TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 9, September
2014, pp. 67
2
5
~ 673
1
DOI: 10.115
9
1
/telkomni
ka.
v
12i9.459
3
6725
Re
cei
v
ed O
c
t
ober 6, 20
13;
Revi
se
d Apr
21, 2014; Accepted Ma
y 6, 2014
Intelligent Train Operation Models Based on Ensemble
Regression Trees
De
w
a
ng Che
n
*, Xiang
y
u
Zeng, Gui
w
e
n
Jia
State Ke
y
L
a
b
o
rator
y
of Rai
l
T
r
affic Control
and Safet
y
, Bei
jing Ji
aoto
ng U
n
iversit
y
,
Beiji
ng, 10
00
4
4
, Chin
a.
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: d
w
c
h
e
n
@b
jtu.edu.cn, 1
112
031
9@b
j
tu.ed
u
.
cn, 1012
505
7
@
bjtu.e
du.cn
A
b
st
r
a
ct
Traditional c
o
ntrol algorithm
s
in Automatic Train Operation (
A
TO) system
have some drawbacks,
such as hi
gh e
nergy co
nsu
m
ption a
nd low
ri
din
g
co
mfort. Co
mbi
n
e
d
w
i
th data mi
ni
ng methods a
nd dri
v
in
g
exper
ienc
e, two Intelli
ge
nt T
r
ain Op
eratio
n (
I
T
O
) model
s fo
r the subw
ay train co
ntrol
are
prop
osed. F
i
rst
l
y
the train
i
n
g
da
ta set w
a
s sorted out
and s
i
eved
out fr
o
m
the rea
l
trai
n
oper
ation
data
set by driv
ers
i
n
Beiji
ng s
ubw
ay
line Y
i
z
h
u
a
n
g
to establ
ish th
e stand
ar
d d
a
t
abas
e. By usin
g Class
ificati
o
n
and R
egr
essi
on
T
r
ees (CART
)
alg
o
rith
m an
d
Bagg
ing
ense
m
b
l
e l
earn
i
n
g
meth
od w
h
ich
base o
n
CART
algor
ith
m
, tw
o
IT
O
mo
de
ls ar
e
du
g o
u
t to r
epr
es
ent the
o
u
tput
of contro
ller
w
i
t
h li
mit
ed s
p
e
e
d
, run
n
in
g ti
me
an
d
grad
ie
nt. In
the train contr
o
l sim
u
lation platform
, ITO m
odels
were com
p
ared with the
traditional PID (Proportional
Integral Derivative) control algorit
hm
of ATO system
s. The simulation
results indicate the propos
ed ITO
mo
de
ls are b
e
tter than PID control
i
n
ener
gy cons
u
m
pti
on, rid
i
ng
comfort an
d
sw
itching times o
f
control
l
er
’
s
ou
tput. F
u
rther
more, the
IT
O mo
de
l w
i
th b
agg
ing
e
n
se
mble
le
arni
ng
meth
od
is
bet
ter
espec
ial
l
y in e
nergy co
nsu
m
ption a
nd ri
din
g
comf
ort.
Ke
y
w
ords
: en
sembl
e
Lear
ni
ng, regress
i
on
trees, data mi
nin
g
, auto
m
ati
c
train oper
ati
on, intel
lig
ent tra
i
n
oper
ation
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
Control te
ch
nology
plays an im
po
rtan
t role i
n
mai
n
taining
saf
e
, reliabl
e, an
d cost-
effective operation of trains. Automatic Train Op
er
ation (ATO
) is resp
on
sible fo
r all the traction
and b
r
a
k
ing
controls.F
aci
ng with
su
ch
real-tim
e dynamic
ope
rat
i
onal requi
re
ments, intelli
gent
control strate
gies
came int
o
play in the 1980
s.
