TELKOM
NIKA
, Vol.12, No
.4, Dece
mbe
r
2014, pp. 10
23~103
0
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v12i4.535
1023
Re
cei
v
ed Au
gust 28, 20
14
; Revi
sed O
c
t
ober 3
0
, 201
4; Acce
pted
No
vem
ber 1
6
,
2014
Complex Optimization Problems Using Highly Efficient
Particle Swarm Optimizer
Kaiy
ou Lei
1
,
Changjiu
Pu*
2
1
Intellige
n
t Soft
w
a
r
e
an
d Softw
a
r
e En
gin
eer
i
ng La
bor
ator
y
,
South
w
e
s
t Uni
v
ersit
y
, Ch
on
g
q
in
g 40
071
5,
Cho
ngq
in
g, Ch
ina
2
Net
w
ork C
ent
er, Chon
gqi
ng
Univers
i
t
y
of Educ
ati
on, Ch
on
gqi
ng 4
000
67,
Cho
ngq
in
g, Ch
ina
*
Corres
p
o
ndi
n
g
author, e-ma
i
l
: pcj88
8
0
289
@sina.c
o
m
A
b
st
r
a
ct
Many e
ngi
ne
er
ing
prob
le
ms
are the c
o
mpl
e
x opti
m
i
z
a
t
io
n pro
b
le
ms w
i
t
h
the l
a
rg
e nu
mb
ers o
f
glo
bal
an
dl
ocal
opti
m
a.
Due
to its co
mplex
i
ty, gen
eral
parti
cle sw
ar
m o
p
ti
mi
z
a
t
i
o
n
meth
o
d
inc
lin
es tow
a
rds
stagnati
on p
h
e
n
o
m
e
na i
n
the
later stag
e of e
v
oluti
on, w
h
ich
leads to
pr
e
m
ature co
nverg
e
n
ce. T
herefor
e, a
hig
h
ly efficie
n
t
particle sw
arm opti
m
i
z
e
r
is prop
ose
d
i
n
this pa
per,
w
h
ich empl
oy the dyn
a
mic
transitio
nstrate
g
y ofin
ertia fac
t
or, search sp
ace b
oun
dary
ands
earchv
e
l
o
citythresho
l
db
a
s
ed o
n
in
divi
d
ual
cognitionin each cycle to plan lar
ge-sc
ale space glob
al search and r
e
fined lo
c
a
l s
earch as
a whole
accord
ing to
the fitness c
h
a
nge
of sw
arm
in o
p
ti
mi
z
a
ti
on
process
of the en
gin
eer
ing
prob
le
ms, an
d
t
o
improve c
onv
e
r
genc
e prec
isio
n, avoi
d pre
m
a
t
ure pro
b
le
m, e
c
ono
mi
z
e
co
mputatio
na
l exp
e
n
ses, an
d obta
i
n
glo
bal o
p
ti
mu
m. Sever
a
l co
mp
lex b
ench
m
ark functions
are use
d
to testify the new
alg
o
rith
m an
d
th
e
results show
ed
clearly the rev
i
sed al
gorith
m
c
an rap
i
dly co
nv
erge at hi
gh q
u
a
lity sol
u
tions.
Ke
y
w
ords
: par
ticle sw
arm o
p
timi
z
e
r, co
mpl
e
x optim
i
z
a
t
i
on
prob
le
m, pre
m
ature conv
erg
e
n
ce
1. Introduc
tion
As a newly d
e
velope
d po
pulation
-
ba
se
d comp
utatio
nal intelligen
ce alg
o
rithm,
Particle
Swarm
O
p
timization
(PSO
) wa
s
origi
nat
ed a
s
a
simul
a
tion of
sim
p
lified soci
al mo
del of
bird
s in
a
floc
k [1]-[4]. The PSO
algorithm
has
les
s
parame
ters, ea
sy impl
e
m
entation, fa
st convergen
ce
spe
ed a
nd ot
her
cha
r
a
c
teristics, is
wid
e
ly use
d
in many
fields,
s
uch as solvin
g
com
b
inato
r
ial
optimizatio
n, fuzzy control, neur
al ne
twork trai
nin
g
, etc. But, the PSO alg
o
rithm
with othe
r
algorithms is also easy to fall into local optim
umin
fa
s
t
c
o
n
v
er
g
e
n
c
e
pr
oc
es
s
,
a
ffe
c
t
ing
th
e
conve
r
ge
nce
pre
c
isi
on, so how to ove
r
come
prematu
r
e convergen
ce, and
im
pro
v
e the accura
cy
of convergen
ce is al
way
s
a hot and difficult pro
b
lem i
n
the re
sea
r
ch field [5]-[11].
