TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.4, April 201
4, pp. 2833 ~ 2
8
4
2
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i4.4736
2833
Re
cei
v
ed Se
ptem
ber 3, 2013; Re
vi
sed
Octob
e
r 25, 2
013; Accepte
d
No
vem
ber
20, 2013
A P2P Traffic Identification Approach Based on
SVM
and BFA
Chun
zhi Wa
ng, Zeqi Wa
ng, Zhi
w
e
i Y
e
*, Hong
w
e
i
Chen
Hub
e
i Un
iversit
y
of T
e
chnolo
g
y
/ Schoo
l of Computer Sci
e
n
c
e, W
uhan, Ch
ina
*Corres
p
o
ndi
n
g
author, e-ma
i
l
:
w
e
izhi
ye
1
2
1
@
16
3.com
A
b
st
r
a
ct
Now
adays n
e
w
peer to peer
(P2P) traffic
with dyna
mic po
rt and encrypte
d
techno
lo
gy mak
e
s th
e
ide
n
tificatio
n
o
f
P2P traffic b
e
co
me
mor
e
a
nd more
d
i
fficult. As one of the opti
m
al cl
assifiers, sup
p
o
rt
vector mach
in
e
(SVM)
h
a
s speci
a
l adva
n
t
ages
w
i
th
avo
i
din
g
l
o
ca
l o
p
timu
m, overc
o
mi
ng
di
me
nsi
o
n
disaster, res
o
lv
ing s
m
a
ll s
a
mp
les a
nd h
i
gh
di
me
nsi
on for P2
P classificati
on
probl
e
m
s. Ho
w
e
ver, to empl
o
y
SVM, the para
m
eters s
e
lecti
o
n of SVM shou
ld be c
onsi
der
ed an
d thus so
me
opti
m
i
z
at
io
n metho
d
s hav
e
bee
n
put forw
ard to
d
eal
w
i
th it, stil
l, it
is
not ful
l
y s
o
lve
d
. He
nce,
in
the
pa
per,
a p
eer to
p
eer
traffi
c
ide
n
tificatio
n
a
ppro
a
ch
bas
ed
on s
u
p
port ve
ctor machi
ne
a
nd b
a
cteri
a
l for
agi
ng
alg
o
rith
m is
pro
pos
ed
fo
r
better id
entific
ation
of P2P n
e
tw
ork traffic.
F
i
rst,
the best para
m
et
ers for SVM are tun
ed w
i
th bacteri
a
l
foragi
ng
alg
o
rit
h
m. S
ubs
equ
e
n
tly, SVM set
w
i
th the best
pa
ram
e
te
rs
i
s
use
d
to
i
d
en
tify
P2P traffic. Finally,
exper
imenta
l
r
e
sults sh
ow
the pro
pose
d
a
p
p
roac
h can
effectively i
m
pr
ov
e the acc
u
racy
of P2P netw
o
rk
traffic identifica
t
ion.
Ke
y
w
ords
:
P2
P traffic identifi
c
ation, bact
e
ria
l
fo
ragi
ng al
gor
ithm, su
pport v
e
ctor mac
h
i
n
e
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
Peer-to
-
p
eer (ab
b
reviated to P2P)
comp
ut
er ne
twork i
s
a
distrib
u
ted
a
pplication
architectu
re
that pa
rtition
s
tasks o
r
work
l
oad
s am
ong pee
rs.
Peers
a
r
e
e
qually
p
r
ivile
ged
partici
pant
s in the appli
c
at
ion. Each co
mputer in
the
netwo
rk i
s
ref
e
rred to as a
node. The o
w
ner
of each com
puter on a P2P network
woul
d set as
ide a portion
of its resou
r
ce
s - such as
pro
c
e
ssi
ng p
o
we
r, disk
storag
e o
r
net
work b
and
wi
dth -to be
m
ade di
re
ctly
available to
other
netwo
rk
parti
cipa
nt, witho
u
t the need f
o
r central
co
ordin
a
tion by
serve
r
s
or st
able ho
sts [1]
.
In
P2P networks, client
s pro
v
ide resou
r
ces, wh
i
c
h m
a
y include b
and
width, sto
r
age
spa
c
e,
and
comp
uting
p
o
we
r [1-3]. T
h
is p
r
o
perty i
s
on
e of the
major
advant
age
s of u
s
in
g P2P net
wo
rks
becau
se it make
s the set
up and ru
nni
ng co
sts very
small for the origi
nal content distri
b
u
tor.
Another characteri
stic of a
P2P net
work is it
s
cap
abili
ty in term
s of
fault-toleran
c
e. Wh
en
a p
e
e
r
goe
s do
wn o
r
is disco
nne
ct
ed from the n
e
twork,
the P
2
P appli
c
atio
n will continu
e
by usin
g other
peers. With t
he wid
e
spre
a
d
use
of peer-to-p
eer
(P
2P
) tech
nolo
g
ie
s, it has o
c
cu
pied the maj
o
rity
of the total Internet traffic
in applications
, su
ch a
s
communi
catio
n
, ente
r
tainm
ent an
d
sha
r
i
ng,
for the advant
age
s of conv
enien
ce, hig
h
-
sp
eed, ri
ch reso
urce
s and
no-cente
r
.
Ho
wever, it b
r
ing
s
challen
ges fo
r conte
n
t
supe
rvisi
o
n and a lot of
probl
ems
ari
s
e at the
same
time. Firstly,
the
decentrali
zati
on
a
nd
ano
nymity of P2P streami
n
g net
wo
rk
made
informatio
nal
resou
r
ce more disp
ersion
and con
c
eal
ment; thus so
me violent an
d blue movie
s
or
video we
re transmitted
arbitrarily, whi
c
h bro
ught ba
d effects to p
eople, juvenil
e
s e
s
pe
cially;
in
addition, de
centrali
zed
net
works
int
r
od
u
c
e ne
w se
curity issu
es be
cau
s
e
they a
r
e d
e
si
gne
d
so
that ea
ch
user i
s
re
spon
sible for controlling t
hei
r
d
a
ta an
d reso
urces. S
e
con
d
ly, harmful
data
can
also be
d
i
stribute
d
o
n
P2P netwo
rks by mo
dify
ing files th
at are already b
e
i
ng di
strib
u
ted
on
the network.
Con
s
e
quently
, it is of grea
t importan
c
e
for a
c
curate
i
dentificatio
n
of traffic that
is
gene
rated by
P2P applications [4-5].
