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.
5
,
Octo
be
r
2020
,
pp.
483
5
~
4843
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v10
i
5
.
pp
4835
-
48
43
4835
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om/i
nd
ex
.ph
p/IJ
ECE
Analysis
of
b
ack
p
r
opag
ation
and
ra
di
al b
asis
f
un
ction
neu
ra
l
n
etworks for
h
andover d
eci
sions
i
n wireles
s commu
nicati
on
Pa
ya
l
M
ahaja
n,
Z
ah
eeru
ddi
n
Depa
rtment
o
f
E
le
c
tri
c
al
Engi
n
eering,
Jam
i
a
Mil
lia
Islamia
Univer
sit
y
,
Ind
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
pr
8
, 2
019
Re
vised
Ma
r
13
,
2020
Accepte
d
Ma
r
3
1
, 202
0
In
m
obil
e
s
y
ste
m
s,
handof
f
is
a
vit
a
l
proc
ess,
ref
err
ing
to
a
proc
ess
of
al
lo
ca
t
ing
an
o
ngoing
c
al
l
fro
m
one
BS
to
anot
her
BS
.
Th
e
handov
er
te
chn
ique
is
ver
y
important
to
m
ai
nta
in
the
Quali
t
y
of
servi
c
e.
Handov
er
al
gorit
hm
s,
b
ase
d
on
n
eur
a
l
n
et
work
s
,
fu
zzy
logi
c
etc.
c
an
be
used
fo
r
the
sam
e
purp
ose
to
ke
ep
Quali
t
y
of
ser
vic
e
as
high
as
poss
ibl
e.
In
thi
s
pap
er,
it
is
proposed
that
bac
k
prop
aga
t
io
n
net
works
and
rad
ial
basis
func
ti
ons
m
a
y
be
used
for
ta
king
h
ando
ver
decision
i
n
wire
le
ss
comm
unic
at
ion
net
works
.
Th
e
p
erf
orm
anc
e
of
t
he
se
c
la
ss
ifiers
i
s
eva
lu
at
e
d
on
the
basis
of
neur
ons
in
hidde
n
lay
er
,
tr
ai
ni
ng
ti
m
e
and
class
ifi
catio
n
ac
cur
acy
.
Th
e
p
roposed
appr
oa
c
h
show
s
tha
t
r
a
dia
l
basis
func
tion
neur
a
l
net
work
give
b
et
t
er
result
s
for
m
aki
ng
handove
r
dec
isions
i
n
wire
le
ss
het
ero
g
ene
ous n
et
works
with cl
a
ss
ifi
ca
ti
on
a
cc
ur
acy
of
90%.
Ke
yw
or
d
s
:
Ba
ck pr
op
a
gati
on n
et
wor
ks
Hand
ov
e
r
N
eu
ral
netw
orks
Q
ualit
y o
f
se
rvi
ce
R
adial
b
asi
s
functi
ons
Copyright
©
202
0
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Paya
l M
ahaj
a
n,
Dep
a
rtm
ent o
f El
ect
rical
En
gi
neer
i
ng,
Jam
ia
Mi
l
li
a Is
lam
ia
U
niv
er
sit
y,
New Del
hi,
India
.
Em
a
il
:
payal
27
m
ahaj
an
@
gm
a
il
.co
m
1.
INTROD
U
CTION
In
tr
odu
ct
io
n
t
o
wireless
c
om
m
un
ic
at
ion
.
W
irel
ess
com
m
un
ic
at
ion
now
a
days
is
t
he
pa
rt
an
d
par
cel
of
our
li
ves
since
it
of
fe
rs
e
la
sti
ci
t
y
and
f
reedom
of
m
ov
in
g
f
ro
m
place
to
a
dif
fer
e
nt
w
hile
not
losin
g
connecti
vity
to
the
netw
ork.
New
i
nteracti
ve
program
servi
ces
request
rat
e
of
u
p
to
hund
red
s
of
m
eg
abi
ts
per
seco
nd
a
nd
so
bigge
r
bands
of
fr
e
qu
e
ncies
are
ne
ede
d
t
o
keep
up
thes
e
new
el
e
vated
data
rate
se
rv
ic
es.
On
t
he
co
ntra
ry,
to
m
ake
m
ob
il
e
no
des
m
ov
e
pr
ob
le
m
free
,
seam
le
ss
han
do
ver
is
ve
r
y
i
m
po
rtant
in
these
netw
orks
[1
]
.
VHO
(v
e
rtic
al
ha
ndover
)
al
gorithm
s
are
ne
cessary
for
the
de
velo
pm
ent
of
the
ne
xt
Ge
ner
at
io
n
HetNets
i.e.
he
te
rogen
e
ous
n
et
w
orks
[2]
.
Web
c
ons
um
e
rs
are
as
su
m
ed
to
possess
world
wide
a
ffi
li
at
ion
within
t
he
f
uture
we
b
s
uppo
rted
Nex
t
Ge
nerat
ion
netw
orks
.
The
a
scensi
on
in
t
he
use
r
re
qu
e
st
for
know
le
dg
e
acce
ss
act
uate
d
the
em
plo
ye
es
too
lo
oked
for
im
pr
ov
e
d
Ra
dio
Access
Te
chnolo
gies
f
or
deliveri
ng
s
er
vi
ces
to
the
co
nsum
ers
add
it
io
nal
e
xp
e
diti
ou
sly
at
higher
sp
ee
d.
T
he
well
-
est
ablis
he
d
T
hir
d
Gen
e
rati
on
net
work
li
k
e
Un
i
ver
sal
M
obil
e
Tel
ecom
m
un
ic
at
ion
s
S
yst
e
m
create
d
it
at
ta
inable
to
s
upply
el
ev
at
ed
qual
it
y
le
vel
wi
t
h
giant
co
ve
rage
howe
ver
m
i
gh
t
sup
ported
s
olely
low
to
m
edium
rate
that
is
insuf
fici
ent
to
acc
om
pl
ish
app
li
cat
io
ns
th
at
d
esi
res
conti
nuous i
ntern
et
conne
ct
io
n [3
]
.
Fo
r
wireless
c
omm
un
ic
at
ion
,
because
of
re
s
tric
te
d
inf
or
m
at
ion
m
easur
e
a
nd
getti
ng
a
good
se
rv
ic
e,
it'
s
vital
to
li
ft
the
c
hannel
r
euse
rati
o,
t
hus
an
acce
s
sibl
e
an
d
e
ff
ect
iv
e
ha
ndover
f
or
m
ula
ought
to
be
dev
el
op
e
d
[
4].
Nex
t
ge
n
wirel
ess
com
m
un
i
cat
ion
netw
orks
ought
to
gua
ra
ntee
the
pe
rp
et
ually
best
c
onne
ct
ed
(A
BC
)
t
o
m
ob
il
e
us
er.
T
he
perform
ance
of
wireless
net
work
is
im
pr
ov
ed
by
us
i
ng
w
el
l
-
order
e
d
ha
ndover
te
chnolo
gies
once
us
er
is
m
ov
in
g
from
on
e
cel
l
to
a
dif
fe
ren
t
[5
]
.
This
new
era
of
c
om
m
un
ic
at
ion
i
nvolv
e
s
the
inter
w
orking
of
the
m
any
te
chnolo
gies
to
c
onfirm
that
the
us
e
r
is
of
te
n
c
onnect
ed
with
the
world
.
R
el
yi
ng
up
on
the
s
or
t
of
a
ppli
cat
ion
the
us
e
r
is
runn
i
ng,
dif
f
eren
t
ha
ndover
decisi
ons
a
re
t
aken
as
a
n
ex
a
m
pl
e
if
a
us
e
r
is
s
har
i
ng
knowl
e
dg
e
on
a
s
ocial
networ
king
web
sit
e,
he
would
li
ke
to
urge
c
onnect
ed
with
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
.
5
,
Oct
ob
e
r 2
020
:
48
35
-
48
43
4836
the
te
chnolo
gy
that
giv
e
him
with
the
ide
nt
ic
al
facil
it
y
wi
th
le
ss
la
te
ncy
and
per
m
it
hi
m
to
us
e
the
serv
ic
e
s
cease
le
ssly
equ
al
ly
on
ce
he
desires
to
hold
an
on
go
i
ng
cal
l
he
w
ould
li
ke
to
u
r
ge
c
onnected
to
t
ho
s
e
serv
ic
e
[
6]
.
