This document describes a proposed artificial neural network (ANN)-based cellular network resource allocation predictive system. The system aims to predict future network traffic volume and determine the optimal quantity of resources to allocate to cells to meet quality of service demands. It was developed using data from a Nigerian cellular network operator and tested against other techniques. The ANN model was formulated as a feed-forward network and performed better than other techniques according to error metrics. The developed application can integrate into cellular networks and predict resource needs based on quality of service parameters.
1. EEE PROJECTS
ARTIFICIAL NEURAL NETWORK-BASED CELLULAR
NETWORK PREDICTIVE SYSTEMFOR RESOURCE
ALLOCATION
ABSTRACT
A cellular network resource allocation predictive system based on artificial neural network (ANN) is presented.
The predictive system is capable ofpredicting the future network traffic volume/intensityin a cell and accurately
determining the optimum quantityof resources to be allocated to the cell to meetQoS demands.The main
objective of this research is to develop a predictive system thatdelivers to the network providers a resource
managementsystem thatis relatively simple,efficientand effective. The ANN based resource allocation
predictive model was developed using data collected from an established cellular network operator in Nigeria.
The data was pre-processed,trained and analysed using the Self-Organizing Map (SOM) and the Neural
Network Toolboxes in a MATLAB environment.The model was formulated as a 3-layer Feed-Forward ANN
network with seven predictors as inputs,a hidden layer and an output variable.After rigourous analysis,the
Conjugate Gradientwith Polak-Ribiere Restarts (CGP) configuration with 14 neurons in the hidden layer was
finally adopted as the model.The performance ofthe model in predicting the future mean traffic in each cell was
further compared with some existing techniques using the cross-validation method.The mean square error
(MSE) and mean average error (MAE) values for the techniques were respectivelyfound to be: single tree
(43.18, 3.70), tree boost(45.26,3.51), multilayer perceptron (44.83,3.81), general regression neural network
(35.35, 3.50), radial basis function (63.01,4.92),general method ofdata handling polynomial network (17616,
54.11), supportvector machine (40.43,3.20), gene expression programming (26.41,3.13),ANN Model (1.60,
1.31). The values obtained showed thatthe prediction capabilityof the developed model was superior to the
existing techniques.The model was then tested through simulation in a MATLAB environmentand the test
results ploughed back into the model for modification and further finer performance improvement.Using the
predicted mean traffic and applying the blocking probabilityas a QoS parameter,the ANN Model computes the
traffic channel(s) to be allocated to each cell. Finally, the model was packaged as an Application software for
integration into the cellular network using the Graphical User Interface DevelopmentEnvironment(GUIDE). The
developed Application can fit easilyinto a cellular network system and it was successfullyused to predict the
number ofchannels needed to service a given cell based on the required QoS parameter values.
