TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.5, May 2014, pp
. 3786 ~ 37
9
1
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i5.5110
3786
Re
cei
v
ed
No
vem
ber 1
1
, 2013; Re
vi
sed
De
cem
ber 2
8
,
2013; Accep
t
ed Jan
uary 1
0
, 2014
A Hybrid Clustering Algorithm Based on Improved
Artificial Fish Swarm
Lin Tian
1
, Li
w
e
i Tian*
2
1
School of Elec
tronic Eng
i
n
eer
ing, She
n
y
a
ng
Univers
i
t
y
,
21 South W
a
n
ghu
a Street, Shen
ya
n
g
, Liao
nin
g
, Chin
a, Ph: +
86159
04
02
016
5
2
Liao
nin
g
Infor
m
ation Integr
at
ion T
e
chnol
og
y E
ngin
eeri
ng R
e
searc
h
Cent
e
r
of Internet of
things
Lia
nhe R
o
a
d
, Shen
ya
n
g
Lia
o
n
in
g, Chin
a, Ph: +
86133
04
03
587
7
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: tianli
w
e
i
@
1
6
3
.com*
A
b
st
r
a
ct
K-me
do
ids cl
u
s
tering
alg
o
rith
m is
use
d
to cl
assify
data,
but
the ap
pro
a
ch i
s
sensitiv
e to t
he i
n
itia
l
selecti
on of the
centers a
nd th
e divi
de
d clust
e
r qu
ality is
n
o
t
high. B
a
sic Artificial F
i
s
h
Sw
arm Al
gorith
m
is
a
new
type of he
uristic sw
arm i
n
telli
ge
nce a
l
g
o
rith
m, but
opti
m
i
z
at
io
n is diffi
cult to get a ve
ry high
precis
i
o
n
due to the ra
n
d
o
m
n
e
ss of the artificia
l
fish beh
avio
r. A no
vel cluster
i
ng
meth
od b
a
se
d on i
m
pr
oved
gl
oba
l
artificial
fish
s
w
arm is
pr
op
o
s
ed
in
this
pa
per
by
ana
ly
z
i
ng th
e
adv
ant
ages
a
nd
dis
a
dvanta
ges
of t
w
o
alg
o
rith
ms, w
h
i
c
h has th
e ab
ili
ty to optimi
z
e
t
he g
l
ob
al cl
usterin
g
effect. T
he result of th
e exper
iment sh
ow
s
that qu
ality of
clusteri
ng is
i
m
prov
e
d
; the o
p
t
ima
l
ce
ntral
p
o
ints a
nd th
e c
l
ear
divis
i
on
of
data
grou
ps a
r
e
obtai
ne
d by the mat
h
e
m
atic
al
mod
e
l co
mbin
g improv
ed fish
sw
arm alg
o
rith
m an
d K-med
o
i
d
s alg
o
rith
m.
Ke
y
w
ords
:
arti
ficial fish sw
ar
m al
gorit
h
m
, K-me
do
ids al
gorit
hm, clust
e
rin
g
ana
lysis
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
Clu
s
ter an
al
ysis is an i
m
porta
nt research
directi
on of data mining; clu
s
t
e
ring i
s
cla
ssifying
da
ta for differe
n
t
pattern
s ba
sed
on th
e
di
fferent chara
c
teri
stics of d
i
fferent obje
c
t
s
[1]. The sam
e
obje
c
ts ha
ve a high si
milarity degr
ee, while o
b
j
e
cts of different grou
ps v
a
ry
greatly from
each othe
r, this form th
e law of
dist
ribu
tion of the ob
je
ct and
co
rrelation bet
we
en
the data [2].
Since data
b
a
s
e coll
ecte
d lots of data,
it
requi
re
s scal
ability of algorithm and cl
u
s
ter
quality. In this pa
per, K
-
m
edoid
s
al
gorit
hm is u
s
e
d
t
o
divide
clu
s
ters by calcul
ating di
stan
ce,
dissimila
rity, squ
a
re
d e
rro
r an
d othe
r
para
m
eter
s, t
h
is al
gorith
m
has strong
robu
stne
ss a
nd
flexibility, but it is su
scepti
b
le to effects of
the outliers an
d local e
x
treme value,
and rand
oml
y
initialize pa
ra
meters play a
deci
s
ive role
on the clu
s
tering re
sults.