To avoidthe p
r
ematu
r
e p
r
o
b
lem and
spe
ed up the con
v
ergen
ce p
r
o
c
e
ss, the
r
ea
re many
approa
che
s
sug
g
e
s
ted b
y
rese
arche
r
s.According
t
o
the re
se
arch results p
u
b
lish
ed in
re
cent
years,
the i
m
provem
ent of
PSO al
gorit
hm mai
n
ly
in
clud
es adj
ust
i
ng al
gorith
m
pa
ramete
rs,
the
improvem
ent
of topol
ogi
cal structu
r
e,
and mixe
d
wi
th other alg
o
rithm, etc [6]-[
12].The p
u
rp
ose
of improvem
ent strate
gie
s
is to b
a
lan
c
e the global
sea
r
ch abilit
y and local
sea
r
ch ability of
particl
es, so as to improve
the performa
n
ce of the alg
o
rithm.
In this pape
r, we modifie
d
the traditio
nal PSO (TP
S
O) algo
rith
m with the dynamic
transitio
n stra
tegy ofinertia
factor, se
arch sp
a
c
e b
o
u
ndary an
dsea
rchvel
ocitythresh
o
ldb
a
se
d on
individual
co
gnitionin
ea
ch cy
cle,which
c
an
bal
an
ce
the glo
bal
se
arch
ability a
nd lo
cal
sea
r
ch
ability of particle
s
, and ha
s an excellent
sea
r
ch pe
rf
ormance to lea
d
the sea
r
ch dire
ction in e
a
rly
conve
r
ge
nce
stage
of se
arch p
r
o
c
e
s
s. Experim
ent
al re
sults
on
several co
mplexben
ch
mark
function
s de
monst
r
ate tha
t
this is a verypromi
s
in
g
w
a
y
to improve the sol
u
tion q
uality and rat
e
of
su
ccess si
gni
ficantly in optimiz
ing
com
p
l
e
x engine
erin
g probl
em
s.
Section 2
gives
some
ba
ckgroun
d kno
w
led
ge of th
e PSO algo
ri
thm. In secti
on 3, the
prop
osed me
thod and the
experimental
design a
r
e
descri
bed in
detail, and
correlative re
sults
are given in
section 4. Fina
lly, the
discu
ssion
s
are dra
w
n in sectio
n 5.
2. Back Gr
o
und
In 199
5, the
pa
rticle
swarm
optimize
r
(PSO
) i
s
a po
pulatio
n
basedal
gorith
m
that
wa
sinvente
d
by Jame
s
Kennedy a
n
d
Ru
ssell Eberh
a
rt,which
wa
s inspire
d
by the so
cial
behavio
rof a
n
imals
su
ch
as fish
scho
o
ling and bi
rd
flockin
g
.Similar to othe
r p
opulatio
n-b
a
sed
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 102
3 – 1030
1024
algorith
m
s, such
as evol
utionary alg
o
rit
h
ms, PSO
can solve a v
a
riety ofdifficult optimizati
on
probl
em
s but
has sho
w
n
a fasterconve
r
gen
ce
rate
than othe
r evolutiona
ry alg
o
rithm
s
on
so
me
probl
em
s. An
othera
d
vanta
ge of PS
O i
s
that it
ha
s
ve
ry
few parame
t
ers
toadj
ust, whi
c
h makes it
particula
rly e
a
sy to im
ple
m
ent [1]. In
PSO, each p
o
tential soluti
on is a
“bi
r
d
”
in th
e sea
r
ch
spa
c
e,
whi
c
h
is
calle
d “pa
r
ticle
”
. Each
particl
e
h
a
s
a fitness valu
e evaluate
d
by the obj
ect
i
ve
function,
and
flies ove
r
th
e solution
sp
ace
with
a
v
e
locity by foll
owin
g the
cu
rre
nt glob
al b
e
st
particl
e and it
s individual b
e
st po
sition. With the di
re
ction
s
of best
particle
s
, all particl
es of th
e
swarm
can e
v
entually land
on the best solution.
The fo
undati
on of PSO
i
s
b
a
sed
on t
he hyp
o
th
e
s
i
s
that so
cial sha
r
ing
of
inf
o
rmatio
n
among
co
nspecifi
c
soffers an evolut
io
n
a
ry advanta
g
e
.In the ori
g
i
nal PSO formula, pa
rticle
i is
denote
d
a
s
X
i
=
(
x
i1
,x
i2
,...,x
iD
), whi
c
h
rep
r
e
s
ent
s a
pote
n
t
ial sol
u
tion t
o
a
pro
b
lem
in D-dim
e
n
s
io
nal
spa
c
e. Ea
ch
particl
e maint
a
ins
a memo
ry of its previous
be
st po
sition
Pbes
t,
and a velo
cit
y
along e
a
ch d
i
mensi
on, re
pre
s
ente
d
a
s
V
i
=
(
v
i1
,v
i2
,...,v
iD
). At each
iteration,
the positio
n of the
particl
e with the best fitness in the se
arch sp
ace, desi
gnate
d
as g, and the
P vector of
the
curre
n
t pa
rticle are
combi
ned to
adju
s
t
the velo
ci
ty along
ea
ch
d
i
mensi
on, a
n
d that velo
city is
then used to comp
ute a ne
w po
sition for the particle.
