Content uploaded by Yevgeni Koucheryavy
Author content
All content in this area was uploaded by Yevgeni Koucheryavy on Mar 21, 2015
Content may be subject to copyright.
IEEE Wireless Communications • December 2014
86 1536-1284/14/$25.00 © 2014 IEEE
Sergey Andreev, Mikhail
Gerasimenko, Olga
Galinina, and Yevgeni
Koucheryavy, are with
Tampere University of
Technology.
Nageen Himayat, Shu-
ping Yeh, and Shilpa Tal-
war are with Intel
Corporation.
MOBILE CONVERGED NETWORKS
RECENT ADVANCES IN
MULTI-RADIO NETWORKING
The rapid expansion of wireless communications
over the last decades has introduced fundamen-
tal changes to “anytime, anywhere” mobile Inter-
net access, as well as posed new challenges for
the research community. In 2011, the fourth gen-
eration of broadband communication standards
was completed and offered aggressive improve-
ments in all aspects of wireless system design,
including system capacity, energy efficiency, and
user quality of service (QoS). As the respective
technologies are being deployed today, the focus
of recent research efforts is shifting to what may
be referred to as fifth generation (5G) wireless
networks.
Given a historical 10-year cycle for every
existing generation, it is expected that 5G sys-
tems will be deployed sometime around 2020.
Whereas there is currently no complete technical
definition of what comes after the state-of-the-
art networking technology, the anticipated com-
munication requirements may already be
understood from the user perspective. Regard-
less of their current location, human users would
like to be connected at all times, taking advan-
tage of the rich set of services provided by con-
temporary multimedia-over-wireless networks.
This creates significant challenges for 5G tech-
nology design, as users’ connectivity experience
should match data rate requirements and be uni-
form no matter where the user is, to whom the
user connects, and what the user’s service needs
are [1].
Unfortunately, contemporary wireless net-
works are currently unable to deliver the desired
ubiquitous connectivity experience. In the first
place, they lack uniformity in data rates and suf-
fer from excessive time delays, or sometimes
even service outage due to poor coverage and
severe interference conditions. While current
technologies have indeed been helpful in coping
with some of these challenges, it is commonly
believed that they will still be insufficient to
meet the anticipated growth in traffic demand
SERGEY ANDREEV, MIKHAIL GERASIMENKO, OLGA GALININA, YEVGENI KOUCHERYAVY,
NAGEEN HIMAYAT, SHU-PING YEH, AND SHILPA TALWAR
ABSTRACT
Heterogeneous multi-radio networks are
emerging network architectures that comprise
hierarchical deployments of increasingly smaller
cells. In these deployments, each user device
may employ multiple radio access technologies
to communicate with network infrastructure.
With the growing numbers of such multi-radio
consumer devices, mobile network operators
seek to leverage spectrum across diverse radio
technologies, thus boosting capacity and enhanc-
ing quality of service. In this article, we review
major challenges in delivering uniform connec-
tivity and service experience to converged multi-
radio heterogeneous deployments. We envision
that multiple radios and associated device/infra-
structure intelligence for their efficient use will
become a fundamental characteristic of future
5G technologies, where the distributed unli-
censed-band network (e.g., WiFi) may take
advantage of the centralized control function
residing in the cellular network (e.g., 3GPP
LTE). Illustrating several available architectural
choices for integrating WiFi and LTE networks,
we specifically focus on interworking within the
radio access network and detail feasible options
for intelligent access network selection. Both
network- and user-centric approaches are con-
sidered, wherein the control rests with the net-
work or the user. In particular, our system-level
simulation results indicate that load-aware user-
centric schemes, which augment SNR measure-
ments with additional information about network
loading, could improve the performance of con-
ventional WiFi-preferred solutions based on
minimum SNR threshold. Comparison with
more advanced network-controlled schemes has
also been completed to confirm attractive practi-
cal benefits of distributed user-centric algo-
rithms. Building on extensive system-wide
simulation data, we also propose novel analytical
space-time methodology for assisted network
selection capturing user traffic dynamics togeth-
er with spatial randomness of multi-radio het-
erogeneous networks.
INTELLIGENT ACCESS NETWORK SELECTION IN
CONVERGED MULTI-RADIO
HETEROGENEOUS NETWORKS
ANDREEV_LAYOUT_Author Layout 12/18/14 4:21 PM Page 86
IEEE Wireless Communications • December 2014 87
(nearly 11-fold over the following 5 years [2])
aggravated by rapid proliferation in types and
numbers of wireless devices. To make matters
worse, billions of diverse machine-type devices
connect to the network, thus reshaping the Inter-
net as we know it today. All these technological
challenges accentuate the need to explore novel
solutions within the context of 5G networks.
MAJOR TRENDS BEHIND 5G TECHNOLOGY
A transformation of mobile user experience
requires revolutionary changes in both network
infrastructure and device architecture, where the
user equipment (UE) is jointly optimized with
the surrounding network context [3]. Many
believe that the only feasible solution to mitigate
the increasing disproportion between the desired
QoS and the limited wireless resources is by
deploying higher densities of femto- and pico-
cells in current cellular architecture. Due to
shorter radio links, smaller cells provide higher
data rates and require less energy for uplink
transmission, especially in urban environments.
However, introducing an increasing number
of serving stations to bridge the capacity gap
incurs extra complexity due to more cumber-
some interference management, higher rental
fees, and increased infrastructure maintenance
costs [4]. More importantly, licensed spectrum
continues to be scarce and expensive, whereas
the traditional methods to improve its efficient
use are approaching their theoretical limits.
Even when additional spectrum is allocated,
these new frequencies are likely to remain frag-
mented and could require diverse transmission
techniques. Consequently, there is a pressing
demand to leverage additional capacity across
multiple radio access technologies (RATs).
As a result, it becomes crucial to aggregate
different radio technologies as part of a common
converged radio network, in a manner transpar-
ent to the end user, and develop techniques that
can efficiently utilize the radio resources avail-
able across different spectral bands potentially
using various RATs [5]. In particular, we expect
that the majority of immediate gains will come
from advanced architectures and protocols that
would leverage the unlicensed spectrum. For
example, mobile users with direct device-to-
device communication capability may take
advantage of their unlicensed-band radios and
cooperate with other proximate users to locally
improve access in a cost-efficient way [6].
