Learning Based Primary User Activity Prediction in Cognitive Radio

Learning Based Primary User Activity Prediction in
Cognitive Radio Networks for Efficient Dynamic
Spectrum Access
Anirudh Agarwal1, Shivangi Dubey1, Mohd. Asif Khan1, Ranjan Gangopadhyay2 and Soumitra Debnath2
Department of ECE
The LNM Institute of Information Technology
Jaipur, India
{a1anirudh, shivangidb1, [email protected] }[email protected]
{ranjan_iitkgp, soumitra_deb}[email protected]
Abstract—Efficient spectrum sensing can be realized by
predicting the future idle times of primary users’ activity in a
cognitive radio network. In dynamic spectrum access, based on a
reliable prediction scheme, a secondary user chooses a channel
with the longest idle time for data transmission. In this paper,
four supervised machine learning techniques, two from ANN, i.e.
Multilayer Perceptron & Recurrent Neural Networks, and two
from Support Vector Machines (SVM), i.e. SVM with Linear
Kernel, SVM with Gaussian Kernel, have been employed to
investigate the prediction of primary activity. Poisson,
Interrupted Poisson and Self-similar traffics are used for the
analysis of licensed user environment. Data generated by each
traffic distribution is used in the training phase individually with
the help of each learning model after which, the testing is done
for the primary activity prediction. The results highlight the
analysis of the learning techniques in accordance with various
traffic statistics, and suggest the best learning model for accurate
primary user activity prediction.
Keywords-machine learning; dynamic spectrum access; data
traffic; spectrum prediction.
I.
INTRODUCTION
Efficient dynamic spectrum access in cognitive radio
networks provides transmission opportunities to the secondary
users (SU) with minimum harmful interference to the primary
users (PU) [1]. For improved spectrum management, SU needs
to even conserve its sensing energy thereby avoiding the black
space during the sensing period. This in turn suggests that
spectrum prediction must be done before spectrum sensing
where SU predicts the primary activity and senses only if the
primary predicted state is idle. After sensing, SU would
choose the channel with the longest idle time for enhanced
dynamic spectrum access (DSA).
In a CR network, it becomes difficult to have the
knowledge of the statistical distribution of primary activity in
advance. In this context, machine learning (ML) proves to be a
powerful tool for a CR system opening up versatile DSA
applications viz. spectrum sensing, spectrum occupancy
modeling, spectrum prediction, traffic pattern prediction,
spectrum scheduling etc [2]. The primary advantage of ML
over other statistical models is that it does not require a-priori
knowledge of the distributions under consideration. In
978-1-5090-1746-1/16/$31.00 © 2016 IEEE
supervised ML, input and output datasets are provided to train
the learning model while there is no provision of the
information of output in the unsupervised ML and hence the
model has to take its decision exclusively based on the input.
The significance of any ML algorithm, whether supervised or
unsupervised, actually depends on the applications where they
have been used. In the context of CR, unsupervised ML
techniques are generally used for threshold adaptation [3],
optimization of transmission parameters [4], power allocation
[5], signal classification [6] etc. On the other end, supervised
ML is used for feature extraction, spectrum prediction [7, 8],
cognitive learning engine [9] and synthesizing system
performance evaluation functions. For CR, mainly artificial
neural networks (ANN) and Support Vector Machines (SVM)
have been investigated in case of supervised ML [10].
Although SVM based feature extraction is done in [11] but
authors have mainly targeted spectrum mobility prediction.
Similarly, Hidden Markov Model (HMM) based channel
status prediction is provided in [12] but there are many
shortcomings in this scheme such as finding the optimal HMM
model and requirement of huge memory space for past
observations. So, application specific work in reference to
DSA has not been explored in sufficient detail and that too, for
comparing the performance of ANN and SVM based ML
techniques. Moreover, most of the mentioned papers are
restricted to one possible traffic model only.
In [12], authors have assumed Poisson traffic of PU
and shown the probability of error in prediction through
Multilayer Perceptron (MLP) only. In this paper, we have
analyzed the performance of four supervised ML techniques,
two from ANN, i.e. MLP & Recurrent Neural Networks
(RNN), and two from Support Vector Machines (SVM), i.e.
