Prediction Based Bandwidth Allocation for Cognitive

LTE Network

Alia Asheralieva

University of Newcastle

School of Electrical Engineering and Computer Science

Callaghan NSW, Australia

alia.asheralieva@uon.edu.au

Abstract— In this paper we present a novel dynamic

bandwidth allocation technique in which different base stations

share the total available spectrum to maximize the quality of

service (QoS) in the network, and show the implementation of

this technique in a cognitive 3rd Generation Partnership Project

Long Term Evolution (3GPP LTE) network. Assuming, that each

base station is characterized by a concave increasing utility and a

positive weight, we conduct a weighted utility maximization

framework, and develop a simple prediction-based bandwidth

allocation algorithm. To deal with heterogeneous network

applications we propose to deploy the approach used in optimal

flow and congestion control (OFC) where the resources are

assigned based on speed of load increase. Using the appropriate

load indictors, the algorithm first identifies the base stations with

increasing (decreasing) load, and then decrease (increase) the

channel utilization of base stations with increased (decreased)

load using weighted proportional fairness criterion.

Keywords— Cognitive Radio Network; LTE; dynamic spectrum

access; traffic prediction

I. INTRODUCTION

Traditional fixed spectrum allocation policy has lead to the scarcity of spectrum and reduced quality of service (QoS) for end-to-end applications. To address this problem, in 2004 a special IEEE working group was set to develop a new 802.22 cognitive radio (CR) standard. It was proposed that the fixed wireless access will be provided by a Wireless Regional Area Network (WRAN) comprising a number of service providers (SPs) with their base stations. SPs will share the total available spectrum among each other to maximize the QoS for their users [1].

Motivated by this concept of a cognitive radio architecture, in this paper we develop a dynamic bandwidth allocation technique in which different base stations share the total available spectrum to maximize the QoS in the network, and implement this techniques in a 3rd Generation Partnership Project Long Term Evolution (3GPP LTE) network. In particular, we assume that each base station (evolved NodeBs) in the network has a concave increasing utility and is assigned a positive weight based on predicted speed of increase of its loads. Based on this assumption, we conduct a weighted utility maximization framework, and derive a simple prediction-based bandwidth allocation algorithm (PRA). To prevent long delays and large losses in a network, PRA first identifies the evolved

NodeBs (eNBs) with increasing (decreasing) load using appropriate load indicator; and then decrease (increase) the channel utilization of eNBs with increased (decreased) load using weighted proportional fairness criterion. The values of load indicators in algorithm are obtained from the following medium access control (MAC) and physical (PHY) layer information: speed of load increase, loss and spectral efficiency of the wireless channels between the users and eNB.

The rest paper is organized as follows. In Section II we outline related work in the domain of dynamic spectrum allocation for CR networks and give a brief description of our contribution in this area. In Section III we present the bandwidth allocation algorithm, prediction technique and optimization criteria used in the network model. In Section IV we provide a detailed performance analysis of the proposed bandwidth allocation algorithm using simulation model developed using the OPNET package [2]. The conclusions are presented in section V.

II. RELATED WORK AND OUR CONTRIBUTION

Firstly introduced with the development of a software defined radio [3], the concept of CR is now considered as a promising way to solve the problem of the spectrum scarcity. CR network has been defined as an intelligent system that has ability to perceive its environment, and then to learn and adapt to the current network conditions [4]. In this context, a CR network should use a so-called cognition cycle, which includes spectrum sensing, channel state estimation, predictive modeling, dynamic spectrum management and power control [1]. A specific area of a CR field, dynamic spectrum access and resource allocation is attracting a growing research interest during the last decade. Consequently, a number of centralized and distributed algorithms have been proposed to solve the problem of dynamic spectrum access and resource allocation in CR networks [5-13]. In these algorithms each user is characterized by its concave increasing utility function, representing the degree of user satisfaction in the network [14].

