You are on page 1of 10

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: +353 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 plans 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 experienced 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 considerable 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 approach 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 cell 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 and 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 hourly 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 resource
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 data [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 weekday
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 cell 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
Neural Network Prediction
0
5
10
15
20
25
30
35
1
1
1
2
1
3
1
4
1
5
1
6
1
7
1
8
1
9
1
1
0
1
1
1
1
1
2
1
1
3
1
1
4
1
1
5
1
1
6
1
Time (hr)
C
h
a
n
n
e
l

D
e
m
a
n
d
Actual
Prediction
Cell group1
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 frequencies 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 frequency 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 constraints 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 =
1
1
1
1
1
1
1
1
1
]
1

¸

nn n n
ij
n
n
x x x
x
x x x
x x x
. . .
. . .
. . .
. . .
. . .
. . . .
2 1
2 22 21
1 12 11
Elements x
ij
(i,j = 1,...,n) represent the frequency separation required 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
] =
if cell i and j cannot use the same frequency
0 : otherwise
1 :
¹
'
¹
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 frequency assignment problem is then defined as, given F
frequencies and N cells each requiring d
i
frequencies, find an NxF frequency assignment
matrix A given by,
C
e
l
l
s
Cells
Cell group1
0010000100000000000
1000000000000000001
0100000001000100000
0000010010000000000
0001000000000010000
0000100000010000000
0000001000000000100
0000100000001000010
0000001000000010100
0010000000100000000
1000000000000000011
0100000100000100000
0000010010000000000
0001000000000001000
0000010010010000000
0001000100000001000
0000100000001000000
0000001001000000100
0010000000100000000
1000000000000010001
Frequencies
A =
1
1
1
1
1
1
1
1
1
]
1

¸

1 . . . . 0 0
. .
. .
. .
. .
0 . . . . 1 0
0 . . . . 0 1
ik a
such that,
A = [a
ik
] =
if cell i is assigned frequency k
0 : otherwise
1 :
¹
'
¹
A frequency assignment is admissible if both traffic and interference constraints are fulfilled.
This implies that:
1.

·
·
F
k
i ik d a
1
for all i.
2. Valid frequencies are assigned to cells according 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 second 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
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.
C
e
l
l
s
Demand Vector: (3,2,4,5,2,3,........................,5,4,3,4)

Resource Vector 1: (1,1,1,1,1,1,.......................,1,1,1,1)
Resource Vector 2: (2,1,3,4,1,2,.......................,4,3,2,3)
Cell group N
Cell group1
1 0 1 1 0
0 1 0 0 1
1 1 1 0 0
. . . . .
. . . . .
. .
. .
. .
. .
. .
. .
. .
. . . . .

¸

1
]
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0010000100000000000
1000000000000000001
0100000001000100000
0000010010000000000
0001000000000010000
0000100000010000000
0000001000000000100
0000100000001000010
0000001000000010100
0010000000100000000
1000000000000000011
0100000100000100000
0000010010000000000
0001000000000001000
0000010010010000000
0001000100000001000
0000100000001000000
0000001001000000100
0010000000100000000
1000000000000010001
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 mutation 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 and 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
the 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 introduce 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 impractical for current GSM networks.
GA Evolution
0
20
40
60
80
100
120
1
4
7
9
3
1
3
9
1
8
5
2
3
1
2
7
7
3
2
3
3
6
9
4
1
5
4
6
1
5
0
7
5
5
3
5
9
9
6
4
5
6
9
1
7
3
7
7
8
3
8
2
9
8
7
5
9
2
1
9
6
7
Generations
C
o
s
t

F
u
n
c
t
i
o
n
00100000000000000010
10000000010000000000
01000000000100001000
00000100000001000000
00010000000000000100
00001000001000000000
00000010000000010000
00001000000010100000
00000010100000000001
00100000001000000000
10000000010000100000
01000000000000001010
00000100100000000000
00010000000100000000
00000100000000010000
00010001000000000110
00001000000010000000
00000010000100010000
00100000000001000000
10000001000000000001
0010000100000000000
1000000000000000001
0100000001000100000
0000010010000000000
0001000000000010000
0000100000010000000
0000001000000000100
0000100000001000010
0000001000000010100
0010000000100000000
1000000000000000011
0100000100000100000
0000010010000000000
0001000000000001000
0000010010010000000
0001000100000001000
0000100000001000000
0000001001000000100
0010000000100000000
1000000000000010001
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 optimal 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 group 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 finds 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 used at the same cells in the
second hour.

