I
AE
S
I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
Ar
t
if
icial
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
Vol.
14
,
No.
4
,
Augus
t
2025
,
pp.
2753
~
2764
I
S
S
N:
2252
-
8938
,
DO
I
:
10
.
11591/i
jai
.
v
14
.i
4
.
pp
27
53
-
2764
2753
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
ai
.
iaes
c
or
e
.
c
om
Op
t
imiz
in
g l
on
g sh
or
t
-
t
e
r
m
m
e
m
o
r
y h
y
p
e
r
p
a
r
am
e
t
e
r
f
o
r
c
r
yp
t
oc
u
r
r
e
n
c
y s
e
n
t
ime
n
t
a
n
al
ysi
s with
swar
m
i
n
t
e
ll
ig
e
n
c
e
al
gor
ith
m
s
Kris
t
ian
E
k
ac
h
an
d
r
a,
Din
ar
Aj
e
n
g
Kris
t
iyan
t
i
I
nf
or
ma
ti
on S
ys
te
ms
S
tu
dy P
r
og
r
a
m, F
a
c
ul
ty
of
E
ngi
ne
e
r
in
g a
n
d
I
nf
or
ma
ti
c
s
, U
ni
ve
r
s
it
a
s
M
ul
ti
me
di
a
N
us
a
nt
a
r
a
, T
a
ng
e
r
a
ng, I
ndone
s
ia
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
J
un
18,
2024
R
e
vis
e
d
F
e
b
26,
2025
Ac
c
e
pted
M
a
r
15,
2025
T
h
i
s
s
t
u
d
y
i
n
v
es
t
i
g
at
e
s
t
h
e
ap
p
l
i
ca
t
i
o
n
o
f
s
w
arm
i
n
t
e
l
l
i
g
e
n
ce
al
g
o
r
i
t
h
ms
,
s
p
ec
i
fi
ca
l
l
y
p
art
i
cl
e
s
w
arm
o
p
t
i
m
i
zat
i
o
n
(PSO
),
an
t
co
l
o
n
y
o
p
t
i
m
i
zat
i
o
n
(A
CO
),
an
d
cat
s
w
arm
o
p
t
i
mi
za
t
i
o
n
(CSO
),
t
o
o
p
t
i
mi
ze
l
o
n
g
s
h
o
rt
-
t
erm
memo
ry
(L
ST
M)
n
et
w
o
r
k
s
fo
r
s
e
n
t
i
men
t
an
al
y
s
i
s
i
n
t
h
e
co
n
t
e
x
t
o
f
cry
p
t
o
c
u
rren
c
y
.
By
l
ev
era
g
i
n
g
t
h
es
e
o
p
t
i
m
i
zat
i
o
n
t
ech
n
i
q
u
e
s
,
w
e
a
i
med
t
o
en
h
a
n
ce
b
o
t
h
t
h
e
accu
ra
cy
an
d
co
m
p
u
t
at
i
o
n
al
effi
c
i
en
c
y
o
f
L
ST
M
mo
d
e
l
s
b
y
fi
n
e
-
t
u
n
i
n
g
cri
t
i
ca
l
h
y
p
er
p
aramet
er
s
,
n
o
t
ab
l
y
t
h
e
n
u
m
b
er
o
f
L
ST
M
u
n
i
t
s
.
T
h
e
s
t
u
d
y
i
n
v
o
l
v
ed
a
c
o
mp
ara
t
i
v
e
an
a
l
y
s
i
s
o
f
L
ST
M
m
o
d
e
l
s
o
p
t
i
m
i
zed
w
i
t
h
each
al
g
o
ri
t
h
m,
e
v
al
u
at
i
n
g
p
erf
o
rman
ce
me
t
ri
c
s
s
u
c
h
a
s
accu
racy
,
l
o
s
s
,
an
d
ex
ecu
t
i
o
n
t
i
me.
Res
u
l
t
s
i
n
d
i
ca
t
e
t
h
a
t
t
h
e
PSO
-
L
ST
M
mo
d
e
l
ach
i
e
v
e
d
t
h
e
h
i
g
h
e
s
t
accu
rac
y
at
8
6
.
0
8
%
an
d
t
h
e
l
o
w
e
s
t
l
o
s
s
at
0
.
5
7
,
w
i
t
h
a
red
u
ce
d
ex
ecu
t
i
o
n
t
i
me
o
f
5
8
.
4
3
s
ec
o
n
d
s
,
o
u
t
p
e
rfo
rm
i
n
g
b
o
t
h
A
CO
-
L
ST
M
an
d
CSO
-
L
ST
M
co
n
f
i
g
u
rat
i
o
n
s
.
T
h
es
e
f
i
n
d
i
n
g
s
u
n
d
er
s
co
re
t
h
e
e
ffect
i
v
en
e
s
s
o
f
PS
O
i
n
t
u
n
i
n
g
L
ST
M
p
arame
t
ers
a
n
d
em
p
h
a
s
i
ze
t
h
e
p
o
t
en
t
i
a
l
o
f
s
w
ar
m
i
n
t
el
l
i
g
en
ce
f
o
r
e
n
h
a
n
ci
n
g
n
eu
ra
l
n
et
w
o
r
k
p
erf
o
r
man
ce
i
n
rea
l
-
t
i
me
s
en
t
i
me
n
t
an
a
l
y
s
i
s
ap
p
l
i
cat
i
o
n
s
.
T
h
i
s
re
s
earch
c
o
n
t
ri
b
u
t
es
t
o
ad
v
a
n
ci
n
g
o
p
t
i
mi
zed
d
eep
l
ear
n
i
n
g
t
ech
n
i
q
u
e
s
i
n
h
i
g
h
d
i
me
n
s
i
o
n
a
l
d
a
t
a
en
v
i
r
o
n
me
n
t
s
,
w
i
t
h
i
m
p
l
i
cat
i
o
n
s
f
o
r
i
mp
r
o
v
i
n
g
cry
p
t
o
cu
rre
n
cy
s
e
n
t
i
me
n
t
p
re
d
i
c
t
i
o
n
s
.
K
e
y
w
o
r
d
s
:
C
r
yptocur
r
e
nc
y
F
e
a
tur
e
s
e
lec
ti
on
L
ong
s
hor
t
-
ter
m
memor
y
ne
t
wor
ks
S
e
nti
ment
a
na
lys
is
S
wa
r
m
int
e
ll
igenc
e
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e.
C
or
r
e
s
pon
din
g
A
u
th
or
:
Dina
r
Aje
ng
Kr
is
ti
ya
nti
I
nf
or
mation
S
ys
tems
S
tudy
P
r
ogr
a
m
,
F
a
c
ult
y
of
E
nginee
r
ing
a
nd
I
n
f
or
matics
Unive
r
s
it
a
s
M
ult
im
e
dia
Nus
a
ntar
a
T
a
nge
r
a
ng,
I
ndone
s
ia
E
mail:
dinar
.
kr
is
ti
ya
nti
@umn
.
a
c
.
id
1.
I
NT
RODU
C
T
I
ON
T
he
global
f
inanc
e
s
e
c
tor
ha
s
incr
e
a
s
ingl
y
tu
r
ne
d
it
s
a
tt
e
nti
on
to
c
r
yptocu
r
r
e
nc
ies
,
a
c
las
s
of
d
igi
tal
a
s
s
e
t
s
known
f
or
r
a
pid
va
lue
f
luctua
ti
ons
a
nd
d
e
c
e
ntr
a
li
z
e
d
c
ontr
ol
mec
ha
nis
ms
.
C
r
yptocur
r
e
nc
i
e
s
ha
ve
ga
ined
wide
s
pr
e
a
d
popular
it
y
a
s
highl
y
volatil
e
d
i
git
a
l
a
s
s
e
ts
that
o
f
f
e
r
both
high
r
is
k
a
nd
potentia
l
r
e
wa
r
d
f
or
inves
tor
s
.
T
he
r
a
pid
a
nd
of
ten
unp
r
e
dicta
ble
f
luctua
ti
ons
in
c
r
yptocur
r
e
nc
y
p
r
ice
s
a
r
e
inf
luenc
e
d
by
va
r
ious
f
a
c
tor
s
,
including
mar
ke
t
s
e
nti
ment,
tec
hn
ologi
c
a
l
a
dva
nc
e
ments
,
r
e
gulator
y
c
ha
nge
s
,
a
nd
in
f
luential
s
oc
ial
media
a
c
ti
vit
ies
.
B
it
c
oin,
the
f
ir
s
t
c
r
ypt
oc
ur
r
e
nc
y,
s
a
w
a
n
e
xt
r
a
or
dinar
y
r
is
e
in
va
lue
in
2017,
incr
e
a
s
ing
by
ove
r
2
,
000%
to
r
e
a
c
h
$20,
000
[
1]
.
C
r
yptocu
r
r
e
nc
ies
li
ke
bit
c
oin
of
f
e
r
s
e
c
ur
e
,
dir
e
c
t
tr
a
ns
a
c
ti
ons
without
int
e
r
media
r
ies
,
f
a
c
il
it
a
ted
by
blockc
ha
in
tec
hnology
[
2]
.
How
e
ve
r
,
they
pr
e
s
e
nt
unique
c
ha
ll
e
nge
s
,
s
uc
h
a
s
high
e
ne
r
gy
c
ons
umpt
ion
due
to
mi
ning
a
c
ti
vit
ies
,
a
nd
ha
ve
be
e
n
a
s
s
oc
iate
d
with
il
li
c
it
a
c
ti
vit
ies
,
whic
h
ha
ve
led
to
r
e
gulato
r
y
r
e
s
pons
e
s
f
r
om
va
r
ious
gove
r
nments
[
3]
,
[
4]
.
T
he
s
e
c
h
a
ll
e
nge
s
c
ontr
ibut
e
to
c
r
yptocur
r
e
nc
y
mar
ke
t
volati
li
ty,
a
s
r
e
gulator
y
ne
ws
a
nd
a
dva
nc
e
ment
s
in
c
r
yptocur
r
e
nc
y
inf
r
a
s
tr
uc
tur
e
(
e
.
g.
,
pr
oof
of
s
take
im
pleme
ntation
s
a
nd
e
xc
ha
nge
-
tr
a
de
d
f
und
(
E
T
F
)
a
ppr
ova
ls
)
c
on
ti
nue
to
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
2025
:
275
3
-
2764
2754
inf
luenc
e
pr
ice
moveme
nts
[
5]
–
[
8]
.
W
it
h
plat
f
or
m
s
li
ke
T
witt
e
r
/X
playing
a
s
igni
f
ica
nt
r
ole
in
s
ha
pi
ng
publi
c
opini
on,
s
e
nti
ment
a
na
lys
is
ha
s
e
mer
ge
d
a
s
a
va
l
ua
ble
tool
f
or
a
s
s
e
s
s
ing
inves
tor
s
e
nti
ment
a
nd
p
r
e
dicting
c
r
yptocur
r
e
nc
y
tr
e
nds
[
9
]
.
How
e
ve
r
,
c
onduc
ti
ng
a
c
c
ur
a
te
s
e
nti
ment
a
na
lys
is
in
thi
s
f
ield
pos
e
s
c
ha
ll
e
nge
s
due
to
the
high
dim
e
ns
ionalit
y
a
nd
nois
y
na
tu
r
e
of
s
oc
ial
media
da
ta,
whic
h
ne
c
e
s
s
it
a
tes
a
dva
nc
e
d
mac
hine
lea
r
ning
models
c
a
pa
ble
of
ha
ndli
ng
s
uc
h
c
ompl
e
x
it
y
[
10]
,
[
11]
.
S
e
nt
im
e
n
t
a
na
lys
is
o
n
s
oc
i
a
l
m
e
d
ia
da
ta
is
t
yp
ic
a
l
ly
c
on
du
c
te
d
us
in
g
th
r
e
e
a
pp
r
oa
c
he
s
:
le
xi
c
o
n
-
ba
s
e
d
,
mac
hine
lea
r
ning
-
ba
s
e
d,
a
nd
hybr
id
methods
.
