TELKOM
NIKA
, Vol.11, No
.3, March 2
0
1
3
, pp. 1465 ~ 1472
ISSN: 2302-4
046
1465
Re
cei
v
ed O
c
t
ober 2
2
, 201
2; Revi
se
d Ja
nuary 17, 20
1
3
; Acce
pted Janua
ry 2
9
, 20
13
SVR-based RPD Approach for Complex Processes and
its Application in Circuit Optimization
Cui Qing’an*
1
, Zhang Yuxue
1
, Cui Nan
2
*, Liu Huihua
1
1
School of man
agem
ent scie
n
c
e and e
n
g
i
ne
erin
g,
Z
hengz
h
ou Un
iversit
y
, Z
hengz
ho
u, Chin
a,
2
Economics a
n
d
Mana
gem
ent
School, W
uha
n Univ
ersit
y
, W
uha
n, Chi
na,
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: cuiqa
@
zzu.e
du.cn
A
b
st
r
a
ct
Durin
g
the life
s
pan of electr
onic pro
ducts,
the
output vol
t
age an
d curre
nt
fluctuate du
e to the
rand
o
m
fluctua
t
ions of parameter values of circ
uit compo
n
ents and envir
on
me
nt
al nois
e
. Extant method
s
of circuit
desi
g
ns, such
as p
a
ra
meter sw
e
e
p
an
d se
nsitivi
t
y analys
is, ar
e har
d to
obta
i
n g
l
ob
al r
o
b
u
s
t
opti
m
i
z
at
ion
of
output c
hara
c
teri
stics. T
h
is pap
er pr
opo
ses a SVR-
b
ased r
o
b
u
st p
a
ra
meter
desi
g
n
appr
oach to r
each g
l
ob
al ci
rcuit opti
m
i
z
at
i
on. F
i
rs
t, the
appr
oach fits an e
m
p
i
rica
l mo
de
l of proc
ess
respo
n
ses
by
usin
g SVR. N
e
xt, it introd
uc
es the fl
uct
uati
ons of c
ontrol
l
abl
e fact
or var
i
atio
ns a
nd n
o
i
s
e
factors into r
e
s
pons
e
mo
del
b
y
prob
abi
lity d
e
n
sity f
unctio
n
s,
and c
a
lc
ulates
process
mea
n
s
and v
a
ri
ance
s
by
inte
grati
on. F
i
nally, it
obta
i
ns
opti
m
al par
ameter
co
mbi
n
ation by mo
del
opti
m
i
z
at
io
n. An e
m
p
i
rica
l st
udy
of the robust d
e
sig
n
of an
inductor-resistor s
e
ries circ
uit
is cond
ucted. T
h
e results
sh
ow
that the prop
os
ed
appr
oach
not
o
n
ly av
oids
the
disa
dvant
age
o
f
ignor
in
g in
ter
a
ctions
betw
e
e
n
factors w
hen
usin
g p
a
ra
met
e
r
sw
eep and se
nsitivity ana
lys
i
s, but
also overco
mes the shortco
m
i
ng
of only ach
i
evi
n
g
non-conti
n
u
o
u
s
opti
m
i
z
at
ion by
T
aguchi
meth
od an
d the li
mi
tation of
obta
i
n
i
ng l
o
cal o
p
ti
mi
z
a
ti
on by D
R
S
M
, and therefo
r
e
,
enh
anc
es the robustn
ess of the circuit outp
u
ts.
Ke
y
w
ords
: circuit opti
m
i
z
at
io
n, SVR, RPD, compl
e
x proce
sses
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
The paramet
er values of the circuit co
mpone
nt
s ra
n
domly fluctua
t
e over time d
u
ring the
lifespa
n of electro
n
ic p
r
od
ucts. The o
u
tput cha
r
a
c
teri
stics of circuit
,
su
ch a
s
voltage and
current,
also fluctu
ate
due to environmental n
o
i
s
e. The
r
ef
ore
,
the key to
prod
uct de
si
gn is to redu
ce
these fluctuat
ions and incr
ease the stability of circuit.
Traditio
nal m
e
thod
s of ci
rcuit
de
sign
in
clu
de p
a
ra
meter
sweep
[1] and se
nsitivity
analysi
s
[2].
Paramete
r swee
p examin
es the infl
ue
nce of the ch
ange of a sp
ecific comp
o
nent
value on
outp
u
ts by fixing the pa
ram
e
ter values
of oth
e
r
comp
one
n
t
s. The meth
od u
s
e
s
a o
n
e
-
factor
rotatio
n
de
sign
wh
ich n
eed
s m
o
re
run
s
a
n
d
ca
nnot ex
amine the
in
teractio
n effe
cts
betwe
en com
pone
nts. Sensitivity
analysis examine
s
the stability
of output voltage or cu
rre
n
t
throug
h the differential tran
sform
a
tion of certai
n
input feature
s
. Ho
wever, neither
of the methods
take
s into accou
n
t the influen
ce of sim
u
ltaneo
us
flu
c
tuation
s
of multiple com
pone
nt para
m
eter
values. Th
ere
f
ore, the outp
u
t stability of circuit after u
s
ing the
s
e
op
timization me
thods n
eed
s t
o
be improved.
