Professional Documents
Culture Documents
Hopfield Neural Network Based On Ant System
Hopfield Neural Network Based On Ant System
com
Jo
urn
al
ofHar
bin
In
sti
tut
e o
f T
echn
olo
gy(Ne
w S
eri
es),V
o1.1
1,No
.3,20
04
Hopf
iel
d neural
net
work based on ant
syst
em
HONG
Bing—
rong,jI
N Fe
i—hu,GUO Qi
洪炳镕 ,
金 飞虎 ,
郭 琦
(S
choo
l o
f Comput
er
Sci
enc
e and
Te
chnol
ogy,Ha
rbi
n I
nst
it
ute
ofTe
chnol
og
y ,Ha
rbi
n 150001,Chi
na,E-
mai
l:Fhj
in@0451.cor
n)
Abst
ract:Hopf
iel
d neur
al net
wor
k i
s a
singl
e l
aver
feedf
orwar
d neural
net
work.Hopf
iel
d net
wor
k r
equi
res
some
eont
rol
par
amet
ers
to be
car
eful
ly
sel
ect
ed.el
se t
he net
wor
k i
s apt
to
conver
ge t
o l
oca
l mi
nimum.An
ant
syst
em
i
s a nat
ure i
nspi
red met
a heur
ist
ic al
gor
ithm .I
t has
been appl
ied
to
sever
al combi
nat
ori
al opt
imi
zat
ion
pr
obl
ems
such as
Tr
avel
ing Sal
esman Pr
obl
em ,Schedul
ing Pr
obl
ems,et
c. Thi
s paper
wi
ll s
how an ant
syst
em may be
used i
n t
un i
ng t
he network
cont
rol
par
ameter
s by
a
group o
f coo
pe r
ated ant
s .
11he major
advant
age of
thi
s net—
work i
s t
o adi
ust
the net
work
paramet
er
s aut
omati
ca
lly.avoi
ding a
blind s
earch f
or the
set
of
cont
rol
parameter
s.
Thi
s net
work was
test
ed on t
wo TSP pr
obl
ems.5
ci
ties
and 10 ci
ties. The r
esul
ts hav
e s
hown an obvi
ous
im—
pr
ovement
.
Key
words:hopf
iel
d net
wor
k;ant
sys
tem ;TSP;combi
nat
orial
opt
imi
zat
ion pr
obl
em
CLC
numbe
r:TP3
9 Do
cument
code:A Ar
ticl
e I
D:1
005-
9Il
3(20
04)
03-
0267-
03
Tr
avel
ing Sal
esman Pr
obl
em (TSP)i
s a cl
ass
of
heuri
sti
c hel
ps f
ind accept
abl
e sol
uti
ons
in t
he ear
ly
combi
nat
ori
l opt
a imi
zat
ion probl
ems. It
has a var
iet
y of
st
ages of t
he search process.
i
mport
ant
appl
icat
ions, most
ly i
n t
he ar
ea of rout
ing TSP probl
em and Hopf
ield Neur
al net
work wi
ll be
and schedul
ing pr
obl
ems. The TSP i
s consider
ed t
o be br
ief
ly present
ed i
n Sect
ion 1. In Sect
ion 2. Hopfi
eld
one of
the st
andard pr
obl
ems i
n t
he oper
ati
on resear
ch/ neural
net
work based on ant
syst
em i
s i
ntroduced. Sec—
management
sci
ence l
iter
ature.
t
ion 3 present
s t
he s
imul
ati
on r
esul
ts and,f
inal
ly,Sec—
Recent
advances
in t
he st
udy of ar
tiici
f al neur
al t
ion 4 summari
zes concl
usi
ons.
net
wor
ks have r
esult
ed,among ot
hers,i
n t
he al
gori
thm
of Hopfi
eld who shows t
he TSP may be sol
ved by neu—
1 TSP and Hopf
iel
d Neural
Net
work
ron—
like comput
ati
onal
net
works. Thi
s work has
at
trac.
