4348 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO.

6, JUNE 2016

Joint Resource Allocation and Relay Selection in

LTE-Advanced Network Using Hybrid Co-Operative
Relaying and Network Coding
Ahmed Zainaldin, Student Member, IEEE, Hassan Halabian, Member, IEEE,
and Ioannis Lambadaris, Member, IEEE

Abstract—The problem of joint resource allocation and relay exchange for network capacity gain to save network resources
selection is studied for bidirectional LTE-advanced relay net- and hence increase the network throughput [2], [3]. CoR and
works. The bidirectional communication between user equipment NC are also used in Two-Way Relay Channels (TWRC) to
(UE) and eNodeB (eNB) is performed via direct transmission, co-
operative relaying (CoR), or a combination of network coding reduce the number of time slots for bidirectional transmission
(NC) and CoR (NC/CoR). In this paper, an enhanced three-time- and increase the spectral efficiency of the system [3].
slot per cycle time-division duplexing (TDD) scheme is proposed
for LTE-Advanced frame architecture to accommodate a hybrid
transmission scheme. More specifically, we formulate the prob- A. Optimization for Relay-Assisted Bidirectional OFDMA
lem of joint resource assignment, relay selection, and bidirectional Cellular Networks
transmission scheme selection as a combinatorial optimization
problem with the objective to maximize the total product of back- There exists a great body of research on the design of net-
log and rate (back-pressure principle). Two approaches are con- work coding and cooperative relaying in wireless OFDMA net-
sidered to solve our combinatorial optimization problem. First, a works [4], [5]. The authors in [4] characterized the rate region
graph-based framework is proposed in which the problem is trans- of the combined direct, NC/CoR and CoR hybrid transmission
formed into a maximum weighted Clique problem (MWCP). In
addition, our problem is also transformed into a three-dimensional schemes in an explicit form. However, in [4] it is assumed
assignment problem (3DAP) which is solved using a hybrid ant that each bidirectional link has a dedicated OFDMA channel.
colony optimization (ACO) algorithm. Using simulations, it is In addition, the model in [4] had the flexibility in uplink and
concluded that the hybrid transmission scheme outperforms all downlink resources to be divided unequally which is not the
conventional nonhybrid schemes. Moreover, the simulation results case for current cellular networks such as LTE. Optimization
confirm that while the two proposed solutions provide similar
results, the ACO algorithm is faster due to its lower complexity. in OFDMA cellular networks has been extensively studied in
the literature [6]–[9]. However, less work studied the optimiza-
Index Terms—Bidirectional Cooperative Relaying, Network
Coding, Optimal Resource Allocation, Ant Colony, Maximum tion of bidirectional OFDMA-based cellular networks using
Weighted Clique Problem. cooperative relaying and network coding [10]–[12]. Most pre-
vious works used the Lagrange dual decomposition method to
I. I NTRODUCTION solve the joint optimization problem by decomposing the sub-
carrier allocation problem into per-subcarrier sub-problems that
R ELAY nodes (RNs) are introduced in LTE-Advanced
networks as a cost effective approach for improving sys-
tem performance, such as cell throughput, power saving and
can be solved independently [11]. The work in [10] proposed
an XOR based cooperative diversity scheme and presented an
coverage extension. RNs can facilitate the transmission by algorithm for joint optimization of relay selection, transmission
employing Amplify-and-Forward (AF) or Decode-and-Forward mode selection, power allocation and subcarrier allocation. In
(DF) [1] schemes. In both cases, the combination of the directly [13], the authors proposed a hierarchical protocol for one-way
transmitted signal and the relayed signal provides a form of and two-way relaying in a two-time-slot TDD mode. In this
cooperative relaying (CoR), which has been shown to improve protocol, the transmission mode of each UE as well as its assist-
the system throughput and power efficiency. Network cod- ing RN is pre-fixed, and the downlink and uplink transmission
ing (NC) is another technique to allow coding capability in modes for each UE are the same. Then, only joint power and
subcarrier allocation is considered and solved by Lagrange dual
Manuscript received January 9, 2015; revised June 26, 2015 and November
19, 2015; accepted February 27, 2016. Date of publication March 9, 2016; date decomposition method. The system studied in [14] operates in
of current version June 7, 2016. The associate editor coordinating the review of frequency-division duplexing (FDD) mode with fixed sets of
this paper and approving it for publication was A. Abrardo. data subcarriers and relay subcarriers. None of these studies
A. Zainaldin and H. Halabian are with Ericsson Canada Inc., Ottawa, ON considered subcarrier pairing which provides additional gains
K2K 2V6, Canada (e-mail: [email protected]; hassan.halabian@
ericsson.com). to the optimization framework.
I. Lambadaris is with the Department of Systems and Computer Engineering,
Carleton University, Ottawa, ON K1S 5B6, Canada (e-mail: ioannis@
sce.carleton.ca). B. Optimization vs. Heuristic Approaches
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. Resource allocation in OFDMA relay-enhanced cellular net-
Digital Object Identifier 10.1109/TWC.2016.2539963 works including power allocation, relay mode selection, and
1536-1276 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Mississippi State University Libraries. Downloaded on January 22,2025 at 20:58:29 UTC from IEEE Xplore. Restrictions apply.
ZAINALDIN et al.: JOINT RESOURCE ALLOCATION AND RELAY SELECTION IN LTE-ADVANCED NETWORK USING COP AND NC 4349

subcarrier scheduling can be formulated into a combinatorial

optimization problem which is difficult to solve in polynomial
time. Exact solution methods, such as branch and bound, are
limited to solving discrete combinatorial problems of small
size. In contrast, for large-scale applications, more research
efforts have been devoted to develop heuristic and metaheuristic
sub-optimal solution methods, including linear programming,
genetic, simulated annealing and tabu search algorithms. Linear
Programming (LP) [15] and Lagrangian relaxation (LR) [10],
[11] have been used extensively in the design of approxima-
tion algorithms for a variety of combinatorial problems. LR
is typically used to relax a set of complex side constraints
and accordingly compensate a penalty term in the objective
function. LR method provides a fast solution which is usually
updated by a sub-gradient method, although it may converge
slowly. Consequently, the quality of final solution is not guar-
anteed. A common heuristic approach is to use randomized
methods such as genetic algorithms (GA). Such methods have
been developed by inspiration from nature and have demon-
strated superior performance both in computational time and Fig. 1. Two-way Relaying Schemes.
accuracy [16]. GA methods are very popular for resource allo-
cation problems but yet to be fully exploited in the context of and thus preferable for LTE-Advanced networks. The Bron-
OFDMA networks [17], [18]. Ant Colony Optimization (ACO), Kerbosch optimal solution is used as a benchmark to evaluate
proposed in [19] to find sub-optimal solutions for NP-hard our sub-optimal algorithms.
problems, is one of the most well-known GA techniques. ACO The rest of the paper is organized as follows. Section II
has also been widely used in OFDMA cellular networks to describes the system description and assumptions for the con-
solve the resource allocation problem [12], [20]. An enhanced sidered TWRC system. Section III describes the optimization
form of ACO is used in this work to solve our three-dimensional framework used in this work. Section IV presents the proposed
NP-hard optimization problem. graph model and Bron-Kerbosch algorithm used to solve our
Our contributions in this paper are summarized as fol- optimization problem. In Section V, we propose our three-
lows: An enhanced three-time-slot per cycle TDD transmission dimensional ACO algorithm as a heuristic approach for our
scheme is proposed for LTE-Advanced frame architecture in optimization problem. In Section VI, we present the simulation
which a dedicated subframe for RN is introduced. The first two results where we show the performance gains of the proposed
subframes in each cycle are used for transmission by eNB only hybrid scheme. Finally, Section VII concludes the paper.
(downlink) and UE only (uplink). However, the RN subframe
is used for both uplink (to eNB) and downlink (to UE) using
OFDMA to be able to use network coding at the RN. Using the
II. S YSTEM D ESCRIPTION AND A SSUMPTIONS
enhanced TDD frame architecture, an optimization framework
is proposed for LTE-Advanced systems to jointly optimize In this paper, the bidirectional LTE-Advanced communica-
resource allocation, transmission scheme selection (direct, CoR tion is modeled by a TWRC channel where the end-point enti-
and NC/CoR) and relay selection. The proposed optimization ties (UE and eNB) can communicate with each other directly
framework is a combinatorial optimization problem maximiz- or through an RN. The RN routes the packets to the UEs and
ing the throughput region of the system and stabilizing the eNB through cooperative diversity or MAC layer network cod-
queues. Our problem is first transformed into a graph-based ing where the nodes are aware of the transmission scheme used.
Maximum Weighted Clique Problem (MWCP). The MWCP In the LTE radio interface, two consecutive time slots create a
problem is first solved using an enhanced Bron-Kerbosch algo- subframe where each time slot spans 0.5 msec. Resources are
rithm [21]. Bron-Kerbosch is a well-known branch and bound allocated to UEs in units of resource blocks (RBs) or Resource
algorithm used in this paper to provide an optimal solution. Block Groups (RBGs). An RB is a set of 12 consecutive sub-
In addition, the MWCP problem is solved using a sequential carriers (180 kHz) over the duration of one slot. The number
greedy heuristic algorithm which provides a fast sub-optimal of RBs included in an RBG is determined by the system band-
solution. Our NP-hard combinatorial optimization problem width. In this paper, resources are allocated to UEs in units of
is then transformed into a 3DAP problem. An enhanced 3- RBGs.
dimensional ACO algorithm is proposed to provide a sub- Cooperative Relaying (CoR): Figure 1(a) shows the CoR
optimal solution for our combinatorial assignment problem. transmission scheme used in this paper. Four network resources
The simulation results show that the proposed hybrid scheme are used to send a pair of packets P1 and P2 between the
outperforms the conventional non-hybrid schemes. Moreover, end-point entities. UE transmits P1 which occupies one RBG
we will observe that the proposed MWCP-based and ACO- (subchannel f 1 ) in the Uplink (UL) subframe and one RBG
based algorithms perform closely while the latter is much faster (subchannel f 2 ) in the Relay subframe to reach the eNB.
Authorized licensed use limited to: Mississippi State University Libraries. Downloaded on January 22,2025 at 20:58:29 UTC from IEEE Xplore. Restrictions apply.
4350 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 6, JUNE 2016

