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.
10
,
No.
3
,
June
2020,
pp. 2
997~
3006
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v10
i
3
.
pp2997
-
30
06
2997
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om/i
nd
ex
.ph
p/IJ
ECE
Optimi
zed BER
fo
r c
h
an
nel equa
lizer usi
ng cucko
o search
and
neural n
etwork
Swati
Katw
al
1
,
Vin
ay Bh
at
i
a
2
1
Depa
rtment
of Electronics a
nd
Com
m
unic
at
ion Engi
ne
eri
ng
,
Baddi
Univ
ersity
of
Emerg
ing
S
ci
en
ce
s
and
T
echnolog
y
,
Ind
ia
2
Depa
rtment of
El
e
ct
roni
cs
and
Com
m
unic
at
ion Engi
ne
eri
ng,
Ch
andi
gar
h
Engi
n
e
eri
ng
Co
ll
eg
e, I
n
dia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
pr
16
, 201
9
Re
vised
Dec
5
,
2019
Accepte
d
Dec
10
, 201
9
The
dig
ital
d
at
a
tra
nsfer
faces
i
ss
ues
reg
ard
ing
Inte
r
-
S
y
m
bol
I
nte
rfe
r
ence
(ISI);
t
her
efo
re
,
the
err
or
r
at
e
be
comes
depe
nden
t
upon
c
hannel
esti
m
at
io
n
and
it
s
equa
l
izat
ion.
Thi
s
pap
er
f
ocuse
s
on
the
de
vel
opm
ent
of
a
m
et
hod
for
opti
m
iz
ing
th
e
c
hanne
l
da
ta
to
i
m
prove
ISI
by
u
ti
lizing
a
sw
arm
int
el
l
ige
n
ce
serie
s
al
gori
thm
te
rm
ed
as
Cuckoo
Se
arc
h
(CS).
The
ad
juste
d
dat
a
throug
h
CS
is
cro
ss
-
val
ida
t
ed
using
Ar
ti
ficia
l
Neura
l
Network
(AN
N
).
Th
e
data
ac
c
ept
an
ce
rate
is
conside
red
with
0
-
10%
m
arg
ina
l
err
or
which
var
ie
s
i
n
the
giv
en
r
ange
with
di
ffe
ren
t
bit
stre
ams
.
Th
e
per
form
ance
e
val
ua
ti
on
o
f
t
he
proposed
a
l
gorit
hm
using
the
Avera
g
e
Bit
Err
or
Rate
(A
-
BER)
and
Loga
ri
thmic
Bit
Err
or
Rat
e
(L
-
BER)
had
show
n
an
over
al
l
imp
rove
m
ent
of
30
-
50%
when co
m
par
ed
with
the
Kalman
filter
ba
sed
al
gor
it
hm
.
Ke
yw
or
d
s
:
ANN
Chan
nel
est
im
at
ion
CS
Eq
ualiz
at
ion
Co
pyright
©
202
0
Instit
ute of
Ad
v
ance
d
Engi
ne
eri
ng
and
Sc
ie
n
ce
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Sw
at
i Kat
wal
,
Dep
a
rtm
ent o
f El
ect
ro
nics
and C
omm
un
ic
ation
En
gin
ee
rin
g
,
Ba
dd
i
U
niv
er
sit
y of
Em
erg
in
g Sci
ences
and
Tech
no
l
og
y,
So
la
n
, In
dia
.
Em
a
il
:
eng
g.s
wati
@yah
oo.c
o.
in
1.
INTROD
U
CTION
Digital
Com
m
un
ic
at
io
n
(D
C
)
pro
vid
es
di
gital
data
tra
ns
fe
r
with h
ig
h
tran
sfer
rate [1
-
2]. Th
e
a
verage
powe
r
of
the
t
ran
sm
it
te
d
sign
al
re
du
ces
as
the
distance
be
tw
een
tra
ns
m
it
te
r
and
recei
ver
s
urges
.
Th
e
no
is
e
com
es
in
the
f
or
m
of
distu
rbance
a
nd
I
nter
-
Sy
m
bo
l
I
nter
fe
ren
ce
(
IS
I
)
in
the
r
ecei
ve
d
sig
nal.
Fi
gure
1
s
how
s
the
occurre
nce
of
ISI
in
the
receive
d
sign
al
.
It
is
seen
that
du
e
to
m
ultip
at
h
pro
pa
gation
of
the
tra
ns
m
itted
sign
al
,
the
stre
ng
t
h
of
the
rec
ei
ved
sig
nal
re
du
ce
s
du
e
to
the
pr
es
ence
of
buil
dings
a
nd
ano
t
her
obsta
cl
e
li
ke
ai
rp
la
ne
and t
r
ees et
c.
In
this
pap
e
r,
t
he
aut
hors
ha
ve
addresse
d
th
e
IS
I
distor
ti
on
s
and
pro
po
se
an
ap
proac
h
to
reduce
this
unwa
nted
p
he
nom
eno
n
to
im
pro
ve
the
reli
abili
ty
of
the
co
m
m
un
ic
at
ion
.
Ba
sic
al
ly
,
there
are
two
m
ai
n
causes
of
occurri
ng
I
SI
.
T
hese
are
non
-
li
near
f
re
quency
betwee
n
the
channels
and
m
ulti
path
pro
pag
at
io
n.
Ma
ny
at
tem
pts
hav
e
been
m
ade
by
the
researc
he
rs
to
reduce
the
in
te
rf
e
ren
ce
e
f
fect.
The
re
are
var
io
us
al
gori
thm
s
and
filt
er
pr
opos
e
d
by
sc
hola
rs
to
c
om
bat
the
interfe
rence
eff
ect
.
For
inst
ance,
Kalm
an
Fil
te
r
(KF)
is
widely
us
e
d
to
re
duc
e
the
ISI
a
nd
channel
inter
fer
e
nce
ef
fects
.
This
filt
er
is
us
e
d
to
li
ne
arize
the
no
nl
i
nea
r
m
od
el
s
[3
]
.
If
the
data
is
well
structu
re
d,
it
resu
lt
s
in
a
balance
d
cha
nnel
est
i
m
at
ion
w
hich
res
ults
in
a
m
ini
m
u
m
er
ror
rate.
Although
K
F
an
d
Exten
de
d
Kal
m
an
Fil
te
r
(EK
F
)
a
re
use
d
world
wide
to
recti
fy
the
cha
nn
el
da
ta
,
it
la
cks
in
a
dap
ti
ve
filt
erin
g
[
4].
Eq
ualiz
ers
are
use
d
in
t
he
ada
ptive
filt
ers.
Fig
ure
2
s
how
s
the
use
of
e
qu
al
iz
er
betwee
n
the
tra
ns
m
itte
r
an
d
t
he
rece
iver.
The
data
is
trans
ferre
d
from
transm
i
t
te
r
to
receiver
.
B
ut,
t
he
tra
nsfer
dat
a
stream
cann
ot
be
kep
t
hom
og
en
ous
t
hroug
hout
[5
-
6].
For
est
i
m
at
ing
the
tim
e
-
var
yi
ng
cha
nn
el
coef
fici
ents
diff
e
ren
t
al
gorithm
s
are
avail
able.
D
ue
to
the
si
m
pl
i
ci
ty
of
the
adap
ti
ve
al
gorithm
,
the
coe
ff
ic
ie
nts
of
the
cha
nne
l
are
est
im
at
e
d
us
in
g
Least
Me
an
Squa
r
e
(LMS
)
al
gorithm
.
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.
10
, No
.
3
,
J
une
2020 :
29
97
-
3006
2998
LMS
ada
ptive
al
gorithm
aim
s
to
m
ini
m
i
ze
the
er
r
or
r
el
yi
ng
on
t
he
netw
ork.
Th
e
filt
er
has
a
on
e
-
ste
p
pr
e
dictor
-
c
orre
ct
or
st
ru
ct
ur
e
a
nd m
ini
m
iz
es t
he
m
ean square est
i
m
at
ion
er
ror
in
each
step
[7
-
9].
This
pap
e
r
c
om
par
es
the
KF
base
d
e
qu
al
iz
er
with
EK
F
ba
sed
for
ti
m
e
-
var
yi
ng
c
ha
nn
el
s
by
up
datin
g
it
s
coef
fici
ents
.By
util
iz
ing
KF
only
it
m
a
y
no
t
be
the
best
so
l
ution
f
or
the
fu
t
ur
e
com
plex
dem
a
nd
s
of
channels
inter
f
eren
ce
,
noise
,
and
r
ed
unda
nc
ie
s.
