T
E
L
KO
M
NI
K
A
,
V
ol
.
14,
N
o.
3,
S
ept
em
ber
20
16,
pp.
94
1~
9
47
I
S
S
N
:
1
693
-
6
930
,
ac
c
r
edi
t
ed
A
b
y
D
IK
T
I,
D
e
c
r
e
e
N
o
:
58/
D
I
K
T
I
/
K
ep/
2013
D
O
I
:
10.
12928/
T
E
LK
O
M
N
I
K
A
.
v
1
4
i
3
.
3713
94
1
R
ec
ei
v
ed
A
p
r
il
6
,
2
01
6
;
R
e
v
i
s
ed
J
une
4
,
201
6
;
A
c
c
ept
ed
J
un
e
2
5
,
20
1
6
P
r
e
di
c
ti
on M
od
e
l
of S
me
l
ti
ng E
ndp
oi
n
t o
f F
umi
ng
F
u
r
n
a
ce
Bas
ed
o
n
G
r
e
y
Ne
u
r
al Net
w
o
r
k
S
o
n
g
Q
i
an
g
*
,
W
u
Y
ao
-
ch
u
n
S
c
hoo
l
of
M
ec
ha
ni
c
al
E
n
gi
n
ee
r
i
ng,
A
ny
ang I
n
s
t
i
t
ut
e
of
T
ec
h
n
ol
ogy
,
H
en
an,
C
hi
n
a
*
C
or
r
es
po
ndi
ng a
ut
hor
,
e
-
m
a
i
l:
1352
3323
305@
1
26.
c
om
A
b
st
r
act
S
i
nc
e
gr
e
y
t
h
eor
y
and
neur
a
l
net
w
or
k
c
oul
d i
m
pr
o
v
e
pr
e
di
c
t
i
on
pr
ec
i
s
i
on,
t
he t
ec
hno
l
ogy
of
c
om
bi
na
t
i
o
n
pr
e
di
c
t
i
o
n
w
as
p
r
opos
ed
i
n
t
h
i
s
s
t
u
dy
.
T
he
n
t
he
al
gor
i
t
hm
w
as
s
i
m
ul
at
ed
b
y
M
at
l
ab
u
s
i
n
g
pr
ac
t
i
c
a
l
d
at
a
of
a
f
um
i
n
g
f
ur
n
ac
e.
T
he
r
es
ul
t
s
r
e
v
ea
l
t
hat
t
h
e
s
m
el
t
i
ng
end
poi
nt
o
f
f
um
i
ng
f
ur
na
c
e
c
ou
l
d
b
e
ac
c
ur
at
e
l
y
pr
edi
c
t
ed w
i
t
h t
hi
s
m
odel
by
r
e
f
er
r
i
n
g t
o s
m
al
l
s
a
m
pl
e and i
n
f
or
m
at
i
on.
It s
h
o
w
s
th
a
t th
e
G
N
N
al
gor
i
t
hm
not
o
nl
y
has
s
t
r
o
ng
gl
oba
l
s
ear
c
h c
a
pabi
l
i
t
y
,
bu
t
al
s
o i
s
ea
s
y
t
o
i
m
pl
em
ent
.
A
S
m
el
t
i
ng E
n
dpo
i
n
t
of
F
um
i
ng
f
or
e
c
a
s
t
i
n
g em
pi
r
i
c
al
e
x
am
pl
e
has
s
h
ow
n t
h
at
c
o
m
pa
r
ed w
i
t
h
ba
c
k
-
pr
opa
gat
i
on
ar
t
i
f
i
c
i
al
n
eur
a
l
net
w
or
k
s
an
d s
i
ngl
e
gr
ay
t
heo
r
y
al
g
or
i
t
hm
,
G
N
N
al
gor
i
t
hm
c
an ac
hi
e
v
e hi
gher
pr
edi
c
t
i
on a
c
c
ur
ac
y
,
bet
t
e
r
c
om
put
at
i
on
al
s
p
eed,
an
d w
hi
c
h i
s
m
or
e s
ui
t
a
bl
e f
or
P
r
edi
c
t
i
on of
S
m
el
t
i
ng
E
nd
p
oi
nt
of
F
um
i
n
g
f
or
ec
as
t
i
ng
.
T
her
ef
or
e,
G
N
N
m
odel
i
s
ef
f
e
c
t
i
v
e w
i
t
h t
h
e a
dv
an
t
age
s
of
hi
g
h pr
e
c
i
s
i
on
,
f
ew
er
s
am
p
l
es
r
equi
r
ed a
nd s
i
m
pl
e
c
a
l
c
u
l
at
i
o
n.
Ke
y
w
o
r
d
s
:
s
m
el
t
i
ng
end
poi
nt
,
gr
ay
neur
al
n
et
w
or
k
,
pr
e
di
c
t
i
o
n,
s
i
nt
er
i
ng
pr
oc
es
s
,
gr
ay
m
od
el
C
o
p
y
r
i
g
h
t
©
20
16 U
n
i
ver
si
t
a
s A
h
mad
D
ah
l
an
.
A
l
l
r
i
g
h
t
s r
eser
ved
.
1
.
I
n
tr
o
d
u
c
ti
o
n
F
um
i
ng f
ur
nac
e c
ons
i
s
t
e
d
of
t
w
o
l
a
y
er
s
w
hi
c
h
w
er
e
c
ool
ed
b
y
c
ol
d
w
at
er
t
o
k
eep t
he
s
af
et
y
of
f
ur
nac
e
s
hel
l
.
T
er
t
i
ar
y
ai
r
or
i
f
i
c
e
w
as
us
ed
t
o
obs
er
v
e
t
he
f
l
am
e
i
n
t
h
e
f
ur
nac
e
-
f
ur
nac
e
t
em
per
at
ur
e and s
m
el
t
i
n
g
s
t
at
us
c
oul
d be
det
er
m
i
ne
d bas
ed
on br
i
ght
nes
s
an
d c
ol
or
of
t
he
f
l
a
m
e.
V
o
l
at
i
l
e
dus
t
s
of
t
he
f
u
m
i
ng
f
ur
nac
e
w
as
gr
adu
al
l
y
c
o
ol
ed
b
y
c
l
i
m
bi
ng
f
l
u
e,
f
ur
nac
e
g
as
c
ool
er
a
nd
ar
r
a
y
c
o
ol
er
i
n
t
he c
l
i
m
bi
ng.
F
i
nal
l
y
,
t
h
e
y
w
er
e r
ec
ov
er
ed
b
y
ba
g dus
t
c
ol
l
ec
t
or
.
F
um
i
ng
f
ur
n
ac
e
s
i
nt
er
i
ng
p
oi
nt
i
s
di
f
f
i
c
ul
t
t
o
be
c
on
t
r
ol
l
ed
us
i
ng
c
o
nv
ent
i
o
nal
m
et
hods
.
I
n r
ec
ent
y
e
ar
s
,
m
os
t
s
t
udi
es
f
oc
us
on us
i
ng
B
P
ne
u
r
al
ne
t
w
or
k
m
odel
t
o
des
c
r
i
be t
he
ent
i
r
e
s
i
nt
er
i
ng
pr
oc
es
s
and
pr
e
di
c
t
t
hes
e
par
am
et
er
s
.
B
P
ne
ur
al
net
w
or
k
has
hi
g
h pr
e
di
c
t
i
on
ac
c
ur
ac
y
and s
om
e def
ec
t
s
.
F
or
ex
am
pl
e,
t
her
e
i
s
n
o d
et
er
m
i
ni
s
t
i
c
r
u
l
e f
or
t
h
e s
e
l
ec
t
i
on
of
B
P
ne
ur
al
net
w
or
k
hi
dd
en
l
a
y
er
;
t
he
net
w
or
k
eas
i
l
y
f
al
l
s
i
nt
o l
o
c
al
m
i
ni
m
u
m
,
w
i
t
h r
e
l
at
i
v
el
y
po
or
s
y
s
t
em
gener
al
i
z
at
i
on
abi
l
i
t
y
.
T
hes
e def
ec
t
s
gr
ea
t
l
y
l
i
m
i
t
i
t
s
a
p
p
lic
a
t
io
n
in
r
e
a
l
-
t
i
m
e pr
edi
c
t
i
on.
