See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/224563030

Cognitive Network Access using Fuzzy Decision Making

Article  in  IEEE Transactions on Wireless Communications · August 2009

DOI: 10.1109/TWC.2009.071103 · Source: IEEE Xplore

CITATIONS READS

61 182

2 authors, including:

Nicola Baldo
CTTC Catalan Telecommunications Technology Centre
57 PUBLICATIONS   1,377 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Cognitive Networks View project

All content following this page was uploaded by Nicola Baldo on 04 June 2014.

The user has requested enhancement of the downloaded file.

 Cognitive Network Access
using Fuzzy Decision Making
Nicola Baldo, Student Member, IEEE, and Michele Zorzi, Fellow, IEEE

Abstract— We consider a scenario in which wireless users cannot be compared to the ones obtained by different solutions
want to connect to the Internet using one of several available designed for other types of radio interfaces. Moreover, while
network access opportunities, possibly using different radio it is possible to design network access schemes to handle a
technologies. We propose a distributed cognitive network access
scheme with the aim of providing the best quality of service with particular set of existing technologies, an ideal network access
respect to both radio link and core network performance and solution would be required to be generic and modular enough
user application requirements. Knowledge of the service quality so as to accommodate even new wireless technologies as they
experienced by active connections is shared, and prospective users are introduced.
use Fuzzy Logic techniques to process cross-layer communication Another issue in wireless network access is that an ideal
quality metrics and to estimate the expected transport-layer per-
formance. These estimates are compared to the Quality of Service solution should be optimal with respect to the end-user per-
requirements of the application using Fuzzy Decision Making spective. In particular, two key aspects should be considered.
techniques to choose the most suitable access opportunity. This First, user-perceived service quality depends on end-to-end
scheme naturally fits into the recently proposed Cognitive Net- performance, in which the wireless link plays an important role
work paradigm in that it defines a cognition process leveraging without however being the only issue. Previous work on access
on end-to-end and cross-layer performance evaluation techniques
as well as information sharing among users; moreover, it offers point selection targeted only scenarios in which the wireless
a significant amount of flexibility and extensibility, thanks to its link is the bottleneck, while in many real life scenarios also the
modularity and its independence from the particular technology core network can exhibit non-ideal performance, thus having
and application being used. The proposed scheme is shown to a non-negligible effect on end-to-end network performance.
outperform state-of-the-art solutions in several multi-technology Second, the optimization of the end-to-end network perfor-
and multi-application scenarios, while at the same time achieving
similar performance to application-specific omniscient schemes mance should take the Quality of Service (QoS) requirements
that we introduce in this paper as a benchmark. of different applications into account. Much previous work
on Access Point Selection focuses on the maximization of
network throughput only; however, applications such as VoIP
I. I NTRODUCTION
and gaming are more strongly affected by other factors such
The recent progress in radio communications today provides as network delay and reliability. In recent years, interest in
several network access technologies for wireless connectivity, providing satisfactory service quality for these applications
e.g., IEEE 802.11, WiMAX and UMTS. At the same time, has grown considerably, due to their increased popularity;
advances in microelectronics allow all these technologies to this in turn has promoted a significant effort by the research
be exploited within a single mobile device equipped with community to find methods and solutions able to enhance
multiple radio interfaces. As a consequence of these facts, the QoS of multimedia applications on various wireless tech-
many new challenges have arisen for the telecommunication nologies, most notably 802.11 [4]–[6] and UMTS [7]–[9].
research community. The one we focus on in this paper is Unfortunately, this QoS-related research has focused almost
wireless network access, i.e., how a user who wants to connect exclusively on optimizing the performance of an already
to the Internet can select, among all the available opportunities, established multimedia communications, and the problem of
the one which yields the best performance. performing a QoS-aware network access decision has not been
Most previous work dealing with this problem considers dealt with so far.
only a specific wireless technology. The most notable example To summarize, the problem of identifying an algorithm for
is 802.11: several valuable solutions have been proposed to wireless network access selection which 1) is independent of
solve the problem of Access Point selection [1]–[3], but all of the radio technology, 2) can accommodate the QoS require-
them rely on 802.11-specific metrics. The consequence is that ments of different applications, and 3) can account not only
these solutions cannot be used with other wireless technolo- for wireless link performance, but also for possible non-ideal
gies and, most importantly, the obtained performance metrics conditions of the entire end-to-end path, is a very challenging
N. Baldo is with the Centre Tecnològic de Telecomunicacions de Catalunya, and still open issue.
Parc Mediterrani de la Tecnologia, Av. Canal Olimpic, 08860 Castelldefels, From this perspective, the recently proposed Cognitive
Barcelona, Spain, email: nbaldo@cttc.es. M. Zorzi is with the Dept. of Network paradigm seems to be a very promising approach.
Information Engineering, University of Padova, via Gradenigo 6/B, 35131
Padova, Italy, email: zorzi@dei.unipd.it. This work was done when N. Baldo In [10], a Cognitive Network is defined as “a network with a
was with the Dept. of Information Engineering, University of Padova. cognitive process that can perceive current network conditions,
This work was partially supported by the European Commission under the and then plan, decide, and act on those conditions”. Cogni-
ARAGORN project (INFSO-ICT-216856)
Part of this work has been presented at the IEEE CogNet 2007 Workshop tive Networking implies the presence of a cognition process
(held in conjunction with IEEE ICC 2007), Glasgow UK, June 24, 2007. which spans both all the layers of the protocol stack and all
 2

the network components of end-to-end communication; this about the best connection available. Based on this knowledge
cognition process is distributed among the different nodes of the previous experience of the first customer, he would
composing the network, which share their knowledge and likely connect to the hotspot, and be satisfied by enjoying
cooperate among themselves. This architecture is promising video streaming with high throughput connectivity. Then a
in that it is potentially able to solve problems which are third user comes in and, under advice from the previous users,
too complex to be handled within a traditional cross-layer connects to the hotspot for a VoIP conversation. It could hap-
approach, while at the same time being capable of learning the pen, however, that she is not satisfied because, e.g., the wired
behavior of different wireless technologies and applications. network serving the hotspot has a considerably high delay,
In this paper, we propose a Cognitive Network approach to or the hotspot itself has become congested; consequently, she
the Wireless Network Access problem. In particular, we pro- might try out UMTS, and possibly be more satisfied because of
pose a knowledge representation framework based on Fuzzy the lower Round Trip Time. Finally, a last user might come in
Logic which enables the implementation of a cognition process with his laptop to play some highly interactive online game.
which is both cross-layer and network-aware; furthermore, we He asks all previous customers about the performance they
exploit knowledge sharing among different devices with the experienced, and he decides to use the UMTS connection,
purpose of achieving a more complete and reliable characteri- since he understands that his application requirements (low
zation of the performance of the whole network. Subsequently, latency) are much closer to the requirements of the VoIP user
we define a network access scheme based on Fuzzy Decision rather than to those of the other two.
Making which allows each user to choose the Access Point Although very simple, this example highlights some im-
which best satisfies its QoS requirements. We stress that portant facts. First of all, while direct experience is an ef-
the modularity of the architecture and the generality of the fective means of inferring the quality of several connection
chosen knowledge representation allow our solution to easily opportunities, exploiting knowledge previously gathered by
provide optimized network access policies for new wireless other users may be a quicker and easier way. Second, the
technologies and applications which outperform state-of-the- choice of a suitable knowledge representation base for per-
art schemes while at the same time minimizing the effort formance evaluation metrics becomes crucial: on one hand
required to accommodate different wireless technologies and it is impossible to find a unique definition of quality, since
applications. different applications can have even conflicting performance
requirements, and so we need different metrics; on the other
hand, we cannot rely on too fine-grained and technology-
II. C OGNITIVE N ETWORK ACCESS
specific metrics, since this could possibly make information
It is worth mentioning that the original definition of Cog- gathered by some users difficult to interpret by others having
nitive Radio by J. Mitola [11] already tries to address the similar but not identical needs. Furthermore, we point out that
problem of wireless network access. Mitola’s Cognitive Radio the performance seen by each user is actually made up of two
has the explicit purpose of “detecting user needs, and provid- components, i.e., radio link and core network performance, and
ing wireless services most adequate to meet them”. The use that these aspects should be evaluated jointly. In the example
of the term cognitive highlights the fact that some Artificial above, suppose the first user leaves the cafe and goes and
Intelligence (AI) needs to be at the heart of the device which is sits on a bench in a nearby park. Previous experience still
to choose and adapt its services to the user’s needs. One of the indicates the hotspot as the most suitable choice. However,
most quoted definitions for AI is “how to make machines do current link quality metrics (e.g., the RSSI indicator) might
things at which, at the moment, humans are better” [12]. So, in show that he is too far from the hotspot to get an acceptable
the context of wireless network access, the goal of a cognitive throughput. Finally, it should be noted that all the available
radio would be to relieve the mobile users from having to information should be interpreted cum grano salis, i.e., not
figure out themselves which is the most satisfactory access only quantitatively but also qualitatively, because performance
opportunity. The key motivation behind the approach proposed reports by other users might be biased, measurements might
in this paper directly follows from these considerations: our be affected by errors, and all this information might be partial
goal is to replace user’s decisions trying to mimic the actual and/or old. 1
decision strategies that a typical human user would adopt. Based on the above discussion, we consider here a scheme
As an example, consider a cafe with a nearby 802.11 in which prospective users have access to a shared knowledge
hotspot, in an area covered by a UMTS provider. A single base that contains information about the service quality expe-
user comes in with his laptop, orders a coffee and looks for rienced by past and present active connections. To overcome
an Internet connection to do some web surfing. He might try the above mentioned issues, we define a generic knowledge
UMTS first, maybe just to realize that it is too slow for what representation framework using Fuzzy Numbers with the aim
he pays for it. Then he might realize that there is an 802.11 of enabling a generic representation of the most relevant
hotspot nearby, connect to it and finally surf the web happily performance metrics of different applications. We propose the
with high throughput. A second user comes in, orders another use of this knowledge representation to build a Cognitive
coffee, and turns on his mobile device to watch a TV program Network Knowledge Database, which is to be filled with
via video streaming from the Internet. Instead of figuring out 1 An important aspect ignored in this paper is how to deal with users that
the quality of the available networks by himself, he might just maliciously provide wrong information to influence other nodes’ decisions.
notice the first user who is surfing happily, and query him This issue is left for future research.
 3

