TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 13, No. 2, Februa
ry 20
15, pp. 369 ~ 378
DOI: 10.115
9
1
/telkomni
ka.
v
13i2.703
2
369
Re
cei
v
ed
No
vem
ber 1
5
, 2014; Re
vi
sed
De
cem
ber 2
8
,
2014; Accep
t
ed Jan
uary 1
4
, 2015
Radial Basis Function Network Learning with Modified
Backpropagation Algorithm
Usman Mu
h
a
mmad Tuk
ur, Siti Mariy
a
m Shamsuddin*
Soft Computin
g Rese
arch Group, F
a
cult
y of
Co
mputi
ng, U
n
iversiti T
e
knol
ogi Ma
la
ysi
a
813
10 Jo
hor, Mala
ysi
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: mtus
maan@
ya
ho
o.com, mari
yam@
utm.my
A
b
st
r
a
ct
Radi
al B
a
sis F
unctio
n
N
e
tw
ork (RBF
N) is a
class of Artifi
cial N
eur
al N
e
tw
ork (ANN) that w
a
s
used
in
many
classificati
on
p
r
obl
ems
in sc
i
ence
an
d e
ngi
neer
ing. B
a
ckp
ropa
gatio
n (BP
)
alg
o
rith
m is
a
lear
nin
g
a
l
gor
ithm t
hat w
a
s w
i
dely
used
in
A
NN. Ho
w
e
ver,
BP
has
ma
jor
disa
dvant
ages
of slow
e
r
r
o
r
r
a
t
e
co
nver
ge
nce
a
n
d a
l
w
a
ys
easi
l
y
stuck
at
the local
m
i
n
i
m
a
.
H
e
n
c
e
,
Modifie
d
BP algorit
h
m
w
a
s propos
ed i
n
this study to
i
m
pr
ove th
e l
e
a
r
nin
g
sp
eed
of
RBF
N
usi
ng
d
i
screti
z
e
d
d
a
ta. C pro
g
ra
mmi
n
g la
ng
uag
e w
a
s
used t
o
d
e
vel
o
p the
pro
g
ra
m for the
prop
o
s
ed
met
hod.
P
e
rformanc
e
measur
e
m
ent of
the meth
od
w
a
s
conducted and
th
e
experimental r
e
s
u
l
t
s
indicate tha
t
our proposed method perform
s better in error ra
te
conver
genc
e
a
nd corr
ect clas
sificatio
n
co
m
p
ared t
o
the
r
e
s
u
lt w
i
th conti
n
uous
datas
et. T
-
test statistical
ana
lysis w
a
s
used
to ch
eck
the si
gn
ifican
ce of th
e
res
u
lts an
d
most
w
e
re foun
d to
be s
a
tisfactor
ily
signific
ant.
Ke
y
w
ords
: ,od
i
fied b
a
ckpro
pa
gatio
n, cost function, di
screti
z
ation, rad
i
al
ba
sis function n
e
t
w
ork
Copy
right
©
2015 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
Artificial Ne
ural Network (A
NN) wa
s dev
elope
d as a p
a
rallel
di
strib
u
ted system i
n
spi
r
ed
by human b
r
ain learning
pro
c
e
ss. ANN lea
r
ning
capa
city to so
lve proble
m
s throug
h trai
ning
make
s it po
pular. Th
ere
are ma
ny types of A
NN su
ch a
s
Ba
ckpro
pag
atio
n Network,
Self-
Orga
nizi
ng M
ap (SO
M
), S
p
ikin
g N
eural
Network, Ra
dial Basi
s F
u
nction Net
w
o
r
k (RBF
N), etc.
RBFN i
s
a cl
ass of ANN t
hat
employs
Radi
al Basi
s Functio
n
s
(RBFs) a
s
a
c
tivation functio
n
s
.
RBFN is a
un
ique type
of
ANN with
onl
y three
layers. It has only
one
hidd
en l
a
yer unli
k
e
oth
e
r
types of ANN that have one or mo
re
hidde
n laye
rs. It is a feed forward and
fully connected
netwo
rk [1]. RBFN
ha
s
several
benefit
s ab
ove it
s p
r
ede
ce
sso
r
s. Some
of
the benefits
i
n
cl
u
d
e
having
simpl
e
network a
r
chite
c
ture, a
b
ility to appro
x
imate better and al
so it
s
algorith
m
s l
e
arn
faster.
RBF
N
s
are u
s
e
d
to solve
pro
b
lem
s
su
ch
a
s
cla
ssif
i
cat
i
on pr
oblem
s,
f
u
n
c
t
i
on
approximatio
n, system
co
ntrol, an
d
tim
e
serie
s
pre
d
i
c
tion. As a
re
sult, RBF
N
it
is
widely u
s
e
d
in
sci
en
ce and
engin
eeri
ng. In this pape
r, we ar
e a
ppl
ying the Mod
i
fied Backpro
pagatio
n (MB
P
)
algorith
m
[2]
with imp
r
ove
d
lea
r
nin
g
pa
ramete
r va
lu
e to trai
n RB
FN
with di
scretized
data. T
hus,
this study att
e
mpts to exp
l
ore the p
e
rf
orma
nce
of RBFN by d
e
termini
ng valu
es for e
r
ror
rate
conve
r
ge
nce
or lea
r
ni
ng ra
te and
co
rre
ct classifica
tio
n
accu
ra
cy of
the network.
The rest of t
h
e
pape
r is o
r
ga
nize
d into se
ction
s
. Sectio
n 2 is
the Propo
sed M
e
th
od; se
ction 3
,
The Re
sea
r
ch
Method. In
section
4, Results a
nd
Discussion
of
the
finding
s a
r
e
given. The
Concl
u
si
on of t
h
e
study is given
in sectio
n 5, and finally the Re
feren
c
e
s
con
s
ulted a
r
e listed in section 6.