A novel bottomup optimiza
t
ion mode Art
i
ficial
Fish S
w
arm Algorit
hm (AFSA) is used in
this pap
er. A
F
SA use
swarm intellig
e
n
ce
of bios
p
here to
solve
optimizatio
n
probl
ems, a
s
a
gene
rali
zed
n
e
ighb
orh
ood
sea
r
ch al
gorit
hm, by mea
n
s of h
e
u
r
isti
c
sea
r
ch
strate
gy, its capa
ci
ty
of tra
cki
ng
chang
es rapid
l
y gives algo
rithm the
a
b
il
ity
of
glob
al optimizatio
n, becau
se of the
cha
r
a
c
teri
stics of
glob
al
converg
e
n
c
e i
t
self, t
he initi
a
l value
can
be
set
as fixed or rand
o
m
allowin
g
pa
rameters to
be set in a
wide
r sco
p
e
[3]. AFSA has
stro
ng
adaptability
and
parall
e
lism, many
beh
aviors com
b
inati
ons can be
selecte
d
due t
o
its goo
d flexibility, and it ca
n
get better op
timization pe
rforman
c
e
whi
c
h ge
net
ic al
gorithm a
nd particl
e swa
r
m optimizatio
n
doe
s not p
o
sse
s
sed. Thi
s
artificial intel
ligen
ce
mo
de
whi
c
h i
s
ba
sed on
biolo
g
i
c
al b
ehavio
r
is
different fro
m
classi
cal
pattern, firstl
y is
to design a singl
e
entity perception, beha
vioral
mech
ani
sm
s, then pla
c
ed
a gro
up of
entities in
th
e enviro
n
me
nt so that th
ey can
solve
the
probl
em in environ
ment
interactio
n [4, 5];
however makin
g
the best re
a
c
tion und
er
the
stimulation of
the environ
ment is the basi
c
i
dea of AFSA. Literature [6] prop
o
s
ed redu
cin
g
the
sea
r
ch field to accele
rate l
o
cal
sea
r
ch o
f
artifici
al fish
individual, b
u
t this optimi
z
ation o
n
ly took
conve
r
ge
nce
spe
ed i
n
to a
c
cou
n
t ma
king
seve
re
limita
t
ion of
swarm
i
ng a
nd foll
o
w
ing
be
havio
rs
of AFs, thus affecting the quality of the optimiz
ation.
[7] introduce
d
the K-mea
n
s algo
rithm
to
spe
ed u
p
the
iteration, but
the pe
rform
a
nce
wa
s
un
s
t
a
b
l
e
be
ca
us
e o
f
ma
n
y
r
and
o
m
pr
oc
es
se
s
in AFSA an
d
it affected
t
he p
r
a
c
tical
appli
c
ation
of
the m
e
thod.
Using
si
mul
a
ted a
nne
ali
n
g
algorith
m
to improve AF
SA
，
the approach in [8] modified p
r
e
y
ing behavio
r to avoid the
degradatio
n
of artificial
fish, alth
oug
h
this hyb
r
id
al
gorithm
overcame
the
sh
ortco
m
ing
wh
ich
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Hyb
r
id Cl
ustering Algo
rithm
Based on
Im
proved Art
i
ficial Fish S
w
arm
(Lin Tian
)
3787
easily fall into local minima
, converg
e
n
c
e time
of the
method was relatively long and it was n
o
t
suitabl
e to a
nalysi
s
hug
e
data. Com
b
i
n
ing AFSA with clu
s
terin
g
analysi
s
alg
o
rithm b
a
se
d
on
grid
an
d d
e
n
s
ity, [9] obtai
ned th
e
num
ber K of
clu
s
t
e
rs a
u
tomatically and
it a
p
p
lied to
a
r
bitrary
sha
pe of
dat
a, better
para
llelism, b
u
t the quality of
ul
timately clu
s
tering
qu
ality wa
s affe
cted
by
the numbe
r a
nd the si
ze of
grids
whi
c
h l
ed to some li
mitations [10]
.