https://takeoffprojects.com/eee-projects
TABLE OF CONTENTS
APPROVAL PAGE.. ii
CERTIFICATION.. iii
DEDICATION.. iv
ACKNOWLEDGEMENTS. v
ABSTRACT. vi
TABLE OF CONTENTS. vii
2. LIST OF FIGURES. x
LIST OF TABLES. xii
LIST OF ACRONYMS. xiii
CHAPTER ONE.. 1
INTRODUCTION.. 1
1.0 STUDY BACKGROUND.. 1
1.1 OBJECTIVES OF THE RESEARCH.. 6
1.2 STATEMENT OF PROBLEM… 6
1.3 NEED FOR RESEARCH.. 6
1.4 SCOPE OF WORK.. 7
1.5 METHODOLOGY.. 7
1.6 THESIS OUTLINE.. 8
CHAPTER TWO.. 9
LITERATURE REVIEW… 9
2.1 CELLULAR NETWORK RESOURCE ALLOCATION.. 9
2.1.1 Allocation of Channel 9
2.1.2 Allocation of Bandwidth.10
2.1.3 Allocation of Frequency. 11
2.2 RESOURCE ALLOCATION TECHNIQUES AND ALGORITHMS. 12
2.2.1 Resource Allocation Techniques using Neural Networks.12
2.2.2 Resource Allocation Techniques using Genetic Algorithms.13
2.2.3 Resource Allocation Techniques using Optimization Methods. 14
2.2.4 Resource Allocation Techniques using EvolutionaryAlgorithms.15
2.2.5 Resource Allocation Techniques using Fuzzy Logic.16
2.3 RESOURCE ALLOCATION PREDICTIVE MODELS. 16
2.4 ARTIFICIAL NEURAL NETWORK (ANN) 19
2.4.1 The Mathematical Model of ANN.. 19
2.4.2 Activation Functions.20
2.4.3 ANN Architectures and Models. 20
3. 2.5 SUMMARY.. 27
CHAPTER THREE.. 28
DEVELOPMENT OF ANN PREDICTIVE MODEL.. 28
3.1 INTRODUCTION.. 28
3.2 METHODOLOGY.. 28
3.2.1 Data Collection and Selection.28
3.2.2 Data Pre-processing.29
3.3 DATA TRAINING.. 31
3.4 DEVELOPMENT OF CELLULAR NETWORK RESOURCE ALLOCATION PREDICTIVE MODEL 32
3.4.1 Back propagation Feed Forward Neural Network Architecture. 32
3.4.2 Developmentofthe Model 33
3.5 PROPOSED MODEL.. 40
3.6 METRICS FOR MEASURING PERFORMANCE.. 41
3.7 MODEL VALIDATION.. 42
3.8 ERLANG-B FORMULA.. 42
CHAPTER FOUR.. 44
SIMULATION RESULTS AND RESULTS ANALYSIS. 44
4.1 INTRODUCTION.. 44
4.2 PRE-PROCESSING RESULTS AND ANALYSIS. 44
4.2.1 Transformation and Visualization.44
4.2.2 Clustering.49
4.3 CONFIGURATION OF THE DEVELOPED MODEL.. 51
4.4 TRAINING PARAMETERS. 61
4.5 VALIDATION OF THE ANN MODEL.. 62
4.6 NETWORK RESOURCE PREDICTION USING THE DEVELOPED ANN MODEL 62
CHAPTER FIVE.. 64
APPLICATION TESTING AND PERFORMANCE ANALYSIS. 64
5.0 INTRODUCTION.. 64
5.1 TESTING THE APPLICATION WITH DIFFERENT INPUT-TARGET SCENARIOS 64
4. 5.1.1 Linear regression.65
5.1.2 Correlation.80
5.2 PREDICTED TRAFFIC AND CHANNEL ALLOCATION.. 85
5.3 APPLICATION PERFORMANCE AND SECTOR CLUSTERING.. 102
5.4 COST OF DEVELOPING THE APPLICATION.. 104
CHAPTER SIX.. 105
MODEL DEPLOYMENT. 105
6.0 INTRODUCTION.. 105
6.1 DEVELOPMENT OF THE ANN-BASED NETWORK RESOURCE ALLOCATION APPLICATION 105
6.1.1 Layout of the Graphical User Interface. 105
6.1.2 Programming ofthe GUI. 108
6.2 PACKAGING THE ANN-BASED NETWORK RESOURCE ALLOCATION APP FOR DEPLOYMENT. ….
113
6.3 INSTALLING THE NETWORK RESOURCE ALLOCATION APPLICATION.. 116
CHAPTER SEVEN.. 120
CONCLUSION.. 120
7.0 INTRODUCTION.. 120
7.1 CONCLUSIONS.120
7.2 CONTRIBUTIONS TO KNOWLEDGE.. 121
7.2.1 Listof Publications.121
7.3 DIRECTIONS FOR FUTURE RESEARCH.. 122
REFERENCES. 123
APPENDIX I. 133
MATLAB Code for Network Resource Allocation Module.133
APPENDIX II. 140
Location Data. 140
APPENDIX III. 142
MATLAB Code for Inverse Erlang-B.. 142
APPENDIX IV.. 143
5. Codebook Vector 143
APPENDIX V.. 146
GSM Logical Channels 146
CHAPTER ONE
INTRODUCTION
1.0 STUDY BACKGROUND
A cellular network is a radio network comprising ofcells which are interconnected usuallyover a large area
spanning several kilometres [118].These cells contain base transceiver stations (BTS) which enables the
transmission and reception ofradio signals to and from mobile user equipmentusuallyreferred to as mobile