The
pa
per
pro
po
ses
th
e
est
ablishm
ent
of
a
hybr
i
d
m
e
t
hod
of
Sw
arm
In
te
ll
i
gen
ce
(CS)
a
nd
m
achine
le
arn
i
ng
based
m
echan
ism
.
C
S
is
a
har
d
thres
hold
opti
m
iz
at
ion
te
chn
iq
ue
a
nd
hen
ce
t
o
recti
f
y
the
bit
strea
m
,
CS
is
utiliz
ed
he
re
[10].
A
ne
w
fitness
functi
on
is
des
ign
e
d,
i
m
ple
m
ented,
and
c
r
os
s
-
vali
dated
us
in
g
fe
ed
f
orwa
rd
ba
ck
pr
op
a
gatio
n
Ne
ural
Net
work
(
NN
).
A
rtific
ia
l
Neural
Net
work
(AN
N)
is
bein
g
us
ed
al
l
ov
e
r
in
di
gital
com
m
un
ic
a
ti
on
[
11
-
12
]
.
Fu
rt
her
m
ulti
-
m
od
el
per
ce
ptr
on
m
od
el
w
hich
is
an
exte
ns
io
n
of
un
i
-
m
od
el
perce
ptr
on
m
od
el
is
al
so
us
e
d
to
equ
al
iz
e
the
c
ha
nn
el
respo
ns
e
[
1
3
]
.
The
propose
d
work
t
ook
t
he
insp
irat
io
n
fro
m
the
util
iz
at
i
on
of
m
ulti
-
m
od
el
per
ce
ptr
on
m
od
e
l
and
ai
m
ed
to
de
sign
a
bette
r
m
ul
ti
-
m
od
el
ne
twork
to
opti
m
iz
e
the
channel
res
pons
e.
ANN
is
com
bin
ed
with
swar
m
intel
li
gen
ce
to
e
qual
iz
e
the
cha
nnel
[1
4
-
15
].
The
li
te
ratu
re
rev
ie
w
is
prese
n
te
d
i
n
se
ct
ion
2.
The
propose
d
m
od
el
and
it
s
validat
io
n
us
i
ng
A
N
N
are
des
cribe
d
in
Se
ct
ion
3.
The
eval
uation
of
the
pro
posed
work
m
od
el
and
a
com
par
is
on
with
KF
al
ong
with
re
su
l
tsi
s
al
so
prese
nted
in
Sect
io
n
4.
T
he
co
ncl
us
io
n
of
the p
a
pe
r
is
pr
e
sented
in
Sect
ion 5
.
Figure
1. Ef
fec
t of ISI
on the
re
cei
ved
sig
nal
Figure
2. Use
of
eq
ualiz
er
t
o rem
ov
e the
noise
level
2.
BACKG
ROU
ND R
E
VIEW
In
a
dig
it
al
com
m
un
ic
at
ion
syst
e
m
,
ban
dwidt
h
is
a
scarce
resou
rce
in
the
m
od
e
r
n
world
.
The
ef
fici
ent
util
iz
at
ion
of
this
res
ource
is
vital
in
these
days.
T
he
data
tran
sfe
r
rate
de
pends
upon
the
util
iz
at
ion
of
ba
ndwi
dth
.
T
he
stren
gt
h
of
the
si
gnal
transm
it
ted
over
the
c
hannelre
duces
du
e
to
the
pr
ese
nce
of
var
i
ous
obsta
cl
es
du
ri
ng
tra
ns
m
issi
on
.Mo
r
eov
e
r,
the
si
gnal
qu
al
it
y
al
so
dam
pen
s
to
a
la
rg
e
extent
due
to
the
presence
of
m
any
oth
er
frequ
e
ncies
in
th
e
channel.
D
ue
to
the
pr
ese
nc
e
of
ob
j
ect
s,
I
SI
a
nd
no
ise
occur
w
it
h
the
input
sign
al
at
the
re
cei
ver
en
d.
Ma
ny
filt
ers
are
us
e
d
to
re
du
ce
these
disturba
nces
.
But
nonlinea
r
filt
ers
are
ins
uffici
ent
to
re
duce
the
le
vel
of
noise
.
T
her
e
f
or
e
,
li
near
filt
ers
su
c
h
as
Ad
a
ptive
filt
ers
are
in
tr
end,
us
ed
to
re
du
ce
the
le
ve
l
of
noise
inten
s
it
y.
Equ
al
iz
ers
are
use
d
i
n
the
se
filt
ers
f
or
str
uctu
ral
dev
el
op
m
ent.
Eq
ualiz
ersar
e
us
e
d
basica
ll
y
for
cha
nnel
equ
al
iz
at
io
n.
It
is
a
pro
cess
w
hich
r
edu
c
e
s
the
inter
fer
e
nc
e
le
vel
to
the
desire
d
a
m
ou
nt.Th
e
pr
ocess
of
c
ha
nn
el
e
qual
iz
at
ion
is
car
ried
out
at
the
destinat
io
n
end
w
hic
h
m
od
i
fies
the
sig
nal
qu
al
it
y
an
d
re
du
ces
t
he
interfe
ren
ce
e
f
fect.
De
ng
e
qual
iz
es
the
noise
le
vel
us
i
ng
the
Co
m
plex
-
Value
d
Mi
ni
m
al
Ra
dial
Ba
sis
Functi
on
Ne
ural
Net
works
[
1
6
]
.
B
esi
des,
researc
hers
an
d
pract
it
ion
e
rs
al
so
de
velo
p
filt
ers
an
d
al
gorith
m
s
to
equal
iz
e
the
no
ise
le
vel
.
The
de
velo
pe
d
filt
er
re
du
c
es
t
he
bit
er
ror
ra
te
by
s
uppress
ing
the
m
ajor
cal
a
m
ities
li
ke
ISI
a
nd
wh
it
e
noise
.
I
n
a
ddit
ion
,
Ca
nd
y
use
s
the
Kalm
an
filt
er
to
achieve
opti
m
u
m
resu
lt
s.
T
his
filt
er
is
us
e
d
with
Re
c
ur
si
ve
le
ast
square
s
an
d
le
ast
m
ean
a
sq
ua
re
wh
ic
h
reduces
the
pe
rfor
m
ance
of
bit
error
rate
[
1
7
]
.
The
ou
tpu
t
sig
nal
ha
ving
disturba
nce
fac
tors
s
uch
as
noise
and
er
r
or
c
an
be
ov
e
rc
ome
by
us
in
g
thes
e
filt
ers.
The
c
hannel
eq
ualiz
at
io
n
is
achieve
d
usi
ng
t
hese
filt
ers
bu
t
t
her
e
is
one
dra
wb
a
ck.
These
ada
ptive
filt
ers
in
non
-
li
near
r
edu
c
e
the
cha
nnel
pe
rfor
m
ance
w
hich
is
a
m
ajo
r
dr
a
w
bac
k.
More
ov
e
r,
a
da
ptive
filt
ers
ge
ner
al
ly
dim
inish
in
non
-
li
near
filt
erin
g
[
18
].
Re
searche
rs
at
tem
pted
to
m
o
dify
the
filt
er
structu
res
to
i
m
pr
ov
e
the
pe
rfor
m
ance
of
t
he
ada
ptive
filt
ers
in
non
-
li
near
filt
erin
g
usi
ng
A
N
N
[
19
]
.
Ther
e
f
or
e,
diff
e
ren
t
al
gorith
m
s
are
dev
el
oped
in
the
li
te
r
at
ur
e
wh
ic
h
revam
ps
the
per
f
orm
ance
of
the
c
hannel
[
20
]
.B
ut
these
m
od
e
ls
do
not
reduce
the
interf
eren
ce
le
vel.Th
e
refo
re
tradit
ion
al
m
et
hods
are
c
om
par
ed
with
the
prese
nt
m
od
el
.
It
is
no
te
d
that
A
NN
pro
vid
es
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
Op
ti
mize
d
B
E
R for
c
hannel
e
qualizer
us
in
g cucko
o searc
h and ne
ural net
work
(
Sw
ati Katw
al
)
2999
bette
r
res
ults
than
tra
diti
on
a
l
m
et
ho
ds
.
