TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.7, July 201
4, pp
. 5476 ~ 54
8
3
DOI: 10.115
9
1
/telkomni
ka.
v
12i7.530
9
5476
Re
cei
v
ed
De
cem
ber 8, 20
13; Re
vised
Febr
uary 13,
2014; Accept
ed Feb
r
ua
ry
25, 2014
An ICVBPNN Algorithm for Time-varying Channel
Trackin
g
and Prediction
Sufang Li, Ming
y
a
n Jiang
*
and Dong
feng Yuan
Schoo
l of Information Sci
enc
e and En
gi
neer
ing, Sha
n
d
ong
Univers
i
t
y
Jinan, 25
01
00
, China
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: jian
g
min
g
y
a
n
@
sdu.e
du.cn
A
b
st
r
a
ct
An i
m
prov
ed c
o
mpl
e
x-val
ued
back pro
p
a
gati
on ne
ura
l
netw
o
rk (ICVBPNN)
algor
ith
m
is pr
opos
e
d
in this p
a
p
e
r. In all
u
si
on to th
e defect of gr
a
d
ie
nt
desce
nt of traditio
nal c
o
mpl
e
x-val
ued
back pro
p
a
gat
io
n
netw
o
rk (CVB
PNN) al
gor
ith
m
, ad
ditiv
e
mo
me
ntu
m
h
a
s
b
een
introd
uce
d
.
It is used for time-varyi
ng c
h
ann
el
tracking
and prediction in
wirele
ss c
o
mm
unication system
and better
application res
u
lt
s ar
e ac
quired.
F
i
rstly, w
i
th the use
of the
le
arni
ng a
b
i
lity o
f
the ne
ural
ne
tw
ork, t
he tracking tra
i
ni
ng
is
started b
a
sed
o
n
the o
b
tain
ed c
han
nel
state i
n
formati
on (
C
SI), thus the
no
n
line
a
r ch
an
nel
mo
de
l is c
onst
r
ucted. Sec
ond
ly,
the unk
nown c
hannel st
ate information
is
predicted us
ing
the ICVBPNN trained
m
ode
l. The simulation
results de
mo
n
s
trate
that
th
e prop
osed
meth
od has
less
es
timate
d
error,
and
ca
n track
the c
h
a
nne
l
mo
r
e
accurate
ly than
the tradition
al
CVBPNN a
nd the Kal
m
an F
ilt
er alg
o
rith
m.
Ke
y
w
ords
: ICVBPNN, CSI, chan
nel tracki
ng
, channe
l pre
d
i
c
tion
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
The
com
p
lex
-
valued back pro
pagation neural
network (furt
her CVBPNN) is a strai
ght
forward generalization of the re
al-valued BPNN. The algorithm
whic
h is
us
ed for CVBPNN is
c
o
mplex-valued error back propagation
(CVEBP, c
a
ll
CVBP for s
h
ort) algorithm, whic
h is
a new
evolutiona
ry algorith
m
p
r
o
posed by
Nitt
a
&
Fu
ruya
i
n
19
91. It
ste
m
s from th
e
real-valu
ed
error
back
propagation algorithm. The
CVBP algorit
hm
has
been used for
c
o
mmunication s
i
gnal
pro
c
e
ssi
ng
a
nd a
daptive
chann
el e
quali
z
ation
[1], cl
a
ssifi
cation
of carotid arte
ry
Do
pple
r
sig
n
als
[2] and surface classification [3]. CVBPNN is
a hi
ghly nonlinear dynamical system
that
exhibits
a
rich
an
d com
p
lex dynami
c
al beh
avior [
4
]. They hav
e bee
n p
r
ove
n
better th
an
traditional
sig
nal
pro
c
e
ssi
ng m
e
thod
s in m
o
deling, p
r
e
d
icting nonli
nea
r an
d chaoti
c
time se
rie
s
and in
a
wid
e
variety of application
s
ra
ng
ing from spe
e
ch p
r
o
c
e
ssi
ng and
chan
nel equ
alizati
on [5]. Until now,
there a
r
e n
o
appli
c
ation
s
of chan
nel tracking a
nd p
r
edictio
n ba
se
d on the
com
p
lex-value
d
b
a
ck
prop
agatio
n neural network (CVBP
NN), we are inte
rested in loo
k
i
ng at the improveme
n
t of the
traditional
CV
BPNN alg
o
rit
h
m and its ap
plicatio
n in wi
rele
ss
comm
unication sy
stem.
The co
re of the tradition
al
CVBPNN al
gorithm
i
s
the gradi
ent de
scent method
[6]. Due
to the defe
c
t of gradi
ent d
e
scent
su
ch
as re
latively low
conve
r
ge
nce
sp
eed [7
] and bei
ng e
a
sy
to get into the local optima, an improv
ed CVBP
NN (ICVBPNN)
algorithm
is
proposed in this
pape
r. In this new
schem
e additive m
o
mentum i
s
prod
uced, which
can
be i
n
allusi
on to
the
defect of gradient descent
of the traditional CVBPNN.
