Access Network Selection based on Fuzzy Logic and Genetic Algorithms

JOURNAL OF ADVANCES IN ARTIFICIAL INTELLIGENCE 1

Access Network Selection based on FuzzyLogic and Genetic Algorithms

Mohammed Alkhawlani and Aladdin AyeshFaculty of Computing Sciences and Engineering

De Montfort UniversityLeicester, LE1 9BH, UK

Email: [email protected], [email protected]

In the next generation of heterogeneous wireless networks (HWNs) a large number of different radio accesstechnologies (RATs) will be integrated into a common network. In this type of networks, selecting the mostoptimal and promising access network (AN) is an important consideration for overall networks stability,resource utilization, user satisfaction, and quality of service (QoS) provisioning. This paper proposes ageneral scheme to solve the access network selection (ANS) problem in the HWN. The proposed schemehas been used to present and design a general multi criteria software assistant (SA) that can consider theuser, operator, and/or the QoS view points. Combined fuzzy logic (FL) and genetic algorithms (GA) havebeen used to give the proposed scheme the required scalability, flexibility, and simplicity. The simulationresults show that the proposed scheme and SA have better and more robust performance over the randombased selection.

F

1 INTRODUCTION

The current radio resource management (RRM) [1], [2],[3] solutions and mechanisms for the wireless networksconsider only the case of a single RAT where mobileusers only can access that RAT and co-existing sub-networks can only be operated independently. CommonRRM (CRRM) [4], [5], [6], [7] and joint RRM (JRRM) [8],[9], [10], [11] are two types of the processes that en-able the management within a single and between thedifferent RATs of the HWN. CRRM is a cooperativeRRM architecture proposed by 3GPP [4], [5] to makeUMTS and GSM/GPRS networks cooperate. It is re-sponsible for coordinating the individual RRM entitiesof each RAT. JRRM is another cooperative RRM archi-tecture and mechanism proposed by IST SCOUT [10]and TRUST [11] projects and is similar to CRRM, butit is not restricted to GSM and UMTS and complementsthe CRRM with additional features and algorithms. BothJRRM and CRRM imply the use of some new RRMmechanisms such as access network selection (ANS),joint admission control (JAC), joint scheduling control(JSC), and vertical handover (VHO). Our work concen-trates on the first mechanism i.e., ANS, which decideshow to select the most suitable access network (AN)based on the discovered accesses, QoS constraints, op-erator policies, user preferences and available systemcapacity and utilization.

Not many algorithms to the ANS problem can beencountered in the literature so far. In [14] there are

different signal thresholds for each AN. Mobile nodecompares the RSS with the signal thresholds and decidesvertical handoff procedure. In [15] mobile nodes keepconnection with the higher throughput network untilRSS goes below a threshold. In [16] service types areconsidered and differentiating service is achieved bydifferent levels of RSS. A network selection algorithmthat is based on the distinct features of voice and datatraffic is proposed in [25]. In this algorithm the new callsprefer the cellular network for voice service and WLANsfor the data service. In general the current conventionalsingle criteria based algorithms such as the proposed[14], [15], [16] are rigid and take only one criterionand can not react easily to the changing environmentconditions and accumulated human knowledge. Usuallythey can not cope with the different view points andgoals of the operators, users, and QoS requirements,which make them inefficient for a multi criteria problemsuch as ANS problem.

Several policy based RAT selection solutions are pro-posed in [17], [18], [26]. General frameworks for a policy-based RAT selection with some specific examples areproposed by X. Gelabert et al. [17]. A network capacitysurfaces policy based admission controller is presentedby K. Murray et al. [18]. A policy-based handover mecha-nism for mobile multihomed hosts is presented in [26] byYlitalo et al. The handover decision is based on explicituser defined policies, and considers different criteriain the order of user defined priority. These solutionsdo not consider the trade-off between criteria of the


ANS problem and does not specify the importance andsensitivity of each criterion to the selection problem.

Several multi-criteria decision making based algo-rithms are proposed [12], [13], [19], [20], [24]. Theseconventional multi criteria based algorithms do not takeinto account the complexities and uncertainties thatarise from the different characteristics and natures ofthe different RATs. For these algorithms it is not easytask to incorporate the accumulated human knowledgeabout the problem and the only method to adapt thealgorithms is to change the criteria weights randomly toget better results. In addition these algorithms do nothave a proper method to address the importance of thedifferent criteria to the ANS.