T
he i
n
i
t
i
al
w
ei
g
ht
s
an
d
net
w
or
k
s
t
r
uc
t
ur
e
of
B
P
n
eur
a
l
net
w
or
k
ar
e
r
andom
l
y
g
i
v
en
v
al
ues
,
s
o
t
he
num
ber
and
w
ei
ght
of
e
v
er
y
t
r
a
i
n
a
r
e
s
l
i
ght
l
y
di
f
f
er
ent
.
T
hat
i
s
,
net
w
or
k
opt
i
m
i
z
at
i
on
i
s
n
ot
u
ni
q
ue,
an
d
t
her
e
w
i
l
l
be
a
l
oc
al
m
i
ni
m
um
v
al
u
e;
bes
i
d
es
,
t
he
i
n
i
t
i
al
w
ei
ght
of
t
r
ai
n
has
bl
i
nd
nes
s
and
s
l
o
w
c
onv
er
g
enc
e.
C
om
bi
ni
ng g
r
a
y
t
h
eor
y
an
d B
P
neur
a
l
n
et
w
or
k
,
t
hi
s
w
or
k
pr
es
ent
ed
a gr
a
y
n
eur
a
l
net
w
or
k
al
gor
i
t
hm
,
a
new
a
l
gor
i
t
hm
bas
ed
on
gr
a
y
m
o
del
an
d
neur
al
ne
t
w
or
k
m
o
del
.
T
hi
s
n
e
w
i
nf
or
m
at
i
on pr
oc
es
s
i
n
g a
n
d pr
e
di
c
t
i
on m
et
hods
t
ak
e ad
v
an
t
ag
e of
t
h
e r
a
ndo
m
nes
s
of
gr
a
y
m
odel
w
e
ak
eni
ng da
t
a,
r
eg
ul
ar
i
t
y
of
ac
c
um
ul
at
i
ng da
t
a
and hi
ghl
y
no
nl
i
ne
ar
of
neur
al
net
w
or
k
.
I
t
has
be
en
w
i
de
l
y
us
ed
i
n
t
he
po
w
er
,
t
r
ans
por
t
,
s
o
c
i
al
,
agr
i
c
ul
t
ur
al
a
n
d
ot
he
r
f
i
el
ds
[
1,
2
].
H
o
w
e
v
er
,
i
t
s
t
i
l
l
r
em
ai
ns
a
bl
a
nk
i
n Met
a
l
l
ur
gi
c
a
l
i
ndus
t
r
y
.
T
hi
s
w
or
k
ut
i
l
i
z
ed
gr
a
y
neur
a
l
net
w
or
k
t
o pr
ed
i
c
t
t
h
e s
m
el
t
i
ng
end
poi
nt
of
s
i
nt
er
i
n
g,
ac
h
i
e
v
i
ng
goo
d r
es
ul
t
s
.
2.
G
r
e
y N
eu
r
al
N
et
w
o
r
k M
o
d
el
2.
1
.
M
o
d
e
l
i
n
g
o
f
G
r
e
y
G
M
(
1,
1)
M
o
d
e
l
G
r
e
y
t
he
or
y
i
s
a
m
et
hod
of
s
t
ud
y
i
ng
s
m
al
l
s
am
pl
e,
po
or
i
nf
or
m
at
i
on
and
t
he
pr
ob
l
em
of
unc
er
t
ai
nt
y
.
I
t
s
obj
ec
t
s
of
s
t
ud
y
r
ef
er
t
o s
m
al
l
s
a
m
pl
es
w
i
t
h s
om
e k
now
n
i
nf
or
m
at
i
on
and
unk
now
n i
nf
or
m
at
i
on,
p
o
or
i
nf
or
m
at
i
on and
unc
er
t
ai
n
t
y
.
T
hr
oug
h dat
a m
i
ni
ng
of
k
now
n
i
nf
or
m
at
i
on,
v
al
uab
l
e
i
nf
or
m
at
i
on
c
an
be
ex
t
r
ac
t
ed
t
o
ac
hi
ev
e
t
he
c
or
r
ec
t
des
c
r
i
pt
i
o
n
of
s
y
s
t
em
beha
v
i
or
[
6]
,
ef
f
ec
t
i
v
e m
oni
t
or
i
ng of
e
v
o
l
ut
i
on
l
a
w
an
d s
y
s
t
em
l
oc
at
i
o
n pr
e
di
c
t
i
on.
C
om
par
ed
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
14
,
N
o
.
3,
S
ept
em
ber
2016
:
9
4
1
–
9
47
942
w
i
t
h
t
r
adi
t
i
ona
l
pr
e
di
c
t
i
on m
et
ho
ds
,
s
t
at
i
s
t
i
c
al
pr
ed
i
c
t
i
o
n m
et
hods
ha
v
e m
an
y
ad
v
ant
a
ges
:
e.
g.
,
i
t
i
s
not
nec
es
s
ar
y
t
o d
et
er
m
i
ne
.
W
h
et
her
pr
edi
c
t
i
v
e
v
ar
i
a
bl
es
ar
e nor
m
al
l
y
d
i
s
t
r
i
b
ut
ed;
r
at
her
t
ha
n l
ar
ge
s
a
m
pl
e
s
t
at
i
s
t
i
c
s
,
i
t
s
pec
i
al
i
z
es
i
n s
m
al
l
s
a
m
pl
e or
poor
i
nf
or
m
at
i
o
n unc
er
t
a
i
nt
y
;
t
he pr
ed
i
c
t
i
on m
odel
w
i
l
l
n
ot
c
han
ge
w
i
t
h t
h
e c
hang
e of
i
nput
v
ar
i
abl
es
.
T
hr
ough gr
a
y
s
eq
uenc
e
gen
er
at
i
o
n,
gr
a
y
s
y
s
t
em
t
heor
y
be
l
i
ev
es
t
hat
i
n s
pi
t
e
of
t
he
c
om
pl
ex
obj
ec
t
i
v
e
r
epr
es
ent
at
i
o
n
an
d
dat
a,
t
he
s
y
s
t
em
s
t
i
l
l
has
t
he
w
ho
l
e f
unc
t
i
on
an
d
i
ne
v
i
t
ab
l
y
s
om
e i
nher
e
nt
l
a
w
s
.
T
he k
e
y
i
s
t
o c
ho
os
e a
pr
oper
w
a
y
t
o
di
g
a
nd
us
e
i
t
.
A
l
l
gr
a
y
s
equenc
e
c
a
n
be
ge
ner
at
ed
b
y
w
eak
eni
n
g
r
and
om
nes
s
w
i
t
h
t
h
e
r
egul
ar
i
t
y
.
U
n
i
f
i
ed
d
i
f
f
er
ent
i
al
eq
uat
i
ons
m
odel
has
hi
g
h
pr
edi
c
t
i
on
ac
c
ur
ac
y
.
G
M
(
1,
1)
m
odel
i
ng
i
s
bas
i
c
al
l
y
t
he c
um
ul
at
i
v
e
gener
a
t
i
o
n of
or
i
gi
na
l
dat
a,
s
o t
hat
t
he ge
ner
at
ed s
eque
nc
e has
a
c
er
t
ai
n
r
eg
ul
ar
i
t
y
.
T
hen
t
he
f
i
t
t
i
ng
c
ur
v
e
c
an
be
ob
t
ai
n
ed
t
hr
o
ugh
t
h
e
di
f
f
er
ent
i
a
l
eq
uat
i
on
m
odel
i
ng,
t
hus
pr
ed
i
c
t
i
n
g t
h
e unk
no
w
n
par
t
of
t
h
e s
y
s
t
e
m
[
3
，
4
].
I
n G
M m
odel
,
onc
e
ac
c
um
u
l
at
i
on
i
s
c
ond
uc
t
ed
on t
he r
a
w
da
t
a t
o ge
ner
at
e 1
-
A
G
O
.
T
he
ac
c
u
m
ul
at
e
d da
t
a
w
i
l
l
h
av
e
c
er
t
ai
n r
egu
l
ar
i
t
y
af
t
er
dat
a
m
i
ni
ng.
T
he or
i
gi
nal
dat
a
)
0
(
X
i
s
not
obv
i
o
us
l
y
r
egu
l
ar
,
w
i
t
h s
w
i
n
gi
n
g de
v
e
l
o
pm
ent
t
r
end.
