Indonesian J
ournal of Ele
c
trical Engin
eering and
Computer Sci
e
nce
Vol. 2, No. 2,
May 2016, pp
. 409 ~ 416
DOI: 10.115
9
1
/ijeecs.v2.i2.pp40
9-4
1
6
409
Re
cei
v
ed
De
cem
ber 2
8
, 2015; Re
vi
sed
April 7, 2016;
Accept
ed Ma
y 1, 2016
A
Novel Membrane Clustering
Algorithm Based on
T
i
ssue-like P
system
Huaning Yan
,
Laisheng Xiang, Xiy
u
L
i
u, Jie Xue
Schoo
l of Man
agem
ent Scie
n
c
e and En
gi
ne
erin
g
Shan
do
ng Nor
m
al Univ
ersit
y
,
Jinan, Ch
in
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: sdxyl
i
u@
16
3.com
A
b
st
r
a
ct
Clusteri
n
g
is a process
of
pa
rti
t
i
o
n
i
n
g
da
ta
poi
n
t
s i
n
to
di
ffe
ren
t
cl
u
s
te
rs d
ue to
th
ei
r sim
i
l
a
rity, a
s
a
pow
erful tech
n
i
qu
e of data
mi
nin
g
, cluster
i
ng is w
i
de
ly u
s
ed in
ma
ny fi
elds. Me
mbr
a
n
e
co
mputi
ng is
a
computing model
abstr
acting
from the biologic
a
l
area, these c
o
m
puting system
s
ar
e
proved to be
s
o
pow
erful th
at they ar
e e
quiv
a
lent w
i
th T
u
rin
g
mach
ines. In
this pa
per, a
mo
difi
ed i
n
vers
ion
particl
e sw
a
r
m
opti
m
i
z
at
ion w
a
s prop
ose
d
, this metho
d
an
d the mutatio
n
a
l mech
anis
m
of genetics
alg
o
rith
m w
e
re us
ed to
combine with t
he tiss
ue-like P
system
, thr
o
ugh these
ev
olut
ionary algor
ithm
s
and th
e P s
ystem
, the idea of
a nov
el
me
mb
rane cl
usteri
ng
alg
o
rith
m co
ul
d co
me
tru
e
. Experi
m
ents w
e
re tested
on
six data s
e
ts, by
compari
n
g
the
clusteri
ng
qua
li
ty w
i
th the GA-K-me
ans
,
PSO
-K-me
ans an
d K-me
ans prov
e
d
the
s
u
p
e
rior
ity
of our metho
d
.
Ke
y
w
ords
: me
mbr
a
n
e
cluster
i
ng a
l
gor
ith
m
, P system, parti
cle sw
arm o
p
ti
mi
z
a
t
i
o
n
, evolu
t
ionary a
l
gor
ith
m
Copy
right
©
2016 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
Membrane
computing i
s
a
recent dom
a
i
n of
natural
comp
uting
started by Gh.
P
ă
un in
1998 a
nd i
s
known as P
system
s or m
e
mbra
ne sy
st
e
m
s [1]. It is a parall
e
l di
stri
buted
comp
uting
model in
spi
r
e
d
from the fu
nction
ality an
d structu
r
e of
living cell
s a
nd the inte
rsection
of the
m
in
the tissue,
organ
and
ne
ural net
work. T
here
a
r
e th
re
e
main
cla
s
ses of
P syste
m
s
i
n
vestigat
ed:
cell-li
ke
P sy
stems, tissue
-li
k
e P
syste
m
s and
neu
ral
-
li
ke P
syste
m
s [2], [3, 4]. These
computin
g
system
s a
r
e
prove
d
to
b
e
so po
we
rf
ul that t
hey
are i
n
some
way
s
eq
uivalent
with T
u
ring
machi
n
e
s
[5], the powe
r
ful global
sea
r
ch
capa
city of membrane
co
mputing is al
so proved [6].
Clu
s
ter an
alysis al
so
calle
d clu
s
terin
g
, is a
typical ex
ample of un
supervi
sed le
a
r
ning, it
aims
at pa
rtitioning
a set
of data o
b
je
cts into
su
b
s
et
s, ea
ch
su
bset calle
d a
cl
uster,
su
ch t
hat
obje
c
ts in a
clu
s
ter a
r
e
si
milar to ea
ch
other, yet dissimil
a
r to o
b
ject
s in oth
e
r cl
uste
rs [7
, 8].
Clu
s
terin
g
a
n
a
lysis ha
s b
een
widely
u
s
ed
in va
riou
s field
s
,
su
ch a
s
ima
ge
pro
c
e
ssi
ng, d
a
ta
analysi
s
, We
b se
archin
g,
market an
alysis
etc.
As
o
ne of the m
o
st cla
s
sical cl
usteri
ng m
e
thod
based on p
a
rtitioning, K-m
ean
s algo
rith
m is an un
su
pervised cl
ustering alg
o
rith
m for cla
ssifyi
ng
data points i
n
to
k
distin
ct grou
ps. It is the most
cla
ssi
cal cl
uste
ri
ng method d
ue to its simple
prin
ciple a
nd easily
appli
c
a
t
ion [9]. However, the final clus
te
ring re
sults of it dep
end heavily o
n
the initial ran
dom
sele
ctio
n su
ch
that th
e K-me
an
s m
e
thod i
s
n
o
t g
uara
n
teed
to
conve
r
ge
to the
global o
p
timu
m and often trapp
ed into th
e local o
p
timum [10].
As a novel
computing m
o
del inspired b
y
biol
ogy, the P system ha
s attra
c
ted m
o
re a
n
d
more people’
s attention. Due to
the
parallel
di
stributed
comput
ing ability of
the P sy
stem, a
numbe
r
of p
apers have appe
are
d
in whi
c
h
P syst
em
are hyb
r
i
d
ize
d
with
ot
her
evolution
a
ry
algorith
m
s li
ke Multi-level
thresholdi
ng
method
s
[11]
, Particle Swarm O
p
timization [12, 13
],
Geneti
c
Algo
rithms [14], Ant Colo
ny Op
timization [1
5
], Quantum I
n
spi
r
ed
EAs [
16], etc. Th
e
s
e
modificatio
n
s obviou
s
ly m
ade
many im
provem
ent
s o
n
the
evolutio
nary
algo
rith
m an
d p
r
om
o
t
ed
the developm
ent of the P system re
sea
r
ch.