the same subframe durations as defined in the LTE standard,

the subframe lengths of the new three-slot cycle is the same as
the length of the combined UL and DL subframes, i.e., the total
number of RBGs are divided equally among the three slots. For
example, for 1.4 MHz LTE carriers, 6 usable RBs (PRBs) per
subframe are available and the number of RBs per RBG group
is 1. Therefore, the 12 RBGs can be divided equally such that
each subframe in our three-time-slot cycle contains 4 RBGs.
Similarly, for 10 MHz carriers (a more common configuration),
each subframe consists of 50 RBs with a maximum RBG size
of 3 RBs. Therefore, the 100 total RBs can be divided such that
Fig. 2. LTE-Advanced Frame Architectures. each subframe contains 11 RBGs. We assume that the RBG size
for uplink is equal to the DL RBG size as defined for Rel-8 in
Similarly, eNB transmits P2 which occupies one RBG (sub- [22]. Moreover, to avoid wasting unused RBGs in the RN sub-
channel f 3 ) in the Downlink (DL) subframe and one RBG frame in direct transmission, we assume that the RBGs in the
(subchannel f 4 ) in the RN subframe to reach the UE. In addi- RN subframe are split equally between UL and DL.
tion, P1 (resp. P2 ) is overheard by eNB (resp. UE) in the UL We denote the link capacity from station x to station y at
(resp. DL) subframe to perform cooperative diversity detection cycle t by C x y (t), which is given by C x y (t) = F(B, h x y (t),
which improves the system efficiency and robustness. Hence, Px , σx y , dx y , α) where C x y (t) is measured in bits/s, B is the
CoR requires that both the relay and destination fully decode bandwidth of the used subchannel, Px is the transmission power
the received data [1]. of node x, and σx y is noise power. h x y (t) is the channel gain
Network Coding/Cooperative Relaying (NC/CoR): of link x→y which is assumed to be independent and identi-
Employing network coding at the RN provides higher spectral cally distributed complex Gaussian and does not change within
efficiency as only 3 network resources are used to send P1 each time-frame (slow fading channels). dx y is the distance of
and P2 between the end-point entities. Figure 1(b) shows the stations x and y and α is the path-loss exponent. If the x→y
NC/CoR transmission scheme. P1 is hopped from UE to RN channel is an M × N MIMO channel (with N transmitting and
using one RBG in the UL subframe at subchannel f 1 . P2 is M receiving antennas), h x y (t) will be a complex matrix. Note
hopped from eNB to RN using one RBG in the DL subframe at that F(·) is a general function to express the channel capacity.
subchannel f 2 . Then, RN performs network coding (bit-wise We use the downscaled and truncated Shannon formula for LTE
XOR) on P1 and P2 as P3 = P1 ⊕ P2 and broadcasts it as one networks proposed in [23] to provide realistic LTE data rates.
packet to both UE and eNB in the RN subframe at subchannel Relays operate in DF mode. Henceforth, we use subscripts U, R
f 3 . Each receiver can then obtain its intended packet by and E, in our formulas to denote UE, RN and eNB, respectively.
XORing the received packet from RN with its own packet. The channel state of the TWRC at cycle t is denoted by vec-
Similar to CoR, P1 (resp. P2 ) is overheard by the eNB (resp. tor C(t) = (CUE (t), CEU (t), CUR (t), CRU (t), CER (t), CRE (t)),
UE) in the UL (resp. DL) subframe where each end-point entity which contains the capacities of all the communication links
will have two versions of its intended packet. Cooperative in the TWRC. According to the LTE Modulation and Coding
diversity detection can then be applied on the two versions of Scheme (MCS) constraint, all resources assigned to one UE
the received packet. should use the same MCS for each subframe. A drawback of
We consider a single cell LTE wireless TDD network in the MCS constraint is that the selection process is based on the
which each transmission frame is divided into a DL subframe lowest SNR of the RBGs assigned to a UE. However, with an
and an UL subframe, both on the same frequency band but efficient RBG assignment algorithm, the impact on the system
in two different time slots. The new proposed frame archi- throughput due to the MCS selection algorithm can be mini-
tecture can be adapted to next-generation networks such as mized. The LTE per-user MCS constraint is not considered in
LTE-Advanced. Figure 2(a) shows a reduced version of a tra- this paper due to space limitation.
ditional LTE-Advanced TDD frame architecture with relay The packet arrival processes to UE and eNB are denoted
support in which a bidirectional cycle is divided into DL and by {AU (t)}∞ ∞
t=1 and {A E (t)}t=1 , respectively. More specifically,
UL subframes using OFDMA and SC-FDMA (Single Carrier- AU (t) and A E (t) are the random variables of the number of
Frequency Division Multiple Access), respectively. This frame packets arrived exogenously to UE and eNB at cycle t. These
architecture is not suitable for the NC transmission scheme due processes are assumed to be stationary with rates U and  E ,
to the introduction of the XOR network coding which requires respectively. The packets are assumed to be of the same length
the use of both UL and DL resources (RBGs) in one trans- L. UE and eNB keep the incoming packets in separate buffers.
mission. Hence, to support relay-assisted CoR and NC, we We assume that these buffers are not limited in size, i.e., we
introduce a new RN subframe which is used to support CoR do not have packet drop in the system. Let Q U (t) and Q E (t)
and/or NC/CoR transmissions. The resulting frame architec- denote the queue lengths of UE and eNB at the end of cycle t,
ture consists of a three-time-slot per cycle frame architecture as respectively. At each cycle t, a centralized scheduler (usually
shown in Figure 2(b). Since the relay node has a larger power in the eNB) observes the queue state of the system as well
capacity than the UE, OFDMA is generally used for transmit- as the channel condition and solves our proposed optimization
ting the network-coded packet. Moreover, in order to maintain problem. At the end of the frame and after serving the packets
Authorized licensed use limited to: Mississippi State University Libraries. Downloaded on January 22,2025 at 20:58:29 UTC from IEEE Xplore. Restrictions apply.
ZAINALDIN et al.: JOINT RESOURCE ALLOCATION AND RELAY SELECTION IN LTE-ADVANCED NETWORK USING COP AND NC 4351

according to the scheduling policy, the new arrivals are added k represents the maximum
and broadcasts it to both of them. Cmin
to the queues and the queue lengths are updated. rate at which both eNB and UE can reliably decode the signals
from RN during the RN subframe.
III. O PTIMIZATION F RAMEWORK
In this section, we present our optimization framework in B. Optimization Problem Formulation
detail. We first describe the achievable bidirectional rate pairs Let L = {1, 2, . . . , L} be the set of RNs, M = {1, 2, . . . , M}
for direct, CoR and NC/CoR transmission modes. the set of UEs and N = {1, 2, . . . , N } the set of RBGs.
Depending on the selected transmission scheme, we define the
i, j,k
A. Achievable Bidirectional Rate Pairs following sets of binary variables for each cycle. ρm,dir ect (t)
1) Direct Transmission: In direct transmission, no relay is indicates whether RBG tuple (i, j, k) in the DL and UL and
used in the bidirectional transmission. However, the RBG in the RN subframes is assigned to UE m using direct transmission
i, j,k
RN subframe is still used for DL and UL. Therefore, the uplink mode. ρm,l,CoR (t) indicates whether RBG tuple (i, j, k) in the
and downlink achievable rate pair for the direct transmission DL, UL and RN subframes is assigned to UE m and RN l using
i, j,k
scheme over the DL, UL and RN subframes for the RBG tuples CoR. ρm,l,N C/CoR (t) indicates whether RBG tuple (i, j, k) in
j,k
(i, j, k) is denoted by (R i,k (t), RU E (t)), in which the DL, UL and RN subframes is assigned to UE m and RN l