In wirel
e
ss co
mmuni
cation
system, ma
n
y
tec
hniqu
es
requi
re the transmitte
r to know the
exact chan
ne
l state info
rm
ation (CSI) in
orde
r to
pl
ay their be
st pe
rforma
nce, th
e beam fo
rmi
ng
pre
c
odi
ng an
d adaptive
modulatio
n a
nd co
ding fo
r
example. I
n
time divisi
on multiplexi
ng
system,
CSI
can
be
a
c
qui
red by th
e
re
cipro
c
ity of
the
uplin
k/do
wnli
nk; in
the f
r
e
quen
cy divi
si
on
multiplexing
s
y
s
t
em, the
CSI es
timated by the rece
ive
r
is
fe
ed
ba
ck
e
d
to
the transmitter us
ing
the feedba
ck
link. Ho
weve
r, due to time-varying
pe
rfo
r
man
c
e of th
e wireless
ch
annel, the
r
e
will
be the time delay error b
e
twee
n the CSI obtained
by the recei
v
er and the
real CSI in th
e
transmitting ti
me, whi
c
h
wil
l
be eve
n
mo
re evid
ent
in
the fast fadi
n
g
chann
els.
To imp
r
ove t
he
system
pe
rfo
r
man
c
e, th
e
comp
en
satio
n
for th
e
CSI delay
will b
e
perfo
rme
d
u
s
ing
the
cha
nnel
predi
ction
te
chni
que
s. F
ading
time-v
arying
ch
an
nel
can
be
de
scribe
d
by Sine
wa
ve
sup
e
rp
ositio
n
process
or auto
r
eg
re
ssive
(AR) proce
s
s with
t
i
me-varyin
g
para
m
eters, so
spe
c
tru
m
e
s
timation plu
s
li
near
predicti
on mod
e
l or
sub
s
p
a
ce alg
o
rithm to p
r
e
d
ict the
CSI [8]
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
An ICVBPNN Algorithm for Time-v
ary
i
ng Chan
nel Trac
k
i
ng and Predic
tion (S
ufang Li)
5477
su
ch
as the
cha
nnel
pred
iction m
e
tho
d
s [9,
10] b
a
se
d o
n
the
Multiple Si
g
nal
Cla
ssifi
ca
tion
(MUSIC) [1
1] and
Estimati
ng Sign
al Pa
rameters via
Rotation
al In
varian
ce
Te
chniqu
e (ESP
RIT)
[12].
Real
-val
ued neu
ral n
e
twork ha
s
b
een wid
e
ly
u
s
ed
in
ch
ann
el predi
ction
becau
se it
h
a
s
better learni
ng ability, the
real
channel
predi
ction based on th
e Rosenblatt’s recurrent
neural
netwo
rk
and
the Vappni
k’s supp
ort ve
ctor m
a
ch
ine
(SVM) [13,
14], for exa
m
ple. Co
mpl
e
x-
valued neu
ral
network ha
s complex st
ructure, whi
c
h
has better l
earni
ng abilit
y, generali
z
at
ion
ability [10] an
d re
du
cing
a
b
ility [15] and
faster
co
nve
r
gen
ce
speed
[16] than the
tradition
al re
al-
valued neu
ral
network, thu
s
ha
s som
e
a
d
vantage
s su
ch a
s
less da
ta is need
ed
for the netwo
rk
training, l
o
we
r comp
utatio
nal
compl
e
xity and
so
on
, whi
c
h
can
be u
s
e
d
for
comm
uni
cati
o
n
sign
al pro
c
e
s
sing [17], su
ch as chann
el
equaliz
ation
[18], chann
el
estimation [1
9] and cha
n
n
e
l
predi
ction [2
0]. Currently, the often u
s
ed meth
o
d
s
f
o
r chan
nel
tracking and p
r
edi
ction
a
r
e as
follows, Kalm
an filter [21],
particl
e filter [22]
an
d L
e
a
s
t Mean
Squa
re
(LMS
) al
g
o
rithm [2
3] a
nd
so on. Pilot signal
s are ne
eded in [21], whi
c
h w
ill lea
d
to a waste of frequen
cy band resources,
and the Kal
m
an filter ne
ed
s lots
of data.
Pilot signal
s
are n
o
t nee
d
ed in [22], bu
t large
r
amo
u
n
t
of cal
c
ulatio
n
is cau
s
ed
b
e
ca
use ra
nd
om varia
b
le i
t
eration i
s
u
s
ed to comple
te the un
kno
w
n
distrib
u
tion. [23] has
slo
w
er conv
ergen
ce
spee
d, an
d it is not ap
plica
b
le in fa
st time varying
sy
st
em.
In orde
r to demon
strate t
he better pe
rf
orman
c
e of ICVBPNN, chann
el tracki
ng and
predi
ction b
a
s
ed o
n
it is introdu
ce
d in
this
pape
r. The metho
d
can avoi
d the trackin
g
a
nd
predi
ction
of
the
real
ch
annel
an
d i
m
agina
ry
ch
annel
respectively. The p
r
opo
se
d m
e
thod
con
s
i
s
ts
of two processe
s.
Firstl
y, with t
he u
s
e
of the
learning
abili
ty of the neu
ral network, th
e
tracking traini
ng is be
gun
based on the
obtained
ch
annel
state informatio
n (CSI), con
s
tru
c
ting
the nonli
nea
r
cha
nnel m
o
d
e
l. Seco
ndly, the forward
p
r
edi
ction to th
e future val
u
e
s
of the m
o
d
e
l
is contin
ued
with the use
of paramete
r
s obtai
ne
d b
y
network tra
i
ning and the
known ch
an
nel
parameters. Compari
ng wi
th
traditional CVBPNN
and the Kalman f
ilter, ICVBPNN algorithm
has
a better predi
ction an
d tracking p
e
rfo
r
ma
nce for th
e time-varyin
g
chann
el.
