Cognitive Network Accessusing Fuzzy Decision Making

Nicola Baldo,Student Member, IEEE, and Michele Zorzi,Fellow, IEEE

Abstract— We consider a scenario in which wireless userswant to connect to the Internet using one of several availablenetwork access opportunities, possibly using different radiotechnologies. We propose a distributed cognitive network accessscheme with the aim of providing the best quality of service withrespect to both radio link and core network performance anduser application requirements. Knowledge of the service qualityexperienced by active connections is shared, and prospective usersuse Fuzzy Logic techniques to process cross-layer communicationquality metrics and to estimate the expected transport-layer per-formance. These estimates are compared to the Quality of Servicerequirements of the application using Fuzzy Decision Makingtechniques to choose the most suitable access opportunity.Thisscheme naturally fits into the recently proposed Cognitive Net-work paradigm in that it defines a cognition process leveragingon end-to-end and cross-layer performance evaluation techniquesas well as information sharing among users; moreover, it offersa significant amount of flexibility and extensibility, thanks to itsmodularity and its independence from the particular technologyand application being used. The proposed scheme is shown tooutperform state-of-the-art solutions in several multi-technologyand multi-application scenarios, while at the same time achievingsimilar performance to application-specific omniscient schemesthat we introduce in this paper as a benchmark.

I. I NTRODUCTION

The recent progress in radio communications today providesseveral network access technologies for wireless connectivity,e.g., IEEE 802.11, WiMAX and UMTS. At the same time,advances in microelectronics allow all these technologiestobe exploited within a single mobile device equipped withmultiple radio interfaces. As a consequence of these facts,many new challenges have arisen for the telecommunicationresearch community. The one we focus on in this paper iswireless network access, i.e., how a user who wants to connectto the Internet can select, among all the available opportunities,the one which yields the best performance.

Most previous work dealing with this problem considersonly a specific wireless technology. The most notable exampleis 802.11: several valuable solutions have been proposed tosolve the problem of Access Point selection [1]–[3], but allofthem rely on 802.11-specific metrics. The consequence is thatthese solutions cannot be used with other wireless technolo-gies and, most importantly, the obtained performance metrics

N. Baldo is with the Centre Tecnologic de Telecomunicacions de Catalunya,Parc Mediterrani de la Tecnologia, Av. Canal Olimpic, 08860Castelldefels,Barcelona, Spain, email: nbaldo@cttc.es. M. Zorzi is with the Dept. ofInformation Engineering, University of Padova, via Gradenigo 6/B, 35131Padova, Italy, email: zorzi@dei.unipd.it. This work was done when N. Baldowas with the Dept. of Information Engineering, University of Padova.

This work was partially supported by the European Commission under theARAGORN project (INFSO-ICT-216856)

Part of this work has been presented at theIEEE CogNet 2007 Workshop(held in conjunction withIEEE ICC 2007), Glasgow UK, June 24, 2007.

cannot be compared to the ones obtained by different solutionsdesigned for other types of radio interfaces. Moreover, whileit is possible to design network access schemes to handle aparticular set of existing technologies, an ideal network accesssolution would be required to be generic and modular enoughso as to accommodate even new wireless technologies as theyare introduced.

Another issue in wireless network access is that an idealsolution should be optimal with respect to the end-user per-spective. In particular, two key aspects should be considered.First, user-perceived service quality depends on end-to-endperformance, in which the wireless link plays an important rolewithout however being the only issue. Previous work on accesspoint selection targeted only scenarios in which the wirelesslink is the bottleneck, while in many real life scenarios also thecore network can exhibit non-ideal performance, thus havinga non-negligible effect on end-to-end network performance.Second, the optimization of the end-to-end network perfor-mance should take the Quality of Service (QoS) requirementsof different applications into account. Much previous workon Access Point Selection focuses on the maximization ofnetwork throughput only; however, applications such as VoIPand gaming are more strongly affected by other factors suchas network delay and reliability. In recent years, interestinproviding satisfactory service quality for these applicationshas grown considerably, due to their increased popularity;this in turn has promoted a significant effort by the researchcommunity to find methods and solutions able to enhancethe QoS of multimedia applications on various wireless tech-nologies, most notably 802.11 [4]–[6] and UMTS [7]–[9].Unfortunately, this QoS-related research has focused almostexclusively on optimizing the performance of an alreadyestablished multimedia communications, and the problem ofperforming a QoS-aware network access decision has not beendealt with so far.

To summarize, the problem of identifying an algorithm forwireless network access selection which 1) is independent ofthe radio technology, 2) can accommodate the QoS require-ments of different applications, and 3) can account not onlyfor wireless link performance, but also for possible non-idealconditions of the entire end-to-end path, is a very challengingand still open issue.

From this perspective, the recently proposed CognitiveNetwork paradigm seems to be a very promising approach.In [10], a Cognitive Network is defined as “a network with acognitive process that can perceive current network conditions,and then plan, decide, and act on those conditions”. Cogni-tive Networking implies the presence of a cognition processwhich spans both all the layers of the protocol stack and all

2

the network components of end-to-end communication; thiscognition process is distributed among the different nodescomposing the network, which share their knowledge andcooperate among themselves. This architecture is promisingin that it is potentially able to solve problems which aretoo complex to be handled within a traditional cross-layerapproach, while at the same time being capable of learning thebehavior of different wireless technologies and applications.

In this paper, we propose a Cognitive Network approach tothe Wireless Network Access problem. In particular, we pro-pose a knowledge representation framework based on FuzzyLogic which enables the implementation of a cognition processwhich is both cross-layer and network-aware; furthermore,weexploit knowledge sharing among different devices with thepurpose of achieving a more complete and reliable characteri-zation of the performance of the whole network. Subsequently,we define a network access scheme based on Fuzzy DecisionMaking which allows each user to choose the Access Pointwhich best satisfies its QoS requirements. We stress thatthe modularity of the architecture and the generality of thechosen knowledge representation allow our solution to easilyprovide optimized network access policies for new wirelesstechnologies and applications which outperform state-of-the-art schemes while at the same time minimizing the effortrequired to accommodate different wireless technologies andapplications.

II. COGNITIVE NETWORK ACCESS

It is worth mentioning that the original definition of Cog-nitive Radio by J. Mitola [11] already tries to address theproblem of wireless network access. Mitola’s Cognitive Radiohas the explicit purpose of “detecting user needs, and provid-ing wireless services most adequate to meet them”. The useof the termcognitivehighlights the fact that some ArtificialIntelligence (AI) needs to be at the heart of the device whichisto choose and adapt its services to the user’s needs. One of themost quoted definitions for AI is “how to make machines dothings at which, at the moment, humans are better” [12]. So, inthe context of wireless network access, the goal of a cognitiveradio would be to relieve the mobile users from having tofigure out themselves which is the most satisfactory accessopportunity. The key motivation behind the approach proposedin this paper directly follows from these considerations: ourgoal is to replace user’s decisions trying to mimic the actualdecision strategies that a typical human user would adopt.

As an example, consider a cafe with a nearby 802.11hotspot, in an area covered by a UMTS provider. A singleuser comes in with his laptop, orders a coffee and looks foran Internet connection to do some web surfing. He might tryUMTS first, maybe just to realize that it is too slow for whathe pays for it. Then he might realize that there is an 802.11hotspot nearby, connect to it and finally surf the web happilywith high throughput. A second user comes in, orders anothercoffee, and turns on his mobile device to watch a TV programvia video streaming from the Internet. Instead of figuring outthe quality of the available networks by himself, he might justnotice the first user who is surfing happily, and query him

about the best connection available. Based on this knowledgeof the previous experience of the first customer, he wouldlikely connect to the hotspot, and be satisfied by enjoyingvideo streaming with high throughput connectivity. Then athird user comes in and, under advice from the previous users,connects to the hotspot for a VoIP conversation. It could hap-pen, however, that she is not satisfied because, e.g., the wirednetwork serving the hotspot has a considerably high delay,or the hotspot itself has become congested; consequently, shemight try out UMTS, and possibly be more satisfied because ofthe lower Round Trip Time. Finally, a last user might come inwith his laptop to play some highly interactive online game.He asks all previous customers about the performance theyexperienced, and he decides to use the UMTS connection,since he understands that his application requirements (lowlatency) are much closer to the requirements of the VoIP userrather than to those of the other two.