In TPSO, the
velocity and p
o
sition of pa
rticle
i
at (
t+
1
)
th iteration a
r
e
updated a
s
follows:
t
id
t
gd
t
id
t
id
t
id
t
id
x
p
r
c
x
p
r
c
v
w
v
2
2
1
1
1
(1)
1
1
t
id
t
id
t
id
v
x
x
(2)
Con
s
tant
s c1 and c2 d
e
termin
e the
relative influen
ce of the
soci
al and
cog
n
ition
comp
one
nts
(lea
rnin
g rates),
whi
c
h
o
ften both a
r
e set to the
sam
e
valu
e
to give e
a
c
h
comp
one
nt e
qual
weig
ht;
r
1
an
d
r
2
are
ran
dom
nu
mbers
uniformly distri
bute
d
in th
e inte
rval
[0,1].A c
ons
t
ant,
v
ma
x
, was u
s
e
d
to limi
t
the velocitie
s
of th
e pa
rti
c
le
s. The
pa
rameter
w
, whic
h
wa
s introd
uced as a
n
ine
r
tia factor,
can dynami
c
al
ly adjust the
velocity over time, grad
u
a
lly
focu
sing the
PSO into a local search [5]
.
To speed
up
the co
nverg
e
n
ce
pro
c
e
s
s
and av
oi
d the
prem
ature problem, Shi p
r
opo
sed
the PSO with linearly decrease fa
ctor method (LDWPSO)
[4],[5].
Suppose
w
ma
x
is the maxi
mum
of inertia fact
or,
w
mi
n
is the minimum of inertia fa
ctor,
run
is the
cu
rre
nt iteration
s
,
run
ma
x
is the
total iteration
s
. The ine
r
tia factor i
s
form
ulated a
s
:
max
min
max
max
run
run
w
w
w
w
(3)
3. A Highly
Efficien
t Parti
c
le S
w
a
r
m
O
p
tim
i
zer (
H
E
PSO)
Due to thecomplexity of a
gr
eat deal global and
local optima,TPSO isrevised as
HEPSO
by fourdynam
ic strategie
s
to
adapt comp
lex optimizati
onproble
m
s.
3.1. D
y
namic
Harmonizati
on Inertia Fa
ctor
w
First of all, the large
r
w
ca
n enha
nce gl
obal search a
b
ilitie
s of PSO, so to expl
ore la
rge
-
scale search
spa
c
e a
nd ra
pidly locate th
e app
ro
ximat
e
positio
n of global o
p
timu
m, the smalle
r
w
can enhance
local
s
earch
abilities of PS
O, part
icles sl
ow down,
deploy
refined l
o
cal
search,
and
obtain glo
bal
optimum. Seco
ndly, the more diffi
cult the optimiza
t
ion probl
em
s are, the more
fortified
the global sea
r
ch
abilities n
eed,
on
ce
lo
cated th
e ap
proximate
po
sition of glo
bal
optimum, the
refine
d lo
ca
l sea
r
ch
will
furthe
r be
strengthe
n to
get glob
al o
p
timum [7]-[1
2].
Therefore,the
w
can
ha
rmo
n
ize
glo
bal
search
an
d lo
ca
l se
arch
aut
omatically, a
v
oid
p
r
ematu
r
e
conve
r
ge
nce
and to rapi
dly gain glob
al o
p
timum.
Acco
rdi
ng to
the
con
c
lu
si
ons ab
ove,
a
ne
w in
ertia
factor de
cline
cu
rve
⑷
for PSO
is
con
s
tru
c
ted,
demon
strated
in Figure 1:
n
run
run
n
w
w
max
max
^
*
exp
*
(4)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Com
p
lex O
p
tim
i
zation Prob
lem
Using Hi
ghly Effici
ent
Particle Swarm
Optim
i
zer (Kaiyou L
e
i)
1025
Whe
r
e
n
is a
con
s
tant larg
er than 1, taken 50 as
initi
a
l value in the pape
r.Evidently, th
e
inertia
facto
r
de
cline
curve of figu
re
1can fo
rcef
ully
se
arch
l
a
rge
-
scale gl
obal sp
ace, and
dynamically tran
sform to refined lo
cal search,
nam
el
y global se
arch abilitie
s a
nd local se
arch
abilities are harmoni
zed base
d on the strategy to adapt demand of
compl
e
x
optimizatio
np
roble
m
s.