L
e
xicon
-
ba
s
e
d
methods
a
r
e
s
uit
a
ble
f
or
uns
upe
r
vis
e
d
da
ta,
while
mac
hine
lea
r
ning
a
ppr
oa
c
he
s
r
e
quir
e
labe
le
d
da
tas
e
ts
[
12]
.
Hybr
id
a
ppr
oa
c
he
s
that
int
e
gr
a
te
lexic
ons
with
mac
hine
lea
r
ning
ha
ve
de
mons
tr
a
ted
im
p
r
ov
e
d
a
c
c
ur
a
c
y,
with
s
tudi
e
s
r
e
por
ti
ng
up
to
10%
ga
ins
ove
r
c
onve
nti
ona
l
methods
[
13]
.
R
e
c
e
nt
r
e
s
e
a
r
c
h
hi
ghli
ghts
the
e
f
f
e
c
ti
ve
ne
s
s
of
de
e
p
lea
r
n
ing
a
lg
or
it
hms
,
pa
r
ti
c
ular
ly
long
s
hor
t
-
ter
m
memo
r
y
(
L
S
T
M
)
n
e
twor
ks
,
whic
h
c
a
n
c
a
ptur
e
tempor
a
l
de
pe
nde
n
c
ies
a
nd
c
ompl
e
x
pa
tt
e
r
ns
in
s
e
que
nti
a
l
da
ta
li
ke
s
oc
ial
media
pos
ts
[
14]
,
[
15]
.
S
tudi
e
s
ha
ve
s
hown
that
L
S
T
M
outper
f
or
ms
t
r
a
dit
ional
mac
hine
lea
r
ning
model
s
in
s
e
nti
ment
a
na
lys
is
,
making
it
a
s
uit
a
ble
c
hoice
f
or
a
na
lyzin
g
high
dim
e
ns
ional
a
nd
nois
y
da
ta
[
11]
,
[
1
6]
,
[
17]
.
De
s
pit
e
L
S
T
M
's
a
dva
ntage
s
,
im
pr
oving
it
s
pe
r
f
o
r
manc
e
f
or
s
e
nti
ment
a
na
lys
is
on
la
r
ge
s
oc
ial
medi
a
da
tas
e
ts
r
e
quir
e
s
f
e
a
tur
e
s
e
lec
ti
on
tec
hnique
s
to
mana
ge
da
ta
dim
e
ns
ionalit
y
a
nd
r
e
duc
e
nois
e
.
How
e
ve
r
,
the
e
f
f
e
c
ti
ve
ne
s
s
of
L
S
T
M
models
is
highl
y
de
pe
nde
nt
on
their
hype
r
pa
r
a
mete
r
s
,
s
uc
h
a
s
the
number
of
L
S
T
M
unit
s
.
T
r
a
dit
ional
methods
of
hype
r
pa
r
a
mete
r
tu
ning
c
a
n
be
ti
me
-
c
ons
umi
ng
a
nd
may
not
a
lwa
ys
yield
opti
mal
c
onf
igur
a
ti
ons
,
e
s
pe
c
ially
in
high
dim
e
ns
ional
s
e
nti
ment
a
na
ly
s
is
tas
k
s
.
I
n
r
e
c
e
nt
ye
a
r
s
,
s
wa
r
m
int
e
ll
igenc
e
a
lgor
it
hms
,
s
uc
h
a
s
pa
r
ti
c
le
s
wa
r
m
op
ti
mi
z
a
ti
on
(
P
S
O)
,
a
nt
c
olony
opt
im
iza
ti
on
(
AC
O)
,
a
nd
c
a
t
s
wa
r
m
opti
mi
z
a
ti
on
(
C
S
O)
,
ha
ve
be
e
n
uti
li
z
e
d
to
i
mpr
ove
opti
mi
z
a
ti
on
pr
oc
e
s
s
e
s
a
c
r
os
s
va
r
ious
dom
a
ins
due
to
their
a
bil
it
y
to
e
f
f
icie
ntl
y
e
xplor
e
lar
ge
s
e
a
r
c
h
s
pa
c
e
s
[
18]
,
[
19]
.
P
r
ior
s
tudi
e
s
ha
ve
us
e
d
s
wa
r
m
int
e
ll
igenc
e
a
lgor
it
hms
to
im
pr
ove
a
c
c
ur
a
c
y
in
mac
hine
lea
r
ning
a
ppli
c
a
ti
ons
,
with
P
S
O
incr
e
a
s
ing
a
c
c
ur
a
c
y
f
or
S
VM
models
f
r
om
78.
70
%
to
86.
20
%
[
20
]
,
a
nd
a
da
pti
ve
pa
r
ti
c
le
s
wa
r
m
opti
mi
z
a
ti
on
(
APS
O)
im
pr
ovin
g
L
S
T
M
a
c
c
ur
a
c
y
f
r
om
95
.
1%
to
97
.
8%
in
s
e
nti
ment
c
la
s
s
if
ica
ti
on
[
21]
.
S
tudi
e
s
c
ompar
ing
P
S
O
a
nd
C
S
O
ha
ve
s
hown
that
C
S
O
c
a
n
de
li
ve
r
e
ve
n
be
tt
e
r
a
c
c
ur
a
c
y
a
nd
f
a
s
ter
pr
oc
e
s
s
ing
ti
mes
in
s
e
nti
ment
a
na
ly
s
is
[
11]
.
How
e
ve
r
,
r
e
s
e
a
r
c
h
int
e
gr
a
ti
ng
thes
e
a
lgor
it
h
ms
s
pe
c
if
ica
ll
y
with
L
S
T
M
f
o
r
c
r
yp
tocur
r
e
nc
y
s
e
nti
ment
a
na
lys
is
r
e
mains
li
mi
ted,
pr
e
s
e
nti
ng
a
ga
p
that
thi
s
s
tudy
a
im
s
to
f
il
l
.
T
his
s
tudy
a
im
s
to
a
d
dr
e
s
s
the
li
mi
tations
of
tr
a
dit
ional
L
S
T
M
tuni
ng
methods
by
e
mpl
oying
P
S
O,
AC
O,
a
nd
C
S
O
a
lgor
it
hms
to
opti
mi
z
e
L
S
T
M
ne
twor
ks
s
pe
c
if
ica
ll
y
f
or
c
r
yptocu
r
r
e
nc
y
s
e
nti
ment
a
na
lys
is
.
B
y
f
ine
-
tun
ing
the
L
S
T
M
un
it
s
a
nd
other
ke
y
hype
r
pa
r
a
mete
r
s
,
we
a
im
to
e
nha
nc
e
the
mod
e
l's
a
c
c
ur
a
c
y,
r
e
duc
e
pr
oc
e
s
s
ing
ti
me,
a
nd
im
pr
ov
e
ove
r
a
ll
pe
r
f
or
manc
e
.
T
he
r
e
mainde
r
o
f
thi
s
pa
pe
r
is
or
ga
nize
d
a
s
f
oll
o
ws
:
s
e
c
ti
on
2
pr
e
s
e
nts
the
methodology,
de
taili
ng
the
da
ta
pr
e
-
pr
oc
e
s
s
ing
s
tep
s
,
the
s
wa
r
m
int
e
ll
igenc
e
-
ba
s
e
d
opti
mi
z
a
ti
on
tec
hniques
a
ppli
e
d,
a
nd
opti
mi
z
ing
L
S
T
M
us
ing
hype
r
pa
r
a
mete
r
tuni
ng.
S
e
c
ti
on
3
d
is
c
us
s
e
s
the
e
xpe
r
im
e
ntal
r
e
s
ult
s
,
including
a
be
s
t
model
pe
r
f
or
manc
e
a
na
lys
is
,
c
onf
us
ion
matr
ix
a
nd
c
las
s
if
ica
t
ion
metr
ics
,
dis
c
us
s
ion,
a
nd
li
mi
tati
ons
a
nd
im
pli
c
a
ti
ons
f
or
f
utur
e
r
e
s
e
a
r
c
h.
F
inally,
s
e
c
ti
on
4
c
onc
ludes
with
a
s
umm
a
r
y
of
the
f
indi
ngs
a
nd
potential
a
ppli
c
a
ti
ons
in
c
r
yptocur
r
e
nc
y
mar
ke
t
a
na
lys
is
.
2.
M
E
T
HO
D
T
he
methodology
of
thi
s
s
tudy,
a
s
il
lus
tr
a
ted
in
F
igur
e
1,
c
ons
is
ts
of
s
e
ve
r
a
l
ke
y
s
tage
s
to
pr
e
pa
r
e
a
nd
opti
mi
z
e
a
n
L
S
T
M
model
f
or
c
r
yptocur
r
e
nc
y
s
e
nti
ment
a
na
lys
is
.
E
a
c
h
s
tage
,
f
r
o
m
da
ta
p
r
e
pr
oc
e
s
s
ing
to
model
e
va
luation,
is
outl
ined
to
f
a
c
il
it
a
te
unde
r
s
tanding
a
nd
r
e
pr
oduc
ibi
li
ty
.
T
h
is
s
tr
uc
tur
e
d
a
ppr
oa
c
h
e
ns
ur
e
s
that
e
a
c
h
c
omponent
of
the
model
de
ve
lopm
e
nt
pr
oc
e
s
s
is
s
y
s
tema
ti
c
a
ll
y
a
ddr
e
s
s
e
d,
lea
ding
to
mor
e
r
e
li
a
ble
a
nd
ins
ight
f
ul
s
e
nti
ment
p
r
e
dictions
.
2.
1.
Dat
a
c
oll
e
c
t
ion
Da
ta
f
or
thi
s
s
tudy
wa
s
c
oll
e
c
ted
f
r
om
T
witt
e
r
/
X
us
ing
the
twe
e
t
-
ha
r
ve
s
t
tool
,
c
onf
igur
e
d
with
pa
r
a
mete
r
s
s
uc
h
a
s
twit
ter
_a
uth_t
oke
n,
s
e
a
r
c
h_ke
ywor
d,
a
nd
li
mi
t
to
s
tr
e
a
ml
ine
da
ta
e
xtr
a
c
ti
on.
T
he
ke
ywor
ds
include
d
ter
ms
l
ike
"
c
r
yptocur
r
e
n
c
y,
"
"
c
r
ypto,
"
a
nd
"
b
it
c
oin,
"
a
s
we
ll
a
s
r
e
late
d
te
r
ms
that
c
a
ptur
e
publi
c
s
e
nti
ment
on
c
r
yptocur
r
e
nc
ies
.
T
h
e
da
ta
c
oll
e
c
ti
on
wa
s
li
mi
ted
to
1,
000
twe
e
ts
pe
r
da
y
to
e
ns
ur
e
a
mana
ge
a
ble
da
tas
e
t
with
a
b
r
oa
d
r
a
nge
of
op
ini
ons
.
Ove
r
the
da
ta
c
oll
e
c
ti
on
pe
r
io
d,
f
r
om
De
c
e
mber
3
1,
2023
,
to
J
a
nua
r
y
31,
2024,
a
tot
a
l
of
9
,
884
twe
e
ts
we
r
e
s
uc
c
e
s
s
f
ull
y
ga
ther
e
d,
pr
o
vidi
ng
a
r
obus
t
da
tas
e
t
f
or
s
e
nti
ment
a
na
lys
is
.
T
his
pe
r
iod
wa
s
s
pe
c
if
ica
ll
y
c
hos
e
n
to
c
a
ptur
e
dis
c
us
s
ions
a
r
ound
the
U.
S
.
S
E
C
’
s
B
T
C
E
T
F
a
ppr
ova
l
on
J
a
nua
r
y
1
0,
2024,
a
n
e
ve
nt
a
nti
c
ipate
d
to
s
igni
f
ica
ntl
y
i
nf
luenc
e
c
r
yptocur
r
e
nc
y
s
e
nti
ment
a
nd
publ
ic
dis
c
us
s
ion
[
22]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
Optimiz
ing
long
s
hor
t
-
ter
m
me
mor
y
hy
pe
r
pa
r
ame
t
e
r
for
c
r
y
ptocur
r
e
nc
y
s
e
nti
me
nt
…
(
K
r
is
ti
an
E
k
ac
handr
a
)
2755
F
igur
e
1.