t
ed t
he at
tent
ion of
many resear
chers t
o t
he
iel
f d of
neu-
r
al net
works. Unf
ortunat
ely,t
he det
ermi
nist
ic nat
ure of
Tr
avel
ing Sal
esman Probl
em i
s a pr
obl
em t
hat
iS
Hopf
iel
d net
work, whi
ch m i
nimi
zes
a gl
obal
ener
gy comput
ati
onal
ly expensi
ve to sol
ve. The premis
e of
TSP
f
uncti
on.is
heavi
ly dependent
on i
nit
ial
val
ues of
pa—
i
s t
ypi
cal
ly t
aken t
o be a t
ravel
ing s
alesman needs t
o
vi
sit
a l
ist
of ci
ties. Each ci
ty only connect
s t
o cer
tai
n
r
amet
ers
and var
iabl
es. Wi
th t
his
na
tur
e,t
he or
igi
nal
net
work i
s apt
to converge t
o l
ocal
mi
nimum ,never
to ot
her
ci
ties.and t
here is
a wei
ght
associ
ated wi
th t
hat
connect
ion. The wei
ght
can be a t
ime.how much f
uel
escape fr
om t
his
st
ati
onar
y poi
nt. Thus.i
t i
s ver
y i
m—
port
ant
to us
e appropr
iat
e ranges of randomi
zed i
nit
ial
i
s consumed,a combinat
ion of
the t
wo or some ot
her
st
ates and of
some mul
tipl
iers. Fur
ther
mor
e when t
he arbi
trar
y condi
tion. The goal
is
to f
ind a pat
h that
al—
TSP i
s si
mul
ated wi
th neur
al net
works.t
he i
nter— r
ela—
l
ows t
he salesman t
o vi
sit
each ci
ty exact
ly once wi
th
t
ionshi
ps among t
he var
iabl
es are t
oo compl
ex to be re—
t
he l
owest
cos
t possi
ble. Of
ten t
he pr
obl
em requi
res
ined deci
f sivel
y. t
hat
he end i
n t
he same ci
ty he began i
n.
In recent
years,Ant
syst
em has been devel
oped t
o In t
he TSP,we are ven n ci
ties and a di
stance
sol
ve several
combi
nat
ori
al opt
imi
zat
ion probl
ems
such ma
tri
x[d
]whe
re
d“i
s t
he
dis
tanc
e be
twe
en
cit
y i
and
as Tr
avel
ing Sal
esman Pr
oblem , Schedul
ing Probl
ems ci
ty TSP i
s a permut
ati
on pr
obl
em , i.e. a feasi
ble
etc. Ant
syst
em i
s a nat
ure i
nspi
red met
a heuri
sti
c al
— s
olut
ion t
o TSP is
a per
m ut
ati
on of
the ven n
cit
ies.
gori
thm . Ant
sys
tem has some proper
ties:posi
tive f
eed—
Hopf
ield Net
works
have been wi
del
y used t
o Sol
ve
back,di
stibut
r ed comput
ati
on,and use of a const
ruc—
a vari
ety of
combi
nat
ori
al opt
imi
zat
ion pr
obl
ems. I
t can
t
ive gr
eedy heur
ist
ic. Posi
tive f
eedback account
s f
or be proved t
hat
a Hopf
iel
d Net
work wi
th di
scret
e out
put
rapi
d di
scover
y of good sol
uti
ons, di
stibut
r ed comput
a— neurons oper
ati
ng i
n an appropr
iat
e m anner 1s
a stabl
e
t
ion avoi
ds premat
ure convergence, and t
he greedy syst
em i
n t
he Lyapunov sense. Hopf
ield Neural
net
is
Recei
ved 2003 —0l—l0
Spo
nsor
ed by
th
e Na
tiona
l Hi
gh—
Technol
og
y 863
Pr
ojec
t(Gr
ant
No.