2. Proposed
M
e
thod
In RBFN trai
ning, clu
s
te
ri
ng algo
rithm
s
are u
s
ed to
determi
ne th
e cent
re an
d weig
ht of
the hidden la
yer. On the other ha
nd, L
east Mea
n
Squares
(LMS
) algo
rithms
are em
ploye
d
in
determi
ning t
he network
weig
hts bet
ween hid
den
and outp
u
t layer [3]. Wei
ghts of net
work
neuron
s thou
can
be a
d
ju
sted have
a val
ue of 1
betwe
en inp
u
t layers an
d hid
den
layers [1]. Th
e
hidde
n layer
node
s dete
r
mine beh
aviour an
d stru
cture
of the netwo
rk. G
a
ussian fun
c
ti
on is
us
ed to ac
tivate the hidden layer [4].
The num
erous synaptic
associations
that link the billi
ons
of neurons in animals are
modified to
u
nderstan
d ne
w thing
s
. Li
kewi
se, an A
N
N be
have
s
in
the sa
me
way by mimicking
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 2, Februa
ry 2015 : 369 – 378
370
the neuron
s in human b
r
ai
n. It modi
fies its weights t
o
enable it l
earn ne
w pattern
s. Theref
ore,
ANN is a
ma
chin
e le
arni
n
g
p
r
o
c
e
s
s in
spired,
by o
u
r
brain.
Every
nod
e of
the
artificial
neu
ron
has an
a
c
tivation fun
c
tion
whi
c
h i
s
i
n
charg
e
of map
p
ing th
e in
pu
ts of th
e n
euron to
its
outp
u
t
[5]. Backp
rop
agation
(BP) algo
rithm h
a
s b
een
wid
e
ly use
d
a
s
training
algo
rithm for ANN in
sup
e
rvised l
e
arnin
g
m
ode
[6]. A study
by [7] rep
o
rte
d
that at
pre
s
ent
ANN h
a
ve num
erou
s real
life applications in ma
chin
e learni
ng an
d other field
s
.
RBFN i
s
train
ed usi
ng un
supervi
sed
an
d sup
e
rvi
s
ed
learni
ng mo
d
e
s. The u
n
su
pervised
mode i
s
impl
emented
bet
wee
n
input to
hidden l
a
ye
rs an
d su
pe
rvised m
ode i
s
i
m
pleme
n
ted f
r
om
hidde
n to out
put layer. Fu
nction
s give
n
by hidde
n ne
uron
s
create
rand
om
starti
ng poi
nt for in
put
pattern
s [8].
Clu
s
terin
g
alg
o
rithm
s
are capabl
e of
finding clu
s
ter ce
ntres that be
st repre
s
ent
s the
distrib
u
tion of
data in the
first
stage
of the trai
ni
ng. S
upervi
sed l
e
a
r
ning i
s
guide
d by the de
si
red
target value
s
for the net
wo
rk. Usin
g mo
dified BP to train the net
wo
rk
with modifi
ed cost fun
c
ti
on
proved by ot
her stu
d
ie
s to have enha
nce
d
ANN le
arnin
g
is imp
l
emented [2] but in our ca
se,
discreti
zed d
a
taset
s
wa
s use
d
whi
c
h al
so wa
s re
po
rted to enhan
ce the spee
d of learnin
g
[9]. In
RBFN, e
a
ch l
a
yer of the
ne
twork p
e
rfo
r
m
s
a
diffe
rent t
a
sks
and th
at is on
e of the
main p
r
obl
em
s
with this
network
.
Therefore, s
e
parating
the processe
s of th
e laye
rs
with vari
ou
s tech
niqu
es i
s
a
good
idea. I
n
this
pap
er,
five stand
ard cl
assi
ficati
on p
r
obl
ems data
s
et was used to te
st
the
prop
osed
alg
o
rithm. T
he
study co
mpa
r
ed the
resu
lt
of the
pro
p
o
s
ed al
gorith
m
: RBF
N
l
earni
ng
with Modified
BP algorithm
(MBP-RB
FN) usi
ng
contin
uou
s and di
screti
zed d
a
ta
sets.
2.1. Backpr
opaga
tion Al
gorithm
Backpropa
ga
tion Algorith
m
(BP) take
s the lead i
n
training A
NN. It is on
e of th
e
sup
e
rvised le
arnin
g
alg
o
rit
h
ms that u
s
e
gradi
ent de
scent te
chni
qu
e to red
u
ce the cost fun
c
ti
on.
Many facto
r
s affect the pe
rforma
nce of
BP algorit
hm,
su
ch a
s
initi
a
l param
eters, initial wei
g
ht,
rate of lea
r
ni
ng, mome
ntu
m
term, si
ze
of net
wo
rk, n
u
mbe
r
of ep
o
c
h, conn
ectio
n
between th
e
units. Th
e le
arnin
g
pe
rformance of BP
depe
nd
s on
n
e
twork
param
eters. A g
ood
choi
ce
of the
s
e
para
m
eters
may greatly improve p
e
rfo
r
man
c
e. It ge
nerally ta
ke
s BP a long time to train [10].
Other BP lea
r
ning
we
akne
sses a
r
e lo
cal minima tr
a
pping a
nd
sl
ow rate of co
nverge
nce [1
1],
[12]. As a result of the
s
e limitations,
rese
a
r
chers have bee
n
workin
g ha
rd
to improve
BP
perfo
rman
ce.
The BP traini
ng pro
c
e
s
s st
arts at the inp
u
t layer (
i
) through the hid
d
en layer (
j
) to output
layer (
k
)
as
explained
in
equatio
ns bel
ow. T
he
net
work a
c
tivation fun
c
tion
a
pplied
is si
g
m
oid
function.
Between laye
r
(i)
an
d layer
(j)
(
1
)
∑
(
2
)
Whe
r
e “
is ou
tput of layer
j,
is output of l
a
yer
i,
is wei
ght co
nne
ctin
g layer
i
to
j
an
d
is
the bias at la
yer
j”
Between laye
r (
j
) an
d layer (
k
)
(
3
)
∑
(
4
)
Whe
r
e “
is ou
tput of layer
k,
is output
of layer
j,
is weig
ht conn
ecting l
a
yer
j to
k
,
a
nd
is the bias at
layer
k”
Between laye
r (
j
) an
d (
k
)
The e
rro
r i
s
comp
uted u
s
ing (5
) to d
e
t
ermine th
e
differen
c
e
s
o
f
desired o
u
tput and
actual o
u
tput. Error i
s
the
n
pro
pag
ated
backward. The propa
gat
i
on sta
r
ts at la
yer
(k
)
to
(j
)
then
finally to layer
(i).