2. ICVBPNN Algorithm
For the ICVB
PNN alg
o
rith
m, we have to defi
ne the
error fun
c
tion
firstly [24]. Beca
use
there a
r
e
no
“greate
r
/le
s
s”
relation
s in th
e co
mplex-va
lued
ca
se, th
e output of th
e error fun
c
tion
must be a re
al-value
d nu
mber in orde
r to make it
p
o
ssible to evaluate the tra
i
ning re
sult a
nd to
guide it int
o
the di
re
ction
of an e
r
ror
re
ductio
n
. To b
egin
with, let’
s
con
s
ide
r
th
e equ
ation
s
f
o
r
informatio
n p
r
ocessin
g
in the thre
e-laye
r network
. It has in
gen
eral N extern
al
inputs a
nd L f
u
lly
interconn
ecte
d hidd
en u
n
its an
d M exte
rnal o
u
tputs.
[]
H
j
iL
N
w
W
is the
com
p
le
x weight ve
ctor
on hidd
en la
yer,
[]
O
kj
M
L
w
W
is the compl
e
x wei
ght vector o
n
the output layer. The in
put
vec
t
or
11
(,
,
)
pp
p
p
N
x
xx
X
and th
e output ve
ctor
11
(,
,
)
pp
p
p
M
Yy
y
y
are all
complex values.
The a
c
tivation of any (o
r
all) of the u
n
i
t
s in
the n
e
twork
ca
n be
consi
der
ed a
s
the output of
the
network
and
all the unit
s
can be
trained to
produce
desired outputs.
ICVBP
NN training
algorithm
can b
e
de
scri
bed a
s
follows:
Step 1: Initialization;
Step 2: Prese
n
t the input value
s
p
X
, and the de
sire
d ou
tput values
p
Y
;
Step 3: Calculate the
net-input valu
es to the
hidd
en
layer u
n
its
h
pj
net
, and the
outp
u
t
s
from the hidd
en layer
pj
i
;
,,
,
,
,
,
,
11
,,
,
,
,
1
()
()
NN
hh
h
h
h
h
h
h
pj
pj
R
p
j
I
j
i
pi
j
j
i
R
pi
R
j
i
I
pi
I
j
R
ii
N
hh
h
ji
R
p
i
I
ji
I
p
i
R
j
I
i
n
e
t
n
et
j
n
et
w
x
w
x
w
x
jw
x
w
x
(1)
Whe
r
e
ji
w
is th
e
com
p
lex wei
ght on th
e
co
nne
ction from
the
th
i
input un
it, and
h
j
is the
bias
term. The
“
h
” sup
e
rscript
refers to qu
an
tities on the
hidde
n layer.
The o
u
tput
of this hid
d
e
n
node i
s
:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5476 – 54
83
5478
,,
,
,
()
(
)
(
)
hh
hh
hh
p
j
p
j
R
pj
I
j
pj
j
p
j
R
j
p
j
I
i
i
j
i
F
n
et
f
n
et
j
f
net
(2)
Whe
r
e the
su
bscript
s
"
R
” a
nd the “
I
” refe
r to quantities on the real p
a
rt and the im
agina
ry part o
f
the values th
ese
sub
s
cript
s
are
written,
respe
c
tively.
Step 4:
Cal
c
u
l
ate the
net-i
nput valu
es,
o
p
k
ne
t
, to e
a
ch o
u
tp
ut layer unit
and th
e o
u
tp
uts,
pk
O
.
,,
,
,
()
(
)
(
)
o
o
oo
oo
p
k
p
k
R
pk
I
k
pk
k
p
k
R
k
p
k
I
O
O
j
O
F
n
et
f
n
et
j
f
n
e
t
(3)
Whe
r
e,
,,
,
,
,
,
,
11
,,
,
,
,
1
()
()
LL
oo
o
o
o
o
o
o
pk
pk
R
p
k
I
kj
pj
k
k
j
R
pi
R
k
j
I
kj
I
k
R
jj
L
oo
o
kj
R
p
i
I
k
j
I
k
j
R
k
I
j
net
n
et
j
n
et
w
i
q
w
i
w
i
q
jw
i
w
i
q
(4)
Whe
r
e ‘
o
’ sup
e
rscript refe
rs to quantities
on the output
layer.
Step 5: Calcu
l
ate the error
terms fo
r the output units.
''
,,
(
)
R
e
()
(
)
I
m
()
oo
o
o
o
p
k
k
pk
R
p
k
p
k
k
pk
I
p
k
p
k
fn
e
t
D
O
j
f
n
e
t
D
O
(5)
And the error
terms fo
r the hidde
n units.
''
,,
()
R
e
(
)
(
)
I
m
(
)
hh
h
h
h
p
j
j
pj
R
p
j
p
j
j
pj
I
p
j
p
j
f
ne
t
D
O
j
f
n
e
t
D
O
(6)
Step 6: Updat
e weig
hts on
the output layer acco
rdi
ng to:
*
(1
)
(
)
(
1
)
oo
o
o
k
j
kj
pk
pj
c
k
j
wt
wt
i
m
wt
(7)
And update
weights o
n
the hidde
n layer
according to:
*
(1
)
(
)
(
1
)
hh
h
h
ji
ji
pj
pi
c
j
i
wt
wt
x
m
wt
(8)
Whe
r
e the l
a
st item on th
e right ha
nd
side i
s
the m
o
mentum te
rm,
c
m
is the mo
mentum fa
ctor
whi
c
h is a
value betwe
en 0 and 1.
