An analytic network process and trapezoidal interval-valued fuzzy technique for order preference by similarity to ideal solution network access selection method

INTERNATIONAL JOURNAL OF COMMUNICATION SYSTEMSInt. J. Commun. Syst. (2014)Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/dac.2833

An analytic network process and trapezoidal interval-valued fuzzytechnique for order preference by similarity to ideal solution

network access selection method

Emmanouil Skondras1,2, Aggeliki Sgora1,2, Angelos Michalas2 andDimitrios D. Vergados1,*,†

1Department of Informatics, University of Piraeus, 80, Karaoli and Dimitriou St., GR-18534, Piraeus, Greece2Department of Informatics Engineering, Technological Educational Institute of Western Macedonia, GR-52100,

Kastoria, Greece

SUMMARY

Next generation wireless networks consist of many heterogeneous access technologies that should supportvarious service types with different quality of service (QoS) constraints, as well as user, requirements andprovider policies. Therefore, the need for network selection mechanisms that consider multiple factors mustbe addressed. In this paper, a network selection method is proposed by applying the analytic network processto estimate the weights of the selection criteria, as well as a fuzzy version of technique for order preferenceby similarity to ideal solution to perform the ranking of network alternatives. The method is applied to aheterogeneous network environment providing different QoS classes and policy characteristics. Each userapplies the method to select the most appropriate network, which satisfies his or her requirements in respectof his or her service-level agreement (SLA). Performance evaluation shows that when the user requests onlyone service, the proposed method performs better compared to the original technique for order preference bysimilarity to ideal solution, as well as the Fuzzy AHP-ELECTRE method. Moreover, the proposed methodcan be applied in cases where a user requires multiple services simultaneously on a device. The sensitivityanalysis of the proposed method shows that it can be properly adjusted to conform to network environmentchanges. Copyright © 2014 John Wiley & Sons, Ltd.

Received 9 January 2014; Revised 17 June 2014; Accepted 18 June 2014

KEY WORDS: network selection; vertical handover; MADM; ANP; TOPSIS; interval value fuzzy numbers

1. INTRODUCTION

Next generation wireless access networks are growing rapidly integrating multiple network tech-nologies aiming to support multimedia services in addition to voice and data with high data ratesand guaranteed QoS [1]. In this context, end users devices (such as mobile phone or netbook) areequipped with multiple radio interfaces allowing connectivity to the most suitable network envi-ronment based on users requirements and operators policies [2, 3]. According to the always bestconnection principle of the fourth generation wireless networks, users of mobile services should beprovided with connectivity to the best access technology at anytime [4, 5]. Therefore, there is a needfor efficient vertical handover (VHO) mechanisms to be applied.

The handover process is supposed to be successful, infrequent, and imperceptible to enabletelecommunication providers meet the QoS requirements of the users [6]. Especially, in the case

*Correspondence to: Dimitrios D. Vergados, Department of Informatics, University of Piraeus, 80, Karaoli and DimitriouSt., GR-18534, Piraeus, Greece.

†E-mail: [email protected]

Copyright © 2014 John Wiley & Sons, Ltd.

E. SKONDRAS ET AL.

of heterogeneous networks, seamless interworking among the different technologies is also needed[7]. Thus, special attention to the VHO process should be given [8].

The VHO procedure consists of three main steps including the handover initiation, the net-work selection, and the handover execution. The initiation step contains the required procedures toidentify the available access networks and select the time of handover in respect of network condi-tions and user mobility. The network selection step is related to the selection of the most appropriatenetwork alternative based on the available network characteristics, user preferences, and applicationsrequirements. Finally, the execution step completes the handover process by seamlessly connect-ing the terminal to the selected network. This paper deals with the network selection step of theVHO process.

Existing handover network selection schemes employ multi attribute decision-making methods(MADM), fuzzy logic, neural networks, and utility functions [9]. However, because the selectionof an access network depends on several parameters with different relative importance, the accessnetwork selection problem is usually looked at from the aspect of multi-criteria analysis and morespecifically by applying different MADM algorithms. In this paper, a network selection method isproposed by employing two MADM algorithms: the analytic network process (ANP), which is anextension of the analytic hierarchy process (AHP) for criteria weights calculation, and a fuzzy ver-sion of the technique for order preference by similarity to ideal solution (TOPSIS) for accomplishingthe ranking of the candidate networks. The proposed method considers network QoS characteristicsand policies, application requirements and different types of users service-level agreements (SLAs)to provide advanced connection services. Linguistic values are used to characterize the performanceof selection criteria, which are represented by interval-valued trapezoidal fuzzy numbers.

Our approach provides the following main contributions:

� It allows complex relationships within and among clusters (in our case, the network QoS char-acteristics and the network policies characteristics) of selection criteria by applying the ANPmethod, which does not use an hierarchical framework as AHP but a network model of depen-dencies. Additionally, ANP eliminates the index consistency requirement of AHP (i.e., in AHPthe relative importance of decision factors need to be redefined in case index consistency valueis more than 0.1).� It can better express imprecise information of performance selection criteria for different

application types and users SLAs by applying linguistic values and interval-valued fuzzy num-bers. Interval value fuzzy numbers are adopted because they can efficiently present uncertaininformation by minimum maximum membership intervals rather than by single membershipvalues.� It performs the selection of the best network access technology by considering contradictory

selection criteria, facilitating the provision of high quality services and at the same time satis-fying different types of users SLAs. This is achieved through a fuzzy version of TOPSIS, thetrapezoidal interval-valued Fuzzy TOPSIS (TFT), introduced in this study. Because TFT usesfuzzy logic, it resolves the case of having several services of different QoS constraints runningsimultaneously on a terminal. Therefore, network selection is performed in a way satisfyingmultiple groups of criteria per user. Furthermore, the ranking abnormality problem experiencedin the original TOPSIS is discarded in a way similar to [10] to avoid inconsistencies when anew network is available or an existing network is removed from the alternatives.

The remainder of the paper is organized as follows: In Section 2, the MADM-related researchliterature is revised. Section 3 presents the proposed network selection method followed in this studyincluding the AHP, the ANP, and the TFT. Section 4 describes a scenario that applies the proposedmethod to accomplish the network selection process. Moreover, a performance evaluation of theproposed method is presented, and its results are discussed compared with the TOPSIS as well as theFuzzy AHP-ELECTRE (FAE) results. Finally, Section 5 concludes our work and presents possiblefuture extensions and plans.

Copyright © 2014 John Wiley & Sons, Ltd. Int. J. Commun. Syst. (2014)DOI: 10.1002/dac

NETWORK ACCESS SELECTION SCHEME BASED ON ANP AND TI-VALUED FUZZY TOPSIS

2. RELATED WORK

Multi attribute decision-making methods are used to select the best alternative among candidatenetworks given a set of criteria with different importance weights. Specifically, MADM algorithmsare able to evaluate criteria of different value ranges, sometimes even contradictory, using multi-criteria analysis. Widely used methods include AHP [11, 12], simple additive weighting (SAW)[12, 13], TOPSIS [12–14], FAE [15], gray relational analysis (GRA) [12, 13], multiplicative expo-nent weighting (MEW) [12, 13], distance to ideal alternative [12], and ANP [16]. Furthermore,various weighting methods are used to provide suitable criteria weights for each alternative. Severalresearch studies use MADM methods for network selection.

Sharma and Khola [14] presented a network selection algorithm based on the TOPSIS algorithm.The proposed algorithm besides the usual parameters (i.e., QoS, bandwidth, and cost) it also takesa prediction of the Received Signal Strength (RSS) into account for the network selection.

