TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.4, April 201
4, pp. 3148 ~ 3
1
5
7
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i4.4785
3148
Re
cei
v
ed Au
gust 28, 20
13
; Revi
sed
No
vem
ber 1
7
, 2013; Accepte
d
De
cem
ber
7, 2013
An Improved Evolutionary Algorithm with Ne
w Gene
tic
Operation for Optimization Problem
Wang Jiekai, Hu Ruikai, Wang
Chao*
Coll
eg
e of Mathematics Sci
e
n
c
e, Harbi
n
Nor
m
al Univ
ersit
y
,
Harbi
n
, Hei
l
on
gjia
ng Prov
inc
e
, P.R. China
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: hsd
w
a
ngc
ha
o@1
26.com
A
b
st
r
a
ct
An improve
d
e
v
oluti
onary a
l
g
o
rith
m (SCAGA) is
propose
d
in this paper f
o
r solvin
g opti
m
i
z
at
io
n
prob
le
m. In or
d
e
r to co
ntrol g
e
netic o
per
ation
s
in a
n
effectiv
e ran
ge, th
e ne
w
algorith
m
r
e
gul
ate b
o
th of t
h
e
crossover prob
abil
i
ty
an
d mut
a
tion prob
ab
ilit
y
w
i
th
t
he
iteration process. In
add
ition, SC
AGA presents
a
new
crossover
strategy that
restricts the cros
s of t
he chro
moso
mes to s
o
me
extent to
protect goo
d g
e
n
e
s
sche
m
a. W
e
also
perfor
m
the sch
e
m
a t
heor
e
m
o
n
th
e al
gor
ith
m
p
r
ocess to
an
a
l
y
z
e
th
e w
o
rki
n
g
mec
h
a
n
is
m of
SCAGA, and
w
e
conc
lu
de
that the new
alg
o
rith
m is ef
fective. Accord
ing to ex
per
i
m
ent
results for
some test functions
an
d
T
SP pro
b
le
ms,
S
C
AGA
hav
e a hig
h
p
e
rformanc
e
i
n
both
c
onstra
i
n
e
d
unco
n
strain
ed opti
m
i
z
at
ion
pr
obl
e
m
s.
Ke
y
w
ords
:
ev
oluti
onary
al
go
rithm, cr
ossov
e
r op
erator,
mu
tation
op
erat
or, crossover
strategy, sche
m
a
theore
m
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
Geneti
c
al
gorithm (GA
for
sho
r
t) i
s
a
ki
nd of
typi
cal
evolutiona
ry
algorith
m
[1]. It is
a
stocha
stic
se
arch alg
o
rith
m for glob
al
optim
izatio
n
based on
pri
n
cipl
e co
me
s from evoluti
o
n
mech
ani
sm i
n
nature.
Wh
en GA is u
s
e
d
to comp
a
r
e
with other traditional alg
o
r
ithms, its uni
que
codi
ng pattern and searching metho
d
can often
be
more effective [2]. As a re
sult, it is
approp
riate f
o
r o
p
timizatio
n
problem
in
compl
e
x
sy
stem [3-5]. Ho
wever,
gen
etic al
gorith
m
al
so
has
som
e
sh
ortco
m
ing
s
such a
s
lo
cal
optimal
solutions, lo
we
r converg
e
n
c
e speed et
c, wh
ich
limit the u
s
e
of gen
etic
alg
o
rithm
se
riou
sly. Fo
r such
pro
b
lem, m
a
ny effective i
m
provem
ent
for
geneti
c
alg
o
ri
thm have be
e
n
pro
p
o
s
ed i
n
re
cent ye
a
r
s.
And mo
st re
sea
r
ch ab
out
GA focu
se
d o
n
the geneti
c
o
perato
r
an
d the
execution
strategy [6-8].
The sta
nda
rd
geneti
c
op
erators
actu
ally pr
ovide
a ki
nd of sto
c
h
a
s
tic
sea
r
ch p
r
ocess
with un
certai
n prob
ability. While the
operat
ors ca
n give each
individual
a cha
n
ce for
optimizatio
n, it also have
a possibility lead into
the
recessio
n of the grou
p [9]. To redu
ce the
blindn
ess of geneti
c
ope
rations, an im
proved a
dapt
ive genetic a
l
gorithm is p
r
opo
sed in this
pape
r. Mea
n
w
hile, the
propo
sed
algo
rithm use a
n
e
w
crossove
r strate
gy so
as to
ma
ke t
h
e
algorith
m
more efficient.
The
re
st of th
is p
ape
r i
s
organi
zed
a
s
fo
llows. In Se
ct
ion 2,
we
introdu
ce th
e im
proved
geneti
c
algo
ri
thm SCAGA. In Section 3,
we p
r
e
s
ent a
theoreti
c
al a
nalysi
s
use schem
a theo
rem
for the al
gori
t
hm. In Section 4, expe
ri
ment
al results both
on b
enchma
r
k fu
nction
and
T
SP
probl
em are given and di
scu
s
sed. Finall
y
, we con
c
lud
e
in Section 5
.
2. Impro
v
ed
Adap
tiv
e
Genetic Algo
rithm
2.1. Adap
tiv
e
Gene
tic Al
gorithm
Due to the e
s
sen
c
e of evolution is a d
y
namic
p
r
o
c
e
ss,
s
o
me r
e
s
ear
che
r
s t
h
in
k relat
e
d
geneti
c
para
m
eters sh
oul
d be reg
u
late
d adaptively
rathe
r
than b
e
ing invari
abl
e. The adapti
v
e
geneti
c
algo
ri
thm (AGA) p
r
opo
sed by Srinvas c
ould
regulate
cro
ssover and m
u
tation pro
babil
i
ty
to get GA more efficient [10]. But it still
has
so
me disadvantages such
as
low convergence rate
and hig
h
pro
bability to local conve
r
ge
n
c
e.In re
cent
years, many rese
arche
r
s
in
various field
s
try
to find more
efficient meth
od to improve
AGA. Some
comp
re
hen
si
ve methods
could be fou
n
d
in
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
An Im
proved
Evolution
a
ry
Algorithm
with Ne
w Gen
e
tic Ope
r
ation f
o
r… (Wa
ng Ji
ekai
)
3149
[11-12]. And
there
are al
so
som
e
works
based o
n
he
u
r
istic me
cha
n
i
sm turn out
be efficie
n
t [13-
14].