In ord
e
r to
attain a bette
r
clu
s
terin
g
result, more im
p
r
oveme
n
ts h
a
ve made
on
cu
rre
nt
clu
s
terin
g
m
e
thod
s. In this pap
er, a
modified in
versio
n parti
cle swa
r
m o
p
timization
wa
s
prop
osed, thi
s
metho
d
an
d the mutati
onal me
ch
an
ism of ge
net
ics
algo
rithm
were u
s
ed
to
combi
ne with
the tissue
-
like P system, Thro
ugh t
he
spe
c
ial d
e
sig
ned tissue
-like P system a
n
d
the evolution
mech
anism and commu
n
i
cation me
ch
anism, the id
ea of the no
vel membra
n
e
clu
s
terin
g
alg
o
rithm
could
reali
z
ed. Th
e
rest of
the
pape
r is o
r
g
anized a
s
fol
l
ows. Section
2
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 25
02-4
752
IJEECS
Vol.
2, No. 2, May 2016 : 409 –
416
410
descri
b
e
s
bri
e
fly partition-based
cluste
r algorith
m
. Section
3 intro
duces ti
ssue-like P sy
ste
m
.
Section 4 prese
n
ts a vari
ant of the PSO al
gorithm
and the pro
posed mem
b
rane
clu
s
teri
ng
algorith
m
. Section5 illu
strat
e
s expe
rime
n
t
al resu
lt
s. Finally, Section
6 make
s con
c
lu
sion
s.
2. Partition-Bas
e
d Clus
ter Algorithm
Clu
s
ter an
alysis i
s
a pro
c
e
ss of pa
rtition
dat
a points i
n
to different clu
s
ters due t
o
their
simila
rity. As a powerful te
chni
que of da
ta mining,
clu
s
terin
g
is wi
d
e
ly used in m
any fields, su
ch
as
bu
sine
ss i
n
telligen
ce,
Web
sea
r
ching, biol
ogi
cal
,
safety a
nd
so
on. M
a
jor ba
sic cl
uste
rin
g
method
s can
be divided int
o
the followin
g
categ
o
ri
e
s
:
partitionin
g
m
e
thod, hierarchi
c
al metho
d
,
den
sity-ba
s
e
d
metho
d
, grid-ba
s
e
d
met
hod [7]. Sup
pose that div
i
ding the
dat
a point
s into
k
clu
s
ters ba
se
d on partition
-ba
s
ed
clu
s
te
ri
ng alg
o
rith
m, the k cluster cente
r
s, z
1
, z
2
, …, z
k
, and
the co
rre
sp
on
ding cl
uste
rin
g
partition
s are C
1
, C
2
, … ,
C
k
re
spe
c
tivel
y
. The data point xi belong
s
to the cluste
r
p if the distan
ce meet
s:
,,
1
,
2
,
p
ip
i
q
x
zx
z
p
q
q
…
,
k,
(1)
Once
clu
s
teri
ng p
a
rtition
s
are fo
rmatted,
ne
w cl
uster
cente
r
s a
r
e
com
puted
by th
e
mean
s of the points in the
corre
s
p
ondin
g
clu
s
ter u
s
in
g
1
j
pj
p
xc
j
zx
n
(2)
Whe
r
e nj is th
e numbe
r of data point
s in
cluste
r j
and
Cj is the cl
ust
e
ring p
a
rtition
of cluste
r j.
In this partitio
n
-ba
s
e
d
clu
s
t
e
ring m
e
thod,
the
cluste
rin
g
metric M i
s
determi
ned a
s
follows:
12
k
1
ji
k
j
i
ix
c
M
CC
C
x
z
,,
…
,
(3)
Obviously, smaller M val
u
e stan
ds fo
r
better
cluste
ri
ng re
sult
s, su
ch that thi
s
cl
usteri
ng
probl
em can
be co
nsi
dere
d
as an o
p
timiza
tion p
r
o
c
e
ss of findin
g
the small
e
r M.
3. The Tissu
e-Like P Sy
s
t
em
Membrane
computing i
s
a
recent dom
a
i
n of
natural
comp
uting
started by Gh.
P
ă
un in
1998. It i
s
al
so
kn
own a
s
membrane
computing
o
r
membrane
systems. In
re
cent ye
ars,
many
different mo
d
e
ls of P
syst
ems
have b
e
en p
r
opo
se
d,
su
ch a
s
cell
-like
P
sy
st
e
m
s,
t
i
s
s
ue
-lik
e P
system, and
spi
k
ing n
eura
l
P system [17-21].