SVM with Linear Kernel (LSVM), SVM with Gaussian
Kernel (GSVM) for the prediction of primary activity as
governed by several well known network traffic models
namely, Poisson, Interrupted Poisson (IP) and Self-similar
(SS) traffic. These traffic models reasonably capture the traffic
characteristics existing in most of the types of the wireless
networks. The motivation behind using these ANN based ML
techniques is their ability of nonlinearity fitness and adaptive
behavior towards varying surrounding environment; while
SVM based ML techniques have been utilized due to their
capability of generalization and proven robustness towards
noise and local outliers[10]. The scope of the present work is
to investigate, how the statistical variation of the primary user
data traffic could impact the accuracy of the prediction
provided by various learning algorithms in estimating the
average length of OFF period of the PU.
The paper is organized as follows: Section II
discusses about the system model and the methodology. The
description of various ML techniques is provided in Section
III. The traffic models and their statistical realization are
described in Section IV. The performance analysis and the
results are shown in Section V. Finally, Section VI concludes
the paper.
II.
SYSTEM MODEL AND METHODOLOGY
In general, the available radio spectrum consists of
multiple primary channels whose activity follows different
traffic patterns. In our model, for simplicity we assume a
single primary channel whose channel state information (CSI)
is used by the system for learning based prediction of future
primary activity. The working flow diagram is given in Fig. 1
and the methodology is described in the following steps:
Step 1. Slot-wise PU data is taken corresponding to different
data traffic models.
Step 2. Based on a particular ML technique, the training is
done using sufficient PU data.
Step 3. Once the model is learned through the training phase,
the future primary activity is predicted.
Step 4. The remaining data, other than that is used for training,
is utilized for testing the prediction accuracy and the mean
square error in estimating the average length of the OFF
) of PU activity is determined.
period (
This model can be used further in multiple primary
channels scenario where the CR would utilize the information
, predicted by the most accurate ML technique, of all
of
the channels and choose the one with the highest value of the
; thereby resulting in an optimized CR sensingtransmission strategy for DSA.
III.
ML PREDICTION TECHNIQUES
In this section, a brief description of the ML prediction
techniques used in the present study is provided.
A.1. Multilayer Perceptron Neural Network Based Prediction
An MLP is a feed-forward network of
simple neurons called perceptrons. It consists of three or more
layers (an input and an output layer with one or more hidden
layers) of nodes in a directed graph. Each node excluding the
nodes at the input layer is a computing unit i.e. perceptron.
The perceptron computes single output from multiple realvalued inputs by forming a linear combination of their
input weights and then putting the output through some
nonlinear activation function. Mathematically, this can be
written as:
+ )(1)
= (
where y is the output vector, is the activation function, w is
the weight vector, x is the input vector and b is the bias.
Channel State
Information
(CSI)
Machine Learning
(ML) Technique
(Training+Testing)
Prediction accuracy
of ML technique(s)
and MSE of
predicted
estimate
Primary activity
Prediction
Figure 1. Methodology
The activation function is often chosen to be the
logistic sigmoid or tangent hyperbolic. The number of hidden
layers and the number of neurons in each layer vary according
to the application.
MLP networks are trained using the backpropagation algorithm (BPA) [13]. In training phase, the error
between the actual and predicted target value is propagated
backward to the network and compute the gradient of the
change in error with respect to changes in weight values;
weights are further adjusted to reduce the error.
For the implementation of the MLP algorithm, we have used
the MATLAB neural network toolbox with 4 inputs in the
input layer, two hidden layers consisting of 15 and 20 neurons
respectively and one neuron in the output layer. We have used
the Tangent Sigmoid and Purelin as the activation functions
respectively for the hidden layer and the output layer neurons.
The learning rate and momentum constant for the gradient
descent method in BPA are taken as 0.2 and 0.9 respectively.
These values are chosen after rigorous cross-verification for
optimality.
A.2. Recurrent Neural Networks based Prediction
An RNN is a special class of ANN obtained by
allowing the cyclical connections between neurons. The only
proviso is that one of the two parts of the input to the network
is the pattern of activation over the network’s own hidden
units at the previous time steps. RNN learns internal
representation of time by employing hidden units and a
learning algorithm in feed-forward network; the hidden units
develop internal representations for the input patterns that
recode those patterns in a way which enable the network to
produce the correct output for a given input. The context units
serve the purpose of storing the previous internal state and the
hidden units map both external input as well as the previous
internal state to the desired output. This property of the Elman
type RNN provides very important advantage in modeling
sequential data. To train RNN, BPA over time is used. The
solution of the back-propagation through time approach is to
"unfold" the recurrent network in time, by stacking identical
copies of the RNN, and redirecting connections within the
network to obtain connections between subsequent copies.
