Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
8
, No
.
6
,
Decem
ber
201
8
, p
p.
4810
~
4822
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v8
i
6
.
pp4810
-
48
22
4810
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
A
N
ovel Neur
ogli
al
A
rchitectu
re fo
r
M
odell
ing
Sin
gu
lar
Perturb
atio
n S
ystem
Sa
mi
a
S
alah,
M’hame
d H
adj
S
adok
, Ab
derrez
ak
Gue
sso
um
Depa
rt
m
ent
o
f
e
l
ec
tron
ic
s,
Unive
rsit
y
Saad
Dahl
a
b
Bli
d
a1,
Alger
i
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ja
n
23
, 2
01
8
Re
vised
Ju
l
5
,
201
8
Accepte
d
J
ul
29
, 2
01
8
Thi
s
work
deve
lops
a
new
m
o
dula
r
arc
h
it
e
ct
ur
e
tha
t
emula
te
s
a
rec
entl
y
-
discove
red
biol
o
gic
a
l
par
adi
gm
.
I
t
origi
n
ates
from
the
hum
an
bra
in
where
th
e
informati
on
flo
ws
al
ong
two
di
ffe
ren
t
pa
t
hwa
y
s
and
is
proc
essed
al
ong
two
ti
m
e
sca
le
s: on
e is a
f
ast
neur
al
n
e
twork
(NN
)
an
d
the
o
the
r
is a
sl
ow ne
twork
ca
l
le
d
th
e
glial
net
work
(GN
).
It
was
found
tha
t
th
e
neur
al
net
work
is
powere
d
an
d
c
ontrol
le
d
b
y
th
e
glial
n
et
work
.
Based
on
our
biol
ogi
cal
knowledge
of
g
li
al
c
el
ls
and
th
e
powerful
con
c
ept
of
m
odula
rity
,
a
nove
l
appr
oac
h
call
ed
a
rti
f
ic
i
al
n
eur
ogl
ia
l
Ne
twork
(AN
GN
)
was
designe
d
and
an
al
gorit
hm
base
d
on
diffe
ren
t
con
ce
pts
of
m
odula
r
ity
w
as
al
so
dev
el
oped
.
The
implementa
t
ion
is
base
d
on
the
noti
on
of
m
ult
i
-
ti
m
e
sca
le
s
y
st
ems
.
Vali
da
ti
on
is
p
e
rform
ed
through
an
as
y
nchr
onou
s
m
ac
hine
(AS
M)
m
odel
ed
in
the
stand
ard
singula
rl
y
per
turbe
d
form
.
W
e
apply
the
geometri
c
al
appr
oac
h
,
b
ase
d
on
Gerschgori
n’
s
ci
rc
le
the
or
em
(GCT),
to
sepa
r
at
e
the
fas
t
and
slow
var
ia
b
le
s,
as
wel
l
as
t
he
sin
gul
ar
p
erturbat
ion
m
et
hod
(SP
M)
to
det
ermine
th
e
r
educ
ed
m
odel
s.
Thi
s
new
arc
h
itect
ur
e
m
ake
s
it
poss
ibl
e
to
obta
in
sm
al
le
r
n
et
works
with le
s
s c
om
ple
xity
an
d
bett
er
per
form
anc
e
.
Ke
yw
or
d:
Ar
ti
fici
al
ne
uro
glial
n
et
w
ork
Async
hro
nous
m
achine
Ger
sc
hgori
n’s
ci
rcle
t
heorem
Glia
l netw
ork
Singular
p
e
rtu
r
bation
m
et
ho
d
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Sam
ia
Salah
,
Dep
a
rtm
ent
of
E
le
ct
ro
nics
,
Un
i
ver
sit
y Saa
d Dah
la
b
Bl
ida
1,
Soum
aa
Road
, B
P 270, Bli
da
,
Alge
ria
,
(+
213)
25 43 3
8 50
Em
a
il
:
sal
ah_
s
a
m
ia
@yaho
o.f
r
1.
INTROD
U
CTION
The
la
st
few
ye
ars
ha
ve
witn
essed
a
trem
en
dous
gr
ow
t
h
in
the
fiel
d
of
in
te
ll
igent
syst
e
m
s
.
In
sp
i
re
d
by
bio
lo
gical
ne
ur
al
net
works
,
on
e
s
uch
s
uc
cess
has
bee
n
achieve
d
in
ev
olu
ti
on
of
arti
f
ic
ia
l
neu
ral
networks
(ANNs).
A
N
Ns
are
c
har
a
ct
erized
by
their
disti
nctiv
e
capab
il
it
ie
s
of
e
xh
i
biti
ng
m
assive
par
al
le
li
s
m
,
gen
e
rali
zat
ion
ab
il
it
y
and
bei
ng
good
f
unct
ion
a
ppr
oxim
a
t
or
s
.
This
re
nd
e
rs
them
us
efu
l
for
so
l
ving
a
va
riet
y
of
pro
blem
s
in
patte
r
n
r
eco
gnit
ion
,
pre
dicti
on,
op
ti
m
iz
ati
on
an
d
a
sso
ci
a
ti
ve
m
e
m
or
y
[1]
,
[
2]
.
A
dd
it
ion
al
ly
,
they
are
al
s
o b
ei
ng
em
plo
ye
d i
n
syst
em
m
od
el
ing
a
nd contr
ol
[
3]
,
[4]
.
These
A
NN
s
,
eff
ic
ie
nt
in
nu
m
ero
us
ap
plica
ti
on
s,
a
re
not
a
s
well
su
it
e
d
f
or
ap
pro
xim
ating
non
-
li
ne
ar
and
high
-
dim
e
ns
io
nal
f
un
ct
io
ns
wit
h
m
ulti
p
le
tim
e
dynam
ic
s
li
ke
the
on
es
in
sin
gu
la
r
per
t
urbati
on
syst
e
m
s
(S
PS
s
)
wh
ic
h
increase
s
the
di
ff
ic
ulti
es
in
sy
stem
m
od
el
ing,
analy
sis
an
d
con
t
ro
ll
er
desi
gn
.An
e
ff
ect
iv
e
way
to
overc
om
e
t
his
pro
blem
is
to
sepa
rate
th
e
or
igi
nal
syst
e
m
sta
te
s
into
su
bsy
ste
m
s
that
change
ra
pidl
y
and
tho
se
that
var
y
slow
ly
on the c
ho
s
en
tim
e scal
e, u
si
ng sin
gu
l
arly
p
ert
urbati
on m
et
ho
d (SP
M).