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 in 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 location 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 implemented and also improved without interference
with existing equipment.
Network Statistics
Frequency Assignment
Fig. 6 Adaptive Radio Resource Manager
OMC
PM
Predictiv
e
RRM
CM
00010000001000000000
00000010000001000000
00001000000100000100
01000000100000000000
10000001000000000000
00000100000010000000
00100000010000000000
01000000100000010000
10000001000000000010
00000100000010000000
00100000010000001000
00010000001000100000
00000010000001000000
00001000000100000000
00100000010000000000
00010000001000100001
00000010000001000000
00001000000100010000
01000000100000000000
10000001000000000010
00010000001000000000
00000010000001000000
00001000000100000100
01000000100000000000
10000001000000000000
00000100000010000000
00100000010000000000
01000000100000001000
10000001000000010000
00000100000010000000
00100000010000100000
00010000001000010000
00000010000001000000
00001000000100000000
00100000010000100000
00010000001000000010
00000010000001000000
00001000000100001000
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, see 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, λ, is Poisson
distributed while the call holding time, 1/µ, 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 performance 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 required 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 blocking 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 this 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 suitable method
of resource assignment for the adaptive radio resource management system.
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
Simulation results using the alternative 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 significantly 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
adaptive 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 Resource
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 call 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 benefits 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 to both terminals and base station equipment.
ACKNOWLEGEMENTS
Average Blocking
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
1
1
7
3
3
4
9
6
5
8
1
9
7
1
1
3
1
2
9
1
4
5
1
6
1
1
7
7
1
9
3
2
0
9
2
2
5
2
4
1
2
5
7
2
7
3
2
8
9
3
0
5
3
2
1
3
3
7
time (hr)
b
l
o
c
k
i
n
g

%
Average Blocking
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
1 25 49 73 97 121 145 169 193 217 241 265 289 313 337 361
time (hr)
b
l
o
c
k
i
n
g

%
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 this 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
Assignment 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] Enrico 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 Assignment
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 Evolutionary 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 Soft 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 Search, 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, June
2001.

Each MFNN consists of three layers with a total of 49 neurons. The system proposed here considers two MFNNs for each cell site. Once the MFNNs are trained. Neural Network Prediction 35 30 Channel Demand 25 20 15 10 5 0 71 1 11 21 31 41 51 61 151 Time (hr) 161 101 111 121 131 141 81 91 Actual Prediction Fig. the neural network resource predictor performs to an excellent degree of accuracy. 1. 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. 2. 2 RESOURCE PREDICTION USING NEURAL NETWORKS As the number of frequencies allocated to each cell in the network depends on the resource predictions. the day and time statistics. Compile statistics of previous 8 weeks from the network. The training and prediction of the MFNNs proceeds as follows: 1. 7. Record the time (time statistic). It has been shown that MFNNs have the capability to predict future values based on training sets composed of sufficient historical data [2]. 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). 4. These statistics constitute the initial training data set. Collect hourly radio resource demand statistics for in-cell and handover calls. One MFNN is used to predict future resources for normal in-cell traffic while the second MFNN predicts guard frequencies for handover calls. while satisfying both the . the method of prediction must be robust enough to track the inherent hourly variations in call traffic. 6. 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. Each MFNN is retrained every 24 hours to maintain accurate predictions. The predicted number of frequencies is assigned to each base station. 5.not impact on the cellular infrastructure. Record whether the demand occurs on a weekday or weekend (day statistic). the channel demand for the next hour in each cell is predicted using the demand statistics from the previous 10 hours. 3. The back-propagation learning algorithm and nonlinear sigmoid activation function are used in the learning process [7]. As shown in Fig. The MFNNs are trained using the data arising from step 2. Translate the resource demand into traffic load measurements in Erlangs and derive the associated radio channel requirements.