service quality information fed back by all users actively using These operators are of particular importance since they cor-
the network. A user willing to set up a new connection can then respond, in predicate logic, to the ¬ (NOT), ∧ (AND) and
retrieve such information and compare it to both application ∨ (OR) operators, which in turn are widely used in AI
requirements and other measurements, in order to assess the techniques such as Fuzzy Control and Fuzzy Decision Making.
expected service quality for each access opportunity; in this Ordinary functions can be extended to act on fuzzy sets by
process, both Fuzzy Logic Inference and Fuzzy Arithmetic are means of the extension principle [18, Sec. 2.3]. Let V and W
used. Finally, the most suitable network access opportunity be ordinary sets, and let A and B be fuzzy sets defined over
must be selected using Fuzzy Decision Making techniques. V and W respectively. For a generic function f : V → W the
The presence of a knowledge representation base including extension principle defines its fuzzy counterpart f : A → B
information belonging to different components of the commu- as
nication system, the cross-layer processing of this information, (f (A)) (w) = sup A(v) (4)
the mechanism by which a node learns from its neighbors’ v:w=f (v)

experience as well as its own history, the use of Fuzzy Logic A fuzzy set is said to be normal if supu∈U A(u) = 1. Finally,
for incomplete knowledge representation, and the use of Fuzzy ∀α ∈ [0, 1], the α-cut αA of a fuzzy set A is defined as
Decision Making to select among access opportunities clearly
α
identify our proposed scheme as belonging to the Cognitive A = {u ∈ U : A(u) ≥ α} (5)
Network Paradigm. α-cuts are important because of the so-called decomposition
We note that Fuzzy Logic has been already proposed for theorem [19], which states that a fuzzy set is uniquely identi-
use in the context of telecommunication system, e.g., for fied by the family of its α-cuts.
QoS routing in wired networks [13], route caching decisions
A fuzzy set X is said to be a Fuzzy Number if the
in wireless ad hoc networks [14], radio resource manage-
following conditions are satisfied: 1) its universe is R; i.e.,
ment [15] and channel selection in cellular networks [16]. An
X : R −→ [0, 1]; 2) X is a normal fuzzy set; 3) ∀ α ∈ (0, 1],
interesting survey on the usage of Fuzzy Logic techniques in
the α-cut αX is a closed interval; 4) the support of X is
the telecommunication field can be found in [17]. The main
bounded, i.e., ∃ a, b ∈ R such that ∀u ∈ / [a, b] X(u) = 0.
difference of our approach is that, unlike previous work, we
Fuzzy Numbers are particularly interesting since they provide
do not consider a Fuzzy Controller implementing a simple
information which is both quantitative and qualitative at the
input-output relationship using logic inference, but rather a
same time, and also because the standard arithmetic operations
Fuzzy Decision Making scheme which works on top of a
can be applied to them. Due to this peculiarity, Fuzzy Numbers
rather complex performance evaluation framework based on
play an important role in many applications such as fuzzy
Fuzzy Arithmetic which spans across the whole protocol stack;
control, fuzzy decision making and approximate reasoning.
moreover, we propose the adoption of Fuzzy Logic not only
The arithmetic operations on fuzzy numbers can be defined
to design a decision strategy which can take into account
thanks to the property that each α-cut of a fuzzy number is
imprecision and uncertainty issues, but also as a technology-
a closed interval. Arithmetic operations on closed intervals
independent knowledge representation which can fit different
[a, b], [c, d] ⊂ R are defined as follows [20]:
radio technologies, network protocols and user applications.
[a, b] + [c, d] = [a + c, b + d] (6)
III. F UZZY L OGIC
[a, b] − [c, d] = [a − d, b − c]
The purpose of this section is to recall the aspects of
[a, b] × [c, d] = [min(ac, ad, bc, bd), max(ac, ad, bc, bd)]
Fuzzy Logic which are referred to throughout the rest of  
this paper. A detailed overview of Fuzzy Logic is however a a b b a a b b
[a, b] / [c, d] = min( , , , ), max( , , , )
out of the scope of this paper, and the interested reader is c d c d c d c d
referred to the abundant literature available on this topic (see Let now A and B be fuzzy numbers. ∀α ∈ [0, 1], using interval
for instance [18]). arithmetic we can then define
Fuzzy Set Theory differs from traditional set theory in that
α
partial membership is allowed, i.e., an element can belong to a (A + B) = αA + αB; α
(A − B) = αA − αB
α
set only to a certain degree. This degree of membership is com- (A × B) = αA × αB; α
(A/B) = αA/αB (7)
monly referred to as the membership value and is represented
using a real value in [0, 1], where 0 and 1 correspond to full Thanks to the decomposition theorem, the definition of a
non-membership and full membership, respectively. Formally, family of α-cuts for the result of each arithmetic operation
a fuzzy set A in a universe U is defined by the membership on fuzzy numbers uniquely identifies the result as a fuzzy
function A : U → [0, 1] so that for each u ∈ U its grade of set. It can be proved that this fuzzy set satisfies the necessary
membership to the set A is given by A(u). For fuzzy sets, requirements to be a fuzzy number.
the standard complement, union and intersection operators are Using a similar approach, if we define the MIN and MAX
defined as operators for closed intervals [a, b], [c, d] ∈ R as