So
m
e
recent
re
search
res
ults
us
in
g
t
he
(
SP
M)
to
a
naly
ze
and
co
ntr
ol
the
SPSs
are
pu
blished
in
[5]
,
[6]
.
H
ow
e
ver,
accurate
a
nd
fa
it
hf
ul
m
at
he
m
a
ti
cal
m
od
el
s
for
th
os
e
syst
em
s
are
usual
ly
diff
ic
ult
to
obta
in
du
e
to
th
e
unce
rtai
nties
an
d
no
nlinearit
ie
s.
In
this
case,
adequate
syst
e
m
identific
at
ion
be
com
es
i
m
po
rtant
a
nd
necessa
ry,
befor
e
a sin
gula
r pert
urbati
on
th
eor
y
-
base
d
c
on
trol sc
hem
e can be
desig
ne
d.
R
ecentl
y
,
resea
rch
us
i
ng
m
ulti
tim
es
scal
e
ne
ur
al
netw
orks
hav
e
been
pr
opose
d
in
li
te
rat
ur
e
to
so
l
ve
the
syst
e
m
identific
at
ion
pro
blem
of
the
nonlinea
r
SP
Ss.
Am
on
g
them
there
are
m
ulti
-
tim
e
-
scal
e
dynam
ic
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A Novel N
e
ur
ogli
al Arc
hitec
ture f
or
M
odel
li
ng S
i
ngular
Pe
rturbati
on
Syst
em
(
Sami
a S
ala
h)
4811
neural
netw
ork
(D
N
N
)
pro
po
s
ed
in
[7]
.
O
r
re
current
ne
ural
netw
ork
(RN
N
)
pro
po
se
d
in
[
8]
.
In
these
pa
per
s
,
trai
ning
m
et
ho
ds
are
base
d
on
a
gra
dient
de
scent
up
dating
al
gorithm
with
fixe
d
“l
ear
ning
gai
n”,
s
uc
h
a
s
bac
k
pro
pag
at
i
on
(B
P)
a
nd
RN
N
al
gorithm
s.
The
m
ai
n
dr
a
w
bac
k
of
these
trai
ni
ng
m
et
hods
is
that
the
c
onve
r
gen
c
e
sp
ee
d
is
us
uall
y
ver
y
sl
ow.
T
o
acce
le
rate
t
he
trai
ning
proc
ess
,r
e
searc
hers
inv
e
sti
gated
t
he
e
xten
ded
K
alm
an
filt
er
(E
KF)
ba
sed
trai
ning
m
et
hods
for
NN
in
[
9]
.
T
he
t
he
or
et
ic
al
analy
si
s
of
E
KF
base
d
trai
ning
al
go
rith
m
requires
t
he
m
od
el
ing
unce
rtai
nty
of
th
e
NN
t
o
be
a
Gau
s
sia
n
proc
ess,
w
hich
m
ay
no
t
be
t
ru
e
in
real
app
li
cat
io
ns
.
So
m
e
oth
er
res
earche
rs
al
s
o
st
ud
ie
d
op
ti
m
al
bounde
d
el
li
psoid
(O
B
E)
al
gorithm
-
base
d
l
earn
i
ng
la
ws
f
or NN
[10
]
-
[
12]
. A
ll
of
these m
et
ho
ds
are c
om
plex
an
d
c
om
pu
ta
ti
onal
ly
intensive.
In
this
pa
per
,
we
pro
pose
a
new
m
ulti
t
i
m
e
-
scal
e
NN
arch
it
ect
ur
e
cal
le
d
"arti
fici
al
neu
r
ogli
al
netw
ork"
(
A
N
GN)
ba
sed
on
the
powe
rful
con
ce
pt
of
"
m
od
ularit
y"
to
so
lve
t
he
pr
ob
le
m
s
of
sin
gu
la
r
per
t
urbati
on
syst
e
m
trai
nin
g.
The
basic
idea
is
to
us
e
t
he
know
le
dg
e
ab
out
the
ne
rvo
us
sy
stem
and
the
hum
an
br
ai
n,
w
her
e
th
e
inform
ation
f
lows
al
ong
tw
o
dif
fer
e
nt
pathw
ay
s
an
d
is
proces
sed
al
ong
two
-
ti
m
e
scales:
one
is
a
fast
-
neural
netw
ork
(NN)
and
the o
the
r
is
a
slo
w
net
wor
k
cal
le
d
the g
li
al
network
(
G
N)
.
It w
as
f
ound
tha
t
the n
e
ural
n
et
work is
powe
r
ed
a
nd contr
olled
by the
glial
netw
ork
[13]
,
[
14
]
.
In
our
ex
pe
rim
ent,
for
a
giv
e
n
ap
plica
ti
on
,
de
pendin
g
on
the
c
om
plexit
y
and
the
ph
ysi
cal
char
act
e
risti
cs
of
the
pro
blem
,
we
d
ivide
our
gl
ob
al
m
od
el
i
nto
tw
o
s
ub
-
m
od
el
s:
slo
w
a
nd
fast
on
e
s
us
ing
th
e
singular
pe
rtu
rb
at
io
n
m
et
hod
(S
PM
).
T
he
first
dif
ficult
y
that
arises
wh
e
n
dec
oupl
ing
va
riables
is
the
identific
at
ion
of
bo
t
h
the
sl
ow
a
nd
fast
m
od
el
var
ia
bl
es.
The
s
olu
ti
on
is
based
o
n
Ge
rsc
hgor
i
n’s
ci
rcle
geo
m
et
ric
theo
rem
(G
CT)
[
15]
.
This
te
c
hniqu
e
m
akes
it
po
s
sible
to
l
oc
at
e
the
ei
genv
al
ues
in
t
he
co
m
plex
plane
within
gro
ups
of
ci
rcle
s.
The
gro
u
pi
ng
of
the
m
od
e
s
is
i
m
m
ediat
e
wh
e
ne
ver
ci
rc
le
s
are
disjoint
,
an
d
afterwa
r
d,
t
he nu
m
ber
of slo
w
a
nd f
ast
m
odes is dete
rm
ine
d.
Vali
dation
of
the
pro
posed
appr
oach
is
c
arr
ie
d
out
on
the
AS
M
m
od
el
,
unde
r
th
e
singularly
per
t
urbed
sta
ndar
d
f
orm
.
Su
bse
quently
,
an
a
lgorit
hm
is
ad
op
te
d
to
te
st
th
e
eff
ect
ive
ness
and
perf
or
m
ance
of
the
pro
posed
ANG
N.