Most of the existing resource allocation strategies for CR networks have been deployed for homogeneous scenarios, and not very efficient in case of heterogeneous network applications. This is due to the fact that all users in the network are characterized by similar utility functions (for instance, user throughput or service rate). Besides, these algorithms are based on current load and channel state information, and do not use

978-1-4673-5939-9/13/$31.00 ©2013 IEEE978-1-4673-5939-9/13/$31.00 ©2013 IEEE

2013 IEEE Wireless Communications and Networking Conference (WCNC): MAC2013 IEEE Wireless Communications and Networking Conference (WCNC): MAC

801

prediction for resource allocation, which also may reduce the efficiency of these schemes (especially for delay-sensitive applications, such as voice and video). Therefore, in this paper we suggest another approach for resource allocation in CR networks. First of all, to deal with heterogeneous network applications we propose to deploy the approach used in optimal flow and congestion control (OFC), where the resources are assigned based on speed of load increase [15, 16]. In this way we will be able to prevent the growth of user queues and minimize the delay for real-time network applications (such as voice and video). Unlike most of the congestion and flow control algorithms where the users are described by simple binary load indicators (congested or not congested node) [15, 16], we use the modified load indicators (MLI) described by more complex functions depending on queue size and the loss in the nodes. To further increase the algorithm efficiency, we allocate the resources based on predicted load and channel state information, which may help to prevent potential delay and loss for the users. Thus, our algorithm deploys a cognition cycle [1] which consists of the collection of the necessary traffic parameters, prediction and dynamic resource allocation.

III. NETWORK MODEL

A. PRA Algorithm

The architecture of the proposed LTE network model is shown in Figure 1. In this architecture n eNBs are connected to the network resource manager (NRM) via Internet Protocol (IP) based links. The NRM is located in the LTE Evolved Packet Core (EPC) and acts both as a gateway for external networks and as a resource controller. The eNBs share the total available bandwidth C via the NRM according to a PRA algorithm characterized by the centralized network control, dynamic bandwidth sharing and prediction-based allocation. To prevent long delays and large losses in a network, the PRA algorithm first identifies the eNBs with increasing (decreasing) load using appropriate load indicator value; and then decrease (increase) the channel utilization of eNBs with increased (decreased) load using weighted proportional fairness optimization (the weights of eNBs are equal to respective load indicators).

External Networks

EPC

eNB1

eNBi

eNBn

NRM

UE1

UE3UE2

UE2

UE1 UE3

UE1UE3

UE2

Fig.1. Proposed network architecture

The following assumptions are made in the proposed network model: 1) uplink and downlink buffers of eNBs have known finite capacities denoted by Q

ULmax and Q

DLmax; 2)

monitoring of parameters, bandwidth allocation and prediction can be performed only discontinuously within fixed time intervals (called the monitoring interval in the paper); 3) the length of the monitoring interval, the length of buffers and the amount of data arrived at the buffers during each monitoring

interval are known; 4) data collected and predicted separately for uplink and downlink directions.

In the algorithm the time axis is partitioned into mutually disjoint intervals {[kΔt, (k+1)Δt]}, k = 0, 1, 2,…. The prediction and resource allocation approach is synchronized with these intervals. At the beginning of each interval [kΔt, (k+1)Δt] the eNBs are allocated some bandwidth based on prediction made from the data collected in the past. In particular, we shall use the following notation: n– the number of nodes (eNBs) in the network; t – integer valued index of a monitoring interval; xi(t) – bandwidth (in MHz) allocated to eNB i at t

th monitoring interval; Qi

UL(t), Qi

DL(t), Qi (t) –length

of the queue (in bits) at eNodeB i at the beginning of the tth

monitoring interval on the uplink, the downlink and the unspecified (general) channels, respectively; Xi

UL(t), Xi

DL(t),

Xi(t) – the number of bits served at eNB i during the monitoring interval t on uplink and downlink, and unspecified (general) channels; Li

UL(t), Li

DL(t), Li(t) – the number of bits arrived to

eNB i at a monitoring interval t on uplink and downlink, and unspecified (general) channels; Di

UL(t), Di

DL(t), Di(t) – the

number of bits dropped at eNB i during the monitoring interval t on uplink and downlink, and unspecified (general) channels.