Figure1. Dyn
a
mic ha
rmo
n
i
z
ation
w
cu
rv
e
Figure 2. Dyn
a
mic tra
n
sfo
r
mation
w
curves
3.2.D
y
namic
Trans
f
orma
tion Inertia Fa
ctor
w
Strateg
y
Global
se
arch an
d lo
cal
search
are
two
key a
s
p
e
ct
s
of PSO ba
se
d on
w
. In
a given time
of se
arch
pro
c
e
ss, it i
s
usua
lly ha
rd to
determi
ne,
when to
end
the la
rge
-
scal
e
glo
bal
sea
r
ch,
and sta
r
t loca
l search,a
nd
gain qui
ck co
nverge
nce [8]-[10].
In figure
2
,p1,
p2,…, pn
are
tran
sform
a
tio
n
poi
nt
s, d1,
d2, …, dn
are
global
co
nve
r
gen
ce
points, th
e al
gorithm
sele
ct a tran
sfo
r
m
a
tion p
o
int fro
m
them,
and
switch
to refined l
o
cal
sea
r
ch
to glob
al
con
v
ergen
ce
poi
nt. The
sel
e
ct
ion of t
r
an
sfo
r
mation
poi
ntis u
s
u
a
lly ha
rd. To
confirm
the
transfo
rmatio
n poi
nt,the a
l
gorithm
is d
e
sig
ned
to
combine
iteration time
s of
cu
rrent gl
ob
al
optimum of functions.If the curren
t global
optimum is not improved
a
fter the search of an interv
al
of definite iterations, the
al
gorithm
switch
to refin
ed l
o
cal
se
arch
with the
smal
ler
n
, or
c
o
n
t
in
ue
curre
n
t global
search
with the cu
rrent
n
. Thecomp
u
ted
equation i
s
d
e
fined a
s
:
n
r
r
n
esle
n
n
p
p
IF
i
gd
k
i
gd
2
1
(5)
Whe
r
e
k
i
gd
p
,
i
gd
p
are th
e
(i
+
k
)th,
i
th
,tak
en values of
t
gd
p
r
e
spec
tively,
k
i
s
an
i
n
terval of
definite iterati
ons.
3.3.D
y
namic
Trans
f
orma
tion Search S
p
ace Bound
ar
y
Strategy
In se
arch
proce
s
s, all
p
a
rticle
s
gath
e
r g
r
a
dually
to the
cu
rrent be
st
reg
i
on, the
algorith
m
is p
r
opitiou
s to q
u
icken
conve
r
gen
ce
be
cau
s
e of the
re
duced
sea
r
ch
spa
c
e, b
u
t, the
global optim
a
may be lost [7]-[10]. In most ca
se
s, the global optim
a
may be hidd
en som
e
whe
r
e
in the gath
e
r
ing a
r
e
a
ne
arby, and th
e effect
ive search a
r
e
a
fo
und i
s
not
easy.To
solv
e
thepro
b
lem, t
he imp
r
oved
algorith
m
not
only red
u
ce
s the
se
arch sp
ace
to qui
cke
n
co
nverg
e
n
c
e,
but also avoi
ds the
prem
ature
pro
b
le
m, esp
e
ci
ally in co
mplex
optimizatio
np
roble
m
s.Th
us, a
dynamic tra
n
s
form
ation
search
sp
ace
boun
dary
strategy is de
sign
ed
ba
se
d on
individ
ual
cog
n
ition. Assume that
a particl
e
flight in the cu
rrent
boun
dary [
b
ma
x
(
i
),
b
mi
n
(
i
)], the
algorith
m
re
du
ce th
e
sea
r
ch
bou
nda
ry if the
curre
n
t op
timum is bett
e
r, othe
rwise, expan
d sea
r
ch
boun
dary in n
e
xt iteration, and in the sa
me br
e
a
th, ra
ndomly initiali
ze the sp
eed
and po
sition
o
f
each p
a
rticl
e
after th
e
k
it
eration
s
. T
h
e
b
ma
x
and
b
mi
n
are the
bo
u
ndary
of the
swarm
in th
e
k
iteration. The
comp
uted eq
uation is d
e
fined a
s
:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 102
3 – 1030
1026
min
max
min
min
min
max
max
max
min
max
min
min
min
max
max
max
i
i
b
b
r
r
i
b
i
b
b
b
r
r
i
b
i
b
else
b
b
r
r
i
b
i
b
b
b
r
r
i
b
i
b
gbest
gbest
IF
2
1
2
1
2
1
2
1
1
1
1
1
1
(6)
3.4. D
y
namic
Transformation Search Velocit
y
ThresholdStra
t
eg
y
Many publi
s
h
ed works
ba
sed
on p
a
ra
meters sele
ction pri
n
ci
ple
s
pointed
out, velocity
threshold
[
v
ma
x
(
i
),
v
mi
n
(
i
)] of a p
a
rti
c
leaff
e
cts the
co
n
v
ergen
ce
p
r
e
c
isi
on
and
sp
eed
of alg
o
rit
h
m
s
t
