1

Load Balancing by Dynamic BBU-RRH Mapping in

a Self-Optimised Cloud Radio Access Network
M.Khan, Student Member, IEEE, Firas A. Sabir, H.S. Al-Raweshidy, Senior Member, IEEE

Abstract—In this paper, a load-balancing scheme is investigated structure. However, it is not preferable for traditional LTE
for C-RAN to optimise its performance subject to physical network due to slow parameter updates. It requires a centralised
resource limitation and users Quality of Service (QoS) demands SON controller to gather information from distributed eNodeBs
constraints. The Remote Radio Heads (RRHs) are allocated to
the Base Band Units (BBUs) dynamically for load balancing which in turn receive feedbacks from the User Equipment (UEs)
and efficient resource utilisation. Dynamic BBU-RRH mapping associated with them.
is formulated as a linear integer-based constrained optimisation However, Cloud Radio Access Network (C-RAN) [5], [6] is a
problem. An Estimation Distribution Discrete Particle Swarm novel architecture which moves the primary signal processing
Optimisation (EDDPSO) algorithm is developed for optimisation
functions performed at the Base Band Units (BBUs) of the
and to achieve a balanced network. Computational results based
on two benchmark problems demonstrate that EDDPSO delivers typical eNodeBs to a centralised location called BBU-Cloud.
optimum performance for small-scale networks. However, a And the Remote Radio Heads (RRHs) with fewer functions
close-optimum is achieved for large-scale networks which is maintains the radio access at the cell sites and can be deployed
99.06% of the optimum value. The EDDPSO algorithm is fast densely with minimum cost. The main features of C-RAN are
and less complex compared to Exhaustive Search (ES) algorithm.
centralised management of computing resources, reconfigura-
Index Terms - Base Band Unit (BBU), Cloud Radio Access tion of spectrum resources, collaborative communications, and
Network (C-RAN), Particle Swarm Optimisation (PSO), Remote Real-time cloud computing on generic platforms [7].
Radio Head (RRH), Self-Optimising Network (SON). C-RAN can redistribute network resources (or capacity)
across a given geographical area with respect to time-varying
traffic load. The main contribution of this paper is to represent
I. INTRODUCTION
a self-optimising C-RAN which can resolve uneven network
In the past few years, the accelerated growth in personal loads. Load balancing in the network is formulated as an
mobile computing devices such as tablets and smartphones, optimisation problem where multiple RRHs are mapped to
along with the upsurging volume of data-intensive services different BBU resources depending on time-varying traffic. The
and mobile applications, has resulted in a massive demand for aggregated BBU resources are allocated to the RRHs according
wireless access and high-speed data transmission. A common to an algorithm. The rest of the paper is organised as follows:
practice to enhance data rate, capacity, and coverage is to den- Section II presents the system model. Section III illustrates
sify the network by increasing Base Stations (BSs) and small the formulation for dynamic RRH allocation problem. Section
cells deployment for greater bandwidth re-use. However, den- IV define the Self-Optimising resource allocation algorithm.
sifying the access network brings huge inefficiencies regarding Computational results of different algorithms are discussed in
Capital (CAPEX) and Operational (OPEX) expenditures due to Section V. Finally, the paper is concluded in Section VI.
additional deployment and maintenance. Moreover, the spatial
distribution of subscribers and capacity demands vary with
time, causing uneven traffic loads and inefficient BSs spectral II. SYSTEM MODEL
and processing resources utilisation. Therefore, it is important
A. C-RAN Architecture
to self-optimise the network resources dynamically on varying
user distribution and their capacity demands. Self-optimisation A self-optimised C-RAN architecture is presented in Fig. 1.
is an important feature of Self-Organising Networks (SON). The BBUs are decoupled from the RRH and migrated to a
However, dynamic inter-cell optimisation itself is a major issue centralised BBU-pool, whereas the RRHs with simple Radio
in SON [1]. Frequency (RF) transmission functionality are left on the cell
Numerous studies on SON have suggested load balancing sites. A SON controller inside the BBU cloud monitors the
as a self-optimisation problem according to different metrics. BBU-pool resource utilisation as well as controls the switch
These include controlling handover parameters, dynamic allo- fabric. Since an optical switch can only support one-to-one
cation of power and channel, Antenna tilt optimisation , Mo- switching, soft switching (one-to-one and one-to-many) is
bility Robustness Optimisation (MRO), mobility load balancing enabled indirectly by using optical splitters and multiplexers
(MLB), and Automatic Neighbor Relation (ANR) [2]. SON can [8]. The SON controller dynamically assigns BBU radio re-
be divided into three different architectures, i.e., a) Centralised, sources to the independent RRHs based on traffic demands.
b) Decentralised, and c) Hybrid [3], [4]. The centralised ar- Note that, there are various possibilities of front-haul exists
chitecture can readily approach a global optimum network for C-RAN [9]. Moreover, optical switches are advantageous

