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.
1588 JOURNAL OF COMPUTERS, VOL. 8, NO. 6, JUNE 2013<br />
functions of the process units are differentiable, which are<br />
usually S type function (Sigmoid function) f ( x ), that is<br />
1<br />
f( x)<br />
= (1)<br />
x<br />
1 + e −<br />
The study process of BP neural network <strong>in</strong>cludes<br />
forward propagation and error back-propagation. If given<br />
some <strong>in</strong>put mode, the BP network will study for every<br />
<strong>in</strong>put mode <strong>in</strong> accordance with the followed methods.<br />
The <strong>in</strong>put mode are transferred from <strong>in</strong>put layers to the<br />
hidden layer units, by which the <strong>in</strong>put mode can be<br />
processed, the new output mode will be transferred to<br />
output layer, that is called forward propagation. If the<br />
output mode is not expected, the error signals will return<br />
along the orig<strong>in</strong> route, connection weights of neurons <strong>in</strong><br />
every layer should be corrected to make the error signals<br />
least, that is error back-propagation. Forward propagation<br />
and back propagation repeatedly, until the expected<br />
output mode can be obta<strong>in</strong>ed<br />
The learn<strong>in</strong>g process of BP network beg<strong>in</strong> from a set<br />
of random weights and thresholds, any selected samples<br />
can be <strong>in</strong>put. The output can be computed by forwardback<br />
method. Usually this error is big, the new weights<br />
and thresholds of the mode must be computed over aga<strong>in</strong><br />
by the back propagation. For all of the samples, the<br />
process should be done repeatedly aga<strong>in</strong> and aga<strong>in</strong>, to get<br />
the appo<strong>in</strong>ted accuracy. In the process of network<br />
operation, the system error and s<strong>in</strong>gle mode error can be<br />
followed. If the network learn<strong>in</strong>g successfully, the system<br />
errors will decrease with <strong>in</strong>creas<strong>in</strong>g of iterative time, at<br />
last converge at a set of steady weights and thresholds.<br />
y1<br />
y2<br />
y3<br />
x1<br />
x2<br />
x3<br />
Figure 1. the BP network model structure with three layers<br />
B. The Mathematical Pr<strong>in</strong>ciple of Back Propagation<br />
Network Model<br />
The propagation formulas for BP network study are<br />
used to adjust the weights and thresholds. In fact, the<br />
network study process is a process <strong>in</strong> which weights and<br />
thresholds of network connection are revised repeatedly<br />
accord<strong>in</strong>g to the propagation formula <strong>in</strong> the direction of<br />
least error. There are some symbol conventions:<br />
O : output of nodei ;<br />
i<br />
net : <strong>in</strong>put of node<br />
j<br />
j ;<br />
w : connected weight from node i to node j ;<br />
ij<br />
θ : threshold of node<br />
j<br />
j ;<br />
y : actual output of node<br />
k<br />
k <strong>in</strong> output layer;<br />
t : expected output of node<br />
k<br />
k <strong>in</strong> output layer.<br />
Obviously, for hidden node j :<br />
net<br />
j<br />
= ∑ wijO<br />
⎫<br />
i ⎪ ⎬<br />
(2)<br />
Oj = f( netj −θ<br />
j)<br />
⎪⎭<br />
In study process of BP algorithm, the errors of every<br />
output node can be computed accord<strong>in</strong>g to the follow<strong>in</strong>g<br />
formula:<br />
1<br />
2<br />
e= ∑ ( tk<br />
− yk)<br />
(3)<br />
2 k<br />
The connection weights can be corrected accord<strong>in</strong>g to<br />
the follow<strong>in</strong>g formula:<br />
w ( t+ 1) = w ( t)<br />
+Δ w<br />
(4)<br />
ij ij ij<br />
w<br />
In the formula, () ij<br />
t wij<br />
( t+ 1)<br />
and are separately<br />
connection weights from node j to node k at time t<br />
andt + 1 Δw<br />
;<br />
ij is variation of connection weights.<br />
In order to improve the connection weights <strong>in</strong> the<br />
Δwij<br />
gradient change direction of error E, can be<br />
computed:<br />
e<br />
Δ wij<br />
=−η ∂<br />
(5)<br />
∂ w<br />
In the formula, η is ga<strong>in</strong> factor,<br />
Thus<br />
Thus<br />
jk<br />
∂e<br />
∂w<br />
jk<br />
∂e<br />
∂e<br />
∂net<br />
=<br />
∂w ∂net ∂w<br />
∂net<br />
jk k jk<br />
∂<br />
can be computed:<br />
k<br />
= ∑ wjkOj = O (6)<br />
j<br />
∂wjk<br />
∂wjk<br />
j<br />
∂<br />
δk<br />
= ∂ net k<br />
∂e<br />
Δ w =− η =−ηδ<br />
O<br />
ij k j<br />
∂wjk<br />
When comput<strong>in</strong>gδ k<br />
, it is essential to dist<strong>in</strong>guish the<br />
output layer nodes and hidden layer nodes. If node k lies<br />
<strong>in</strong> output layer, thus:<br />
∂e<br />
∂e<br />
∂yk<br />
δk<br />
= =<br />
∂netk ∂yk ∂netk<br />
Because of<br />
∂ e<br />
∂y<br />
=− ( t −<br />
k<br />
k<br />
yk)<br />
= f ′( netk<br />
)<br />
∂yk<br />
∂netk<br />
Thus<br />
δk =−( tk −yk) f′<br />
( netk)<br />
⎫⎪ ⎬ (8)<br />
Δ wjk = η( tk − yk) f′<br />
( netk)<br />
Oj⎪⎭<br />
k<br />
(7)<br />
© 2013 ACADEMY PUBLISHER