Automatic Planning of 3G UMTS All-IP Release 4 Networks with Realistic Trac

by

Mohammad Reza Pasandideh

A thesis submitted to The Faculty of Graduate and Postdoctoral Aairs in partial fulllment of the requirements for the degree of Master of Applied Science in Electrical and Computer Engineering

OttawaCarleton Institute for Electrical and Computer Engineering (OCIECE) Department of Systems and Computer Engineering Carleton University

Ottawa, Ontario, Canada, K1S 5B6 Copyright 2011, Mohammad Reza Pasandideh

The undersigned recommend to the Faculty of Graduate and Postdoctoral Aairs acceptance of the thesis

Automatic Planning of 3G UMTS All-IP Release 4 Networks with Realistic Trac

submitted by

Mohammad Reza Pasandideh, B.A.Sc.

in partial fulllment of the requirements for the degree of Master of Applied Science in Electrical and Computer Engineering

Chair, Howard Schwartz, Department of Systems and Computer Engineering

Thesis Supervisor, Marc St-Hilaire

Carleton University January 2011

Abstract
This thesis focuses on designing automatic planning tools for the planning problem of 3rd Generation (3G) Universal Mobile Telecommunication System (UMTS) all-IP Release 4 networks. A new mathematical model for the design problem of such architecture was proposed. The main advantage of the proposed model is to incorporate a realistic trac prole taken from real live networks. Two approximate algorithms based on the local search and tabu search principles are adopted to solve the problem. Numerical results show that good solutions are found with the proposed heuristics. Results demonstrate that the local search algorithm produces solutions that are, on average, at 4.98% of the optimal solution, and in the worst case at 11.31% of the optimal solution. Better solutions are obtained using the tabu search algorithm. Indeed, tabu search is able to provide solutions with an average gap of 2.82% and a maximum gap of 7.51% from the optimal solution.

iii

Acknowledgments
I wish to express my deepest gratitude to Prof. Marc St-Hilaire, my supervisor, for many suggestions and constant supports during this thesis work. I would have been lost without them. His tireless and detailed involvement in my research has been an invaluable asset for me. I would also like to thank my thesis readers Prof. Amiya Nayak and Prof. Thomas Kunz for their time and constructive comments. I would like to thank my parents, Eynollah (RIP) and Rezwan Pasandideh, and my sisters; Shamsi, Shahnaz, Shahla and Sholeh for their encouragement and love. This thesis would not have been possible without their countless sacrices. I am very thankful to my brothers Alireza and Morteza for their continuous presence and support. I have been very fortunate to have so many friends who have always supported me and enriched my life. To name a few, I would like to thank Masud Naaseri, Behzad Khalvati, Farzad Khalvati, Nader Javani, Payvand Parvizi, Yasin Miyar, Armando Espana, Arash Shokrani, Reza Yazdani, Bagher Khanloo, Ali, Reza and Saeed Roshani, Hamid Sakhidel, Amir Kharrazi, Siamak kashanichi, and Rostam Shirmardian. I am grateful to Mohammad Zulhasnine, Sebastian Szyszkowicz and Frank Cervantes for very fruitful discussions that helped me to solve some of my research problems. I would also like to thank my ex-colleagues in Nokia Siemens Network Company; Hamid Dabir, Sanaz Alizadeh, Zakaria Hadi and Rogin Thomas for their support during my research.

iv

I am deeply grateful to my wife, Narges Karimi Tabar, for all the love, support and encouragement that she has given me. She always gave me hope and energy to complete this thesis. Finally, I would like to dedicate this thesis to my father, Eynollah, may his soul rest in peace.

To my father, Eynollah (may your soul rest in peace)

vi

Contents

1 Introduction 1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1.1 1.1.2 1.1.3 3G UMTS Networks . . . . . . . . . . . . . . . . . . . . . . . Network Planning Process . . . . . . . . . . . . . . . . . . . . Network Planning Techniques . . . . . . . . . . . . . . . . . . 1.1.3.1 1.1.3.2 Exact Algorithms . . . . . . . . . . . . . . . . . . . . Approximate Algorithms . . . . . . . . . . . . . . . .

1 2 2 6 8 9 10 11 13 13 15 17 18 18 19 20

1.2 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 Research Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.6 Thesis Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Literature Review on the Planning of UMTS Networks 2.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Evolution of Cellular Networks 2.2.1 . . . . . . . . . . . . . . . . . . . . .

1G Cellular Networks . . . . . . . . . . . . . . . . . . . . . . . vii

2.2.2 2.2.3 2.2.4 2.2.5 2.2.6 2.2.7

2G Cellular Networks . . . . . . . . . . . . . . . . . . . . . . . 2.5G Cellular Networks . . . . . . . . . . . . . . . . . . . . . . 2.75G Cellular Networks . . . . . . . . . . . . . . . . . . . . . 3G Cellular Networks . . . . . . . . . . . . . . . . . . . . . . . 3G Transitional Cellular Networks . . . . . . . . . . . . . . . . 4G Cellular Networks . . . . . . . . . . . . . . . . . . . . . . .

21 23 25 26 28 29 31 31 33 34 35 36 37 38 41 43 46 47 50 51 52 55 55

2.3 Sequential Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.1 The Cell Planning Sub-Problem . . . . . . . . . . . . . . . . . 2.3.1.1 2.3.1.2 2.3.1.3 2.3.1.4 2.3.1.5 2.3.1.6 2.3.2 Cell Planning Objectives . . . . . . . . . . . . . . . . Cell Planning Inputs and Outputs . . . . . . . . . . Trac Modeling of Mobile Users . . . . . . . . . . . Air Interface Propagation Models . . . . . . . . . . . Air Interface Power Control . . . . . . . . . . . . . . Cell Planning Algorithms . . . . . . . . . . . . . . .

The Access Network Planning Sub-Problem . . . . . . . . . . 2.3.2.1 2.3.2.2 Cost-eective Access Networks . . . . . . . . . . . . Reliable Access Networks . . . . . . . . . . . . . . .

2.3.3 2.3.4 2.4 2.5

The Core Network Planning Sub-Problem . . . . . . . . . . . Remarks on Sequential approach . . . . . . . . . . . . . . . .

Global Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Section Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3 Network Model Formulation and Planning Tools 3.1 Basic Concepts in Optimization . . . . . . . . . . . . . . . . . . . . .

viii

3.2 Global Planning Problem: Inputs, Outputs and Objectives . . . . . . 3.3 Model Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.1 3.3.2 Node B Trac Record . . . . . . . . . . . . . . . . . . . . . . Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.2.1 3.3.2.2 3.3.2.3 3.3.2.4 Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . Decision Variables . . . . . . . . . . . . . . . . . . . Trac variables . . . . . . . . . . . . . . . . . . . . . Cost Parameters . . . . . . . . . . . . . . . . . . . .

58 59 60 62 62 64 66 67 69 69 81 82 82 84 87 92 92 97 98 98 99

3.4 Cost Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6 Exact Methods for the Design Problem . . . . . . . . . . . . . . . . . 3.7 Heuristics for the Design Problem . . . . . . . . . . . . . . . . . . . . 3.7.1 3.7.2 3.7.3 Solution Representation and Initial Solution . . . . . . . . . . Local Search Heuristic . . . . . . . . . . . . . . . . . . . . . . Tabu Search Heuristic . . . . . . . . . . . . . . . . . . . . . .

4 Experiment Design and Result Analysis 4.1 Experiment Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.1 4.1.2 Implementation of the Exact Algorithm . . . . . . . . . . . . . Implementation of Approximate Algorithms . . . . . . . . . . 4.1.2.1 4.1.2.2 Local Search . . . . . . . . . . . . . . . . . . . . . . Tabu Search . . . . . . . . . . . . . . . . . . . . . . .

4.2 Illustrative Example . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 4.3 Result Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

ix

4.3.1 4.3.2

Numerical Analysis . . . . . . . . . . . . . . . . . . . . . . . . 105 Statistical Analysis . . . . . . . . . . . . . . . . . . . . . . . . 111

4.4 Section Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 5 Conclusions and Future Work Appendix A Appendix B References 114 117 121 133

List of Tables
3.1 Typical example of trac record . . . . . . . . . . . . . . . . . . . . . 3.2 Structure of the tabu list size . . . . . . . . . . . . . . . . . . . . . . 4.1 The subscriber trac prole . . . . . . . . . . . . . . . . . . . . . . . 4.2 The planning parameters . . . . . . . . . . . . . . . . . . . . . . . . . 4.3 Features of the Node B types . . . . . . . . . . . . . . . . . . . . . . 4.4 Features of the RNC types . . . . . . . . . . . . . . . . . . . . . . . . 4.5 Features of the MGW types . . . . . . . . . . . . . . . . . . . . . . . 4.6 Features of the MSS types . . . . . . . . . . . . . . . . . . . . . . . . 4.7 Features of the SGSN types . . . . . . . . . . . . . . . . . . . . . . . 4.8 Features of the GGSN types . . . . . . . . . . . . . . . . . . . . . . . 4.9 Features of the router types . . . . . . . . . . . . . . . . . . . . . . . 4.10 Costs of the link and interface types . . . . . . . . . . . . . . . . . . . 61 89 93 93 95 95 95 95 96 96 96 96

4.11 Problem sizes - set#1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 4.12 Problem sizes - set#2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 4.13 Numerical results for the rst instance of set#1 . . . . . . . . . . . . 107 4.14 Numerical results for the rst instance of set#2 . . . . . . . . . . . . 108 4.15 Solution gap comparison for the total simulations . . . . . . . . . . . 112 xi

4.16 CPU time comparison for the total simulations . . . . . . . . . . . . . 112 B.1 Numerical results for the second instance of set#1 . . . . . . . . . . . 121 B.2 Numerical results for the second instance of set#2 . . . . . . . . . . . 123 B.3 Numerical results for the third instance of set#1 . . . . . . . . . . . . 125 B.4 Numerical results for the third instance of set#2 . . . . . . . . . . . . 127 B.5 Numerical results for the fourth instance of set#1 . . . . . . . . . . . 129 B.6 Numerical results for the fourth instance of set#2 . . . . . . . . . . . 131

xii

List of Figures
1.1 3GPP UMTS Release 4 network architecture [1] . . . . . . . . . . . . 1.2 3GPP UMTS all-IP Release 4 architecture . . . . . . . . . . . . . . . 1.3 Network planning process [2] . . . . . . . . . . . . . . . . . . . . . . . 2.1 Architecture of 2.5G cellular networks . . . . . . . . . . . . . . . . . . 2.2 IMT-2000 radio interface standards [3] . . . . . . . . . . . . . . . . . 2.3 Sequential approach for UMTS planning sub-problems [4] . . . . . . . 2.4 Cell planning sub-problem [4] . . . . . . . . . . . . . . . . . . . . . . 2.5 Access network planning sub-problem . . . . . . . . . . . . . . . . . . 2.6 Core network planning sub-problem . . . . . . . . . . . . . . . . . . . 2.7 Global approach for the UMTS planning problem [4] . . . . . . . . . 3.1 Global minimum vs. local minimum . . . . . . . . . . . . . . . . . . . 3.2 Graphical representation of the model formulation . . . . . . . . . . . 3.3 Example of the NE-Install vector . . . . . . . . . . . . . . . . . . . . 3.4 Example of the NE-Type vector . . . . . . . . . . . . . . . . . . . . . 3.5 Local search algorithm [5] . . . . . . . . . . . . . . . . . . . . . . . . 3.6 Tabu search algorithm [6] . . . . . . . . . . . . . . . . . . . . . . . . 3 5 6 24 27 31 32 41 48 51 58 62 83 83 85 88

xiii

4.1 Potential locations for the illustrated example . . . . . . . . . . . . . 101 4.2 Solution for the illustrated problem obtained with LS . . . . . . . . . 102 4.3 Solution for the illustrated problem obtained with TS . . . . . . . . . 103 4.4 Solution for the illustrated problem obtained with CPLEX . . . . . . 104 4.5 Solution comparison for the rst instance of set#1 (cost) . . . . . . . 109 4.6 Solution comparison for the rst instance of set#2 (cost) . . . . . . . 109 4.7 Solution comparison for the rst instance of set#1 (CPU) . . . . . . 110 4.8 Solution comparison for the rst instance of set#2 (CPU) . . . . . . 110 B.1 Solution comparison for the second instance of set#1 (cost) . . . . . . 122 B.2 Solution comparison for the second instance of set#1 (CPU) . . . . . 122 B.3 Solution comparison for the second instance of set#2 (cost) . . . . . . 124 B.4 Solution comparison for the second instance of set#2 (CPU) . . . . . 124 B.5 Solution comparison for the third instance of set#1 (cost) . . . . . . 126 B.6 Solution comparison for the third instance of set#1 (CPU) . . . . . . 126 B.7 Solution comparison for the third instance of set#2 (cost) . . . . . . 128 B.8 Solution comparison for the third instance of set#2(CPU) . . . . . . 128 B.9 Solution comparison for the fourth instance of set#1 (cost) . . . . . . 130 B.10 Solution comparison for the fourth instance of set#1 (CPU) . . . . . 130 B.11 Solution comparison for the fourth instance of set#2 (cost) . . . . . . 132 B.12 Solution comparison for the fourth instance of set#2 (CPU) . . . . . 132

xiv

Acronyms
1G 2G 3G 3GPP 3GPP2 4G ACO ACP AMC AMPS AMR ATM AUC BH BHCA BSC BS First Generation Second Generation Third Generation 3rd Generation Partnership Project 3rd Generation Partnership Project 2 Fourth Generation Ant Colony Optimization Automatic Cell Planning Adaptive Modulation and Coding Advanced Mobile Phone Service Adaptive Multi Rate Asynchronous Transfer Mode Authentication Center Busy Hour Busy Hour Call Attempt Base Station Controller Base Station xv

BTS CAPEX CDMA CEPT CN CSD CS EDGE EIR ETSI FDD FDM FM GGSN GMSK GPRS GSM HARQ HLR HSCSD HSDPA HSUPA IETF IMS

Base Transiver Station Capital Expenditure Code Division Multiple Access Conference on European Post and Telecommunications Core Network Circuit Switch Data Circuit Switching Enhanced Data Rate for GSM Evolution Equipment Identity Register European Telecommunications Standards Institute Frequency Division Duplex Frequency Division Multiplexing Frequency Modulation Gateway GPRS Support Node Gaussian Minimum Shift Keying General Packet Radio Service Global Systems for Mobile Hybrid Automatic Repeat Request Home Location Register High Speed Circuit Switch Data High Speed Downlink Packet Access High Speed Uplink Packet Access Internet Engineering Task Force IP Multimedia Subsystem xvi

IMT IP IS ITU-T IWO LP LS LTE MGW MIMO MIP MMS MPLS MS MSC MSS MU NE NMT NP-hard OFDMA OPEX P2P PC

International Mobile Telecommunications Internet Protocol Interim Standard International Telecommunication Union-Telecommunication Invasive Weed Optimization Linear Programming Local Search Long Term Evolution Media Gateway Multi-Input Multi-Output Mixed Integer Programming Multi Media Service Multi Protocol Label Switching Mobile Subscriber Mobile Switching Center Mobile Switching Server Mobile User Network Element Nordic Mobile Telephony Non-deterministic Polynomial-time hard Orthogonal Frequency Division Multiple Access Operational Expenditure Point to Point Power Control xvii

PDP PS PSK PSTN QAM QoS QPSK RAN RF RNC RTT SGSN SIR SMS TACS TDD TDMA TIA TL TS UMTS UTRAN VAS VLR

Packet Data Protocol Packet Switching Phase Shift Keying Public Switch Telephony Network Quadrature Amplitude Modulation Quality of Service Quadrature Phase-Shift Keying Radio Access Network Radio Frequency Radio Network Controller Radio Transmission Technology Serving GPRS Support Node Signal to Interface Ratio Short Message Service Total Access Communications Systems Time Division Duplex Time Division Multiple Access Telecommunications Industry Associations Time Limit Tabu Search Universal Mobile Telecommunication System Universal Terrestrial Radio Access Network Value Added Service Visitor Location Register xviii

WCDMA WiMAX

Wideband Code Division Multiple Access Worldwide Interoperability for Microwave Access

xix

Chapter 1 Introduction
Wireless communications have a key role in human life. Innovations in microelectronics serve other sciences and result in development of various markets including telecommunications. The growth can be seen from the growth of telecom companies and increasing competition among service providers. Over the recent years, with presence of the Internet Protocol (IP) technology and popularity of modern cell phones, demand for new and hi-speed services has been increasing rapidly. Extraordinary success of wireless industry, and specically cellular networks, has motivated academia and industry researchers to look for optimized planning methods for such networks. Standardization organizations, in turns, have converged several researches and studies into so-called standards. Wireless cellular networks have unbelievably spread across the globe during the last two decades and currently, 3rd Generation (3G) Universal Mobile Telecommunication System (UMTS) networks are very popular in the world. 3G cellular systems are very exible, but more complex and costly compared to the older systems which 1

make the design and planning of such networks very challenging. In this context, the competitive market of cellular networks mandates operators to capitalize on ecient design tools. Planning tools are used to optimize networks and keep both operators and users satised. On one side, users expect to have seamless access to dierent high quality services with aordable prices. On the other side, operators expect to have an always-operational network with high number of users, capacity and quality with low Capital Expenditure (CAPEX) and Operational Expenditure (OPEX). In this thesis, we are interested in the design problem of third generation UMTS all-IP Release 4.0 networks. As a result, this introduction chapter is divided as follows. The rst sub-section will provide background information on relevant topics. Then, the problem statement will be exposed followed by the research objectives and the proposed methodology. Finally, the main contributions will be outlined and an overview of the remaining parts of the thesis will be given.