Th
e wei
ghts a
r
e adj
u
s
ted to de
cre
a
se the
error
as it is ba
ckward p
r
op
agat
ed
throug
h the la
yers.
∑
–
(
5
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Radi
al Basi
s Functio
n
Net
w
ork Le
arnin
g
with Modifie
d
… (Usm
an
Muham
m
ad Tuku
r)
371
From the e
r
ro
r com
puted,
BP is used from (
k
) to (
j
) a
s
in Equation
(6) a
nd (7
).
1
∆
1
(
6
)
∆
1
∆
(
7
)
With,
1
(
8
)
Whe
r
e “
rep
r
ese
n
t the we
ight from layer
k
to layer
j
at time
t,
∆
repre
s
e
n
t the
weig
ht chan
g
e
,
rep
r
e
s
ent the learnin
g
rate,
repre
s
e
n
t momentu
m
rate,
repre
s
ent erro
r at
node
k,
rep
r
e
s
ent the a
c
tu
al netwo
rk
ou
tput at layer
j,
repre
s
e
n
t the actual n
e
twork
output
at layer
k,
rep
r
esent the target output value at layer
k.”
Equation
s
(9
) and (10
)
are for backwa
r
d
cal
c
ulatio
ns from layer (
j
) to (
i
).
1
∆
1
(
9
)
∆
1
∆
(
1
0
)
With,
1
∑
(
1
1
)
(
1
2
)
∑
(
1
3
)
Whe
r
e “
is th
e weig
ht from
j
to
i
at time
t,
∆
is the
weight adj
u
s
tment,
is the
learni
ng rate,
is
the momentum rate,
is error
at no
de
j,
is e
r
ror
at node
k,
is the
actual
netwo
rk o
u
tp
ut at node
i,
is the actu
al
network out
put at node
j,
is the actual netwo
rk
output at nod
e
k,
is the wei
ght con
n
e
c
te
d betwe
en no
de
j
and
k,
is the bias of no
de
k.”
This p
r
o
c
ed
u
r
e is
reite
r
ate
d
till converg
enc
e is attai
ned. Du
rin
g
trainin
g
, stan
dard BP
norm
a
lly use
s
lea
r
ning
rat
e
whi
c
h is th
e amount of
corre
c
tion a
p
p
lied to adju
s
t weights at t
h
e
time of training and mome
n
t
um term. It is
used to set the rate of adj
ustment an
d is always withi
n
the range of [0, 1]. The small value of learni
ng ra
te will make the
changes of
weight very sm
all.
Momentum
t
e
rm i
s
th
e a
m
ount of
pre
v
ious
co
rrecti
ve term th
at
sho
u
ld
be
re
membe
r
ed.
It is
use
d
to spe
e
d
up the prop
agation, an
d is within
[0.1,
1.0] rang
e. Hen
c
e, the
s
e
two paramet
ers
are u
s
ed g
e
n
e
rally to moni
tor the adju
s
tment of the weight [6] and [13].
Sensitivity of
BP to the learning
rate val
ue ma
ke
s it
runs v
e
ry
slo
w
.
That
is
wh
y
B
P
is
alway
s
lo
cke
d in local mi
nima. But, optimum conv
erge
nce spe
ed of BP ca
n be o
b
taine
d
by
tuning the le
arnin
g
pa
ram
e
ter. A study
by [14]
prop
ose
d
an a
d
a
p
tive learni
ng
rate ba
se
d
on
Barzil
ai and
Borwein lea
r
ning rate upd
ate for ste
e
p
e
st de
scent
method, an
d
the re
sult sh
owe
d
that the p
r
op
ose
d
meth
od
improve
s
th
e
conve
r
ge
nc
e
rates of BP
algorithm.
The
r
efore, from
th
e
previou
s
stu
d
ies, it sh
ows that
in BP learni
ng, it is e
s
sential to pick the le
arnin
g
con
s
tant
cor
r
e
c
t
l
y
.
2.2. Cos
t
Func
tio
n
The erro
r cal
c
ulatio
ns em
ployed to trai
n an AN
N are signifi
cant [15]. Also, [16] report
e
d
that output error m
e
a
s
u
r
e
m
ent gives
u
s
a
way to up
date the
weig
hts iteratively
to redu
ce
error.
The
co
st fun
c
tion i
s
inte
rcha
nge
ably
calle
d obj
ecti
ve or
error functio
n
. Mea
n
Squa
re E
r
ror
(MSE) is
wid
e
ly used in
weight adju
s
tm
ent in BP training.
∑
(
1
4
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 2, Februa
ry 2015 : 369 – 378
372
Whe
r
e,
Desi
red r
e
p
re
se
nt target value
and
Ac
tual
re
pre
s
ent
real
value gen
era
t
ed by netwo
rk.
The “e
rror
sig
nal”
is define
d
as:
1
(
1
5
)
Whe
r
e,
T
k
re
pre
s
ent
targe
t
output and
O
k
rep
r
e
s
ent
the actual o
u
tput of the network.
Ho
wever, in t
h
is
study, mo
dified cost fu
nction
(MM
)
i
n
[2] defined
in (16
)
is
use
d
with
MBP-RBF
N
in place of the commo
nly used MSE.
∑
(
1
6
)
With,
(
1
7
)
Whe
r
e “
, and,
is the error
at
(k
)
,
is the t
a
rget valu
e at
(k
),
is the a
c
tivation of
(k
)
”
Applying pa
rtial derivative
s
u
s
ing
chai
n rule
, for
weight upd
atin
g, an erro
r signal is
gene
rated fo
r MBP for the
output layer a
s
,
(
1
8
)
While MBP error
s
i
gnal for hidden layer is
s
a
me as
SBP,
∑
′
(
1
9
)
Whe
r
e “
is the weig
ht of conne
ction,
a sigmoi
d functi
on of (1/1
+e
– 2
x
)”.
2.3.