A
f
t
er
ac
c
um
ul
at
i
on
ge
ner
at
i
on
,
r
a
w
da
t
a
w
i
l
l
c
ont
ai
n m
or
e
obv
i
o
us
r
egu
l
ar
i
t
y
.
A
s
s
um
i
ng t
i
m
e s
eque
nc
e
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
n
x
x
x
x
0
0
0
0
,..
....,
2
,
1
=
as
a f
i
r
s
t
-
or
der
ac
c
u
m
ul
at
i
v
e
ge
ner
at
ed 1
-
A
G
O
,
a n
e
w
dat
a
s
equ
en
c
e
c
an b
e o
bt
a
i
ne
d t
hr
o
ugh
onc
e
ac
c
u
m
ul
at
i
on of
:
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
∑
=
=
=
k
i
k
x
x
w
h
ere
n
x
x
x
x
0
0
1
1
1
1
1
,
,......,
2
,
1
κ
(
1
)
A
f
t
er
c
ons
t
r
uc
t
i
ng
a f
i
r
s
t
-
or
der
l
i
ne
ar
di
f
f
er
ent
i
al
equ
at
i
on,
t
he
w
hi
t
eni
ng
di
f
f
er
ent
i
a
l
eq
uat
i
ons
c
an be
obt
ai
n
ed:
(
)
(
)
u
x
a
dt
x
d
=
+
1
1
(
2
)
T
he c
ol
um
n of
l
eas
t
s
quar
e
es
t
i
m
at
i
on p
ar
am
et
er
c
an
s
ol
v
e a
and
u:
(
)
Y
B
B
B
T
u
a
N
T
1
−
=
=
α
,
(
3)
W
h
er
e
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
−
+
−
+
−
+
−
=
1
1
5
.
0
1
2
3
5
.
0
1
1
2
5
.
0
1
1
1
1
1
1
n
x
n
x
x
x
x
x
B
,
and
(
)
(
)
(
)
(
)
(
)
(
)
=
n
x
x
x
Y
N
0
0
0
3
2
.
T
i
m
e r
es
pons
e f
unc
t
i
on s
eque
nc
e ex
pr
es
s
i
on of
G
M (
1,
1)
m
odel
w
i
l
l
be o
bt
ai
n
ed,
nam
el
y
t
h
e gr
a
y
pr
edi
c
t
i
on
m
odel
of
(
)
x
1
.
(
)
(
)
(
)
(
)
a
u
e
a
u
x
k
x
ak
+
−
=
+
−
1
1
ˆ
0
1
(
4)
W
h
er
e a i
s
t
he
de
v
e
l
opm
en
t
f
ac
t
or
;
b t
he
am
ount
of
gr
a
y
ef
f
ec
t
.
G
r
e
y
pr
e
di
c
t
i
on m
odel
of
(
)
x
0
is
:
(
)
(
)
(
)
(
)
(
)
)
,
,
2
,
1
(
,
1
1
1
ˆ
0
0
=
−
−
=
+
−
k
e
a
u
x
e
k
x
a
k
a
(
5
)
(
)
x
1
(
)
x
0
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
I
S
S
N
:
1
693
-
6
930
P
r
ed
i
c
t
i
o
n
Mode
l
of
S
mel
t
i
n
g E
n
dpo
i
n
t
of
F
um
i
ng F
ur
n
ac
e B
as
e
d o
n G
r
ey
…
(
S
on
g Q
i
a
ng
)
943
2.
2
.
E
s
ta
b
l
i
s
h
m
e
n
t o
f
G
r
e
y N
eu
r
al
N
et
w
o
r
k
M
o
d
el
S
et
or
i
g
i
n
al
dat
a
as
)
(
t
x
,
t
he
o
nc
e
-
c
um
ul
at
i
v
e
da
t
a
i
s
obt
a
i
ne
d as
)
(
t
y
;
pr
ed
i
c
t
i
o
n
r
es
ul
t
as
)
(
t
z
.
T
hen
t
he
di
f
f
er
ent
i
a
l
equ
at
i
on
ex
pr
es
s
i
on
of
gr
a
y
n
eur
a
l
net
w
or
k
of
a
par
am
et
er
i
s
:
n
n
y
b
y
b
y
b
ay
dt
dy
1
3
2
2
1
1
1
−
+
+
+
=
+
W
h
er
e
n
y
y
y
,
,
,
2
1
ar
e
s
y
s
t
em
i
nput
v
ec
t
or
s
;
i
s
t
he
s
y
s
t
em
out
pu
t
v
ar
i
ab
l
e;
t
he
ot
h
er
s
ar
e
c
oef
f
i
c
i
ent
s
of
di
f
f
er
ent
i
al
e
quat
i
on
.
T
hen t
he t
i
m
e r
es
pons
e f
or
m
ul
a i
s
:
=
d
T
hr
ough c
om
pl
ex
t
r
ans
f
or
m
at
i
on an
d m
appi
n
g,
an
ex
t
end
ed
B
P
n
eur
a
l
n
et
w
or
k
w
i
l
l
bec
om
e
a
gr
a
y
ne
ur
al
net
w
or
k
w
i
t
h
n
i
np
ut
par
am
et
er
s
and
o
ne
out
p
ut
par
am
et
er
.
G
r
e
y
neur
a
l
net
w
or
k
c
ons
i
s
t
s
of
f
our
l
a
y
er
s
-
LA
,
LB
,
LC
a
nd
LD
,
t
h
u
s
det
er
m
i
ni
ng
t
he
c
onn
ec
t
i
on
w
ei
ght
an
d
e
rro
r.
)
(
)
(
...
)
(
)
(
))
(
)
(
...
)
(
)
0
(
(
)
(
1
3
2
2
1
1
3
2
1
1
t
y
a
t
y
t
y
a
b
t
y
a
b
e
t
y
a
t
y
t
y
a
b
a
y
y
t
z
n
n
at
n
n
−
−
−
+
+
+
+
−
−
−
=
)
1
(
)
1
1
2
1
1
)
0
(
)
)
0
(
((
)
(
1
1
a
t
a
t
a
t
e
e
d
e
y
d
y
t
z
−
−
−
+
×
+
×
+
+
×
−
−
=
(
6
)
2.
3
.
A
r
t
i
f
i
ci
a
l
N
eu
r
al
N
et
w
o
r
k (
A
N
N
)
I
n r
ec
ent
y
e
ar
s
,
m
uc
h r
es
ear
c
h has
be
en c
on
duc
t
e
d on t
h
e app
l
i
c
at
i
on
o
f
a
r
t
if
ic
ia
l
i
nt
e
l
l
i
g
enc
e t
ec
h
ni
ques
t
o
f
or
ec
as
t
i
ng pr
ob
l
em
s
.
H
o
w
e
v
er
,
t
h
e m
odel
t
hat
has
r
ec
ei
v
ed
ex
t
ens
i
v
e a
t
t
en
t
i
o
n i
s
u
ndo
ubt
e
dl
y
t
h
e A
N
N
,
c
i
t
e
d as
am
ong t
he m
os
t
pow
er
f
ul
c
o
m
put
at
i
ona
l
t
ool
s
ev
er
de
v
e
l
op
ed.
F
ig
ur
e
1
.
A
r
c
hi
t
ec
t
ur
e of
t
w
o l
a
y
er
s
B
P
neur
a
l
n
et
w
or
k
us
ed i
n t
h
e s
t
ud
y
B
P
ne
ur
al
net
w
or
k
is
a
m
u
lt
i
-
l
a
y
er
ar
c
hi
t
ec
t
ur
.
F
or
t
h
e t
w
o
l
a
y
er
s
B
P
ne
t
w
or
k
us
ed i
n
t
hi
s
s
t
ud
y
(
F
i
g
ur
e
1)
,
t
h
e
t
r
ans
f
er
f
unc
t
i
on
of
neur
on
i
n
hi
d
de
n
l
a
y
er
i
s
t
he
s
i
gm
oi
d
f
unc
t
i
o
n
an
d
t
he t
r
a
ns
f
er
f
unc
t
i
on of
neur
on i
n o
ut
pu
t
l
a
y
er
i
s
a
l
i
near
f
unc
t
i
on.
)
(
ex
p
1
1
x
B
j
u
−
+
=
(
7
)
Lev
enb
er
g
-
Mar
q
uar
dt
r
u
l
e
w
as
us
ed t
o t
r
a
i
n t
h
e t
w
o
-
l
a
y
e
r
B
P
net
w
or
k
.