1.1

Background

In this part, we rst review 3G UMTS networks. Then, we outline the planning steps of cellular networks. Finally, dierent planning techniques in the area of cellular networks are discussed.

1.1.1

3G UMTS Networks

3G networks are designed for improved voice quality, high-speed Internet and multimedia services. The standards for 3G networks were drafted by International Telecommunication Union (ITU) and 3G UMTS networks were developed under 3rd Gener2

Figure 1.1: 3GPP UMTS Release 4 network architecture [1] ation Partnership Project (3GPP) standards. 3GPP standards are structured as Releases whereas each Release is called to a set of main improvements respect to the old Release. The rst generation of 3GPP UMTS networks, known as Release 99 [7], was deployed in 2000 for the rst time. Later, 3GPP introduced Release 4 networks with major improvements respect to the Release 99. The architecture of 3GPP UMTS Release 4 networks is shown in Figure 1.1. UMTS networks are composed of two main sub-networks: the Radio Access Network (RAN), also known as Universal Terrestrial Radio Access Network (UTRAN) and the Core Network (CN). UTRAN is the geographical coverage area of UMTS network, comprising of a set of cells wherein Mobile Users (MUs) can successfully enjoy wireless services. In UMTS networks, the rst contact point of the MUs with 3

cellular network is known as Node B. The Node B equipped with antennas has similar functionalities of 2G BSs, but with new features like soft handover derived from WCDMA technology. Another network element in UTRAN is Radio Network Controller (RNC), which has the similar role of BSC in 2G networks. RNC provides connectivity between one or more Node Bs and the core network, as well as radio resource and mobility management [8]. The core network in Release 4 is divided into CS and PS domains. Unlike Release 99, the signaling and voice tracs in Release 4 are handled by separate network elements. Basically, the MSC is broken into an MSC Server (MSS) for control plane functionalities and a Media Gateway (MGW) for user plane functionalities. The control plane includes application protocols, mobility management and call control logic for circuit-switched calls, established between RNC and MSS, while the user plane corresponds to the voice streams of circuit-switched calls, established between RNC and MGW [1]. Splitting MSC into MSS and MGW makes Release 4 scalable, reliable and cost-eective with respect to Release 99 [9]. The functionalities of SGSN and GGSN in Release 4 networks are similar to those of 2.5G. The network transport protocol in Release 99 was Asynchronous Transfer Mode (ATM). With the advent of IP and mechanisms like Multi Protocol Label Switching (MPLS), IP became a suitable replacement for costly ATM solutions. IP/MPLS was used in Release 4 core backhaul for the rst time and was extended to UTRAN in next Releases. By so doing, cellular networks started enjoying at-IP architecture wherein all nodes can reach each other via IP connectivity [10]. One of the main advantages of using at-IP architecture is the capability to transport dierent trac types over a common IP/MPLS network which could save cost for operators. The 4

Figure 1.2: 3GPP UMTS all-IP Release 4 architecture edge router and core router are the new elements which are installed in UTRAN and core sites respectively for packet routing purposes. The idea of mixing at-IP or all-IP architecture with Release 4 network was very attractive and was appreciated by many operators. The advantage of easily adaptation of all-IP Release 4 networks to state-of-the-art networks made it so popular in the world. A typical architecture for the 3GPP UMTS all-IP Release 4 network is shown in Figure 1.2. 3G UMTS all-IP Release 4 network will be used as the model in this thesis.

Network Design Inputs

Network Design Process

Network Performance Analysis

Final Design

Figure 1.3: Network planning process [2]

1.1.2

Network Planning Process

Due to the highly complex nature of cellular networks, it is imperative to follow a structured and organized design process. A step by step UMTS network design can be found in reference [11]. Network Planning is an iterative process and can be summarized in three steps as show in Figure 1.3. Network Design Inputs: The rst step in network planning is to gather network inputs and requirements. This involves collecting extensive amount of information like anticipated trac volume and types, trac distribution, equipment types and cost, candidate technologies and providers, reliability of the technologies, constraints and limitations, utilization, resources and so on. In the case of a green eld network (i.e. there is no existing infrastructure), collecting design inputs could be a time consuming task and the unavailable inputs should be predicted or generated. For an existing network, design (or 6

upgrade) inputs can be collected easier than green eld networks, as many data can be obtained from existing resources. Network Design Process: This step requires exploring various design techniques and algorithms to create a network topology. It comprises of realizing each network element type, number and location, link and interface types, connectivity, as well as trac routing. Due to the large number of potential combinations, especially for large size networks, this step can not be done manually and is usually done by a planning tool. The network design inputs are given to the specic tool and the tool produces a topology for the network in the output. Depending on planning tool, extra information like cost of the network can also be part of the output. Network Performance Analysis: Also known as optimization, network performance analysis is the last step in planning to evaluate the topology developed in design phase. The evaluation is done based on certain criteria such as cost, reliability, coverage, capacity, etc. The good solution is used as benchmark for further renements. This step requires repeating the planning from the beginning either by revised input data (i.e. revised trac forecast) or dierent design approach. After a pre-dened number of iterations, the nal design is produced. The nal design is assumed to be the best solution among several iterations. Because network design involves exploring various alternatives, it is impossible to do it manually or without computer-based planning tools. Automatic planning tools involve optimization techniques and will be discussed in the next section. 7

Producing planning tools has been the interest of some researchers. Allen et al. [12] introduced and used CDS-SmartPlan tool in a case study for UMTS network design in London. Tutschku et al. [13] presented ICEPT tool based on an integrated design approach. A comprehensive research on radio planning was done by Eisenblatter et al. under the EU-project of MOMENTUM [14]. The planning tools usually are supported by industry and become commercial. Mentum Planet 5 [15], Atoll [16] and Capesso [17] are popular planning tools for cellular networks. A list of commercial planning tools can be found in reference [18].

1.1.3

Network Planning Techniques

Early generations of cellular networks were simple and relatively small in size and it was possible to design them manually. Manual design processes are not ecient and prone to human error. On the other hand, complex and large size networks required huge amount of computations which ultimately oblige cellular operators to look for automatic computer based planning tools. There are two major methods to approach a network design problem. The rst approach is simulation and is used whenever the network design problem cannot be expressed in analytical or mathematical form. To do the simulation, a tool is required to build a simplied representation of the network. Then, the behavior of the simulated network is investigated under various inputs and parameters. Simulation sometimes requires a large amount of time and money, but for some cases, it might be the only solution. The alternative approach to the network design problem is expressing the problem in a logical and mathematical way which can be tackled by algorithms. Algorithms 8

are the methods, used for solving a problem within a nite number of steps [2]. The formulation of a mathematical model for the network is the prime task which characterizes the problem by using the inputs, objectives and constraints. Then, an algorithm is employed to solve the problem by performing tremendous amount of calculations to nd the solutions. Algorithms are embedded inside the network planning tools. There are many ways to classify algorithm, but in the context of combinatorial optimization, algorithms are classied in exact (complete) and approximate categories. Combinatorial optimization is a method to solve combinatorial problems. The intention is to nd a minimum or maximum for a function where the set of feasible solutions is discrete or can be reduced to discrete.

1.1.3.1

Exact Algorithms

Exact algorithms are used to nd a solution by searching all potential solutions in the search space within a bounded time. However, for a very large search spaces, exact algorithms are not ecient and require high Central Processing Unit (CPU) time. CPU time is often referred to the time required to process a problem and nd the solution. In fact, exact algorithms might need exponential computational time in practical cases [19]. Linear Programming (LP) and Integer Programming are examples of exact algorithms. LP is a mathematical method for optimization of linear object function. If the decision variables are required to be integer, the LP is called Linear Integer Programming or Integer Programming in short [20]. In this thesis, CPLEX Mixed Integer Optimizer 12.1 (see reference [21] for more information about CPLEX) is used. The algorithm used in CPLEX is branch and 9

cut. Details on the branch and cut algorithm will be described later in chapter 3. 1.1.3.2 Approximate Algorithms

Approximate algorithms are also known as heuristics. As their names implies, approximate algorithms provide relatively good solutions for optimization problems. Approximate algorithms sacrice the guarantee of nding optimal solutions for the sake of providing good solutions in a signicantly reduced amount of time [19]. A general class of heuristics, known as metaheuristics, are used in combinatorial optimization. Local Search Methods The local search method, denoted as LS, starts from a random initial solution and try to replace the current solution by a better solution in an appropriately dened neighborhood of the current solution [19]. A neighborhood is a new solution, achieved by applying a move over the current solution. Each move is performed if the resulting solution is better (in terms of cost) than the current solution. LS algorithm is trapped in the rst local minimum, which may or may not be a global minimum. Metaheuristics Metaheuristics are a class of approximate algorithms introduced in 1970s. With the objective of eectively explore the search space, metaheuristics try to combine basic heuristic methods in a higher level frameworks. The class of metaheuristics includes, but is not restricted to, Ant Colony Optimization (ACO), Evolutionary Computation (EC), Iterated Local Search (ILS), Simulated Annealing (SA) and Tabu Search (TS) [99]. The tabu search is an advanced algorithm which is able to nd global minimum. 10

Details of TS algorithm will be described later in chapter 3. Information about the other metaheuristics can be found in reference [19].

1.2

Problem Statement

The cellular industry is growing. Rapid increase in the demand for data services has pushed wireless operators to invest in new technologies. Operators capitalize a major portion of their money in their network infrastructure to be able to oer new services with high quality and lower rates. To survive in such a competitive market, they look for network planning tools which can design an optimized network with low cost. Furthermore, an optimized network requires less maintenance cost, meaning more saving. The primary goal of UMTS network planning tools is to provide an optimum topology for the network. Many parameters and factors are involved to select optimum node location, node and link types, topology and so on. Several models and algorithms have been proposed to solve the planning problem of UMTS networks. However, due to the high complexity of the problem, most of the proposals only focus on a portion of the network. In fact, the entire planning problem is decomposed into three sub-problems: the cell planning sub-problem, the access network planning subproblem and the core network planning sub-problem. There are two main approaches to tackle the planning problem: the sequential and the global. In the sequential approach, each sub-problem is solved separately. Although the sequential approach reduces the problem complexity, it neglects the interactions between sub-problems. Moreover, combining the partial solutions obtained by the sequential approach might 11

not result in an optimal solution for the whole network planning problem. In the global approach, the three sub-problems are considered together and solved jointly. Although a global model is more complex, but it provides optimal solutions. The planning problem of UMTS network is NP-hard [22]. NP-hard problems are so complex that is unlikely to nd an exact solution in polynomial time [23]. This means that computing the optimal solution can be very long or even unfeasible. As a result, approximate methods, which provide good solutions in less CPU time, are preferred to the exact methods which require longer CPU time. Most UMTS network models proposed in the literature deal with Release 99. Since then, new releases (or versions), such as Release 4, have been standardized. It is therefore important to consider these new releases. Moreover, most planning tools consider a very simple trac model. In order to have a realistic planning tools, realistic trac prole that capture all aspects of the voice and data tracs (e.g. Busy Hour Call Attempt (BHCA), simultaneous calls, bandwidth, signaling trac, etc.) must be considered. The advantage of using IP technology within Release 4 network architecture, makes it more cost-optimized and exible. Furthermore, the realistic trac prole reects more attributes from the realistic networks. As a result, the planning tool developed in this thesis, would be more appropriate to be used by operators dealing with real networks.

12

1.3

Research Objectives

The main objective of this thesis is to develop automatic planning tools based on exact and approximate algorithms in order to solve the planning problem of 3G UMTS all-IP Release 4 networks with realistic trac prole. More specically, we aim to: 1. Propose a mathematical model in order to solve the planning problem of 3G UMTS all-IP Release 4 networks (at the expense of computational complexity). 2. Develop approximate algorithms (based on local search and tabu search) in order to solve the planning problem of 3G UMTS all-IP Release 4 networks. This should allow us to solve larger instances of the problem. 3. Compare the performance of the approximate algorithms with respect to a reference value.

1.4

Methodology

The inherent complexity of the cellular networks design problem obliges network planners to take a well-disciplined approach. Problem modeling and solving the model is an exhaustive task, as many parameters and variables are involved. The methodology used to approach the planning problem in this thesis is described below: 1. Study cellular networks technologies: In order to acquire a good knowledge about cellular networks technologies, the evolution of such networks was studied from the beginning up to now. 13

2. Select technology for the thesis model: Reviewing the literature on the area of cellular networks design problem, it was understood that, as of today, the design problem of 3G UMTS all-IP Release 4 networks has not been tackled by academia and could be a potential planning assignment. Another motivation was the fact that such architecture is adaptable to the next generation networks. To make the planning tool usable in industry, a realistic trac prole, commonly used by operators was also an important aspect to consider. 3. Dene the network planning problem: The architecture of the thesis model was extracted from 3GPP standards and topology of the live networks. The planning problem of 3G UMTS Release 4 networks was determined. The objective function is composed of cost of the nodes, links and interfaces and is expected to be the minimum, subject to a series of constraints. 4. Construct a mathematical model: Given the planning problem in previous step and by using a set of decision variables and parameters, the characteristics of the network elements, links and interfaces including their capacity, type and cost were determined and mathematical model was constructed. 5. Implement exact algorithm: The linear programming method was used to solve the thesis problem. Due to high number of constraints and complexity of the problem, we used C++ programming language to translate objective function, constraints and bounds into an LP le format. (For more information on LP le format, see [21]). Finally CPLEX was used to solve the LP problem. 6. Study optimization methods: The planning problem of UMTS networks is NP-

14

hard problem. NP-hard problem have not exact solution in a polynomial time. Linear integer programming used as an optimization method to nd the exact solutions. Approximate solutions are also obtained from two approximate algorithms: local search and tabu search. 7. Implement approximate algorithms: A time-ecient way to tackle NP-hard problem is to use approximate algorithms. To solve the problem, we implemented LS and TS algorithms by using C++ programming language. Both algorithms were incorporated within the same C++ le. First LS was used to solve the problem and nd solutions. Later the solution found by LS was given to TS as an initial solution. TS performed more searches and found better solutions in terms of cost. 8. Compare solutions: A set of problems were solved by CPLEX, LS and TS methods. The cost and CPU time of the solutions were calculated and compared together. In fact, the approximate solutions found by LS and TS were evaluated respecting to the exact solutions found by CPLEX. Finally, the eciency of the LS and TS algorithms to solve the thesis planning problem were investigated and analyzed.

1.5

Contributions

The main contributions of this thesis are summarized as follows: 1. Modeling UMTS Release 4.0 networks with realistic trac prole: A mathematical model for the design problem of such architecture was proposed and 15

analyzed. The results of this work were published in the following conference paper: M. R. Pasandideh and M. St-Hilaire, Automatic Planning of UMTS Release 4.0 Networks using Realistic Trac, in Proc. IEEE International World of Wireless Mobile and Multimedia Networks (WoWMoM) Symposium., 2010, pp. 1-9. 2. Modeling UMTS All-IP Release 4 networks with realistic trac prole: An extension to the model proposed in 1 was to include an all-IP infrastructure. The results of this work were presented to the following conferences: PASANDIDEH, M. R. and ST-HILAIRE, M., Optimization Model for the Design of All-IP UMTS Release 4.0 Networks , Optimization Days, Montreal, May 2010. PASANDIDEH, M. R. and ST-HILAIRE, M., On the Planning Problem of All-IP UMTS Release 4.0 Networks with Realistic Trac , 10th INFORMS Telecommunications Conference, Montreal, May 2010. 3. Proposing a local search heuristic for the design problem of UMTS all-IP Release 4 networks with realistic trac prole: A local search heuristics was developed to solve the design problem of such architecture. 4. Proposing a tabu search heuristic for the design problem of UMTS all-IP Release 4 networks with realistic trac prole, solution comparison and analysis: A tabu search heuristic was developed to solve the design problem of such ar-

16

chitecture. The eciency of the heuristic was compared with the local search and the reference model. Results will be submitted to a journal paper.

1.6

Thesis Overview

The reminder of the thesis is organized as follows: In Chapter 2, we present a selective review on cellular network planning problems found in the literature. In Chapter 3, we rst present the mathematical formulation of the problem. This model will be used to nd optimal solutions. In Chapter 4, we present the simulation results and compare the dierent algorithms. In Chapter 5, we conclude the work and suggest future directions.

17

Chapter 2 Literature Review on the Planning of UMTS Networks

2.1 Background

The planning problem of UMTS networks is a complex task. In order to reduce the complexity, the planning problem is usually decomposed into three sub-problems: the cell planning sub-problem, the access network planning subproblem and the core network planning sub-problem. Scholars have mainly two approaches to tackle these sub-problems: sequential and global. In the sequential approach, each sub-problem is considered separately and is solved solely; while the global approach involves more than one sub-problem at a time. In this chapter, rst we study the evolution of cellular networks. Then, we review the literature in the area of UMTS network planning in the context of sequential and global approaches. First, we describe the sequential approach and each sub-problem 18

in detail. We present the major works in solving the three sub-problems in terms of methods and algorithms. Then, we discuss the global approach and present the related works in literature.