The traditio
n
a
l K-medoi
ds has greate
r
ability of
local search, but i
s
very sen
s
iti
v
e to the
initial cluster
centers and
easily
falling
into local optimum, if out
li
ers are randomly sel
e
cted as
the initial centers, the whol
e qualit
y of cl
assification will decline.
AFSA is less
sensitive to initial
values,
even
if its gl
obal
optimizatio
n,
has ba
d
con
v
ergen
ce
an
d sl
ower ite
r
ation rate in
late
perio
d. Aimin
g
at the
adva
n
tage
s an
d d
i
sadva
n
t
age
s of both
algo
rithms, this pa
per
pre
s
e
n
ts
a
global o
p
timization idea to i
m
prove K-m
e
doid
s
clu
s
teri
ng algo
rithm
based on AF
SA, the result
o
f
the test
on
a
small
data
se
t sho
w
s that
the imp
r
ove
d
algorith
m
o
b
tains
cl
ear cla
ssifi
cation
s a
nd
better pe
rformance.
2. Clustering
M
odel
12
N
X
=
(
x
,x
,...x
)
as the N d
a
ta sampl
e
s, x is the data repre
s
e
n
tative point, C
i
is
an
arbitrary clu
s
t
e
r, O
i
is the center of the cl
uster
C
i
, (j=1,2…,k). Algorit
hm is present
ed as follo
ws.
S
e
lect
ed
k o
b
ject
s in s
e
t
X
as t
he initi
a
l ce
nters arbitrarily (O
1
,O
2
,…O
i
…O
k
), assign
ed
the remaini
n
g
data
exce
pt f
o
r
rep
r
e
s
enta
t
ive cente
r
s b
y
the p
r
oximit
y
prin
ciple
to each clu
s
ter; in
each clu
s
ter
(C
i
), cho
s
e a
noncentral
p
o
int O
j
ran
d
o
m
ly, calculati
ng total co
st
∆
E after us
ing
non-ce
nter i
n
stead
of the
origin
al cente
r
poi
nt; If
∆
E<0, then
re
pl
ace th
e o
r
igin
al O
i
with a n
on-
cente
r
O
j
, performing the a
bove step
s re
peatedly until
k c
ente
r
s i
s
fixed [11,
12]. Co
st
function
is
use
d
to evalu
a
te the clu
s
te
ring qu
ality improve
d
. The function i
s
defined a
s
fol
l
ows:
21
EE
E
(1)
∆
E represen
ts the
ch
an
ge of
ab
sol
u
te erro
r
st
anda
rd, E
2
refers to the s
u
m of
dissimila
rity degre
e
b
e
twe
en rep
r
e
s
ent
ative point
s a
nd
cente
r
p
o
i
n
ts in
the
sa
me cl
uste
r
after
repla
c
in
g the
centers, an
d E
1
repre
s
e
n
ts the dissi
m
ilarity
degree before re
placi
ng [13, 14].
Cal
c
ulate
∆
E, if
∆
E<0, the effect of clust
e
ring h
a
s b
e
e
n
improve
d
, then u
s
e the n
e
w center.
3. Optimized
AFSA
3.1. Descrip
tion of the
Ba
sic Behav
i
ors
Population o
f
AFs is N, individual state of AF:
12
(,
,
.
.
.
)
n
Ff
f
f
,[where f
i
is
optimizatio
n
variable
s
], th
e large
s
t mo
ving ste
p
is
Step, vision i
s
Visual, te
st time of p
r
ey
ing
behavio
r is T
r
y_num
ber, crowd facto
r
is
δ
, food co
nsi
s
ten
c
e
()
Yf
F
(Y is the value of
obje
c
tive function).
3.1.1.
Pre
y
in
g Behav
i
or
As one of the
basi
c
habit
s
of AF, the main prin
cipl
e is finding the
area
whe
r
e t
here i
s
a
large
food
co
nce
n
tration
b
y
sen
s
e
of si
ght an
d ta
ste
.