978-1-5386-0643-8/17/$31.00 ©2017 IEEE

Authorized licensed use limited to: Indian Institute of Technology Hyderabad. Downloaded on April 07,2021 at 15:05:53 UTC from IEEE Xplore. Restrictions apply.
 2

BBU Cloud III. DYNAMIC BBU-RRH CONFIGURATION AND

Switch Fabric FORMULATION
Optical

BBU pool
Electronic
or optical
Splitters
&
Radio over
Fibre
For a self-optimised C-RAN introduced in Section II and
Multiplexer
Fig. 1, it is critical to balance the load amongst BBUs by
switch(s) (RoF)

Optical Fibre
proper BBU-RRH mapping. If the BBU-RRH configuration at
S/M RoF
𝐁𝐁𝐔𝟏
𝐑𝐑𝐇𝟏

SON
controller
𝐁𝐁𝐔𝟐 S S/M
RoF
time t is known then it is important to modify the BBU-RRH
configuration at time t+1 to balance the load across the BBUs.
𝐑𝐑𝐇𝟐
W RoF
: I :
: T
:
: 𝐑𝐑𝐇𝟑
Note that, the time between t and t + 1 is longer than that of a
:
C
H S/M
:
: subframe (i.e., one millisecond) and is called the load balancing
𝐁𝐁𝐔𝑴−𝟏
RoF
cycle. A user location vector l={l1 , l2 , l3 , ..., lk } is utilised,
𝐁𝐁𝐔𝑴
S/M
RoF
𝐑𝐑𝐇𝑵−𝟏 where lk 2 {1, 2, 3, ..., N }, lk = n if the uplink received
power of user k at RRHn is higher than at all other RRHs.
Switch control
𝐑𝐑𝐇𝑵
To identify the BBU-RRH allocation, a new RRH allocation
vector R is defined such that R = {R1 , R2 , R3 , ..., RN },
Rn 2 {1, 2, 3, ..., M }, and Rn = m if RRHn is associated with
BBUm . If the user location vector l is given, then the problem
Fig. 1. Cloud Radio Access Network represented as SON
is to identify the new RRH allocation vector R.