Shi and Zhu [11] employed two MADM methods combined with the group decision-making algo-rithm to perform network selection. The proposed procedure defines two types of weights, namelythe objective weights, which consider the current attributes of candidate networks and the subjectiveweights specified according to the subscribers and traffic class preferences. The objective weightsvector is determined using the entropy weighting method while the subjective weights vector isevaluated using the AHP. Then, the group decision-making method employs both vector types toproduce a synthesized vector, whereas the ranking of alternatives is the sum of the product of the nor-malized attribute values with their respective weights. The compatibility of the integrated decisionis finally checked to ensure the effectiveness of the proposed solution. Results showed that the pro-posed method reduces the number of handoffs and improves QoS characteristics of conversationaland interactive traffic flows compared with entropy weighting and GRA approaches.

Lassoued et al. [13] described an evaluation framework of VHO mechanisms, which emulatesapplication characteristics, mobile terminals context, and user and operators preferences. The modelprovides user traces containing information about the location of the users and the QoS performanceof the networks. Current network characteristics are obtained from a mobility simulator emulatingnetwork access technologies, location of access points, and user mobility. The proposed method-ology is used to compare the efficiency of various MADM network selection algorithms includingSAW, TOPSIS, GRA, MEW, and their own proposed scheme called Ubique [17] in a dynamic envi-ronment. Simulation results showed that the examined algorithms achieve good performance, whileUbique is less flexible to changes of delay and cost criteria weights than the other approaches.

Lahby et al. [12] proposed a network selection scheme, which is based on the AHP method andthe Mahalanobis distance. Mahalanobis distance is used to measure the distance of alternatives fromthe correlation of criteria so that the optimal network satisfying the QoS, security, and cost criteria isselected. According to simulation results, both the ranking abnormality problem and the number ofhandoffs in the proposed method are reduced compared with the decision algorithms SAW, MEW,TOPSIS, and distance to ideal alternative.

Lahby et al. [16] proposed a technique for network selection using ANP to estimate the weightsof selection criteria and GRA to rank the alternative networks. Selection criteria include networkrelated attributes while the preference of users is expressed by evaluating different criteria weightsthrough the ANP for each access network. Accordingly, the ANP evaluates the criteria weights ofeach access network separately based on users preferences; in that way, unique criteria weights existfor each network. Simulation results indicated that this method reduces both the ranking abnormalityproblem and the number of handoffs compared with other method variants.

Sheng-mei et al. [18] presented a network selection algorithm making use of the AHP and theentropy weight method to evaluate the weights of network and user related criteria. The candidateaccess networks are identified on the basis of their signal-to-interference-plus-noise ratio (SINR)values. TOPSIS is used for the final ranking of the network alternatives. The proposed methodachieved higher throughput and reduced number of vertical handoffs for various traffic classes com-pared with combined SINR-based vertical handoff [19] and multi-dimensional adaptive SINR-basedvertical handoff [20] algorithms.


E. SKONDRAS ET AL.

Alkhawlani et al. [21] proposed a VHO decision system, which integrates fuzzy logic and TOP-SIS method. Network and user related criteria are each processed by parallel fuzzy logic control(FLC) subsystems, and consequently, TOPSIS is applied to perform the selection of the best networkchoice. Simulation results showed that the proposed solution reduces handover rate and handoverfailure while it increases the percentage of users assigned to networks of their preference, as wellas, the utilization of inexpensive networks.

Alkhawlani and Mohsen [22] presented a network selection system suitable for tightly coupledwireless network environments consisting of two modules. The user software module evaluates thebest network alternative based on selection criteria set by the user including reliability, security, bat-tery power, and price. The operator software module resides at the coordinator of the radio accesstechnologies and performs the final selection decision. It takes into account the network choiceproposed by the user as well as criteria imposed by the operator such as network policies, QoScharacteristics, system capacity, and utilization. The operator module initially uses the FLC subsys-tems of [18] to evaluate the performance of criteria and finally the AHP method to assess the FLCsubsystems outputs and select the best possible network. Simulation results show that the proposednetwork selection scheme achieves better performance in terms of user preferences satisfaction, QoSfulfillment, and operator benefits improvement, than four different reference algorithms performing(i) random selection, (ii) selection based on terminal speed, (iii) selection based on service type, and(iv) selection based on the availability of resources, respectively.

Vasu et al. [23] proposed a fuzzy rule based decision algorithm for vertical handoff in wire-less heterogeneous networks. The algorithm uses QoS performance values as decision parameters,while triangular fuzzy membership functions are used for the fuzzification of the input parametersand the defuzzification of the output result. For the evaluation of the proposed model, a non-birthMarkov chain with states corresponding to available access networks is used. Simulation experi-ments comparing the proposed approach against various MADM methods demonstrated that themethod presented improves the performance of delay sensitive applications.

The use of fuzzy logic for network selection requires the definition of logic rules from specialistswith thorough knowledge of the behavior of the available access networks in various conditions.Furthermore, as the number of selection criteria and the available networks increase, rules becomemore complex, struggling to define effective policies and evaluate the best alternative. Accordingly,the use of fuzzy logic based solutions is limited to handover decision schemes with reduced numberof networks and selection criteria.

Some network selection methods combine fuzzy logic with neural networks to rate the alternativeaccess networks. Accordingly, Gowrishankar et al. [24] created an artificial neural network multi-criteria decision analysis system, which performs network selection using network related attributesexpressed either in crisp or in fuzzy linguistic values. Sensitivity analysis among the proposed solu-tion, the TOPSIS and the SAW methods, is carried out in a network environment consisting offour overlaid networks, where weights of different criteria change and connections of four traffictypes exist. Results show that the proposed method is less stable than TOPSIS but more stable thanSAW in respect to criteria weights changes. Neural network approaches replace the complex logicrules of fuzzy logic approaches, but they still suffer from scalability issues because of the requiredlarge number of the processing elements at their hidden layers as the complexity of criteria and thenumber of networks increase.

Several network selection schemes make use of utility/cost functions to provide performancemetrics for different types of criteria. Rodriguez et al. [25] use a cost function for the networkselection that includes the rules and policies for selecting the best candidate network or for adapt-ing ongoing session parameters. Wu et al. [26] used a set of utility functions to quantify selectioncriteria including the link quality (RSS), battery power, average throughput, network delay, mone-tary cost, and application type. The relative weights of criteria are calculated according to the AHPmethod. Consequently, the candidate networks are ranked using the weighted product method. Sim-ulation results show that the proposed scheme improves network performance and reduces powerconsumption of users terminals. In the approach of Wang et al. [27], the concepts of fuzzy logic,neural network, and utility functions are combined to perform network selection. The proposedmethod uses a fuzzy neural network, which obtains network, user, and terminal related input crite-



ria and evaluates the performance of each access network. Attributes of criteria are defined throughutility functions and processed through the fuzzification, interference, and defuzzification layersof the neural network. A fuzzy version of the particle swarm optimization is used for neural net-work training; however, it is not clear how expected network performance degrees are specifiedduring the learning process. Simulation results show that the proposed method achieves better per-formance in terms of access blocking probability, packet drop probability, and average throughputof access networks compared with other network selection algorithms including GRA, AHP, andgame theoretic.

Generally, there is a rate of uncertainty in characterizing performance measurements as well asrates of influence of performance metrics. Therefore, fuzzy MADM methods expressing uncertainquantities by fuzzy numbers have received the interest of many researchers in decision theory. Inparticular, several fuzzy MADM network selection methods are suggested utilizing linguistic vari-ables, triangular fuzzy numbers, trapezoidal fuzzy numbers, etc. to model network attributes andtheir respective weights.