Curre
n
t stat
e of AF i
s
F
j
,
sele
ct
a
st
at
e
F
j
rand
omly aro
und current l
o
catio
n
withi
n
its vis
ual field, in the pro
c
e
ss of seeki
ng optim
al
s
o
lution, if
ij
YY
, then F
j
is a
b
e
tter state th
an the curren
t one and m
o
ve one
step
to this
dire
ction,
def
ault choo
se
a
ne
w
state
an
d jud
ge
agai
n
,
test Try_number time
s repeate
d
ly, if
still
unabl
e to get a better solution then move
a rando
m ste
p
[15, 16].
3.1.2. S
w
a
r
m
i
ng Behav
i
or
To en
su
re th
e survival of
fish p
opul
ations
, AF
will
gather to the
ce
nter
of a
d
jacent
partne
r
s. F
i
still corresponds to the
current
state,
perceive the
AF num
ber nf nearby and its
central lo
cati
on F
c
.if satis
f
ied
/
cf
i
Yn
Y
,
i
t
mea
n
s t
h
e po
sit
i
on
wa
s le
ss
con
g
e
s
t
i
on
level, more fo
od, then step
forwa
r
d to F
c
, or impleme
n
t preying be
ha
vior [17].
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3786 – 37
91
3788
3.1.3.
Follo
w
i
ng Behav
i
or
In nature, wh
en one or a few fishe
s
ha
ve
explored food, its neigh
bors will follo
w swarm
to rea
c
h the
food positio
n [18]. Perception of the
best state
F
j
within the vis
i
on, s
a
tisfied
j
/
fi
Yn
Y
which displa
y the location
was le
ss cro
w
ing d
egree, more foo
d
, then ma
ke a
step to F
j
, or do preyin
g be
havior.
3.2. Impro
v
e
m
ent of
AFS
A
(a) In preyin
g behavio
r, whe
n
a state
of randomly
sele
cted F
j
doe
s not sati
sfy the
moving
con
d
i
t
ion it will
ch
oose rand
om
behavio
r, th
at
is difficult
to obtain
hig
h
preci
s
io
n,
AFs
sea
r
ching
ne
arby th
e gl
ob
al extrem
e p
o
i
nts
circui
tou
s
ly at ana
pha
se of
conve
r
g
e
nce,
whi
c
h
le
ad
to an invalid cal
c
ulatio
n. In this pape
r,
when p
r
eyin
g failed, AFs choo
se to m
o
ve a step to
a
better value compa
r
ing
with the bulletin
board re
co
rd
s:
(1
)
(
)
[
(1
)
(
)
]
i
i
bet
t
e
r
i
Fk
Fk
S
t
e
p
F
k
Fk
(2)
F
i
(k
+1
) an
d
F
i
(k) de
note
respe
c
tively curre
n
t position and n
e
x
t position a
fter the
movement, F
better i
s
the
better
state
re
co
rde
d
by
bulletin
boa
rd, comp
arin
g with
ra
ndo
m
method it giv
e
s the p
o
ssib
ility of a better forward a
n
d
thus jum
p
ou
t of local opti
m
a, preve
n
tin
g
AFs in the local con
c
u
s
sion
at a standstill
.
(b)
In AFSA, the pa
ramete
r cro
w
din
g
fa
ctor
δ
i
s
to av
oid ove
r
cro
w
ding of AF
an
d
δ
is
a
fixed value i
n
glob
al alg
o
rithm, this
approa
ch th
a
t
make
δ
a
con
s
tant
will
lead to
mut
ual
exclu
s
ion
bet
wee
n
individu
als
whi
c
h a
r
e
adja
c
ent to g
l
obal o
p
timization solution
, so AFs
ca
n
not
gather to ext
r
eme p
o
ints
accurately an
d cont
ra
st crowdi
ng condi
tion afte
r eve
r
y iteration
will
increa
se the
comp
utationa
l cost
too. Improve
d
method define
s
the initial con
gestio
n
facto
r
δ
=0.75, whe
n
Try_num
be
r =
18
0,
igno
ri
ng the
cong
e
s
tion fa
ctor n
a
mely
1
f
n
in init
ia
l
stage
s, it ne
eds to limit the si
ze of a
r
tificial
fish,
but in the la
tter part fish
es have al
re
ady
gathered in
optimum, def
ault
δ
can redu
ce calcul
ation amou
nt and execution time of the
algorith
m
, in t
h
is
way n
o
t o
n
ly doe
s it im
prove
s
the
op
eration
efficie
n
cy of AFSA
but al
so h
a
s
no
effect on co
n
v
ergen
ce.