Traditio
nally, netwo
rk traffic ca
n be ea
sily
identified by detecti
ng
the port num
bers of
that traffic
,
as
mos
t
of the traffic
in the
Inte
rnet u
s
e
s
stan
dard p
o
rt numbe
rs [6]
.
Yet, as m
a
ny
newly-eme
rg
ed P2P a
ppli
c
ation
s
u
s
in
g
dynami
c
po
rt
numbe
rs, m
a
sq
ueradin
g
techni
que
s,
a
n
d
payload
en
cryption to avo
i
d dete
c
tion,
the cl
a
s
sical
approa
che
s
based o
n
p
o
r
t mappi
ng
a
n
d
payload
an
alysis
a
r
e
ineff
e
ctive. Th
us the ta
sk of
i
dentifying P
2
P traffic i
s
becoming
m
o
re
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: 2833 – 2
842
2834
chall
engin
g
a
nd a l
o
t of work ha
s
been
done
to
effe
ctively identify P2P data t
r
affic, which i
s
mainly divide
d into four cat
egori
e
s.
(1) Port num
ber ba
se
d appro
a
che
s
, whi
c
h we
re the first tech
nique
s to de
tect P2P
traffic. In the early eme
r
ge
nce of P2P, the por
t
-
ba
se
d method is
succe
ssful b
e
cause many well-
kno
w
n ap
plications have
spe
c
ific po
rt
numbers
(a
ssi
gne
d by IANA). Ho
we
ver, in order to
circumve
nt d
e
tection, fixe
d po
rt nu
mb
ers a
r
e
unre
liable. Th
erefore, m
o
re
a
nd mo
re
P2
P
appli
c
ation
s
n
e
ver use po
rt-ba
s
ed m
e
tho
d
s any mo
re [7].
(2) De
ep Pa
cket Inspe
c
tion (DPI) m
e
thod, wh
ich tries to j
udg
e
wheth
e
r it
ha
s fou
n
d
kno
w
n P2P
chara
c
te
risti
c
s throug
h byte
-by-byte
sc
an of mess
age
c
ontents
for
data traffics
. DP
I
techn
o
logie
s
have a high d
egre
e
of exactness and the
y
can judg
e
specifi
c
appli
c
ation types, b
u
t
they execute
in a lo
w sp
ee
d, can
do n
o
thing to
e
n
crypted data
an
d ne
w P2P a
pplication
s
wi
th
unkno
wn ch
a
r
acte
ri
stics, and have a high mainta
in
cost. Beca
u
s
e mo
st P2P protocol
s
are
prop
rieta
r
y a
nd
reverse
en
ginee
ring
is n
eede
d, an
d th
e meth
od i
s
n
o
t able
to
han
dle
with
bra
n
d
-
new ap
plications that
use
un
kn
o
w
n
P
2
P protocols [8]. With
th
e devel
opme
n
t of a
n
ti-ide
ntify
techn
o
logie
s
for P2P software, sin
g
le u
s
e of
DPI tech
nologi
es may
not satisfy the requi
rem
e
n
t
.
(3)
Deep Traffic
Ins
p
ec
tion (DFI) method, wh
i
c
h trie
s to judg
e whether it ha
s
satisfie
d
P2P traffic feature
s
thro
ugh stati
s
tical analys
i
s
of data traffic [9]. Com
pare
d
with
DPI
techn
o
logie
s
,
DFI exe
c
ut
e in a hi
gh
spe
ed,
are e
ffective to encrypte
d
dat
a and
ne
w
P2P
appli
c
ation
s
with u
n
kno
w
n cha
r
a
c
teri
stics, a
n
d
hav
e a
rel
a
tively low mai
n
tai
n
cost,
but t
heir
exactne
s
s i
s
lower than
that of
DPI,
and
ca
nnot
j
udge
spe
c
ific appli
c
atio
n t
y
pes.
Cu
rre
n
t
ly,
studie
s
an
d p
r
odu
cts b
a
se
d on DPI tech
nologi
es a
r
e
comp
arably more.
(4) Ma
chin
e l
earni
ng
ba
se
d identificatio
n. As
a
matt
er of fa
ct, to i
dentify P2P traffic is a
probl
em bel
o
ngs to
pattern re
cog
n
ition
,
natura
lly all
kind
s of cl
a
ssifi
cation m
e
thod
could
be
applie
d to P
2
P traffic id
e
n
tification p
r
oblem,
su
ch
as Baye
sia
n
deci
s
io
n, C4.
5
de
ci
sion t
r
ee,
Neu
r
al
Net
w
o
r
k
etc [1
0-1
2
]. The
s
e m
e
tho
d
s
use trai
nin
g
data to
e
s
ta
blish
a
cla
ssif
i
cation
mod
e
l,
whi
c
h is u
s
e
d
to generate a cla
ssifie
r
to
classi
fy the u
n
kn
own data
set. Ho
weve
r, many machi
n
e
learni
ng m
e
thod
s have
their in
heren
t wea
k
ne
sse
s
. Especi
ally in this
real-time traffic
identificatio
n, the accuracy
and efficie
n
cy
of
the methods are not we
ll meeting the
demand.
Suppo
rt Vect
or Ma
chin
e (SVM) is one
of
the most promi
s
ing a
nd po
werful
machi
n
e
learni
ng
met
hod
s
for cla
ssifi
cation a
nd
re
gressio
n
pro
b
lem
s
of
small sa
mples and high
dimen
s
ion
s
. It was initially
pre
s
ente
d
by
Vapnik i
n
the
last de
cad
e
of the 20th century ba
se
d
on
statistical lea
r
ning the
o
ry a
nd structu
r
al
risk
minimi
zat
i
on pri
n
ci
ple [
13]. It has be
en proved to
be
very su
ccessf
ul in many ap
plicatio
ns
su
ch as
ha
nd
writ
ten digit re
co
gnition, imag
e cla
ssifi
catio
n
,
face detection, object det
ection, text classifica
tion [14]. Since SVM has an excellent abilit
y to
solve
bina
ry cla
s
sificatio
n
problem
s
and th
e ma
i
n
pu
rpo
s
e
o
f
P2P traffic identificatio
n is
accurately cl
a
ssifying
two
classe
s: P2P
and
non
-P
2P
traffic, the
r
e
are
so
me
stu
d
ies which h
a
v
e
prop
osed P2
P traffic identification ba
sed o
n
SVM
[15]. While it is noteworthy that
the
para
m
eters h
a
ve influen
ce
on the g
ene
ralizatio
n
pe
rforma
nce of SVM, some
work
ha
s do
ne
to
deal with the
para
m
eter p
r
oblem
s of SVM [16-17].