In fact, how
to sele
ct app
rop
r
iate com
pone
nt para
m
eter value
s
whi
c
h ma
ke
s ci
rcuit
output chara
c
teri
stics in
sensit
ive to
compon
ent va
riation
s
and
environ
menta
l
noise ca
n
be
con
s
id
ere
d
a
s
a typical
ro
bust pa
ram
e
ter de
sig
n
(RPD) p
r
obl
em. Moreove
r
, the feature of t
he
relation
shi
p
betwee
n
circui
t inputs and outputs
is co
mplex nonlin
earity, not only because the
circuit con
s
i
s
ts of various compon
ents (e.g., re
sisto
r
, capa
citor, inducto
r, and powe
r) that have
very different
electri
c
al pe
rforma
nce, but also
be
ca
use the
r
e a
r
e rand
om flu
c
tuation
s
of the
para
m
eter val
ues d
u
rin
g
the lifespa
n of electroni
c pro
duct
s
.
Traditional robust design met
hods have limited capability to dea
l with such complex
pro
c
e
s
ses. T
agu
chi metho
d
can o
p
timize the pro
c
e
ss only at certai
n factor level
s
, but it fails to
gain continuo
us optimi
z
ati
on. Dual Respon
se
Surfa
c
e Methodol
og
y (DRSM
)
is appli
c
able to
the
optimizatio
n of the simple
processe
s
whi
c
h
ca
n b
e
fitted by s
e
co
nd-order
polynomial
s
.
Both
method
s ca
n
not effectively deal with the influen
ce
of the rando
m fluctuation
s
of compo
n
ent
para
m
eter val
ues o
n
the ou
tputs.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 3, March 20
13 : 1465 – 1
472
1466
This pap
er p
r
opo
se
s a Suppo
rt Vecto
r
Reg
r
essio
n
(SVR) base
d
RPD approach for
circuit optimization. The ap
proa
ch u
s
e
s
the followi
n
g
step
s to achi
eve the
optimization of circuit
comp
one
nt para
m
eter v
a
lue
s
. Firstly
,
it es
tablish
e
s an em
pirical mod
e
l of the comp
le
x
relation
shi
p
betwe
en ci
rcuit compo
n
e
n
t param
ete
r
values and
the outputs by using SVR.
Secon
d
ly, it
descri
b
e
s
the fluctuatio
n
s
of compone
nt paramete
r
values by
using no
rmal or
uniform p
r
ob
ability distrib
u
tions, and
descri
b
e
s
the influence o
f
compon
ent
paramete
r
value
fluctuation
s
a
nd process n
o
is
e on circuit
outputs by
u
s
ing
joi
n
t probability distri
butions. Finally, it
finds the parameter value
s
that
minimize ci
rcuit output fluctuatio
ns by using
parall
e
l gra
d
ient
desce
nt meth
od. In the foll
owin
g sectio
n
s
, the
p
ape
r f
i
rst reviews
e
x
tant RPD a
p
p
roa
c
h
e
s and
gives a
bri
e
f introd
uction
of SVR the
o
ry, and
th
e
n
de
scribe
s
the algo
rithm
step
s of th
e
approa
ch. After that, the p
aper dem
on
strates th
e e
ffectivene
ss of
the app
ro
ach by a case
study
of the optimization of indu
ctor-resi
s
to
r se
ries
circuit.
2. Brief Intro
duction to T
h
eories o
f
RPD
2.1. Taguchi
Metho
d
It was Tag
u
chi [3] who first introdu
ced t
he co
ncept of RPD. Tagu
chi method
cat
egori
z
e
s
process parameters into controllabl
e factors an
d noise factors. It incr
eases the stability of
pro
c
e
s
ses by
sele
cting ap
prop
riate
con
t
rollable fa
cto
r
levels that make q
uality characte
ri
stics
(i.e., “respon
se
s”) of pro
c
ess output
s i
n
se
nsitive to the variation
s
of controllabl
e factors and
the
influen
ce of noise facto
r
s. Signal-to
-
Noi
s
e Rati
o (S
NR) analy
s
is is one of the
main method
s to
achi
eve RP
D. The ba
si
c id
ea of this
me
thod is to
(1
)
plan a
nd run
experim
ents
by usin
g inn
e
r
-
outer array; (2) cal
c
ulate
SNR a
c
cordi
ng to t
he quality chara
c
teri
stics; and (3) find the optimal
controllabl
e factor l
e
vels t
hat minimize
the fluc
tuati
ons
of the resp
on
se
s. SNR i
s
calcul
ated
according to
the goal
s of
different type
s of qu
a
lity chara
c
te
risti
c
s. Tagu
chi me
thod cl
assifie
s
quality ch
ara
c
teri
stics into
three type
s:
the sm
a
ller
-
th
e-bette
r (STB
), the la
rge
r-t
he-b
e
tter
(LT
B
),
and the no
minal-th
e
-b
est (NTB) ch
aracteri
stics. It use
s
SNR
to reflect the robu
stne
ss of
respon
se
s, a
nd sea
r
ch for the factor level
combin
ation that maximize
s the
SNR, there
b
y
achi
eving RP
D.