A
NN
is
popula
r
to
c
om
bat
c
om
plex
pro
blem
s.
Ar
ti
fici
al
neural
netw
orks
a
re
us
e
d
to
e
qu
al
i
ze
the
le
vel
of
no
ise
.
I
f
the
channel
s
uffers
the
no
-
n
-
eq
ua
li
zi
ng
pro
ble
m
then
equ
al
iz
at
io
n
is
basical
ly
us
ed
to
in
ve
rse
the
filt
er
resu
lt
s.
The
tra
diti
on
al
m
et
ho
ds
us
e
the
in
ver
se
filt
erin
g
te
chn
iq
ue
t
o
c
om
bat
the
dist
urba
nces.
The
process
of
in
ve
rse
filt
erin
g
is
us
e
d
at
the
f
ront
en
d
of
the
r
ecei
ver
in
th
e
di
gital
com
m
un
ic
at
ion
syst
e
m
.
The
tradit
io
nal
m
e
t
hods
us
e
d
to
c
om
bat
the
time
-
va
ryi
ng
dist
or
ti
ons
and
noise
.
In
add
it
io
n,
t
he
c
har
act
erist
ic
s
of
t
he
filt
er
a
r
e
ad
j
us
te
d
in
su
c
h
a
way
th
at
the
sig
nal
le
vel
is
m
ai
ntained.
T
he
noise
le
vel
and
IS
I
re
duces
us
in
g
a
da
ptive
li
near
filt
ers
[
21]
.
T
he
se
li
near
filt
er
s
play
an
im
po
rtant
r
ole
in
dete
rm
i
ning
t
he
c
har
a
ct
erist
ic
of
th
e
receiv
er.
D
ue
to
adv
anc
e
m
ent
in
te
chn
ology,
adv
a
nce
d
li
near
filt
ers
were
dev
el
ope
d
by
m
any
scho
la
rs
an
d
pra
ct
it
ion
ers.
I
n
order
to
ac
hieve,
fas
t
conve
rg
e
nce
r
at
e,
ANNar
e
use
d
t
o
e
qu
al
iz
e
the
noise
le
ve
l.
ANN
im
pr
oves
the
perfor
m
ance
of
th
e
chann
el
by so
l
ving c
om
plex
prob
le
m
s.
A
n
e
ural
arc
hitec
ture
is
d
e
ve
lop
e
d usin
g
li
near fil
te
rin
g
t
echn
i
qu
e
s.
In
the
ca
se
of
ne
ur
al
netw
orkin
g,
the
pr
ob
le
m
of
inv
e
rse
filt
e
rin
g
r
edu
ce
s
to
a
great
extent
.
This
filt
erin
g
t
echn
i
qu
e
can
be
us
ed
at
al
l
the
c
hannels.
A
m
ult
il
evel
chan
nel
e
qual
iz
er
reduces
t
he
le
vel
of
the
no
ise
by
re
d
uci
ng
th
e
m
e
an
square
e
rro
r
an
d
non
-
li
ne
ar
com
m
un
ic
ation
cha
nnel
s
can
be
ha
ndle
d
easi
ly
us
in
g
th
e
ne
ural
net
works.
The
netw
ork
ou
tc
om
es
resul
t
in
a
fast
c
onve
rg
e
nce
ra
te
and
fast
s
pe
ed
of
the
netw
ork.
In
a
dd
it
io
n,
m
ul
ti
le
vel
per
ceptr
on
c
om
bats
non
-
li
nea
r
op
ti
m
iz
at
ion
prob
le
m
s.
M
or
e
over,
the
pr
op
e
rtie
s
of
A
NN
are
orga
nized
in
suc
h
a
way
t
hat
optim
iz
ing
pr
ob
le
m
s
can
be
so
l
ved
us
i
ng
li
near
equ
al
i
zi
ng
[
22
].
The
in
pu
t
a
nd
outp
ut
m
app
in
g
a
re
s
or
te
d
by
usi
ng
e
qual
iz
er
at
t
he
fron
t
e
nd
of
the
m
ai
n
receiver
.
The
dev
el
op
e
d
e
qual
iz
er
determ
i
nes
the
tim
e
-
var
yi
ng
coe
ff
i
ci
ents
of
the
sign
al
.
F
ur
t
herm
or
e,
the
coe
ff
ic
ie
nt
s
are
est
im
at
e
d
usi
ng
the
tra
ining
seq
ue
nce
s.
I
n
the
netw
orks,
t
he
blind
equ
al
iz
er
is
tr
ai
ned
us
in
g
the
trai
ni
ng
se
qu
e
nces
.
These
trai
ni
ng
sequ
e
nces
are
al
so
us
e
d
in
adap
ti
ve
filt
ering.
A
direct
sear
c
h
al
gorithm
is
i
m
ple
m
ented
us
in
g
the
ne
ur
al
netw
ork
to
co
nver
ge
the
outc
om
e
s
[
23
]
.
Ba
loc
h
us
e
s
the
pe
rce
ptr
on A
N
N
i
n
c
om
bin
at
ion
wit
h
the
sim
ulate
d
network
to
com
bat
the
I
SI. An
er
ror
back prop
a
gatio
n
al
gorithm
is
us
ed
for
pe
rfor
m
ing
blin
d
opti
m
iz
at
ion
.Th
e
optim
iz
at
ion
te
c
hn
i
qu
e
im
pr
ov
es
the
per
f
orm
ance
of
the
ne
tw
ork
.
I
n
a
dd
it
io
n,
th
e
de
velo
pe
d
a
l
gorithm
achieves
t
he
desire
d
resu
lt
by
re
du
ci
ng
the
le
vel
of
the noise.
So
m
e
scho
la
rs
al
so
use
fitne
ss
f
un
ct
io
ns
f
or
c
onsist
ent
r
esults.
T
he
fit
ness
functi
on
us
in
g
neural
netw
orks
devel
op
the
filt
ers
us
i
ng
ti
m
e
-
va
ryi
ng
c
ompone
nts.
T
he
new
st
ru
ct
ured
filt
e
rs
have
hi
gh
com
pu
ta
ti
on
al
com
plexity
and
high
data
s
pe
ed
rate.
T
he
da
ta
sy
m
bo
ls
ar
e
extracte
d
with
a
fast
c
onve
r
gen
ce
rate.
T
he
a
dd
it
ion
of
w
hite
G
aussian
noise
and
ISI
doesn
’
t
aff
ect
the
sig
nal
stre
ng
t
h
usi
ng
fitness
f
unct
ion
.
More
ov
e
r,
t
he
sign
al
is
eas
il
y
detect
ed
in
t
he
prese
nce
of
these
inter
fere
nces.
T
he
IS
I
com
po
ne
nt
al
m
os
t
appr
oach
es
t
o
zero
with
t
he
ne
wly
devel
op
e
d
str
uctu
res
of
the
filt
er.
More
over
,
Katwal
al
so
us
es
the
A
rtific
ia
l
Neural
Net
work
(AN
N)
for
adap
ti
ve
cha
nn
el
e
qu
al
iz
at
ion
us
i
ng
4
-
Qu
ad
ratu
re
Am
plit
ud
e
Modula
ti
on
(
QA
M
)
si
gn
al
in
t
he
dig
it
al
com
m
un
ic
at
i
on
syst
em
[2
4
]
.Th
e
well
-
struct
ur
e
d
filt
ers
usi
ng
the
la
te
st
m
od
el
s
li
ke
ANN
and
sim
ulate
d
netw
ork
e
qu
al
i
ze
the
no
nline
ar
net
works
[
2
5
]
.
H
oweve
r,
c
ast
ing
su
c
h
a
netw
ork
re
qu
i
res
hi
gh
ba
nd
wi
dth
a
nd
netw
ork
m
us
t
be
a
sin
gle
la
ye
r.
The
no
n
-
li
near
m
od
el
s
are
le
ss
conve
rg
e
d
to
t
he
re
qu
i
red
outp
ut
[26]
.
T
he
refor
e
,
an
at
tem
pt
has
bee
n
m
ade
to
converge
the
non
-
li
near
m
od
el
s
into
lin
ear
m
od
el
s
usi
ng
m
ini
m
a
l
rad
ia
l
neural
ne
tworks
[
2
7
].