863—2—4—1—2)
・
267 ・
维普资讯 http://www.cqvip.com
J
our
nal
o
f Ha
rbi
n I
nst
it
ute
of
Tec
hno
log
y(Ne
w Se
ri
es),V
o1.1
1,No
.3,2
004
usual
ly descr
ibed by ener
gy f
unct
ion i
n whi
ch t
he l
ow.
2 Appl
ying Ant
Syst
em to the Hopf
iel
d Neural
est
energ
y st
ate cor
responds t
o t
he opt
imum sol
uti
on.
Net
work
In the case ofn
citi
es pr
obl
em .make an × n
neu.
r
on mat
rix. For example. i
llust
rat
e a 5 x 5 matr
ix i
n For
par
amet
ers C14,t
ake al
—
l possi
ble val
ues t
o
Fi
g.1. In t
his
net
wor
k each neur
on i
s connect
ed t
o al
l compose a set
Bc(i=
5
. 1~ 4).As
sume
that
ther
e a
re
ot
her
neur
ons
f
ir
f ed
neur
on
symbol=“1” and
not
ir
f ed M ant
s i
n t
he const
ruct
ed al
gor
ithm . These ant
s st
ar
t
三℃ J
o (
{g =^I
一
=s
ymbol“0”).I
n t
his
mat
rix.a
row
is
the
number of from t
he am nes
/t
t
/ o
search
\ o r
f food.ever
y ant
choosi
ng
t
he ci
ty and a column i
s the number
of t
he vi
sit. I
f a
r
andoml
y an el
ement
f
rom ever
y setBc(i= 1~4)i
n .
ci
ty was
vi
sit
ed.t
he coresponding neuron i
n t
his
ma.
t
urn accor o o
ding t o
o t mo
he a o 奴
unt f pher
o omones core-
t
rix i
s f
ired as shown i
n Fi
g.2.
sponding t
o ever
y el
ement
in the set
and r
oulet
te al
go・
ri
thm . W hen ano
o
ant
f
ini
s o
ohes
c
h
眦 oosi
ng the el
ement
s i
n
f
our set
s,it
arri
ves at
a f
ood sour
ce,and t
unes phero—
● O o O O
i
)
O 0 O O ● mo
ne el
eme
nts
accor
ding t
o r
ules
f
or adj
ust
ing t
he a
—
O O O ● O
mount
of
pheromone.
0 ● o O O
O O ● O O As
sume
that
the
set
Br
have
Ni
(i= 1 ~ 4)el
e.
symb0l● i
s f
ire
d neuron ment .)(
s,P(Bc J= 1~N )denot
.
ing
itsj
th
element
s,
svmbol O i
s not
fir
ed neuron whose corespondi
ng amount
of pher
omones i
s denot
ed
pher
P(Bc
,.
).As
sume
Inc
denot
es
the
incr
ement
of
t
he
Fi
g.1
5 ×5 m atri
x pheromone amount
cor
respondi
ng t
o t
he el
ement
whi
ch
an ant
chooses dur
ing i
ts goi
ng f
rom ant
nest
to f
ood
source.
The st
eps that
ant
s use t
o sear
ch are t
he opt
imal
2 paramet
ers
for
a Hopf
iel
d net
wor
k:
St
ep 1: I
nit
ial
izi
ng t
he amount
of
pher
omone
pher
P(Bc
)o
f each
el
ement
P(Bc
)of
al
l t
he s
ets
l
B
c and t
heM ant
s st
art
fr
om t
he nest.
t
our rout
e
St
ep 2:Ant
s choose el
ement
s i
n each set
in t
urn
accordi
ng t
o path choosi
ng r
ule:
Fi
g.2 A t
our
rout
e and m atr
ix
Pat
h choosi
ng r
ule: f
or set
Bc .
