Adaptive Radio Resource Management for
GSM using Neural Networks and Genetic
Algorithms
Ken Murray and Dirk Pesch
Adaptive Wireless Systems Group
Department of Electronic Engineering
Cork Institute of Technology, Cork, Ireland
Tel. +353 21 4326100
FAX: +35
3 21 4326625
E

Mail: {kmurray,
dpesch}@cit.ie
Abstract
With the introduction of 2.5G services, the tele

traffic demand on current GSM networks is expected to grow
exponentially. In this paper we propose a new method
of increasing network capacity by introducing an
adaptive radio resource management system into a typical GSM network. The adaptive radio resource
management system predicts future resource demands for each cell in the network using neural networks
(NNs).
Frequency assignment is then performed using a genetic algorithm (GA). Two methods of frequency
assignment using GAs are presented. The first method obtains optimal solutions in terms of resource
requirements, but produces different frequency assignment p
lans for each assignment cycle. The second
method retains the majority of the frequency assignment plan from each assignment cycle, but at times does
not produce the most optimal solutions.
1
INTRODUCTION
Cellular networks around the world have experien
ced exponential increase in service
demand due to the introduction of 2.5G services such as General Packet Radio Service
(GPRS), High Speed Circuit Switched Data (HSCSD) and Enhanced Data Rates for GSM
Evolution (EDGE). These services result in a considera
ble load on the available radio
resources due to their highly dynamic tele

traffic variations. To maximise system resources,
a more flexible resource management system is needed. A plethora of concepts and
schemes, which attempt to introduce adaptation in
form of dynamic resource allocation
(DRA), have been proposed over the last two decades [1]. Nearly all of the proposed
schemes try to adapt in real

time in a reactive manner to the variation in traffic demands.
This results in either a significant control
signalling load on the fixed network infrastructure
or the requirement for distributed control, which results in considerable change to the
configuration and operation of both terminals and base station equipment.
As an alternative to the reactive approac
h to resource allocation, we consider a proactive
approach, which is based on resource demand prediction using multi

layered feed

forward
neural networks (MFNNs) located within the OMC of a typical cellular network. Resource
predictions are made for each c
ell based on previous load characteristics and the system
resource allocation adapted accordingly using GA techniques. Various other schemes
incorporating neural networks for DRA have been proposed, [2,3,4,5,6] however they carry
high computational costs a
nd function on a real

time basis and thus produce high signalling
loads, which makes their incorporation into current cellular network infrastructures
infeasible. The system proposed here does not operate in real

time, but in a pro

active
manner through ho
urly predictions, which coincide with the typical performance reporting
cycle of existing GSM networks. The centralised implementation within the OMC is ideal
as it does not impact on the cellular infrastructure, the required performance management
data is
located at this location and need not be sent across the network for the operation of
the adaptive resource management system.
2
RESOURCE PREDICTION USING NEURAL NETWORKS
As the number of frequencies allocated to each cell in the network depends on the re
source
predictions, the method of prediction must be robust enough to track the inherent hourly
variations in call traffic. It has been shown that MFNNs have the capability to predict future
values based on training sets composed of sufficient historical d
ata [2]. The system
proposed here considers two MFNNs for each cell site. One MFNN is used to predict future
resources for normal in

cell traffic while the second MFNN predicts guard frequencies for
handover calls. Each MFNN consists of three layers with a
total of 49 neurons. The back

propagation learning algorithm and nonlinear sigmoid activation function are used in the
learning process [7]. The training and prediction of the MFNNs proceeds as follows:
1.
Collect hourly radio resource demand statistics for
in

cell and handover calls. Compile
statistics of previous 8 weeks from the network.
2.
Translate the resource demand into traffic load measurements in Erlangs and derive the
associated radio channel requirements. Record whether the demand occurs on a
weekda
y or weekend (day statistic). Record the time (time statistic). These statistics
constitute the initial training data set.
3.
The MFNNs are trained using the data arising from step 2.
4.
Once the MFNNs are trained, the channel demand for the next hour in each ce
ll is
predicted using the demand statistics from the previous 10 hours, the day and time
statistics.
5.
The predicted number of frequencies is assigned to each base station.
6.
The training set of 8 weeks is updated to contain the statistics for the current hour
(assuming the network gathers statistics at least every 60 minutes).
7.
Each MFNN is retrained every 24 hours to maintain accurate predictions.
As shown in Fig. 1, the neural network resource predictor performs to an excellent degree
of accuracy.
Fig. 1
Neural Network Resource Prediction
3
FREQUENCY ASSIGNMENT USING A GA
An important issue in the design of a cellular network is to determine a spectrum