T
his
ne
w
arc
hitec
tur
e
has
m
ade
it
po
s
sible
to
ob
t
ai
n
netw
orks
of
co
ns
ide
ra
bly
sm
a
ll
er
siz
e
with
sim
ple
structu
res
w
hich
ha
ve
a
str
ong
nonline
ar
appr
ox
im
at
ion
capa
bili
ty
and
w
hich
e
na
bles
it
to
m
od
el
nonline
ar
sin
gula
rly
pe
rturbe
d
syst
e
m
s
m
or
e
accur
at
el
y
with
le
ss
com
pu
ta
ti
on
c
om
plexity
,
co
m
par
ed
to the c
onve
nti
on
al
ne
ur
al
net
work m
od
el
.
2.
S
INGUL
AR P
ERTU
RBATI
ON MET
HO
D
This
m
et
ho
d
i
s
us
e
d
f
or
m
ulti
-
tim
e
scal
e
syst
e
m
s
that
can
be
re
duced
to
the
sta
nd
a
rd
form
of
equ
at
io
n
(
3)
by
the
deter
m
inati
on
of
the
pa
rasit
ic
te
r
m
ε
.
Con
si
der
the
sta
te
m
od
el
o
f
a
li
near
syst
e
m
of
dim
ension
n:
{
̇
=
+
=
}
(1)
Ev
olv
in
g
ac
co
r
di
ng
t
o
two
-
ti
m
e
scal
es,
it
c
an
be
dec
ouple
d
i
nto
t
wo
slo
w
an
d
fast
su
bsy
stem
s.
The
sta
te
vector
X
con
ta
in
s
al
l
the
sta
te
var
ia
bles
cor
re
spo
nd
i
ng
to
the
dyna
m
ic
e
lem
ents.
If
x
is
the
set
of
sta
te
var
ia
bles of sl
ow ele
m
ents, and
z is
the set
of
f
ast
elem
ents
the m
od
el
is
wri
tt
en
,
2
1
22
21
12
11
U
B
B
z
x
A
A
A
A
z
x
(2)
z
x
C
C
y
,
2
1
with
:
-
(
0
)
=
0
,
(
0
)
=
0
-
2
22
21
,
,
B
A
A
-
are
ver
y l
a
rge com
par
ed
to
1
12
11
,
,
B
A
A
.
We
intr
oduce
t
he
pa
ram
et
er
ε
to
norm
al
iz
e
ou
r
m
od
el
,
we
wr
it
e:
21
*
21
A
A
,
22
*
22
A
A
,
.
2
*
2
B
B
can
be give
n by:
0
12
0
1
22
L
A
A
A
with
:
21
1
22
0
A
A
L
an
d
0
1
12
11
0
L
A
A
A
By
assum
ing
that
m
at
rix
A
22
is
inv
e
rtible
,
the
sta
te
e
qu
at
ion
i
n
t
he
sta
ndar
d
si
ngularl
y
per
t
urbe
d
form
w
it
h
ε as
the p
e
rtu
rb
at
i
on
par
am
et
er is then
w
ritt
en
as
:
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4810
-
4822
4812
,
*
2
1
*
22
*
21
12
11
U
B
B
z
x
A
A
A
A
z
x
(
3)
2.1.
Slow a
nd fa
st re
duced
mo
de
ls
The
sl
ow r
e
du
ced m
od
el
is
de
te
rm
ined
fro
m
eq
. (
3)
by conside
rin
g
that
:
0
s
s
s
s
s
s
s
s
s
s
s
s
s
u
B
x
A
A
z
u
D
x
C
y
u
B
x
A
x
2
21
1
22
(
4)
wh
e
re
s
s
s
s
y
u
z
x
,
,
,
are t
he sl
ow
com
pone
nts
of
t
he varia
bles
y
u
z
x
,
,
,
resp
ect
ive
ly
,
with
:
2
1
22
2
21
1
22
2
1
2
1
22
12
1
21
1
22
12
11
B
A
C
D
A
A
C
C
C
B
A
A
B
B
A
A
A
A
A
s
s
s
s
(
5)
with:
0
0
x
t
x
s
.T
he
init
ia
l value
of the
slow
com
pone
nts
s
z
is
:
0
12
1
22
0
t
x
A
A
t
z
s
s
wh
ic
h
is
ge
neral
ly
diff
ere
nt
f
rom
0
z
.
The
fast
var
ia
bles
z
cannot
the
refor
e
be
ap
pro
xim
a
te
d
by
s
z
in
th
e
tim
e
interval
,
0
T
.
We
int
rod
uce
the
c
orrecti
ve
t
erm
f
z
,
de
fine
d
by
:
s
f
z
z
z
wh
ic
h
repr
esents
ra
pi
d
changes
in
z
.Th
e
refor
e
, th
e
bo
unda
ry lay
er e
quat
ion, e
xpress
ed
in
the
dilat
ed
ti
m
e
fo
ll
ows:
(
dt
dz
f
d
dz
f
):
(
6)
The
fast re
duc
ed
m
od
el
is the
n wr
it
te
n:
0
21
1
22
0
0
2
2
22
x
A
A
z
t
z
z
C
y
u
B
z
A
d
dz
f
f
f
f
f
f
(
7)
3.
I
DENTIF
IC
A
TION
OF TH
E GEOMET
R
IC DYN
AMI
CS
Sett
ing
the
pre
vious
sta
ndar
d
fo
rm
assum
es
:
a)
Kno
wled
ge
of
the
ei
ge
nv
al
ues
to
dete
r
m
ine
the
siz
e
of the sl
ow and
f
ast
eige
nvect
or
s
.
b)
A suit
a
ble gr
ouping
of slo
w
m
od
es
a
nd f
ast
m
od
es
.
Our
at
te
ntio
n
is
f
ocu
s
ed
on
geo
m
et
rical
m
e
tho
ds
incl
ud
i
ng
the
ci
r
cl
es
of
ge
rsc
hgori
n.
T
he
local
iz
at
ion
of
the
ei
g
e
nval
ue
s
on
t
he
c
om
plex
plane
m
akes
it
possible
to
put
a
syst
em
i
n
t
he
sta
nd
a
rd
for
m
without
ha
ving
to
cal
culat
e
t
hese
ei
ge
nv
al
ue
s.