PID (Propo
rtional I
n
tegral
De
rivative) cont
rol
has
been
wid
e
ly use
d
in in
d
u
strial
co
ntro
l
system be
cau
s
e of
it
s simple
stru
cture
and
ro
bust
perfo
rman
ce,
also it
ca
n
be u
s
e
d
in
A
T
O
system
[1], Geneti
c
a
l
gorithm
(GA) is propo
se
d
to
optimize
train
movement
s
usin
g ap
pro
p
r
iate
coa
s
t control th
at ca
n be inte
grated withi
n
AT
O
system
s [2] or con
s
tru
c
te
d optimal trai
n driv
ing
stra
tegy [3]. Least sq
uare est
i
mation an
d an
adaptive
net
work
ba
sed
fuzzy inferen
c
e sy
stem
(A
NFIS)
we
re
p
r
esented
to e
s
timate the
train
station pa
rki
n
g erro
r in urb
an rail tran
sit [4-5].
However, these
cont
rol method
s a
r
e limited to the
traditional
co
ntrol theo
ry, whi
c
h takes
control ac
cu
racy as the
main goal in
the process o
f
tracking
ope
ration spee
d curve, in o
r
de
r to make
the
real
spe
ed
curve a
s
close
to the optim
al
one
as po
ssi
b
le. Mo
reove
r
the
controller
of AT
O
systems ne
ed
freq
uently
switchi
ng i
n
t
he
pro
c
e
ss
of train ope
ration,
which is n
o
t con
d
u
c
tive to the riding
co
mfort and e
n
e
rgy saving, the
life of controll
er is al
so g
r
e
a
tly reduced i
n
the meanti
m
e.
Different from
ATO system
s, experi
e
n
c
e
d
dr
ivers
can
operate the tr
ain to the
sp
ecified
locatio
n
o
n
ti
me
with a
fe
w time
of h
a
ndle
ch
angi
n
g
smoothly.
A larg
e a
m
ou
nt of the
data
are
gene
rated
by
huma
n
d
r
ivers in th
e p
r
oce
s
s of m
a
nual train
co
ntrol. From
t
he
calculatio
n of
actual data,
we ca
n find that
manual d
r
iving by experien
c
e
d
driv
ers i
s
better
than automat
ic
driving in e
n
e
rgy con
s
um
ption and
ridi
ng comfort. In
ord
e
r to g
e
t
better co
ntrol effect, we
are
trying to find
the intelligent
train op
eration (ITO
)
mo
d
e
l, the driving
model mi
nin
g
from the l
a
rge
amount of ma
nual drivin
g d
a
ta by data mining techniq
ues [6].
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 9, September 20
14: 67
25 – 673
1
6726
2. Collection
of Field Data
There are two basi
c
drivin
g mode in su
bway
train, a
u
tomatic drivi
ng mode an
d
manual
driving mod
e
. In this paper,
field data in two day
s (20 times) from Yizhu
ang Lin
e
Beijing Subway
were collecte
d
. We choo
se one blo
c
k
from Xiaoh
o
ngmen
statio
n to Xiaocu
n
station a
s
an
example. F
o
r this bl
ock,
we collect
112
35 g
r
ou
ps
of
sampl
e
s in t
w
o
days, a
n
d
ch
oo
se 8
u
s
eful
attributes i
n
each sa
mpl
e
. The eig
h
t attri
butes
a
r
e limited
sp
eed, gradie
n
t, train spe
e
d
,
remai
n
ing tim
e
, remainin
g distan
ce, cha
nging value o
f
next limited
spe
ed, remai
n
ing dista
n
ce
of
next limited speed a
nd co
ntrolle
r’s o
u
tp
ut (from -1 t
o
1, positive is tra
c
tion, ne
gative is bra
k
ing,
zero is idle
ru
nning
). The fi
rst
seven
attributes
are
use
d
as i
nput va
riable
s
, and th
e last attri
but
e
is used a
s
the output varia
b
le.