2.2

Evolution of Cellular Networks

History of mobile telephony dates back to the 1920s with the use of radiotelephony by the police department in United States. The initial equipment were bulky and phones were not dealing well with obstacles and buildings. Introducing Frequency Modulation (FM) in 1930s made some progress and helped radio communications in battleeld during World War II. The rst mobile telephony was introduced in 1940s with limited capacity and manoeuvre. Mobile communications development continued for years to become commercial as we have it today. Terminology of generation is used to dierentiate the signicant technology improvement in cellular networks which in turn, resulted in major changes in the wireless industry. The rst generation (1G) of cellular networks was introduced in late 1970s, which was followed by the second generation (2G) in early 1990s, the third generation (3G) in early 2000 and the fourth generation (4G) nowadays. Changes from analog to digital technology, implementing new multiplexing and access techniques, employing new codes and frequencies, introducing IP as a substitution for legacy transmission methods and many other innovations resulted in networks with more services, higher capacity, speed and security. In the following sub-sections, we explain dierent generations of cellular networks and discuss their specications.

19

2.2.1

1G Cellular Networks

The trial system of what is known today as the rst generation (1G) of cellular networks was implemented in Chicago in 1978 and commercially launched in 1983. The technology used in the system was known as Advanced Mobile Phone Service (AMPS) and operated in 800 MHZ band. Japan launched a commercial AMPS system in 1979 and later in 1981, the Nordic Mobile Telephony (NMT) technology based network, operating in 450 MHZ and 900 MHZ bands, was launched in Nordic counties. A modied version of AMPS called Total Access Communications Systems (TACS) was also deployed in UK in 900 MHZ band. Many countries followed along and mobile communications spread out over the world. 1G systems used Frequency Division Multiplexing (FDM) technology to divide predened spectrum into portions named channels. Each channel was able to serve one user at a time. All these technologies formed 1G cellular networks which were only oering analog voice service [3]. The network geographical area is divided in small sectors, each called a cell. Derived from this concept, the technology was named cellular and the phones were called cell phones. The rst generations of cellular networks were incompatible with each others as each network had its own standards. Handsets were expensive and networks had limited capacity and mobility. Moreover, networks had diculties with frequency use, security, roaming, power and so on. Such drawbacks resulted in a very low penetration of 1G cellular networks and mandated a signicant eort to develop the second generation networks.

20

2.2.2

2G Cellular Networks

The second generation (2G) of cellular networks were based on digital communications and were rst deployed in early 1990s. Shifting technology from analog to digital improved the quality, capacity, cost, power, speed, security and quantity of services. Like 1G cellular networks, several types of technologies were developed for the second generation. Depending on the multiplexing technique, 2G cellular networks are divided into two main groups: Time Division Multiple Access (TDMA) and Code Division Multiple Access (CDMA) based cellular networks. Global Systems for Mobile communications (GSM) and Interim Standards 136 (IS-136) are key 2G systems based on TDMA and Interim Standards 95 (IS-95) is a famous 2G system based on CDMA. More information on 2G systems is provided in the following. Global Systems for Mobile communications Despite introducing NMT in Europe, there were several incompatible variants of analog 1G systems deployed across Europe. The need for development of a Europewide digital cellular network was announced during the Conference on European Post and Telecommunications (CEPT) in 1982. CEPT established GSM organization to harmonize all European systems. Several researches and tests were done by GSM group before the foundation of the European Telecommunications Standards Institute (ETSI) in mid 1980s. ETSIs technical groups nalized the rst set of specications for GSM networks in 1989 and the rst GSM network was launched in 1991. Introducing a digital technology, GSM oered new services like Short Message Service (SMS) and moderate Circuit Switch Data (CSD) services beside a better voice service. GSM became rapidly popular in Europe and was deployed in many countries over the world.

21

GSM is based on TDMA and operates in 900 MHZ and 1800 MHZ (850 MHZ and 1900 MHZ in North America). For the 900 MHZ, the uplink frequency band is 935960 MHZ and the downlink frequency band is 890-915 MHZ. Thus, the bandwidth for both uplink and downlink is 25 MHZ which is able to oer 124 carriers with channel spacing of 200 KHZ. In GSM, each Radio Frequency (RF) channel caters 8 speech channels. Techniques like cell sizing and splitting, power control and frequency reuse are applied to increase GSM networks capacity [3]. Interim Standards 136 With the advent of digital technology and application of TDM technology, AMPS which was an analog system based of FDM, was enhanced to the rst Digital AMPS or D-AMPS, known as IS-54B in 1990. Using TDM technique, each 30 KHZ channel became able to cater three speech channels which tripled the capacity. Only voice channels of IS-54B system were digital and still control channels were analog. Despite oering more capacity, analog control channels were limiting number of services. This problem was solved in the next version of IS-54B known as IS-136 and published by Telecommunications Industry Associations (TIA) in 1994. IS-136 operates both in 800 MHZ and 1900 MHZ and adds a number of features to the original IS-54B, including SMS and CSD service [3]. IS-95 CDMA Both GSM and IS-136 use TDMA to share a single RF channel and increase network capacity, but there is another way to increase capacity and that is CDMA. In CDMA, users share the same 1.25 MHZ channel at the same time, but they employ dierent pseudo-random codes. These codes are known in receiver and the original signal is 22

decoded. IS-95 is a 2G cellular system based on CDMA (also known as CDMA2000 ), developed by Qualcomm in United States and introduced in 1989. IS-95 system was able to oer data rate of 14.4 Kbps, with improved voice quality, coverage and security. IS-95 operates in the 800 MHZ band in North America [3].

2.2.3

2.5G Cellular Networks

2G cellular networks were designed based on Circuit Switching (CS) and were capable to oer good voice services, but low CSD rates (up to 14.4 Kbps). By using multiple number of 14.4 Kbps time slots, GSM successfully introduced High Speed Circuit Switch Data (HSCSD) which could provide data rate of 57.6 Kbps. The problem of HSCSD was a reduction in scarce voice channels and this became a motivation to introduce Packet Switching (PS); a technology for faster data services like Multi Media Service (MMS) and Internet communications. PS technology enhanced 2G cellular networks to 2.5G through adding a PS domain to the existing CS domain. Voice trac in 2.5G systems uses circuit switching like 2G systems, but data trac is based on packet switching. Packet switching allocates radio resources on demand, meaning that resources are utilized only when user is actually sending or receiving data. This allows ecient use of scarce radio resources and rather than dedicating a radio channel to a mobile data user for a xed period of time, the available radio resources can be concurrently shared between multiple users. PS technology is known as CDMA2000 one times Radio Transmission Technology (CDMA2000 1xRTT) in CDMA standards which can provide speed of 144 Kbps. Similarly in GSM standards, PS technology is known as General Packet Radio Service (GPRS), providing speed of 171 Kbps. GPRS uses Gaussian Minimum Shift Keying 23

Figure 2.1: Architecture of 2.5G cellular networks (GMSK) as the modulation scheme [3]. Implementing PS domain in 2.5G network requires introducing new Network Elements (NEs). As shown in Figure 2.1., the general architecture of 2.5G cellular networks is composed of Radio Network and Core Network. Radio Network comprises of Mobile Subscriber (MS),Base Station (BS) and Base Station Controller (BSC). As mentioned earlier, the 2.5G core network is divided in CS and PS domains from functional point of view in which Mobile Switching Center (MSC) performs circuit switching functions in CS domain, while Serving GPRS Support Node (SGSN) and Gateway GPRS Support Node (GGSN) perform packet switching in PS domain. MSC and GGSN also act as a gateway to the Public Switch Telephony Network (PSTN) switch and external data networks (e.g. Internet) respectively. The information of all mobile subscribers currently administrated by the associated MSC is stored in the 24

Visitor Location Register (VLR) which is usually a built-in unit of MSC. Another key NE is the Home Location Register (HLR), acting as a database of subscriber proles. HLR could perform identity check for subscribers and mobile handsets if it is equipped with an Authentication Center (AUC) and an Equipment Identity Register (EIR), otherwise theses functionalities are performed by separate NEs. The Value Added Services (VAS) services such as SMS and voice mail services require integrating additional NEs to the network.

2.2.4

2.75G Cellular Networks

Enhanced Data Rate for GSM Evolution (EDGE) is an improvement over GPRS data rate by using an ecient modulation scheme. EDGE uses 8-Phase Shift Keying (8PSK) as modulation scheme and coexists with GMSK that is used for GPRS. EDGE provides speed of 384 Kbps which is three times more than that of GPRS, but the major advantage of EDGE is its low cost of upgrade. Major change is in software and only minor hardware changes in BS are required to upgrade a 2.5G network to 2.75G [3]. Despite high data rate in 2.75G cellular networks, they lack higher capacities and global roaming. These decits led to keep research and enhancement to bring some convergence for the next generation of cellular networks by reducing the main cellular technologies.

25

2.2.5

3G Cellular Networks

In the absence of global standards, dierent regional standards for cellular networks were set in Europe, Japan, China, Korea, North America and other countries. In order to harmonize dierent variants of existing standards and promoting the next generation of cellular networks, the International Telecommunication UnionTelecommunication (ITU-T) dened a set of specications for 3G networks, under the International Mobile Telecommunications 2000 (IMT-2000) [3]. The followings are the most important specications of IMT-2000: Global standard and exible with next generation of wireless systems; High quality and compatible within IMT-2000 and other xed networks; Worldwide roaming and common frequency band with improved spectrum eciency; High speed packet data rate: 2 Mbps for xed users, 384 Kbps for pedestrian and 144 Kbps for vehicular trac. Based on the above criteria, ITU-T approved ve radio interfaces for IMT-2000 in 1999 and Worldwide Interoperability for Microwave Access (WiMAX) was added as a sixth candidate in 2007. The various radio systems under the IMT-2000 are shown in Figure 2.2. In 1998, ETSI and the standardization organizations from Japan, China and Korea formed a consortium known as 3rd Generation Partnership Project (3GPP) to regulate global specications for 3G systems in scope of ITUs IMT-2000 specications.

26

IMT-2000

IMT-DS W-CDMA (UMTS)

IMT-MC CDMA2000

IMT-TC TD-CDMA and TD-SCDMA (UMTS)

IMT-SC EDGE IS-136

IMT-FT DECT

IP-OFDMA WiMAX

CDMA

TDMA

FDMA

OFDMA

Figure 2.2: IMT-2000 radio interface standards [3] Similarly, in the same year and for the same purpose, TIA and three telecommunications standardization organizations from Japan, China and Korea collaborated on the 3rd Generation Partnership Project 2 (3GPP2) [3]. The bottleneck of 2G cellular networks was low data transmission rate. The next generation of cellular networks would require innovations in radio access techniques to balance growing data applications with higher rates. By the time ITU dened IMT2000 specications for 3G networks, new radio access technologies such as Wideband CDMA (WCDMA) with 5 MHZ bandwidth and CDMA2000 with 1.25 MHZ bandwidth, both beneting Frequency Division Duplex (FDD) and Time Division Duplex (TDD) solutions were introduced. Exploiting such techniques, 3GPP2 introduced CDMA2000 1X-Enhanced Version Data Optimized CDMA2000 (1X-EV/DO) in 1999 as the rst 3G network with data rate as high as 2.4 Mbps in downlink and 153 Kbps in uplink [3].

27

On the other hand, 3GPP mixed the core of GSM networks with WCDMA-based radio access technology and introduced the initial generation of 3G networks, known as UMTS Release 99 in 1999. Later, 3GPP improved the architecture of the core networks and with separating voice and signaling introduced Release 4. By this time, IP-based transport technologies were mature and at-IP architecture was promoted and implemented in Release 4 networks. The above discussed 3G standards by 3GPP and 3GPP2 are in compliance with the IMT-2000 requirement. However research and innovations for next generation of cellular networks continued towards higher data rates with lower latency. A brief study over those enhancements is presented in the next section.

2.2.6

3G Transitional Cellular Networks

3G cellular networks have been improved gradually by 3GPP and 3GPP2 towards higher data rates and various services. New spatial multiplexing techniques like Multi-Input and Multi-Output (MIMO), higher order modulation, ecient scheduling, turbo-coding, Adaptive Modulation and Coding (AMC), Hybrid Automatic Repeat Request (HARQ) and many other innovations made extensions to 3G standards. These extensions are sometimes known as 3.5G, 3.75G and 3.9G series of standards. Following CDMA2000(1X-EV/DO) and with exploiting new techniques in radio access interface, 3GPP2 introduced EV/DO Rev A with speed of 3.1 Mbps and EV/DO Rev B with speed of 14.7 Mbps. On the other hand, 3GPP introduced Release 5 in 2002 promoting IP Multimedia Subsystem (IMS) and High Speed Downlink Packet Access (HSDPA) [24]. Employing HARQ and AMC, HSDPA is able to oer up to 14 Mbps in downlink and 384 Kbps in uplink directions [25]. To increase 28

data throughput in uplink, 3GPP introduced Release 6 [26], known as High Speed Uplink Packet Access (HSUPA) with 5.76 Mbps data rate in uplink. Combination of HSDPA and HSUPA is known as High Speed Packet Access (HSPA) which oers 14.4 Mbps in downlink and 5.76 Mbps in uplink [27]. As of today, HSPA is a popular and operational network in the globe. 3GPP continued development of the standards by employing Quadrature Phase-Shift Keying (QPSK) and Quadrature Amplitude Modulation (QAM) techniques in Release 7 [28]. Release 7 was able to oer 28 Mbps in downlink and 11.5 Mbps in uplink. The Release 7 is often called evolved HSPA or HSPA+. Beside 3GPP and 3GPP2, WiMAX introduced release 1.0 prole for 3G cellular networks in 2007 and ocially entered in competition for 3G and beyond standards. WiMAX release 1.0 was able to meet IMT-2000 specications, providing speed of 46 Mbps in downlink and 4 Mbps in uplink. The roadmap of cellular networks from 3G has now been extended to 4G. Discussion in the next section is on the proposals for 4G standards.

2.2.7

4G Cellular Networks

The fourth generation (4G) of cellular networks is the successor to 2G and 3G standard families which supports all-IP packet switching and ultra-broadband access networks. Similar to IMT-2000, ITU has dened IMT-Advanced specications for 4G networks which any 4G candidate standard should meet those requirements to be ocially recognized and announced to the cellular world by ITU. IMT-Advanced systems encompass several attributes, including higher spectral eciencies, scalable bandwidth up to 40 MHZ and higher data rates. 29

In November 2008, Qualcomm which is the main member of 3GPP2 forum, surprisingly announced ending the development of CDMA standards and decided to adapt 3GPP standards. Dispensing 3GPP2 with development of 4G standards, the competition between 3GPP and WiMAX was highly raised. 3GPP started Long Term Evolution (LTE) project to prepare specications for pre-4G standards. LTE is based on all-IP architecture and it is designed to support voice in PS domain. 3GPP Release 8 [29], based on Orthogonal Frequency Division Multiple Access (OFDMA), is the rst set of LTE standard. Using FDD, TDD and MIMO, Release 8 provides 42.2 Mbps in downlink and 11.5 Mbps in uplink within scalable bandwidth of 1.4 MHZ to 20 MHZ. LTE was deployed rst in Finland and Norway in December 2009. 3GPP enriched LTE specications with Release 9 and Release 10 which are able to oer speed of 84 Mbps and 168 Mbps in downlink respectively and the same speed of 23 Mbps in uplink. 3GPP Release 10 is the rst 4G standards for LTE-Advanced. LTE-Advanced has been designed to meet IMT-Advanced requirements and it has been shown that LTE-Advanced exceeds requirements of IMT-Advanced. On the other hand, WiMAX introduced IEEE 802.16m series of specications, also known as Release 2.0 which is able to oer 120 Mbps in downlink and 60 Mbps in uplink direction. In October 2010, ITU announced both LTE-Advanced and WiMAX as the ocial standards for 4G cellular networks.

30

Input Cell Planning Sub-Problem

Input Access Network Planning Sub-Problem

Input Core Network Planning Sub-Problem Final Solution

Figure 2.3: Sequential approach for UMTS planning sub-problems [4]

2.3

Sequential Approach

In a sequential (or decomposition) approach, the planning problem of UMTS network is divided in three sub-problems [30]: 1. The cell planning sub-problem; 2. The access network planning sub-problem; 3. The core network planning sub-problem. Beside the input of each sub-problem, the output of the previous sub-problem is also used as input for the next sub-problem. As shown in Figure 2.3, the output of the cell planning is used as input for the access network sub-problem. In a similar way, the output of the access network sub-problem is given as input for core network sub-problem. The nal solution is a topology which satises all three sub-problems. Each sub-problem has been widely explored from dierent perspective. In the following sub-sections, each sub-problem is explained and the major works in solving them are presented.