Chamodrakas and Martakos [10] proposed a method that considers network conditions, QoSconstraints, and energy consumption requirements for network selection criteria. User preferencesindicating the relative importance of criteria in different applications are expressed using linguis-tic expressions, which are transformed to triangular fuzzy numbers. The graded mean integrationmethod is used for the defuzzification of fuzzy numbers into crisp values. Furthermore, utilityfunctions are used to model QoS requirements and energy consumption characteristics of differentapplications. The fuzzy set representation version of TOPSIS is used to combine selection crite-ria and weights to perform the rating of the available networks. The fuzzy set representation ofTOPSIS resolves possible inconsistencies because of conflicting criteria such as bandwidth andenergy consumption. Simulation results show that the proposed method accomplishes a trade-offbetween QoS requirements and energy consumption.

Sasirekha and Ilanzkumaran [28] described two methods to perform network selection. Initially,both methods use a fuzzy version of the AHP technique to obtain the weights of selection crite-ria specifying networks performance. The relative importance matrix resulting from the pairwisecomparison of criteria is fuzzified using triangular fuzzy numbers with membership functions rep-resenting the scale of importance of five levels. Then, the relative importance values are turned intocrisp values using the geometric mean operator while the rest of the steps of the AHP method follow.Subsequently, the former network selection method uses TOPSIS to evaluate the best alternativenetwork based on the weights from AHP and the criteria values of each alternative network. The lat-ter method combines the fuzzy AHP with VIKOR method, which has less complexity and performsequally well as TOPSIS. Evaluation examples are given illustrating that both methods succeed toselect the best network alternative.

Kaleem et al. [29] presented a VHO decision algorithm, which is based on network performancemeasurements to evaluate, firstly, the necessity of making a handoff and, secondly, the best networkalternative in case that handoff is required. To determine the handoff decision, a handoff factoris evaluated and compared with a constant threshold. Network selection is performed using fuzzyTOPSIS. User preferences are defined in the form of criteria weights, while ratings of selectioncriteria and criteria weights are expressed as trapezoidal or triangular fuzzy numbers. Numericalexamples and simulation experiments present the competence of the proposed approach for varioustraffic classes in heterogeneous network access technologies.

Lahby et al. [30] compared the weighting algorithms of AHP, fuzzy AHP, ANP, and fuzzy ANPfor assigning weights to network dependent criteria used by MADM algorithms performing networkselection. To evaluate the effects of the weighting algorithms, the TOPSIS method is used. Resultsshow that all algorithms achieve similar results concerning the network selected. However, the rank-ing abnormality of TOPSIS is reduced, when the ANP weighting method is used for background,conversational, and interactive traffic classes, as well as for streaming traffic.

Zhang [31] performed an analysis of MADM methods for handover decision. Uncertain linguisticterms of decision criteria such as sojourn time and seamlessness are converted to fuzzy data whichin turn are converted to crisp values. SAW and TOPSIS are suggested to perform the final rankingof the candidate networks while results from the sensitivity analysis of these methods conclude


E. SKONDRAS ET AL.

that TOPSIS is more sensitive to the criteria performance and their weights. Moreover, the paperidentifies the handover decision case in which several applications are running simultaneously on aterminal as a group decision problem, although its solution is not being addressed.

3. THE PROPOSED NETWORK SELECTION METHOD

Our proposed method consists of two MADM algorithms: the ANP to calculate the relative impor-tance of the selection criteria and the TFT to accomplish the ranking of the candidate networks. Itshould be noted that the TFT represents the performance of selection criteria using interval-valuedtrapezoidal fuzzy numbers. The following of this section presents the algorithms that this methodemploys, as well as an overview of the interval-valued trapezoidal fuzzy numbers that TFT algorithmutilizes.

3.1. The analytic network process

The ANP was also introduced by Saaty [32] to deal with decision problems that criteria and alter-natives depend on each other. ANP is actually the generalization of the AHP. A decision problemthat is analyzed with the ANP can be designed either as a control hierarchy or as a nonhierarchicalnetwork. Nodes of the network represent components (or clusters) of the system while arcs denoteinteractions between them. All interactions and feedbacks within clusters are called inner depen-dencies, while interactions and feedbacks between clusters are called outer dependencies. The ANPis composed of four major steps [33]:

Step 1. Model construction and problem structuring: During this step, the problem is analyzedand decomposed into a rational system, such as a network.

Step 2. Pairwise comparison matrices and priority vectors: During this step, the pairwisecomparison matrix, as in AHP, is derived using Saaty’s nine-point importance scale(Table I).

Step 3. Supermatrix formation: During this step, the supermatrix of the ANP model is con-structed to represent the inner and outer dependencies of the network. It is actually apartitioned matrix, where each matrix segment represents a relationship between twoclusters in the network. To construct the supermatrix, the local priority vectors obtainedin step 2 are grouped and placed in the appropriate positions in a supermatrix based onthe flow of influence from one cluster to another, or from a cluster to itself, as in theloop. Then, the supermatrix is transformed to a stochastic one, the weighted superma-trix. Finally, the weighted supermatrix is raised to limiting powers until all the entriesconverge to calculate the overall priorities, and thus, the cumulative influence of each ele-ment on every other element with which it interacts is obtained [34]. At this point, all thecolumns of the new matrix, the limit supermatrix, are the same, and their values show theglobal priority of each element of network.

For example, if we assume a network with n clusters, where each cluster Ck; k D1; 2; ; n; and has mn elements, denoted as ek1; ek2; ; ekmk , then the standard form for asupermatrix can be expressed as

Table I. Analytic hierarchy process.

Importance Definition

1 Equal importance3 Moderate importance5 Strong importance7 Very strong importance9 Extreme importance2, 4, 6, 8 Intermediate values



W D

C1 : : : Ck : : : Cne11 : : : e1m1 : : : ek1 : : : ekmk : : : en1 : : : enmn2

666666666666666666664

3777777777777777777775

e11

C1::: W11 : : : W1k : : : W1n

e1m1:::

::::::

::::::

:::

ek1

Ck::: Wk1 : : : Wkk : : : Wkn

ekmk:::

::::::

::::::

:::

en1

Cn::: Wn1 : : : Wnk : : : Wnn

enmn

(1)

Step 4. Selection of the best alternatives: If the supermatrix formed in step 3 covers the wholenetwork, then the priority weights of the alternatives can be found in the column of alter-natives in the normalized supermatrix. Otherwise, additional calculations are required inorder to obtain the overall priorities of the alternatives. The alternative with the largestoverall priority should be selected, as it is the best alternative as determined by thecalculations made using matrix operations.

3.2. The trapezoidal interval-valued fuzzy numbers

The concept of fuzzy logic was introduced by Zadeh [35] and is used to make a decision fromindeterminate and approximate information. A fuzzy number is represented by a set of real valuesrepresenting an uncertain quantity and a convex normalized continuous function, which estimatesthe degree of membership for each value in the subset. Triangular or trapezoidal fuzzy numbers arefrequently used to represent uncertain information. A trapezoidal fuzzy number can be defined as avector x D .x1; x2; x3; x4; v OA/ with membership function:

�.x/ D

8<ˆ:

x�x1x2�x1

; if x1 6 x < x2Iv OA; if x2 6 x 6 x3Ix�x4x3�x4

; if x3 < x 6 x4I0; otherwise.

(2)

where x1 < x2 < x3 < x4 and v OA 2 Œ0; 1�.An interval-valued fuzzy number (IVFN) introduced by Sambuc [36] is defined as AD ŒAL; AU �

consisting of the lower AL and the upper AU fuzzy numbers. IVFNs replace the crisp membershipvalues by intervals in Œ0; 1�. They were proposed because fuzzy information can be better expressedby intervals than by single values. Liu and Jin [37] and Cornelis et al. [38] suggest that IVFNsare useful in multiple criteria decision-making problems and particularly in cases where attributevalues are in the form of linguistic expressions. Therefore, Ashtiani et al. [39] propose an extensionof the fuzzy TOPSIS method using interval-valued triangular fuzzy numbers. Moreover, Liu andJin [37] propose a decision-making method using weighted geometric aggregation operators onattribute values expressed in the form of interval-valued trapezoidal fuzzy numbers. According tothe definition in [39], an IVFN A is defined as follows:

A D®�x;��LA.x/; �

UA .x/

��¯(3)

�LA.x/; �UA .x/ W X ! Œ0; 1�8x 2 X;�LA.x/ < �

UA .x/ (4)


E. SKONDRAS ET AL.

Figure 1. The interval-valued trapezoidal fuzzy numbers.