P
r
opos
e
d
a
r
c
hit
e
c
tur
e
2.
2.
Dat
a
p
r
e
-
p
r
oc
e
s
s
in
g
T
he
c
oll
e
c
ted
T
witt
e
r
/X
da
ta
r
e
quir
e
d
e
xtens
ive
pr
e
-
pr
oc
e
s
s
ing
to
e
ns
ur
e
c
ons
is
tenc
y
a
nd
r
e
duc
e
nois
e
,
ther
e
by
im
pr
ov
ing
the
a
c
c
ur
a
c
y
of
the
s
e
nti
ment
a
na
lys
is
model.
E
a
c
h
s
tage
in
thi
s
pr
oc
e
s
s
is
outl
ined
in
f
ol
lowing
s
ubs
e
c
ti
on.
T
his
pr
oc
e
s
s
include
s
da
ta
c
lea
ning,
text
nor
maliza
ti
on
,
a
nd
r
e
moval
o
f
ir
r
e
leva
nt
e
leme
nts
to
pr
oduc
e
higher
qua
li
ty
input
f
o
r
the
m
ode
l.
2.
2.
1.
Re
m
ove
UR
L
s
,
m
e
n
t
io
n
s
,
h
as
t
ags
,
s
p
e
c
ial
c
h
ar
ac
t
e
r
s
,
an
d
n
u
m
b
e
r
T
he
f
ir
s
t
s
tep
invol
ve
d
r
e
movi
ng
UR
L
s
,
mentions
,
ha
s
htags
,
s
pe
c
ial
c
ha
r
a
c
ter
s
,
a
nd
number
s
f
r
om
the
twe
e
ts
.
UR
L
s
a
nd
mentions
(
e
.
g.
,
h
tt
ps
:/
/exa
mpl
e
.
c
om,
@us
e
r
na
me)
of
ten
c
ontain
non
-
s
e
nti
ment
-
be
a
r
ing
inf
or
mation
,
while
ha
s
htags
a
nd
s
pe
c
ial
c
ha
r
a
c
te
r
s
a
dd
nois
e
to
the
da
ta.
Numbe
r
s
we
r
e
a
ls
o
r
e
moved
to
f
oc
us
the
a
na
lys
is
on
textua
l
c
ontent
r
e
leva
nt
to
s
e
nti
ment.
F
or
e
xa
mpl
e
,
the
twe
e
t
“
C
he
c
k
thi
s
out!
htt
ps
:/
/cr
ypto.
c
om
@c
r
ypto
#
bit
c
oin
123”
would
be
r
e
duc
e
d
to
“
C
h
e
c
k
thi
s
out
bit
c
oin
”
[
23]
.
2.
2.
2.
Re
m
ove
p
u
n
c
t
u
a
t
ion
a
n
d
c
on
ve
r
t
t
o
lowe
r
c
as
e
All
punc
tuation
wa
s
r
e
moved
,
a
nd
the
text
wa
s
c
onve
r
ted
to
lowe
r
c
a
s
e
.
T
his
s
tep
e
ns
ur
e
s
that
s
im
il
a
r
wo
r
ds
with
dif
f
e
r
e
nt
c
a
s
e
s
(
e
.
g.
,
“
B
it
c
oin”
a
nd
“
bit
c
oin”)
a
r
e
tr
e
a
ted
uni
f
or
ml
y
,
r
e
duc
ing
va
r
iabili
ty
a
nd
voc
a
bular
y
s
ize
.
F
or
ins
tanc
e
,
the
text
“
B
it
c
oin,
the
f
utur
e
!
”
be
c
omes
“
bit
c
oin
the
f
utur
e
,
”
e
ns
ur
ing
c
ons
is
tenc
y
a
c
r
os
s
the
da
tas
e
t
[
23]
,
[
24]
.
2.
2.
3.
T
ok
e
n
izat
ion
T
oke
niza
ti
on
wa
s
a
ppli
e
d
to
s
pli
t
e
a
c
h
twe
e
t
int
o
indi
vidual
wor
ds
or
tokens
,
making
it
e
a
s
ier
f
or
the
model
to
a
na
lyze
textua
l
da
ta
mor
e
e
f
f
e
c
ti
ve
ly.
T
his
p
r
oc
e
s
s
invol
ve
s
br
e
a
king
down
s
e
ntenc
e
s
int
o
s
maller
c
omponents
,
s
uc
h
a
s
wor
ds
or
s
ymbol
s
,
whic
h
a
ll
ows
mac
hine
lea
r
ning
a
lgor
it
hms
to
ha
ndle
a
nd
int
e
r
pr
e
t
the
text
s
ys
tema
ti
c
a
ll
y.
F
or
in
s
tanc
e
,
the
s
e
ntenc
e
“
B
it
c
oin
is
the
f
utur
e
”
would
be
tokeni
z
e
d
int
o
[
“
bit
c
oin”,
“
is
”
,
“
the”
,
“
f
utur
e
”
]
,
e
na
bli
ng
the
mo
de
l
to
e
va
luate
e
a
c
h
token
s
e
pa
r
a
tely
a
nd
identif
y
pa
tt
e
r
ns
or
s
e
nti
ments
a
s
s
oc
iate
d
with
indi
vidual
wor
ds
[
25
]
.
2.
2.
4.
Re
m
ove
s
t
op
wor
d
s
C
o
m
m
o
n
l
y
u
s
e
d
w
o
r
d
s
t
h
a
t
d
o
n
o
t
c
o
n
t
r
i
b
u
t
e
t
o
s
e
n
t
i
m
e
n
t
,
k
n
o
w
n
a
s
s
t
o
p
w
o
r
d
s
(
e
.
g
.
,
"
i
s
,
"
"
a
n
d
,
"
"
t
h
e
"
)
,
we
r
e
r
e
moved.
T
his
s
tep
r
e
duc
e
s
da
ta
c
ompl
e
xit
y
by
a
ll
owing
the
model
to
f
oc
us
on
s
e
nti
ment
-
r
e
leva
nt
wor
ds
.
Af
ter
s
top
wor
d
r
e
moval,
the
phr
a
s
e
“
B
it
c
oin
is
the
f
utur
e
”
would
r
e
tain
only
[
"
bit
c
oin"
,
"
f
utur
e
"
]
,
highl
ight
ing
the
c
o
r
e
s
e
nti
ment
-
be
a
r
ing
wor
ds
[
23]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
2025
:
275
3
-
2764
2756
2.
2.
5.
L
e
m
m
a
t
izat
ion
L
e
mm
a
ti
z
a
ti
on
wa
s
a
ppli
e
d
to
c
onve
r
t
wo
r
ds
int
o
their
ba
s
e
f
or
ms
,
e
ns
ur
ing
that
va
r
iations
of
the
s
a
me
wor
d
a
r
e
tr
e
a
ted
a
s
one
.
F
or
ins
tanc
e
,
“
r
is
i
ng”
a
nd
“
r
os
e
”
we
r
e
both
c
onve
r
ted
to
“
r
is
e
.
”
T
his
s
tep
im
pr
ove
s
the
model’
s
a
bil
it
y
to
r
e
c
ognize
di
f
f
e
r
e
nt
f
or
ms
of
the
s
a
me
wor
d
,
the
r
e
by
e
nha
nc
ing
c
o
ns
is
tenc
y
a
nd
r
e
duc
ing
dim
e
ns
ionalit
y
in
the
da
ta
[
23]
.
2.
2.
6.
Re
m
ove
wor
d
s
wi
t
h
leng
t
h
<
3
L
a
s
tl
y,
wor
ds
with
f
e
we
r
than
thr
e
e
c
ha
r
a
c
ter
s
we
r
e
r
e
moved,
a
s
they
typi
c
a
ll
y
pr
ovide
li
mi
ted
s
e
nti
ment
inf
or
mation
a
nd
c
a
n
a
dd
nois
e
.
S
hor
t
w
or
ds
s
uc
h
a
s
“
it
”
a
nd
“
a
n”
we
r
e
e
xc
luded
to
s
tr
e
a
ml
ine
the
da
tas
e
t
f
ur
ther
a
nd
f
oc
us
the
a
na
lys
is
on
mea
ningf
ul
ter
ms
.
T
his
f
il
ter
ing
im
pr
ove
s
the
model
’
s
e
f
f
ic
ienc
y
by
r
e
duc
ing
unne
c
e
s
s
a
r
y
da
ta
e
leme
nts
[
26]
.
2.
3.
Dat
a
lab
e
li
n
g
Da
ta
wa
s
labe
led
us
ing
the
va
lenc
e
a
wa
r
e
dictiona
r
y
a
nd
s
e
nti
ment
r
e
a
s
one
r
(
VA
DE
R
)
,
a
s
e
nti
ment
a
na
lys
is
tool
that
pe
r
f
o
r
ms
we
ll
on
s
oc
ial
media
d
a
ta
by
a
c
c
ounti
ng
f
o
r
e
mot
icons
,
a
bb
r
e
viation
s
,
a
nd
other
inf
or
mal
langua
ge
f
e
a
tur
e
s
c
omm
only
us
e
d
on
T
witt
e
r
/X.
VA
DE
R
a
s
s
igns
s
e
nti
ment
s
c
or
e
s
that
c
a
tegor
ize
e
a
c
h
twe
e
t
a
s
pos
it
ive,
ne
ga
ti
ve
,
o
r
ne
utr
a
l
[
26]
.
F
or
thi
s
s
tudy,
only
pos
it
ive
a
nd
ne
ga
ti
ve
s
e
nti
me
nts
we
r
e
r
e
taine
d,
a
s
thes
e
a
r
e
mos
t
indi
c
a
ti
ve
of
the
buy
-
or
-
s
e
ll
de
c
is
ions
in
c
r
yptocur
r
e
nc
y
tr
a
ding,
while
ne
utr
a
l
s
e
nti
ments
we
r
e
e
xc
luded
to
maintain
a
f
oc
us
on
di
r
e
c
ti
ona
l
s
e
nti
ment
that
im
pa
c
ts
tr
a
ding
be
ha
vior
[
27]
,
[
28]
.
T
a
ble
1
r
e
pr
e
s
e
nts
a
s
a
mpl
e
of
the
labe
led
da
ta
us
ing
the
VA
DE
R
on
top
5
da
ta
th
a
t
ha
ve
unde
r
gone
da
ta
pr
e
-
pr
oc
e
s
s
ing.
T
a
ble
1.
L
a
be
led
da
ta
us
ing
VA
DE
R
P
r
oc
e
s
s
e
d
t
e
xt
V
A
D
E
R
s
e
nt
im
e
nt
B
T
C
hone
s
tl
y dont
th
in
k ma
tt
e
r
t
a
r
ge
t
ge
t
pa
s
t
a
th
dr
op ba
c
k
s
ti
ll
bul
li
s
h c
lo
s
e
P
os
it
iv
e
of
f
ic
ia
l
ha
ppy ne
w
ye
a
r
c
r
ypt
o c
omm
uni
ty
pr
a
yi
ng ma
y ye
a
r
br
in
g gr
e
e
n c
a
ndl
e
a
mp e
ve
r
y c
oi
n
P
os
it
iv
e
f
in
a
nc
ia
l
ma
r
ke
t
a
na
ly
s
t
c
r
ypt
oc
ur
r
e
nc
y bl
oc
kc
ha
in
a
mp w
e
b r
e
s
e
a
r
c
he
r
N
e
ga
ti
ve
mone
y br
oke
n l
ow
i
nt
e
r
e
s
t
r
a
te
f
a
ke
mone
y l
e
a
d pe
opl
e
t
r
e
a
t
r
e
a
l
e
s
ta
te
i
nve
s
tm
e
nt
e
nt
e
r
e
d c
ha
t
N
e
ga
ti
ve
a
bs
ol
ut
e
ly
out
r
a
ge
ous
ba
nkma
nf
r
ie
ds
p
a
c
e
s
s
e
nt
ia
ll
y
pa
id
f
r
a
nc
is
c
onol
e
s
de
moc
r
a
ti
c
pr
im
a
r
y
c
a
mpa
ig
n s
to
le
n c
r
ypt
o f
und c
onol
e
e
ke
d n
a
r
r
ow
w
in
f
ix
one
he
ld
a
c
c
ount
a
bl
e
N
e
ga
ti
ve
2.