T
su
da
de
velo
p
s
the
i
m
pr
oved
NLMS
al
gorithm
fo
r
channel
eq
ualiz
at
ion
.T
he
non
-
li
near
m
od
el
s
are
co
nv
e
r
ged
into
li
near
op
ti
m
iz
ed
m
od
el
s
wh
i
c
h
pro
vid
e
the
c
on
sist
e
nt
ou
tc
om
e
s
[
2
8
].
Th
e
li
near
m
od
e
ls
us
ed
in
the
networ
k
eq
ua
li
ze
the
no
ise
le
vel.
The
de
velo
ped
filt
ers
reduc
e
the
i
nterf
e
re
nc
e
le
vel
in
bot
h
non
-
sta
ti
on
a
ry
an
d
sta
ti
on
ary
net
works.
In
this
pap
e
r,
a
n
e
ff
ic
ie
nt
te
ch
nique
is
de
velo
ped.
A
n
at
te
m
pt
has
been
m
ade
to
pro
po
se
t
he
m
ulti
-
m
od
el
network
t
o
achieve
the
de
sired
re
su
lt
.
T
he
cha
nnel
res
pons
e
is
opti
m
iz
ed
us
ing
li
near
filt
ering.
The
ne
ural
ne
twork
m
et
ho
ds
ar
e
com
bin
ed
with
so
ft
com
pu
ti
ng
based
op
ti
m
izati
on
te
ch
niqu
e.
The
sw
arm
intel
li
gen
ce
m
od
el
is
com
bin
ed
with
ANN
to
re
duc
e
the
IS
I
in
t
he
sign
al
.
Mo
re
over
,
eff
ect
iv
e
and
desire
d
res
ults
are
achiev
ed
by
i
m
ple
m
ent
ing
the nov
el
tech
niq
ue
. A
no
vel strate
gy is introd
uced
in
this p
a
per
usi
ng opti
m
iz
at
ion
techni
qu
es
.
The
ch
an
nels
are
eq
ualiz
ed
us
in
g
the
dev
e
lop
e
d
m
od
el
s
and
al
gorithm
s
.
A
n
eff
ic
ie
nt
m
od
el
is
dev
el
op
e
d
i
n
this
pap
e
r
to
i
m
pr
ov
e
the
pe
rfor
m
ance
of
BER
.
An
o
ptim
al
chan
nel
BER
is
achiev
ed
us
i
ng
t
he
f
it
ness
functi
on.
T
he
dev
el
oped
a
lgorit
hm
equ
al
iz
es
the
no
is
e
le
vel
us
in
g
li
near
opti
m
iz
at
ion
te
ch
niq
ue
s
.
The
eq
ualiz
at
ion
m
od
el
s
are
dev
el
op
e
d
f
or
sta
ti
on
ary
an
d
non
-
sta
ti
ona
ry
channels.
I
n
ad
diti
on,
the
fitness
f
unct
io
n
is
de
sign
e
d
us
in
g
swar
m
op
ti
m
iz
at
ion
te
chn
iq
ues
to
re
duce
the
ISI.T
he
sim
ulati
on
resu
lt
s
areprese
nted
i
n t
his
pap
e
r wh
i
ch pr
oves the
c
onf
or
m
ity of
t
he
d
e
velo
ped a
ppr
oac
h.
3.
PROP
OSE
D WOR
K
MO
D
EL
The
pro
po
se
d
structu
re
esse
nt
ia
lly
aim
s
to
red
uce
the
B
it
Error
Ra
te
(B
ER)
of
the
dat
a
an
d
it
al
so
fo
c
us
es
on al
gorithm
s to
eq
ua
li
ze the cha
nn
el
. Th
is
sect
ion i
s d
i
vid
e
d
int
o t
wo p
a
rts as
fo
ll
ow
s:
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.
10
, No
.
3
,
J
une
2020 :
29
97
-
3006
3000
3.1.
CS
algorit
hm
fo
r
ch
an
nel e
q
ua
li
z
at
ion
CS
is
to
be
a
pp
li
ed
ove
r
th
e
data
wh
ic
h
is
to
be
tra
nsfer
red.
T
he
da
t
a
is
trans
ferr
ed
th
r
ough
a cha
nn
el
,
whi
ch
m
od
el
for b
oth
sta
ti
on
a
ry
and no
ns
ta
ti
onary has
b
ee
n p
r
esented
in
t
he
s
equ
el
.
3.2.
Equaliz
at
ion
for
s
tatio
na
r
y chan
nel
Discrete
-
ti
m
e
filt
er
with
ad
diti
ve
w
hite
Ga
ussi
an
noise
(
A
WGN)
an
d
Ra
l
ei
gh
Fa
di
ng
(
R
F)
is
uti
li
zed
on
e
b
y
one
, a
nd the
outp
ut
of
the ch
a
nnel
z
(
n
)
wi
th m
easur
em
ent n
oise
v
(
n
)
is
z
(
n
)
=
C
T
U
(
n
)
+
v
(
n
)
(1)
As
sho
wn
i
n
(
1),
c
=
[
c
0
,
c
1
,
c
2
,
.
.
,
c
M
−
1
]
T
is
the
cha
nnel
co
-
e
ff
ic
ie
nt
,
su
bsc
r
ipt
T
denotes
the
tra
ns
pose
a
nd
input sig
nal
ve
ct
or
is
giv
e
n b
y
U
(
n
)
=
[
u
(
n
)
,
u
(
n
−
1
)
,
.
.
,
u
(
n
−
M
+
1
)
]
T
(2)
If
we
wan
t
to
represe
nt
this
trans
ver
sal
filt
er
into
the
sta
te
sp
ace
eq
uation,
then
we
c
an
co
ns
ide
r
the
tim
e
delayed
input
as
sta
te
s.
By
con
sid
erin
g
the
tim
e
delay
ed
input
as
sta
te
var
ia
bles,
the
sta
te
sp
ace
m
od
el
o
f
t
he
c
hannel ca
n be
ob
ta
ine
d by
[
6
,
7
].
X
(
n
+
1
)
=
AX
(
n
)
+
Bu
(
n
)
+
w
(
n
)
(3)
z
(
n
)
=
C
T
X
(
n
)
+
v
(
n
)
(4)
X
(
n
)
=
[
x
(
n
)
,
x
(
n
−
1
)
,
…
x
(
n
−
M
+
1
)
]
T
are
the stat
e
va
ria
bles o
f
t
he
s
ta
te
m
od
el
,
w
(
n)
is
t
he
proces
s
noise
and v(
n)
is t
he m
easur
em
ent noise. T
he st
at
e trans
fer
,
in
pu
t
and outp
ut m
atr
ic
es are
(5)
=
[
0
,
1
,
2
,
…
.
,
−
1
]
(6
)
Since
at
the
receiving
e
nd
th
e
act
ual
sta
te
i
s
unknown
a
nd
act
ua
l
input
so
we
will
m
od
el
a
filt
er
wh
e
re
in
pu
t,
a
s
well
as
or
igi
nal
sta
te
s,
are
est
i
m
at
ed.
The
est
i
m
a
te
d
input
and
est
im
ated
sta
te
equ
at
ion
s
a
re
represe
nted
as:
X
n
+
1
=
A
X
n
+
B
u
n
(7)
z
̂
(
n
)
=
C
T
X
̂
(
n
)
(8)
As
s
how
n
i
n
(
7)
X
̂
(
n
)
is est
i
m
at
ed
sta
te
an
d i
n (
6)
u
̂
(
n
)
is est
im
a
te
d
input.
3.3.
Equaliz
er a
f
or
non
-
s
tatio
nary ch
annel
The
syst
em
fo
r
a
tim
e
-
inv
aria
nt
cha
nn
el
w
hose
c
oeffici
ents
C
are
kn
own
is
con
si
der
e
d
first.
But
i
n
pr
act
ic
al
cases,
m
os
t
of
the
ch
ann
el
is
tim
e
-
var
yi
ng.