over electronic switches regarding cost, power, and data rate. A. Key Performance Indicator for Load Fairness Index
However, they may incur longer reconfiguration times. Let ⌘m (t) be the load of a BBU at time period (t), which is
represented as
PK
Im,k (t)NRB,K (t)
B. link model ⌘m (t) = k=1 (4)
PRB
As shown in Fig. 1, the BBU-pool aggregates N number of Where Im,k is a binary indicator such that Im,k = 1 if
BBUs with M number of RRHs covering the entire geograph- user
ical area. K is the number of users in the network such that PM k is served by BBUm . However, an important constraint
m=1 Im,k = 1, 8k defines that each user k is served by only
Kni = Gni [ Bni , where Gni and Bni are the Guaranteed one BBU at time period t. Note that, all BBUs are assigned
Bit Rate and Best Effort users of cell i served by RRHn , the
respectively. Each user reports its Channel Quality Information PKsame number of PRBs. Another important constraint is that
k=1 Im,k (t)NRB, K (t) < PRB , 8m, which states that the number
(CQI) to the serving BBU for Physical Resource Blocks (PRBs) of PRBs assigned to users served by the same BBU should
assignment. Based on the CQI received at the BBU, the Signal- not exceed the BBU PRB limitation. Moreover, the amount of
to-Noise-and-Interference Ratio of user Kni at the pth PRB resource shortage (or PRB shortage) in the network based on
of sub-frame ⌧ is given as users PRB hdemand can be iestimated as Resource Shortage =
PM
m=1 max (⌘m (t) 1), 0 ⇥ 100 To measure how balanced
Gkni ,p (⌧ )⇢ni ,p (⌧ ) the network is, a load fairness index is monitored by evaluating
kni ,p (⌧ ) = P (1)
N0 + nj 2N,nj 6=ni Gknj ,p (⌧ )⇢nj ,p (⌧ )
the load distribution in all cells. The Jain’s fairness Index [10]
regarding network load can be given as
where ⇢ni ,p (⌧ ) and Gkni ,p (⌧ ) is the transmit power and ⇣P ⌘2
M
channel gain between RRHn of cell i and user k at pth PRB of m=1 ⌘ m (t)
(t) = ⇣P ⌘ (5)
subframe ⌧ and N0 is the power of Additive White Gaussian M 2
|M| m=1 ⌘m (t)
Noise per PRB. Then the average spectral efficiency of user
kni at a time instance t is given as (t) ranges between the interval [ M
1
, 1]. A higher (t) value
represents a more balanced load distribution among the BBUs.
8 9 Another objective is to maximise the value of (t).
1 < X X ⇥ ⇤=
#kni (t) = log2 1 + kni ,p (⌧ ) (2)
PRB .N⌧ : ; B. Key Performance Indicator for Handover Index
⌧ 2[t 1,t] p2PRB
Network transition to a new BBU-RRH configuration may
where PRB is the total number of PRBs assigned to a BBU require significant handovers. Unnecessary handovers in the
and N⌧ is the number of sub-frames between two BBU- system are undesirable and lead to performance degradation.
RRH configuration times. To keep up with the requirement Assigning an RRH to a different BBU at a particular time
throughput k , the number of PRBs required by user Kni at a requires an inter-BBU handover of all users associated with
time period t can be calculated by multiplying the achievable the RRH, which cause increased signalling overhead in the
user throughput to the PRB bandwidth (i.e., 180 KHz per PRB) core network. Therefore, handovers needed for a new BBU-
RRH mapping transition should be at a minimum possible. In
k (t) this paper, a handover index is monitored and defined as
NRB,K (t) = (3) !
pBW .#kni (t) PM PK
1 m=1 k=1 Im,k (t) Im,k (t)
h(t) = (6)
where pBW represents the bandwidth of a PRB. 2 K

Authorized licensed use limited to: Indian Institute of Technology Hyderabad. Downloaded on April 07,2021 at 15:05:53 UTC from IEEE Xplore. Restrictions apply.
 3