Ho
wever, the
problem h
a
s not been fully re
solved.
The Bacte
r
ial
Foragi
ng Algorithm
(abb
reviate
d
to BG) is
a re
cently devel
o
ped
sw
arm in
telligen
ce alg
o
rithm b
a
sed
on the foragi
ng
behavio
r of E. coli bacte
ria. It is a
simple
stoch
a
stic gl
obal
optimizatio
n techni
que. T
he
optimizatio
n probl
em search spa
c
e cou
l
d be model
e
d
as a soci
al
foraging e
n
vironm
ent wh
e
r
e
grou
ps of pa
rameters com
m
unicate co
o
peratively
for
finding soluti
ons to difficul
t
problem
s. T
h
is
idea
co
uld b
e
ap
plied to
optimal p
a
ra
meters of SV
M. So we
ca
n u
s
e BG
alg
o
rithm to
find
the
optimal pa
ra
meters of SVM for the iden
tification of P2P traffic.
The re
st of the pape
r is organi
zed a
s
follows.
In secti
on 2 we sim
p
ly describe th
e basi
c
prin
ciple of suppo
rt vector machin
e. The idea of
b
a
cteri
a
l forag
i
ng algo
rithm
is explained
in
se
ction
3. Ho
w to
empl
oy
BG alg
o
rithm
to find
the
o
p
timal pa
ram
e
ters of SVM
is ill
ust
r
ated
in
se
ction 4. In
se
ction 5, th
e perfo
rma
n
ce of
pro
p
o
s
e
d
app
roa
c
h i
s
evalu
a
ted
on the real P
2
P
traffic data a
nd is
comp
a
r
ed
with existing techni
qu
es. Some
co
nclu
sio
n
s a
r
e
also p
r
ovide
d
towards
the concl
uding se
ction.
2. The Basic
Principle of Support Vec
t
or Mac
h
ine
Suppo
rt Vect
or Ma
chi
ne i
s
a
cla
s
sification an
d re
gre
ssi
on
pre
d
iction to
ol t
hat uses
machi
ne lea
r
ning theo
ry to maximize p
r
edictive a
c
curacy whil
e aut
omatically av
oiding ove
r
fitting
to the data,
whi
c
h i
s
an
active pa
rt
of t
he ma
chi
ne lea
r
nin
g
resea
r
ch aro
und the
wo
rl
d
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A P2P Traffic Identification
Approa
ch Ba
sed o
n
SVM and BFA (Ch
unzhi Wa
ng)
2835
Cla
ssifying
d
a
ta is a
co
m
m
on ta
sk i
n
machi
ne le
arning. A supp
ort vecto
r
ma
chin
e con
s
tru
c
ts a
hyperpl
ane
o
r
set of hype
rplane
s in
a hi
gh o
r
infinite
dimen
s
ion
a
l
spa
c
e,
whi
c
h
can
be
used
for
cla
ssifi
cation
and othe
r tasks.
Suppo
sed
th
at so
me
given data
poi
nts
each b
e
lo
n
g
t
o
on
e of t
w
o
cla
s
ses,
and
the go
al
is
to decide whi
c
h class
a
new sampl
e
point will
be i
n
. There
are
many
hyperpl
anes that mi
ght
cla
ssify the
data. Intuitively, a go
od
sep
a
ratio
n
i
s
achieved
by the hyp
e
rpl
ane th
at ha
s the
large
s
t di
stan
ce to the n
e
a
r
est training
data point
of
any cla
s
s, si
nce in
gen
eral the larger t
he
margi
n
the lower the g
ene
ralizatio
n error of the cl
assifier. If such a h
y
perpla
ne exi
s
ts, it is kno
w
n
as the maxim
u
m-ma
rgi
n
h
y
perpla
ne an
d the linear
cl
assifier it defines i
s
kn
own
as a maximu
m
margi
n
classifier; or equival
ently,
the perception of optimal stability.
Let us have
a data
set
,,
1
,
,
ii
x
yi
l
of example
s
whe
r
e
1,
1
i
y
and
d
i
x
R
whe
r
e
i
x
is an a
r
bitra
r
y
d
a
ta point
a
n
d
i
y
its
corre
s
p
ondin
g
bip
o
la
r la
b
e
l. Let
us al
so defin
e
a
linear de
ci
sio
n
surfa
c
e
by
the eq
uation
0
fx
x
b
.The o
r
igin
al
formulatio
n o
f
the SVM
algorith
m
se
eks a line
a
r
deci
s
io
n su
rf
ace
whi
c
h m
a
ximize
s the
margin
bet
wee
n
the cl
o
s
e
s
t
positive and.
negative exa
m
ples. Thi
s
may be achi
e
v
ed throug
h the minimization of the pen
alty
term
2
||
|
|
/2
. It yields
ii
i
i
y
x
with the co
nst
r
aint
0
ii
i
y
, and
0,
1
,
,
i
Ci
l
whe
r
e
C i
s
a re
gula
r
i
z
ation vari
abl
e call
ed trad
e-off pa
ram
e
ter or
Penalty factor. The para
m
eters
i
can
be found af
ter the following q
uad
rat
i
c optimizatio
n
probl
em is m
a
ximized:
1
2
D
ii
j
i
j
i
j
i
Ly
y
x
x
(1)
The data ex
ample
s
wh
ose corre
s
p
ond
ing
i
values are not ze
ro a
r
e call
ed sup
port
v
e
ct
or
s.
Instead of co
nsid
erin
g the
input spa
c
e,
we may con
s
ide
r
a given
augmente
d
spa
c
e by
repla
c
in
g the
inne
r
pro
d
u
c
t of Equatio
n
(1
) by
the
d
o
t produ
ct
,
ij
i
j
Kxx
x
x
whic
h
yields:
,
1
,
2
D
ii
j
i
j
i
j
ii
j
L
yy
K
x
x
(
2
)
Whe
r
e fu
ncti
on
,
K
xy
x
y
rep
r
e
s
e
n
t
s semi-defini
t
e ke
rnel. S
o
me of th
e
cl
assical
SVM kern
els
are Lin
e
a
r
Kernel, Polynom
ial Kernel a
n
d
RBF Kernel.