In this p
ape
r, we p
r
e
s
ent
a ne
w im
proved ad
aptive gen
etic
alg
o
rithm S
C
AG
A. The
prop
osed ne
w algo
rithm is based on
adju
s
ting mut
a
tion and cro
s
sover proba
bility to regulate
the evolution
process. Before iteratio
n, SCAGA
use
a new initiali
zation meth
o
d
to make su
re
that the initiali
zed
pop
ulatio
n ca
n di
stribu
te in
the solut
i
on spa
c
e ev
enly. And the
algorith
m
cou
l
d
optimize
crossover and m
u
tation probability dynam
i
c
ally by i
m
proved
cr
ossov
e
r and
mutati
on
operators. A
seri
es of ne
w met
hod
s to
regulate th
e g
enetic
ope
ra
t
o
rs dynami
c
a
lly were ap
pli
e
d
so a
s
to improve the perfo
rman
ce of the
propo
se
d alg
o
rithm.
2.2. Population Initializati
on
In gen
eral,
way of gen
etic algo
rithm fo
r pop
ulation i
n
itialization
is com
p
letely random.
Ran
domly initial populatio
n
tends to make indivi
dual
s unevenly dist
ributed in the
solutio
n
spa
c
e.
It is ea
sy to
cause the i
m
b
a
lan
c
ed
evol
ution of th
e
a
l
gorithm
in th
e iteratio
n p
r
oce
s
s a
nd
re
sult
in local opti
m
al solutio
n
s. In order to overco
me this problem,
we use a
new meth
od
with
uniform
soluti
on sp
ace to make the init
i
a
l individual
s can b
e
evenl
y distributed.
Step 1: Divide the solutio
n
spa
c
e avera
gely into
N su
bsp
a
ces a
c
co
rding to the p
opulatio
n si
ze
.
Step 2: Each sub
s
p
a
ce ge
nerate
sub
-
in
dividual
s with
completely random
way.
Step 3: Comp
ose all the
su
b-individ
ual
s and get the in
itial populatio
n.
The in
dividual
s of initial popul
ation con
s
i
s
t o
f
A(0)=
{
A
1
(0)
1
,A
2
(0)
1
,…,A
n
(0
)
1
}
∪
{A
1
(0
)
2
,A
2
(0)
2
,…,A
n
(0
)
2
}
∪∪
…{
A
1
(0
)
N
,A
2
(0)
N
,…,A
n
(0)
N
}
∈
R
N
. This
approa
ch
ca
n improve th
e diversity o
f
populatio
n
on the p
r
em
ise of initiali
zing
pop
ulati
o
n
rand
omly. And it can impro
v
es the co
nver
ge
nce pe
rforma
nce of propo
se
d algo
rithm.
2.3. The Improv
ement of Cros
sov
e
r and Muta
tion
Opera
t
or
Acco
rdi
ng to
the schem
a theorem [1], the natu
r
e of t
he gen
etic al
gorithm
sho
u
l
d
be the
repla
c
e
m
ent
of the origin
a
l
sch
ema a
n
d
the form
ati
on of excell
e
n
t sch
ema. I
n
the pro
c
e
s
s o
f
geneti
c
op
eration, it sho
u
l
d ke
ep a
ne
w sch
e
ma a
s
mu
ch
as
p
o
ssible i
n
th
e early time
of
popul
ation evolution, in order to mainta
in populat
io
n
diversity. In
the later peri
od of populat
ion
evolution, it
should
try to
maintain
an
a
ppro
p
ri
ate m
ode
and
p
r
ev
ent the
de
struction, to
p
r
e
v
ent
the algorithm
from
premat
urity. Individuals
hav
e
a greater probability of high adaptability
to
contai
n the f
i
ne sch
e
ma,
so they a
r
e
more
su
ita
b
l
e for the
rel
a
tively small
cro
s
sove
r a
n
d
mutation p
r
o
bability du
rin
g
the
evoluti
on. On
t
he
contra
ry, low fitness in
dividuals not
only is
unlikely to contain fine schema, but they may
have
a possibility to dam
age the fine schem
a by
hybridi
z
ing
wi
th individu
als co
ntaine
d fin
e
sch
e
ma. T
herefo
r
e, th
e
low fitne
s
s a
r
e mo
re
suita
b
le
for small
e
r
cro
s
sove
r probability. For the almo
st
average fitness, we
ca
n increa
se their
crossover and mutation probab
ility properly to
expl
ore the
potential fine
schema i
n
these
individual
s.
In this pape
r we uses fo
rmula (1
) as
follo
w that d
e
crea
se
s pro
g
re
ssively while the
algorith
m
iterating:
ma
x
si
n(
(
1
)
)
2
ge
n
x
ge
n
(1)
Whe
r
e
ge
n
rep
r
e
s
ent th
e cu
rrent ev
olution g
ene
ration,
gen
ma
x
rep
r
e
s
ent t
he preset tota
l
evolution ge
n
e
ration.
More formally, let
N
be t
he po
pulatio
n scale,
a
b
e
the p
e
rcen
tage that a
particula
r
individual fitn
ess a
c
count
s for the
sum
of all individ
ual’s,
F
and
F
avg
represen
t the individu
al
fitness a
nd a
v
erage fitne
s
s of popul
atio
n.