In
the
case
of
GCT,
t
he
gro
up
i
ng
of
m
od
e
s
is
i
m
m
ediat
e
as
so
on
as
the
ei
genvalue
s
are
ci
rcu
m
scribed
in
dis
joint
ci
rc
le
s
.
The
ge
ome
tric
m
e
tho
d
ba
sed
on
GCT
for
t
he
sel
ect
ion
a
nd s
epar
at
io
n of di
ff
e
ren
t t
im
e scal
es is re
pr
ese
nt
ed
as
fo
ll
ows:
3.1.
Gersch
go
ri
n
’s
circ
le
s t
he
ore
m (
G
CT)
No
ti
ng
n
j
i
a
ij
.
..
.
.
1
,
,
as the el
e
m
ents o
f
the
s
ta
te
m
at
rix
A
,
,
are
expre
ssed
b
y
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A Novel N
e
ur
ogli
al Arc
hitec
ture f
or
M
odel
li
ng S
i
ngular
Pe
rturbati
on
Syst
em
(
Sami
a S
ala
h)
4813
,
1
n
i
j
j
ij
i
a
p
n
i
.,
,
..
.
..
.
..
.
2
,
1
(8)
,
1
n
i
j
i
ij
i
a
Q
n
j
.,
,
..
.
..
.
.
..
2
,
1
(9)
Geo
m
et
rical
se
par
at
io
n
of
the
diff
e
ren
t
dyna
m
ic
m
od
es
is
base
d
on
the
a
pp
li
cat
io
n
of
the
f
ollow
i
ng
two
the
or
em
s.
These
the
or
e
m
s
ob
ta
ined
f
ro
m
Ger
sch
go
rin
gi
v
e
a
loc
al
iz
at
ion
of
th
e
e
igen
values
on
t
he
com
plex
pla
ne
.
3.1.1.
Theorem
1
All
the
ei
genv
al
ues
of
a
m
atr
ix
of
ar
bitra
r
y
ran
k
n,
a
re
con
ta
ine
d
in
n
ci
rcle
bundle
s
centere
d
at
11
,
22
,
…
.
.
,
and
rad
ii
1
,
2
,
…
.
.
,
fo
r
the
li
nes
or
1
,
2
,
…
.
.
,
f
or
the
c
olu
m
ns
,
w
hich
ar
e
ob
ta
ine
d by s
um
m
ing
the m
od
ules
of the
off
-
dia
gonal term
s appeari
ng in
t
he
sam
e li
ne
or
colum
n:
=
=
.
3.1.2.
Theorem
2
Wh
e
n
a
gro
up
of
k
li
ne
-
ci
rcle
s
(o
r
k
colum
n
-
ci
rcles)
is
co
m
ple
te
ly
disj
oi
nt
from
the
othe
r
ci
rcles,
it
con
ta
in
s
k
ei
ge
nv
al
ues
[
16
]
.
Wh
e
n
a
gro
up
of
k
ci
rcles
is
com
plete
ly
disj
oi
nt
from
the
oth
e
r
ci
rcles,
it
can
be
sai
d
that t
he
syst
e
m
then
h
as at l
east
two
-
ti
m
e
scal
es. W
he
ther
this g
rou
p
of
circl
es is to the r
ig
ht o
r
t
o
the left
of
the
oth
e
r
ci
r
cl
es,
we
can
de
te
rm
ine
the
k
s
low
m
od
es
co
r
respo
nd
i
ng
t
o
these
k
ci
rcles
or
resp
ect
ively
the
k
fast
m
od
es.
Ea
ch
ci
rcle
re
pr
es
ents
a
sta
te
of
t
he
syst
em
.
It
is
then
po
s
sible
to
gi
ve
an
a
de
quat
e
pa
rtit
ion
of
the
m
od
el
.
Gen
e
r
al
ly
,
this
dire
ct
m
et
ho
d
does
not
m
ake
i
t
po
ssi
ble
to
con
cl
ud
e
im
mediat
el
y
in
al
l
cases.
Daup
hin
-
Tan
guy
[
17]
then
pr
opos
es
the
us
e
of tran
sf
or
m
at
i
on
s:
)
1
,....,
1
,
,
1
,...,
1
(
k
k
d
i
a
g
S
,
n
k
,
,
2
,
1
;
these
par
am
et
ers
al
low
the
va
riat
ion
of
the
ci
rcles
siz
e
a
nd
thei
r
optim
iz
at
ion
l
ead
to
ci
rcles
of
m
i
nim
u
m
rad
ii
.
H
ow
e
ver, this m
et
ho
d d
oes n
ot sep
a
rat
e all
syst
e
m
s.
3.1.3.
Chan
ging th
e
radius si
z
e
Let
the m
at
rix
be:
)
1
,....,
1
,
,
1
,...,
1
(
k
k
d
i
a
g
S
,
n
k
,
,
2
,
1
(
10)
The
c
hange
of
base
X
S
X
k
'
le
ads
to
a
ne
w
sta
te
m
at
rix.
The
rad
ii
k
R
1
an
d
ck
R
becom
e
k
k
R
1
and
k
ck
R
res
pecti
vely
[18]
.
I
f
the
oper
at
io
n
is
re
peated
severa
l
tim
e
s,
the
a
ggre
ga
te
d
trans
form
ation
is:
1
S
A
S
A
X
S
X
with:
k
k
S
S
(
11)
If
t
her
e
are
t
wo d
is
j
oi
n
t set
s
of circl
es, th
en
th
e p
e
rm
utati
on
m
at
rix
is:
1
P
A
P
A
X
P
X
(
12)
3.1.4.
Movin
g Circ
le
C
en
ters
In
or
der
to
im
pr
ov
e
the
s
epar
at
io
n
of
dynam
ic
s,
it
is
so
m
e
tim
es
necessary
to
introdu
ce
a
disp
la
cem
ent of the
circl
es
whic
h
is c
har
act
er
iz
ed
by t
he foll
ow
i
ng tra
ns
f
orm
at
ion
[
18
]
:
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4810
-
4822
4814
ij
l
n
l
J
B
I
T
.
.
(
13)
On
ly
the
el
e
m
ents
of
li
ne
i
and
colum
n
j
cha
nge,
the
centers
of
the
ci
rcles
i
and
j
are
sh
ifte
d
f
ro
m
ii
a
and
jj
a
to
ji
l
ii
a
B
a
and
ji
l
jj
a
B
a
,
res
pecti
vely
.