Furthermore, as cell sizes shrink, the foot-
prints of cellular, local, and personal area net-
works are increasingly overlapping. This creates
an attractive opportunity to simultaneously utilize
multiple RATs for improved wireless connectivity.
We thus believe that intelligent multi-RAT cou-
pling will efficiently leverage performance bene-
fits across several dimensions of diversity,
including spatial, temporal, frequency, interfer-
ence, load, and others. In future 5G networks,
both short- and long-range technologies may need
to work cooperatively and exploit the intricate
interactions between the device and the network,
as well as between the devices themselves, to real-
ize the desired uniform user experience [7].
Consequently, the incentive to efficiently
coordinate between the alternative RATs is
growing stronger, and we envision that multiple
radios together with the associated device/system
intelligence for their efficient use will become a
fundamental characteristic of next-generation
networks [8]. More specifically, the distributed
unlicensed-band network (e.g., wireless local
area network, WLAN) may take advantage of
the centralized control function residing in the
cellular network to effectively perform dynamic
multi-RAT network association, and hence pro-
vide beyond-additive gains in network capacity
and user connectivity experience.
SCOPE AND CORE NOVELTY OF
CURRENT RESEARCH
According to the above, there is currently an
increasing shift toward tighter interworking
between different RATs. To this end, our
research campaign is targeting joint RAT assign-
ment, selection, and scheduling algorithms,
which provide significant improvement in overall
system performance. In what follows, our focus
is set on integration between multiple RATs
within heterogeneous network architecture. As
our case study, we consider convergence of
WLAN-based small cells with operator-managed
cellular deployment to illustrate feasible archi-
tectural options for integration and their associ-
ated performance benefits. Consequently, we
seek to explore the potential of a diverse range
of devices requiring connectivity at different
scales to augment system capacity and improve
user connectivity experience.
We emphasize that interworking between
WLAN and cellular networks has already been
considered in the past, but largely from the per-
spective of internetwork (vertical) handoff [9].
The cellular standards community, represented
by the Third Generation Partnership Project
(3GPP), has also been involved in developing
specifications that address cellular/WLAN inter-
working for a number of years. Several new
study and work items have recently emerged to
develop specifications toward tighter integration
of WLAN with cellular networks. The areas of
investigation range from solutions for trusted
access to 3GPP services with WLAN devices,
seamless mobility between 3GPP and WLAN
technologies, and support for the access network
discovery and selection function (ANDSF).
While much of this effort has focused on loose
interworking solutions only requiring changes
within the core network, there has been a recent
shift in 3GPP Release 12 to address interwork-
ing within the radio access network (RAN) [10].
This emerging trend is driven by the need to
support better QoS on unlicensed spectrum as
demanded by a consortium of network operators
who have introduced stringent requirements for
carrier-grade WiFi. The WLAN community has
also responded with new initiatives such as Hot
Spot 2.0, as well as a novel high-efficiency
WLAN standardization effort by the IEEE
802.11 working group. Hence, it is timely to
investigate RAN-based integration solutions,
which assume increased cooperation between
3GPP Long Term Evolution (LTE) and WiFi
radio technologies. Along these lines, our work
details several intelligent network selection
Even when additional
spectrum is allocated,
these new frequencies
are likely to remain frag-
mented and could
require diverse transmis-
sion techniques. Conse-
quently, there is a
pressing demand to
leverage additional
capacity across multiple
radio access
technologies.
ANDREEV_LAYOUT_Author Layout 12/18/14 4:21 PM Page 87
IEEE Wireless Communications • December 2014
88
mechanisms that deliver significant gains in over-
all system performance and user QoS. We
address both network- and user-centric
approaches, wherein the control of how different
radio technologies are utilized rests with the net-
work or the user, respectively.
ENABLING ARCHITECTURES AND
ALGORITHMS FOR CONVERGED
HETEROGENEOUS NETWORKS
As argued previously, the capacity and connec-
tivity limitations faced by future 5G networks
will continue to drive the need for closer integra-
tion across different RATs. Along these lines,
Fig. 1a illustrates our vision of an operator’s
multi-RAT heterogeneous network (HetNet). It
features a hierarchical deployment of wide-area
macrocells for ubiquitous coverage, connectivity,
and seamless mobility augmented with an over-
lay tier of inexpensive low-power smaller cells
(picos, femtos, WiFi access points, integrated
WiFi-LTE modules, etc.) to enhance capacity by
moving infrastructure closer to the users in areas
with higher traffic demand.
Whereas the trend toward the use of WLAN
in conjunction with cellular networks has
emerged from the near-term need of operators
to relieve congestion in cellular networks, the
use of WiFi is expected to remain an integral
part of operators’ long-term strategy to address
future capacity needs. In the simplest case, no
cooperation between WiFi and cellular RANs is
available, and users are left to determine how
the two RATs are utilized. However, when WiFi
Figure 1. Topology and architecture of a converged heterogeneous network.
LTE macro
base station
Integrated
WiFi-LTE
small cell
eNB
MRCF
A. Application layer integration
Proprietary interface between
applications (application-specific)
B. Core-network-based integration
Existing solutions: SaMOG, IFOM, etc.
Semi-static ANDSF-based policy and
resource management
Mobility anchor at P-GW
(a)
(b)
WiFi
Indoor
Indoor
Outdoor
Outdoor
Pico BS
LTE-WiFi
LTE femto
access point
Integrated LTE-WiFi
base station
WiFi access point
Client cooperation
WiFi
LTE
Multi-radio UE
Application
Connection
manager
MRCF
WiFi
WiFi AP
EPC = Evolved packet core
Policy, charging
and control
Serving
gateway
AAA
server
MME
(mobility
mgmt)
HSS
PDN gateway
(home agent)
Standalone
(distributed APs)
WiFi
gateway
controller
LTE
Multi-RAT small cells allow for
easy cooperation across RATs
Multi-radio policy
ANDSF
RAN:
radio access network
MRCF: Multi-RAT coordination function
To core network
Offload to
Internet
Content
server
Direct
Internet
access
C. RAN-based integration
Dynamic radio resource management through
cross-RAT cooperation in the radio network
3GPP RAN anchored mobility and control: exploit
cellular link as anchor for improving handoff and
reliability/QoS of WLAN link
Operator
services (IMS,
Internet)
Whereas the trend
toward the use of WLAN
in conjunction with cellu-
lar networks has
emerged from the near-
term need of operators
to relieve congestion on
cellular networks, the
use of WiFi is expected
to remain an integral
part of operators’ long-
term strategy to address
future capacity needs.