Apparently, whe
n
a i
s
cl
ose
r
to
1/
N
,
we
can
see t
he pa
rticular
individual fitn
ess
F
is
clo
s
er t
o
F
avg
, thus (1
+Na
)
is cl
oser to
2. Acco
rdin
g
to the mentio
ned ab
ove, at this time th
e
cro
s
sove
r p
r
obability
sho
u
ld b
e
hig
h
e
r
. Conve
r
sely, wh
en th
e
F
is mo
re
far f
r
om
F
avg
, nam
ely
(
1+Na
) is
clo
s
er to 1, the
cro
s
sover
probability sh
o
u
ld be
small
e
r. The follo
wing fo
rmula
are
use
d
in this p
aper:
(1
)
yl
n
N
a
(2)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 3148 – 3
157
3150
In orde
r to regulate the
crossove
r pro
bab
ility and
mutation pro
bability in an
effecti
v
e
rang
e, we set up follo
wi
ng up
dating
formula fo
r
cross
pro
babil
i
ty accordi
n
g
to the
sigm
od
function:
ma
x
s
i
n
(
(1
)
)
l
n
(1
)
2
gen
zx
y
N
a
ge
n
(3)
1
'
1
c
z
p
e
(4)
Therefore, the crossover proba
bility of two individual
s is:
12
''
2
cc
c
pp
p
(5)
Becau
s
e m
u
tation ope
rato
r ca
n distu
r
b
the per
fo
rm
ance of the algorith
m
to a ce
rtain
extent, the determin
a
tion
of mutation p
r
oba
bility is
very importa
nt. In general,
the sele
ction
o
f
mutation probability is m
a
i
n
ly got according to expe
ri
ence, so it
s reliability is oft
en suspectable.
Accordi
ng to
the followi
ng formula,
the mutation probability
could be
adjusted
dynamically i
n
a
better way.
ma
x
s
i
n(
(
1
)
)
0.
5
2
m
ge
n
p
ge
n
(6)
Where pm
represent the curr
ent mutation probability
.
The value of
pm will decrease gradually
with
the iteration process. In this way, it ca
n
avoid getting into the local optimum in the early
peri
o
d
,
and it can also
improve the conve
r
ge
nce
perfo
rman
ce
of SCAGA in the later pe
rio
d
.
2.4. The Improv
ement of Cros
sov
e
r Strateg
y
Cro
s
sove
r op
erato
r
also h
a
s
a great influ
enc
e on the converg
e
n
c
e rate of GA. However,
traditional
cro
s
sover op
erat
ion often has
a certai
n blin
dne
ss. The of
fsprin
g individ
uals
which are
formed
by crossover of p
a
r
ental
ch
rom
o
som
e
s may
discard o
r
d
e
s
tru
c
t the fin
e
gene
s
sche
ma
existed i
n
pa
rental
individ
uals.
The
r
efo
r
e, in
ad
ditio
n
to th
e a
d
ju
stable
cro
s
so
ver a
nd
muta
tion
probability, this paper
al
so presents a
new
crossover stra
tegy based on
our
rel
a
ted works [15
-
16]. So that the fit schem
a
existed i
n
pa
rental
gen
er
a
t
ion is
prote
c
t
ed to tra
n
sfe
r
to offsp
r
ing
as
far as
poss
ible.
Definition 1 (Codi
ng Di
sta
n
ce
): Suppo
se that two individual
s are
enco
ded in
binary
respe
c
tively,
and the codin
g
length is
L
,
we say the co
ding di
stan
ce
of
X
and
Y
is
:
,
1
()
L
X
Ym
m
m
kc
(7)
Whe
r
e
k
is th
e given wei
g
h
t,
,,
,,
1,
0,
X
mY
m
m
X
mY
m
aa
c
aa
Definition 2 (Codi
ng simil
a
rity): The c
o
d
i
ng simila
rity betwe
en two
individual
s
X
and
Y
can b
e
com
p
uted with formula (8
).
,
1
(,
)
/
L
X
Ym
m
s
SX
Y
k
(8)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
An Im
proved
Evolution
a
ry
Algorithm
with Ne
w Gen
e
tic Ope
r
ation f
o
r… (Wa
ng Ji
ekai
)
3151
Obviou
sly, th
e value of coding simil
a
rity of any two in
dividual
s is b
e
twee
n [0,1].
In orde
r
to control the cro
s
sove
r op
eration, he
re
we u
s
e a critical value a
s
referen
c
e.
Definition 3 (Standard Cro
s
sover Poin
t): The stand
ard cro
s
sove
r p
o
int (
sc
p
) is a critical
value to de
cide weathe
r
the cro
s
sove
r op
eratio
n
sho
u
ld b
e
i
m
pleme
n
ted,
and it
can
be
comp
uted a
s
follows:
ma
x
ma
x
2
3
g
en
g
e
n
scp
ge
n
(9)
From the formula (9
),
sc
p
will contin
uou
sly incre
a
se
with the iteration of geneti
c
algorith
m
. Based o
n
above
definition, the new
cro
s
so
ver strate
gy can be de
scrib
ed as:
The valu
e of
codi
ng
simila
rity between
i
ndividual
s i
s
relatively lowe
r in the
pri
o
r
stage
of
SCAGA. In o
r
de
r to en
su
re that the out
standi
ng g
e
n
e
tic sch
e
ma
will not b
e
broke
n
for th
e full
evolution of chromo
som
e
group,
scp
value sh
o
u
l
d
be relativ
e
ly low co
ntrolled
crosso
ver
operation. On
the co
ntra
ry, in t
he late
st
age
s of SCA
G
A, the difference bet
wee
n
individu
als
can
be very small
,
so the
sc
p
shall
also in
crea
se. Acco
rding to formu
l
a (9
), the in
terse
c
tion
of
th
e
dynamic
cont
rol stand
ards can
hel
p imp
r
ove the
effici
ence a
nd
co
n
v
ergen
ce
pe
rf
orma
nce of t
h
e
algorith
m
.