T
hu
s
to
pe
rm
it
the
us
er
to
e
njo
y
the
co
nt
inu
al
ser
vices
an
i
m
pr
ove
d
hand
off
m
echan
ism
is
nee
ded.
Fi
gure
draw
n
bel
ow
sho
ws
t
he
inter
work
i
ng
of
va
rio
us
te
c
h
nolo
gies.
A
new
desi
gn
c
on
ce
pt
,
i.e.
the
Heter
ogene
ous
Ne
tw
orks,
was
intr
oduce
d
i
n
LTE
to
m
eet
de
m
and
f
or
both
cel
lular
netw
ork
c
apacit
y
and
c
overa
ge.
HetNets
'
idea
is
the
de
ploym
ent
of
sev
e
ral
sm
a
ll
er
cells
under
m
acro
cel
l
area
to
exte
nd
cov
e
ra
ge or
i
nc
rease ca
pacit
y i
n
s
om
e h
igh
-
dem
and
a
reas.
HetNets
is
a
m
ajor
cha
nge
in
the
desig
n
of
c
el
lular
netw
ork
s,
offe
rin
g
incr
ease
in
netw
ork
capaci
ty
and
en
ha
nced
cov
e
ra
ge.
HetNets
re
fer
s
to
m
ul
ti
-
acce
ss
netw
orks
with
the
sam
e
UE
(
i.e.
LTE
with
W
i
-
Fi
)
and
m
ay
ref
er
to
hiera
rch
ic
a
l
sect
or
str
uctur
es
w
he
re
m
ulti
ple
c
el
l
str
uctu
res
with
t
he
sam
e
rad
i
o
acce
ss
sta
nd
a
rds
are
us
e
d
(m
acro
c
el
l).
Re
centl
y,
du
e
t
o
the
rap
i
d
de
velo
pm
ent
of
t
he
cu
rr
e
nt
In
te
r
net
a
nd
m
ob
il
e
com
m
un
ic
at
ion
s
i
ndus
try
,
m
ob
il
e
network
i
ng
syst
em
s
ha
ve
bee
n
de
sign
e
d
with
gr
eat
er
in
fr
ast
ru
ct
ur
e
com
plexit
y
and
great
er
div
e
r
sit
y
of
associat
ed
de
vices
a
nd
resou
rces,
as
well
as
m
or
e
dynam
ic
netwo
r
king
form
ation
s.
In
su
ch
em
erg
ing
heter
og
e
ne
ou
s
m
ob
il
e
ne
tworks
(H
et
Nets),
t
her
e
a
re
a
la
rg
e
num
ber
of
te
chn
ic
al
chal
le
ng
es
over
t
he
c
om
plica
ted
syst
em
reso
urces
fo
c
us
i
ng
on
the
ef
fici
ent
m
anag
e
m
ent,
orga
nizat
ion
.
The
wor
king
of
t
he
ef
fici
en
t
wireless
c
omm
un
ic
at
ion
syst
e
m
is
the
m
ai
n
f
ocus
thes
e
days.
Ma
ny
auth
ors
hav
e
co
ntribut
ed
in
t
he
im
pr
ov
em
ent
an
d
i
nnovat
ion
of
t
he
wireless
c
om
m
un
ic
at
ion
s
yst
e
m
s
.
In
t
his
wor
k,
w
e
hav
e
s
ugge
ste
d
the
us
e
of
Ra
dial
Ba
sis
an
d
Ba
c
k
prop
a
gatio
n
neural
to
ols
f
or
m
akin
g
hand
ov
e
r
deci
sion
s
.
The
res
ults
are
deduc
ed
from
the
Ma
tl
ab
fu
nctio
ns
an
d
it
is
s
een
that
RB
F
wo
r
ks
eff
ic
ie
ntly
than
BPN
N
f
or
m
aking
hand
over
decisi
ons
in
wireless
he
te
ro
gen
e
ous
ne
tworks
.
The
s
e
three
featur
e
s ar
e
c
overe
d by the
Het
erogeno
us
wireless
Netw
ork
s:
-
In
te
ll
igent
a
nd
seam
le
ss
netw
ork
-
wide
acc
ess
—
Het
Net
can
im
pr
ove
t
he
netw
ork
s
wi
tc
hin
g
procedu
res
to
delive
r
a
s
ea
m
le
ss
m
ov
em
ent
exp
e
rien
ce
as
us
e
rs
m
ov
e
betwee
n
diff
e
re
nt
Heter
og
e
ne
ous
wire
le
ss
netw
orks.
F
ur
t
her
m
or
e,
HetN
et
s
can
pro
vide
de
vices
with
the
necessa
ry
netw
ork
inf
or
m
at
ion
to
e
na
ble
aut
he
ntica
ti
on
and access t
o v
ario
us
netw
ork
s.
-
Con
sist
e
nt
Net
work
E
xperien
ce
Q
ualit
y
—
HetNets
al
lo
w
diff
e
re
nt
net
w
orks
t
o
e
xc
hange
i
nfor
m
at
ion
on
loads
s
uc
h
as
data
traf
fic
an
d
la
te
ncy.
S
uc
h
capa
bili
ti
es
enab
le
a
m
ob
il
e
dev
ic
e
to
i
de
ntify
the
netw
ork
that
b
est
su
it
s
the
us
e
r'
s
m
ob
il
it
y
need
s.
T
his
re
su
lt
s
i
n
a
co
ns
ist
ent
qual
it
y
of
us
e
r
exp
e
rience
acr
oss
m
ul
ti
ple n
et
w
orks
and
bette
r dist
rib
ution o
f l
oad
s
acr
os
s
di
ff
e
ren
t
netw
orks.
-
Inn
ov
at
ive
a
nd
Dynam
ic
Re
s
ource
Ma
na
ge
m
ent
—
two
t
ypes
of
intel
li
gen
ce
are
in
v
ol
ved
in
dy
nam
ic
resou
rce
m
anag
em
ent:
sp
ect
r
um
agili
ty
and
serv
ic
e
agili
ty
.
Re
co
nf
i
gurab
l
e
ra
dio
with
co
gn
it
ive
f
un
ct
i
ons
and
opport
un
i
sti
c
acce
ss
to
the
sp
ect
ru
m
need
s
s
pectrum
agility
.
On
the
oth
er
ha
nd,
ser
vice
ag
il
ity
requires
virtu
a
li
zat
ion
te
chn
i
qu
e
s
to
ci
rc
umvent
c
om
pu
ta
ti
on
al
netw
orkin
g
issues
.
Coll
e
ct
ively
,
dynam
ic
resou
rce m
anag
em
ent all
ow
s
scal
able, re
-
c
onfig
ur
a
ble a
nd a
dap
ta
ble
networks.
A
struc
ture
of
div
er
se
wirel
ess
acce
ss
net
works
is
Heter
og
e
ne
ous
W
ir
el
ess
Netw
ork
is
sh
own
i
n
F
igure
1
.
W
it
h
reg
ar
d
to
co
ve
rag
e
of
net
wor
k,
te
ch
no
l
og
y,
band
width,
capaci
ty
,
and
net
work
op
e
rato
r,
VHO
is
done
[
7
].
A
Ver
ti
cal
Ha
ndoff
(
VHO
)
occ
urred
i
n
s
uch
a
n
at
m
os
ph
ere
wh
e
re
a
m
ob
il
e
te
rm
inal
(MT)
wi
t
h
a
nu
m
ber
of
ne
twork
inter
fac
es
le
ave
a
wireless
network
a
nd
c
onnects
to
oth
er
wireless
network
[8
]
.
But
on
wh
at
basis
to
perform
,
wh
er
e
to
per
f
orm
,
wh
e
n
to
pe
rform
the
Ver
ti
cal
Handover
a
re
the
key
qu
e
sti
on
s
tha
t
m
us
t
be
determ
ined
to
car
ry
ou
t
a
n
ef
fici
ent
VHO
f
or
m
axim
iz
ing
the
s
up
po
s
ed
QoS
(
Q
ualit
y
of
Ser
vi
ce
)
an
d
QoE
(
qual
it
y
of
e
xp
e
rience
)
pe
rceive
d
by
th
e
us
e
r
[
1].
Sel
ect
ion
of
s
uitable
hand
ov
e
r
de
ci
sion
at
tri
bu
t
e
an
d
assessm
ent
m
e
thod
play
a v
it
al
r
ole i
n
this
re
gard.
Figure
1
.
Heter
og
e
ne
ous
w
irel
ess n
et
wor
ks
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
An
alysis of
bac
k p
r
opag
ation
and
r
adial
basis f
unct
ion ne
ural netw
or
ks f
or
hand
over
....