It is ne
ce
ssary to get the
data
with go
o
d
pe
rform
a
n
c
e from
the
o
b
tained
ma
ssive field
data. The
rea
s
on i
s
, some
of the data i
s
prod
uced
by
high-l
e
vel d
r
ivers
but som
e
is p
r
od
uced
by
middle
-
level
or lo
w-level
drivers. M
o
re
over, th
e
dri
v
ers may b
e
affected
by
psycholo
g
ical
o
r
physi
cal
con
d
i
tions, for
exa
m
ple, long
time op
eratio
n
may lead to
fatigue an
d p
r
essu
re. T
hat
is
to say
we
ne
ed to pi
ck o
u
t
the data
wit
h
low en
ergy
co
nsumption
,
high
riding
comfo
r
t, and
low
runni
ng time
error.
Acco
rdi
ng to
the statisti
cal
results for th
e m
anu
al dri
v
ing data set, the variation
of time
error is [-5.2
7.8](s), the variati
on of switching time
s of controll
er
’
s
output is [4
16], the variation
of impingem
e
n
t rate is [0.087 0.146](m/s
3
) an
d the variation of en
er
gy co
nsum
ption is [197.
32
216.63](J). By trial-an
d-e
r
ror, we set the followi
ng fo
ur rule
s to se
lect data
whi
c
h
satisfying
the
all rules.
Rule 1 Tim
e
error is
within
5
s
;
Rule 2 Swit
ch
ing times of controlle
r’
s out
put is within 1
0
times;
Rule 3 Impin
gement rate is within 0.1
2
m/s
3
;
Rule 4 Ene
r
g
y
consumptio
n is within 2
1
0
J.
Thro
ugh the
above rul
e
s,
7312 g
r
ou
ps
of sa
mple
s a
r
e sorte
d
out for data mini
n
g
.
3. Regre
ssio
n
Trees an
d its Improv
e
m
ent b
y
Ens
e
mble Learn
i
ng
3.1. Classific
a
tion and
Re
gression Tr
e
es Algorithm
Due to the m
u
lti-varia
b
le a
nd larg
e amo
unt of data for the re
gre
s
sion p
r
obl
em
in this
pape
r, tra
d
itional
reg
r
e
s
si
on m
e
thod
s
are
not
appl
i
c
abl
e. So
we
use
CART (Cla
ssifi
cation
an
d
Reg
r
e
ssi
on T
r
ee
s) al
gorith
m
to solve it.
CART
algo
rit
h
m [7] wa
s
p
r
opo
se
d by B
r
eima
n, Frie
d
m
an, Ol
she
n
and Ston
e i
n
198
4
.
The l
e
tters
CART in
dicate
that trees m
a
y be
u
s
ed
not only
to
cl
assify entitie
s into
a
di
screte
numbe
r
of group
s, but
also a
s
a
n
alte
rn
ative app
ro
a
c
h to
reg
r
e
ssi
o
n
an
alysi
s
in
whi
c
h th
e val
ue
of a re
sp
on
se (d
epen
dent
) varia
b
le i
s
to be e
s
timate
d, given the v
a
lue of e
a
ch
variable i
n
a
set
of explanatory (indepe
nde
nt) variabl
es.
But single
CA
RT alg
o
rithm
can
not a
c
hie
v
e be
tter results. To imp
r
o
v
e its pe
rform
ance ,
an en
sembl
e
learni
ng alg
o
rithm is use
d
with it.
3.2. Eesemble Learning
In 1997,T.G.
Dietteri
ch, au
thority in the field of machi
ne lea
r
nin
g
, put en
sembl
e
learni
ng
in the first pla
c
e of fou
r
re
search di
re
ctio
ns of
ma
chin
e learning [8]
.
Resemble l
e
arnin
g
is o
ne
of
the research
hot spot
s in machi
ne lea
r
ning in
re
cent
years, and the main a
c
hi
evements a
r
e
as
follows: Bagg
ing [9], Boosting [10], Ran
dom Forest [11] and so o
n
. The main idea of re
sem
b
le
learni
ng i
s
t
r
aining
multipl
e
wea
k
le
arni
ng
syst
em
s a
nd combi
n
ing
the re
sults
i
n
a ce
rtain
way,
whi
c
h can si
gnificantly improve
the generali
z
ation ability of t
he learni
ng sy
stems. The brief
introdu
ction o
f
bagging al
g
o
rithm is
sho
w
n a
s
follow.