A(u) = 1 − A(u) (1) MIN([a, b] , [c, d]) = [min(a, c), min(b, d)]
(A ∩ B)(u) = min(A(u), B(u)) (2) MAX([a, b] , [c, d]) = [max(a, c), max(b, d)] (8)
(A ∪ B)(u) = max(A(u), B(u)) (3) we can then define the MIN and MAX operators for fuzzy
 4

numbers A and B using α-cut decomposition: modules properly, the overall architecture is technology-
α α α independent, thanks to its modularity and to the use of an
(MIN(A, B)) = MIN( A, B)
abstract and generic knowledge representation base in the
α
(MAX(A, B)) = MAX(αA, αB) (9) definition of the interfaces between modules. In the rest of
this section, we discuss the technology-independent architec-
IV. K NOWLEDGE R EPRESENTATION , C ROSS - LAYER ture; an example implementation of the technology-dependent
I NFORMATION P ROCESSING AND I NFORMATION S HARING components will be provided in Section V.
In a previous study [21], we discussed several issues Each cognitive user performs at a given time one of two cog-
concerning the representation of cross-layer information in nitive activities. The first one, Access Point Characterization,
wireless devices. In particular, we concluded that using a is carried out by users having ongoing communications, and
generalized and technology-independent knowledge base is consists in processing radio link and end-to-end performance
advisable, and that techniques such as Fuzzy Numbers can measurements to obtain an estimate of the core network
further enhance the correct usability of information by provid- performance for the access point in use; this information is
ing a mixed quantitative/qualitative representation. Following shared with other users by means of the cognitive network
this approach, in this study we have chosen to characterize database. The second activity is Access Point Selection, which
performance in terms of throughput, delay and reliability is performed by users willing to start a new communication;
(defined as the success ratio of packet transmissions), where it consists of estimating the application-layer communication
each of these items is represented using a fuzzy number. This performance by combining the estimated performance at the
characterization is applied to different components of the com- radio link with the Access Point Characterization obtained
munication performance: wireless link quality measurements from the Cognitive Network Database, and subsequently se-
performed by each user, wired network performance reported lecting the Access Point which is expected to better satisfy
by all users to the Cognitive Network Database, estimation the application requirements. Each of these activities will be
of end-to-end network-layer and transport layer performance, explained in the following subsections.
and QoS requirements of user applications.
We introduce the notation beforehand, in order to make the Access Point Characterization – This process is repre-
discussion easier to understand in spite of the heterogeneity sented in the left side of Figure 1. When a communication is
and vast number of variables used. We denote throughput, being performed over the radio link, the radio link module
delay and reliability metrics with t, d and r; the subscripts provides instantaneous measurements for the metrics ťl , dˇl
l, n, e, t and a denote respectively radio link, core network, and řl . For the same communication, the transport layer
end-to-end, transport and application metrics. Finally, for a module measures the end-to-end network layer performance it
generic metric x, we denote its measured value with x̌, its is perceiving, which we denote with ťe , dˇe and ře . We define
estimated value with x̂, and its fuzzy representation with x̃. core network performance as the performance of the part of
Our knowledge representation and cross-layer information the network beyond the Access Point. In our architecture,
processing architecture is represented in Figure 1. Some mod- this is the information which is to be shared among users to
ules are expected to make use of technology-dependent in- support cooperative access point characterization effectively.
formation to measure and/or estimate communication quality, Unfortunately, core network performance cannot be measured
and represent it using the technology-independent throughput, directly. For this reason, we propose to measure the core
delay and reliability metrics. One of such modules is the radio network throughput, delay and reliability (ťn , dˇn and řn ,
link module, which is in charge of providing the radio link respectively) by comparing the measured end-to-end network-
performance measurements ťl , dˇl and řl for ongoing communi- layer performance with the performance at the radio link layer
cations, and of expressing the estimated radio link performance and at the transport layer. In detail, the delay and reliability
for access point selection using the fuzzy metrics t̃l , d˜l and r̃l . are calculated as dˇn = dˇe − dˇl and řn = ře /rˇl , respectively.
Another use of technology-dependent information is within the For the core network throughput, the most reasonable way to
transport module, where a characterization of the performance measure it is by evaluating the measured end-to-end network-
provided to the application by the transport layer as a function layer throughput ťe . However, we must consider that if the
of the end-to-end network layer performance is needed; we core network is not the bottleneck this method can result in
suppose this characterization can be expressed using the three a severe underestimation; this can happen, for instance, when
functions ft (de , re , te ), fd (de , re , te ) and fr (de , re , te ), which the bottleneck is at the wireless link, or when throughput is
provide respectively application layer throughput, delay and limited by the transport layer (e.g., due to high round trip time
reliability as a function of end-to-end network-layer perfor- in TCP) or the application layer. To overcome these issues, we
mance (either measured or estimated). Finally, application QoS define the following estimate of core network throughput:
(
requirements are represented using fuzzy sets. We denote with ťe if ťe < ťl and ťe < ft (dˇe , ře , ťl )
t̃a , d˜a , r̃a the fuzzy sets representing respectively satisfactory ťn = (10)
Tmax otherwise
throughput, delay and reliability for a particular application;
the membership functions of these sets are to be determined where Tmax is a suitably large number. Using all these calcu-
a priori using application-specific knowledge. lations, users with ongoing communications can periodically
According to the principles discussed in [21], while compute core network performance measurements and upload
technology-specific knowledge is needed to design these them to the Cognitive Network Database. The database will
 5

User

ξ user-perceived
service quality

application–specific
Application Layer QoS requirements

t̃a d˜a r̃a transport layer

performance

protocol-specific protocol-specific
Transport Layer transport layer model
Transport Layer transport layer model

ťe dˇe ře end-to-end network t̃e d˜e r̃e end-to-end network
layer performance layer performance

Network Layer Network Layer

ťl dˇl řl wireless link ťn dˇn řn core network t̃l d˜l r̃l wireless link t̃n d˜n r̃n core network
performance performance performance performance

technology-specific cognitive network technology-specific cognitive network

Radio Link radio link model Core Network database
Radio Link radio link model Core Network database

Access Point Characterization Access Point Selection

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

therefore be populated with a performance characterization of on, we explicitly include the index of the AP in the notation
all available access networks. whenever needed to avoid confusion.
Access Point Selection – This process is represented in First of all, the expected network-layer end-to-end per-
the right side of Figure 1. When a user wants to start a new formance t̂e (i), dˆe (i) and r̂e (i) for each Access Point i
communication, it estimates the service quality which can be is determined by combining radio link and core network
provided by all the available access points, and selects the performance as follows:2
most suitable one. This estimation is obtained for each AP by t̃e (i) = MIN(t̃l (i), t̃n (i));
processing the shared access point performance characteriza-
d˜e (i) = d˜l (i) + d˜n (i)
tion provided by the cognitive network database, and the radio
link performance estimation provided by the relative radio link r̃e (i) = r̃l (i) × r̃n (i) (11)
module within the particular user being considered. Radio Then, transport-layer performance is derived applying (4) to
link performance metrics are represented using the fuzzy the functions ft , fd and fr :
numbers t̃l , d˜l and r̃l ; these metrics are to be provided by the
radio link module based on technology specific measurements t̃t (i) = ft (d˜e (i), r̃e (i), t̃e (i))
(such as RSSI, interference, mobility. . . ), and the fuzzification d˜t (i) = fd (d˜e (i), r̃e (i), t̃e (i))
process is intended to account for imprecision and inaccuracy r̃t (i) = fr (d˜e (i), r̃e (i), t̃e (i)) (12)
in the measurements. Core network performance metrics are
represented by the fuzzy numbers t̃n , d˜n and r̃n ; the fuzzi- The fuzzy metrics just defined provide an estimate of the
fication process is intended to represent the uncertainty due communication performance which will be provided to the
to differences in the measurements performed by different application. By comparing them with t̃a , d˜a , r̃a we can derive
users. To this aim, we chose to represent the resulting metrics the values ζt (i), ζd (i), ζr (i) ∈ [0, 1], which represent to what
using triangular fuzzy numbers, where maximum membership degree the connection through Access Point i is expected
is attained by the mean µ of the available measurements, and to satisfy each performance requirement of the application.
the support of the membership function is [µ − 2σ, µ + 2σ], Different techniques can be used for this purpose, the most
where σ is the standard deviation of the measurements [22]. straightforward being maximum membership [22]:
A rectangular sliding window is used to select only the more
ζt (i) = max(t̃t (i) ∩ t̃a )(x)
recent measurements, so that in the fuzzification process older x∈R
measurements are discarded. ζd (i) = max(d˜t (i) ∩ d˜a )(x)
x∈R
All the fuzzy performance metrics just introduced are
processed using Fuzzy Arithmetic in order to evaluate the 2 we note that (11) is based on the assumption that the radio link and the
communication quality expected from each AP. From now core network performances are independent.
 6