This gives a feed-forward network, which is amenable to
BPA. For faster convergence, an extended Kalman filtering is
used to estimate the weights of RNN [14].
In this work, we exploit the use of very simple Elman
RNN, containing the input layer with 4 neurons, a single
hidden layer with 6 neurons and an output layer with one
neuron.
B. Linear and Gaussian Support Vector Machines Based
Prediction
SVM is a maximum margin discriminative classifier,
meaning that it learns a decision boundary that maximizes the
distance between samples of the two classes, given a kernel.
The distance between a sample and the learned decision
boundary can be used to make the SVM a "soft" classifier.
In the present implementation, we have used the linear kernel
based SVM (LSVM) and the Gaussian kernel based SVM
(GSVM). The training feature and the response vectors can be
represented as Z = (Ti, xi) where Ti ∊ {-1, 1}. The two classes
are separated by a hyperplane denoted as H: wTx + b = ε,
where w are the weight vectors, b is the bias and =
∑
is a slack variable vector whose 1-norm is the penalty
term. The two classes, separated by two hyperplanes, are given
by:
1 ℎ + >1−
(2)
=
−1 ℎ + < −1 +
With a soft margin, the optimization problem for the
SVM can be defined as follows:
min
( , , )∊
‖ ‖ +
∑
(3)
+ )>1− ,
= 1,2, … , .
where n denotes the space dimension; m is the total number of
feature vectors, C > 0 is a regularization parameter that
and the
balances the weight of the penalty term ∑
margin maximization term ‖ ‖ [15].
For training and testing purposes, we have utilized
the widely used library for SVMs, LIBSVM [16] where the
algorithm, with 4 as the input order, is iterated until the
minimum tolerance value (taken 0.0001 in this work) is
achieved.
. . (
IV.
DATA TRAFFIC PREDICTION
This section deals with three types of data traffics for
our analysis for predicting the average length of the OFF
period duration of PU traffic.
A. Poisson Traffic
The Poisson model is one of the most widely used
traffic models which has traditionally been used to model the
voice data. In a Poisson process, the inter-arrival times are
exponentially distributed with a rate parameter λ. The
distribution has its mean and variance both equal to λ.
The probability density function of the Poisson traffic
model is given by
(4)
( ) =
where λ is the rate parameter.
B. Interrupted Poisson Traffic
The Interrupted Poisson Process (IPP) is a two state
process. This kind of traffic is found in computers, e-mails,
etc., i.e., there is heavy traffic for some time and then no
traffic for some time. In one state, traffic arrival occurs as per
Poisson process with mean arrival rate λ1 and in other state,
the arrival occurs as per Poisson process with mean arrival rate
λ2. Therefore, IPP is a switched Poisson process with λ2 as
zero. The system may switch from one state to another
alternatively or it may choose the state as per certain
probability. Here the packets are generated by one Poisson
source of mean arrival rate λ1, which is independent of the
system state.
C. Self-similar Traffic
Self-similar (SS) processes are the stochastic processes
that exhibit the phenomenon of self-similarity. Self-similarity
is a well known feature in the Internet traffic. It is
characterized by long range dependence of traffic, burstiness
and correlation over varying time scales.
We have generated the self-similar traffic with the
help of Pareto distribution. For an infinite number of sources,
this results in an SS process. The probability density function
of this distribution is given by
( ) =
,
≥ (5)
where β > 0 and α > 0 are respectively the shape and the scale
parameters of the distribution. The mean value of the
for the shape parameter value of β
distribution is ( ) =
> 1. Moreover, the variance of the random variable x is infinite
if β ≤ 2. SS traffic is quantified by the Hurst (H) parameter
given by H = (3- β)/2. For self-similarity, H ∊ (0.5, 1). If H =
0.5, the traffic is Poisson and an increase in value of H,
increases the burstiness in the traffic.