(c) In ord
e
r to
solve the pro
b
lem of ce
nt
e
r
s of K-m
edoi
ds by AFSA, whe
n
swa
r
mi
ng and
followin
g
beh
avior failed, p
r
eying be
havi
o
r is ca
rried
out, thus increasi
ng the co
nverge
nce time
and
com
puta
t
ion. So we
rene
w the
be
havior
as fo
ll
ows: su
bstitu
te ran
dom
swim for preying
behavio
r afte
r failing
in m
o
vement. An
d the
step
i
s
adaptive step
-si
z
e.
T
he
m
e
thod overco
me
s
the proble
m
t
hat AFs ag
gregated
at lo
cal soluti
on
an
d misse
d
the
global
one
s
a
nd e
nha
nce t
h
e
quality of solu
tions.
4.
A H
y
brid Clustering
Algo
rithm Bas
e
d
on Impro
v
ed ASFA
4.1.
Defini
tions o
f
Improv
ed A
F
SA
Definition
1:
(a
daptive
st
ep-size
of A
F
)
Ad
aptive
step-si
ze
re
p
r
esents the
moving
distan
ce of AF chan
ging
with iterati
ons.
Adaptive step
-si
z
e is d
e
fine
d as:
1
()
ii
FF
S
t
e
p
R
a
n
d
(3)
Definition 2:
(clu
ste
r
ing ev
aluation
crite
r
ion
)
Obje
ctive functio
n
m
easure
s
di
ssi
m
ilarity
betwe
en rep
r
esentative points
an
d obje
c
ts,
whi
c
h
mea
n
s the
com
p
a
c
t deg
re
e of
dat
a
distrib
u
tion b
e
twee
n cla
sses, the obje
c
t
i
ve function is defined a
s
:
1
j
k
j
jX
C
EX
O
(4)
4.2. The Procedure o
f
th
e Mixed
Clus
tering Base
d
on Optimize
d AFSA
Step 1: Initialize the initial value of AF
para
mete
rs, cal
c
ulate
fo
od consi
s
ten
c
e at
cu
rre
nt
positio
n by objective fun
c
tion;
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Hyb
r
id Cl
ustering Algo
rithm
Based on
Im
proved Art
i
ficial Fish S
w
arm
(Lin Tian
)
3789
Step 2: Ca
rry
out the alg
o
rithm throu
gh
behavio
r’s
co
ndition, up
dat
e the lo
cation
of AFs
by preying, swarming a
nd
followin
g
beh
aviors, d
a
ta
d
ensity refe
r to food con
c
e
n
tration; contra
st
food con
s
i
s
te
nce withi
n
vision distan
ce t
o
sele
ct
solut
i
on, with its state reco
rde
d
in the bulletin
board, finally fishe
s
gathe
r in
the area
s o
f
high data de
nsity;
Step 3: Ea
ch state
of A
F
re
pre
s
e
n
ts a d
e
ci
sion
variable, and
the
fitne
s
s value
i
s
comp
uted by
objective fun
c
tion, evaluat
e optimizat
io
n deg
ree a
n
d
record; repe
at 2) 3), up
d
a
te
the locatio
n
informatio
n of AFs until the termin
ation co
ndition is met
;
Step 4: According to bulleti
n board informati
on and t
he locatio
n
of fishes, choo
se input
para
m
eters f
o
r K-me
doid
s
, namely th
e initial cent
er an
d the numbe
r of cl
usters; u
s
ing
K-
medoid
s
fo
r
clu
s
ter
analy
s
is until m
eet
ing minim
u
m
within
-cl
a
ss
scatter of
dat
a. The
minim
u
m
within-c
lass
M is
pres
ented as
follows
:
m
in
ME
(5)
The flowcha
r
t
sho
w
s p
r
o
c
e
dure of ap
pro
a
ch in Fig
u
re 1:
Figure 1. Flowchart of Clu
s
terin
g
Algorit
hm Base
d on
Optimized A
F
SA
5. Simulation
Simulation data includ
e 3
00 3D data; runni
n
g
enviro
n
ment for experim
ent: Pentium(R),
3.00G; Prog
ramming
envi
r
onm
ent: Ma
tlab7.12.0
(R2011
a); AFS
A
paramete
r
s a
r
e
set a
s
follows: Step is 0.2, Visu
al is 100,
δ
i
s
0.75, Try_
numbe
r(ite
rat
i
on times) is 200,N (th
e
total
numbe
r of AF) is 50.