2.5. Impro
v
e
d
Adap
tiv
e
Gene
tic Alg
o
rithm
The main ope
ratio
n
s of the prop
ose
d
algo
rith
m SCAGA ca
n be sum
m
ari
z
ed a
s
follo
ws:
Step 1: Initialize po
pulatio
n
,
initial popula
t
ion con
s
i
s
t of
A(0)=
{
A
1
(0
)
1
,A
2
(0)
1
,…,A
n
(0
)
1
}
∪
{A
1
(0
)
2
,
A
2
(0)
2
,…,A
n
(0
)
2
}
∪∪
…{
A
1
(0)
N
,A
2
(0)
N
,…,A
n
(0)
N
}
∈
R
N
;
Step 2: Evaluation of the fitness value of
each in
dividu
al;
Step 3: Selection operator:
Perf
orm sele
ction op
erato
r
and get
A(t)={A
1
(t),A
2
(t),…,A
n*N
(t)}
;
Step 4: Cro
s
sover ope
rato
r: Perform the cro
s
sove
r op
eration a
nd g
e
t
A’
(
t
)={
A’
1
(
t
)
,A
’
2
(
t
)
,
…
,A
’
n*
N
(
t
)}, if
the cro
s
so
ver ope
ration
con
d
ition is
satisfied, wh
ere the
cro
s
sove
r pro
bability can b
e
cal
c
ulate
d
according to formul
a (3
)(4
)
and (5
);
Step 5: Mutation ope
rato
r: Perform the
mutation ope
ration acco
rdi
ng to formula
(6) a
nd turn
individual
s into
A’
’
(
t
)
={A’
’
1
(t)
,
A’
’
2
(t
)
,,
…A
’
’
n*N
(t)}
;
Step 6: Repe
at Step3~ Step5 until a st
o
pping
crite
r
io
n is sati
sfied;
Step 7: Output the best so
lution of all individual
s.
3. The Sche
ma Theorem
Analy
s
is for SCAG
A
The sche
ma
theore
m
ha
s been a effici
e
n
t method for analyzin
g ge
netic alg
o
rith
m. The
schema th
eo
rem p
r
edi
cts
that gro
w
th o
f
high fi
tness has a g
ood
effect on alg
o
rithm p
r
o
c
e
ss,
and in
dicates sho
w
crosso
ver and
muta
tion effect
on
the propag
ation of ge
netic
schema. In th
is
se
ction, we a
nalyze p
r
op
o
s
ed al
gorith
m
with sc
hema
theore
m
. The
symbols u
s
e
d
is as follo
ws.
H
: a present schema
f(H)
: the fitness value of sche
ma
f
: the average
fitness valu
e of populatio
n
H
f
: the average
fitness valu
e of sch
ema
H
s
≤
scp
C
l
one
th
e
pa
re
ntal
Ch
r
o
m
o
s
o
me
it
se
lf
N
o
Ye
s
Perform t
h
e
Crosso
ver ope
ration
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 3148 – 3
157
3152
m
(
H,gen
+1)
: the number
of expected o
ffspring
creat
ed in gen
erati
on
gen
+1
of schem
a
H
m
(
H,gen)
: the number
of sch
ema
H
i
n
gene
ration
gen
+1
δ
(H
)
: the defining length of sch
e
ma
H
O(H)
: the order of
schema
H
l
: the length of the enco
ded
solutio
n
M(H,g
en)
: the number
of solution of
gene
ration g
e
n
of sch
ema
H
Suppo
se that
the sele
ction
operato
r
hav
e impl
emente
d
, we here to con
s
ide
r
the
effect
of cro
s
sover
operator a
nd
mutation ope
rator.
Acco
rdi
ng to
the sch
e
m
a
theo
rem, t
he
nu
mbe
r
of expecte
d
offspri
ng
created in
gene
ration
gen
+
1
of sch
e
m
a
H
is given b
y
:
()
(
)
(,
1
)
(,
)
[
1
]
1
c
f
HH
mH
g
e
n
m
H
g
e
n
p
fl
(10)
The pa
ramet
e
r
P
c
is
given by (5), (6) and (7).
To sim
p
lify a
nalysi
s
p
r
o
c
e
dure,
we
assume that P’
c1
is e
qual
s to
P’
c2
, therefore, after
sub
s
tituting for Pc in form
ula (10
)
, we
could get:
ma
x
s
i
n
(
(1
)
)
l
n
(1
)
2
()
1
(
)
(,
1
)
(,
)
[
1
]
1
1
ge
n
Na
ge
n
f
HH
mH
g
e
n
m
H
g
e
n
fl
e
(11)
No
w we
coul
d get the
M(H,gen
)
, and to estimate its value, we co
nsid
er to acq
u
ire the
summ
ation of
m
(
H,gen)
ov
er all sol
u
tion
s.