I
t
w
as
de
v
el
o
ped
and t
r
a
i
ne
d t
o f
i
t
f
unc
t
i
o
ns
and m
a
k
e ex
t
r
apol
at
i
on.
T
w
o l
e
ar
ni
ng pr
oc
e
dur
es
ar
e
i
nc
l
u
ded
i
n
B
P
(
)
x
0
(
)
(
)
1
ˆ
0
+
k
x
1
y
)
(
)
(
)
(
)
(
1
3
2
2
1
t
y
a
t
y
t
y
a
b
t
y
a
b
n
n
−
+
+
+
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
14
,
N
o
.
3,
S
ept
em
ber
2016
:
9
4
1
–
9
47
944
net
w
or
k
t
r
ai
ni
ng
[
4
]
.
T
he f
i
r
s
t
one i
s
t
he pos
i
t
i
v
e pr
o
paga
t
i
o
n pr
oc
es
s
i
n
w
h
i
c
h
i
nput
s
i
gna
l
i
s
t
r
ans
f
er
r
ed l
a
y
er
b
y
l
a
y
er
and pr
ac
t
i
c
a
l
out
put
s
of
ev
er
y
n
eur
o
ns
ar
e c
o
m
put
e
d;
t
he s
ec
ond
pr
oc
edur
e
i
s
t
h
e
b
ac
k
-
pr
opagat
i
on
i
n
w
h
i
c
h
t
he
er
r
or
s
bet
w
een
pr
ac
t
i
c
a
l
a
nd
ex
p
ec
t
ed
o
ut
p
ut
s
ar
e
pr
ogr
es
s
i
v
e
l
y
c
om
put
ed
l
a
y
er
b
y
l
a
y
er
,
and
w
e
i
gh
t
s
ar
e adj
us
t
e
d ac
c
or
di
ng t
o
t
he
er
r
or
s
[
5
].
2.
4
.
G
r
a
y
N
e
u
r
a
l
N
e
tw
o
r
k
C
o
m
b
i
n
a
ti
o
n
M
o
d
e
l
G
r
e
y
n
eur
a
l
net
w
or
k
m
odel
i
s
t
he
i
nt
e
gr
at
i
on
of
t
w
o al
gor
i
t
hm
s
-
gr
a
y
m
odel
and
neur
a
l
net
w
or
k
,
s
o i
t
has
t
he
ad
v
a
nt
ag
es
of
bot
h.
Mod
el
i
ng
a
ppr
oac
h
of
gr
a
y
n
eur
a
l
n
et
w
o
rk
i
s
:
F
i
rs
t
l
y
,
af
t
er
es
t
abl
i
s
h
i
ng
a
G
M
(
1,
1)
m
odel
of
v
ar
i
ab
l
es
,
t
h
e
pr
edi
c
t
e
d
v
a
l
u
e
of
r
aw
s
eq
uenc
e
da
t
a
c
an
be o
bt
a
i
ne
d.
T
her
e i
s
a
c
er
t
ai
n
de
v
i
at
i
on
bet
w
een
t
he pr
e
di
c
t
e
d
v
al
ue
and
t
h
e or
i
g
i
n
al
dat
a,
w
her
e
i
n t
h
e or
i
gi
nal
s
equ
e
nc
es
al
s
o h
av
e c
er
t
a
i
n r
e
l
a
t
i
ons
h
i
p.
W
e
m
a
y
n
ot
b
e a
bl
e t
o f
i
gur
e
out
t
hi
s
r
e
l
at
i
ons
h
i
p.
T
hes
e
as
s
oc
i
at
i
on
and
de
v
i
at
i
on
ar
e
c
ons
i
der
ed
i
n
t
he
neur
al
n
et
w
or
k
m
odel
:
pr
edi
c
t
i
v
e v
a
l
u
e of
G
M (
1,
1)
i
s
r
egar
ded as
t
h
e i
n
pu
t
s
a
m
pl
e;
t
he ac
t
u
al
v
a
l
ue
as
t
he out
p
ut
s
a
m
pl
es
of
neur
a
l
ne
t
w
or
k
s
.
C
er
t
ai
n
net
w
or
k
s
t
r
uc
t
ur
e
i
s
ad
opt
ed
t
o
t
r
ai
n
t
h
e
n
et
w
or
k
,
t
her
eb
y
obt
a
i
n
i
ng t
h
e r
i
gh
t
v
a
l
ues
a
nd t
hr
es
ho
l
d v
al
ues
of
c
or
r
es
pond
i
n
g nod
es
.
T
he pr
edi
c
t
i
v
e v
al
u
es
of
G
M (
1,
1)
m
odel
on
t
he
n
ex
t
on
e m
o
m
ent
or
s
ev
er
al
m
o
m
ent
s
ar
e r
eg
ar
de
d
as
t
he
i
n
put
of
neur
a
l
net
w
or
k
,
and t
he c
or
r
es
pondi
ng ou
t
put
i
s
t
h
e f
i
nal
pr
e
di
c
t
i
v
e
v
a
l
ue of
t
he nex
t
m
o
m
ent
[
7
].
G
r
e
y
n
eur
a
l
net
w
or
k
pr
i
m
ar
i
l
y
c
o
ns
i
s
t
s
of
i
nput
l
a
y
er
,
h
i
dden
l
a
y
er
a
nd
out
put
l
a
y
er
,
w
i
t
h
s
o
m
e
dev
i
at
i
on
and
at
l
eas
t
one
S
-
t
y
p
e
h
i
d
den
l
a
y
er
and
l
i
n
ear
o
ut
pu
t
l
a
y
er
.
T
h
e
net
w
or
k
ha
s
t
he c
har
ac
t
er
i
s
t
i
c
of
appr
ox
i
m
at
i
ng an
y
r
at
i
on
al
f
unc
t
i
o
n,
s
i
m
ul
at
i
n
g t
h
e r
el
at
i
ons
hi
p bet
w
een
t
he
s
equenc
e
dat
a b
y
t
r
a
i
n
i
ng
t
he n
eur
a
l
ne
t
w
or
k
[
8
].
I
t
i
s
s
upp
os
ed
t
h
at
t
h
er
e
ar
e
m
s
a
m
pl
es
of
i
nt
er
r
el
at
e
d
dat
a
c
o
l
um
ns
,
and
eac
h
c
ol
um
n
c
ont
ai
ns
n d
at
a.
G
r
e
y
neur
al
n
et
w
or
k
pr
edi
c
t
i
on m
odel
i
s
es
t
ab
l
i
s
h
ed
as
f
ol
l
o
w
s
:
1)
T
he
m
r
aw
d
at
a s
e
que
nc
e
w
as
us
e
d t
o
es
t
ab
l
i
s
h c
or
r
es
pond
i
ng
G
M (
1,
1)
m
odel
;
2)
T
he m
m
odel
s
w
er
e us
e
d t
o
pr
ed
i
c
t
t
h
e s
ec
o
nd t
o n
-
t
h dat
a of
e
ac
h c
o
l
um
n,
ob
t
ai
n
i
n
g
m
dat
a s
equenc
es
P
w
i
t
h t
he l
eng
t
h of
n
-
1;
3)
P
v
al
ues
of
dat
a
s
eq
ue
nc
e
w
er
e
r
e
gar
d
ed
as
t
he
i
nput
v
ec
t
or
of
neur
a
l
n
et
w
or
k
s
;
T
as
t
he
out
put
v
ec
t
or
of
neu
r
al
n
et
w
or
k
;
t
he
net
w
or
k
s
t
r
uc
t
ur
e,
i
ni
t
i
a
l
w
ei
ght
s
a
nd
t
hr
es
hol
ds
ar
e
s
et
;
4)
T
he pr
edi
c
t
i
on ac
c
ur
ac
y
of
neur
al
net
w
or
k
w
as
al
s
o s
et
t
o t
r
ai
n t
he B
P
net
w
or
k
.
