Download Full Issue in PDF - Academy Publisher
Download Full Issue in PDF - Academy Publisher
Download Full Issue in PDF - Academy Publisher
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
1452 JOURNAL OF COMPUTERS, VOL. 8, NO. 6, JUNE 2013<br />
M of the adaptive RBF neural network filter<strong>in</strong>g is<br />
determ<strong>in</strong>ed.<br />
The dimension m of chaotic time series is<br />
calculated by the way of G- P algorithms, and<br />
the delay time τ is calculated by the selfcorrelation<br />
method. For the overall description<br />
of the dynamics characteristic of the orig<strong>in</strong>al<br />
system by the Takens' delay-coord<strong>in</strong>ate phase<br />
reconstruct theory, a chaotic series demand<br />
m≥ 2d<br />
+ 1 variances at least, so the number of<br />
the <strong>in</strong>put nerve cells of the adaptive RBF neural<br />
network filter<strong>in</strong>g is M = m ; The reconstruction<br />
phase space vector number is 200, Then, the<br />
200 phase space vectors to make a simple<br />
normalized, the normalized as<br />
[ x() t −mean( x())]/[max( t x()) t − m<strong>in</strong>( x())]<br />
t<br />
,<br />
t = 1, 2, 200 , and mak<strong>in</strong>g the value is owned by a range<br />
of -1 / 2 to 1/2.<br />
Step2) The adaptive filter<strong>in</strong>g is <strong>in</strong>itialized and the<br />
weights are vested the <strong>in</strong>itial values. RBF neural network<br />
vector weight<strong>in</strong>g parameters w is <strong>in</strong>itialized, where the<br />
weight vector w <strong>in</strong> each component take random function<br />
between 0 and 1; and the learn<strong>in</strong>g rate η is <strong>in</strong>itialized at<br />
the same time, where η = 0.0002 . β and γ are the<br />
learn<strong>in</strong>g rate adjustment factors, 0< β < 1, γ > 1 , for<br />
example, β = 0.75, γ = 1.05 .<br />
Step3) Us<strong>in</strong>g the above the <strong>in</strong>itialization network and<br />
the pretreatment traffic flow time series, the first tra<strong>in</strong><strong>in</strong>g<br />
network is carried out.<br />
Step4) The error is calculated. If the error is <strong>in</strong> the<br />
scope of the permission, the error is calculated and it<br />
turns <strong>in</strong>to Step4), otherwise it cont<strong>in</strong>ues; the error<br />
function formula:<br />
200<br />
1 2<br />
E( θ ) = ( y( t) − y( t))<br />
(9)<br />
∑<br />
2 t = 1<br />
Set the maximum error is E max<br />
= 0.035 , if E < Emax<br />
,<br />
the storage RBF neural network parameter use w ;<br />
otherwise, then a second tra<strong>in</strong><strong>in</strong>g network will be<br />
required.<br />
Step5) Adjust the adaptive learn<strong>in</strong>g rate If A previous<br />
tra<strong>in</strong><strong>in</strong>g error is recorded as En<br />
− 1<br />
, the current error is<br />
recorded as E n<br />
, then Calculate the ratio of E<br />
n<br />
to En<br />
− 1<br />
,<br />
En<br />
Sett<strong>in</strong>g constants k = 1.04 , if > k = 1.04 , then<br />
En<br />
− 1<br />
substitute βη for η to reduce learn<strong>in</strong>g rate; otherwise,<br />
replace η with γη to <strong>in</strong>crease learn<strong>in</strong>g rate.<br />
Step6) In the adaptive RBF neural network filter<strong>in</strong>g for<br />
the chaotic time series prediction <strong>in</strong> Figure 1,<br />
x( k) = x( t)<br />
t = 1, 2, , N is the <strong>in</strong>put, yk ˆ( ) = xt ˆ( ) is the<br />
output.<br />
Introduce nonl<strong>in</strong>ear feedback <strong>in</strong>to the weight<strong>in</strong>g<br />