The o
b
t
ained compu
t
ing system
s
have prove
d
to
be
so p
o
werf
ul that it is
e
quivalent
with Tu
ring
ma
chine
s
[22]. T
he P
system
s a
r
e
a cl
ass of
distrib
u
ted p
a
r
allel
comp
uting devi
c
es
of a bioc
hemi
c
al type [23], membrane
co
mputing n
o
w
is a
hot cro
s
s-di
sciplin
e topi
c
whi
c
h involve
s
comp
uter
scien
c
e, m
a
th
ematics, biol
ogy and
artifi
cial
intelligen
ce,
etc. One P
system mai
n
ly con
s
i
s
ts of
th
ree p
a
rt
s: me
mbra
ne st
ru
cture, multisets of
obje
c
ts
and
evolution
rule
s [1]. In tissu
e
-like P
syst
em, memb
ra
nes or cells
are
pla
c
ed
a
s
the
node
s of a graph. Th
e ne
t of nodes d
eals
with
symbols a
nd communi
cate
s symbols al
o
ng
cha
nnel
s spe
c
ified in adva
n
ce. Th
e co
mmuni
cation
among
cell
s i
s
ba
sed o
n
symport/antipo
r
t
rule
s [24]. Symport
rule
s
move obje
c
t
s
acro
ss
a
me
mbra
ne tog
e
ther i
n
on
e direction,
whe
r
e
a
s
antiport
rul
e
s move
obje
c
t
s
a
c
ro
ss a
m
e
mbrane
in
o
ppo
site di
re
ctions [25]. Th
e tissue
-like
P
system
whi
c
h co
ntain
s
m eleme
n
tary membra
ne
s is
sh
own i
n
Figu
re 1
and d
e
scri
be
d as
follows
:
1m
1
m
0
=,
,
,
'
,
)
R
RR
i
(,
…
,,
…
(4)
(1)
i
(1
≤
i
≤
m) is
a
fi
nite set of obje
c
ts in elementa
r
y membran
e
i ;
(2) R
i
(1
≤
i
≤
m) is a
fi
nite set of evolution rule
s in el
ementa
r
y membra
ne i;
(3)
R’ i
s
a
fi
ni
te set of com
m
unication
ru
les
with the two follo
wing f
o
rm
s: antipo
r
t rule: (
݅
,
ܼ
/
ܼ
',
݆
),
݅
,
݆
= 1,
2,
…, m ,
݅
݆
. Th
e rule i
s
use
d
to
comm
un
icate th
e o
b
j
e
cts bet
wee
n
an
eleme
n
ta
r
y
Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS
ISSN:
2502-4
752
A Novel M
e
m
b
ran
e
Clu
s
te
ring Algorithm
Based o
n
Tissue
-like P system
(Xiyu Li
u)
411
membrane
a
nd its n
e
igh
b
o
r; symp
ort rule: (
݅
,
ܼ
/
ߣ
, 0),
݅
=
1, 2,
…, m. The
rule i
s
u
s
ed
to
comm
uni
cate
the best obje
c
ts bet
wee
n
membrane
s a
nd the memb
rane
with the environ
ment.
(4) i
0
indi
cate
s the output region of the system.
F
i
gure 1.
The desi
gne
d tissue-li
ke P syst
em
The st
ru
cture
and the
spe
c
ial me
ch
ani
sm of this tissue-li
ke P sy
stem are th
e key point
of the me
mb
rane
cl
uste
ri
ng al
gorith
m
. The
obje
c
ts of the P
sy
stem a
r
e
re
pre
s
ent
as the
possibl
e clu
s
t
e
ring
ce
nters,
su
ch t
hat th
e dimen
s
io
n
of each obje
c
t is
k x d
, in which the clust
e
r
numbe
r is
k
and the dim
ensi
on of the data point
s
is d. In this tissue
-like P systems,
each
elementa
r
y membrane in
depe
ndently
work in a m
a
xi
mally parall
e
l way. The
comp
utation
of a
t
i
ssu
e-li
ke
P
sy
st
em
is
a
seq
uen
ce
of
comp
ut
ing
st
eps,
whi
c
h
starts from
the
m el
eme
n
ta
ry
membrane
s contai
ning w
1
, … , w
m
. At each
cal
c
ul
ation ste
p
, o
ne o
r
mo
re
rules
are ap
pl
ied
utilizing
the multisets of
object
s
.
.When the computation met the st
op criteri
on;
the final results
are sho
w
ed i
n
the output region.
4. The Propo
sed Membra
ne Clus
terin
g
Algorithm
4.1. An Improv
ed
In
v
e
rsi
on Particle Sw
arm Optimi
z
a
tion
Particle
swa
r
m optimi
z
at
ion is
a p
opulat
io
n-b
a
sed sto
c
h
a
sti
c
optimal
a
l
gorithm
prop
osed by
Kennedy a
n
d
Eberha
rt in
1995. Th
e b
a
si
c idea
of this alg
o
rithm
simulate
s bi
rd
flocki
ng or fish schooli
ng b
ehavior to bui
ld a self-e
volv
ing system [1
8]. As a repre
s
entative of the
swarm intelli
g
ence algo
rith
ms, pa
rticle
swarm optim
i
z
ation algo
rith
m is u
s
ed to
solve
continu
ous
optimizatio
n probl
em
s ori
g
inally [26].
In this
algori
t
hm, each p
a
rticle
wa
s consi
dered a
s
a
potential solu
tion to an opti
m
al que
stion,
these p
a
rticl
e
s flied at a
specifi
c
sp
eed
and then
adju
s
t
spe
ed
with th
eir o
w
n
and t
he othe
r p
a
rti
c
le
s’ flying ex
perie
nce, eve
r
y parti
cle h
a
s
an
obje
c
tive
function to d
e
termin
e its fitness whi
c
h
wa
s used fo
r evaluate itse
lf, during this iteration cou
r
se,
the parti
cle fl
ied toward t
he be
st po
sit
i
on, and th
e
optimizatio
n
probl
em ten
d
s to o
b
tain
its
optimal soluti
on, The findi
ng process o
f
local b
e
st
a
nd glo
bal be
st that are co
mputed in
every
iteration give
s the p
o
tentia
l to this meth
od to re
ach th
e most o
p
tim
a
l sol
u
tion[27
]. Particle swarm
optimizatio
n
also
ha
s som
e
dra
w
b
a
cks, for exampl
e,
it is ea
sy tra
p
into the lo
cal optimum
a
nd
be ab
se
nce
of popul
ation
diversity. Some certain
m
odi
fi
catio
n
s
h
a
ve bee
n ma
de in
ba
sic P
S
O
algorith
m
, for instan
ce, in
ertia wei
ght, reverse
sea
r
chin
g, para
m
eter modifi
cat
i
on and
so fo
rth
[28]. In this paper, we pro
pose a modifi
ed inve
rsion
particl
e swa
r
m optimizatio
n (MIPSO) b
a
se
d
on the i
dea
o
f
avoiding th
e
wo
rst
po
sitio
n
in o
r
de
r to
improve th
e
exploratio
n
capa
city and t
h
e
diversity of
the
whol
e
swarm [2
9]. Su
ppo
se th
at
N pa
rticle
s
exist in
D-dime
nsio
nal
se
arching
spa
c
e, the pa
rticle up
date
s
its velocity and po
sition in
our metho
d
as follo
ws:
1
12
()(
)
()(
)
kk
k
k
id
id
iw
i
d
w
i
d
v
v
c
r
and
p
x
c
r
a
nd
pg
x
(5)
11
kk
k
id
id
id
xx
v
(6)
whe
r
e d = 1,2
,
…,
D; i = 1,2, …,
N, v
k+1
id
is the
i-th
particl
e’s ne
w velo
city at the
kth
iteration
step,
the rang
e of Vid is [-Vma
x
,Vmax],
such per-set value preve
n
ts t
he pa
rticle from
0
w
1
w
2
w
m
……
1
2
m
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 25
02-4
752
IJEECS
Vol.