efficient
and conflict free allocation of frequencies among the cell sites, while satisfying both the
traffic demands and interference constraints. As resource predictions are not made in real

time, a GA embedded in the resource management system is used to find the frequency
assignment plan for a particular demand scenario. We now present two methods of
frequency assignment using GA techniques. The first approach produces an excellent
degree of optimality in terms of resource requirements but suffers from large variations in
successive frequency assignment plans. The second approach ensures that frequenci
es
assigned to a cell include most of the frequencies assigned to that cell in the previous hour,
however, this approach does not produce the most optimal solutions.
3.1
RESOURCE OPTIMISATION USING A GA
Various schemes incorporating GAs to solve the frequenc
y assignment problem have been
proposed in [8,9,10,11]. Such techniques achieve optimality in the range of (80

90%)
depending on the problem complexity. By combining a GA with a local search technique,
we can achieve 96% and 99% optimality for in

cell and
handover frequency assignment
problems respectively. The resource optimisation scheme is presented here.
In the proposed adaptive radio resource management system, the resource predictors make
hourly resource demand predictions. The interference constraint
s for the network can be
represented by an
n
x
n
compatibility matrix
C
, where
n
is the number of cells in the
system.
C
=
Elements x
ij
(
i
,
j
= 1,...,
n
) represent the frequency separation re
quired between frequencies
assigned to cells
i
and
j,
respectively, necessary to maintain interference below a certain
threshold. Using this matrix, it is possible to represent co

channel interference by choosing
values for x
ij
such that,
C
= [x
ij
] =
The traffic demands can be represented by the demand vector
D
, with elements d
i
(
i
=
1,2...
n
) representing the number of required frequencies at cell
i
, the resource predictors fill
out this vector at the end of each hour. The freq
uency assignment problem is then defined
as, given F frequencies and N cells each requiring d
i
frequencies, find an
N
x
F
frequency
assignment matrix
A
given by,
Cells
Cells
Frequencies
A =
such that,
A
= [a
ik
] =
A frequency assignment is admissible if both traffic and interference constraints are
fulfilled. This implies that:
1.
for all
i.
2.
Valid frequencies are assigned to cells acco
rding to the compatibility matrix,
C
.
Two resource vectors of length
n
are derived from the demand vector
D
as shown in Fig. 2.
The first resource vector assigns one frequency to each cell site, so as to sustain a Broadcast
Control Channel (BCCH). The sec
ond resource vector assigns the remaining frequencies.
The GA takes each resource vector and finds a frequency assignment plan using the
minimum number of resources required for that particular demand. The GA works with a
population of 80 individuals, each
individual is represented as follows,
Fig. 2
Determination of Resource Vectors
Cells
Demand Vector:
(3,2,4,5,2,3,........................,5,4,3,4)
Re獯畲se⁖ec瑯爠ㄺ
⠱ⰱⰱⰱⰱⰱⰮ⸮⸮⸮⸮⸮⸮⸮⸮⸮⸮⸮ⰱⰱⰱⰱ,
Re獯畲se⁖ec瑯爠㈺
⠲ⰱⰳⰴⰱⰲⰮ⸮⸮⸮⸮⸮⸮⸮⸮⸮⸮⸮ⰴⰳⰲⰳ
)
Cell group N
Cell
group1
The number of binary ones in each cell group corresponds to the frequency requirements in
the resource vector, i.e. the number of frequencies to
be assigned to the cell.
The roulette wheel selection algorithm is used to generate the parents for the new
population [12]. The new population is created using the standard multi