station (MS) such as mobile phones.These cells together provide radio coverage over a given geographical
area.
The architecture for mobile cellular network is mainlydivided into three subsystems:the MS, BTS, and network
[119]. It can be further structured into a number of sections:Network and Switching Subsystem (NSS),
Operations SupportSystem (OSS), servers,Operation and Maintenance (O & M). Each subsystem performs its
separate functions which are linked together by logical and physical channels to enable full operational capability
of the system.
The MS otherwise called mobile phone or ‘handset’ is the part of mobile cellular network thatthe subscriber uses
to communicate.Itconsists mainlyof the hardware and subscriber identitymodule (SIM) [119]. The hardware
comprises ofall the electronics needed to generate,transmit,receive and process signals between the MS and
BTS. The SIM provides the information that identifies the user to the network using the international mobile
subscriber identity(IMSI) system.
The base station subsystem (BSS) consists ofthe base transceiver station (BTS) and base station controller
(BSC) [120]. The BTS uses antennas,which are made up of transmitters and receivers,for direct communication
with the MS through a special interface.The BSC manages the radio resources and controls a group of BTSs
and also manages handovers and the allocation ofchannels in a network.
The network subsystem provides overall control and interfacing ofthe whole mobile network.It comprises mainly
of the Mobile Switching Centre (MSC) which acts like a normal switching node in a telephone exchange [119].It
also,performs other tasks for a mobile phone like registration,authentication,call location and routing using the
Visitor Location Register (VLR),Home Location Register (HLR), EquipmentIdentity Register (EIR) and
Authentication Centre (AuC).
Other importantelements contained in the network subsystem are the Gateway Mobile Switching Centre (GMSC)
and Message Service Gateway (SMS-G). The GMSC terminates call initiallyrouted to the network without
knowledge ofthe location of the mobile phone while the SMS-G handles and directs messages in different
directions [121].
Established cellular technologies existin the telecom industryand they include the Global System for Mobile
Communication (GSM), and code division multiple access (CDMA). GSM is the mostwidelyused wireless
cellular technology[122]. Its family of technologies include General PacketRadio Service (GPRS), Enhanced
Data for GSM Evolution (EDGE), Universal Mobile Telecommunication System (UMTS), High Speed Packet
Access (HSPA), and more recently Long Term Evolution (LTE) network [122].
The major players in the telecom industryare the regulators,operators,and subscribers [137].Regulators
provide the framework and general rules ofoperation in the industry.Such regulators include the International
Telecommunication Union (ITU),and several other regulatory bodies in differentcountries ofthe world.The
6. Nigerian Communication Commission (NCC) is the main regulatory body saddled with the responsibilityof
supervising the operation ofmobile cellular network operators in Nigeria.
Cellular network operators provide services to users thatsubscribe to their network.They are majorlyclassified
into mobile and fixed wireless operations.Several operators existin differentcountries,a listof such operators
are available [1, 2]. In Nigeria,the major operators are in the area of GSM and they comprise ofAirtel, MTN,
GLO, and Etisalat.The fifth mobile operator,Mtel, is now moribund [3].