The
re
sultin
g de
ci
sion
frontie
r i
s
,
ii
i
i
f
xy
K
x
x
b
where
,
ii
x
an
d
b
rep
r
e
s
ent re
spectively
the
ith
su
ppo
rt vect
or, its corre
s
pondi
ng m
u
ltiplier
and
the
hyperplan
e
bias. Fo
r nonsepa
ra
ble cla
s
se
s, the L2-SVM variant minimize
s the fun
c
tion
22
1
1/
|
|
|
|
/
2
2
i
i
C
which lead
s to the maximization of:
,
1
,
2
Di
i
j
i
j
i
j
ii
j
L
yy
K
x
x
(
3
)
With,
1
,,
Kx
x
K
x
x
ij
ij
C
i
f
ij
,
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: 2833 – 2
842
2836
,,
Kx
x
K
x
x
ij
ij
if
ij
,
Und
e
r the co
nstrai
nts
0
ii
i
y
and
0,
1
,
,
i
il
.
The effe
ctiveness of SVM
depe
nd
s o
n
the sel
e
ct
io
n
of ke
rnel, the
ke
rnel'
s
para
m
eters,
and soft marg
in para
m
eter.
The mo
st co
mmon kern
el function
s u
s
e
d
in SVM are as follo
ws:
Linea
r ke
rn
el function:
(,
)
T
K
xx
x
x
ij
i
j
(4)
Polynomial kernel fun
c
tion
:
(,
)
(
1
)
T
Kx
x
x
x
ij
i
j
(5)
RBF ke
rnel fu
nction:
2
||
||
(
,
)
e
xp(
)
2
xx
ij
Kx
x
ij
(
6
)
The paramet
ers
in
kernel
function
reflect the
ch
ar
act
e
risti
c
of train
i
ng d
a
ta, an
d
grea
t
effect on the perfo
rman
ce
of the SVM.
Penalty factor
C
determine
s the trade-off co
st betwee
n
minimizi
ng th
e traini
ng
error a
nd
minim
i
zing
the
m
o
del’s comple
xity. Whether the valu
e i
s
too
big o
r
small
can redu
ce
th
e gen
eralization of SVM
. I
n
re
al a
pplica
t
ions, mo
st of
paramete
r
s
are
sele
cted
emp
i
rically by tryi
ng a finite
nu
mber
of
pa
ra
meter valu
es and
sele
ctin
g those that
get
the lea
s
t te
st
erro
r. Exce
p
t
for
con
s
u
m
i
ng e
n
o
r
mou
s
time, such trial an
d e
r
ror
pro
c
ed
ures for
sele
cting
the
para
m
eters o
f
SVM often f
a
il to o
b
tain
the b
e
st
pe
rfo
r
man
c
e
a
s
it i
s
im
pre
c
i
s
e
a
nd
the re
sult is u
n
relia
ble. In p
r
acti
ce, g
r
id search
i
s
a rat
her effici
ent for go
od pa
ra
meters of SVM,
however, it is only suitable
for adju
s
tme
n
t of ve
ry few param
eters
and do
es n
o
t perform well
in
pra
c
tice
be
ca
use
it is com
p
lex in
com
p
utation a
nd ti
me
con
s
umi
n
g. It is
a vital
step to
optimi
z
e
the paramete
r
s
of SVM for a goo
d pe
rfo
r
man
c
e i
n
ha
ndling
a lea
r
n
i
ng task a
s
th
e perfo
rma
n
ce
of SVM will b
e
weakene
d i
f
these
pa
ra
meters
a
r
e
n
o
t pro
perly
chosen. Th
e p
aper propo
se
s a
BG algorithm to optimiz
e
parameters
C
and
autom
atically, the
prin
ciple
of
BG will
b
e
illustrated in the next section.
3. Bacterial Foraging Al
gorithm
Re
cently, re
sea
r
che
s
of optimal
foraging
of b
a
c
teria
h
a
ve
been
u
s
ed
for
solvin
g
optimizatio
n probl
em
s.
Th
e
fora
ging be
havior of
Escheri
c
hia coli, whi
c
h
i
s
a common
type
o
f
bacte
ria, i
s
consi
dered in t
h
e
research [
18]. The E. coli ba
cteria
th
at are
pre
s
e
n
t
in the inte
stines
have a fora
ging st
rateg
y
governe
d
by four
proce
s
se
s, na
mely, chemo
t
axis, swa
r
m
i
ng,
rep
r
od
uctio
n
, and elimin
ation and di
sp
ersal [19].
Chem
otaxis:
This p
r
o
c
e
ss
achi
eves th
ro
ugh swimmi
n
g
and tumbli
n
g
as
sho
w
n i
n
Figure
1. Dep
endin
g
upon the
rot
a
tion of the flagella in
ea
ch ba
cteriu
m, it decid
es
wh
ether it shoul
d
move in
a p
r
edefine
d
di
re
ction
(swimm
ing)
or th
e b
a
cteri
u
m. To
rep
r
e
s
ent
a
tumble, a
u
n
it
length ran
d
o
m
directio
n,
j
say, is generated; th
is
will be used to
def
ine the di
rection
of
movement after a tumble. In particular,
1,
,
,
,
ii
j
kl
j
k
l
C
i
j
(7)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A P2P Traffic Identification
Approa
ch Ba
sed o
n
SVM and BFA (Ch
unzhi Wa
ng)
2837
Whe
r
e
,,
i
j
kl
rep
r
e
s
ents the
i
th bacte
rium
at
j
th ch
emota
c
tic
k
th reprodu
ctive,
and
l
th elimination and
dispersal
step.
Ci
is the si
ze
of the step t
a
ke
n in the
rand
om
dire
ction spe
c
ified by the tumble. “C” is
termed a
s
the
“run le
ngth u
n
it”.
Figure 1. Swimming an
d Tumbling of E.coli
Fi
gure 2. Flowchart of Bacterial Foragin
g
Algorithm
It is al
ways
desi
r
ed
that
the ba
cte
r
iu
m that
h
a
s searche
d
the
optimum
pat
h of foo
d
sho
u
ld try to attract othe
r bacte
ria so that t
hey reach
the desired
place
more rapidly. Swarming
make
s the
b
a
cteri
a
cong
regate into g
r
oup
s and
he
nce m
o
ve a
s
con
c
ent
ric
p
a
tterns
of gro
ups
with high b
a
ct
erial de
nsity. Mathemati
c
al
ly, swarmi
ng
can b
e
rep
r
e
s
ented by:
1
,,
,
S
ii
CC
C
C
i
JJ
j
k
l
(8)
2
11
ex
p
p
S
i
attra
c
t
a
ttra
c
t
m
m
im
d
2
11
ex
p
p
S
i
r
e
p
e
ll
en
t
r
ep
e
l
l
e
n
t
m
m
im
h
W
h
er
e
,,
,
CC
JP
j
k
l
is th
e cost fu
ncti
on valu
e to
be a
dde
d to
the a
c
tual
co
st
function
to b
e
minimi
zed
to pr
es
en
t a time
va
r
y
in
g c
o
s
t
fu
nc
ti
on. “S” is
the total number of
bacte
ria.