So
to
i
m
ple
m
ent
su
ch
c
hannels
in
th
e
above
schem
e
it
i
s
require
d
to
esti
m
at
e the ch
a
nnel
co
ef
fici
ents
C =
C
̂
(
n
)
us
in
g
the
LMS al
gorith
m
, th
us
(5) i
s a
s,
z
̂
(
n
)
=
C
̂
T
X
̂
(
n
)
(9)
This
e
stim
at
ed
outp
ut
z
̂
(
n
)
)
is
co
m
par
ed
with
a
ct
ual
ou
t
pu
t
z
(
n
)
and
th
e
est
im
a
te
d
sta
te
s
are
up
dated
i
n
the equat
io
n:
X
n
+
1
=
X
n
+
K
[
z
(
n
−
z
n
)
]
(10)
This
is
the
time
wh
e
n
the
data
is
go
ing
to
be
received
by
the
receive
r.
T
he
eq
ualiz
er
play
s
it
s
part
and
CS
is
intr
oduce
d
in
the
process
.
Figur
e
3
represents
the
arch
it
ect
ur
e
of
the
eq
ualiz
er
wh
e
re
CS
is
to
be
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
Op
ti
mize
d
B
E
R for
c
hannel
e
qualizer
us
in
g cucko
o searc
h and ne
ural net
work
(
Sw
ati Katw
al
)
3001
app
li
ed
.
w
k
and
v
k
re
pr
es
e
nt
the
proces
s
noise
and
m
easur
e
m
ent
no
ise
.
I
n
the
diff
e
re
nce
eq
uatio
n,
the
non
-
li
nea
r
functi
on
f
relat
es
the
sta
te
at
the
pr
e
vious
ti
m
e
ste
p
k
−
1
to
th
e
sta
te
at
the
cu
rr
e
nt
ti
m
e
ste
p
k
.
In the m
easur
e
m
ent eq
uati
on
s
, th
e
non
-
li
nea
r
fun
ct
io
n rela
t
es the s
ta
te
t
o
th
e m
easur
e
m
ent
z
k
.
Figure
3. Str
uc
ture of
the
e
qu
al
iz
er
If
the
value
of
tw
o
ra
ndom
var
ia
bles
is
no
t
kn
own,
then
we
can
appr
ox
im
at
e
t
he
sta
te
an
d
m
easur
em
ent vec
tor
as:
x
k
=
f
k
−
1
,
u
k
(11)
U
ti
li
zi
ng
EK
F
z
k
=
h
(
x
k
,
v
k
)
(12)
As
sho
wn
in
(
12),
x
k
is
a
post
erior
i
est
im
a
te
of
the
sta
te
.
The
ne
w
go
ve
rn
i
ng
e
qu
at
io
ns
that
li
near
iz
e
an
est
i
m
at
e abo
ut
(10) an
d (
11)
a
re as f
ollow
i
ng:
x
k
~
=
f
(
x
k
−
1
,
u
k
,
0
)
(13)
z
k
~
=
h
(
x
k
,
0
)
(14)
wh
e
re
x
k
and
z
k
t
he
act
ual
sta
te
and
m
easurem
ent
vector
s
,
x
k
~
an
d
z
k
~
are
the
appr
ox
im
at
e
s
ta
te
and
m
easur
em
ent
vecto
rs,
x
k
~
is
an
apost
erio
ri
est
im
at
e
of
the
sta
te
at
ste
pK
,
A,W,
H,
V
ar
e
th
e
m
a
trix
of
pa
rtia
l
der
i
vatives
of
x,w,
x,v. N
ow the
pr
e
dicti
on a
nd m
easur
em
e
nt erro
rs
a
re:
x
k
~
x
k
~
+
A
(
x
k
−
1
−
x
k
~
−
1
)
+
W
W
k
−
1
(15)
z
k
~
z
k
~
+
H
(
x
k
−
1
−
x
k
~
)
+
V
V
k
(16)
w
he
re
∈
k
an
d
n
k
are
in
dep
e
nde
nt
r
andom
var
ia
bles
hav
i
ng
ze
ro
m
ean.
T
he
dat
a
pa
sses
th
rou
gh
the
e
qual
iz
e
r
and
gets
into
t
he
cha
nnel
.
Th
e
syst
e
m
m
ai
n
ly
op
erates
in
decisi
on
di
rect
ed
m
od
e
wh
e
n
act
ual
transm
i
ssio
n
and
recei
ving
i
nfor
m
at
ion
ta
ke
s
place.
I
n
the
decisi
on
direct
ed
m
od
e
the
outp
ut
of
CS
is
us
e
d
as
i
nput
f
or
an
d
the
aptive
filt
er.
We
are
al
re
ady
us
i
ng
(
)
an
d
̂
(
)
as
input
f
or
trai
ning
an
d
de
ci
sion
directe
d
m
od
e
t
o
adap
ti
ve
filt
er
at
receivi
ng
end.T
he
sim
i
l
ar
in
pu
t
ca
n
be
pro
vid
e
d
to
KF
i.e
.
in
de
ci
sion
directed
m
ode
wh
e
reas
in
tra
ining
m
od
e
[
20]
.Th
e
data
m
us
t
sat
isfy
the
dem
and
s
of
CS
to
get
thr
ough.
T
he
str
uc
ture
is
as foll
ows.
Algori
th
m
1:
Ap
pli
ca
tio
n
of C
S
Inp
ut : Bi
t Stre
a
m
(BS)
Ou
t
put:
Optim
al
Bi
t
Stream
Total
_E
gg_S
l
ot
s=4;
Divid
e
the
B
S
into
E
gg_s
l
ot
Tim
e
Segm
ents.
//
Di
vi
din
g
the
bit
stream
into
fr
am
es
to
pr
oc
ess
/
/
the d
at
a
sm
oo
thly
eggspe
rslot=(
durati
on)/T
otal_
Eg
g_Slots;
init
ia
lc
ou
nter=
1; //
Tak
i
ng
a c
ounter
to kee
p op
ti
m
al
//
eg
gs o
nly
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.
10
, No
.
3
,
J
une
2020 :
29
97
-
3006
3002
finalco
unte
r=e
gg
s
pe
rslot
foreach
egg i
n Total
_E
gg_S
l
ot
s
Eg
gcou
nter=in
it
ia
lc
ou
nter:fi
na
lc
ounter
//
T
he
The//
enc
ounter
w
il
l st
art f
ro
m
1
a
nd
will
g
o t
o 4
Eg
gs
insl
ot=B
S(
Eg
gc
ounte
r
);//
Ex
tract
i
ng d
at
a from
Bi
t
//
Strea
m
Total
Eg
gs
=B
S
;
Fit
value=Cuc
kooFi
tness
(Egg
sinslot);/
/ Passi
ng the
data t
o
t
he fit
ness funct
ion
of Cuc
koo
Searc
h
En
d
F
or
En
d Alg
or
it
hm
CS is a hard thresh
old
al
gorithm
. Th
e entire
bitst
ream
r
eceiv
ed
is
div
i
ded
into fo
ur
e
gg slots. T
he
e
gg
va
lue i
s
passe
d
t
o
the
f
it
ness
f
unct
ion
of
t
he
CS
w
he
n
t
he
c
ucko
o
is
out.
T
he
fitness
f
un
ct
io
n
of
the
CS
is
de
fin
e
d
as foll
ows.
Let
Nc
be
a
natur
al
c
hange
in
the
st
r
uctu
re
of
the
eg
gs
w
hen
the
cuc
koo
bi
rd
is
out.
The
ef
fecti
ve
ne
ss
of
the
data
is
evaluated
in
co
ntrast
t
o
the
oth
e
r
da
ta
‘D
e
’
w
hic
h
is
prese
nt
in
that
env
i
ronm
ent. H
ence
the C
uc
koo Searc
h Al
gorithm
u
ti
li
ze
s both
the
Nc
a
nd D
e
in
it
s
fitness
functi
on:
De
=
∑
(
d
)
/
n
n
k
=
0
(17)
w
he
re
D
is t
he data
in
t
he bit
stream
, n
is the
total
num
ber
of
bits in t
he
ti
m
e
stream
if
D
∗
N
C
<
D
e
∗
N
c
(18)
Othe
rw
ise
:
I
f
t
he
c
urren
t
data
el
e
m
ent
D
ste
ps
in
with
t
he
Nc
a
nd
sti
ll
is
sm
a
ll
er
than
th
e
overall
ot
her
data
pr
ese
nt
in
the
s
yst
e
m
,
then
th
e
cucko
o
will
keep
t
his
eg
g
e
lse
the
eg
g
wil
l
be
dum
ped
th
at
m
eans
the
da
ta
bit
will
be
forw
a
r
ded
el
se
the
da
ta
bit
will
be
dum
ped
.