Although very simple, this example highlights some im-portant facts. First of all, while direct experience is an ef-fective means of inferring the quality of several connectionopportunities, exploiting knowledge previously gatheredbyother users may be a quicker and easier way. Second, thechoice of a suitable knowledge representation base for per-formance evaluation metrics becomes crucial: on one handit is impossible to find a unique definition of quality, sincedifferent applications can have even conflicting performancerequirements, and so we need different metrics; on the otherhand, we cannot rely on too fine-grained and technology-specific metrics, since this could possibly make informationgathered by some users difficult to interpret by others havingsimilar but not identical needs. Furthermore, we point out thatthe performance seen by each user is actually made up of twocomponents, i.e., radio link and core network performance,andthat these aspects should be evaluated jointly. In the exampleabove, suppose the first user leaves the cafe and goes andsits on a bench in a nearby park. Previous experience stillindicates the hotspot as the most suitable choice. However,current link quality metrics (e.g., the RSSI indicator) mightshow that he is too far from the hotspot to get an acceptablethroughput. Finally, it should be noted that all the availableinformation should be interpretedcum grano salis, i.e., notonly quantitatively but also qualitatively, because performancereports by other users might be biased, measurements mightbe affected by errors, and all this information might be partialand/or old.1

Based on the above discussion, we consider here a schemein which prospective users have access to a shared knowledgebase that contains information about the service quality expe-rienced by past and present active connections. To overcomethe above mentioned issues, we define a generic knowledgerepresentation framework using Fuzzy Numbers with the aimof enabling a generic representation of the most relevantperformance metrics of different applications. We proposetheuse of this knowledge representation to build a CognitiveNetwork Knowledge Database, which is to be filled with

1An important aspect ignored in this paper is how to deal with users thatmaliciously provide wrong information to influence other nodes’ decisions.This issue is left for future research.

3

service quality information fed back by all users actively usingthe network. A user willing to set up a new connection can thenretrieve such information and compare it to both applicationrequirements and other measurements, in order to assess theexpected service quality for each access opportunity; in thisprocess, both Fuzzy Logic Inference and Fuzzy Arithmetic areused. Finally, the most suitable network access opportunitymust be selected using Fuzzy Decision Making techniques.The presence of a knowledge representation base includinginformation belonging to different components of the commu-nication system, the cross-layer processing of this information,the mechanism by which a node learns from its neighbors’experience as well as its own history, the use of Fuzzy Logicfor incomplete knowledge representation, and the use of FuzzyDecision Making to select among access opportunities clearlyidentify our proposed scheme as belonging to the CognitiveNetwork Paradigm.

We note that Fuzzy Logic has been already proposed foruse in the context of telecommunication system, e.g., forQoS routing in wired networks [13], route caching decisionsin wireless ad hoc networks [14], radio resource manage-ment [15] and channel selection in cellular networks [16]. Aninteresting survey on the usage of Fuzzy Logic techniques inthe telecommunication field can be found in [17]. The maindifference of our approach is that, unlike previous work, wedo not consider a Fuzzy Controller implementing a simpleinput-output relationship using logic inference, but rather aFuzzy Decision Making scheme which works on top of arather complex performance evaluation framework based onFuzzy Arithmetic which spans across the whole protocol stack;moreover, we propose the adoption of Fuzzy Logic not onlyto design a decision strategy which can take into accountimprecision and uncertainty issues, but also as a technology-independent knowledge representation which can fit differentradio technologies, network protocols and user applications.

III. F UZZY LOGIC

The purpose of this section is to recall the aspects ofFuzzy Logic which are referred to throughout the rest ofthis paper. A detailed overview of Fuzzy Logic is howeverout of the scope of this paper, and the interested reader isreferred to the abundant literature available on this topic(seefor instance [18]).

Fuzzy Set Theory differs from traditional set theory in thatpartial membership is allowed, i.e., an element can belong to aset only to a certain degree. This degree of membership is com-monly referred to as themembership valueand is representedusing a real value in[0, 1], where0 and1 correspond to fullnon-membership and full membership, respectively. Formally,a fuzzy setA in a universeU is defined by the membershipfunction A : U → [0, 1] so that for eachu ∈ U its grade ofmembership to the setA is given byA(u). For fuzzy sets,the standard complement, union and intersection operatorsaredefined as

A(u) = 1 − A(u) (1)

(A ∩ B)(u) = min(A(u), B(u)) (2)

(A ∪ B)(u) = max(A(u), B(u)) (3)

These operators are of particular importance since they cor-respond, in predicate logic, to the¬ (NOT), ∧ (AND) and∨ (OR) operators, which in turn are widely used in AItechniques such as Fuzzy Control and Fuzzy Decision Making.

Ordinary functions can be extended to act on fuzzy sets bymeans of theextension principle[18, Sec. 2.3]. LetV andWbe ordinary sets, and letA andB be fuzzy sets defined overV andW respectively. For a generic functionf : V → W theextension principle defines its fuzzy counterpartf : A → Bas

(f(A)) (w) = supv:w=f(v)

A(v) (4)

A fuzzy set is said to be normal ifsupu∈U A(u) = 1. Finally,∀α ∈ [0, 1], theα-cut αA of a fuzzy setA is defined as

αA = {u ∈ U : A(u) ≥ α} (5)

α-cuts are important because of the so-calleddecompositiontheorem[19], which states that a fuzzy set is uniquely identi-fied by the family of itsα-cuts.

A fuzzy set X is said to be a Fuzzy Number if thefollowing conditions are satisfied: 1) its universe isR; i.e.,X : R −→ [0, 1]; 2) X is a normal fuzzy set; 3)∀α ∈ (0, 1],the α-cut αX is a closed interval; 4) the support ofX isbounded, i.e.,∃ a, b ∈ R such that∀u /∈ [a, b] X(u) = 0.Fuzzy Numbers are particularly interesting since they provideinformation which is both quantitative and qualitative at thesame time, and also because the standard arithmetic operationscan be applied to them. Due to this peculiarity, Fuzzy Numbersplay an important role in many applications such as fuzzycontrol, fuzzy decision making and approximate reasoning.

The arithmetic operations on fuzzy numbers can be definedthanks to the property that eachα-cut of a fuzzy number isa closed interval. Arithmetic operations on closed intervals[a, b], [c, d] ⊂ R are defined as follows [20]:

[a, b] + [c, d] = [a + c, b + d] (6)

[a, b] − [c, d] = [a − d, b − c]

[a, b] × [c, d] = [min(ac, ad, bc, bd), max(ac, ad, bc, bd)]

[a, b] / [c, d] =

[

min(a

c,a

d,b

c,b

d), max(

a

c,a

d,b

c,b

d)

]

Let nowA andB be fuzzy numbers.∀α ∈ [0, 1], using intervalarithmetic we can then define

α(A + B) = αA + αB; α(A − B) = αA − αBα(A × B) = αA × αB; α(A/B) = αA/αB (7)

Thanks to the decomposition theorem, the definition of afamily of α-cuts for the result of each arithmetic operationon fuzzy numbers uniquely identifies the result as a fuzzyset. It can be proved that this fuzzy set satisfies the necessaryrequirements to be a fuzzy number.

Using a similar approach, if we define the MIN and MAXoperators for closed intervals[a, b], [c, d] ∈ R as

MIN([a, b] , [c, d]) = [min(a, c), min(b, d)]

MAX ([a, b] , [c, d]) = [max(a, c), max(b, d)] (8)

we can then define the MIN and MAX operators for fuzzy

4

numbersA andB usingα-cut decomposition:

α(MIN(A, B)) = MIN(αA, αB)α(MAX (A, B)) = MAX (αA, αB) (9)

IV. K NOWLEDGE REPRESENTATION, CROSS-LAYER

INFORMATION PROCESSING ANDINFORMATION SHARING

In a previous study [21], we discussed several issuesconcerning the representation of cross-layer informationinwireless devices. In particular, we concluded that using ageneralized and technology-independent knowledge base isadvisable, and that techniques such as Fuzzy Numbers canfurther enhance the correct usability of information by provid-ing a mixed quantitative/qualitative representation. Followingthis approach, in this study we have chosen to characterizeperformance in terms ofthroughput, delay and reliability(defined as the success ratio of packet transmissions), whereeach of these items is represented using a fuzzy number. Thischaracterization is applied to different components of thecom-munication performance: wireless link quality measurementsperformed by each user, wired network performance reportedby all users to the Cognitive Network Database, estimationof end-to-end network-layer and transport layer performance,and QoS requirements of user applications.