Radial Basis
Function
Ne
t
w
o
r
k
RBFs a
r
e u
s
ed as a
c
tivation function
s i
n
the hidden
layer of RBF
N
. The link betwe
en
inputs to hi
dd
en layer i
s
no
t weighted
while the lin
k b
e
twee
n hidd
e
n
to output la
yer is
weig
hted.
The hidd
en la
yer neu
ron
s
p
r
ovide
func
tions
set to makeup ra
ndo
m
basi
s
for input patterns
[17]
.
Figure 1. The
RBF Netw
ork Structu
r
e [1
8]
The output y
j
(
i
) of the
j
th
hidden layer in a
RBFN mo
del
, is as [4]:
Ф
,
,
(
2
0
)
1
,
2
,
…..,
1
,
2
,
…
,
Her
e
k
represents th
e RBF
s
a
pplied,
whi
l
e
C
k
∈
R
m
,
σ
k
∈
R
m
are
ce
ntre a
nd
widt
h value ve
cto
r
s
of RBF, given as:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Radi
al Basi
s Functio
n
Net
w
ork Le
arnin
g
with Modifie
d
… (Usm
an
Muham
m
ad Tuku
r)
373
…
∈
,
,….
(21
)
…
∈
,
,
…
,
,
(22
)
Whi
c
h are we
ights of RBF
s
linked to
jth
output.
RBFN impl
e
m
entation is repre
s
e
n
ted in
two
layered
netwo
rk form. Non-line
a
r transfo
rm
is implante
d
by the input layer. Wh
ile i
n
the hidden
layer, each
term
Φ
k
(.) is an a
c
tivation
function. The
output layer o
n
the ot
her h
and ap
plies li
near tran
sformation.
Φ
x,
c
,
σ
∐
ϕ
x
,c
,
σ
(
2
3
)
Also, the most common o
p
tion for
(.) is t
he Gau
s
sian
form belo
w
in
Equation (1
9
)
.
ϕ
x
,c
,
σ
e
x
p
x
c
/
2
σ
(
2
4
)
Equation (20) output beco
m
es (25)
∑
e
xp
∑
/2
(
2
5
)
A study by [1
9] pro
p
o
s
ed
a ne
w tre
e
b
a
se
d
RBF
N
model com
b
i
n
ing
lo
gisti
c
regre
s
sion,
aimed to
enh
ance its
cla
s
sification p
e
rfo
r
man
c
e
by
in
put data p
r
e
p
r
ocessin
g
u
s
i
ng RBF
N
fra
m
e.
The p
r
op
ose
d
metho
d
wa
s bi
nary
cla
ssification a
nd
wa
s ea
sy to
gene
rali
ze fo
r many p
r
obl
e
m
s.
The mod
e
l was teste
d
wit
h
data obtai
n
ed from “hydraulic fractu
rin
g
in Oil and
Gas
well
s”, a
nd
the results a
r
e sup
e
rio
r
to logisti
c
reg
r
e
s
sion.
Data fro
m
1
200 individ
u
a
ls
was
coll
ected
by [20], for empiri
cal com
pari
s
o
n
of the
netwo
rk mod
e
ls. Fa
sting
plasm
a
glu
c
o
s
e
(FPG
) a
s
crite
r
ia fo
r cl
a
ssifying
a p
a
tient as a di
ab
etic
wa
s u
s
e
d
in
sele
cting
dia
betic
patient
s. Ri
sk
pa
ram
e
ters em
ploy
ed a
r
e
gen
de
r, ag
e a
nd fa
mily
history of di
a
betes; a
nd
many othe
r factors a
r
e al
so
con
s
ide
r
e
d
. The stu
d
y use
d
RBF
N
in
diagn
osin
g o
f
diabetes m
e
llitus an
d the re
sult
s
compa
r
ed
with MLP Network a
nd logi
stic
reg
r
e
ssi
on. T
he results
obt
ained
proved
that RBF
N
ha
s a
better
pe
rforman
c
e
co
mpared to
other
model
s.
Also, [21] applied brea
st can
c
e
r
and g
ene to RBFN with Gro
w
in
g and Pruni
n
g
(GAP)
algorith
m
. [22] use
d
RBF
N
for tree
s sp
ecie
s re
cogni
tion and
cla
s
sificatio
n
for f
o
re
st invento
r
ies
to estimate the wood q
uant
ity in a given
forest
a
r
ea. T
r
ee
s diamete
r
, thickn
ess b
a
rk, g
r
owth o
f
diamete
r
an
d
height a
r
e th
e pa
ramete
rs use
d
. Tree
clusteri
ng
with
diverse in
put
wa
s ta
ken by
the netwo
rk. The re
sult
s improve
d
foreca
sting
te
ch
nique
s in forest invento
r
ie
s. [23] applie
d
RBF
Network for time
s
e
ries
fo
re
ca
stin
g in
a
hybrid
syste
m
com
p
risi
ng
of ma
ny su
b-RBF
N
s
whi
c
h produ
ced a faste
r
training
spee
d
with
slig
htly better gene
rali
zation
cap
abil
i
ty.
2.4. Discre
t
izatio
n
Discretizatio
n
acco
rdin
g to
[24] ca
n be
d
e
fined a
s
one
way of
red
u
cing data
o
r
ch
angin
g
origin
al co
ntinuou
s attrib
ute into discrete attri
bute. Di
scretization p
r
oces
s h
e
lp
s in data re
du
ctio
n
and si
mplification, as a
re
sult, the data
beco
m
e
s
ea
sier to u
nde
rstand, u
s
ed
and explai
ne
d. In
addition a
c
hi
eving, faster
and a
c
curate
training p
r
o
c
ess [25, 26]. Many studie
s
on discretiza
tion
algorith
m
s
were te
sted in
Data minin
g
and kno
w
led
ge discove
r
y domain, to sh
ow that it has the
potential
to redu
ce data and
im
provin
g
predi
ctiv
e
accuracy. T
h
e alg
o
rithm
s
are
“sup
ervi
sed
versu
s
un
su
p
e
rvise
d
, glob
al versu
s
local and dyna
mic versus
static [27]”.
Also, [28], explained
that based o
n
differe
nt theoreti
c
al o
r
igi
n
s, vari
ou
s
method
s for
discreti
zation
in literatu
r
e
are
identified
su
ch a
s
statistics, Boolean
rea
s
oni
ng,
clu
s
tering
techniq
ues an
d e
n
tropy calculatio
ns.