O�A.x/ D��LA.x/; �

UA .x/

�(5)

A D ¹.x; O�A.x//º; x 2 .�1;1/ (6)

In particular, the interval-valued trapezoidal fuzzy number defined in [40] is the most general formof fuzzy number (Figure 1) and can be represented as A D ŒAL; AU � D

��xL1 ; x

L2 ; x

L3 ; x

L4 ; vAL

�;�

xU1 ; xU2 ; x

U3 ; x

U4 ; vAU

��, where 0 6 xL1 6 xL2 6 xL3 6 xL4 6 1, 0 6 xU1 6 xU2 6 xU3 6 xU4 6 1,

0 6 vAL 6 vAU 6 1 and AL � AU . The operational rules of the interval-valued trapezoidal fuzzynumbers are defined in [40].

3.3. The trapezoidal interval-valued fuzzy TOPSIS algorithm

The Technique for order preference by similarity to ideal solution (TOPSIS) introduced by Hwangand Yoon [41] is based on the concept that the best alternative should have the shortest distance fromthe positive ideal solution and the longer distance from the negative ideal solution. In the presentwork, network selection is performed using a proposed fuzzy version of TOPSIS, namely TFT. Thismethod assumes that the linguistic values of criteria attributes are represented by interval-valuedtrapezoidal fuzzy numbers.

Suppose A D ¹A1; A2; : : : ; Anº is the set of possible alternatives, C D ¹C1; C2; : : : ; Cnº is theset of criteria, and w1; w2; : : : ; wm are the weights of each criterion. The steps of the method areas follows:

Step 1. Construction of the decision matrix: Each xij element of the n � m decision matrixD is an interval-valued trapezoidal fuzzy number, which expresses the performance ofalternative i for criterion j . Thus,

D D

C1 : : : CmA1 x11 : : : x1m:::

:::: : :

:::

An xn1 : : : xnm

(7)

where xij Dh�xLij1; x

Lij2; x

Lij3; x

Lij4; v

Lij

�;�xUij1; x

Uij2; x

Uij3; x

Uij4; v

Uij

�iIn case there are Q decision makers, the decision matrix and the criteria weights

include the average of the performance values and weights, respectively, of the decisionmakers. Hence, assuming that for the k-th decision maker, xijk is the performance ofalternative i for criterion j , and wjk is the importance weight for criterion j ; the averageof the performance values and weights are given by



xij D1

Q

QXkD1

xijk

D

240@ 1

Q

QXkD1

xLijk1;1

Q

QXkD1

xLijk2;1

Q

QXkD1

xLijk3;1

Q

QXkD1

xLijk4; vLijk

1A ;

0@ 1

Q

QXkD1

xUijk1;1

Q

QXkD1

xUijk2;1

Q

QXkD1

xUijk3;1

Q

QXkD1

xUijk4; vUijk

1A35

(8)

and

wj D1

Q

QXkD1

wjk : (9)

Step 2. Normalization of the decision matrix: Consider that ˝b is the set of benefits attributesand˝c is the set of costs attributes. Then, the elements of the normalized decision matrixare computed as

(a)

rij D

" xLij1

bj;xLij2

bj;xLij3

bj;xLij4

bj; vLij

!;

xUij1

bj;xUij2

bj;xUij3

bj;xUij4

bj; vUij

!#(10)

where bj D maxi xUij4 for each j 2 ˝b .(b)

rij D

" cj

xLij4;cj

xLij3;cj

xLij2;cj

xLij1; vLij

!;

cj

xUij4;cj

xUij3;cj

xUij2;cj

xUij1; vUij

!#(11)

where cj D mini xLij4 for each j 2 ˝c .

Step 3. Construction of the weighted normalized decision matrix: The weighted normalized deci-sion matrix is constructed by multiplying each element of the normalized decision matrixrij with the respective weight wj according to the formula.

uij D��rLij1 � wj ; r

Lij2 � wj ; r

Lij3 � wj ; r

Lij4 � wj ; v

Lij

�;�

rUij1 � wj ; rUij2 � wj ; r

Uij3 � wj ; r

Uij4 � wj ; v

Uij

�� (12)

Step 4. Determination of the positive and negative ideal solution: The positive ideal solution isdefined as

XC Dh�xCLij1 ; x

CLij2 ; x

CLij3 ; x

CLij4 ; v

CLij

�;�xCUij1 ; x

CUij2 ; x

CUij3 ; x

CUij4 ; v

CUij

�i

D

" î

uLij1;î

uLij2;î

uLij3;î

uLij4; vLij

!;

î

uUij1;î

uUij2;î

uUij3;î

uUij4; vUij

!# (13)

whereVi

� maxi in case j 2 ˝b andVi

� mini in case j 2 ˝c .


E. SKONDRAS ET AL.

The negative ideal solutions are defined accordingly as

X� D��x�Lij1 ; x

�Lij2 ; x

�Lij3 ; x

�Lij4 ; v

�Lij

�;�x�Uij1 ; x

�Uij2 ; x

�Uij3 ; x

�Uij4 ; v

�Uij

��D

" _i

uLij1;_i

uLij2;_i

uLij3;_i

uLij4; vLij

!;

_i

uUij1;_i

uUij2;_i

uUij3;_i

uUij4; vUij

!# (14)

whereWi � mini in case j 2 ˝b and

Wi � maxi in case j 2 ˝c .

Step 5. Measurement of the distance of each alternative from the ideal solutions: The distancesof each alternative from the positive ideal solution are evaluated as follows:

dCi1 D

mXjD1

²1

4

��uLij1 � x

CLij1

�2C�uLij2 � x

CLij2

�2C�uLij3 � x

CLij3

�2C�uLij4 � x

CLij4

�2³ 12(15)

dCi2 D

mXjD1

²1

4

��uUij1 � x

CUij1

�2C�uUij2 � x

CUij2

�2C�uUij3 � x

CUij3

�2C�uUij4 � x

CUij4

�2³ 12(16)

Likewise, the distances of each alternative from the negative ideal solution are estimated

d�i1 D

mXjD1

²1

4

h�uLij1 � x

�Lij1

�2C�uLij2 � x

�Lij2

�2C�uLij3 � x

�Lij3

�2C�uLij4 � x

�Lij4

�2i³ 12(17)

d�i2 D

mXjD1

²1

4

h�uUij1 � x

�Uij1

�2C�uUij2 � x

�Uij2

�2C�uUij3 � x

�Uij3

�2C�uUij4 � x

�Uij4

�2i³ 12(18)

Consequently, similar to [39], the distance of the alternatives from the positive and nega-tive ideal solutions are expressed by intervals such as ŒdCi1 ; d

Ci2 � and Œd�i1; d

�i2�, instead of

single values. In this way, less information is lost.Step 6. Calculation of the relative closeness: The relative closeness of the distances from the

ideal solutions are computed as

RCi1 Dd�i1

dCi1 C d�i1

(19)

and

RCi2 Dd�i2

dCi2 C d�i2

(20)

The compound relative closeness is obtained from the average of the aforementionedvalues

RCi DRCi1 CRCi2

2(21)

Step 7. Alternatives ranking: The alternatives are ranked according to their RCi values. The bestalternative is the one with the highest RCi value.



Tabl

eII

.Q

oScl

ass

map

ping

and

serv

ice-

leve

lagr

eem

ents

.

LTE

QC

IW

iMA

XIE

EE

802.