The
cho
ic
e
of
l
B
can
be
m
ade
in
su
ch
a
way
that
0
)
(
2
ji
l
ii
jj
l
ij
ij
a
B
a
a
B
a
X
If
se
ver
al
ci
r
cl
es
intersect
,
the
te
rm
s
,...
2
,
1
l
B
l
are
cal
culat
ed
in
th
e
sam
e w
ay
, s
o
t
he
final tra
ns
f
orm
ation
is:
1
T
A
T
A
X
T
X
with:
T
l
T
,
(
14)
If
t
wo gr
oups
of circl
es a
re
dis
j
oi
nt,
the
perm
utati
on
m
at
rix
P
is again:
1
P
A
P
A
X
P
X
(15)
4.
CONCE
PT
O
F MOD
ULA
R
ITY
The
ap
plica
ti
on
of
co
nce
pt
of
m
od
ularit
y
to
def
ine
t
he
ne
w
arc
hitec
ture
of
sm
al
l
arti
fic
ia
l
netwo
r
ks
involves
the
fol
lowing
four st
eps:
4.1.
The dec
ompos
ition
The
dec
om
po
s
it
ion
of
a
ta
s
k
into
subta
s
ks
is
the
first
ste
p
toward
the
a
ppli
cat
ion
of
m
odularit
y.
It
can
be
done
on
t
he
in
put
sp
ace
(ho
rizo
ntal
decom
po
sit
ion
)
or
on
t
he
in
put
var
ia
bles
(v
e
rtic
al
dec
om
po
sit
io
n)
[
19]
.
4.2.
Org
an
iz
at
io
n
of t
he
m
od
ul
ar
a
rchi
tectu
r
e
.
The
interc
onne
ct
ion
of
the
m
od
ules
can
be
pa
rall
el
or
in
series.
I
n
the
pa
rall
el
arch
it
ect
ur
e,
al
l
m
od
ules
proce
ss
their
in
f
or
m
at
io
n
sim
ultaneousl
y.
The
gl
ob
al
ou
t
pu
t
i
nvol
ves
s
om
e
m
odules
or
al
l
of
them
,
dep
e
ndin
g
on
t
he
ap
plica
ti
on.
The
c
ooperat
ion
li
nk
betwee
n
the
m
od
ules
wh
ic
h
can
be
of
ty
pe
"a
nd"
or
of
ty
pe
"o
r"
[
20]
.
4.3.
Nature
of lear
ning
The
orga
nizat
ion
of
NN
s
in
a
m
od
ular
a
rc
hitec
ture
m
akes
le
arn
i
ng
m
or
e
di
f
ficult
.
T
he
m
od
ules
of
su
c
h
a
n
arc
hite
ct
ur
e ca
n fo
ll
ow
diff
e
re
nt learni
ng pro
ce
sse
s.
4.3.1.
Indepen
den
t
l
earnin
g
Trainin
g
m
odul
es
ind
e
pe
nd
e
nt
ly
see
m
s
to
be
the
sim
plest
way.
This
s
ugge
sts
that
the
othe
r
m
od
ule
s
of
t
he
arc
hitec
ture
do
not
pa
rtic
ipate
in
le
arn
i
ng.
The
i
nteracti
on
betwe
en
the
m
od
ule
s
then
occ
ur
s
on
l
y
durin
g resti
tuti
on phase
[
21
]
.
4.3.2.
Coop
er
ati
ve
l
earnin
g
The
pro
po
se
d
idea
is
to
us
e
a
global
m
et
ho
d
to
trai
n
al
l
m
od
ules
at
the
sam
e
tim
e.
It
is
necessa
ry
then
t
o hav
e
a
fixe
d
arc
hitec
ture
, det
erm
ined
in
adva
nce. An exam
ple is
giv
e
n by ME
[
22
]
.
4.4.
Co
m
munic
at
i
on
between
m
od
ules
The
te
ch
nique
s
for
cal
culat
ing
the
ov
e
rall
ou
tp
ut
of
a
m
ul
ti
-
networ
k
arch
it
ect
ur
e
a
re
div
e
rsely
var
ie
d,
am
on
g wh
ic
h
is t
he
te
chn
i
qu
e
of
w
e
igh
te
d v
otes
[23]
. A
w
ei
gh
t i
s
the
n
ass
ociat
ed
with eac
h
cl
assifi
er
represe
nting
a
m
easur
e
of
pe
r
form
ance.
Anothe
r
te
chn
i
que
is
to
m
ini
m
i
ze
the
m
ean
square
e
rror
(M
S
E)
of
the g
l
ob
al
outp
ut.
5.
PROPE
SED
NEU
ROGL
IAL N
ET
W
O
RK A
RCHIT
ECTU
RE
The
a
rch
it
ect
ure
a
dopted
f
or
our
ANG
N
is
par
tl
y
base
d
on
the
co
nce
pts
of
m
od
ularit
y
an
d
rem
ai
ns
ver
y
cl
os
e
to
t
he
ab
ov
e
-
m
entione
d
arc
hitec
ture
"M
E".
I
n
this
arch
it
ect
ur
e,
a
nu
m
ber
of
N
Ns
(e
xp
e
rts)
are
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A Novel N
e
ur
ogli
al Arc
hitec
ture f
or
M
odel
li
ng S
i
ngular
Pe
rturbati
on
Syst
em
(
Sami
a S
ala
h)
4815
su
pe
r
vised
by
a
GN
(F
i
g
ure
1).
The
glial
su
pe
r
visor
net
w
ork
determ
ines
the
wei
gh
ti
ng
coeffic
ie
nts
of
each
exp
e
rt’s
par
ti
ci
pation
acc
ordi
ng
to
t
he
input.
Th
e
A
NGN
use
s
the
"divi
de
and
c
onquer"
strat
egy
in
w
hich
the
respo
ns
es
f
r
om
the
ex
per
t NNs
are
com
bin
ed
into
a
sin
gle
r
apid
respo
ns
e. Th
e
la
tt
er
is
ag
gr
e
gated
t
o
the
slo
w
respo
ns
e
of
the
glial
super
visor
netw
ork,
res
ulti
ng
in
the
ove
rall
re
sp
onse
of
our
syst
e
m
.
Algo
rithm
dev
el
op
e
d
in
t
his
arti
cl
e
is
based
on
softm
ax
f
un
ct
io
n.
In
this
al
go
rithm
,
the
super
visor
GN
e
valuate
s
the
perform
ance o
f
each
e
xpert
N
N
acc
or
ding t
o t
he
in
put a
nd s
el
ect
s the b
e
st
of them
to
be
a
ct
ivate
d.