In the
simul
a
tion, it cl
assif
y
the d
a
ta b
y
two
hybrid
clu
s
terin
g
alg
o
rithm
s
, com
pari
s
on
results of the
appro
a
ch this pap
er p
r
op
ose
d
and b
a
s
ic hyb
r
id cl
u
s
terin
g
algo
ri
thm. Operati
on
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
0
46
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3786 – 37
91
3790
result of classic hyb
r
id m
e
t
hod shown
in Figure 2,
Figure 3 sh
ows perfo
r
m
a
nce of improved
approa
ch.
Figure 2. Optimization G
r
a
ph of Basic
Clu
s
terin
g
Algorithm Ba
se
d on AFSA
Figure 3. Optimization G
r
a
p
h of Improve
d
Method
AFs find the centers in the
3D
data, a
s
shown in Figu
re 2 agg
reg
a
tion effect is n
o
t clea
r,
a few individ
uals m
o
ves
to local
clu
s
ters; o
p
timiza
tion re
sult a
pproxim
ate to global
dat
a
-
intensive a
r
e
a
s that can b
e
see
n
from t
he iteratio
n ro
ute in Figu
re
3; comp
ari
s
o
n
of perfo
rma
n
ce
sho
w
s the edge of clu
s
te
rs is mo
re o
b
vious by im
proved m
e
th
od on the sa
me con
d
ition,
the
aggregatio
n o
f
position is
cl
ose
r
so
that
we can obtai
n a highe
r accura
cy of the division to ve
rify
the advantag
es of this alg
o
rithm.
Table 1. The
Re
sults of T
w
o Algorithm
s
Total Numbe
r
of
AF
Iteration Times
Iteration Time /m
s
Correct R
ate
Method in [6]
50
200
762
89
Proposed Metho
d
50
200
685
93
It is
sho
w
n i
n
Tabl
e 1
the
prop
osed
me
thod
redu
ce
d
not o
n
ly the
i
t
eration
time
but al
so
cal
c
ulatio
n a
m
ount on the
same
con
d
ition,
and the a
c
cura
cy is al
so improve
d
.
4. Conclu
sion
Hybrid
clu
s
te
ring i
s
wid
e
ly applied in d
e
ci
sion p
r
obl
em and e
a
rl
y warnin
g at
current
resea
r
ch. Co
mpari
s
o
n
of e
x
perime
n
tal result
s
sho
w
s
improve
d
AF
SA hybrid
clu
s
terin
g
al
gorit
hm
make
similar data
gathe
r obviou
s
, the
mod
e
l i
s
m
o
re
sta
b
le
an
d a
c
curate th
an the
old
o
ne,
disting
u
ish sample
s p
r
e
c
i
s
ely while
al
so im
provin
g
the cl
uste
r
quality and
obtainin
g
bet
t
er
cente
r
s
with
clea
r divisi
on,
redu
cin
g
co
mputation a
m
ount is
also a
brea
kth
r
ou
g
h
. The mo
del
o
f
mode
rn intelli
gen
ce alg
o
rit
h
m ba
sed o
n
animal a
u
tonomo
u
s
bo
dy combi
n
e
s
K-medoi
d
s,
this
novel metho
d
avoids th
e weakne
ss
of d
epen
den
cy o
n
Clu
s
te
r initi
a
lizatio
n, and
overcome
s
slow
iteration
spe
e
d
in late
peri
od; its go
od
parall
e
lis
m
can be
effectively applie
d in
variou
s field
s
, it
also play
s a
major rol
e
in kno
w
led
g
e
discover
y, informatio
n forecast an
d d
e
ci
sion an
alysis.