point crossover and
a modified mutation operator with probabilities 0.6 and
0.0014 respectively. With the
standard mutation operator, each bit is flipped with a predefined probability. In the
algorithm described above, flipping a bit would result in a change to the number of
frequencies been assigned to a cell, therefore, the muta
tion operator used here swaps bits
within a cell group. Only bits that are different are swapped
–
swapping two 1’s or two 0’s
results in no change to an individual and is therefore not considered to be a mutation. The
mutation probability is increased by
approximately 15% if the GA converges to local
minima where the evaluation of the cost function remains at a constant value over a number
of generations.
Crossover can only occur at the end of a cell group, as crossing within a cell group would
alter the
number of frequencies being assigned to that cell.
In order to improve the performance and increase the rate of convergence for difficult
frequency assignment problems, a local search routine has been developed. The search
algorithm finds the cell group a
nd bit position in the solution that is preventing the cost
reaching zero. That bit is then swapped with the other bits in the cell group and the cost of
the solution evaluated. This process continues until the cost reaches zero (optimal solution)
or all t
he bits in the cell group have been swapped.
Fig. 3 illustrates the performance of the GA for a 49 cellular network with a cluster size of
7. In this scenario each cell requires one frequency, the GA reaches this optimal solution by
using 7 frequencies.
Fig. 3
GA Evolutionary Performance
Although this approach produces optimal solutions, the frequency assignment plan for
consecutive hours can be significantly different as shown in Fig. 4. Such large changes to
the frequency assignment plan would intro
duce a considerable amount of inter

cell
handovers into the system as each cell site would be required to change its allocated
frequencies every hour. To execute such large amounts of frequency redeployments is also
very time consuming and would be impract
ical for current GSM networks.
Fig. 4
Consecutive frequency assignment plans for 20 cells using GA resource
optimisation
3.2
ALTERNATIVE APPROACH TO FREQUENCY ASSIGNMENT USING A GA
To overcome the problem of resource optimisation
discussed in section 3.1, we propose an
alternative method of frequency assignment using a GA. This approach does not produce
the most optimal solutions in terms of resource requirements, but ensures that frequencies
allocated to a cell include most of the
frequencies allocated to that cell in the previous hour,
thus minimising inter

cell handover. A number of binary arrays of length
n
are created from
the demand vector,
D
. A binary 1 within an array denotes a cell that requires a frequency.
The first array
represents those cells requiring at least one frequency, the second array for
those requiring at least two frequencies and so on. Each array is then passed to the GA,
which finds the minimum number of frequencies for that demand. Since the GA finds
optima
l solutions for each array separately, the overall solution may be sub

optimal,
however, it does ensure that cells can maintain the majority of frequencies from hour to
hour, as such changes will only be reflected in the last one or two binary arrays, thus
minimizing inter

cell handovers. The optimal solutions from each array are augmented to
create the final frequency assignment plan.
The GA works with an initial population of size 40, each individual in the population is
represented as follows:
Cell gro
up 1
Cell group N
(1,0,0,0,0,0,0,0,1,0,0,0,0,0,….,0,0,0,1,0,0,0)
Each cell group has an initial length of 7, as this is the maximum number of frequencies
required for the first binary array (assuming a cluster size of 7). If the GA find
s a valid
assignment for seven frequencies, a solution is sought for six and so on until no better
solution can be found. Two consecutive frequency assignment plans are shown in Fig. 5,
note how most of the frequencies assigned in the first hour are also u
sed at the same cells in
the second hour.
001000000000000000
1
0
100000000
1
0000000000
01000000000
1
0000
1
000
0000010000000
1
000000
00010000000000000
1
00
0000100000
1
000000000
000000100000000
1
0000
00001000000010
1
00000
00000010
1
000000000
0
1
00100000001000000000
100000000
1
0000
1
00000
0100000000000000
1
0
1
0
00000100100000000000
00010000000
1
00000000
000001000000000
1
0000
00010001000000000
1
10
00001000000010000000
00000010000
1
000
1
0000
0010000000000
1
000000
1000000
1
00000000000
1
0010000
1
00000000000
100000000000000000
1
010000000
1
000
1
00000
00000100
1
0000000000
00010000000
000
1
0000
00001000000
1
0000000
0000001000000000
1
00
00001000000010000
1
0
00000010000000
1
0
1
00
0010000000100000000
10000000000000000
11
0100000
1
00000
1
00000
0000010010000000000
000100000000000
1
000
00000100
1
0010000000
000100010000000
1
000
000010000000100
0000
000000100
1
000000
1
00
0010000000
1
00000000
10000000000000
1
000
1
Fig. 5
Consecutive frequency assignment plans for 20 cells using GA frequency
assignment
4
SYSTEM IMPLEMENTATION IN GSM
The proposed adaptive radio resource management system is integrated i
n the OMC
between the performance management (PM) tool and the configuration management (CM)
tool, see Fig. 6.
As all the network performance data required for the system is available
at this location, no additional signalling load is generated. The OMC l
ocation also has the
advantage that the system can be implemented without the need of any hardware or
software updates to mobile and base station equipment. The non