Several multi criteria based algorithms with the aidof artificial intelligence tools such as fuzzy logic, neuralnetworks, and genetic algorithms [21], [22], [23] are suf-fering from scalability and modularity problems. Theycan not cope easily with the increased numbers of RATsand criteria in the HWN. These algorithms take all inputsfrom the different RATs at once to one fuzzy logic block,so they suffer from scalability and complexity problemswhen more RATs or membership functions added dueto the exponential increase on the number of inferencerules.

It is worth to mention that most of the above algo-rithms do not addressed clearly the roles of the users andthe operators in the selection process. Furthermore someof the current algorithms could not differentiate betweenthe initial selection process for the new service requestand the selection process during the VHO process.

The main contribution of this paper is the develop-ment of a new class of ANS algorithms that are based onhybrid parallel fuzzy logic (FL) based decision, multiplecriteria decision making (MCDM) systems, and geneticalgorithms (GA). This class of algorithms represents thefirst attempt to develop adaptive, flexible, and scalableANS algorithms that are utilize the advantages of hy-brid parallel FL decision making systems and MCDMsystems. Up to our knowledge no utilization for GA orparallel FL based decision abilities for ANS has beencarried out so far.

Our proposed scheme to solve the ANS problem isdescribed in section 2. Software assistant (SA) based onthe proposed scheme in co-existed CDMA and TDMAbased systems is presented in section 3. The simulationmodels and performance evaluation are carried out insection 4. Conclusions are presented in section 5.

2 ANS PROBLEM SOLVING SCHEME

More intelligent and robust solutions than the existedones are needed. In general any ANS solution has totake into its account the following general requirements.

• The solution has to solve the ANS problem in asimple way to give a reasonable and acceptabledelay before the decision appears.

• The solution has to cope with the different viewpoints and goals of the operators and the users andto give both parties the right and fair role in theselection process.

• The solution has to react to the changing environ-ment conditions and accumulated human knowl-edge about the problem.

• The solution has to allow any type of inputs andto be applicable to any type of ANs.

• The solution has to be able to handle the increasingnumber of RATs and it also has to be able to handlea large number of criteria.

The ANS procedure can be divided into three phases,initiation phase, decision phase, and execution phase.The objective of the initiation phase is to recognizethe need for ANS and subsequently initiate it and findout the required information and measurements for thedecision phase. In the decision phase a comparison ofthe information and the measurements calculated froma variety of sources including networks measurements,QoS requirements, user preferences and operator poli-cies. This comparison lead to the identification of thebest available AN according to the defined performanceevaluation metrics. In the execution phase, the user isassigned to the selected network in the decision phase.

A generic scheme to solve the ANS problem andany other similar selection problem is shown in Fig-ure 1. The scheme decision phase consists of threemain components, the first component contains a setof small parallel fuzzy logic (FL) based subsystems, thesecond component is a multiple criteria decision making(MCDM) system and the third component is a geneticalgorithm (GA) based component to assign a suitableweights for the criteria in the second component. Thescheme decision phase can be described in more detailsas follows.

• The heterogeneous wireless environment containsup to n ANs (RAT1, RAT2, .., RATn) and theframework has to select the most promising oneor to rank the RATs according to their suitability.

• The selection depends on multiple criteria up to i(c1, c2, .., ci). Different type of criteria can be mea-sured from different sources to cover the differentview points of the users, the operators, the appli-cations, and the network conditions. Each criterionis measured then passed to its FL based controlsubsystem in the first component.

• Every FL based subsystem gives an initial score foreach RAT that reflects the suitability of that RATaccording the FL subsystem criterion. The differentsets of scores (d1, d2, .., di) are sent to the MCDMin the second component.

• The GA component assigns a suitable weight (w1,w2,.., wi) for each initial decision according to theobjective function that is specified by the operatoraccording to the importance and sensitivities ofANS criteria to the different characteristics of a


wireless heterogeneous environment.• Using the initial scores coming from the first com-

ponent and the weights that are assigned manuallyor using the third component, the MCDM willselect the most promising AN or will rank theavailable RATs according to there suitability.

Our proposed algorithm utilizes the advantages ofparallel FL control and MCDM, and GA. The follow-ing subsections summarize the main advantages andstrengths of our proposed ANS.

2.1 The advantages of the application of FLto the ANS problemThe framework utilizes the advantages of the applicationof FL to the ANS problem that can be summarized asfollows.

• Because of the HWN complex nature, any proposedANS has to allow the solution of the ANS problemin a simple way and to avoid the complex ana-lytical and mathematical equations used to in theconventional solutions. FL is conceptually easy tounderstand and easy to apply. The mathematicalconcepts behind fuzzy reasoning are very simple.The simplicity of the FL helps in providing therequired simple ANS solution.

