svm and anfis as channel selection models for the spectrum ... · svm and anfis as channel...
TRANSCRIPT
Contemporary Engineering Sciences, Vol. 10, 2017, no. 10, 475 - 502
HIKARI Ltd, www.m-hikari.com https://doi.org/10.12988/ces.2017.7438
SVM and ANFIS as Channel Selection
Models for the Spectrum Decision Stage in
Cognitive Radio Networks
Danilo Alfonso López Sarmiento
Universidad Distrital Francisco José de Caldas
Bogotá, Colombia
Leydy Johana Hernández Viveros
Universidad Internacional de La Rioja
Madrid, Spain
Edwin Rivas Trujillo
Universidad Distrital Francisco José de Caldas
Bogotá, Colombia
Copyright © 2017 Danilo Alfonso López Sarmiento et al. This article is distributed under the
Creative Commons Attribution License, which permits unrestricted use, distribution, and
reproduction in any medium, provided the original work is properly cited.
Abstract
The spectrum decision stage in cognitive radio (CR) wireless networks is in
charge of selecting and assigning channels to secondary users (SUs); an optimum
selection depends on: primary user (PU) characterization, the way in which SU
requests are processed at the base station (BS), and the way in which available
spectrum resources are managed before they are assigned. This paper proposes the
use and evaluation of the artificial-intelligence-based Support Vector Machine
(SVM) and Adaptative Neuro-Fuzzy Inference System (ANFIS) methodologies in
order to appropriately classify potential resources (free spectrum in licensed
frequency bands) available on the network so that they are assigned to SUs in an
opportunistic fashion according to quality of service (QoS) requirements, In the
characterization sub-stage, the methodology includes the use of a database with
real information on PU behavior in multiple spectrum bands; the generation via
476 Danilo Alfonso López Sarmiento et al.
simulation of potential SUs with Best Effort (BE) or Real Time (RT) service
quality criteria for the opportunistic assignment of free spectrum; and the
development and simulation of the entire system on MATLAB. One of the
relevant contributions is the determination of whether the use of methods based on
supervised learning is a good option for improving resource assignment
performance in CR. The main conclusion is that implementing a proactive
strategy in the spectrum decision and channel classification stage optimizes the
performance of a cognitive network because the time needed to assign spectrum
resources to SUs is reduced, thus reducing the amount of collisions, which implies
a more efficient data transmission.
Keywords: ANFIS, cognitive radio, proactive strategy, reactive strategy, base
station, spectrum decision, SVM
1 Introduction
The spectrum decision stage in CR is critical because it's in charge of selecting the
best channels available for SU data transmission according to QoS requirements,
seeking to minimize potential collisions between PUs and SUs; success depends
on various factors, including, on the one hand, having an algorithm that is able to
predict with a high level of accuracy future instants in time in which a PU isn't
using the licensed spectrum so that it may exploited in an opportunistic fashion by
non-licensed users (for data transport from a source to a destination) and, on the
other hand, appropriately choosing the way in which SU-channel pairs are
selected and assigned. In turn, such assignment involves defining the way in
which requests made by SUs to the BS are handled, and establishing the manner
in which available CR channels are classified. The method for processing requests
from cognitive nodes can be implemented with a reactive strategy (in which
requests are handled when they reach the BS), or a proactive strategy (in which
the characteristics of the required channels are known before the arrival of
re-quests at the BS by calculating the probability of, or predicting, the arrival of
an SU). Organizing free or under-utilized channels involves classifying them
according to variables such as the available bandwidth, the transmission delay, the
estimated channel availability time, the probability of channel availability, and
other metrics. The above discussion gives rise to the following question: can the
spectrum band selection and assignment process be optimized so as to reduce the
amount of collisions between PUs and SUs? In comparison with other
state-of-the-art proposals, the differentiating elements of this paper’s approach are:
the development of a spectrum selection and assignment model based on PU
characterization using neural networks; the use of a proactive strategy; and the
classification of free frequencies with ANFIS and SVM, seeking to reduce the
time needed to efficiently assign available resources. It’s important to point out
that the main goal of this paper is to socialize the results of using ANFIS and
SVM as spectrum classifiers in the GSM band, and the impact of such use on system organization, without neglecting the other aspects that influence the improve-
SVM and ANFIS as channel selection models 477
ment of the spectrum decision stage in CR. The rest of this paper is organized as
follows: a section with a review of the state of the art; a description of the block
diagram for the proposed model; a description of the characterization (or
forecasting) algorithm of spectrum use; a description of the way in which the
proactive strategy for SU request arrival was modeled; the design of the SVM and
ANFIS classifiers; and an explanation of the channel assignment algorithm. Lastly,
the results obtained are analyzed, a discussion section is included, and the
conclusions are laid out.
2 Scientific Review
Spectrum decision with a centralized topology refers to the ability of a BS to
select and assign the best spectrum band available on a CRN (Cognitive Radio
Network) avoiding interferences with the PU; to begin with, this implies the
characterization of the PU because there is no guarantee of channel availability for
full transmission by a SU [1]. Multiple proposals exist for modeling and
forecasting PUs, such as those described in [2], [3], [4], [5], [6], [7], [8], [9], [10],
[11], [12], [13], [14]; a general summary of the main methodologies used is
shown in (Fig. 1).
Figure 1. Classification of models for the characterization of PUs. Modified from
[14].
According to [15], [16], it’s necessary to continue to propose and design
alternatives that improve the efficiency of prediction models that reduce to a greater
extent potential collisions between PUs and SUs, which would optimize the system and increase the probability of a large-scale real implementation of the CRN paradigm.
Modeling and/or prediction of PUs
Based on Queuing
theory
Learning-based
Based on time series
Additional methodologies
Bayesian
networks
Neural networks
SVM
M/G/1
Bernoulli
processes
ARIMA
Based on statistics
Poisson modelling
First-difference filter clustering
Based on
measured data
478 Danilo Alfonso López Sarmiento et al.
Intelligent algorithms such as LSTM (Long Short-Term Memory), ANFIS
(Adaptative Neuro-Fuzzy Inference System), Reinforcement Learning, and
Decision Tree are viable options for use in CR on account of their ability to learn.
Another criterion in band selection is the manner in which a BS processes requests
by SUs (Fig. 2). In the reactive strategy, the negotiation of characteristics that
governs the sending of data is processed after the arrival of a request; this is a
conventional method used in the state of the art.