We introduce the notation beforehand, in order to make thediscussion easier to understand in spite of the heterogeneityand vast number of variables used. We denote throughput,delay and reliability metrics witht, d and r; the subscriptsl, n, e, t and a denote respectively radio link, core network,end-to-end, transport and application metrics. Finally, for ageneric metricx, we denote its measured value withx, itsestimated value withx, and its fuzzy representation withx.

Our knowledge representation and cross-layer informationprocessing architecture is represented in Figure 1. Some mod-ules are expected to make use of technology-dependent in-formation to measure and/or estimate communication quality,and represent it using the technology-independent throughput,delay and reliability metrics. One of such modules is the radiolink module, which is in charge of providing the radio linkperformance measurementstl, dl andrl for ongoing communi-cations, and of expressing the estimated radio link performancefor access point selection using the fuzzy metricstl, dl andrl.Another use of technology-dependent information is withinthetransport module, where a characterization of the performanceprovided to the application by the transport layer as a functionof the end-to-end network layer performance is needed; wesuppose this characterization can be expressed using the threefunctionsft(de, re, te), fd(de, re, te) andfr(de, re, te), whichprovide respectively application layer throughput, delayandreliability as a function of end-to-end network-layer perfor-mance (either measured or estimated). Finally, application QoSrequirements are represented using fuzzy sets. We denote withta, da, ra the fuzzy sets representing respectively satisfactorythroughput, delay and reliability for a particular application;the membership functions of these sets are to be determineda priori using application-specific knowledge.

According to the principles discussed in [21], whiletechnology-specific knowledge is needed to design these

modules properly, the overall architecture is technology-independent, thanks to its modularity and to the use of anabstract and generic knowledge representation base in thedefinition of the interfaces between modules. In the rest ofthis section, we discuss the technology-independent architec-ture; an example implementation of the technology-dependentcomponents will be provided in Section V.

Each cognitive user performs at a given time one of two cog-nitive activities. The first one, Access Point Characterization,is carried out by users having ongoing communications, andconsists in processing radio link and end-to-end performancemeasurements to obtain an estimate of the core networkperformance for the access point in use; this information isshared with other users by means of the cognitive networkdatabase. The second activity is Access Point Selection, whichis performed by users willing to start a new communication;it consists of estimating the application-layer communicationperformance by combining the estimated performance at theradio link with the Access Point Characterization obtainedfrom the Cognitive Network Database, and subsequently se-lecting the Access Point which is expected to better satisfythe application requirements. Each of these activities will beexplained in the following subsections.

Access Point Characterization– This process is repre-sented in the left side of Figure 1. When a communication isbeing performed over the radio link, the radio link moduleprovides instantaneous measurements for the metricstl, dl

and rl. For the same communication, the transport layermodule measures the end-to-end network layer performance itis perceiving, which we denote withte, de and re. We definecore network performance as the performance of the part ofthe network beyond the Access Point. In our architecture,this is the information which is to be shared among users tosupport cooperative access point characterization effectively.Unfortunately, core network performance cannot be measureddirectly. For this reason, we propose to measure the corenetwork throughput, delay and reliability (tn, dn and rn,respectively) by comparing the measured end-to-end network-layer performance with the performance at the radio link layerand at the transport layer. In detail, the delay and reliabilityare calculated asdn = de − dl and rn = re/rl, respectively.For the core network throughput, the most reasonable way tomeasure it is by evaluating the measured end-to-end network-layer throughputte. However, we must consider that if thecore network is not the bottleneck this method can result ina severe underestimation; this can happen, for instance, whenthe bottleneck is at the wireless link, or when throughput islimited by the transport layer (e.g., due to high round trip timein TCP) or the application layer. To overcome these issues, wedefine the following estimate of core network throughput:

tn =

{

te if te < tl and te < ft(de, re, tl)

Tmax otherwise(10)

whereTmax is a suitably large number. Using all these calcu-lations, users with ongoing communications can periodicallycompute core network performance measurements and uploadthem to the Cognitive Network Database. The database will

5

Access Point Characterization Access Point Selection

Network Layer

Transport Layer

Radio Link Core Network

User

Network Layer

Transport Layer

Radio Link

Application Layer

Core Networkradio link model

technology-specific

radio link model

technology-specific

end-to-end networklayer performance

wireless linkperformance

core networkperformance

transport layer model

protocol-specific

cognitive network

database

end-to-end networklayer performance

transport layerperformance

wireless linkperformance

core networkperformance

user-perceivedservice quality

ξ

transport layer model

protocol-specific

application–specific

QoS requirements

cognitive network

database

te de re

tl dl rl tn dn rn tl dl rl tn dn rn

te de re

ta da ra

Fig. 1. Knowledge representation and cross-layer information processing for Cognitive Network Access

therefore be populated with a performance characterization ofall available access networks.

Access Point Selection– This process is represented inthe right side of Figure 1. When a user wants to start a newcommunication, it estimates the service quality which can beprovided by all the available access points, and selects themost suitable one. This estimation is obtained for each AP byprocessing the shared access point performance characteriza-tion provided by the cognitive network database, and the radiolink performance estimation provided by the relative radiolinkmodule within the particular user being considered. Radiolink performance metrics are represented using the fuzzynumberstl, dl and rl; these metrics are to be provided by theradio link module based on technology specific measurements(such as RSSI, interference, mobility. . . ), and the fuzzificationprocess is intended to account for imprecision and inaccuracyin the measurements. Core network performance metrics arerepresented by the fuzzy numberstn, dn and rn; the fuzzi-fication process is intended to represent the uncertainty dueto differences in the measurements performed by differentusers. To this aim, we chose to represent the resulting metricsusing triangular fuzzy numbers, where maximum membershipis attained by the meanµ of the available measurements, andthe support of the membership function is[µ − 2σ, µ + 2σ],whereσ is the standard deviation of the measurements [22].A rectangular sliding window is used to select only the morerecent measurements, so that in the fuzzification process oldermeasurements are discarded.

All the fuzzy performance metrics just introduced areprocessed using Fuzzy Arithmetic in order to evaluate thecommunication quality expected from each AP. From now

on, we explicitly include the index of the AP in the notationwhenever needed to avoid confusion.

First of all, the expected network-layer end-to-end per-formance te(i), de(i) and re(i) for each Access Pointiis determined by combining radio link and core networkperformance as follows:2

te(i) = MIN(tl(i), tn(i));

de(i) = dl(i) + dn(i)

re(i) = rl(i) × rn(i) (11)

Then, transport-layer performance is derived applying (4)tothe functionsft, fd andfr:

tt(i) = ft(de(i), re(i), te(i))

dt(i) = fd(de(i), re(i), te(i))

rt(i) = fr(de(i), re(i), te(i)) (12)

The fuzzy metrics just defined provide an estimate of thecommunication performance which will be provided to theapplication. By comparing them withta, da, ra we can derivethe valuesζt(i), ζd(i), ζr(i) ∈ [0, 1], which represent to whatdegree the connection through Access Pointi is expectedto satisfy each performance requirement of the application.Different techniques can be used for this purpose, the moststraightforward being maximum membership [22]:

ζt(i) = maxx∈R

(tt(i) ∩ ta)(x)

ζd(i) = maxx∈R

(dt(i) ∩ da)(x)

2we note that (11) is based on the assumption that the radio link and thecore network performances are independent.

6

ζr(i) = maxx∈R

(rt(i) ∩ ra)(x) (13)

We obtain an overall measure of the fitnessξi of Access Pointi to meet the user needs, by calculating the highest degree towhich all application requirements are jointly satisfied, i.e.,

ξi = min(ζt(i), ζd(i), ζr(i)) (14)

and choose the access opportunity for whichξi is maximum.