Equal Frequ
e
n
cy Binnin
g
(EFB) Equal
Width Bi
nni
n
g
(EWB
) and
Entropy-b
ased Di
screti
za
tion
(EBD)
are th
e most g
ene
ral on
es. In t
h
is p
ape
r, M
D
L di
screti
za
tion algo
rith
m in Ro
ug
h
Se
t
Toolk
i
t
f
o
r
D
a
ta
Anal
y
s
is
i
s
us
ed to disc
retiz
e
all input datas
et
s
.
2.5. Classific
a
tio
n
Problem
RBFN have
been
ap
plied
effectively i
n
ma
ny
field
s
fo
r dive
rse
pu
rpo
s
e
s
su
ch
as in
cla
ssifi
cation.
Cla
ssifi
catio
n
is
appli
ed i
n
virt
ually all
circu
m
stan
ces
whe
r
e
co
rrelation
bet
ween
inputs an
d o
u
tputs
exist. It can
comp
ute several
po
ssible
re
sult
s
across vario
u
s
p
a
ra
meters, if
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 2, Februa
ry 2015 : 369 – 378
374
we have
eno
ugh info
rmati
on from the
p
a
st [16]. Al
so,
[29] in a stu
d
y prop
osed
Enhan
cem
e
n
t
o
f
RBFN
with F
u
zzy-OS
D al
gorithm fo
r o
n
line sy
stem
that classify rada
r pul
se.
It needs qui
ck
netwo
rk retra
i
ning
when n
e
w
un
se
en
e
m
itters
a
r
e d
e
tected. Th
e
netwo
rk
performan
ce
wa
s
sup
e
rio
r
to T
h
ree
-
Pha
s
e
OSD m
e
thod
whe
n
it was evaluated
a
nd comp
are
d
. Also, adju
s
t
i
ng
RBF units
wit
h
great
care in t
he
network will i
m
prove
per
formance
in less tr
ai
ning cy
cles, whi
c
h
is v
i
tal for rea
l
-time sy
ste
m
s.
2.6.
Net
w
o
r
k
Arc
h
itec
ture for each Datase
t
In this stu
d
y, Modified B
P
algorithm
wa
s ap
plied
to enha
nce
RBFN l
e
a
r
n
i
ng with
discreti
zed d
a
taset. The result
s of MBP-RBF
N
with
continuo
us
and discretized datasets
were
comp
ared to ascertai
n the perfo
rman
ce
of our
pro
p
o
s
ed method. Defining netwo
rk a
r
chitectu
re
mean
s
cho
o
sing
suitable
numbe
r of l
a
yers
and
no
d
e
s fo
r e
a
ch
dataset. Tabl
e 1 d
epi
cts t
h
e
netwo
rk a
r
chi
t
ecture fo
r this study. Opti
mal netwo
rk desi
gn come
s from carefu
l sele
ction of the
numbe
r of hi
dden n
e
u
r
on
s. Gen
e
rali
za
tion cap
a
city,
training time
and complex
i
ty is affected
by
the si
ze
of th
e hid
den
laye
r. Tho
ugh,
th
ere
is
no
sta
ndard
rule f
o
r
d
e
t
e
r
m
i
n
i
n
g
optim
al
num
ber
o
f
h
i
d
d
e
n
n
o
d
e
s
[
3
0
]
.
Man
y
techn
i
q
ues
h
a
v
e
b
een p
u
t
forw
ard
b
y
rese
archers
on
h
o
w
bes
t,
h
i
d
d
e
n
nod
es can be d
e
termined.
Acco
rdi
ng to
[31] the perf
o
rma
n
ce of a
n
ANN i
s
strongly influen
ced
by the n
e
tworks’
structure and its paramete
rs. If the net
work structure i
s
too
simple,
it will lower the classification
cap
a
city of the netwo
rk, he
nce affe
cting
the final
re
sul
t. Howeve
r, if the netwo
rk
stru
cture is to
o
compl
e
x it will lower the l
earni
ng spee
d and ultimat
e
ly the final
error may increa
se. Also the
momentum,
l
earni
ng rate para
m
eters will
al
so a
ffect net
work l
e
arnin
g
spe
e
d
and
accu
ra
cy. If
suitabl
e
network stru
ctu
r
e and
pa
ram
e
ters are
u
s
e
d
it will get the best result. The bre
a
kdo
w
n of
the a
r
chitectu
ral
co
nfiguration fo
r e
a
ch d
a
taset i
s
give
n in
Tabl
e 1,
with th
ree
lea
r
ning
rates of
0.5, 0.1 and
0.7 re
spe
c
tively. A consta
nt moment
u
m
of 0.9 is maintaine
d
. The minim
u
m error
rate is
set to
0.005. In the experim
ent
s k-fo
ld d
a
ta
partition me
thod was u
s
ed. For
each
k
experim
ents,
k-1
fold
was
employed
for network l
earning
whil
e
what is left is f
o
r te
sting. T
h
e
value of
k=5.
The
stoppi
n
g
co
ndition
s f
o
r e
a
ch ru
n f
o
r all th
e dat
aset
s a
r
e
rea
c
hin
g
a mi
ni
mum
error of 0.005
or maximum
of 10000 itera
t
ions re
ached
.
Table 1. Data
set su
mma
ry and archite
c
t
u
ral configu
r
a
t
ion
XOR
Balloon
Iris
Cancer
Ionosphere
Input
3
4
4
9
34
Hidden
2
2
3
3
2
Output
1
1
3
1
1
Instances 8
16
150
699
351
2.7. Modified
BP
-based
RBF
Net
w
o
r
k T
r
ai
ning Algorithm
The algo
rithm
is here
b
y pre
s
ente
d
in det
a
il, and part o
f
it is adopted
from [32].
Algorithm
1. Initialize
net
work
2.
Forward pa
ss by inserting i
nput and
d
e
si
red outp
u
t, and com
pute n
e
twork outp
u
ts
layer by layer.
3.