In the proactive strategy, channel selection and assignment is decided before the
arrival of a SU by reserving the resource; the disadvantage is there is a dependency
on the accuracy of the model used to estimate the arrival of a SU for a specific
application (BE or RT); if the arrival probability calculation or the prediction is
poor, the amount of hits will be very low, and the efficiency that is obtained will be
insufficient.
Figure 2. Execution strategies for SU requests.
With regards to the selection of the best spectrum (SU-channel pairs), there has
been research using the reactive strategy, as shown in [5], [11], [17], [18], [19],
[20}, [21], [22], [23], [24], [25], [26], [27], [28], [29], [30], [31], but it's not
known if the proactive scenario has been applied to the spectrum decision stage
[15], which would probably optimize performance. (Fig. 3) shows the main
spectrum access algorithms that are used. They are distributed according to their
operating criteria (based on learning, multi-criteria decision-making, and various
methodologies).
3 Proposed Model for Channel Selection in CR
The general diagram (Fig. 4) of the model implementing a proactive strategy at
the BS starts on the upper portion with the processing of the database that contains
information about the real behavior of PUs in the frequency bands, expressed as
power levels (dB). The data are averaged and discretized for each channel,
considering that a logical 1 represents the presence of a PU, and a logical 0
represents the absence of a PU or free spectrum.
Processing of request at
the BS
Reactive
strategy
Proactive
strategy
SVM and ANFIS as channel selection models 479
Figure 3. Classification of methodologies for spectrum selection.
Figure 4. General diagram of the channel selection model for CR.
Algorithms for the selection and assignment of spectrum
channels
Based on multicriterion
decision making
Additional methodologies
Learning-based
Neural networks
Genetic models
Reinforcement
learning
AHP
FAHP
Adaptive methods
Bayesian
inference
Diffuse systems
Game theory
TOPSIS
Ant colony
Bio-inspired
optimization
algorithms
Decision tree
Channel availability
database
Power as a function
of time and
frequency
For each
band calculate
: Characterization of
PUs with Long Short-Term Memory
(LSTM)
Bandwidth
Probability of
channel availability
Estimated
availability time
Delay
Ranking of
channels
ANFIS
SUs database
Obtaining of
SU-Channel/Channels pairs for each
instant of time
(channel
reservation)
Channel
assignment
algorithms
Better channels
for more
demanding SUs
Evaluation metrics
Results (graphical
representation)
System for predicting the
arrival of SUs requesting a
RT or BE service from the
reactive model
Real data (arrival of SUs and state of spectrum bands)
Discretization of
power levels
Presentation segments with
characteristics
of SUs
SU requirements with “best effort”
characteristics (1 or
3 channels)
Identifier
Historical records
Message length
Identifier
Historical records
Delay
Bandwidth
SU requirements
with “real time”
characteristics (5 or 7 channels)
SVM
Yes
No
480 Danilo Alfonso López Sarmiento et al.
Subsequently, the PU's use of the spectrum is characterized (modeled and
predicted). For each channel of the GSM band, the following variables are
obtained: bandwidth (ab), probability of availability (pd), average availability
time (tp) [32], and delay (r); these are variables that determine transmission
capability, and make it possible to establish (through the design of a ranking
algorithm based on ANFIS and SVM (Section 4)) what spectrum or spectra are
best for assignment to SUs according to the required QoS. The other essential
component of the system in Fig. 4 includes the description of the proactive
strategy (Section 3.2), which estimates the probability of appearance of, or
predicts, an SUi requesting channels that support BE or RT applications (based on
the history in the SU database), and is able to select in advance the channel that
will be assigned in the very near future.
3.1 PU Characterization Algorithm
The prediction algorithm was built using Long-Short Term Memory (LSTM) [33],
[34], a type of recurrent neural network grounded on the Deep Learning concept
[35], [36], based on the PU's history of spectrum use. The selection of this type of
learning paradigm was made seeking to reduce the prediction error in the frequency
bands. (Fig. 5) shows the algorithm's general flowchart, and (Fig. 6) shows the
training algorithm, reaching accuracy above 83%, a value that exceeds the state of
the art.
Figure 5. LSTM algorithm flowchart.
Start
Enter historical data
discretized PU
Determine the
number of neurons in the
input layer
Determine the number of layers
and neurons of the
LSTM
Train the LSTM
neural networks
Take performance measurement metrics
Validate results
Is this desired
performance?
End
Build the LSTM
recurrent networks
Yes
No
No
Yes
SVM and ANFIS as channel selection models 481
3.2 Algorithm to Predict the Arrival of SU Requests at the Base Station [37]
The algorithm for calculating the probability of, or predicting, the arrival of the
next SU at the central station of the CRN can be found in [37], and is based on the
construction of a Multilayer Perceptron Neural Network (MLPNN) based on
historical data on the use of the available spectrum.
Figure 6. Training algorithm for a LSTM recurrent network.
The structure of the neural network is comprised of seven neurons (Fig. 7), three
of which are in the input layer, which receives data; two in the hidden layer,
where the learning process is generated; and two in the output layer. The input
𝑥(𝑖) and output 𝑦(𝑖) variables are itemized in Equation 1. In particular, in the
case of the input variables, the first column refers to the arrival of BE-type
requests; the second column refers to RT-type requirements; the third column
refers to the requested bandwidth (depending on the need, 1 to 7 channels may be
assigned according to the availability and application type).
Initialize the weights for each
neuron
Read the training
examples
Is this the end of the list?
End
Obtain the LSTM recurrent network’s
output for the training
example
Is this the expected
output?
Correct the
weighting of the LSTM
recurrent
network
Start
(1)
482 Danilo Alfonso López Sarmiento et al.
The outputs [from] 𝑦(𝑖) provide the probability of arrival of the next SU with
BE or RT criteria to be processed by the BS. It’s worth noting that the matrix that
describes the output for the bandwidth variable isn’t shown, but is described in
[37].
Figure 7. Neural network structure for the calculation of the probability of BE and
RT requests only.