rongly [9]-[11]. Large
v
ma
x
(
i
)incre
ases th
e sea
r
ch regi
on, enha
nc
in
g global sea
r
ch capa
bility,
as
well as
sm
all
v
ma
x
(
i
) de
c
r
e
a
se
s t
h
e
sea
r
ch
r
egio
n
,
a
d
just
in
g
sea
r
ch
dire
ct
ion
of
ea
ch
part
i
cle
freque
ncy. T
hus, adyna
mi
c tran
sform
a
tion sea
r
ch
velocitythre
sh
oldstrategy is design
ed ba
sed
on individual
cog
n
ition. Th
e
v
ma
x
and
v
mi
n
are the thre
shol
d of the swarm in the
k
iteration
s
, th
e
comp
uted eq
uation is d
e
fined a
s
:
min
max
min
min
min
max
max
max
min
max
min
min
min
max
max
max
i
i
v
v
r
r
i
v
i
v
v
v
r
r
i
v
i
v
else
v
v
r
r
i
v
i
v
v
v
r
r
i
v
i
v
gbest
gbest
IF
2
1
2
1
2
1
2
1
1
1
1
1
1
(7)
Accordi
ng to
the above m
e
thods, TPSO is
modified
as HEPSO, whi
c
hhas the excellent
s
e
arch performance to optimiz
e
c
o
mplex probl
ems
.
The flow of
the HEPSOalgorithm is
as
follows
:
Step1.
Set algorithm
param
eters;
Step2.
Ran
domly initialize the
sp
e
ed and p
o
siti
on of each pa
rticle;
Step3.
Evaluate the fitness of e
a
c
h pa
rticl
e
a
nd dete
r
mine
the initial values of the in
dividual
and glo
bal be
st positio
ns:
t
id
p
and
t
gd
p
;
Step4.
Upd
a
te veloci
ty and positio
n usin
g (1
), (2) and
(4);
Step5.
Evaluate the fitness a
nd d
e
termin
e the curr
e
n
t value
s
of the indivi
dual an
d glo
bal be
st
positio
ns:
t
id
p
and
t
gd
p
;
Step6.
Dete
ct
the
gbe
st
i
,
gbest
i+1
and
gbest
i+k
, to dynamically tran
sform
w,
sea
r
c
h
spac
e
boun
dary an
d
velocity threshold u
s
ing (5), (6)an
d
(7
);
Step7.
Ran
domly initialize the
spe
ed and p
o
siti
on after the
k
iteration
s
;
Step8.
Loop to Ste
p
4 and
re
pea
t until a give
n maximum i
t
eration
num
ber i
s
attain
e
d
or th
e
conve
r
ge
nce crite
r
ion i
s
sa
tisfied.
4. Computati
onal Experi
ments
4.1.Testing F
unction
s
To tes
t
the
HEPSO and
c
o
mpare it with ot
her tec
h
niques
in the
literature, we adopt
large
variety
of ben
ch
ma
rk fu
nctio
n
s
[8]-[16], amo
ng which mo
st fun
c
tion
s
are
multimod
al,
abno
rmal o
r
comp
utationa
l time con
s
u
m
ing, and
ca
n hardly get
favorable
re
sults by cu
rre
nt
optimizatio
n
algorith
m
. Due to limited
spa
c
e,
we
only sel
e
ct f
our
rep
r
e
s
en
tative functio
n
s
optimizatio
n result
s to list in the pape
r.
500
500
sin
)
(
1
1
i
n
i
i
i
x
x
x
x
f
(8)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Com
p
lex O
p
tim
i
zation Prob
lem
Using Hi
ghly Effici
ent
Particle Swarm
Optim
i
zer (Kaiyou L
e
i)
1027
12
.
5
12
.
5
]
10
)
2
cos(
10
[
)
(
1
2
2
i
i
n
i
i
x
x
x
x
f
(9)
32
32
20
))
2
cos(
1
exp(
)
2
1
5
1
exp(
20
)
(
1
1
3
i
n
i
i
n
i
i
x
e
x
n
x
n
x
f
(10)
600
600
1
)
cos(
4000
1
)
(
1
1
2
4
i
n
i
n
i
i
i
x
i
x
x
x
f
(11)
4.2. Algorith
m
Parameter
Setting
Paramete
rs u
s
ed
in
ou
r al
gorithm
a
r
e
set to b
e
: lea
r
ning
rate
c
1
=c
2
=2;
w
=
0
.7,
w
ma
x
=1,
w
mi
n
=
0
.1; maximumiterations
ru
n
ma
x
=30000; iteration
s
k
=1
000
(30
0
for
f
1
(
x
)); p
opulatio
n size is
100; spee
d a
nd po
sition
of parti
cles are
limited in
defi
n
ition area of
function
s; take functio
n
f
1
(
x
),
f
2
(
x
),
f
3
(
x
),
f
4
(
x
) a
s
fitness value. Stop rule is: |
f
best
─
f
min|
≤
10
-4
(
f
best an
d
f
min are the glo
bal
optimum and
the cu
rrent getting
o
p
tim
u
m).