.  . We now present two methods of frequency assignment using GA techniques. the resource predictors make hourly resource demand predictions.. necessary to maintain interference below a certain threshold. .  . . .  .  xnn   Cells C = xij . where n is the number of cells in the system.traffic demands and interference constraints. Elements xij (i... a GA embedded in the resource management system is used to find the frequency assignment plan for a particular demand scenario.  .  xn1 xn 2  . Such techniques achieve optimality in the range of (80-90%) depending on the problem complexity. .9.   . find an NxF frequency assignment matrix A given by.2. the resource predictors fill out this vector at the end of each hour.  .n) represent the frequency separation required between frequencies assigned to cells i and j. .  1 : if cell i and j cannot use the same frequency C = [xij] =   0 : otherwise The traffic demands can be represented by the demand vector D. . The first approach produces an excellent degree of optimality in terms of resource requirements but suffers from large variations in successive frequency assignment plans. given F frequencies and N cells each requiring di frequencies. however. Cells  x11 x12  x 21 x 22   .1 RESOURCE OPTIMISATION USING A GA Various schemes incorporating GAs to solve the frequency assignment problem have been proposed in [8. The frequency assignment problem is then defined as. By combining a GA with a local search technique. x1n  x2n  . As resource predictions are not made in realtime. The resource optimisation scheme is presented here.11]. The interference constraints for the network can be represented by an n x n compatibility matrix C..n) representing the number of required frequencies at cell i. .  . 3.j = 1. respectively.. with elements di ( i = 1. this approach does not produce the most optimal solutions. Using this matrix. In the proposed adaptive radio resource management system. it is possible to represent co-channel interference by choosing values for xij such that. we can achieve 96% and 99% optimality for in-cell and handover frequency assignment problems respectively. . The second approach ensures that frequencies assigned to a cell include most of the frequencies assigned to that cell in the previous hour.10.   .

Valid frequencies are assigned to cells according to the compatibility matrix. i.2. .. .  .. .2.1.. 2 (3.  0 0  .4.....1.  1 : if cell i is assigned frequency k A = [aik] =   0 : otherwise A frequency assignment is admissible if both traffic and interference constraints are fulfilled...1. ..  . . ...  ..5. ...A = Frequencies 1 0 .  aik . ... .. The GA works with a population of 80 individuals.2.1. This implies that: 1. C..3..4) ⇓ (1. . .. . The GA takes each resource vector and finds a frequency assignment plan using the minimum number of resources required for that particular demand.4. .. .  ....1. . ..e.. ∑a k =1 F ik = di for all i..1.  . 0 1  .3.3.1.. .. so as to sustain a Broadcast Control Channel (BCCH).1) (2. . ..  . . 2..... .  0  Demand Vector: Resource Vector 1: Resource Vector 2: Fig. .3) Determination of Resource Vectors The number of binary ones in each cell group corresponds to the frequency requirements in the resource vector.... 2.5. each individual is represented as follows.. ..... .1.3. 0 0 . . 1  such that.  . ....  Cell group N 1 1 1 0  Cell group1 1 0 1 1 Cells ..1......4....2.. .. 0 1 0 0  . Two resource vectors of length n are derived from the demand vector D as shown in Fig. ... The second resource vector assigns the remaining frequencies. ..  ..  ...1...  .... . 0 1 . .... The first resource vector assigns one frequency to each cell site... the number of frequencies to be assigned to the cell. .4.