(
Pay
al
Ma
haja
n)
4837
Hand
ov
e
r
ref
e
r
s
to
the
proce
dure
w
he
n
an
Mob
il
e
Term
inal(M
T)
c
hange
it
s
at
ta
ch
m
ent
po
i
nt
f
ro
m
on
e
access
point(
AP
)
to othe
r
, as
s
how
n
i
n
F
igure
2
.
-
Horizo
ntal Ha
ndover
-
Ver
ti
cal
H
a
ndover
Horizo
ntal
ha
ndove
r
occ
ur
w
hen
a
MT
(m
ob
il
e
te
rm
inal)
m
ov
es
within
t
he
sam
e
wireless
netw
ork
from
on
e
cel
l
to
ano
the
r.
Horizo
ntal
ha
ndoffs
[9
]
can
al
so
be
known
as
I
nter
-
B
SC
handoff,
I
ntra
-
B
SC
hand
off
an
d
Int
er
-
Msc
ha
ndoff
in
t
he
cel
lul
ar
wireless
syst
e
m
.
The
reas
ons
f
or
w
hich
horiz
on
ta
l
ha
nd
ov
e
rs
are
done
a
re:
poor
si
gn
al
q
ualit
y
(RSS)
or
sign
al
l
os
s,
assi
gn
m
ent
of
bandw
i
dth
,
loa
d
ba
la
ncing
an
d
Mob
il
e
Term
inal
sp
eed.
T
he
HHO
(hor
iz
on
ta
l
ha
ndover
)
pr
oce
ss
co
ns
ist
s
of
three
forem
os
t
ste
ps
:
init
ia
ti
on
of
the h
a
ndoff
, as
sign
m
ent o
f
th
e cha
nn
el
,
a
nd
execu
ti
on
of th
e h
a
ndover
/ c
onnecti
on t
ransf
er [4].
Ver
ti
cal
ha
ndov
e
r
can
be
done
in
th
ree
ph
ase
s
cal
le
d
Netw
ork
disc
ov
e
ry
,
Ha
ndover
decisi
on
,
and
e
xec
utio
n
of
Hand
over.
Netw
ork
disco
ver
y
is
a
pr
oce
ss
of
fin
ding
th
e
appr
opriat
e
ne
twork
i
n
su
c
h
a
way
that
the
netw
ork
m
eet
s
us
er
pr
e
f
ere
nces
wi
th
the
require
d
ser
vice
qual
it
y
[2
]
.
T
oo
earl
y
hand
ov
e
r
r
es
ults
in
redu
nd
a
nt
handove
r,
wh
il
e
s
erv
ic
e
qual
it
y
i
s
re
du
ce
d
as
a
resu
lt
of
la
te
ha
ndover
.
Last
l
y,
the
tra
ns
fe
r
sh
oul
d
be
e
xecu
te
d
at
an op
ti
m
al
t
i
m
e. Se
ver
al
alg
os ha
ve been
an
t
ic
ipate
d
f
or v
e
r
ti
cal
h
ando
ver
(VHO) m
echan
ism
s
on
the
basis
of
nu
m
ber
of
para
m
et
ers,
nam
ely
the
received
sign
al
stre
ng
t
h
(RSS)
,
c
onnec
ti
on
costs,
a
vai
la
ble
band
width,
an
d
s
peed
of
the
m
ob
il
e
sta
t
ion
,
serv
ic
e
qual
it
y,
and
sig
nal
to
inter
fer
e
nce
no
ise
rati
o,
gu
aran
t
ee
delay
of
ha
ndover
a
nd
batte
ry
con
s
um
ption
.
The
ob
j
ect
ive
of
the
nex
t
ge
ner
at
io
n
com
m
un
ic
at
io
n
syst
em
s
is
to off
e
r use
rs,
a large
r
a
nge
of
op
ti
ons
[
10
].
Figure
2
.
Ha
nd
ov
e
r
i
n hetero
ge
neous
n
et
wor
ks
2.
RELATE
D
W
ORK
Ma
ny
aut
hors
hav
e
co
ntri
bu
t
ed
in
the
fiel
d
of
wi
reless
c
omm
un
ic
at
ion
.
So
m
e
of
the
work
w
hic
h
helpe
d
us
in
th
is
wo
r
k
are
s
how
n
he
re.
I
n
[
11
]
,
the
a
uthor
s
gav
e
a
m
e
thod
by
w
hich
sel
ect
ion
of
the
optim
u
m
ta
rg
et
net
work
from
the
availab
le
netw
orks
can
be
done
.
This
ap
proac
h
is
desig
ned
t
o
m
axi
m
iz
e
the
degre
e
of
c
us
t
om
er
sat
isfact
ion
by
sel
ect
ing
the
"
best"
net
work
am
on
g
num
ero
us
c
and
i
date
net
w
orks
as
the
ha
ndover
ta
rg
et
netw
ork.
In
[
12
]
,
t
he
auth
or
s
sho
we
d
that
for
sm
oo
t
h
c
onnecti
vi
ty
in
heterog
eneou
s
wireles
s
netw
orks,
t
he
m
ai
n
obj
ect
ive
of
hand
off
is
ch
oo
si
ng
the
be
st
ta
rg
et
net
w
ork
f
r
om
the
existi
ng
netw
ork.
T
o
a
ccom
plish
this
go
al
,
t
hey
des
ign
e
d
a
ta
r
get
netw
ork
sel
ect
ion
al
go
rithm
t
o
im
pr
ov
e
t
he
degree
of
cu
stom
er
sat
isfact
ion
.
T
he
te
c
hn
i
qu
e
f
ocuses
on
a
dwel
l
ti
m
e
an
d
e
stim
at
i
on
of
the
re
cei
ved
sign
al
st
rength
.
In
[
13
]
,
t
he
a
uthors
s
uggest
ed
a
DS
-
CD
MA
cel
lular
ne
twork
tra
ff
ic
m
od
el
co
m
bin
ing
bo
t
h
soft
capaci
ty
and
s
of
t
hand
ov
e
r.
Netw
ork
ou
t
put
is
then
m
eas
ur
e
d
f
or
the
bl
ock
i
ng
of
cal
ls.
I
n
[
14
]
,
the
a
uthor
s
sh
owe
d
that
ba
sed
on
a p
re
-
processe
d
un
der
s
ta
nd
in
g
of
t
he
perform
ance
pro
file
s
of
t
he
li
nk
s
,
ne
ural
net
works
essenti
al
ly
m
a
ke
im
po
rtant
hand
off
c
ho
ic
es
with
t
he
as
sessm
ent
of
the
qu
al
it
y
of
the
li
nk
s
,
inst
ead
of
com
par
ing
t
he
relat
ive
inte
ns
it
y
of
t
he
s
ign
al
.
T
he
res
ults
of
t
he
st
ud
y
a
nd
sim
u
la
ti
on
s
sho
we
d
that
the
sugg
e
ste
d
al
gorithm
enh
ances
use
r
-
pe
rc
ei
ved
qu
al
it
y
in
a
V
oIP
ap
plica
ti
on
com
m
un
ic
at
ion
e
nv
i
ronm
ent
by
reducin
g
both
the
am
ou
nt
of
pac
ket
lost
and
unwa
nted
ha
ndoff
s.
In
[
15
]
,
a
ha
ndover
al
gorit
hm
is
pro
po
se
d,
a
nd
it
s o
utput i
s stud
ie
d
in
pap
e
r.
Hand
off
decisi
on
s a
re u
s
ually
b
ased
on
si
gnal
stren
gt
h,
and m
any
arti
fici
al
intel
lig
ence
tech
niqu
es su
c
h
a
s F
uz
zy
Logic,
ANN
a
nd so on
ar
e u
ti
li
zed lat
el
y.
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
.
5
,
Oct
ob
e
r 2
020
:
48
35
-
48
43
4838
3.
INTROD
U
CTION
T
O
NE
URAL
NETWOR
KS
T
O
OLS
U
SED
F
OR
HAN
DOVER
DE
CIS
I
ON
MEC
HANIS
M
In
this
pap
e
r,
a
com
par
ison
of
ha
ndoff
m
echan
ism
us
ing
diff
e
ren
t
t
oo
ls
of
neural
netw
orks
is
pr
ese
nted
.