The MLPNN that was designed is multilayer and unidirectional, indicating that
each constituent layer receives the information from the preceding layer and
transmits it to the next layer. The training or learning of the network that was
constructed consists of obtaining a vector 𝜃 that meets the conditions of the
preceding layer until reaching the exit [38]. The mathematical foundation for the
prediction model [37] is the feed-forward concept described by Equation 2
𝐴(𝑖) = 𝑔(𝜃(𝑖−1)𝑇𝐴(𝑖−1)) (2)
Where 𝑇 corresponds to the transposed matrix; 𝐴(𝑖)identifies the output of the
neuron at the layer that needs to be calculated; 𝐴(𝑖−1) represents the output of the
preceding layer; 𝑖 = 1, 2, … . , n; 𝐴(𝑖) = 𝑋; 𝑔 identifies the sigmoid function.
Considering that the weights matrix is identified as 𝜃(1), the transition from the
input layer to the intermediate layer can be described by Equations 3 and 4.
SVM and ANFIS as channel selection models 483
𝜃(1) = [
𝜃11(1)
𝜃12(1)
𝜃21(1)
𝜃22(1)
𝜃31(1)
𝜃32(1)
] (3)
𝐴(1) = 𝑋 = [
𝑥1
𝑥2
𝑥3
] (4)
Replacing the value of 𝜃(1) and 𝐴(1) in Equation 2 yields the transition from the
output of the input layer and the input of the intermediate layer [37] of Fig. 7 (see
Equation 5).
𝐴(2) = [𝑔(𝑥1𝜃11
(1)+ 𝑥2𝜃21
(1)+ 𝑥3𝜃31
(1))
𝑔(𝑥1𝜃12(1)
+ 𝑥2𝜃22(1)
+ 𝑥3𝜃32(1)
)] (5)
Simplifying the above matrix yields (Equation 6):
𝐴(2) = [𝑎1
(2)
𝑎2(2)
] (6)
Just like in the above case, using the weights matrix 𝜃(2), as a reference, the
matrices describing the transition from the hidden layer to the output layer are
determined as described in Equations 7 and 8.
𝜃(2) = [𝜃11
(2)𝜃12
(2)
𝜃21(2)
𝜃22(2)
] (8)
𝐴(3) = [𝑔(𝑎𝜃11
(2)+ 𝑎2𝜃21
(2))
𝑔(𝑎1𝜃12(2)
+ 𝑎2𝜃22(2)
)] (9)
Where the vector 𝐴(3) corresponds to the neural network's output; in this case, a
dimensional matrix 𝑅2.
Part of the pseudo code developed for MLPNN learning is shown in Fig. 8, where
the existence of weights matrices Theta_1 and Theta_2 is initially assumed, and
training examples are then taken until finding appropriate values for said variables
so that the error is reduced to the minimum possible. The results after executing
the algorithm show accuracy in the arrival of the next SU (with QoS criteria)
above 80% [37].
484 Danilo Alfonso López Sarmiento et al.
Algoritmo MLPNN
𝑑𝑒𝑙𝑡𝑎_𝑎𝑐𝑐𝑢𝑚_1 = 𝑧𝑒𝑟𝑜𝑠 (𝑠𝑖𝑧𝑒 (𝑇ℎ𝑒𝑡𝑎1));
𝑑𝑒𝑙𝑡𝑎_𝑎𝑐𝑐𝑢𝑚_2 = 𝑧𝑒𝑟𝑜𝑠 (𝑠𝑖𝑧𝑒 (𝑇ℎ𝑒𝑡𝑎2));
for 𝑡 = 1: 𝑚 do
𝑎_1 = 𝑋(𝑡,: );
𝑧_2 = 𝑎_1 ∗ 𝑇ℎ𝑒𝑡𝑎1’; 𝑎_2 = [1 𝑠𝑖𝑔𝑚𝑜𝑖𝑑𝑒 (𝑧_2)]; 𝑧_3 = 𝑎_2 ∗ 𝑇ℎ𝑒𝑡𝑎2’; 𝑎_3 = 𝑠𝑖𝑔𝑚𝑜𝑖𝑑𝑒 (𝑧_3);
𝑦_𝑖 = 𝑧𝑒𝑟𝑜𝑠 (1, 𝐾);
𝑦_𝑖(𝑦(𝑡)) = 1; 𝑑𝑒𝑙𝑡𝑎_3 = 𝑎_3 − 𝑦_𝑖; 𝑑𝑒𝑙𝑡𝑎_2 = 𝑑𝑒𝑙𝑡𝑎_3 ∗ 𝑇ℎ𝑒𝑡𝑎2 .∗ 𝑠𝑖𝑔𝑚𝑜𝑖𝑑𝑒𝐺𝑟𝑎𝑑𝑖𝑒𝑛𝑡 ([1 𝑧_2)]; 𝑑𝑒𝑙𝑡𝑎_𝑎𝑐𝑐𝑢𝑚_1 = 𝑑𝑒𝑙𝑡𝑎_𝑎𝑐𝑐𝑢𝑚_1 + 𝑑𝑒𝑙𝑡𝑎_2 (2: 𝑒𝑛𝑑)′ ∗ 𝑎_1;
𝑑𝑒𝑙𝑡𝑎_𝑎𝑐𝑐𝑢𝑚_2 = 𝑑𝑒𝑙𝑡𝑎_𝑎𝑐𝑐𝑢𝑚_2 + 𝑑𝑒𝑙𝑡𝑎_3′ ∗ 𝑎_2;
end;
𝑇ℎ𝑒𝑡𝑎1_𝑔𝑟𝑎𝑛𝑑 = 𝑑𝑒𝑙𝑡𝑎_𝑎𝑐𝑐𝑢𝑚_1 / 𝑚
𝑇ℎ𝑒𝑡𝑎2_𝑔𝑟𝑎𝑛𝑑 = 𝑑𝑒𝑙𝑡𝑎_𝑎𝑐𝑐𝑢𝑚_2 / 𝑚
Figure 8. MLPNN pseudo code.
4 Design of SVM and ANFIS Classifiers
The classifier responsible for generating the ranking of channels in each time slot
of the simulation of the entire global system (Fig. 4) classifies the bands in
descending order, from the one with the best characteristics to the one with the
least optimal characteristics. Two classification algorithms that use the ANFIS
and SVM methodologies, and are catalogued as supervised learning paradigms,
were implemented with the purpose of determining which of the two has a better
performance in a real-life implementation for a cognitive radio wireless network
with a centralized topology under the design conditions laid out in this paper.