The number of bits served at eNB i during the monitoring interval t on uplink and downlink direction can be found from the spectral efficiencies of the uplink and downlink LTE channels SE

UL, SE

DL (in bits per second (bps)/Hz) using

expression:

DL

iDL

iUL

iUL

iSE

ttxtX

SE

ttxtX

)()( ,

)()(

The proposed PRA algorithm can be described as follows: Input: The eNBs monitor n

i

DL

i

UL

itQtQ

1)}1(),1({

and

n

i

DL

i

UL

itLtL

1)}1(),1({

for 1 ≤ i ≤ n, and send this information to

the NRM using the S1 interface. Prediction: Using the input information collected up to t

th

monitoring interval, the NRM computes the prediction )1(ˆ tL

of )1( tL . The details of the prediction algorithm are given in

Section III.B. Weight generation: Based on the predicted load values

n

i

DL

i

UL

itLtL

1)}1(ˆ),1(ˆ{

and buffer lengths for each eNB the NRM

generates the weights

)(

)()1(ˆ)(

)(

)()1(ˆ)(

tQ

SE

ttxtLtQ

tQ

SE

ttxtLtQ

DL

i

DL

iDL

i

DL

i

UL

i

UL

iUL

i

UL

i

i

where [x]+ denotes the max(0, x). These weights are used later

in weighted proportional fairness problem. The justification behind using the expression (2) is given in Section III.C. Optimization: Using the generated weights ωi, the NRM calculates the bandwidth xi(t+1), that will be allocated to each eNB in the next monitoring interval [t+1, t+2] based on the weighted proportional fairness criterion, and transmit these values to corresponding eNBs via the S1 interface. The details are given in Section III.C.

802

Output: Received bandwidth allocation values xi(t+1) are assigned to the physical interfaces of the corresponding eNBs at the beginning of the (t + 1)

th monitoring interval.

Since the sizes of the node buffers are limited, the length of node buffer Q(t) is a nonlinear function of t, which can be readily verified from the Lindley’s equation [17] for a finite buffer length given by:

)1()1()1()()1()1()(,min

)(

maxtDtXtLtQtXtLtQQ

ttQ

where Qmax is the maximal buffer size of the queue. This step makes Q(t) inappropriate to use in the linear prediction. Therefore, instead of predicting the nonlinear parameter Q(t), we predict the traffic arrived to each eNB, L(t).

B. Prediction Technique

The amount of data which will arrive at the buffers of eNBs at the next monitoring interval are predicted separately for the uplink and the downlink of each eNB using the recursive least squares (RLS) algorithm [18] with exponential forgetting applied to order p autoregressive time-series model [18]

TT

)1()1()()1(

),()()1()1(

ptLtLtLt

tetttL

The RLS algorithm has been chosen as a real-time, adaptive algorithm, suitable for tracking time-varying parameters of a bursty traffic source in a wireless network. This algorithm is also very simple, has a low dependency on priory data and modest memory requirements [18]. To generate the prediction of )1(ˆ tL using the RLS technique, the unknown

time-varying parameter θ is estimated recursively according to the algorithm presented below [18]:

).1()1()(ˆ)1(ˆ

),(ˆ)1()1()1(

),1()1()1(

,)1()()1(

)()1()1()()(

1)1(

tetKtθtθ

tθttLte

ttPtK

ttPt

tPtttPtPtP

T

T

T

Here K is the gain vector showing how much the value of e will modify different elements of θ. Vector e is the prediction error, and λ is a forgetting factor used to discount the previous measurements. If it is necessary to make prediction algorithm more adaptive in case the characteristics of arrival process at nodes of the network have high variation over time, a smaller is λ is used. This increases the rate at which the past information is forgotten (usually one takes λ between 0.95 and 1) [18]. Without any priori information the initial values of θ and P are

commonly set as: ,)0( ,0)0(ˆ ρIPθ where ρ is a “big” number

[18]. The AR(p) model, deployed for parameter estimation

generates smallest prediction error among other time-series models when tested in bursty traffic environments. The accuracy of the RLS prediction using the AR(1) model for “real life” traffic traces has been demonstrated in [19, 20]. In the appropriate order of the autoregressive part, p = 1 was

chosen by minimizing the Akaike Information Criterion (AIC) given by [21]

)(1

,2)(log 1

2 teN

VpθVNAICN

t

NNN

where VN is a loss function, N represents the total number of recursions [19].