(,
)
1
(,
1
)
(,
1
)
MH
g
e
n
i
MH
g
e
n
m
H
g
e
n
(12)
From formula
(12), we co
ul
d get:
ma
x
(,
)
sin
(
(
1
)
)
l
n
(
1
)
1
2
()
1
(
)
(,
1
)
(,
)
[
1
]
1
1
MH
g
e
n
ge
n
Na
i
ge
n
f
HH
MH
g
e
n
m
H
g
e
n
fl
e
(13)
From the ine
quality (13
)
, we get:
ma
x
(,
)
s
i
n
(
(1
)
)
l
n
(1
)
1
2
()
1
(
)
(,
1
)
[
1
]
1
1
MH
g
e
n
ge
n
Na
i
ge
n
f
HH
MH
g
e
n
fl
e
(14)
Con
s
id
er that
H
gen
H
M
i
f
gen
H
M
H
f
,
,
1
, we can m
o
d
i
fy (14) as:
ma
x
sin
(
(
1
)
)
l
n
(
1
)
2
()
1
(
)
(,
1
)
(,
)
[
1
]
1
1
ge
n
Na
ge
n
f
HH
MH
g
e
n
M
H
g
e
n
fl
e
(15)
No
w co
nsi
der the cro
s
sove
r strate
gy pro
posed in (9
), (10
)
in se
ctio
n 2.4,
We could rea
rra
nge terms
of formula (1
5) as follo
ws:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
An Im
proved
Evolution
a
ry
Algorithm
with Ne
w Gen
e
tic Ope
r
ation f
o
r… (Wa
ng Ji
ekai
)
3153
ma
x
ma
x
max
ma
x
s
i
n
(
(1
)
)
l
n
(1
)
ma
x
2
(2
)
()
(,
1
)
(,
)
3
2
()
()
(,
)
3(
1
)
1
gen
Na
gen
ge
n
g
e
n
fH
MH
g
e
n
M
H
g
e
n
af
ge
n
ge
n
g
e
n
fH
H
MH
g
e
n
af
ge
n
l
e
(16)
Formul
a (16
)
rep
r
e
s
ent
s t
he sch
e
ma t
heorem
influ
enced by p
r
o
posed
crosso
ver op
erato
r
and
cro
s
sov
e
r st
r
a
t
egy
in S
C
A
G
A
.
After that we could
sim
p
ly get form
o
f
the sch
e
ma
theorem
of
SCAGA effe
cted by
mutation ope
ration as follo
ws d
ue to mu
tation prob
abi
lity used in (8
).
ma
x
max
ma
x
max
s
i
n
(
(1
)
)
l
n
(1
)
max
2
(2
)
[
1
(
)
]
()
(,
1
)
(
,
)
3
2
()
[
1
()
]
()
(,
)
3(
1
)
1
m
m
gen
Na
ge
n
ge
n
g
e
n
O
H
p
fH
MH
g
e
n
M
H
g
e
n
af
gen
ge
n
g
e
n
H
OH
p
fH
MH
g
e
n
af
g
e
n
l
e
(17)
Since
5
.
0
2
1
sin
max
gen
gen
P
m
, the schema theo
re
m for SCAGA
could b
e
gen
erali
z
ed to:
ma
x
ma
x
max
ma
x
max
ma
x
s
i
n
(
(1
)
)
l
n
(1
)
ma
x
2
(
2
)[
1
(
)
s
i
n
((
1
)
)
0
.
5
]
2
()
(,
1
)
(,
)
3
(
)
[1
(
)
s
i
n
(
(1
)
)
0
.
5
]
2
2
()
(,
)
3(
1
)
1
ge
n
Na
gen
ge
n
ge
n
g
en
O
H
ge
n
fH
MH
g
e
n
M
H
g
e
n
af
ge
n
ge
n
HO
H
ge
n
ge
n
g
en
fH
MH
g
e
n
af
ge
n
l
e
(18)
Since the sch
e
ma theo
rem
for SGA can
be de
scribe
d as follo
ws:
()
()
(,
1
)
(
,
)
[
1
(
)
]
1
cm
fH
H
mH
g
e
n
m
H
g
e
n
p
O
H
p
fl
(19)
Con
s
id
er (13), we get:
(,
)
1
()
()
(,
1
)
(,
)
[
1
(
)
]
1
MH
g
e
n
SG
A
c
m
i
fH
H
M
Hg
e
n
m
H
g
e
n
p
O
H
p
fl
(20
)
To com
pare
M(
H
,
ge
n+
1
)
SC
A
G
A
and
M(H,gen+1)
SGA
, we coul
d divide
(18) by (20),
or we
coul
d
simplify the computation a
s
follows:
()
(1
)
(
1
(
)
)
(,
1
)
1
()
(,
1
)
(1
)
(
1
(
)
)
1
()
()
1(
)
(
)
11
()
()
1(
)
(
)
11
IA
G
A
S
C
A
G
A
SG
A
S
G
A
S
C
AG
A
S
CAG
A
S
C
AG
A
SCAG
A
SG
A
S
G
A
SG
A
S
G
A
IA
G
A
cm
SC
A
G
A
SG
A
cm
cm
c
m
cm
c
m
cm
H
pO
H
p
MH
g
e
n
l
H
MH
g
e
n
pO
H
p
l
H
H
pp
O
H
p
O
H
p
ll
HH
pp
O
H
p
O
H
p
ll
pp
I
A
GA
I
A
GA
I
A
GA
S
G
A
S
GA
S
G
A
S
GA
cm
cm
c
m
pp
pp
p
p
(21)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 3148 – 3
157
3154
After s
u
bs
tituting for
P
m
an
d
P
c
of SCAGA in formula (21), we
coul
d
get:
(,
1
)
1
(,
1
)
SC
AGA
SG
A
MH
g
e
n
MH
g
e
n
(22)
Equivalently, from (2
2) we can imm
ediat
ely get:
(,
1
)
(,
1
)
IAGA
SGA
MH
g
e
n
M
H
g
e
n
(23)
Because the probability of so
lution di
sruption
with
high fitness
can be small
e
r than
solutio
n
with
lowe
r fitness,
we could o
b
s
erve
th
at propo
sed S
C
AGA use hi
gh
fitness valu
e
to
prom
ote sche
ma.
Mea
n
whi
l
e,
SCAGA could also
g
e
t sche
ma i
n
crease
steadily
. From
(23),
we
can
see SCA
G
A is more efficient than SG
A according
to the sche
m
a theore
m
.