• The data, information, and measurements that haveto be taken into account in the ANS are in generalvery dissimilar, imprecise, contradictory, and com-ing from different sources. For example, some inputparameters such as the received signal strength, thecell loads, and the SNR from different RATs, arenot directly comparable. FL is tolerant of imprecisedata and it allows imprecise and contradictoryinputs. As a result of that, a FL based solution hasbeen thought to be a good candidate for reachingsuitable ANS decisions from such imprecise anddissimilar information.

• FLC reconciles the conflicting objectives of the op-erators, users and applications about the ANS.

• ANS solution has to be able to response to thechanging conditions of the HWN environments andthe accumulated experience of the operators andusers. FL based solution is easy to modify by tuningand adjusting the inference rules and membershipfunctions.

• The use of linguistic variables and inference rulesmake the FL based solution similar to the wayhumans think, which makes it more intelligent andpowerful than the conventional algorithm. This alsocan help in easy incorporation for the accumulatedoperator and user knowledge about the ANS prob-lem.

2.2 Advantages of using Parallel FLC to theANS problemApplying the parallel FLC rather than the traditionalFLC has some advantages when solving the ANS prob-

lem. The advantages of the application of parallel FL tothe ANS problem can be summarized as follows.

• The expected huge complexity involved in the het-erogeneous environment is greatly reduced usingthe idea of parallel FLC that reduce the number ofneeded inference rules and the complexity of eachrule. We have avoided the exponential increase ofthe rule base, and thus the increase in the com-plexity of the decision, when the number of inputvariables increases. To illustrate this point assumethat we have two RATs only, five inputs per RAT toone FL system and every input has 3 membershipfunctions, then we will have up to 310 inferencerules. In our scheme if we use five parallel FLCwith the same inputs and number of networks, wewill have 5*32 rules only.

• The idea of the parallel FLC reduces the complexityof the inference rules used in the fuzzy basedsolutions. To illustrate this point assume that wehave five inputs per RAT to one fuzzy controlsystem in HWN that have two RATs. The mostcomplex inference rule will be as follows IF C1/N1AND C1/N2 AND C1/N3 AND C1/N4 ANDC1/N5 AND C2/N1 AND C2/N2 AND C2/N3AND C2/N4 AND C2/N5 AND C3/N1 ANDC3/N2 AND C3/N3 AND C3/N4 AND C3/N5AND C4/N1 AND C4/N2 AND C4/N3 ANDC4/N4 AND C4/N5 AND C5/N1 AND C5/N2AND C5/N3 AND C5/N4 AND C5/N5 THEN D1AND D2 AND D3 AND D4 AND D5, where Ci/Njmeans network j rating value for criteria i and Djmeans initial score for network j. However usingour parallel FLC, the most complex inference rulewill be as follows IF C1/N1 AND C1/N2 ANDC1/N3 AND C1/N4 AND C1/N5 THEN D1 ANDD2 AND D3 AND D4 AND D5.

• The output of the parallel FLC are in the range [0, 1]provide already normalize inputs (i.e., RATs ratingagainst criteria) for the MCDM. This provides greatadvantage over the current MCDM based solutionthat need to normalize their input data with someinappropriate methods.

2.3 The advantages of the application ofMCDM to the ANS problemThe framework utilizes the advantages of the applicationof MCDM to the ANS problem that can be summarizedas follows.

• In a very complex and uncertain environmentssuch as HWN, MCDM can sufficiently reduce theuncertainty and doubt about the alternatives andallows a reasonable choice to be made from amongthem.

• ANS problem is a multi-criteria problem in natureand the flexibility and complementary ANS multi-criteria have to be utilized to provide a solution that


FIGURE 1. The proposed solving scheme for access network selection problem based on FL, MCDM, and GA

can cope with the different view points and goals.If suitable weights have been assigned for each cri-terion using proper optimization methods such asGA, the user satisfaction average and the operatorbenefit can be maximized at the same time. Thismakes the proposed solution more efficient than thetraditional algorithms that are either user-centric oroperator-centric.

2.4 The advantages of the application of GAto the ANS problemIn our study the GA is used to overcome one of theMCDM drawbacks, which is the finding of suitableweights for the different criteria. The following pointssummarize the main advantages of applying the GA tothe ANS problem.

• GA can deal with the large number of variablesand the complex search space included on the ANScriteria weights with high probability of successin finding optimal, near optimal or at least goodsolution.

• GA can handle the different constraints and objec-tives of the ANS criteria weights.

• GA doesnt require derivative information, whichcan help out in the user-centric algorithm wherenot enough data is existed due to the limitation ofstorage in the mobile terminal.

• GA is less likely to be trapped by local optimalminima or maxima.