where Im,k (t) is a binary indicator of users association in 0.729, 2.05, and 2.05, respectively [12]. Equation (8) shows
previous BBU-RRH configuration i.e., Im,k (t)=1, if user k is that each particle partially depends on its previous experience
served by BBUm in previous BBU-RRH configuration. (personal memory) and partially on the social experience (social
This paper defines a QoS function which is a weighted memory). The memory helps each particle to probe the solution
combination of the Key Performance Indicators (KPIs) set in space somewhere around its own best position and the global
(5) and (6). The multiple KPI objectives are combined into a best position. However, this might lead to searching the same or
single QoS objective function with constraints and defined as approximately the same solution space region again and again
without any noticeable progress. Therefore, the distribution of
Max QoS(t) = ↵ (t) h(t) favourable zones in the solution space is estimated to help the
PSO escape from the local optima. Information obtained during
K
X the optimisation process is exploited to place the particles
s.t. C1 : Im,k (t)NRB,K (t)  PRB , 8m 2 {1, 2, ...M}
within favourable regions. The idea is taken from Ant Colony
k=1
Optimisation [13] in which the solution space (or search space)
M
X (7) is explored based on an archive of solutions maintained during
C2 : Im,k (t) = 1, 8k 2 {1, 2, ..., K}
the optimisation process. The EDDPSO supports an array of
m=1
pbests to estimate the distribution. In EDDPSO, the size of
M
X solution archive equals the swarm size.
C3 : Im,n (t) = 1, 8n 2 {1, 2, ..., N}
m=1
After the velocity update execution, each particle’s position
PM is updated according to a new policy, i.e., the pbests solution
The constraint m=1 Im,n (t) = 1, 8n shows that an RRHn archive is sorted in descending order of the quality of solutions.
can be served P by a single BBUm at a particular time t. And A weight wj is associated with each particle solution Rj
M
the constraint m=1 Im,k (t) = 1, 8k shows that each user is according to the following Gaussian function [13]
served by a singleP BBU at a particular time t. Another important
K (rank(j) 1)2
constraint is that k=1 Im,k (t)NRB,K (t)  PRB , 8m, which states 1
wj = p e 2q 2 | |2 (10)
that the number of PRBs assigned to users served by the same q| | 2⇡
BBU should not exceed the BBU PRB limitation. The ↵ and
| | is the swarm size. The particle with highest solution quality
are control parameters of the QoS function. The main objective
is assigned the highest weight whereas the weight of other
is to maximise the QoS function.
particles decreases exponentially with respect to their rank. The
probability of selecting a particle from the solution archive is
IV. S ELF - OPTIMISED C LOUD R ADIO ACCESS N ETWORK
given as:
In this section, an Evolutionary Algorithm (EA) is proposed wl
⇢l = P | | (11)
which serves as an optimiser for the SON controller and iden-
j=1 wj
tifies optimum BBU-RRH configuration. Network information
is utilised to analyse the QoS of current BBU-RRH setup and The standard deviation li of the ith component of the lth
for other possible BBU-RRH configurations. selected particle from the archive is given by
| |
A. Estimation distribution discrete particle swarm optimisation X |xi xil |
i a
l =⇠ (12)
(EDDPSO) a=1
| | 1
Particle Swarm Optimisation (PSO) algorithm is inspired by
The parameter ⇠ is utilised to balance the exploration be-
the social behaviour of flocking organisms [11]. PSO method
haviour of PSO. The lower the value of ⇠, the higher con-
uses Swarm Intelligence for solving optimisation problems.
vergence rate of the algorithm. Note that, the optimisation
Each particle in a swarm represents the BBU-RRH allocation
problem in this paper is discrete, i.e., the BBU-RRH allocation
solution R. A particle j in iteration I has a position xIj and
solution should be real-valued (RRH allocation vector R is
velocity vjI . The fitness function defined in (7) determines the
real-valued). Therefore, an estimation distribution of ’discrete’
quality of each particle. The particle velocities are updated
particle swarm optimisation is applied to the QoS minimisation
at each iteration based on the historical best position (pbest)
problem defined in (7). A pseudo-code of EDDPSO is shown
visited by the particle itself and the best position ever visited by
in Algorithm 1.
any particle in the swarm (gbest) at any iteration. The velocity
The Gaussian functions defined in (10), (11), and (12) are
update rules for each particle j are given as
used to move the particles to a new position. If the particle
vjI+1 = (vjI + '1 U1 (0, 1) ⇤ (pbest xIj ) movement is successful, the algorithm proceeds as normal,
(8) however, in case of unsuccessful particle flow, the value of j th
+ '2 U2 (0, 1) ⇤ (gbest xIj ))
component of particle i is sampled with respect to its neigh-
with bourhood. The sampling is done using a normal probability
2k
= p (9) density function with mean µ and standard deviation . where
2 ' '2 4' µ equals the value of the j th component.
where '1 and '2 are constants known as the cognitive and 1 (x µ)2
g(x, µ, ) = p e 2 2 (13)
social coefficients, respectively. ' = '1 + '2 and ' > 2⇡
4, k 2 [0, 1]. The values for , '1 , and '2 are set to

Authorized licensed use limited to: Indian Institute of Technology Hyderabad. Downloaded on April 07,2021 at 15:05:53 UTC from IEEE Xplore. Restrictions apply.
 4