11e

Req

uire

dR

equi

red

Req

uire

dR

equi

red

(typ

e/pr

iori

ty)

QoS

clas

sQ

oScl

ass

thro

ughp

utpa

cket

loss

dela

y(m

s)jit

ter

(ms)

Serv

ices

1(G

BR

/2)

UG

S/er

tPS

(802

.16e

–802

.16

m)

AC

_VO

200

Kbp

s10�2

100

50V

oIP,

CV

ideo

,BSt

ream

ing,

RT

Gam

ing,

Web

3(G

BR

/3)

UG

SA

C_V

O25

0K

bps

10�3

5040

CV

ideo

,BSt

ream

ing,

RT

Gam

ing,

Web

2(G

BR

/4)

UG

SA

C_V

I8

Mbp

s10�3

6550

CV

ideo

,BSt

ream

ing,

Web

4(G

BR

/5)

rtPS

AC

_VI

8M

bps

10�5

6560

CV

ideo

,BSt

ream

ing,

Web

6(N

on-G

BR

/6)

nrtP

SA

C_B

E2.

5M

bps

10�5

200

N/A

BSt

ream

ing,

Web

7(N

on-G

BR

/7)

nrtP

SA

C_B

E2

Mbp

s10�5

160

100

BSt

ream

ing,

Web

8(N

on-G

BR

/8)

BE

AC

_BE

1.5

Mbp

s10�3

300

N/A

Web

9(N

on-G

BR

/9)

BE

AC

_BE

1.5

Mbp

s10�5

300

N/A

LTE

,lon

g-te

rmev

olut

ion;

QC

I,Q

oScl

ass

indi

cato

r;U

GS,

unso

licite

dgr

ants

ervi

ce;G

BR

,gua

rant

eed

bitr

ate.


E. SKONDRAS ET AL.

4. SIMULATION SETUP AND RESULTS

In our experiments, we consider a heterogeneous network environment consisting of a number oflong-term evolution (LTE), WiMAX, and WiFi networks. Each network can provide at least one ofthe following five service types: Voice-over-Internet protocol (VoIP), conversational video (CVideo),buffered streaming (BStreaming), real time gaming (RTGaming), and Web browsing. In order toallow service continuity, QoS mapping among the QoS classes of the different access technologiesis required. Table II shows this mapping relation among the different technologies.

Four SLAs are defined, with SLA1 having the highest service priority and SLA4 having the lowestservice priority. SLA1 supports all service types, as well as provides the best values for QoS andpolicy decision criteria. SLA2 supports less service types, by not providing support for the VoIP andreal time gaming services. Additionally, it provides slightly worse decision criteria values than thoseoffered by the SLA1. SLA3 supports only the buffered streaming and the Web browsing servicesand satisfactory QoS characteristics and policies. Whereas the low price SLA4 supports only theWeb browsing service while providing acceptable decision criteria values.

Network selection weightsper service & SLA

Network QoSCharacteristics

Network PolicyCharacteristics

Throughput

Delay

Jitter

Packet Loss

Service Reliability

Security

Price

Goal

Criteria Groups

Criteria

Figure 2. The analytic network process network model.

Throughput

Delay Jitter

Packet loss

ServiceReliability

Price

Security

Figure 3. Relations of criteria.

Table III. The analytic network process supermatrix for SLA1 Voice-over-Internet protocol service.

Throughput Delay Jitter Packet loss Price Reliability Security

Throughput 0.015625 0.015625 0.015625 0.015625 0.015625 0.015625 0.015625Delay 0.328125 0.328125 0.328125 0.328125 0.328125 0.328125 0.328125Jitter 0.328125 0.328125 0.328125 0.328125 0.328125 0.328125 0.328125Packet loss 0.328125 0.328125 0.328125 0.328125 0.328125 0.328125 0.328125Price 0.05 0.05 0.05 0.05 0.019607 0.05 0.0625Reliability 0.95 0.95 0.95 0.95 0.759804 0.95 0Security 0 0 0 0 0.220588 0 0.9375



The ANP method is applied in order to estimate the weights of network selection criteria perservice type and SLA. Figure 2 depicts the ANP network model. Criteria are classified into twogroups, namely the QoS and the policy characteristics. The QoS characteristics group contains net-work performance related criteria including throughput, delay, jitter, and packet loss while the policycharacteristics group contains operator defined rules such as price, security, and service reliabil-ity. Service reliability determines the ability for service constraints satisfaction and optimization ofperformance when a network is congested. Pairwise comparison decision matrices are created on

Table IV. The analytic network process weighted supermatrix for SLA1 Voice-over-Internet protocol service.


Throughput 0.0078125 0.0078125 0.0078125 0.0078125 0.0078125 0.0078125 0.0078125Delay 0.164062 0.164062 0.164062 0.164062 0.164062 0.164062 0.164062Jitter 0.164062 0.164062 0.164062 0.164062 0.164062 0.164062 0.164062Packet loss 0.164062 0.164062 0.164062 0.164062 0.164062 0.164062 0.164062Price 0.025 0.025 0.025 0.025 0.00980392 0.025 0.03125Reliability 0.475 0.475 0.475 0.475 0.379902 0.475 0Security 0 0 0 0 0.110294 0 0.46875

Table V. The analytic network process limit supermatrix for SLA1 Voice-over-Internet protocol service.


Throughput 0.0078125 0.0078125 0.0078125 0.0078125 0.0078125 0.0078125 0.0078125Delay 0.164062 0.164062 0.164062 0.164062 0.164062 0.164062 0.164062Jitter 0.164062 0.164062 0.164062 0.164062 0.164062 0.164062 0.164062Packet loss 0.164062 0.164062 0.164062 0.164062 0.164062 0.164062 0.164062Price 0.0246573 0.0246573 0.0246573 0.0246573 0.0246573 0.0246573 0.0246573Reliability 0.470224 0.470224 0.470224 0.470224 0.470224 0.470224 0.470224Security 0.0051191 0.0051191 0.0051191 0.0051191 0.0051191 0.0051191 0.0051191

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

VoIP Conversational Video Buffered Streaming Real Time Gaming Web

Wei

ght

SLA1

ThroughputDelayJitterPacket LossPriceReliabilitySecurity

Figure 4. Criteria weights for SLA1.


E. SKONDRAS ET AL.

the basis of relations among the seven selection criteria depicted in Figure 3. Then, these pairwisecomparison decision matrices are used to evaluate the priority vectors of criteria and form the super-matrix per service type and SLA. Subsequently, the weighted supermatrices and, finally, the limitsupermatrices are obtained. Indicatively, for the SLA1 VoIP service, the initial, the weighted, andthe limit supermatrices are presented in Tables III–V, respectively.

The criteria weights per service and SLA obtained by the limit supermatrices are presented inFigures 4–7. As illustrated, the weights are proportional to the constraints of each service as wellas to the agreements of each SLA. In particular, the weight of the price criterion is low for SLA1,

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

Conversational Video Buffered Streaming Web

Wei

ght

SLA2



0

0.1

0.2

0.3

0.4

0.5

WebBuffered Streaming

Wei

ght

SLA3


Figure 6. Criteria Weights for SLA3.



0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

Web

Wei

ght

SLA4



Table VI. Linguistic terms and the corresponding interval-valued trapezoidalfuzzy numbers.