V. CASE STUDY: FILE TRANSFER ANDVOIP OVER WLANAND UMTS

In this section, we describe how our proposed scheme canbe implemented in some practical cases. In particular, we willconsider the scenario of file transfer and VoIP applicationsover WLAN and UMTS radio links. For this scenario, we willprovide a possible implementation of the technology-specificcomponents which are present in our proposed framework.For the majority of the fuzzy metrics, we adhere to the wide-spread practice of using triangular membership functions [18],since they provide a good tradeoff between expressivity andsimplicity. We note that the characterizations provided inthefollowing sections are not intended to be the best possible;rather, our focus is on showing the implementability of ourarchitecture, and to provide support for the performance eval-uation which will be presented in Section VII.

Application requirements – For file transfer, the member-ship function ta(u) we use for the throughput requirementis logarithmically increasing from0 to 1 in (103, 108); wepreferred a logarithmic increase over a linear increase sinceit yields a more meaningful representation of how user sat-isfaction increases with throughput. Furthermore, we assumethat file transfer poses no delay constraints (da(u) = 1 ∀u)but requires strict reliability (ra(u) = 1 for u = 1, and0 otherwise). For VoIP, we adopt the commonly recognizedreference values for G.711 speech quality. In particular, theone-way delay is considered excellent if< 0.15 s and poor if> 0.45 s, so we define the round-trip delay requirementda(u)as linearly decreasing from1 to 0 in (0.3, 0.9); the packet lossrate is considered acceptable if< 0.05, and consequently wedefine ra(u) as linearly increasing from0 to 1 in (0.95, 1);finally, the throughput requirementta(u) is chosen as linearlyincreasing from0 to 1 for throughput ranging from64000(G.711 bitrate) to80000 in order to provide some margin.We note that, thanks to the expressivity of the proposed fuzzyknowledge representation base, it is straightforward to providesupport for other applications: all that is needed is to providea proper fuzzy definition of their QoS requirements.

Radio link performance – For 802.11 access, we proposethe following characterization derived from the AP capacitymetric presented in [1] which refers to the case where thedownlink communication path is the bottleneck of the ongoingcommunications in the cell (a realistic assumption for most802.11 scenarios). In these conditions, it has been shown[23] that in the long term each user gets a fair share of theavailable cell capacity, due to the fact that the AP is almostthe only node contending for the channel. LetA be the setof users already associated with the AP, and letτj be thetime required for the transmission of a packet from the AP to

user j ∈ A. If we make the further assumption that packetlosses are negligible, a lower bound forτj can be easilycalculated once the packet size and the modulation schemeused by each userj ∈ A are known. Supposing that theAP uses a simple round-robin scheduling policy, it will serveall its users in an intervalT =

∑

j∈A τj , and the averagethroughputtl experienced by a particular user is given bytl = s/T , wheres is the payload size of the packets addressedto the user being considered. In the case of a user whichis not already associated with the AP, we redefinetl as thethroughput he might expect when associated with the AP astl = s/(T + τ), whereτ is the transmission time of the userunder consideration. Due to the assumptions made,tl is anoptimistic estimate of the radio link throughput. To account forthis estimation bias, we define the fuzzy metrictl as havinga triangular membership function with support(0.5tl, tl) andpeak at0.75tl. Using the same assumptions, we can determinea lower bound on the radio link contribution to the averageround trip time asdl = 0.5T + τ , where we have accountedfor the average waiting time for a random packet arrival at theAP in the downlink, and for the lowest possible transmissiontime (no contention, no retransmissions) in the uplink. Again,to account for the estimation bias we define the fuzzy metricdl

with a triangular membership function with support(dl, 1.5dl)and peak at1.25dl.

For the reliability metric, we consider the error probabilityof a SDU frame, for which we calculate the lower boundpl =prmax+1 and upper boundpl = (p + q)rmax+1, wherep isthe frame error rate (due to SNR only) for the modulationscheme being used,rmax is the MAC retransmission limit andq is a worst case estimate of the collision probability (in oursimulations we used the fixed valueq = 0.3; more accurateestimators could be used, such as the one in [24]). We thendefine the lower and upper reliability bounds asrl = 1 − pl

and rl = 1 − pl, respectively. Finally we define the fuzzymetric rl using a symmetric triangular membership functionwith support(rl, rl).

We point out that the type of information needed for thisscheme is likely to be possessed by the Access Point, and canbe forwarded to the clients, e.g., using information framessuchas those specified in the 802.11k and 802.11e protocols [25],[26]. Alternatively, each user could privately monitor eachwireless channel and get equivalent performance metrics bydirect measurements, though at a higher computational cost.

For UMTS access, we consider a Release 4 system inwhich data transmission is performed over a Dedicated Chan-nel (DCH) using Acknowledged Mode (AM) at the RadioLink Control (RLC) layer. For convenience we define thefollowing: K is the number of PDUs per SDU,I is thelength of the interleaving,TPDU is the duration of a PDUtransmission,TRLC = 2(2I + TPDU ) is the RLC RoundTrip Time, m = ⌊TRLC/TPDU⌋ is the number of PDUs perRLC Round Trip Time,L is the maximum allowed numberof transmission attempts for a single PDU. Furthermore, wesuppose an estimate of the PDU error probabilityp to be

7

available3 with a known confidence interval ofσp. We definep = max(p − σp, 0) and p = min(p + σp, 1). For a givenspreading factorx and link directiony (uplink/downlink),a DCH has a well-defined data throughputtl(x, y) [27].This is the maximum throughput which can be achieved inideal conditions; however, in typical conditions the actuallyachieved throughput can be slightly lower due to PDU lossesand the Selective Repeat ARQ used in AM. To account forthis factor, we define the fuzzy metrictl as having a triangularmembership with support((1 − p)tl, (1 − p)tl) and peak at(1 − p)tl. For the reliability, we consider the widely adoptedblock-fading model, according to which SDU losses occurwith probabilityf(z) = 1−(1−zL)K , wherez is a given PDUerror probability. Consequently, for the fuzzy reliability metricrl we use a triangular membership function with support(f(p), f(p)) and peak atf(p).

Finally, for the delay, we use the heuristic proposed in [28]which provides a very good approximation of the complemen-tary cumulative distribution functionccdf(t, z)4 of the SDUdelay t as a function of the PDU error probabilityz. Letdl = {t | ccdf(t, p) = 0.5}, dl = {t | ccdf(t, p) = 0.95}

and dl = {t | ccdf(t, p) = 0.05}. The fuzzy delay metricdl is defined as having a triangular membership with support(dl, dl) and peak atdl.

Transport layer performance – Since the applicationsconsidered in this case study are file transfer and VoIP, weneed to provide a proper characterization of the impact on theperformance of the TCP and UDP/RTP transport protocol. ForRTP/UDP we suppose that no particular ARQ/FEC scheme isin place; this choice yieldsfr(de, re, te) = re, fd(de, re, te) =de andft(de, re, te) = ηte, whereη ≤ 1 accounts for protocoloverhead. For TCP, we havefr(de, re, te) = 1 thanks to TCP’sreliability; for the throughput we adopt the formulation in[29],i.e.,

ft(de, re, te) = min (te, ηL min (W0/2de, β)) (15)

β =

(

2de

√

2pe

3 + αde min

(

1, 3√

3bpe

8

))−1

whereW0 is the maximum window size,pe = 1−re is the end-to-end packet error rate,L is the packet size,b is the delayedACK parameter,2de is the Round Trip Time, andαde is anestimate of the retransmission timeout obtained by scalingtheend-to-end delay. Furthermore, since we did not impose anydelay constraints for the file transfer application, we provideno formulation of fd(de, re, te) for the TCP case, as it isnot needed for our implementation. Finally, we note that themodel in [29] might not be very accurate in wireless scenarios.However, the use of a fuzzy representation and processingof the inputs of the model ensures that our transport layercharacterization is able to provide an effective and meaningfulrepresentation of the range and likelihood of the performanceexperienced on top of TCP, even in scenarios in which themodel itself might not be very accurate.

3We note that in a real systemp can be evaluated as a function of the SINRmeasurements which are commonly performed by UMTS devices as part ofthe inner loop power control.