The mome
ntum term can rai
s
e the
standa
rd b
a
ck
prop
agatio
n a
l
gorithm
spe
e
d
by introdu
ci
ng the stabilit
y in the weigh
t
update.
Next, the ICVBPNN alg
o
r
ithm is
use
d
for
chan
ne
l tracking
an
d pre
d
ictio
n
. For the
ICVBPNN al
gorithm,
th
e activation
fu
n
c
tion (AF)
m
u
st be used
in its
com
p
lex version. T
h
e
compl
e
xactiv
ation functio
n
F for the network in this p
aper i
s
:
()
(
)
(
)
R
I
F
xf
x
j
f
x
.
(
9
)
The lin
ea
r fu
nction
is take
n a
s
o
u
tput l
a
yer AF,
and
the first-o
r
d
e
r
d
e
rivative o
f
it is
a
scalar. Th
e hi
dden laye
r AF is the sigm
oid functio
n
.
A
2
()
1
1
h
x
fx
e
.
(
1
0
)
The first
-
orde
r derivative of
the above formula is:
A
'
A2
A
()
1
(1
)
x
h
x
e
fx
e
.
(
1
1
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
An ICVBPNN Algorithm for Time-v
ary
i
ng Chan
nel Trac
k
i
ng and Predic
tion (S
ufang Li)
5479
3. The Adap
tiv
e
Channel Model
For a na
rrowband fadi
ng chann
el,
the sampled
re
cei
v
ed sign
al
()
rk
is given by:
()
(
)
()
(
)
rk
c
k
b
k
n
k
(12
)
Whe
r
e
ck
is o
b
tained
by sa
m
p
ling the
co
m
p
lex-value
d
fading
ch
ann
e
l
ct
at the time i
n
stant
of
T
tk
b
and
T
b
is the
data sym
bol
duratio
n,
bk
is
the
k
th
tra
n
smitted symbo
l
value, whil
e
nk
is
acomplex
-valued
di
screte AW
GN p
r
ocess
havin
g a va
rian
ce
of
0
/2
N
per di
me
nsio
n.
The ch
ann
el para
m
eter e
s
timation
()
ck
is obtained fro
m
the ch
annel e
s
timator. In order to have
a goo
d an
al
ysis, we a
ssume the
ch
a
nnel e
s
timati
on is a
ccu
ra
te, namely,
()
(
)
ck
ck
. The
probl
em to
b
e
re
solve
d
i
s
to pro
d
u
c
e th
e D-ste
p
forward
ch
ann
el
predi
ction
val
ues with th
e
use
of the current
chan
nel pa
ra
meters and th
e observed v
a
lue
s
.
Figure 1. System Block
Di
agra
m
The chan
nel
predi
ction i
s
divided into t
w
o
step
s,
whi
c
h i
s
sh
own i
n
Figu
re 1. Fi
rstly, the
c
u
rr
en
t c
h
an
ne
l C
S
I
ck
is estimated usi
ng the ICVBPNN,
namely, the cha
nnel net
work m
odel
is traine
d and
the nonlinea
r chann
el tracking mo
del
is const
r
u
c
ted; se
con
d
ly, the D-ste
p
forwa
r
d
predi
ction b
e
g
ins ma
kin
g
use of the pa
ramete
rs
o
b
tained by ICV
BPNN traini
n
g
and the kn
own
cha
nnel p
a
ra
meters
,1
,
1
ck
ck
c
k
P
.
4. ICVBPNN Tracking a
n
d Prediction
Model
4.1. ICVBPNN Trac
king Model
The tracke
r i
s
e
s
sentially
at
ype ofon
e
-
step
p
r
edi
ctor
whi
c
h h
a
s do
ctor traini
ng. The
unkno
wn sig
nal
ck
is lo
ok
ed
as
th
e des
ir
ed
s
i
g
n
a
l
,
th
e
in
pu
t s
i
gn
a
l
s
[1
,
(
2
)
,
,
ck
c
k
()
]
T
ck
P
den
otes the
compl
e
x-valu
ed in
put ve
ctor
of at the
time ind
e
x
k
. T
he a
r
chitectu
re of t
h
e
ICVBPNN tra
cki
ng mo
del i
s
sho
w
n in Fi
gure
2. The
b
a
si
c idea i
s
th
at weight valu
es a
r
e u
pdate
d
based o
n
the
error
ba
ck
propag
ation to t
he hid
den
l
a
yer u
n
til the weight value
s
unchan
ged, t
hen
the optimal p
a
ram
e
ters is
obtaine
d. The
back pro
pag
ation error i
s
:
(
)
()
()
e
k
ck
ck
.
(
1
3
)
Then the
cost
function of:
2
1
()
()
2
Ek
e
k
(14
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5476 – 54
83
5480
Is invoked
by a training algorithm
to
generate the updated
ICVBPNN weights until a
satisfa
c
to
rily
low mea
n
sq
u
a
re e
rro
r (MS
E
) is obtain
e
d
.