• GA works with numerically generated data, ex-perimental data, or analytical functions. This cangive different options when designing the ANSalgorithm.

3 ANS SA FOR CO-EXISTED CDMA AND TDMABASED SYSTEMS

User software assistant (USA) based on our schemewith suitable input criteria can be used in the userequipment. Operator software assistant (OSA) based onthis scheme with suitable input criteria can be used inthe operator CRRM or JRRM entity. Based on the schemeproposed on the last section, ANS software assistants forusers and operators can be developed. Since both OSAand USA are based on the same scheme and the maindifference will be in the used input criteria, a general SAis presented on this section. The proposed SA deals witha wireless environment that contains a CDMA-basedand a TDMA-based wireless networks. The details ofthe general SA is shown in Figure 2. The componentsof the SA can be summarized in the following threesubsections.

3.1 The FL based control component

Our SA contains four FL based subsystems. Each sub-system considers one of the ANS important criteria. TheRSS subsystem considers the received signal strengthcriteria. The MSS subsystem considers the mobile station


FIGURE 2. ANS SA for co-existed CDMA and TDMA systems

speed criteria. The ST subsystem considers the servicetype criteria. The UP subsystem considers the user pref-erence and price criteria. RSS subsystem has two inputvariables RSS1 to describe the received signal strengthfrom CDMA based network and RSS2 to describe thereceived signal strength from TDMA based network.MSS subsystem has only one input variable MSS todescribe the mobile station speed. ST subsystem has twoinput variables, the first is ”DelayReqc” to describe theone-way delay needed for the required service and thesecond is ”RateReqc” to describe the bit rate needed forthe required service. UP subsystem has only one inputvariable ”Price” to describe the user preferred price.Every input variable has three membership functions{Low, Medium, High}. Figure 3 shows the membershipfunctions of the input variables.

Every subsystem has two output variables, one vari-able to describe the probability of acceptance for the newuser in the CDMA network and the other variable todescribe the probability of acceptance for the new userin the TDMA network. Each output variable has fourmembership functions {TR (Totally Reject), PR (Proba-bility Reject), PA (Probability Accept), and TA (TotallyAccept)}. The subsystems output variables are ”RSSc1”and ”RSSc2” for RSS subsystem, ”MSSc1” and ”MSSc2”for MSS subsystem, ”STc1” and ”STc2” for ST subsystem,and ”UPc1” and ”UPc2” for UP subsystem. Figure 4shows ”STc1” variable with its membership functions asa sample for the output variables.

3.2 The MCDM component

Enhanced version of simple multi attribute rating tech-nique (SMART) has been used. SMART is one of thesimplest MCDM and most efficient methods. The rank-ing value xj of alternative Aj is obtained simply as theweighted algebraic mean of the utility values associatedwith it i.e., aij according to equation 1.

xj =∑m

i=1 wiaij∑mi=1 wi

j = 1, 2, ....n (1)

FIGURE 3. Membership functions of the input variables

SMART employs relatively uncomplicated andstraightforward manipulation method, which makes itstronger and easier to use in a hybrid and more complexmodels such as the proposed one in this paper. With theaid of both FL and GA, SMART has all the capabilitiesrequired to address the specific considerations thatare involved in the ANS decision making process.SMART can be quickly and easily understood by theinexperienced decision makers. However, to get betterunderstanding and deeper insights into the selectiondecision making, other advanced MCDM methods canbe used in the future work and compared with SMARTmethod.

In our algorithm there are two alternatives for theMCDM, one is a CDMA based network and the other is a


FIGURE 4. Membership functions for ”STc1” output variable

TDMA based network. The input criteria of the MCDMare the outputs of the FL based control subsystems inthe first component. The GA component assigns theweight wi for criteria i to reflect its relative importance.The criteria with more importance to the operator anduser can be assigned higher weight using the objectivefunction of the GA specified by the operator. Since all theoutputs of FL subsystems are in the range [0, 1], there isnot be any need to scale the criteria performance againstalternatives.

The ranking value of CDMA network xc and theranking value of TDMA network xt can be calculatedas follows.

xc =MSSc1 ∗Wv + RSSc1 ∗Ws + STc1 ∗Wt + UPc1 ∗Wu

TW(2)

xt =MSSc2 ∗Wv + RSSc2 ∗Ws + STc2 ∗Wt + UPc2 ∗Wu

TW(3)

where Wv is the assigned weight for the mobile stationspeed criteria. Ws is the assigned weight for the receivedsignal strength criteria. Wt is the assigned weight for theservice type criteria. Wu is the assigned weight for theuser preferred price criteria. TW is the total weight andis calculated using 4.