2, No. 2, May 2016 : 409 –
416
412
flying out of the sea
r
ch
sp
ace;
k i
s
the
iteration
n
u
m
ber; xk+1i
d
is the
i-th pa
rticle’
s
po
sition
of
the k+1 times, based on i
t
s previou
s
p
o
sition
an
d n
e
w velocity a
t
the last iteration;
ω
is
inertia
weig
ht; piw is the wo
rst fitness value
of the parti
cle i
t
self achieve
d
so fa
r, and
pgd
is th
e wo
rst
global fitne
ss the wh
ole
swarm attaine
d
so fa
r; c
1
, c
2
are l
earnin
g
factors;
ra
n
d
()
are
ra
ndom
numbe
rs unif
o
rmly distri
bu
ted in the ran
ge [0, 1
]. The performan
ce
of each pa
rti
c
le is eval
uat
ed
according to
a pre
define
d
fitness fun
c
t
i
on, whi
c
h i
s
usu
a
lly prop
ortional to th
e co
st functi
on
asso
ciated
with the probl
em. Whe
n
finding the
o
p
t
imum or me
eting the ma
ximal iteratio
n
numbe
rs, the sea
r
ching p
r
o
c
e
ss h
a
lted.
In this metho
d
, we con
s
id
er the pa
rticl
e
sea
r
chin
g for the be
st condition by
combinin
g
the worst flying exp
e
rie
n
ce of itself and
the
wors
t flying exp
e
rie
n
ce of the
wh
ol
e po
pulation,
the
particl
e adj
usts its velocity
by flying towards th
e co
ntrary di
re
ction of
the wo
rst po
sition, t
he
fitness valu
e
is ado
pted to
evaluate the
particl
e.
Lea
rning from th
e wo
rst expe
rien
ce a
nd ru
n
away from it is the ba
sis id
ea of this imp
r
oved alg
o
rith
m.
4.2. The Ev
olution Rule
s and Commu
nicati
on Rules of th
e Tis
s
ue-Like P Sy
stem
In each
ele
m
entary me
mbra
ne of the tiss
ue
-like
P system, n numb
e
r o
b
ject
s are
contai
ned. Every obje
c
t re
pre
s
ent
s the
k po
ssi
ble
cl
u
s
ter
cente
r
s,
evolution rule
s are appli
e
d
to
evolve the
obje
c
ts, an
d
the evoluti
on me
ch
ani
sm is d
e
si
gn
ed ba
se
d o
n
pa
rticle
swarm
optimizatio
n, its improve
d
versi
on MIPSO
and mutati
on rule
s of ge
neratio
n algo
rithm.
In the p
r
o
p
o
s
ed
tissue
-like P
system,
ea
ch
eleme
n
tary o
w
n
s
t
he
same
n
numbe
r
obje
c
ts, d
u
rin
g
the
evolutio
n, these
obje
c
ts
are
rand
o
m
ly divided
i
n
to two
pa
rts ro
ughly, o
n
e
of
the
pa
rts exe
c
ute stan
dard
parti
cl
e
swa
r
m optimi
z
atio
n while th
e ot
her
ca
rry
out
the MIPSO, the
two optimi
z
e t
he obje
c
t
s
through
by finding the be
st
p
o
sition
and a
v
oid the wo
rst position. In t
h
is
work, mutatio
n
mechani
sm
in the bina
ry codin
g
ge
ne
ration al
gorit
hm is al
so u
s
ed to
coevol
ve
the obj
ect
s
, a
nd it ma
ke
s
mutation a
ccordin
g to th
e
possibl
e pm,
if the value
of
a mutatio
n
p
o
int
at dimensi
on
j is v, after mutation the value be
come
s
that
,0
'
,0
vv
v
v
vv
(7)
In the above formula, the signs
“+” or “
−
” o
c
cur wi
th equal pro
bability, and
ߜ
is real
numbe
r g
ene
rated
with uni
form di
stributi
on in t
he
ran
ge [0, 1]. The
optimizatio
n
pro
c
e
ss
halte
d
until it meets the maximum
iteration crite
r
ia.
The commu
ni
cation
rule
s i
n
the desi
gne
d P sy
stem m
a
inly have two types: antip
ort rule:
(
݅
,
ܼ
/
ܼ
',
݆
) ,
݅
,
݆
= 1,
2,
. . .
, m ,
݅
݆
, includ
ing m is the numbe
r of elementa
r
y membra
ne
s, this
rule i
s
u
s
e
d
to co
mmuni
cate th
e obj
ects bet
wee
n
an
eleme
n
tary mem
b
rane a
nd it
s
two
neigh
bor; sy
mport rul
e
: (
݅
,
ܼ
/
ߣ
, 0),
݅
= 1, 2, .
. . , m, this rul
e
is u
s
ed to comm
u
n
icate the o
b
jects
betwe
en me
mbra
ne and
the environ
ment, of wh
ich m is the
number of the elementa
r
y
membrane
s.