V.
PERFORMANCE EVALUATION, RESULTS AND
DISCUSSION
For evaluating the performance of learning based
estimate
prediction techniques for reliable prediction of
in different traffic scenarios, we have used two performance
measures i.e. probability of error in predicting the busy state,
Proberr(busy state), and the mean-square error in predicting the
average length of the OFF duration of primary activity, viz.
. The significance of Proberr(busy state) lies in
MSE of the interference caused by the CR user to PU. More
Proberr(busy state) would lead to more tendency of CR user to
violate the interference constraint. However, the overall
utilization efficiency of SU with better channel allocation
. Moreover, the
strategy, is governed by MSE of
for all the ML
prediction accuracy and MSE of
techniques are calculated and depicted in Table 1.
The traffic intensity for random ON-OFF data traffic
is defined as:
(6)
+
where TON and TOFF are respectively the average time for
which the primary user is busy and idle.
The mean-square error in predicting the average
length of OFF duration of primary is calculated as:
=
=
1
(
( )−
( )) (7)
Figure 2. Proberr(busy state) vs. ρ for different data traffic using different ML
techniques .
Figure 4. MSE of
techniques.
Figure 3. Proberr(busy state) vs. ρ for SS traffic using different ML techniques.
Figure 5. MSE of
where N is the total number of simulation intervals,
(n) is the average length of OFF duration of PU
activity in predicted data at the nth simulation interval and
(n) is the average length of OFF duration
similarly,
of PU activity in original data at the nth simulation interval.
The other performance metrics i.e., the prediction
accuracy (PA) is defined as:
= (1 − ) × 100(8)
where Pe is the overall probability of error, i.e. when the busy
state is predicted as idle and vice versa, in the prediction by an
ML technique. The prediction analysis is done using30% of
the primary user data for training while the rest of the data is
utilized for testing the trained model. The total number of
traffic slots for primary data is taken as 50000 in this work.
The results are shown for 30% training length for different
traffic intensities and all data traffic models. Moreover, as the
characteristics of particular data traffic might change with
time, we have evaluated the performance of all the considered
parameters after averaging over sufficient number (50 in this
work) of simulation intervals.
Figures 2 and 3 depict the probability of error in
predicting the busy state of the primary for different traffic
types utilizing the four learning schemes. It can be observed
that for all types of data traffics, the probability of error in
predicting the busy state decreases as we increase the traffic
intensity. As
increases, the number of times the channel
being ON also increases, hence the probability of predicting
busy as idle tends to decrease. Moreover, it is found that the
MLP prediction technique performs slightly better than the
other three techniques in this case.
, for
Fig. 4 shows the variation of MSE of
Poisson and IP traffic against . The decreasing nature is
attributed to the fact that, there are less number of transitions
from busy state to idle and vice-versa with increase in . This
leads to more dependency of future states on the present and
previous states, thereby suggesting a decrease in the prediction
error and an improvement in the prediction accuracy.
However, in this case, LSVM based prediction turns out to be
.
the best in predicting
With the same reasoning, similar pattern was expected
for bursty SS traffic in Fig. 5. But it is observed that prediction
is not so accurate for high traffic intensity. This may be due to
the heavy burstiness and strong OFF period correlation in the
data traffic. Nevertheless, LSVM is found to perform
vs. ρ for different traffic using different ML
vs. ρ for SS traffic using different ML techniques.
Table 1. Comparison of ML Techniques for different training lengths and traffic intensities in 3 traffic scenarios
Percentage of Training data
Traffic Intensity
0.5
30%
0.8
ML Technique
MSE
PA
MSE
PA
MLP
2.032e-1
91.558
2.034e-1
87.040
1.195e-1
85.104
LSVM
8.828e-7
92.850
2.233e-6
87.470
5.817e-7
85.398
RNN
4.117e-5
91.925
4.710e-3
87.395
3.296
85.012
GSVM
2.012e-1
91.890
2.055e-1
87.290
6.597e-7
85.368
MLP
5.360e-4
92.645
5.223e-3
89.950
2.140
86.746
LSVM
7.382e-8
92.931
1e-7
90.003
4.082e-7
87.068
RNN
1.050e-7
92.924
6.052e-7
89.990
2.141
85.920
GSVM
1.088e-7
92.919
7.330e-7
89.988
8.850e-7
87.054
CONCLUSION AND FUTURE WORK
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
REFERENCES
[2]
SS
PA
The present work highlights the applicability of
learning based models for enhanced dynamic spectrum access
in different traffic scenarios, enabling accurate prediction of
primary OFF period, which is an important parameter for
characterizing the spectrum utilization. Among the learning
models considered, LSVM has been found to behave
consistently and accurately for all the traffic patterns.