Ho
wever, the
convergen
ce
spee
d issue
re
mai
n
s to be
improved a
n
d
resea
r
ched.
Ackn
o
w
l
e
dg
ements
This wo
rk wa
s sup
porte
d by
the
Li
aoni
ng Nation
al Natural
S
c
ien
c
e Foun
datio
n
(Gra
nt
No. 201
3020
011), the Internatio
nal S&T Coop
erati
o
n Prog
ram
of China (IS
T
CP) un
de
r Gran
t
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Hyb
r
id Cl
ustering Algo
rithm
Based on
Im
proved Art
i
ficial Fish S
w
arm
(Lin Tian
)
3791
2011
DFA9
18
10-5
a
nd P
r
o
g
ram
for Ne
w
Centu
r
y E
x
cell
ent Tale
nts
in
Universi
ty of Mini
stry of
Educatio
n of Chin
a unde
r
Grant
NCET
-
12-1
012.
Referen
ces
[1]
Jia
w
e
i
Ha
n, Kamber M. Data Minin
g
: Conc
e
p
ts and T
e
chni
ques. 20
07; 25
2-27
2.
[2]
Ning
x
ia
Xi
a, Yi
dan S
u
.
An Efficient K-me
do
i
d
s Cluster
in
g Algorit
hm.
Jour
nal of Ap
plic
ati
on Res
earc
h
of Computers
.
201
0; 27(1
2
): 4517-
451
9.
[3]
Xi
nmin
T
ao, Jing
Xu, L
i
b
i
a
o
Ya
ng. A
n
I
m
pr
o
v
ed
Hy
bri
d
Alg
o
r
i
t
h
m
Ba
se
d on
Parti
c
l
e
S
w
arm
Optimizatio
n
a
nd K-mea
n
s Algorit
hm.
Journ
a
l of Electron
i
cs &Informati
o
n
T
e
chno
logy.
2010; 32(
1):
92-9
4
.
[4]
Xi
ao
hua
W
a
n
g
, Jie S
hen,
Ron
gbo
W
ang
. A Ne
w
H
y
b
r
id Al
gorithm
Based
on
Ant
Col
o
n
y
an
d
Clusteri
ng.
Jou
r
nal of Ha
ng
z
h
ou Di
an
z
i
U
n
iv
ersity
. 2010; 3
0
(1): 26-2
7
.
[5]
Yichu
an S
hao,
Xi
ng
jia
Yao,
Li
w
e
i T
i
an, H
ann
ing
Ch
en.
A Multi-sw
arm Optimi
z
e
r for
Distrib
ute
d
Decisi
on Mak
i
n
g
in Virtua
l Enterprise R
i
sk Mana
ge
me
nt
. 20
12.
[6]
Yuju
n F
an, D
o
ngd
on
g W
ang,
Mingm
ing
Su
n. Improved
A
r
tificial F
i
s
h
S
w
a
rm Al
gorith
m
.
Journa
l of
Cho
ngq
in
g Nor
m
a
l
Univ
ersity (Natural Sc
ien
c
e Editio
n)
. 20
07; 24(3): 2
4
-2
6.
[7]
Bai
Liu,Yo
ng
q
uan
Z
hou. A
H
y
brid
Cl
usteri
ng A
l
gor
ithm
Based
on
Artificial
F
i
sh S
w
a
rm Algor
ithm.
Journ
a
l of Co
mputer Eng
i
n
eeri
ng an
d App
lica
t
ions.
200
8; 44
(18): 136-
13
8.
[8] Jia
Liu.
Improv
ed Artifici
al F
i
s
h
S
w
arm A
l
gor
i
t
hm and Its Ap
plicati
ons
in F
u
nction Optim
i
z
a
tion.