EU  using NC/CoR.
1 1 k
R i,k (t) = C i
EU (t) + C (t) , We assume each RBG can only operate in one of the
EU
3 2 EU
  three transmission modes. Our objective is to maximize the
j,k 1 1 total product of backlog and achievable rate by allocating
RU E (t) = CUi E (t) + CUk E (t) , (1)
3 2 RBG tuples optimally, selecting the best RNs and transmis-
where the factor 13 is a result of the use of the proposed three- sion schemes for each bidirectional communication per-cycle.
time-slot frame architecture. The objective function is based on the back-pressure princi-
2) CoR Transmission: According to [1], the uplink and ple which provides throughput-optimality and queue stability
downlink achievable rate pair for the CoR scheme over the DL, [25]. Therefore, the proposed hybrid transmission scheme can
UL and RN subframes for the RBG tuples (i, j, k) is denoted stabilize the system for all the traffic rates inside the net-
j,k work stability region, which is proved based on applying the
by (R i,k
EU (t), RU E (t)), in which
   Lyapunov stability technique [26]. The objective function is
1
R i,k
EU (t) = min C i
ER (t), D F i,k
γ i
EU (t), γ 1
ef f (t) , formulated as the maximization Problem (4) subject to the con-
3 straints (5)–(7), which are shown at the bottom of the next page.
1   
j,k j j
RU E (t) = min CU R (t), D F j,k γU E (t), γe2f f (t) , (2) Q Um (t) and Q E m (t) are the queue lengths associated to UE m
3 in the UE and eNB at the end of cycle t, respectively.
where D F(γ1 (t), γ2 (t)) is the capacity of the cooperative trans- Problem (4) is a characterization of the maximum weighted
mission using the DF scheme where γ1 (t) and γ2 (t) are the rate (total product of backlog and achievable downlink/uplink
SNR of source-destination and relay-destination links, respec- rates) of each UE-eNB bidirectional communication over the
tively. Since in our frame architecture, the RBGs in the RN direct, CoR and NC/CoR transmission schemes. The deci-
subframe are split for UL and DL transmissions, we need to use sion binary variables in constraints (5)–(7) indicate that each
effective SNR for the relay links reflected in (2) as γe1f f (t) = RBG tuple (i, j, k) can only be assigned to one UE m in
 
direct transmission and one RN l in CoR and NC/CoR per-
1 + γ RUk (t) − 1 and γ 2 (t) = 1 + γ k (t) − 1 [24]. The
ef f RE cycle. Moreover, the bidirectional communication session on
D F(γ1 (t), γ2 (t)) for the case of Shannon capacity formula is each RBG tuple (i, j, k) can only operate in one of the three
given by a log function as log(1 + γ1 (t) + γ2 (t)). We only use transmission schemes.
three RBGs for the bidirectional communication to complete
the four transmissions. Equation (2) assumes that the coopera-
tive rate requires that both the relay and destination fully decode C. Modified Optimization Problem
the received data [1].
3) NC/CoR Transmission: According to [5], the uplink and Problem (4) is a combinatorial optimization problem whose
downlink achievable rate pair for the NC/CoR scheme over computational complexity increases drastically with the num-
the DL, UL and RN subframes for the RBG tuples (i, j, k) is ber of binary variables [27]. By observing the first summation
denoted by (R i,k
j,k term of problem (4), it is easy to find that there is at most one
EU (t), RU E (t)), in which
  non-zero element for a given RBG tuple (i, j, k) due to the con-
1
R i,k
EU (t) = min C Ei R (t), C EU i
(t) + Cmink
(t) , straints (5)–(7). This implies that among the M UEs, at most
3
1   one UE m can occupy the RBG tuple for direct transmission
j,k j j
RU E (t) = min CU R (t), CU E (t) + Cmin k
(t) , (3) in the DL, UL and RN subframes. By observing the second
3 and third summation terms of problem (4), one also finds at

k = min C k , C k

where Cmin RU R E . In the RN subframe, RN most one non-zero element for a given RBG tuple for CoR and
combines the received signals (bitwise XOR) from UE and eNB NC/CoR, respectively. This implies that the RBG tuple can only

Authorized licensed use limited to: Mississippi State University Libraries. Downloaded on January 22,2025 at 20:58:29 UTC from IEEE Xplore. Restrictions apply.
4352 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 6, JUNE 2016

be assigned to one relay node l to help one UE m for CoR or IV. G RAPH -BASED O PTIMIZATION
NC/CoR. Conventional approaches to solve mixed-integer combinato-
Based on the above observations, we define W1 (i, j, k) in
rial problems, such as exhaustive search, are computationally
(9), which is shown at the bottom of the page, as the weight inefficient since our optimization problem needs to be called
function for the RBG tuple (i, j, k) in direct transmission and at each cycle. We propose to solve our problem defined in (8)
W2 (i, j, k) in (10), which is shown at the bottom of the page,
by transforming it into a Maximum Weighted Clique Problem
as the weight function for the RBG tuple (i, j, k) in CoR and (MWCP). A clique is a fully connected subgraph of a graph
NC/CoR. Then, for maximizing the system total objective and a maximum clique is the clique with the largest number of
function, Problem (4) can be represented as the maximization
vertices in a given graph. In other words, a clique is a subset
Problem (8) subject to the new constraints (11)–(13), which S of vertices in a graph G such that each pair of vertices in S
are shown at the bottom of the page. The term m  in the first is connected by an edge. The maximum clique problem is the
summation term represents the UE that maximizes the term in
problem of finding, in a given graph, the clique with the largest
(9). The terms m  , l  and s  in the second summation represent number of vertices. A Maximum Weighted Clique Problem
the UE index, RN index and transmission scheme (CoR or
(MWCP) is a clique of a weighted graph with maximum total
NC/CoR) that maximizes the term in (10), respectively. m  ,
vertex weights.
l  and s  are functions of RBGs (i, j, k). However, for better Finding all cliques using exhaustive search is expensive as
presentation of indicator variable ρ, we omit the RBG indices
the number of cliques can grow exponentially with every node
from m  , l  and s  . The new problem boils down to the RBG
(vertex) added. The Bron-Kerbosch algorithm [21] is an algo-
assignment problem in the DL and UL and RN subframes for rithm used to solve the maximal clique problem. However, the
direct, CoR and NC/CoR transmissions.
Bron-Kerbosch algorithm for clique search is only investigated

 
j,k i, j,k
max R i,k
EU,m,dir ect (t)Q E m (t − 1) + RU E,m,dir ect (t)Q U m (t − 1) ρm,dir ect (4)
m∈M,i, j,k∈N
 
j,k i, j,k
+ R i,k
EU,m,l,CoR (t)Q E m (t − 1) + RU E,m,l,CoR (t)Q Um (t − 1) ρm,l,CoR
m∈M,l∈L,i, j,k∈N
 
j,k i, j,k
+ R i,k
EU,m,l,N C/CoR (t)Q E m (t − 1) + RU E,m,l,N C/CoR (t)Q Um (t − 1) ρm,l,N C/CoR
m∈M,l∈L,i, j,k∈N
 
i, j,k i, j,k i, j,k
s.t. ρm,dir ect + ρm,l,CoR + ρm,l,N C/CoR ≤ 1, ∀i ∈ N (5)
m∈M, j,k∈N m∈M,l∈L, j,k∈N
 
i, j,k i, j,k i, j,k
ρm,dir ect + ρm,l,CoR + ρm,l,N C/CoR ≤ 1, ∀j ∈ N (6)
m∈M,i,k∈N m∈M,l∈L,i,k∈N
 
i, j,k i, j,k i, j,k
ρm,dir ect + ρm,l,CoR + ρm,l,N C/CoR ≤ 1, ∀k ∈ N (7)
m∈M,i, j∈N m∈M,l∈L,i, j∈N

i, j,k i, j,k
max W1 (i, j, k)ρm  ,dir ect + W2 (i, j, k)ρm  ,l  ,s  (8)
i, j,k∈N i, j,k∈N
 
j,k
W1 (i, j, k) = max R i,k
EU,m,dir ect (t)Q E m (t − 1) + RU E,m,dir ect (t)Q U m (t − 1) (9)
m∈M
 
j,k
W2 (i, j, k) = max R i,k
EU,m,l,CoR (t)Q E m (t − 1) + RU E,m,l,CoR (t)Q U m (t − 1) , (10)
m∈M,l∈L
 
j,k
R i,k
EU,m,l,N C/CoR (t)Q E m (t − 1) + RU E,m,l,N C/CoR (t)Q Um (t − 1)

i, j,k i, j,k
s.t. ρm  ,dir ect + ρm  ,l  ,s  ≤ 1, ∀i ∈ N (11)
j,k∈N j,k∈N

i, j,k i, j,k
ρm  ,dir ect + ρm  ,l  ,s  ≤ 1, ∀j ∈ N (12)
i,k∈N i,k∈N

i, j,k i, j,k
ρm  ,dir ect + ρm  ,l  ,s  ≤ 1, ∀k ∈ N (13)
i, j∈N i, j∈N

Authorized licensed use limited to: Mississippi State University Libraries. Downloaded on January 22,2025 at 20:58:29 UTC from IEEE Xplore. Restrictions apply.
ZAINALDIN et al.: JOINT RESOURCE ALLOCATION AND RELAY SELECTION IN LTE-ADVANCED NETWORK USING COP AND NC 4353

