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.
1450 JOURNAL OF COMPUTERS, VOL. 8, NO. 6, JUNE 2013<br />
τ <strong>in</strong>cludes autocorrelation function method, multiple<br />
correlation function method, mutual <strong>in</strong>formation method.<br />
Embedd<strong>in</strong>g dimension m is calculated by the methods of<br />
GP algorithm, pseudo-nearest-po<strong>in</strong>t method, correlation<br />
<strong>in</strong>tegral method and Cao method.<br />
The chaotic time series prediction is based on the<br />
Takens' delay-coord<strong>in</strong>ate phase reconstruct theory. If the<br />
time series of one of the variables is available, based on<br />
the fact that the <strong>in</strong>teraction between the variables is such<br />
that every component conta<strong>in</strong>s <strong>in</strong>formation on the<br />
complex dynamics of the system, a smooth function can<br />
be found to model the portraits of time series. If the<br />
chaotic time series are{ x()<br />
t }, then the reconstruct state<br />
vector is<br />
x( t) = ( x( t), x( t+ τ ), , x( t+ ( m−1) τ ))<br />
where m ( m = 2,3, ) is called the embedd<strong>in</strong>g<br />
dimension ( m= 2d<br />
+ 1 , d is called the freedom of<br />
dynamics of the system), and τ is the delay time. The<br />
predictive reconstruct of chaotic series is a <strong>in</strong>verse<br />
problem to the dynamics of the system essentially. There<br />
exists a smooth function def<strong>in</strong>ed on the reconstructed<br />
m<br />
manifold <strong>in</strong> R to <strong>in</strong>terpret the<br />
dynamics x( t+ T) = F( x( t))<br />
, where T ( T > 0) is forward<br />
predictive step length, and F()<br />
⋅ is the reconstructed<br />
predictive model.<br />
B. RBF Neural Network Function Approximation Theory<br />
Takens embedd<strong>in</strong>g theorem states that there is a<br />
smooth mapp<strong>in</strong>g F of the F makes:<br />
x( t+ τ ) = F( x( t))<br />
(1)<br />
that is,<br />
xt ( + τ), xt (), , xt ( −( m− 2) τ) = F{[ xt (), xt ( −τ), , xt ( −( m−1) τ]}<br />
For purposes of calculation, equation (1) can be rewritten<br />
as:<br />
xt ( + τ ) = Fxt [ ( ), xt ( −τ), , xt ( −( m−1) τ]<br />
(2)<br />
where, f is the mapp<strong>in</strong>g from R M to R L . Chaos theory<br />
suggests that the chaotic time series is short-term forecast,<br />
and the essence of prediction is how to get a good<br />
approximation f on the function f . Chaotic time series<br />
determ<strong>in</strong>ed by the <strong>in</strong>ternal regularity, this regularity<br />
comes from the non-l<strong>in</strong>ear, it exhibits the time series <strong>in</strong><br />
the time delay state, this feature makes the system seem<br />
to have some k<strong>in</strong>d of memory capacity. The same time, it<br />
is difficult to demonstrate such a regularity by us<strong>in</strong>g the<br />
analytic methods; this type of <strong>in</strong>formation process<strong>in</strong>g<br />
happens to be the neural network, and the Kolmogorov<br />
cont<strong>in</strong>uity theorem <strong>in</strong> the neural networks theory<br />
provides a theoretical guarantee for the neural network<br />
nonl<strong>in</strong>ear function approximation.<br />
Theorem (Kolmogorov cont<strong>in</strong>uity theorem) Let ϕ ( x)<br />
be a non-constant and bounded monotonically <strong>in</strong>creas<strong>in</strong>g<br />
a cont<strong>in</strong>uous function; M is a compact sub-set of R n ,<br />
and f( x) = f( x1, x2, , x n<br />
) is the cont<strong>in</strong>uous real value<br />
function on M , then for ∀ ε > 0 , exists a positive <strong>in</strong>teger<br />
N and real numbers C , makes:<br />
<br />
N n<br />
f( x , x , , x ) = Cϕ( ϖ x −θ<br />
) (3)<br />
1 2<br />
∑<br />
∑<br />
n i ij j j<br />
i= 1 j=<br />
1<br />
meet:<br />
<br />
max f( x , x , , x ) − f( x , x , , x ) < ε (4)<br />
M<br />
1 2 n 1 2<br />
By the above theorem, the nonl<strong>in</strong>ear time series<br />
prediction process us<strong>in</strong>g neural network can be<br />
considered as dynamic reconfiguration, which is an<br />
<strong>in</strong>verse process. Namely, the existence of a three-layer<br />
network, the hidden unit output function, the network<br />
<strong>in</strong>put and output function is l<strong>in</strong>ear, three-layer network<br />
<strong>in</strong>put and output relation f can approximate p.<br />
Therefore, the theorem from mathematics is to ensure<br />
the feasibility of chaotic time series prediction by neural<br />
network.<br />
C. Realized Architecture of Adaptive RBF Neural<br />
Network Filter<strong>in</strong>g Predictive Model<br />
After reconstruct<strong>in</strong>g the phase space, the RBF neural<br />
networks adopt three layers networks of Figure 1. Where<br />
the <strong>in</strong>put layer has m nerve cells, the first layer feed to<br />
the second layer directly and it do not need the power<br />
process<strong>in</strong>g. r i<br />
( i = 1, 2, , L ) is the reference vector and<br />
ϖ<br />
k<br />
( i = 1, 2, , L ) is the adjustable parameters <strong>in</strong> the<br />
adaptive RBF neural network filter<strong>in</strong>g. Thus, the adaptive<br />
RBF neural network filter<strong>in</strong>g is more flexible <strong>in</strong> study<strong>in</strong>g<br />
the nonl<strong>in</strong>ear functions. The differentiation between the<br />
networks and the traditional neural networks is that the<br />
activation function is a RBF function but not the Sigmoid<br />
function. The activation function usually choose the<br />
Gauss function, the spl<strong>in</strong>e function f ( di<br />
( k )) ,<br />
where d ( k) = x( k) − r( k)<br />
. In the adaptive RBF neural<br />
i<br />
network filter<strong>in</strong>g, yk ( ) is expressed as<br />
i<br />
2<br />
L−1<br />
yˆ( k) = f ( ∑ ϖ ( k) f( d ( k)))<br />
,<br />
i=<br />
0<br />
i = 0, 2, , L−1,<br />
where f () 2<br />
⋅ is the activation function of output signal.<br />
i<br />
x(k)<br />
−1<br />
z<br />
−1<br />
z<br />
−1<br />
z<br />
r 0<br />
r 1<br />
r 2<br />
r L<br />
i<br />
ϖ 0<br />
( k)<br />
ϖ 1<br />
( k)<br />
ϖ 2<br />
( k)<br />
(k) ϖ L<br />
n<br />
∑<br />
yˆ ( k)<br />
<strong>in</strong>put hidden layer output<br />
Figure 1. Structure of adaptive RBF neural network filter<strong>in</strong>g<br />
Generally, the learn<strong>in</strong>g of the RBF neural network<br />
filter<strong>in</strong>g has three steps. If the gradient method and the<br />
Gauss activation function are adopted, the regulate<br />
formulas of RBF are shown as:<br />
© 2013 ACADEMY PUBLISHER