ANDREEV_LAYOUT_Author Layout 12/18/14 4:21 PM Page 88
IEEE Wireless Communications • December 2014 89
is managed as part of an operator’s RAN, an
increased level of cooperation between WLAN
and 3GPP infrastructures may become feasible.
For instance, one may envisage an architec-
ture where integrated WiFi-LTE small cells
enable full cooperation between the two RATs,
allowing for WiFi to simply become a “virtual
carrier” anchored on the 3GPP radio network.
We note that multi-RAT small cells with co-
located WiFi and 3GPP interfaces are an emerg-
ing industry trend for lowering deployment costs
by leveraging common infrastructure across mul-
tiple RATs. However, given that such deploy-
ments are presently not common, current
standardization efforts aim to improve UE-cen-
tric interworking architectures while assuming
only limited cooperation or assistance across a
multi-RAT network.
OPTIONS FOR INTEGRATING WIFIWITH3GPP LTE
We continue by illustrating various architectural
choices for integrating WiFi and LTE networks
in Fig. 1b. These generally offer different mech-
anisms to implement important operations
required for multi-RAT integration, including
RAT discovery, RAT selection or assignment,
control of multi-RAT radio resource manage-
ment (RRM), protocols for inter-RAT mobility
or session transfers, and so on.
Application Layer Integration
— In Fig. 1b, case A
corresponds to the application or higher-layer
integration architecture. Accordingly, there is a
proprietary or higher-layer interface allowing the
UE and the content server to communicate
directly by exchanging information over multiple
RATs. As no coordination at the network layer
is involved, such solutions are typically simple
and have already been explored in the context of
improving over-the-top applications. This choice
of architecture is beneficial for boosting user
quality of experience (QoE), but it remains
largely application-dependent and may not fully
account for underlying network conditions, espe-
cially when such conditions vary dynamically.
Core-Network-Based Integration
— Furthermore,
case B summarizes recent solutions proposed by
3GPP for cellular/WLAN integration based on
interworking within the core network. Accord-
ingly, ANDSF assists in discovery of WiFi access
points and may also specify policies for network
selection, but the overall network selection deci-
sion remains in control of the UE. Therefore, it
can combine the local radio link state informa-
tion, operator policies, and user preferences to
make a decision that improves user QoE.
There are a number of benefits with this inte-
gration option, as it can more adequately
account for both operator policies and user pref-
erences. However, the performance of corre-
sponding control procedures may still be rather
limited. This is due to the fact that the UE may
only have local knowledge of the network condi-
tions and is thus likely to make greedy decisions,
ultimately hurting overall system performance.
Whereas the UE can be made to report its per-
ceived radio link state to the core network, such
information exchange cannot be updated dynam-
ically due to prohibitive levels of associated sig-
naling overhead. Hence, when wireless channel
conditions change dynamically, local RRM
directly on the RAN layer may deliver higher
QoS. Therefore, advanced architectures allowing
multi-RAT integration within the RAN are of
increasing interest today, as they employ net-
work-wide knowledge of radio link conditions.
RAN-Based Integration
— Finally, case C details the
emerging RAN-based 3GPP/WLAN integration
architecture. Here, UE assistance may facilitate
information exchange between cellular and
WLAN infrastructures; for that matter, a dedi-
cated interface may be introduced. The available
levels of cooperation within the RAN are con-
strained by the capacity of the intercell/inter-
RAT backhaul links. When high-capacity
backhaul or integrated multi-RAT small cells are
available, full cooperation across multiple RATs
may become available, thus enabling more
dynamic RRM for improved system and user
performance.
In addition, the cellular RAT may be
employed as a mobility and control anchor: a
user thus utilizes 3GPP protocols for transfer-
ring sessions to multi-radio small cells and then
uses local switching to steer sessions to/from
WLAN with low latency. The benefits of this
solution are obvious, as adaptations to dynamic
variations in interference conditions can easily
be performed without undesired session inter-
ruptions and packet drops. Furthermore, user
and operator preferences may be accounted for
through appropriate feedback by the UE or via a
suitable configuration of the RAN by the opera-
tors.
In summary, the degrees of cooperation with-
in the RAN can range from exploiting simple
assistance information (e.g., network loading) by
the radio network to tight coupling and joint/cen-
tralized RAN-based RRM. In what follows, we
describe the various levels of cross-RAT cooper-
ation options across a multi-RAT HetNet and
then characterize the associated performance
benefits. We pay particular attention to the
more practical case when only limited assistance
across multi-RAT network is available to users,
in contrast to significantly more complex net-
work-controlled approaches requiring higher sig-
naling and computation overheads.
ALGORITHMS FOR
RADIO RESOURCE MANAGEMENT
In what follows, we detail various options for uti-
lizing and managing multi-RAT radio resources
available in the network. Both user- and net-
work-controlled (or assisted) RRM may be con-
sidered for the range of architectural options
described above. For application- or core-net-
work-based integration (options A and B), only
UE-based RRM schemes may be feasible. A
richer set of choices is available for RAN-based
multi-RAT integration (option C), which
depends on the degree of inter-RAT coopera-
tion achieved with different RAN topologies.
Generally, RAN can play a major role in
multi-RAT resource management across the
HetNet. Even if RAN does not directly control
the RRM decisions, it may provide optimized
We pay particular atten-
tion to the more practi-
cal case when only
limited assistance across
a multi-RAT network is
available to users, in
contrast to significantly
more complex network-
controlled approaches
requiring higher signal-
ing and computation
overheads.
ANDREEV_LAYOUT_Author Layout 12/18/14 4:21 PM Page 89
IEEE Wireless Communications • December 2014
90
network assistance to enable better decisions by
the UE. In virtual RAN architectures, where the
mobility and control anchor is moved from the
core network to the RAN, more dynamic RRM
with fast session transfers between RATs
(dynamic switching) may become feasible. For
integrated multi-RAT small cells or where the
delay between the interfaces is negligible, tighter
cooperation involving joint RAT scheduling may
also be enabled.
We continue by introducing specific RRM
schemes that are investigated in our research.
They range from typical implementations used
by UEs today, where the UE always prefers to
connect to the less expensive WiFi network if it
is available (WiFi-preferred), to more intelligent
cross-RAT access network selection for con-
verged HetNets.