T
his
arc
hitec
ture
f
ram
e
is
fu
rthe
r
cr
os
s
-
va
li
dated
by
ANN,
t
o
prove
that
the
fit
nes
s
functi
on
of
t
he
CS
is
s
uffic
ie
nt
in
this
c
on
trast
.
A
N
N
is
a
three
-
la
ye
r
str
uctu
re
nam
ely input,
hidden
and a
n ou
t
pu
t l
ay
er
.
3.4.
Cro
s
s
-
va
li
d
at
i
on
using
ANN
The
in
put
la
ye
r
of
the
Neural
netw
ork
inta
ke
s
the
ra
w
dat
a
at
the
fron
t
.
ANN
do
es
not
unde
rstan
d
wh
at
it
is
pr
ov
i
ded
to
it
if
it
is
no
t
conve
rted
into
a
su
btype
wh
ic
h
is
underst
ood
by
it
s
arch
it
ect
ur
e
an
d
hen
ce
it
util
iz
es
the
sigm
oid
functi
on
a
nd
weig
ht
conve
rted
t
o
gen
e
rate
inter
m
ediat
e
la
ye
r
data
of
it
s
pro
cessi
ng.
This
pap
e
r
c
onside
rs
s
uper
vi
sed
m
achine
le
arn
in
g
for
it
s
proce
ssin
g.
The
a
rch
it
ect
ure
ta
kes
the
e
ntire
op
ti
m
iz
ed
bit
patte
rn
as
i
nput
with
a
dd
it
iv
e
la
bel
seri
es
sta
rting
from
1
to
n
w
he
re
n
is
a
total
nu
m
ber
of
the
bit
se
qu
e
nc
e.
He
nce
if
the
bitst
ream
is
Fi
gure
4(
a
)
the
n
it
s
la
bel
would
be
Fi
gure
4(b
)
.
BSU
is
opti
m
iz
ed
bit
strea
m
and
L
is
ta
rg
et
la
bel.N
e
ural
Network
is
util
iz
ed
in
su
c
h
a
m
ann
e
r
that
every
bits
trea
m
is
trai
ne
d
us
in
g
Leve
nbe
rg
arc
hitec
ture
of
A
N
N.
Ta
ble
1
represents
t
he
sim
ulatio
n
arch
it
ect
ure
of
ANN.
A
total
of
50
0
epo
c
hs
is
pro
vid
e
d
f
or
eve
ry
sim
ulatio
n
of
A
NN.
Al
gorithm
2
ex
p
l
ai
ns
the
w
ork
ing
of
the
A
NN
i
n
the pr
opos
e
d
st
ru
ct
ur
e.
(a)
(
b)
Figure
4
(a). In
pu
t l
ay
er
d
at
a
,
(b).
I
nput lay
er
label
Table
1.
Ne
ur
a
l
structu
re
Total Nu
m
b
e
r
o
f
Neuron
s
10
-
30
Architectu
re
T
y
p
e
Feed
Forwa
rd Bac
k
Pr
o
p
ag
atio
n
Perf
o
r
m
an
ce
Meas
u
re
Mean Sq
u
are Er
ror
Feed
in
g
Par
a
m
e
ter
s
Gradien
t,
Ti
m
e
,
Ite
ration
s
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
Op
ti
mize
d
B
E
R for
c
hannel
e
qualizer
us
in
g cucko
o searc
h and ne
ural net
work
(
Sw
ati Katw
al
)
3003
Algori
th
m
2:
WorkF
low
of
ANN
1.
Net=Trai
n_Neural(BS
U,
L
,10
-
30); //
Init
ia
te
d
Trai
ning
Me
chan
ism
f
or optim
iz
ed
Bi
t /
/
//
Stream
(BSU
)
wit
h
it
s corre
sp
on
ding la
bel
L.
//
Th
e trai
ned s
tructu
re
is sa
ve
d
in
a str
uctu
re
n
am
e
//
“net”
2.
Net.trai
np
a
r
a
m
.ep
oc
hs
=
500//
500 t
raini
ng it
erati
ons
3.
If sat
isfie
d(
Gr
a
dient)
// I
f t
he
G
ra
dient is
sat
isfie
d,
t
hen
stop fee
ding in
the
forw
a
r
d direct
ion
4.
Ba
c
kT
rack()
; /
/ B
ackTr
ac
k and Fin
d
m
os
t
su
it
able MSE
5.
En
d If
6.
Fin
d
M
os
t S
uitable
I
te
rati
on
base
d on MS
E;
7.
Sim
ulate
(BS
U,
Net); //
Sim
ulate
BSU
w
it
h
st
or
e
d
str
uct
ur
e
En
d Alg
or
it
hm
Algorithm
2
is
a
co
m
bin
at
ion
of
Feed
F
orward
an
d
Ba
ck
Prop
a
ga
ti
on
Algorit
hm
in
wh
ic
h
the
Feedin
g
in
forw
a
r
ding
dir
ect
ion
ta
kes
pl
ace
un
ti
l
ei
ther
the
pr
opa
gatio
n
it
erati
on
bec
om
es
equ
al
to
a
total
n
um
ber
of pr
ovide
d
it
erati
on
or
the
gr
a
dient
of
the
ANN
is
no
t sat
isfie
d. T
his p
r
ocess
is f
ollow
e
d by the Bac
k
Pr
opa
gatio
n
m
et
hod
whose
s
at
isfyi
ng
par
a
m
et
er
is
MSE.
The
sim
ulator
r
olls
bac
k
to
fin
d
the
le
ast
MS
E
base
d
on
the
unde
rstan
ding
of
the
gra
dient.
O
nce
t
he
A
N
N
is
trai
ned
,
eac
h
opti
m
iz
ed
BSU
is
cl
assifi
e
d
with
the
trai
ne
d
str
uc
ture.
If
t
he
re
su
lt
la
bel
of
th
at
BSU
is
not
equ
al
to
the
La
bel
set
L
the
n
the
m
ai
n
bit
value
of
that
optim
iz
ed
bit
va
lue
is o
pt
i
m
iz
ed
again.
Figure 5
a
nd
F
igure
6
re
pr
ese
nts
th
e
trai
ning
arch
it
ect
ure of
ANN
and Bac
kpr
opa
gation.
Figure
5. Trai
ni
ng
m
od
el
of ANN
Figure
6. Ba
ck
pro
pa
gation
As
s
how
n
i
n
Fi
gure 6
,
the n
et
work b
ack
pro
pag
at
es
in
t
he
netw
ork.
A
N
N
co
ns
ide
rs
s
ome
data
as
te
st
data
out
of
th
e
pro
vid
e
d
tra
ining
data
e
ve
n
at
the
trai
nin
g
phase.
T
an
sig
is
the
trai
ning
f
unct
ion
wh
ic
h
gen
e
rates
the
weig
ht
W
of
th
e
trai
nin
g
data
D
co
ns
ide
rin
g
a
and
b
as
ar
bi
trary
const
ants.
In
a
si
m
il
ar
fa
sh
io
n,
pureli
n
functi
on
is
util
iz
ed
at
the
sim
ulati
o
n
f
unct
ion.
T
he
ANN
util
iz
es
sig
m
oid
trai
ning
f
unct
ion
to
trai
n
the d
at
a
an
d
al
so
t
o
create
a c
ro
ss
-
validat
ion data
for
te
sti
ng. Ne
ur
al
a
rch
i
te
ct
ur
e is s
how
n
in
Fig
ure
7.
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.
10
, No
.
3
,
J
une
2020 :
29
97
-
3006
3004
Figure
7.
Ne
ural
arch
it
ect
ure
4.
RESU
LT
S
A
ND
DI
SCUS
S
ION
Fo
ll
owin
g para
m
et
ers
are a
naly
zed b
a
sed
on t
he
pr
opos
e
d w
ork.
4.1.
A
-
BER
It
is
the
ave
ra
ge
bit
er
ror
ra
te
of
the
netw
ork
w
hic
h
is
gen
e
rated
a
fter
10000
sim
ulati
on
r
ounds
.