The subscribers are the end users ofthe services provided by the mobile operators.All over the world the
number ofsubscribers has been increasing steadily.Taking Nigeria as a typical example,mobile cellular network
was introduced in 2001 and since then figures from the NCC show tremendous growth ofconnected lines
(subscribers).As at March 2014 connected lines stand at124,884,842 for GSM, 2,039,391 for CDMA, and
172,963 for Fixed/Fixed Wireless from only 266,461 connected subscribers atinception [3].The tele density grew
from a mere 0.73 to 92.24 within the period.With a population ofabout170 million people and an annual growth
projection of 2.54 %, the number ofsubscribers is bound to rise [4,5]. Similar results were obtained in other
developing countries [123].
Over the years, several advances have been made in cellular communication technology,with increasingly
capable user terminals,and expanding range ofmobile applications.This growth has led to significantincrease in
voice and data traffic with different Quality of Service (QoS) requirements. Understanding the characteristics of
the traffic is importantfor network design,traffic modelling,resource planning and resource control.The ITU and
other regulatoryagencies have laid out the requirements and framework for effective service delivery in mobile
cellular networks [126].
Provision of services to subscribers by cellular network providers require adequate resources such as bandwidth,
radio links,traffic channels, which are usuallyscarce and expensive.For this reason,the regulators and
operators ofmobile wireless cellular networks have continuallybeen looking for ways to improve on network QoS
standard.This they do by constantly searching for the mosteffective, efficient and proactive cellular network
resource allocation scheme thatwill deliver the bestQoS values to network subscribers [128,129,132].
The ITU Recommendation E.507 provides an overview of the existing mathematical techniques for modelling and
forecasting.Such techniques include curve fitting, Autoregressive Integrated Moving Average (ARIMA), state
space models with Kalman filtering,regression and econometric models [6].It also describes methods for the
evaluation of the forecasting models and for the choice of the mostappropriate method in each case,depending
on the available data,length of the forecastperiod,etc [6].
There are many resource allocation techniques existing in literature.Majorly, these are based on fixed or
dynamic allocation schemes.In fixed allocation systems the resources are predetermined and are assigned
manuallyby operators [7, 8]. In dynamic allocation schemes the allocation ofresources is varied according to
needs and demands and this is usuallyachieved by means ofan algorithm [9,10, 11].
In recent times,Artificial Neural Network (ANN) algorithms have been used withoutcomplexities in other fields for
the analysis ofsimilar problems in related areas using data sets [12 – 27]. It is on this basis thatthe ANN
technique which greatly reduces the complexity in data processing and analysis withoutany infringementon
system’s precision is intended to be applied in the realization of a predictive model thatcan be used for cellular
network resource allocation.
A predictive model is a statistical model created to forecastfuture behaviour based on presentand/or existing
indicators [124].Predictive models are made up ofa number of predictors,which are variable factors that are
likely to influence future behaviour or results.Predictive modelling,therefore,is an area of data mining concerned
with forecasting probabilities and trends [125].In predictive modelling,data is collected for the relevant
predictors,a model is formulated,predictions are made and the model is validated (or revised) as additional data
becomes available.The model mayemploya simple linear equation,statistics,machine learning,neural
computing,robotics,computational mathematics,and artificial intelligence techniques or a complexneural
network, mapped outby sophisticated software.
Predictive modelling explores all the data setinstead ofa narrow subsetto bring out meaningful relationships and
patterns,it can be applied in various areas such as customer relationship management,capacityplanning,
change management,disaster recovery, securitymanagement,engineering,meteorologyand city planning [28].
7. There are two major types of predictive modelling approaches – those with supervised learning and the others
with unsupervised learning.In supervised learning,predictive models are created using a setof trained data that
contains results upon which the prediction will be based [117].The techniques used for this type of learning
include classification,regression,and time-series analysis.The classification identifies groups within the data and
associates anynew data with a group,regression uses pastvalues to predict future values,while time -series
uses time to predictseasonal variances.