The ba
gging
algorith
m
wa
s propo
se
d b
y
Breiman in
1996 [9]. Du
ri
ng the traini
n
g
pha
se
of the algo
rith
m, we repe
atedly sam
p
le
s the origi
nal
t
r
ainin
g
sampl
e
s
with re
plication so that
we
get a new
set
of training sa
mples.
Select trainin
g
sample
s from the
origi
n
al set of trai
ning
sam
p
le
s at ra
ndom; t
r
ain th
e
sampl
e
s u
s
in
g the given base lea
r
ni
ng algorith
m
,
then we ca
n get a model; put back the train
i
ng
sampl
e
s. Re
peat k
time
s, so
th
at
we ca
n
get a
set of
k
model
s. A
s
fo
r the
re
gression
proble
m
,
we can obtai
n a final forecasting mo
del
as follo
ws:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Intelligent Tra
i
n Operation
Model
s Base
d on En
sem
b
l
e
Reg
r
e
ssi
on
Tree
s (Dewa
ng Ch
en)
6727
1
1
()
()
K
ik
i
k
F
xf
x
K
(1)
The b
aggi
ng
algo
rithm i
s
the mo
st si
mple
a
nd int
u
itive ensem
ble lea
r
nin
g
method.
Theo
retically, when u
s
in
g the Baggin
g
a
l
gorithm,
ab
o
u
t 36.8% of the sa
mple
s
will not app
e
a
r in
the new
set o
f
training sa
m
p
les ave
r
ag
el
y as we resa
mple ea
ch time.
3.3. Algorith
m
s Implementa
tion
The ba
gging
ensemble le
a
r
ning al
go
rith
m can b
e
used for re
gression a
nalysi
s
.
In this
pape
r,
CART
algo
rithm
is
use
d
a
s
the
wea
k
l
e
a
r
nin
g
ma
chi
ne of baggi
ng algo
rithm,
we
call this
algorith
m
as
B-CART, and
we can al
so
use a
sin
g
le
CART
algo
rithm for an
alyzing. He
re we
use
these t
w
o m
e
thod
s to mi
ne the IT
O
model
s fro
m
the data. Af
ter that, we
will an
alyze
and
comp
are the results.
B-CART is a
method inte
g
r
ating the
reg
r
essio
n
tree
s,
and it ca
n set the iteratio
n times.
Here, we
set
the iteration
times ra
ngin
g
from 1
to 1
00. The me
a
n
absolute e
r
ror i
s
sho
w
n
in
Figure 1.
Figure 1. Mean Absol
u
te Erro
r Ch
ang
e
s
with Iteratio
n
Times
In Figure 1,
comp
ared
wi
th the re
sult
s of
CART
algorith
m
,
B-CART algo
rit
h
m
can
alway
s
b
e
su
perio
r to
it. T
he results
obt
ained
are
coi
n
cid
ent to th
e
theoretical
a
nalysi
s
. With
th
e
iteration tim
e
s increasing,
the ca
lculating speed of th
e algorithm
will be
greatly reduced. So
we
set the iterati
on times a
s
5
0
in our si
mul
a
tion.
4. ITO Model and its Simulating Platform
4.1. Opera
t
ion Requirem
e
nts fo
r each
Stage
Safety is important in the t
r
ain o
peratio
n,
and the m
o
st ba
sic
req
u
irem
ent for i
t
is that
the ru
nnin
g
speed
should
not exceed
th
e limited
sp
e
ed. So
we
di
vide the train
ope
ration
int
o
4
stage
s, a
c
cel
e
ration
stage
, idle
run
n
in
g sta
ge,
de
celeratio
n
sta
ge
a
n
d
stop
ping stag
e. The
requi
rem
ents
for each stag
e are differen
t.
Accel
e
ration
stage a
nd Idl
e
runni
ng sta
ge: If
ma
x
0.
95
*
VV
, then
0
a
.
De
cele
ration stage:
If
ma
x
0.