It is relatively plain and
co
nvenien
ce to
sele
ct the o
p
timal factor
level com
b
in
ation by
usin
g SNR. But the metho
d
has m
any li
mitations
a
s
well. First, it requires i
n
formation ab
out
th
e
approximate rang
e of factor levels in advan
ce
and
needs man
y
runs to obtain a satisfied
s
o
lution [4]. Sec
o
nd, SNR los
e
s
a lot of informati
on
that is rel
a
te
d to pro
c
e
s
s
feature
s
[5]. The
third, it can o
p
timize only at certain fact
or leve
ls, whi
c
h lead
s to a satisfied sol
u
tion rather than
an optimal solution. Therefore, many schola
r
s a
n
d
experts pro
pose variou
s appro
a
che
s
to
improve
RP
D method
amo
ng which the
rep
r
e
s
ent
ati
v
e is the
DRSM pro
p
o
s
ed
by Vining a
nd
Myers
[6].
2.2. DRSM
DRSM [7, 8] i
n
vestigate
s
p
r
ocess o
p
tim
u
ms
step by
step by
sequ
entially adopt
ing first
-
orde
r polyno
m
ial modelin
g and stee
pe
st ascent
opti
m
ization, an
d
then fits pro
c
e
ss me
an a
nd
pro
c
e
ss va
ria
n
ce m
odel
s
by using
se
cond-order
p
o
l
y
nomial in a
relatively sm
all ran
ge of the
factors. After
that, it obtains an
optimal
solutio
n
by m
i
nimizin
g
the
varian
ce u
n
d
e
r the
co
nstraint
of mean ta
rg
et. Unlike Ta
guchi metho
d
, DRSM
su
ccessfully co
mbine
s
pa
ra
meter d
e
si
gn
and
reg
r
e
ssi
on a
nalysi
s
, and
obtain
s
re
gre
ssi
on
mo
del
betwe
en respon
se
s and
factors throu
g
h
experim
ental
desi
gn. It ca
n attain the
continuo
us
opt
imization
of factor level
s
, and i
s
the m
a
in
method for
RPD.
DRSM [9],
however, ha
s its limitati
ons
wh
en
it
is applied t
o
RPD
with compl
e
x
pro
c
e
s
ses. F
i
rstly, seco
n
d
-o
rde
r
polynomial
s
fail
to fit the complex nonlin
ear rel
a
tion
ship
betwe
en fact
ors an
d re
sp
onses. Seco
ndly, the result
of optimization is sen
s
i
t
ive to
the initial
values. It may obtain local
rather than
global optim
i
z
ation when i
napp
rop
r
iate
initial values are
sele
cted. Th
e third, DRS
M
doe
sn’t co
nsid
er the
i
n
fluen
ce of fa
ctor variatio
ns on re
sp
on
ses,
therefo
r
e, is n
o
t a real ro
bu
st para
m
eter
desi
gn.
2.3. SVR
SVR is a sm
all-sampl
e
ba
sed ap
proximate
statistical learni
ng a
ppro
a
ch pro
p
o
se
d by
Vapnik [10]. It can establi
s
h no
npa
ram
e
tric mo
del
s that meet the cha
r
a
c
teri
stics of pro
c
e
ss
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
SVR-ba
s
e
d
RPD Appro
a
ch
for Com
p
lex
Proces
se
s an
d Its Applicati
on … (Qi
nga
n Cui)
1467
unde
r the
co
nstrai
nt of sm
all sa
mple. It has
bee
n wi
d
e
ly use
d
to m
odel
compl
e
x pro
c
e
s
se
s [1
1-
13]. The ba
sic prin
cipl
es of
SVR are a
s
follows:
let
n
R
x
and
y
R
deno
te the input variabl
es
Vector a
nd
output varia
b
le of a pr
oce
s
s; sup
p
o
se that the
function
()
yf
x
is
unkno
wn, the task of model fitting is to us
e the
data from the indepe
nde
nt and identical
distrib
u
tion e
m
piri
cal sam
p
le S
11
2
2
{
(
,
)
,(
,
)
,
.