The pre
d
ictio
n
pro
c
e
ss m
odel belo
n
g
s
to a f
eed fo
rwa
r
d tra
c
kin
g
stru
cture, that is to
say
,
(
)
g
(
1
)
,(
2
)
,(
)
ck
ck
ck
ck
P
(
1
5
)
Whe
r
e
g
refers to ICVBPNN algorithm.
Figure 2. Cha
nnel Trackin
g
Model
4.2. ICVBPNN Prediction Model
Synchroni
ze the
training netwo
rk
p
a
ra
meters of the tracking p
r
oce
s
s
H
W
,
O
W
to th
e
backward p
r
edictio
n neu
ral network m
odel, an
d th
e
n
the predi
ction proc
ess i
s
op
erate
d
. The
predi
ction
p
r
oce
s
s m
odel
belo
n
g
s
to
a recursiv
elyforward p
r
edi
ction
structu
r
e which
can
be
s
e
en in Figure 1, that is
to
s
a
y,
(
)
g
(
1
)
,(
2
)
,(
)
,
1
ck
i
c
k
i
ck
i
c
k
i
D
(
1
6
)
Whe
r
e
g
refers to ICVBPNN algorithm.
The back propa
gation erro
r is:
()
()
(
)
,
1
ek
c
k
c
k
i
i
D
.
(
1
7
)
Then the
cost
function is:
2
1
()
()
2
Ek
e
k
.
(18)
In this
se
ctio
n, it is explai
ned the
re
sul
t
s of
research and
at the
same
time i
s
given the
comp
re
hen
si
ve discussio
n
.
Results can
be prese
n
te
d in figure
s
,
grap
hs, tabl
e
s
and
others
tha
t
make the
rea
der un
de
rsta
nd ea
sily [2,
5]. The discu
ssi
on can be
made in seve
ral su
b-ch
apt
ers.
5. Simulation Resul
t
s
Ja
ke
s
cha
n
n
e
l mod
e
l [25]
is u
s
e
d
to ve
rify the ch
ann
el tra
c
king
an
d p
r
edi
ction
a
l
gorithm
based on the
prop
osed ICV
BPNN algo
rit
h
m. Assum
e
that the maximum Dop
p
le
r frequen
cy sh
ift
6.4
d
f
Hz
, ca
rrie
r
f
r
eq
uen
cy is 2.3
G
Hz; the
Do
ppler
no
rmali
z
ation
coefficient is 0.32
0. The
architec
ture of the ICVBPNN in this
paper is
2-
80-1, whic
h means there
ar
e
2 nodes
in the input
layer, 80 no
d
e
s in the hi
d
den laye
r, 1 node in th
e o
u
tput layer. Since the
hidd
en layer h
a
s
the
function of st
oring the p
r
e
d
icted in
fo
rm
ation, the nu
mber of which can
not be
much little, in this
pape
r it is 80
. It cannot be
too larg
e or t
oo sm
all; oth
e
rwi
s
e th
e al
gorithm
will b
e
diverg
ed. T
h
e
Figure 3
den
otes th
e
conv
erge
nce
curv
es fo
r diffe
re
nt hidd
en lay
e
r n
ode
s. Th
e a
c
tivating facto
r
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
An ICVBPNN Algorithm for Time-v
ary
i
ng Chan
nel Trac
k
i
ng and Predic
tion (S
ufang Li)
5481
of the hidden
layer activatio
n
function i
s
2 and the
line
a
r sl
ope fa
cto
r
of the outpu
t layer is 3. For
different nu
m
ber of th
e del
ayed ste
p
s, t
he tra
cki
ng e
r
ror i
s
differen
t, which i
s
sh
own i
n
Figu
re
4.
The greater t
he delaye
d
st
ep, t
he bigge
r the predi
ctio
n error.
Figure 3. The
Trackin
g
Error ca
used by
Different Nu
mber of Hi
dd
en No
de
s
Figure 4. The
Trackin
g
Error ca
used by
Different Nu
mber of Dela
yed Step
For
3
D
whi
c
h
m
ean
s the
del
ayed
step i
s
3, t
he ob
se
rv
ed
signal
that
is the
del
ayed
CSI is the
d
o
tted line
a
s
sh
own in
Fi
gure
5. Fi
rstly, the ob
se
rved si
gnal
is train
ed
by the
ICVBPNN to
con
s
tru
c
t th
e nonlin
ear
cha
nnel tracking n
e
two
r
k model. Figu
re 5
sho
w
s
the
training
resul
t
s. Obvio
u
sly
the
estimati
on
er
ror bet
wee
n
the training re
sult
s and the ob
se
rved
sign
als is le
ss tha
n
that between th
e traini
ng re
sults a
nd the ideal CSI. Thus the D-ste
p
predi
ction
m
u
st b
e
p
r
oce
s
sed u
s
in
g t
he optim
al
p
a
ram
e
ters o
b
t
ained by th
e
netwo
rk trai
ning
and the
kno
w
n chan
nel p
a
ram
e
ters. T
he predi
ction
result of the compl
e
x cha
nnel is
sho
w
n in
the Figure 6.
The algorithm for c
h
annel predic
t
ion is c
o
mpared with
traditional CVBPNN
and
Kalman Filte
r
[21] in Figu
re 7. The Kal
m
an Filt
e
r
u
s
es the A
R
ch
annel m
odel
and the
ord
e
r is 8,
12
re
spe
c
tively.The hig
her the
ord
e
r th
e smalle
r
the
predi
ction
error. B
u
t with
the in
crea
se
d
numbe
r
of th
e o
r
de
r, the
a
m
ount
of the
data
req
u
ir
e
d
is rel
a
tively increa
sed.