4.1 ANFIS Classifier
To obtain a ranking with ANFIS, a model was implemented in MATLAB [39]
with Takagi-Sugeno type inference [40]. Fig. 9 describes the neuro-diffuse
network scheme. The channel characteristics are found in the input layer; the
membership functions are defined in the next layer (high and low for each input
(see Fig. 10)); the third and fourth layers calculate the rules and establish the
Takagi-Sugeno type inference; the fifth layer performs the summation of the
outputs in order to obtain the scores for each channel and determine the ranking.
SVM and ANFIS as channel selection models 485
Figure 9. ANFIS network architecture.
The metrics X1, X2, X3, X4 in (Fig. 9) are 𝑡𝑝, 𝑝𝑑, 𝑟 and 𝑎𝑏, and are defined
previously; for each parameter, two Gaussian membership functions (high in red
color and low in blue color) are defined, and are described as shown in (Fig. 10).
Figure 10. Membership functions that were defined.
As was described in the previous graph, the membership functions are defined in
the same way for each input; the function that represents the low level (in color
blue) has a maximum of 0, and the function that represents the high level (color
red) has a maximum of 1, so that when the channel data are entered and then
evaluated in the membership functions, coherent results are obtained. The
decision was made to work with only two linguistic labels in order to limit the
amount of computing resources required by the system, and thus reduce the
algorithm's compilation time. The flowchart of the neuro-diffuse model that was
used is described and explained in (Fig. 11).
No
Yes
No
Yes
Yes
No
486 Danilo Alfonso López Sarmiento et al.
Figure 11. Flowchart for the ranking algorithm with ANFIS.
4.2 SVM Classifier
The procedure to implement the SVM technique was to establish a decision
hyperplane that, according to the ranking's expected behavior, would suitable for
the requirements; subsequently, the channels were classified according only to the
characteristics of the SU applications (BE and RT). From there, the MATLAB
Toolbox [41] was used to obtain the parameters for the SVM. With the data, each
channel characteristic was evaluated in the equation for the generated hyperplane
Start
Define membership
functions (MFs) and initial
sequent parameters
Read input data
Evaluate data in
corresponding functions
Calculate product T-norm
for each pair rules
Normalize rules
Calculate rules’
contributions (Sugeno)
Add rules’ contributions
to obtain scores
Channels classification by
score in descending order
Channel assigment
simulation End
Define random vectors for
MFs and sequent
Subtract vectors from
initial values
Add vectors to the initial
values
Obtain new result
¿Did the results
improve?
Validate vectors
Obtain new results
¿Did the
results
improve?
Results evaluation
¿Look for a
better result?
Discard vectors
Yes
No
SVM and ANFIS as channel selection models 487
(Equation 9), which yields a score for each free channel and therefore the respective
final ranking. The description of the algorithm's block diagram can be found in
(Fig. 12).
Figure 12. Flowchart for the ranking algorithm with SVM.
−79.46 ∗ 𝑡𝑝 − 58.26 ∗ 𝑝𝑑 + 142.84 ∗ 𝑟 + 0 ∗ 𝑎𝑏 + 30.7329 = 0 (9)
The operation of the SVM that was implemented (Fig. 12) has two main processes:
the training of the network (which implicitly contains the linear separator or
hyperplane), and the decision stage. The training process consists of finding a
decision function capable of separating the spectrum bands (as was mentioned
previously) making use of the optimization function of Equation 10 [42], [43]. In
the decision stage, given a new vector 𝑥, the class in which each spectrum channel
belongs is determined (this allows the ranking to be classified in descending order).
𝑚í𝑛
𝑊, 𝑏, 𝜉
1
2𝑊𝑇𝑊 + 𝐶 ∑ 𝜉𝑖
𝑛
𝑖=1
(10)
Whereas (𝑊𝑇 𝜙 ( 𝑥𝑖 ) +𝑏 ) ≥ 1 − 𝜉𝑖 , 𝜉𝑖 ≥ 0 ; and where, according to [44],
(𝑥1,𝑦1), (𝑥2, 𝑦2) … (𝑥𝑛,𝑦𝑛) corresponds to the training set; 𝑥𝑖 is the vectorial
representation of the training examples; 𝑦𝑖 , the label that represents the
corresponding class, 𝑛 , the number of individuals; 𝑊𝑇 𝜙(𝑥𝑖)+𝑏 identifies the
separation hyperplane; 𝑊 is the weights vector; B is the bias; 𝜙 is the kernel
function that allows taking data to an N-dimensional space in which they may be
separable; 𝐶 is a parameter that controls the balance between machine complexity
and the number of points not separable by a a hyperplane, 𝜉𝑖is the surplus variable
that measures the deviation of a point 𝑥𝑖 from the point of separation
𝑊𝑇 𝜙(𝑥𝑖)+𝑏.
Start
Obtain SVM variables
Define hyperplane according to
requirements
Evaluate parameters of each channel in
the hyperplane equation Label the channels (BE or RT)
SVM training
Define the scores for each
channel for the ranking (decision
at the SVM)
Classify the channel list by score
in descending order
Acccept
results?
End
488 Danilo Alfonso López Sarmiento et al.
The decision hyperplane obtained with the SVM during network training contains
three of the four characteristics defined by each spectrum, and is shown in (Fig. 13).
Fig. 13. Decision hyperplane obtained for the variables 𝑡𝑝, 𝑝𝑑, 𝑟.
A detailed analysis makes it possible to infer that the spectrum of frequencies with
the greatest 𝑡𝑝 , 𝑝𝑑 and a low 𝑟 are labeled as channels for supporting the
transmissions of RT applications (blue-colored points under the linear separator),
while the rest of the points (red-colored points located above the plane) are labeled
as channels for the transport of BE applications.
It's necessary to take into account that SUs admitted by the BS in the assignment
process deliver the best quality channels to the RT SUs first, and the worst quality
channels to the BE SUs; if the range of frequencies in a category is exhausted, an
SU may use the spectrum of the other category, subject to, in the worst case, a
reduction in QoS.
5 Channel Selection and Assignment
The flowchart in Fig. 14 describes the operation of the spectrum assignment stage.
The algorithm starts by obtaining and loading the results generated by the PU
characterization and SU arrival probability models for a period of time 𝑇
(defined in terms of the number of time slots); it also defines the variable 𝑡 = 1,
which corresponds to the time slot in which the simulation is being performed,
and the variable 𝑖 = 1 , which identifies the SUi that is undergoing the early
spectrum assignment process. Subsequently, the grey-colored sub-block is
executed, as described by the flowchart in (Fig. 15).