A
f
t
er
t
he
t
r
ai
ni
n
g
w
as
qua
l
i
f
i
e
d,
w
e
c
oul
d
o
bt
a
i
n
a
s
er
i
es
of
w
e
i
ght
s
an
d
t
hr
es
ho
l
ds
c
or
r
es
pondi
ng
t
o
eac
h n
od
e;
5)
G
M
(
1
,
1)
m
odel
es
t
ab
l
i
s
hed
i
n
t
h
e
f
i
r
s
t
s
t
ep
w
as
u
s
ed
t
o
pr
e
di
c
t
t
he
v
al
ue
of
f
ut
ur
e
t
i
m
e.
T
hes
e
pr
edi
c
t
e
d
v
al
u
es
w
er
e r
egar
d
ed as
i
nput
of
net
w
or
k
f
or
s
i
m
ul
at
i
on,
t
hus
obt
ai
n
i
n
g
t
he c
or
r
es
po
ndi
ng
out
put
,
t
he r
es
ul
t
s
of
gr
a
y
neur
al
ne
t
w
or
k
pr
edi
c
t
i
v
e
m
odel
;
3
.
S
i
m
u
l
a
ti
o
n
R
e
s
u
l
ts
a
n
d
A
n
a
l
ys
i
s
3.
1
.
I
n
p
u
t a
n
d
O
u
tp
u
t L
a
y
e
r
D
e
s
i
g
n
T
he
f
i
r
s
t
f
u
m
i
ng f
ur
nac
e of
a l
ead
-
z
i
nc
s
m
el
t
er
pl
ant
has
t
w
o f
u
m
i
ng f
ur
nac
es
,
w
i
t
h a
m
oni
t
or
i
ng
s
y
s
t
em
dev
e
l
o
ped
b
y
t
he
C
i
t
ec
t
.
T
he
j
udgm
ent
v
ar
i
a
bl
es
of
s
m
el
t
i
ng
s
i
nt
er
i
n
g
en
dp
oi
nt
i
nc
l
ud
e c
ol
d
i
n
put
(
us
ual
l
y
t
w
o
w
ar
eh
ous
es
or
a
w
ar
eho
us
e
ha
l
f
)
,
s
m
el
t
i
ng t
i
m
e af
t
e
r
f
eedi
ng
,
c
oa
l
c
on
v
er
t
er
f
r
e
quenc
y
(
c
oa
l
f
r
equenc
y
)
a
nd t
hr
e
e o
ut
l
et
t
em
per
at
ur
e.
F
ur
t
h
er
m
or
e
,
s
i
nc
e
t
he
f
l
am
e
i
m
age
i
n
t
hr
ee
out
l
et
s
has
obv
i
o
us
f
eat
ur
es
at
v
ar
i
ous
s
t
ages
.
A
hi
g
h
-
def
i
n
i
t
i
on
di
g
i
t
a
l
v
i
de
o
c
am
er
a w
as
i
ns
t
al
l
ed at
t
he s
i
t
e,
s
o
i
m
ages
of
t
hr
ee out
l
et
s
c
an
be c
apt
ur
ed a
nd
ana
l
y
z
e
d at
al
l
s
t
ag
es
of
t
he s
m
el
t
i
ng pr
oc
es
s
.
T
hes
e pr
o
v
i
d
ed c
o
nv
eni
enc
e f
o
r
m
ul
t
i
-
s
ens
or
dat
a
f
us
i
on.
F
i
r
s
t
l
y
,
t
hi
s
w
o
r
k
c
onduc
t
ed
f
us
i
on
s
i
m
ul
at
i
on
on
f
our
v
ar
i
a
bl
es
-
c
oal
i
nput
,
s
m
el
t
i
ng
t
i
m
e,
c
oal
f
r
equenc
y
a
nd o
ut
l
e
t
t
em
per
at
ur
e,
w
h
i
c
h c
an be di
r
ec
t
l
y
obt
ai
n
ed f
r
om
t
he m
oni
t
or
i
n
g
s
y
s
t
em
.
T
hen
t
he
i
m
age
f
eat
ur
e
of
t
hr
ee
out
l
et
s
w
as
al
s
o
o
bv
i
ous
,
s
o
br
i
g
ht
nes
s
of
t
he
i
m
age
w
as
adde
d
on t
he bas
i
s
of
f
our
v
ar
i
ab
l
es
.
F
i
na
l
l
y
,
f
us
i
o
n s
i
m
ul
at
i
on
w
as
c
onduc
t
e
d on t
h
es
e f
i
v
e
v
ar
i
ab
l
es
.
T
he t
oo
l
of
s
i
m
ul
at
i
o
n
w
as
MA
T
LA
B
6.
5
[
5
].
R
es
ul
t
of
abo
v
e f
us
i
o
n s
i
m
ul
at
i
o
n o
n f
our
v
ar
i
a
bl
es
s
h
o
w
ed
t
hat
t
h
e j
udgm
ent
of
f
um
i
ng
f
ur
nac
e s
m
el
t
i
n
g e
ndp
oi
n
t
bec
am
e
m
or
e ef
f
ec
t
i
v
e.
B
u
t
t
he
i
m
age f
eat
ur
e
of
t
hr
e
e out
l
et
s
al
s
o
had
gr
eat
i
m
pac
t
on s
m
el
t
i
ng e
ndp
oi
nt
,
s
o t
hi
s
w
or
k
adopt
ed t
he br
i
gh
t
nes
s
of
i
m
age
as
ano
t
her
v
ar
i
ab
l
e.
T
he f
us
i
on s
i
m
ul
at
i
on
w
as
c
ond
uc
t
ed
on t
hes
e f
i
v
e
v
ar
i
ab
l
es
[
7
].
T
hi
s
w
or
k
es
t
abl
i
s
hed
G
M
(
1,
1)
pr
edi
c
t
i
on m
odel
on
i
nput
v
ar
i
abl
es
r
el
at
ed t
o
f
u
m
i
ng
f
ur
nac
e
s
m
el
t
i
ng
en
dpo
i
nt
s
,
obt
a
i
ni
ng
s
e
v
er
a
l
pr
e
di
c
t
i
v
e
v
a
l
u
e
as
t
h
e
i
npu
t
of
B
P
n
eur
a
l
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
I
S
S
N
:
1
693
-
6
930
P
r
ed
i
c
t
i
o
n
Mode
l
of
S
mel
t
i
n
g E
n
dpo
i
n
t
of
F
um
i
ng F
ur
n
ac
e B
as
e
d o
n G
r
ey
…
(
S
on
g Q
i
a
ng
)
945
net
w
or
k
.
U
s
i
ng
a
hi
dde
n
l
a
y
e
r
,
t
he
t
r
ans
f
er
f
unc
t
i
on
i
s
(
0,
1)
S
-
t
y
pe
f
unc
t
i
o
n
(
)
e
x
f
x
−
+
=
1
1
; th
e
out
p
ut
i
s
t
he t
i
m
e
f
r
o
m
f
u
m
i
ng
f
ur
nac
e s
m
el
t
i
ng en
dpo
i
nt
.
G
r
a
y
n
eur
a
l
net
w
o
r
k
i
s
ut
i
l
i
z
ed t
o
pr
edi
c
t
f
u
m
i
ng f
ur
nac
e s
m
el
t
i
ng
en
dpo
i
nt
,
w
h
i
c
h
i
s
one
of
t
he
m
os
t
i
m
por
t
an
t
per
f
or
m
anc
e
i
nd
i
c
at
or
s
i
n
s
i
n
t
er
i
ng
pr
o
d
uc
t
i
on
.
I
n
t
he
ent
i
r
e
s
i
nt
er
i
n
g
pr
od
uc
t
i
o
n
pr
oc
es
s
,
v
ar
i
a
bl
es
r
e
l
at
ed
t
o
t
he e
nd
of
t
he s
i
nt
er
i
ng
s
m
el
t
i
ng
i
nf
l
u
enc
e s
h
ou
l
d
be c
ar
ef
ul
l
y
s
el
ec
t
ed,
s
o
w
e c
a
n det
er
m
i
ne t
he
i
np
ut
v
ar
i
ab
l
es
of
gr
a
y
ne
u
r
al
n
et
w
or
k
.
60
dat
a s
et
s
of
i
npu
t
v
ar
i
ab
l
es
ar
e
s
t
or
e
d i
n t
he
ex
c
el
dat
a
bas
e a
nd
em
bedded
i
n Mat
l
a
b6.
5.
I
n
Ma
t
l
a
b6.
