svm and anfis as channel selection models for the spectrum ... · svm and anfis as channel...

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


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


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


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


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)


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.


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].


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.


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


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


(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


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


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


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)


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.


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.


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


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


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.


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


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

https://doi.org/10.21817/ijet/2016/v8i4/160804043

https://doi.org/10.1109/tnet.2010.2065031

https://doi.org/10.1109/jsac.2007.070408

https://doi.org/10.1109/tmc.2010.147

https://doi.org/10.1109/pimrc.2008.4699912


[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


https://doi.org/10.1109/90.392383

https://doi.org/10.1109/twc.2012.071612.110010

https://doi.org/10.1109/infcom.2013.6566941

https://doi.org/10.1109/mwc.2013.6507399


https://doi.org/10.1109/glocom.2010.5683361

https://doi.org/10.1016/j.jnca.2014.04.001

https://doi.org/10.4067/s0718-07642015000100004

https://doi.org/10.1109/tvt.2010.2043968


[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.


[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

https://doi.org/10.1504/ijmissp.2014.066438


https://doi.org/10.12988/ces.2017.69154


https://doi.org/10.1109/csqrwc.2011.6037122

https://doi.org/10.1109/tvt.2009.2029552

https://doi.org/10.1109/mcom.2008.4481339

https://doi.org/10.1109/icct.2011.6157854

http://www.m-hikari.com/ces/ces2017/ces1-4-2017/p/pedrazaCES1-4-2017.pdf



https://doi.org/10.12988/ces.2017.68149


[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.


[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


https://doi.org/10.1145/1592568.1592573

https://doi.org/10.1109/surv.2012.111412.00160


https://doi.org/10.1109/iswcs.2010.5624500

https://doi.org/10.1186/s13638-015-0445-8

https://doi.org/10.1016/j.neunet.2005.06.042

https://doi.org/10.1109/icassp.2013.6638947


[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.

https://doi.org/10.1109/access.2016.2604397

https://doi.org/10.1109/taslp.2016.2520371

https://doi.org/10.1109/iscit.2015.7458335

https://doi.org/10.1109/21.256541




[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

svm and anfis as channel selection models for the spectrum ... · svm and anfis as channel...

Documents