Algorithm 1. Basic Bron-Kerbosch Algorithm

BronKerbosch (R, P, X )
1 if P = φ AND X = φ then
2 report R as a maximal clique
–
3 for each vertex v in P do
4 BronKerbosch (R ∪ v, P ∩ N (v), X ∩ N (v))
5 P ← P \ {v}
6 – X ← X ∪ {v}
Fig. 3. MWCP Graph with N = 2 where i ∈ N , j ∈ N and k ∈ N .
Algorithm 2. Enhanced Bron-Kerbosch with pivoting
for graphs without weights. In this section, an enhanced Bron- (MWCP)
Kerbosch algorithm is used to solve our combinatorial MWCP
BronKerbosch (R, P, X )
problem in a weighted graph.
1 if P = φ AND X = φ then
if W (R  ) ≤ W (R) then
A. Maximum Weighted Clique Problem (MWCP) 2 report R as the maximum weighted clique, i.e., R  = R
– –
Let G = (V, E) be an undirected graph with vertex set V and 3 for each vertex v in P \ N (u) do
edge set E ⊂ V × V . A clique C of G is a subset of V such 4 BronKerbosch (R ∪ v, P ∩ N (v), X ∩ N (v))
that every two vertices are pairwise adjacent, i.e., ∀u, v ∈ C we 5 P ← P \ {v}
have u, v ∈ E. A clique is maximal if it is not contained in any 6 X ← X ∪ {v}
other clique. A clique is maximum if its cardinality is the largest –
among all the cliques of the graph. The maximal clique problem
(MCP) is the problem of finding the maximal cliques for a given as they were before the choice, remove the vertex from P and
graph G. add it to X . In this algorithm, set X will control the avoidance
A generalization of the MCP problem is the Maximum of reporting duplicate cliques.
Weighted Clique Problem (MWCP). A weighted graph G is
defined by G = (V, E, W ), where W is the vertex weighting
function. Each vertex v is associated with a positive weight C. Enhanced Bron–Kerbosch Algorithm for MWCP With
ωv . For a clique C of G, the weight is defined by W (C) = Pivoting
v∈C ωv . Then, the MWCP problem is to determine a clique The basic Bron-Kerbosch procedure described in
C  of maximum weight, i.e., ∀C ∈ C, W (C  ) ≥ W (C) where Algorithm 1 is inefficient in the case of graphs with many
C is the set of all possible cliques of the graph. Figure 3 shows non-maximal cliques since it makes a recursive call for every
an example of a sample weighted graph used in this work. clique (maximal or not). Bron and Kerbosch also introduced
a second version of the standard algorithm involving a pivot
vertex u chosen from P. Any maximal clique of X either
B. Basic Bron–Kerbosch Algorithm includes the pivot vertex u or one of the vertices not adjacent
Algorithm 1 outlines the standard Bron-Kerbosch procedure to u, i.e., v ∈ P \ N (u). Therefore, only the pivot vertex u and
for enumerating the maximal unweighted cliques of a given vertices from P not adjacent to u need to be considered as
graph G using a recursive backtracking algorithm. The algo- expansions of the current clique R in each recursive call of the
rithm operates on three disjoint subsets R, P, and X of vertices Bron-Kerbosch algorithm. Vertex v from P adjacent to u can
from V . The set R contains the vertices belonging to the cur- be skipped, because any clique containing v must also contain
rent clique. Set P keeps all prospective vertices, each of which u. Such a clique will be discovered in a subsequent recursive
is connected to all vertices of R. Vertices from P are used for call once u has been added to R. Algorithm 2 presents the
expanding the current clique R. Finally, the set X contains all Bron-Kerbosch procedure with pivoting and an enhancement
vertices that can no longer be used for completion of R, because for enumerating all maximal weight cliques of G. To find the
all maximal cliques containing these vertices have already been maximum weighted clique then we need to compare the weight
reported. The Bron-Kerbosch algorithm is initially called with of any recently found clique with the global maximum weight.
R = X = φ and P = V , i.e., BronKerbosch (φ, V, φ). N (v) Initially the global maximum clique, denoted by R  , is an
indicates the neighbors of the vertex v. empty set with weight zero. As the algorithm progresses, the
The algorithm operates as follows: Pick a vertex v from P to maximum weighted clique and its weight are updated.
expand. Add v to R and remove its non-neighbors from P and
X . Then pick another vertex from the new P set and repeat the
process. Continue until P is empty. Once P is empty, if X is D. MWCP Grapth Model
empty then report the content of R as a new maximal clique. It is shown that Problem (8) is equivalent to a Maximum
If it is not then R contains a subset of an already found clique. Weighted Clique Problem (MWCP). We introduce an arbitrary
Now backtrack to the last vertex picked and restore P, R and X undirected and weighted graph G = (V, E, W ) where V is the
Authorized licensed use limited to: Mississippi State University Libraries. Downloaded on January 22,2025 at 20:58:29 UTC from IEEE Xplore. Restrictions apply.
4354 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 6, JUNE 2016

set of vertices, E is the set of edges and W is the weighting shown in Subsection IV-E, the complexity grows exponentially
function. with the size of the graph. Using heuristic techniques is a pop-
Proposition 1: Our hybrid optimization problem defined ular alternative which is less bleak than attempting to solve
in (8) is equivalent to a maximum-weighted clique problem the MCP exactly. Such methods can be practical since they
(MWCP). are able to provide sub-optimal solutions of acceptable qual-
Proof: Let G = (V, E, W ) be a graph in which each ver- ity with reasonable computing efforts. A greedy algorithm is an
tex represents an RBG tuple (i, j, k) with RBGs i for DL algorithm that follows a technique of making a locally-optimal
subframe, j for UL subframe and k for RN subframe. Two decision at each stage with the expectation of finding a global
vertices intersect if they share at least one common index ele- optimum solution for an optimization problem [29]. We imple-
ment at a given subframe, and are otherwise said to be disjoint. mented a greedy algorithm which builds a maximal clique by
Two vertices will have an edge if they are disjoint. For exam- the repeated addition of a vertex into a partial clique, starting
ple (1, 1, 0) and (1, 0, 0) intersect but (1, 0, 0) and (0, 1, 1) are from an empty clique. The algorithm constructs feasible solu-
disjoint. Each vertex has a corresponding weight, denoting the tions from scratch by making, in each step, the most favorable
maximum product of backlog and achievable rate. Specifically, choice for a decision variable. Each choice depends on the deci-
the weighting functions defined in (9) and (10) are found over sions made in the previous steps, but the effects that this choice
all possible assignments of this vertex’s RBGs to all UEs and will have in the following steps are unknown. Such choices can
RNs, and then the UE and/or RN with the maximum weight thus be viewed as local decision rules that generally lead to
function are selected since the RBG tuple (i, j, k) can only be sub-optimal solutions since future decisions may have a large
assigned to one UE in direct transmission and one RN in CoR impact on the resulting quality of the solution. The sequential
and NC/CoR, i.e., Wv (i, j, k) = max{W1 (i, j, k), W2 (i, j, k)}. greedy heuristics are very fast by nature, but their quality is gen-
Our optimization problem is to find the optimal strategy to erally unsatisfactory for large graph sizes. The run-time of our
choose the transmission scheme and assign RN and RBG tuples algorithm is limited to a pre-defined duration.
to each UE in order to maximize the aggregated total product
of backlog and achievable bidirectional rate. Equivalently, we V. ACO-BASED A LGORITHM
can also interpret it as to find the optimal strategy to select dis-
joint RBG combinations and assign RNs and UEs to them to Many heuristic methods currently used in combinatorial opti-
maximize our objective function. Therefore, jointly optimizing mization are inspired by natural behaviors. Examples of such
the RBG assignment, transmission scheme selection and RN methods are genetic algorithms, simulated annealing and neu-
selection for our maximization problem is to find a subset C ral networks. The Ant Colony Optimization (ACO) belongs to
of pairwise adjacent vertices in the graph G having the largest the class of biologically inspired heuristics. The basic idea is
total weight, i.e., the so called Maximum Weighted Clique. An to imitate the cooperative behavior of ant colonies in order to
example of a complete graph for our problem for set of RBGs solve combinatorial optimization problems within a reasonable
with N = 2 where i, j, k ∈ {0, 1} is shown in Figure 3.  amount of time. The ACO is a general purpose heuristic (meta-
heuristic) algorithm that has been proposed in [30] and achieved
widespread success in solving different optimization problems.
E. Complexity
The total number of distinct vertices in our three-time-slot A. Generic ACO Algorithm
TDD frame architecture is N 3 , i.e., |V | = N 3 where |V | is
the cardinality of the set V . The complexity of the weighting The ant colony algorithm is an algorithm for finding opti-
function (9), corresponding to the direct transmission scheme, mal paths based on the behaviour of ants searching for food
is O(M) while the complexity of the weighting function (10) sources using the shortest path. During their trips back, a chemi-
is O(2M L) where the factor 2 corresponds to the CoR and cal trail (pheromone) is left on the ground to guide the other ants
NC/CoR transmission schemes. Therefore, the total complex- towards the food source. For one ant, the path is chosen accord-
ity of the defined weighting functions is O(M N 3 + 2M L N 3 ). ing to the quantity of pheromone. Furthermore, pheromones
It is shown in [28] that the worst-case time complexity for have a decreasing action over time, and the quantity left by one
generating all maximal cliques of an undirected graph in ant depends on the amount of food found and the number of
|V | ants using this trail. As illustrated in Figure 4, when ants face
the Bron-Kerbosch algorithm is O(3 3 ), where |V | is the an obstacle, there is an equal probability for each ant to choose
number of vertices in the graph. Therefore, the total com- the left or right path. As the right trail is shorter than the left
plexity of our algorithm for solving Problem (8) is O(M N 3 one and so requires less travel time, it will end up with higher
N3
+ 2M L N 3 + 3 3 ). level of pheromone. The more the ants take the right path, the
higher the pheromone trail is. This principle of communicating
ants has been used as a framework for solving combinatorial
F. Sequential Greedy Allocation for MWCP Problem optimization problems. Algorithm 3 presents the generic ACO
Exact approaches for the Maximum Clique Problem algorithm. The first step consists mainly of the initialization of
(MCP) (e.g., branch and bound) are algorithms which guarantee the pheromone trail. In the iteration step, each ant constructs a
finding an optimal solution. However, they require a run-time complete solution to the problem according to a probabilistic
that often grows exceeding large with respect to graph size. As state transition rule. The state transition rule depends mainly on
Authorized licensed use limited to: Mississippi State University Libraries. Downloaded on January 22,2025 at 20:58:29 UTC from IEEE Xplore. Restrictions apply.
ZAINALDIN et al.: JOINT RESOURCE ALLOCATION AND RELAY SELECTION IN LTE-ADVANCED NETWORK USING COP AND NC 4355