formal to adopt Chaos Mechanisms, due to the nonl<strong>in</strong>ear<br />
feedback is vector form of weight<strong>in</strong>g variables. In order<br />
to facilitate understand<strong>in</strong>g, respectively, gives the vector<br />
w and its weight<strong>in</strong>g formal, as follows.<br />
Note<br />
Δ w l ( t+ 1) = w l ( t+ 1) −w l ( t)<br />
,<br />
ji ji ji<br />
which represents the current value of weight<strong>in</strong>g variables,<br />
then<br />
1<br />
Δ w l ( t+ 1) = w l ( t+ 1) − w l () t = −ηδ l+<br />
() t x<br />
l () t<br />
ji ji ji j i<br />
In order to speed up the learn<strong>in</strong>g process, <strong>in</strong> the right to<br />
l<br />
jo<strong>in</strong> a momentum term αΔw () t , then<br />
l 1<br />
( 1) l +<br />
( ) l ( ) l<br />
ji<br />
t ηδ<br />
j<br />
t<br />
i<br />
t α<br />
ji<br />
( t)<br />
ji<br />
Δ w + = − x + Δw (10)<br />
where α is <strong>in</strong>ertia factor. As a constant, the weight of<br />
amendments is l<strong>in</strong>ear, not <strong>in</strong>troduce chaos mechanism.<br />
then we Introduce a nonl<strong>in</strong>ear feedback (chaos<br />
mechanism on the right):<br />
1<br />
Δ w l ( t+ 1) = − ηδ l+<br />
( t) x l ( t) + g( Δ w l ( t+<br />
1)) (11)<br />
ji j i ji<br />
Expand this equation <strong>in</strong>to scalar form as follow:<br />
l l+<br />
1 l l<br />
⎧Δ wji ( t+ 1) = − ηδ<br />
j<br />
( t) xi ( t) + g( Δwji<br />
( t))<br />
⎪<br />
l l+<br />
1<br />
l l<br />
⎪Δ wji ( t+ 1 + τ) =− ηδ<br />
j<br />
( t+ τ) xi ( t+ τ) + g( Δ wji<br />
( t+<br />
τ))<br />
⎪<br />
l l+<br />
1<br />
l l<br />
⎪Δ wji ( t+ 1+ 2 τ ) = − ηδ<br />
j<br />
( t+ 2) xi ( t+ 2 τ ) + g( Δ wji<br />
( t+<br />
2 τ ))<br />
⎨<br />
⎪<br />
⎪ l l+<br />
1<br />
l<br />
⎪<br />
Δ wji ( t+ 1 + ( m− 1) τ ) =− ηδ<br />
j<br />
( t+ ( m− 1) τ ) xi<br />
( t+ ( m−1) τ )<br />
⎪ l<br />
⎩<br />
+ g(<br />
Δwji<br />
( t+ ( m−1) τ ))<br />
(12)<br />
where, feedback can take a variety of vector functions,<br />
for example:<br />
2<br />
g( x) = tanh( px)exp( − qx )<br />
or<br />
g( x) = pxexp( − q x)<br />
,<br />
<strong>in</strong> the study, p = 0.7 , q = 0.1.<br />
Step7) Us<strong>in</strong>g the new learn<strong>in</strong>g rate <strong>in</strong> Step5) and RBF<br />
network parameters with nonl<strong>in</strong>ear feedback <strong>in</strong> Step6) to<br />
calculate the new value, and tra<strong>in</strong> network aga<strong>in</strong>, then get<br />
the error and enter <strong>in</strong>to Step4), repeated tra<strong>in</strong><strong>in</strong>g until the<br />
relative error <strong>in</strong> traffic meet E < Emax<br />
.<br />
Step8) Output of each stored network parameters and<br />
tra<strong>in</strong><strong>in</strong>g error curve.<br />
V. EXAMPLE ANALYSIS AND CONCLUSIONS<br />
A. Model and Data<br />
In this paper, the chaotic time series is the object of<br />
study of the numerical simulation <strong>in</strong> Lorenz dynamic<br />
system. In 1963, the meteorologist Lorenz describe the<br />
evolution of the weather by three-dimensional<br />
autonomous equations; when the parameter σ = 10 ,<br />
8<br />
r = 28 , b = , the long-term changes <strong>in</strong> the weather<br />
3<br />
unpredictable, that is, the system presents a chaotic state,<br />
and for the first time given a strange attractor. The<br />
attractors are shown <strong>in</strong> Figure 2 (a), Figure 2 (b), Figure 2<br />
(c) and Figure 2 (d):<br />
© 2013 ACADEMY PUBLISHER