4 Simulations
4.1. Functio
n
Test
In orde
r to test the perfo
rman
ce of SCAGA, three
commo
nly used multim
o
dal test
function a
r
e
perfo
rmed i
n
this se
ction.
Both function
s and th
eir v
a
riabl
e ra
nge
are summa
ri
zed
as
follows
:
24
2
2
2
1
min
(
,
)
(
4
2
.
1
/
3
)
(
4
4
)
f
xy
x
x
x
x
y
y
y
Whe
n
the ran
ge of
f
1
is
x
∈
(-100,1
0
0
)
,
y
∈
(-1
00,10
0), the optimum is
-1.031
628.
55
2
11
mi
n
(
,
)
{
c
o
s
[(
1
)
]}
{
c
o
s
[(
1
)
]
}
ii
f
x
y
ii
x
i
i
i
y
i
Whe
n
the ran
ge of
f
2
is
x
∈
(-10,10
),
y
∈
(-10,10), the o
p
timum is -1
8
6
.7309.
22
2
2
2
2
3
m
a
x
(
,
)
0
.
5
(
sin
0
.
5
)
/
(
1
0.
00
1
(
)
)
fx
y
x
y
x
y
Whe
n
the ran
ge of
f
3
is
x
∈
(-1
00,10
0),
y
∈
(-10
0,100
), the optimum is 1.
3
12
4
3
11
2
sin
(
2
)
sin
(
2
)
min
(
)
()
x
x
f
xx
x
Whe
n
the ran
ge of
f
4
is
x
1
∈
(0,10),
x
2
∈
(0,
10), and the
constraine
d co
ndition is:
2
11
2
()
1
0
gx
x
x
,
2
21
2
()
1
(
4
)
0
gx
x
x
,
The optimum
is
(
1
.2
279713
,
4
.
2453733
)
x
,
(
)
0.
095825
fx
.
5
1
mi
n
(
(
)
)
n
n
i
i
f
nx
Whe
n
the ran
ge of
f
5
is
n=
10,
xi
∈
(0,1
),
i
=1,...,
n
, and the con
s
trai
ned
conditio
n
is:
2
1
1
()
1
0
n
i
i
hx
x
,
the optimu
m
is
1
/
(
1
,
...,
)
x
ni
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
An Im
proved
Evolution
a
ry
Algorithm
with Ne
w Gen
e
tic Ope
r
ation f
o
r… (Wa
ng Ji
ekai
)
3155
22
2
61
2
3
mi
n
(
(
1
0
0
(
5
)
(
5)
(
5
)
)
/
1
0
0
))
fx
x
x
Whe
n
the ran
ge of
f
6
is
xi
∈
(0,10),
i
=1,2,3
, and the con
s
train
ed cond
ition is:
22
2
12
3
(
)
(
)
(
)
(
)
0.
0625
0
gx
x
p
x
q
x
r
,
,
,
1
,
2
,
...,
9
pqr
The optimum
is
(
5
,5
,5
)
x
,
()
1
fx
.
To evaluate the perfo
rma
n
c
e of the pro
pos
ed alg
o
rit
h
m, SGA, AGA and PSO
is use
d
for comp
ari
s
ons.
We
test
ed a
bove fu
nction
s fo
r 5
0
times to gi
ve a comp
arison
of ave
r
age
solutio
n
and
optimal sol
u
tion for the
s
e four alg
o
rithm
s
,
The optimi
z
at
ion re
sults a
r
e listed in Ta
ble 1.
Table 1. The
Tested
Com
p
arison Ta
ble
of SGA, AGA, PSO and SCAGA for Te
st Fun
c
tion
s
Function
SG
A
AG
A
PSO
SCAG
A
optimal
mean
optimal
mean
optimal
mean
optimal mean
f
1
-0.9622
-0.9453
-0.9934
-0.9681
-1.0314
-0.1012
-1.0312
-1.0229
f
2
-186.7250
-186.7010
-186.7300
-186.7180
-186.7280
-186.7100
-186.7300
-186.7240
f
3
0.9999
0.9999
0.9999
0.9999-
0.9999
0.9999
0.9999
0.9999
f
4
-0.0958
-0.0966
-0.0958
-0.0967
-0.0958
-0.0947
-0.0958
-0.0958
f
5
-1.0000
-0.9866
-1.0000
-0.9972
-1.0000
-1.0000
-1.0000
-1.0000
f
6
-1.0000
0.9466
-1.0000
0.9878
-1.0000
-0.9997
-1.0000
-1.0000
From T
able 1
,
the optimiza
t
ion re
sults of
SCAGA for
f
1-6
, including
both of co
nst
r
aine
d
and un
con
s
trained optimi
z
ation pro
b
le
m, are better than those i
n
SGA and AGA and PSO
algorith
m
[10]. For
f
1
(Cam
el function
), comp
ared th
e mean an
d optimal re
sult
s with SGA and
AGA and PSO algo
rithm, except the
re
sults
with t
he
mean
solutio
n
of PSO is sl
ightly better tha
n
those
of SCA
G
A, other
rel
a
ted sim
u
lati
on re
sult
s of t
he propo
se
d
algorith
m
are
better than
ot
he
r
algorith
m
for comp
ari
s
o
n
.
We use
f
2
(
Sh
ubert
fun
c
tion
) a
nd
f
3
(
S
c
haffer
fun
c
tion)
to test
conve
r
gen
ce
pe
rformance
of SCAGA. The com
p
a
r
iso
n
of average
results of
SG
A, AGA, and
SCAGA are li
sted in Tabl
e 2.