Fig
ure
1
.
Ne
urog
li
al
n
et
w
ork a
rch
it
ect
ure
As
il
lustrate
d
i
n
Fig
ure
1,
our
global
A
NGN
is
com
po
sed
of
K
fast
N
Ns
a
nd
a
slo
w
supe
rv
is
or
G
N
.
The
vecto
r
of
the
in
pu
ts
is
di
vid
e
d
int
o
tw
o
vecto
rs
X
s
an
d
X
f
rep
rese
ntin
g,
resp
ect
i
vely
,
the
slo
w
in
pu
t
s
an
d
the
fast
inputs
of
the
net
wor
k.
T
he
vecto
r
X
s
is
assigned
to
the
su
pe
r
vi
so
r
net
wor
k
and
the
vecto
r
X
f
is
ass
ign
e
d
t
o
t
he
va
rio
us
e
xpert
s.
T
he
respo
nse
s
of
the
e
xp
e
r
t
m
od
ules
are
com
bin
ed
t
o
f
or
m
the
fast
outp
ut.
The
supe
rv
is
or
GN
has
tw
o
outp
uts,
the
first
on
e,
wh
ic
h
is
us
e
d
for
co
ntr
ol
and
super
visi
on
of
the
e
xp
e
r
ts
by
sel
ect
ing
the
m
os
t
su
it
able
netw
ork
an
d
de
act
ivati
ng
the
oth
ers
for
ea
ch
input
vect
or.
The
sec
ond
ou
tp
ut
represe
nts
the
slow
res
ponse
of
t
he
GN,
this
outp
ut
is
aggre
gated
at
the
fast
ou
t
pu
t
to
f
or
m
the
global
respo
ns
e
of
th
e
ANGN.
The
structu
re
of
the
global
A
N
GN
is
cl
os
e
to
m
ulti
-
m
od
el
appr
oach.
I
nd
ee
d,
eac
h
exp
e
rt
netw
ork
is
spe
ci
al
iz
ed
in
a
pr
eci
se
s
ub
-
pro
blem
,
a
ve
rtic
al
deco
m
posit
ion
of
t
he
i
nput
var
ia
bles
into
fast an
d sl
ow i
nputs,
as
well
as a
horizo
ntal
deco
m
po
sit
io
n of t
he fast
in
pu
t space
X
f
a
nd
slow
X
s
is use
d.
5.1.
Algori
th
m
b
ase
d on func
tion so
ft
m
ax
In
t
he
A
N
GN,
the
vect
or
of
the
fast
in
puts
X
f
is
sect
ion
e
d
bot
h
seq
ue
nt
ia
ll
y
and
in
pa
rall
el
into
K
vecto
rs
X
f1
,
X
f2
,….,
X
fK
.
T
hes
e
K
vecto
rs
X
fi
const
it
ute
the
r
especti
ve
in
puts
of
t
he
K
e
xperts.
T
he
slo
w
input
vecto
r
X
S
is
al
so
sect
io
ned
i
n
t
he
sam
e
way
into
K
vecto
rs
X
S1
,
X
S2
,…,
X
S
K
wh
ic
h
a
re
a
ppli
ed
c
on
sec
ut
ively
to
the s
up
e
rv
is
or
GN.
In
this
al
gorith
m
,
each
vect
or
of
t
he
fast
in
puts
(
)
,
(
=
1
,
2
,…,K)
is
a
pp
li
ed
to
al
l
e
xp
e
rts
at
t
he
sam
e
t
i
m
e.
These
m
od
ules
le
arn
dif
fer
e
nt
e
xam
ples
from
the
le
ar
ning
ba
se
an
d
s
pecial
iz
e
in
s
pecific
gro
ups
of
res
ponse
s
that
are
then
w
ei
gh
te
d
by
the
su
pe
rv
is
or
G
N
accor
ding
to
their
abs
olu
t
e
diff
ere
nces
a
nd
th
e
desire
d
res
po
ns
e.
T
he
e
xp
ert
w
ho
se
res
pons
e
is
t
he
cl
os
est
to
the
desire
d
re
spon
s
e
will
ha
ve
the
highest
weig
ht.
The
s
uper
vis
or
G
N,
w
hose
i
nput
is
the
vect
or
of
slo
w
i
nput
s
(
)
,
e
valuates
t
he
pe
rfor
m
ance
of
eac
h
exp
e
rt
acc
ordi
ng
to
the
i
nput
an
d
sel
ect
s
th
e
best
one
t
o
be
act
ivate
d.
T
hi
s
al
gorithm
shows
m
any
si
m
i
la
riti
e
s
t
o
the
on
e
de
ve
lop
e
d
by
Jaco
b
[
22]
in
the
a
r
chite
ct
ur
e
"m
i
xtures
of
e
xp
e
r
ts"
.
These
sim
i
la
riti
es
are
relat
ed
to
the s
up
e
rv
isi
on and select
io
n of ex
per
ts
, howeve
r
t
her
e a
r
e thr
ee
m
ai
n
dif
fe
ren
ces:
In
t
he
"m
ixtur
e
of
e
xperts"
a
rch
it
ect
ure,
the
super
visor
an
d
the
e
xp
e
rts
ha
ve
the
sam
e
tim
e
scal
e.
Fo
r
ou
r
ANG
N,
t
he
tw
o
ty
pes
of m
odules
work on t
wo d
i
ff
e
ren
t t
i
m
e scal
es, slow a
nd f
ast
.
The
ta
sk
of
the
su
pe
rv
is
or
in
the
ME
app
roach
is
to
su
perv
ise
and
c
on
tr
ol
the
com
petition
of
ex
per
ts.
I
n
our
ap
proac
h,
the
GN
s
uper
vi
ses
and
c
on
t
r
ols
the
com
petit
ion
of
e
xpert
s
as
well
as
con
t
rib
uting
to
the
ov
e
rall
r
es
ponse
of the
syst
em
b
y
pro
vid
in
g
t
he
sl
ow r
es
ponse
.
1
2
.
.
Exp
ert
n
etwo
rk
k
∑
Exp
ert
n
etwo
rk 1
ex
p
ert
1
Glial
n
etwo
rk
∑
Exp
ert
n
etwo
rk 2
=
1
,
2
,
…
,
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4810
-
4822
4816
Fo
r
the
sel
ect
ion,
the
e
xperts
'
weigh
ti
ng
i
s
bin
a
ry
(0
or
1).