4i.e., the probability that a SDU experiences a delay greaterthan t

Cognitive Network Database Implementation– It is to benoted that in this paper we do not make any assumption on thetype of architecture to be used for data upload and retrieval;instead, we just assume that the Cognitive Network Databaseis available to all users. In practice, different approachescould be adopted, each one having its own benefits anddrawbacks5 [30]. This assumption is justified by the need tovalidate the proposed cognitive network approach while notprecluding any possible implementation. A detailed study ofthe architecture of the Cognitive Network Knowledge Baseand the related trade-offs is left as a future research topic.

VI. OTHER NETWORK ACCESSDECISION SCHEMES

In this section we introduce other network access schemesfor comparison purposes. The first two schemes, Highest RSSIand Link Capacity, represent the current state-of-the-artinwireless network access; however, as we will discuss in thissection and show in Section VII, they present some obvioussub-optimality with respect to our cognitive scheme – e.g.,they do not consider the performance of the part of the networkbeyond the AP – and therefore they cannot provide any insighton how close the performance of our cognitive scheme isto a hypothetical optimal performance. To address this issue,we introduce two additional schemes, Network Capacity andLow Delay, that we explicitly design for comparison purposes.These two last schemes are application-specific, and exploita priori knowledge of the network topology as well as ofthe characteristics of the traffic generated by the users. Inmost network deployments this information is not known;moreover, even when this knowledge is available, e.g., innetworks designed to fit tight QoS requirements, it is normallynot communicated to the end users. As a consequence, theNetwork Capacity and Low Delay strategies introduced in thissection do not appear very well suited for implementation.Furthermore we note that, in general, relying on application-specific solutions requires a significant design effort for everynew application to be supported.

Unlike the Network Capacity and Low Delay schemes, ourcognitive solution has the remarkable features that 1) it isindependent of the particular application being used, and 2) thecharacterization of the performance of the wireless networkis obtained by means of information sharing performed bythe users, instead of usinga priori knowledge of the networkcharacteristics. Clearly, these features come at the priceofsome sub-optimality in the resulting access point selectionperformance. In order to quantify this degree of sub-optimality,we designed the Network Capacity and Low Delay schemes tobe as close to the optimal as possible, so that their performancecan be used as a reference.

Highest RSSI scheme– This access scheme is very oftenimplemented in real devices, due to its simplicity. The fun-damental assumption behind it is that a higher RSSI allows

5For example, a centralized approach such as storing the information relatedon each access opportunity on the Access Point itself could be very efficient;however this approach may have serious security problems, since a maliciousoperator could deliberately alter the information in the database to get moresubscribers and increase its own revenue. Keeping the Cognitive NetworkDatabase distributed among users would be more robust in this sense, thoughat an increased computational/storage cost and communication overhead.

8

a higher rate modulation scheme to be used, and thereforeyields higher throughput and lower communication delay. Letρi be the Received Signal Strength Indicator (RSSI) seen fromAccess Pointi by the user under consideration. The HighestRSSI approach consists of choosing the access opportunitywhich maximizesρi.

For the multi-technology case, we consider the case inwhich one or more 802.11 APs are present together with atmost one UMTS AP. A simple but reasonable policy would beto prefer 802.11 whenever available, following the commonsense that 802.11 provides a higher throughput and lowerprice connection. In this case, we assume a SNR threshold∆ is defined, representing the minimum SNR required forsuccessful communication with an 802.11 AP. If no 802.11 APis reachable with SNR> ∆, then the UMTS AP is selected,otherwise one of the 802.11 APs is selected using the HighestRSSI selection method.

Link Capacity scheme – Using the throughput estimategiven in Section V, a new userk can evaluate the throughputtl(i) he could get from each available APi. It has beenproposed in [1] to perform 802.11 AP selection based exclu-sively on this metric, which aims at achieving an even loadbalancing between different APs. This is clear once we notethat, with reference to the expression oftl for 802.11 providedin Section V,s does not depend on the AP being considered,and the chosen AP is the one that minimizesT +τ . If τ ≪ T ,this is the AP with the lowest load, and hence highest residualcapacity. For this reason we refer to this scheme asLinkCapacity.

Network Capacity scheme– The Network Capacity accessscheme is explicitly designed to maximize TCP throughput.We adopt the same metrics used for our Cognitive scheme,except that explicit knowledge of core network performanceis employed instead of measurements retrieved from the Cog-nitive Network Database. More in detail, suppose we knowa priori the nominal bandwidthB(i), delayD(i) and packeterror rateP (i) of the link between the fixed host and APi. Lettl(i), dl(i) andpl(i) be the expected radio link throughput andlower bounds for the delay and packet error rate, as defined inSection V. Furthermore, letNi,TCP andNi,CBR be the numberof TCP and CBR users associated with APi, respectively,and let the generic CBR applicationk = 1 . . .Ni,CBR havebandwidthBCBR,k. For each access opportunityi, the end-to-end network-layer performance can be estimated as

te(i) = min(

tl(i),(

B(i) −∑Ni,CBR

k=1BCBR,k

)

/Ni,TCP

)

de(i) = dl(i) + D(i) + B(i)/s

re(i) = 1 − P (i)pl(i) (16)

wheres is the packet size in use by the user performing the ac-cess decision. Then, using (15), we can calculate the expectedTCP throughput from APi as tt(i) = ft(de(i), re(i), te(i)),and select the AP which maximizestt(i).

Low Delay scheme– This scheme is explicitly gearedtowards real-time applications, such as Voice over IP, VideoConferencing, and other interactive applications whose servicequality is heavily influenced by packet errors and communica-

tion delays. As before, we use the same radio link performancemetrics used for the Cognitive scheme, and we replace the roleof the Cognitive Network Database by explicit knowledge ofcore network performance. In particular, we reuse the end-to-end network performance metrics derived in (16) for theNetwork Capacity scheme, and we choose the access oppor-tunity which satisfies the following minimization problem:

mini

de(i) : re(i) ≥ 0.95 (17)

where the value0.95 refers to the minimum packet successrate required by the G.711 codec, and is coherent with thedefinition of the fuzzy QoS requirements that we proposed inSection V for VoIP applications. The solution of (17) is notnecessarily optimal for all real-time communications, since acertain amount of delay might be tolerable and throughputrequirements are not considered. However, this scheme isreasonably efficient for applications such as VoIP, whichhave very low throughput requirements together with a well-defined maximum packet error rate for acceptable servicequality (typically, 0.05 for the G.711 codec), and in whichthe satisfaction level of the end user is inversely related to theexperienced end-to-end delay.

VII. PERFORMANCEEVALUATION

Performance evaluation of the proposed access scheme wascarried out using NS-Miracle [31], which is a multi-interfacecross-layer extension of the well-known NS simulator [32].We simulated a square area of30 × 30 m2 with two AccessPoints placed on one side of the square, andn randomlyplaced users. Each AP is connected to a fixed host with adedicated symmetrical link, whose bandwidth has a specifiedfixed value according to the considered scenario. Dependingon the scenario, either two 802.11 APs or one 802.11 and oneUMTS AP were used.6 For 802.11 communications, both thewireless users and the APs use 802.11g with a rate adaptationscheme which consists of selecting the modulation schemebased on the experienced SNR in order to achieve a targetPacket Error Rate≤ 0.01. For UMTS, we used a spreadingfactor of8 in both downlink and uplink, which corresponds toa bit rate of456 kbps in downlink and of240 kbps in uplink.

In the following, we present simulation results highlightingthe different performance of the schemes outlined in theprevious section. We consider several scenarios in order toevaluate different issues relevant to the network access se-lection problem, such as load balancing on the radio link,core network performance degradation, satisfaction of QoSrequirements, and interactions in mixed traffic situations. Allreported results, unless explicitly stated, have been obtainedaveraging 100 independent simulation runs in order to achievethe necessary statistical confidence. In several cases, we willbe interested in the fairness of a metricxi for the set of users

6We actually carried out performance evaluation for scenarios with 3 and 4APs as well. In general, the behavior of the Access Schemes under evaluationwas very similar to the one observed in the two AP scenarios, but theanalysis of the results is much more complex due to the highernumber ofenvironmental variables to consider. Since scenarios withmore than 2 APsdo not give any significant insight compared to scenarios with 2 APs, theyhave been omitted for the sake of brevity.