T
he
ru
n
n
ing
enviro
n
m
ent i
s
: MAT
L
AB7.0, Pent
ium
IV 2GHz
CPU, 256M RA
M, Win XP OS.
4.3.Experimental Results
The te
sting
function
s i
s
run5
0 time
s based o
n
T
PSO, LDWP
SO and
HE
PSO, the
comp
ari
s
o
n
of statistical
results of 2
0
-
100
0 dime
n
s
ion
s
fun
c
tio
n
s a
r
e
sho
w
n in table 1
-
2
,
respe
c
tively. In addition, t
he datum
of l
i
terature
[12]
(MAGA) i
s
li
kewi
se li
sted i
n
table 1. 1
0
0
0
-
1000
0 dimen
s
ion
s
fun
c
tio
n
s a
r
e test with MAGA,TPSO, LDWPS
O
and HEPS
O based on t
he
sampli
ng inte
rval 500, each testing function run
s
20 times yet, the statistical re
sults a
r
e sh
o
w
n
in table 1-2 a
nd figure 3
-
6,
resp
ectively.
Table 1. Re
sults of 20-1
0
0
00 dimen
s
io
n
s
functio
n
s av
erag
e co
nvergen
ce iteratio
ns
n Function
Stopcriterion
Algorithm
MAG
A
TPSO
LDWPSO
HEPSO
20
f
1
(x
)
f
2
(x
)
f
3
(x
)
f
4
(x
)
10
-4
2483
4301
3583
2566
2120
685
731
936
1872
377
608
1873
802
102
77
1632
100
f
1
(x
)
f
2
(x
)
f
3
(x
)
f
4
(x
)
10
-4
5443
10265
5410
4447
3743
934
872
1169
2659
865
864
948
867
209
151
2346
200
f
1
(x
)
f
2
(x
)
f
3
(x
)
f
4
(x
)
10
-4
7284
14867
6061
5483
6562
1183
1063
1349
4458
962
947
1135
887
1278
165
2678
400
f
1
(x
)
f
2
(x
)
f
3
(x
)
f
4
(x
)
10
-4
12368
17939
6615
6249
10348
1461
1564
1723
7659
1123
1062
1587
1586
1743
245
2356
103
f
1
(x
)
f
2
(x
)
f
3
(x
)
f
4
(x
)
10
-4
22827
20083
7288
7358
15617
2834
2034
2327
13457
1260
1143
4562
2654
1668
389
2698
2×10
3
f
1
(x
)
f
2
(x
)
f
3
(x
)
f
4
(x
)
10
-4
45621
17521
7156
14578
27435
6533
2867
4021
18652
4534
1145
2523
10845
3532
452
2034
4×10
3
f
1
(x
)
f
2
(x
)
f
3
(x
)
f
4
(x
)
10
-4
90453
13067
7611
12034
70123
8545
3823
6701
40032
4065
1945
4967
15034
2056
624
2506
6×10
3
f
1
(x
)
f
2
(x
)
f
3
(x
)
f
4
(x
)
10
-4
130067
13166
7607
16223
85324
9022
4712
8156
43136
4517
2055
3578
25156
2533
1156
3156
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 102
3 – 1030
1028
n Function
Stopcriterion
Algorithm
MAG
A
TPSO
LDWPSO
HEPSO
8×10
3
f
1
(x
)
f
2
(x
)
f
3
(x
)
f
4
(x
)
10
-4
125245
14523
7921
18234
81489
85434
4567
7067
61378
6556
3467
4523
18000
3500
568
3512
10
4
f
1
(x
)
f
2
(x
)
f
3
(x
)
f
4
(x
)
10
-4
198745
19807
7992
27983
130679
14332
6621
15357
85045
8523
4434
13534
41289
4668
1745
6123
Table 2.Com
pari
s
on
re
sult
s of 20-1000
0
dimensi
o
n
s
functio
n
s ave
r
age converge
nce rate (%)
n
f
1
(x
)
f
2
(x
)
TPSO
LDWPSO HEPSO TPSO
LDWPSO
HEPSO
20 82.2
89.7
100
100
100
100
100 52.7
65.8
100
84.4
100
100
200 33.5
53.5
100
66.3
90.7
100
400 26.6
45.1
98.1
43.8
65.2
100
10
3
6.8 22.1
93.2
31.2
54.3
100
2×10
3
5.5
19.8
84.3
24.2
43.6
89.8
4×10
3
4.2
16.8
69.1
22.3
39.7
84.4
6×10
3
2.9
14.4
60.1
19.7
32.6
78.3
8×10
3
1.8
12.8
53.6
28.5
30.5
70.1
10
4
1.3
11.6
44.7
16.3
25.9
64.8
n
f
3
(x