User-Centric Approaches
— The simplest threshold-
based algorithm serves as our baseline user-cen-
tric network selection scheme. With this solution,
a UE is continuously monitoring the signaling
messages from the neighboring WiFi access
points (APs) to obtain timely signal-to-noise
ratio (SNR) information. When a particular
SNR value exceeds a predefined threshold
(which we set equal to 40 dB as discussed in
3GPP), the user starts steering its traffic to the
respective WiFi AP. Otherwise, it keeps trans-
mitting on the LTE network (Fig. 2a).
Naturally, such behavior is an automatized
version of what a human user would do: whenev-
er a hotspot with a reliable signal is available,
UEs switch to WiFi to enjoy higher data rates
and reduce expenses associated with paid cellu-
lar traffic. Alternative user-centric algorithms
include schemes based on preferring WiFi if cer-
tain minimum performance (coverage, QoS, etc.)
is available, as well as solutions where the UE is
able to transmit on both RATs without any intel-
ligent coordination across them.
RAN-Assisted Approaches
— Due to its simplicity, the
baseline WiFi-preferred (SNR-threshold) scheme
may experience limitations in dense interference-
limited scenarios typical of modern urban deploy-
ments. For instance, a hotspot AP may experi-
ence overload conditions when a significant num-
ber of users try to steer their traffic through it.
Moreover, nomadic WiFi users, such as those
with laptops, could consume most of the WLAN
capacity. To make matters worse, the WiFi medi-
um access is contention-based, which results in
nonlinear degradation of the throughput perfor-
mance with increasing number of users.
Therefore, the load-agnostic SNR-threshold
scheme is not expected to remain effective in
environments with varying load. In such situa-
tions, UEs may attempt to combine SNR knowl-
edge with additional knowledge of the loading
information from the network infrastructure
(cellular/WLAN). While accounting for WiFi
load would certainly improve performance
beyond the SNR-threshold scheme, it is easy to
envision scenarios where accounting for WiFi
load only will be insufficient. Hence, we focus
our further investigation on schemes that
account for both LTE and WiFi loading, and
compare them with existing network-based
schemes that have been standardized in 3GPP
for small cell offload. Our proposed load-aware
scheme works as follows (see simplified time dia-
gram in Fig. 2b).
Throughput estimation: A user attempts to
listen on both interfaces in order to monitor the
SNR information in its neighborhood and esti-
mate its expected throughput. For WiFi, such
estimation is conducted based on predicted net-
work capacity divided by number of UEs con-
nected to a particular AP (as advertised by the
AP through the load indicators in the beacon
frames) as well as accounting for several weight-
ing factors (SNR, contention, etc.). The motiva-
tion behind the SNR weighting is to exclude APs
with low signal quality. Another coefficient may
account for the contention-based nature of WiFi
channel access and include signaling overheads
as well as collision losses. For LTE, throughput
prediction may simply be built on the scheduler
advertisements by a base station (BS or eNodeB)
and the used power control.
Randomization: A user may select the net-
work with the highest expected throughput value
Figure 2. Alternative network selection algorithms for HetNets.
(a)
WiFi AP
UE
(b)
WiFi AP eNodeBeNodeB
UE
(c)
WiFi AP eNodeB
Bias value
UE
AP/eNodeB SNR
measurements
SNR threshold
Yes
Yes
Yes
No
No
Broadcast load
measurements
Broadcast load
measurements
No
Switch
to/stay
at AP
Switch to new
AP/eNodeB
mi+=1
Stay at old
AP/eNodeB
Switch
to/stay
at eNodeB
AP/eNodeB SNR
measurements
*Et_new/Et_old-expected throughput on new/old interface (IF) ** Hysteresis is taken into account
Measurement window
AP/eNodeB SNR
measurements
RSSI_LTE< RSSI_WiFi+bias**
Yes N o
Switch
to/stay
at AP
Switch
to/stay
at eNodeB
AP/eNodeB SNR
measurements
Measurement window
Measurement window
Et_new>Et_old?*
Et_new>Et_old?*
rand(0..1)>pmi+1
ANDREEV_LAYOUT_Author Layout 12/18/14 4:21 PM Page 90
IEEE Wireless Communications • December 2014 91
probabilistically, rand(0..1) < pmi+1, where miis
the number of recent connections to this AP/BS
and pis the number in (0, 1) that represents the
reconnection probability. The proper use of p
reduces the number of concurrent reconnections
to the same AP/BS, which will prevent uncon-
trollable hopping from one interface to another.
If a network reselection occurs, miis increment-
ed for AP/BS i. Other users take this informa-
tion into account by dividing their expected
throughput value for this AP/BS by mi+ 1. This
allows for dynamic control of reselections on
both networks.
Hysteresis: To additionally decrease the num-
ber of cell border switchings, an appropriate hys-
teresis value should be added to the current
expected throughput value.
Filtering throughput estimations: Further
improvement in throughput estimates is obtained
through averaging. After each measurement win-
dow, the actual throughput obtained over this
period may be filtered with a moving average fil-
ter. The resultant value, which combines the
measured and predicted throughput, is then
used as the expected throughput value for this
AP/BS. This averaging is made to achieve more
reliability, which could suffer due to contention-
based channel access.
In summary, RAN-assisted approaches
employ network assistance from the RAN to
improve UE-based RAT selection decisions.
Network assistance can be very simple in that
the RAN may transmit certain assistance param-
eters (e.g., network load, utilization, expected
resource allocation), but with increased cross-
RAT cooperation, RAN assistance may also be
improved.
RAN-Controlled Approaches
— The above two net-
work selection schemes are user-centric in nature.
Hence, they may still result in suboptimal system-
wide performance, which may otherwise be
improved through network-based centralized
mechanisms. Consequently, RAN-controlled
approaches place the control of the RRM in the
radio network so that the BS can assign the UEs
to use certain RATs. Such network control may
be distributed across base stations, or may utilize
a central RRM entity that manages radio
resources across several cells/RATs.
Below we consider the conventional cell-
range extension schemes applied in cellular net-
works to steer users to small cells employing a
network-optimized received signal strength indi-
cation (RSSI) bias value. We use the RSSI bias
to increase/decrease the effective WiFi AP cov-
erage area depending on the network capacity
expectations. One limitation of this method is
that the optimal bias value needs to be adapted
based on network-wide knowledge of user distri-
bution. For example, our results show that the
optimal bias depends on the user deployment
model as well as the interference levels in the
network, which may not always be available as
typically WiFi cells may not have a direct inter-
face to a cellular BS. In what follows, we evalu-
ate RSSI-based cell range extension with bias
values optimized for the target scenario. We also
use hysteresis for the RSSI-based algorithm. The
time diagram of this method is shown in Fig. 2c.