The search algorithm finds the cell group and bit position in the solution that is preventing the cost reaching zero. 3 GA Evolutionary Performance Although this approach produces optimal solutions. the mutation operator used here swaps bits within a cell group. In order to improve the performance and increase the rate of convergence for difficult frequency assignment problems. 00000010000000010000 0000001000000000100 00001000000010100000 00000010100000000001 00100000001000000000 10000000010000100000 01000000000000001010 00000100100000000000 00010000000100000000 00000100000000010000 00010001000000000110 00001000000010000000 00000010000100010000 00100000000001000000 10000001000000000001 0000100000001000010 0000001000000010100 0010000000100000000 1000000000000000011 0100000100000100000 0000010010000000000 0001000000000001000 0000010010010000000 0001000100000001000 0000100000001000000 0000001001000000100 0010000000100000000 1000000000000010001 . This process continues until the cost reaches zero (optimal solution) or all the bits in the cell group have been swapped. Crossover can only occur at the end of a cell group. With the standard mutation operator. 4. GA Evolution 120 100 Cost Function 80 60 40 20 0 277 737 1 47 231 323 461 645 875 139 369 415 507 553 783 93 829 921 185 599 691 967 Generations Fig. Fig. each bit is flipped with a predefined probability. a local search routine has been developed.The roulette wheel selection algorithm is used to generate the parents for the new population [12]. In this scenario each cell requires one frequency. To execute such large amounts of redeployments is 00001000001000000000 0000100000010000000 consuming and would be impractical for current GSM networks. therefore. 3 illustrates the performance of the GA for a 49 cellular network with a cluster size of 7. That bit is then swapped with the other bits in the cell group and the cost of the solution evaluated. as crossing within a cell group would alter the number of frequencies being assigned to that cell.0014 respectively. 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. In the algorithm described above. the GA reaches this optimal solution by using 7 frequencies. 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.6 and 0. the frequency assignment plan for 00100000000000000010 0010000100000000000 consecutive hours can be significantly different as shown in Fig. flipping a bit would result in a change to the number of frequencies been assigned to a cell. Such large changes to the 10000000010000000000 1000000000000000001 0100000001000100000 frequency assignment 01000000000100001000 a considerable amount of inter-cell handovers plan would introduce 00000100000001000000 0000010010000000000 into the system as each cell site would be required to change its allocated frequencies every 00010000000000000100 frequency 0001000000000010000 also very time hour. The new population is created using the standard multi-point crossover and a modified mutation operator with probabilities 0.

Since the GA finds optimal solutions for each array separately. Each array is then passed to the GA.0) Each cell group has an initial length of 7. each individual in the population is represented as follows: Cell group 1 Cell group N (1. the second array for those requiring at least two frequencies and so on.1. .1. 5. thus minimizing inter-cell handovers. D.0. we propose an alternative method of frequency assignment using a GA.Fig. as this is the maximum number of frequencies required for the first binary array (assuming a cluster size of 7). The GA works with an initial population of size 40. The optimal solutions from each array are augmented to create the final frequency assignment plan.0. A binary 1 within an array denotes a cell that requires a frequency.1. Two consecutive frequency assignment plans are shown in Fig. which finds the minimum number of frequencies for that demand.0. 4 Consecutive frequency assignment plans for 20 cells using GA resource optimisation 3. This approach does not produce the most optimal solutions in terms of resource requirements. as such changes will only be reflected in the last one or two binary arrays. it does ensure that cells can maintain the majority of frequencies from hour to hour. a solution is sought for six and so on until no better solution can be found.0.2 ALTERNATIVE APPROACH TO FREQUENCY ASSIGNMENT USING A GA To overcome the problem of resource optimisation discussed in section 3.0. A number of binary arrays of length n are created from the demand vector.….0..0. The first array represents those cells requiring at least one frequency.0.0. thus minimising inter-cell handover.0.0. however.0.0.0. note how most of the frequencies assigned in the first hour are also used at the same cells in the second hour.0.0. but ensures that frequencies allocated to a cell include most of the frequencies allocated to that cell in the previous hour. If the GA finds a valid assignment for seven frequencies.0. the overall solution may be sub-optimal.