5,
t
he
i
m
por
t
w
i
z
ar
d c
an
eas
i
l
y
c
al
l
o
ut
t
he
dat
a
i
n t
h
e ex
c
el
da
t
ab
as
e:
s
i
m
pl
y
t
y
p
i
ng t
he da
t
ab
as
e nam
e i
n t
he
w
i
ndo
w
c
a
n c
al
l
o
ut
t
he r
i
ght
dat
a
bas
e.
3.
2
.
T
r
a
i
n
i
n
g
S
a
m
p
l
e
N
o
r
m
a
l
i
z
a
ti
o
n
a
n
d
N
et
w
o
r
k S
et
U
p
1)
T
r
ai
ni
ng
dat
a
i
s
t
h
e ac
t
ual
pr
oduc
t
i
on r
ec
or
d of
a
l
ea
d
-
z
i
nc
pl
ant
f
r
om
Mar
c
h 1 t
o
Mar
c
h 31 i
n 201
2.
A
c
c
or
di
ng t
o t
he r
e
qui
r
em
ent
s
,
s
t
abl
e 60 s
et
s
of
dat
a
w
er
e
s
el
ec
t
ed,
w
i
t
h
bet
t
er
c
o
nt
r
ol
ef
f
ec
t
[
10
].
2)
T
o
f
ac
i
l
i
t
at
e
net
w
or
k
l
ear
ni
ng
a
nd
s
pee
d
up
c
onv
er
ge
nc
e
s
peed
,
nor
m
al
i
z
at
i
on
pr
oc
es
s
w
as
c
on
duc
t
ed
o
n t
he ac
t
ual
s
am
pl
e d
at
a,
di
v
i
d
i
n
g t
he
ac
t
ua
l
ph
y
s
i
c
al
v
ar
i
a
bl
es
as
v
a
l
ues
i
n [
-
1,
1]
.
3)
G
r
a
y
ne
ur
al
net
w
or
k
pr
edi
c
t
i
ng
pr
oc
es
s
w
as
w
r
i
t
t
en
us
i
ng
Mat
l
ab
pr
ogr
a
m
m
i
ng
l
an
gua
ge,
w
i
t
h
t
h
e pr
ed
i
c
t
i
on
ac
c
ur
ac
y
of
0.
01.
T
hi
s
ac
c
ur
ac
y
f
u
l
l
y
m
et
t
he
pr
oduc
t
i
on
of
s
i
nt
er
i
ng.
T
he
m
ax
i
m
u
m
n
um
ber
of
t
r
ai
ni
ng
w
as
1
0,
0
00
t
i
m
es
,
and
l
ear
ni
n
g
r
at
e
=
0.
7.
T
hr
ee
-
l
a
y
er
B
P
n
eur
a
l
n
et
w
or
k
adopt
ed o
ne s
i
ng
l
e
hi
d
den
l
a
y
er
,
s
o t
he t
r
a
ns
f
er
f
unc
t
i
on
of
hi
d
den
l
a
y
er
a
nd o
ut
p
ut
l
a
y
er
w
er
e l
o
gar
i
t
hm
i
c
S
i
gm
oi
d t
r
a
n
s
f
er
f
unc
t
i
on
and
pos
i
t
i
v
e
l
i
ne
ar
t
r
ans
f
er
f
unc
t
i
on;
num
ber
of
neur
ons
i
n t
he
hi
d
de
n l
a
y
er
w
as
50;
t
he num
ber
of
neur
on
s
i
n
t
he out
pu
t
l
a
y
er
1;
t
he t
r
a
i
n
i
ng
and
ad
apt
i
v
e
adj
us
t
m
ent
f
unc
t
i
on
w
as
el
as
t
i
c
i
t
y
bac
k
-
pr
opag
at
i
o
n a
l
gor
i
t
hm
.
T
he hi
dd
en
l
a
y
er
and
o
ut
put
l
a
y
er
t
r
a
ns
f
er
f
unc
t
i
on,
as
w
e
l
l
as
t
he
num
ber
o
f
hi
dde
n l
a
y
e
r
neur
o
ns
,
w
as
d
et
er
m
i
ned t
hr
oug
h a
num
ber
of
r
epeat
ed t
r
a
i
n
i
ng
c
om
par
i
s
ons
,
b
as
ed o
n t
h
e r
u
l
e
of
f
as
t
er
t
r
ai
ni
ng a
nd be
t
t
er
pr
edi
c
t
ed o
ut
p
ut
.
S
o t
he ar
c
hi
t
ec
t
ur
e of
f
u
m
i
ng f
ur
nac
e s
m
el
t
i
ng
e
nd
neur
a
l
n
et
w
or
k
w
as
5
×
50
×
1.
Mat
h
em
at
i
c
al
ex
pr
es
s
i
o
n of
MS
E
m
ean s
quar
e
er
r
or
f
unc
t
i
on
i
s
:
M
S
E=
22
11
11
(
)
(
)
NN
i
ii
ii
e
t
a
NN
=
=
=
−
∑∑
MS
E
d
ev
i
at
i
on
i
n t
r
a
i
ni
ng
pr
oc
es
s
c
ur
v
e
i
s
s
ho
w
n
i
n
F
i
gur
e
2;
t
he
ac
t
ua
l
v
a
l
ue
T
and
pr
edi
c
t
e
d v
a
l
u
es
of
t
he c
o
nt
r
as
t
c
ur
v
e ar
e s
ho
w
n i
n F
i
gur
e 3.
T
he hor
i
z
ont
al
a
x
i
s
r
epr
es
ent
s
s
m
el
t
i
ng
t
i
m
e (
m
i
nut
es
)
;
t
h
e or
di
nat
e t
h
e o
ut
pu
t
of
f
us
i
on s
y
s
t
em
-
t
he
t
i
m
e f
r
o
m
s
m
el
t
i
ng
en
d
(
m
i
nut
es
)
.
I
n
add
i
t
i
on,
f
um
i
ng f
ur
nac
e
s
t
at
e
dat
a of
t
en t
i
m
es
s
m
el
t
i
ng pr
oc
es
s
ar
e c
ol
l
ec
t
e
d
f
r
o
m
t
he s
c
ene
as
t
h
e t
es
t
s
a
m
pl
e,
t
hus
obt
a
i
n
i
ng
pr
e
di
c
t
i
v
e
ou
t
put
t
hr
o
ugh
s
i
m
ul
at
i
o
n.
C
ont
r
as
t
c
ur
v
e
of
pr
edi
c
t
ed
ou
t
put
a
nd
ac
t
ua
l
out
p
ut
of
t
wo
f
ur
nac
es
i
s
s
how
n
i
n
F
i
g
ur
e
3
and
4
,
w
her
e
hor
i
z
ont
al
ax
i
s
r
epr
es
e
nt
s
s
m
el
t
i
ng t
i
m
e;
t
he
or
d
i
nat
e t
he t
i
m
e f
r
o
m
t
he s
m
el
t
i
n
g e
nd.
F
ig
ur
e
2
.
M
S
E
c
ha
nges
gr
a
ph
0
50
100
150
200
10
-2
10
-1
10
0
10
1
10
2
10
3
10
4
223 E
poc
hs
T
r
ai
ni
ng-
B
l
ue G
oal
-
B
l
ac
k
P
er
f
or
m
anc
e i
s
0.
0984837,
G
oal
i
s
0.
1
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
14
,
N
o
.
3,
S
ept
em
ber
2016
:
9
4
1
–
9
47
946
F
ig
ur
e
3
.
A
c
t
ua
l
o
ut
pu
t
an
d
pr
edi
c
t
ed o
ut
p
ut
c
u
r
ve
F
ig
ur
e
4
.
T
ar
get
v
ec
t
or
and
s
i
m
ul
at
i
on
out
put
c
u
r
ve
3.
3
.
P
r
e
d
i
c
ti
o
n
w
i
th
G
r
a
y
S
y
st
em
M
o
d
el
a
n
d
C
o
m
b
i
n
e
d
N
e
u
r
a
l
N
e
tw
o
r
k
M
o
d
e
l
Let
1
λ
be gr
a
y
pr
e
di
c
t
i
on
v
al
u
e,
2
λ
be t
he pr
e
di
c
t
i
on
v
a
l
ue
b
y
B
P
ne
ur
al
n
et
w
or
k
,
w
h
i
l
e
c
λ
be
pr
ed
i
c
t
i
o
n
v
al
ue
b
y
op
t
i
m
al
c
om
bi
ned m
odel
.