More generally, network-controlled schemes
may utilize proprietary or standardized inter-
faces between cells/RATs. Distributed network-
controlled schemes have recently been discussed
as part of the 3GPP study on WLAN/3GPP RAN
interworking. Here, the network establishes cer-
tain triggers for UE to report measurement on
their local radio environment. The final RAT
selection decisions are then made by the 3GPP
BS based on UE measurement reports. Other
examples of centralized RAN-controlled archi-
tecture is the emerging dual connectivity, or
“anchor/booster” architecture, where the UE
always maintains a control link to the macrocell
tier, and the macrocell centrally manages the
user offload to smaller cells. Hence, the macro-
cell can centrally determine the optimal offload
mechanisms.
ANALYZING INTELLIGENT
ACCESS NETWORK SELECTION
In what follows, we concentrate on the impor-
tant problem of network selection between LTE
and WiFi RATs [11], assuming that WLAN is a
part of an operator deployed and managed
multi-RAT HetNet. We target feasible practical
extensions to improve the performance of UE-
centric network selection schemes. To be consis-
tent with current network deployments, we
consider distributed small cell overlay with
stand alone WiFi APs, assuming that there is no
interface between the WiFi and 3GPP radio net-
works [10]. Additionally, we discuss benefits of
deploying integrated WiFi-LTE small cells and
quantify the respective performance gains.
In particular, we investigate distributed RAT
selection schemes that account for network load-
ing information across the LTE and WiFi tech-
nologies, and compare them with solutions that
only rely on signal strength measurements. We
also benchmark the performance of UE-centric
RAT selection with optimized network-based
load balancing mechanisms. Intuitively, network-
centric solutions may seem to offer better per-
formance than UE-based approaches as
network-wide radio link information across users
can be employed to develop optimum RAT
assignment algorithms. However, with distribut-
ed architectures assuming no direct cooperation
between LTE and WiFi RATs, such solutions
may only be implemented through extensive UE
feedback, which could result in significant over-
heads. UE-centric RAT selection may also be
preferred as the UE can better account for user
preferences and application QoE.
SYSTEM-LEVEL
EVALUATION SCENARIO AND RESULTS
In the course of this study, we have developed
an advanced system-level simulator (SLS) that
mimics a complete LTE-WiFi system deploy-
ment compatible with the 3GPP LTE Release 10
and IEEE 802.11-2012 specifications. Presently,
neither free nor commercially available simula-
tion tools are readily applicable for evaluating
heterogeneous multi-RAT systems, as they lack
the necessary features, as well as the scalability to
In summary, RAN-assist-
ed approaches employ
network assistance from
the RAN to improve UE-
based RAT selection deci-
sions. Network
assistance can be very
simple in that RAN may
transmit certain assis-
tance parameters, but
with increased cross-RAT
cooperation RAN assis-
tance may also be
improved.
ANDREEV_LAYOUT_Author Layout 12/18/14 4:21 PM Page 91
IEEE Wireless Communications • December 2014
92
adequately capture the dependences between the
studied variables. In contrast, our SLS is a flexible
tool designed to support diverse deployment
strategies, traffic models, channel characteristics,
and wireless protocols. It comprises several soft-
ware modules modeling the deployment of wire-
less infrastructure and user devices, control events
related to transmission of signals between several
distinct types of transmitters and receivers,
abstractions for wireless channels, mechanisms for
collecting measurements, statistics for quantifying
system performance, and so on.
Below we construct a multi-RAT simulation
model representative of an urban deployment,
where WiFi small cells are overlaid on top of the
3GPP cellular network. Outdoor deployments
are considered based on recommendations in
[12]. A brief summary of the parameters is pro-
vided in Table 1. Specifically, we consider a
loaded (full-buffer) WiFi network with WLAN
APs uniformly distributed across the cellular
coverage area. Most UEs cluster around the
APs, which recreates a hotspot area (airport,
restaurant, shopping mall, or university campus)
with many bandwidth-hungry users loading the
WiFi network. Moreover, around one third of
UEs are still deployed uniformly across the cel-
lular network mimicking regular mobile users.
While this scenario may not be characteristic of
all practical urban conditions, it represents a
harmonized 3GPP vision of a characteristic Het-
Net deployment.
The major expected outcome of leveraging
WiFi small cells is efficient offloading of cellular
user traffic resulting in significant user benefits.
For that reason, our primary metric of interest is
the uplink UE throughput (in contrast to many
existing studies concentrating on downlink per-
formance), which, in turn, determines the overall
system capacity. The cumulative distribution
function (CDF) of individual user throughput
comparing performance between SNR-threshold
(WiFi-preferred) and our proposed load-aware
(RAN-assisted) scheme is shown in Fig. 3a. The
results indicate that the load-aware scheme gives
visible benefits at the cell edge (e.g., over 75 per-
cent of improvement is observed in the 5 percent
quantile), as well as some improvement in the
average throughput for integrated deployments
(i.e., with co-located WiFi-LTE interfaces).
Energy efficiency (EE) is also becoming
increasingly important for 5G wireless systems
due to the limited battery resources of mobile
clients [13], and we confirm significant gains in
bits per Joule for both distributed (19 percent)
and integrated (29 percent) scenarios in Fig. 3b.
Furthermore, as QoS may be equally important,
we also account for fairness between the users,
which indicates how large the deviation between
actual user throughput and the cell-average per-
formance is. In terms of fairness, Jain’s index
(table, Fig. 3b) of the load-aware scheme
(0.72/0.63) is also higher than that for the SNR-
threshold scheme (0.65/0.54). The stability of
UE-centric schemes is another very important
aspect of UE-centric RAT selection, as excessive
ping-ponging between RATs is undesirable due
to the overhead and latency of mobility proto-
cols as well as EE considerations. In Fig. 3b (see
table), we additionally report the number of cel-
lular/WLAN reconnections (in number of recon-
nections per second) and employ hysteresis
mechanisms (an optimized 3 dB value has been
used in our experiments) to improve perfor-
mance.