)
f
4
(x
)
TPSO LDWPSO
HEPSO
TPSO
LDWPSO
HEPSO
20
87.3 100
100
69.4
90.6
100
100 61.8
91.2
100
50.7
72.5
100
200 40.6
72.6
100
30.1
53.8
95.7
400 33.9
54.5
95.3
20.8
38.8
90.3
10
3
14.1
47.3
90.6
7.2
24.3
89.5
2×10
3
8.7 22.8
84.6
23.5
23.5
78.2
4×10
3
7.1
20.4
68.4
21.7
17.3
58.4
6×10
3
5.4 15.6
60.1
18.5
14.2
55.1
8×10
3
4.9 14.1
57.8
16.9
12.3
50.4
10
4
3.6 12.4
53.6
15.3
10.5
43.6
Figure 3. The
convergen
ce
result
s of f
1
(x
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Com
p
lex O
p
tim
i
zation Prob
lem
Using Hi
ghly Effici
ent
Particle Swarm
Optim
i
zer (Kaiyou L
e
i)
1029
Figure 4. The
c
onve
r
ge
nce results of f
2
(x
)
Figure 5.The
c
onve
r
ge
nce results of f
3
(x
)
Figure 6. The
c
onve
r
ge
nce results of f
4
(x
)
5. Conclusio
n
The exp
e
rim
ental results of Tabl
e 1
-
2
ca
n
de
du
ce th
at the
effectivene
ss of the
HEPSOalgori
thm based on individual cognitioni
s vali
dated,whi
c
h guide particles to search in t
he
more
effective are
a
thro
ug
h dynami
c
ad
justmentt
he search spa
c
e,
provi
de stabl
e
co
nverg
e
n
c
e,
resulting in hi
gher
su
ccess rate and a
c
cura
cy of
conv
erge
nce. The
algorithm run
s
cla
s
sical PSO
only, so to ke
eps its
simple
and ea
sy ch
ara
c
teri
stic.
The experimental results
of Figure 3-6 s
h
ow that the HEPSOalgor
ithm has
exc
e
llent
sea
r
ch pe
rfo
r
man
c
e, e
s
p
e
cially compl
e
x enginee
ri
ng pro
b
lem
s
.
As the dimensi
o
n
s
of the
function
s g
r
o
w
fleetly, the increa
se of th
e aver
a
ge co
nverge
nce st
eps i
s
sl
ow,
so the al
gorit
hm
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 102
3 – 1030
1030
has
rapi
d co
nverge
nce sp
eed a
nd can
avoid premat
ure.In ad
ditio
n
, it can ea
sil
y
be applie
d
to
large a
nd mo
re co
mplex practical multi-o
b
jective optim
ization p
r
obl
e
m
s.
Ackn
o
w
l
e
dg
ements
This work is
suppo
rted by Key Project o
f
Chine
s
e Min
i
stry of Educa
t
ion (104
262
).
Referen
ces
[1]
J. Kenne
d
y
,
RC. Eber
hart
. Particle swarm
optim
i
z
a
tion
. Proceedings of IEEE International
Confer
ence
on
Neura
l
Net
w
or
ks. Piscata
w
a
y, USA. 1995: 1
942-
194
8.
[2]
X
.
Hu, RC. Eberhart, YH. Shi.
Engi
neer
in
g o
p
timi
z
a
ti
on w
i
t
h
p
a
rticle sw
ar
m.
Proc
ee
din
g
s
of the IEEE
S
w
a
rm Intell
ig
ence S
y
mp
osi
u
m. Indian
ap
oli
s
, Indiana. 2
0
0
3
: 53-57.
[3]
Clerc, J. Ke
nn
ed
y. T
he partic
l
e s
w
a
rm: e
x
p
l
osio
n,
stabi
lit
y
,
an
d co
nverg
e
n
ce i
n
a mu
lti-dime
nsio
n
a
l
complex
space.
IEEE
Transac
tions on Evo
l
uti
onary C
o
mput
ation
. 20
02; 6(
1): 58-73.
[4]
YH. Shi, RC.
Eberh
a
rt.
Emp
i
rical stu
d
y of
particl
e sw
arm opti
m
i
z
at
ion
.
Procee
din
g
s o
f
the IEE
E
Con
g
ress on E
v
oluti
onar
y C
o
mputatio
n.
Piscata
w
a
y, USA.
1999: 1
945-
19
50.
[5]
YH. Shi, RC. Eberh
a
rt.