“p”
is the
num
be
r pa
ram
e
ters to be
optim
i
z
ed
that a
r
e
pre
s
e
n
t in e
a
ch
ba
cteri
u
m.
,,
at
t
r
ac
t
a
t
t
r
a
c
t
r
e
pe
l
e
nt
dh
and
re
pelent
are different
coefficient
s that are to be chosen
judici
ou
sly.
Rep
r
od
uctio
n
:
The least he
althy bacteri
a
die,
and the other he
althie
st bacte
ria ea
ch split
into two
ba
ct
eria,
whi
c
h
are pla
c
e
d
in
th
e same
lo
cati
on. Thi
s
m
a
kes th
e p
opul
a
t
ion of b
a
cte
r
i
a
con
s
tant.
Elimination a
nd Di
sp
ersal:
It is po
ssi
bl
e that
in the local
environment, the li
fe of a
popul
ation of bacte
ria
chan
ges eith
er g
r
adually by
co
nsum
ption of
nutrient
s or
sudde
nly due to
some other i
n
fluence. Ev
ents
can kill
or di
sperse
all the ba
cteri
a
in a region. They have
the
effect of possibly destroyin
g the che
m
ot
actis p
r
og
re
ss, but in cont
rast, they also assist it, since
disp
ersal ma
y place ba
ct
eria ne
ar g
o
od food
sou
r
ce
s. Elimina
t
ion and di
spersal hel
ps in
redu
cin
g
the
beh
avior
of stag
nation
(being t
r
ap
pe
d in a
p
r
em
ature
sol
u
tio
n
point
or lo
cal
optima).
The
pro
c
e
s
s
of ba
cteria
foraging
alg
o
rith
m
for
solvin
g
the optimi
z
ati
on p
r
obl
em i
n
clud
es:
(1) En
co
de t
he sol
u
tion
s
for the probl
em; (2)
De
si
gn evaluatio
n
function; (3) Gene
rate ini
t
ia
l
solutio
n
pop
u
l
ation; (4) O
p
timize pa
ram
e
te
rs by u
s
ing
the intera
ctio
n betwe
en group
s.
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: 2833 – 2
842
2838
Suppo
sed
Nc, Nre a
nd
Ne
d den
ote ma
ximum
ch
em
otaxis, maximum re
produ
ction, an
d
maximum eli
m
ination
-
di
sp
ersal respe
c
tively.
The optimizatio
n
pro
c
ed
ure of
the ba
cteri
a
l
foragin
g
algo
rithm is descri
bed bri
e
fly as following Fig
u
re 2.
4. SVM Param
e
ters O
p
ti
m
i
zation B
a
sed on B
G
4.1. The principle of opti
m
al parameters of SVM
The
perfo
rm
ance of
SVM mainly
ref
e
rred to
the
gene
rali
zatio
n
ability. As
stated i
n
se
ction 2, pe
nalty factor
C and
kernel f
unctio
n
pa
ra
meters
exert a co
nsid
erab
le influen
ce o
n
the gene
rali
zation ability of SVM. The kernel fun
c
tion
paramete
r
s determi
ne the mappin
g
of the
origin
al spa
c
e to high
dim
ensi
onal
sp
a
c
e a
nd the
value of p
enal
ty factor can
adju
s
t the e
r
ror
and
compl
e
xity. As the value of ea
ch pa
ramete
r,
too big o
r
too sm
all, al
l hampe
rs the
gene
rali
zatio
n
of SVM,
the optimi
z
at
ion of
para
m
eters i
s
i
m
porta
nt to
achieving
good
gene
rali
zatio
n
ability in p
r
acti
ce. Thi
s
study p
r
op
o
s
e
s
to em
pl
oy BG algo
ri
thm to optim
ize
para
m
eters C
and
automatically. T
he main ide
a
of applyin
g
BG to se
arch the be
st
para
m
eters p
a
ir ( C a
n
d
) of SVM is as b
e
low.
Each p
o
sition
vector of the
bacte
ria sta
nds
a
can
d
id
ate paramete
r
s p
a
ir for SV
M. The
initial populat
ion is gen
era
t
ed with
N
number of sol
u
tions an
d ea
ch solution i
s
a D-dim
e
n
s
io
n
vector, he
re
D is set to
2 that each solu
tio
n
repre
s
e
n
ts 2
-
D ca
ndid
a
te
param
eters.
i
X
rep
r
e
s
ent
s the i-th ba
cteria
position in th
e popul
ation
whi
c
h de
note
s
a ca
ndid
a
te
param
eter p
a
ir
and
it
s
fitne
s
s can
b
e
me
asu
r
ed by fitness functi
o
n
,
with defined movement rule
s, the virtual
bacte
rium m
o
ves in the
search pl
ace and up
date
t
heir p
o
sitio
n
with predefin
ed rul
e
s, till the
virtual bacteri
u
m meets their e
nd
condi
tion, the al
gorithm
will
terminate and
output the best
positio
n as th
e optimal pa
rameters for S
V
M.
4.2. The Implementa
tion
of the Prop
o
sed Me
thod
The RBF
ke
rnel fun
c
tion i
s
taken a
s
th
e ke
rnel fu
nction. The pa
rameters n
e
e
d
s to be
optimize
d
are
C
and
. The basi
c
ste
p
s a
r
e state
d
as f
o
llows:
Step 1: Read
data S from
file, then divide it into two gro
u
p
s
S1 (traini
ng set)
and S2
(testing
set).
Suppo
sed
th
e si
ze
of ba
ct
erium
as
N,
a
nd rand
omly
gene
rated
N
grou
ps set of
{C
,
} to initial loca
tion of the position of ba
cterium;
Step 2: Desi
gn fitness ev
aluation fun
c
tion
(,
)
fit
ne
ss
f
C
. Accordi
ng to the value of
C
an
d
, train the SVM mod
e
l with g
r
ou
p
S1, then co
n
s
ide
r
op
po
site value of te
sting accu
ra
cy
with gro
up S2
as the fitness.
Step 3: Execute the loop o
f
chemotaxis,
reprodu
ction and
elimin
ation-di
sp
ersal.
Step 4: En
co
de, defin
e the
po
sition
of b
a
cteri
u
m
with bes
t
fitness
as
the optimal C*or
*.