T
he pr
ed
i
c
t
i
on
er
r
or
s
ar
e
1
η
,
2
η
an
d
c
η
r
es
pec
t
i
v
el
y
.
T
he c
or
r
es
po
ndi
ng
w
ei
g
ht
e
d c
oef
f
i
c
i
ent
s
ar
e
1
ω
,
2
ω
and
c
ω
,
and
1
2
1
=
+
ω
ω
.
2
2
1
1
η
ϖ
η
ϖ
η
+
=
c
(
12)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
2
1
1
1
2
2
1
1
2
1
2
1
2
1
2
2
2
1
2
1
2
2
1
1
,
)
1
(
2
)
1
(
2
,
2
)
(
η
η
ω
ω
η
ω
η
ω
η
η
ω
ω
η
ω
η
ω
η
ω
η
ω
η
Cov
V
ar
V
ar
Cov
V
ar
V
ar
V
ar
V
ar
c
−
+
−
+
=
+
+
=
+
=
(
13)
As
t
o
1
ω
,
i
n or
d
er
t
o
det
er
m
i
ne t
he f
u
nc
t
i
on
al
m
i
ni
mu
m v
a
l
u
e
,
le
t
:
(
)
0
1
=
∂
∂
ω
η
c
V
ar
and
,
bec
a
us
e
(
)
0
,
2
1
=
η
η
Cov
Let
(
)
(
)
12
2
11
1
,
γ
η
γ
η
=
=
V
ar
V
ar
T
hen t
he
w
ei
ght
e
d c
oef
f
i
c
i
e
nt
s
of
c
o
m
bi
ned
pr
ed
i
c
t
i
on
ar
e
:
12
11
11
1
12
11
12
1
,
γ
γ
γ
ω
γ
γ
γ
ω
+
=
+
=
(
14)
I
n F
i
g
ur
e 2
,
w
h
en t
he gr
a
y
neur
a
l
n
et
w
or
k
i
s
t
r
ai
n
ed
t
o 30
0 s
t
eps
,
t
he s
y
s
t
em
out
p
ut
m
ean
s
quar
e
er
r
or
w
i
l
l
r
ea
c
h
0.
01%
,
t
he
n
t
he
t
r
ai
n
s
t
ops
.
I
n
F
i
gur
e
3
,
af
t
er
t
he
neur
a
l
net
w
or
k
t
r
ai
n
i
ng
w
as
c
om
pl
et
e,
t
h
e
pr
edi
c
t
i
v
e v
al
ue of
t
r
ai
n
i
n
g
s
a
m
pl
es
c
oul
d
w
e
l
l
f
i
t
ac
t
u
a
l
pr
oj
ec
t
v
al
u
e.
I
n F
i
gur
e
4,
t
he
de
v
i
at
i
on
bet
w
e
en pr
ed
i
c
t
i
o
n
c
ur
v
e and t
he
ac
t
u
al
c
ur
v
e
w
as
s
m
al
l
,
w
i
t
h
t
w
o
al
t
er
n
at
i
v
e
s
m
el
t
i
ng
f
ur
nac
e
s
t
at
us
dat
a
as
t
he
t
es
t
s
am
pl
e.
T
hei
r
bas
i
c
t
r
ends
w
er
e
t
he
s
am
e:
t
he ac
t
u
al
out
p
ut
c
ur
v
e
w
a
s
r
el
at
i
v
e
l
y
s
t
r
a
i
g
ht
,
w
h
i
l
e
pr
edi
c
t
e
d o
ut
pu
t
c
ur
v
e
w
a
s
s
l
i
gh
t
l
y
ben
t
.
D
ev
i
at
i
on
of
t
he pr
ed
i
c
t
ed
out
p
ut
an
d ac
t
ua
l
ou
t
put
w
as
about
10 m
i
nut
es
,
nam
el
y
t
he f
us
i
on
s
y
s
t
em
has
been
ab
l
e
t
o m
or
e ac
c
ur
at
e
l
y
d
et
er
m
i
ne s
m
el
t
i
ng e
ndp
oi
nt
.
T
her
ef
or
e
,
t
he
pr
ed
i
c
t
i
on
al
g
or
i
t
hm
has
hi
g
h ef
f
i
c
i
en
c
y
and
pr
ed
i
c
t
i
o
n ac
c
ur
ac
y
.
I
n F
i
g
ur
e 4,
t
he
hor
i
z
o
nt
a
l
ax
i
s
w
as
t
he
s
m
el
t
i
ng t
i
m
e,
and or
di
nat
e t
he t
i
m
e f
r
o
m
t
he end of
t
he s
m
el
t
i
ng.
C
om
par
ed t
o pr
e
v
i
o
us
s
i
m
ul
at
i
on r
es
u
l
t
s
,
t
h
e s
i
m
ul
at
i
on r
es
u
l
t
s
w
er
e i
m
pr
ov
e
d
w
i
t
h t
h
e ad
ded
br
i
g
ht
n
es
s
of
t
he i
m
age.
H
o
w
e
v
er
,
t
he
i
m
pr
ov
em
ent
w
as
n
ot
c
l
ear
,
s
o
f
ur
t
he
r
s
t
udi
es
w
er
e nee
de
d.
B
es
i
des
,
ot
h
er
f
eat
ur
es
af
f
ec
t
i
ng t
he j
u
dg
m
ent
of
s
m
el
t
i
ng en
dpo
i
nt
s
hou
l
d b
e i
nt
r
od
uc
ed.
60
70
80
90
100
110
120
130
0
10
20
30
40
50
60
70
60
70
80
90
100
110
120
0
10
20
30
40
50
60
70
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
I
S
S
N
:
1
693
-
6
930
P
r
ed
i
c
t
i
o
n
Mode
l
of
S
mel
t
i
n
g E
n
dpo
i
n
t
of
F
um
i
ng F
ur
n
ac
e B
as
e
d o
n G
r
ey
…
(
S
on
g Q
i
a
ng
)
947
4
.
C
o
n
c
l
u
s
i
o
n
I
n
or
d
er
t
o
get
a
bet
t
er
s
ol
ut
i
o
n
of
a
ppl
i
c
at
i
on,
c
om
bi
ni
n
g
pr
e
di
c
t
i
ng
m
odel
of
gr
a
y
G
M
(
1,
1)
and
B
P
neur
al
net
w
or
k
w
as
appl
i
ed
i
n
f
u
m
i
ng
f
ur
nac
e
s
m
el
t
i
ng
en
dp
oi
nt
,
w
i
t
h
h
i
gh
pr
ec
i
s
i
o
n.
I
t
c
an f
ur
t
her
adj
us
t
t
he pr
oduc
t
i
on pr
oc
es
s
t
i
m
e as
a quant
i
t
at
i
v
e bas
i
s
t
o i
m
pr
ov
e t
he
r
ef
i
ni
ng
c
y
c
l
e,
pr
oduc
t
q
ua
l
i
t
y
and
y
i
el
d,
t
h
us
l
a
y
i
ng
a
s
ol
i
d
f
ou
ndat
i
o
n
f
or
f
ur
t
her
ener
g
y
-
s
av
in
g
and gr
e
en s
t
ee
l
.
G
r
e
y
n
eur
al
net
w
or
k
,
a new
i
nf
or
m
at
i
on pr
oc
es
s
i
ng
and pr
e
di
c
t
i
on m
ode,
t
ak
e
s
f
ul
l
ad
v
ant
a
ge
of
t
he r
a
ndo
m
nes
s
of
t
he gr
a
y
m
odel
w
eak
eni
ng
dat
a,
s
h
o
w
i
ng t
he
hi
g
h r
e
gul
ar
i
t
y
of
ac
c
u
m
ul
at
i
n
g
d
at
a
an
d
neur
a
l
n
et
w
or
k
non
-
pr
e
di
c
t
i
on
m
et
hod.
T
hi
s
n
e
w
,
pr
ac
t
i
c
al
and
h
i
gh
-
ac
c
ur
ac
y
pr
ed
i
c
t
i
o
n
al
gor
i
t
hm
s
houl
d
b
e
pr
om
ot
ed
an
d
f
ur
t
her
s
t
udi
e
d,
gr
ad
ua
l
l
y
appl
i
e
d
i
n
t
he
m
et
al
l
ur
gi
c
a
l
i
nd
us
t
r
y
.