I
n
the
ME
a
ppr
oac
h
wei
gh
t
s
are
prob
a
bili
ti
es
betwee
n 0 a
nd
1.
Be
fore
present
ing
our
al
gorit
hm
,
it
is
wo
rth
no
ti
ng
that
th
e
exp
e
rts’
le
ar
ning
are
c
oope
rati
ve,
w
hic
h
m
eans
ex
pe
rts learn
sim
ultaneousl
y
an
d
div
i
de
the
ta
sk
du
r
ing
t
he
le
ar
nin
g
process
.
T
he
weig
hts o
f
t
he
exp
e
rt
and
th
os
e
of
the
G
N,
f
or
the
sel
ect
ed
vecto
r,
are
up
dated
at
the
sa
m
e
tim
e
by
pr
opaga
ti
on
.
T
he
le
arni
ng
of
exp
e
rts a
nd the
GN is ca
rr
ie
d ou
t
sim
ultaneou
sly
b
y
fo
ll
owi
ng these
steps:
1.
The
se
pa
rati
on of in
put vect
or X
i
nto
both
slow an
d fast
ve
ct
or
s:
and
.
2.
Each
vecto
r
(
)
,
(
=
1
,
2
,…,K)
is i
nten
de
d for all
expe
r
ts.
3.
Each
vecto
r
(
)
,
(
=
1
,
2
,
…,K) is i
nten
de
d for the
s
up
e
rv
is
or GN.
4.
The
le
ar
ni
ng of the
GN t
o ob
t
ai
n
the
desir
ed
sl
ow
res
pons
e
corres
pondin
g t
o
the i
nput.
5.
The
sel
ect
ion
of
t
he
ex
per
t
ac
cordin
g
to
the
value
of
the
prob
a
bili
ty
(
/
)
and
sel
ect
ion
of
t
he
it
h
ex
pe
rt
by
e
valuati
ng the sl
ow in
pu
t
(
)
.
6.
The
outp
ut
of
exp
e
rt
i
repres
ents
the
co
ndit
ion
al
ave
rag
e
of
the
desire
d
respo
ns
e
with
resp
ect
the
in
put
and the e
xpert
netw
ork.
The
le
ar
ni
ng algorit
hm
o
f
th
e
ANG
N
arc
hite
ct
ur
e
.
1.
I
niti
al
iz
at
io
n of t
he
sy
nap
t
ic
w
ei
ghts
of
e
xp
e
rts a
nd the
GN.
2.
f
or each
slo
w
in
put vect
or
(
)
:
2.1 Ca
lc
ulate
for
=
1
,
2
,
…
f
or
=
1
,
2
,
…
,
e
nd
for
2.2.
Re
peat ste
p 2.1
un
ti
l t
he
al
gorithm
con
ve
rg
es
.
2.3. If m
ax
i
g
t
han
1
i
g
el
se
g
i
=
O
2.4.
)
(
)
(
1
)
(
1
)
(
(
f
i
s
i
K
i
i
f
i
K
i
i
f
i
y
y
y
g
y
Y
Aggregati
on
of the tw
o sl
ow a
nd f
ast
outp
uts
end f
or
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
1
2
)
(
2
)
(
2
1
)
(
)
(
)
(
)
(
1
)
(
1
)
(
1
1
2
1
e
x
p
2
1
e
x
p
)
(
)
(
]
,...,
,
[
)
(
)
(
))
(
e
x
p
(
))
(
e
x
p
(
)
(
)
(
)
(
s
i
i
i
i
i
f
i
m
f
i
i
m
f
i
m
f
i
m
f
i
m
f
m
f
i
s
i
m
s
i
m
s
i
m
s
i
m
s
i
m
s
m
s
i
m
s
i
m
s
i
m
s
i
k
j
f
i
f
j
f
i
f
i
i
T
q
i
i
i
f
i
m
f
i
T
f
i
m
f
i
k
j
j
i
i
i
T
s
i
i
X
k
g
k
h
k
a
k
a
X
k
e
k
h
k
W
k
W
k
y
k
e
X
k
e
k
W
k
W
k
y
k
d
k
e
k
W
X
y
y
d
g
y
d
n
g
k
h
y
y
y
k
y
k
W
X
y
k
u
k
u
k
g
k
a
X
k
u
d
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A Novel N
e
ur
ogli
al Arc
hitec
ture f
or
M
odel
li
ng S
i
ngular
Pe
rturbati
on
Syst
em
(
Sami
a S
ala
h)
4817
6.
APPLI
CA
TI
ON
The
A
SM
is
a
highly
coupled
nonlinear
c
omplex
syst
em
an
d
is
a
ty
pical
e
xam
ple
of
two
tim
e
-
scal
es
syst
e
m
.
The
per
f
or
m
ance
of
t
he
pro
posed
A
NGN
arc
hitec
ture
is
assesse
d
on
both
t
he
re
du
ce
d,
sl
ow
a
nd
fast
m
od
el
s
of
the
m
achine.
The
sta
te
m
od
el
of
the
in
duct
ion
m
achine
in
the
sta
ti
on
a
ry
c
oord
i
nate
syst
em
(
,
)
can
be writt
en a
s:
0
.
1
1
1
2
2
s
r
s
r
r
r
s
r
S
r
s
v
I
T
R
B
T
R
T
B
I
T
dt
d
(16)
r
T
s
s
r
em
J
R
L
B
P
T
2
(17)
Or:
.
1
1
1
2
2
I
T
R
B
T
R
T
B
I
T
A
rp
r
rp
sp
r
sp
with
:
r
rp
s
sp
T
T
T
T
,
,
=
=
(
1
−
2
)
/
(
∗
)
)
Applic
at
ion
of
the
GCT
t
o
th
e
sta
te
m
at
rix
A
res
ults
in
tw
o
ci
rcles
wh
ic
h
inters
ect
(F
i
g
ure
2a)
.
A
change i
n
th
e s
iz
e o
f
rays is
c
arr
ie
d o
ut
by tr
ansfo
rm
ation
1.
R
B
I
P
P
r
r
s
0
0
,
2
1
1
1
To
get:
.
1
1
1
1
2
2
2
2
2
1
J
I
T
I
T
I
T
I
T
A
rp
rp
sp
sp
(18)
The
ce
nters
of
the tw
o
ci
rcles
(F
ig
ur
e
2a
)
a
re
r
el
ocate
d by:
,
0
,
2
2
2
2
1
2
2
I
I
I
P
P
And
.