2.3.1

The Cell Planning Sub-Problem

Cell planning is the process of connecting all mobile users to the Node Bs in a specic geographical area. Figure 2.4 shows a typical cell planning assignment. Cell planning 31

Figure 2.4: Cell planning sub-problem [4] in 3G UMTS networks is dierent form that in 2G networks [31, 32]. Cell planning in 2G networks (like GSM) is divided in two steps: coverage and capacity planning. During the coverage planning phase, dierent propagation techniques are used to place BSs in locations where the maximum number of users can receive an acceptable level of signal power. Signal to Interface Ratio (SIR) is a signal quality factor which should be greater than a predened threshold in 2G systems. Capacity planning, also known as frequency planning, is the process of channel (frequency) assignment to the BSs in order to minimize the interference in the network while being able to re-use those frequencies in other cells. Unlike 2G networks, coverage and capacity planning in UMTS network should be done concurrently [33]. Using WCDMA technology in the air interface, mobile users in UMTS network share the whole spectrum, therefore no frequency planning is strictly required [34], but the capacity planning remains a valid and complex task. The main dierences between GSM and UMTS radio network are explained by Neubauer and Toeltsch [31] and Ramzi [32]. 32

2.3.1.1

Cell Planning Objectives

The objective of the cell planning sub-problem depends on the interests of network planners. The following objectives may be the target for a cell planning sub-problem: 1. Minimize network cost; 2. Maximize capacity; 3. Maximize coverage; 4. Maximize signal quality; 5. Minimize electromagnetic eld level. Some of the above objectives are conicting with each other. For example, maximizing the coverage and capacity requires deploying more Node Bs, which in turn, increases the network cost. Another example of contradiction happens when the signal power is increased for maximizing signal quality, but that results in higher electromagnetic eld level. If more than one criterion is considered during the cell planning, then multi-objective functions are dened. A multi-objective function can be produced in either linear and/or weighted combinations of the single objectives. The weighted combination is inspired by the Pareto rule [2]. The problem is formulated by a set of decision variables (parameter space/vector) and a set of objective functions (objective vectors). A certain weight (between 0 and 1) is given to each objective function in the set. When the objective vectors can not be improved in one objective criterion without degrading the other objective criteria, the optimum solution set is found. The weighted multi-objective function provides more exibility 33

as it is possible to put more (less) emphasis on a given objective by assigning higher (lower) weight to it. In linear combination, the objective function is composed of dierent objectives with the same weight. Theil et al. [35] introduce a linear combination of dierent objectives such as cost, interference, coverage, balanced trac and so on. They propose a SA algorithm to solve the site selection problem. Their planning tool can be tuned for speed or quality. Wu and Pierre [36] consider location, maximum transmitted power, antenna height, assignment of demand nodes with BSs and number of BSs as the set of decision variables. Their problem objective is to minimize the cost of BSs, minimize the total emitted power by all users and maximize the total number of active connections. More cell planning multi-objective functions can be found in Maple et al. [37], Jamaa et al. [38] and Choi et al. [39]. Minimizing electromagnetic eld level as an objective for the cell planning problem has been the interest of Crainic et al. [40]. They considered radio protection constraints, handover and downlink capacity constraints and formulated ve objective functions, scaling ve functions of electrical eld levels.

2.3.1.2

Cell Planning Inputs and Outputs

As stated by Amaldi et al. [41], Theil et al. [35] and Lauther et. al [30], dierent inputs are required to solve the cell planning sub-problem. Usually, the following inputs must be known [4]: 1. The potential locations where Node Bs can be installed. Some geographical constraints are applied to restrict the location selection;

34

2. The types (or models) of Node Bs, which includes, but not restricted to, the cost and capacity (e.g. power, sensitivity, switch fabric capacity, interfaces, etc.); 3. The user distributions and their required amount of trac (e.g. voice and data); 4. The coverage and propagation prediction. Various planning algorithms are used to solve cell planning sub-problem. Each algorithm may consider one or more of the objectives mentioned previously. The goal of the cell planning sub-problem is to provide one or more of the following as output: 1. The optimal number of Node Bs; 2. The best locations to install Node Bs; 3. The types of Node Bs; 4. The conguration (height, sector orientation, tilt, power, etc.) of Node Bs; 5. The assignment of mobile users to Node Bs. For the modeling of the cell planning sub-problem, it is required to know how to represent users (or trac) in the model. In the following sub-section trac modeling and related issues are discussed.

2.3.1.3

Trac Modeling of Mobile Users

UMTS networks provide voice and data services for mobile users. It is important to decide how to represent mobile users in the cell planning sub-problem. A basic model could be to represent a user with a point in the cell. For unknown trac distribution, 35

a regular point grid can be used. Dealing with practical cases, as the number of users is high, a clustering or agglomeration technique is required to reduce the complexity. The cluster of users is often called trac node [34] or test point [42]. A trac node or test point represents several mobile users. It is also important to consider the trac (link) direction. Trac direction can be uplink (from user to Node B) or downlink (from Node B to user). Uplink direction is used when planners deal with symmetric trac like voice services. However, if the network is designed to provide data services, downlink direction is more appropriate [8], because downlink is highly utilized for services like web browsing and Internet downloads. The type of area which is aimed to be planned is also required to be known. The area can be rural, urban, sub urban, dense urban and so on. Each of these areas has specic characteristics which need to be taken in account during cell planning assignment.

2.3.1.4

Air Interface Propagation Models

Dealing with radio frequencies, cell planners should consider signal propagation, signal power, fading, as well as interference [43]. Radio propagation prediction model is a set of mathematical expressions, diagrams, and algorithms used to represent the characteristics of a given environment [44]. Dierent models have been created to predict the path loss between transmitter and receiver in dierent environments. The application of each model varies according to frequency, link range, terrain type, land use, land cover and many other parameters. Propagation models are either empirical (referred to as statistical) or deterministic (referred to as theoretical). 36

Okumura [45] developed an empirical model derived from eld measurements in Tokyo. This model is simple and applicable to urban, sub-urban and rural areas. Later, Hata [46] proposed an empirical model for propagation loss on the basis of Okumura curves. Okumura-Hata model does not consider terrain prole, reection and shadowing. COST231 [47] is another propagation model which extends OkumuraHata model to higher frequencies (1.5 GHZ to 2 GHZ). COST231 takes rooftop height, street width and orientations into account. Ray tracing model is a deterministic 2D or 3D physical model which considers geographical information of the planning area. Due to the complexity of the ray tracing model and the vast data requirement for accurate prediction of radio propagation, the model is used just for critical scenarios. More information on propagation models can be found in reference [22].

2.3.1.5

Air Interface Power Control

The coverage and capacity planning of UMTS network should be done mutually. The capacity of each cell is based on the actual interference level which depends on the emitted power [41]. In UMTS networks, the power of the Node B is shared among all the cell users and the allocated power to a given user depends on its distance from the Node B. The cell size is not xed and depends on the number of users, level of interference and their distance form the Node B. Air interface in UMTS systems is self-interference, meaning that cell interference level is increased as it is overloaded by users. With an increase in interference level, users located at the edge of the cell are detached from the parent Node B and this in turn, results in decrease of cell size. Such users will be covered by neighbor cells. On the other hand, when calldrops occur, interference decreases for the remaining users and cell is expanded. This 37

phenomenon is called cell breathing. Cell breathing is the result of constant changes in the coverage area with respect to amount of trac. It is important to keep the transmission power of Node Bs and users at the minimum levels to minimize interference and guarantee adequate quality at the receiver [48]. SIR in UMTS networks is highly aected by the trac distribution in the whole area and unlike 2G networks, SIR should be equal to a given threshold. In summary, the cell capacity and coverage depends on number of users and their distribution, as well as Power Control (PC) mechanisms. The PC mechanisms are based on either the received power or estimated SIR [41]. More information about PC can be found in references [8, 49, 50].

2.3.1.6

Cell Planning Algorithms

The cell planning sub-problem has been studied from dierent perspectives: link direction (uplink/downlink), optimal location and conguration of Node Bs, channel assignment, etc. The cell planning sub-problem is NP-hard [41] and most planning methods are based on approximate algorithms. Considering dierent constraints, several algorithms are used to solve the cell planning sub-problem. Zdunek and Ignor [51] use a newly introduced Invasive Weed Optimization (IWO) algorithm [52] to solve optimal locations of BSs, their pilot powers and channel assignments in UMTS mobile networks. They show that the IWO algorithm outperforms the algorithms like evolutionary strategies and genetic algorithms. A polynomial time approximation scheme, proposed by Galota et al. [53], deals with nding the location of BSs. The goal is to keep the network cost below a certain limit, provide coverage for a maximum number of users and keep the interference low. Their model does 38

not deal with SIR and PC mechanisms. Eisenblatter et al. [54] describe a model for the location and conguration of BSs in UMTS network. They focus on the modeling of the conguration problem using mixed-integer variables and linear constraints. Under the MOMENTUM project [14], their modeling tool produces exible network congurations. Chamaret et al. [55] propose a mathematical model for a given trac demand area and potential location of BSs to solve the radio network problem using graph theory. They try to achieve a large coverage area and minimize the network cost with minimum number of BSs. Amaldi et al., in a rst set of papers [41, 5658], investigate the location problem of BS in the uplink direction. Considering a set of candidate locations for BSs, trac distribution and propagation model, they propose two discrete mathematical programming model algorithms to nd the optimal location of BSs. The uplink optimization is applicable when we deal with symmetric services like voice. The most important constraints in uplink planning are emitted power of mobile users and sensitivity of BSs. They introduce two PC mechanisms. In the rst PC mechanism which is based on SIR level, power is adjusted in order to sustain SIR level in a target level. The second PC mechanism is based on controlling received power. This scheme adjusts the power such that, each channel receives the same target power value. The former mechanism is more complex, since the power emitted by each user depends on that of emitted by all other users. Since the problem in NP-hard, they proposed three greedy algorithms, followed by a tabu search heuristic to solve the problem and nd sub-optimal solutions. Amaldi et al. [59] in a collaboration with Signori [60], enhance their models for the uplink direction to nd the optimal location of BSs, as well as their conguration 39

(height, tilt and sector orientation). They consider SIR based PC as quality measure and a tabu search heuristic. A similar solution was proposed by Wu and Pierre [36] for the downlink direction. They introduced heuristic and stochastic optimization technique. Amaldi et al., in a second set of papers [61,62], investigate the problem in downlink direction. The major constraints in downlink direction are the maximum power of BSs and the sensitivity of the user devices. They presented an integer programming model which takes power-based and SIR-based PC mechanisms as quality measure. They proposed randomized greedy algorithms followed by a tabu search heuristic. Amaldi et al., in third set of papers [63, 64], consider both uplink and downlink (voice and data). They present a mathematical programming model (a combination of previous models) to optimize the location and conguration of BSs. The goal is to maximize the coverage and minimize the cost. They also propose a tabu search algorithm and a simpler PC mechanism to reduce the computational time. Jamma et al. [65] propose manual and automatic design strategies to optimize antenna parameters and common channel powers. Their design depends on the objectives dened by the operator. Jamma et al. [66] also deal with capacity and coverage optimization, using multi-objective functions. Finally, Jamma et al. [67] propose a steered optimization strategy in UMTS networks in order to answer the following two questions: i ) how local should an optimization problem be dened and ii ) which stations should be considered for possible modications to guarantee a robust and ecient optimization. They found out that too local problem reduces the eciency of planning tool and too large problem increases number of BSs and network cost. In a dierent work, Sohn and Jo [68] tackle the location problem and assign40

Figure 2.5: Access network planning sub-problem ment of mobile users to appropriate BSs in 3G W-CDMA uplink environment. The authors propose a constraint satisfaction model and apply dierent techniques like variable ordering and value ordering to nd good optimal solutions. Instead of cost minimization, the objective of their model is to minimize the total transmitted power. Once the location and conguration of BSs are known, the next step is to study BS assignment to the higher level. This involves investigation of access network subproblem.

2.3.2

The Access Network Planning Sub-Problem

The main elements of the access network are the Node Bs and the RNCs as shown in Figure 2.3. In order to plan a good access network, the following inputs are usually needed: 1. The physical location of Node Bs (either given or obtained from the cell planning sub-problem); 41

2. The trac demand passing through each Node B (either given or obtained from the cell planning sub-problem); 3. The set of potential locations to install RNCs; 4. The dierent types of RNCs; 5. The dierent types of links to connect Node Bs to RNCs; 6. The handover frequency between adjacent cells. Depending on the planners decision, the Node Bs might connect internally to each other based on some interconnection policies. This is also true for the RNCs. By so doing, the access network sub-problem is more extended and will include the trunks among Node Bs with themselves, as well as RNCs with themselves. In a tree interconnection, the Node Bs are either directly connected to RNCs or cascaded. Other types of topologies are star, ring and mesh. The interested reader on access network topologies can nd more information in reference [69]. Given the above inputs and the type of topology, the access network planning sub-problem aims to nd one or more of the following as output: 1. The optimal number of RNCs; 2. The best location to install RNCs; 3. The type of RNCs; 4. The link topology and type between Node Bs; 5. The link topology and type between RNCs; 42

6. The link topology and type between Node Bs and RNCs; 7. The trac (volume and type) passing through each RNC. The objective function is usually cost minimization, but other objectives such as reliability or combination of cost and reliability could be considered.

2.3.2.1

Cost-eective Access Networks

The cost of the access network includes the cost of RNCs, interconnection links and interfaces. Depending on the access network topology, the cost might vary. As a result, it is important to evaluate the cost subject to the topology. Harmatos et al. [70] propose an algorithmic network topology optimization method to simultaneously nd the optimum number of location of RNCs, as well as the transmission network between BSs and RNCs. In order to solve the NP-hard sub-problem, their method uses a combination of SA and greedy algorithm to minimize the cost. They also consider a degree constraint on the number of BSs that can be supported by one RNC. In a second paper, Harmatos et al. [71] found the bottleneck in their previous algorithm [70], which was the tree topology of the access network based on simple greedy algorithm. Because of the greedy principle, in many cases, the algorithm was not able to build the access tree correctly, causing a signicant rise in cost. They modied their algorithm to provide more cost-eective access network topology for one RNC. The objective is to nd the cost-optimal interconnection of BSs to their dedicated RNC, considering topological limitations, constraints and the originating trac of BSs. The authors state that, although their optimization model and process is working for UMTS network, it is also applicable to any multi-constrained capacitated tree 43

optimization problem with non-linear cost function. Lauther et al. [30] approach the access planning sub-problem as a clustering problem. They try to nd the optimal number and size of clusters for a set of BSs to minimize the cost. Given the location of BSs, they present two clustering procedures based on proximity graph. The rst method is based on tree generation and cutting. The idea is to build a tree in the rst step. In the second step, the tree is cut into sub-trees (clusters). The rst step is based on an algorithm like Prim [72] or Kruskal [73], while the second step is based on the generation of sub-trees starting form the leaves. Initially, each Node B forms its own cluster. Then, two clusters are merged per iteration if the cost of the access network is reduced. Another clustering approach is also presented in a paper by Godor and Magyar [74]. They aggregate the user trac in multi-level tree-like fashion using some intermediate concentrator nodes. Considering several constraints, the NP-hard problem is solved by heuristic algorithms to minimize the cost. Krendzel et al. [75] consider the problem of physical links ring conguration between BSs in 4G network. Considering planning constraints and using dynamic programming, they try to minimize the cost of the ring conguration. In another paper, Juttner et al. [76] propose two network design methods to nd the cost-optimal number and location of RNCs and their connection to BSs in tree topology, while respecting a number of constraints. First, a global algorithm combines a metaheuristic technique with the solution of a specic b-matching problem. Then, the tree structure made by the rst method is improved by the second method, which uses a combination of Lagrangian lower bound with branch-and-bound. They demonstrate the eectiveness of their algorithms in reducing the cost by a number of test cases. 44

Constraint-based optimization of the access network sub-problem was considered by Wu and Pierre. In [77], they propose a model to optimally nd the number and location of RNCs and solve the assignment of Node Bs to selected RNCs. Constraints like number of Node Bs supported by one RNC, number of interfaces on the RNC, the amount of trac supported by one RNC, as well as handover volume between adjacent cells are taken into consideration. Greedy heuristic algorithms, TS and SA, are explored in the proposed model to minimize the cost. Wu and Pierre, in [78], used a three-staged hybrid constraint-based approach. In the rst step, good feasible solutions are found and then improved by local search in the second step. Such solutions are considered as the upper bound. In the last step, the solution is rened by constraint optimization technique. They state that the obtained solutions can be used as initial solutions for heuristics. Minimizing handover cost has been investigated in a series of cell-to-switch papers [7982]. The idea is to reduce the number of handovers between two adjacent cells by linking both cells to the same RNC. Bu et al. [83] investigate the access planning problem from a dierent perspective. Usually, Point to Point (P2P) transmission links (E1 and/or T1) used in 3G access network are not optimal in case of asymmetric and bursty trac. The authors propose to use a 802.16 (WiMAX) based radio access networks to transmit data from Node Bs to RNCs. They design the access network with minimum number of 802.16 links upon position of BSs and RNCs. Charnsripinyo [84] considers the design problem of 3G access network while maintaining an acceptable level of quality of service. The problem was formulated as a Mixed Integer Programming (MIP) model to minimize the cost. 45

2.3.2.2

Reliable Access Networks

Network reliability (also known also as survivability) describes the ability of the network to function and not to disturb the services during and after a failure. The need for seamless connectivity has been a motivation for many researchers to explore new techniques for network reliability. Tripper et al. [85] introduce a framework to study wireless access network survivability, restoration techniques and metrics for quantifying network survivability. Cellular networks are very vulnerable to failure. Failure can happen either on node level (BSs, RNC, MSS, etc.) or link level. Simulation results on dierent types of failure scenarios in a GSM network shows that after a failure, mobility of users worsens network performance. For example, in the case of a BS failure, users will try to connect to the adjacent BS and that degrades the overall network performance. Charnsripinyo and Tipper [86] proposed an optimization based model for the design of survivable 3G wireless access backhaul networks in a mesh topology. Using a two-phase algorithm, the authors rst design a network with a minimum cost, considering Quality of Service (QoS) and then update the topology to satisfy survivability constraints. They also propose a heuristic, based on the iterative minimum cost routing to scale the design with real world networks. Increasing reliability level imposes more cost to the network. There is a balance (best trade o) between cost and reliability and in fact, higher level of reliability will obtrudes higher cost to the network. Aiming to create a balance between reliability and cost, Szlovencsak et al. [87] introduce two algorithms. The rst algorithm modies the cost-minimum tree as produced in [70, 71], while respecting reliability constraints

46

and retains the tree structure. In the second algorithm, dierent links are added to the most vulnerable parts of the topology to have a more reliable network. Krendzel et al. [88] study cost and reliability of 4G RAN in a ring topology. They estimate cost and reliability in dierent congurations and state that considering cost and reliability, the most preferable topology for 4G RAN is a multi-ring. Once the access planning sub-problem is solved and the number, type, location and trac of each RNC in known, the next step is to deal with the core planning sub-problem.