Algorithm 1: Estimation Distribution Discrete Particle algorithm. The Swarm size (| |) for P1 and P2 are set to 2⇥103 .
Swarm Optimisation Algorithm Which is < 1% of the entire solution space. The maximum
1 for i = 1 to | | do number of iterations is 200. The parameters '1 = '2 = 2.05,
2 Generate Particle i with random position and velocity = 0.729, q = 0.1, and ⇠ = 0.85. This paper uses MATLAB
3 end as a simulation platform for computational results.
4 Initialise all pbesti and gbest The QoS, load fairness and handover indexes, and the
5 I=0 average network load for both P1 and P2 over 200 iterations
6 while I < Imax do are shown in Fig. 2, Fig. 3, and Fig. 4. Note that, the optimum
7 for i = 1 to | | do values are obtained by exhaustively searching for all BBU-RRH
8 if f (xi ) > pbesti then configurations (i.e., |M ||N | ) using ES algorithm. Where M is
9 pi = x i the number of BBUs, and N is the number of RRHs in the
10 pbesti = f (xi ) benchmark problems. The ES algorithm does not depend on
11 end the number of iterations. The optimal values are necessary to
12 if pbesti > gbest then analyse the improvement achieved at each iteration. This paper
13 gbest = pbesti defines convergence rate (CR) of the algorithm as the number
14 s = pi of times an optimal or best solution is achieved after a total
15 end number of iterations are performed.
16 end Fig. 2 demonstrates the convergence of EDDPSO towards the
17 Rank all pbesti w.r.t quality optimal solution. In P1 , the CR is 0.935, where the optimum
18 Compute weights using equation (10) solution is obtained 187 times over 200 iterations. In P2 , the
19 Compute selection probability ⇢ using equation (11) CR is 0.135. However, the optimum solution is not achieved by
20 for i = 1 to | | do EDDPSO but the best value given is 99.06% of the optimum
21 value and is produced 27 times over 200 iterations. The
EDDPSO algorithm performs 173 ⇥| | fitness evaluations to
viI+1 = (viI + '1 U1 (0, 1) ⇤ (pbesti xIi ) reach the best solution whereas the ES algorithm performs 749 ,
+ '2 U2 (0, 1) ⇤ (gbestj xIi )) which is too enormous.
Fig. 3 shows the EDDPSO convergence towards optimum
0
xi = xi + v i load fairness index. For P1 , the optimum is achieved after 11th
iteration. In P2 , although the optimum is not achieved over 200
22 Choose particle l from solution archive highest
iterations. However, the best load fairness index value is 98%
probability
of the optimum value and is reached after 169 ⇥ | | fitness
23 Compute li from p equation (12)
0 evaluations which are too small compared to the extensive 749
24 P robmove = li 2⇡ (xi ) (|M ||N | ) evaluations of ES algorithm. Fig. 4 shows that the
25 if U (0, 1) < P robmove then handover index converges to the optimal and best values after
0
26 xi = xi 11th and 171th iteration in P1 and P2 , respectively. Although,
27 else the optimal value for handover index is not achieved in P2 ,
28 for j = 1 to n do however, the best value achieved is 99% of the optimum value
29 xij = g(xij ) given by ES. Note that, the EDDPSO can achieve optimum
30 end solution for P2 if the number of iterations (Imax ) or the
31 end Swarm size (| |) is increased. The ↵ and control parameters
32 end in (7) are selected by performing an exhaustive search (ES)
33 I =I +1 algorithm to identify the optimal BBU-RRH setting for P1 .
34 end Both ↵ and values are orderly set to 0, 0.1, ..., 1 with
a constraint ↵ +  1. An optimal BBU-RRH setting is
found using ES algorithm for each pair of ↵ and . Setting
V. C OMPUTATIONAL R ESULTS AND A NALYSIS values for ↵ > 0.8 results into improper BBU-RRH mapping
which implies that maximising network load balance is overly
Two benchmark problems, P1 and P2 are considered such that considered compared to minimising handovers, which results in
P1 consists of 5 BBUs serving 19 RRHs and P2 consists of 7 increased resource shortage. This paper considers ↵ = 0.8 and
BBUs serving 49 RRHs. The RRHs are uniformly distributed, = 0.1, assigning a 10% weight to handover minimisation.
and a fixed distance of 90 meters is set between two adjacent
RRHs. Note that, the actual distances between RRHs are deter-
VI. C ONCLUSION
mined by coverage area, users density and other environment
related factors. However, in this paper distances between RRHs BBU-RRH allocation in C-RAN is examined with an aim to
are 90 meters as a reasonable example when small cells are balance network load dynamically. Proper BBU-RRH allocation
considered. A bandwidth of 20 MHz (100 PRBs) and a path is formulated as a linear integer-based programming problem
loss exponent of 3.76 is considered. However, different path which maximises the network QoS and load fairness index
loss models can be utilised. An exhaustive search for the while minimising the handovers. The EDDPSO algorithm is
optimum solution is obtained for both benchmark problems via developed to resolve the BBU-RRH configuration problem,
ES algorithm and compared to results obtained by EDDPSO and its performance is compared to ES algorithm. Based on

Authorized licensed use limited to: Indian Institute of Technology Hyderabad. Downloaded on April 07,2021 at 15:05:53 UTC from IEEE Xplore. Restrictions apply.
 5

two benchmark problems, the EDDPSO algorithm delivers 0.44

noticeably faster convergence than ES algorithm. The proposed EDDPSO, 19 RRH 5 BBUs
ES, 19 RRH 5 BBUs

concept can be extended to develop Network Function Virtual-

EDDPSO, 49 RRH 7 BBUs
0.42 ES, 49 RRH 7 BBUs

isation (NFV) and Software Defined Network (SDN) solutions

for C-RAN, which are important concepts that can help to 0.4

implement virtualisation of baseband resources.