T
he
w
hol
e i
ndus
t
r
y
w
i
l
l
be
nef
i
t
m
or
e
f
r
o
m
t
hi
s
hi
gh
t
ec
hn
ol
og
y
.
A
c
k
n
o
w
l
e
d
g
e
m
e
n
t
s
T
hank
s
f
or
K
e
y
P
r
oj
ec
t
s
of
H
ena
n U
n
i
v
er
s
i
t
i
es
(
pr
oj
ec
t
num
ber
:
16A
510
013)
.
R
ef
er
en
ces
[1
]
Y
an Lu,
D
ongx
i
ao N
i
u,
B
i
ngj
i
e
Li
,
M
i
n
Y
u.
C
os
t
F
or
e
c
as
t
i
ng
M
odel
of
T
r
ans
m
i
s
s
i
on P
r
oj
ec
t
bas
ed
o
n
t
h
e
PSO
-
B
P
M
et
hod.
T
EL
KO
M
N
I
KA
T
el
ec
om
m
uni
c
at
i
on C
om
put
i
ng E
l
ec
t
r
on
i
c
s
and C
o
nt
r
ol
.
20
1
4;
12(
4
)
:
773
-
7
78.
[2
]
Li
u L
i
pi
ng,
S
unj
i
n S
h
eng,
Y
i
n
J
in
g
-
t
a
o,
L
i
an
g N
a.
P
r
edi
c
t
i
on
and R
eal
i
z
at
i
on
of
D
O
i
n S
ew
age
T
r
eat
m
ent
B
as
ed
o
n
M
ac
hi
n
e
V
i
s
i
on
and
B
P
N
eur
al
N
e
t
w
or
k
.
T
EL
KO
M
N
I
KA
T
el
ec
o
m
m
uni
c
at
i
on
C
om
put
i
ng E
l
e
c
t
r
on
i
c
s
an
d C
o
nt
r
ol
.
201
4
;
12(
4)
:
89
0
-
8
96.
[3
]
C
heng Y
ong
m
i
ng
.
O
n
i
nt
el
l
i
g
e
nc
e
opt
i
m
i
z
at
i
on
al
gor
i
t
h
m
a
n
d
i
t
s
ap
pl
i
c
at
i
on
i
n
c
om
m
u
ni
c
at
i
on
.
P
h
D
T
h
e
si
s.
S
han
dong
U
ni
v
er
s
i
t
y
;
2010.
[4
]
S
han X
i
ao
j
uan
.
O
n
t
he
app
l
i
c
at
i
on
of
i
nt
e
l
l
i
gent
c
om
p
ut
i
n
g i
n
ne
t
w
or
k
opt
i
m
i
z
at
i
on
.
Ph
D
T
h
e
s
is
.
S
hando
ng U
n
i
v
er
s
i
t
y
;
20
0
7.
[5
]
Z
hou
J
un
he.
O
n
D
N
A
e
nc
o
di
ng
ba
s
ed
on
hy
br
i
d
o
pt
i
m
i
z
at
i
on
a
l
gor
i
t
h
m
an
d
A
F
S
A
.
P
hD
T
h
e
s
i
s.
Z
hengz
ho
u U
ni
v
er
s
i
t
y
;
20
07.
[6
]
Li
Z
hi
w
u.
I
m
pr
ov
em
e
nt
of
A
S
F
A
and
i
t
s
ap
pl
i
c
at
i
on
i
n
w
i
r
el
es
s
s
en
s
or
c
ov
er
age
opt
i
m
i
z
a
t
i
on
.
P
h
D
T
h
e
si
s.
H
unan
U
ni
v
er
s
i
t
y
;
201
2.
[7
]
J
i
an
g M
i
ngy
an,
Y
uan D
ongf
en
g.
S
y
s
t
em
de
s
i
g
n of
ener
gy
ef
f
i
c
i
ent
-
b
as
e
d
w
i
r
el
es
s
s
en
s
or
net
w
o
rk
s
.
C
om
put
er
S
ys
t
e
m
s
.
201
0
;
7
(1
)
:
7
-
12.
[8
]
Yu
X
H
.
C
an
ba
c
k
pr
opa
gat
i
on
er
r
or
s
ur
f
ac
e
not
hav
e
l
oc
a
l
m
i
ni
m
s
.
N
e
ur
al
N
et
w
or
k
s
.
2
008
;
12(
3
)
:
1
009
-
1
021
.
[9
]
Li
S
ongy
i
n
g,
Z
he
n J
u
nl
i
.
F
or
w
ar
d m
u
l
t
i
l
ay
er
neur
a
l
net
w
or
k
f
uz
z
y
adapt
i
v
e al
gor
i
t
hm
.
A
c
t
a E
l
ec
t
r
oni
c
a
S
in
ic
a
.
20
0
9
;
23(
2)
:
1
-
6
.
[1
0
]
W
a
n
g
Z
h
eng
ou
.
A
v
al
i
d m
ut
i
l
ay
er
B
P
al
gor
i
t
hm
of
c
hang
e s
c
al
e
.
J
ou
r
n
al
o
f
T
i
anJ
i
n U
ni
v
er
s
i
t
y
.
200
9
;
29(
3)
:
364
-
36
9
.
[1
1
]
R
J
ac
obs
.
I
n
c
r
ea
s
ed R
at
es
o
f
C
onv
er
gen
c
e t
hr
ou
gh L
ear
n
i
ng R
at
e A
da
pt
at
i
on
.
N
eur
al
N
et
w
or
k
s
.
2014
;
1
(
4
):
3
1
-
38
.
[1
2
]
C
haam
b
ou
s
C
.
C
on
j
uga
t
e
gr
a
di
ent
al
g
or
i
t
hm
f
or
ef
f
i
c
i
ent
t
r
ai
ni
ng of
ar
t
i
f
i
c
i
al
neur
al
n
et
w
or
k
s
.
I
EEE
Pro
c
.
, P
a
r
t G
.
20
12
;
139
(
4
):
3
0
1
-
310
.
[1
3
]
Z
HA
NG
X
i
a
o
-
l
ong.
F
or
ec
as
t
i
ng m
et
h
od an
d appl
i
c
a
t
i
o
n of
B
T
P
bas
ed on neur
a
l
net
w
or
k
.
P
h
D
T
h
es
i
s
.
C
ent
r
a
l
S
out
h U
ni
v
er
s
i
t
y
;
2010.
[1
4
]
Y
an Lu,
D
ongx
i
ao N
i
u,
B
i
ngj
i
e
.
C
os
t
F
or
e
c
a
s
t
i
n
g M
odel
of
T
r
ans
m
i
s
s
i
on P
r
oj
ec
t
b
as
e
d on
t
he P
S
O
-
B
P
M
et
hod.
T
EL
KO
M
N
I
KA
T
el
ec
om
m
uni
c
at
i
on C
om
put
i
ng E
l
ec
t
r
o
ni
c
s
and C
o
nt
r
ol
.
2014
;
1
2(
4)
:
773
-
778.
[1
5
]
B
udi
m
an
P
A
,
R
ohm
an,
K
e
n
P
ar
am
ay
udha,
A
s
e
p
Y
udi
H
er
c
uad
i
.
A
N
ov
e
l
S
c
h
em
e
of
S
pee
c
h
E
nhan
c
em
ent
u
s
i
n
g P
ow
er
S
pec
t
r
al
S
ubt
r
ac
t
i
on
-
Mu
l
t
i
-
La
y
er
P
er
c
ept
r
o
n N
et
w
or
k
.
T
EL
KO
M
N
I
KA
T
el
e
c
om
m
uni
c
at
i
o
n C
om
put
i
n
g E
l
e
c
t
r
on
i
c
s
an
d C
ont
r
ol
.
201
6;
16(
1
)
:
1
81
-
18
6.
[1
6
]
Z
hou H
o
ng.
G
r
ay
ne
ur
al
net
w
or
k
an
d i
t
s
a
ppl
i
c
at
i
on
i
n t
he a
s
s
e
s
s
m
en
t
of
c
on
c
r
et
e
s
t
r
uc
t
u
r
es
u
s
i
n
g
.
P
h
D
T
h
e
si
s.
20
09.
Evaluation Warning : The document was created with Spire.PDF for Python.