No
Yes
Yes
No
No
Yes
BE RT
SVM and ANFIS as channel selection models 489
Figure 14. Flowchart for channel selection and assignment in the proactive system.
Figure 15. Assignment and reservation of channels according to QoS.
After that, a determination is made whether the spectrum allows a successful
transmission; if so, the SU-channel pairs (i.e., the resources) are reserved and
stored. Lastly, a real transmission is made in the exact times according to the
order of arrival of the estimated SUs.
Start
Obtain SU arrival and spectrum
availibity predictions
Conditions for a
successful Tx?
End
t=1
i=1
Assign and reserve channel/channels
for SU according to predictons and
QoS
Discard
transmission_ti
Store channel
assignment data
for future Tx
i=i+1 1
1
i=N
t=T
t=t+1
2
Execute Tx with
stored data
2
Update spectrum
occupation data
Start
Load rankings of
channels
SU_i
type:
Assign and reserve the first channels in
the BE ranking to
SU_i
Update spectrum
occupation data
End
Assign and reserve
the first channels in the RT ranking to
SU_i
490 Danilo Alfonso López Sarmiento et al.
6 Results
The performance evaluation was performed by comparing the proactive strategy
with ANFIS classification, the reactive strategy without classification (its
operation has been described for the arrival of SUs) as shown in the grey
sub-block of Fig. 4 and without the existence of ranking), and the proactive
strategy with SVM when the arrival rate is exponential with a maximum growth
of up to 100 requests per time slot (equivalent to 290 ms), in which case the
cognitive network isn’t able to meet the transmission requirements of secondary
users. The number of channels the applications can request depends on the
bandwidth they require – 1-3 for BE or 5-7 for RT. The following metrics were
evaluated during the simulation: the probability of total success in the
transmissions performed (Equation 11), RT success probability (Equation 12) and
BE success probability (Equation 13), delay error (Equation 14), spectrum
distribution assignment error (Equation 15), successful transmissions (Te),
throughput (Th), and Jain Index (Ij) [45].
𝑃𝑒 =#𝑇𝑥𝑒
#𝑇𝑥𝑇 (11)
Where #𝑇𝑥𝑒 is the number of successful transmissions, and #𝑇𝑥𝑇 , the total
number of transmissions
𝑃𝑒(𝑅𝑇) =#𝑇𝑥𝑒(𝑅𝑇)
#𝑇𝑥𝑇(𝑅𝑇) (12)
Where #𝑇𝑥𝑒(𝑅𝑇) is the number of successful RT transmissions, and #𝑇𝑥𝑇(𝑅𝑇)s the
total number of RT transmissions
𝑃𝑒(𝐵𝐸) =#𝑇𝑥𝑒(𝐵𝐸)
#𝑇𝑥𝑇(𝐵𝐸) (13)
Where #𝑇𝑥𝑒(𝐵𝐸) is the number of successful BE transmissions, and #𝑇𝑥𝑇(𝐵𝐸) is
the total number of BE transmissions
𝐸𝑟 =∑ (𝑟𝑜 − 𝑟𝑒)𝑇𝑥𝑒
𝑇𝑥𝑒 (14)
Where 𝑇𝑥𝑒 is the number of successful transmissions, 𝑟𝑜 the delay obtained (the
greatest delay among assigned channels), and 𝑟𝑒 , the expected delay (according to
SU requirements).
𝐸𝑎 =∑ |𝑐ℎ𝑠 − 𝑐ℎ𝑎|𝑇𝑥𝑇
𝑡𝑠𝑇 (15)
SVM and ANFIS as channel selection models 491
Where 𝑇𝑥𝑇 is the total number of transmissions, #𝑐ℎ𝑠, the number of requested
channels, #𝑐ℎ𝑎 , the number of assigned channels, 𝑡𝑠𝑇 , the total number of
transmitted time slots.
The results obtained for PU traffic located in the GSM band (Uplink) are shown
between Fig. 16 and Fig. 23, where the green-colored line represents the reactive
scenario without ranking (R-no rank), the red color represents the proactive
scenario with ANFIS (P-Anfis), and the green color represents the proactive
scenario with SVM (P-Svm). To begin with, there is evidence of a rapid descent
in the performance of the metrics analyzed with the passage of time, due to the
increase in the number of SUs, which leads to increasingly more SUs not having
access to the system; this is a condition that allows an objective evaluation
because the network is subject to critical or saturation conditions. Fig. 16 shows
the relationship between the number of successful transmissions (for RT and BE)
and the total number of transmission attempts (Equation 11). It is observed that at
the beginning, the system with no ranking has a success probability close to zero
because at the start of the simulation the system repeatedly assigns low quality
channels (due to the elevated presence of PUs and/or low QoS criteria),and since
there is a low number of SUs at the start, the system leaves the rest of the spectrum
(including the best) unassigned, causing SUs to repeatedly fail in their transmission
attempts; this is a trend that is maintained during the first 225 seconds.
Figure 16. Probability of success in RT and BE transmissions.
The proactive scenarios based on ANFIS and SVM exhibit a better performance;
the results are broken down for RT traffic (Fig. 17) and BE traffic (Fig. 18), where
it’s established that RT flows are assigned the best spectrum and attain success
percentages of up to 100% for the first 100 seconds, at which time the system
becomes congested because of spectrum saturation due to the presences of PUs.
492 Danilo Alfonso López Sarmiento et al.
Figure 17. Probability of success in RT transmissions.
A detailed analysis of Fig. 18 reveals that from the 215th second on, the probability
of success of the reactive model without ranking improves and even exceeds the
others, because a greater number of SUs arrive at the BS, so that a wider variety of
the spectrum starts to be assigned, and since the majority of the best channels are
free, there is a notable increase in the number of successful transmissions in the
system without ranking.
Figure 18. Probability of success in BE transmissions.
In Fig. 19, the delay error (normalized with respect to the number of successful
transmissions and sensible to RT data flow), defined as the parameter that measures
the excess additional time imposed by the network on SUs that manage to transmit,
with respect to the delay expected by each SU (Equation 14), reveals that in the
reactive scenario without ranking, the value is zero for up to 120 seconds because
there are no successful RT transmissions (as established in Fig. 17), which is
detrimental to applications of the same type; nevertheless, when RT transmissions
begin to be accepted, the value of the delay error variable increases vertically for
approximately 20 seconds, thus generating from that instant an irregular trend due
to the characteristics of the available spectrum. The ANFIS- and SVM-based
models have a growing trend over time, which is to be expected because the
processing of traffic with QoS criteria prevails over BE.