He
r
e,
Ba
ck
pr
op
a
gation
a
nd
Ra
dial
Ba
sis
Functi
ons
to
ols
of
neural
net
works
are
us
e
d
an
d
a
com
par
iso
n
betwee
n
th
e
e
ff
ic
ie
ncy
of
both
t
he
to
ols
i
s
presente
d
de
pendin
g
upon
their
r
ole
in
ta
king
the
hand
over
decisi
ons
with
the
us
e
of
m
a
tl
ab
cod
i
ng.
In
this
sect
ion
,
a
n
intr
oductio
n
to
bo
th
t
he
to
ols
is
giv
e
n
a
nd the
n i
n
the
ne
xt sect
ion
a
co
m
par
is
on b
et
ween t
he
m
is
pr
ese
nted
.
3
.1
.
B
ack pr
opagat
i
on
neu
ral ne
twork
Ba
ck
propagat
ion
,
a
c
om
pr
es
sion
f
or
"i
n
re
ver
se
pr
olifera
ti
on
of
er
rors
",
is
a
ty
pical
strat
egy
to
trai
n
the
a
rtific
ia
l
neural
syst
em
s.
Fr
om
a
per
fect
ou
t
pu
t,
th
e
syst
e
m
gains
from
var
io
us
i
nfor
m
at
ion
s
our
ces
i.e.
in
puts,
j
ust
li
ke
the
m
anner
in
w
hich
a
child
fig
ur
es
out
a
way
to
ac
knowle
dge
a
dog
.
BP
N
N
i.e.
ba
c
k
pro
pag
at
io
n
ne
ur
a
l
netw
ork
is
a
super
vis
ed
le
ar
ning
st
rategy,
a
nd
is
a
sp
ec
ulati
on
of
the
delta
ru
le
.
BPNN
nee
ds
a
dataset
of
the
pe
rf
ect
outpu
t
f
or
a
fe
w
inputs
t
hat
m
a
kes
up
the
tra
ining
set
.
It
is
m
os
t
us
ef
ul
for
fee
d
for
ward
syst
e
m
s.
Ba
ck
pr
opa
gation
ne
cessi
ta
te
s
the
act
ivati
on
operate
us
e
d
by
arti
fici
al
neur
on
s
(
or
"
hubs
"
)
be
di
fferentia
ble
[9
].
The
ta
rg
et
of
a
supe
rv
ise
d
le
arn
i
ng
al
go
rithm
ic
pr
ogra
m
is
to
search
a
functi
on
w
hich
su
it
s
best,
a
c
ollec
ti
on
of
in
put
to
i
ts
pr
eci
se
outp
ut.
Var
ie
ty
of
i
nput
(i/p
)
a
nd
ou
t
pu
t
(o
/
p)
patte
rn
is
effor
tl
essly
w
el
l
-
read
by
sin
gle
-
la
ye
r
ne
ura
l
networks
ad
di
ti
on
al
ly
ref
erre
d
to
as
pe
rcept
rons.
Th
ough,
these
per
ce
ptr
on
s
c
an'
t
be
trai
ned
for
m
any
relat
ively
si
m
ple
patte
rn
s
for
ex
a
m
ple
tho
se
w
ho
a
re
li
near
ly
ind
i
vis
ible
.
F
or
insta
nce,
a
hu
m
an
bein
g
wil
l
cl
assify
an
pictu
re
of
an
a
nim
a
l
by
rec
ognizin
g
m
an
y
char
act
e
risti
cs
li
ke
the
feel
of
the
s
kin,
(whethe
r
it
'
s
feath
ered,
f
urry,
sc
al
ed)
,
num
ber
of
li
m
bs
,
the
anim
a
l
siz
e
an
d
lots
of
al
te
rn
at
ive
c
ha
racteri
sti
cs
[
16
]
.
A
sin
gle
-
la
ye
r
ed
ne
ural
net
work,
ho
wev
e
r
m
us
t
le
arn
a
f
un
ct
io
n
t
hat
ou
t
pu
ts
a
la
be
l
exclusively
us
in
g
the
inten
sit
y
of
the
pix
el
s
within
the
im
age
.
A
m
ulti
-
la
ye
red
net
work
ove
rco
m
e
t
his
const
raint
becaus
e
i
t
will
turn
ou
t
inn
e
r
re
pr
es
ent
at
ion
s
a
nd
get
t
raine
d
f
or
va
riou
s
c
har
act
eris
ti
cs
in
ever
y
la
ye
r
[1]
.
The
pri
m
ary
la
ye
r
can
be
r
e
sp
onsi
ble
f
or
l
earn
i
ng
t
he
ori
entat
ion
s
of
li
nes
victim
isa
tio
n
the
in
puts
f
ro
m
the
each
sing
le
picture
el
em
ent
within
the
im
age
[
17
]
.
T
he
seco
nd
la
ye
r
m
igh
t
j
oi
n
the
op
ti
ons
or
ch
aracte
risti
cs
le
arn
e
d
within
the
1
st
la
ye
r
and
obta
in
trai
ne
d
to
ac
knowle
dge
s
ha
pes
li
ke
ci
rcles
.
W
it
h
e
ver
y
hi
gh
e
r
la
ye
r,
it
le
arn
s
add
it
io
nal
c
ha
racteri
sti
cs.
T
he
ta
rg
et
a
nd
i
ns
pi
rati
on
f
or
de
velo
pm
ent
of
the
rea
r
pro
pag
at
io
n
al
gorith
m
(BPNN)
is
t
o
s
eek
out
a
way
to
coac
h
m
ulti
-
la
ye
red
ne
ur
al
netw
orks
s
uch
it
will
le
arn
an
y
ran
dom
m
ap
ping
of in
pu
t t
o ou
t
pu
t
Dif
fer
e
ntiable
acti
vatio
n
f
un
ct
io
ns
[1
8
]
.
The
BP
NN
sea
rch
t
he
le
ast
va
lue
of
the
e
rror
functi
on
in
we
igh
t
s
pace
by
usi
ng
t
he
gra
dient
desce
nt
m
et
ho
d.
Th
e
a
m
al
ga
m
at
ion
of
wei
gh
ts
t
hat
re
du
c
es
th
e
error
f
un
ct
i
on
is
sup
pose
d
to
be
a
n
a
nswer
of
the lea
rn
i
ng
prob
le
m
. A
s the
m
et
ho
d need
s
cal
culat
io
n of
t
he
er
ror
f
unct
ion
gr
a
die
nt at ever
y st
ep o
f
it
erati
on,
on
e
s
houl
d
en
su
re
th
e
per
m
anen
ce
a
nd
di
ff
e
ren
ti
abili
ty
of
er
r
or
f
unct
ion.
A
pp
a
re
ntly
on
e
nee
d
to
util
iz
e
a
ty
pe
of
act
ivati
on
f
un
ct
io
n
apar
t
f
ro
m
the
ste
p
functi
on
wh
ic
h
is
us
ed
i
n
perce
ptr
ons,
b
ecau
se
the
co
m
po
sit
e
functi
on
create
d
by
c
onsist
en
t
per
ce
ptr
on
s
is
irre
gu
la
r.
Out
of
al
l,
the
po
pu
la
r
act
ivati
on
f
unct
ions
f
or
bac
k
pro
pag
at
io
n ne
tworks
is sigm
oid
f
un
ct
io
n [19]
.
s
c:
giv
e
n b
y t
he
ex
pressi
on
(
)
=
1
1
+
−
The
co
ns
ta
nt
c
m
a
y
be
ch
ose
n
in
discrim
inate
ly
and
it
s
reciprocal
1/c
is
known
as
th
e
tem
per
at
ur
e
par
am
et
er
in
r
andom
neural
netw
orks.
T
he
sh
a
pe
of
t
he
sigm
oid
cha
ng
e
s
in
ste
p
wh
e
n
t
he
va
lue
of
c
increases
.
T
he
hig
he
r
t
he
va
lues
of
c,
it
br
in
gs
the
sigm
oid
sh
a
pe
li
ke
that
of
th
e
ste
p
functi
on
an
d
with
the lim
it
c=
1,
the si
gm
oid
con
verge t
o
ste
p
f
un
ct
io
n
at
ori
gi
n.
Fig
ure
3
s
hows
a
rch
it
ect
ur
e
of
bac
k
prop
a
gatio
n.
M
ulti
-
la
ye
r,
Fee
d
f
orwa
rd
net
works
have
the
featu
res
s
uc
h
as:
T
hey
m
us
t
ha
ve
in
any
case
one
hidd
en
la
ye
r.
It
sho
uld
ha
ve
unit
s
wh
ic
h
are
non
-
li
near
.