Unsupervised learning on the other hand does not use previous results to train its models;rather it uses
descriptive statistics to examine the natural patterns existing within the data [124, 125]. The techniques used
here include clustering,and association.Clustering identifies groups of similar behaviour within the data, while
association identifies the relationship among various groups within the data.
In general,predictive models can be builtby supervised learning and/or unsupervised learning using several
techniques [125].The models can be implemented using a variety of algorithms suited for different data or
problems.Many software packages are available thatimplementthis models and algorithms to find the best
combination thatworks.These software packages can be classified into proprietaryand open source.Proprietary
software are those licensed by companies and are usuallyvery expensive;open source software on the other
hand can be freely downloaded from the internetunder the GNU agreement.Examples ofproprietarysoftware
include MATLAB, Statistica,MapleSim, Mathematica,IBM SPSS Modeler, SAS Enterprise Miner, and Microsoft
SQL Server [29 – 35]. Open Source software includes Knime,Orange,Weka,R, and RapidMiner [36 – 40].
The developmentofa cellular network predictive system will be based on a number ofpredictors,which are
variable factors that influence subscribers’ resource demand behaviour in the future using historical data.The
use of this data is imperative because the intrinsic behaviour ofa network is usually embedded in the data
collected from the network over time [41].
The multidimensional data collected will be analysed simultaneouslyto observe emerging network trends and
performance.Such predictions and also the requirementto analyse multidimensional data simultaneouslyhas an
overwhelming influence on the traditional data analysis methods.This results in complexdata processing and
data analysis thatare usuallydifficult to track and consequentlyleads to erroneous resource allocation decisions.
Hence the choice of ANN becomes advantageous as itshows excellentperformances in similar circumstances.
The task, therefore,will be the developmentof a predictive model that is based on very influential predictors
which are well known Key Performance Indicators (KPIs) like call setup success rate (cssr), drop call rate
(dcr), standalone dedicated channel (SDCCH) blocking rate (sdcchblk),SDCCH loss rate (sdcchloss),handover
success rate (hosr),call setup blocking rate (callsetblk),traffic channel blocking rate (tchblk),traffic channelmean
traffic (tchmean).
This were selected based on a meticulous analysis ofthe GSM logical channels from where the KPIs were
abstracted.The multidimensional data collected will be analysed simultaneouslyto observe emergin g network
trends and performance.Such predictions and also the requirementto analyse multidimensional data
simultaneouslyhas an overwhelming influence on the traditional data analysis methods.This results in complex
data processing and data analysis thatare usuallydifficultto track and consequentlyleads to erroneous resource
allocation decisions.Hence the choice of ANN becomes advantageous as itshows excellentperformances in
similar circumstances.
1.1 OBJECTIVES OF THE RESEARCH
The objective of this research is to develop a cellular network predictive system that,
1. provides the required QoS parameters values to the network subscribers atrelatively affordable price.
2. maximizes the utilization of cellular network resources and therebymaxim izing revenue for the network
providers.
3. is capable of being integrated into a typical mobile wireless cellular network with ease.
4. responds to random network resource demands instantlyand with precision.
8. 5. delivers to the network providers a resource managementsystem thatis relatively simple,efficientand
effective.
1.2 STATEMENT OF PROBLEM
The growing number ofsubscribers and increased access to mobile user terminals and/or devices has caused a
strain on the usage of network resources required to provide satisfactoryQoS needs such as bandwidth,traffic
channels,and radio links.Mostresource allocation schemes do notaddress the issue ofresource utilization and
availability in terms of physical resources,these schemes tend to focus on time rather than the resources even
though time too is critical. A balance is therefore required between making resource available in the rightquantity
when needed.This saves costas resource will be provided based on actual need.These resources are scarce,
limited and expensive;hence it becomes imperative that network resources be properlyand intellige ntlyallocate
for optimal performance.