4.
Dat
a
s
p
li
t
t
i
n
g
T
he
labe
led
da
tas
e
t
wa
s
s
pli
t
int
o
t
r
a
ini
ng
a
nd
tes
ti
ng
s
e
ts
,
with
80%
o
f
the
da
ta
a
ll
oc
a
ted
f
o
r
tr
a
ini
ng
a
nd
20%
f
or
tes
ti
ng.
T
his
s
pli
t
e
na
bles
the
e
va
luation
of
model
ge
ne
r
a
li
z
a
bil
it
y
.
R
e
f
e
r
r
ing
to
s
tudy
that
c
ompar
e
d
the
r
a
ti
os
o
f
tr
a
ini
ng
a
nd
tes
ti
ng
d
a
ta
in
s
e
nti
ment
a
na
lys
is
f
or
c
r
yptocur
r
e
nc
ies
us
ing
twe
e
t
da
ta,
the
r
a
ti
o
of
80:20
yielde
d
the
be
s
t
pe
r
f
or
man
c
e
c
ompar
e
d
to
r
a
ti
os
of
90:10
a
nd
70:30
[
29]
.
T
h
e
tr
a
ini
ng
da
ta
is
us
e
d
to
tr
a
in
the
c
las
s
if
ica
ti
on
model
f
or
both
the
objec
ti
ve
f
unc
ti
on
a
nd
s
e
nti
ment
c
las
s
if
ica
ti
on,
while
the
tes
ti
ng
da
ta
is
us
e
d
to
va
li
da
t
e
the
tr
a
ined
model.
2.
5.
F
e
a
t
u
r
e
s
e
lec
t
ion
u
s
in
g
s
war
m
in
t
e
ll
igence
algorit
h
m
s
T
he
f
e
a
tur
e
s
e
lec
ti
on
s
tage
in
s
e
nti
ment
a
na
lys
is
is
us
e
d
to
c
hoos
e
r
e
leva
nt
f
e
a
tur
e
s
be
c
a
us
e
da
ta
e
xtr
a
c
ted
f
r
o
m
s
oc
ial
media
ge
ne
r
a
ll
y
ha
s
high
di
mens
ional
c
ha
r
a
c
ter
is
ti
c
s
.
S
wa
r
m
in
telli
ge
nc
e
is
o
ne
of
the
c
omponents
of
f
e
a
tur
e
s
e
lec
ti
on
that
u
ti
li
z
e
s
th
e
hybr
id
method
.
Due
to
it
s
s
uit
a
bil
it
y
f
or
the
r
e
s
e
a
r
c
h
objec
ti
ve
,
thi
s
s
tudy
e
mpl
oys
the
hybr
id
method
o
f
f
e
a
tur
e
s
e
lec
ti
on
ba
s
e
d
on
s
wa
r
m
int
e
ll
igenc
e
to
opti
mi
z
e
the
L
S
T
M
model.
S
wa
r
m
in
telli
ge
nc
e
a
lgor
it
hm
s
P
S
O,
AC
O,
a
nd
C
S
O
we
r
e
e
mp
loyed
to
opt
i
mi
z
e
the
number
of
L
S
T
M
unit
s
.
E
a
c
h
a
lgor
it
hm
a
im
e
d
t
o
identif
y
a
n
idea
l
c
onf
igur
a
ti
on
that
maximi
z
e
s
a
c
c
ur
a
c
y
while
r
e
duc
ing
c
omput
a
ti
ona
l
ti
me.
T
he
hype
r
pa
r
a
mete
r
tuni
ng
p
r
oc
e
s
s
f
oc
us
e
d
on
the
L
S
T
M
unit
s
,
a
s
thi
s
pa
r
a
mete
r
c
r
it
ica
ll
y
i
mpac
ts
the
model's
a
bil
it
y
to
c
a
ptur
e
s
e
que
nti
a
l
pa
tt
e
r
ns
in
s
e
nti
ment
da
ta.
Ho
we
ve
r
,
to
ge
t
the
L
S
T
M
uni
ts
we
mus
t
us
e
the
objec
ti
ve
f
un
c
ti
on.
T
he
objec
ti
ve
f
unc
ti
on
is
de
s
igned
to
tr
a
in
t
he
L
S
T
M
m
ode
l
int
e
ll
igently
,
a
nd
it
ha
s
be
e
n
p
r
ove
n
to
e
f
f
e
c
ti
ve
ly
a
djus
t
the
we
ight
s
in
the
L
S
T
M
model,
mi
nim
izing
los
s
a
nd
f
a
c
il
it
a
ti
ng
e
f
f
icie
nt
model
lea
r
ning
[
30
]
.
2.
5.
1.
P
ar
t
icle
s
war
m
op
t
im
izat
io
n
Optim
iza
ti
on
tec
hnique
ins
pir
e
d
by
the
s
oc
ial
be
ha
vior
of
b
ir
ds
or
f
is
h
whe
n
s
e
a
r
c
hing
f
o
r
f
o
od
.
B
ir
ds
do
not
know
the
e
xa
c
t
loca
ti
on
of
f
ood
,
s
o
they
ge
ne
r
a
ll
y
f
oll
ow
other
b
ir
ds
c
ons
ider
e
d
c
los
e
to
the
f
ood
s
our
c
e
[
31]
.
I
n
P
S
O,
e
a
c
h
bir
d
is
r
e
f
e
r
r
e
d
to
a
s
a
pa
r
ti
c
le,
a
nd
e
a
c
h
pa
r
ti
c
le
ha
s
a
f
it
ne
s
s
f
unc
ti
on
(
s
qua
r
e
of
e
r
r
o
r
)
.
A
gr
oup
o
f
pa
r
ti
c
les
is
known
a
s
a
s
wa
r
m.
P
S
O
a
lgor
it
hm
is
s
hown
in
P
s
e
udoc
ode
1
.
T
he
wor
king
p
r
inciple
o
f
P
S
O
is
a
s
f
o
ll
ows
:
in
e
a
c
h
it
e
r
a
ti
on
,
the
a
lgor
it
hm
f
ir
s
t
f
inds
the
be
s
t
s
olut
ion
f
ound
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
Optimiz
ing
long
s
hor
t
-
ter
m
me
mor
y
hy
pe
r
pa
r
ame
t
e
r
for
c
r
y
ptocur
r
e
nc
y
s
e
nti
me
nt
…
(
K
r
is
ti
an
E
k
ac
handr
a
)
2757
withi
n
the
s
wa
r
m,
whic
h
is
s
tor
e
d
a
s
pe
r
s
ona
l
be
s
t
(
pbe
s
t)
.
T
he
n
,
global
be
s
t
(
gbe
s
t)
is
upda
ted
to
be
the
be
s
t
s
olut
ion
f
ound
a
c
r
os
s
a
ll
i
ter
a
ti
ons
.
T
he
dis
c
ove
r
y
of
pbe
s
t
a
nd
gbe
s
t
is
de
ter
m
ined
by
(
1)
[
31]
.
{
⃗
⇐
⃗
+
⃗
⃗
⃗
(
0
,
1
)
⊗
(
⃗
−
⃗
)
+
⃗
⃗
⃗
(
0
,
2
)
⊗
(
⃗
−
⃗
)
⃗
⇐
⃗
ⅈ
+
⃗
(
1)
W
he
r
e
⃗
r
e
pr
e
s
e
nts
the
pa
r
ti
c
le's
ve
locity
,
⃗
r
e
pr
e
s
e
nts
the
pa
r
ti
c
le's
pos
it
ion,
⃗
⃗
⃗
(
0
,
ⅈ
)
s
igni
f
ies
a
s
e
que
nc
e
of
unif
or
ml
y
dis
tr
ibut
e
d
r
a
ndom
numbe
r
s
be
twe
e
n
0
a
nd
,
in
f
r
e
s
hly
ge
ne
r
a
ted
f
or
e
a
c
h
pa
r
ti
c
le
a
t
e
ve
r
y
it
e
r
a
ti
on,
⊗
de
notes
the
ope
r
a
ti
on
o
f
mul
ti
plyi
ng
e
leme
nts
c
or
r
e
s
pondingl
y
[
32]
.
T
o
e
ns
ur
e
ba
lanc
e
,
e
a
c
h
ve
locity
ve
c
tor
c
omponent
⃗
is
c
onf
ined
withi
n
the
mi
nim
um
a
nd
maximum
ve
locity
thr
e
s
holds
,
dono
ted
a
s
[
−
,
+
]
[
33]
.
P
s
e
udoc
ode
1.
P
a
r
ti
c
le
s
wa
r
m
opti
mi
z
a
ti
on
a
lgo
r
it
hm
[
33]
1)
S
e
t
up
a
n
a
r
r
a
y
o
f
pa
r
ti
c
les
with
r
a
ndom
c
oor
dinat
e
s
a
nd
s
pe
e
ds
a
c
r
os
s
‘
D’
dim
e
ns
ions
.
2)
B
e
gin
it
e
r
a
ti
on.
3)
F
or
e
a
c
h
pa
r
ti
c
le
with
in
the
it
e
r
a
ti
on
,
de
ter
mi
ne
t
he
va
lue
of
the
tar
ge
ted
opti
mi
z
a
ti
on
f
unc
ti
on
in
‘
D’
va
r
iable
s
.
4)
As
s
e
s
s
the
f
it
ne
s
s
of
the
pa
r
ti
c
le
a
nd
c
ompar
e
it
with
it
s
be
s
t
r
e
c
or
de
d
pos
it
ion
(
pbe
s
t)
.
I
f
the
ne
w
e
va
luation
is
s
upe
r
ior
,
upda
te
pbe
s
t
to
thi
s
ne
we
r
mea
s
ur
e
me
nt
a
nd
r
e
c
or
d
the
pa
r
ti
c
le’
s
c
ur
r
e
nt
c
oor
dinate
s
a
s
it
s
be
s
t
s
pot
withi
n
the
‘
D’
d
im
e
ns
ional
gr
id.
5)
R
e
c
o
gn
ize
t
he
mos
t
s
u
c
c
e
s
s
f
ul
pa
r
t
ic
le
in
t
he
v
ic
in
i
ty
a
n
d
a
s
s
ig
n
t
he
in
de
x
o
f
th
is
p
a
r
t
ic
le
t
o
a
va
r
iab
le
.
6)
M
odif
y
e
a
c
h
pa
r
ti
c
le’
s
mot
ion
a
nd
loca
ti
on
us
i
ng
(
1)
,
whic
h
incor
por
a
tes
the
be
s
t
pos
it
ions
identif
ie
d
by
the
indi
vidual
pa
r
t
icle
a
nd
it
s
ne
ighbo
r
s
.
7)
P
e
r
s
is
t
with
the
it
e
r
a
ti
on
unti
l
a
c
e
r
tain
-
r
e
quir
e
me
nts
is
f
ulf
il
led,
whic
h
c
ould
be
a
n
a
c
c
e
ptable
leve
l
of
f
it
ne
s
s
or
a
c
e
il
ing
on
it
e
r
a
ti
on
c
ounts
.
8)
T
e
r
mi
na
te
the
it
e
r
a
ti
on
loop
.
I
n
the
c
ontext
o
f
f
e
a
tur
e
s
e
lec
ti
on,
the
P
S
O
a
lgor
i
thm
is
de
s
igned
to
f
ind
a
n
opti
mal
s
ubs
e
t
of
f
e
a
t
ur
e
s
that
im
pr
ove
s
the
model's
pe
r
f
or
manc
e
by
r
e
duc
ing
the
da
ta's
dim
e
ns
ionalit
y
while
maintaining
or
e
n
ha
nc
ing
c
las
s
if
ica
ti
on
a
c
c
ur
a
c
y
[
34]
.