2.3.3

The Core Network Planning Sub-Problem

The core network is the central part of UMTS network. The core network is responsible for trac switching, providing QoS, mobility management, network security and billing [89]. As described in Chapter 1, the core network consists of CS and PS domains. The key elements of CS domain are MGW and MSS, responsible for switching and controlling functions respectively. PS domains key elements are SGSN and GGSN which are responsible for packet switching. Figure 2.6 shows a typical core network planning sub-problem. The core planning sub-problem supposes that the following inputs are known: 1. The physical location of RNCs (either given or obtained from the access planning sub-problem); 2. The trac demand (volume and type) passing through each RNC (either given or obtained from the access planning sub-problem); 3. The potential location of core NEs; 47

Figure 2.6: Core network planning sub-problem 4. The dierent types of core NEs; 5. The dierent types of links to connect RNCs to core NEs. Depending on the network planner, the topology of the backbone network could be a ring, a full mesh, a mesh or a layered structure format. In the ring topology, each NE is directly attached to the backhaul ring. Full mesh topology provides pointto-point communication such that each NE is able to communicate to any other NE directly. The mesh topology is a limited version of the full mesh, whereas due to some restrictions, not every NE can communicate directly to another NE. For fast growing networks, maintaining a mesh or full mesh topologies becomes an exhaustive task. To solve this sub-problem, the layered structure was introduced. A layered structure 48

does not provide direct link between all NEs. A tandem layer, as the nucleus of the layered structure is dened. The tandem layer is composed of a series of tandem (transit) nodes, usually connected in full mesh. Then, all NEs in the core network are connected to at least one of the tandem nodes. Ouyang and Fallah [90] state that a layered structure has many advantages compared to full mesh topology. Given that the above inputs are available and the type of topology is decided, the core network planning sub-problem aims to nd one or more of the following as output: 1. The optimal number of core NEs; 2. The best location to install core NEs; 3. The type of core NEs; 4. The link topology and type between RNCs and core NEs; 5. The link topology and type between core NEs; 6. The trac (volume and type) passing through core NEs. The objective function is usually cost minimization, but other objectives like reliability could be considered. Not many researches have been concentrated on the core network planning sub-problem. The reason could be the similarity of this sub-problem to the wired network planning problem. Shalak et al. [89] present a model for UMTS network architecture and discuss the required changes for upgrading core network from GSM to UMTS. They outline network planning steps and compare the products of dierent vendors in packet switch network. 49

Ricciato et al. [91] deal with the assignment of RNCs to SGSNs based on measured data. The optimization goals are to balance the number of RNC among the available SGSNs and minimize the inter-SGSN routing area updates. Required inputs are taken from live network and the objective function is solved by linear integer programming methods. While they focus on GPRS, they state that their approach can be applied to UMTS networks. Harmatos et al. [92] deal with the interconnection of RNCs, placement of MGWs and planning core network. They split the problem in two parts. The rst problem is interconnection of the RNCs which belong to the same UTRAN and the placement and selection of a MGW to connect to core network. The second problem is interconnection of MGWs together in backbone through IP or ATM network. The objective is to design a fault-tolerant network with cost-optimal routing.

2.3.4

Remarks on Sequential approach

The sequential approach used to solve the design problem of UMTS networks has many advantages, but some disadvantages. The sequential approach reduces the complexity of the problem by splitting the problem into three smaller sub-problems. By so doing, it is possible to include more details in each sub-problem for better planning. On the contrary, solving each sub-problem independently from the other sub-problems may result in local optimization, because interactions between subproblems are not taken into account. Combining the result of sub-problems does not guarantee a nal optimal solution. There is no integration technique developed yet to incorporate all partial solutions in order to obtain a global solution. Therefore, a global view from the network is required to dene a global problem. 50

Input

Cell, Access and Core Network Planning Sub-Problems

Final Solution

Figure 2.7: Global approach for the UMTS planning problem [4]

2.4

Global Approach

As mentioned earlier, the sequential approach breaks down the UMTS planning problem in three sub-problems and solves them solely. As shown in Figure 2.7, a global (also called integrated) approach considers more than one sub-problem at a time and solves them jointly. Since all interactions between the sub-problems are taken into account, a global approach has the advantage of providing a solution close to the global optimal, but at the expense of increasing problem complexity. The global problem of UMTS networks which is composed of three NP-hard sub-problems is also an NP-hard problem [4]. The objective of the global approach is similar to the objective of the sequential approach. Network cost minimization is the main concern, while considering network performance. Researches on the global approach are mainly divided into three directions: i ) cell and access networks, ii ) access and core networks and iii ) the whole network (i.e. cell, access and core). Zhang et al. [93] proposed a global approach to solve the UTRAN planning problem. Their model nds the number and location of Node Bs and RNCs, as well as their interconnections in order to minimize the cost. Chamberland and Pierre [94] consider access and core network planning sub-problems. Given the BSs locations, their model nds the location and types of BSCs and MSCs, types of links and topol-

51

ogy of the network. Since such sub-problem is NP-hard, the authors propose a TS algorithm and compare the results with a proposed lower bound. While the model is targeted to GSM networks, it can be also applied to UMTS networks with minor modications. In another paper, Chamberland [95] investigates the update problem in UMTS network. Considering an update in BSs subsystem, the expansion model accommodates the new BSs into the network. The model determines the optimal access and core networks and considers network performance issues like call and handover blocking. The author proposes a mathematical formulation of the problem, as well as a heuristic based on the TS principle. Recently St-Hilaire et al. [96] proposed a global approach in which the three subproblems are considered simultaneously. The authors developed a mathematical programming model to plan UMTS networks in the uplink direction. Through a detailed example, they compared their integrated approach with the sequential approach. They proposed two heuristics based on local search and tabu search to solve the NPhard problem [97]. Furthermore, St-Hilaire et al. [98] proposed a global model for the expansion problem of UMTS networks as an extension to their previous works. They state that this model can also be used for green eld networks. They also present numerical results based on branch and bound implementation.

2.5

Section Remarks

The purpose of solving the design problem of UMTS networks is to nd an optimum topology for the network which satises all desired constraints like cost, reliability, performance and so on. Such an optimum topology is favorable for operators, as it can 52

save money and attract more subscribers. The planning problem of UMTS networks is complex and composed of three sub-problems: the cell planning sub-problem, the access network sub-problem and the core network sub-problem. There are two main approaches to solve planning problem of UMTS networks: the sequential and the global. In the sequential approach, the three sub-problems are tackled sequentially. Since each sub-problem is less complex than the initial problem, more details can be considered in each sub-problem. As a result, solving sub-problems is easier than solving the whole planning problem. However, since each sub-problem is solved independently from other sub-problems, the combination of the optimal solution of each sub-problem (if obtained), might not result in an optimal solution for the whole network planning problem. A global approach deals with more than one sub-problem simultaneously and considers all interactions between the sub-problems. The global problem has the advantage of nding good solutions which are closer to the global minimum. The global problem is NP-hard and is more complex compared to three sub-problems. To nd approximate solutions for global planning of UMTS networks in a polynomial time, heuristics need to be dened. It has been proven by scholars that dierent adaptations of heuristics are eectively able to solve the planning problem of cellular networks. The focus of literature in the area of UMTS network planning has been over Release 99 architecture. However, the planning problem of Release 4 networks has not been considered so far. Such architecture is the base model for higher 3GPP network releases and through accommodating new multimedia servers in the core network; the thesis model can be easily adopted to support new releases. 53

Moreover, the all-IP architecture has been also mixed with Release 4 networks. Furthermore a realistic trac prole has been considered. All together, it is expected that the planning algorithm proposed in this thesis would be useful for operators to plan real networks.

54

Chapter 3 Network Model Formulation and Planning Tools

In this chapter, we rst present the step by step mathematical formulation of 3G UMTS all-IP Release 4 networks. Then, two planning tools are presented in detail and the design steps are explained. The rst planning tool is based on exact algorithm on the second planning tool is based on local search and tabu search. Before presenting the mathematical formulation, we need to describe some basic concepts in optimization for better understanding. Such denitions will be used throughout this chapter.

3.1

Basic Concepts in Optimization

Any optimization problem P , is described as triple (S, , F ) where: The objective function F is a mathematical function of the decision variables 55

to express the objective of a problem. The search space S is the domain in which every element could be a candidate solution for the problem. The search space is dened over a nite set of decision variables. The constraints are the set of conditions applied over the decision variables, which in fact are applied on the objective function [19]. Constraints reduce the search space by discarding infeasible solutions. Structural constraints deal with equality or inequality (, =, ) , while nonnegativity constraints represent the positive domain of the decision variables. It is also necessary to explain the following denitions [19]: The decision variables are the unknown quantities, dened by decision-maker to construct the objective function. If decision variables have discrete domains, the problem is called discrete optimization or combinatorial optimization problem and in case of continuous domain, the problem is known as continues optimization. The linear (objective) function is a mathematical function in which the decision variables appear only in rst degree, meaning that variables are not multiplied to each other. Variables are indeed multiplied by constant and are combined only by addition and subtraction. The feasible area is an area within the search space of the problem in which any solution can satisfy all of the dened constraints, bounds and integrality restrictions. 56

 The feasible solution implies to any solution belonging to the feasible area. The neighborhood structure is a function that assigns a set of neighbors N (s) S to every s S . A neighbor is achieved by transformation of a solution. The transformation operator is called move. The local optimum is a solution s S , which for s N (s ) : f (s ) f (s) for minimization problem or f (s ) f (s) for maximization problem. The global optimum is a solution s S , which for s S : f (s ) f (s) for minimization problem or f (s ) f (s) for maximization problem. The sub-optimal (near-optimal) is a solution s S , which f (s ) f ( s) f (s ) for minimization problem or f (s ) f ( s) f (s ) for maximization problem. Sub-optimal solution is also known as good solution in literature. The NP-hard (Non-deterministic Polynomial-time hard) problem is a problem which has not exact solution in polynomial time. Global minimum and local minimum are shown in Figure 3.1. The local minimum is the value of the function at a certain point in its domain, where the value is less than or equal to the values at all other points in the immediate vicinity, while the global minimum is the lowest point over the entire domain of the function.

57

Objective Function

Local Minimum

Global Minimum

Solution Space

Figure 3.1: Global minimum vs. local minimum

3.2

Global Planning Problem: Inputs, Outputs and Objectives

In order to formulate the model and dene the planning problem of 3G UMTS all-IP Release 4 networks, we suppose the following information is known: The location of Node Bs; The potential locations of all NEs (i.e. RNCs, MGWs, MSSs, SGSNs, GGSNs, edge routers/CS routers/PS routers); The dierent types of NEs and links, their capacities and their costs; The installation cost for given types of facilities (including for instance: oor space, cables, racks, electric installations, labor, etc.); The subscriber trac prole; 58

 The planning parameters; The number of subscribers for each Node B. Our focus is on the development of automatic planning tools for the design of 3G UMTS all-IP Release networks. More precisely, the following problems are provided in the output: Selecting the number, the location and the type of NEs (i.e. RNCs, MGWs, MSSs, SGSNs, GGSNs, edge routers/CS routers/PS routers); Selecting the number and the type of links and interfaces; Designing the network topology. The objective of the planning problem in this thesis is to design a network with minimum cost (including cost of nodes, links, interfaces, installation, etc.) while having an acceptable level of quality. Such objective is favorite of operators along with maximizing prot. Prot maximization involves dierent streams, policies, management and marketing techniques which are beyond the scope of this thesis.

3.3

Model Formulation

The model formulation is the process of translating the characteristics of the nodes, links and interfaces; including their capacity, type and cost into a mathematical format. In this thesis, we consider each Node B as a trac node or test point. The all-IP Release 4 networks involve dierent NEs for voice and data services in which

59

every NE deals with specic aspects of the trac. Therefore, to consider all aspects of the trac, we need to produce a trac record for each Node B.

3.3.1

Node B Trac Record

Node B trac record is a set of dierent aspects for the voice and data tracs (i.e. BHCA, simultaneous calls, bandwidth, signaling trac, etc.). Considering Node Bs subscribers, such record is calculated using subscriber trac prole and planning parameters. (An example of subscriber trac prole and planning parameters is presented in Chapter 4). This information is given by the operator and might vary subject to trac forecast and planning strategies. The trac records for ve Node Bs are shown in Table 3.1. The Node B trac record includes the number of covered subscribers, and the following information for those subscribers: BHCA; Channels for CS trac; Bandwidth required for the CS trac; Bandwidth required for the PS trac; Attached subscribers in Busy Hour (BH); Simultaneous Packet Data Protocol (PDP) contexts; Signaling messages bandwidth for the CS trac.

60

Table 3.1: Typical example of trac record

Node B index 1 2 3 4 5 Subscriber (c0 ) 11, 180 8, 089 13, 950 16, 123 5, 266 BHCA (c1 ) 7, 090 5, 130 8, 847 10, 224 3, 340 CS Channels (c2 ) 257 191 316 363 129 CS BW (Mbps) (c3 ) 6 5 8 9 3 PS BW (Mbps) (c4 ) 2 1 2 2 1 Attached Subscribers (c5 ) 2, 236 1, 617 2, 790 3, 224 1, 053 Simultaneous PDP context (c6 ) 280 203 349 404 132 Signaling (Mbps) (c7 ) 0.56 0.40 0.70 0.81 0.26

The rst column of the table refers to the Node B index. The number of subscribers covered by the Node B is randomly generated and tabulated in column two. BHCA, in column three, shows the number of calls made by Node B subscribes at BH. The number of CS trac channels and the associated CS trac bandwidth are shown in columns four and ve respectively. PS trac bandwidth in column six, represents the bandwidth required for data services of the Node B subscribers. We suppose that a specic ratio of subscribers use data services (attached subscribers) in BH, and furthermore, a portion of these subscribers have PDP context. The PDP context is a set of data structure dened on SGSN and GGSN, which contains the subscriber session information when the subscriber has an active session. Columns seven and eight show BH attached subscribers and simultaneous PDP context. If a subscriber decides to use packet data services, it must rst attach and then activate a PDP context. Finally, the last column gives the signaling load for the CS trac. When the trac records are calculated, we need to design a network which can fulll those trac aspects for the subscribers. In order to accomplish this goal, in the next section, we rst dene a set of mathematical notions, a set of decision variables and parameters to deal with the characteristics of the network model.

61

Figure 3.2: Graphical representation of the model formulation

3.3.2

Notation

The following notation, used throughout this thesis, is composed of sets, decision variables, trac variables and cost parameters. Figure 3.2 represents, in a graphical way, a summarized version of the notation. 3.3.2.1 Sets

B , the set of Node Bs (already installed); c b , the trac entry of type c C for a Node B, (b B ); C , the set of trac characteristics such that C = {c0 , c1 , ..., c7 } where: 62

 c0 : Number of subscribers; c1 : BHCA; c2 : CS trac channel; c3 : CS trac bandwidth (Mbps); c4 : PS trac bandwidth (Mbps); c5 : BH attached subscribers; c6 : Simultaneous PDP context; c7 : Signaling message bandwidth (Mbps) for the CS trac. M0 is the set of links and interface types that can be used to connect the Node B to edge router and edge router to CS router/PS router; M1 is the set of links and interface types that can be used to connect the edge router to RNC, CS router to MGW/MSS and PS router to SGSN/GGSN; m : Capacity (in Mbps) of the link (interface) of type M0 or M1 ; S1 , S2 and S3 are respectively the sets of potential sites to install the RNCs/edge routers, the MGWs/MSSs/CS routers and the SGSNs/GGSNs/PS routers; T0 , T1 , T2 , T3 , T4 , T5 and T6 are respectively the sets of the Node B types, RNC types, MGW types, MSS types, SGSN types, GGSN types and edge router/CS router/PS Router types where: SU B t : Subscriber capacity for a node of type t; BHCAt : BHCA capacity for a node of type t; 63

 t : CS trac capacity (in channel) for a node of type t; t : Switch fabric capacity (in Mbps) for a node of type t; P DP t : Simultaneous PDP context capacity for a node of type t;
t t t t t t t t nt 0 , n1 , n2 , n3 , n4 , n5 , n6 , n7 and n8 are respectively the maximum number

of interfaces that can be installed in the type t of the Node B, the edge router, the RNC, the CS router, the PS router, the MGW, the SGSN, the MSS and the GGSN.

