Вы находитесь на странице: 1из 25

MOGADES: Multi-Objective Genetic Algorithm

with Distributed Environment Scheme


Intelligent Systems Design LaboratoryDoshisha UniversityKyoto Japan

Jiro KAMIURA
Tomoyuki HIROYASU,
Mitsunori MIKI,
Shinya WATANABE
Doshisha Univ., Kyoto Japan
2
Multi-objective Optimization Problems : MOPs
In the optimization problems, when there are several objective
functions, the problems are called multi-objective problems.
f
1
(x)
f
2

(
x
)

Objective function
Constraints
Gi(x)<0 ( i = 1, 2, , k)
F={f1(x), f2(x), , fm(x)}
X={x1, x2, . , xn}
non-dominated solutions
Design variable
Solving MOPs needs huge calculation costs,
so we need the parallel model for solving MOPs.
f1(x) : Minimize
f2(x) : Minimize
Doshisha Univ., Kyoto Japan
3
Multi-Objective Genetic Algorithms : MOGAs
VEGA : Schaffer (1985)
MOGA : Fonseca (1993)
SPEA2 : Zitzler (2001)
NPGA2 : Erickson, Mayer, Horn (2001)
NSGA-II : Deb, Goel (2001)
Typical method on MOGAs
Genetic Algorithm for solving MOPs
None of all is parallel model
Doshisha Univ., Kyoto Japan
4
MOGADES : Multi-Objective Genetic Algorithm
with Distributed Environment Scheme
Distributed Genetic Algorithm
(enable to implement on parallel computers)
Unification of objective functions using a weighted-sum
Adaptive change of the weight parameters
Neighborhood migration
Archive of the excellent solutions
Features
Proposed method : MOGADES
Doshisha Univ., Kyoto Japan
5
Distributed Genetic Algorithm : DGA (Tanese 89)
Migration: Exchange of individuals
among islands
DGA can show better performance than single population GAs
in solving single objective problems.
A population is divided into smaller
subpopulations (islands)
One of the parallel models of GAs
Canonical GA is performed in each
island
Distributed Environment Scheme
(Miki 1999): the environment (that
is crossover rate, mutation rate, and
so on) in each island are different.
Doshisha Univ., Kyoto Japan
6
Assignment of fitness using the weighted-sum of each
objective function
Using Distributed Environment Scheme :
Weight parameters are different in each island.
Unification of the objective functions
f
2
(
x
)

f1(x)
) (
1
x f w i i
k
i

=
= >
k
i
k i w w
1
1 , 0
Fitness value =
:the number of objective functions
k
:the weight parameter of the ith objective function
i w
:the value of the ith objective function
) (x i f
The searching
directions
Doshisha Univ., Kyoto Japan
7
Assignment of weight parameters
The weight values are arranged equally from 0.0 to 1.0.
e.g.) 2 objective functions, 5 islands
island 1 w 2 w
1.0 0.0 1
0.75 0.25 2
0.5 0.5 3
0.25 0.75 4
0.0 1.0 5
searching
direction
f
2
(
x
)

f1(x)
Doshisha Univ., Kyoto Japan
8
Adaptive change of the weight parameters
f
2
(
x
)

f1(x)
e.g.) 2 objective functions, 3 islands
To get good distributed non-dominated solutions.
Performed in the migration phase.
f
2
(
x
)

f1(x)
Change
Island 1
Island 2
Island 3
Island 2
Island 1
Island 3
distance
d1
d2
Doshisha Univ., Kyoto Japan
9
Neighborhood migration
Exchange individuals with neighborhood islands.
The weight values of islands change.
i w
Step 1. Sort islands by . i w ( changes for each migration phase.)
i
neighborhood
3
island
2, 4
Step 2. Migrate with neighborhood islands.
Step 3. Change the weight values of each islands.
Doshisha Univ., Kyoto Japan
10
Archive of the excellent solutions
Archive of
the non-dominated solutions
the solutions which have good fitness
: non-dominated solutions
: solutions which have good fitness
: searching direction
f
2
(
x
)

f1(x)
: individuals
Doshisha Univ., Kyoto Japan
11
The overview of MOGADES
f
2
(
x
)

f1(x)
searching direction
changed weight
neighborhood
migration
non-dominated
archive
individual
island
Doshisha Univ., Kyoto Japan
12
ZDT4
Continuous
2 objective functions
10 design variables
Multi-modal
] 5 , 5 [ ] 1 , 0 [
) 4 cos( 10 91 ) (
) (
1 ) ( ) ( min
) ( min
1
10
2
2
1
2
1 1
e e
+ =
|
|
.
|