We also account for the performance of an
optimized cell range extension (RAN-controlled)
scheme based on RSSI bias, where the network-
wide optimization is expected to result in
improved performance. The main feature of the
considered cell range extension scheme is that it
increases the effective WiFi/LTE small cell
radius with respect to the bias level. This could
work well in the scenario with uniformly
deployed UEs, but in the clustered case the
interference between WiFi users needs to be
considered as well, which is what our load-aware
scheme does explicitly. To this end, we perform
Figure 3. Comparing SNR-threshold (WiFi-preferred) and our load-aware (RAN-assisted) network selection schemes
0.50
5
0
10
15
1 1.5 2 2.5
Individual throughput, Mb/s
5
(a)
0
20
0
CDF, %
40
60
80
100
10 15 20 25
SNR-threshold, distributed
SNR-threshold, integrated
Load-aware, distributed
Load-aware, integrated
SNR-threshold, distributed
SNR-threshold, integrated
Load-aware, distributed
Load-aware, integrated
Individual energy efficiency, Mb/J
5
(b)
0
20
0
CDF, %
40
60
80
100
10 15
SNR_thr
distributed 4.21 0.91 0.65 0.28 2.72
Tpt,
mean
Tpt,
5%
Jain’s
index
No. re-
connect
EE,
mean
Load-aware
distributed 4.20 1.60 0.72 0.45 3.26
SNR_thr
integrated 5.36 0.69 0.54 0.28 2.95
Load-aware
integrated 5.69 1.30 0.63 0.45 3.81
ANDREEV_LAYOUT_Author Layout 12/18/14 4:21 PM Page 92
IEEE Wireless Communications • December 2014 93
optimization of small cell offloading bias based
on network-wide knowledge of user distribution
in Fig. 4a.
However, from Fig. 4b we learn that even
with a network controlled bias value (the opti-
mal value of 14 dB is chosen), the individual
user throughput is very close to that in the load-
aware case (and even smaller at the cell edge).
In more detail, Fig. 4b (see bar chart) also high-
lights the average percentage of time spent by
users on each interface. It may be seen from our
results that the load-aware scheme is effective in
utilizing the available WLAN capacity while effi-
ciently balancing capacities across the LTE
macro and pico tiers.
ANALYTICAL SPACE-TIME METHODOLOGY
FOR CONVERGED HETNETS
The above performance results addressed loaded
multi-RAT HetNets, but such networks may also
be substantially underutilized during off-peak
hours. Hence, the load on a heterogeneous
deployment can vary significantly, and it is cru-
cial to capture network dynamics explicitly when
modeling HetNet performance. However, given
the associated complexity, dynamic systems have
not been studied as broadly as their static coun-
terparts with a fixed set of active users. Conse-
quently, our proposed analytical methodology
suggests assessing flow-level network perfor-
mance enabling user, traffic, and environment
dynamics.
Recently, we have made progress along these
lines and have results demonstrating that the
locations of the network users relative to each
other highly impact the overall system perfor-
mance [14]. Indeed, given that users are not reg-
ularly spaced, there may be a high degree of
spatial randomness that needs to be captured
explicitly. We thus adopt a range of random spa-
tial models where user locations are drawn from
a particular realization of a random process.
Coupling such topological randomness with sys-
tem dynamics requires a fundamental difference
in characterizing user signal power and interfer-
ence. Fortunately, the field of stochastic geome-
try provides us with a rich set of powerful results
and analytical tools that can capture the net-
work-wide performance of a random user
deployment [15].
More specifically, every data flow in a dynam-
ic network may generally represent a stream of
packets corresponding to a new file transfer,
web-page browsing, or real-time voice/video ses-
sion. As an example, consider an isolated cell of
a macro network with radius Rencompassing a
macro BS together with several distributed pico
BSs and WLAN APs. All the BSs/APs are capa-
ble of serving uplink data from their wireless
users concurrently. The considered traffic is
characteristic of real-time sessions with some
target bit rate. Based on the recent 3GPP speci-
fications, we further assume non-overlapping
frequency bands for all three tiers. However, all
WLAN/pico links share the frequency bands of
their respective tiers and thus interfere, whereas
the macro tier may be considered interference-
free (with appropriate intercell power control).
To explicitly model topological randomness
in our network (Fig. 5a), we employ several
stochastic processes and, to this end, adopt a
number of simplifications based on a Poisson
point process (PPP). The key novelty of this
approach is that we consider a space-time PPP
with the rate function L(x, t), where xŒR2is the
spatial component, and tŒR+is the time com-
ponent. While random network topology is the
primary focus of our model, we also couple it
with flow-level system dynamics. This involves an
appropriate queuing model, where the session
arrives and leaves the system after being served
(the service time is determined by the random
session length). When a new session arrives or a
served session leaves the system, the centralized
assisting entity in the RAN performs admission
and power control on all tiers by deciding
whether the session would be admitted to a par-
ticular tier or not and/or advising on the users’
transmit powers.
Our general system model is illustrated in
Table 1. Important simulation parameters. ITU: International Telecommu-
nication Union.
Parameter Value
LTE/WiFi configuration 10 MHz FDD/20 MHz
Macrocell layout 7 cells, 3 sectors each
LTE signaling mode 2 out of 20 special subframes,
short CP, 10 ms frame
Inter-site distance (ISD) 500 m
LTE macro antenna configuration 1 ×2 (diversity reception)
UE to eNodeB/pico/AP pathloss ITU UMa/UMi
eNodeB antenna gain 14 dB
eNodeB/AP/UE maximum power 43/20/(23/20 LTE/WiFi) dBm
LTE power control Fractional (a= 1.0) [12]
WiFi power/rate control Max-power/ARQ
UE/eNodeB/AP antenna height 1.5/25/10 m
UE noise figure/feeder loss 9 dB/0 dB
Feedback/control channel errors None
Traffic model Full-buffer
Number of UEs/APs 30/4 per macrocell (3 sectors)
AP/UE deployment type Uniform/clustered (4b in [12])
AP/UE-eNodeB, AP/UE-UE distance > 75/35 m, 40/10 m
WiFi MPDU 1500 bytes
Modeling time 3 s
Number of trials per experiment 30
ANDREEV_LAYOUT_Author Layout 12/18/14 4:21 PM Page 93
IEEE Wireless Communications • December 2014
94
Fig. 5b representing areas of the macro, pico,
and WLAN tiers together with the correspond-
ing users and infrastructure nodes. We consider
the following cascade network selection when a
new session arrives into the system. First, the
RAN-based network selection assistance entity
attempts to offload the newly arrived session
onto the nearest WLAN AP by performing cen-
trally managed WLAN admission control. If the
session is accepted on the WLAN tier, it is
served there without interruption until it suc-
cessfully leaves the system. Otherwise, if this ses-
sion cannot be admitted onto the WLAN, the
pico network admission control is executed, and
either the session is accepted on the pico tier
and served by the nearest pico BS, or the macro
network itself attempts to serve this session.