.
.
,(
,
)
}
ll
Sy
y
y
xx
x
,
,
ii
yy
xx
(1)
to find the op
timum functio
n
0
(,
)
f
x
in the fun
c
tion set
{(
,
)
}
f
x
whi
c
h
minimizes th
e exce
pted
risk of pre
d
ict
i
on:
(
)
(
,
(,
)
)
(,
)
RL
y
f
d
F
y
xx
(2)
whe
r
e
is the
gene
rali
zed
para
m
eter,
a
nd
(,
(
,
)
)
Ly
f
x
is the lo
ss-fun
ction
wh
ich d
e
fined
a
s
the
inse
nsitiv
e function:
0
(,
(
,
)
)
(
,
)
0,
(,
)
(,
)
,
(,
)
yf
y
f
yf
yf
yf
Lx
x
x
x
x
(3)
If the relation
ship
between
input
x
an
d o
u
tput
y
i
s
o
n
li
nearity,
x
i
s
fi
rst m
appe
d o
n
to a hi
gh lin
ear
dimen
s
ion
a
l feature
sp
ace
using
nonli
n
ear m
appin
g
function
T
(
x
), and then
a linear m
odel i
s
obtaine
d in the feature
spa
c
e:
(,
)
(
)
f
Tb
xw
w
x
. (4)
Then SVR m
odel fitting is formul
ated a
s
the followin
g
optimizatio
n probl
em:
2
*
,
1
*
*
1
mi
n
(
)
2
.(
(
)
)
((
)
)
,0
,
1
,
2
,
,
n
ii
b
i
ii
i
ii
i
ii
C
st
T
b
y
yT
b
il
w
w
wx
wx
(5)
Whe
r
e
i
,
*
i
is th
e non-neg
ative sla
ck vari
ab
les an
d C is t
he pen
alty paramete
r
.
The dual p
r
o
b
lem of equat
ion (6
) is:
*
,1
11
1
1
mi
n
(
)
(
)
(
,
)
2
()
(
)
.(
)
0
0
,
,
1
,
2
,
,
,
n
ii
jj
i
j
ij
ll
ii
i
i
i
ii
l
ii
i
i
i
aa
aa
k
aa
y
a
a
s
ta
a
a
a
C
i
l
xx
, (6)
with the sol
u
tion of
*
(,
)
aa
*T
11
(,
,
,
,
)
*
nn
aa
a
a
. (7)
In equation (7
),
(,
'
)
k
xx
=
()
(
)
TT
xx
'
is the kernel
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 3, March 20
13 : 1465 – 1
472
1468
function
whi
c
h ca
n redu
ce
the complex
of ope
rati
on
in the hi
gh di
mensi
on fe
ature
sp
ace after
mappin
g
. Fin
a
lly, the SVR
fitting model become
s
*
1
*
1
*
1
(,
)
(
)
(
)
()
(
,
)
()
(
,
)
n
ii
i
i
n
ii
i
i
n
ji
i
i
i
f
ba
a
k
b
aa
k
by
a
a
k
xw
x
x
x
wx
x
xx
(8)
3. SVR Base
d RPD
3.1. Basic Id
eas
For
RPD of
compl
e
x processe
s, esta
bl
ishin
g
th
e
emp
i
r
i
c
a
l mode
ls
o
f
pr
oc
ess
me
an
s
and va
rian
ce
s is on
e of t
he key ste
p
s. Specific
ally, the empi
ri
ca
l model
s n
e
e
d
to meet t
w
o
requi
rem
ents.
First, the m
odel
s sh
ould
reflect the
complex rel
a
tionship bet
we
en facto
r
s a
nd
respon
se
s in
the wh
ole ran
ge of facto
r
s. Seco
n
d
, the
sampl
e
ne
ed
ed for m
odeli
ng shoul
d be
as
small as p
o
ssible to redu
ce
the cost of optimizat
ion. T
herefo
r
e, SVR model
s tha
t
are suitable
fo
r
small
-
sample
global mo
del
ing ca
n be se
lected a
s
the
basi
c
form of
empiri
cal mo
dels. Mo
reov
er,
SVR model
s have analytic form
which help
s
the
con
s
eq
uent
comp
uting
and optimi
z
i
ng.
Therefore,
using SVR fo
r
modelin
g ha
s its natu
r
al
a
d
vantage fo
r
robu
st d
e
si
gn
of circuit
wit
h
rand
om p
a
ra
meter vari
atio
ns a
nd p
r
o
c
e
ss
noi
se. The
cha
r
a
c
teri
sti
cs
of SVR en
able u
s
to ad
opt
spa
c
e filling
design
s
su
ch as unifo
rm design
[1
4] with large interval and LHS design.