\
|
=
=

=
i
i
i i
x x
x x x g
x g
x
x g x f
x x f
t
Test Problems
Doshisha Univ., Kyoto Japan
13
KUR
Continuous
2 objective functions
100 design variables
Multi-modal
( )
100 , , 1 , ] 5 , 5 [
) sin( 5 | | ) ( min
)) 2 . 0 exp( 10 ( ) ( min
3 8 . 0
2
100
1
2
1
2
1
= = e
+ =
+ =

=
+
n n i x
x x x f
x x x f
i
i i
i
i i

Test Problems
Doshisha Univ., Kyoto Japan
14
Objectives
Constraints
0/1 Knapsack Problem (750items 3knapsacks)
Combination problem
3 , 2 , 1 ) ( max
750
1
,
e =

=
i x p x f
j
j j i i

=
s
750
1
,
j
i j j i
c x w
1 , 0 ) , , , (
750 2 1
e =
j
x x x x x
p
i,j
= profit of item j according to knapsack i
Test Problems
w
i,j
= weight of item j according to knapsack i
c
i,
= capacity of knapsack i
Doshisha Univ., Kyoto Japan
15
Applied models and Parameters
Applied models
Crossover
2 points crossover
Mutation
bit flip
Migration Interval
10 generations
SPEA2
NSGA-II
MOGADES
population size 100(10islands)
crossover rate 1.0
mutation rate 1/(chromosome length)
Parameters
terminal condition 50000
250(25islands)
1000000
number of trials
30
ZDT4 KP750-3
chromosome length
KUR
200 2000 750
100000
Doshisha Univ., Kyoto Japan
16
ZDT4
MOGADES is superior to
NSGA-II and SPEA2
Doshisha Univ., Kyoto Japan
17
KUR
MOGADES is superior to
NSGA-II and SPEA2
Doshisha Univ., Kyoto Japan
18
KP750-3
MOGADES is superior to
NSGA-II and SPEA2
18
Doshisha Univ., Kyoto Japan
19
We proposed a new model of MOGA.
MOGADES: Multi-Objective Genetic Algorithm with
Distributed Environment Scheme
Conclusion
MOGADES was compared to SPEA2 and NSGA-II in
3 test functions.
In all of the test functions in which we compared to,
MOGADES derives the good results.
MOGADES is good model for solving MOPs.
MOGADES is based on Distributed Genetic Algorithm
which is one of the parallel models, so MOGADES is
the parallel model, too.
Doshisha Univ., Kyoto Japan
20
Doshisha Univ., Kyoto Japan
21
ZDT6
Continuous
2 objective functions
10 design variables
Non-convex
] 5 , 5 [ ] 1 , 0 [
1
9 1 ) (
1 ) ( ) ( min
) 6 ( sin ) 4 exp( 1 ) ( min
1
25 . 0
10
2
2
2
1
2
1
6
1 1
e e
|
|
|
|
.
|

\
|

+ =
|
|
.
|

\
|
|
|
.
|

\
|
=
=

=
i
i
i
x x
N
x
x g
g
f
x g x f
x x x f t
Test Problems
Doshisha Univ., Kyoto Japan
22
Assignment of weight parameters
As many as possible, the weight values are arranged
equally from 0.0 to 1.0.
In the rest of the islands, the weight values are
assigned randomly.
e.g.) 3 objective functions
6 islands
1 w
2 w 3 w
8 islands
10 islands
Random
0
1
0
1
0
1
0.5, 0.0, 0.5
0.5, 0.5, 0.0
0.0, 0.5, 0.5
0.0, 1.0, 0.0
1.0, 0.0, 0.0
0.0, 0.0, 1.0
Doshisha Univ., Kyoto Japan
23
The flow of MOGADES
initialization
evaluation ( includes reservation of the excellent solutions)
selection for reproduction
crossover
mutation
selection for survival
evaluation
neighborhood migration
migration interval
terminal check
end
0 P
0 E
: population
t P
t E
: excellent solutions
t C
t t t C E P +
2 individuals are selected from Pt + Et
by tournament selection
: parents
t C'
: offsprings
1 + t E
1 ' + + t t t P C C
t C'
2 individuals are sampled without
replacement from Ct + Ct
and replace bad 2 individuals of Pt.
Doshisha Univ., Kyoto Japan
24
Adaptive change of the weight parameters
Weight values are changed by following equation
) 1 , ( ) , 1 (
) , 1 (
) 1 (
) 1 , ( ) , 1 (
) 1 , (
) 1 ( '
+

+
+
+
+
+
+
=
n n n n
n n
n
n n n n
n n
n n
d d
d
w
d d
d
w w
n w
) , ( b a d : distance between islands a and b.
: weight value of nth island.
f
2
(
x
)

f1(x)
f
2
(
x
)

f1(x)
Change
Island 1
Island 2
Island 3
d(1,2
)
d(2,3
)
Doshisha Univ., Kyoto Japan
25

Вам также может понравиться