The lab
e
l 0
stand
s for th
e output
regi
on nam
ely e
n
vironm
ent. The role of the
comm
uni
cati
on me
ch
ani
sm is to
com
m
unicate the
optimal
obj
ect of
each
membrane
a
n
d
eventually attain the glob
al optimum whi
c
h sto
r
e
s
in the output re
gi
on. The refe
rence point is
the
fitness o
b
je
ction that is the M value,
4.3. Descrip
tion of the Pr
oposed M
e
mbrane
Cluste
ring Algorith
m
In this pap
er,
a new m
e
m
b
ran
e
cl
uste
ri
ng algo
rithm
based on p
a
rtitioning clu
s
t
e
ring i
s
prop
osed through utilizi
n
g
the rules of
evolut
ionary
algorithm, communi
catio
n
rule
s and
the
spe
c
ial fram
e
w
ork withi
n
the distri
buted
paralle
l com
putation fram
ewo
r
k p
r
ovid
ed by tissue
-
l
i
ke
P system. T
he obj
ect
s
of
each
cell a
r
e re
pre
s
e
n
te
d as the p
o
ssible
clu
s
te
r
cente
r
s. P
a
rti
c
le
swarm
optimi
z
ation
alg
o
rit
h
m a
nd
a m
o
dified
reve
rse
pa
rticle
swa
r
m optimi
z
atio
n alg
o
rithm
a
r
e
applie
d a
s
ev
olution m
e
ch
anism
to evol
ve the obj
ect
s
. Mutation
mech
ani
sm o
f
the gen
erati
on
algorith
m
is
also im
po
rte
d
as
co
evol
ution ru
les i
n
order to
improve th
e
diversity of t
h
e
popul
ation. Communi
catio
n
rul
e
s
between
cell
s o
r
betwe
en
cell
s an
d e
n
viro
nment i
s
u
s
e
d
to
excha
nge th
e
obje
c
t and
a
rrive to attain
the optimum
. This n
e
w
m
e
mbrane
clu
s
tering
algo
rith
m
can
autom
atic
sea
r
ching
the o
p
timum
clu
s
ter
ce
nt
e
r
s
and
arrive
at better cl
u
s
ter
re
sult
s, the
c
l
us
te
r
i
ng
a
l
go
r
i
th
m is
de
sc
r
i
be
d
as
fo
llo
ws
:
Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS
ISSN:
2502-4
752
A Novel M
e
m
b
ran
e
Clu
s
te
ring Algorithm
Based o
n
Tissue
-like P system
(Xiyu Li
u)
413
Step 1.
Gen
e
rate
n i
n
itia
l obje
c
ts (th
e
po
ssible
cluster
cente
r
s) fo
r e
a
ch of
the
m
elementa
r
y m
e
mbrane, th
e
n
di
stingui
sh
clu
s
terin
g
reg
i
ons a
c
cordi
n
g to th
e exi
s
ting fo
rmula
a
nd
cal
c
ulate
the new clu
s
ter
centers.
Step 2. Initial a group
of i
ndividual
s ra
ndomly, divid
e
the in
dividuals i
n
to two
portio
n
s
with the
equ
a
l
scale, initial
the velo
city and t
he
po
sitio
n
, set th
e mu
tation proba
bi
lity pm, fitness
function f.
Step 3. Execute the p
a
rti
c
le
swarm
o
p
timi
zation
a
nd the
modif
i
ed inve
rsi
o
n
parti
cle
swarm o
p
timization, mutati
on rule
s a
r
e a
pplied
simulta
neou
sly.
Step 4. Co
mmuni
cation
rule
s are
use
d
to exchang
e the o
p
timum between the
elementa
r
y membrane
s an
d elementa
r
y membrane
wi
th environm
e
n
t.
Step 5. Rep
eat these ste
p
s u
n
til attain t
he maxim
u
m iteratio
n
crite
r
ia, an
d
the final
global o
p
timu
m is sh
owe
d
in the environ
ment.
5. Experiments and
Res
u
lts
The p
r
op
ose
d
memb
ran
e
clu
s
terin
g
a
l
gorithm i
s
a
s
sesse
d
on
six data
set
s
an
d
comp
ared wit
h
K-mean
s a
l
gorithm, PSO-K-m
ean
s
a
l
gorithm a
nd
GA-K-m
ean
s algorithm [3
0,
31]. For illust
rating the results of
the proposed algori
thm
exactly,
30 runs were executed when
applying on
e
of the tested algorithm
s.
These
six
data sets in
cl
uding two a
r
tificial data sets
named A
r
t1, Art2 and fou
r
real life d
a
ta sets
name
d
Iris, Ca
ncer, Vowel,
Wine a
r
e used. All da
ta
sets except
Art1 and Art2 are
availabl
e at ftp://ftp.i
cs.uci.edu/pu
b/machi
n
e-learni
ngdatabases/.
Table 1
sum
m
ari
z
es the characteristi
c
s of these dat
a sets. Data sets Art 1, Art 2 is illustrat
e
in
Figure 2, Figure 3.
Table 1. Ch
aracteri
stics of
the data sets
con
s
id
ere
d
Name
of
NO.
of
NO.
of
Siz
e
of da
ta se
t
Date se
t
classes
features
(size of classe
s in parenth
e
ses)
Art1
3
3
900 ( 300,
300, 3
00 )
Art2
2
3
900 ( 450,
450 )
Iris
3
4
150 ( 50, 5
0
, 50 )
Cancer
2
9
683 ( 444,
239 )
Vowel
6
3
871 ( 72, 8
9
. 172
, 151, 207, 18
0 )
Wine
6
13
178 ( 59, 7
1
, 48 )
Figure 2. Art 1
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 25
02-4
752
IJEECS
Vol.
2, No. 2, May 2016 : 409 –
416
414
Figure 3. Art 2
The value
s
of the param
eter in these te
sted alg
o
rith
ms are sh
own in Table 2.
Table 2. Para
meter setting
Parameter
ω
c
1
c
2
P
m
M
Value
0.9~0.4
1.2 1.2 0.05
1000
ω
is the in
ert
i
a wei
ght, the value i
s
be
tween
0.4 an
d 0.9, sm
all
value imp
r
ov
es lo
cal
conve
r
ge
nce
cap
a
city and
the larg
e valu
e
improve
s
gl
obal conve
r
g
ence ca
pa
city; c
1
and c
2
ar
e
learni
ng factor; pm
is mutation pr
obability; M is the m
a
ximum iterat
ion. In order
to investigate
the effect of
e
l
ementa
r
y me
mbra
ne n
u
m
bers, four di
fferent ti
ssue-li
ke P
system
with 4, 8, 1
6
,
20
elementa
r
y m
e
mbrane
s a
r
e de
sign
ed. In order to
ov
ercome th
e e
ffects of a
c
ci
dental fa
ctors, we
cho
o
se the a
v
erage valu
e
of the every whol
e 30
tim
e
s run
s
. The
M values of these P syste
m
s
are sho
w
n in
Table 3.