Con
s
e
quently
, not only the sampl
e
point
s sel
e
cte
d
ca
n cove
r the whole fea
s
ible
zon
e
, but also
the sampl
e
is relatively sm
all.
After modeli
n
g and
sa
mple
point
sele
cting ap
pr
o
a
ch
are
determin
ed, the next
step is to
con
s
id
er h
o
w to deal
with the fluctu
ation
s
of cont
rolla
ble facto
r
vari
ations
and
no
ise fa
ctors, b
o
th
of which influen
ce the resp
on
se out
puts. Ho
we
v
e
r, most of the existing
studie
s
lay their
empha
si
s on the influence
of noise facto
r
s. Little
effort has been put
to
examine the influence o
f
the variation
s
of controll
able fa
c
t
ors
[15, 16]. In f
a
c
t, the fluc
t
uation
s
of controlla
ble fa
ctor
variation
s
ha
ve certai
n sta
t
istical p
a
tterns (e.g., in a
circuit, the re
sista
n
ce of a
resi
sto
r
no
rm
ally
or unifo
rmly
and rand
oml
y
fluctuates
arou
nd it
s n
o
minal
valu
e
)
,
so we ca
n
use pro
babi
lity
den
sity functions
(e.g., multi-dime
nsio
nal unifo
rm distrib
u
tion a
nd multi-dim
ensi
onal n
o
rma
l
distrib
u
tion, etc.) to describ
e the fluctuations. In
addition, the chang
e of noise factors is typicall
y
norm
a
l distri
buted. The
r
ef
ore, in mod
e
l
i
ng, we
can first e
s
tablish the singl
e re
spo
n
se mod
e
ls
among p
r
o
c
e
ss
re
spon
se
s, cont
rollabl
e factors,
an
d noise fact
ors, an
d the
n
introdu
ce t
he
fluctuation
s
of controllabl
e factor va
riations
a
nd
noise facto
r
s into respon
se mo
del
s by
prob
ability de
nsity functio
n
,
and calculat
e pro
c
e
s
s
me
ans
and va
ria
n
ce
s by inte
g
r
ation. Th
e la
st
step is to find multiple minimum points o
f
proce
s
s variances by usi
ng method
s such a
s
parall
e
l
gradi
ent de
scent [17], and determi
ne th
e optimal pa
rameter
com
b
ination a
c
cord
ing to the value
of process m
ean
s.
3.2. Algorith
m
Steps of S
V
R Bas
e
d RPD Appr
oac
h
Based o
n
the
theoretical a
nalyse
s
ab
ove,
we pro
p
o
s
e the SVR-ba
s
ed
RPD for
circuit
optimizatio
n as follo
ws:
Step 1: select the
co
ntrollabl
e factors
'
12
,,
m
x
xx
x
and noise facto
r
s
'
12
,,
n
zz
z
z
, then decid
e the rang
e of each factor
b
y
prior kno
w
l
edge of the p
r
ocess;
Step 2: run
uniform de
si
gn or uniform grid desi
g
n experime
n
ts with singl
e
arrays of
controllabl
e factors an
d n
o
ise fa
ctors to get
raw d
a
ta, and then standardize the sampl
e
dat
a as
follows
:
11
1
,,
,
,
,
,
,
,
,
ll
l
l
l
l
Sy
y
X
Z
y
Y
xz
xz
x
z
. (9)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
SVR-ba
s
e
d
RPD Appro
a
ch
for Com
p
lex
Proces
se
s an
d Its Applicati
on … (Qi
nga
n Cui)
1469
Step 3: choo
se Gau
s
s function as the kernel fun
c
tion
and set appropriate pa
ra
meters of
C and
, and then fit the SVR model of p
r
oce
s
s output
*2
1
[
;
]
[
;]
,
[
;]
,
~
0
,
l
ii
i
i
yf
k
b
xz
xz
xz
(10
)
Step 4: set the probability densi
ty of controllable factor va
riations and noise factors
according to
prio
r kn
owl
e
d
ge.
Specifically,
assume the
controllabl
e factor
vari
atio
ns follow the
normal dist
ri
bution,
and let
x
pi
denotes the variation of
x
i
, then the prob
ability density functions of the controll
ab
le
fac
t
or variations
turn into
2
2
()
1
(
)
e
xp[
]
,
1
,
2
,
2
2
pi
i
ip
i
pi
pi
xx
gx
i
m
(11
)
Assu
me noi
se factors foll
ow the norm
a
l di
stributio
n
with the
followin
g
prob
a
b
ility
den
sity functions:
2
2
()
1
()
e
x
p
[
]
,
1
,
2
,
2
2
jj
jj
j
j
z
f
zj
n
(12
)
and a
s
sume
the inde
pe
nden
ce
both
among
different
cont
rol
l
able fa
ctors and
betwe
en cont
rollabl
e
facto
r
s and noise
factors,
the
mean, vari
a
n
ce
and m
e
an squa
re e
r
ror
(
MSE
) of the output
y
a
r
e d
enoted a
s
1
1
1
1
111
2
11
1
1
1
1
1
2
()
;
;
;
;
()
()
;
;
;
;
()
()
pm
p
m
n
n
n
p
p
m
p
m
nnn
p
Ey
y
g
x
x
g
x
x
f
z
f
z
d
d
Dy
y
E
y
g
x
x
g
x
x
f
z
f
z
d
d
MS
E
E
y
t
D
y
xz
xz
(13
)
Step 5: minimize
D
(
y
) or
MSE
(
y
) by using the con
c
u
rre
nt gradie
n
t descent algo
rithm
or geneti
c
alg
o
rithm to get the optimal le
vels of
contro
llable factors
and
therefore
reach the RPD
of the pro
c
ess.