Fig. 4. Ants Facing an Obstacle. Fig. 5. 3DAP Cost Matrix.

where I , J and K are disjoint sets with |I | = |J | = |K | = N .

Algorithm 3. Generic ACO Algorithm A cost ci jk exists for each assignment of i, j and k, and the goal
Step 1: Initialization is to minimize the total cost after all the elements in I , J and K
- Initialize the pheromone trail are assigned. Figure 5 shows the three dimensional cost matrix
Step 2: Iteration used in 3DAP.
for each Ant m do Our RBG assignment problem defined in (8) is a three-
- Solution construction using pheromone trail dimensional (RBG tuple (i, j, k)) problem. The problem can
– - Update the pheromone trail be revisited as a 3DAP problem by defining
- Continue until stopping criterion
ci jk = max{W1 (i, j, k), W2 (i, j, k)}, (15)

the state of the pheromone. Once all ants generate a solution, where W1 (i, j, k) and W2 (i, j, k) are given by (9) and (10),
a global pheromone updating rule is applied in two phases; an respectively and each element ci jk corresponds to the maximum
evaporation phase where a fraction of the pheromone evapo- product of the achievable rate and backlog for each of the trans-
rates, and an intensification (re-enforcement) phase where each mission schemes. Hence, our problem is equivalent to revenue
ant deposits an amount of pheromone. This process is iterated maximization 3DAP.
until a stopping criterion is met. HAS-QAP algorithm is selected as the basis to the proposed
3DAP-ACO algorithm. Normally in standard ACO algorithms,
the pheromone trails are used to generate completely new solu-
B. Three-Dimensional Assignment Problem-ACO (3DAP- tions. However in HAS-QAP, pheromone trails are used to
ACO) Algorithm modify the current solutions by keeping a global best solution
In this section, the 3DAP-ACO algorithm is proposed to to shorten the finding process. The algorithm also performs
solve our optimization problem defined in (8). The 3DAP-ACO an additional local search phase to help improving the con-
algorithm is a modified version of the Hybrid Ant System- vergence speed. The idea of intensification and diversification
Quadratic Assignment Problem (HAS-QAP) [31]. HAS-QAP were devised to strengthen or weaken the pheromone trail,
is an application of the ant system to the QAP problem which respectively.
is one of the fundamental combinatorial problems. Our mod-
ified optimization problem defined in (8) is a 3DAP problem C. Enhanced Hybrid 3DAP-ACO Algorithm
based on the RBG tuple assignment. The proposed algorithm is
an enhanced algorithm based on HAS-QAP to solve our 3DAP Similar to HAS-QAP, in 3DAP-ACO each ant a of the m
optimization problem. 3DAP can be stated as a binary integer ants is associated with an integer permutation (solution) π a .
programming problem as follows: Each solution π a is an N × 3 matrix that determines the RBG
assignments in each subframe. It is also called permutation as
min ci jk xi jk it is being updated during the ACO algorithm. Modifications
i∈I j∈J k∈K based on the pheromone trails are then applied to each per-
mutation. The solutions found so far are then optimized using
s.t. xi jk = 1, ∀i ∈ I a 2-exchange local search method. Each pheromone trail is
j∈J k∈K
represented in the algorithm by a pheromone matrix τ . The
xi jk = 1, ∀ j ∈ J pheromone matrix can be seen as a shared memory holding
i∈I k∈K the desirability of the assignments of the best found solutions.
xi jk = 1, ∀k ∈ K The pheromone levels are updated globally after each iteration
based on the best solution found. The major difference between
i∈I j∈J
3DAP-ACO and HAS-QAP is that 3DAP-ACO requires two
xi jk ∈ {0, 1} , ∀(i, j, k) ∈ I × J × K (14) N × N pheromone matrices rather than one, where N is the
Authorized licensed use limited to: Mississippi State University Libraries. Downloaded on January 22,2025 at 20:58:29 UTC from IEEE Xplore. Restrictions apply.
4356 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 6, JUNE 2016

number of elements per dimension (number of RBGs per Algorithm 4. 3DAP-ACO Algorithm
subframe). The first pheromone matrix τ (1) represents the desir-
ability (strength) of the first and second dimension assignments Problem Initialization:
(RBGs of the first and second subframes) and the second Generate m random initial permutations π 1 , . . . , π m associ-
pheromone matrix τ (2) represents the desirability of the sec- ated to m ants
ond and third dimension assignments (RBGs of the second and Improve π 1 , . . . , π m with 2-exchange local search procedure
third subframes). Let π  be the best solution
The pheromone levels are also used for exploration and Initialize the pheromone trail matrices τ (1) and τ (2)
exploitation. Exploration is a stochastic process in which a Activate intensification
solution to the problem is made probabilistically. Exploitation Main Loop:
utilizes the pheromone levels to choose a path that maximizes for iter = 1 to I max do
the sum of pheromone levels between available edge assign- for a = 1 to m do
ments. After building a new solution, our algorithm updates /*pheromone swap procedure */ for swaps = 1 toR do
the pheromone trails. First, the pheromone trails are decreased Select a random element index as 1st swap position
globally to simulate the evaporation of pheromone. Then, the Generate a random number r
pheromone trails corresponding to the best solution obtained if r ≤ q then
so far are intensified, taking into consideration the strength /* exploit pheromone trails */
of the solution. This makes the search more aggressive and Based on maximizing pheromone sum do:
requires less time to reach good solutions. Intensification and Choose an element from the 2nd column to
diversification are used to decide where each ant starts its solu- swap
tion. Intensification is used to explore the neighborhood of – Choose an element from the 3rd column to swap
good solutions. When intensification is active in an iteration, else
the ant comes back to the previous solution if the previous Do the default actions /* explore */
solution is better. In the case where the new solution is bet- Based on a probabilistic method do:
ter, the ant simply continues working on its current solution. Choose an element from the 2nd column to
Such a process imposes a risk of an early convergence of the swap
algorithm. Therefore, a diversification mechanism implements – Choose an element from the 3rd column to swap
a partial restart of the algorithm when the solutions seem not Perform swaps on π a within 2nd and 3rd column
– and get π
a
to be improving any more. It consists of a re-initialization
– Apply 2-exchange local search to π to get π
a a
of both pheromone trails matrices. For a detailed descrip-
tion of the HAS-QAP algorithm, the reader is encouraged to for a = 1 to m do
consult [31]. if intensification is active then
The complete architecture of the modified 3DAP ACO π a (iter + 1) ← best permutation between
heuristic system can be found in Algorithm 4. A brief expla- π a (iter ) and π a (iter )
–
nation of the parameters and the algorithm is provided as else
follows. π a (iter + 1) ← π a (iter )
– –
• m indicates the number of ants in the system, where each
if ∀aπ a (iter + 1) = π a (iter ) then
ant is shown by the parameter a.
/* No more improvement to the solution */
• τ (1) indicates the first (N × N ) pheromone matrix which
deactivate intensification
represents the strength of the first and second dimension –
assignments (RBGs of the DL and UL subframes). if ∃a f (π a (iter )) > f (π  ) then
• τ (2) indicates the second (N × N ) pheromone matrix Update π  , the best solution found so far
which represents the strength of the second and third activate intensification
–
dimension assignments (RBGs of the UL and RN sub- pheromone trail updating
frames). Update pheromone trail matrices τ (1) and τ (2) by evapora-
• π indicates a 3D assignment permutation solution which tion and re-enforcement
is represented by an N × 3 matrix. The elements of diversification
the matrix represent the RBG number assigned. The if S iterations performed without improving π  then
3 columns represents the DL, UL and RN subframes, – – Perform a diversification
respectively.
• π a indicates the assignment solution of ant a. π a indi-
cates the permutation solution of ant a after performing 1) Initialization Phase: The initial solution, assigned for
R swaps during the swap procedure. π a indicates the each ant a, is randomly generated and goes through a fast
permutation solution of ant a after performing the local 2-exchange local search procedure as will be explained later.
search procedure. π  indicates the best assignment solu- All the elements of the pheromone matrices τ (1) and τ (2) are
tion found so far. initially set to the same initial scalar τ0 and the best initial
• f (π  ) indicates the objective function of the best solution solution to the 3DAP is stored in π  . Similar to [31], τ0 is set to
permutation. τ0 = Q× f1(π  ) , where Q is a fixed parameter.
Authorized licensed use limited to: Mississippi State University Libraries. Downloaded on January 22,2025 at 20:58:29 UTC from IEEE Xplore. Restrictions apply.
ZAINALDIN et al.: JOINT RESOURCE ALLOCATION AND RELAY SELECTION IN LTE-ADVANCED NETWORK USING COP AND NC 4357