Eventually, if the session cannot be admitted
onto the macro tier either, it is considered per-
manently blocked and leaves the system
unserved.
In Fig. 5c, we detail the overall blocking
probabilities for the converged HetNet as well as
for the three tiers in, macro, pico, and WLAN,
individually. Our main observation is that with
two additional overlay tiers, HetNet perfor-
mance improves significantly over what can be
achieved in the macro-only networks (cellular
baseline). Remarkably, we actually witness visi-
ble performance improvement even with only a
few additional infrastructure nodes, such as two
WLAN APs and two pico BSs in this example.
Therefore, we believe that multiple RATs and
the associated network selection intelligence for
their efficient use will become a characteristic
feature of future 5G HetNets.
MAIN TAKEAWAYS AND THE
WAY FORWARD
In summary, this article reviews major challenges
in delivering uniform connectivity and service
experience to future heterogeneous 5G net-
works. It discusses several architecture choices
and associated algorithms for intelligent access
network selection in multi-RAT HetNet deploy-
ments, both when the control of how radios are
utilized rests with the network and when it rests
with the user. In particular, it compares simulat-
ed performance of RAN-assisted load-aware
network selection schemes with the convention-
al/existing UE- and network-based solutions
employed in current systems. We primarily focus
on uplink performance as it has not been fully
addressed in past literature.
The main advantages of load-aware schemes
stem from the fact that the SNR-threshold
(WiFi-preferred) scheme, as well as the network-
centric cell range extension scheme, do not
explicitly account for the loading and interfer-
ence on WiFi APs typically encountered in clus-
tered UE deployments. Our results show that
the load-aware user-centric scheme, which aug-
ments SNR measurements with additional infor-
mation about network load, could improve the
performance of a WiFi-preferred scheme based
on minimum SNR threshold. We observed over
75 percent improvement in 5 percent cell edge
throughput as well as significant gains in energy
efficiency for both distributed and integrated
deployment scenarios.
Comparison with more advanced network-
controlled schemes has also been completed
across various heterogeneous deployments to
confirm attractive practical benefits of dis-
tributed user-centric solutions. Next steps
include further investigation of UE-based algo-
rithms while explicitly considering load varia-
tion in the network and accounting for
application-layer statistics. System behavior in
the presence of uncoordinated (rogue) WiFi
interference must also be accounted for, and
the hysteresis mechanism may further be
improved to combat the uncertainty in estimat-
ing user throughput.
Building on the system-wide simulation data,
we also propose a novel dynamic methodology
for RAN-assisted network selection capturing
the spatial randomness of HetNets together with
Figure 4. Comparing our load-aware (RAN-assisted) and RSSI-based (RAN-controlled) network selection schemes.
Bias value, dB
-5-10
1
0
Individual throughput, Mb/s
2
3
4
5
6
7
0 5 10 15 20 25 30
Individual throughput, Mb/s
0
(b)(a)
20
0
CDF, %
40
60
80
100
5 10 15 20 25
RSSI-based, distributed
RSSI-based, integrated
Load-aware, distributed
Load-aware, integrated
RSSI-based0
10
0
Percentage of time
20
30
40
50
60
LTE macro
LTE pico
WiFi
Load-aware
Mean, distributed scenario
Mean, integrated scenario
50%, distributed scenario
50%, integrated scenario
ANDREEV_LAYOUT_Author Layout 12/18/14 4:21 PM Page 94
IEEE Wireless Communications • December 2014 95
unsaturated uplink traffic from its users. Our
stochastic-geometry-based analysis enables in-
depth characterization of dynamic interactions
between macro and pico cellular networks, as
well as WLAN, mindful of user QoS and based
on intelligent RAT selection/assignment. Going
further, we expect our space-time methodology
to be capable of encompassing other technolo-
gies beyond LTE and WiFi, as well as additional
use cases beyond simple aggregation of capacity
across unlicensed bands.
More generally, studying the ultimate capac-
ity of multi-radio wireless networks remains an
open problem in the field of information theo-
ry, and stochastic geometry has the potential to
shed light on it given that it can explicitly cap-
ture new interference situations and hence the
achievable data rates. This challenging objec-
tive may require novel advanced analytical tools
to interconnect and apply techniques and meth-
ods coming from the area of point processes,
probability theory, queuing theory, and percola-
tion theory, as well as modern engineering
insights.
ACKNOWLEDGMENT
This work is supported by Intel Corporation,
GETA, and the IoT SRA program of Digile,
funded by Tekes. The work of the first author is
supported with a Postdoctoral Researcher grant
by the Academy of Finland, as well as with a
Jorma Ollila grant by Nokia Foundation.
Figure 5. Illustration of the proposed space-time methodology for multi-RAT networks.
Arrival rate, s-1
20
0.2
0
Blocking probability
0.4
0.6
0.8
1
4 6 8 10 12
Distance, m
500
5
0
Throughput, Mb/s
10
25
30
20
15
100 150 200 350 400300250
Blocked
(b)
(a)
(c)
WLAN user
WLAN user
Arrived user
Pico user
Macro user Macro user
Macro user
Macro user
Macro user
Pico user
Arrived user
Arrived user
WLAN
access point
WLAN
access point Transmission
Interference
WLAN
access point
Pico
base station
Pico
base station
Pico
base station
Pico blocking probability
WLAN blocking probability
Macro blocking probability
Total blocking probability
Macro (baseline)
LTE
WiFi
Macro
base station
Blocked
Session blocked
Session served
ANDREEV_LAYOUT_Author Layout 12/18/14 4:21 PM Page 95
IEEE Wireless Communications • December 2014
96
REFERENCES
[1] B. Bangerter et al., “Networks and Devices for the 5G
era,” IEEE Commun. Mag., vol. 52, 2014, pp. 90–96.
[2] Cisco Visual Networking Index: Global Mobile Data Traf-
fic Forecast Update, 2013–2018, Feb. 2014.
[3] S.-P. Yeh et al., “Capacity and Coverage Enhancement
in Heterogeneous Networks,” IEEE Wireless Commun.,
vol. 18, 2011, pp. 32–38.