4. Case Stud
y
In orde
r to demon
strate t
he effectiven
ess of
the ap
proa
ch
we p
r
opo
sed a
bov
e, a case
study of the
optimizatio
n
of indu
ct
or-re
sisto
r
seri
es
circuit is con
ducte
d in thi
s
se
ction. Fig
u
r
e 1
sho
w
s the ci
rcuit diag
ram,
whi
c
h in
clud
e
s
an ind
u
cto
r
L
,
a resi
st
or
R
and A
C
po
wer
with volta
g
e
V
and frequ
e
n
cy
f
.
Figure 1. Inductor-resi
sto
r
seri
es
circuit
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 3, March 20
13 : 1465 – 1
472
1470
The rel
a
tion
ship between t
he circuit out
put
I
and
L
,
R
,
V
,
f
follows
2
2
2
fL
R
V
I
(14
)
Acco
rdi
ng to the prio
r kn
o
w
led
ge, the controllabl
e factors a
r
e
x
1
(
L
)and
x
2
(
R
), and
the noise factors are
z
1
(
V
) and
z
2
(
f
),
with the ra
nge
s
12
[
2
m
H
,
6
m
H
],
[9
,
1
1
]
xx
,
12
[
9
5
V
,
105
V
]
,
[
45Hz
,
55Hz
]
zz
, All the
factors are st
anda
rdi
z
ed i
n
to [-1,1] before
optimizatio
n.
The goal of
the circuit op
timization is
to rea
c
h the
mean value
of
I
by 10A, and
mean
while,
minimize the fluctuation of
I
caused by the variation
s
of
x
1
,
x
2
, and
z
1
and
z
2
.
Firstly, we ru
n 100 times uniform de
sign
experiment
s to get the sample set of
S
, and
then fit the SVR model a
c
cording to ste
p
3 in se
ction
3.2:
100
*
1
,,
ii
i
i
yf
x
z
k
x
x
b
(15
)
whe
r
e
y
i
s
the output ch
aracteri
stic d
e
n
o
tes a
s
I
.
Secon
d
ly, we assu
me that the
distrib
u
tion of variation of
x
1
,
x
2
,which are den
ote
d
b
y
x
p
1
and
x
p
2
,
follow the n
o
rmal distri
buti
on as
2
2
()
1
(
)
e
xp[
]
2
2
0.01
1,
2
ip
i
ii
pi
pi
pi
pi
i
xp
gx
p
x
i
(16
)
and the di
stri
bution of
z
1
,
z
2
follow the
norm
a
l distri
b
u
tion as
2
2
1(
)
(
)
e
xp[
]
2
2
0.1
0
1,
2
ii
i
i
i
i
i
z
fz
u
i
(17
)
Then we get
E
(
y
),
D
(
y
)
and
MSE
as
follows
11
1
2
2
2
1
1
2
2
1
2
1
2
2
11
1
2
2
2
1
1
2
2
1
2
1
2
2
(
)
;
;
;0
;0
(
)
(
)
;
;
;0
;0
()
1
0
()
pp
p
p
pp
p
p
Ey
y
g
x
x
g
x
x
f
z
f
z
d
x
d
x
d
z
d
z
Dy
y
E
y
g
x
x
g
x
x
f
z
f
z
d
x
d
x
d
z
d
z
MS
E
E
y
D
y
(18
)
Thirdly, after
integratio
n, we get t
he m
e
sh pl
ot of the
relation
shi
p
of
E
(
y
),
D
(
y
) and
MSE
, shown i
n
Figure 2.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
SVR-ba
s
e
d
RPD Appro
a
ch
for Com
p
lex
Proces
se
s an
d Its Applicati
on … (Qi
nga
n Cui)
1471
-1
0
1
-1
0
1
9
10
11
x
1
E(
y)
x
2
-1
0
1
-1
0
1
-0.
0
5
0
0.