Journal
of Shiji
a
z
h
u
a
n
g
Institute of Railw
ay T
e
chn
o
l
ogy
. 201
1; 10(
3): 33-36
[9] Xu
de
ng
He,Li
ang
do
ng
Qu.
Artificial F
i
s
h
Sw
arm Cl
uste
ring A
l
gor
ith
m
. Appl
icatio
n
Rese
arch o
f
Comp
uters. 20
09; 26(1
0
): 366
6-36
68
[10] Hon
g
W
e
i
Z
hao
.
A Resource D
i
scovery Mec
h
anis
m
o
n
Clo
u
d
Co
mp
uting S
ystem
. 20
12.
[11] Juan
yi
ng
Xi
e,
Shua
i Jian
g.
A Simp
le a
n
d
F
a
st Algorith
m
for Globa
l K-me
ans Cl
us
tering
.Sec
on
d
Internatio
na
l Workshop o
n
Educati
on T
e
chno
log
y
a
nd
Comp
uter Sci
ence (ET
C
S). IEEE.
Wuhan
.
201
0; (2); 36-4
0
.
[12]
Bingr
u Yan
g
, Dan
y
a
ng C
ao.
An Impro
v
ed K-
med
o
i
d
s Clusteri
ng
Algorit
h
m
.2nd
Internatio
na
l
Confer
ence
on
Computer a
n
d
Automation E
ngi
neer
in
g (ICCAE). IEEE Singa
pore. 2
010:
132-1
35.
[13]
Lip
o
W
,
Xiuj
u F
.
Data Minin
g
W
i
th Computat
ion
a
l Intell
ig
en
ce. 2005.
[14]
Yuzhe
n
Z
h
a
o
, Xi
yu
L
i
u, Hu
a
Z
hang. T
he K-Medo
ids C
l
ust
e
rin
g
Alg
o
rithm
w
i
t
h
Membr
a
n
e
Comp
utin
g.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
. 201
3;
11(4)
:
20
50-2
057.
[15]
Doa
a
M Ati
a
, F
a
ten
H F
a
hm
y
,
Nin
et M A
h
me
d,
Hass
en
T
Dorrah. A
Ne
w
Contro
l a
n
d
De
sign
of PE
M
F
uel Cel
l
S
y
st
em Po
w
e
r
ed
Diffused Air A
e
ratio
n
S
y
ste
m
.
T
E
LKOMN
I
KA Indonesi
a
n Journ
a
l of
Electrical E
ngi
neer
ing
. 2
012;
10(2):
29
1-3
0
2
[16]
Chu
x
i
a
o
L
i
, Z
h
u Yi
ng, S
h
i
Ju
nT
ao, Song
Ji
Qing.
Meth
od
of Image
Se
g
m
e
n
tatio
n
B
a
s
ed
on
F
u
zz
y
C
-
Means
Cl
uste
ring A
l
g
o
rith
m an
d Artifici
al
F
i
sh Sw
ar
m Alg
o
rith
m
. In
ternatio
nal
C
o
nferenc
e o
n
Intelli
gent C
o
mputin
g an
d In
tegrated S
y
stem
s (ICISS).
IE
EE Press. 2010:
254-2
57..
[17]
Yongm
ing
C
h
e
ng, Mi
ng
ya
n J
i
ang,
Do
ngfen
g
Yua
n
.
N
o
vel
Clusteri
n
g
Alg
o
rith
ms base
d
on
I
m
pr
ove
d
Artificial F
i
s
h
Sw
arm Al
gorit
hm
. S
i
xth Inte
rnatio
nal
Co
nferenc
e o
n
F
u
z
z
y
S
y
stems
a
nd Kn
o
w
l
e
d
g
e
Discover
y
. IEE
E
Press. 2009;
(3): 141-14
5
[18]
Don
g
fen
g
Yua
n
, Ming
ya
n Jia
ng. Artificial F
i
sh
S
w
a
rm Alg
o
rithm an
d Its Applic
ations.
201
2; 43-6
6
:
110-
130
Evaluation Warning : The document was created with Spire.PDF for Python.