B
a
ck
wa
rd
p
a
ss
,
,
4. Paramete
rs
u
pdate
1
∆
1
(
2
1
)
∆
1
(
2
2
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Radi
al Basi
s Functio
n
Net
w
ork Le
arnin
g
with Modifie
d
… (Usm
an
Muham
m
ad Tuku
r)
375
With,
1
(
2
3
)
e
x
p
(
2
4
)
Whe
r
e “
is th
e wei
ght fro
m
k
to
j
at
time
t,
∆
is the
wei
ght ch
ang
e,
is the l
earning
rate,
is e
r
ror
the at
k,
is th
e actu
al network
output at
j,
is the actu
al netwo
rk ou
tput at
k,
is the target o
u
tput value at
k”
.
1
∆
1
(
2
5
)
∆
1
(
2
6
)
Whe
r
e “
is the centre from
j
to
i
at
time
t,
∆
is the centre chan
ge,
is the learni
ng rate,
is the erro
r at
k,
is the actual network output at
j,
is the width at
j,
is the wei
ght
con
n
e
c
ted be
tween
j
and
k
and
is the input
j
to
i
.”
1
∆
1
(
2
7
)
∆
1
(
2
8
)
Whe
r
e “
is the
width of
j
at ti
me
t,
∆
is the wi
dth cha
nge,
is the learning
rate
is er
ro
r
at
k,
is the actual netwo
rk
output at
j,
is
the width at
j,
is the weig
ht con
n
e
c
ted be
tween
j
and
k,
is the
input at
i
and whe
r
e
,
,
are lea
r
nin
g
ra
te factor
s
in the range [0; 1].”
5.
Rep
eat the al
gorithm for all
training inp
u
ts.
Figure 2. Modified BP-ba
sed RB
F Network T
r
aini
ng
Flowcha
r
t
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 2, Februa
ry 2015 : 369 – 378
376
3. Resul
t
s
and
Discus
s
ion
The
expe
rim
ents are in two
pa
rt
s.
Pa
rt 1, is experim
e
n
t
fo
r
RBF
N
trainin
g
an
d testing
w
i
t
h
M
B
P
a
l
g
o
r
i
t
h
m
using
continuous d
a
taset.
Part 2
,
is
an e
x
periment
for R
B
F
N
t
r
a
i
n
i
n
g
and
te
s
t
in
g
w
i
th
MBP algorith
m
usin
g di
scretized
dat
a
s
e
t. The
classifi
cation
accu
ra
cy and e
r
ror
rate
on
t
e
st
re
sul
t
s
a
r
e
com
p
ared.
Re
sult
summ
ar
y fo
r MBP-RBFN with both
continuo
us
an
d
discreti
zed d
a
taset
s
are gi
ven in the table and figures below.
Table 2. Re
sult summa
ry for MBP-RBF
N
with differe
nt types of dataset
Dataset
MBP-RBFN w
i
th continuous
dataset
MBP-RBFN w
i
th discretized
dataset
Classifi
cation
Ac
c
u
rac
y
(%
)
Error rate
Classifi
cation
Accur
a
cy
(%)
Error
rate
XOR
56.25
0.0876
59.90
0.0145
Balloon 54.55
0.0419
51.31
0.0374
Iris 82.95
0.4411
83.78
0.5276
Cancer
88.93
0.1081
86.86
0.0343
Ionosphere
59.17
0.1339
78.27
0.1046
Figure 3.
Cla
ssif
c
ation
Re
sult for MBP-
RBFN
From
Table
2 an
d Fig
u
re
3, the
resu
lts com
p
arison
s
b
e
tween contin
u
ous and
discreti
zed d
a
taset
s
for MBP-RBF
N
i
ndicate
that MBP-RBF
N
with discretized data
s
ets
has
better cla
s
sification results in XOR, Iris and
Inonsph
ere while on
the other ha
nd; MBP-RB
FN
with co
ntinuo
us data
s
et
s h
a
s bette
r re
su
lts in Ballon a
nd Ca
ncer.
Also, MBP-RBFN with di
screti
zed
data
s
ets
has
th
e least
conve
r
g
ence erro
r ra
tes in all
the five datasets ex
cept
in Iris, wh
ere MBP-
RBF
N
with conti
nuou
s data
s
et has the l
east
conve
r
ge
nce
error
rate
as
sho
w
n i
n
Ta
b
l
e 2 an
d Fig
u
e
4. Thi
s
p
r
ov
es the
cl
aims of [9] and ot
her
resea
r
chers
who
said di
scretization i
m
prove
s
th
e
cla
s
sificatio
n
pe
rform
a
n
c
e
and
fast
er
conve
r
ge
nce
of error
rate
s. The results a
bove ar
e ba
cked
by the st
atistical te
st condu
cted
whi
c
h
sho
w
e
d
that the t-test for a
ll the results
wa
s st
atisti
ca
lly significant
except for Bo
lloon whi
c
h
was
statistically not significant. We ca
n su
mmari
se
he
re that MBP-RBFN
with discretized dat
aset
has b
e
t
t
e
r re
s
u
lt
s in most
o
f
t
he case
s.
0
10
20
30
40
50
60
70
80
90
100
X
o
r
B
alloon
Iris
Cancer
Ionosphere
MBP-RBFN w
i
th
continuous
dataset
MBP-RBFN w
i
th
discretized
dataset
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Radi
al Basi
s Functio
n
Net
w
ork Le
arnin
g
with Modifie
d
… (Usm
an
Muham
m
ad Tuku
r)
377
Figur
e 4. Erro
r rate for MB
P-RBF
N
wi
th
contin
uou
s a
nd discretize
d dataset
The o
b
je
ct
iv
e of
t
h
is
st
u
d
y
wa
s t
o
t
e
st
if
traini
ng
of RBF
N
wit
h
modifie
d
B
P
usin
g
discreti
zed d
a
taset woul
d
enha
nce
th
e
cla
s
sificati
on
a
c
cura
cy
and small error rate
s o
f
conve
r
ge
nce.