ζr (i) = max(r̃t (i) ∩ r̃a )(x) (13) user j ∈ A. If we make the further assumption that packet
x∈R
losses are negligible, a lower bound for τj can be easily
We obtain an overall measure of the fitness ξi of Access Point calculated once the packet size and the modulation scheme
i to meet the user needs, by calculating the highest degree to used by each user j ∈ A are known. Supposing that the
which all application requirements are jointly satisfied, i.e., AP uses a simple round-robin scheduling policy, it will serve
P
ξi = min(ζt (i), ζd (i), ζr (i)) (14) all its users in an interval T = j∈A τj , and the average
throughput tl experienced by a particular user is given by
and choose the access opportunity for which ξi is maximum. tl = s/T , where s is the payload size of the packets addressed
to the user being considered. In the case of a user which
V. C ASE STUDY: F ILE T RANSFER AND VO IP OVER WLAN is not already associated with the AP, we redefine tl as the
AND UMTS throughput he might expect when associated with the AP as
In this section, we describe how our proposed scheme can tl = s/(T + τ ), where τ is the transmission time of the user
be implemented in some practical cases. In particular, we will under consideration. Due to the assumptions made, tl is an
consider the scenario of file transfer and VoIP applications optimistic estimate of the radio link throughput. To account for
over WLAN and UMTS radio links. For this scenario, we will this estimation bias, we define the fuzzy metric t̃l as having
provide a possible implementation of the technology-specific a triangular membership function with support (0.5tl , tl ) and
components which are present in our proposed framework. peak at 0.75tl . Using the same assumptions, we can determine
For the majority of the fuzzy metrics, we adhere to the wide- a lower bound on the radio link contribution to the average
spread practice of using triangular membership functions [18], round trip time as dl = 0.5T + τ , where we have accounted
since they provide a good tradeoff between expressivity and for the average waiting time for a random packet arrival at the
simplicity. We note that the characterizations provided in the AP in the downlink, and for the lowest possible transmission
following sections are not intended to be the best possible; time (no contention, no retransmissions) in the uplink. Again,
rather, our focus is on showing the implementability of our to account for the estimation bias we define the fuzzy metric d˜l
architecture, and to provide support for the performance eval- with a triangular membership function with support (dl , 1.5dl )
uation which will be presented in Section VII. and peak at 1.25dl .
Application requirements – For file transfer, the member-
ship function t̃a (u) we use for the throughput requirement For the reliability metric, we consider the error probability
is logarithmically increasing from 0 to 1 in (103 , 108 ); we of a SDU frame, for which we calculate the lower bound pl =
preferred a logarithmic increase over a linear increase since prmax +1 and upper bound pl = (p + q)rmax +1 , where p is
it yields a more meaningful representation of how user sat- the frame error rate (due to SNR only) for the modulation
isfaction increases with throughput. Furthermore, we assume scheme being used, rmax is the MAC retransmission limit and
that file transfer poses no delay constraints (d˜a (u) = 1 ∀u) q is a worst case estimate of the collision probability (in our
but requires strict reliability (r̃a (u) = 1 for u = 1, and simulations we used the fixed value q = 0.3; more accurate
0 otherwise). For VoIP, we adopt the commonly recognized estimators could be used, such as the one in [24]). We then
reference values for G.711 speech quality. In particular, the define the lower and upper reliability bounds as rl = 1 − pl
one-way delay is considered excellent if < 0.15 s and poor if and rl = 1 − pl , respectively. Finally we define the fuzzy
> 0.45 s, so we define the round-trip delay requirement d˜a (u) metric r̃l using a symmetric triangular membership function
as linearly decreasing from 1 to 0 in (0.3, 0.9); the packet loss with support (rl , rl ).
rate is considered acceptable if < 0.05, and consequently we
define r̃a (u) as linearly increasing from 0 to 1 in (0.95, 1); We point out that the type of information needed for this
finally, the throughput requirement t̃a (u) is chosen as linearly scheme is likely to be possessed by the Access Point, and can
increasing from 0 to 1 for throughput ranging from 64000 be forwarded to the clients, e.g., using information frames such
(G.711 bitrate) to 80000 in order to provide some margin. as those specified in the 802.11k and 802.11e protocols [25],
We note that, thanks to the expressivity of the proposed fuzzy [26]. Alternatively, each user could privately monitor each
knowledge representation base, it is straightforward to provide wireless channel and get equivalent performance metrics by
support for other applications: all that is needed is to provide direct measurements, though at a higher computational cost.
a proper fuzzy definition of their QoS requirements.
Radio link performance – For 802.11 access, we propose For UMTS access, we consider a Release 4 system in
the following characterization derived from the AP capacity which data transmission is performed over a Dedicated Chan-
metric presented in [1] which refers to the case where the nel (DCH) using Acknowledged Mode (AM) at the Radio
downlink communication path is the bottleneck of the ongoing Link Control (RLC) layer. For convenience we define the
communications in the cell (a realistic assumption for most following: K is the number of PDUs per SDU, I is the
802.11 scenarios). In these conditions, it has been shown length of the interleaving, TP DU is the duration of a PDU
[23] that in the long term each user gets a fair share of the transmission, TRLC = 2(2I + TP DU ) is the RLC Round
available cell capacity, due to the fact that the AP is almost Trip Time, m = ⌊TRLC /TP DU ⌋ is the number of PDUs per
the only node contending for the channel. Let A be the set RLC Round Trip Time, L is the maximum allowed number
of users already associated with the AP, and let τj be the of transmission attempts for a single PDU. Furthermore, we
time required for the transmission of a packet from the AP to suppose an estimate of the PDU error probability p to be
 7