Although, a single primary channel case has been considered
here but the same analysis would easily be extended for the
case of multiple channels of PU, leading to the most accurate
and improved dynamic channel allocation to SU.
Further work on the model based learning through
deep belief based recurrent neural networks are currently
under investigation for prediction of the spatio-temporal
spectrum occupancy probability map using real time spectrum
measurement data for efficient DSA applications in CR
systems.
[1]
IP
MSE
uniformly and reliably for the SS traffic.
Finally, Table 1 provides a comparison of the
performance of ML techniques under various simulated
and prediction accuracy of
conditions in terms of MSE of
the algorithms under consideration. It is clearly observed that
even for training length as low as 30%, LSVM outperforms all
the other techniques for different data traffic considered.
VI.
Poisson
Höyhtyä, Marko, Sofie Pollin, and Aarne Mämmelä. "Improving the
performance of cognitive radios through classification, learning, and
predictive channel selection." Advances in Electronics and
Telecommunications 2.4 (2011): 28-38.
Azmat, Freeha, Yunfei Chen, and Nigel Stocks. "Analysis of Spectrum
Occupancy Using Machine Learning Algorithms." arXiv preprint
arXiv:1503.07104 (2015).
[14]
[15]
[16]
Gong, Shimin, et al. "Threshold-learning in local spectrum sensing of
cognitive radio." Vehicular Technology Conference, 2009. VTC Spring
2009. IEEE 69th. IEEE, 2009.
Zhao, Guodong, et al. "Spatial spectrum holes for cognitive radio with
directional transmission." Global Telecommunications Conference,
2008. IEEE GLOBECOM 2008. IEEE. IEEE, 2008.
Li, Liying, et al. "Simplified relay selection and power allocation in
cooperative cognitive radio systems." Wireless Communications, IEEE
Transactions on 10.1 (2011): 33-36.
Clancy, T. Charles, Awais Khawar, and Timothy R. Newman. "Robust
signal
classification
using
unsupervised
learning." Wireless
Communications, IEEE Transactions on 10.4 (2011): 1289-1299.
Xing, Xiaoshuang, et al. "Spectrum prediction in cognitive radio
networks."Wireless Communications, IEEE 20.2 (2013): 90-96.
He, An, et al. "A survey of artificial intelligence for cognitive
radios."Vehicular Technology, IEEE Transactions on 59.4 (2010): 15781592.
Dong, Xu, et al. "A learner based on neural network for cognitive
radio."Communication Technology (ICCT), 2010 12th IEEE
International Conference on. IEEE, 2010.
Bkassiny, Mario, Yang Li, and Sudharman K. Jayaweera. "A survey on
machine-learning techniques in cognitive radios." Communications
Surveys & Tutorials, IEEE 15.3 (2013): 1136-1159.
Wang, Yao, et al. "SVM-Based Spectrum Mobility Prediction Scheme in
Mobile Cognitive Radio Networks." The Scientific World Journal 2014
(2014).
Tumuluru, Vamsi Krishna, Ping Wang, and Dusit Niyato. "A neural
network based spectrum prediction scheme for cognitive
radio." Communications (ICC), 2010 IEEE International Conference on.
IEEE, 2010.
Rumelhart, David E., Geoffrey E. Hinton, and Ronald J. Williams.
"Learning representations by back-propagating errors." Cognitive
modeling 5 (1988): 3.
Williams, Ronald J. "Training recurrent networks using the extended
Kalman filter." Neural Networks, 1992. IJCNN., International Joint
Conference on. Vol. 4. IEEE, 1992.
Burges, Christopher JC. "A tutorial on support vector machines for
pattern recognition." Data mining and knowledge discovery 2.2 (1998):
121-167.
http://www. csie.ntu.edu.tw/~cjlin/libsvm.