invasive nature of the
proposed concept is one of its major advantages in that it can be im
plemented and also
improved without interference with existing equipment.
Network Statistics
Frequency Assignment
Fig. 6
Adaptive Radio Resource Manager
OMC
PM
Pr
edictive
RRM
CM
000100
00001000000000
00000010000001000000
00001000000100000100
01000000100000000000
10000001000000000000
00000100000010000000
00100000010000000000
010000001000000
1
0000
100000010000000000
1
0
00000100000010000000
0010000001000000
1
000
00010000001000
1
00000
0000001000
0001000000
00001000000100000000
00100000010000000000
00010000001000100001
00000010000001000000
000010000001000
1
0000
01000000100000000000
10000001000000000010
00010000001000000000
00000010000001000000
00001000000100000100
01000000100000000000
10000001000000000000
00000100000010000000
00100000010000000000
0100000010000000
1
000
1000
00010000000
1
0000
00000100000010000000
00100000010000
1
00000
000100000010000
1
0000
00000010000001000000
00001000000100000000
00100000010000100000
00010000001000000010
00000010000001000000
0000100000010000
1
000
01000000100000000000
10000001000000000010
5
NETWORK SIMULATION AND RESULTS
Two simulation models have been developed
–
a Fixed
Channel Allocation (FCA) model,
which is used in current GSM networks and a model based on the proposed adaptive
resource management system with the embedded GA frequency assignment schemes. Both
network models contain 49 cells with a cluster size of 7, s
ee Fig. 7. Each cell has a total of
18 neighbours with wraparound performed at the borders in the x and y planes. The load in
each model is non

uniformly distributed across the network. The call arrival rate,
, is
Poisson distributed while the call holdin
g time, 1/
, has a mean of 3 minutes. The number
of handover arrivals in each cell will depend on the load in each of the six surrounding
cells, the handover arrival rate is therefore taken to be 5% of the sum of the call arrivals in
these cells.
The perfo
rmance of both models is measured by the number of frequencies required to
maintain the call blocking below 2% throughout the network.
Fig. 7
49

cellular structure
5.1
FCA NETWORK MODEL
Each cell in the FCA model were assigned the requi
red number of frequencies so as to
maintain the call blocking below 2%. The network was monitored over a 2

week simulation
period. Performance measurements were averaged over 700 simulation runs so as to
achieve accurate results. Fig. 8 shows the average b
locking for cell site 2. A total of 29
frequencies were required in the network for new call arrivals, while 14 guard frequencies
achieved a call dropping rate of zero.
5.2
ADAPTIVE RADIO RESOURCE MANAGEMENT
The same call traffic scenario was used in thi
s model. Unlike the FCA concept, cells are
assigned frequencies based on resource predictions and the GA frequency assignment
schemes discussed in section 3. Using the resource optimisation GA, a total of 22
frequencies were required for new call arrivals,
producing a resource gain of 24%. Since
this scheme results in unpractical frequency assignment plans, it is not considered to be a
x
1
2
4
3
5
12
11
10
9
8
15
14
13
7
6
43
36
29
22
21
20
19
18
17
16
23
24
30
37
44
31
38
25
26
32
27
35
28
42
49
45
39
46
33
40
47
34
41
48
y
suitable method of resource assignment for the adaptive radio resource management
system.
Simulation results using the alte
rnative GA frequency assignment scheme show resource
requirements of 23 frequencies, giving a resource gain of 20.7% when compared with the
equivalent FCA network. This result is comparable to current DRA schemes [1], but with
the advantage of significantl
y less complexity and no additional signalling load.
It was observed however that no resource gain could be obtained from adaptive guard
channel allocation. As each cells handover call arrival rate depends on the six surrounding
cells, handover calls tend
to be more uniformly distributed than new call arrivals. The
average call blocking over 700 simulation runs is shown in Fig. 9. By comparing Figs. 8
and 9, the similarities in call blocking at peak times can be seen. The additional blocking in
the adaptiv
e network arises because each cell is allocated just the required number of
frequencies for the next hour, thus maximising the system resources. The simulation results
are summarised in Table 1.
Table 1.
Simulation Results
Network Model
FCA
Adaptive Re
source
Management, GA
resource optimisation
Adaptive Resource
Management, GA
Frequency Assignment
Frequency Requirements for new
call arrivals
29
22
23
Frequency Requirements for
handover calls
14
14
14
Resource Gain