Linguistic term Interval-valued trapezoidal fuzzy number

Absolutely poor [(0.0, 0.0, 0.0, 0.0, 0.8), (0.0, 0.0, 0.0, 0.0, 1)]Very poor [(0.01, 0.02, 0.03, 0.07, 0.8), (0.0, 0.01, 0.05, 0.08, 1)]Poor [(0.04, 0.1, 0.18, 0.23, 0.8), (0.02, 0.08, 0.2, 0.25, 1)]Medium poor [(0.17, 0.22, 0.36, 0.42, 0.8), (0.14, 0.18, 0.38, 0.45, 1)]Medium [(0.32, 0.41, 0.58, 0.65, 0.8), (0.28, 0.38, 0.6, 0.7, 1)]Medium good [(0.58, 0.63, 0.8, 0.86, 0.8), (0.5, 0.6, 0.9, 0.92, 1)]Good [(0.72, 0.78, 0.92, 0.97, 0.8), (0.7, 0.75, 0.95, 0.98, 1)]Very good [(0.93, 0.98, 1, 1, 0.8), (0.9, 0.95, 1, 1, 1)]Absolutely good [(1, 1, 1, 1, 0.8), (1, 1, 1, 1, 1)]

Table VII. Relation of the network QoS characteristics and linguistic terms for Voice-over-Internet protocol.

Linguistic term Throughput range (Kbps) Delay range (ms) Jitter range (ms) Packet loss range

Absolutely poor 6 164 > 116 > 65 > 0:4Very poor 165–174 111–115 55–64 > 0:2–0.4Poor 175–184 106–110 45–54 >10�1–<0.2Medium poor 185–194 100–105 40–44 10�1

Medium 195–204 95–99 35–49 10�2

Medium good 205–214 86–94 30–34 10�3

Good 215–224 66–85 25–29 10�4

Very good 225–239 41–65 20–24 10�5

Absolutely good > 240 6 40 6 20 6 10�6

in which the service reliability and the network QoS characteristics are considered as the mostimportant factors. In SLA2, the price criterion is more important than in SLA1, thus the respectiveweight is greater than that of SLA1. Consequently, the weights of the service reliability and QoScharacteristics criteria in SLA2 are lower compared to the relative weights of SLA1. In SLA3,


E. SKONDRAS ET AL.

Table VIII. The available networks of SLA1 and SLA2.

SLA Service Network Throughput Delay Jitter Packet loss Price Service reliability Security

SLA1 VoIP LTE 1 MG AG VG VG VP AG VGLTE 2 AG MG AG MG VP VG AGWiMAX 1 M M MP AG P VG AGWiMAX 2 G G G G P AG VGWiFi 1 VG VG MG AG MP G GWiFi 2 MG M MG VG MP MG MGWiFi 3 M MP M AG MP G G

CVideo LTE 1 MP MG VG G AP AG VGLTE 2 AG AG AG VG AP VG AGWiMAX 1 MP M MG AG AP VG AGWiMAX 2 MG MG G AG VP AG VGWiFi 1 M MG M VG P G GWiFi 2 VG VG VG AG P MG MGWiFi 3 G G M VG P G G

BStreaming LTE 1 M G VG VG AP AG VGLTE 2 VG VG AG AG AP VG AGWiMAX 1 M MG MG VG VP VG AGWiMAX 2 MG G MG G P AG VGWiFi 1 VG G M AG P G GWiFi 2 AG AG G VG P MG MGWiFi 3 G VG VG AG MP G G

RTGaming LTE 1 G AG AG VG VP AG VGLTE 2 G MG VG AG VP VG AGWiMAX 1 MP MG G AG P VG AGWiMAX 2 VG AG AG VG VP AG VGWiFi 1 AG VG VG VG VP G GWiFi 2 M M MG AG MP MG MGWiFi 3 P M M AG MP G G

Web LTE 1 AG AG AG AG VP AG VGLTE 2 MG M G VG MP VG AGWiMAX 1 G M G AG P VG AGWiMAX 2 VG G VG AG P AG VGWiFi 1 MG MP MG VG MP G GWiFi 2 VG G M VG MP MG MGWiFi 3 AG VG AG AG MP G G

SLA2 CVideo LTE 1 MG G VG AG MP G GLTE 2 MP M MG VG M G GWiMAX 1 M MG G AG MP MG MGWiMAX 2 MP M M AG M MG MGWiFi 2 G VG VG AG MG G MWiFi 3 MP G M VG MG P M

BStreaming LTE 1 M G G VG MP G GLTE 2 MG MG AG G MP G GWiMAX 1 M MG MP AG MP MG MGWiMAX 2 G G MG VG MP MG MGWiFi 1 G VG MG AG MP MP MPWiFi 2 AG AG VG VG MP M MWiFi 3 MG VG VG AG M P M

Web LTE 2 M MP MG VG M G GWiMAX 1 MG M G AG MG MG MGWiMAX 2 VG G AG AG M MG MGWiFi 1 MG MP M VG MG MP MPWiFi 2 MG M G VG MG M MWiFi 3 VG VG AG AG MG P M

AG, absolutely poor; VP, very poor; P, poor; MP, medium poor; M, medium; MG, medium good; G, good; VG,very good; AG, absolutely good; SLA, service-level agreement; LTE, long-term evolution.



the weights of price and service reliability criteria are balanced as they are almost of equivalentimportance. Finally, in SLA4, the price is the most important criterion resulting in a high estimatedweight value.

Ranking of the networks alternatives is performed using the TFT algorithm described in Section 3.The weights of network selection criteria are obtained from Figures 4–7. The linguistic terms forthe criteria attributes are represented by interval-valued trapezoidal fuzzy numbers as shown inTable VI. Network policy specifications are expressed directly using linguistic terms. Additionally,crisp values of network QoS characteristics are converted into linguistic terms, which correspond tospecific ranges of values per service type. Specifically, Table VII presents a relative example for theVoIP service, illustrating the correspondence between ranges of network QoS characteristics valuesand linguistic terms.

The available-candidate networks in our simulations at the time of network selection per serviceand SLA, as well as, their specifications expressed by linguistic terms, are depicted in Tables VIIIand IX.

The case of having several services of different QoS constraints running at the user site is beingaddressed, and network selection is performed in a way satisfying multiple groups of criteria peruser. Specifically, we consider the case where nine users need to select a network that satisfiesthe requirements of their services as presented in Table X and at the same time comply with theirrespective SLA agreements. To achieve this goal, the proposed TFT algorithm is applied for eachuser, and the available networks are ranked as shown in Figure 8. The positive and negative idealsolutions are represented by unary and null trapezoidal fuzzy numbers, respectively, to eliminate theranking abnormality problem.

From the obtained results, it is clear that the ranking of the network alternatives is in accordancewith the users expectations. For example, user 1 requiring increased QoS provisioning selects LTE 1network, which guarantees the best QoS characteristics and service reliability. As Figure 8 depicts,LTE 1 achieves higher ranking than the other networks, because of the high values of the QoScharacteristics and service reliability factors bearing higher importance according to the relativeANP weights in SLA1. On the contrary, user 9, whose prior selection criterion is the price of the

Table IX. The available networks of SLA3 and SLA4.

SLA Service Network Throughput Delay Jitter Packet loss Price Service reliability Security

SLA3 BStreaming LTE 1 M MG G VG MG MP MPLTE 2 G G M AG MG M MWiMAX 1 M G MP VG MG M MWiFi 1 G G MG AG G VP PWiFi 2 G AG G VG MG VP PWiFi 3 MG VG MG AG G VP P

Web LTE 1 MG MP M G G MP MPLTE 2 M M G VG G M MWiMAX 1 MG M M AG G M MWiMAX 2 VP M AG AG VG P MPWiFi 1 MG MP M AG G VP PWiFi 2 AP AP VP G VG VP P

SLA4 Web LTE 1 MP M M VG VG P PLTE 2 M M G MG VG P MPWiMAX 1 VP P M AG AG VP VPWiMAX 2 P MP MP G VG VP PWiFi 1 MG G M G AG AP APWiFi 2 AP AP VP G AG AP VPWiFi 3 AP VP P AG AG AP VP

AG, absolutely poor; VP, very poor; P, poor; MP, medium poor; M, medium; MG, medium good; G, good; VG,very good; AG, absolutely good; SLA, service-level agreement; LTE, long-term evolution.