TW = Wv + Ws + Wt + Wu (4)

3.3 The GA componentThe GA has been used in our scheme to help theusers or network operator to find suitable values for theweights {Wv , Ws, Wt, Wu} in the offline mode. Besidethe GA advantages that are summarized on subsection2.4, our decision to use GA was based on the natureof our objective functions that have several dynamicand stochastic components, where any other derivative-based optimization method can not perform well. An-other important issue that encouraged the selection ofGA to our problem is the high interaction betweendifferent variables. At the end usually the GA is not

used to find an optimal solution and most of the timeit is considered as a very promising option to find goodand acceptable solution. It is worthy to mention that GAis not used on every selection decision. It is optionalcomponent on the scheme where the user can use it ornot in offline mode prior to the selection process.

Since the ANS is a multi criteria problem in nature,different objectives need to be optimized. In this paper,three different objective functions are proposed to coverthe different and opposite objectives and requirementsof the users, QoS and operators. The proposed objectivefunctions are outlined in the following points.

• The first objective is to maximize the percentage ofsatisfied users i.e., users who are assigned to a net-work of their preference (Pu): since the algorithmconsiders multiple criteria, it may assign a user to anetwork different from his/her original preference.Pu is considered appropriate for performance eval-uation from the user point of view. To maximize Pu

the objective function ObjFun1 shown in Figure 5is used.

FIGURE 5. Pesudeo-code for objective function ObjFun1

• The second objective is to maximize the percentageof users assigned to the networks with strongersignal strength (Pq): since the algorithm considersmultiple criteria, it may assign a user to a networkwith weaker signal strength. Pq is considered as asimple indicator for performance evaluation fromthe QoS point of view. To maximize Pq the objectivefunction ObjFun2 shown in Figure 6 is used.

• The third objective is to achieve the load balancingbetween networks: Po is the percentage between thenumber of users in both networks. It is used as abasic indicator for load balancing. If the resourcesof both networks has the same importance andcost for network operator, this objective is consid-ered appropriate for performance evaluation fromnetwork operator point of view. To achieve loadbalancing the objective function ObjFun3 shown inFigure 7 is used.

The weights of the input criteria {Wv , Ws, Wt, Wu}have been encoded using real encoding method. The



length of the real-valued encoding is 4-real floating num-bers. The length of each floating number is dependingon the internal precision and round-off used by thecomputer to define the precision of the floating numbers.

The results achieved by extensive comparisons of GAperformance as a function of the different GA parameters(i.e., the population size, the mutation rate and thecrossover fraction) and the different GA operators (i.e.,the selection and crossover operators) have been sum-marized by R. Haupt and S. Haupt in [27]. They haveconcluded that crossover fraction, selection operator, andcrossover operator are not of much importance. On theother side the population size and the mutation rate havethe most significant impact on the ability of the GA tofind better minimum value for objective function.

Based on that, several sets of experiments to deter-mine good population size and mutation rate for ourproposed objective functions have been carried out. Eachexperiment has been repeated several iterations whilethe performance is recorded. The results from all theiterations are then combined by calculating the average(and standard deviation) for each experiment. Usuallytesting GAs includes mainly two issues; how far theresult obtained by GA is from the benchmark results,


which can be measured by average fitness value. Thesecond issue is how fast GA is in finding the bestsolution, which can be measured by the total numberof function evaluations. Since our GA is working offline,only the average best fitness value achieved by the GA isused. The average best fitness and the standard deviationare plotted against the tested parameters. When studyinga parameter all the other parameters and operators arekept constant.

Only a sample for the experiments results achieved isgiven in this paper as shown on Figure 8. The lower plotin Figure 8 shows the means and standard deviations ofthe ObjFun1 function best fitness values over 20 runs, foreach of the values of the population sizes between 10 and100 with step size equals to 10 individuals. The upperplot shows a color-coded display of the best fitness val-ues in each run where dark color indicates better results.For ObjFun1 fitness function, setting the population sizeto 20 individuals or more yields better result. Based onthese results and the other experiments results for otherobjective functions, the population size chosen in ourGAs has been set to 20. Other sets of experiments havebeen carried out to determine a suitable mutation rate.Based on the experiments, the mutation rate has beenset to 0.1.

In addition our GA uses a roulette selection function.To ensure elitism, the number of individuals that areguaranteed to survive to the next generation has beenset to 4 individuals. It uses two point crossover functionand a uniform mutation function. The GA terminates ifthe maximum number of the GA iterations reaches to 300iterations, if there is no improvement in the best fitnessvalue for number of consecutive generations equals to100 generation, or if there is no improvement in thebest fitness value for an interval of time equals to 300seconds.