3.3.2.2

Decision Variables

bi1 bi1 v01 , a 01 variable such that v01 = 1 if and only if the Node B (b B ) is

connected to the edge router installed at site i1 S1 ;

bi1 m v01 , the number of links of type m M0 connecting the Node B (b B ) to

the edge router installed at site i1 S1 ;

i1 i2 i1 i2 v11 , a 01 variable such that v11 = 1 if and only if the edge router installed

at site i1 S1 is connected to the RNC installed at site i2 S1 ;

i1 i2 m v11 , the number of links of type m M1 connecting the edge router installed

at site i1 S1 to the RNC installed at site i2 S1 ;

i1 j1 i1 j1 v12 , a 01 variable such that v12 = 1 if and only if the edge router installed

at site i1 S1 is connected to the CS router installed at site j1 S2 ;

i1 j1 m v12 , the number of links of type m M0 connecting the edge router installed

at site i1 S1 , to the CS router installed at site j1 S2 ; 64

i 1 k1 i 1 k1 v13 , a 01 variable such that v13 = 1 if and only if the edge router installed

at site i1 S1 is connected to the PS router installed at site k1 S3 ;

i 1 k1 m v13 , the number of links of type m M0 connecting the edge router installed

at site i1 S1 to the PS router installed at site k1 S3 ;

j1 j2 j1 j2 v21 , a 01 variable such that v21 = 1 if and only if the CS router installed at

site j1 S2 is connected to the MGW installed at site j2 S2 ;

j1 j2 m v21 , the number of links of type m M1 connecting the CS router installed

at site j1 S2 to the MGW installed at site j2 S2 ;

j1 j2 j1 j2 v22 , a 01 variable such that v22 = 1 if and only if the CS router installed at

site j1 S2 is connected to the MSS installed at site j2 S2 ;

j1 j2 m v22 , the number of links of type m M1 connecting the CS router installed

at site j1 S2 to the MSS installed at site j2 S2 ;

k1 k2 k1 k2 v31 , a 01 variable such that v31 = 1 if and only if the PS router installed at

site k1 S3 is connected to the SGSN installed at site k2 S3 ;

k1 k2 m v31 , the number of links of type m M1 connecting the PS router installed

at site k1 S3 to the SGSN installed at site k2 S3 ;

k1 k2 k1 k2 v32 , a 01 variable such that v31 = 1 if and only if the PS router installed at

site k1 S3 is connected to the GGSN installed at site k2 S3 ;

k1 k2 mm v32 , the number of links of type m M1 connecting the PS router installed

at site k1 S3 to the GGSN installed at site k2 S3 ; 65

i2 t 2t xi 1 , a 01 variable such that x1 = 1 if and only if an RNC of type t T1 is

installed at site i2 S1 ;
j2 t 2t xj 2 , a 01 variable such that x2 = 1 if and only if a MGW of type t T2 is

installed at site j1 S2 ;
j2 t 2t xj 3 , a 01 variable such that x3 = 1 if and only if an MSS of type t T3 is

installed at site j2 S2 ;
k2 t 2t xk 4 , a 01 variable such that x4 = 1 if and only if an SGSN of type t T4 is

installed at site k1 S3 ;
k2 t 2t xk 5 , a 01 variable such that x5 = 1 if and only if a GGSN of type t T5 is

installed at site k2 S3 ;
i1 t 1t xi 6 , a 01 variable such that x6 = 1 if and only if an edge router of type t T6

is installed at site i1 S1 ;
j1 t 1t xj 7 , a 01 variable such that x7 = 1 if and only if an CS router of type t T6

is installed at site j1 S2 ;
k1 t 1t xk 8 , a 01 variable such that x8 = 1 if and only if an PS router of type t T6

is installed at site k1 S3 . 3.3.2.3 Trac variables

cbi1 f01 , the trac entry c C on the link from the Node B installed at site b B

to an edge router installed at site i1 S1 ;

66

ci1 i2 f11 , the trac entry c C on the link from the edge router installed at site

i1 S1 to an RNC installed at site i2 S1 ;

ci1 j1 f12 , the trac entry c C on the link from the edge router installed at site

i1 S1 to a CS router installed at site j1 S2 ;

ci1 k1 f13 ,, the trac entry c C on the link from the edge router installed at site

i1 S1 to a PS router installed at site k1 S3 ;

cj1 j2 f21 , the trac entry c C on the link from the CS router installed at site

j1 S2 to a MGW installed at site j2 S2 ;

cj1 j2 f22 , the trac entry c C on the link from the CS router installed at site

j1 S2 to an MSS installed at site j2 S2 ;

ck1 k2 f31 , the trac entry c C on the link from the PS router installed at site

k1 S3 to an SGSN installed at site k2 S3 ;

ck1 k2 f32 , the trac entry c C on the link from the PS router installed at site

k1 S3 to a GGSN installed at site k2 S3 . 3.3.2.4 Cost Parameters

1m abi 01 , the link and interface costs (including installation cost) for connecting a

Node B installed at site b B to an edge router installed at site i1 S1 through a link and interface of type m M0 ;
1 i2 m ai , the link and interface costs (including installation cost) for connecting an 11

edge router installed at site i1 S1 to an RNC installed at site i2 S1 through 67

a link and interface of type m M1 ;

1 j1 m ai , the link and interface costs (including installation cost) for connecting 12

an edge router installed at site i1 S1 to a CS router installed at site j1 S2 through a link and interface of type m M0 ;
1 k1 m ai , the link and interface costs (including installation cost) for connecting 13

an edge router installed at site i1 S1 to a PS router installed at site k1 S3 through a link and interface of type m M0 ;
1 j2 m , the link and interface costs (including installation cost) for connecting a aj 21

CS router installed at site j1 S2 to a MGW installed at site j2 S2 through a link and interface of type m M1 ;
1 j2 m aj , the link and interface costs (including installation cost) for connecting a 22

CS router installed at site j1 S2 to an MSS installed at site j2 S2 through a link and interface of type m M1 ;
1 k2 m ak , the link and interface costs (including installation cost) for connecting a 31

PS router installed at site k1 S3 to an SGSN installed at site k2 S3 through a link and interface of type m M1 ;
1 k2 m , the link and interface costs (including installation cost) for connecting a ak 32

PS router installed at site k1 S3 to a GGSN installed at site k2 S3 through a link and interface of type m M1 ;
t t t t t t t bt 1 , b2 , b3 , b4 , b5 , b6 , b7 and b8 are respectively the cost (including installation

cost) of an RNC of type t T1 , of a MGW of type t T2 , of an MSS of type 68

t T3 , of an SGSN of type t T4 , of a GGSN of type t T5 , of an edge router of type t T6 , of a CS router of type t T6 and of a PS router of type t T6 .

3.4

Cost Function

The cost function, representing the total cost of the network, is composed of two terms: the cost of the links and interfaces and the cost of the nodes. The cost of the links and interfaces, noted by CL , is given by the following equation: CL ( v ) = + + +
bB i1 S1 mM0
1 m bi1 m + abi 01 v01

i1 S1 i2 S1 mM1

i1 i2 m i1 i2 m v11 a11

1 j1 m i1 j1 m ai v12 + 12

1 k1 m i 1 k1 m v13 ai 13

i1 S1 j1 S2 mM0 j1 j2 m j1 j2 m a21 v21

i1 S1 k1 S3 mM0

(3.1)
1 j2 m j1 j2 m aj v22 22

j1 S2 j2 S2 mM1

j1 S2 j2 S2 mM1
1 k2 m k1 k2 m + v31 ak 31

1 k2 m k1 k2 m v32 ak 32

k1 S3 k2 S3 mM1

k1 S3 k2 S3 mM1

The cost of the nodes, noted by CN , is given by the following equation: CN (x) = +
tT5 tT1

bt 1 bt 5

i2 S1

2t xi 1 +

tT2

bt 2 bt 6

j2 S2

2t xj 2 +

tT3

bt 3 bt 7

j2 S2

2t xj 3 +

tT4

bt 4 bt 8

k2 S3

2t xk 4

2t xk 5

1t xi 6

1t xj 7

(3.2)
1t xk 8

k2 S3

tT6

i1 S1

tT6

j1 S2

tT6

k1 S3

3.5

The Model

The model for the 3G UMTS all-IP Release 4 networks planning problem, denoted as ALLIPR4PP, can now be given. 69

ALLIPR4PP: min (CL (v) + CN (x)) subject to the following constraints: 1. Uniqueness constraints impose that at most one type of a NE is installed at the site. RNC, MGW, MSS, SGSN, GGSN, edge router, CS router and PS router type uniqueness constraints are as follows: RNC-Type uniqueness constraints
tT1

(3.3)

2t xi 1 1 ( i 2 S1 )

(3.4)

MGW-Type uniqueness constraints

tT2

2t xj 2 1 (j2 S2 )

(3.5)

MSS-Type uniqueness constraints

tT3

2t xj 3 1 (j2 S2 )

(3.6)

SGSN-Type uniqueness constraints

tT4

2t xk 4 1 (k 2 S3 )

(3.7)

70

GGSN-Type uniqueness constraints

tT5

2t xk 5 (k2 S3 )

(3.8)

Edge router-Type uniqueness constraints

tT6

1t xi 6 1 ( i 1 S1 )

(3.9)

CS router-Type uniqueness constraints

tT6

1t xj 7 1 (j1 S2 )

(3.10)

PS router-Type uniqueness constraints

tT6

1t xk 8 1 (k 1 S3 )

(3.11)

2. Assignment constraints impose that a lower layer node is assigned to only one upper layer node. Node B, edge router, CS router and PS router assignment constraints are as follows: Node B assignment constraints
i1 S1

bi1 v01 = 1 (b B )

(3.12)

71

Edge router assignment constraints

i2 S1

i1 i2 v11 =

tT6

i1 t x6 (i1 S1 )

(3.13)

j1 S2

i1 j1 v12 =

tT6

i1 t x6 (i1 S1 )

(3.14)

k1 S3

i1 k1 v13 =

tT6

1t xi 6 (i1 S1 )

(3.15)

CS router assignment constraints

j2 S2

j1 j2 v21 =

tT6

1t xj 7 (j1 S2 )

(3.16)

j2 S2

j1 j2 v22 =

tT6

1t xj 7 (j1 S2 )

(3.17)

PS router assignment constraints

k2 S3

k1 k2 v31 =

tT6

1t xk 8 (k1 S3 )

(3.18)

k2 S3

k1 k2 v32 =

tT6

1t xk 8 (k1 S3 )

(3.19)

3. Capacity constraints at the interface level impose that the number of links from lower layer nodes to upper layer node should not exceed the maximum number of interfaces on the upper layer node. Obviously, the link is only established if the upper layer node is installed. Capacity constraints of edge router, RNC, CS router, MGW, MSS, PS router, SGSN and GGSN at the interface level are as follows: 72

Edge router capacity constraints at the interface level

mM0 bB bi1 m v01 +

mM1 i2 S1

i1 i2 m v11 +

mM0 j1 S2 i1 t nt 1 x6

i1 j1 m v12 +

i1 k1 m v13

tT6

(3.20)

(i1 S1 )

mM0 k1 S3

RNC capacity constraints at the interface level

mM1 i1 S1

i1 i2 m v11

tT1

i1 t nt 2 x1 (i2 S1 )

(3.21)

CS Router capacity constraints at the interface level

mM0 i1 S1

i1 j1 m v12 +

mM1 j2 S2

j1 j2 m v21 +

mM1 j2 S2

j1 j2 m v22

tT6

j1 t nt 3 x7 (j1 S2 )

(3.22) MGW capacity constraints at the interface level

mM1 j1 S2

j1 j2 m v21

tT2

j2 t nt 5 x2 (j2 S2 )

(3.23)

MSS capacity constraints at the interface level

mM1 j1 S2

j1 j2 m v22

tT3

j2 t nt 7 x3 (j2 S2 )

(3.24)

PS Router capacity constraints at the interface level

mM0 i1 S1

i 1 k1 m v13 +

mM1 k2 S3

k1 k2 m v31 +

mM1 k2 S3

k1 k2 m v32

tT6

k1 t nt 4 x8 (k1 S3 )

(3.25) 73

SGSN capacity constraints at the interface level

mM1 k1 S3

k1 k2 m v31

tT4

k2 t nt 6 x4 (k2 S3 )

(3.26)

GGSN capacity constraints at the interface level

mM1 k1 S3

k1 k2 m v32

tT4

k2 t nt 8 x5 (k2 S3 )

(3.27)

4. Capacity constraints at the node level impose that the aggregated trac entry of type c C on a node should not exceed its capacity statement. Capacity constraints of edge router, RNC, CS router, MGW, MSS, PS router, SGSN and GGSN at the node level are as follows: Edge router capacity constraints at the node level
mM0

bB

bi1 m v01

tT6

1t t xi 6 (i1 S1 )

(3.28)

RNC capacity constraints at the node level

i1 S1

ci1 i2 f11

tT1

2t SU B t xi 1 (c {c0 }, i2 S1 )

(3.29)

i1 S1

ci1 i2 f11

tT1

i2 t BHCAt x1 (c {c1 }, i2 S1 )

(3.30)

mM1

i1 S1

i1 i2 m v11

tT1

i2 t (i2 S1 ) t x1

(3.31)

74

CS router capacity constraints at the node level

mM0

i1 S1

i1 j1 m v12

tT6

1t t xi 7 (j1 S2 )

(3.32)

MGW capacity constraints at the node level

j1 S2

cj1 j2 f21

tT2

2t BHCAt xj 2 (c {c1 }, j2 S2 )

(3.33)

j1 S2

cj1 j2 f21

tT2

2t t xj 2 (c {c2 }, j2 S2 )

(3.34)

mM1

j1 S2

j1 j2 m v21

tT2

2t t xj 2 (j2 S2 )

(3.35)

MSS capacity constraints at the node level

j1 S2

cj1 j2 f22

tT3

2t SU B t xj 3 (c {c0 }, j2 S2 )

(3.36)

j1 S2

cj1 j2 f22

tT3

2t BHCAt xj 3 (c {c1 }, j2 S2 )

(3.37)

j1 S2

cj1 j2 f22

tT3

2t t xj 3 (c {c2 }, j2 S2 )

(3.38)

mM1

j1 S2

j1 j2 m v22

tT3

2t t xj 3 (j2 S2 )

(3.39)

PS router capacity constraints at the node level

mM0

i1 S1

i1 k1 m v13

tT6

1t t xk 8 (k1 S3 )

(3.40)

75

SGSN capacity constraints at the node level

k1 S3

ck1 k2 f31

tT4

2t SU B t xk 4 (c {c5 }, k2 S3 )

(3.41)

k1 S3

ck1 k2 f31

tT4

2t P DP t xk 4 (c {c6 }, k2 S3 )

(3.42)

mM1

k1 S3

k1 k2 m v31

tT4

2t t xk 4 (k2 S3 )

(3.43)

GGSN capacity constraints at the node level

k1 S3

ck1 k2 f32

tT5

2t P DP t xk 5 (c {c6 }, k2 S3 )

(3.44)

mM1

k1 S3

k1 k2 m v32

tT5

2t t xk 5 (k2 S3 )

(3.45)

5. Capacity constraints at the link level impose that the trac entry of type c C of a node should not exceed its total links capacities to upper layer node. Capacity constraints of interconnection links for the thesis model are as follows: Node B-edge router link capacity constraints
cbi1 c bi1 c bi1 (f01 + f01 + f01 )

mM0

bi1 m m v01

(3.46)

(c {c3 }, c {c4 }, c {c7 }, i1 S1 , b B )

76

Edge router-RNC link capacity constraints

ci1 i2 c i1 i2 c i1 i2 (f11 + f11 + f11 )

mM1

i1 i2 m m v11

(3.47)

(c {c3 }, c {c4 }, c {c7 }, i1 , i2 S1 ) Edge router-CS router link capacity constraints

ci1 j1 c i1 j1 (f12 + f12 )

mM0

i1 j1 m m v12

(3.48)

( c {c 3 } , c { c 7 } , i 1 S1 , j 1 S2 ) CS router-MGW link capacity constraints

mM1

cj1 j2 f21

j1 j2 m m v21 (c {c3 }, j1 , j2 S2 )

(3.49)

CS router-MSS link capacity constraints

mM1

cj1 j2 f22

j1 j2 m m v22 (c {c7 }, j1 , j2 S2 )