24%
20.7%
Fig. 8
Average ca
ll blocking in FCA
Fig. 9
Average call blocking in
network
adaptive network
6
CONCLUSION
An adaptive radio resource management system for current GSM networks has been
proposed. Simulation results have shown resource benefit
s of up to 20.7% when compared
with an equivalent FCA network. Using frequency deployment mechanisms such as those
discussed in [13], this approach can achieve self

configuring cellular networks without the
need of additional signalling loads and changes t
o both terminals and base station
equipment.
ACKNOWLEGEMENTS
The authors acknowledge the financial support of Enterprise Ireland and Motorola’s
European Cellular Infrastructure Division under grant AR/2000/36 in the funding of the
work reported in thi
s paper.
REFERENCES
[1]
I. Katzela, M. Naghshineh, “Channel Assignment Schemes: A Comprehensive Survey”, IEEE Personal
Communications, June 1996.
[2]
Peter T. H. Chan, Marimuthu Palaniswarni, and David Everitt, “Neural Network

Based Dynamic Channel
Assig
nment for Cellular Mobile Communication Systems”,
IEEE Trans. Veh. Technol
., vol. 43, pp. 279

288, May 1994.
[3]
Dietmar Kunz, “Channel Assignment for Cellular Radio Using Neural Networks”,
IEEE Trans. Veh.
Technol
., vol. 40, pp. 188

193, Feb. 1991.
[4]
En
rico Del Re, Romano Fantacci, and Luca Ronga, “A Dynamic Channel Allocation Technique Based on
Hopfield Neural Networks”,
IEEE Trans. Veh. Technol
., vol. 45, pp. 26

32, Feb. 1996.
[5]
Harilaos G. Sandalidis, Peter P. Stavroulakis, and Joe Rodriguez

Tellez,
“Borrowing Channel
Assignment Strategies Based on Heuristic Techniques for Cellular Systems”,
IEEE Trans. Neural
Networks
, vol. 10, pp. 176

181, Jan. 1999.
[6]
Nobuo Funabiki, and Yoshiyasu Takefuji, “A Neural Network Parallel Algorithm for Channel
Assign
ment Problems in Cellular Radio Networks”,
IEEE Trans. Veh. Technol
., vol. 41, pp. 430

437,
Nov. 1992.
[7]
S. Haykin, “Neural Networks: A Comprehensive Foundation”, Prentice

Hall, 1994.
[8]
W. K. Lai, George G. Coghill, “Channel Assignment Through Evol
utionary Optimization”,
IEEE Trans.
Veh. Technol
., vol. 45, pp. 91

96, Feb. 1996.
[9]
Goutam Chaktaborty and Basabi Chakraborty, “A Genetic Algorithm Approach to Solve Channel
Assignment Problem in Cellular Radio Networks”,
IEEE Midnight

Sun Workshop on So
ft Computing
Methods in Industrial Applications
, Kuusamo, Finland, June 16
th
–
18
th
, 1999.
[10]
Dirk Beckmann and Ulrich Killat, “A New Strategy for the Application of Genetic Algorithms to the
Channel Assignment Problem”,
IEEE Trans. Veh. Technol
., vol. 48
, pp. 1261

1269, July 1999.
[11]
Chiu Y. Ngo, and Victor O. K. Li, “Fixed Channel Assignment in Cellular Radio Networks Using a
Modified Genetic Algorithm”,
IEEE Trans. Veh. Technol
., vol. 47, pp. 163

172, Feb. 1998.
[12] Goldberg, “Genetic Algorithm in Se
arch, Optimization and Machine Learning”, Addison Wesley, 1999.
[13]
M. Perez

Carbonell, D. Pesch, P. Stephens “Optimum Frequency Deployment in Cellular Mobile
Networks using Genetic Algorithms”, Irish Signals and Systems Conference, Maynooth, Ireland, Jun
e
2001.
Comments 0
Log in to post a comment