4. Conclusio
n
In this study
, MBP algorithm was u
s
ed to train RBFN
with both co
ntinu
ous a
nd
discreti
zed d
a
taset
s
. For
uniform
com
pari
s
on in thi
s
study, sa
m
e
para
m
eters were
used i
n
all
experim
ents for the five dataset
s. F
i
n
a
lly,
inve
stigation is don
e by compari
ng
results of MBP-
RBFN
u
s
ing
both contin
uo
us and
di
scre
tized dat
a
s
et and
fo
r ea
ch dataset
an
d
t
he
o
u
tco
m
e
i
s
an improved
result for MBP-RBF
N
with
discreti
z
ed datas
e
t which is
better than MBP-RBFN with
contin
uou
s d
a
taset which also ju
stified the study carri
ed out.
Ackn
o
w
l
e
dg
ements
Authors
wou
l
d like to th
ank Soft Co
mputi
ng Re
search Group
(SCRG
)
Un
iversit
i
Tekn
ologi Ma
laysia and
UTM Big Data
Centre for
suppo
rting this study fr
om ince
ption to the
end of the stu
d
y.
Referen
ces
[1]
Che
n
Y. App
l
i
c
ation
of a N
o
n-lin
er C
l
a
ssifi
er Mod
e
l
base
d
on
SVR-RB
F
Algorithm. i
n
Power a
n
d
Energy Engineering
Confer
ence (APPEEC), 2010 Asia-P
acific
. IEEE. 2010
.
[2]
Shamsu
ddi
n S
M
, MN Sul
a
im
an, M D
a
rus.
A
n
i
m
pr
ove
d
err
o
r sig
n
a
l
for th
e b
a
ckpro
pa
ga
tion
mode
l f
o
r
classificati
on p
r
obl
ems.
Intern
ation
a
l Jo
urna
l of Computer
M
a
thematics. 20
01; 76(3): 2
97-
305.
[3]
Yu D, J Gomm, D Williams,
A recursive orth
ogo
nal l
east sq
uares a
l
gor
ith
m
for traini
ng
RBF
netw
o
rks.
Neur
al Proces
sing L
e
tters. 1997;
5
(3): 1
67-
176.
[4]
Qasem SN, SM
Shamsud
d
i
n
.
Hybrid L
ear
nin
g
Enha
nce
m
e
n
t of RBF
Netw
ork Base
d on Particl
e
Swarm
Optimi
z
a
tion
.
Adva
nce
s
in Neur
al Net
w
orks–ISNN 2
009
. Spri
ng
er. 200
9; 19-2
9
.
[5] Chang
WY.
Es
timati
on
of th
e
state of c
harg
e
for a
LF
P
bat
t
e
ry us
ing
a
hy
brid
metho
d
th
at co
mb
in
es
a
RBF
neur
al
ne
tw
ork, an OLS alg
o
rith
m
an
d
AGA.
Internati
ona
l Jo
urna
l of
Electrica
l
Po
wer & En
erg
y
S
y
stems. 20
13
.
53
: 603-6
11.
[6]
Z
w
eiri YH, et
al.
A new
thre
e-term
back
p
r
opa
gati
on a
l
g
o
r
ithm w
i
th co
n
v
erge
nce a
n
a
l
ysis
.
Robotics
and Auto
matio
n
. Proceed
in
gs
. ICRA'
02.
IEEE Internatio
nal
Confer
ence
on
. 2002. IEEE.
[7]
Shahpaz
ov VL, VB Velev, LA Doukovska.
Desig
n
and
a
p
p
licati
o
n
of Art
i
ficial
Ne
ura
l
N
e
tw
orks for
pred
icting th
e
valu
es of in
dex
es on th
e Bul
g
aria
n Stock ma
rket
.
Signal Pr
ocessing Sy
m
p
osium
(SPS),
201
3
. IEEE. 2013.
[8]
Qu M, et al.
A
u
tomatic so
lar
flare d
e
tectio
n
usin
g MLP, RB
F
,
and SVM.
S
o
lar P
h
ysics. 2
003;
21
7
(1
)
:
157-
172.
[9]
Len
g W
Y
, SM
Shams
udd
in.
W
r
iter id
entifi
c
ation
for C
h
i
nese
ha
ndw
riti
ng.
Int. J. Advance. Soft
Comp
ut. Appl. 201
0; 2(2): 142
-173.
0
0.
1
0.
2
0.
3
0.
4
0.
5
0.
6
XO
R
B
alloon
I
r
is
C
a
n
c
e
r
I
on
osp
h
e
re
Error
rat
e
of
MBP
‐
RBFN
with
continuous
dataset
Error
rat
e
of
MBP
‐
RBFN
with
discretized
dataset
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 2, Februa
ry 2015 : 369 – 378
378
[10]
Cho T
H
, RW
Con
ners, PA Araman.
F
a
st backpr
opa
gati
on le
arni
ng us
i
ng steep activ
a
tion functi
on
s
and a
u
to
matic
w
e
ight reinitia
li
z
a
tio
n
.
Systems, Man, and
Cy
ber
netics. Decisio
n
Aid
i
ng
for Compl
e
x
System
s, Conferenc
e Proceedings.,
IEEE In
ternational Conference on
IEEE. 1991.
[11]
Cui L, C W
a
n
g
, B Yang,
Applic
atio
n of RBF
neural n
e
tw
ork im
prov
ed
by PSO algorithm i
n
faul
t
dia
gnos
is.
Jour
nal of T
heoreti
c
al an
d App
lie
d Info
rmation T
e
chn
o
lo
g
y
. 201
2; 46(1): p. 268
-273.
[12] Yufeng L,
X
C
hao, F
Yaozu.
T
he Co
mp
aris
on of RBF
and
BP Neural Ne
tw
ork in Decou
p
lin
g of DT
G.
201
0.
[13]
Sirat M, C T
a
lbot.
App
licati
o
n of artificia
l
neur
al n
e
tw
orks to fracture ana
lysis at th
e Äspö H
R
L,
Sw
eden: fractu
re sets cl
assifi
cation.
Inter
nat
ion
a
l J
ourn
a
l
o
f
Rock Mec
h
a
n
ics a
nd M
i
ni
n
g
Scie
nces,
200
1; 38(5): 62
1-63
9.
[14]
Plag
ian
a
kos V
,
D Sotiropo
ul
os, M Vrahati
s
.
A non
mon
o
t
one b
a
ckpro
p
agati
on trai
nin
g
metho
d
for
neur
al netw
o
rk
s.