A modifi
ed p
a
rticle
sw
arm opti
m
i
z
er
. Proceedings of the IEEE Congress o
n
Evoluti
onar
y C
o
mputati
on.
Ne
w
York. 1
998:
69-7
3
.
[6]
Che
n
D, W
a
n
g
GF
, Chen Z
Y
.
T
he in
ertia w
e
ig
ht self-a
dap
ting i
n
PSO
. Procee
din
g
s of t
he Sev
ent
h
w
o
rld C
o
n
g
res
s
on intel
lig
ent
Contro
l and
au
tomation. Ne
w
York. 2008: 5
3
13-5
316.
[7]
Ange
lin
e, P.
Using
sel
e
ctio
n to i
m
pr
ove
particl
e sw
arm opti
m
i
z
at
io
n
. Procee
din
g
s o
f
IJCNN’9
9
.
W
a
shin
gton, U
SA. 1999: 84-
8
9
.
[8]
Kai
y
o
u
L
e
i, Y
u
hui Qi
u, Yi
He
.
A new
a
dapti
v
e w
e
ll-ch
ose
n
inerti
a w
e
i
ght
strategy to
au
tomatic
a
l
l
y
har
mo
ni
z
e
gl
o
bal and loc
a
l
search abi
lit
y
in
p
a
rticle sw
arm
o
p
timi
zation
. Proc
ee
din
g
s of 1st
Internatio
na
l S
y
mp
osi
u
m on
S
y
stems a
nd
Contro
l in
A
e
ro
space and Astrona
utics.Harb
i
n
Chi
na.
2
0
0
6
:
342-
346.
[9]
Che
n
Bi
ng-ru
i, F
eng
Xia-ti
ng
. Particle
s
w
ar
m opt
imiz
atio
n
w
i
th c
ontract
ed r
ang
es
of
both s
earc
h
space a
nd ve
lo
cit
y
.
Jour
nal of
Northe
astern U
n
iversity (Nat
ur
al Scie
nce)
. 20
05; 5: 488-
491.
[10]
Y.Shi, RC. Eberhart.
Para
me
ter selectio
n in
particle sw
ar
mo
pti
m
i
z
a
t
i
o
n
.
Proceed
in
gs of theSeve
n
t
h
Annu
al Co
nfer
ence o
n
Evol
ution
a
r
y
Pr
ogra
mming. Ne
w
Y
o
rk. 1998; 1
4
4
7
: 591–
60
0.
[11]
W
ang Ju
n-
w
e
i
,
W
ang Di
ng-
w
e
i. E
x
p
e
rime
nts and
an
al
ysis on
inerti
a
w
e
i
ght in
pa
rticle s
w
ar
m
optimiz
ation.
J
ourn
a
l of Syste
m
s En
gin
eeri
n
g
. 2005; 2: 18
4
-
198.
[12]
Z
hong W
e
i-ca
i, Xue Mi
ng-zh
i, et
al. Multi-agent gen
etic alg
o
rithm for sup
e
r-hig
h dim
ens
ion
comp
l
e
x
functions optimization.
Natural
Science Pro
g
r
e
ss
. 2003; 1
0
: 107
8-10
83.
[13]
Li Bi
ng-
yu,
Xi
ao Yu
n-shi,
W
ang L
e
i. A
h
y
bri
d
partic
l
e
s
w
a
rm o
p
timi
zation
alg
o
rith
m for solvi
n
g
compl
e
x functi
ons
w
i
th h
i
g
h
d
i
mensi
ons.
Info
rmati
on a
nd C
ontrol
. 20
04; 1:
27-30.
[14]
Hui
w
a
ng, Z
h
ij
ian
W
u
, S
Ra
hnma
y
a
n
. E
n
h
ance
d
opp
ositi
on
bas
ed
diffe
rentia
l ev
ol
utio
n for s
o
lvi
n
g
hig
h
-dim
ensi
o
n
a
l conti
nuo
us
optimizati
o
n
probl
ems.
Soft Compti
ng-
A F
u
sion of F
ound
atio
ns,
Method
olo
g
i
e
s and Ap
plic
atio
ns
. 2011; 1
5
(1
1): 2127-
21
40.
[15]
Li W
en-fen
g
, Lia
ng
Xi
ao-l
e
i,
Z
hang Yu. R
e
searc
h
on P
S
O
w
i
th cl
uste
rs and het
ero
gen
eit
y
.
Acta
Electron
ica Sin
i
ca
. 201
2; 40(1
1
): 2194-
21
99.
[16]
Y Shi, L Gao, G Z
hang. Cell
ular
p
a
rticles
w
a
rm optimiz
a
t
ion.
Information Sciences
. 20
11; 181(
20):
446
0-44
93.
Evaluation Warning : The document was created with Spire.PDF for Python.