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 in the OMC between the performance management (PM) tool and the configuration management (CM) tool. Network Statistics PM Predictiv e OMC RRM CM Frequency Assignment Fig. The OMC location 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. As all the network performance data required for the system is available at this location.00010000001000000000 00000010000001000000 00001000000100000100 01000000100000000000 10000001000000000000 00000100000010000000 00100000010000000000 01000000100000010000 10000001000000000010 00000100000010000000 00100000010000001000 00010000001000100000 00000010000001000000 00001000000100000000 00100000010000000000 00010000001000100001 00000010000001000000 00001000000100010000 01000000100000000000 10000001000000000010 00010000001000000000 00000010000001000000 00001000000100000100 01000000100000000000 10000001000000000000 00000100000010000000 00100000010000000000 01000000100000001000 10000001000000010000 00000100000010000000 00100000010000100000 00010000001000010000 00000010000001000000 00001000000100000000 00100000010000100000 00010000001000000010 00000010000001000000 00001000000100001000 01000000100000000000 10000001000000000010 Fig. The non-invasive nature of the proposed concept is one of its major advantages in that it can be implemented and also improved without interference with existing equipment. see Fig. no additional signalling load is generated. 6 Adaptive Radio Resource Manager . 6.

has a mean of 3 minutes. 48 45 38 31 24 17 10 3 46 39 32 25 18 11 4 47 40 33 26 19 12 5 41 34 27 20 13 6 49 42 35 28 21 14 7 y 44 43 36 29 22 15 8 1 37 30 23 16 9 2 x Fig. 1/µ. while 14 guard frequencies achieved a call dropping rate of zero. Performance measurements were averaged over 700 simulation runs so as to achieve accurate results. Unlike the FCA concept. producing a resource gain of 24%. Both network models contain 49 cells with a cluster size of 7. Using the resource optimisation GA. A total of 29 frequencies were required in the network for new call arrivals. the handover arrival rate is therefore taken to be 5% of the sum of the call arrivals in these cells. Since this scheme results in unpractical frequency assignment plans. 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. 7 5. a total of 22 frequencies were required for new call arrivals. 8 shows the average blocking for cell site 2.2 ADAPTIVE RADIO RESOURCE MANAGEMENT The same call traffic scenario was used in this model. Fig. 5. see Fig. λ. The performance of both models is measured by the number of frequencies required to maintain the call blocking below 2% throughout the network. 7. . it is not considered to be a suitable method of resource assignment for the adaptive radio resource management system. The number of handover arrivals in each cell will depend on the load in each of the six surrounding cells. is Poisson distributed while the call holding time. The load in each model is non-uniformly distributed across the network. The network was monitored over a 2-week simulation period.1 49-cellular structure FCA NETWORK MODEL Each cell in the FCA model were assigned the required number of frequencies so as to maintain the call blocking below 2%. The call arrival rate. cells are assigned frequencies based on resource predictions and the GA frequency assignment schemes discussed in section 3. Each cell has a total of 18 neighbours with wraparound performed at the borders in the x and y planes.5 NETWORK SIMULATION AND RESULTS Two simulation models have been developed – a Fixed Channel Allocation (FCA) model.