available3 with a known confidence interval of σp . We define Cognitive Network Database Implementation – It is to be
p = max(p − σp , 0) and p = min(p + σp , 1). For a given noted that in this paper we do not make any assumption on the
spreading factor x and link direction y (uplink/downlink), type of architecture to be used for data upload and retrieval;
a DCH has a well-defined data throughput tl (x, y) [27]. instead, we just assume that the Cognitive Network Database
This is the maximum throughput which can be achieved in is available to all users. In practice, different approaches
ideal conditions; however, in typical conditions the actually could be adopted, each one having its own benefits and
achieved throughput can be slightly lower due to PDU losses drawbacks5 [30]. This assumption is justified by the need to
and the Selective Repeat ARQ used in AM. To account for validate the proposed cognitive network approach while not
this factor, we define the fuzzy metric t̃l as having a triangular precluding any possible implementation. A detailed study of
membership with support ((1 − p)tl , (1 − p)tl ) and peak at the architecture of the Cognitive Network Knowledge Base
(1 − p)tl . For the reliability, we consider the widely adopted and the related trade-offs is left as a future research topic.
block-fading model, according to which SDU losses occur
with probability f (z) = 1−(1−z L)K , where z is a given PDU VI. OTHER N ETWORK ACCESS D ECISION S CHEMES
error probability. Consequently, for the fuzzy reliability metric In this section we introduce other network access schemes
r̃l we use a triangular membership function with support for comparison purposes. The first two schemes, Highest RSSI
(f (p), f (p)) and peak at f (p). and Link Capacity, represent the current state-of-the-art in
Finally, for the delay, we use the heuristic proposed in [28] wireless network access; however, as we will discuss in this
which provides a very good approximation of the complemen- section and show in Section VII, they present some obvious
tary cumulative distribution function ccdf (t, z)4 of the SDU sub-optimality with respect to our cognitive scheme – e.g.,
delay t as a function of the PDU error probability z. Let they do not consider the performance of the part of the network
dl = {t | ccdf (t, p) = 0.5}, dl = {t | ccdf (t, p) = 0.95} beyond the AP – and therefore they cannot provide any insight
and dl = {t | ccdf (t, p) = 0.05}. The fuzzy delay metric on how close the performance of our cognitive scheme is
d˜l is defined as having a triangular membership with support to a hypothetical optimal performance. To address this issue,
(dl , dl ) and peak at dl . we introduce two additional schemes, Network Capacity and
Transport layer performance – Since the applications Low Delay, that we explicitly design for comparison purposes.
considered in this case study are file transfer and VoIP, we These two last schemes are application-specific, and exploit
need to provide a proper characterization of the impact on the a priori knowledge of the network topology as well as of
performance of the TCP and UDP/RTP transport protocol. For the characteristics of the traffic generated by the users. In
RTP/UDP we suppose that no particular ARQ/FEC scheme is most network deployments this information is not known;
in place; this choice yields fr (de , re , te ) = re , fd (de , re , te ) = moreover, even when this knowledge is available, e.g., in
de and ft (de , re , te ) = ηte , where η ≤ 1 accounts for protocol networks designed to fit tight QoS requirements, it is normally
overhead. For TCP, we have fr (de , re , te ) = 1 thanks to TCP’s not communicated to the end users. As a consequence, the
reliability; for the throughput we adopt the formulation in [29], Network Capacity and Low Delay strategies introduced in this
i.e., section do not appear very well suited for implementation.
Furthermore we note that, in general, relying on application-
ft (de , re , te ) = min (te , ηL min (W0 /2de , β)) (15) specific solutions requires a significant design effort for every
−1
new application to be supported.
 q  q
β = 2de 2p3e + αde min 1, 3 3bp 8
e
Unlike the Network Capacity and Low Delay schemes, our
cognitive solution has the remarkable features that 1) it is
where W0 is the maximum window size, pe = 1−re is the end- independent of the particular application being used, and 2) the
to-end packet error rate, L is the packet size, b is the delayed characterization of the performance of the wireless network
ACK parameter, 2de is the Round Trip Time, and αde is an is obtained by means of information sharing performed by
estimate of the retransmission timeout obtained by scaling the the users, instead of using a priori knowledge of the network
end-to-end delay. Furthermore, since we did not impose any characteristics. Clearly, these features come at the price of
delay constraints for the file transfer application, we provide some sub-optimality in the resulting access point selection
no formulation of fd (de , re , te ) for the TCP case, as it is performance. In order to quantify this degree of sub-optimality,
not needed for our implementation. Finally, we note that the we designed the Network Capacity and Low Delay schemes to
model in [29] might not be very accurate in wireless scenarios. be as close to the optimal as possible, so that their performance
However, the use of a fuzzy representation and processing can be used as a reference.
of the inputs of the model ensures that our transport layer Highest RSSI scheme – This access scheme is very often
characterization is able to provide an effective and meaningful implemented in real devices, due to its simplicity. The fun-
representation of the range and likelihood of the performance damental assumption behind it is that a higher RSSI allows
experienced on top of TCP, even in scenarios in which the
5 For example, a centralized approach such as storing the information related
model itself might not be very accurate.
on each access opportunity on the Access Point itself could be very efficient;
however this approach may have serious security problems, since a malicious
3 We note that in a real system p can be evaluated as a function of the SINR
operator could deliberately alter the information in the database to get more
measurements which are commonly performed by UMTS devices as part of subscribers and increase its own revenue. Keeping the Cognitive Network
the inner loop power control. Database distributed among users would be more robust in this sense, though
4 i.e., the probability that a SDU experiences a delay greater than t at an increased computational/storage cost and communication overhead.
 8

a higher rate modulation scheme to be used, and therefore tion delays. As before, we use the same radio link performance
yields higher throughput and lower communication delay. Let metrics used for the Cognitive scheme, and we replace the role
ρi be the Received Signal Strength Indicator (RSSI) seen from of the Cognitive Network Database by explicit knowledge of
Access Point i by the user under consideration. The Highest core network performance. In particular, we reuse the end-
RSSI approach consists of choosing the access opportunity to-end network performance metrics derived in (16) for the
which maximizes ρi . Network Capacity scheme, and we choose the access oppor-
For the multi-technology case, we consider the case in tunity which satisfies the following minimization problem:
which one or more 802.11 APs are present together with at min dˆe (i) : r̂e (i) ≥ 0.95 (17)
most one UMTS AP. A simple but reasonable policy would be i
to prefer 802.11 whenever available, following the common where the value 0.95 refers to the minimum packet success
sense that 802.11 provides a higher throughput and lower rate required by the G.711 codec, and is coherent with the
price connection. In this case, we assume a SNR threshold definition of the fuzzy QoS requirements that we proposed in
∆ is defined, representing the minimum SNR required for Section V for VoIP applications. The solution of (17) is not
successful communication with an 802.11 AP. If no 802.11 AP necessarily optimal for all real-time communications, since a
is reachable with SNR > ∆, then the UMTS AP is selected, certain amount of delay might be tolerable and throughput
otherwise one of the 802.11 APs is selected using the Highest requirements are not considered. However, this scheme is
RSSI selection method. reasonably efficient for applications such as VoIP, which
Link Capacity scheme – Using the throughput estimate have very low throughput requirements together with a well-
given in Section V, a new user k can evaluate the throughput defined maximum packet error rate for acceptable service
tl (i) he could get from each available AP i. It has been quality (typically, 0.05 for the G.711 codec), and in which
proposed in [1] to perform 802.11 AP selection based exclu- the satisfaction level of the end user is inversely related to the
sively on this metric, which aims at achieving an even load experienced end-to-end delay.
balancing between different APs. This is clear once we note
that, with reference to the expression of tl for 802.11 provided VII. P ERFORMANCE E VALUATION
in Section V, s does not depend on the AP being considered,
Performance evaluation of the proposed access scheme was
and the chosen AP is the one that minimizes T + τ . If τ ≪ T ,
carried out using NS-Miracle [31], which is a multi-interface
this is the AP with the lowest load, and hence highest residual
cross-layer extension of the well-known NS simulator [32].
capacity. For this reason we refer to this scheme as Link
We simulated a square area of 30 × 30 m2 with two Access
Capacity.
Points placed on one side of the square, and n randomly
Network Capacity scheme – The Network Capacity access placed users. Each AP is connected to a fixed host with a
scheme is explicitly designed to maximize TCP throughput. dedicated symmetrical link, whose bandwidth has a specified
We adopt the same metrics used for our Cognitive scheme, fixed value according to the considered scenario. Depending
except that explicit knowledge of core network performance on the scenario, either two 802.11 APs or one 802.11 and one
is employed instead of measurements retrieved from the Cog- UMTS AP were used.6 For 802.11 communications, both the
nitive Network Database. More in detail, suppose we know wireless users and the APs use 802.11g with a rate adaptation
a priori the nominal bandwidth B(i), delay D(i) and packet scheme which consists of selecting the modulation scheme
error rate P (i) of the link between the fixed host and AP i. Let based on the experienced SNR in order to achieve a target
tl (i), dl (i) and pl (i) be the expected radio link throughput and Packet Error Rate ≤ 0.01. For UMTS, we used a spreading
lower bounds for the delay and packet error rate, as defined in factor of 8 in both downlink and uplink, which corresponds to
Section V. Furthermore, let Ni,TCP and Ni,CBR be the number a bit rate of 456 kbps in downlink and of 240 kbps in uplink.
of TCP and CBR users associated with AP i, respectively, In the following, we present simulation results highlighting
and let the generic CBR application k = 1 . . . Ni,CBR have the different performance of the schemes outlined in the
bandwidth BCBR,k . For each access opportunity i, the end-to- previous section. We consider several scenarios in order to
end network-layer performance can be estimated as evaluate different issues relevant to the network access se-
  XNi,CBR   lection problem, such as load balancing on the radio link,
t̂e (i) = min tl (i), B(i) − BCBR,k /Ni,TCP
k=1 core network performance degradation, satisfaction of QoS
dˆe (i) = dl (i) + D(i) + B(i)/s requirements, and interactions in mixed traffic situations. All
reported results, unless explicitly stated, have been obtained
r̂e (i) = 1 − P (i)pl (i) (16)
averaging 100 independent simulation runs in order to achieve
where s is the packet size in use by the user performing the ac- the necessary statistical confidence. In several cases, we will
cess decision. Then, using (15), we can calculate the expected be interested in the fairness of a metric xi for the set of users
TCP throughput from AP i as tt (i) = ft (dˆe (i), r̂e (i), t̂e (i)),
6 We actually carried out performance evaluation for scenarios with 3 and 4
and select the AP which maximizes tt (i).
APs as well. In general, the behavior of the Access Schemes under evaluation
Low Delay scheme – This scheme is explicitly geared was very similar to the one observed in the two AP scenarios, but the
towards real-time applications, such as Voice over IP, Video analysis of the results is much more complex due to the higher number of
environmental variables to consider. Since scenarios with more than 2 APs
Conferencing, and other interactive applications whose service do not give any significant insight compared to scenarios with 2 APs, they
quality is heavily influenced by packet errors and communica- have been omitted for the sake of brevity.
 9