2.
5.
2.
Ant
c
olon
y
op
t
i
m
izat
ion
Optim
iza
ti
on
tec
hnique
ins
pir
e
d
by
the
f
or
a
ging
be
ha
vior
of
a
nts
.
As
a
nts
s
e
a
r
c
h
f
or
f
ood
,
they
lea
ve
phe
r
omone
tr
a
il
s
that
s
e
r
ve
a
s
a
r
oute
to
guide
them
ba
c
k
to
the
ne
s
t
[
31]
.
T
he
number
of
a
nts
tr
a
ve
ll
ing
thr
ough
that
pa
th
in
f
luenc
e
s
the
de
ns
it
y
of
phe
r
omone
de
pos
it
ion
a
nd
e
va
por
a
ti
on.
T
he
qua
li
ty
a
nd
qua
nti
ty
of
f
ood
br
ought
by
the
a
nts
a
ls
o
a
f
f
e
c
t
phe
r
omone
de
pos
it
ion.
T
he
r
e
f
or
e
,
the
a
nts
c
a
n
identif
y
the
opti
mal
pa
th
by
f
ol
lowing
the
tr
a
il
wi
th
the
maximum
phe
r
omone
de
ns
it
y
.
T
he
dis
c
ove
r
y
o
f
the
opti
mal
pa
t
h
(
2)
a
nd
the
upda
te
of
the
phe
r
omone
(
3)
by
the
a
nts
a
r
e
de
t
e
r
mi
ne
d
by
the
f
oll
owing
e
qua
ti
ons
[
31]
.
(
)
=
[
]
⋅
[
(
)
]
∈
(
)
[
]
[
(
)
]
∀
∈
(
)
(
2)
←
(
1
−
)
+
⋅
{
←
}
⋅
(
)
(
3)
He
r
e
is
the
e
xplana
ti
on
of
phe
r
omone
de
pos
it
ion
in
whic
h
it
r
e
pr
e
s
e
nts
the
phe
r
omone
de
pos
it
ion
a
t
the
ⅈ
ℎ
node
.
is
a
n
opti
ona
l
we
ighi
ng
f
unc
ti
on,
r
e
p
r
e
s
e
nts
e
a
c
h
f
e
a
s
ibl
e
s
olut
ion,
a
nd
a
r
e
pos
it
ive
pa
r
a
mete
r
.
On
the
other
ha
nd,
phe
r
o
mone
upda
ti
o
n
is
the
s
olut
ion
us
e
d
f
or
phe
r
omone
upda
te.
is
the
we
ight
of
s
olut
ion
,
is
the
e
va
por
a
ti
on
c
ons
tant,
(
)
is
the
qua
li
ty
f
unc
ti
on
.
B
a
s
e
d
on
thes
e
e
qua
ti
ons
,
P
s
e
udoc
ode
2
is
the
AC
O
a
lgor
it
hm
.
P
s
e
udoc
ode
2.
Ant
c
olony
opti
mi
z
a
ti
on
a
lgo
r
it
hm
[
35]
1)
I
nit
ialize
phe
r
omone
t
r
a
il
s
2)
W
hil
e
(
ter
mi
na
ti
on
c
r
it
e
r
ia
not
met
)
do
3)
F
or
e
a
c
h
a
nt
4)
B
uil
d
a
s
olut
ion
pa
th
ba
s
e
d
on
phe
r
omone
t
r
a
il
s
a
n
d
he
ur
is
ti
c
inf
o
r
mation
(
2)
5)
C
a
lcula
te
the
f
it
ne
s
s
of
the
s
olut
ion
6)
Upda
te
the
loca
l
phe
r
omone
tr
a
il
(
3)
7)
E
nd
it
e
r
a
ti
on
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
2025
:
275
3
-
2764
2758
8)
Upda
te
the
global
phe
r
omone
tr
a
il
ba
s
e
d
on
the
be
s
t
s
olut
ion
f
ound
9)
E
nd
I
n
f
e
a
tur
e
s
e
lec
ti
on,
AC
O
is
us
e
d
to
f
ind
a
s
ubs
e
t
of
f
e
a
tur
e
s
that
yields
the
be
s
t
pe
r
f
o
r
manc
e
f
or
the
model
by
mi
mi
c
king
how
a
nts
f
ind
the
s
hor
tes
t
p
a
th
f
r
om
the
ne
s
t
to
a
f
ood
s
our
c
e
[
19]
.
Vi
r
tual
a
nts
it
e
r
a
te
thr
ough
the
f
e
a
tur
e
s
,
c
ons
tr
uc
ti
ng
s
olut
ions
by
s
e
lec
ti
ng
f
e
a
tur
e
s
ba
s
e
d
on
pr
oba
bil
it
ies
inf
luenc
e
d
by
phe
r
omone
leve
ls
.
T
his
inc
r
e
a
s
e
s
the
li
ke
li
hood
o
f
s
e
lec
ti
ng
f
e
a
tur
e
s
that
c
ontr
ibut
e
pos
it
ively
to
i
mpr
ove
d
c
las
s
if
ica
ti
on
a
c
c
ur
a
c
y
[
36]
.
2.
5.
3.
Cat
s
war
m
op
t
i
m
izat
ion
Optim
iza
t
ion
tec
hn
ique
ins
pi
r
e
d
by
t
he
be
ha
vio
r
of
c
a
ts
,
u
ti
li
z
i
ng
tw
o
s
pe
c
i
f
ic
be
ha
v
io
r
s
kno
wn
a
s
the
s
e
e
ki
ng
mo
de
a
nd
t
r
a
c
ing
mod
e
[
11]
.
E
a
c
h
c
a
t
ha
s
a
pos
i
ti
o
n
c
ons
is
ti
n
g
of
M
dim
e
ns
ions
in
t
h
e
s
e
a
r
c
h
s
pa
c
e
,
w
he
r
e
e
a
c
h
dim
e
ns
ion
ha
s
i
ts
ve
l
oc
it
y.
T
he
f
i
tnes
s
va
l
ue
r
e
pr
e
s
e
nts
h
ow
we
ll
a
s
e
t
o
f
s
ol
uti
o
ns
(
c
a
ts
)
pe
r
f
or
ms
.
Ad
dit
iona
ll
y
,
the
r
e
is
a
f
la
g
us
e
d
to
c
las
s
if
y
c
a
ts
in
to
s
e
e
kin
g
mode
o
r
tr
a
c
ing
m
ode
.
T
he
wor
k
ing
pr
in
c
ipl
e
o
f
C
S
O
in
volves
de
te
r
m
ini
n
g
th
e
nu
mbe
r
of
c
a
ts
inv
olve
d
i
n
e
a
c
h
it
e
r
a
t
ion
a
nd
r
un
ning
the
m
thr
o
ugh
the
a
l
go
r
it
hm
.
T
he
be
s
t
c
a
t
i
n
e
a
c
h
i
ter
a
ti
o
n
is
s
tor
e
d
in
mem
or
y,
a
nd
the
c
a
t
in
the
f
i
na
l
it
e
r
a
t
ion
r
e
p
r
e
s
e
nts
the
f
ina
l
s
ol
uti
on
.
T
he
C
S
O
a
lgo
r
it
hm
a
im
s
to
f
in
d
t
he
opti
mal
s
o
lut
ion
in
the
s
e
a
r
c
h
s
pa
c
e
b
y
u
ti
l
izin
g
t
he
s
e
e
king
a
nd
tr
a
c
ing
be
h
a
vio
r
s
ins
p
ir
e
d
b
y
c
a
ts
.
I
n
the
s
e
e
kin
g
mode
,
c
a
ts
r
a
ndo
ml
y
e
xp
lo
r
e
or
o
bs
e
r
ve
the
ir
s
ur
r
o
und
ings
to
f
i
nd
be
tt
e
r
pos
i
ti
ons
.
On
the
othe
r
ha
nd
,
i
n
the
tr
a
c
ing
mo
de
,
c
a
ts
mo
ve
t
owa
r
ds
t
he
ta
r
ge
t
wi
th
a
ma
thema
ti
c
a
l
ly
c
a
lc
ulate
d
ve
loc
it
y
.
T
he
t
r
a
c
in
g
mode
C
S
O
a
l
gor
it
h
m
is
e
xp
r
e
s
s
e
d
in
(
4
)
a
nd
(
5
)
[
3
7]
.
=
(
1
+
∙
)
∙
(
4)
ⅈ
=
|
−
|
−
,
ℎ
0
<
ⅈ
<
(
5)
T
he
s
e
e
king
mode
o
f
C
S
O
m
im
ics
the
r
e
s
ti
ng
be
h
a
vior
o
f
c
a
ts
.
T
he
r
e
a
r
e
f
our
im
po
r
tant
pa
r
a
mete
r
s
in
thi
s
mode:
s
e
e
king
memor
y
poo
l
(
S
M
P
)
,
s
e
e
king
r
a
nge
of
the
s
e
lec
ted
dim
e
ns
ion
(
S
R
D)
,
c
ount
of
dim
e
ns
ions
to
c
ha
nge
(
C
D
C
)
,
a
nd
s
e
lf
-
pos
it
ion
c
ons
ider
ing
(
S
P
C
)
,
whic
h
a
r
e
manua
ll
y
s
e
t
va
lues
.
I
n
e
a
c
h
it
e
r
a
ti
on
of
C
S
O,
r
a
ndoml
y
s
e
lec
t
C
DC
dim
e
ns
io
ns
to
be
mut
a
ted.
Add
or
s
ubtr
a
c
t
a
r
a
ndom
va
lu
e
withi
n
S
R
D
f
r
om
the
c
ur
r
e
nt
va
lue,
r
e
plac
ing
the
old
pos
it
ion
with
the
ne
w
pos
it
ion
,
a
s
s
hown
in
(
4)
.
He
r
e
,
r
e
pr
e
s
e
nts
the
ne
xt
pos
it
ion,
de
notes
the
c
a
t
index,
mea
ns
the
dim
e
ns
ion,
a
nd
is
a
r
a
ndom
number
in
the
int
e
r
va
l
be
twe
e
n
0
a
nd
1.
B
a
s
e
d
on
pr
oba
bil
it
ies
,
s
e
lec
t
one
of
the
c
a
ndidate
point
s
to
be
the
f
o
ll
owing
pos
it
ion
f
o
r
the
c
a
t.
T
he
c
a
ndidate
point
s
with
a
highe
r
f
it
ne
s
s
va
lue
a
r
e
mor
e
l
ikely
to
be
c
hos
e
n,
a
s
s
hown
in
(
5)
.
How
e
ve
r
,
if
a
ll
f
it
ne
s
s
va
lues
a
r
e
e
qua
l,
s
e
t
the
pr
oba
bil
it
y
of
s
e
lec
ti
ng
e
a
c
h
c
a
ndidate
point
to
1
.
I
f
the
goa
l
is
mi
nim
iza
ti
on,
s
e
t
=
other
wis
e
,
if
the
goa
l
is
maximi
z
a
ti
on,
s
pe
c
if
y
=
.
P
s
e
udoc
ode
3
is
the
C
S
O
a
lgor
it
hm
in
s
e
e
king
mode.
T
he
s
e
e
king
mode
C
S
O
a
lgor
it
hm
is
e
xpr
e
s
s
e
d
in
(
6)
a
nd
(
7)
[
37
]
:
,
=
,
+
1
1
(
,
−
,
)
,
ℎ
=
1
,
2
,
.
.
.
,
(
6)
,
=
,
+
,
(
7)
P
s
e
udoc
ode
3.
C
a
t
s
wa
r
m
opti
mi
z
a
ti
on
a
lgor
it
hm
i
n
s
e
e
king
mode
[
37]
1)
C
r
e
a
te
ins
tanc
e
s
of
the
c
ur
r
e
nt
pos
it
ion
of
the
c
a
t,
whe
r
e
=
.