C. Weight Generation

To prevent long packet delays and buffer overflows in the network the PRA algorithm should be able to 1) identify the nodes with increasing (decreasing) load (length of buffers); 2) increase (decrease) the channel utilization of the nodes with decreasing (increasing) load. These can be achieved by applying the weighted proportional fairness criterion proposed by Kelly in [22] for bandwidth allocation in the network as described below.

Consider a network model as a set of n eNBs, each characterized by some positive load indicator ωi, and sharing the total available bandwidth C. Let xi(t+1) be the bandwidth allocated to node i, at t

th monitoring interval. For simplicity we

write xi = xi(t+1) in this section. Then a weighted proportionally fair allocation should maximize the weighted sum of logarithmic bandwidth allocations:

Cxxg

nixxg

xxf

n

i

in

ii

n

i

ii

1

1

1

0:)(

1 ,0:)( subject to

log)( maximize

It is clear, that the problem presented in (7) has a unique optimum, because the objective here is an increasing, strictly concave, and continuously differentiable function over a convex feasibility region. The solution of the problem will give such allocation in which the bandwidth of the nodes will be assigned proportionally to their load indicators.

The main question now is how to derive a suitable load indicator, which can be used to detect the nodes with increasing (decreasing) load. For instance, in congestion avoidance and control algorithm proposed by Jacobson in [15], the load in different nodes were chosen based on the assumption that in an uncongested node the length of buffer is non-increasing as:

1 ,)(

)1(

i

i

i

itQ

tQ while in congested node the length of buffer

experiences a multiplicative increase as: 1 ,)(

)1(

i

i

i

itQ

tQ .

Since the bandwidth assigned to eNBs will affect both uplink and downlink channels, the node weights ωi should take into account both uplink and downlink load indicators γi

UL and γi

DL,

can be assigned as .DL

i

UL

ii

However, the load indicator γi used in [14] can only be applied for the nodes with infinite buffers. Applying the same algorithm in the nodes with finite buffer size might fail to detect the overload situation. Indeed, if at some point the buffer of a eNB reaches its maximum value Qmax, and does not decrease over the time, then the load indicator γi will remain to be equal 1, indicating the uncongested node. Therefore, in

803

order to prevent the buffer overflows, we propose to modify the load control indicator γi used in [14] by adding the amount of data dropped to the numerator as shown in equation:

)(

)1()1(mod

tQ

tDtQ

i

ii

i

With the increasing traffic arrival at some point the buffer of certain eNB(s) will reach its maximum length Qmax. At this point the non-zero amount of lost data will indicate that the bandwidth allocated to this eNB is not sufficient to prevent packet losses due to the buffer overflow. Combining equations (3) and (1) with the equation (8), we obtain following new expression for the uplink and downlink modified load indicators γi

ULmod and γi

DLmod that can detect the buffer lost due

to congestions at eNB’s.

)(

)()1()(

)(

)()1()(

modmod

tQ

SE

ttxtLtQ

tQ

SE

ttxtLtQ

DL

i

DL

iDL

i

DL

i

UL

i

UL

iUL

i

UL

i

DL

i

UL

ii

The expression (9) contains true values of )1( tL . In

reality, however, we use the estimated values )1(ˆ tL (as shown

in equation (2)), because the values )1( tL are not available at

the moment of allocation (i.e. at tth

monitoring interval). Since the objective function in (7) is continuously differentiable concave function over a convex feasibility region, the unique optimal bandwidth allocation vector T

nxxx **

1

* ,..., can be found

from the necessary and sufficient Karush–Kuhn–Tucker (KKT) conditions [23] given by:

...)( ,...)( :where

11 ,0 ,0)(

1 ,0)( ,0)(

0)( )(

T

1

T

1

***

*

1

*

1

1

***

n

ii

i

n

iii

ni

n

i

ii

x

g

x

gxg

x

f

x

fxf

niμxgμ

nixgxg

xgμxf

iμ are non-negative Lagrange multipliers, associated with i

th

constraint in (7); *

iμ are the values of the Lagrange multipliers

associated the optimal stationary point. Using (7), conditions in (10) are equivalent to the following system of equations:

11 ,0 ,0)( ,0

1 ,0 ,0 ,0

*

1

**

1

1

*

****

1

*

*

niμCxμCx

nixμxμμx

i

n

i

in

n

i

i

iini

i

ii

From the optimality conditions (11) we get:

niμμ

xin

i

i

1 , **

1

*

Using (12) in complementary slackness conditions of (11) gives:

0)( ,1 ,0 **

1

1

*

1**

1

*

Cμμμniμμ

μin

n

i

in

in

ii

Now if 0*

1

nμ then (12) implies: .1 , 0

*

* niμ

xi

i

i

However, (11) means µi* ≥ 0 and ωi ≥ 0, 1 ≤ i ≤ n by

construction. Hence, if µn+1* = 0, this contradicts (11). Thus,

we conclude that µn+1* > 0, using which in (13) gives:

)( **

1

1

in

n

i

iμμC

Putting (14) into (12) gives us:

niCxn

j

jii

1 , 1

*

which clearly satisfies the feasibility constraints of (11) (because of non-negativity of weight ωi). The Lagrange multipliers corresponding to constraints in (7) are equal:

1 ,0

1 ,0

1

*

niC

ni

μn

j

ji

IV. SIMULATION RESULTS

In this section we present a performance analysis of the proposed PRA algorithm for cognitive LTE network comprising a number of eNBs supporting traffic in different cells. The network model used in simulations is illustrated on Figure 1. The simulation model consists of seven eNBs communicating with the Backbone Server through NRM through a 1Gbit/s data rate IP links. Each eNB serves a number of fixed user equipments (UEs), randomly positioned in the system area with a 1km radius. The simulation model was developed using the OPNET modeling package [2].

To evaluate the performance of the proposed PRA algorithm with the modified load indicator (MLI) γi

mod, we

compare its performance with the performance of two bandwidth allocation schemes: bandwidth allocation with the conventional load indicator (LI) γi

and equal fixed bandwidth

allocation (FA). The features of different bandwidth allocation techniques used for performance evaluation in this paper are listed in Table I. The parameters of LTE network model used in simulations are provided in Table II. The total available bandwidth in all simulations is C = 35 MHz. The length of monitoring time interval Δt = 10sec is calculated using the

equation )(sup1

tfEti

ni

, where E{fi(t)} is mean packet round-

trip-delay at eNB i.

TABLE I. BANDWIDTH ALLOCATION SCHEMES IN SIMULATIONS

Scheme Optimization Criteria Weight Generation

FA Fixed Bandwidth Allocation nii

1 ,1

MLCI Weighted Proportional Fairness niDL

i

UL

ii 1 ,modmod

LCI Weighted Proportional Fairness niDL

i

UL

ii 1 ,

804

TABLE II. SIMULATION PARAMETERS

Parameter Value

PHY profile Operation mode FDD

Cyclic Prefix Type Normal (7 Symbols/

Slot)

Carrier frequency 2GHz

Subcarrier spacing 15kHz

Modulation and Coding

Scheme

16QAM with coding

rate = 0.601563

EPC Bearer Definitions 348kbit/s (Non-GBR)

BSR Parameters Periodic Timer 5 subframes

Retransmission Timer 2560 subframes

Packet

Scheduling

Proportional Fair Throughput

(PF)

Performance of the proposed PRA algorithm is observed

using two scenarios. In the first homogeneous scenario, referred as “voice”, each eNB served only VoIP users. In the second heterogeneous scenario, referred as “mix”, the traffic of eNBs is represented by a “mixed” traffic comprising VoIP, video and data users. The overall number of users of each type is in proportion to 2:2:3 for VoIP, video, and data, respectively. The VoIP traffic is generated by using the G.723.1 (12.2 Kbps) codec with a voice payload size 40 bytes and a voice payload interval 30 ms. Each VoIP user might be either in active (talk-spurts period) or inactive (silent period) state. The durations of the talk-spurts and silent periods are exponentially distributed with 0.65s and 0.352s means, respectively. Video services are simulated using a high resolution video model with a constant frame size equal 6250 bytes and exponentially distributed frame inter-arrival intervals (with mean equal 0.5s). Data users in simulations are HTTP1.1 users generating pages or images with exponential page inter-arrival intervals (mean equal 60sec). It is assumed that one page consists of one object, whereas one image consists of five objects. The object size is constant and equal 1000 bytes. To show the consistent performance of the proposed PRA algorithm for different traffic loads we show that: 1) the fairness of PRA algorithm with MLI is higher (or at least not smaller) than the fairness of other bandwidth allocation schemes (LI and FA); 2) the service performance of PRA algorithm with MLI is better that the performance of LI and FA, which implies that the target performance metrics (delay and loss) in the MLI technique should be smaller than LCI and FA techniques for any traffic mix and load in the network.