Table 3. Co
m
pari
s
on of the
M values of diffe
rent-elem
entary memb
rane tissu
e
-li
k
e P system
s
Data se
t
4
8
16
20
Art
1
628.3298
627.2565
624.4532
634.0342
Art
2
546.7652
545.3545
539.2868
539.7662
Iris
99.9238
98.7596
98.6324
98.9328
Cancer
3253.3342
3052.5372
3249.6575
3250.0442
Vowel
149315.409
6
149311.874
1
149309.486
3
149309.822
3
Wine
16313.6521
16306.4536
16293.5662
16293.7976
Table
3 indi
cates that ti
ssue-li
ke P
syst
em
with
1
6
elementa
r
y membrane
s behave
s
better than th
e others
with the small
e
r M
value in the
s
e six test dat
a set
s
. For th
e true life dat
a
set of
Ca
ncer, the
pe
rfo
r
man
c
e
of th
e 16
elem
en
tary memb
ra
ne P
system
is
324
9.657
5,
sup
e
rio
r
to the other 32
51.
3342, 30
52.5
372 an
d 325
0
.
0442 obvio
u
s
ly.
Gene
rally, more p
a
rticl
e
s i
n
the swarm
coul
d enh
an
cing the searching range
a
nd may
lead
s b
e
tter
optimizatio
n
result
s. To
in
quiry the
a
p
p
r
op
riate
num
ber of the
pa
rticle
s, differe
nt
scale of the
swarms
are
d
e
sig
ned fo
r t
he expe
ri
me
n
t. Table 4
sh
ows the dive
rse
re
sults
of the
different pa
rticle scal
es.
Table 4. Co
m
pari
s
on of the
M values of
the different p
a
rticle
scale swarms
Data
se
t
20 50 100
200
400
Art
1
629.2378
624.9662
624.5342
624.4532
624.4532
Art
2
550.4574
546.6753
539.7662
539.2868
539.2868
Iris
104.2438
101.7456
99.9458
98.6324
98.6324
Cancer
3257.5664
3053.8698
3252.6572
3249.6575
3249.6575
Vowel
149320.403
3
149314.654
1
149310.622
2
149309.486
3
149309.486
3
Wine
16317.3496
16311.4581
16295.2803
16293.5662
16293.5662
Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS
ISSN:
2502-4
752
A Novel M
e
m
b
ran
e
Clu
s
te
ring Algorithm
Based o
n
Tissue
-like P system
(Xiyu Li
u)
415
Table 4
shows that when
t
he num
ber
of particl
es
attains 2
00, the
clustering quality will
be better th
a
n
the sm
all-scale
swarm
s
.
For the
Win
e
data set, the M value is
1629
3.566
2 with
the pa
rticl
e
scale
is 2
00,
obviou
s
ly sm
aller t
han
the
sm
aller
scal
e with
the
va
lue 1
631
7.34
96
,
1631
1.458
1, 1629
5.280
3. Too m
any p
a
rticle
s
ma
y
doe
s n
o
hel
p
in imp
r
oving
the
clu
s
terin
g
quality, the
cl
usteri
ng
re
sul
t
s have
no
ch
ange
when t
h
e num
be
r rose from
20
0 to
400,
and
thu
s
the approp
ria
t
e numbe
r co
uld define 2
0
0
in our mem
b
ran
e
clu
s
te
ri
ng method.
Table 5. Co
m
pari
s
on of the
M value of the
tested alg
o
rithms on the
six data set
s
(%)
Data se
t
G
A
PSO
K-me
ans
MC
A
Art
1
649.9533
639.2434
649.1573
623.3662
Art
2
542.0232
541.6982
552.3472
539.0848
Iris 98.9734
98.0531
106.1347
97.0324
Cancer
3249.8713
3048.2311
3257.6725
3248.5472
Vowel
149310.090
8
149309.749
2
149336.964
6
149307.422
3
Wine
16293.3142
16292.0419
16318.4763
16290.7976
Acco
rdi
ng to
the re
sults
of the ch
art in T
able
5, we co
uld
con
c
lude that,
in
most of
run
s
of these alg
o
rithm
s
, the prop
osed me
mbra
ne clu
s
t
e
ring meth
od
is supe
rio
r
to other test
ed
algorith
m
. Fo
r Art 1, the value of the
MCA is 6
23.
3662, mu
ch l
e
ss than the
PSO which
is
639.23
34, the GA which is 649.9
533 a
nd the K-
me
ans alg
o
rith
m which is 6
49.157
3. The
M
value on the
Vowel is 1
4
9
307.42
23, sm
aller t
han the
K-mean
s, GA
-K-me
a
n
s
an
d PSO-K-me
ans
with 14
933
6.9645,
1493
0
9
.7492
and
1
4931
0.090
8, respe
c
tively. 30 time
s run
s
of e
a
ch te
sted
algorith
m
lo
ws the
co
nting
ency
of the
experim
ent a
nd en
han
ce
the pe
rsua
si
on of the
te
st.
Thro
ugh
co
m
parin
g an
d a
nalyzin
g, the
better
cl
u
s
tering
quality
of the prop
ose
d
mem
b
rane
clu
s
terin
g
alg
o
rithm could
be prove
d
.
6. Conclusio
n
s
At pre
s
ent th
ere
are m
any novel te
ch
ni
que
s
committ
ed to i
m
proving the
data
cl
usteri
ng.
This
pap
er fo
cu
se
s on
re
al
izing
a mem
b
rane
cl
uste
rin
g
algo
rithm b
a
se
d on
a d
e
s
ign
ed tissu
e
-
like P
system
. With the
structure of th
e
tissu
e-li
ke P
system, p
a
rti
c
le
swarm op
timization
an
d a
n
modified i
n
versi
on
parti
cl
e swa
r
m o
p
timization
are
applie
d a
s
the evolution
rule
s
and t
h
e
mutation me
cha
n
ism
of the gen
eratio
n algo
rithm
i
s
u
s
ed a
s
th
e su
pplem
en
t of the evolution
rule
s an
d e
n
han
ce the
di
versity of the
popul
ation
simultaneo
usly
, commu
nication rules which
contai
n antip
ort rule
s an
d symp
ort
rules
are em
ployed to
co
mmuni
cate t
he optim
um, ou
r
prop
osed m
e
mbrane
clu
s
terin
g
algo
ri
thm behave
s
supe
rio
r
to the some
other cl
uste
ri
ng
algorith
m
, the perfe
ct performan
ce
re
sult of t
he simulation exp
e
rime
nt prov
e this clai
m. The
further
work con
c
entrate
upon usi
n
g
more evolu
t
ionary algo
rithm base
d
on memb
ran
e
comp
uting in
orde
r to improve the clu
s
teri
ng m
e
thod
s and e
nha
nce the clu
s
teri
ng quality.