4 THE PERFORMANCE EVALUATIONThe used simulation models and some simulation resultsare presented in this section.

FIGURE 8. The effect of population size on objective functionObjFun1


4.1 The simulation models

Our proposed scheme and SA are evaluated using thesimulation approach. MATLAB mathematical softwareand a set of functions called RUNE [1] have been usedfor the simulation. The system model considers thecoexistence of a CDMA based WWAN network withseven macro cells with omni directional antenna and cellradius = 1000m and a TDMA based WLAN networkwith twelve micro cells with omni directional antennaand cell radius = 500m. In the system environment, eachmobile has a velocity and is moved with a randomdistance and a random direction at defined time steps.The velocity is a vector quantity with magnitude anddirection. The velocity of the ith mobile is updatedaccording to 5.

Ui = Ui ∗ P +√

1− P 2 ∗ Vm ∗X (5)

where Ui is the complex speed [m/s]. P is the correlationof the velocity between time steps. X is Rayleigh dis-tributed magnitude with mean 1 and a random direction.Vm is the mean speed of mobiles. Figure 9 shows theusers, in blue, over the simulated environment. Fourtypes of services are considered in our simulation, thevoice calls, the low bit rate real time video telephony,the high bit rate streaming video, and the non-real timedata traffic. The traffic is modeled according to Poissonprocess. The main call holding time is assumed to be 50seconds.

The propagation model used in this paper can bedescribed in logarithmic scale as in equation 6.

G = GD + GF + GR + GA (6)

The equation 6 contains four components. The first com-ponent is the distance attenuation GD, which is given bythe Okumura-Hata formula 7

GD = −{β + 10α ∗ log(d)} (7)

FIGURE 9. The users over the simulated environment

d is the distance to the transmitter, β is a constant usedto model the effect of carrier frequency, the antenna sizeand other physical parameters. The parameter α is thedistance attenuation coefficient. β has been set to -28 dBand α has been set to 3.5. The second component is theshadow fading GF , which is modeled as a log-normaldistribution with standard deviation of 6 dB and 0 dBmean. The third component is the Rayleigh fading GR,which is a Rayleigh distributed. The forth component isthe antenna gain GA.

4.2 The simulation results

This subsection shows some simulation results and com-pares the performance of our proposed solution to threedifferent reference ANS algorithms. The first algorithm isa service-type based selection algorithm where high bitservices with low propagation delay requirements aresent to the WLAN and the low bit rate services withthe high propagation delay requirements are sent to theWWAN. The second algorithm is a terminal-speed basedselection algorithm where high speed users are sent tothe large-coverage network (i.e., WWAN) and the lowspeed users are sent to the small-coverage network (i.e.,WLAN). The third algorithm is a random based selectionalgorithm where the users are assigned randomly tothe two networks. All solutions have been simulated,evaluated and compared for the three different objectivesaddressed in subsection 3.3. Several runs of simulationhave been carried out. Each run uses a different numberof users.

Figure 10 shows Pu values in all solutions. The hori-zontal axis shows the number of users while the verticalaxis shows the Pu values. From both Figure 10 andthe numerical samples for Pu values shown in Table 1,the great improvement in the number of the satisfiedusers in our solution can be seen. For example with840 users in the environment, the percentage of satisfiedusers with all different reference solutions is around 50%while the same percentage in our proposed solution isaround 80.7%. In general our proposed solution achievesaround 31% enhancement over the different reference al-gorithms. In Figure 11, the lower plot shows the weightsassigned by the GA for the different criteria using thefirst objective function. The upper plot shows the bestfunction value in each generation versus iteration num-ber.

Figure 12 shows Pq values in all solutions. The hori-zontal axis shows the number of users while the verticalaxis shows the Pq values. From both Figure 12 and thenumerical samples for Pq values shown in the Table 2,the great improvement in the percentage of the usersassigned to networks with stronger received signal canbe seen. For example with 840 users in the environmentPq with the different reference algorithms is around 50%while the same percentage in our proposed solution isaround 64%. In average our proposed solution achievesaround 26.5% enhancement over the random based se-


FIGURE 10. The percentage of satisfied users (Pu) in the proposed algorithm against the three references algorithms

TABLE 1Pu values in all solutions

Users The proposed Random based Service type based Terminal-Speed basedNo. solution Pu selection Pu selection Pu Selection Pu

125 0.744 0.560 0.600 0.520223 0.839 0.520 0.507 0.489318 0.833 0.522 0.487 0.481431 0.789 0.510 0.527 0.501539 0.785 0.484 0.501 0.486840 0.807 0.502 0.507 0.4981070 0.798 0.467 0.517 0.4711269 0.790 0.519 0.502 0.5011393 0.825 0.503 0.490 0.4931634 0.860 0.488 0.511 0.5231735 0.867 0.493 0.485 0.502

FIGURE 11. The weights assigned by GA for the 1st objective

lection. Figure 13 shows the weights assigned by the GAfor the second objective.