In this paper we introduce a new definition of fairness of bandwidth allocation. According to this definition, a bandwidth is allocated fairly if all eNBs provide similar service performance to the UEs, i.e. the user-perceived QoS does not depend on their location in the network (serving eNB). To evaluate the fairness of bandwidth allocation defined in this way, we introduce two new indices of fairness: the weighted delay index of fairness (WDIF) and the weighted loss index of fairness (WLIF) given by:

T

kn

i i

i

n

i i

i

T

kn

i i

i

n

i i

i

tDn

tD

TWLIF

tWn

tW

TWDIF

0

1

2

2

1

0

1

2

2

1

)(

)(

1 ,

)(

)(

1

where Wi(t) and Di(t) are the average packet delay and loss in eNB i at t

th monitoring interval, respectively; ωi is the weight

assigned to eNB i at tth

monitoring interval; n is the total number of eNBs; T is the duration of simulation (counted in terms of the number of monitoring time intervals). The WDIF and the WLIF express the “degree of similarity” of delay and loss in different eNBs. These indices are derived using the well-known Jain's index of fairness using delay and loss instead of throughput for fairness evaluation [25]. Figure 2 shows the values of WDIF and WLIF in a scenario with 80 voice users per eNB. Results show that the highest value of fairness is achieved in PRA with MLI. The impact of increased fairness with MLI on user-perceived QoS in different eNBs is shown on Figure 3: it follows, that with MLI the users of different eNBs experience relatively similar delays and losses, while with the other techniques (LI and FA) the user-perceived QoS highly depends on their serving eNBs.

Fig.2. WDIF/WLIF in a scenario with 80 voice users per eNB

Fig. 3. Packet delay and loss in a scenario with 80 voice users per eNB

The next figures further examine the performance of the resource allocation techniques for different traffic mixes by incorporating data and video traffic sources along with the voice traffic. Graphs below show the total (uplink and downlink) packet transmission and queuing delays and the (uplink) delays due to uplink packet scheduling (Figure 4) in scenarios with voice users. Figure 5 shows the packet transmission and queuing delays and the (uplink) delays due to uplink packet scheduling, respectively in scenarios with voice, video and data applications. From these graphs it follows that a PRA algorithm reduces not only transmission and queuing delays, but also the delay due to uplink packet scheduling for the users in both scenarios.

805

Fig. 4. Packet transmission and queuing delays and delays due to uplink

packet scheduling in scenarios with voice applications

Fig. 5. Packet transmission and queuing delays and delays due to uplink

packet scheduling in scenarios with mixed user applications

REFERENCES

[1] OPNET website: www.opnet.com

[2] J. Mittola and G.Q. MaQuire, “Cognitive Radio: Making Software Radios More Personal”, IEEE Personal Communications, vol. 6(4), 1999, pp. 13-18.

[3] S. Haykin, “Cognitive Radio: Brain-Empowered Wireless Communications”, IEEE JSAC, vol. 23(2), 2005, pp. 201-220.

[4] D. Cavdar, H. B. Yilmaz, T. Tugcu, F. Alagöz, “Analytical Modeling and Resource Planning for Cognitive Radio Systems”. Wirel. Commun. Mob. Comput. (2010). DOI: 10.1002/wcm.961.

[5] S. Glisic, B. Lorenzo (2009), “Advanced Wireless Networks: Cognitive, Cooperative & Opportunistic 4G Technology”, 2nd Edition, John Wiley & Sons Ltd., 892 pp.

[6] D. Gözüpek, F. Alagöz, “Genetic algorithm-based scheduling in cognitive radio networks under interference temperature constraints”. Int. J. Commun. Syst. (2010). DOI: 10.1002/dac.1152.