9
*
VV
, then
0.
5
a
. Where
V
i
s
spee
d,
V
ma
x
is limited
spe
ed,
a
is
c
o
ntroller’s
output.
4.2. Simulation Platform
The ITO m
o
del si
mulatio
n
platform
is establi
s
h
ed
with Matla
b
Simulink. T
h
e platform
inclu
d
e
s
five module
s
, inp
u
t module, ge
nerato
r
mo
du
le, controller
module, a
c
tu
ator mod
u
le
and
displ
a
y modul
e. Figure 2 is the stru
cture
grap
h of ITO model.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 9, September 20
14: 67
25 – 673
1
6728
Figure 2. Structure
Gra
ph o
f
ITO Model
5. Simulation and Ev
aluation
We
also
ch
o
o
se
blo
c
k
1
as
an
examp
l
e. The
di
sta
n
ce
of
blo
c
k
1 is 1
341m,
and th
e
stand
ard
run
n
ing time is 1
05s. Th
e re
gression
m
odel
s mine
d by CART and B
-
CART are use
d
in
the ITO mod
e
ls, and the regre
s
sion m
o
del mined by
CART a
nd B-CART a
r
e sh
own in Fig
u
re
3
and Figu
re 4.
Figure 3. Reg
r
essio
n
Mod
e
l
of CART Alg
o
rithm
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Intelligent Tra
i
n Operation
Model
s Base
d on En
sem
b
l
e
Reg
r
e
ssi
on
Tree
s (Dewa
ng Ch
en)
6729
Figure 4. Reg
r
essio
n
Mod
e
l
of B-CART
Algorithm
As
can be
seen i
n
Fi
gire 3
and Figure 4,
the x1,
x
2,...,x7 repr
esent limited speed,
gradi
ent, trai
n sp
eed,
rem
a
ining tim
e
, remainin
g di
st
ance, ch
angi
ng value
of n
e
xt limited sp
eed
and remaini
n
g dista
n
ce of next limited spe
ed the
s
e seve
n inpu
t variable
s
resp
ectively. The
reg
r
e
ssi
on m
odel of B-CA
RT alg
o
rithm
is big
ger
th
an
the re
gre
s
si
on mod
e
l of
CART
algo
rithm,
the num
bers
of leaf no
de i
n
thes
e two
model
s a
r
e 1
9
and
30
re
spectively. Fro
m
the root no
de to
each leaf no
de co
rrespon
ds to a rul
e
, that is to
say,
there are 19
rule
s in reg
r
ession mo
del
of
CART al
gorit
hm and 30
rul
e
s in re
gres
si
on model of
B-CART algo
rithm.
The
com
pari
s
on
of
spe
e
d
an
d
controller’
s o
u
tput
und
er
CA
RT algo
rithm,
B-CA
RT
algorith
m
and
PID control a
r
e sh
own in the Figu
re 5 a
nd Figu
re 6.
Figure 5. Co
mpari
s
o
n
of Speed Curve
s
Figur
e 6. Co
mpari
s
o
n
of Controlle
r’s O
u
tput
Figure 5
sh
o
w
s the o
p
e
r
at
ion of ITO
mo
del wi
th
CA
RT or B
-
CA
RT
is
smooth
e
r t
han PID
control, and
there
is
a lon
g
time of idle
runni
n
g
in the
middle
stage
, so the
ene
rgy con
s
um
ption
will be le
ss.
In Figure 6
,
the switchi
ng times
of
controlle
r’s o
u
tput in ITO model re
du
ce
s
obviou
s
ly. Th
e pe
rform
a
n
c
e of ITO
mod
e
l with
CA
RT
and B
-
CA
RT
are
very
simi
lar
with ma
nu
al
driving. We
simulate 1
0
0
times using
ITO model with CART o
r
B-CART re
spe
c
tively, and
cal
c
ulate th
e
averag
e valu
e. The fun
c
tio
n
com
p
a
r
ison
s of PID control and ITO
m
odel
with CA
RT
or B-CA
RT a
r
e sho
w
n in T
able 1.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 9, September 20
14: 67
25 – 673
1
6730
Table 1.