All
la
ye
rs
m
us
t
be
c
onnecte
d
betwee
n
unit
s
in
tw
o
seq
ue
nt
ia
l
la
ye
rs
.I
t
is
not
re
quired
t
hat
outp
ut
unit
s
m
us
t
hav
e
in
put
un
i
ts
equ
al
in
nu
m
ber
.
Nu
m
ber
of
Hidden
unit
s
per
la
ye
r
do
es
not
de
pend
upon
the
nu
m
ber
of
input o
r
outp
ut unit
s [20
]
.
3.1.1
.
Le
arnin
g
in
b
ack pro
pag
at
i
on
networ
ks
a.
Learn
i
ng in ba
ck pr
opagati
on
n
et
w
orks
Learn
i
ng
i
n
ba
ck
pro
pa
gatio
n
netw
orks:
W
e
igh
ts
init
ia
ll
y
t
ake
sm
al
l
and
rando
m
value
s
for
eac
h
trai
ning
pai
r
(i
nput
/
outp
ut).
Each
in
put
uni
t
transm
it
s
it
s
value
t
o
eve
ry
hidden
unit
.
E
ver
y
hidden
un
it
sets
it
s input sig
nal
and act
ivate
s it
s outp
ut sig
nal
cal
culat
ion
fun
ct
ion
.
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
An
alysis of
bac
k p
r
opag
ation
and
r
adial
basis f
unct
ion ne
ural netw
or
ks f
or
hand
over
....
(
Pay
al
Ma
haja
n)
4839
b.
Ba
ck pr
op
a
gati
on s
ta
ge
Each
ou
t
pu
t
ca
lc
ulate
s
it
s
te
rm
of
error
,
it
s
own
te
rm
of
w
ei
gh
t
c
orrecti
on
a
nd
it
s
te
rm
of
c
orrecti
on
and
se
nd
s
it
do
wn
to
t
he
la
ye
r
belo
w.
E
ver
y
hidden
unit
s
um
s
up
it
s
delta
input
f
ro
m
above
a
nd
m
ulti
pli
es
by
it
s acti
vation f
un
ct
io
n de
rivat
ive; i
t al
so
calc
ulate
s it
s own
weig
ht correcti
ng and
bias c
orrecti
ng term
.
c.
Adjusti
ng the
w
ei
ghts
Each
ou
t
put
(
o/p)
unit
up
da
te
s
firstly
it
s
weig
hts
an
d
t
he
n
bias
an
d
e
ve
ry
hidde
n
un
i
t
updates
it
s
weig
hts
a
nd
th
e
n
bias
te
rm
.
Ever
y
t
rainin
g
seq
ue
nce
is
know
n
as
an
ep
och.
T
he
weig
hts
are
re
orga
ni
sed
i
n
ever
y cy
cl
e
Figure
3
.
A
rch
i
te
ct
ur
e
of
b
ac
k
pro
pa
gation
3
.2
.
In
tr
od
uc
tion t
o
r
ad
ial
ba
sis
fu
nc
tio
n
s
Moody
a
nd
Da
rk
e
n
a
nd
Br
oo
m
head
an
d
Lo
we,
w
e
re
the
fi
rst
pe
rs
ons
to
m
ake
us
e
of
t
he
ra
dial
basis
functi
on
(RBF
)
in
t
he
syst
e
m
of
neural
ne
tworks
(
NN).
RB
F
net
wor
ks
a
re
the
fee
d
-
forwar
d
networks
pr
act
ic
ed
with
the
us
e
of
s
uper
vised
t
raini
ng
al
gorithm
.
These
a
re
m
os
tl
y
well
-
order
e
d
with
an
i
nd
i
vidual
hidden
la
ye
r
of
un
it
s
w
hose
act
ivati
on
f
un
c
ti
on
is
carefu
ll
y
sel
ec
te
d
a
m
o
ng
a
cl
us
te
r
of
fu
nctio
ns
kn
own
as
basis
functi
on.
Thou
gh
they
are
si
m
il
ar
to
BPNN
in
m
any
structur
es
,
Ra
dial
Ba
sis
Fu
nction
netw
or
ks
ha
s
a
nu
m
ber
of
a
dv
a
ntage
s.
T
he
y
gen
e
rall
y
le
arn
faster
t
han
B
PN
N
.
Ra
dia
l
basis
f
un
ct
io
ns
are
rece
nt
ways
of
appr
ox
im
at
ing
sever
al
m
ulti
v
ariable
f
un
ct
io
ns
.
T
hey
have
been
recog
nized,
exam
ined
a
nd
te
ste
d
f
or
s
ever
al
ye
ars
now
[
21
].
The
hidde
n
unit
s
of
fe
r
f
un
ct
ion
s
that
sit
uate
a
ran
d
om
bas
e
fo
r
the
in
pu
t desig
n
[
22
]
.
T
her
e
is
a
non
-
li
nea
r
co
nv
e
rsion
f
ro
m
input
sp
ace
to
hidden
un
it
an
d
li
near
co
nv
e
r
sion
f
ro
m
hid
de
n
unit
to
the
ou
tp
ut
un
it
.
The
hi
dden
la
ye
r
functi
on
s
pr
oduce
si
gn
i
ficant
non
-
zero
res
pons
e
on
ly
at
ti
m
es
t
he
in
put
dr
op
s
withi
n
a
sli
gh
tl
y
lim
i
te
d
re
gion
of
the
in
pu
t
[
23
]
.
E
ver
y
hidde
n
un
it
has
it
s
se
pa
rate
rece
ptive
fiel
d
in
input
sp
ace
[24
].
T
he
arc
hitec
ture of
Radial
Basi
s
Functi
ons is
di
sp
la
ye
d
i
n
the
foll
ow
i
ng Fig
ur
e
4.
Figure
4
.
A
rch
i
te
ct
ur
e
of
ra
dial basis
functi
on
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
.
5
,
Oct
ob
e
r 2
020
:
48
35
-
48
43
4840
The
m
ai
n
featu
res of
RB
F ar
e
[25
]
:
-
These a
re t
wo lay
ered
feed
-
f
or
ward
netw
orks
.
-
The hid
de
n nodes
pro
vid
e a
s
et
o
f
Radial
Ba
sis Funct
io
ns
(
e.g
.
G
a
us
sia
n f
un
ct
io
ns).
-
The o
utput n
od
es provide
li
ne
ar c
on
ci
se
f
un
c
ti
on
s as
in
B
P
NN.
-
The
net
work
t
rainin
g
ha
ve
two
st
a
ges:
fir
stl
y
the
weight
s
fr
om
the
inp
ut
la
ye
rs
to
hidden
la
ye
r
a
re
sp
eci
fied
,
a
nd
then
the
weig
ht
s
from
the
hi
dden
la
ye
r
t
o
outp
ut
la
ye
r
a
re
sp
eci
fied.
T
he
trai
ning
is
rea
ll
y
fast. T
hese
net
works ar
e
bette
r
f
or
do
i
ng inte
rpolat
ion.
3
.3
.
D
iffere
n
ce betw
een
t
h
e two
too
ls
Com
par
ison
in
the
perform
ance
of
BP
N
N
an
d
RB
F
network
s
fa
sci
na
te
d
the
thou
gh
t
fu
l
ness
of
the
m
any
research
e
rs.
Ma
ny
r
esearche
rs
f
ou
nd
that
t
he
RB
F
def
eat
e
d
the
BPNN
in r
eco
gn
it
io
n
of
voic
e
wh
il
e
so
m
e
oth
er
res
earche
rs
fou
nd
that
Mult
i
-
la
y
er
pe
rce
ptr
on
i.
e.
BPN
N
is
s
uperi
or
t
o
RB
F
in
fa
ult
tolera
nc
e
an
d
al
locat
i
on
of
r
eso
ur
ces
a
pp
li
cat
ion
s
[
26
].
Additi
on
al
ly
appr
ov
i
ng
the
ve
ry
fact
that
BPNN
with
sigm
oid
act
ivati
on
func
ti
on
is
s
uperio
r
f
or
cl
as
sific
at
ion
a
pp
li
cat
ion
s
[
27
]
.
B
P
NN
with
si
gm
oid
act
ivati
on
di
dn’t
so
le
ly
at
ta
in
th
e
si
m
plest
m
e
an
s
q.
er
ror,
ho
wev
e
r
c
onjoint
ly
wish
ed
the
m
ini
m
u
m
train
in
g
ti
m
e
in
on
-
line
le
arn
in
g
in
al
l
dataset
s
[28
].