700000 1
RSSI
Link Capacity
600000 Network Capacity
Cognitive 0.8
500000

Throughput Fairness
Throughput (bit/s)

0.6
400000

300000
0.4

200000
0.2 RSSI
100000 Link Capacity
Network Capacity
Cognitive
0 0
4 6 8 10 12 14 16 18 0.01 0.1 1
Average number of active Nodes Bandwidth Ratio

Fig. 2. Throughput for Scenario 1 Fig. 4. Throughput Fairness for Scenario 2

250000
AP1, RSSI
14 AP2, RSSI
AP1, Link Capacity

Average number of active TCP users

200000 12 AP2, Link Capacity
AP1, Network Capacity
AP2, Network Capacity
10 AP1, Cognitive
Throughput (bit/s)

AP2, Cognitive
150000
8

100000 6

4
50000 RSSI
Link Capacity 2
Network Capacity
Cognitive
0 0
0.01 0.1 1 0.01 0.1 1
Bandwidth Ratio Bandwidth Ratio

Fig. 3. Average Throughput for Scenario 2 Fig. 5. AP usage for Scenario 2

iP= 1, . . . , n; P
for this purpose, we use Jain’s index, defined as consider core network performance at all (Highest RSSI, Link
( i xi )2 /(n i xi 2 ) Capacity).
Scenario 1: Load balancing on the radio links – The Scenario 2: Asymmetric core network performance –
purpose of this scenario is to evaluate the load balancing In this scenario we examine the performance of the different
capabilities of the different access schemes being considered. access schemes in response to asymmetries in core network
We simulated a scenario in which the links connecting the bandwidth: the bandwidth of the link connecting the second
two APs with the fixed host have the same bandwidth (10 AP to the fixed host is only a fraction of the 10 Mbps
Mbps). The APs are placed so that the RSSI seen by all users bandwidth which is available to the first AP. In each simula-
from one AP is always slightly better than from the other tion, we have n TCP users uniformly distributed with respect
AP. The results, reported in Figure 2 for different numbers to the APs. We show only the results for n = 15, as a
of TCP users, show that in such a situation the Highest qualitatively similar behavior was observed for other values
RSSI scheme suffers a severe performance degradation due of n. The obtained behavior for different bandwidth ratios is
to unbalanced load at the APs. All other schemes achieve reported in Figures 3 and 4: although the throughput averaged
a similar performance, with the Cognitive scheme achieving among all users seems similar across the different schemes,
a slight throughput improvement over the others when there in asymmetric situations (low bandwidth ratio) the RSSI and
are enough users (n ≥ 6) to provide sufficient statistical Link Capacity schemes exhibit a significantly lower degree
confidence for the performance estimation provided by the of throughput fairness compared to the other schemes. Both
Cognitive Network Database. The performance improvement the Cognitive and the Network Capacity schemes provide
is due to the fact that using network performance measure- good throughput and good fairness in all cases. As evident
ments fed back from all users allows to account for events such from Figure 5, this is due to the fact that the RSSI and
as, for instance, increased Round Trip Time due to downlink Link Capacity schemes assign on average half of the users
congestion at the AP; this type of performance degradation is to AP2 in spite of the fact that it can offer significantly
neglected by the other schemes because they consider a priori lower throughput compared to AP1, whereas the Network
knowledge only (Network Capacity scheme) or they do not Capacity and Cognitive schemes are able to properly adapt
 10

500000 14
RSSI
450000 Link Capacity
12 Low Delay
400000 Cognitive

10
TCP Throughput (bit/s)

350000

Round Trip Time (s)

300000
8
250000
6
200000

150000 4
100000 RSSI
Link Capacity 2
50000 Network Capacity
Cognitive
0 0
1e+05 1e+06 1e+07 0.01 0.1 1
backhaul link bandwidth for 802.11 AP (bit/s) Bandwidth Ratio

Fig. 6. Throughput for Scenario 2b Fig. 7. Average Round Trip Time for Scenario 3

700000
to the bandwidth differences. RSSI
Link Capacity
We also examined other types of asymmetries in core 600000 Network Capacity
Cognitive
network performance, e.g., in terms of delay or packet error
500000
rate. The obtained performance is similar to what observed

Throughput (bit/s)
for the varying bandwidth case: the Highest RSSI and Link 400000
Capacity schemes fail to recognize core network performance
degradation and result in severe throughput differences among 300000

users, while the Cognitive and Network Capacity schemes 200000

provide significantly better fairness.
Scenario 2b: Multi-technology load balancing – For this 100000
scenario, we used an 802.11 AP and a UMTS base station co-
0
located in the center of one of the sides of the square area in 0.01 0.1 1
which 15 users are randomly placed. The backhaul link of the Bandwidth Ratio
UMTS AP was configured with a bandwidth of 100 Mbps. We
Fig. 8. TCP throughput for Scenario 4
ran several simulations varying the backhaul link bandwidth
of the 802.11 AP.
The key point of this scenario is that the performance of the
UMTS access is radio-link limited, while for the 802.11 access backhaul link bandwidth becomes lower.
it is core-network limited; in particular, when the backhaul link Scenario 3: Real-time applications – This scenario is
of the 802.11 AP is not a bottleneck, the overall throughput designed to compare the performance of the access schemes
achievable with 802.11 is greater than with UMTS due to with respect to real-time applications. The topology of this
the greater radio link capacity. Moreover, while UMTS can scenario is the same as in Scenario 2 as far as the asym-
offer almost the same performance regardless of the number metry in core network performance is concerned; the only
of users (as long as they are not located at the border of difference is the use of VoIP connections instead of TCP
the cell and their number is below the interference-limited file transfers. The results, reported in Figure 7, show again
capacity of the cell), the performance in an 802.11 cell and that the RSSI and Link Capacity schemes provide poor delay
in its backbone is heavily influenced by the number of users, performance, due to their inability to choose the AP based on
due respectively to the contention-based medium access and the bandwidth of the backhaul link. On the other hand, the
the limited bandwidth available. Cognitive scheme achieves almost the same performance as
The performance obtained by the different schemes in the Low Delay scheme, which achieves the best performance
this scenario is reported in Figure 6. The RSSI and Link among all schemes thanks to its perfect knowledge of the
Capacity schemes always select the 802.11 AP, resulting in network parameters. (In this scenario, where delay rather than
poor throughput performance as the bandwidth of the backhaul throughput is the main application constraint, we use Low
link becomes low. The behavior of the RSSI scheme is due to Delay instead of Network Capacity.)
the fact that in the chosen topology the 802.11 AP is always Scenario 4: Mixed traffic types – In this scenario, we
reachable with sufficient RSSI to perform communication, consider the case in which the two traffic types coexist in the
while the Link Capacity scheme always chooses the 802.11 same area (6 TCP users and 7 VoIP users in the results shown),
AP because the expected radio link throughput is always and share the same access resources. The purpose of this study
higher for 802.11 than for UMTS. On the other hand, the is to investigate the interactions between the two traffic classes
Network Capacity and Cognitive schemes are successful in and to understand how the transmission resources are shared.
progressively preferring the UMTS AP as the 802.11 AP Figures 8 and 9 show the TCP throughput and VoIP delay
 11