9

CognitiveNetwork Capacity

Link CapacityRSSI

Average number of active Nodes

Th

rou

gh

pu

t(b

it/s)

1816141210864

700000

600000

500000

400000

300000

200000

100000

0

Fig. 2. Throughput for Scenario 1

CognitiveNetwork Capacity

Link CapacityRSSI

Bandwidth Ratio

Th

rou

gh

pu

t(b

it/s)

10.10.01

250000

200000

150000

100000

50000

0

Fig. 3. Average Throughput for Scenario 2

i = 1, . . . , n; for this purpose, we use Jain’s index, defined as(∑

i xi)2/(n

∑

i xi2)

Scenario 1: Load balancing on the radio links– Thepurpose of this scenario is to evaluate the load balancingcapabilities of the different access schemes being considered.We simulated a scenario in which the links connecting thetwo APs with the fixed host have the same bandwidth (10Mbps). The APs are placed so that the RSSI seen by all usersfrom one AP is always slightly better than from the otherAP. The results, reported in Figure 2 for different numbersof TCP users, show that in such a situation the HighestRSSI scheme suffers a severe performance degradation dueto unbalanced load at the APs. All other schemes achievea similar performance, with the Cognitive scheme achievinga slight throughput improvement over the others when thereare enough users (n ≥ 6) to provide sufficient statisticalconfidence for the performance estimation provided by theCognitive Network Database. The performance improvementis due to the fact that using network performance measure-ments fed back from all users allows to account for events suchas, for instance, increased Round Trip Time due to downlinkcongestion at the AP; this type of performance degradation isneglected by the other schemes because they considera prioriknowledge only (Network Capacity scheme) or they do not

CognitiveNetwork Capacity

Link CapacityRSSI

Bandwidth Ratio

Th

rou

gh

pu

tF

airn

ess

10.10.01

1

0.8

0.6

0.4

0.2

0

Fig. 4. Throughput Fairness for Scenario 2

AP2, CognitiveAP1, Cognitive

AP2, Network CapacityAP1, Network Capacity

AP2, Link CapacityAP1, Link Capacity

AP2, RSSIAP1, RSSI

Bandwidth Ratio

Ave

rag

en

um

ber

of

activ

eT

CP

use

rs

10.10.01

14

12

10

8

6

4

2

0

Fig. 5. AP usage for Scenario 2

consider core network performance at all (Highest RSSI, LinkCapacity).

Scenario 2: Asymmetric core network performance–In this scenario we examine the performance of the differentaccess schemes in response to asymmetries in core networkbandwidth: the bandwidth of the link connecting the secondAP to the fixed host is only a fraction of the 10 Mbpsbandwidth which is available to the first AP. In each simula-tion, we haven TCP users uniformly distributed with respectto the APs. We show only the results forn = 15, as aqualitatively similar behavior was observed for other valuesof n. The obtained behavior for different bandwidth ratios isreported in Figures 3 and 4: although the throughput averagedamong all users seems similar across the different schemes,in asymmetric situations (low bandwidth ratio) the RSSI andLink Capacity schemes exhibit a significantly lower degreeof throughput fairness compared to the other schemes. Boththe Cognitive and the Network Capacity schemes providegood throughput and good fairness in all cases. As evidentfrom Figure 5, this is due to the fact that the RSSI andLink Capacity schemes assign on average half of the usersto AP2 in spite of the fact that it can offer significantlylower throughput compared to AP1, whereas the NetworkCapacity and Cognitive schemes are able to properly adapt

10

CognitiveNetwork Capacity

Link CapacityRSSI

backhaul link bandwidth for 802.11 AP (bit/s)

TC

PT

hro

ug

hp

ut

(bit/

s)

1e+071e+061e+05

500000

450000

400000

350000

300000

250000

200000

150000

100000

50000

0

Fig. 6. Throughput for Scenario 2b

to the bandwidth differences.We also examined other types of asymmetries in core

network performance, e.g., in terms of delay or packet errorrate. The obtained performance is similar to what observedfor the varying bandwidth case: the Highest RSSI and LinkCapacity schemes fail to recognize core network performancedegradation and result in severe throughput differences amongusers, while the Cognitive and Network Capacity schemesprovide significantly better fairness.

Scenario 2b: Multi-technology load balancing– For thisscenario, we used an 802.11 AP and a UMTS base station co-located in the center of one of the sides of the square area inwhich 15 users are randomly placed. The backhaul link of theUMTS AP was configured with a bandwidth of 100 Mbps. Weran several simulations varying the backhaul link bandwidthof the 802.11 AP.

The key point of this scenario is that the performance of theUMTS access is radio-link limited, while for the 802.11 accessit is core-network limited; in particular, when the backhaul linkof the 802.11 AP is not a bottleneck, the overall throughputachievable with 802.11 is greater than with UMTS due tothe greater radio link capacity. Moreover, while UMTS canoffer almost the same performance regardless of the numberof users (as long as they are not located at the border ofthe cell and their number is below the interference-limitedcapacity of the cell), the performance in an 802.11 cell andin its backbone is heavily influenced by the number of users,due respectively to the contention-based medium access andthe limited bandwidth available.

The performance obtained by the different schemes inthis scenario is reported in Figure 6. The RSSI and LinkCapacity schemes always select the 802.11 AP, resulting inpoor throughput performance as the bandwidth of the backhaullink becomes low. The behavior of the RSSI scheme is due tothe fact that in the chosen topology the 802.11 AP is alwaysreachable with sufficient RSSI to perform communication,while the Link Capacity scheme always chooses the 802.11AP because the expected radio link throughput is alwayshigher for 802.11 than for UMTS. On the other hand, theNetwork Capacity and Cognitive schemes are successful inprogressively preferring the UMTS AP as the 802.11 AP

CognitiveLow Delay

Link CapacityRSSI

Bandwidth Ratio

Ro

un

dT

rip

Tim

e(s

)

10.10.01

14

12

10

8

6

4

2

0

Fig. 7. Average Round Trip Time for Scenario 3

CognitiveNetwork Capacity

Link CapacityRSSI

Bandwidth Ratio

Th

rou

gh

pu

t(b

it/s)

10.10.01

700000

600000

500000

400000

300000

200000

100000

0

Fig. 8. TCP throughput for Scenario 4

backhaul link bandwidth becomes lower.Scenario 3: Real-time applications– This scenario is

designed to compare the performance of the access schemeswith respect to real-time applications. The topology of thisscenario is the same as in Scenario 2 as far as the asym-metry in core network performance is concerned; the onlydifference is the use of VoIP connections instead of TCPfile transfers. The results, reported in Figure 7, show againthat the RSSI and Link Capacity schemes provide poor delayperformance, due to their inability to choose the AP based onthe bandwidth of the backhaul link. On the other hand, theCognitive scheme achieves almost the same performance asthe Low Delay scheme, which achieves the best performanceamong all schemes thanks to its perfect knowledge of thenetwork parameters. (In this scenario, where delay rather thanthroughput is the main application constraint, we use LowDelay instead of Network Capacity.)

Scenario 4: Mixed traffic types – In this scenario, weconsider the case in which the two traffic types coexist in thesame area (6 TCP users and 7 VoIP users in the results shown),and share the same access resources. The purpose of this studyis to investigate the interactions between the two traffic classesand to understand how the transmission resources are shared.Figures 8 and 9 show the TCP throughput and VoIP delay

11

CognitiveLowest DelayLink Capacity

RSSI

Bandwidth Ratio

Ro

un

dT

rip

Tim

e(s

)

10.10.01

14

12

10

8

6

4

2

0

Fig. 9. VoIP Round Trip Time for Scenario 4

AP2, CognitiveAP1, Cognitive

AP2, Network CapacityAP1, Network Capacity

AP2, Link CapacityAP1, Link Capacity

AP2, RSSIAP1, RSSI

Bandwidth Ratio

Ave

rag

en

um

ber

of

activ

eT

CP

use

rs

10.10.01

14

12

10

8

6

4

2

0

Fig. 10. AP usage for Scenario 4 (TCP users)

performance of the two classes of users. As expected, theHighest RSSI scheme performs poorly in both cases, due to itscomplete unawareness of the core network asymmetries, as inscenarios 2 and 3. The other schemes trade off the throughputof TCP users and the delay of VoIP users in different ways.For example, the Cognitive scheme tends to keep a sufficientlylow delay, at the expense of a somewhat lower TCP throughputfor data users. On the other hand, the Link Capacity schemeprovides better TCP throughput but totally unacceptable VoIPdelay.