I
f
S
P
C
is
a
t
r
ue
c
ondit
ion
,
let
=
(
−
1
)
,
maintain
the
c
ur
r
e
nt
pos
it
ion
a
s
a
n
op
ti
on
a
mong
the
pos
s
ibl
e
c
a
ndidate
s
.
2)
F
or
e
a
c
h
ins
tanc
e
,
a
c
c
or
ding
to
C
DC
,
r
a
ndoml
y
in
c
r
e
a
s
e
or
de
c
r
e
a
s
e
S
R
D
pe
r
c
e
nts
of
the
e
xis
ti
ng
va
lues
a
nd
r
e
plac
e
the
f
o
r
mer
one
s
.
3)
C
a
lcula
te
the
f
it
ne
s
s
va
lues
of
a
ll
c
a
ndidate
point
s
.
4)
I
f
the
c
a
s
e
whe
r
e
not
a
ll
f
it
ne
s
s
va
lue
a
r
e
identica
l,
c
a
lcula
te
the
s
e
lec
ti
ng
pr
oba
bil
it
y
of
e
a
c
h
c
a
ndi
da
te
point
by
(
5)
o
ther
wis
e
s
e
t
a
ll
the
s
e
lec
ti
ng
p
r
oba
bil
it
y
of
e
a
c
h
c
a
ndidate
point
be
1
.
5)
R
a
ndoml
y
pick
the
point
to
move
to
f
r
om
the
c
a
nd
idate
point
s
,
a
nd
r
e
plac
e
the
pos
it
ion
o
f
the
c
a
t
.
I
n
the
f
ir
s
t
it
e
r
a
ti
on
o
f
the
tr
a
c
ing
mode,
the
ve
locity
va
lues
a
r
e
r
a
ndoml
y
a
s
s
igned
f
or
a
ll
dim
e
ns
ions
of
the
c
a
t's
pos
it
ion.
How
e
ve
r
,
the
ve
locity
va
lues
ne
e
d
to
be
upda
ted
f
o
r
e
a
c
h
di
mens
ion
a
c
c
or
ding
to
(
6
)
f
or
the
s
ubs
e
que
nt
s
teps
.
I
f
the
ve
locity
e
xc
e
e
ds
the
maximum
a
ll
owe
d
va
lue
,
it
is
s
e
t
to
the
maximum
ve
locity
.
Upda
te
the
c
a
t's
pos
it
ion
ba
s
e
d
on
(
7)
.
P
s
e
udoc
ode
4
is
the
C
S
O
a
lgor
it
h
m
i
n
tr
a
c
ing
mode
r
e
f
e
r
r
ing
to
(
6)
a
nd
(
7)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
Optimiz
ing
long
s
hor
t
-
ter
m
me
mor
y
hy
pe
r
pa
r
ame
t
e
r
for
c
r
y
ptocur
r
e
nc
y
s
e
nti
me
nt
…
(
K
r
is
ti
an
E
k
ac
handr
a
)
2759
P
s
e
udoc
ode
4.
C
a
t
s
wa
r
m
opti
mi
z
a
ti
on
a
lgor
it
hm
i
n
tr
a
c
ing
mode
[
38]
1)
Upda
te
the
ve
lociti
e
s
f
or
e
ve
r
y
dim
e
ns
ion
,
a
s
s
hown
in
(
6
)
.
2)
Ve
r
if
y
that
the
ve
lociti
e
s
a
r
e
withi
n
the
maxi
mum
a
ll
owe
d
ve
lociti
e
s
r
a
nge
.
Adjus
t
a
ny
ve
l
oc
it
y
e
xc
e
e
ding
thi
s
r
a
nge
ba
c
k
to
the
maximum
li
mi
t.
3)
Adjus
t
the
loca
ti
on
of
the
c
a
t
,
a
c
c
or
ding
to
(
7)
.
B
a
s
e
d
on
two
modes
in
the
C
S
O
a
lgor
it
h
m,
na
mely,
s
e
e
king
mode
a
nd
t
r
a
c
ing
mode.
P
s
e
udoc
ode
5
is
the
c
ombi
na
ti
on
of
the
two
modes
.
P
s
e
udoc
ode
5.
C
a
t
s
wa
r
m
opti
mi
z
a
ti
on
a
lgor
it
hm
[
38]
1)
I
nit
ialize
by
c
r
e
a
ti
ng
‘
N’
c
a
ts
in
the
a
lgor
it
hm
2)
P
lac
e
the
c
a
ts
r
a
ndoml
y
withi
n
a
n
‘
M
’
-
dim
e
ns
ional
s
e
a
r
c
h
a
r
e
a
,
a
s
s
ig
ning
ve
lociti
e
s
that
a
r
e
withi
n
the
pr
e
de
f
ined
maximum
bounds
.
R
a
ndoml
y
de
ter
mi
n
e
a
number
o
f
c
a
ts
to
e
nga
ge
in
tr
a
c
ing
mode
ba
s
e
d
on
the
M
ixi
ng
R
a
ti
o
(
M
R
)
,
pos
it
ioni
ng
the
r
e
s
t
in
s
e
e
king
mode.
3)
C
omput
e
the
f
it
ne
s
s
f
o
r
e
a
c
h
c
a
t
us
ing
the
f
it
n
e
s
s
f
unc
ti
on
w
hich
mea
s
ur
e
s
their
p
r
oxim
it
y
to
the
objec
ti
ve
,
a
nd
memor
ize
the
loca
ti
on
of
the
mos
t
o
pti
mal
c
a
t
.
4)
R
e
loca
te
the
c
a
ts
ba
s
e
d
on
their
a
s
s
igned
modes
:
thos
e
in
s
e
e
king
mode
unde
r
go
a
di
f
f
e
r
e
nt
pr
oc
e
s
s
,
while
thos
e
in
tr
a
c
ing
mode
a
djus
t
thei
r
ve
loc
ity
a
nd
pos
it
ion
a
c
c
or
ding
to
s
pe
c
if
ic
f
o
r
mul
a
s
.
5)
S
e
lec
ti
ve
ly
s
witch
a
number
of
the
c
a
ts
ba
c
k
to
tr
a
c
ing
mode
a
s
pe
r
the
M
R
,
a
nd
the
r
e
mainde
r
c
onti
nue
in
s
e
e
king
mode.
6)
C
he
c
k
if
the
e
nd
c
ondit
ions
of
the
a
lgo
r
it
hm
ha
v
e
be
e
n
met;
if
s
o,
s
top
the
a
lgor
it
hm
,
other
wis
e
c
yc
le
thr
ough
s
teps
3
to
5
a
ga
in.
I
n
the
c
ontext
of
f
e
a
tur
e
s
e
lec
ti
on,
the
C
S
O
a
lgo
r
it
hm
ope
r
a
ted
by
e
xplor
ing
the
f
e
a
tur
e
s
pa
c
e
to
dis
c
ove
r
the
opti
mal
f
e
a
tur
e
c
ombi
na
ti
ons
.
T
h
r
ou
gh
it
e
r
a
ti
ons
be
twe
e
n
the
s
e
e
king
a
nd
tr
a
c
ing
mo
de
s
,
C
S
O
a
da
pti
ve
ly
e
xplor
e
d
a
nd
uti
l
ize
d
inf
or
mation
f
r
om
the
f
e
a
tur
e
s
pa
c
e
to
identif
y
f
e
a
tur
e
s
ubs
e
ts
that
pr
ovided
the
be
s
t
pe
r
f
or
manc
e
f
or
the
us
e
d
model
.
I
n
t
his
pr
oc
e
s
s
,
the
a
lgor
it
h
m
a
im
e
d
to
ba
lanc
e
e
x
plor
a
ti
on
(
s
e
a
r
c
hing
f
or
di
f
f
e
r
e
nt
f
e
a
tur
e
c
ombi
na
ti
ons
)
a
nd
e
x
ploi
tation
(
f
o
ll
owing
pr
om
is
ing
pos
it
ions
)
to
o
btain
a
n
opti
mal
s
olut
ion.
2.
6.
Op
t
im
izin
g
lon
g
s
h
or
t
-
t
e
r
m
m
e
m
or
y
u
s
in
g
h
yp
e
r
p
ar
am
e
t
e
r
t
u
n
i
n
g
T
he
hype
r
pa
r
a
mete
r
tuni
ng
pr
oc
e
s
s
wa
s
s
pe
c
if
ica
ll
y
dir
e
c
ted
towa
r
ds
opti
mi
z
ing
the
L
S
T
M
model
to
im
p
r
ove
it
s
pe
r
f
or
manc
e
in
s
e
nti
ment
a
na
lys
is
.
T
he
p
r
im
a
r
y
hype
r
pa
r
a
mete
r
a
djus
ted
in
thi
s
s
tudy
wa
s
the
number
of
L
S
T
M
unit
s
,
whic
h
de
ter
mi
ne
s
the
di
mens
ional
it
y
of
the
c
e
ll
s
tate
withi
n
the
model
a
nd
dir
e
c
tl
y
inf
luenc
e
s
it
s
a
bil
it
y
to
c
a
ptur
e
s
e
que
nti
a
l
de
pe
nde
nc
ies
in
the
da
ta.
An
L
S
T
M
model
wa
s
c
onf
i
gur
e
d
to
c
las
s
if
y
the
s
e
nti
ment
of
the
pr
e
p
r
oc
e
s
s
e
d
c
r
yptocur
r
e
nc
y
-
r
e
late
d
da
ta.
T
he
opt
im
ize
d
model
c
onf
ig
ur
a
ti
ons
de
ter
mi
ne
d
by
P
S
O,
AC
O,
a
nd
C
S
O
we
r
e
c
o
mpar
e
d
a
ga
ins
t
a
ba
s
e
li
ne
L
S
T
M
without
op
ti
mi
z
a
ti
on.
E
va
luation
metr
ics
include
d
a
c
c
ur
a
c
y,
los
s
,
a
nd
e
xe
c
uti
on
ti
me,
whic
h
pr
ovide
ins
ight
s
int
o
the
model’
s
e
f
f
e
c
ti
ve
ne
s
s
a
nd
e
f
f
icie
nc
y.
T
he
L
S
T
M
model’
s
pe
r
f
or
m
a
nc
e
wa
s
e
va
luate
d
us
ing
a
c
c
ur
a
c
y,
los
s
,
a
nd
e
xe
c
uti
on
ti
me.
Ac
c
ur
a
c
y
mea
s
ur
e
s
the
pe
r
c
e
ntage
of
c
o
r
r
e
c
t
p
r
e
dictions
;
los
s
indi
c
a
tes
model
c
on
ve
r
ge
nc
e
a
nd
e
xe
c
uti
on
ti
me
a
s
s
e
s
s
e
s
c
omput
a
ti
ona
l
e
f
f
icie
nc
y.
E
a
c
h
metr
ic
wa
s
c
ompar
e
d
a
c
r
os
s
the
P
S
O
-
L
S
T
M
,
AC
O
-
L
S
T
M
,
C
S
O
-
L
S
T
M
,
a
nd
ba
s
e
li
ne
L
S
T
M
models
to
de
te
r
mi
ne
the
mos
t
e
f
f
e
c
ti
ve
opti
mi
z
a
ti
on
tec
hnique.
T
he
opti
mal
c
on
f
igur
a
ti
on
o
f
the
m
ode
l
wa
s
de
ter
mi
ne
d
thr
ough
hype
r
pa
r
a
mete
r
t
uning,
a
s
s
umm
a
r
ize
d
in
T
a
ble
2
.
T
a
ble
2.
Hype
r
pa
r
a
mete
r
tun
ing
A
lg
or
it
hm
P
a
r
a
me
te
r
s
V
a
lu
e
s
L
S
T
M
E
mbe
ddi
ng i
nput
di
me
ns
io
n
7818
E
mbe
ddi
ng output
di
me
ns
io
n
300
E
mbe
ddi
ng i
nput
l
e
ngt
h
25
L
S
T
M
uni
t
256 (
opt
im
iz
e
d by e
a
c
h s
w
a
r
m i
nt
e
ll
ig
e
nc
e
a
lg
or
it
hms
)
L
S
T
M
dr
opout
0.2
L
S
T
M
r
e
c
ur
r
e
nt
dr
opout
0.2
D
e
ns
e
c
la
s
s
e
s
2
D
e
ns
e
a
c
ti
va
ti
on
s
ig
moi
d
O
pt
im
iz
e
r
A
da
m
P
S
O
, A
C
O
, C
S
O
n_pa
r
ti
c
le
, n_a
nt
s
, n_c
a
ts
15
num_i
te
r
a
ti
ons
50
lb
;
ub
16;
256
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
2025
:
275
3
-
2764
2760
3.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
3.