Ap
a
rt
from
the
si
m
il
arit
ie
s
of
top
ol
og
ie
s,
R
BF
networ
ks
and
BP
NN
pe
r
f
or
m
s
in
a
un
iq
ue
way
.
RB
F
networks
are
m
od
est
than
BPN
N
wh
i
ch
m
igh
t
ha
ve
m
or
e
than
3
la
ye
rs,
s
o
the
tra
ining
m
et
ho
d
is
usua
ll
y
qu
ic
ke
r
tha
n
BP
NN.
It
is
vital
to
fi
x
acc
ur
at
e
ea
rly
sta
te
s
f
or
RB
F
net
works w
he
reas
BPN
N
us
e
a
rb
it
ra
rily
create
d
par
am
et
ers
pri
m
arily
[2
1
].
The
cl
assifi
ca
ti
on
proce
dures
f
or
RB
F
ne
tworks
an
d
B
PNN
are
total
ly
different
.
T
he
RB
F
netw
ork
pri
m
aril
y
based
m
od
el
s
are
li
near
in
pa
ram
e
te
rs
an
d
thus
assure
conve
rg
e
nce
to
their
ideal
va
lues
for
spe
ci
fic
net
work
a
r
chite
ct
ur
e
.
Gro
wth
of
the
RB
F
netw
ork
m
od
el
s
o
needs
f
e
we
r
s
a
m
ples
an
d
fa
ult
an
d
th
us,
a
sm
aller
am
o
un
t
of
ti
m
e
and
ef
f
or
t
,
t
ha
n
that
de
sired
by
BP
appr
oach
[
29
]
.
4.
PROP
OSE
D APP
ROAC
H
In
this
pa
pe
r,
hand
ov
e
r
te
ch
nique
base
d
on
two
diff
e
re
nt
too
ls
of
ne
ural
network
s
is
sh
own
a
nd
the
be
st
am
on
g
them
is
sel
ect
ed.
Wh
e
n
t
he
res
ults
f
or
both
t
he
to
ols
wer
e
draw
n,
it
was
seen
that
RB
F
is
bette
r
than
Ba
ck
pro
pa
gation
of
ne
ural
networks
in
m
any
aspects.
Let
us
ex
plain
this
fr
om
the
resul
ts
and
si
m
ulati
on
s
of
the
to
ols.
We
a
re
usi
ng
tw
o
m
et
hods
of
ne
ural
netw
ork
s
i.e
.
Ba
ck
pro
pag
at
ion
ne
ur
al
net
works
(BPNN)
an
d
rad
ia
l
ba
sis
f
un
ct
io
n
(RBF)
neural
net
work
m
et
ho
d
f
or
ha
ndover
de
ci
sion
s
i
n
w
irel
ess
com
m
un
ic
at
ion
netw
orks.
Re
su
lt
s
of
im
ple
m
entat
ion
ha
ve
bee
n
su
m
m
arized
in
the
T
able
1.
F
ro
m
t
he
T
able
1,
it
is
seen
tha
t
cl
assifi
cat
ion
a
ccur
acy
a
nd
tra
ining
ti
m
e
increases
with
inc
r
ease
in
num
ber
of
neurons
i
n
hidden
la
ye
r
in
case
of
bo
t
h
BPN
N
and
RB
FNN.
Table
1
s
how
that
cl
assifi
cat
i
on
acc
ur
acy
of
RB
FN
N
is
be
tt
er
than
BPN
N
but
the
tim
e
ta
ken
by
RB
F
is
m
or
e
than
t
hat
of
B
PN
N
.
Wh
e
n
nu
m
ber
of
hidden
la
ye
r
ne
uron
s
are
50
f
or
give
n
dataset
of
2500
trai
ning
in
put
sa
m
ples,
cor
r
esp
onding
cl
assifi
cat
ion
accu
racy
for
BPN
N
is
88
%
an
d
trai
ni
ng
tim
e
is
on
ly
0.32
m
inu
te
s
an
d
acc
ur
acy
of
RB
FN
N
is
58
%
,
trai
ni
ng
ti
m
e
is
0.
06
m
inu
te
s
.
B
ut
w
he
n
nu
m
ber
of
hidde
n
la
y
er
ne
uro
ns
i
nc
reases
t
o
50
0,
cl
assi
ficat
ion
acc
ur
acy
for
BP
NN
is
83%
a
nd
trai
ni
ng
tim
e
increases
gr
a
dual
ly
to
2
m
in
utes
an
d
27
se
conds
c
orres
pond
i
ng
acc
urac
y
in
case
of
R
BFNN
is
90
a
nd
ti
m
e
ta
ken
is
11
m
i
nu
te
s
a
nd
55
s
econds
.
W
hen
nu
m
ber
of
hi
dd
e
n
la
ye
r
ne
uro
ns
inc
reases
to
500,
accu
ra
cy
of
RB
FN
N
inc
rea
ses
and
it
s
correspo
nd
i
ng
value
of
MSE
de
crease
(
0.03).
The
res
ults
for
the
neural
networ
k
too
ls
a
re
cal
c
ulate
d
a
fter
runn
i
ng
the
M
A
TLAB
c
odes
assigne
d
t
o
ba
ck
propagati
on
an
d
Ra
d
ia
l
Ba
sis
functi
ons
respec
ti
vely
.
Thes
e
co
des
a
re
gi
ven
i
n
t
he
a
ppen
dix.
Ta
ble
1
gi
ves
th
e
perform
ance
of
bo
t
h
the to
ols
discu
ssed
a
bove
.
Table
1.
Sho
wing co
m
par
iso
n betwee
n
t
he
tr
ai
nin
g t
im
es ca
lc
ulate
d usin
g R
BF an
d
b
ack
pro
pag
at
io
n
Serial
n
u
m
b
er
n
u
m
b
er
of
n
eu
ron
s
Accurac
y
BPNN (%
)
Accurac
y
RB
F (
%)
Tr
ain
in
g
ti
m
e
BPNN in
m
in
u
tes
Tr
ain
in
g
T
i
m
e
RB
F in
m
in
u
tes
1
50
80
58
0
0
:3
2
0
0
:0
6
2
100
79
64
0
0
:5
0
0
0
:1
1
3
150
81
68
0
0
:5
6
0
0
:1
9
4
200
83
71
0
1
:0
7
0
0
:2
8
5
250
79
78
0
1
:1
8
0
1
:5
7
6
300
87
81
0
1
:3
0
0
3
:5
3
7
350
82
85
0
1
:4
2
0
5
:4
3
8
400
79
86
0
1
:5
3
0
7
:4
2
9
450
83
88
0
2
:1
5
0
9
:5
5
10
500
83
90
0
2
:1
7
1
1
:5
5
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
An
alysis of
bac
k p
r
opag
ation
and
r
adial
basis f
unct
ion ne
ural netw
or
ks f
or
hand
over
....
(
Pay
al
Ma
haja
n)
4841
Figure
5
s
how
s
an
im
po
rtant
par
am
et
er
w
hich
helps
in
dec
iding
the
best
hand
ov
e
r
decis
ion
s
.
I
n
t
hi
s
era
wh
e
re
al
l
t
he
te
ch
nolo
gie
s
are
i
nter
wor
king
ti
m
e
m
an
agem
ent
in
ha
ndover
decisi
ons
is
ve
ry
im
p
or
ta
nt.
As
the
us
e
r
t
od
ay
wa
nts
to
en
j
oy
t
he
un
interr
up
te
d
se
r
vices
w
hile
m
ov
i
ng
f
ro
m
one
place
to
a
no
t
her.
The
gr
a
ph
dra
wn
a
bove
pro
ve
s
that
the
ti
m
e
ta
ken
by
RB
F
in
ha
ndover
is
pr
act
ic
al
ly
no
t
reali
sable
as
th
ere
is
a
pa
us
e
of
up
to
t
welve
m
i
nu
te
s
f
or
50
0
neur
o
ns.
In
th
e
m
eantim
e,
al
l
the
ser
vices
to
w
hich
t
he
us
er
i
s
connecte
d
a
re
at
a
halt
wh
ere
as
the
m
axi
m
um
t
i
m
e
ta
ken
by
the
BPN
N
netw
orks
f
or
m
axi
m
u
m
nu
m
ber
of
neur
on
s
is
f
our
m
inu
te
s
for
500
ne
urons.
