Prediction of Spectrum based on Improved RBF Neural Network
in Cognitive Radio
Shibing Zhang
1
, Jinming Hu
1
, Zhihua Bao
1
and Jianrong Wu
2
1
School of Electronics and Information, Nantong University, Nantong, China
2
Library, Nantong University, Nantong, China
Keywords: Cognitive Radio, Spectrum Prediction, K-means Clustering, RBF Neural Network.
Abstract: Spectrum prediction is a key technology of cognitive radio, which can help unlicensed users to determine
whether the licensed user’s spectrum is idle. Based on radial-basis function (RBF) neural network, this
paper proposed a spectrum prediction algorithm with K-means clustering algorithm (K-RBF). This
algorithm could predict the spectrum holes according to the historical information of the licensed user’s
spectrum. It not only increases the veracity of spectrum sensing, but also improves the efficiency of
spectrum sensing. Simulation results showed that this prediction algorithm can predict the spectrum
accessing of the licensed user accurately and the prediction error is only one-third of that of the RBF neural
network.
1 INTRODUCTION
The electromagnetic spectrum has been exclusively
allocated to different wireless services by
government, although some of the frequency bands
in the spectrum are unoccupied most of the time or
only partially occupied. How to maximize the use of
the existing spectrum resources is an urgent problem
to be solved. Cognitive radio (CR) is a kind of
intelligent spectrum sharing technology, which can
rely on artificial intelligence support to adjust the
transmission parameters (such as transmission
power, data rate, carrier frequency, etc.). CR can
effectively use idle spectrum and greatly reduce the
restriction to the development of wireless
technology by the spectrum and the limitation of
bandwidth.
CR network is composed of two parts of the
users – licensed user (also known as primary user)
and unlicensed user (also known as second user). In
each time slot, the unlicensed user must perceive the
short-term activities of the licensed user and access
slot when it is idle (the idle time slot is also known
as spectrum holes). To minimize the interference to
licensed users, unlicensed users need a reliable
spectrum sensing mechanism. Spectrum prediction
is important to effective spectrum of CR network
and has become a hot topic in CR. A prediction
model using sliding window was established to
predict licensed users’ future spectrum activity [1].
This model sets a threshold value through the
adaptive filter. The frequency band which is lower
than the threshold will be set to be unreliable and do
not allow unlicensed users to access this band. A
multilayer perceptron for spectrum prediction was
proposed [2]. However, the multilayer perceptron
uses traditional unconstrained minimization method
to achieve minimization of the error function.
Therefore, it inevitably has local minima problem.
Subsequent studies have proposed ON-OFF,
Blackman window, POMDP, and other prediction
mechanisms (Federal Communications Commission,
2002); (Acharya et al., 2006); (Jianli et al., 2011). In
(Marko et al., 2008), a dynamic spectrum access
algorithm based on probably density estimation was
proposed to predict channel state with flexibility and
availability.
However, when the CR node sensing the
spectrum, it will detect the whole spectrum
concerned every time. It will consume a lot of
network resources. We addressed the problem in this
paper and proposed a spectrum prediction algorithm
using radial-basis function (RBF) neural network
based on K-means clustering algorithm (K-RBF). In
the algorithm, the spectrum holes are predicted
according to the licensed users’s historical
information. Then, appropriate spectrum bands are
chosen for the unlicensed users to detect. It can
243
Zhang S., Hu J., Bao Z. and Wu J..
Prediction of Spectrum based on Improved RBF Neural Network in Cognitive Radio.
DOI: 10.5220/0004537002430247
In Proceedings of the 10th International Conference on Signal Processing and Multimedia Applications and 10th International Conference on Wireless
Information Networks and Systems (WINSYS-2013), pages 243-247
ISBN: 978-989-8565-74-7
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
greatly reduce the resources consumed in spectrum
sensing.
The rest of the paper is organized as follows. In
Section II, we present the system model of spectrum
prediction with RBF neural network. In Section III,
we propose the improved prediction algorithm of
spectrum. In Section IV, we provide the simulations
for the improved prediction algorithm and
demonstrate the effect of spectrum prediction.
Finally, Section V concludes this paper.
2 SYSTEM MODEL
Whether the licensed user’s spectrum is idle can be
modelled as a binary series prediction problem. We
design the binary series predictor using neural
networks. Neural networks are nonlinear parametric
models which create a mapping function between
the input and output data.
The most basic form of RBF neural network is a
three-layer forward network, which includes the
input layer, the hidden layer, and the output layer, as
shown in Figure 1. The input layer has some source
nodes (perception units) connecting to the external
environment, while the hidden layer has a variable
number of neurons (the optimal number is
determined by the training process). The neurons in
the hidden layer contain Gaussian transfer functions
whose outputs are inversely proportional to the
distance from the centre of the neuron. The output
layer produces response to the input mode.