E. SKONDRAS ET AL.

Table X. The required services per user.

User SLA Required services

1 SLA1 -VoIP2 SLA1 -VoIP,-RTGaming, -BStreaming, -Web3 SLA1 -VoIP, -RTGaming4 SLA1 -RTGaming5 SLA2 -CVideo6 SLA2 -BStreaming7 SLA3 -BStreaming, -Web8 SLA3 -Web9 SLA4 -Web

SLA, service-level agreement; VoIP, Voice-over-Internet protocol.

0

0.05

0.1

0.15

0.2

User 1 User 2 User 3 User 4 User 5 User 6 User 7 User 8 User 9

Net

wor

k S

core

Trapezoidal Fuzzy Topsis Results

LTE 1LTE 2WiMAX 1WiMAX 2WiFi 1WiFi 2WiFi 3

Figure 8. The TFT results.

service, selects the WiFi 1 network, which satisfies his or her requirements in respect of his or herSLA agreement.

4.1. Performance evaluation of the TFT algorithm

The performance of TFT algorithm was evaluated against the original TOPSIS method, as well as,the method presented in [15], the FAE method. The FAE method calculates the criteria weightsusing the fuzzy AHP and performs the network selection by applying the ELECTRE algorithm. Weconsider the scenario of the nine users of Table X. A critical weakness of the TOPSIS and FAEis that they do not support users with more than one service. In these cases, the TOPSIS and FAEmethods consider only the most demanding service of the user. Specifically, for users 2 and 3, theyapplied only for the VoIP service; for user 7, it is applied only for the BStreaming service; and forthe rest of the users, the methods are applied, respectively, for each single user service defined inTable X.

Table XI presents the networks classification performed by the proposed TFT, the TOPSIS, andthe FAE algorithms, respectively. From the analysis of the results, we conclude that when a userhas only one service, the methods usually provide similar results. However, when a user requiresmultiple services, the TFT accomplishes more reliable results than the TOPSIS and FAE, because



Tabl

eX

I.N

etw

orks

clas

sific

atio

nin

resp

ecto

ftr

apez

oida

lint

erva

l-va

lued

Fuzz

yT

OPS

IS(T

FT),

tech

niqu

efo

ror

der

pref

eren

ceby

sim

ilari

tyto

idea

lsol

utio

n(T

OPS

IS)

(T),

and

Fuzz

yA

HP–

EL

EC

TR

E(F

AE

)re

sults

.

Net

wor

ksU

ser

1U

ser

2U

ser

3U

ser

4U

ser

5U

ser

6U

ser

7U

ser

8U

ser

9

Met

hod

TFT

TFA

ET

FTT

FAE

TFT

TFA

ET

FTT

FAE

TFT

TFA

ET

FTT

FAE

TFT

TFA

ET

FTT

FAE

TFT

TFA

E

LTE

11

11

21

11

11

21

14

22

42

54

25

21

44

64

LTE

23

33

33

33

33

44

22

36

11

41

31

32

53

34

WiM

AX

15

55

55

55

55

55

43

43

54

72

14

13

16

73

WiM

AX

22

24

12

42

24

12

15

54

23

2—

——

55

35

54

WiF

i14

62

46

24

62

33

3—

——

66

33

53

44

21

11

WiF

i26

46

74

66

46

67

51

11

35

15

42

66

67

44

WiF

i37

76

67

67

76

76

66

65

77

6—

——

——

—2

22

The

bold

valu

esre

pres

entt

hebe

stal

tern

ativ

epr

ovid

edby

each

met

hod.


E. SKONDRAS ET AL.

Figure 9. TFT’s networks ranking in case of networks environment changes.



it considers the weights of each service. For example, the results concerning user 1 using only theVoIP service are similar for TFT and TOPSIS methods with the exception of the evaluation sequenceof the WiFi 1 and WiFi 2 networks. Also, FAE accomplishes quite similar network rates with theTFT and TOPSIS methods for this user. Nevertheless, TFT succeeds more reliable results for user4, compared with TOPSIS and FAE methods. In this case, only the RTGaming service is used, andthe most important criteria are service reliability, throughput, and delay. TFT selects the WiMAX 2network, which provides AG for service reliability, VG for throughput, and AG for delay criterion.On the other hand, TOPSIS selects the LTE 1 network, which has similar values with the WiMAX 2for service reliability and delay criteria but worse performance for throughout criterion by providingG instead of VG. Moreover, FAE does not provide a clear choice for user 4 and results to equalevaluation sequence for both WiMAX 2 and LTE 1 networks. Finally, the classification of networksobtained from the three methods is quite different for user 7 who requests both BStreaming andWeb browsing services, and the TFT accomplishes more reliable results by taking into account theweights of both services.

4.2. Sensitivity analysis of the TFT

In this section, the sensitivity of the TFT is evaluated when the number of the available access net-works changes frequently. Particularly, we consider three different network configuration scenariosfor the users defined in Table X. In the first scenario, all networks defined in Tables VIII and IX areavailable. In the second and third scenarios, the LTE 1 and the WiFi 2 networks, respectively, arenot reachable. The graphs of Figure 9 include three column types of different pattern indicating theranking of network alternatives in each case. Particularly, in the first case, user 1 selects the LTE 1network. In the second case, the remaining networks improve their ranking order thus user 1 selectsthe WiMAX 2 network. Furthermore, in the third case, only the last rated WiFi 3 network increasesits rank, because the WiFi 2 network preceded WiFi 3 in the other two cases. Similar behavior isobserved in the ranking of network alternatives for the other users. From the aforementioned analy-sis, we conclude that ranking results of the proposed method are normally adjusted with respect tothe heterogeneous network environment changes, highlighting thus the methods sensitivity.

5. CONCLUSIONS

Network selection in heterogeneous networks is a complex task because it takes into accountdifferent parameters with different relative importance, such as the network and the applicationcharacteristics, the user preferences, and the service cost. This paper presents a network selectionmethod that takes into account the network QoS characteristics policies, application requirements,and different types of users SLAs to select the optimal network that will satisfy simultaneously allthe applications’ requirements and user’s preferences running on a mobile user’s device.

More specifically, the proposed method employs two MADM algorithms: the ANP for criteriaweights calculation and the TFT for accomplishing the overall rating of the network technologies.The ANP is selected to determine the relative importance and the dependence of the criteria. Asselection criteria, we consider the network QoS parameters, service constraints, user requirements,and provider policies. These criteria are easily configured and represented by interval-valued trape-zoidal fuzzy numbers. Then, the TFT algorithm is applied to calculate the overall rating of theavailable networks.

Performance evaluation of the TFT showed that when a user has only one service, it providessimilar results to the original TOPSIS and FAE methods. However, when a user requires multipleservices, the TFT performs better by satisfying multiple groups of criteria per user because theoriginal TOPSIS and FAE methods cannot support more than one services. Furthermore, accordingto the sensitivity analysis of results, it is showed that the described method does not suffer from theranking abnormality problem; thus, the results are normally adjusted to the heterogeneous networkenvironment changes.

Our future work will be focused on the design of a complete solution for the VHO process withthe proposed method as the main mechanism for the network selection step.


E. SKONDRAS ET AL.

REFERENCES

1. Khanjari SA, Arafeh B, Day K, Alzeidi N. Bandwidth borrowing-based qos approach for adaptive call admis-sion control in multiclass traffic wireless cellular networks. International Journal of Communication Systems 2013;26(7):811–831.

2. Ma X, Liu J, Jiang H. On the design of algorithms for mobile multimedia systems: a survey. International Journal ofCommunication Systems 2011; 24(10):1330–1339.

3. Ou S, Pan H, Li F. Heterogeneous wireless access technology and its impact on forming and maintaining friendshipthrough mobile social networks. International Journal of Communication Systems 2012; 25(10):1300–1312.