,
an ant
chooses
Let
V4
denot
e t
he s
tat
e of
the
neur
on i
n r
ow
A, r
ando
mly
its
jt
h el
ement
ac
cor
ding t
o t
he pr
obabi
li
ty.
col
umn . Hopf
iel
d pr
oceeds
furt
her t
o devel
opuul
ine P (B ) = pher P(B ,)
(5) .
an ener
y funct
g ion:
E P :E1+E2+E3,
(1)
∑i
=
p
her
P(
1
Bc
i
j)I
St
ep3: r
epeat
Step 2,al
l ant
s arr
ive at
the f
ood
E =c
∑ ∑ ∑ A i
J≠ i
+c
∑ ∑ ∑
I
B≠ {
,
source.
(2)
St For an ant, after it
ep4: f
ini
shes
it
s searchi
ng
E
=c,
(∑ ∑ 一n
) ,
(
3) par
am et
er i
s , t
ad
jus
ts
the
amount
of
pher
omones
core・
o the chosen el
ement
s accordi
ng t
o t
he f
ol-
A t
spondi
ng t
E3= C4dA
B (V +
l+V 一
1).
(4)
l
owing r
ule.
The subscr
ipts
on V’s ar
e t
o be t
reat
ed modul
o n
Rule f
or
adjusti
ng the amount
of pheromone:Af ter
t
o t
ake car
e of
connect
ing t
he l
ast
ci
ty i
n t
he t
our
to t
he
ants
ini
f s
h the i
r one cycle,e very one of them adjusts
ir
f st
one. The par
amet
ers C
is are t
o be chosen.
the
amount of pheromone according
to equation(6)
Mi
nimi
zing t
he t
erm El
for
ces
the net
work t
o a
pherP(B哳)(t+ 4) = P ・phe rP(B )(t) +
st
ate wi
th no mor
e t
han one neur
on t
urned on i
n any
Inc(Bc ), (6)
r
ow or
col
umn——a requi
rement
for
the sol
uti
on t
o be a
where
P i
s a coei ci
f ent
s uch that(1一p)r epresent
s t
he
permut
ati
on. The t
erm
E,is
mi
nimi
zed when t
her
e ar
e
evapor
ati
on of
trai
l bet
ween t
ime and
t+ 4,
exact
ly n neur
ons
turned on. Thi
s i
s usef
ul t
o make
sure t
hat
neur
ons
do not
converge t
o t
he t
rivi
al“al
1.ze—
l
nc(
B。
)=∑I
nc
(B嘶),
(
7)
ro” state .
However. t
his
term can be repl
aced by a
gl
obal—bi
as t
erm f
eedi
ng al
l t
he neur
ons
to pr
oduce t
he
wher
e I
nc (Bc
)i
s t
he s
ubs
tance (pher
omone
in r
eal
same ef
fect
. The t
erm E t
akes car
e of t
he act
ual
di
s— ant
s)quant
it
y l
aid
on
par
ame
ter
P(Bc
..
)by
the
kt
h ant
tance m l
‘nl
‘nn’zatl
‘
on .
bet
ween t
ime t
and t+ 4;i
t i
s gi
ven by
Inc
(B ) =
・
268 ・
维普资讯 http://www.cqvip.com
J
our
nal
o
fHar
bin
In
sti
tut
e o
f T
echn
olo
gy(Ne
w S
eri
es),V
o1.11,No
.3,20
04
r
i
f
ktb ant
uses
par
amet
er尸(B ),
Tab
3 Si
mul at
ion
Resu
lts
{E i
n i
ts
to
ur(
bet
ween
time t
and
t+4),
【
。
th wi .
(8)
wh
ere
Q
is
a
pos
it
iv
e c
ons
tan
t,i
t i
s us
ed
to
adj
ust
th
e
s
peed
of
ad
jus
ting
the
amount
of
pher
omone.
St
ep 5: The above st
eps cont
inue, unt
il al
l ant
s
conver
ge on t
he sanl
e pat
h or a generati
on number
is
l
ager t
han t
he gi
ven number.
心
t
he same as
the opt
imal
tour. W hi
le usi
ng a st
andar
d
Hopf
iel
d neur
al net
wor
k,onl
y 37 va
lid s
olut
ions
coul
d
be
obt
ained
in 50 expe
iment
r s, and t
he bes
t s
olut
ion
has a
length of
3.63.5% l
onger
than t
he opt
imal
tour.