Ackn
o
w
l
e
dg
ements
Proje
c
t supp
orted
by
Na
tional
Natura
l Scie
nce F
ound
ation
of Chi
na
(6
14
7223
1,
6117
0038,
6
1502
283
),
Ji
nan
City in
depe
ndent
i
nnovation
pl
an p
r
oj
ect i
n
Colleg
e
a
nd
Universitie
s
, Chin
a (201
4
0120
2),
Mi
ni
stry
of
edu
cation of Hu
manities a
n
d
social
sci
e
n
ce
resea
r
ch proj
ect, Chin
a (1
2YJA63
015
2), Social
Scie
nce Fu
nd Project of Shan
dong Province,
Chin
a (11
C
G
L
J2
2).
Referen
ces
[1]
G
P
ǎ
un, Roze
nber
g and A
Salom
aa.
T
h
e
Oxford Hand
book of Me
mbranc
e Co
mp
uting
. Ox
ford
Univers
i
t
y
Pr
es
s, Ne
w
Y
o
rk. 2010.
[2]
G
P
ǎ
un. Computing
w
i
th
membranes.
Jour
nal
of C
o
mput
er a
nd Syste
m
Scie
nces
. 20
00
; 6
1
(
1
)
: 1
08-
143.
[3]
C Martin-Vide,
J Pazos, Gh P
ǎ
un, A R
o
d
r
igu
e
z-Paton.
T
i
ssue P s
y
st
ems.
T
heoreti
c
al C
o
mpute
r
Scienc
e
. 200
3; 296(2): 29
5—
326.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 25
02-4
752
IJEECS
Vol.
2, No. 2, May 2016 : 409 –
416
416
[4]
M Ionescu, Gh P
ǎ
un, T
Yokomori. Spiki
ng n
eura
l
P s
y
stem
s.
F
unda
ment
a
Informatic
ae
. 200
6; 71(2-
3): 279—3
08.
[5]
G
P
ǎ
un,
Roz
e
nber
g G
an
d S
a
lom
aa A.
M
e
mbr
a
n
e
C
o
mp
ut
ing
. O
x
for
d
Univers
i
t
y
Pres
s, Ne
w
York
.
201
0.
[6]
G Z
hang, J Cheng, M Gheor
g
he an
d Q Men
g
. “A h
y
br
id p
p
r
oach b
a
se
d o
n
differenti
a
l e
v
oluti
on a
n
d
tissue m
e
mbra
ne
ystems
for
solvin
g c
onstr
ain
ed m
a
n
u
fac
t
uring
par
amet
er o
p
timizati
on
pro
b
lems”.
Appl
ied S
o
ft Computi
ng Jo
urnal
. 20
13; 13(
3
)
: 1528–
15
42.
[7]
J Han, M Kam
ber. Data Mining:
Conce
p
ts an
d T
e
chni
qu
es
. Morga
n
Kaufm
ann, San F
r
a
n
c
isco. 200
1.
[8]
N Kinos
hita,
Y Endo. On
Objective-B
a
se
d Ro
ugh
Har
d
an
d F
u
zz
y
c-Means C
l
ust
e
rin
g
.
J. of
Advanced Com
p
utational In
t
e
lligence and In
telligent Inform
atics (JACIII)
. 2014; 19(
3): 29-35.
[9]
ZT
ang, H Z
han
g. Improv
ed K-m
eans
Clust
er
in
g
Algorit
hm Ba
sed
on Ge
n
e
tic Al
gorithm
.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2014; 1
2
(3): 1
917 - 19
23.
[10]
R Maitra,
A P
e
terson, A
Gh
os
h. A s
y
stem
atic ev
alu
a
tio
n
of d
i
ffe
rent m
e
thods
for
initi
a
lizi
n
g
the
K-
means cl
usteri
ng al
gorithm.
IEEE Trans. Knowl. Data Eng
. 201
0.
[11]
H Pen
g
, J W
a
n
g
, MJ Pérez-Ji
ménez, P S
h
i.
A nove
l
ima
ge
thresho
l
di
ng m
e
thod
bas
ed
o
n
membr
a
n
e
computi
ng a
n
d
fuzz
y
entro
p
y
.
Journal of Intelligent & Fu
z
z
y
System
s
. 20
13
; 24(2): 229
–23
7.
[12]
W
Pan –T
sao
.
Combi
n
in
g
PSO cluster a
nd n
o
n
lin
ear
mapp
ing
al
gor
ithm to p
e
rfor
m clusteri
n
g
performa
nce
a
nal
ysis:
take t
he
enter
prise
fina
ncia
l
a
l
armi
ng
as e
x
amp
l
e
.
Qualit
y &
Qu
antit
y
.
20
11
;
45(6): 12
91-
13
02.
[13]
MY Ch
eng,
K
Y
Hu
an
g, HM
Ch
en. K-m
e
a
n
s p
a
rticle
s
w
arm o
p
timizati
on
w
i
th
emb
e
dde
d c
haoti
c
search
for so
l
v
ing
multi
d
ime
n
sio
nal
pr
obl
e
m
s.
Appl
ied
M
a
the
m
atics
a
n
d
C
o
mputati
o
n
. 201
2; 2
19:
309
1-30
99.
[14]
Md Anis
ur
Ra
hman, M
d
Z
a
hid
u
l Isl
a
m. A
h
y
bri
d
c
l
uster
i
ng t
e
chn
i
q
ue
combi
n
in
g a
n
o
vel
ge
neti
c
alg
o
rithm w
i
th
K-Means.
Kno
w
ledge-
bas
ed Systems
. 20
14
; 71: 345-3
65.