0
5
x
1
D(
y
)
x
2
-1
0
1
-1
0
1
-1
0
1
2
3
x
1
MSE
x
2
Figure 2. Surface of
E
(
y
),
D
(
y
) and
MSE
As can be
se
en from
Figu
re 2, within th
e ra
n
ge of
co
ntrollabl
e fact
ors, th
e variet
ies of
E
(
y
),
D
(
y
) an
d
MSE
are typically co
mpl
e
x, espe
cially for
MSE
, who has
several
local minimu
m.
If we optimize the circuit by using DRS
M
, it will fa
ll i
n
to a certain l
o
cal optimum
inevitably. While
by using the
prop
osed ap
p
r
oa
ch, all the
local o
p
timu
m can b
e
rea
c
he
d throu
g
h
parallel g
r
a
d
i
ent
sea
r
ch (a
s sh
own in Ta
ble
1), hen
ce we can g
e
t the global optimu
m
con
s
eq
uenti
a
lly.
Table 1. Mini
mum of
MSE
with co
rrespo
nding
E
(
y
),
D
(
y
),
x
1
, and
x
2
x
1
x
2
E
(
y
)
D
(
y
)
MS
E
0.7895
-0.2632
9.9429
0.000134
0.028916
0.8947
-0.3684
10.0054
0.000239
0.027638
0.6842
-0.1579
9.8688
0.003216
0.035053
0.1579
-0.2632
10.1868
0.002748
0.037414
Table 1 sho
w
s that
MSE
has several lo
cal optimum
s, the global optimum of wh
ich is
0.0276
38, where the
co
ntrollabl
e factors
x
1
=0.8
9
74 and
x
2
=-0.3684. After mappi
ng the
controllabl
e factors into their initial ran
ges, we
get the optimal value
s
of
the
indu
ctor-re
si
stor
seri
es cir
c
uit
:
L
=5.7
894m
H,
R
=9.6
316
Ω
, and the corre
s
po
ndin
g
I
=1
0.0054A .
5. Conclusio
n
This pap
er propo
se
s a SVR based RP
D approa
ch
for the robu
st para
m
eter de
sign of
circuit. The
a
ppro
a
ch ove
r
come
s the
shortcomin
g o
f
ignorin
g int
e
ra
ction
s
bet
wee
n
facto
r
s in
para
m
eter
swee
p and se
nsitivity analysis meth
od
s,
avoids the limitation of failure to achie
v
e
contin
uou
s o
p
timization b
y
using Tagu
chi method,
and preve
n
ts to obtain local optimizatio
n by
usin
g DRSM.
The pape
r d
e
scrib
e
s d
e
ta
iled appli
c
at
i
on step
s of the app
roa
c
h,
and take
s in
to
accou
n
t the influen
ce of input variatio
ns on
re
sp
o
n
se
s, whi
c
h
make the ex
perim
ents m
o
re
reali
s
tic and
the optimization step
s more re
as
on
able. The re
sults of
the ca
se study also
demon
strate the appli
c
abilit
y and the effectivene
ss
of the app
roa
c
h i
n
circuit opti
m
ization.
Ackn
o
w
l
e
dg
ement
This research
was
sup
port
ed by the Nat
i
onal Natural
Scien
c
e Fo
un
dation of Chi
na
unde
r Grant 7117
1180
an
d Grant 7
127
2225.
Referen
ces
[1]
Cristian Mateos, Elina Pacini
, Carlos García Garino. An ACO-
inspire
d
algorithm for
minimizin
g
w
e
ig
hted fl
o
w
ti
me in cl
ou
d-ba
sed p
a
rameter
s
w
e
e
p
e
x
p
e
ri
ments.
Advanc
es in E
ngi
ne
eri
ng Softw
are
.
201
3; 56: 38-5
0
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 3, March 20
13 : 1465 – 1
472
1472
[2]
V Gregor
y W
e
i
r
s, James R Kamm, Laura P
S
w
i
l
er, et
a
l
. Sensitivit
y a
n
a
l
ysis techni
qu
es
app
lie
d to a
s
y
stem of h
y
p
e
r
bolic c
onserv
a
tion la
w
s
.
R
e
li
a
b
ility En
gin
eeri
ng & System S
a
fety
. 2012; 10
7: 157-1
70.
[3]
T
aguchi G. S
y
stem of experi
m
ental des
ign:
E
ngine
erin
g methods to op
timize qua
lit
y
and min
i
mise
cost.
W
h
ite Plains
, Ne
w
York:
UNIPUB/Kraus
Internation
a
l.
198
7.
[4]
P Angelo
p
o
u
lo
s and C Kouko
u
vin
o
s. Some r
obust param
eter desig
ns from orthogon
al a
rra
y
s
.
Journal
of Appli
ed Stati
s
tics
. 2008; 35(
12): 139
9–
140
8.
[5]
Nair VN, Shoe
maker AC.