Handover Index
0.38

0.36
0.7

0.34
0.6

0.32
0.5
Quality of Service (QoS)

0.3
0.4 0 20 40 60 80 100 120 140 160 180 200
EDDPSO, 19 RRH 5 BBUs
ES, 19 RRH 5 BBUs
Iterations
EDDPSO, 49 RRH 7 BBUs
ES, 49 RRH 7 BBUs
0.3

Fig. 4. Handover Index values for ES, and EDDPSO in benchmark problem
0.2 P1 and P2

0.1

[6] M. Peng, Y. Sun, X. Li, Z. Mao, and C. Wang, “Recent advances in

0 cloud radio access networks: System architectures, key techniques, and
open issues,” IEEE Communications Surveys & Tutorials, vol. 18, no. 3,
0 20 40 60 80 100 120 140 160 180 200
Iterations
pp. 2282–2308, 2016.
[7] D. Pompili, A. Hajisami, and T. X. Tran, “Elastic resource utilization
Fig. 2. Quality of Service (QoS) values for ES, and EDDPSO in benchmark framework for high capacity and energy efficiency in cloud RAN,” IEEE
problem P1 and P2 Communications Magazine, vol. 54, no. 1, pp. 26–32, 2016.
[8] K. Sundaresan, M. Y. Arslan, S. Singh, S. Rangarajan, and S. V.
Krishnamurthy, “Fluidnet: a flexible cloud-based radio access network
for small cells,” IEEE/ACM Transactions on Networking, vol. 24, no. 2,
1 pp. 915–928, 2016.
[9] M. Peng, C. Wang, V. Lau, and H. V. Poor, “Fronthaul-constrained
cloud radio access networks: Insights and challenges,” IEEE Wireless
0.95
Communications, vol. 22, no. 2, pp. 152–160, 2015.
EDDPSO, 19 RRH 5 BBUs [10] A. B. Sediq, R. H. Gohary, R. Schoenen, and H. Yanikomeroglu, “Optimal
ES, 19 RRH 5 BBUs
EDDPSO, 49 RRH 7 BBUs
tradeoff between sum-rate efficiency and Jain’s fairness Index in resource
allocation,” IEEE Transactions on Wireless Communications, vol. 12,
0.9
ES, 49 RRH 7 BBUs
Load Fairness Index

no. 7, pp. 3496–3509, 2013.

0.85
[11] J. Kennedy, “Particle swarm optimization,” in Encyclopedia of machine
learning. Springer, 2011, pp. 760–766.
[12] I. C. Trelea, “The particle swarm optimization algorithm: convergence
0.8 analysis and parameter selection,” Information processing letters, vol. 85,
no. 6, pp. 317–325, 2003.
[13] T. Liao, K. Socha, and M. A. M. de Oca, “Ant colony optimization for
0.75
mixed-variable optimization problems,” IEEE Transactions on Evolution-
ary Computation, vol. 18, no. 4, pp. 503–518, 2014.
0.7
0 20 40 60 80 100 120 140 160 180 200
Iterations

Fig. 3. Load Fairness Index values for ES, and EDDPSO in benchmark
problem P1 and P2

R EFERENCES
[1] E. U. T. R. A. Network, “Self-configuring and self-optimizing network
(son) use cases and solutions,” Third Generation Partnership Project
(3GPP) specification TR 36.902 V9.3.1, vol. 36, 2009.
[2] O. G. Aliu, A. Imran, M. A. Imran, and B. Evans, “A survey of self
organisation in future cellular networks,” IEEE Communications Surveys
& Tutorials, vol. 15, no. 1, pp. 336–361, 2013.
[3] A. Misra and K. K. Sarma, “Self-Organization and Optimization in Het-
erogenous Networks,” Interference Mitigation and Energy Management
in 5G Heterogeneous Cellular Networks, p. 246, 2016.
[4] J. Li, J. Zeng, X. Su, W. Luo, and J. Wang, “Self-optimization of coverage
and capacity in LTE networks based on central control and decentralized
fuzzy Q-learning,” International Journal of Distributed Sensor Networks,
vol. 8, no. 8, p. 878595, 2012.
[5] T. Q. Quek, M. Peng, O. Simeone, and W. Yu, Cloud radio access net-
works: Principles, technologies, and applications. Cambridge University
Press, 2017.

Authorized licensed use limited to: Indian Institute of Technology Hyderabad. Downloaded on April 07,2021 at 15:05:53 UTC from IEEE Xplore. Restrictions apply.