Th
e p
r
edi
ction
e
rro
r
of the ICVBPNN i
s
le
ss th
an Kalman
Fi
lter wh
en th
e
orde
r the
ch
annel m
odel i
s
eq
ual o
r
le
ss
than 12.
(a)
Real p
a
rt trac
kin
g
(b) Ima
ge pa
rt tracki
ng
Figure 5. Tra
cki
ng Perfo
r
mance of the Compl
e
x Cha
nnel
3900
3920
3940
3960
3
980
4
000
0
0.
2
0.
4
0.
6
0.
8
1
N
u
m
b
er
of
I
t
e
r
at
i
ons
T
r
ac
k
i
n
g
er
r
o
r
H
i
dd
en Lay
e
r
=
2
0
H
i
dd
en Lay
e
r
=
5
0
H
i
dd
en Lay
e
r
=
8
0
H
i
dd
en Lay
e
r
=
1
2
0
3900
39
20
394
0
3960
3980
4000
0
0.2
0.4
0.6
0.8
1
N
u
m
ber
of
i
t
er
ati
ons
P
r
edi
c
t
i
on er
r
o
r
N
o
P
r
edi
c
t
i
o
n
D=
3
D=
2
D=
1
3900
3920
3940
3960
3
980
4
000
-2
0
2
4
N
u
m
ber
of
i
t
er
a
t
i
o
ns
C
h
an
ne
l
V
a
l
u
e
R
e
a
l
P
a
r
t
s
o
f
I
deal
C
h
a
nnel
I
d
eal
C
han
nel
V
a
l
u
e
D
e
l
a
y
ed 3 S
y
m
b
o
l
s
T
r
ac
k
i
ng
C
h
a
nnel
3900
3920
3940
3960
3
980
4
000
-2
0
2
4
N
u
m
ber
of
i
t
er
a
t
i
o
ns
C
h
an
ne
l
V
a
l
u
e
I
m
a
ge P
a
r
t
s
of
C
han
nel
V
a
l
u
e
I
d
eal
C
han
nel
V
a
l
u
e
D
e
l
a
y
e
d
3
S
y
m
b
ol
s
T
r
ac
k
i
ng
C
h
a
nnel
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5476 – 54
83
5482
The running t
i
me of the ICVBPNN and
CVBPNN i
s
6.140s, 6.105s. As
for the Kalman
Filter, whe
n
the cha
nnel
orde
r is 8 a
nd
12 the runnin
g
time is 35.81
20
s and 40.4
9
1
5
s
respe
c
tively.
Obviou
sly, ICVBPNN nee
d
s
less
data a
nd run
n
ing ti
me than the Kalman Filter.
Figure 7 also sho
w
s that
the ICVBPNN has b
e
tter stability and less tra
c
kin
g
erro
r than
the
traditional
CVBPNN. BER analysi
s
for
BPSK modulat
ion with 2×2 MIMO channel is shown in
Figure 8, fro
m
which we
can se
e t
he BER is improved after predi
ction.
(a)
Real p
a
rt predi
ction
(b) Ima
ge pa
rt predictio
n
Figure 6. Pre
d
iction Pe
rformance of the Compl
e
x Cha
nnel
Figure 7. Co
mpari
s
o
n
of the Tra
c
king
Erro
r
betwe
en Kal
m
an Filter, CVBPNN and
ICVBPNN
Figure 8. BER Plot for 2×2 MIMO Cha
nnel
with ICVBPNN Track
i
ng and Predic
t
ion
6. Conclusio
n
In this pa
per,
a narro
wba
n
d
fading
cha
nnel tra
c
king
and p
r
edi
ctio
n algo
rithm b
a
se
d on
ICVBPNN i
s
proposed. Bit error
rate
analysi
s
for B
PSK modulat
ion in Gaussi
an white noi
se
Ja
ke
s chan
n
e
l verifies th
e co
rrectn
ess of
the p
r
o
posed al
gorit
hm. The
si
mulation
re
sults
demonstrate
that the ICV
BP
NN
has better stability than tradi
tional
CVBPNN. The ICVBP
NN
training
sch
e
m
e
conve
r
ge
s fa
ster than
tradi
tional
CV
BPNN and
Kalman Filter and nee
ds
le
ss
runni
ng tim
e
. For the high-speed parallel perf
orm
a
nce of the ICVBPNN,
the running
speed
increa
se
s an
d real
-time proce
s
sing of
time-varyin
g
chann
el is a
c
hi
eved.
390
0
3
920
394
0
3
960
398
0
4
000
-2
0
2
4
N
u
m
b
er
o
f
i
t
er
ati
o
n
s
C
hanne
l
V
a
l
u
e
R
e
a
l
P
a
r
t
s
of
C
h
an
ne
l
V
a
l
u
e
D
e
s
i
r
ed C
h
an
ne
l
P
r
ed
i
c
t
i
o
n
C
h
an
nel
3900
3920
3940
3960
3
980
4
000
-2
0
2
4
N
u
m
ber
of
i
t
er
a
t
i
o
ns
C
h
an
ne
l
V
a
l
u
e
I
m
a
ge P
a
r
t
s
of
C
han
nel
V
a
l
u
e
D
e
s
i
r
e
d
C
h
a
nnel
P
r
ed
i
c
ti
on C
hann
el
3900
3920
3940
3960
3
980
4
000
0
0.