Let us assume that it has N input nodes, M
hidden nodes, and one output node. Consider that in
the RBF neural network structure, the network input
vector is
12
,,,
T
N
x
xxX
(1)
The network radial base vector is
12
,,,
T
M
hh h H
(2)
where
j
h is the Gaussian basis function
2
2
exp 1, 2, ,
2
j
j
j
hjM
b
XC
(3)
and
j
C
and
j
b
are the center and width of the
j
th neuron in the hidden layer, respectively,
.
denotes the Euclidean distance,
12
,,,
T
jjj jN
cc c
C
(4)
The network base width vector B can be given as
12
,, ,
T
M
bb b B
(5)
The network weight vector is
12
,,,
T
M
ww w W
(6)
The output of the network is
11 2 2mMM
yk whwh wh
(7)
The RBF is used as a hidden unit “base” and
constitutes the hidden layer space. The input vector
is transformed in the hidden layer and low-
dimensional model input data are transformed to the
high-dimensional space, making the linear
inseparable problem in low-dimensional space
become linear separable in high-dimensional space.
However, the initial values of the centers of hidden
layer nodes and the width of base function will
affect the prediction ability of the network.
Therefore, selecting appropriate values for the two
initial parameters can improve the prediction
accuracy of the network. In this study, we have used
K-means clustering algorithm to obtain the values of
the centers of the hidden layer nodes and the width
of the base function, then construct and train a more
accurate RBF neural network.
3 SPECTRUM PREDICTION
3.1 K-means Clustering Algorithm
K-means algorithm (Zhao et al., 2007) is a clustering
algorithm based on the sum of error square criterion.
First, it randomly selects K points from the data as
the initial cluster center. Then, it calculates the
distance from each sample to each center of the
clusters and assigns samples to categories whose
cluster center is nearest to them. Subsequently, it
calculates the average of each newly formed cluster
data to obtain a new cluster center. If there is no
change in the adjustments between the adjacent two
cluster centers, then it is the end of sample adjusts
and clustering function is converged. If there are
some changes, then the allocation and update steps
are repeated until the clustering function converges.
A characteristic of this algorithm is to examine
whether the classification of each sample is correct
in each iteration. If the classification is not correct,
then it should be adjusted. After adjusting the whole
WINSYS2013-InternationalConferenceonWirelessInformationNetworksandSystems
244
1
x
2
x
N
x
1
h
2
h
M
h
1
w
2
w
M
w
m
y
Figure 1: RBF neural network structure.
sample, the cluster center is modified, and the next
iteration is carried out.
3.2 K-RBF Prediction Algorithm
Assume that K is the number of iterations of the
network, the K
th
iteration clustering center
is
12
,,
M
ck ck c k , and the corresponding
clustering domain is
12
,,
M
wk wk w k .
Through the K-means algorithm, the center of
hidden layer nodes C and base function width B of
RBF neural network can be determined.
First, the first M samples are chosen from the
whole samples input as the initial cluster centers.
The center values of the M samples cannot be same.
Then, set
1k . The distance d from the selected
samples and the cluster centers is calculated as
, 1,2, , , 1,2, ,
ji
dXcki Mj N
(8)
where
j
X is the input sample. The samples are
assigned based on the minimum distance rule. When
min , 1, 2, ,
ji
i
iXcki M
,
j
X
is
assigned to category i (
kwX
ij
). After
classification, the new cluster centers of all
categories are recalculated as
1
1,1,2,
i
i
xwk
ck x i M
N
(9)
If
kckc
ii
1 , then the classification is
repeated and the steps are updated. If
kckc
ii
1 , each hidden node’s base width
i
b is determined according to the distance between
each clustering centers. The expression of
i
b is
ii
db
(10)
where
i
d is the distance between the i
th
clustering
center from the centers of the other sample data,
which is the nearest to i
th
clustering center, and
is
the overlap coefficient.
i
d is expressed as
min
iji
i
dcck
(11)
Then, the hidden nodes’ output is calculated by the
Gaussian basis function, according to (3).
The K-RBF prediction algorithm above is
summed as following:
(1) Initialization setting. Select the first M input
samples from samples. The values of the centers
of the h sample cannot be the same. Assume that
the number of iterations is
1k .
(2) Calculate the distance d from the selected input
samples to the clustering center according to (8).
(3) Classify the input samples
j
X according to the
minimum distance rule.
(4) Recalculate the new cluster center according to
(9). If the two cluster centers are not equal,
repeat the classification and the update steps.
(5) Obtain the distance between each cluster centers
according to (11), and determine the base width
vectors according to (10). The outputs of the
hidden nodes are obtained according to (3) and
(7).