SVM and ANFIS as channel selection models 493
Figure 19. Delay error.
With regards to the assignment error (Equation 15), which is a variable that is
normalized with respect to the total number of time slots necessary for
transmission, and is defined as the absolute value of the difference between the
number of channels requested by an SU and the number of channels actually
assigned, a superior trend is observed with ANFIS during the simulation because
performance is better when classifying or generating the ranking, closely followed
by SVM, even though the value of the metric increases over time in both
scenarios as a result of characterization error (Section 3.1) and an increase in the
rate of arrival of SUs, which leads to failures in the calculation of the probability
of cognitive users (Section 3.2). The performance of the model without ranking is
the most atypical and volatile, and has the worst behavior; the value for the first
time slots is zero because the system attempts to assign the spectrum without any
kind of classification (which implies a greater presence of PUs and worse QoS).
Fig. 20. Assignment error.
With regards to successful transmissions (Fig. 21), it’s important to begin by
pointing out that the action of the three models increases or, in the worst of cases, is
constant, because the records delivered by the algorithms are cumulative; therefore,
it’s observed that, during a significant portion of the simulation, the proactive
methodologies are more efficient than the reactive ones; however, from the 200th
second on, the value starts to increase at a higher rate until the amount of completed
transmissions is greater than with ANFIS and SVM. This happens because the
systems with designed rankings organize the channel-SU pairs such that the best
494 Danilo Alfonso López Sarmiento et al.
spectra are assigned to the most demanding users (RT) and eventually, if there is
spectrum saturation (as observed in Fig. 16 to 18), the available bands, even if they
are the “best”, cannot meet the required conditions, which causes transmissions to
fail; on the other hand, the system without ranking will possibly assign said
channels to less demanding SUs who can complete their transmission (BE users).
This rationale is supported by Fig. 17, where it's evident the system without ranking
exhibits a greater probability of success for BE from the 300th second.
Figure 21. Successful transmissions.
Fig. 22 shows a comparison of system throughput for the three scenarios; it’s
evident that, on average, they are similar.
Figure 22. Throughput.
A detailed analysis suggests that in the first moments of the simulation the system
without ranking exhibits a throughput close to null, because it repeatedly assigns
channels that don’t meet SU requirements, and leaves the best spectra free; this
course of action is maintained until they can be assigned (thanks to greater
spectrum availability combined with a high number of SUs), thus abruptly
increasing throughput. The opposite occurs with ANFIS and SVM, where what are
classified as the best channels are assigned first, and the bands with less QoS,
greater presence of PUs, and greater congestion are relegated for later.
Another important performance metric that is rarely found in the state of the art is
the fairness level, which is based on allowing all users to have the possibility of
SVM and ANFIS as channel selection models 495
using available resources, thus reducing the potential for monopolizing them. Fig.
23 shows this criterion has more equity for ANFIS until the system can meet the
QoS criteria.
Figure 23. Jain Index.
7 Discussion
The preceding analysis evaluates and validates the performance of the proactive
strategy vs. the reactive strategy based on the condition that the amount of available
resources cannot meet the needs of all SUs; however, said analysis doesn’t allow
establishing with certainty which of the algorithms (ANFIS or SVM) is a better
option for use in a proactive strategy in CR.
Based on Figs. 24 and 25, it can be conjectured that, although SVM is more
efficient in terms of the number of collisions between PUs and SUs, it requires
more robust hardware for real-life implementation, which could prove to be a
significant disadvantage because a greater energy consumption would be required
to keep the wireless network infrastructure operating; this is contrary to what
happens with ANFIS, where global performance is more efficient for most
variables, but is subject to a greater amount of collisions, which is also a critical
condition for licensed users due to the interferences generated, although it’s less
expensive in terms of energy consumption.
Figure 24. Collisions between PUs and SUs.
496 Danilo Alfonso López Sarmiento et al.
Figure 25. Training time needed for the classification algorithms.
On the other hand, using as a reference the problem formulation and the research
question (Section 1), and supported by Fig. 26, it can be conclusively stated that the
use of the proactive strategy reduces the time needed to process SU requests, but
depends on an optimal characterization model, an accurate calculation of SU
arrival, and an appropriate classification method.
Figure 26. Processing time for SU requests after training the learning algorithms.
8 Conclusions
This paper poses the possibility of replacing the reactive strategy, which is
commonly used for [handling] channel requests, processing and assignment in
cognitive networks (in the spectrum decision stage), by a proactive strategy that
anticipates the arrival of secondary users, so that channels to be assigned in the
future are reserved early and in advance, using a spectrum classification system
based on ANFIS and SVM learning methodologies for prioritizing the transmission
SVM and ANFIS as channel selection models 497
of Real Time applications (vulnerable to delays) over Best Effort applications.
Based on the analysis of the results, the conclusion is that the use of a proactive
strategy is more efficient than the use of a reactive strategy in terms of processing
time at the BS, thus optimizing the spectrum selection stage.
After evaluating the proposed spectrum decision system (which integrates the
sub-stages of PU characterization, obtaining of the arrival probability of SUs
requesting QoS criteria, and early establishment and assignment of
SU-channel/channels pairs) under network saturation conditions, similar trends
were found in most of the responses received when using SVM or ANFIS, in
contrast to the collision amount (Fig. 24) and algorithm execution time variables
during the training stage (Fig. 25); depending on which of the two is willing to
sacrifice, SVM or ANFIS benefit spectrum assignment in cognitive radio wireless
networks.
References
[1] D. López, J. Ordoñez, E. Rivas, User characterization through dynamic
Bayesian networks in cognitive radio wireless networks, International Journal
of Engineering and Technology (IJET), 8 (2016), no. 4, 1771-1783.
https://doi.org/10.21817/ijet/2016/v8i4/160804043
[2] B. Canberk, I. Akyildiz, and S. Oktug, Primary user activity modeling using
first difference filter clustering and correlation in cognitive radio networks,
IEEE/ACM Transactions on Networking, 19 (2011), no. 1, 170-183.
https://doi.org/10.1109/tnet.2010.2065031
[3] M. Hoyhtya, S. Pollin, A. Mammela, Improving the performance of cognitive
radios through classification learning and predictive channel selection,
Advances in Electronics and Telecommunications, 2 (2011), No. 4, 28-38.