Each
sim
ulatio
n
rou
nd
ge
ne
rates
a
bit
error
rate
valu
e
wh
ic
h
is
store
d
in
the
s
i
m
ulati
on
arr
a
y
and
a
com
par
ison
of
pr
opos
e
d
w
ork
an
d
Kal
m
a
n
filt
er
resu
lt
s
are
her
e
by
co
m
par
ed.
A
-
BE
R
is
no
te
d
after
ever
y
10000 si
m
ulatio
n r
ound.
Aver
age Bit
Erro
r
R
at
ecan be cal
cu
la
te
d
as foll
ow
s:
A
−
BER
=
∑
BER
n
−
10000
k
=
0
n
(19)
Figure
8
pr
e
se
nts
the
A
-
BER
for
Propose
d
work
m
od
el
ev
al
uated
agai
ns
t
Kalm
an
Fil
te
r
Mod
el
. I
t
is
observ
e
d
that
our
pr
opose
d
m
od
el
de
m
on
strat
ed
le
s
ser
a
ver
a
ge
bit
erro
r
r
at
e
as
c
om
par
ed
t
o
the
Kalm
an
Fil
te
r
Mo
del
wh
e
n
sim
ul
at
ed ov
e
r 1
0000 s
i
m
ulati
on
s.
Figure
8. A
-
B
ER vs
total
nu
m
ber
o
f si
m
ula
ti
on
s
4.2.
L
-
BER
It
is
the
log
ari
thm
scal
e
of
the
A
-
BER
bu
t
it
is
ta
ken
ag
ai
n
the
SN
R
va
lue
of
the
c
om
m
un
ic
at
ion
channel.
Fig
ure 9
s
how
s the
c
om
par
ison o
f L
-
BER
a
nd S
N
R. Th
e
sim
ul
ation
grap
h
s
ho
ws
that
L
-
BER
p
lott
ed
against
the
sig
nal
to
noise
rati
o
al
so
s
howe
d
the
effe
ct
iven
ess
of
our
pro
pose
d
desi
gn.
T
he
res
ults
sho
w
that
pro
po
se
d
hybr
id
is
m
or
e
eff
ect
ive
in
im
pro
ving
the
I
SI
as
c
om
par
ed
to
th
e
Ka
l
m
an
Fil
te
r
Mod
el
.
In
bo
t
h
cases,
the
pro
posed
s
ol
ution
ha
s
sho
wn
im
pr
ovem
e
nt
by
a
m
arg
in
of
30
-
50%.
T
he
m
axi
m
u
m
A
-
BER
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
Op
ti
mize
d
B
E
R for
c
hannel
e
qualizer
us
in
g cucko
o searc
h and ne
ural net
work
(
Sw
ati Katw
al
)
3005
for
the
pr
opose
d
s
olu
ti
on
is
no
te
d
to
be
0.
003999
w
her
e
as
it
is
0.0
09899
f
or
Kalm
an
Fil
te
r.
T
he
propos
e
d
m
od
el
al
so
evaluated
MSE
at
each
trai
ning
inter
val
and
the
resu
lt
s
ar
e
dep
ic
te
d
in
Figure
10.
T
he
gr
a
ph
sh
ows
that
r
oo
t
m
ean
sq
uar
e
changes
with
each
sim
ulatio
n
an
d
the
grap
h
com
par
es
th
e
ob
se
rv
e
d
er
r
or
f
or
the 10
sim
ulati
on
s
.
Figure
9. L
-
BE
R vs SNR
Figure
10. T
rai
ning MSE
5.
CONCL
US
I
O
N
This
pap
e
r
pr
e
sents
a
n
en
ha
nc
e
m
ent
of
t
he
filt
ering
te
c
hn
i
qu
e
util
iz
ing
t
he
CS
a
nd
A
NN.
A
ne
w
fitness
f
un
ct
io
n
is
desig
ned
i
n
the
CS
an
d
ANN
has
act
e
d
li
ke
a
cro
ss
va
li
dator
to
the
C
S.
In
order
t
o
va
li
date
the
propose
d
a
lgorit
hm
,
bo
th
the
tr
ai
ning
t
echn
i
qu
e
s
li
ke
Feed
F
orwa
rd
an
d
Ba
c
k
Pro
pag
at
io
n
ha
ve
be
e
n
i
m
ple
m
ented.
MSE
is
consi
der
e
d
as
the
par
am
et
er
of
cro
ss
-
validat
io
n
w
herea
s
the
propose
d
work
i
s
com
par
ed
with
the
Kalm
an
Fil
te
r
an
d
is
evaluate
d
on
the
ba
se
of
A
-
BER
a
nd
L
-
BER
.
An
e
nh
a
nce
i
m
pr
ovem
ent
i
n
ISI
ra
ngin
g
from
30
%
to
50%
has
been
achieve
d
w
he
n
the
data
have
been
sim
ulatio
n
on
10000 si
m
ulatio
ns.
REFERE
NCE
S
[1]
P.
S.
Henr
y
,
et
al.
,
“
Tra
nsm
ission
devi
ce
with
cha
nne
l
equa
l
iza
ti
on
and
cont
rol
and
m
et
hods
fo
r
use
the
rewit
h,
”
Unite
d
S
tat
es
Pa
te
nt
Application
,
no.
9,
pp.
670
-
68
0,
2017
.
[2]
R.
Marti
nek
,
e
t
al.
,
“
Adapti
v
e
o
pti
m
iz
ation
of
c
ontrol
par
ame
te
r
s
for
fee
d
-
forwa
rd
software
def
i
ned
equa
l
izati
on
,
”
Wirel
ess P
ersonal
Comm
unic
at
i
ons
,
vol. 95, no.
4,
pp
.
4001
-
401
1,
Aug
.
2017.
[3]
G.
Das,
et
al.
,
“
Artifi
cial
ne
ura
l
n
et
work
tr
ai
ned
b
y
p
a
rtic
le
sw
arm
opt
i
m
iz
at
ion
for
n
on
-
li
ne
ar
cha
nn
el
equa
l
iz
a
ti
on
,
”
E
xpe
rt S
yste
ms
wi
th
App
li
ca
ti
ons
,
vol.
41
,
no
.
7
,
pp
.
3491
-
3496
,
Jun
.
2014
.
[4]
S.
Bal
o
ch,
et
al
.
,
“
Chann
el
Equ
al
i
za
t
ion
Us
ing
Multi
lay
e
r
Per
c
ept
ron
Ne
tworks,”
M
ehran
Uni
ve
rs
it
y
Re
s
earc
h
Journal
o
f
Engi
n
ee
ring a
nd
Tech
nology
,
vol
.
31
,
no.
3
,
Jul.
2012
[5]
X.
Yang
and
Suash
Deb,
"Cuc
koo
Sear
ch
v
ia
Lé
v
y
fl
ight
s,"
2
009
World
Con
gress
on
Nature
&
Bi
ologi
cal
l
y
Inspired
Comput
ing
(
NaBIC)
,
Coim
bat
ore
,
pp.
21
0
-
214,
2009
.
[6]
K.
M.
S.
Huq,
e
t
al.
,
“
Channel
Equa
liza
ti
on
in
Digit
al
Tra
nsm
i
ss
ion,
”
Int
ernational
Journal
of
Computer
Sci
en
ce
and
Information Sec
urit
y
(
IJCSI
S
)
,
US
A,
vol.
4
,
n
o.
1&2
,
Aug
.
20
09.
[7]
M.
Ene
scu,
et
a
l.
,
“
Adapti
v
e
Eq
ual
i
za
t
ion
of
Tim
e
-
Var
y
ing
MI
MO
cha
nnel
s,”
Signal
Proce
ss
i
ng
,
vol.
85
,
no.
1,
pp.
8
1
-
93
,
Jan
.
2
005.
[8]
R.
A.
Bouje
m
aa
and
S.
Marc
o
s,
“
Para
ll
el
Ka
l
m
an
Filt
eri
ng
f
or
Optimal
S
y
m
bol
-
By
-
S
y
m
bo
l
Esti
m
at
ion
in
an
Equ
al
i
zation Conte
xt
,
”
S
ignal P
roce
ss
ing
,
E
lsevi
er
,
vo
l
.
85,
no
.
6
,
pp
.
1125
-
11
38,
Jun.
2005.