The procedu
re for de
scribi
ng pro
p
o
s
ed
BG-SVM is a
s
follows:
Step 1: Initialize B
G
wit
h
pop
ulation
size. Set the nu
mbe
r
of bacte
rium
and it
s
dimen
s
ion
an
d othe
r (l
=0,
k=0, j=0). Evaluate the
fitness value
of
each ba
cte
r
ia
. Take
the
cross
validation error of the SVM training set as fitness valu
e.
Step.2 Elimination-di
sp
ersal loop: l=l
+
1
Step.3 Rep
r
o
ductio
n
loop:
k=k+1
Step.4 Chem
otaxis loop: j
=
j+1
Step.5: Executive chemot
axis ope
ratio
n
Step.6 If j<
Nc, turn to s
t
ep 4
Step.7 Execu
t
ive repro
d
u
c
tion ope
ration
Step.8 If k
<
Nre, turn to s
t
ep 3
Step.9 Execu
t
ive elimination-di
spe
r
sal o
peratio
n
Step.10 If l<
Ned, turn to s
t
ep 2, els
e
finish
Step.11 Rep
eat the
step
4-7
until a
value of
the
fitness fun
c
tion conve
r
ge
s o
r
the
numbe
r of iteration rea
c
he
d.
After conve
r
g
i
ng, the globa
l best obje
c
t is fed in to SVM classifie
r
for testing.
5. Experiment and Dis
c
u
ssion
To test the e
ffectiveness
of the pro
p
o
s
ed me
tho
d
, some real ca
mpus P
2
P traffic data
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A P2P Traffic Identification
Approa
ch Ba
sed o
n
SVM and BFA (Ch
unzhi Wa
ng)
2839
are u
s
e
d
to evaluate the
performan
ce of the pro
posed BG-S
VM model for the pa
ra
meter
optimizatio
n
of SVM an
d t
r
affic i
dentifat
ion of
P2
P; m
o
reove
r
, the
perfo
rman
ce
i
s
com
pared
with
some
comm
o
n
ly used algo
rithms p
r
e
s
en
ted in
the lite
r
ature, which
are GA, PSO and GA-P
SO
algorith
m
. The data of
P2P netwo
rk traffic is
co
llected from
netwo
rk l
abo
ratory of Hu
bei
Univ
er
sit
y
of
Te
chn
o
logy
.
We
colle
ct
e
d
mo
re
than
300
sampl
e
data
an
d
ch
ose
11
featu
r
es
whi
c
h sho
w
e
d
in Table 1 for ou
r experi
m
ents.
Table 1. The
De
scription of
Feature
s
Feat
ures
Explana
t
io
n
Duration
The du
ration of
traffic
T
y
pe
IP or IP-Port
TCP-I
O
The ratio of sen
d
i
ng and receiving TCP packets
UDP-I
O
The ratio of sen
d
i
ng and receiving UDP packets
All-IO
The ratio of sen
d
i
ng and receiving all packets
Avg-speed
The averag
e spe
ed of traffic
Avg-packets
The averag
e pac
kets size of traffic
Avg-TCP/UDP
The b
y
te ratio of
TCP and U
D
P av
erage packets size
TCP/UDP
The b
y
te ratio of
TCP and U
D
P tr
affic
TCP-p
r
o
The pro
portion of
TCP in traffic
UDP-pr
o
The pro
portion of
UDP in traffic
The
colle
cted
data
of n
e
twork traffic i
s
divided i
n
to training
set an
d testin
g
set,
and
put
into SVM to
classify for P2
P identificatio
n. We
cho
s
e
nine-te
nths fo
r trai
ning
dat
a an
d o
ne-te
nth
for testing d
a
t
a to get better expe
rime
n
t
al
results. T
he ra
nge
d of para
m
eters
C and
is form
10
2
to
10
2
on the d
a
taset. Th
e
main pa
ram
e
ters
used for these
app
ro
ach
e
s
are: t
he initial
popul
ation fo
r thre
e algo
rithms i
s
the same, that is
20, and all th
ese al
go
rithm
s
will te
rmina
t
e
after bein
g
ex
ecute
d
100 ti
mes. Mo
reov
er, the cro
s
so
ver rate fo
r G
A
is 0.4 an
d mutation rate for
GA is 0.0
1
. And for PSO,
C1
=2.0 a
nd
C2
=2.0, the
value of ine
r
tia wei
ght is
set as 1. F
o
r
BG,
Ped=0.2
5
.
Figure 3-Fi
gu
re 5 show av
erag
e cla
s
sification
a
c
cura
cy and be
st classificatio
n
a
c
cura
cy
respe
c
tively by three app
ro
ach
e
s: GA-S
VM, PSO-SVM and BG-S
VM.
Figure 3. Cla
ssifi
cation by
GA Algorithm
(100,
20)
Figure 4. Cla
ssifi
cation by
PSO Algorith
m
(100, 20
)
Figure.5. Cla
ssifi
cation by
BG Algorithm
(200,50
)
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: 2833 – 2
842
2840
From expe
ri
ment one, the be
st cla
s
sifi
catio
n
accura
cy and a
v
erage
cla
s
sification
accuracy
obt
ained
by the
prop
osed B
G
-SVM alg
o
rith
m are hig
h
e
r
comp
ared
with the
GA-SV
M
approa
ch an
d PSO-SVM approa
ch, wh
ich are illustra
ted in Figs.
3
-5: the be
st cro
ss valid
ation
accuracy
(BCVA) of GA i
s
92.
306
9% an
d the
avera
g
e
cro
s
s valid
ation a
c
curacy (ACVA) of
GA
is a ra
nge from 90.5% to
91.5%; the BCVA
of PSO is 89.7
8
8
7
% and the
ACVA of PSO
fluctuate
s
m
a
rke
d
ly
a
bove and belo
w
87
%;
while
the
BCVA of BG
is 9
3
.662%
a
nd the
ACVA
of
BG fluctuate
s
na
rro
wly b
e
twee
n 89% to 91%. Mo
reover, a
s
we
can see
in
Figure.3, the GA
method
stop
s evolution a
fter 20
th
gen
eration. So
we chan
ged
popul
ation si
ze from 2
0
to 50,
maximum ge
neratio
n from
100 to 200, a
nd sh
owed re
sults a
s
follo
wing
F
i
gu
re
.6-
F
ig
ur
e
8
.