14
RSSI AP1, RSSI
Link Capacity 14 AP2, RSSI
12 Lowest Delay AP1, Link Capacity

Average number of active CBR users

Cognitive 12 AP2, Link Capacity
AP1, Network Capacity
10 AP2, Network Capacity
Round Trip Time (s)

10 AP1, Cognitive
AP2, Cognitive
8
8

6
6

4 4

2 2

0 0
0.01 0.1 1 0.01 0.1 1
Bandwidth Ratio Bandwidth Ratio

Fig. 9. VoIP Round Trip Time for Scenario 4 Fig. 11. AP usage for Scenario 4 (VoIP users)

700000
AP1, RSSI
14 AP2, RSSI
AP1, Link Capacity 600000
Average number of active TCP users

12 AP2, Link Capacity

AP1, Network Capacity
AP2, Network Capacity 500000

TCP Throughput (bit/s)

10 AP1, Cognitive
AP2, Cognitive
400000
8

300000
6

4 200000

RSSI
2 100000 Link Capacity
Network Capacity
Cognitive
0 0
0.01 0.1 1 1e+05 1e+06 1e+07
Bandwidth Ratio backhaul link bandwidth for 802.11 AP (bit/s)

Fig. 10. AP usage for Scenario 4 (TCP users) Fig. 12. TCP Throughput for Scenario 4b

performance of the two classes of users. As expected, the performance. This is a result of the blindness of the Low
Highest RSSI scheme performs poorly in both cases, due to its Delay strategy that, while knowing the network parameters
complete unawareness of the core network asymmetries, as in a priori, is unable to react to events such as sudden increases
scenarios 2 and 3. The other schemes trade off the throughput in delay due to the AP queues being filled by the relatively
of TCP users and the delay of VoIP users in different ways. aggressive behavior of TCP flows. As the bandwidth of AP2
For example, the Cognitive scheme tends to keep a sufficiently is further increased, congestion at AP2 is relieved, and VoIP
low delay, at the expense of a somewhat lower TCP throughput performance significantly improves. At this point, some TCP
for data users. On the other hand, the Link Capacity scheme users also try to move towards AP2, but this in fact results
provides better TCP throughput but totally unacceptable VoIP in poorer overall performance due to the complex interactions
delay. among the coexisting traffic flows. On the other hand, the
The behavior of the centralized schemes (Network Capacity Cognitive scheme, while possibly showing a slightly inferior
for TCP users and Low Delay for VoIP users, according to overall performance for some classes of users, has a much
the performance requirements of the two applications) needs more stable behavior and a significantly better fairness than
some more detailed explanation. For extreme asymmetry, all the other schemes.
users are connected to the “good” Access Point, AP1 (see Scenario 4b: Multi-technology multi-application – This
Figures 10 and 11). As the bandwidth of AP2 increases, at is similar to Scenario 4 but with a UMTS AP in place of
some point (400 kbps in the Figures) there is a sharp increase one of the 802.11 APs, as done for Scenario 2b. The results
of the TCP throughput. In fact, such bandwidth value can are reported in Figures 12 and 13. As expected, the RSSI
serve very well a single VoIP connection, and therefore each and Link Capacity schemes always choose the 802.11 AP,
VoIP user decides to move from AP1 to AP2 (note that these thus resulting in poor performance for both TCP and VoIP
decisions are made simultaneously with no awareness of other flows. The joint usage of the Network Capacity and Low
users’ intentions). The result is that much more bandwidth Delay scheme results in the VoIP flows always selecting
becomes available to TCP users (who stay at AP1) whereas UMTS to minimize the communication delay, and in the TCP
AP2 becomes congested, which leads to unacceptable VoIP flows distributing between UMTS and 802.11 to maximize
 12

throughput performance. The Cognitive scheme has a similar 30

RSSI
behavior, with the difference that for high values of the Link Capacity
Low Delay
25
backhaul link bandwidth a small fraction of VoIP users select Cognitive

CBR Round Trip Time (s)

the 802.11 AP; this results in some fluctuations in the AP
20
usage by TCP users which resembles the one already observed
in Scenario 4, although smaller in magnitude, and has the same 15
explanation provided in the previous section.
Control traffic overhead – To quantify the control traffic 10
overhead for the Cognitive scheme, we consider a simple im-
plementation with a centralized Cognitive Network Database 5
reachable through any of the APs. We assume that a QoS
report consists of a 20 byte packet, 12 of which are used for 0
1e+05 1e+06 1e+07
the throughput, delay and reliability metrics (a 32 bit floating backhaul link bandwidth for 802.11 AP (bit/s)
point number for each, just to be very conservative), 32 bytes
to identify the AP (i.e., interface type, operating frequency, Fig. 13. VoIP Round Trip Time for Scenario 4b
network name, and so on) and 8 bytes reserved for protocol-
related information. Such a report protocol would likely work
on top of UDP/IP, which requires additional 20 + 8 byte and UMTS) were considered to evaluate the capability of
headers. Consequently, a QoS report packet just below the IP the different schemes to cope with heterogeneous wireless
layer consists of 80 bytes. A query to the Cognitive Network technologies.
Database could consist only of the 8 bytes of protocol-related The results have shown that the Cognitive access scheme
information. The response provided by the Cognitive Network proposed in this paper may perform significantly better than
Database would include a 72 byte field for each AP being state-of-the-art schemes, in terms of both overall performance
reported, plus the 8 bytes of protocol-related information. If and fairness. Also, in most cases, the Cognitive scheme has
we suppose that the performance of at most 10 APs can proven capable of achieving performance similar to the ref-
be included for each response, we obtain a packet size of erence omniscient application-specific schemes. The fact that
36 bytes for the query and of 756 bytes for the response. the Cognitive scheme achieved this by exploiting information
In our simulations we used a reporting interval of 2 s, and shared by users rather than omniscience, while at the same
an average data flow duration of 12.5 s. As a consequence, time offering a modular and flexible design which can easily
for each user, the bandwidth required is 352 bit/s for QoS integrate new wireless technologies and applications, confirms
reports and 506.88 bit/s for queries/responses to and from the that the cognitive network approach we proposed in this paper
Cognitive Network Database. Although these figures do not for wireless network access is effective, and worthy of further
take into account the overhead introduced by the MAC and investigation. Some interesting directions for future research
PHY layers, it is clear that an implementation of the solution include a detailed analysis of the extent to which the use
proposed in this paper would require almost negligible control of fuzzy logic is effective in handling the imprecision and
traffic overhead. uncertainty of the performance characterization of wireless
networks, as well as broadening the analysis presented in
this paper to include applications with more complex QoS
VIII. C ONCLUSIONS requirements, such as for example video communications.
In this paper we have proposed a Cognitive Network Access
scheme in which the network performance reported by all R EFERENCES
previous users is compared to application requirements and [1] K. Sundaresan and K. Papagiannaki, “The need for cross-layer infor-
expected radio link performance in order to help new users mation in access point selection,” in Internet Measurement Conference,
to choose the best access opportunity. The framework features Rio De Janeiro, Brazil, Oct 2006.
[2] A. Nicholson, et al., “Improved access point selection,” in Int’l Confer-
a modular design and a generic and technology-independent ence on Mobile Systems, Applications and Services, Uppsala, Sweden,
knowledge representation based on fuzzy logic, which together June 2006.
facilitate the integration of different wireless technologies and [3] V. Mhatre and K. Papagiannaki, “Using smart triggers for improved user
performance in 802.11 wireless networks,” in Int’l Conference on Mobile
applications. Systems, Applications and Services, Uppsala, Sweden, June 2006.
The performance of the proposed scheme has been evaluated [4] D. Deng and H. Yen, “Quality-of-service provisioning system for
by means of simulation and compared with both state-of- multimedia transmission in IEEE 802.11 wireless LANs,” IEEE Journal
on Selected Areas in Communications, vol. 23, no. 6, pp. 1240–1252,
the-art access schemes known from the literature and omni- 2005.
scient application-specific schemes that we introduced as per- [5] M. Matsumoto and T. Itoh, “QoS-guarantee Method for Public Wireless
formance benchmarks. The performance evaluation involved LAN Access Environments,” in Int’l Conference on Wireless Networks,
Communications and Mobile Computing, Maui, USA, June 2005.
scenarios explicitly designed to highlight specific issues, in [6] J. Al-Karaki and J. Chang, “A simple distributed access control scheme
particular load balancing on the radio link, load balancing in for supporting QoS in IEEE 802.11 wireless LANs,” in IEEE Wireless
the whole network, satisfaction of real-time QoS requirements, Communications and Networking Conference, vol. 1, March 2004.
[7] R. Koodli and M. Puuskari, “Supporting packet-data QoS in next
and coexistence of different traffic types. Both single radio generation cellular networks,” IEEE Communications Magazine, vol. 39,
technology (802.11) and multi-technology scenarios (802.11 no. 2, pp. 180–188, 2001.
 13