[15]
GX Z
han
g, JM
Chen
g, M Gheorg
he.
An a
p
p
roxi
mat
e
al
go
rithm c
o
mbi
n
in
g P ystems
an
d ant col
o
n
y
opti
m
i
z
at
ion
for trave
lin
g s
a
les
m
an
pro
b
l
e
ms
.
Proce
edi
ngs
of the
8t
h ra
instormi
ng
W
eek
o
n
Membra
ne Co
mputin
g, ISBN
978-
84-6
14-
23
57-6, 32
1-3
40.
[16]
GX Z
han
g, M
Gheorg
he, CZ
W
u
. A quantu
m-inspir
ed ev
o
l
utio
nar
y
al
gori
t
hm based o
n
p s
y
stems for
a class of com
b
in
atoria
l optim
izatio
n.
F
unda
me
nta Infor
m
at
icae
. 20
08; 87:
93–1
16.
[17]
T
O
Ishdorj, A Lepor
ati, L Pan,
X Z
e
n
g
, an
d
X
Z
hang. “D
eter
ministic so
luti
o
n
s to QSAT
and Q3SAT
by
spikin
g ne
ural
P
s
y
stems w
i
t
h
pre-com
p
uted
r
e
so
urce
s”.
T
heoretic
al
Co
mputer
Sc
ienc
e
. 2
010
;
411(
25): 23
45–
235
8.
[18]
M Couce
i
ro, P Ghamisi. Particle S
w
a
r
m
Optimization.
Spring
er Brie
fs in Appli
ed
Scienc
es an
d
T
e
chno
logy
. 2
015: 1-1
0
.
[19]
L
Pan and C
Mart
ı
n-Vi
de. “S
olvin
g
multi
d
im
ensi
ona
l 0-1 k
naps
ack pro
b
l
e
m b
y
P s
y
ste
m
s
w
i
t
h
in
pu
t
and activ
e
me
mbran
e
s”.
Jour
nal of Para
ll
el and D
i
stribute
d
Computi
n
g
. 20
05; 65(1
2
): 157
8– 15
84.
[20]
T
O
Ishdorj, A
L
epor
ati,
P
Lin
q
i
ang,
an
d J W
a
ng. “
So
lvin
g
N
P
-Co
m
pl
ete
pr
obl
e
m
s by
spik
ing
ne
ura
l
p
system
s with
budding r
u
les
”
.
in Proc
ee
din
g
s of th
e 1
0
th
Internati
o
n
a
l
Confer
ence
o
n
Membra
n
e
Comp
uting. 2
0
09: 335
–3
53.
[21]
L Pan,
X Z
e
n
g
,
X Z
h
a
ng,
a
n
d
Y Jia
ng. “Spi
king
neur
al P
s
y
stems
w
i
t
h
w
e
ig
hted s
y
n
a
pses”.
Ne
ura
l
Processi
ng Let
ters
. 2012; 35(
1): 13–2
7.
[22]
G P
ǎ
un, Roze
nber
g, A Salo
maa.
Membrane Com
p
uting
.
Oxford U
n
ivers
i
t
y
Press, Ne
w
York. 2010.
[23]
L Pan, T
O
Ish
dorj. P s
y
stem
s
w
i
t
h
activ
e
membra
nes a
nd se
parati
on
rules.
Jo
urna
l of
Univ
ersa
l
Co
mp
uter Scie
nce
. 200
4; 10(
5): 630-6
49.
[24]
R F
r
eu
nd, G P
ˇ
aun
an
d MJ
P
´
erez-Jim´e
nez
. T
i
ssue-like P
s
y
stems
w
i
th
c
han
nel-stat
e
s.
T
heoretic
al
Co
mp
uter Scie
nce
. 200
5; 330
(1): 101-1
16.
[25]
A
P
ǎ
un, G P
ǎ
un. T
he p
o
w
e
r
of comm
unic
a
tion:
P s
y
stem
s
w
i
th s
y
m
port/antip
ort.
New
Gen
C
o
mput
.
200
2; 20: 295
–
305.
[26]
J Kenne
d
y
, R
Eberhart. Particle s
w
arm o
p
timizati
on.
Internati
ona
l Con
f
erence o
n
N
eura
l
IEEE
Netw
orks
. 199
5: 1942
–1
94
8.
[27]
JJ Jami
an, M
W
Mustafa, H
Mokhlis,
MA B
ahar
udi
n. Impl
i
m
entatio
n of
E
v
oluti
onar
y Par
t
icle
S
w
a
r
m
Optimizatio
n
i
n
Distri
bute
d
Generati
on S
i
zing.
Inter
nati
ona
l Jo
urna
l
of Electric
al
a
nd C
o
mp
uter
Engi
neer
in
g (IJECE)
. 2012; 2(
1): 137-1
46.
[28]
G Z
hang,
F
Z
hou,
X H
uan
g
et a
l
. A
h
y
bri
d
mu
lti-s
w
arm
partic
l
e
optimi
z
ation
to s
o
lve
constra
i
n
e
d
optimiz
ation pr
obl
ems.
Journ
a
l of Univ
ersal
Co
mp
uter Scie
nce
. 200
5; 8(1
3
): 1821-
18
41.
[29]
CM
Ya
ng, D Simon.
A ne
w p
a
r
ti
cl
e
swa
r
m
o
p
t
imi
z
ati
on te
ch
ni
qu
e
. In:
Proce
edi
ngs
of the
18t
h
intern
ation
a
l co
nferenc
e on s
ystems engi
ne
e
r
ing (ISCEn
g’0
5
). 2005: 1
64–
169.
[30]
YT
Kao, E Z
a
hara a
nd IW
Kao. “A
h
y
brid
ized a
ppro
a
ch
to data cluste
ring”.
Expert S
ystem
s with
Appl
icatio
ns
. 2
008; 34(
3): 175
4–1
76
2.
[31]
U Maul
ik a
nd
S Band
yo
p
a
d
h
y
a
y
. Ge
net
ic al
gorithm base
d
clusteri
ng
tec
h
niq
ue.
Pattern Recognition
.
200
0; 33(9): 14
55-1
465.
Evaluation Warning : The document was created with Spire.PDF for Python.