[9]
F.Orderud,
“
Com
par
ison
of
K
al
m
an
Filt
er
Es
ti
m
at
ion
Appro
ac
hes
for
State
Space
Models
with
Nonline
a
r
Mea
surem
ent
s,”
In
Proceedi
ng
of
Scandi
na
vi
an
Confe
renc
e
on
Si
mulati
on
and
M
odel
ing
,
pp
.
1
-
8
,
2005.
[10]
J.W
.
Mark,
“
A
Note
on
the
Mo
difi
ed
KalmanFi
lt
er
for
ch
annel
Equa
liza
ti
on
,
”
in
Proceedi
ngs
of
the
I
E
EE
,
vo
l.
6
1,
no.
4
,
pp
.
481
-
4
82,
Apr.
1973.
[11]
H.
N.
Kim
and
W
.
J.
Song,
“
An
Adapti
v
e
IIR
E
qual
i
ze
r
for
Non
-
Minim
um
-
Phas
e
ch
anne
l
,
”
ICS
P
'98.
1998
Fou
rth
Inte
rnational
Co
nfe
renc
e
on
Sign
al
Proc
essing
(
Cat.
No
.
98TH
8344)
,
Bei
ji
ng
,
Ch
in
a,
vo
l. 1, pp. 441
-
444,
19
98
.
[12]
J.
C.
Patra
and
R.
N.
Pal,
“
A
func
ti
ona
l
li
n
k
art
ificial
neu
ral
net
work
for
ada
pti
v
e
cha
n
nel
equalizat
ion
,
”
Signal
Proce
ss
in
g
,
vol
.
43
,
no
.
2
,
pp.
181
-
195
,
Ma
y
1995.
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.
10
, No
.
3
,
J
une
2020 :
29
97
-
3006
3006
[13]
T.
Shim
amura,
et
al.
,
“
Equa
l
izati
on
of
Ti
m
e
-
Vari
ant
Com
m
unic
a
ti
on
Channel
via
Chann
el
Es
ti
m
at
ion
Base
d
Approac
hes,
”
1996
IEE
E
Int
ernati
onal
Conf
ere
nce
on
A
coust
ic
s,
Spe
ec
h,
an
d
Signal
Proces
sing
Confe
renc
e
Proce
ed
ings
,
At
l
ant
a
,
GA
,
US
A,
vol.
3
,
pp
.
1703
-
1706,
1996
.
[14]
P.
Grohan
and
S.
Marc
os,
“
Struct
ure
s
and
Per
form
anc
es
of
Sever
a
l
Adapti
v
e
Kalman
Equa
l
i
ze
rs,”
1996
IEEE
Digit
al
S
ignal P
roce
ss
ing
Workshop P
roce
edi
ngs
,
Lo
en, Norwa
y
,
pp.
454
-
457
,
199
6.
[15]
S.
Mcla
ughli
n
,
“
Adapti
ve
Equalizat
ion
vi
a
Kal
m
an
Filt
eri
ng
T
ec
hniqu
es,
”
in
I
EE
E
Pro
ce
ed
in
gs
F
-
R
adar
and
Signal
Proce
ss
in
g
,
vol
.
138
,
no
.
4
,
pp
.
388
-
396
,
A
ug.
1991
.
[16]
Deng
J.,
et
al.
,
“
Comm
unic
at
io
n
Channe
l
Equalizat
ion
Us
ing
Com
ple
x
-
Valued
Minim
al
Radi
al
Basis
Functi
o
n
Neura
l
N
et
work
s,”
in
IE
EE Tr
ansacti
ons on
N
eural
Ne
tworks
,
vo
l
.
13
,
no
.
3
,
pp
.
6
87
-
696,
Ma
y
20
02.
[17]
J.V.
Cand
y
,
“
Model
-
base
d
signa
l proce
ss
ing
,
”
Joh
n
Wiley
&
Sons
,
vol.
36
,
Oc
t
.
200
5.
[18]
S.
O.
Ha
y
k
in, “A
dapt
ive Filter Theor
y
,
”
Pren
ti
c
e
Hal
l, E
ag
le wood Cl
iffs,
NJ,
2
nd
Ed.,
1991
.
[19]
G.R.
Patra,
et
a
l.
,
“
Non
li
n
ea
r
cha
nne
l
equ
al
i
z
at
ion
for
d
igi
t
al
comm
unic
at
ion
using
DE
tra
in
ed
func
t
iona
l
link
art
if
ic
i
al ne
ura
l net
works
,
”
Inte
r
nati
onal
Con
f
.
o
n
Conte
mpor
ary
Computing, Heidelbe
rg,
vol
.
2
,
pp.
403
-
14
,
201
1.
[20]
P.
S.
R
.
Den
iz
,
“
Adapti
ve
Filt
ering a
nd
Pra
ctical Im
ple
m
ent
at
ion
,
”
Springer
,
Jan
.
2008.
[21]
S.
McL
aughl
in
,
et
al
.
,
“
A
Perform
anc
e
Stud
y
of
Thre
e
Adapt
iv
Equa
lizer
envi
ron
m
ent
,
”
IEEE
Int
ernati
onal
Conf
.
on
Comm
unic
ations
,
World
Prosperity Thr
ough
Comm
unic
ati
ons
,
Boston,
MA
,
U
SA
,
vol. 1, pp. 1
93
-
197,
1989
.
[22]
K.
Burse,
R.
N
.
Yada
v,
S.C.
Shr
iva
stav
a,
“
Chan
nel
Equalizat
ion
using
neur
al
ne
twork,
”i
n
IE
EE
Tr
ansacti
ons
on
Syste
ms
,
Man
,
a
nd
Cybe
rne
ti
cs
,
Part
C
(
Appl
i
cations and
R
evie
w
s)
,
vol.
40
,
no
.
3
,
pp.
352
-
357,
Ma
y
2010.
[23]
S.
Panda,
P.
K.
Mohapa
tra,
“
A
new
tra
ini
ng
sch
eme
for
neur
al
net
works
and
ap
pli
c
at
ion
in
No
n
-
Li
ne
ar
Channel
Equa
liza
ti
on
,
”
A
ppli
ed
Softcomputing
,
vol. 27
,
p
p.
47
–
52
,
Feb
.
2
015.
[24]
S.
Katwa
l
et
al
.
,
“
Non
Li
nea
r
Channe
l
Equ
al
i
z
at
ion
Us
ing
Artifi
cial
Neura
l
Network,
”
RI
EE
C
E
,
vol.
3,
no
.
1,
pp.
333
-
334
,
20
16.
[25]
J.
C.
Patr
a,
e
t
al
.
,
“
Nonline
ar
Cha
nnel
Equ
aliz
at
io
n
with
QA
M
Sig
nal
Us
ing
Ch
eb
yshev
Artificial
N
eur
al
Network,”
Proc
.
2005
IE
E
E
Int
ernati
onal
Joi
nt
Confe
ren
c
e
on
N
eural
N
etwor
ks
,
Montre
al
,
Que.,
vol
.
5
,
pp
.
3214
-
3219,
200
5.
[26]
W
.
D.
W
eng,
et
a
l
,
“
The
Design
o
f
an
SC
FN
N Ba
s
ed
Nonlinear
Ch
anne
l Equa
l
ize
r,”
Journal
of
In
fo
rm
ati
on
Sci
en
ce
and
Engi
n
ee
ring
, v
ol
.
21
,
no
.
4
,
p
p.
695
-
709
,
Jul
.
2005.
[27]
C.
P.
Kum
ar,
et
al.
,
“
Nonlinear
cha
nne
l
equaliza
ti
on
using
m
ini
m
al
rad
ial
basis
func
ti
on
n
eur
al
net
works
,
”
Pro
c
.
Inst.
E
lect. E
ng
.
Vi
sion I
mage
Si
gnal
Proc
essing
,
vol. 147, pp. 42
8
-
435,
2000
.
[28]
Y.
Tsuda
,
et
al.
,
"A
n
improved
N
LMS a
lgori
thm for
cha
nnel
equ
alization,
"
2002
IEE
E
Int
ernati
on
al
Symposium o
n
Circui
ts and
Sys
te
ms
.
Proc
ee
d
in
gs (
Cat.
No.
02C
H37353)
,
Phoenix
-
Scott
sdal
e, A
Z,
US
A,
pp
.
V
-
V,
2002.
Evaluation Warning : The document was created with Spire.PDF for Python.