Figure 6. Cla
ssifi
cation by
GA Algorithm
(100,
20)
Figure 7. Cla
ssifi
cation by
PSO Algorith
m
(200, 50
)
Figure 8. Cla
ssifi
cation by
BG Algorithm
(200, 50
)
From
expe
ri
ment two, the b
e
st
cla
s
sifica
tio
n
a
ccura
cy a
nd
a
v
erage
cla
s
sification
accuracy o
b
tained by the
prop
osed BG
algorithm
are also
better
than PSO al
gorithm a
nd
BG
algorith
m
aft
e
r ch
angin
g
popul
ation si
ze and
maxi
mu
m g
ene
rat
i
on. From Fi
g
u
re
6 to
Figu
re
8,
the BCVA of
GA incre
a
ses to 92.6
804%
and
the
A
C
V
A
of GA
still range
s from
9
0
.5% to 91.5
%
;
two indexe
s
of PSO are n
early the sam
e
with
expe
ri
ment one; wh
ile the BCVA of BG gro
w
s
to
94.014
1% an
d the ACVA of BG is more
stable.
Finally
, the superi
o
r
of propo
se
d BG method can
be proved b
y
further exp
e
rime
nt com
pare
d
with
G
A
-PSO algo
ri
thm [20-2
1
] whi
c
h
combi
nes
good p
o
ints o
f
GA algorith
m
and PSO a
l
gorithm, an
d sho
w
e
d
in Figure 9
-
1
0
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A P2P Traffic Identification
Approa
ch Ba
sed o
n
SVM and BFA (Ch
unzhi Wa
ng)
2841
Figure 9. Cla
ssifi
cation by
BG Algorithm
(200,
20)
Figure 10. Cl
assificatio
n
b
y
GA-PSO Algorithm
(200, 20
)
As we
can
seen i
n
Fig
u
re 9
and
Fig
u
re
10,
we fi
nd eve
n
the
best
cross validation
accuracy of BG algorithm
is the same
as
GA-PSO
algorithm,
more
over, the averag
e cross
validation
accura
cy of th
e f
o
rme
r
i
s
better th
an th
e la
tter, whi
c
h
sh
ows that th
e
BG algo
rithm
is
more
sta
b
le t
hat the
GA-P
SO algo
rithm
.
Ba
sed
on
the a
bove ex
perim
ents,
we may
con
c
l
ude
that BG-SVM method ha
s the potential t
o
be useful in
classification
for P2P traffic identificatio
n.
6. Conclusio
n
In this p
ape
r,
a P2P traffic identificatio
n
app
roa
c
h
is
develop
ed b
a
s
ed
on SVM
and BG
algorith
m
. An
d we h
a
ve te
sted th
e p
r
o
p
o
se
d m
e
thod
on
re
al P2P
datasets
and
comp
ared th
em
with
seve
ral
existing
techniqu
es.
The
expe
rim
enta
l
re
sult
s in
di
cate th
at the
propo
sed
B
G
algorith
m
is f
easi
b
le to opt
imize the
pa
rameters
for S
V
M, which te
stifies that th
e novel BG
-SVM
model
can yi
eld p
r
omi
s
in
g
re
sult
s. In
all
,
ideally, the
prop
osed
me
thod h
a
s the
high
acc
u
r
a
cy o
f
identifying P2P traffics. Ho
wever, b
e
cau
s
e of t
he limitation of the e
x
perime
n
tal e
n
vironm
ent, the
P2P traffic ca
ptured i
n
this pape
r is f
r
o
m
cam
p
u
s
ne
twork, the testing data u
s
e
d
in the p
r
o
c
ess
still cannot cover all
the factors t
hat
can
affect test results. Mor
eover,
in general,
different
traffic
feature
of P2P has
differe
nt effect on
cl
assificatio
n
of
the flow, in t
he pa
per,
all
the features
are
set
with the
same
weight
value; h
o
w
to set
prope
r wei
ght valu
e for the
s
e f
eature
s
is worth
further studyi
ng.
Ackn
o
w
l
e
dg
ements
This
wo
rk i
s
sup
porte
d by
Natural Scie
nce F
oun
dati
on of China
(6117
0135
61
2022
87)
and
Natu
ral
Scien
c
e
Fou
n
dation
of Hub
e
i Provin
ce
of
Chi
na
(20
1
1
CDB0
7
5
)
, the
Key Proj
ect f
o
r
Scientific an
d
Techn
o
logi
cal Re
sea
r
ch of Educ
ation
Depa
rtment
of Hubei Province in
Chi
na
(No. D201
11
409, No. D2
0121
409
) an
d the Key
Proje
c
t for Scientific an
d Tech
nolo
g
i
c
al
Re
sea
r
ch of Wuh
an City in Chin
a (20
1
2104
2113
4).
Referen
ces
[1]
Yin Chuny
ong, Sun Ruxia,
Yang
Lei.
Node-based s
a
m
p
ling P2P
bot
detection.
TELKOMNIKA
Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
ng.
2012; 1
0
(5): 1
117-
112
2.
[2]
Yu F
anp
en
g, Z
hang S
i
do
ng,
Z
hou
Xi
n. Incentive M
e
cha
n
ism Alg
o
rith
m for P2P F
i
l
e
S
y
st
em
i
n
Camp
us N
e
t
w
orks.
T
E
LKOMNIKA Indo
nesi
an J
ourn
a
l
of Electrical
En
gi
neer
ing.
20
12;
10(6): 14
77-
148
4.
[3]
Shen
Xu
emin,
Yu H
eath
e
r, B
u
ford J
o
h
n
, Ak
on M
u
rsal
in. H
and
bo
ok of P
e
er-to-Peer
Net
w
o
r
kin
g
(
1
st
ed.). Ne
w
York
: Springer. 20
0
9
; 118.
[4] DS
Wallach.
A
Survey
of P
e
er-to-Peer
Sec
u
rity Issues.
P
r
ocee
din
g
s
of Internati
o
n
a
l S
y
mposi
u
m
o
n
Soft
w
a
re Secu
rit
y
. 20
02; 42-
5
7
.
[5]
Yan H
ua, D
ah-
Ming
Chi
u
, Joh
n
CS L
u
i.
Profi
l
i
ng an
d
id
entifi
c
at
ion
of P2P t
r
affic.
Compute
r
Netw
orks.
200
9; 53: 849-
863.
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: 2833 – 2
842
2842
[6]
Hon
g
w
e
i Che
n
,
Xin Z
hou, F
a
ngp
ing Yo
u, Hui Xu, Chu
n
zhi
W
ang, Z
h
i
w
e
i
Ye. A SVM Me
thod for P2P
T
r
affic Identification b
a
se
d
on
Multipl
e
T
r
affic Mode.
Journ
a
l
of Netw
orks.
2
010; 5(1
1
): 138
1-13
88.
[7]
Li Ju
n, Z
h
a
ng
Shun
yi,
Li
u Sh
i
don
g,
Xua
n
Y
e
.