2) Pheromone Swap Procedure: The pheromone swap pro- TABLE I

cedure is performed by each ant on its own permutation. ACO S YSTEM PARAMETERS
Parameter R controls the number of 2-exchange swaps. The two
element swap is performed to check if the solution improves
the current solution. The swap procedure continues until all
R swaps are considered. The two elements chosen to be
swapped in each pheromone matrix are chosen according to the
following rule:
An index r (between 1 and N ) is chosen randomly and
used to perform the swap with the first and second dimensions.
Depending on the value of r , the indices s1 and s2 (s1 , s2 = r )
can be chosen by employing one of two different policies: where α1 is a parameter that controls the evaporation of the
• Exploiting the pheromone trail with probability q. Indices pheromone trail. Then, the pheromone trails of the best solution
(1) (1) π  are re-enforced using the equations defined in (17).
s1 and s2 are chosen such that τr π a + τs π a is maxi-
s1 ,2 1 r,2

mized for τ (1) and

(2)
τr π a
(2)
+ τs π a is maximized for τ (2) . τ (1) (iter + 1) = τ (1) (iter + 1) + ( τ ) N ×N ,
s ,3 2 r,3
τ (2) (iter + 1) = τ (2) (iter + 1) + ( τ ) N ×N ,
2
• Exploring the pheromone trail with probability (17)
1 − q. Indices s1 and s2 are chosen with probability
(1) (1)
τr π a +τs π a
(2) (2)
τr π a +τs π a
where ( τ ) N ×N is an N × N matrix whose elements are iden-
s1 ,2
1 r,2
 and s1 ,3
1 r,3
 for τ (1) and tically equal to τ which is the amount of deposited pheromone
α2
 ) . The parameter α2 controls the amount of
(1) (1) (2) (2)
u =r τrπ a +τuπ a u =r τrπ a +τuπ a and is given by f (π
u,2 r,2 u,3 r,3
new pheromone deposits (re-enforcement).
τ (2) , respectively.
6) Diversification Phase: The diversification helps on gen-
After the selection of the indices r , s1 and s2 , the elements
erating new starting solutions for the ants (only one of the m
of the current solution π a (i.e., πr,2a and π a
s1 ,2 in the second ants keeps the global best solution). The pheromone matrices
column, πr,3 and πs2 ,3 in the third column) are swapped if
a a
are re-initialized for the other (m − 1) ants by using a new
improved to produce a new permutation solution π a .
random permutation. It is activated if no new best solution
3) 2-Exchange Local Search Procedure: Local search con-
is generated during the last S iterations, where S is a fixed
sists of applying a complete neighborhood examination. An
parameter.
exhaustive 2-exchange local search is also performed dur-
As it is observed, the ACO algorithm is controlled through a
ing solution initialization phase. This procedure systematically
set of parameters. The factors along with descriptions of each
examines exchanging two element positions (RBG indices). If
parameter and default values are listed in Table I and are chosen
an improving swap is found the solution is updated and the
according to the default values in [31].
local search continues until all possible swaps are considered.
Local search does not necessarily reach a local optimum, but is
fast and may produce different solutions when starting with the D. Complexity
same initial solution. Comparison of the different local search
methods is out of the scope of this work. The R pheromone The complexity of the HAS-QAP algorithm is defined in
swaps result in all m ants to have a new modified solution π a [31]. The most time consuming part of the algorithm is the local
in addition to the initial solution π a . After the local search search procedure with complexity O(N 3 ) and it is repeated
procedure, the modified solution π a is changed to π a . I max m times. Therefore, the total complexity of HAS-QAP is
4) Intensification Phase: The intensification mechanism O(I max m N 3 ), where m is the number of permutations (ants).
explores the neighborhood of the best solution so far. Since our proposed 3DAP-ACO algorithm uses 2 pheromone
Intensification is active as long as at least one ant is capa- matrices, the total complexity is O(2I max m N 3 ). Similar to
ble of improving the current global best solution. The inten- section IV, the total complexity of the defined weighting func-
sification is deactivated after iteration iter if the modified tions is O(M N 2 + 2M L N 3 ). Therefore, the total complexity
solutions π a (iter ) and π a (iter ) do not improve the current of problem (8) using Algorithm 4 is O(M N 3 + 2M L N 3 +
solution π a (iter ) for all m ants. If a better solution is found, 2I max m N 3 ).
π a (iter + 1) is set to π a (iter ) and intensification remains
active. E. Fast-3DAP-ACO
5) Pheromone Trail Updating: The two pheromone matri-
ces are updated according to the global best solution π  , which As shown in the Subsection V-D, the proposed 3DAP-ACO
leads to a faster convergence of the algorithm. Before doing so, algorithm solves our three-dimensional resource assignment
the usual pheromone evaporation has to be realized by choosing problem in polynomial time. However, the most consuming
parameter α1 and using the equations defined in (16). parts of the algorithm are the local search and the number of
algorithm iterations I max followed by all ants. Since a proper
τ (1) (iter + 1) = (1 − α1 )τ (1) (iter ), sub-optimal solution can be achieved with lower number of
iterations, we developed another version of the proposed ACO
τ (2) (iter + 1) = (1 − α1 )τ (2) (iter ), (16) algorithm called Fast-3DAP-ACO. Fast-3DAP-ACO limits the
Authorized licensed use limited to: Mississippi State University Libraries. Downloaded on January 22,2025 at 20:58:29 UTC from IEEE Xplore. Restrictions apply.
4358 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 6, JUNE 2016

TABLE II network stability region. As expected, it is shown that our pro-

S IMULATION PARAMETERS IN THE S YSTEM posed hybrid transmission scheme outperforms direct, CoR and
NC/CoR.
It is also shown in Figure 6(a) that the hybrid scheme
under a SISO configuration provides 2 Mb/cycle (46%) and
1 Mb/cycle (19%) improvements over direct and NC/CoR
schemes. Similar performance results for MIMO simulations
are shown in Figure 7(a). However, the improvements are
even higher for MIMO in which the hybrid scheme provides
9 Mb/cycle (80%) and 4 Mb/cycle (25%) improvement over the
direct and NC/CoR schemes.