(3.50)

Edge router-PS router link capacity constraints

mM0

ci1 k1 f12

i1 k1 m m v12 (c {c4 }, i1 S1 , k1 S3 )

(3.51)

PS router-SGSN link capacity constraints

mM1

ck1 k2 f31

k1 k2 m m v31 (c {c4 }, k1 , k2 S3 )

(3.52)

77

PS router-GGSN link capacity constraints

mM1

ck1 k2 f32

k1 k2 m m v32 (c {c4 }, k1 , k2 S3 )

(3.53)

6. Trac ow conservation constraints state that sum of the input trac type c C to a node should be equal to the output trac of the same type from the node.

bi1 cbi1 c b v01 = f01 (c {c0 , . . . , c7 }, b B, i1 S1 )

(3.54)

bB

cbi1 f01 =

i2 S1

ci1 i2 f11 (c {c0 , c1 , c3 , c4 , c7 }, i1 S1 )

(3.55)

bB

cbi1 f01 =

j1 S2

ci1 j1 f12 (c {c0 , . . . , c3 , c7 }, i1 S1 )

(3.56)

bB

cbi1 f01 =

k1 S3

ci1 k1 f13 (c {c4 , c5 , c6 }, i1 S1 )

(3.57)

i1 S1

ci1 j1 f12 =

j2 S2

cj1 j2 f21 (c {c1 , c2 , c3 }, j1 S2 )

(3.58)

i1 S1

ci1 j1 f12 =

j2 S2

cj1 j2 f22 (c {c0 , c1 , c2 , c7 }, j1 S2 )

(3.59)

i1 S1

ci1 k1 f13 =

k2 S3

ck1 k2 f31 ( c {c 4 , c 5 , c 6 } , k 1 S2 )

(3.60)

i1 S1

ci1 k1 = f13

k2 S3

ck1 k2 (c {c4 , c6 }, k1 S2 ) f32

(3.61)

7. Additional constraints state that if a lower layer node is connected to an upper layer node, there will be at least one link between them, but the number

78

of links should not exceed the maximum number of interfaces on the lower layer node.
t t Note that nt 1 , n3 and n4 are total number of interface on the edge router, CS

router and PS router. Total number of router interfaces is divided between the groups of interconnection links. For this purpose, a K coecient is dened by the network planner. The value of K would vary for each group of links connecting to the same router.
mM0

bi1 v01

bi1 m v01 (b B, i1 S1 )

(3.62)

bi1 v01 max{nt 0}

mM0

bi1 m v01 (b B, i1 S1 , t T0 )

(3.63)

bi1 Kv01 max{nt 1}

mM0

bi1 m v01 (b B, i1 S1 , t T6 )

(3.64)

i1 i2 v11

mM1

i1 i2 m (i1 , i2 S1 ) v11

(3.65)

i1 i2 Kv11 max{nt 1}

mM1

i1 i2 m v11 (i1 , i2 S1 , t T6 )

(3.66)

i1 j1 v12

mM0

i1 j1 m v12 (i1 S1 , j1 S2 )

(3.67)

i1 j1 Kv12 max{nt 1}

mM0

i1 j1 m v12 (i1 S1 , j1 S2 , t T6 )

(3.68)

i1 j1 Kv12 max{nt 1}

mM0

i1 j1 m v12 (i1 S1 , j1 S2 , t T6 )

(3.69)

i1 k1 v13

mM0

i1 k1 m v13 (i1 S1 , k1 S3 )

(3.70)

79

i1 k1 Kv13 max{nt 1}

mM0

i1 k1 m v13 (i1 S1 , k1 S3 , t T6 )

(3.71)

j1 j2 v21

mM1

j1 j2 m (j1 , j2 S2 ) v21

(3.72)

j1 j2 Kv21 max{nt 3}

mM1

j1 j2 m v21 (j1 , j2 S2 , t T6 )

(3.73)

j1 j2 v22

mM1

j1 j2 m v22 (j1 , j2 S2 )

(3.74)

j1 j2 Kv22 max{nt 3}

mM1

j1 j2 m v22 (j1 , j2 S2 , t T6 )

(3.75)

k1 k2 v31

mM1

k1 k2 m v31 ( k 1 , k 2 S3 )

(3.76)

k1 k2 Kv31 max{nt 4}

mM1

k1 k2 m v31 (k1 , k2 S3 , t T6 )

(3.77)

k1 k2 v32

mM1

k1 k2 m ( k 1 , k 2 S3 ) v32

(3.78)

k1 k2 Kv32 max{nt 4}

mM1

k1 k2 m v32 (k1 , k2 S3 , t T6 )

(3.79)

8. Non-negativity constraints identify the domain of the variables used in the model. They state that the domain of trac variables is set of all positive real numbers. f R+ |C |(|S1 |(|B |+|S1 |+|S2 |+|S3 |)+|S2 |
2 +|S |2 ) 3

(3.80)

9. Integrality constraints state that the domain of the following decision variables is set of binary numbers and can hold either 0 or 1. v B|S1 |(|B |+|S1 |+|S2 |+|S3 |)+|S2 | 80
2 +|S |2 3

(3.81)

x B|S1 |(|T1 |+|T6 |)+|S2 |(|T2 |+|T3 |+|T6 |)+|S3 |(|T4 |+|T5 |+|T6 |)

(3.82)

The following variables are dened as general variables, stating that domain is set of all natural numbers. v N|M0 ||S1 |(|B |+|S2 |+|S3 |)+|M1 |(|S2 |
2 +|S |2 ) 3

(3.83)

The ALLIPR4PP is classied as NP-hard. As a result, we hope to nd optimal solutions for the problem by using exact algorithms within a predened nite amount of time.

3.6

Exact Methods for the Design Problem

A linear programming method is proposed for the implementation of the exact algorithm. We translate the objective function, constraints and other inputs into linear programming le format which can be read by a solver. In this thesis, we used CPLEX solver. The implemented algorithm in CPLEX is branch and cut. In branch and cut algorithm, a series of continues sub-problems are solved. To manage those sub-problems, CPLEX builds a tree in which each sub-problem is dened as a node. The root of the tree is a continuous relaxation of the original problem. In case the solution to the relaxation has any fractional variables, CPLEX will try to nd cuts in the rst step. Cuts are constraints that cut away areas of the feasible region of the relaxation which contain fractional solutions. When an integer feasible solution is found or no further cuts are found, CPLEX continues with branch and bound algorithm in the second step. Branching includes splitting the search space into smaller 81

spaces, while bounding is the procedure of computing minimum and maximum of the function within the subsets of the search space [19, 21]. Since ALLIPR4PP is NP-hard and exact algorithms might require longer computational time to nd optimal solutions, we concentrate our eorts on the development of ecient heuristics to nd sub-optimal solutions.

3.7

Heuristics for the Design Problem

In this section, we propose a local search and a tabu search heuristics to nd good solutions for the global planning problem within reasonable amount of time. Both heuristics start from an initial solution and explore the search space to nd a better solution. Before introducing these heuristics, we need to describe the solution representation and initial solution.

3.7.1

Solution Representation and Initial Solution

The initial step of the LS and TS algorithms is to determine an initial solution, but rst we need to nd a way to represent the solutions. For this purpose, we use the binary and the non-binary encoding schemes. In the binary encoding scheme, a decision variable is represented by 0 or 1. The binary scheme is good for dual state cases, but is not ecient (suitable) for multi state cases. For such cases, non-binary schemes are used in which the decision variables may hold more than two values. In the UMTS network planning problem, the binary coding scheme is used to indicate if the corresponding site (i.e. potential location) is installed with a NE or not. On the other hand, the non-binary scheme is used to represent dierent types 82

Edge Router 1 0 1 0 1 0 1

RNC 0 1 1

CS Router 0 1 0 0

MGW 1 0 1

MSS 0 1

PS Router 0 1

SGSN 1 0

GGSN 1 1

Edge Router installed

CS Router not installed

MSS installed

SGSN not installed

Figure 3.3: Example of the NE-Install vector

Edge Router 1 0 2 0 2 0 2 RNC 0 1 1 CS Router 0 2 0 2 MGW 1 0 0 MSS 0 2 PS Router 0 2 SGSN 2 0 GGSN 1 1

Edge Router of type 2

RNC of type 1

MSS of type 0

SGSN of type 2

Figure 3.4: Example of the NE-Type vector of NEs, installed in each site. In order to represent a solution, rst we need to know whether a NE is installed on a specic site and then determine its type. For this purpose, we dened a vector, called NE-Install, where the size of the vector is equal to the number of potential sites for all NE. Figure 3.3 shows an example of the NEInstall vector with 5 potential sites for each edge router/RNC, 3 potential sites for CS Router/MGW/MSS and 2 potential sites for each PS router/SGSN/GGSN. A value of 1 means that the corresponding node is installed on that site whereas the value 0 indicates that no NE is installed in that site. The next step is to determine the type of all NEs, if they are already installed in a given site. As we deal with more than two types of NEs, we used non-binary scheme. For this purpose, a second vector, called NE-Type was dened. As shown in Figure 3.4, this vector has the same size as the NE-Install vector. For example, in

83

this thesis, each NE has three types. Therefore, each site in the NE-Type vector can hold any of the dened values (0, 1 or 2), representing the type of the corresponding NE installed on that site. We use a combination of the NE-Install and the NE-Type vectors to represent a solution, including the initial solution.

3.7.2

Local Search Heuristic

Local search heuristic starts with an initial solution and explores the neighborhood to nd a better solution. LS can sometimes provide the global optimum. However, it will be trapped in the rst local minimum. The local search heuristic algorithm used in this thesis is shown in Figure 3.5. The rst step is to nd an initial solution to start with. In this thesis, we suppose that in the initial solution, all NEs (i.e. RNCs, MGWs, MSSs, SGSNs, GGSNs, edge routers/CS routers/PS routers) are installed in all potential locations with their most powerful type. To design interconnecting links, we break down the global problem into eight sub-problems: the Node B to edge router, the edge router to RNC, the edge router to CS router, the edge router to PS router, the CS router to MGW, the CS router to MSS, the PS router to SGSN and the PS router to GGSN. Each of these assignment sub-problems is NP-hard and to solve them, we use shortest augmentation path algorithm LAPJV [99] as proposed in [95]. The LAPJV algorithm, developed by Jonker and Volgenant, contains new initialization routines and a special implementation of Dijkstras [100] shortest path method. An implementation of LAPJV algorithm in C++ is used in this thesis. Objective of the eight assignment sub-problems is to minimize the cost of the interconnecting links between the lower and the higher level nodes. The trac volume 84

Find an initial solution s

Solution s is feasible

No No feasible solution exist

Yes Evaluate the cost function of the current solution f(s) Set best_cost := f(s)

Explore the neighborhood of the current solution : N(s) Set the best neighbourhood solution t' := argmin {f(t) : t in N(s)} Set cost_temp := f(t')

cost_temp < best_cost

No

Local minimum found Return the best solution

Yes s := t' Set best_cost := cost_temp

Figure 3.5: Local search algorithm [5] passing from a node determines the number and type of its interconnecting links and interfaces. Starting from the rst sub-problem, LAPJV uses the cost matrix of the interconnection links among the lower layer to the upper layer node to nd the cost-optimal interconnections. LAPJV repeats the same task for the rest of the subproblems. At the end, an initial solution is found for the whole planning problem. If the initial solution does not respect all constraints, then it is not feasible and the 85

program will terminate. Otherwise, the cost of the feasible solution is calculated and is stored as the best cost found so far. The solution is also stored as the current solution. The second step consists of exploring the neighborhood of the current solution. This type of operation is called move. A move is transformation of the current solution to produce a new solution. The following types of move can be made from a current solution: 1. Remove a NE (i.e. RNCs, MGWs, MSSs, SGSNs, GGSNs, edge routers/CS routers/PS routers) that is already installed. 2. Change the type of a NE (i.e. RNCs, MGWs, MSSs, SGSNs, GGSNs, edge routers/CS routers/PS routers) that is already installed. Starting from the rst sub-problem, a new solution is created by a move. Considering the new solution, the number and type of the interconnecting links and interfaces for the entire problem are recalculated. Then, the feasibility check is done. If the new solution is not feasible, the corresponding move is returned and another move is performed. But if the new solution is feasible, its cost is calculated and is stored in cost temp. If value of cost temp is less that the best cost calculated in the previous step, the new solution is accepted and is considered as the current solution. In the similar way, cost temp is also taken as the best cost. Exploring the neighborhood of the current solution is a continuous duty of LS, in the hope of nding a better solution (i.e. a solution with a lower cost). If a move results in a feasible solution with higher cost than the current solution, the algorithm will suppose that no further improvement can be made. Therefore, a local minimum 86

is found and ultimately, the current solution is considered as the nal solution.

3.7.3

Tabu Search Heuristic

The tabu search was rst formalized by Fred Glover and Manuel Laguna in 1986 [101]. TS uses short term memory in the form of so-called tabu list and aspiration criteria to avoid cycling and escape from local minimum. Tabu list guides the search direction by storing the features of recently visited solutions to prevent returning to them. The solutions from the neighborhood of the current solution, which their features are currently found in tabu list are excluded. Such action allows for the possibility of losing unvisited high quality solutions. In order to overcome this problem, the aspiration criteria are dened. Aspiration criteria allow to accept a tabu move despite violating the tabu condition. However, the included solution should be better than the best solution found so far [19]. The tabu search heuristic presented in this thesis is shown in Figure 3.6. Similar to LS, the global planning problem is broken into eight sub-problems. The procedure to nd the initial solution is the same as mentioned for the LS. However, a better way consists to use the nal solution produced by LS as the initial solution for TS. The initial solution (current solution) and its cost (best cost) are kept in memory for further improvements. The next step is to determine the best move in the neighborhood of the current solution. The following types of move can be made from a current solution: 1. Remove a NE (i.e. RNCs, MGWs, MSSs, SGSNs, GGSNs, edge routers/CS routers/PS routers) that is already installed

87

Find an initial solution s

Evaluate the cost function of the current solution f(s) Set best_cost := f(s) Set num_start := 1

Set num_iter := 1

Explore the neighborhood of the current solution : N(s) Set the best neighborhood solution s' := argmin {f(t) : t in N(s)} considering the tabu list and aspiration criteria Set cost_temp := f(t)

Update tabu list

cost_temp < best_cost

Yes

Update the best solution Set best_cost := cost_temp

No

s := t' num_iter := num_iter + 1

Update tabu list

Yes

num_iter <= max_iter No

Yes

num_start <= max_start

No Return best solution found

Figure 3.6: Tabu search algorithm [6] 88

Table 3.2: Structure of the tabu list size

NE Edge router RNC CS router MGW MSS PS router SGSN GGSN Tabu tenure No. of potential locations for the edge router L1 + No. of potential locations for the RNC L2 + No. of potential locations for the CS router L3 + No. of potential locations for the MGW L4 + No. of potential locations for the MSS L5 + No. of potential locations for the PS router L6 + No. of potential locations for the SGSN L7 + No. of potential locations for the GGSN Tabu tenure representation L1 L2 L3 L4 L5 L6 L7 L8

2. Add a NE (i.e.

RNCs, MGWs, MSSs, SGSNs, GGSNs, edge routers/CS

routers/PS routers) of type t. 3. Change the type of a NE (i.e. RNCs, MGWs, MSSs, SGSNs, GGSNs, edge routers/CS routers/PS routers) that is already installed We have dened the tabu list as a non-binary vector with a variable length. The length of the tabu list, know as tabu tenure, depends on the NE which the move is made over. For example, in the rst sub-problem, the move is applied on the edge router and the tabu tenure is equal to the number of potential locations for the edge routers. As we proceed with the second sub-problem, the tabu tenure will grow and will include the number of potential locations for the RNCs plus the number of the potential locations for the edge routers. The tabu tenure grows up as the algorithm makes moves in the subsequent sub-problem. Table 3.2 shows the size of the tabu list, when a move is made over an specic NE. For each iteration, TS determines the best move while considering the tabu list and the aspiration criteria. Among all possible neighbors, the best move is the one 89

which produces a solution with lower cost than the best cost in the neighborhood. However, the best move also can be an inferior solution (not been in tabu list) which can be accepted temporary in order to prevents entrapment in a local minimum. If the resulting solution has a lower cost than the best cost, it is taken as the best solution found so far. Otherwise, if the solution is not in tabu list, it is accepted even though the solution quality is not better than the best found so far. At the start of the algorithm, the tabu list is empty, but if a new solution is accepted, it is added to tabu list. The purpose of such an action is to avoid cycling and revisiting instances that have been seen in the recent past. Any solution is kept in tabu list for a given number of iteration. To simplify the denition, we call it tabu life time. TS conguration parameters are problem specic and are acquired through multiple tests. Selecting a proper duration for tabu conditions is so important. It was shown that assigning a very small number to tabu life time, causes cycling, but as this number is increased, the probability of visiting several good solutions increases. However, the tabu life time should not be a very large number, because it then becomes less probable to nd good local minimum for lack of available moves [102]. It is also shown that randomly selecting the tabu life time can direct the search towards good solutions [19]. The number of iterations is randomly chosen between a lower bound (L) and an upper bound (U). Due to this randomness, running the algorithm for multiple times can yield dierent solutions. As a result, a multi-start TS algorithm is proposed [97]. TS can be congured to stop either at the end of the search or by some conditions. Examples of stopping conditions can be that the neighborhood is empty, the maximum number of solutions to be explored is xed or the total number of iteration is reached. 90

After each iteration, the life time of solutions in the tabu list is decreased by one. During the overall search process, the current solution will be compared with the best solution found so far to keep the best solution up to date. When the whole search process ends, the best solution over the whole search process will be the nal output. Tabu search is a metaheuristic technique that must be adapted to the problem at hand for it to be ecient. The choice of moves that generate the neighborhoods is problem-specic. Dierent implementations can be generated by varying the denition and the structure of the tabu list, the aspiration criteria and other attributes [103]. In the next chapter, the simulation results based on these algorithms are presented and analyzed.