1
0.
2
0.
3
0.
4
0.
5
N
u
m
b
er
of
I
t
e
r
at
i
ons
P
r
e
d
i
c
t
i
on
er
r
o
r
Ka
lm
a
n
F
i
lt
e
r
,
P
=8
Ka
lm
a
n
F
i
lt
e
r
,
P
=1
2
CV
B
P
N
N
ICV
B
P
N
N
0
5
10
15
20
25
10
-4
10
-3
10
-2
10
-1
SN
R
/
d
B
BER
I
deal
c
h
ann
el
(
T
x
=
R
x
=
2
)
T
r
ac
k
i
n
g
c
han
nel
(
T
x
=
R
x
=
2
)
P
r
ed
i
c
t
i
on c
h
a
nnel
(
T
x
=
R
x
=
2
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
An ICVBPNN Algorithm for Time-v
ary
i
ng Chan
nel Trac
k
i
ng and Predic
tion (S
ufang Li)
5483
Referen
ces
[1]
Y Che
o
l
w
o
o
,
H Daes
ik. No
nlin
ear
bli
nd
equ
aliz
atio
n s
c
hemes
usin
g
comple
x-v
a
l
u
ed multi
l
a
y
e
r
feedfor
w
a
rd n
e
u
ral n
e
t
w
orks.
IEEE Transactions on
Neur
al Networks
. 199
8; 9: 1442-
145
5.
[2]
M Ce
yla
n
, R Ce
yla
n
, F Dirge
nali, S Kara, Y
Ozba
y
.
Cl
assi
fication of caro
tid arter
y
D
o
p
p
l
er sig
nals i
n
the e
a
rl
y p
h
a
s
e of ath
e
ros
c
lerosis
usi
n
g
compl
e
x-val
u
ed artific
i
al
n
e
u
ral
net
w
o
rk.
Co
mp
uters i
n
Biol
ogy an
d Medici
ne
. 2
007;
37; 28-3
6
.
[3]
A Prashanth,
PK Kalra, NS V
y
as.
Surfac
e classificati
on
using ANN a
nd co
mpl
e
x-va
lue
d
neur
a
l
netw
o
rk.
Proce
edi
ngs of th
e 9
t
h Internati
ona
l
Confer
enc
e
o
n
Ne
ural Inf
o
rmation Pr
oces
sing. 2
0
0
2
; 3:
109
4-10
98.
[4]
H Jin, W
Jun.
Globa
l Stabi
lit
y of Compl
e
x-V
a
lu
ed R
e
curre
nt Neur
al N
e
t
w
orks
w
i
th T
i
me-Dela
y
s.
IEEE
T
r
ansactio
n
s o
n
Neur
al Netw
orks and L
ear
n
i
ng Syste
m
s
. 2
012; 23; 8
53-8
65.
[5]
A Hirose, S Yo
shid
a. Genera
l
i
z
ation C
haract
e
rist
ics of Com
p
le
x-V
a
lu
ed F
e
edfor
w
a
rd Ne
u
r
al Net
w
orks
in Re
latio
n
to Sign
al Co
her
enc
e.
IEEE Transactions on N
eur
al Netw
orks an
d Lear
ni
ng Sys
t
ems
. 20
12
;
23; 541-
55
1.
[6]
T
Nitta. An Extensio
n of t
h
e
Back-Pro
pag
a
t
ion
A
l
gor
ithm
to Com
p
le
x N
u
mbers. N
eur
a
l
Net
w
o
r
ks.
199
7; 10: 139
1
-
141
5.
[7]
K Burse, A P
a
ndey
,
A S
o
mk
u
w
ar.
C
onv
erg
ence A
n
a
l
ysis
of Co
mp
lex V
a
lu
ed M
u
ltip
lic
ative N
eur
al
Netw
ork for Various Activati
on
F
unctions.
In
te
rn
a
t
io
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
Comp
u
t
a
t
i
o
n
a
l
In
te
l
l
i
g
e
n
c
e
and
Commun
i
cati
o
n
Net
w
orks (CI
CN). 201
1: 279
-282.
[8]
A Duel-
H
al
le
n, H Hal
l
en, Y T
ung-Sh
en
g. Lon
g rang
e pr
edict
ion a
nd r
educ
e
d
feed
back for
mobil
e
ra
di
o
ada
ptive OF
DM s
y
stems.
IEEE Transactions on Wireless
Communic
a
tions
. 2006; 5; 27
23-2
733.
[9]
HA a J
P
D
e
l
m
as. MUSIC-
Like
Estimatio
n
of
Directi
on
of Arriva
l. IE
EE T
r
ansactio
n
s On S
i
gn
al
Processi
ng. 20
06; 54: 26
78-2
690.
[10]
A Hirose. N
a
tu
re of compl
e
x num
b
e
r an
d co
mple
x-va
lu
ed
neur
al n
e
t
w
ork
s
.
F
r
ontiers of Electrical
an
d
Electron
ic Engi
neer
ing i
n
Ch
in
a
. 2011; 6: 17
1
-
180.
[11]
CR Berger, Z
Shengli, JC Preisig,
P Willett. S
parse Channel Estimation
fo
r Multicarrier Under
w
a
t
e
r
Acoustic Com
m
unic
a
tion: F
r
om Subsp
a
ce
Methods to C
o
mpresse
d Sen
s
ing.
