012 32 .

.

.

.

     .

6413 0.5602 Tab.0021 0.0034 0.0021 0.4803 0.7825 0.0028 MOPSO 0.Comparison for KUR Test Function Convergence Metric Best Worst Mean NSGA-II 0.1: Results of the Convergence Metric for the KUR Test Function Diversity Metric Best Worst Mean NSGA-II 0.0041 0.3344 0.4399 MOPSO 0.2: Results of the Diversity Metric for the KUR Test Function ‘–ƒŽ—„‡”‘ˆˆ‹–‡••‡˜ƒŽ—ƒ–‹‘•™ƒ••‡––‘ 2/13/2012 .0026 Tab.

.

.

.

     33 .

II a) Coverage b) Lifetime 2/13/2012 .Layout Optimization for a Wireless Sensor Network using NSGA .

.

.

.

     34 .

Wireless Sensor Network (WSN) Data Processing Unit (DPU) Node 1 High Energy Communication Node (HECN) Node 6 Node 9 Node 7 Node 8 Node 4 Node 5 Node 2 Node 3 Example of a WSN where sensor nodes are communicating with the DPU through HECN 2/13/2012 .

.

.

.

     35 .

..Area of the ROI 2/13/2012 .. 7i !1.Total number of nodes A . N Ai C! A Ai ..Area covered by the ith node N .Optimization of Coverage  Coverage is defined as the ratio of the union of areas covered by each node and the area of the entire ROI.

.

.

.

     36 .

 In every sensing cycle.Optimization of Lifetime  The lifetime of the whole network is the time until one of the participating nodes run out of energy. the data from every node is routed to HECN through a route of minimum weight Tfailure Lifetime ! Tmax Tfailure = maximum number of sensing cycles before failure of any node Tmax = maximum number of possible sensing cycles .

.

.

.

     .

Competing Objectives Lifetime Coverage HECN ‡ try to arrange the nodes as close as possible to the HECN for maximizing lifetime 2/13/2012 ‡ try to spread out the nodes for maximizing coverage 38 .

.

.

.

     .

9 0.5 20 20 2 2/13/2012 .Simulation Parameters Parameters of NSGA-II Number of chromosomes Number of generations Crossover Probability Mutation Probability Distribution index for crossover Distribution index for mutation Tour size 100 50 0.

.

.

.

     39 .

8 0.2 0.3 0.6 0.4 0.9 1 Coverage  Pareto Front obtained for a WSN with 10 sensors.1 0 0 0.3 0.1 0.2 0. 100 chromosomes and 50 generations 2/13/2012 .7 0.NSGANSGA-II Results Pareto optimal front 1 0.7 Lifetime 0.6 0.9 0.4 0.5 0.8 0.5 0.

.

.

.

     40 .

3335 Lifetime = 0.999 6 Best Lifetime 4 4 2 2 0 0 -2 -2 -4 -4 -6 -6 -4 -2 0 2 4 6 -6 -6 -4 -2 0 2 4 HECN 6 Coverage = 0.249 6 Best Coverage 4 2 0 -2 -4 -6 -6 -4 -2 0 2 4 2/13/2012 .NSGANSGA-II Results (Cont¶d) 6 Initial Disconnect Coverage = 0.63709 Network HECN Coverage = 0.5353 Lifetime = 0.

.

.

.

     HEC N 6 41 .

Radial Basis Function Network x1 x2 x3 x4 Input Layer  Accuracy  Complexity of the model .

.

.

.

     1 w0 w1 w2 w3 w4 w5 Ö di Hidden Layer Output Layer 42 .

Multiobjective Problem Formulation Structure determination of RBF network can be considered as the multiobjective optimization problem concerning with accuracy and complexity of the model. f1 ! M f 2 ! mse 1 n Ö ! § di  d i n i!1 n .

2 ¨M ¸¸ 1 ¨ © di  © § w jJ xi  c j  w0 ¹ ¹ ! § © ¹¹ n i!1 © ª j !1 ºº ª .

w : Weight vector of the RBF network c j : Center vector of the RBF network J . 2 here M is total number of basis functions (centers) in RBF network. d i: Desired output Ö d i : Estimated output during the training of RBF network.

. : Gaussian Function J x  cj .

1 ! exp¨  2W © ª 2 x  cj 2 ¸ ¹ º 43 where W is the spread of the Gaussian function. .

.

.

.

     .

the position of gene value ³1´ indicate the center position Desired Output 0 0 0 +1 0 0 of the basis function (selected center) and number of0³1´ genes in chromosome indicates the number of basis functions (number of centers).Structure selection of RBF network Selected centers Gaussian Distribution Input Data Points Chromosome Input Data Points 1 +1 0 0 0 0 +1 +1 0 0 1 -1 +1 +1 0 1 +1 -1 +1 +1 0 0 +1 -1 +1 0 0 0 +1 -1 1 0 0 0 +1 In the chromosome. 44 .

.

.

.

     .

MSE changes. ‡ The performance of an RBF network critically depends upon the chosen centers. 10 M e a n E q u a r e E r r o r (lo g s c a le ) 0 Pareto Front for 13-element Barker Code 10 -10 10 -20 10 -30 10 -40 2 4 6 8 10 12 14 16 18 Numbers of Centers 20 22 24 25 45 .Pareto Fronts ‡ This Pareto Front shows that for the different number of centers.

.

.

.

     .

3. Comput. 2. Deb. K. pp. ³Multiobjective function optimization using nondominated sorting genetic algorithms.References 1. Fall 1995. S. Z. Apr. IEEE Transaction on Evolutionary Computation. Agarwal. June 2004 46 . Nafploitis. Agrawal.´ Evol. pp.. Horn. ³A niched Pareto genetic algorithm for multiobjective optimization. Comput. 8. 3. ³Handling multiple objectives with particle swarm optimization´. Srinivas and K. Goldberg. IEEE. no. 82±87 .´ in Proceedings of the First IEEE Conference on Evolutionary Computation.W. E.´ in Complex Syst. and D. J. A. 4. vol. vol.´ Evol. 1995. Deb. vol. 149±172. Gregorio Toscano Pulido.´ A fast and elitist multi-objective genetic algorithm: NSGA-II´. Michalewicz. pp. Piscataway. N. 256±279. B. Meyarivan.. No. Comput. 115±148. Ed. Pratap. and Maximino Salazar Lechuga. Evol. Carlos A. 8. K. 2. Corne. ³Approximating the nondominated front using the Pareto archived evolution strategy. and T. 1994. Coello Coello. ³Simulated binary crossover for continuous search space.2002. 221±248. Member. 6(2). 9. N. 2000. vol. 3. J. Deb and R... pp. D. pp. 181-197.. Knowles and D. NJ: IEEE Press.

.

.

.

     5. 6. 2/13/2012 .

THANK YOU 2/13/2012 .

.

.

.

     47 .