Table 2. Co
m
parin
g of Con
v
ergen
ce Pe
rforman
c
e of
SGA, AGA, a
nd SCAGA
Algorithm Function
Populat
ion Time Gene
ration
SG
A
f
2
100 48
58
f
2
200 50
55
f
3
300 42
96
f
3
500 45
93
AG
A
f
2
100 50
54
f
2
200 50
52
f
3
300 43
92
f
3
500 44
88
SCAG
A
f
2
100 50
50
f
2
200 50
49
f
3
300 42
89
f
3
500 47
87
We u
s
e SG
A, AGA, and SCAGA to test
Schaf
fer
function
and
Shub
ert
function
indep
ende
ntly for 50
time
s. Form
the
co
mpari
s
o
n
result from
Tabl
e 2, the
conv
erge
nce time
s of
SCAGA are
more tha
n
SGA and SCAGA while t
he co
nverg
e
n
ce g
ene
rati
on in averag
e of
SCAGA is le
ss than SGA a
nd AGA.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 3148 – 3
1
57
3156
Form the e
x
perime
n
t above, we can se
e SCAGA has a
better co
n
v
ergen
ce
perfo
rman
ce
than SGA and SCAGA,
and can o
b
tain better
optimal sol
u
tion than oth
e
r
comp
ared alg
o
rithm.
4.2.
TSP
Pro
b
lem Test
The travelin
g sale
sm
an
probl
em (
TSP
) ha
s attracte
d attention from
many
mathemati
c
ia
ns
and
com
puter
scienti
s
ts a
s
a
NP
-compl
ete p
r
o
b
lem. It invol
v
es findi
n
g t
h
e
sho
r
test
Ha
miltonian cy
cle in a com
p
lete dire
ct
e
d
grap
h, an
d it is a go
od gro
und t
o
test
perfo
rman
ce
of optimizatio
n algorith
m
.
Firstly we ch
oose to test
30-city and
1
05-city
TSP
probl
em to
compa
r
e p
r
efe
r
en
ce of
SCAGA with
both SGA and AGA.
Table 3. Perf
ormance of SGA, AGA, and SCAGA for TSP
Cities Average
Optimal
Genes
Pop.
Size
SG
A
AG
A
SCAG
A
SG
A
AG
A
SCAG
A
30(424.0
) 442.1
430.2
428.1
0
7
9
100
1000
105(1438
3) 16344.3
14801.4
14689.3
0
4
4
500
2000
We u
s
e p
opul
ation si
ze a
s
1000 a
nd 2
0
0
0
for 30
-city a
nd 105
-
city
TSP
res
p
ec
tiv
e
ly. For
both p
r
oble
m
s, SCAGA
re
pre
s
ent
a bet
ter pe
rfor
m
a
n
c
e tha
n
AGA
and SGA
on
the average
of
tour le
ngth. A
nd SCAGA l
o
cated th
e o
p
timal solution f
o
r 9 tim
e
s i
n
30-city pro
b
le
m and
4 time
s
in 105
-city problem, while
AGA located
the optimal solution for 7 ti
mes in 3
0
-cit
y problem
an
d 4
times in 105
-city proble
m
.
S
e
con
d
ly
,
we
sele
ct
8
TSP
instan
ce
s to evaluate the
effectivene
ss
of SCAGA. Figure 1
is a co
mpa
r
ison with different algorith
m
for the
perce
ntage deviati
ons of the av
erag
e sol
u
tio
n
to
the best kno
w
n solution.
Figure 1. Percenta
ge Deviations
of the
Average Solu
tion in each
TSP
Datas
e
t of Three
Algorithm
s
As we can
see in Fi
gure
1, the propo
sed
S
C
AGA gets relativel
y
smalle
r
pe
rcenta
ge
deviation
s than both SG
A and AGA. That indica
te SCAGA is a efficient
algorithm for
optimizatio
n probl
em.
5. Conclusio
n
In this p
ape
r,
we
present
a ne
w evol
utionary
algo
rith
m SCAGA b
a
s
ed
on th
e i
m
prove
d
geneti
c
op
eration, whi
c
h
adopt
s a d
y
namic m
e
th
od to re
gula
t
e cro
s
sove
r and mutati
on
prob
ability. Solution of G
A
can often
be infe
cted
by its ran
d
o
m
ly initializin
g popul
ation
and
geneti
c
p
a
ra
meter. T
he i
m
provem
ent
in this
pa
pe
r co
uld
ma
ke
the S
C
AGA
obtain
g
r
eat
e
r
optimizatio
n
perfo
rman
ce
and convergen
ce pe
rf
ormance. Mea
n
whil
e, we
prop
ose a n
e
w
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
An Im
proved
Evolution
a
ry
Algorithm
with Ne
w Gen
e
tic Ope
r
ation f
o
r… (Wa
ng Ji
ekai
)
3157
cro
s
sove
r st
rategy to dete
r
mine
wh
ethe
r to take
cro
s
sover ope
rati
on. The
dyna
mic
control for
the crossove
r operation
co
uld help
i
m
prove the conv
erge
nce rate
of
the algo
rith
m. Acco
rdin
g
to
the schem
a theorem an
al
ysis for S
C
A
G
A, we c
oul
d see th
at SCAGA is m
o
re efficie
n
t than
SGA.In Section 4,
we al
so co
mpa
r
e t
he optim
i
z
ati
on pe
rforma
nce
of SGA
G
A with
sev
e
ral
typical alg
o
rit
h
m in 6
different test fun
c
tions
and
TSP
probl
em
s. T
herefo
r
e, it can be
co
ncl
u
ded
that SCAGA
is hig
h
ly imp
r
oved
and i
s
efficient
to
solve both
co
nstrai
ned
an
d un
con
s
trai
ned
optimizatio
n probl
em.
Referen
ces
[1]
JH Holl
an
d. Adaptatio
n in N
a
tural Ar
tifici
al S
y
stems. MIT
Press. 1975: 1-
1
7
.