The behavior of the centralized schemes (Network Capacityfor TCP users and Low Delay for VoIP users, according tothe performance requirements of the two applications) needssome more detailed explanation. For extreme asymmetry, allusers are connected to the “good” Access Point, AP1 (seeFigures 10 and 11). As the bandwidth of AP2 increases, atsome point (400 kbps in the Figures) there is a sharp increaseof the TCP throughput. In fact, such bandwidth value canserve very well a single VoIP connection, and therefore eachVoIP user decides to move from AP1 to AP2 (note that thesedecisions are made simultaneously with no awareness of otherusers’ intentions). The result is that much more bandwidthbecomes available to TCP users (who stay at AP1) whereasAP2 becomes congested, which leads to unacceptable VoIP

AP2, CognitiveAP1, Cognitive

AP2, Network CapacityAP1, Network Capacity

AP2, Link CapacityAP1, Link Capacity

AP2, RSSIAP1, RSSI

Bandwidth Ratio

Ave

rag

en

um

ber

of

activ

eC

BR

use

rs

10.10.01

14

12

10

8

6

4

2

0

Fig. 11. AP usage for Scenario 4 (VoIP users)

CognitiveNetwork Capacity

Link CapacityRSSI

backhaul link bandwidth for 802.11 AP (bit/s)

TC

PT

hro

ug

hp

ut

(bit/

s)

1e+071e+061e+05

700000

600000

500000

400000

300000

200000

100000

0

Fig. 12. TCP Throughput for Scenario 4b

performance. This is a result of the blindness of the LowDelay strategy that, while knowing the network parametersa priori, is unable to react to events such as sudden increasesin delay due to the AP queues being filled by the relativelyaggressive behavior of TCP flows. As the bandwidth of AP2is further increased, congestion at AP2 is relieved, and VoIPperformance significantly improves. At this point, some TCPusers also try to move towards AP2, but this in fact resultsin poorer overall performance due to the complex interactionsamong the coexisting traffic flows. On the other hand, theCognitive scheme, while possibly showing a slightly inferioroverall performance for some classes of users, has a muchmore stable behavior and a significantly better fairness thanthe other schemes.

Scenario 4b: Multi-technology multi-application – Thisis similar to Scenario 4 but with a UMTS AP in place ofone of the 802.11 APs, as done for Scenario 2b. The resultsare reported in Figures 12 and 13. As expected, the RSSIand Link Capacity schemes always choose the 802.11 AP,thus resulting in poor performance for both TCP and VoIPflows. The joint usage of the Network Capacity and LowDelay scheme results in the VoIP flows always selectingUMTS to minimize the communication delay, and in the TCPflows distributing between UMTS and 802.11 to maximize

12

throughput performance. The Cognitive scheme has a similarbehavior, with the difference that for high values of thebackhaul link bandwidth a small fraction of VoIP users selectthe 802.11 AP; this results in some fluctuations in the APusage by TCP users which resembles the one already observedin Scenario 4, although smaller in magnitude, and has the sameexplanation provided in the previous section.

Control traffic overhead – To quantify the control trafficoverhead for the Cognitive scheme, we consider a simple im-plementation with a centralized Cognitive Network Databasereachable through any of the APs. We assume that a QoSreport consists of a 20 byte packet, 12 of which are used forthe throughput, delay and reliability metrics (a 32 bit floatingpoint number for each, just to be very conservative), 32 bytesto identify the AP (i.e., interface type, operating frequency,network name, and so on) and 8 bytes reserved for protocol-related information. Such a report protocol would likely workon top of UDP/IP, which requires additional 20 + 8 byteheaders. Consequently, a QoS report packet just below the IPlayer consists of 80 bytes. A query to the Cognitive NetworkDatabase could consist only of the 8 bytes of protocol-relatedinformation. The response provided by the Cognitive NetworkDatabase would include a 72 byte field for each AP beingreported, plus the 8 bytes of protocol-related information. Ifwe suppose that the performance of at most 10 APs canbe included for each response, we obtain a packet size of36 bytes for the query and of 756 bytes for the response.In our simulations we used a reporting interval of2 s, andan average data flow duration of12.5 s. As a consequence,for each user, the bandwidth required is352 bit/s for QoSreports and506.88 bit/s for queries/responses to and from theCognitive Network Database. Although these figures do nottake into account the overhead introduced by the MAC andPHY layers, it is clear that an implementation of the solutionproposed in this paper would require almost negligible controltraffic overhead.

VIII. C ONCLUSIONS

In this paper we have proposed a Cognitive Network Accessscheme in which the network performance reported by allprevious users is compared to application requirements andexpected radio link performance in order to help new usersto choose the best access opportunity. The framework featuresa modular design and a generic and technology-independentknowledge representation based on fuzzy logic, which togetherfacilitate the integration of different wireless technologies andapplications.

The performance of the proposed scheme has been evaluatedby means of simulation and compared with both state-of-the-art access schemes known from the literature and omni-scient application-specific schemes that we introduced as per-formance benchmarks. The performance evaluation involvedscenarios explicitly designed to highlight specific issues, inparticular load balancing on the radio link, load balancinginthe whole network, satisfaction of real-time QoS requirements,and coexistence of different traffic types. Both single radiotechnology (802.11) and multi-technology scenarios (802.11

CognitiveLow Delay

Link CapacityRSSI

backhaul link bandwidth for 802.11 AP (bit/s)

CB

RR

ou

nd

Tri

pT

ime

(s)

1e+071e+061e+05

30

25

20

15

10

5

0

Fig. 13. VoIP Round Trip Time for Scenario 4b

and UMTS) were considered to evaluate the capability ofthe different schemes to cope with heterogeneous wirelesstechnologies.

The results have shown that the Cognitive access schemeproposed in this paper may perform significantly better thanstate-of-the-art schemes, in terms of both overall performanceand fairness. Also, in most cases, the Cognitive scheme hasproven capable of achieving performance similar to the ref-erence omniscient application-specific schemes. The fact thatthe Cognitive scheme achieved this by exploiting informationshared by users rather than omniscience, while at the sametime offering a modular and flexible design which can easilyintegrate new wireless technologies and applications, confirmsthat the cognitive network approach we proposed in this paperfor wireless network access is effective, and worthy of furtherinvestigation. Some interesting directions for future researchinclude a detailed analysis of the extent to which the useof fuzzy logic is effective in handling the imprecision anduncertainty of the performance characterization of wirelessnetworks, as well as broadening the analysis presented inthis paper to include applications with more complex QoSrequirements, such as for example video communications.

REFERENCES

[1] K. Sundaresan and K. Papagiannaki, “The need for cross-layer infor-mation in access point selection,” inInternet Measurement Conference,Rio De Janeiro, Brazil, Oct 2006.

[2] A. Nicholson,et al., “Improved access point selection,” inInt’l Confer-ence on Mobile Systems, Applications and Services, Uppsala, Sweden,June 2006.

[3] V. Mhatre and K. Papagiannaki, “Using smart triggers forimproved userperformance in 802.11 wireless networks,” inInt’l Conference on MobileSystems, Applications and Services, Uppsala, Sweden, June 2006.

[4] D. Deng and H. Yen, “Quality-of-service provisioning system formultimedia transmission in IEEE 802.11 wireless LANs,”IEEE Journalon Selected Areas in Communications, vol. 23, no. 6, pp. 1240–1252,2005.

[5] M. Matsumoto and T. Itoh, “QoS-guarantee Method for Public WirelessLAN Access Environments,” inInt’l Conference on Wireless Networks,Communications and Mobile Computing, Maui, USA, June 2005.

[6] J. Al-Karaki and J. Chang, “A simple distributed access control schemefor supporting QoS in IEEE 802.11 wireless LANs,” inIEEE WirelessCommunications and Networking Conference, vol. 1, March 2004.

