Nsga-Ii

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
Multiobjective 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.
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)
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 -
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 Pareto 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
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.
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
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
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
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
-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
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
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
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

!
m
m m
m m
D C
f f
i f i f
i
min max
. .
] 1 [ ] 1 [
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
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
" !
2/13/2012 17
Runs a tournament 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

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.

"
e
!

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
!
!
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
!
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
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
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
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 ?
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
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 nondominated 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
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 ()= 20
Distribution index for mutation () = 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
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

MOPSO
Test Problem : SCH Test Problem : SCH
2/13/2012 26
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
Total numbei of fitness evaluations was set to
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

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

MOPSO
Test Problem : DEB Test Problem : DEB- -11
2/13/2012 28
Best 0.0066 0.0070
Worst 0.5140 0.1664
Mean 0.0078 0.0079
Best 0.3467 0.5112
Worst 0.5140 0.7168
Mean 0.4243 0.5938
DEB-1 Test Function
DEB-1 Test Function
Comparison for DEB Comparison for DEB- -1 Test Function 1 Test Function
2/13/2012 29
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

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

NAGA-II
Test Problem : DEB Test Problem : DEB- -22
2/13/2012 30
Best 0.0449 0.0515
Worst 0.0559 0.0725
Mean 0.0516 0.0608
Best 0.7248 0.6800
Worst 0.7939 0.7582
Mean 0.7597 0.7193
DEB-2 Test Function
DEB-2 Test Function
Comparison for DEB Comparison for DEB- -2 Test Function 2 Test Function
2/13/2012 31
-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

MOPSO
Test Problem : KUR Test Problem : KUR
2/13/2012 32
Best 0.0021 0.0021
Worst 0.0041 0.0034
Mean 0.0028 0.0026
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
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
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
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
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
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
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
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
NSGA NSGA- -II Results (Contd) II Results (Contd)
-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
Input Layer Hidden Layer Output Layer

Radial Basis Function Network Radial Basis Function Network
Accuracy
Complexity of the model
i
d
1
x
2
x
3
x
4
x
1
w
2
w
3
w
4
w
5
w
0
w
1
42
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
! !
!
!
!
!
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
i
d
w
j
c
. J
43

!
2
2
2
1
exp
j j
c x c x
W
J
where is the spread of the Gaussian function.
W
Desired Output
0 0 0 +1 0 0 0
In the chromosome, the position of gene value 1 indicate the center position
of the basis function (selected center) and number of 1 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
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
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, Simulated binary crossover for continuous search space, in
Complex Syst., vol. 9, pp. 115148., Apr. 1995.
3. N. Srinivas and K. Deb, Multiobjective function optimization using nondominated sorting
genetic algorithms, Evol. Comput., vol. 2, no. 3, pp. 221248, Fall 1995.
4. J. Horn, N. Nafploitis, and D. E. Goldberg, A 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. 8287 , 1994.
5. J. D. Knowles and D.W. Corne, Approximating the nondominated front using the Pareto
archived evolution strategy, Evol. Comput., vol. 8, pp. 149172, 2000.
6. Carlos A. Coello Coello, Member, IEEE, Gregorio Toscano Pulido, and Maximino Salazar
Lechuga, Handling multiple objectives with particle swarm optimization, Evol. Comput.,
vol. 8, pp. 256279, No. 3, June 2004
2/13/2012 46
THANK YOU
2/13/2012 47

Nsga-Ii

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Nsga-Ii

Uploaded by

Copyright:

Available Formats

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

Dean Academic Affairs Dean Academic Affairs

Input Layer Hidden Layer Output Layer

You might also like