B. Impact of Number of RNs on Average System Capacity

number of iterations whenever the time required to solve our
The average system capacity performance is computed for
3DAP-ACO algorithm reaches a threshold (LTE frame duration
our proposed algorithm by increasing the number of RNs. The
in our case).
number of RNs is assumed to be 1-8 and the UEs are ran-
domly placed across the network. Figure 6(b) shows the average
system capacity of the proposed optimization algorithm for dif-
VI. S IMULATIONS AND D ISCUSSIONS
ferent number of RNs for SISO configuration. It is observed
We developed our own C++ simulator to implement the that increasing the number of RNs makes the average sta-
3DAP-ACO and Bron-Kerbosch algorithms. The TDD frame ble throughput increase. However, increasing the number of
architecture is simulated for an LTE-Advanced system in which RNs above 6 relays does not provide significant impact on
UEs are randomly distributed and moving over a 10Km x 10Km the average system capacity. This is because most of the UEs
plane. RNs are placed at 50% of the eNB transmission range. across the cell will be covered by the relay nodes and hence
UE and eNB arrival processes (U and  E , respectively) are the maximum reachable performance will eventually saturate.
modeled as Poisson processes. Simulation experiments have Therefore, adding more RNs would increase the operator cost
been conducted under various scenarios in order to assess the without providing any gain. Figure 7(b) shows similar results
performance of the proposed hybrid scheme, i.e., transmission for MIMO configuration.
scheme selection, relay selection and RBG resource assignment
in terms of stable end-to-end throughput and queue backlog
(average queueing delay) for SISO and MIMO configura- C. Optimization Algorithms Comparison
tions. Each data point collected is averaged over 10,000 cycles
We compare the optimal solution (Bron-Kerbosch) with
(three-time-slot) and averaged over 10 simulation trials using
3DAP-ACO, Fast-3DAP-ACO and Greedy sub-optimal solu-
a different random seed (UE position). The detailed simulation
tions to our resource allocation problem. It can be noticed in
parameters are shown in Table II. The simulations consider a
Figure 6 that solving our hybrid assignment problem using
1.4 MHz bandwidth carrier (N = 4) with 10 UEs and 10 MHz
3DAP-ACO, Fast-3DAP-ACO and Greedy algorithms provide
bandwidth (N = 11) with 15 UEs (a more typical configura-
approximate results to the Bron-Kerbosch MWCP algorithm for
tion for LTE networks). In addition, we compare our results
N = 4. The run-time of our Greedy algorithm is limited to the
with our Greedy Allocation algorithm in Subsection IV-F and
LTE frame duration (10 msec). It provides results closest to the
Fast-3DAP-ACO algorithm in Subsection V-E.
optimal solution since it is dealing with a small graph (low num-
ber of vertices) to solve the MWCP problem. In addition, both
ACO and Fast-3DAP-ACO algorithms produce results close to
A. Average System Queue Backlog Evaluation
the optimal result with a slight degradation in Fast-3DAP-ACO
We first compute the average queue backlog performance, because the number of iterations is limited to the LTE frame
which is analogous to the end-to-end delay performance of the duration. The same simulations are performed for 10MHz LTE
all transmission schemes (direct, CoR, NC/CoR and hybrid) carrier bandwidth with 15 UEs and N = 11 and is shown in
under a random network topology using 4 relay nodes (RNs), Figure 8. Both 3DAP-ACO and Fast-3DAP-ACO have similar
10 UEs and N = 4 using the optimal Bron-Kerbosch MWCP performances while the Greedy algorithm produces unsatisfac-
algorithm. Figures 6(a) and 7(a) show the performance plots tory results because the graph size is large (113 ) and we limit
which represent the average system queue length of the differ- the execution of the algorithm to a frame duration. Therefore,
ent transmission schemes for SISO and MIMO configurations, the proposed ACO based algorithms produce the best results
respectively. It can be shown that as we increase the traffic rate in more general LTE scenarios while the Greedy algorithm can
for the UE, the average queue length also increases. However, still be used when the number of RBGs is relatively small. The
at some traffic rate, the average queue length tends to infinity. Bron-Kerbosch was not feasible due to the exponential increase
Therefore, adding more traffic would make the system unsta- in complexity and cannot be used to solve such problems with
ble and impose significant delay. This traffic rate marks the large graph sizes.
Authorized licensed use limited to: Mississippi State University Libraries. Downloaded on January 22,2025 at 20:58:29 UTC from IEEE Xplore. Restrictions apply.
ZAINALDIN et al.: JOINT RESOURCE ALLOCATION AND RELAY SELECTION IN LTE-ADVANCED NETWORK USING COP AND NC 4359

Fig. 6. Network Capacity and Delay Performance of Different Transmission Schemes (SISO).

Fig. 7. Network Capacity and Delay Performance of Different Transmission Schemes (MIMO).

Fig. 8. ACO vs. Greedy for N = 11 for SISO. Fig. 9. Average Simulation Time per Cycle.

and Fast-3DAP-ACO algorithms for 4 RNs. Bron-Kerbosch,

D. Optimal vs Sub-Optimal Algorithm Simulation Time Greedy and Fast-3DAP-ACO algorithms produce good results
Figure 9 shows the simulation times in milliseconds for when N = 4 with simulation times 6, 3 and 10 msec, respec-
the hybrid transmission scheme defined in problem (8) using tively. The 3DAP-ACO takes more time (63 msec) because
our enhanced Bron-Kerbosch MWCP, Greedy, 3DAP-ACO of the local search algorithm and the iterations needed by the
Authorized licensed use limited to: Mississippi State University Libraries. Downloaded on January 22,2025 at 20:58:29 UTC from IEEE Xplore. Restrictions apply.
4360 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 6, JUNE 2016