[7] W. Wang, W. Wang, Q. Lu, T. Peng, “An uplink resource allocation scheme for OFDMA-based cognitive radio networks”. Int. J. Commun. Syst. (2009). Vol. 22(5), pp. 603-623.

[8] D. T. Ngo, C. Tellambura, H. H. Nguyen, “Resource Allocation for OFDM-Based Cognitive Radio Multicast Networks”. In Proc. WCNC’09. April 2009, pp. 1-6.

[9] J. Neel, R.M. Buehrer, B.H. Reed, R.P. Gilles, Game theoretic analysis of a network of cognitive radios, In Proc.MWSCAS-2002, 2002, vol 3, pp. 409-412.

[10] R.W. Thomas, Cognitive Networks, PhD thesis, Virginia Polytechnic Institute and State University, 2007.

[11] M. Felegyhazi, M. Cagalj, S.S. Bidokhti, J.P. Hubaux, Noncooperative multi-radio channel allocation in wireless networks, In Proc. INFOCOM’07, 2007, pp. 1442-1450.

[12] G. He, S. Gault, M. Debbah, E. Altman, Distributed power allocation game for uplink ofdm systems, In Proc. WiOPT, 2008, pp. 515-521.

[13] J. Jin, W. H. Wang, M. Palaniswami, Utility max-min fair resource allocation for communication networks with multipath routing, Comp. Comm., vol. 32 (17), 2009, pp. 1802-1809.

[14] V. Jacobson, “Congestion Avoidance and Control”. In Proc. SIGCOMM’88. Aug. 1988, Vol. 18(4), pp. 314-329.

[15] J. H. Mo and J. Walrand, “Fair End-to-End Window-based Congestion Control”, IEEE/ACM Transactions on Networking. Vol.9, pp.556-567.

[16] D. V. Lindley (1952), "The theory of queues with a single server", Mathematical Proceedings of the Cambridge Philosophical Society, 48 (2), pp. 277–289.

[17] T. Soderstrom and P. Stoica. System identification. Prentice Hall International (UK) Ltd, 1989.

[18] A. Asheralieva et.al., “A Predictive Network Resource Allocation Technique for Cognitive Wireless Networks”, in Proc. ICSPCS’10, Dec. 2010, pp. 1 - 9.

[19] A. Asheralieva, J. Y. Khan, K. Mahata, “Traffic Prediction Based Packet Transmission Priority Technique in an Infrastructure Wireless Network”, in Proc. WCNC’11, March 2011, pp. 404 – 409.

[20] H. Akaike. Fitting autoregressive models for prediction. Ann. Inst. Statist. Math., Vol. 21, pp. 243-247, 1969.

[21] F. Kelly, “Charging and rate control for elastic traffic”, Europ. Trans. Telecom, Vol 8 (1997), pp. 33-37.

[22] H. W. Kuhn, A. W. Tucker (1951), “Nonlinear programming". In Proc. 2nd Berkeley Symposium, Berkeley: University of California Press. pp. 481–492.

[23] H. Wang, D. Jiang “Performance Comparison of Control-less Scheduling Policies for VoIP in LTE UL”, in Proc. WCNC 2008, pp. 2497-2501.

[24] A.Zalesky, “Teletraffic Performance for All-Optical Network and Their Analysis”, PhD thesis, University of Melbourne, Australia, 2006.

[25] R. Jain, D. Chiu, and W. Hawe, A Quantitative Measure of Fairness and Discrimination for Resource Allocation in Shared Computer Systems, DEC Research Report TR-301, 1984.

[26] “LTE for UMTS: Evolution to LTE-Advanced” by Harri Holma and Antti Toskala, John Wiley and Sons (2011). 576 pp.

[27] A.Larmo et. al., “The LTE link-layer design”. IEEE Comm. Mag. 47 (4), pp. 52-59, 2009.

[28] 3GPP TS 36.321. E-UTRA; MAC protocol specification (Release 8).

[29] P. Kela et.al., “Dynamic Packet Scheduling Performance in UTRA Long Term Evolution Downlink,” in Proc. ISWPC’08, May 2008.

[30] 3GPP TS 36.211. Physical Channels and Modulation (Release 8).

806