1.3 NEED FOR RESEARCH
The regulators and operators ofmobile wireless cellular networks have both come to settle with continually
looking for ways to improve on network Quality of Service (QoS) standard by continuous search for network
resource allocation schemes thatwill deliver the bestQoS values to network subscribers.A proactive resource
allocation scheme is required to predictthe future behaviour of the network and properlyallocate resources
efficiently in the face of competing demands.These allocations should be dynamic enough to simultaneouslylook
at other QoS requirements in order to react appropriatelyto the usual random network resource demands.
Therefore, the need to develop a cellular network resource allocation predictive system capable ofpredicting the
future network traffic volume/intensityand accurately determining the optimum quantityof resources to be
allocated becomes necessary.Proper allocation of these resources will minimize call drops [110].
1.4 SCOPE OF WORK
The scope of this thesis will be limited to the application ofneural network to the developmentof resource
predictive model thatpredicts future mean traffic intensityin a cellular network from its historical data.From the
predictions made,channels will be allocated to cells within the network in advance to efficiently service the
predicted traffic.
Finally, the model will be implemented bydeveloping an Application software that can be deployed in a cellular
network environmentto predictthe future mean traffic and to allocate channels thatwill be required to service the
predicted traffic intensity based on the inverse Erlang-B formula.
1.5 METHODOLOGY
The methodologyadopted a research and developmentdesign.The ANN based cellular network resource
allocation predictive system model will be developed using data collected,for a period of twelve (12) months,
from an established typical cellular network operator in Nsukka,Nigeria,.The historical data will be pre -
processed,trained and analysed using the Self-Organizing Map (SOM) Toolbox [114] and the Neural Network
Toolbox [116] in a MATLAB environment.The model is formulated as an n-layer Feed-Forward ANN with seven
predictors as inputs,a hidden layer and an outputvariable. The predictors are: cssr, dcr,
sdcchblk,sdcchloss,hosr,callsetblk,and tchblk.The application ofup to seven predictors as inputis to further
improve on the accuracy of the system’s prediction capabilitywhich,in turn, will lead to accurate resource
allocation decisions.The target outcome will be based on tchmean.
Rigorous analysis will be carried outto selecta suitable and bestperforming algorithm with adequate number of
neurons which will be adopted as the model.The inverse Erlang-B formula will be used in the model to determine
the amountof resource required to meetQoS requirements.The performance ofthe developed ANN-based
Model in predicting the future mean traffic in each sector will be validated by comparing itwith some e xisting
techniques using the 10 fold cross-validation method.This validation is importantto assess how the results will
generalize to an independentdata set.The Mean Squared Error (MSE), Mean Absolute Error (MAE) and
regression analysis will be used as indicators to measure the performance.
The model will then be tested through simulation in a MATLAB environmentand the testresults ploughed back
into the model for modification and further finer performance improvement.Finally, the model will be implemented
9. as an Application software for integration and deploymentin the cellular network using the Graphical User
Interface DevelopmentEnvironment(GUIDE) in MATLAB.
1.6 THESIS OUTLINE
The restof the thesis is organized as follows:Chapter 2 presents the literature review of cellular network
resource allocation,techniques,algorithms and models.Related research works based on network resource
predictive models are also presented.In Chapter 3, the methodologyand processes leading to the development
of the ANN based predictive model are presented.Starting with the description ofthe area of study, through how
the data was collected,pre-processed,trained and ended with the developmentand validation of ANN-based
model based on detailed analysis.Chapter 4 presents simulation results and results analysis ofthe results
obtained from some processes in Chapter 3. These includes results from pre-processing,training,and the
developed model.The testing of the developed model and statistical analysis of its performances in different
input-targetscenarios are presented in Chapter 5.Chapter 6, presents the steps leading to the development,
packaging and deploymentofthe Application software that can be installed on a network system for the purpose
of predicting future mean traffic of an existing network and the number oftraffic channels required to adequately
service the needs based on given QoS requirements.In Chapter 7, conclusions,recommendations and direction
for future work are presented.