The results of Po produced by all solutions for thethird objective are shown in Table 3. Our proposedsolution achieves good and comparable results againstthe results achieved by other solutions. Figure 14 showsthe weights assigned by the GA for the third objective.

5 CONCLUSIONS

A general scheme to solve the multi criteria ANS prob-lem has been presented in this paper. The proposedscheme is scalable and is able to handle any numberof RATs with a large set of criteria. The scheme cancope with the different and contrast view points andgoals of the operator and users and it can react to the


FIGURE 12. The percentage Pq values in the proposed algorithm against the three references algorithms

TABLE 2Pq values in all solutions

Users The proposed Random based Service type based Terminal-Speed basedNo. solution Pq selection Pq selection Pq Selection Pq

125 0.720 0.456 0.504 0.440223 0.798 0.565 0.484 0.520318 0.830 0.503 0.503 0.472431 0.831 0.506 0.538 0.490539 0.844 0.486 0.542 0.464840 0.640 0.510 0.507 0.5051070 0.750 0.473 0.494 0.4561269 0.797 0.502 0.490 0.4531393 0.819 0.497 0.498 0.4731634 0.598 0.505 0.492 0.4741735 0.622 0.519 0.482 0.476

TABLE 3Po values in all solutions

Users The proposed Random based Service type based Terminal-Speed basedNo. solution Po selection Po selection Po Selection Po

125 1.341 1.155 0.984 0.944223 1.159 1.139 0.848 0.741318 1.196 0.989 0.935 0.866431 1.078 0.927 1.025 0.882539 1.355 0.968 0.900 0.855840 1.169 1.053 1.175 0.9391070 1.232 1.007 0.856 0.8821269 1.178 1.130 1.046 0.8561393 1.063 1.048 0.911 0.8701634 1.212 0.937 1.030 0.8881735 1.171 0.924 1.175 0.877


FIGURE 13. The weights assigned by GA for the 2nd objective

FIGURE 14. The weights assigned by GA for the 3rd objective

accumulated human knowledge about the problem. Theproposed scheme has been used to present and designa multi criteria ANS solution that considered the user,the QoS, and the operator view points. The simulationresults show that the proposed solution has a better andmore robust performance over the reference solutions.

Different aspects of our work have to be developed.More criteria can be included such as resources avail-ability and Signal-to-Noise Ratio (SNR). Multi-objectiveoptimization method can be used to find an optimumweights jointly rather than using the single objective GAthat optimize each objective weight independently. Theused rules and membership functions of the fuzzy sub-systems are still subjective, so they can be built or tunedusing a suitable learning method such as the geneticalgorithms or the neural networks. Our algorithm canbe compared with more reference algorithms.

REFERENCES

[1] J. Zander and S. Kim, Radio Resource Management forWireless Networks, Artech House, 2001.

[2] J. Perez-Romero, O. Sallent, R. Agusti, and M. Angel,Radio Resource Management Strategies in UMTS, John Wileyand Sons, 2005.

[3] M. Cardei, I. Cardei, and D. Du, Resource Management inWireless Networking, Springer Science and Business Media,2005.

[4] 3GPP TR 25.881 v5.0.0, Improvement of RRM across RNSand RNS/BSS (Release 5).

[5] 3GPP TR 25.891 v 0.3.0, Improvement of RRM across RNSand RNS/BSS (Post Rel-5).

[6] A. Tlli, P. Hakalin, H. Holma, Performance evaluation ofcommon radio resource management (CRRM), Proceedingsof the IEEE ICC 2002), vol. 5, pp. 3429-3433, April 2002.

[7] IST Evolutionary Strategies for Radio Resource Manage-ment in Cellular Heterogeneous Networks (EVEREST)Project, www.everest-ist.upc.es.

[8] IST End to End Reconfigurability (E2R) Project,http://e2r.motlabs.com.

[9] IST Wireless World Initiative New Radio (WINNER)Project, www.ist-winner.org.

[10] IST Smart user-centric communications environment(SCOUT)Project, www.ist-scout.org.

[11] IST Transparently Reconfigurable Ubiquitous Terminal(TRUST) Project, www.ist-trust.org.