T
houg
h
the
ef
fici
ency
pro
vid
e
d
by
RB
F
is
m
uch
m
or
e
than
th
at
of
BPNN b
ut t
he fact
can
not be
neg
le
ct
e
d
that t
he
ti
m
e taken
by
RB
F is v
e
ry large
.
Figure
5.
T
rai
ni
ng
ti
m
e fo
r
BPN
N
a
nd RB
F
The
Fig
ure
6
s
hows
the
c
om
par
iso
n
bet
ween
the
accuracy o
f
RB
F
an
d
BPNN
a
nd
it
is
cl
early
sh
ow
n
that
the
eff
ic
ie
ncy
of
RB
F
is
bette
r
than
that
of
BPN
N.
F
or
500
ne
uro
ns
the
accu
racy
for
RB
F
reaches
up
t
o
90% whereas t
hat o
f
BP
NN
i
s j
us
t 8
3%
.In takin
g
ha
ndove
r
decisi
ons,
ac
cur
acy
p
la
ys a vita
l ro
le
. A
cc
ur
acy
in
case
of
ha
ndov
er
m
eans
how
accuratel
y
the
hand
ov
e
r
decisi
on
s
a
re
ta
ke
n
wh
e
n
us
er
e
nte
rs
the
c
ov
e
ra
ge
area
of
new
base
st
at
ion
or
the
use
r
inten
ds
to
m
ake
us
e
of
ot
her
se
rv
ic
es
r
at
her
tha
n
that
serv
ic
e
wh
ic
h
he
is
c
urren
tl
y u
sin
g.
Figure
6
.
G
raph
betwee
n
cl
as
sific
at
ion
acc
uracy
o
f
RB
F,
B
PNN
v/s num
ber
of
ne
uro
ns
5.
CONCL
US
I
O
N
The
ha
ndove
r
decisi
ons
can
be
ta
ke
n
us
i
ng
conve
ntion
al
so
urces
a
nd
s
oft
com
pu
ti
ng
t
echn
i
qu
e
s.
In
this
pa
pe
r,
it
is
sh
own
that
wh
e
n
the
ne
ural
netw
ork
to
ols
BPN
N
an
d
RB
F
are
us
ed
fo
r
ta
ki
ng
handove
r
decisi
ons,
RB
F
netw
orks
is
pro
ved
to
be
be
tt
er
than
BPN
N
in
te
rm
s
of
accuracy
but
ti
m
e
ta
ken
by
RB
F
is
m
uch
m
or
e
than
that
of
BP
N
N.
T
he
perf
orm
ance
of
t
hese
cl
assifi
ers
is
evaluated
on
th
e
basis
of
nu
m
ber
of
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
.
5
,
Oct
ob
e
r 2
020
:
48
35
-
48
43
4842
neur
on
s
i
n
hidden
la
ye
r
,
trai
ni
ng
ti
m
e
and
cl
assifi
cat
ion
ac
cur
acy
.
This
pa
per
s
how
s
tha
t
RB
F
neural
ne
twor
k
giv
es b
et
te
r
res
ults
f
or
h
a
ndov
er
decisi
on
i
n
wireless
c
ommun
ic
at
io
n
netw
orks
with
cl
ass
ifi
cat
ion
acc
uracy
of
90%.
Th
us
,
it
can
be
co
nclu
de
d
that
though
RB
F
pr
ovid
es
m
or
e
accuracy
bu
t
it
ta
kes
a
huge
tim
e
fo
r
m
akin
g
the
decisi
ons.
The
f
uture
w
ork
s
uggeste
d
in
this
pa
per
are
ot
her
ne
ur
al
netw
ork
te
chn
i
qu
e
s
w
hi
ch
can
pr
act
ic
al
ly
m
a
ke
possible
the
us
e
of
neural
netw
orks
i
n
ha
ndover
decisi
o
ns
a
nd
these
te
chn
i
qu
e
s
ca
n
be
C4
.
5
decisi
on tree
, Nave
Bay
es, B
ay
es n
et
.
REFERE
NCE
S
[1]
B.
Van
Quang,
et
al
.
,
“
A
surve
y
on
handof
fs
-
le
ss
ons
for
60
GH
z
base
d
wire
le
ss
s
y
stems
,”
IEEE
Comm
unic
ati
ons
Surve
ys
&
Tutor
ial
s,
vo
l.
14
,
no.
1
,
pp
.
64
-
86
,
20
12
.
[2]
X.
Yan,
e
t
al.,
“
A
surve
y
of
ver
ti
c
al
handov
er
d
ec
ision
al
gori
th
m
s
in
Fourth
G
ene
ra
ti
on
he
te
ro
gene
ous
wire
l
ess
net
works
,”
Computer
Ne
tworks
,
vol.
54
,
no
.
11
,
p
p.
1848
-
1863
,
2
010
.
[3]
M.
M.
A.
Khan,
et
a
l.,
“
Inte
rwo
rking
bet
we
en
W
iMAX
and
U
MT
S
to
provide
sea
m
le
ss
services
,”
Int
ernati
on
al
Journal
of
the P
hysic
al
Scienc
es
,
vol.
5
,
no.
6
,
pp.
852
-
864
,
2010
.
[4]
C.
H.
Lee
and
C
.
J
.
Yu
,
“
An
int
ellige
n
t
han
doff
al
gori
thm
for
wire
le
ss
co
m
m
unic
at
ion
s
y
stems
using
gre
y
pre
di
c
ti
on
and
f
uzzy
de
ci
sion
s
y
stem
,”
2004
IEEE
Inte
rnationa
l
Confe
renc
e
on
Net
working
,
Sen
sing
and
Control,
v
ol.
1
,
pp
.
541
-
5
46,
2004
.
[5]
L.
C
.
Paul,
“
Handoff/
Handove
r
Mec
hani
sm
for
Mobili
t
y
Im
prove
m
ent
in
W
ire
le
ss
Comm
unic
a
ti
on
,”
Global
Journal
of
R
ese
a
rche
s
in Engi
n
eer
ing
,
vol
.
13
,
no
.
16
,
2014
.
[6]
P
.
Maha
ja
n,
e
t
al
.
,
“
A
Survey
on
Handove
r
bet
wee
n
Diffe
ren
t
H
y
brid
W
i
rel
ess
Networks
,
”
Inte
rnat
ional
conf
ere
n
ce of
m
athe
mati
cal
and
engi
ne
ering
sci
e
nce
s
,
2014
.
[7]
P
.
Maha
ja
n,
e
t
al
.
,
“
Handove
r
base
d
on
Bac
k
propa
gation
,
”
In
te
rnational
Mul
t
it
rack
Confe
ren
ce
on
Sci
en
ce,
Engi
ne
ering
and
Technical
Innov
ati
ons
,
2014
[8]
P
.
Maha
ja
n
,
et
a
l.
,
“
Handove
r
D
ec
ision
Me
cha
ni
sm
in
Inte
rworking
Te
chno
logi
es
Us
ing
Radi
al
B
asis
Functi
ons
,
”
IAE
S
In
te
rnation
al
Journal
of
Art
if
icial
Intelli
g
en
ce
,
v
ol
.
3
,
n
o
.
2
,
pp.
79
-
83
,
2014
.
[9]
M.
T.
Haga
n
a
nd
M
.
B.
Men
haj
,
“
Tra
in
ing
fee
d
forward
net
work
s
with
the
Marqua
rdt
a
lgori
thm
,”
IE
E
E
Tr
ansacti
ons on Neural
Ne
tworks,
vol
.
5
,
no
.
6
,
pp
.
989
-
993
,
1994
.
[10]
O.
A.
Oludola
p
o,
et
a
l.,
“
Com
p
ari
ng
per
form
an
ce
of
MLP
an
d
RBF
neur
al
netw
ork
m
odel
s
for
pre
dic
t
ing
Sout
h
Afric
a’s energ
y
consum
pti
on
,”
J
ournal
of Ene
rg
y
in
South
ern
Afr
ic
a
,
vo
l.
23
,
no.
3
,
pp.
40
-
46,
20
12
.
[11]
J.
Mada
an
,
e
t
a
l
.
,
“
A
new
m
et
h
od
for
handof
f
t
arg
et
n
et
work
s
el
e
ct
ion
,”
Inte
rn
ati
onal
Journal
of
Robotics
and
Aut
omation
(
IJR
A)
,
vol. 8, no. 1,
pp.
36
-
43
,
2019
.
[12]
J.