● ●
ant
nest
f
ood s
our
ce
Fi
g.3 Sel
ecti
ng
parem et
ers
Fi
g.4 Best
tour resul
ts
3 Si
e ul
r ati
on Res
ult
s
4 Concl
usi
on
The above met
hod has been t
est
ed on t
wo TSP
pr
obl
ems.5
ci
tie
s and 1
0 ci
ti
es. Coor
dinat
es i
n t
hes
e Hopf
iel
d neural
net
work based on ant
syst
em i
s
pr
obl
ems ar
e shown i
n Tabs.1
and 2. T
h e result
s ob.
present
ed in this paper. Hopf
ield neural netwo
rk has
t
ained
af
ter
50
tes
ts
are
sho
wn i
n Ta
b.3.The
best
tour
been widely used t
o sol
ve opt
imizati
on problems. How.
resul
ts are shown i
n Fi
g. 4. They ar
e compared wi
th ev
er,t
he cont
rol
par
amet
ers
for
thi
s wor
k have
to
be
t
he r
esul
ts usi
ng st
andard Hopf
iel
d neura】net
work. I
n ear
eful
ly
sel
ect
ed. or
el
se t
he net
wor
k i
s apt
to
con.
t
he 5 ci
ties probl
em .t
he sol
uti
on f
rom our
al
gor
ithm i
s verge t
o l
ocal
mi
nimum . In t
his paper,t
hese par
ame-
t
he same as a St
andar
d Hopf
iel
d net
work. In t
he 10 t
ers coul
d be obt
ained by a group of cooperat
ing work-
ci
ties
pr
obl
em 。t
his
met
hod can pr
oduce 48 (96% )
i
ng ant
s. Thi
s net
work has
been demonst
rat
ed i
n TSP
vali
d sol
uti
ons.and t
he best
tour
has
a l
engt
h of
3.46.
pr
obl
em of 5 ci
ties and 10 ci
ties. T
h e resul
ts have
shown a si
gni
fcant
improvement
usi
ng t
he al
gor
ithm.
Tab.1
Coordi
nat
es i
n 5 ci
tis
e
Ref
erences:
1
J J
D0RI
G0
M,MANI
EZZ0
V,C0L0RNI
A.Ant
sys
tem:
o
pti
miz
ati
on
by
a c
olon
y o
f c
oop
era
tin
g a
gen
t[J]
.IEEE
Tr
ansact
ions
on Syst
ems,M an,and Cybernet
ics, 1996.
26(1):29—41.
2 l
1 C0LORNI
A. Heur
ist
ics
from na
tur
e f
or har
d combi
nat
o.
r
il
a o
pti
miz
ati
on p
rob
lems[J].I
nt Tr
ans
in
Opn
l Re
s,
I996,3(1):
1—21.
Tab.2 Co
o rdi
nat
es i
n 10 ci
tis
e
{3
f C0RNE D,D0RI G0 M,GL0VER F.New Id
eas i
n Op.
y ti
miz
ati
on[M].Ne w
York:Mc
Graw—Hi l
1,1999.
O O.3oo I
4 HOPFI
l ELD
J J,TANK D W.Neura
l c
omp uta
tio
n of
de.
O.4l8 0.237
c
isi
ons
in
opt
i ̄
・i
fz
ati
on
pro
blems[
J].Bi
olo
gica
l Cy
ber
net
—
j
cs,1985,52:141— 152.
0.627 0.09l
:
5 j
WI
LS0N
G V,PAWLEY
G S.On
st
abi
li
ty
of
the
tr
aye
1.
0.732 O.3oo
1
ing sal
esman pr
obl
em a
lgor
ithm of Hopf
ie/
d and t
ank
1.045 0.6l
4
l
J
1.Bi
o Cv
ber
n,l
998,
58:
63~70.
0.962 0.948
0.836 l,l36
0.627 865
0.690 0.384
0.3l4 0.405
・
269 ・