The
Functio
n
Co
mpari
s
o
n
s of
PID Contro
l, ITO model
with CART o
r
B-CART
PID
Control
CART
B-CART
e
t
(s
)
0.8
1.25
1.98
T
c
14 6.34
5.56
I
r
0.416
0.201
0.191
E
(J
)
236.7
206.48
201.52
Whe
r
e
e
t
is time error,
T
c
is swit
ching times of cont
roller’s output,
I
r
is impingem
ent rate,
E
is ene
rgy.
Cal
c
ulation of
impingem
ent
rate:
1
1
1
1
n
ii
i
aa
I
nt
(2)
The
smalle
r t
he impi
ngem
ent rate, the
better the
ridi
ng comfort
(
C
r
), wh
ere
a
i
i
s
output of
controlle
r. Ca
lculatio
n of energy con
s
u
m
ption:
Fv
d
t
m
a
v
d
t
Ea
v
d
t
mm
(3)
Whe
r
e
a
is t
he accele
rati
on,
v
is sp
e
ed, this form
ula is a ro
ug
h cal
c
ulation,
just for
comp
ari
s
o
n
. In Table 1,
co
mpared with
PID cont
rol, the ITO mod
e
l with CAR
T
or B-
CART
h
a
s
some
a
d
vant
age
s: the
T
c
and
E
re
du
ce
obviou
s
ly, a
nd
C
r
b
e
com
e
s better; althoug
h
e
t
re
d
u
ce
s
slightly, it still
meet the requirem
ents.
In Figu
re
5
a
nd Fi
gure 6,
the results
of
ITO m
odel
with
CART
a
nd B-CA
RT
are
very
simila
r. From
the fou
r
co
mpari
s
o
n
s, t
hey ju
st
hav
e a little diffe
ren
c
e. T
he I
T
O mo
del
wi
th B-
CART p
e
rfo
r
ms better on
T
c
,
C
r
and
E
.
Above all, we
use B-CA
RT
as the data mining alg
o
rit
h
m
of ITO model.
In order to verify the generality of
ITO model
, the other blo
c
ks of Yizhua
ng Lin
e
Beijing
Subway a
r
e simulate
d too
.
Given that the sp
ac
e of pape
r is limited, we don’t
list the results
here. Th
e si
mulation resu
lts are si
milar with block
1, that is to say, the ITO model ha
s achi
e
v
ed
good p
e
rfo
r
m
ance in all blo
c
ks. The
r
efo
r
e,
the general
ity of
ITO model is very go
od.
6. Conclusio
n
In this pa
per,
data in ma
n
ual drivin
g of
ex
cellent d
r
i
v
ers a
r
e
colle
cted a
nd filtered, an
d
then the stan
dard d
a
taba
se is esta
blish
ed. Th
ro
ugh t
w
o data mini
ng algo
rithm
s
, two ITO models
are
dug
o
u
t. We
sim
u
late
all the
blo
c
ks of Yizhua
ng
Line B
e
ijing
Subway. F
r
o
m
the
re
sult
s, IT
O
model
s
have
achieved
g
ood
pe
rfor
m
ance com
pared with
PID
co
nt
rol, esp
e
cially
i
n
ridi
ng
comfo
r
t and
energy co
nsumption. As
for the tw
o d
a
ta
mining al
gorithm
s,
B-CART algo
rithm
perfo
rms b
e
tter, so
we cho
o
se thi
s
algo
rithm as the d
a
ta mining al
gorithm of IT
O model.
There are
some issue
s
of this work need
to be
further re
se
arched, such
as th
e
robu
stne
ss o
f
ITO mod
e
l, the ad
aptab
ility of
the ITO mod
e
l for
steep
gradie
n
t and
co
mpl
e
x
limited spe
e
d
.
Moreove
r
, more a
d
van
c
ed data mini
ng algo
rithm
s
are worth fu
rther
studyin
g for
both simul
a
tion and real
-world sce
n
a
r
io
s.