Verm
a,
et
al.,
“
Ta
rge
t
Netwo
rk
Selecti
on
Al
gorit
hm
base
d
o
n
Requi
r
ed
Dw
el
l
Ti
m
e
Esti
m
a
ti
on
,”
Indon
esia
n
Journal
of
Elec
t
rical
Engi
ne
erin
g
and
Informati
c
s (
IJE
EI)
,
vol.
6,
no.
2
,
pp.
172
-
1
82
,
2018
.
[13]
S
.
L.
Su,
et
al
.
,
“
Perfor
m
anc
e
a
naly
s
is
of
soft
h
andof
f
in
CDMA
ce
ll
ul
ar
net
w
orks
,”
I
EE
E
Jou
rnal
on
sele
ct
ed
areas i
n
commu
nic
ati
ons
,
vol
.
1
4,
no
.
9
,
pp
.
176
2
-
1769
,
1996
.
[14]
T.
Li
n
,
e
t
al
.
,
“
A
neur
al
-
ne
twork
-
base
d
con
text
-
awa
re
hando
ff
al
gor
it
hm
fo
r
m
ult
imedia
c
om
puti
ng
,”
AC
M
Tr
ansacti
ons on Multi
media
Com
puti
ng,
Comm
unic
ati
ons
,
and
Ap
pli
cations (
TO
M
M)
,
vol.
4
,
no
.
3
,
pp.
1
-
23
,
2008.
[15]
P.
P.
Bha
tt
a
cha
r
y
a
,
“
Applicat
ion
of
ar
ti
fi
ci
a
l
n
e
ura
l
n
et
work
in
cellular
han
dof
f
m
ana
gement
,”
i
n
In
te
rnationa
l
Confe
renc
e
on
C
om
putat
ional Int
el
li
g
ence
and
M
ult
imedi
a
App
li
c
ati
ons (
ICCIMA 2007)
,
pp.
237
-
2
41.
2007
.
[16]
Sati
shkum
ar,
“N
eur
al ne
tworks
:
a
c
la
ss
room
appr
oac
h
,
”
MC Gr
a
w Hil
ls c
om
panies
,
2004
.
[17]
S.
Jos
hi
and
M
.
Borse
,
“
De
te
c
ti
on
and
Pre
dic
ti
on
of
Di
a
bet
es
Me
ll
it
us
Us
ing
Bac
k
-
Propaga
t
ion
Neur
al
Network
,”
2016
Inte
rnational
Co
nfe
renc
e
on
Micr
o
-
El
ectronics
and
Tele
communi
cat
ion
Eng
ine
eri
ng
(
ICMETE
)
,
pp.
110
-
113
,
20
16.
[18]
P.
W
ang,
et
al
.
,
“
Apply
ing
ba
c
k
-
propa
gation
neur
al
ne
twork
to
pre
dic
t
bus
tra
ffic
,”
i
n
2016
12th
Inte
rnation
al
Confe
renc
e
on
N
atural
Computat
ion,
Fuzzy
Syst
e
ms
and
Knowle
d
ge
Discov
ery
(
ICNC
-
FSKD)
,
pp.
752
-
756
,
2016
.
[19]
P
.
Maha
ja
n,
e
t
al
.
,
“
Handove
r
base
d
on
Bac
k
propa
gation
,
”
In
te
rnational
Mul
t
it
rack
Confe
ren
ce
on
Sci
en
ce,
Engi
ne
ering
and
Techn
i
cal Innovations
,
2014
.
[20]
H.
Le
ung
and
S
.
Ha
y
k
in
,
“
The
complex
bac
k
p
ropa
gation
al
gor
it
hm
,”
IEEE
Tr
ansacti
ons
on
Signal
Proce
ss
ing,
vol.
39
,
no
.
9
,
pp
.
2101
-
2104
,
19
91
[21]
M.
D.
Buhm
ann
,
“
Radi
al ba
sis fu
nct
ions
,”
Acta
N
umerica
,
vo
l.
9
,
pp.
1
-
38
,
2000
.
[22]
L
.
Behe
r
a
,
“
Modul
1
Le
ct
ur
e
6:
Radi
al
Basis
Functi
on
Networks
,”
India
n
I
nstit
ute
of
Tech
nolog
y
,
Kanpur
,
pp.
1
-
23
.
[23]
M.
Mohammadi,
et
al.
,
“
A
har
dware
arc
hi
tect
ure
for
rad
ia
l
basis
func
ti
on
neur
al
ne
twork
cl
assifie
r
,”
IEEE
Tr
ansacti
ons on Paralle
l
and
Dis
tribut
ed
Syste
ms
,
vol
.
29
,
no
.
3
,
p
p.
481
-
495
,
201
8.
[24]
Y.
Shi,
et
al.,
“
Acti
on
Antic
ip
at
ion
with
RBF
Kerne
li
z
ed
Fe
at
ure
Mapping
RNN
,”
Europea
n
Confe
renc
e
o
n
Computer
Vi
sion
,
2018
[25]
J
.
A.
Bull
ina
ri
a
,
“
Radi
al
Basis
Functi
on
Networks:
Introduc
tion
,
”
Introductio
n
to
neur
al
netw
ork:
le
ct
ur
e
12,
pp.
1
-
12
,
2004
[26]
R.
A.
Fin
an,
et
a
l.
,
“
Com
par
ison
of
m
ult
il
a
y
e
r
an
d
rad
i
al
b
asis
fu
nct
ion
n
eur
a
l
ne
t
works
for
te
xt
-
d
epe
nden
t
spea
k
e
r
rec
ogni
ti
on
,”
Procc
ed
ings o
f
Int
ernati
onal
Confer
enc
e
on
Neural
Net
works,
v
o
l. 4,
pp.
1
99
2
-
1
99
7,
1996
.
[27]
T.
Xie,
et
al.,
“
Com
par
ison
bet
wee
n
Tra
di
ti
ona
l
Neura
l
Networ
ks
and
Radi
al
Basis
Functi
on
Networks
,”
IEEE
Inte
rnational
Sy
mpos
ium
on
Industrial
Elec
troni
cs
,
pp
.
1194
-
119
9
,
2011
.
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
An
alysis of
bac
k p
r
opag
ation
and
r
adial
basis f
unct
ion ne
ural netw
or
ks f
or
hand
over
....
(
Pay
al
Ma
haja
n)
4843
[28]
E.
A.
M.
A.
Shenouda,
“
A
quant
itati
v
e
compari
son
o
f
diffe
r
ent
MLP
ac
ti
v
a
ti
on
func
t
ions
in
cl
assificat
ion
,”
Adv
anc
es
in
Neu
ral Ne
tworks
,
pp
.
849
-
857
,
20
06
.
[29]
R.
B.
San
tos,
e
t
al
.
,
“
Com
par
ison
Bet
wee
n
Multi
lay
er
Fe
ed
forw
ard
Neura
l
Netw
orks
and
a
R
adial
Basis
Func
ti
o
n
Network
to
Detect
and
Locat
e
L
ea
ks
in
Pip
el
in
es
Tra
nsporting
Gas
,”
Chem
ic
a
l
E
ng
ine
ering
Tr
ans
act
ions,
vol.
32
,
pp.
1375
-
1380
,
2013
BIOGR
AP
H
I
ES
OF
A
UTH
ORS
Pay
a
l
Mahajan
,
Rese
arc
h
Sch
ola
r
in
D
epa
rtm
ent
of
El
e
ct
ri
cal
E
ngin
ee
ring
,
J
amia
Mill
ia
Islamia
Univ
ersity
.
Resea
r
ch
i
nt
ere
sts:
fu
zzy
log
ic
,
wire
l
ess
comm
unic
at
ion
,
m
obil
e
n
et
works
,
neur
al ne
tworks,
opti
m
izati
on
t
echnique
s.
Em
ail:
pa
y
a
l27mahaj
a
n@gm
ai
l.
com
Pr
of.
Z
aheer
uddin
,
Senior
p
rofe
ss
or
in
Depa
rtment
of
El
ec
tr
ical
E
ngin
ee
ring
,
Jam
ia
Mill
ia
Islamia
Univer
si
t
y
.
Rese
arc
h
in
t
ere
sts:
fuz
z
y
log
ic
,
soft
computi
n,
m
obil
e
net
wo
rks,
art
if
ic
i
al
Inte
lligen
ce ne
ur
al
n
et
works
,
op
tim
iz
at
ion
techniq
ues
.
Em
ai
l:
za
hee
ruddin@j
m
i
.
a
c.
in
Evaluation Warning : The document was created with Spire.PDF for Python.