This result is comparable to current DRA schemes [1].06 0.02 0 177 273 1 49 17 33 65 113 129 145 161 209 225 241 257 305 321 337 81 97 193 289 time (hr) blocking % 0.14 blocking % 0. the similarities in call blocking at peak times can be seen.04 0. As each cells handover call arrival rate depends on the six surrounding cells. but with the advantage of significantly less complexity and no additional signalling load.2 0.08 0.18 0.Simulation results using the alternative GA frequency assignment scheme show resource requirements of 23 frequencies. this approach can achieve self-configuring cellular networks without the need of additional signalling loads and changes to both terminals and base station equipment.02 0 1 25 49 73 97 Average Blocking 121 145 169 193 217 241 265 289 313 337 361 time (hr) Fig.14 0. giving a resource gain of 20. 9 Average call blocking in adaptive network CONCLUSION An adaptive radio resource management system for current GSM networks has been proposed. Table 1.7% Average Blocking 0. GA resource optimisation 22 14 24% Adaptive Resource Management.06 0.12 0.1 0.7% when compared with the equivalent FCA network.12 0. The additional blocking in the adaptive network arises because each cell is allocated just the required number of frequencies for the next hour.2 0.18 0. GA Frequency Assignment 23 14 20. It was observed however that no resource gain could be obtained from adaptive guard channel allocation.08 0. The average call blocking over 700 simulation runs is shown in Fig.7% when compared with an equivalent FCA network. By comparing Figs. thus maximising the system resources. 8 6 Average call blocking in FCA network Fig. Simulation results have shown resource benefits of up to 20. The simulation results are summarised in Table 1. 9. ACKNOWLEGEMENTS .1 0. 8 and 9.16 0. handover calls tend to be more uniformly distributed than new call arrivals. Using frequency deployment mechanisms such as those discussed in [13].16 0. Network Model Frequency Requirements for new call arrivals Frequency Requirements for handover calls Resource Gain Simulation Results FCA 29 14 Adaptive Resource Management.04 0.

Veh.. [9] Goutam Chaktaborty and Basabi Chakraborty. George G. 1994. vol. Maynooth. June 1996. 40. Technol. Ireland. “A New Strategy for the Application of Genetic Algorithms to the Channel Assignment Problem”... Nov. [3] Dietmar Kunz. pp. 279288. 45. Technol. pp. July 1999. 1998. vol. Stavroulakis. 1999. [6] Nobuo Funabiki. pp. Irish Signals and Systems Conference. Optimization and Machine Learning”. IEEE Trans. Lai. 1999. and Yoshiyasu Takefuji. [4] Enrico Del Re. 176-181. 1991. “Neural Networks: A Comprehensive Foundation”. pp. 47. [5] Harilaos G. Veh. H. [12] Goldberg. IEEE Midnight-Sun Workshop on Soft Computing Methods in Industrial Applications. “A Neural Network Parallel Algorithm for Channel Assignment Problems in Cellular Radio Networks”. . Stephens “Optimum Frequency Deployment in Cellular Mobile Networks using Genetic Algorithms”. 1999. pp.. and Joe Rodriguez-Tellez. 48. IEEE Trans. K. Technol. Sandalidis. 91-96. Veh. [7] S. 1261-1269. June 2001. and Victor O. K. 26-32. 1992. vol. 45. Neural Networks. Feb. 1996. May 1994. Romano Fantacci.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 this paper. Technol. “Channel Assignment Through Evolutionary Optimization”. 430-437.. Finland. [8] W. [11] Chiu Y. “Channel Assignment Schemes: A Comprehensive Survey”. Coghill. Chan. [2] Peter T. pp. pp. Feb. M. 163-172. vol. Technol. Veh. vol. vol. IEEE Trans. Haykin. “Channel Assignment for Cellular Radio Using Neural Networks”. Technol. “A Dynamic Channel Allocation Technique Based on Hopfield Neural Networks”. Veh. IEEE Trans. Li. Veh. “Fixed Channel Assignment in Cellular Radio Networks Using a Modified Genetic Algorithm”. [10] Dirk Beckmann and Ulrich Killat. Pesch. Perez-Carbonell. 188-193. IEEE Trans. June 16th –18th. Prentice-Hall. “A Genetic Algorithm Approach to Solve Channel Assignment Problem in Cellular Radio Networks”. 43. pp. vol. 41. “Borrowing Channel Assignment Strategies Based on Heuristic Techniques for Cellular Systems”.. Feb. and Luca Ronga. “Genetic Algorithm in Search. Veh. Feb. Peter P. Kuusamo. Naghshineh. Marimuthu Palaniswarni. [13] M. and David Everitt. 10. Technol. “Neural Network-Based Dynamic Channel Assignment for Cellular Mobile Communication Systems”. Ngo. D. 1996. Addison Wesley. IEEE Trans. REFERENCES [1] I. IEEE Personal Communications. Jan. IEEE Trans. P. vol. Katzela.. IEEE Trans.