Dept. of Mathematics, Un
iv
. of
Patras,
T
e
chnic
a
l Re
port, 199
8; 98-0
4
.
[15] Ed
w
a
r
d
R.
A
n
Introductio
n
to
Neur
al
N
e
tw
orks.
A W
h
ite p
a
per. Vis
ual
Nu
merics Inc., Un
ited States
o
f
America, 20
04.
[16]
Rimer M,
T
Ma
rtinez,
CB3: an
adaptiv
e error
function for ba
ckprop
agati
on
traini
ng.
Neur
al
processin
g
letters. 2006; 2
4
(1): 81-9
2
.
[17]
Qasem S, S
Shamsu
ddi
n,
Genera
l
i
z
a
t
i
o
n
impr
ove
m
ent of radia
l
basis
function n
e
tw
ork base
d
o
n
mu
lti-ob
jectiv
e particl
e sw
arm
opti
m
i
z
at
ion.
J. Artif.
Intell. 2010;
3
(1).
[18]
Arisari
y
a
w
o
n
g
S, S Char
oen
sean
g.
Dyn
a
m
ic self-or
gan
i
z
ed l
earn
i
n
g
fo
r opti
m
i
z
i
n
g t
h
e co
mp
lexity
grow
th of r
a
dial
b
a
sis
fu
nction
n
eura
l
netw
o
rks
.
Industrial Technology. IEEE I
C
IT'02. IEEE
Internatio
na
l C
onfere
n
ce o
n
. 200
2.
[19] Akbil
g
ic
O.
Bin
a
ry Class
ificati
on for Hydr
aul
i
c
F
r
acturing
Operati
ons i
n
Oil
& Gas W
e
lls via T
r
ee Base
d
Log
istic RBF Netw
orks.
European Jo
urn
a
l of Pure & Appl
ie
d
Mathematics. 201
3; 6(4).
[20] Venkates
an
P, S
Anitha.
Ap
pli
c
ation
of a rad
i
al bas
is functi
o
n
ne
ural
netw
o
rk for diag
nos
is
of diab
etes
me
llitus.
CU
R
R
ENT
SCIENCE-BANGALOR
E
. 2006; 91(
9): 1195.
[21]
Z
hang R, et al.
An efficient se
que
ntial RBF
n
e
tw
ork for bio-me
dic
a
l class
i
fi
cation pr
obl
e
m
s
. in
Neura
l
Networks.
Proceedings. IEEE International
Joint Conferenc
e
on. 2004.
[22]
Castell
a
n
o
s A, A Martinez B
l
anco, V Pa
le
n
c
ia.
App
licati
o
ns of rad
i
al
ba
sis ne
ural
net
w
o
rks for area
forest.
2007.
[23]
Hua
ng R
b
, Y
m
Che
u
n
g
.
A fast i
m
pl
e
m
ent
ation
of rad
i
a
l
basis
fu
nctio
n
netw
o
rks w
i
th app
licati
on t
o
tim
e
series for
e
casting
.
Intell
i
gent D
a
ta E
n
g
i
ne
erin
g a
nd A
u
tomated
Le
ar
nin
g
. Spri
nger.
200
3; 14
3-
150.
[24]
Gohari
an N,
D Gr
ossman, an
d N. Ra
ju.
Ext
end
ing
the
und
ergra
duat
e co
mp
uter sci
enc
e curric
u
lu
m t
o
inclu
de d
a
ta mi
nin
g
.
Informati
on T
e
chn
o
lo
gy
: Codin
g
and C
o
mputi
ng.
Proc
eed
ings IEEE. IT
CC 2004.
Internatio
na
l C
onfere
n
ce o
n
. 200
4.
[25]
Liu H, et a
l
.
Di
screti
z
at
io
n: An en
abl
ing tec
hni
que.
D
a
ta
minin
g
a
nd kn
o
w
le
dg
e disc
o
v
er
y
.
2
0
0
2
; 6(4
)
:
393-
423.
[26] Dou
ghert
y
J,
R Kohav
i, M Sahami.
Su
p
e
rvise
d
an
d u
n
sup
e
rvise
d
di
screti
z
at
io
n of
continu
o
u
s
features
. in
ICML
. 1995.
[27]
Agre G, S Peev.
On
su
pe
rvised and un
superv
i
sed dis
c
reti
z
a
t
i
o
n
.
C
y
bern
e
tics A
nd Information
T
e
chnolog
ies.
200
2; 2(2): 43-
57.
[28]
Saad P,
et al.
Rou
gh
set on trade
mark i
m
a
ges
for
n
eur
al netw
o
rk
classif
i
er.
Internati
o
n
a
l j
ourn
a
l
of
computer math
ematics. 200
2; 79(7): 78
9-7
9
6
.
[29]
Montazer
GA, H K
hos
hni
at, V F
a
thi.
Imp
r
ove
m
e
n
t of
RBF
Ne
ural
Netw
orks Usi
n
g F
u
zz
y
-
OSD
Algorit
h
m
in an
Online R
a
d
a
r Pulse C
l
assific
a
tion Syste
m
.
Appl
ied S
o
ft Computi
ng. 20
1
3
.
[30]
Kim GH, et al
.
Neural netw
o
rk mo
del i
n
c
o
rpor
ating a g
enetic
a
l
gor
ith
m
in esti
matin
g
constructio
n
costs.
Buildi
ng
and Env
i
ro
nme
n
t. 2004; 39(
11
): 1333-1
3
4
0
.
[31]
Yi H, G Ji, H Z
heng.
Optimal Parameters
of Bp Netw
ork for Characte
r Recog
n
itio
n
. in
Industria
l
Contro
l and El
ectronics En
gi
neer
ing (ICICE
E)
. IEEE. International
Confer
ence on. 2012.
[32]
Vakil-B
aghm
is
heh MT
, N
Paveši
ć
.
T
r
ai
nin
g
RBF
netw
o
rks
w
i
th
selectiv
e bac
kprop
agati
o
n
.
Neur
ocomp
u
tin
g
. 2004; 6
2
: 39
-64.
Evaluation Warning : The document was created with Spire.PDF for Python.