[4] C. Chou, S. Shankar, H. Kim, G. Shin, What and how much to gain by spectrum
agility?, IEEE Journal on Selected Areas in Communications, 25 (2007), no. 3,
576-588. https://doi.org/10.1109/jsac.2007.070408
[5] W-Y. Lee, I. Akyildiz, A spectrum decision framework for cognitive radio
networks, IEEE Transactions on Mobile Computing, 10 (2011), No. 2, 161-174.
https://doi.org/10.1109/tmc.2010.147
[6] J. Riihijarvi, P. Mahonene, M. Wellens, M. Gordziel, Characterization and
modeling of spectrum for dynamic spectrum access with spatial statistics and
random fields, IEEE 19th International Symposium on Personal, Indoor and
Mobile Radio Communications, 2008, 1-6, Cannes, France,
https://doi.org/10.1109/pimrc.2008.4699912
498 Danilo Alfonso López Sarmiento et al.
[7] A. Daoud, M. Alanyali, D. Starobinski, Secondary pricing of spectrum in
cellular CDMA networks, 2nd IEEE International Symposium Access Networks
(DySPAN), 2007, 535-542, Dublin, Ireland.
https://doi.org/10.1109/dyspan.2007.76
[8] V. Paxon, S. Floyd, Wide area traffic: The failure of Poisson modeling, IEEE/
ACM Transactions on Networking, 3 (1995), no. 3, 226-244.
https://doi.org/10.1109/90.392383
[9] M. Khabazian, S. Aissa, N. Tadayon, Performance modeling of a two-tier
primary-secondary networks operated with IEEE 802.11 DCF mechanism,
IEEE Transaction Wireless Communications, 11 (2012), no. 9, 3047-3057.
https://doi.org/10.1109/twc.2012.071612.110010
[10] X. Xing, T. Jing, Y. Huo, H. Li, X. Cheng, Channel quality prediction based on
Bayesian Inference in Cognitive Radio Networks, IEEE Conference on
Computer Communications, 2013, Turin, Italy.
https://doi.org/10.1109/infcom.2013.6566941
[11] X. Xing, T. Jing, Y. Huo, X. Cheng, Spectrum prediction in cognitive radio
network, IEEE Wireless Communications, 20 (2013), no. 2, 90-96.
https://doi.org/10.1109/mwc.2013.6507399
[12] W. Saad, Z. Han, H. Poor, T. Basar, J. Song, A cooperative Bayesian
non-parametric framework for primary user activity monitoring in cognitive
radio networks, IEEE Journal on Selected Areas in Communications, 30
(2012), no. 9, 1815-1822. https://doi.org/10.1109/jsac.2012.121027
[13] H. Li, L. Qiu, A graphical framework for spectrum modeling and decision
making in cognitive radio network, IEEE Global Communications Conference,
2010, Miami, Florida. https://doi.org/10.1109/glocom.2010.5683361
[14] Y. Saleem, M. Rehmani, Primary radio user activity models for cognitive radio
networks: A survey, Journal of Networks and Computer Applications, 43
(2014), 1-16. https://doi.org/10.1016/j.jnca.2014.04.001
[15] D. López, E. Trujillo and O. Gualdron, Elementos fundamentales que
componen la radio cognitiva y asignación de bandas espectrales, Journal
Información Tecnológica, 26 (2015), no. 1, 23-40.
https://doi.org/10.4067/s0718-07642015000100004
[16] A. He, K. Bae, T. Newman, J. Gaeddert, K. Kim, R. Menon, L. Morales, J.
Neel, Y. Zhao, J. Reed, W. Tranter, A survey of artificial intelligence for
cognitive radios, IEEE Transactions on Vehicular Technology, 59 (2010), no. 4,
1578-1592. https://doi.org/10.1109/tvt.2010.2043968
SVM and ANFIS as channel selection models 499
[17] M. Jeet, H. Verma, M. Uddin, QoS provisioning in channel allocation strategy
in cognitive radio using fuzzy logic, International Journal Machine
Intelligence and Sensory Signal Processing, 1 (2014), no. 3, 215.
https://doi.org/10.1504/ijmissp.2014.066438
[18] D. Niyato, E. Hossain, Competitive pricing for spectrum sharing in cognitive
radio networks: Dynamic game inefficiency of Nash equilibrium, and collusion,
IEEE Journal Selected Areas in Communications, 26 (2008), no. 1, 192-202.
https://doi.org/10.1109/jsac.2008.080117
[19] C. Hernandez, L. Pedraza, E. Rodriguez, Multivariable adaptive handoff
spectral model for cognitive radio networks, Contemporary Engineering
Sciences, 10 (2017), no. 2, 39-72. https://doi.org/10.12988/ces.2017.69154
[20] D. Xu, D. Jung., and X. Liu, Optimal bandwidth selection in multichannel
cognitive radio networks: How much is too much?, IEEE Int. Symposium for
New Frontiers in Dynamic Spectrum Access Networks (DySPAN), 2008, 1-11,
Illinois, USA. https://doi.org/10.1109/dyspan.2008.45
[21] Z. Zhitao, W. Lifeng, D. Guoru, W. Shengqun, Enhanced spectrum decision
based on the combination of sensing and prediction, Cross Strait
Quad-Regional Radio Science and Wireless Technology Conference
(CSQRWC), 2011, 985-989, Heilongjiang, China.
https://doi.org/10.1109/csqrwc.2011.6037122
[22] M. Timmers, S. Pollin, A. Dejonghe, L. Van der Perre, F. Atthoor, A
Distributed Multichannel MAC Protocol for Multihop Cognitive Radio
Networks, IEEE Transactions on Vehicular Technology, 59 (2010), no. 1,
446-459. https://doi.org/10.1109/tvt.2009.2029552
[23] I. Akyildiz, W.Y Lee, M. Vuran, S. Mohanty, A survey on spectrum
management in cognitive radio networks, IEEE Communications Magazine, 46
(2008), no. 4, 40–48. https://doi.org/10.1109/mcom.2008.4481339
[24] X. Cheng, M. Jiang, Cognitive radio spectrum assignment based on artificial
bee colony algorithm, IEEE 13th International Conference on Communication
Technology, 2011, 161–164, Jinang, China.