4 SIMULATION AND ANALYSIS
The channel state is divided into two types:
occupation (in the binary sequence using “1”) and
idle (with “0”). In the simulation, we use the m
sequence and Gold sequence to simulate the channel
state occupied by licensed user respectively. We
took the first 350 data to train the neural network
PredictionofSpectrumbasedonImprovedRBFNeuralNetworkinCognitiveRadio
245
and the last 70 data as the test data to test the neural
network. First, we used m sequence to simulate. The
comparison of the prediction data with the actual
data is shown in Figure 2, and the prediction error is
presented in Figure 3. They showed that K-RBF can
not only accurately predict spectrum occupancy state
but the prediction error is also very small.
0 20 40 60
0
0.2
0.4
0.6
0.8
1
1.2
1.4
m sequence
Spectrum occupancy stat
e
K-RBF prediction
actual data
Figure 2: Comparison of prediction data with actual data
under m sequence.
0 20 40 60
-5
0
5
10
15
x 10
-15
m sequence
Error
Figure 3: Prediction error of m sequence.
0 20 40 60
0
0.2
0.4
0.6
0.8
1
1.2
1.4
Gold sequence
Spectrum occupancy state
K-RBF prediction
actual data
Figure 4: Comparison of prediction data with actual data
under Gold sequence.
0 20 40 60
-2
0
2
4
6
8
x 10
-15
Gold sequence
Error
Figure 5: Prediction error of Gold sequence.
0 20 40 60
-5
0
5
10
15
x 10
-15
msequence
Erro
r
K-RBF
RBF
Figure 6: Prediction error comparison.
To verify the robustness of K-RBF algorithm to
predict the spectrum occupancy state, we simulated
the K-RBF algorithm with Gold sequence also. The
comparison of the predicted data and the actual data
is shown in Figure 4, and the prediction error is
presented in Figure 5.
From the simulation results of Figure4 and
Figure 5, it was found that the prediction error of
Gold sequence is also very small. It implies that the
K-RBF algorithm has the robustness to randomicity
of licensed user accessing spectrum.
Figure 6 compares the prediction error of m
sequence simulated by RBF with K-RBF. It showed
that the prediction accuracy of the K-RBF is higher
than that of RBF neural network. And, the prediction
error of the K-RBF was found to be only one-third
of that of the RBF neural network.
5 CONCLUSIONS
In this paper, we proposed a spectrum prediction
WINSYS2013-InternationalConferenceonWirelessInformationNetworksandSystems
246
algorithm: K-RBF algorithm. We used K-means
clustering algorithm to obtain the hidden nodes
center and base function width of the RBF neural
network. Then, we train and form more appropriate
neural network. Through simulation, we found that
K-RBF algorithm can achieve better predicting
precision. Thus, we can use the prediction
information to sensing the licensed user spectrum
more simply. And, it will reduce the resource
consumed.
ACKNOWLEDGEMENTS
This work is supported by the National Science
Foundation of China under grant 61071086 and the
National Science Foundation of Nantong University
under grant 11Z061.
REFERENCES
Federal Communications Commission. “Spectrum policy
task force,” ET Docket no. 02-135, Nov. 2002.
Acharya P. A., Singh S., Zheng H., (2006) Reliable open
spectrum communications through proactive spectrum
access. Proc of the 1st International Workshop on
Technology and Policy for Accessing Spectrum
(TAPAS06):1-8.
Zhao Jianli, Wang Mingwei, Yuan Jinsha, (2011). Based
on neural network spectrum prediction of cognitive
radio. 2011 International Conference on Electronics,
Communication and Control (ICECC): 762-765.
Marko H., Sofue P., Aarne M., (2008). Performance
improvement with predictive channel selection for
cognitive radios. Proc of the 1st International
Workshop on Cognitive Radio and Advanced
Spectrum Management.1-5
Stefan G., Lang T., Brain M., (2008). Interference-aware
OFDMA resource allocation: a predictive approach.
Proc of IEEE Military Communications Conference.
1-5.
Zhao Q., Tong L., Swami A. et al., (2007). Decentralized
cognitive MAC for opportunistic spectrum access in
Ad hoc networks: a POMDP framework. IEEE J
Select Areas Communication: Special Issue Adaptive,
Spectrum Agile Cognitive Wireless Networks,
25(3):589-600
S. Broomhead, D. Lowee, (1988). Multivariable function
interpolation and adaptive networks. Complex system.
(2): 321~355.
Khaled Alsabti, Sanjay Ranka, Vineet Singh, (1997). An
efficient K-means clustering algorithm. Electrical
Engineering and Computer Science. 1(1):43-39.
PredictionofSpectrumbasedonImprovedRBFNeuralNetworkinCognitiveRadio
247