1.
E
xp
e
r
im
e
n
t
al
r
e
s
u
lt
s
T
a
ble
3
pr
e
s
e
nts
a
c
ompar
is
on
o
f
e
a
c
h
L
S
T
M
model
c
onf
igu
r
a
ti
on,
including
a
s
tanda
r
d
L
S
T
M
model
without
opti
mi
z
a
ti
on
a
nd
L
S
T
M
models
opti
mi
z
e
d
us
ing
P
S
O
,
AC
O,
a
nd
C
S
O
a
lgor
i
th
ms
.
E
a
c
h
c
onf
igur
a
ti
on
ba
s
e
li
ne
,
P
S
O
-
L
S
T
M
,
AC
O
-
L
S
T
M
,
a
nd
C
S
O
-
L
S
T
M
wa
s
e
va
luate
d
ba
s
e
d
on
a
c
c
ur
a
c
y,
los
s
,
a
nd
e
xe
c
uti
on
ti
me.
T
he
r
e
s
ult
s
o
f
thi
s
tuni
ng
pr
oc
e
s
s
,
guided
by
e
a
c
h
s
wa
r
m
int
e
ll
igenc
e
a
lgor
it
hm,
de
mons
tr
a
ted
that
a
djus
ti
ng
the
L
S
T
M
unit
s
us
in
g
s
wa
r
m
-
ba
s
e
d
opti
mi
z
a
ti
on
s
igni
f
ica
ntl
y
im
pr
ove
d
model
pe
r
f
or
manc
e
.
E
a
c
h
a
lgor
it
hm
p
r
ovided
a
unique
p
e
r
s
pe
c
ti
ve
on
opti
mal
L
S
T
M
unit
s
e
lec
ti
on,
r
e
f
lec
ti
ng
the
s
tr
e
ngths
of
s
wa
r
m
int
e
ll
igenc
e
in
hype
r
pa
r
a
mete
r
opti
mi
z
a
ti
on.
T
a
ble
3.
C
ompar
is
on
of
s
e
nti
ment
c
las
s
if
ica
ti
on
m
ode
ls
L
S
T
M
opt
im
iz
e
r
N
um L
S
T
M
uni
t
L
os
s
A
c
c
ur
a
c
y
E
xe
c
ut
io
n t
im
e
(
s
)
-
256
0.930019
0.853225
139.465759
PSO
16
0.570487
0.860843
58.430918
A
C
O
16
0.602706
0.853225
56.374625
C
S
O
29
0.662189
0.859319
65.443925
As
s
hown
in
T
a
ble
3,
the
P
S
O
-
opti
mi
z
e
d
L
S
T
M
a
c
hieve
d
the
highes
t
a
c
c
ur
a
c
y
a
t
86.
08%
with
the
lowe
s
t
los
s
va
lue
of
0.
570487
a
nd
a
r
e
latively
low
e
xe
c
uti
on
ti
me
of
58.
43
s
e
c
onds
.
C
ompar
e
d
to
the
ba
s
e
li
ne
L
S
T
M
model
without
opt
im
iza
ti
on,
whic
h
ha
s
a
los
s
va
lu
e
of
0
.
930019
a
nd
a
n
a
c
c
ur
a
c
y
of
85.
32%
,
the
PSO
-
L
S
T
M
de
mons
tr
a
tes
im
pr
ove
d
pe
r
f
or
manc
e
,
e
s
pe
c
ially
in
ter
ms
of
a
c
c
ur
a
c
y
a
nd
e
f
f
icie
nc
y.
T
he
P
S
O
opti
mi
z
e
r
e
f
f
e
c
ti
ve
ly
identi
f
ied
a
c
onf
igur
a
t
ion
with
only
16
L
S
T
M
unit
s
,
ther
e
by
ba
lanc
in
g
model
pe
r
f
or
manc
e
a
nd
e
xe
c
uti
on
ti
me.
T
he
AC
O
a
nd
C
S
O
a
lgor
it
hms
a
ls
o
a
c
hieve
d
c
ompar
a
ble
r
e
s
ult
s
,
though
their
a
c
c
ur
a
c
y
a
nd
los
s
va
lues
we
r
e
s
li
ghtl
y
low
e
r
than
thos
e
of
P
S
O.
Ne
ve
r
thele
s
s
,
both
AC
O
a
nd
C
S
O
s
igni
f
ica
ntl
y
r
e
duc
e
d
e
xe
c
uti
on
ti
me
r
e
lative
to
th
e
non
-
opti
mi
z
e
d
model,
de
mons
tr
a
ti
ng
the
e
f
f
e
c
ti
ve
ne
s
s
of
s
wa
r
m
int
e
ll
igenc
e
a
lgor
it
hms
in
a
c
c
e
ler
a
ti
ng
the
t
r
a
ini
ng
pr
oc
e
s
s
.
3.
2.
M
od
e
l
p
e
r
f
or
m
an
c
e
an
alys
is
T
he
pe
r
f
or
manc
e
of
the
P
S
O
-
L
S
T
M
model
dur
ing
tr
a
ini
ng
a
nd
va
li
da
ti
on
is
s
hown
in
F
igur
e
2.
T
he
lef
t
plot
i
ll
us
tr
a
tes
the
model
a
c
c
ur
a
c
y
a
c
r
os
s
e
poc
hs
,
with
the
tr
a
in
ing
a
c
c
ur
a
c
y
incr
e
a
s
ing
to
ne
a
r
l
y
100%
by
the
e
nd
of
the
tr
a
ini
ng
e
poc
hs
.
T
he
va
li
da
ti
on
a
c
c
ur
a
c
y
s
tabili
z
e
s
a
r
ound
85%
,
indi
c
a
ti
ng
a
potent
ial
is
s
ue
with
ove
r
f
it
ti
ng.
T
his
r
e
s
ult
s
ugge
s
ts
that
th
e
mo
de
l
mi
ght
be
lea
r
ning
f
e
a
tur
e
s
s
pe
c
if
ic
to
the
tr
a
i
ning
da
ta
that
do
not
ge
ne
r
a
li
z
e
we
ll
to
uns
e
e
n
da
ta,
whic
h
c
ould
im
pa
c
t
it
s
e
f
f
e
c
ti
ve
ne
s
s
in
r
e
a
l
-
wor
ld
a
ppli
c
a
ti
ons
.
F
igur
e
2.
M
ode
l
a
c
c
ur
a
c
y
a
nd
model
los
s
of
L
S
T
M
us
ing
the
P
S
O
a
lgo
r
it
hm
T
he
r
ight
plot
s
hows
the
tr
a
ini
ng
a
nd
va
li
da
ti
on
los
s
ove
r
the
e
poc
hs
,
whe
r
e
the
tr
a
ini
ng
los
s
de
c
r
e
a
s
e
s
r
a
pidl
y,
ne
a
r
ing
z
e
r
o
by
the
f
inal
e
poc
h.
I
n
c
ont
r
a
s
t,
the
va
li
da
ti
on
los
s
ini
ti
a
ll
y
de
c
r
e
a
s
e
s
but
then
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
Optimiz
ing
long
s
hor
t
-
ter
m
me
mor
y
hy
pe
r
pa
r
ame
t
e
r
for
c
r
y
ptocur
r
e
nc
y
s
e
nti
me
nt
…
(
K
r
is
ti
an
E
k
ac
handr
a
)
2761
be
gins
to
r
is
e
s
li
ghtl
y
a
f
ter
the
thi
r
d
e
poc
h.
T
hi
s
pa
tt
e
r
n
of
incr
e
a
s
ing
va
li
da
ti
on
los
s
a
longs
ide
de
c
r
e
a
s
ing
tr
a
ini
ng
los
s
indi
c
a
tes
that
the
model
may
be
mem
or
izing
the
t
r
a
ini
ng
da
ta
r
a
ther
than
lea
r
ning
ge
ne
r
a
li
z
a
ble
pa
tt
e
r
ns
.
T
o
a
ddr
e
s
s
thi
s
is
s
ue
in
f
utur
e
e
xpe
r
im
e
nts
,
r
e
gular
iza
ti
on
tec
hniques
s
uc
h
a
s
dr
opout
c
ould
be
e
mpl
oye
d
to
pr
e
ve
nt
ove
r
f
it
ti
ng
a
nd
im
p
r
ove
the
m
ode
l's
r
obus
tnes
s
.
3.
3.
Conf
u
s
ion
m
a
t
r
ix
an
d
c
las
s
if
icat
ion
m
e
t
r
ic
s
F
igur
e
3
pr
e
s
e
nts
the
c
onf
us
ion
matr
ix
f
or
the
P
S
O
-
L
S
T
M
model,
whic
h
r
e
ve
a
ls
that
the
model
c
or
r
e
c
tl
y
c
las
s
if
ied
831
ins
tanc
e
s
a
s
ne
ga
ti
ve
a
nd
864
ins
tanc
e
s
a
s
pos
it
ive.
How
e
ve
r
,
it
a
ls
o
pr
oduc
e
d
147
f
a
ls
e
pos
it
ives
a
nd
127
f
a
ls
e
ne
ga
ti
ve
s
.
T
h
e
s
e
r
e
s
ult
s
a
ll
ow
us
to
c
a
lcula
te
im
por
tant
c
las
s
if
ica
ti
on
metr
ics
that
e
va
luate
the
model's
pe
r
f
or
manc
e
.
T
he
ove
r
a
ll
a
c
c
ur
a
c
y
o
f
the
model
is
85.
5%
,
r
e
f
le
c
ti
ng
it
s
a
bil
it
y
to
c
or
r
e
c
tl
y
pr
e
dict
both
pos
it
ive
a
nd
ne
g
a
ti
ve
c
las
s
e
s
in
mos
t
c
a
s
e
s
.
T
he
pr
e
c
is
ion
f
or
the
pos
it
ive
c
las
s
,
whic
h
mea
s
ur
e
s
the
pr
opor
ti
on
of
c
or
r
e
c
t
pos
it
ive
pr
e
dictions
out
of
a
ll
pr
e
dicte
d
pos
it
ives
,
s
tands
a
t
a
ppr
oxim
a
tely
85.
5%
.
T
his
high
pr
e
c
is
ion
indi
c
a
t
e
s
that
the
model
is
ge
ne
r
a
ll
y
r
e
li
a
ble
in
identif
y
ing
tr
ue
pos
it
ives
,
mi
nim
izing
the
oc
c
ur
r
e
nc
e
of
f
a
ls
e
a
lar
ms
.
F
igur
e
3.
C
onf
us
ion
matr
ix
of
L
S
T
M
us
ing
the
P
S
O
a
lgor
it
hm
I
n
te
r
ms
of
r
e
c
a
ll
,
whic
h
a
s
s
e
s
s
e
s
the
model's
s
e
ns
it
ivi
ty
to
c
or
r
e
c
tl
y
identi
f
y
a
c
tual
pos
it
ive
c
a
s
e
s
,
the
P
S
O
-
L
S
T
M
model
de
mons
tr
a
tes
a
s
tr
ong
c
a
p
a
bil
it
y
in
de
tec
ti
ng
pos
it
ive
ins
tanc
e
s
.
How
e
ve
r
,
t
he
r
e
a
r
e
s
ti
ll
c
a
s
e
s
wh
e
r
e
the
model
f
a
il
s
to
c
a
ptur
e
s
ome
p
os
it
ive
e
xa
mpl
e
s
,
a
s
e
viden
c
e
d
by
the
f
a
ls
e
ne
ga
ti
v
e
s
in
the
c
onf
us
ion
matr
ix.