Active P2P T
r
affic Identific
ati
on T
e
c
hni
qu
e.
Internatio
na
l
Confer
ence
on
Computati
o
n
a
l
Intellig
enc
e an
d Securit
y
. 20
0
7
; 37-42.
[8]
Jing
ang Z
h
e
n
g
,
Yabin
Xu. Ide
n
tificatio
n
of ne
t
w
ork traffic ba
sed on su
pp
ort vector machin
e.
Advance
d
Co
mp
uter T
h
e
o
ry and En
gi
ne
erin
g (ICACT
E).
2010; 3: 286-
290.
[9]
Yabi
n Xu a
nd J
i
ng
ang Z
h
en
g. I
dentificati
on o
f
Net
w
o
r
k T
r
affi
c Based o
n
Ra
dial B
a
sis F
unc
tion Ne
ur
a
l
Net
w
ork Intell
i
gent Com
putin
g and Informat
i
on Scie
nce C
o
mmunic
a
tio
n
s
.
Comp
uter an
d Informati
o
n
Scienc
e
. 201
1; 134: 17
3-17
9.
[10]
Rui W
a
ng, Y
a
ng
Liu, Y
ue-
xi
ang
Ya
ng, H
a
i
-
lon
g
W
a
n
g
.
A
New
Meth
od
for P2P T
r
affic
Identific
atio
n
Based o
n
Su
pport Vector
Machi
ne.
Internatio
nal C
onfe
r
ence o
n
Artificial Intel
lig
enc
e & Machin
e
Lear
nin
g
Co
nferenc
e (AIML0
6). Sharm El Sheik
h
, Eg
ypt. 2006; 58-
63.
[11]
Shen
F
u
ke,
C
han
ge
Pan,
R
en
Xia
o
li.
Res
earch
of P
2
P
T
r
affic Identific
ation
Bas
e
d
o
n
BP
Ne
ura
l
Net
w
ork.
Intell
i
gent Infor
m
ati
o
n Hidi
ng a
nd M
u
lti
m
e
d
ia Si
gn
al Process
i
ng
.
200
7; 75-7
8
.
[12]
Gu Yiran, W
a
ng S
uop
in
g. T
r
a
ffic Identific
a
t
ion Met
hod
fo
r Specific
P2P
Base
d o
n
Mu
ltila
yer
T
r
ee
Combi
nati
on
Classific
a
tio
n
b
y
BP-LVQ N
eura
l
-Net
w
o
rk.
Informatio
n T
e
chn
o
lo
gy an
d
Applic
ation
s
(IFITA)
. 2010; 34-3
8
.
[13]
Don
g
Seo
ng Ki
m, Ha-Nam Ng
u
y
en, Jo
ng So
u Park
. Geneti
c
algor
ithm to i
m
prove SVM b
a
sed n
e
t
w
or
k
intrusi
on d
e
tec
t
ion s
y
stem.
A
d
vanc
ed Infor
m
ati
on
Netw
orking
an
d App
l
i
c
ations (AINA)
.
2005;
155
-
158.
[14]
Mathias
M Ad
anko
n
, Mo
ham
ed
Cher
iet. Mo
del
sel
e
ct
io
n f
o
r the
LS-SVM
. Appl
icatio
n to
ha
nd
w
r
itin
g
recog
n
itio
n.
Pattern Recognition.
200
9; 42: 3
264-
327
0.
[15]
Xi
ao
Li Z
h
ang,
Xu
e F
e
n
g
C
h
e
n
, Z
heng J
i
a
H
e
. An
ACO-bas
ed a
l
gor
ithm fo
r param
eter op
timizatio
n
o
f
supp
ort vector machi
nes.
Exp
e
rt Systems w
i
th Appl
icatio
ns.
201
0; 37: 661
8
-
662
8.
[16]
Xu
an
yu
Li
u, C
hen
g Sh
ao, H
a
if
en
g Ma, Re
nxi Li
u. Optim
a
l e
a
rth press
u
re b
a
la
nce c
o
ntrol for shi
e
l
d
tunne
lin
g bas
e
d
on LS-SVM a
nd PSO.
Automati
on i
n
Cons
truction.
201
1; 20: 321-
32
7.
[17]
Z
hen
yu
an Ji
a, Jian
w
e
i Ma, F
u
ji W
a
n
g
, W
e
i Liu. H
y
bri
d
of simulat
ed an
n
eali
ng a
nd SV
M for h
y
dra
u
l
i
c
valve ch
aracter
i
stics pred
ictio
n
.
Expert Systems w
i
th Appl
ic
ations.
20
11; 3
8
: 8030-
80
36.
[18]
Nasir A
N
K, Tokhi MO, Gh
a
n
i
NM, A
h
ma
d MA. A
nove
l
h
y
br
id
spir
al-
d
ynam
ics b
a
cterial-for
a
g
i
n
g
alg
o
rithm for gl
oba
l optimiz
ati
on
w
i
th
app
lic
a
t
ion to contro
l
desi
g
n
. Co
mpu
t
ationa
l Intell
ig
ence (UK
C
I)
.
201
2; 1-7.
[19]
Das S, Dasgu
p
ta S, Bis
w
a
s
A, Abraham A, K
onar A. On Stabil
i
t
y
of the
Chemotactic
D
y
namics i
n
Ba
cte
r
i
a
l
-
Fo
rag
i
ng
Op
ti
mi
za
ti
o
n
Alg
o
r
i
t
h
m
. Sy
ste
m
s, Man
a
n
d
Cy
be
rne
t
i
cs. Pa
rt A:
Sy
ste
m
s and
Huma
ns. 200
9; 670-6
79.
[20]
Qian Y
uLi
an
g,
Z
hang
Ha
o.
F
a
ult d
i
ag
nos
is fo
r ge
nerat
or u
n
i
t
base
d
on
RB
F
neur
al
netw
o
rk opti
m
i
z
e
d
by GA-PSO.
8th Internatio
na
l Confer
ence
on
Na
tural C
o
mp
utation. 2
012;
233-
236.
[21]
Kim Dong H
w
a, Park Ji
n.
Loss mi
ni
mi
z
a
ti
on co
ntrol
of inductio
n
motor usi
n
g
GA-PSO
. 9t
h
Internatio
na
l Confer
ence
o
n
Kno
w
l
e
dge-
Based Inte
lli
g
ent Informatio
n
and E
ngi
ne
erin
g S
y
stems
.
200
5; 222-
227.
Evaluation Warning : The document was created with Spire.PDF for Python.