4. Gelabert X, Sallent O, Prez-Romero J, Agust R. Performance evaluation of radio access selection strategies inconstrained multi-access/multi-service wireless networks. Computer Networks 2011; 55(1):173–192.

5. Verma R, Singh N. GRA based network selection in heterogeneous wireless networks. Wireless Personal Communi-cations 2013; 72(2):1437–1452.

6. Halgamuge MN, Vu H, Ramamohanarao K, Zukerman M. A call quality performance measure for handoffalgorithms. International Journal of Communication Systems 2011; 24(3):363–383.

7. Wu J-S, Yang S-F, Hwang B-J. A terminal-controlled vertical handover decision scheme in ieee 802.21-enabledheterogeneous wireless networks. International Journal of Communication Systems 2009; 22(7):819–834.

8. Radhika K, Redd AV. Vertical handoff decision using game theory approach for multi-mode mobile terminals in nextgeneration wireless networks. International Journal of Computer Applications 2011; 36:31–37.

9. Wang L, Kuo G-S. Mathematical modeling for network selection in heterogeneous wireless networks—a tutorial.Communications Surveys Tutorials, IEEE 2013; 15(1):271–292.

10. Chamodrakas I, Martakos D. A utility-based fuzzy TOPSIS method for energy efficient network selection inheterogeneous wireless networks. Applied Soft Computing 2012; 12(7):1929–1938.

11. Shi Z, Zhu Q. Network selection based on multiple attribute decision making and group decision making for het-erogeneous wireless networks. The Journal of China Universities of Posts and Telecommunications 2012; 19(5):92–114.

12. Lahby M, Leghris C, Adib A. New multi access selection method based on mahalanobis distance. AppliedMathematical Sciences 2012; 6(53-56):2745–2760.

13. Lassoued I, Bonnin J-M, Ben Hamouda Z, Belghith A. A methodology for evaluating vertical handoff decisionmechanisms. 7th International Conference on Networking, ICN 2008, Cancun, Mexico, 2008; 377–384.

14. Sharma M, Khola R. Pre decision based handoff in multi network environment. Advances in Computer Science,Engineering & Applications, New Delhi, India, 2012; 609–616.

15. Charilas DE, Markaki OI, Psarras J, Constantinou P. Application of fuzzy AHP and ELECTRE to network selection.1st International Conference on Mobile Lightweight Wireless Systems, Athens, Greece, 2009; 63–73.

16. Lahby M, Leghris C, Adib A. New multi access selection method using differentiated weight of access inter-face. International Conference on Communications and Information Technology (ICCIT), Beirut, Lebanon, 2012;237–242.

17. Suciu L, Bonnin JM, Guillouard K, Stévant B. Towards a highly adaptable user-centric terminal architecture. The 7thInternational Symposium on Wireless Personal Multimedia Communications (WPMC 2004), Abano, Italie, 2004.

18. Sheng-mei L, Su P, Ming-hai X. An improved topsis vertical handoff algorithm for heterogeneousm wirelessnetworks. 12th IEEE International Conference on Communication Technology (ICCT), Nanjing, China, 2010;750–754.

19. Yang K, Gondal I, Qiu B, Dooley LS. Combined SINR based vertical handoff algorithm for next generation het-erogeneous wireless networks. Global Telecommunications Conference, GLOBECOM’07, Washington, USA, 2007;4483–4487.

20. Yang K, Gondal I, Qiu Bin. Multi-dimensional adaptive sinr based vertical handoff for heterogeneous wirelessnetworks. Communications Letters, IEEE 2008; 12(6):438–440.

21. Alkhawlani MM, Alsalem KA, Hussein AA. Multi-criteria vertical handover by topsis and fuzzy logic. InternationalConference on Communications and Information Technology (ICCIT), Aqaba, Jordan, 2011; 96–102.

22. Alkhawlani MM, Mohsen AM. Hybrid approach for radio network selection in heterogeneous wireless networks.International Journal of Advanced Science and Technology 2012; 44:33–48.

23. Vasu K, Maheshwari S, Mahapatra S, Kumar CS. QoS aware fuzzy rule based vertical handoff decision algorithmfor wireless heterogeneous networks. National Conference on Communications (NCC), Bangalore, India, 2011; 1–5.

24. Gowrishankar RBH, Sekhar GG, Satyanarayana P, et al. Performance evaluation of vertical handoff in wirelessdata networks. 4th International Conference on Wireless Communications, Networking and Mobile Computing,WICOM’08, Dalian, China, 2008; 1–4.

25. Rodriguez J, Tsagaropoulos M, Politis I, Kotsopoulos S, Dagiuklas T. A middleware architecture supporting seamlessand secure multimedia services across an intertechnology radio access network. Wireless Communications, IEEE2009; 16(5):24–31.

26. Wu Q, Li W, Wang R, Yu P. An access network selection mechanism for heterogeneous wireless environments.Journal of Computational Information Systems 2013; 9(5):1799–1807.

27. Wang H, Wang Z, Feng G, LV H, Chen X, Zhu Q. Intelligent access selection in cognitive networks: a fuzzy neuralnetwork approach. Journal of Computational Information Systems 2012; 8(21):8877–8884.



28. Sasirekha V, Ilanzkumaran M. Heterogeneous wireless network selection using FAHP integrated with topsis andvikor. International Conference on Pattern Recognition, Informatics and Medical Engineering (PRIME), Salem,India, 2013; 399–407.

29. Kaleem F, Mehbodniya A, Yen KK, Adachi F. Application of fuzzy topsis for weighting the system attributes inoverlay networks. 14th Asia-Pacific Network Operations and Management Symposium (APNOMS12), 2012; 1–6.

30. Lahby M, Leghris C, Adib A. A survey and comparison study on weighting algorithms for access network selection.9th Annual Conference on Wireless On-Demand Network Systems and Services (WONS), Courmayeur, Italy, 2012;35–38.

31. Zhang W. Handover decision using fuzzy MADM in heterogeneous networks. Wireless Communications andNetworking Conference, WCNC 2004, Atlanta, USA, Vol. 2, 2004; 653–658.

32. Saaty TL. Decision making with dependence and feedback: the analytic network process, 1996.33. Yüksel I, Dagdeviren M. Using the analytic network process (ANP) in a swot analysis—a case study for a textile

firm. Information Sciences 2007; 177(16):3364–3382.34. Saaty TL, Vargas LG. Diagnosis with dependent symptoms: Bayes theorem and the analytic hierarchy process.

Operations Research 1998; 46(4):491–502.35. Zadeh LA. Fuzzy sets. Information and Control 1965; 8(3):338–353.36. Sambuc R. Fonctions and Floues: Application a l’aide au Diagnostic en Pathologie Thyroidienne. Faculté de

Médecine de Marseille: Marseille, France, 1975.37. Liu P, Jin F. A multi-attribute group decision-making method based on weighted geometric aggregation operators of

interval-valued trapezoidal fuzzy numbers. Applied Mathematical Modelling 2012; 36(6):2498–2509.38. Cornelis C, Deschrijver G, Kerre EE. Advances and challenges in interval-valued fuzzy logic. Fuzzy Sets and Systems

2006; 157(5):622–627.39. Ashtiani B, Haghighirad F, Makui A, et al. Extension of fuzzy topsis method based on interval-valued fuzzy sets.

Applied Soft Computing 2009; 9(2):457–461.40. Wei S-H, Chen S-M. Fuzzy risk analysis based on interval-valued fuzzy numbers. Expert Systems with Applications

2009; 36(2):2285–2299.41. Hwang C-L, Yoon K. Multiple Attribute Decision Making. Springer-Verlag: Berlin, New York, 1981.


An analytic network process and trapezoidal interval-valued fuzzy technique for order preference by similarity to ideal solution network access selection method

Documents