[2]
Goldb
e
rg
DE.
Genetic
Al
go
rithms in
Se
ar
ch, Optimizati
on
an
d Mac
h
i
ne
Le
arni
ng.
Rea
d
in
g. MA:
Addis
on W
e
sle
y
. 19
89.
[3]
Chu
ng-C
h
e
ng
Lu, Vi
ncent F
Yu. Data
env
e
l
opm
ent a
nal
ysis for ev
alu
a
ti
ng th
e effici
en
c
y
of ge
neti
c
alg
o
rithms o
n
solvin
g the v
e
hicle r
outi
ng p
r
obl
em
w
i
th s
o
ft time
w
i
nd
o
w
s
.
C
o
mput
ers & Industria
l
Engi
neer
in
g
. 2012; 2(6
3
): 520
-529.
[4]
A Serdar T
a
san, Mitsuo Gen. A genetic alg
o
ri
thm bas
ed appr
oach to v
ehicl
e routi
ng
prob
lem
w
i
t
h
simulta
neo
us p
i
ck-up a
nd d
e
li
veries.
Co
mput
ers & Industrial
Engin
eeri
n
g
. 2
012; 3(6
2
): 755
-761.
[5]
José A Caste
l
l
anos-Garz
ón,
F
e
rnan
do Díaz
. An ev
oluti
o
n
a
r
y
com
putati
o
n
a
l mod
e
l a
p
p
l
i
ed to clust
e
r
ana
l
y
sis of DN
A microarra
y d
a
ta.
Expert Systems w
i
th Appl
i
c
ations
. 20
13;
7(40): 25
75-
25
91.
[6] D
T
h
ierens.
A
daptiv
e
mutati
on rate
contro
l
st
rategys in
g
enetic
al
gorith
m
s
. Proc
eed
in
gs of the
20
02
Con
g
ress on E
v
oluti
onar
y C
o
mputatio
n CEC
.
2002
[7]
MM Lotfi, R
T
a
vakkoli-Mog
had
dam. A
ge
netic a
l
g
o
rithm
usin
g pr
iorit
y
-base
d
e
n
co
di
ng
w
i
t
h
n
e
w
oper
ators for fixe
d char
ge tra
n
sportati
on pr
o
b
lems.
App
lie
d
Soft Comp
utin
g
. 2013; 5(
13): 271
1-27
26.
[8]
L Cor
d
e
lla,
C
De Stefa
no, F
F
ontan
ell
a
, A
Ma
rcel
li Ev
o
G
eneS
ys. A
n
e
w
ev
ol
ution
a
r
y
ap
proac
h to
grap
h ge
nerati
on.
Appl
ie
d Soft Comp
utin
g
. 2013; 4(1
3
): 192
2-19
38.
[9]
David
H W
o
lp
e
r
t, W
illiam G M
a
crea
d
y
.
No
Free
Lunc
h T
heo
rems for Optim
i
zatio
n
.
IEEE Transactions
on Evol
utio
nar
y Computati
on.
1997; 4(1): 6
7
-
82.
[10]
M Srinivas, L
Patnaik. Ad
apt
ive Prob
abi
liti
e
s of Crossover
and
Mutati
on
in Genetic Al
g
o
rithm.
IEEE
Trans. On Syst
em
s
,
Ma
n
,
a
nd Cyb
e
rnetics
.
1994; 24(
4): 656-6
66.
[11]
Lich
eng
Jia
o
,
Lei W
a
ng. A
Novel
Gen
e
tic
Algor
ithm Ba
sed o
n
Immu
nit
y
.
IEEE Transactions on
Systems, Man,
and Cyb
e
rn
eti
cs-
Part A: Systems a
nd H
u
ma
ns
. 30; 5.
[12]
Clau
d
i
o
Comi
s, Da Ronco,
Ernesto Beni
ni.
A Simple
x Crossover b
a
sed ev
oluti
o
n
a
r
y
a
l
gor
ith
m
inclu
d
i
ng the g
enetic d
i
versit
y as objectiv
e
.
Appli
ed Soft Co
mputin
g. 201
3; 4(13): 210
4-2
123.
[13]
Siva Venk
ad
es
h, Gerrit Hoog
enb
oom. W
a
lter Po
tter and
Ron
a
ld McC
l
e
ndo
n.A gen
etic
algor
ithm to
refine in
put dat
a selecti
on for air temperat
ur
e
predicti
on us
ing artifici
al n
e
u
ral net
w
o
rks.
Appl
ied Soft
Co
mp
uting.
2
0
13; 5(13): 2
253
-226
0.
[14]
Adam
L C
hen,
Dan
i
e
l
H M
a
rtinez. A
he
urist
i
c
metho
d
bas
ed on gen
etic alg
o
rithm
for
t
he base
lin
e-
prod
uct desi
g
n
.
Expert Systems w
i
th Appl
ic
ations
. 20
12; 5
(
39): 582
9-5
8
3
7
.
[15] Jiekai
W
a
ng.
Impr
ove
d
Gene
tic Algorith
m
B
a
sed o
n
the Sma
ll Grou
p Par
a
lle
l
. International Workshop
on F
u
ture Com
m
unic
a
tion a
n
d
Net
w
o
r
ki
ng. IW
F
CN. 2011.
[16]
CAI Mingd
i, YAO Huanmi
n
,
W
A
NG Jiekai,
etc.
A Novel E
v
oluti
onary A
l
g
o
rith
m w
i
th Improve
d
Genetic
Operator a
nd
Crossov
e
r Strategy
. T
he 2013
2nd Intern
atio
nal C
onfer
ence
on Informatio
n
T
e
chnol
og
y
and Ma
na
gem
ent Innov
ation.
ICIT
MI. 2013.
Evaluation Warning : The document was created with Spire.PDF for Python.