https://doi.org/10.1109/icct.2011.6157854
[25] Luis F. Pedraza, Cesar A. Hernandez, E. Rodriguez, A spectral opportunities
forecasting method in a mobile network based on the integration of COST 231
Walfisch-Ikegami and wavelet neural models, Contemporary Engineering
Sciences, 10 (2017), no. 3, 113-128.
https://doi.org/10.12988/ces.2017.68149
500 Danilo Alfonso López Sarmiento et al.
[26] G. Ko, A. Franklins, S-J. You, J-S Pak, M-S. Song, C-J. Kim, Channel
management in IEEE 802.22 WRAN systems, IEEE Communications
Magazine, 48 (2010), no. 9, 88-94.
https://doi.org/10.1109/mcom.2010.5560592
[27] P. Bahl, R. Chandra, T. Moscibroda, R. Murty, M. Welsh, White space
networking with WiFi like connectivity, SIGCOMM Conference on
Applications, Technologies, Architectures and Protocols for Computer
Communications, 2009, 1-12, Barcelona, Spain.
https://doi.org/10.1145/1592568.1592573
[28] M. Masonta, M. Mzyece, N. Ntlatlapa, Spectrum decision in cognitive radio
networks: A survey, IEEE Communications Surveys & Tutorials, 15 (2013), no.
3, 1088–1107. https://doi.org/10.1109/surv.2012.111412.00160
[29] M. Matinmikko, J. Del Ser, T. Rauma, M. Mustonen, Fuzzy-Logic based
framework for spectrum, availability assessment in cognitive radio systems,
IEEE Journal on Selected Areas in Communications, 31 (2013), no. 11, 2173–
2184. https://doi.org/10.1109/jsac.2013.131117
[30] M. Petrova, P. Mahonen, A. Osuna, Multi-Class classification of analog and
digital signals in cognitive radios using Support Vector Machines, 7th
International Symposium on Wireless Communication Systems, 2010, 986–990,
United Kingdom. https://doi.org/10.1109/iswcs.2010.5624500
[31] C. Hernández, L. Pedraza, F. Martinez, Algorithms for spectrum allocations in
cognitive radio networks, Journal Tecnura, 20 (2016), no. 48, 69-88.
http://revistas.udistrital.edu.co/ojs/index.php/Tecnura/article/view/10550/1203
6
[32] C. Hernández, C. Salgado, H. López, E. Rodriguez, Multivariable algorithm
for dynamic channel selection in cognitive radio networks, Journal on Wireless
Communications and Networking, 2015 (2015), no. 1, 1-17.
https://doi.org/10.1186/s13638-015-0445-8
[33] A. Graves, J. Schmidhuber, Framewise phoneme classification with
bidirectional LSTM and other neural network architectures, Journal Neural
Network, 18 (2005), no. 5-6, 602-610.
https://doi.org/10.1016/j.neunet.2005.06.042
[34] A. Graves, A. Mohamed, G. Hinton, Speech recognition with deep recurrent
neural networks, IEEE International Conference on Acoustic Speech and
Signal Processing, 2013, 6645-6649, Vancouver, Canada.
https://doi.org/10.1109/icassp.2013.6638947
SVM and ANFIS as channel selection models 501
[35] B. Sun, H. Feng, K. Chen, X. Zhu, A deep learning framework of quantized
compressed sensing for wireless neural recording, IEEE Access, 4 (2016),
5169-5178. https://doi.org/10.1109/access.2016.2604397
[36] H. Palangi, L. Deng, Y. Shen, J. Gao, X. He, J. Chen, X. Song, X, R. Ward,
Deep sentence embedding using Long Short-Term Memory Networks:
Analysis and application to information retrieval, IEEE/ACM Transactions on
Audio, Speech, and Language Processing, 24 (2016), no. 4, 694-707.
https://doi.org/10.1109/taslp.2016.2520371
[37] D. López, E. Rivas, L. Pedraza, Estimating SUs arrival for channel selection in
cognitive radio, ARPN Journal of Engineering and Applied Sciences, 12
(2017), 2159-2162.
http://www.arpnjournals.org/jeas/research_papers/rp_2017/jeas_0417_5875.pd
f
[38] J. Liu, X. Jing, S. Sun, X. Wang, D. Cheng, H. Huang, Multilayer
Feed-forward Neural Network learning based Dynamic Chinese restaurant
model for dynamic spectrum access in cognitive radio networks, 15th
International Symposium on , Communications and Information Technologies
(ISCIT), 2015, 173-176, Nara, Japan.
https://doi.org/10.1109/iscit.2015.7458335
[39] Matlab, Matworks, Adaptive neuro-fuzzy modeling.
https://www.mathworks.com/help/fuzzy/adaptive-neuro-fuzzy-inference-syste
ms.html, 2016.
[40] J. Jang, ANFIS: Adaptive-Network-Based Fuzzy Inference System, IEEE
Transactions on Systems, Man, and Cybernetics, 23 (1993), no. 3, 665-685.
https://doi.org/10.1109/21.256541
[41] Matlab, Matworks, Train support vector machine classifier.
https://www.mathworks.com/help/stats/svmtrain.html, 2016.
[42] J. Shawe, S. Sun, A review of optimization methodologies in support vector
machine, Journal Neurocomputing, 74 (2011), no. 17, 3609-3618.
https://doi.org/10.1016/j.neucom.2011.06.026
[43] J. Chorowski, J. Wang, J. Zurada, Review and performance comparison of
SVM -and ELM- based classifiers, Journal Neurocomputing, 128 (2014),
507-5016. https://doi.org/10.1016/j.neucom.2013.08.009
[44] F. Camacho, Sistema de Clasificación de Péptidos Antibacterianos Utilizando
Máquinas de Soporte Vectorial, System Engineer Thesis, Dept. of Engineering,
Universidad Industrial de Santander, Bucaramanga, Colombia, 2013.
502 Danilo Alfonso López Sarmiento et al.
[45] R. Jain, D-M Chiu, W. Hawe, A quantitative measure of fairness and
discrimination for resource allocation and shared computer system, Digital
Equipment Corporation, 1984, 1-38.
http://www.cse.wustl.edu/~jain/papers/ftp/fairness.pdf
Received: May 2, 2017; Published: May 29, 2017