[7] R. Koodli and M. Puuskari, “Supporting packet-data QoS in nextgeneration cellular networks,”IEEE Communications Magazine, vol. 39,no. 2, pp. 180–188, 2001.

13

[8] S. Baey, M. Dumas, and M. Dumas, “QoS tuning and resource shar-ing for UMTS WCDMA multiservice mobile,”IEEE Transactions onMobile Computing, vol. 1, no. 3, pp. 221–235, 2002.

[9] R. Chakravorty, I. Pratt, and J. Crowcroft, “A frameworkfor dynamicSLA-based QoS control for UMTS,”IEEE Wireless Communications,vol. 10, no. 5, pp. 30–37, 2003.

[10] R. W. Thomas, D. H. Friend, L. A. DaSilva, and A. B. MacKenzie,“Cognitive Networks: Adaptation and Learning to Achieve End-toEndPerformance Objectives,”IEEE Communications Magazine, vol. 44,no. 12, pp. 51–57, 2006.

[11] J. Mitola, “Cognitive radio: an integrated agent architecture for softwaredefined radio,” Ph.D. dissertation, Royal Institute of Technology (KTH),2000.

[12] K. Knight and E. Rich,Artificial Intelligence. McGraw-Hill, 1994.[13] R. Zhang and K. Long, “A Fuzzy Routing Mechanism in Next Genera-

tion Networks,” in IASTED Int’l Conference on Intelligent Systems andControl, Tsukuba City, Japan, October 2002.

[14] S. Rea and D. Pesch, “Multi-metric routing decisions for ad hocnetworks using fuzzy logic,” inInt’l Symposium on Wireless Communi-cation Systems, Mauritius, 2004.

[15] L. Giupponi, R. Agusti, J. Perez-Romero, and O. Sallent, “Joint radioresource management algorithm for multi-RAT networks,” inIEEEGlobecom, St. Louis, USA, November 2005.

[16] M. Abdul-Haleem, K. Cheung, and J. Chuang, “Aggressivefuzzy dis-tributed dynamic channel assignment algorithm,”IEEE Int’l Conferenceon Communications, vol. 1, 1995.

[17] S. Ghosh, Q. Razouqi, H. Schumacher, and A. Celmins, “A survey ofrecent advances in fuzzy logic in telecommunications networks and newchallenges,”IEEE Transactions on Fuzzy Systems, vol. 6, no. 3, pp. 443–447, 1998.

[18] G. Klir and B. Yuan,Fuzzy sets and fuzzy logic: theory and applications.Prentice-Hall PTR, 1995.

[19] L. Zadeh, “Similarity relations and fuzzy orderings,”Information Sci-ences, vol. 3, no. 2, pp. 177–200, 1971.

[20] R. Moore,Interval analysis. Prentice-Hall Englewood Cliffs, NJ, 1966.[21] N. Baldo and M. Zorzi, “Fuzzy Logic for Cross-layer Optimization in

Cognitive Radio Networks,”IEEE Communications Magazine, vol. 46,no. 4, pp. 64–71, 2008.

[22] H. R. Berenji, “Fuzzy logic controllers,” inAn Introduction to FuzzyLogic Applications in Intelligent Systems, R. R. Yager and L. A. Zadeh,Eds. Boston: Kluwer, 1992, pp. 69–96.

[23] S. Choi, K. Park, and C. Kim, “On the performance characteristics ofWLANs: revisited,” Int’l Conference on Measurement and Modeling ofComputer Systems, 2005.

[24] N. Baldo,et al., “GORA: Goodput Optimal Rate Adaptation for 802.11using Medium Status Estimation,”IEEE Int’l Conference on Communi-cations, May 2008.

[25] Wireless LAN Medium Access Control (MAC) and Physical Layer(PHY)Specifications Amendment 1: Radio Resource Measurement of WirelessLANs, IEEE Std. 802.11k, 2008.

[26] Wireless LAN Medium Access Control (MAC) and Physical Layer(PHY)Specification, Amendment 8: Medium Access Control (MAC) Quality ofService Enhancements, IEEE Std. 802.11e, 2005.

[27] H. Holma and A. Toskala,WCDMA for UMTS. John Wiley and Sons,2004.

[28] M. Rossi and M. Zorzi, “Analysis and heuristics for the characterizationof selective repeat ARQ delay statistics over wireless channels,” IEEETransactions on Vehicular Technology, vol. 52, no. 5, pp. 1365–1377,2003.

[29] J. Padhye, V. Firoiu, D. Towsley, and J. Kurose, “Modeling TCP Renoperformance: a simple model and its empirical validation,”IEEE/ACMTransactions on Networking, vol. 8, no. 2, pp. 133–145, 2000.

[30] B. Manoj, R. Rao, and M. Zorzi, “Architectures and Protocols for NextGeneration Cognitive Networking,” inCognitive Wireless Networks:Concepts, Methodologies and Visions, M. Katz and F. Fitzek, Eds.Springer, 2007.

[31] N. Baldo, et al., “ns2-MIRACLE: a Modular Framework for Multi-Technology and Cross-Layer Support in Network Simulator 2,” Int’lWorkshop on Network Simulation Tools, October 2007.

[32] S. McCanne and S. Floyd, “NS-2 Network Simulator.” [Online].Available: http://www.isi.edu/nsnam/ns/

Nicola Baldo (S’07) was born in Rovigo, Italy,in 1981. He received his Laurea (BE) and LaureaSpecialistica (ME) degree in TelecommunicationsEngineering in 2003 and 2005, respectively, fromthe University of Ferrara, Italy, and the PhD inInformation Engineering from the University ofPadova, Italy, in 2009. In summer 2003 he was aninternship student at Ericsson Eurolab Deutschland,Aachen, Germany, working on Wireless MultimediaCommunications. In 2005 he was on leave at theSTMicroelectronics Advanced System Technology

group, Agrate Brianza (MI), Italy, working on Cross-layer Optimization forVoIP over Wireless LAN. In 2008 he was on leave at the California Institutefor Telecommunications and Information Technology (Calit2), University ofCalifornia, San Diego, USA, working on Cognitive Networks and SoftwareDefined Radio. Since February 2009 he has been a post-doc researcher atthe Centre Tecnologic de Telecomunicacions de Catalunya,Barcelona, Spain.His research interests include Cognitive Radio and Networks, Cross-layerOptimization, Multimedia Communications and Network Simulation Tools.

Michele Zorzi (S’89, M’95, SM’98, F’07) wasborn in Venice, Italy, in 1966. He received theLaurea degree and the Ph.D. degree in ElectricalEngineering from the University of Padova, Italy, in1990 and 1994, respectively. During the AcademicYear 1992/93, he was on leave at the University ofCalifornia, San Diego (UCSD), attending graduatecourses and doing research on multiple access inmobile radio networks. In 1993, he joined the facultyof the Dipartimento di Elettronica e Informazione,Politecnico di Milano, Italy. After spending three

years with the Center for Wireless Communications at UCSD, in 1998 hejoined the School of Engineering of the University of Ferrara, Italy, and in2003 joined the Department of Information Engineering of the Universityof Padova, Italy, where he is currently a Professor. His present researchinterests include performance evaluation in mobile communications systems,random access in mobile radio networks, ad hoc and sensor networks, energyconstrained communications protocols, and cognitive radio and networks.

Dr. Zorzi was the Editor-In-Chief of the IEEE WIRELESSCOMMUNICA -TIONS MAGAZINE from 2003 to 2005, is currently the Editor-In-Chief of theIEEE TRANSACTIONS ON COMMUNICATIONS, and serves on the SteeringCommittee of the IEEE TRANSACTIONS ONMOBILE COMPUTING, and onthe Editorial Boards of the WILEY JOURNAL OF WIRELESSCOMMUNICA -TIONS AND MOBILE COMPUTING and the ACM/URSI/KLUWER JOURNAL

OF WIRELESSNETWORKS. He was also guest editor for special issues in theIEEE PERSONAL COMMUNICATIONS MAGAZINE (Energy Management inPersonal Communications Systems) and the IEEE JOURNAL ON SELECTED

AREAS IN COMMUNICATIONS (Multi-media Network Radios).