1
1
0
2
2
2
2
2
2
J
I
T
J
I
T
I
T
A
sp
sp
sp
The ne
w
li
ne
c
ircl
es are s
ti
ll
double
a
nd d
is
jo
int (Fi
g
ure
2
b);
the
final tra
nsf
or
m
at
ion
is:
r
s
P
P
2
1
,
R
B
I
I
P
P
r
2
2
2
1
0
(
19)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4810
-
4822
4818
(a)
(
b)
Fig
ure
2
.
(a
)
C
i
rcles that
inter
sect
;
(b)
Disjointed Circl
es
In
this
case,
t
he
slow
an
d
fas
t
com
po
nen
ts
are
easi
ly
identifie
d.
W
e
ca
n
then
a
pp
ly
SP
M
to
dev
el
op
the slo
w
a
nd f
a
st su
bm
od
el
s.
By
putt
ing
:
z
x
We get:
(20
)
z
J
x
L
p
T
T
sp
em
2
This
f
orm
is
st
and
a
r
d,
the
fl
ux
x
is
slo
w
a
nd
the
flu
x
z
is
fast.
By
dec
om
po
sing
the
fl
ux
e
s,
the
volt
ages
a
nd
the
to
rque,
w
e
ob
ta
in:
f
s
s
s
s
f
s
s
v
t
v
v
z
t
z
z
t
x
x
(21)
f
em
s
em
em
T
T
T
The red
uce
d
sl
ow m
od
el
is th
en:
s
s
s
s
s
s
r
r
s
s
v
v
T
T
dt
d
1
1
1
0
0
1
1
,
1
s
s
s
s
s
s
s
s
s
em
v
v
L
pT
T
-
2
5
0
-
2
0
0
-
1
5
0
-
1
0
0
-
5
0
0
-
2
0
0
-
1
5
0
-
1
0
0
-
5
0
0
50
100
150
200
Re
Im
-
2
0
0
-
1
0
0
0
100
-
3
0
0
-
2
0
0
-
1
0
0
0
100
200
300
Re
Im
A
c
t
u
a
l
E
i
g
e
n
v
a
l
u
e
s
A
c
t
u
a
l
E
i
g
e
n
v
a
l
u
e
s
,
0
0
.
1
1
.
1
0
2
2
2
2
2
s
r
sp
rp
sp
v
R
B
I
I
z
x
I
T
I
T
I
T
z
x
dt
d
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A Novel N
e
ur
ogli
al Arc
hitec
ture f
or
M
odel
li
ng S
i
ngular
Pe
rturbati
on
Syst
em
(
Sami
a S
ala
h)
4819
The red
uce
d
fa
st m
od
el
is:
,
1
1
1
s
s
sp
s
s
sp
s
s
v
T
L
I
f
s
r
s
f
r
s
s
sp
s
s
r
s
r
I
R
B
L
v
T
R
B
1
1
1
1
6.1.
Results
and
discussi
on.
The
ANGN_
MAS
m
od
el
is
com
po
sed
of
two
groups
of
inputs
v
sα
,
v
sβ
,
wh
ic
h
are
dec
om
po
sed
i
nt
o
two
sl
ow
i
nput
s
v
sα(
s
)
,
v
sβ
(
s
)
and
tw
o
fa
st
inputs
v
sα(
f
)
,
v
sβ(
f
)
.
The
ANG
N_M
AS
c
on
sist
s
of
fou
r
ex
pe
rts
and
a
glial
su
pe
rv
is
or
netw
ork
.
The
four
e
xp
e
rts
a
nd
t
he
glial
ne
twork
hav
e
si
m
il
ar
arch
it
ectu
res
,
co
ns
ist
in
g
of
an
input
la
ye
r
of
four
neurons
a
nd
a
n
ou
t
put
l
ay
er
of
one
ne
uro
n.
The
ef
fe
ct
iveness
of
th
e
ANGN
al
gor
it
h
m
pro
po
se
d
i
n
th
is
pap
e
r
is
il
lu
strat
ed
by
the
perform
ance
ind
e
x
root
m
ea
n
s
qu
a
re
(RMS)
value.
T
he
RM
S
of
the stat
es er
ror i
s calc
ulate
d
as
:
RM
S=
√
(
∑
2
(
)
=
1
)
⁄
wh
e
re
n
is
the
nu
m
ber
of
sim
ulati
on
ste
ps
,
a
nd
(
)
is
the
diff
e
ren
ce
betwee
n
the
sta
te
var
ia
bl
es
of
t
he
m
od
e
l
and
t
h
e
true
sy
stem
at
the
it
h
ste
p.
Fig
ure
3
(
a)
an
d
(
b)
s
ho
w
the
go
od
c
onve
r
gen
ce of
t
his
le
arn
i
ng
al
gorithm
.
The
m
ini
m
u
m
value,
w
hich
i
s
ve
ry
cl
os
e
to
zer
o
occurs
a
f
te
r
only
tw
o
it
erati
on
s
.
T
he
a
lgorit
hm
con
ve
rge
d
perfect
ly
and
quic
kly.
(a)
(b)
Fig
ure
3. (a
)
E
vo
l
ution o
f
t
he gli
al
n
et
w
ork
;
(b)
E
voluti
on
of e
xp
e
rts
Fo
r
a
c
om
par
ison
with
the
al
gorithm
s
in
[7]
,
we
ch
oose
the
sam
e
par
am
et
e
rs
of
the
induct
ion
m
oto
r
.
The
sim
ulati
on
resu
lt
s
are
pre
sented
in
Fi
g
ure
4
and
5.
T
he
RM
S
values
f
or
sta
te
var
ia
bles
are
pr
ese
nte
d
in
Table
1.
2
4
0
0
.
2
0
.
4
0
.
6
0
.
8
1
x
1
0
-26
i
t
e
r
a
t
i
o
n
R
M
S
2
4
0
0
.
2
0
.
4
0
.
6
0
.
8
x
1
0
-28
X
:
2
Y
:
7
.
5
5
e
-
0
3
0
i
t
e
r
a
t
i
o
n
R
M
S
E
x
p
e
r
t
1
E
x
p
e
r
t
2
E
x
p
e
r
t
3
E
x
p
e
r
t
4
f
s
s
f
s
s
f
em
I
I
p
T
f
s
f
s
s
f
s
f
s
s
s
f
s
f
s
v
v
L
I
I
T
T
I
I
dt
d
1
1
0
0
1
Evaluation Warning : The document was created with Spire.PDF for Python.