Ackn
o
w
l
e
dg
ements
This work is partially su
pporte
d by the
Natio
nal High Te
ch
no
logy Re
sea
r
ch an
d
Develo
pment
Program (“8
63” Program) of China
un
d
e
r grant 201
2
AA11280
0, by New Sci
enti
f
ic
Star Prog
ram
of Beijing
u
nder grant 2
010B01
5,
by
the Fun
dam
ental Resea
r
ch F
und
s fo
r the
Central Univ
ersitie
s
un
de
r 2012
JBM0
1
6
, by t
he indepen
dent re
search p
r
oje
c
t
from the State
Key Laborato
r
y of Rail Traf
fic Cont
rol
an
d Safety under grant RSC2011Z
T00
1
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Intelligent Tra
i
n Operation
Model
s Base
d on En
sem
b
l
e
Reg
r
e
ssi
on
Tree
s (Dewa
ng Ch
en)
6731
Referen
ces
[1]
XX Ch
en, Y Z
hang, H Hu
ang.
T
r
ain speed co
ntrol al
gorithm
base
d
on PID contro
l
l
er an
d singl
e-n
euro
n
PID controll
er.
Co
mp
uter soci
ety
. 2010; 1: 1
07-1
10.
[2]
CS Chang, SS Sim.
Optimis
i
n
g
train
mov
e
ments throu
gh c
oast contro
l usi
ng g
enetic
al
g
o
rith
ms
. Proc.
IEEE Electric Po
w
e
r Ap
plic
atio
ns. 1997; 1
44: 65–
73.
[3]
SH Han, YS B
y
e
n
, JH Ba
ek, T
K
An, SG Lee, HJ Park.
An opti
m
al a
u
tomatic train
o
perati
on (AT
O
)
control us
ing
g
enetic a
l
g
o
rith
ms
. Proc. IEEE Regi
on 1
0
Co
nf.
T
E
NCON. 1999; 1: 36
0–3
6
2
.
[4]
DW
Chen, T
T
ang, CH Gao, RQ Mu. Research o
n
the
error estimati
o
n
mode
ls an
d onli
ne l
ear
nin
g
alg
o
rithms for train statio
n par
king i
n
urba
n rail trans
it.
Chin
. Railw
ay Sci.,
201
0; 31(6): 12
2-12
7.
[5]
DW
Chen, CH
Gao. Soft computin
g meth
ods ap
pli
ed to
train station
parki
ng in
urb
an rai
l
transit.
Appl
ied S
o
ft Computi
ng.
20
1
2
; 12(2): 75
9-7
67.
[6]
F
Eibe, W
Ian. Data Mini
ng: Practical Mac
h
ine L
earn
i
n
g
T
ools a
nd T
e
chniq
ues.
Mo
rg
an
Ka
u
f
m
a
nn.
200
5.
[7]
L Breima
n, JH
F
r
iedman, RA
Olshen, CJ S
t
one.
Class
ific
ation a
nd R
e
g
r
essio
n
T
r
ees. Belmont, CA:
W
ads
w
o
rth.1
9
84.
[8]
T
G
Dietterich. Machi
ne Le
arn
i
ng R
e
searc
h
: F
our Cu
rre
nt Directio
n
s. AI Magaz
ine. 1
997;
18(4): 97-1
36.
[9]
L Breima
n. Baggi
ng pre
d
ictor
s
.
Ma
ch
i
n
e
Le
arn
i
ng
.
199
6; 24
(2): 123-1
40.
[10]
RE Schap
ire, Y F
r
eund, P Bart
let. Boostin
g
the margin: a
ne
w
e
x
p
l
a
nati
on for the effe
ctiveness o
f
voting meth
ods
.
T
he Annals of
Statistics
. 1998; 26(5): 16
51-
168
6.
[11]
L Breima
n. Ra
ndom forests.
Machi
ne Le
arn
i
ng.
20
01; 45(
1
)
: 5-32.
Evaluation Warning : The document was created with Spire.PDF for Python.