T
he role of expe
riment
ati
on in qua
lit
y
e
n
g
i
ne
erin
g: a revie
w
of
T
aguchi’s
contrib
u
tion.
D
e
sig
nan
d an
aly
s
is of industri
a
l
experi
m
ents
. NY: Marcel De
kker. 1990; 2
4
7–2
77.
[6]
GG Vining, M
y
ers, RH. Com
b
ini
ng T
aguc
hi
and
Re
sp
ons
e Surface P
h
il
osop
hies: A D
ual R
e
sp
ons
e
Appro
a
ch.
Jour
nal of qu
al
ity techno
logy
. 1
990
; 22(1): 38-45.
[7]
AB Shaib
u
, B
y
ung Ra
e Ch
o. Another vie
w
of dual respo
n
se surface m
ode
lin
g and o
p
timizati
on i
n
robust par
amet
er desi
gn.
Int J Adv Manuf T
e
chno
l
. 200
9; 41: 631–
64.
[8]
Sangm
un Shi
n
, Ngu
y
en Kho
a
Viet
T
r
uong,
Byun
g Rae C
ho, Sung Ho
o
n
Hong. Dev
e
l
opme
n
t of a
sequ
enti
a
l rob
u
st–toler
anc
e desi
gn mode
l
fo
r a destructive qual
it
y
character
i
stic.
Co
mp
uters &
Industria
l Engi
neer
ing
. 2
011;
60(4): 77
7-7
8
9
.
[9]
Cui Qing'
a
n
, He Z
hen, Cui Nan. SVM-base
d
RSM model fitting appr
oac
h.
Journal of Mana
ge
me
n
t
Scienc
es in Ch
ina
. 20
08; 11(
1
)
: 31-41.
[10]
Vapn
ik VN. T
h
e nature of stat
istical le
arn
i
ng
theor
y. Ne
w
Y
o
rk: Spring
er-V
erla
g. 199
5.
[11]
J Acevedo-R
o
dríguez, S Mal
don
ad
o-Bascó
n, S Laf
uente-
Arro
yo, et al.
Comp
utation
a
l
loa
d
reducti
on
in dec
isio
n fun
c
tions usi
ng su
pport vector m
a
chi
nes.
Sig
nal
Processin
g
. 2009: 89(
10): 20
66-2
071.
[12]
Jenn
y T
e
rzic,
CR Nag
a
raj
ah,
Muhammad Al
amgi
r. F
l
uid le
vel measur
em
ent
in d
y
namic
environm
ents
usin
g a singl
e
ultrasonic se
nsor and
Su
p
port Vector Machi
ne (SVM).
Sensors and
Actuators A:
Physical
. 201
0; 161(1
–2): 27
8
-
287.
[13]
Jae Pil H
w
a
n
g
, Seongk
eun
Park, Euntai Kim.
A ne
w
w
e
i
ghte
d
ap
proac
h to imbala
n
ced d
a
t
a
classificati
on p
r
obl
em via su
p
port vector
mach
i
n
e
w
i
th
qu
ad
ra
tic cost function.
Expert S
ystem
s with
Appl
icatio
ns
. 2
011; 38(
7): 858
0-85
85.
[14]
F
ang KT
and DKJ Lin. U
n
ifo
rm desig
n
in c
o
mputer a
nd
p
h
y
s
i
cal
e
x
p
e
ri
ments,
in
T
he Grammar of
T
e
chno
logy D
e
velo
p
m
ent
. 20
08: 105-
12
5.
[15]
H Evange
lar
a
s
,
E Kolaiti, C
Kouko
u
vin
o
s. Robust par
amet
er desig
n: Optimizatio
n
of combin
ed arra
y
appr
oach w
i
t
h
orthog
on
al
arra
ys.
Jour
nal of
Statistica
l Pla
n
n
in
g an
dInfere
n
ce
. 200
6; 369
8-37
09.
[16]
Pisvimol Ch
at
sirirun
g
ru
ang,
Masami Mi
yaka
w
a
. App
lic
ation of gen
e
t
ic algorithm
to numerica
l
exper
iment in robustp
aramet
er desig
n
for sign
al
multi-re
spons
e pr
obl
e
m
.
Internatio
n
a
l Jo
urn
a
l of
Mana
ge
me
nt Scienc
e an
d En
gin
eeri
ng Ma
n
age
ment
. 200
9
;
4(1): 49-59.
[17]
Cui Qi
ng'
an.
Globa
l optim
iz
at
ion
of proc
e
ss param
eters
for mult
i-e
x
tr
eme q
ual
it
y
c
haracter
i
stics.
Journ
a
l of Man
age
ment Scie
n
c
es in Ch
ina
. 2
012; 15(
9): 46-
57.
Evaluation Warning : The document was created with Spire.PDF for Python.