[12] F. Bari and V. Leung, ”Application of ELECTRE to net-work selection in a heterogeneous wireless network envi-ronment,”, Proceedings of the IEEE Wireless Communicationsand Networking Conference (WCNC 2007), Page(s):3810 -3815, March 2007.

[13] F. Bari and V. Leung, ”Multi-Attribute network selectionby iterative TOPSIS for heterogeneous wireless access,”,Proceedings of the IEEE Wireless Communications and Net-working Conference (WCNC 2007), pp.3810 - 3815, March2007.

[14] Y. Gwon, D. Funato, and A. Takeshita, ”Adaptive ap-proach for locally optimized IP handoffs across hetero-geneous wireless networks,” 4th International Workshop onMobile and Wireless Communications Netwoor heterogeneouswireless access, pp. 475-479, Sept. 2002.

[15] M. Ylianttila, M. Pande, J. Makela, and P. Mahonen,”Optimization scheme for mobile users performing ver-tical handoffs between IEEE 802.11 and GPRS/EDGEnetworks,” Proceedings of IEEE Global TelecommunicationConference (Globecom 2001), pp.25-29, Nov. 2001.

[16] H. Park, S. Yoon, T. Kim, J. Park, M. Do, and J. Lee,”Vertical Handoff Procedure and Algorithm betweenIEEE802.11 WLAN and CDMA Cellular Network,” 7thCDMA International Conference 2003, pp. 103-112, Seoul,Korea, Nov. 2003.

[17] X. Gelabert, J. Prez-Romero, O. Sallent, R. Agusti, andF. Casadevall, “Radio resource management in hetero-geneous networks,” Proceedings of the 3rd InternationalWorking Conference on Performance Modelling and Evaluationof Heterogeneous Networks, UK, 2005.

[18] K. Murray and D. Pesch, “Policy based access manage-ment and handover control in heterogeneous wireless net-works,” Proceedings of the Vehicular Technology Conference,2004, vol.5, pp.3319- 3323, Sep. 2004.

[19] T. Al-Gizawi, K. Peppas, D. Axiotis, E. Protonotarios, andF. Lazarakis, “Interoperability criteria, mechanisms, andevaluation of system performance for transparently in-teroperating WLAN and UMTS-HSDPA networks,” IEEENetwork, vol.19, pp.66-72, July-Aug. 2005.

[20] R. Pintenet, J. Gosteau, T. Al-Gizawi, F. Lazarakis, K.Peppas, A. Alexiou, “Evaluation of interoperability mech-anisms for coexisting HSDPA and WLAN enhanced withMTMR techniques,” Proceedings of the Vehicular TechnologyConference 2003, vol.3, pp.1807- 1811, Oct. 2003.


[21] L. Giupponi, R. Agusti, J. Perez-Romero, and O. Sallent,“A novel joint radio resource management approach withreinforcement learning,” Proceedings of the Performance,Computing, and Communications Conference (IPCCC 2005),pp.621- 626, April 2005.

[22] R. Agusti, O. Sallent, J. Prez-Romero, and L. Giupponi, “Afuzzy neural based approach for joint radio resource man-agement in a B3G framework,” Proceedings of the 1st Inter-national. Conference on QoS in Heterogeneous Wired/WirelessNetworks, pp.216 - 224, USA, Oct. 2004.

[23] A. Wilson, A. Lenaghan, and R. Malyan, “Optimizingwireless access network selection to maintain QoS inheterogeneous wireless environments,” Proceedings of theWireless Personal Multimedia Communications 2005, Den-mark, Sep. 2005.

[24] L. Liang, H. Wang, P. Zhan, ”Net utility-based networkselection scheme in CDMA cellular/WLAN integratednetworks,”, Proceedings of the IEEE Wireless Communica-tions and Networking Conference (WCNC 2007), pp. 3313 -3317, March 2007.

[25] W. Song, H Jiang, W. H. Zhuang et al., ”Resource manage-ment for QoS support in cellular/WLAN interworking,”IEEE Network, vol. 19, no. 5, pp. 1218, Oct. 2005.

[26] J. Ylitalo, T. Jokikyyny, T. Kauppinen, A. Tuominen, J.Laine, ”Dynamic network interface selection in mobilehosts,” Proceedings of the 36th Annual Hawaii InternationalConference on System Sciences (HICSS 03), pp. 315-324,Hilton Waikaloa Village, USA, Jan. 2003.

[27] R. Haupt and S. Haupt, Practical Genetic Algorithms,John Wiley and Sons, 2nd Edition, 2004.

Access Network Selection based on Fuzzy Logic and Genetic Algorithms

Documents

Transcript of Access Network Selection based on Fuzzy Logic and Genetic Algorithms