[4] P. Marsch et al., “Future Mobile Communication Net-
works: Challenges in the Design and Operation,” IEEE
Vehic. Tech. Mag., vol. 7, 2012, pp. 16–23.
[5] J. Andrews et al., “An Overview of Load Balancing in
HetNets: Old Myths and Open Problems,” IEEE Wireless
Commun., vol. 21, no. 2, 2014, pp. 18–25.
[6] S. Andreev et al., “Cellular Traffic Offloading onto Net-
work-Assisted Device-to-Device Connections,” IEEE
Commun. Mag., vol. 52, no. 4, 2014, pp. 20–31.
[7] R. Baldemair et al., “Evolving Wireless Communications:
Addressing the Challenges and Expectations of the Future,”
IEEE Vehic. Tech. Mag., vol. 8, 2013, pp. 24–30.
[8] J. Andrews, “Seven Ways that HetNets Are a Cellular
Paradigm Shift,” IEEE Commun. Mag., vol. 51, 2013,
pp. 136–44.
[9] M. Bennis et al., “When Cellular Meets WiFi in Wireless
Small Cell Networks,” IEEE Commun. Mag., vol. 51,
2013, pp. 44–50.
[10] 3GPP TR 37.842, “3GPP/WLAN RAN Interworking Study
Item Report,” 2013.
[11] L. Wang and G.-S. Kuo, “Mathematical Modeling for
Network Selection in Heterogeneous Wireless Networks
— A Tutorial,” IEEE Commun. Surveys & Tutorials, vol.
15, 2013, pp. 271–92.
[12] 3GPP TR 36.814, “Further Advancements for E-UTRA
Physical Layer Aspects,” 2010.
[13] S. Navaratnarajah et al., “Energy Efficiency in Hetero-
geneous Wireless Access Networks,” IEEE Wireless Com-
mun., vol. 20, 2013, pp. 37–43.
[14] O. Galinina et al., “Capturing Spatial Randomness of Het-
erogeneous Cellular/WLAN Deployments with Dynamic
Traffic,” IEEE JSAC, vol. 32, no. 6, 2014, pp. 1083–99.
[15] H. ElSawy et al., “Stochastic Geometry for Modeling,
Analysis, and Design of Multi-Tier and Cognitive Cellu-
lar Wireless Networks: A Survey,” IEEE Comm. Surveys
& Tutorials, vol. 5, 2013, pp. 996–1019.
BIOGRAPHIES
SERGEY ANDREEV is a senior research scientist in the Depart-
ment of Electronics and Communications Engineering at
Tampere University of Technology (TUT), Finland. He
received his Specialist degree (2006) in information security
and his Cand.Sc. degree (2009) in wireless communications
from St. Petersburg State University of Aerospace Instru-
mentation, Russia, as well as his Ph.D. degree (2012) in
technology from Tampere University of Technology, Fin-
land. He has (co-)authored more than 80 published
research works. His research interests include wireless com-
munications, energy efficiency, heterogeneous networking,
cooperative communications, and machine-to-machine
applications.
MIKHAIL GERASIMENKO is a research assistant at TUT in the
Department of Electronics and Communications Engineer-
ing. He received his Specialist degree from St. Petersburg
University of Telecommunications in 2011. In 2013 he
obtained his M.Sc. degree from Tampere University of
Technology. He started his academic career in 2011 and in
recent years he appeared as a (co)-author on several scien-
tific journal and conference publications, as well as several
patents. Moreover, he also acted as a reviewer and partici-
pated in education activities. His main subjects of interest
are wireless communications, machine-type communica-
tions, and heterogeneous networks.
OLGA GALININA is a Ph.D. candidate in the Department of
Electronics and Communications Engineering at TUT. She
received her B.Sc. and M.Sc. degrees in applied mathemat-
ics from the Department of Applied Mathematics, Faculty
of Mechanics and Physics, St. Petersburg State Polytechni-
cal University. She has published work on mathematical
problems in the novel telecommunication protocols in
internationally recognized journals and high-level peer-
reviewed conferences. Her research interests include
applied mathematics and statistics; queueing theory and
its applications; wireless networking and energy efficient
systems, and machine-to-machine and device-to-device
communication.
YEVGENI KOUCHERYAVY is a full orofessor and laboratory
director at the Department of Electronics and Communica-
tions Engineering of TUT. He received his Ph.D. degree
(2004) from TUT. He is the author of numerous publica-
tions in the field of advanced wired and wireless network-
ing and communications. His current research interests
include various aspects of heterogeneous wireless commu-
nication networks and systems, the Internet of Things and
its standardization, as well as nanocommunications. He is
an Associate Technical Editor of IEEE Communications Mag-
azine and an Editor of IEEE Communications Surveys and
Tutorials.
NAGEEN HIMAYAT is a senior research scientist with Intel
Labs, where she performs research on broadband wireless
systems, including heterogeneous networks, cross-layer
radio resource management, MIMO-OFDM techniques, and
optimizations for M2M communications. She has over 15
years of research and development experience in the tele-
com industry. She obtained her B.S.E.E. from Rice Universi-
ty and her Ph.D. in electrical engineering from the
University of Pennsylvania in 1989 and 1994, respectively.
SHU-PING YEH is a research scientist in the Wireless Commu-
nications Laboratory at Intel. She received her M.S. and
Ph.D. from Stanford University in 2005 and 2010, respec-
tively, and her B.S. from National Taiwan University in
2003, all in electrical engineering. Her current research
focus includes interference mitigation in multitier networks
utilizing multi-antenna techniques, machine-to-machine
communications, and interworking of multiple radio access
technologies within a network.
SHILPA TALWAR is a principal engineer in the Wireless Com-
munications Laboratory at Intel, where she is conducting
research on mobile broadband technologies. She has over
15 years of experience in wireless. Prior to Intel, she held
several senior technical positions in the wireless industry.
She graduated from Stanford University in 1996 with a
Ph.D. in applied mathematics and an M.S. in electrical
engineering. She is the author of numerous technical publi-
cations and patents.
Our stochastic-geometry-
based analysis enables
in-depth characterization
of dynamic interactions
between macro and pico
cellular networks, as
well as WLAN, mindful
of user QoS and based
on intelligent RAT selec-
tion/assignment.
ANDREEV_LAYOUT_Author Layout 12/18/14 4:21 PM Page 96