ants. The Greedy allocation algorithm provided the best results R EFERENCES
when the graph is small. The case is different with N = 11. [1] J. Laneman, D. Tse, and G. W. Wornell, “Cooperative diversity in wire-
The Bron-Kerbosch algorithm is not feasible and unable to less networks: Efficient protocols and outage behavior,” IEEE Trans. Inf.
solve our resource allocation problem. This Bron-Kerbosch Theory, vol. 50, no. 12, pp. 3062–3080, Dec. 2004.
[2] R. Ahlswede, N. Cai, S.-Y. Li, and R. Yeung, “Network information
algorithm is required to go through all the vertices of the undi- flow,” IEEE Trans. Inf. Theory, vol. 46, no. 4, pp. 1204–1216, Jul. 2000.
rected graph. It can also be shown that as the number of RBGs [3] S. Katti, H. Rahul, W. Hu, D. Katabi, M. Medard, and J. Crowcroft,
increase, the assignment problem takes significantly longer to “XORs in the air: Practical wireless network coding,” in Proc. ACM
SIGCOMM, Pisa, Italy, Sep. 2006, pp. 243–254.
be solved using the Bron-Kerbosch algorithm. This is because [4] H. Halabian, A. Zainaldin, and I. Lambadaris, “Optimal joint resource
to get the maximal weighted clique, the calculation needs to allocation and power control in bidirectional relaying networks,” IEEE
happen for each UE and RN for all vertices which results in Trans. Veh. Technol., vol. 63, no. 9, pp. 4520–4535, Nov. 2014.
[5] S. J. Kim, P. Mitran, and V. Tarokh, “Performance bounds for bidirec-
a longer time to solve the assignment problem. This results tional coded cooperation protocols,” IEEE Trans. Inf. Theory, vol. 54,
in a significant increase in the simulation time because the no. 11, pp. 5235–5241, Nov. 2008.
complexity of the algorithm is exponentially dependent on the [6] G. Song and Y. Li, “Cross-layer optimization for OFDM wire-
less networks-part I: Theoretical framework,” IEEE Trans. Wireless
number vertices as shown in Section IV. 3DAP-ACO is able Commun., vol. 4, no. 2, pp. 614–624, Mar. 2005.
to solve the assignment problem but with increased delays [7] G. Song and Y. Li, “Cross-layer optimization for OFDM wire-
(600 msec). Therefore, applying Fast-3DAP and Greedy algo- less networks-part II: Algorithm development,” IEEE Trans. Wireless
Commun., vol. 4, no. 2, pp. 625–634, Mar. 2005.
rithms with limiting the the complexity to the frame duration [8] P. Tarasak and S. Sun, “Resource allocation for downlink two-user
(10 msec) is important. However, as mentioned in section VI-C, OFDMA systems with wireless network coding,” in Proc. 12th Int.
the Greedy allocation is not a feasible algorithm to be deployed Workshop Signal Process. Adv. Wireless Commun, Jun. 2011, pp. 261–
265.
when the number of RBGs is higher. It is important to [9] D. Zhang, Y. Wang, and J. Lu, “Qos aware relay selection and subcar-
note that the simulations are performed in a personal laptop rier allocation in cooperative OFDMA systems,” IEEE Commun. Lett.,
while in real environments, eNBs have high processing and vol. 14, no. 4, pp. 294–296, Apr. 2010.
[10] H. Xu and B. Li, “Xor-assisted cooperative diversity in OFDMA
scheduling power. wireless networks: Optimization framework and approximation algo-
rithms,” in Proc. IEEE INFOCOM, Rio de Janeiro, Brazil, Apr.
2009, pp. 2141–2149.
[11] C. K. Ho, R. Zhang, and Y.-C. Liang, “Two-way relaying over OFDM:
Optimized tone permutation and power allocation,” in Proc. IEEE Int.
Conf. Commun. (ICC’08), Beijing, China, May 2008, pp. 3908–3912.
VII. C ONCLUSION [12] Y. Liu, M. Tao, B. Li, and H. Shen, “Optimization framework and
graph-based approach for relay-assisted bidirectional OFDMA cellular
In this paper, we studied the impact of cooperative relaying networks,” IEEE Trans. Wireless Commun., vol. 9, no. 11, pp. 3490–3500,
(CoR) and Network Coding (NC) on the performance of relay- Nov. 2010.
assisted LTE-advanced networks. An enhanced three-time-slot [13] K. Jitvanichphaibool, R. Zhang, and Y.-C. Liang, “Optimal resource allo-
cation for two-way relay-assisted OFDMA,” IEEE Trans. Veh. Technol.,
per cycle frame architecture is proposed to support direct, CoR, vol. 58, no. 7, pp. 3311–3321, Feb. 2009.
and NC/CoR bidirectional transmission schemes. A combina- [14] M. Tao, Y.-C. Liang, and F. Zhang, “Resource allocation for delay dif-
torial optimization problem was formulated to jointly optimize ferentiated traffic in multiuser OFDM systems,” IEEE Trans. Wireless
Commun., vol. 7, no. 6, pp. 2190–2201, Jun. 2008.
resource block group (RBG) allocation, transmission scheme [15] I. Kim, I.-S. Park, and Y. H. Lee, “Use of linear programming for dynamic
selection and relay selection with the objective of maximiz- subcarrier and bit allocation in multiuser OFDM,” IEEE Trans. Veh.
ing the system total product of backlog and achievable rate. Technol., vol. 55, no. 4, pp. 1195–1207, Jul. 2006.
[16] A. Sinha and T. Kirubarajan, “A randomized heuristic approach for mul-
An equivalent problem was then formulated for the origi- tidimensional association in target tracking,” in Proc. SPIE Signal Data
nal problem for which an enhanced optimal Bron-Kerbosch Process. Small Targets, Apr. 2004, pp. 202–210.
MWCP algorithm was employed to solve the MWCP prob- [17] X. Gao and M. Naraghi-Pour, “Computationally efficient resource allo-
cation for multiuser OFDM systems,” in Proc. IEEE Wireless Commun.
lem. In addition, sub-optimal solutions were proposed based on Netw. Conf. (WCNC’06), Las Vegas, NV, USA, Apr. 2006, pp. 804–809.
an enhanced 3-dimensional Ant Colony Optimization (3DAP- [18] Y. Yu and W. Zhou, “Resource allocation for OFDMA system based on
ACO) algorithm as well as a Greedy heuristic to solve the genetic algorithm,” in Proc. Int. Workshop Cross Layer Des. (IWCLD’07),
Jinan, China, Sep. 2007, pp. 65–69.
MWCP problem. Simulation results confirm that our proposed [19] M. Dorigo, M. Birattari, and T. Stützle, “Ant colony optimization—
algorithm provides the least average system queue backlog Artificial ants as a computational intelligence technique,” IEEE Comput.
which maps to the improved throughput and end-to-end delay Intell. Mag., vol. 1, no. 4, pp. 28–39, Nov. 2006.
[20] H. Ahmadi, Y. H. Chew, and C. C. Chai, “Multicell multiuser OFDMA
performance. Moreover, the enhanced 3DAP-ACO sub-optimal dynamic resource allocation using ant colony optimization,” in Proc.
algorithm significantly reduces the complexity and simulation IEEE 73rd Veh. Technol. Conf. (VTC Spring’11), Budapest, Hungary,
time compared to the optimal solution with minimal effect May 2011, pp. 1–5.
[21] C. Bron and J. Kerbosch, “Algorithm 457: Finding all cliques of an
on our simulation results for the most common deployments. undirected graph,” Commun. ACM, vol. 16, no. 9, pp. 575–577, Sep.
A possible future direction would be to consider distributed 1973.
resource allocation algorithms in order to reduce the strin- [22] 3GPP, “Physical layer procedures,” 3rd Generation Partnership Project
(3GPP), TS 36.213, Sep. 2009 [Online]. Available: http://www.3gpp.
gent time requirement to exchange information and the high org/ftp/Specs/html-info/36213.htm.
feedback complexity of centralized algorithms [32]. Moreover, [23] O. Osterbo, “Scheduling and capacity estimation in LTE,” Adv. Electron.
efficient power allocation schemes should be adapted to the Telecommun., vol. 2, no. 3, pp. 31–42, Sep. 2011.
[24] E. M. Elsheikh and K.-K. Wong, “Optimizing time and power alloca-
resource allocation problem by taking the LTE MCS constraint tion for cooperation diversity in a decode-and-forward three-node relay
into account. channel,” J. Commun., vol. 3, no. 2, pp. 43–52, Apr. 2008.

Authorized licensed use limited to: Mississippi State University Libraries. Downloaded on January 22,2025 at 20:58:29 UTC from IEEE Xplore. Restrictions apply.
ZAINALDIN et al.: JOINT RESOURCE ALLOCATION AND RELAY SELECTION IN LTE-ADVANCED NETWORK USING COP AND NC 4361

[25] L. Tassiulas and A. Ephremides, “Stability properties of constrained Ioannis Lambadaris (M’94) was born in
queueing systems and scheduling policies for maximum throughput in Thessaloniki, Greece. He received the Diploma
multihop radio networks,” IEEE Trans. Autom. Control, vol. 37, no. 12, degree in electrical engineering from the Polytechnic
pp. 1936–1949, Dec. 1992. School, Aristotle University of Thessaloniki,
[26] L. Georgiadis, M. J. Neely, and L. Tassiulas, Resource Allocation and Thessaloniki, Greece, in 1984, the M.Sc. degree in
Cross Layer Control in Wireless Networks. Hanover, MA, USA: Now engineering from Brown University, Providence, RI,
Publisher, 2006. USA, in 1985, and the Ph.D. degree in electrical
[27] C. H. Papadimitriou and K. Steiglitz, Combinatorial Optimization: engineering from the Department of Electrical
Algorithms and Complexity. Mineola, NY, USA: Courier Corp., 1998. Engineering, Systems Research Center (SRC),
[28] E. Tomita, A. Tanaka, and H. Takahashi, “The worst-case time complex- Institute for Systems Research (ISR), University of
ity for generating all maximal cliques and computational experiments,” Maryland, College Park, MD, USA, in 1991. After
Theor. Comput. Sci., vol. 363, no. 1, pp. 28–42, Oct. 2006. finishing his graduate education, he was a Research Associate with Concordia
[29] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein, Introduction to University, Montreal, QC, Canada, from 1991 to 1992. Since September 1992,
Algorithms, 3rd ed. Cambridge, MA, USA: MIT Press, 2009. he has been with the Department of Systems and Computer Engineering,
[30] M. Dorigo, V. Maniezzo, and A. Colorni, “Ant system: Optimization Carleton University, Ottawa, ON, Canada, where he is currently a Professor.
by a colony of cooperating agents,” IEEE Trans. Syst., Man, Cybern. B, His research interests include applied stochastic processes and their application
Cybern., vol. 26, no. 1, pp. 29–41, Feb. 1996. for modeling and performance analysis of computer communication networks
[31] L. M. Gambardella, E. Taillard, and M. Dorigo, “Ant colonies for the and wireless networks, quality-of-service (QoS) control for IP and evolving
quadratic assignment problem,” J. Oper. Res. Soc., vol. 50, no. 2, pp. 167– optical networks architectures and stochastic control/optimization in emerging
176, Feb. 1999. wireless networks. His research is done in close collaboration with his students
[32] M. Moretti, A. Abrardo, and M. Belleschi, “On the convergence and opti- and colleagues in the Broadband Networks Laboratory. He was the recipient of
mality of reweighted message passing for channel assignment problems,” a Fellowship from the National Fellowship Foundation of Greece (1980–1984)
IEEE Signal Process. Lett., vol. 21, no. 11, pp. 1428–1432, Nov. 2014. during his undergraduate studies, and a Fulbright Fellowship (1984–1985)
for graduate studies in the U.S. He was also the recipient of the Technical
Chamber of Greece Award (ranked first in graduating class). While at Carleton
Ahmed Zainaldin (S’13) received the B.Eng. and University, he received the Premiers Research Excellence Award, and the
M.Sc. degrees in electrical and computer engineer- Carleton University Research Excellence Award (2000–2001), for his research
ing from Carleton University, Ottawa, ON, Canada, achievements in the area of modeling and performance analysis of computer
in 2005 and 2008, respectively. He is currently pur- networks.
suing the Ph.D. degree at Carleton University. His
research interests include wireless communications,
network coding, and optimization techniques in net-
work design.

Hassan Halabian (S’10–M’12) received the B.Sc.

and M.Sc. degrees in electrical engineering from
Isfahan University of Technology, Isfahan, Iran, in
2005 and 2008, respectively, and the Ph.D. degree
in electrical and computer engineering from Carleton
University, Ottawa, ON, Canada, in 2012. His
research interests include stochastic network opti-
mization for wireless networks, queueing systems,
and information theory.

Authorized licensed use limited to: Mississippi State University Libraries. Downloaded on January 22,2025 at 20:58:29 UTC from IEEE Xplore. Restrictions apply.