IEEE Transactions
on
Sign
al Process
i
ng
. 20
10; 58:
170
8-17
21.
[12]
NY a B Friedla
nder. As
y
m
pto
p
icp
e
rformanc
e
ana
l
y
sis of ESPRIT, Higher
Order ESPRIT, and Virtua
l
ESPRIT Algorit
hm.
IEEE Transaction on Signal Pric
ession
. 199
6; 44: 253
7
-
255
0.
[13]
B Vis
w
e
s
w
a
r
a
n, T
Kiran.
C
han
nel
pr
edict
ion
bas
ed
po
w
e
r control
in
W
-
CDMA sy
stems
. First
Internatio
na
l C
onfere
n
ce o
n
(Conf. Publ. N
o
. 471
3G Mob
i
l
e
Commu
nicati
on T
e
chnol
ogi
es). 200
0: 41
-
45.
[14]
NI Sap
ank
ev
yc
h, R S
ankar. T
i
me S
e
ries
Pre
d
ictio
n
Us
ing
S
upp
ort Vector
Machi
nes: A S
u
rve
y
.
IEEE
Co
mp
utation
a
l Intelli
genc
e
Ma
ga
z
i
n
e
. 20
09; 4: 24-38.
[15]
M
Koba
yas
h
i. Exce
ptio
nal R
educ
ibi
lit
y
of Compl
e
x-Va
lu
ed Ne
ural
Net
w
o
r
ks.
IEEE Transactions on
Neur
al Netw
or
ks
. 2010; 21: 1
060-
107
2.
[16]
H Z
h
a
ng,
D
Xu, Z
W
ang,
Co
nverg
ence
of
a
n
On
l
i
ne
Sp
lit-
C
ompl
e
x
Gra
d
i
ent Al
gorit
hm f
o
r C
o
mpl
e
x-
Valu
ed Ne
ura
l
Net
w
orks.
Disc
r
ete Dyna
mics in Natur
e
an
d Society.
201
0; 1-27.
[17]
A Hirose, S Yo
shid
a. Genera
l
i
z
ation C
haract
e
ristics of Com
p
le
x-V
a
lu
ed F
e
edfor
w
a
rd Ne
u
r
al Net
w
orks
in Re
latio
n
to Sign
al Co
her
enc
e.
IEEE Transactions on N
eur
al Netw
orks an
d Lear
ni
ng Sys
t
ems
. 20
12
;
23: 541-
55
1.
[18]
HQ Z
hao,
XP
Z
eng, Z
Y
He,
W
D
Jin, T
R
Li
.
Compl
e
x-val
u
ed p
i
p
e
li
ne
d d
e
cisio
n
fe
edb
a
ck recurre
nt
neur
al net
w
o
rk
for non-li
ne
ar chan
nel
equ
ali
z
ation.
IET
Commu
n
icati
ons
.
201
2; 6: 1082-
109
6.
[19]
Z
Xia
o
b
o
, W
Xi
ao
don
g. Ch
a
nne
l estimati
o
n
for
OF
DM systems us
in
g
ada
ptive ra
di
al
basis fu
nction
net
w
o
rks.
IEEE Transactions
on Vehic
u
lar Technology
. 200
3; 52: 48-5
9
.
[20]
G Su Lee, DP Mandic. Non
l
i
near a
daptiv
e pred
iction of c
o
mpl
e
x-val
ued
signals b
y
co
mple
x-va
lu
e
d
PRNN.
IEEE Transactions on Signal Process
i
ng
. 20
05; 53:
182
7-18
36.
[21]
Y Z
heng. A no
vel cha
n
n
e
l est
i
matio
n
an
d tra
cking meth
od f
o
r
w
i
re
less OF
DM s
y
stems b
a
sed
on p
ilot
s
and Ka
lma
n
filterin
g.
IEEE Tr
ansactions on Cons
um
er Electronics.
200
3; 49: 275-
28
3.
[22]
JIANG MY, YUAN DF
. Blind
tracking bas
e
d
on Mont
e C
a
rlo a
l
gor
ithm over MIMO fast time-var
y
i
n
g
chan
nels.
Jo
ur
nal o
n
Co
mmu
n
icati
ons
. 20
07
; 28: 40-44.
[23] JF
Galdin
o, EL
Pinto, MS d
e
Alenc
ar. Ana
l
y
t
ic
al
performa
n
c
e of the
LMS
alg
o
rithm o
n
th
e estimati
on
of w
i
de
se
nse stationar
y
ch
an
nels.
IEEE Transactions on Communic
a
tions
.
2004; 52: 98
2
-
991.
[24]
A Min
i
n. C
o
mp
le
x V
a
lu
ed
Re
current
Neur
al
Net
w
ork: F
r
om
Architectur
e
t
o
T
r
ainin
g
.
Jo
u
r
nal
of S
i
gn
a
l
and Infor
m
ati
o
n Processi
ng
. 201
2; 03: 192-
197.
[25]
YR Z
hen
g,
X
Che
ngsh
an. S
i
mulati
on m
ode
ls
w
i
th
correct
statistical
pro
perties f
o
r Ra
yl
eig
h
fa
din
g
chan
nels.
IEE
E
T
r
ansactions
on Co
mmu
n
ic
ations
. 20
03; 5
1
; 920-9
28.
Evaluation Warning : The document was created with Spire.PDF for Python.