F
inally
,
the
F
1
-
s
c
or
e
,
whic
h
ba
lanc
e
s
pr
e
c
is
ion
a
nd
r
e
c
a
ll
int
o
a
s
ingl
e
metr
ic,
pr
ovides
a
c
ompr
e
he
ns
ive
view
of
the
model’
s
c
las
s
if
ica
ti
o
n
pe
r
f
or
manc
e
.
T
he
F
1
-
s
c
or
e
is
e
s
pe
c
ially
us
e
f
ul
in
c
a
s
e
s
whe
r
e
ther
e
is
a
tr
a
de
-
of
f
be
twe
e
n
pr
e
c
is
ion
a
nd
r
e
c
a
ll
,
a
s
it
r
e
f
lec
ts
the
model’
s
e
f
f
e
c
ti
ve
ne
s
s
in
maintaining
both
a
high
p
r
e
c
is
ion
a
nd
a
s
tr
ong
r
e
c
a
ll
.
Ove
r
a
ll
,
thes
e
metr
ics
c
onf
ir
m
that
the
P
S
O
-
L
S
T
M
model
pe
r
f
or
ms
we
ll
,
though
it
e
xhibi
ts
a
mi
nor
tende
nc
y
to
mi
s
c
las
s
if
y
c
e
r
tain
ins
tanc
e
s
,
pa
r
ti
c
ular
ly
whe
n
dis
ti
nguis
hing
be
twe
e
n
s
im
il
a
r
ne
ga
ti
ve
a
nd
pos
it
ive
c
a
s
e
s
.
3.
4.
Dis
c
u
s
s
ion
S
wa
r
m
int
e
ll
igenc
e
a
lgor
it
hms
e
f
f
e
c
ti
ve
ly
opti
mi
z
e
d
the
L
S
T
M
model
by
f
ine
-
tuni
ng
the
number
of
unit
s
in
the
L
S
T
M
laye
r
.
T
his
opti
mi
z
a
ti
on,
pa
r
ti
c
ular
ly
thr
ough
P
S
O
,
a
ll
owe
d
f
or
a
s
igni
f
ica
nt
r
e
du
c
ti
on
in
e
xe
c
uti
on
ti
me
without
c
ompr
omi
s
ing
model
a
c
c
ur
a
c
y.
S
uc
h
im
pr
ove
ments
unde
r
s
c
or
e
the
potential
of
s
wa
r
m
int
e
ll
igenc
e
in
de
e
p
lea
r
ning
a
ppli
c
a
ti
ons
,
e
s
pe
c
ially
f
or
ta
s
ks
invol
ving
high
d
im
e
ns
ional
da
ta
li
ke
c
r
yptocur
r
e
nc
y
s
e
nti
ment
a
na
lys
is
.
3.
5.
L
im
i
t
at
ion
s
an
d
i
m
p
li
c
at
ion
s
f
or
f
u
t
u
r
e
r
e
s
e
ar
c
h
W
hil
e
the
opti
m
ize
d
models
s
how
pr
o
mi
s
ing
r
e
s
ult
s
,
the
s
tudy
is
li
mi
ted
to
c
r
yptocur
r
e
nc
y
s
e
nti
ment
da
ta
a
nd
a
f
ixed
s
e
t
o
f
s
wa
r
m
int
e
ll
igen
c
e
a
lgor
it
hms
.
F
utu
r
e
r
e
s
e
a
r
c
h
c
ould
e
xpa
nd
by
in
tegr
a
ti
ng
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
4
,
Augus
t
2025
:
275
3
-
2764
2762
hybr
id
opti
m
iza
ti
on
tec
hniques
a
nd
tes
ti
ng
a
c
r
os
s
diver
s
e
s
e
nti
ment
a
na
lys
is
tas
ks
.
T
he
c
ur
r
e
nt
f
in
dings
lay
the
gr
oundwor
k
f
or
f
ur
ther
e
xplor
a
ti
on
of
s
wa
r
m
i
ntelli
ge
nc
e
in
ne
ur
a
l
n
e
two
r
k
opti
mi
z
a
ti
on
.
4.
CONC
L
USI
ON
T
his
s
tudy
e
xplor
e
s
the
e
f
f
e
c
ti
ve
ne
s
s
of
int
e
gr
a
t
ing
s
wa
r
m
int
e
ll
igenc
e
a
lgor
it
hms
na
mely
P
S
O,
AC
O,
a
nd
C
S
O
with
L
S
T
M
ne
twor
ks
f
or
s
e
nti
men
t
a
na
lys
is
tas
ks
.
E
a
c
h
of
thes
e
opti
mi
z
a
ti
on
tec
hniq
ue
s
wa
s
e
mpl
oye
d
to
f
ine
-
tune
t
he
L
S
T
M
model,
s
pe
c
if
ica
ll
y
a
djus
ti
ng
the
number
of
L
S
T
M
unit
s
to
e
nha
nc
e
pe
r
f
or
manc
e
met
r
ics
.
C
ompar
a
ti
ve
a
na
lys
is
r
e
v
e
a
ls
that
the
P
S
O
-
L
S
T
M
model
outper
f
o
r
med
both
the
AC
O
-
L
S
T
M
a
nd
C
S
O
-
L
S
T
M
models
,
a
c
hieving
t
he
highes
t
a
c
c
ur
a
c
y
of
86.
08%
a
nd
the
lowe
s
t
los
s
of
0.
57
,
a
longs
ide
the
s
hor
tes
t
e
xe
c
uti
on
ti
me
of
58.
43
s
e
c
onds
.
T
he
s
e
r
e
s
ult
s
s
ugg
e
s
t
that
P
S
O
opti
mi
z
a
ti
on
e
f
f
e
c
ti
ve
ly
e
nha
nc
e
s
L
S
T
M
model
pe
r
f
or
manc
e
,
de
li
ve
r
ing
s
upe
r
ior
a
c
c
ur
a
c
y
a
nd
f
a
s
ter
pr
oc
e
s
s
ing
ti
mes
c
ompar
e
d
to
the
other
s
wa
r
m
int
e
ll
igenc
e
a
lgor
i
t
hms
us
e
d
in
th
is
s
tudy.
T
he
a
na
lys
is
a
ls
o
unde
r
s
c
or
e
s
the
va
lue
of
us
ing
s
wa
r
m
int
e
ll
igenc
e
a
lgor
it
hms
in
d
e
e
p
lea
r
ning
c
ontexts
.
B
y
a
pplyi
ng
thes
e
opti
mi
z
a
ti
ons
,
it
wa
s
pos
s
ibl
e
to
r
e
f
ine
the
L
S
T
M
a
r
c
hit
e
c
tur
e
,
lea
d
ing
to
s
igni
f
ica
nt
i
mpr
ove
ments
in
s
e
nti
ment
c
las
s
if
ica
ti
on
a
c
c
ur
a
c
y
a
nd
c
omput
a
ti
ona
l
e
f
f
icie
nc
y.
Additi
on
a
ll
y,
the
us
e
of
the
P
S
O
a
lgor
i
thm
de
mons
tr
a
te
d
r
obus
t
pa
r
a
mete
r
tuni
ng
c
a
pa
bil
it
ies
,
p
r
ovidi
ng
a
ba
lanc
e
be
twe
e
n
model
c
ompl
e
xit
y
a
nd
a
c
c
ur
a
c
y
that
is
s
uit
a
ble
f
or
s
e
nti
ment
a
na
ly
s
is
a
ppli
c
a
ti
ons
.
Give
n
the
volatil
e
a
nd
s
e
nti
ment
-
dr
iven
na
tur
e
of
c
r
yptoc
ur
r
e
nc
y
mar
ke
ts
,
a
c
c
ur
a
te
a
nd
e
f
f
icie
nt
s
e
nti
ment
a
na
lys
is
models
a
r
e
va
luable
f
o
r
unde
r
s
tanding
publ
ic
s
e
nti
ment
a
nd
making
da
ta
-
dr
iven
p
r
e
dictions
.
T
he
int
e
gr
a
ti
on
of
P
S
O
with
L
S
T
M
ne
twor
ks
of
f
e
r
s
pr
omi
s
ing
potential
f
or
r
e
a
l
-
ti
me
c
r
yptocu
r
r
e
nc
y
s
e
nti
ment
a
na
lys
is
,
whic
h
c
ould
s
uppor
t
be
tt
e
r
de
c
is
ion
-
making
in
tr
a
ding
a
nd
inves
tm
e
nt.
F
utu
r
e
r
e
s
e
a
r
c
h
c
ould
e
xplor
e
the
a
ppl
ica
ti
on
of
other
opti
mi
z
a
ti
on
tec
hniques
to
f
u
r
ther
im
pr
ove
m
ode
l
pe
r
f
or
manc
e
in
thi
s
domain
.
AC
KNOWL
E
DGM
E
N
T
S
W
e
a
c
knowle
dge
the
s
uppor
t
r
e
c
e
ived
f
r
om
Uni
ve
r
s
it
a
s
M
ult
im
e
dia
Nus
a
ntar
a
f
r
om
c
a
r
r
ying
out
r
e
s
e
a
r
c
h
wor
k
of
thi
s
pa
pe
r
.
F
UN
DI
NG
I
NF
ORM
AT
I
ON
T
he
a
uthor
s
would
li
ke
to
thank
the
Unive
r
s
it
a
s
M
ult
im
e
dia
Nus
a
ntar
a
f
or
s
uppor
ti
ng
a
nd
f
unding
thi
s
r
e
s
e
a
r
c
h
f
unde
d
by
the
I
nter
na
l
R
e
s
e
a
r
c
h
Gr
a
n
t
Unive
r
s
it
a
s
M
ult
im
e
dia
Nus
a
nt
a
r
a
2024
with
the
c
ontr
a
c
t
number
0015
-
RD
-
L
P
P
M
-
U
M
N/P
-
I
NT
/VI
/2024.
AU
T
HO
R
CONT
RI
B
U
T
I
ONS
S
T
AT
E
M
E
N
T
T
his
jour
na
l
us
e
s
the
C
ontr
i
buto
r
R
oles
T
a
xo
nomy
(
C
R
e
diT
)
to
r
e
c
ognize
indi
vidual
a
uthor
c
ontr
ibut
ions
,
r
e
duc
e
a
utho
r
s
hip
dis
putes
,
a
nd
f
a
c
il
it
a
te
c
oll
a
bor
a
ti
on.
Nam
e
of
Au
t
h
or
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
Kr
is
ti
a
n
E
ka
c
ha
ndr
a
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Dina
r
Aje
ng
Kr
is
ti
ya
nti
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
C
:
C
onc
e
pt
ua
li
z
a
ti
on
M
:
M
e
th
odol
ogy
So
:
So
f
twa
r
e
Va
:
Va
li
da
ti
on
Fo
:
Fo
r
ma
l
a
na
ly
s
is
I
:
I
nve
s
ti
ga
ti
on
R
:
R
e
s
our
c
e
s
D
:
D
a
ta
C
ur
a
ti
on
O
:
W
r
it
in
g
-
O
r
ig
in
a
l
D
r
a
f
t
E
:
W
r
it
in
g
-
R
e
vi
e
w
&
E
di
ti
ng
Vi
:
Vi
s
ua
li
z
a
ti
on
Su
:
Su
pe
r
vi
s
io
n
P
:
P
r
oj
e
c
t
a
dmi
ni
s
tr
a
ti
on
Fu
:
Fu
ndi
ng a
c
qui
s
it
io
n
CONF
L
I
CT
OF
I
NT
E
RE
S
T
S
T
AT
E
M
E
N
T
Author
s
s
tate
no
c
onf
li
c
t
of
int
e
r
e
s
t.
I
NF
ORM
E
D
CONSE
NT
W
e
ha
ve
obtaine
d
inf
or
med
c
ons
e
nt
f
r
om
a
ll
ind
ivi
dua
ls
include
d
in
thi
s
s
tudy.
Evaluation Warning : The document was created with Spire.PDF for Python.