[8] S. Baey, M. Dumas, and M. Dumas, “QoS tuning and resource shar- Nicola Baldo (S’07) was born in Rovigo, Italy,
ing for UMTS WCDMA multiservice mobile,” IEEE Transactions on in 1981. He received his Laurea (BE) and Laurea
Mobile Computing, vol. 1, no. 3, pp. 221–235, 2002. Specialistica (ME) degree in Telecommunications
[9] R. Chakravorty, I. Pratt, and J. Crowcroft, “A framework for dynamic Engineering in 2003 and 2005, respectively, from
SLA-based QoS control for UMTS,” IEEE Wireless Communications, the University of Ferrara, Italy, and the PhD in
vol. 10, no. 5, pp. 30–37, 2003. Information Engineering from the University of
[10] R. W. Thomas, D. H. Friend, L. A. DaSilva, and A. B. MacKenzie, Padova, Italy, in 2009. In summer 2003 he was an
“Cognitive Networks: Adaptation and Learning to Achieve End-toEnd internship student at Ericsson Eurolab Deutschland,
Performance Objectives,” IEEE Communications Magazine, vol. 44, Aachen, Germany, working on Wireless Multimedia
no. 12, pp. 51–57, 2006. Communications. In 2005 he was on leave at the
[11] J. Mitola, “Cognitive radio: an integrated agent architecture for software STMicroelectronics Advanced System Technology
defined radio,” Ph.D. dissertation, Royal Institute of Technology (KTH), group, Agrate Brianza (MI), Italy, working on Cross-layer Optimization for
2000. VoIP over Wireless LAN. In 2008 he was on leave at the California Institute
[12] K. Knight and E. Rich, Artificial Intelligence. McGraw-Hill, 1994. for Telecommunications and Information Technology (Calit2), University of
[13] R. Zhang and K. Long, “A Fuzzy Routing Mechanism in Next Genera- California, San Diego, USA, working on Cognitive Networks and Software
tion Networks,” in IASTED Int’l Conference on Intelligent Systems and Defined Radio. Since February 2009 he has been a post-doc researcher at
Control, Tsukuba City, Japan, October 2002. the Centre Tecnològic de Telecomunicacions de Catalunya, Barcelona, Spain.
[14] S. Rea and D. Pesch, “Multi-metric routing decisions for ad hoc His research interests include Cognitive Radio and Networks, Cross-layer
networks using fuzzy logic,” in Int’l Symposium on Wireless Communi- Optimization, Multimedia Communications and Network Simulation Tools.
cation Systems, Mauritius, 2004.
[15] L. Giupponi, R. Agusti, J. Pérez-Romero, and O. Sallent, “Joint radio
resource management algorithm for multi-RAT networks,” in IEEE
Globecom, St. Louis, USA, November 2005.
[16] M. Abdul-Haleem, K. Cheung, and J. Chuang, “Aggressive fuzzy dis-
tributed dynamic channel assignment algorithm,” IEEE Int’l Conference
on Communications, vol. 1, 1995.
[17] S. Ghosh, Q. Razouqi, H. Schumacher, and A. Celmins, “A survey of
recent advances in fuzzy logic in telecommunications networks and new
challenges,” 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, Michele Zorzi (S’89, M’95, SM’98, F’07) was
no. 4, pp. 64–71, 2008. born in Venice, Italy, in 1966. He received the
[22] H. R. Berenji, “Fuzzy logic controllers,” in An Introduction to Fuzzy Laurea degree and the Ph.D. degree in Electrical
Logic Applications in Intelligent Systems, R. R. Yager and L. A. Zadeh, Engineering from the University of Padova, Italy, in
Eds. Boston: Kluwer, 1992, pp. 69–96. 1990 and 1994, respectively. During the Academic
[23] S. Choi, K. Park, and C. Kim, “On the performance characteristics of Year 1992/93, he was on leave at the University of
WLANs: revisited,” Int’l Conference on Measurement and Modeling of California, San Diego (UCSD), attending graduate
Computer Systems, 2005. courses and doing research on multiple access in
[24] N. Baldo, et al., “GORA: Goodput Optimal Rate Adaptation for 802.11 mobile radio networks. In 1993, he joined the faculty
using Medium Status Estimation,” IEEE Int’l Conference on Communi- of the Dipartimento di Elettronica e Informazione,
cations, May 2008. Politecnico di Milano, Italy. After spending three
[25] Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) years with the Center for Wireless Communications at UCSD, in 1998 he
Specifications Amendment 1: Radio Resource Measurement of Wireless joined the School of Engineering of the University of Ferrara, Italy, and in
LANs, IEEE Std. 802.11k, 2008. 2003 joined the Department of Information Engineering of the University
[26] Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) of Padova, Italy, where he is currently a Professor. His present research
Specification, Amendment 8: Medium Access Control (MAC) Quality of interests include performance evaluation in mobile communications systems,
Service Enhancements, IEEE Std. 802.11e, 2005. random access in mobile radio networks, ad hoc and sensor networks, energy
[27] H. Holma and A. Toskala, WCDMA for UMTS. John Wiley and Sons, constrained communications protocols, and cognitive radio and networks.
2004. Dr. Zorzi was the Editor-In-Chief of the IEEE W IRELESS C OMMUNICA -
[28] M. Rossi and M. Zorzi, “Analysis and heuristics for the characterization TIONS M AGAZINE from 2003 to 2005, is currently the Editor-In-Chief of the
of selective repeat ARQ delay statistics over wireless channels,” IEEE IEEE T RANSACTIONS ON C OMMUNICATIONS , and serves on the Steering
Transactions on Vehicular Technology, vol. 52, no. 5, pp. 1365–1377, Committee of the IEEE T RANSACTIONS ON M OBILE C OMPUTING, and on
2003. the Editorial Boards of the W ILEY J OURNAL OF W IRELESS C OMMUNICA -
[29] J. Padhye, V. Firoiu, D. Towsley, and J. Kurose, “Modeling TCP Reno TIONS AND M OBILE C OMPUTING and the ACM/URSI/K LUWER J OURNAL
performance: a simple model and its empirical validation,” IEEE/ACM OF W IRELESS N ETWORKS . He was also guest editor for special issues in the
Transactions on Networking, vol. 8, no. 2, pp. 133–145, 2000. IEEE P ERSONAL C OMMUNICATIONS M AGAZINE (Energy Management in
[30] B. Manoj, R. Rao, and M. Zorzi, “Architectures and Protocols for Next Personal Communications Systems) and the IEEE J OURNAL ON S ELECTED
Generation Cognitive Networking,” in Cognitive Wireless Networks: A REAS IN C OMMUNICATIONS (Multi-media Network Radios).
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’l
Workshop on Network Simulation Tools, October 2007.
[32] S. McCanne and S. Floyd, “NS-2 Network Simulator.” [Online].
Available: http://www.isi.edu/nsnam/ns/

View publication stats