91

Chapter 4 Experiment Design and Result Analysis

In this chapter, we will rst describe how the simulations were developed. We introduce simulation environments and data used to run the simulations. Later, we explain the simulation procedures and present the results and analysis.

4.1

Experiment Design

In this thesis, we supposed that Node Bs are already installed and their number and locations are known. We also assume that the potential location for all other NEs (i.e. RNCs, MGWs, MSSs, SGSNs, GGSNs, edge routers, CS routers, PS routers) is known. For trac modeling, we suppose that the subscriber trac prole and the planning parameters are known. This information is usually provided by the operator

92

Table 4.1: The subscriber trac prole

Circuit services Average rate Monthly usage Activity factor Voice AMR 12.2 Kbps 300 Min 0.5 Call Type CS Data 16 Kbps 64 Kbps 10 Min 200 Mbps 1 0.4 PS Data 144 Kbps 100 Mbps 0.4

384 Kbps 50 Mbps 0.4

Table 4.2: The planning parameters

AMR rate in active mode AMR rate in silent mode Mean holding time Soft handover ratio IP over head Attached subscriber ratio Active BH PDP ratio Average PDP duration Retransmission factor Iub interface blocking factor Iub utilization factor Busy days per month Busy days call ratio Busy hour call ratio 18.9 Kbps 4.5 Kbps 120 seconds 0.4 0.38 0.2 0.5 900 seconds 0.25 0.02 0.75 22 0.9 0.1

and might vary subject to trac forecast and planning strategies. The subscriber trac prole and the trac planning parameters used in this thesis are respectively presented in Tables 4.1 and 4.2. These values are considered based on practical work experience of the author with mobile operators. The Adaptive Multi Rate (AMR) codec is used by voice users. The activity factor of a trac channel is dened as the percentage of time, where trac is present in the channel. We suppose that the activity factor for voice is 0.5 (and 0.5 silence), while for CS-data and PS-data trac the activity factor is 1 and 0.4 respectively. It

93

is supposed that network operators dene a monthly limited amount for voice and data trac per subscriber. The network is supposed to be designed to deliver the required capacity for the specic number of subscribers. We rst started with the subscriber allocation for each Node B. We assigned a random number of subscribers to each Node B in a range of 5,000 to 20,000 subscribers. Then, using the subscriber trac prole and trac planning parameters, we created a trac record for each Node B. A sample trac record was introduced in Chapter 3. The detail calculations of the Node B trac records are presented in Appendix A. We supposed that three types are available for each NE (type A, B and C) and four dierent types of links/interfaces can be used. It is also supposed that the specications such as the processing capacity, the interface capacity and cost are given for each type of equipment. Similarly, we assumed that the capacity and cost for all types of the links/interfaces are provided. The cost includes all related costs such as oor space, cables, racks, electric installations, labor, etc. The specications for the nodes, links and interfaces used in this thesis are shown in Tables 4.3 to 4.10. These values are considered based on practical work experience of the author with mobile operators. The up-to-date values for these parameters in real market might vary in time with what presented in this thesis. Using a C++ based program, we created a text le containing the number and location of all NEs (i.e. Node Bs, RNCs, MGWs, MSSs, SGSNs, GGSNs, edge routers, CS routers, PS routers), as well as the trac records for each Node B. This le is used as input for both exact and approximate algorithms. A MATLAB le was also created to graphically depict the potential locations of all NEs. 94

Table 4.3: Features of the Node B types

Type A No. of edge router interfaces 10

Table 4.4: Features of the RNC types

Type A Capacity (channels) Capacity (BHCA) Capacity (Mbps) No. of edge router interfaces Cost ($) 181, 000 147, 000 1, 950 12 100, 000 Type B 284, 000 231, 000 2, 800 16 150, 000 Type C 463, 000 443, 000 3, 600 32 250, 000

Table 4.5: Features of the MGW types

Type A Capacity (channels) Capacity (BHCA) No. of CS router interfaces Cost ($) 4, 048 200, 000 12 200, 000 Type B 12, 600 600, 000 16 350, 000 Type C 22, 000 900, 000 32 500, 000

Table 4.6: Features of the MSS types

Type A Capacity (Mbps) Capacity (channels) Capacity (BHCA) Capacity (subscribers) No. of CS router interfaces Cost ($) 500 15, 000 400, 000 400, 000 12 300, 000 Type B 1, 00 20, 000 600, 000 600, 000 16 500, 000 Type C 2, 000 41, 500 1, 000, 000 1, 000, 000 32 600, 000

95

Table 4.7: Features of the SGSN types

Type A Capacity (Mbps) Capacity (PDP context) Capacity (subscribers) No. of PS router interfaces Cost ($) 500 480, 000 500, 000 12 200, 000 Type B 750 720, 000 750, 000 16 300, 000 Type C 1, 000 960, 000 1, 000, 000 32 400, 000

Table 4.8: Features of the GGSN types

Type A Capacity (Mbps) Capacity (PDP context) No. of PS router interfaces Cost ($) 333 333, 000 12 300, 000 Type B 666 666, 000 16 400, 000 Type C 1, 000 960, 000 32 500, 000

Table 4.9: Features of the router types

Type A Capacity (Mbps) No. of interfaces Cost ($) 2, 000 32 80, 000 Type B 4, 000 48 120, 000 Type C 8, 000 64 160, 000

Table 4.10: Costs of the link and interface types

Type DS-1 DS-3 FE GE Interface Cost ($) 300 1, 500 300 1, 500 Link Cost ($/km) 700 1, 500 700 1, 500

96

4.1.1

Implementation of the Exact Algorithm

We used the linear programming method for the implementation of the exact algorithm to solve the planning problem of 3G UMTS all-IP Release 4 networks. For this purpose, we developed a C++ based program to minimize the constructed cost function subjects to the corresponding constraints. Using the input le (explained in previous section), the problem (including the cost function, the constraints and all inputs) was properly converted into a linear programming (.lp) le format which can be read by the CPLEX solver. We used CPLEX solver 12.1 with all default setting to solve the problem, under a Linux server with one 2.39 GHZ CPU (AMD Opteron Processor 250 family) and 16 GB RAM. Considering the nature of the problem and the high number of constraints, we expected a long computation time. As a result, we set up a Time Limit (TL) of 24 hours. This means that after 24 hours of computation time, CPLEX will return the best solution found so far. If the problem can be solved within the time limit, then the output of the CPLEX solver is the optimal solution. The solution consists of the number, the location and the type of all NEs, links, interfaces, as well as the topology of access and core networks. The solution also provides the cost and the CPU time for the problem. These costs will be used as benchmark for comparison with the results from the approximate algorithms. For the simplicity of depiction, a C++ program converts the solution into a MATLAB le to produce a graphical topology.

97

4.1.2

Implementation of Approximate Algorithms

Among dierent heuristics, local search and tabu search were implemented to solve the planning problem of 3G UMTS all-IP Release 4 networks. We decided to use the local optimum solution found by LS as the initial solution for TS. For this purpose, we implemented LS and TS both in the same C++ based program, such that, a local minimum is rst found by LS. Then, the quality of such solution is improved by TS with the hope of nding a global optimum. The program aims to minimize the cost function as presented in equation 3.3. We start from the initial solution in which NEs are installed in all their potential locations with their most powerful type. It means that the all elements of NE-Install and NE-Type vectors are respectively equal to 1 and 2. To design the interconnecting links, we break down the ALLIPR4PP into eight sub-problems: the Node B to edge router, the edge router to RNC, the edge router to CS router, the edge router to PS router, the CS router to MGW, the CS router to MSS, the PS router to SGSN and the PS router to GGSN.

4.1.2.1

Local Search

Beside the number, type and location of NEs, we require to nd the cost-optimal interconnecting topology as a part of the initial solution. We started from the rst sub-problem. The link type and capacity between each Node B and edge router is determined based on amount of Node B trac. Node B trac is ultimately accumulated on the selected edge routers. Knowing the number and types of the links, a cost matrix of interconnection links between the Node Bs and the edge routers is

98

constructed, while considering corresponding constraints. The cost matrix contains several possible assignment combinations and is used as input for LAPJV function to nd the cost-optimal interconnections between all Node Bs and the edge routers. Once the solution of the rst sub-problem is found, the upcoming sub-problems are similarly tackled. Once the eight sub-problems are solved, the obtained solution will be the current solution. The cost of the current solution is then calculated and is stored as the best cost so far. The second step consists of exploring the neighborhood of the current solution. For this purpose, a move is made and and is accepted, if it has a lower cost than the best cost found so far. Otherwise, LS will suppose that no further improvement can be made and will stop. At this point a local minimum is found and ultimately, the current solution will be the nal solution. To execute LS, we used Linux g++ compiler with the same platform as before.

4.1.2.2

Tabu Search

The local minimum found by LS is used as the initial (current) solution for TS. The current solution and its cost are respectively known as the best solution and the best cost. For further improvement, TS determines the best move in the neighborhood of the current solution. The best move is chosen based on tabu list and aspiration criteria. TS performs three types of moves: i ) remove a NE, ii ) add a NE and iii ) change the type of a NE. We implemented add(remove) a NE by changing the corresponding location in NE-Install vector to 1(0), while the type change of a NE is done by modifying the corresponding location in NE-Type vector to the new type value. 99

We congured TS with the following parameters: Tabu life time in tabu list randomly chosen between 5 and 9; Maximum number of iterations (max iter)=100; Maximum number of start (max start)=2. To chose the tabu life time, dierent values are tested and the optimum results were obtained with a random number between 5 to 9. To x a value for the number of iteration, we repeated the search for several times. In fact, the main improvement was happened in the early iterations. We set an upper bound of 100 iterations to t for all set of problems. It was also seen that, with running the tabu program for more than 2 times, minor quality improvement is achieved over some problems, in the expense of longer CPU time. TS performs 100 iterations for 2 times. At each iteration, the current solution is compared with the best solution found so far. When the whole search is completed, the best solution is returned as the nal solution. The solution consists of the number, the location and the type of all NEs, type and number of the links and interfaces, as well as the topology of access and core networks. The solution also includes the cost and the CPU time for the solution. Similar to LS, we used Linux g++ compiler with the same platform as before.

4.2

Illustrative Example

In this section, we select a sample problem and attempt to solve it using the exact and the approximate algorithms. We use local search, tabu search and integer 100

20

Node B Edge Router RNC CS Router MGW MSS PS Router SGSN GGSN

15

Distance (km)

10

10 Distance (km)

15

20

Figure 4.1: Potential locations for the illustrated example programming methods. The example problem depicted in Figure 4.1, comprises of 45 installed Node Bs, 5 potential sites to install the edge routers (RNCs), 5 potential sites to install the CS routers (MGWs/MSSs) and 5 potential sites to install the PS routers (SGSNs/GGSNs) (i.e. |S1 | = 5, |S2 | = 5, |S3 | = 5). The input for the problem consists of randomly generated locations for these NEs, as well as trac records for the Node Bs. In order to design the network, three types of NEs (i.e. RNCs, MGWs, MSSs, SGSNs, GGSNs, edge routers, CS routers, PS routers) and four types of links (interfaces) are available with the features presented in Tables 4.3 to 4.10. In the rst step, we used the local search heuristic to solve the example problem. 101

20

15

Distance (km)

Node B RNC MGW MSS SGSN GGSN Edge Router CS Router PS Router

10

10 Distance (km)

15

20

Figure 4.2: Solution for the illustrated problem obtained with LS Figure 4.2 shows the designed network with the LS method. The network comprises of two RNCs (one of type B and one of type C ), one MGW (of type C ), one MSS (of type B ), one SGSN (of type A), one GGSN (of type A), two edge routers (both of type B ), one CS routers (of type A) and one PS routers (of type A). The network cost and CPU time are $2,975,949 and 743 seconds respectively. In the second step, tabu search was used to nd a solution with lower cost. As shown in Figure 4.3 the network designed by tabu search includes two RNCs (one of type A and one of type C ), one MGW (of type C ), one MSS (of type B ), one SGSN (of type A), one GGSN (of type A), three edge routers (all of type A), one CS router (of type A) and one PS router (of type A). The network cost and CPU time are $2,862,127 and 2,427 seconds respectively. The designed network with TS has an 102

20

15

Distance (km)

Node B RNC MGW MSS SGSN GGSN Edge Router CS Router PS Router

10

10 Distance (km)

15

20

Figure 4.3: Solution for the illustrated problem obtained with TS optimized topology with lower cost compared to LS. In the last step, we used the integer programing method. To solve the example, the CPLEX Mixed Integer Optimizer 12.1 was used with all the default settings. CPLEX spent 16,319 seconds to complete and the returned objective value is $2,833,406 after exploration of 44,538,161 iterations and 1,078,504 nodes. Figure 4.4 is the output of CPLEX, showing the topology and conguration with two RNCs (one of type A and one of type C ), one MGW (of type C ), one MSS (of type B ), one SGSN (of type A), one GGSN (of type A), two edge routers (one of type A and one of type B ), one CS router (of type A) and one PS router (of type A). The gap of the solution provided by local search is 5.03% from the solution provided by CPLEX, while the solution provided by TS has gap of 1.01% from the 103

20

Node B RNC MGW MSS SGSN

15

GGSN Edge Router CS Router

Distance (km)

PS Router

10

10 Distance (km)

15

20

Figure 4.4: Solution for the illustrated problem obtained with CPLEX optimal solution.

4.3

Result Analysis

In this section, we present the experiment results and analyze the performance of the proposed algorithms. Two problem sets of 15 sizes, each set with four dierent instances (120 problems in total) were randomly generated within an area. The size of the area for small and large size problems are respectively 100 km2 and 400 km2 .

104

Table 4.11: Problem sizes - set#1

Problem No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Area (km2 ) 100 100 100 100 400 400 400 400 400 400 400 400 400 400 400 No. of Node B 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 No. of edge router, RNC (S1 ) 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 No. of CS router, MGW, MSS (S2 ) 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 No. of PS router, SGSN, GGSN (S3 ) 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5

4.3.1

Numerical Analysis

Table 4.11 shows the 15 problem sizes, where the rst and second columns represent the problem number and the area size. The third column shows the number of Node Bs that are already installed. The number of potential locations for the edge router (RNC) is presented in column four. The fth column shows the potential locations for CS router (MGW/MSS). Finally, the number of potential locations for PS router (SGSN/GGSN) are presented in column six. In this set of problems, we intended to represent real live networks where the problems complexity is mainly proportional to the number Node Bs. A more complex set of problems are presented in Table 4.12 in which the potential locations for edge router (RNC) is increased from 5 to 10. For each problem sets, four instances were randomly generated and solved by

105

Table 4.12: Problem sizes - set#2

Problem No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Area (km2 ) 100 100 100 100 400 400 400 400 400 400 400 400 400 400 400 No. of Node B 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 No. of edge router, RNC (S1 ) 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 No. of CS router, MGW, MSS (S2 ) 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 No. of PS router, SGSN, GGSN (S3 ) 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5

CPLEX, LS and TS. Tables 4.13 and 4.14 show the numerical results for the rst instance of the problem set#1 and set#2. The rst column shows the problem number which corresponds to the rst column of Tables 4.11 and 4.12. The following two columns provide the results obtained by CPLEX solver. Columns 4 and 5 provide the best results and the corresponding CPU time obtained with the local search. Column 6 shows the gap (expressed in percentage) between the solution values obtained by the local search and the optimal solution. The following three columns show the best solutions, the corresponding computation time, as well as the gap for the tabu search. As mentioned earlier, a CPU time limit of 24 hours (86,400 seconds) is set for CPLEX. Since the problems are NP-hard, CPLEX is not able to solve some of the problems in the dened TL. Once the TL is reached, CPLEX stops and returns the best solution found so far. It should be noted that problems facing the time limit are 106

Table 4.13: Numerical results for the rst instance of set#1

CPLEX Problem No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Cost ($) 1, 462, 335 1, 499, 105 1, 760, 827 1, 850, 368 2, 100, 229 2, 444, 225 2, 514, 815 2, 788, 062 2, 877, 196 3, 133, 179 3, 233, 417 3, 274, 512 3, 320, 597 3, 576, 809 3, 551, 323 CPU (s) 152 54 1, 435 945 1, 893 4, 026 434 650 898 3, 839 48, 275 13, 291 30, 260 45, 666 78, 102 Local Search Cost ($) 1, 499, 944 1, 520, 426 1, 776, 100 1, 974, 095 2, 221, 985 2, 596, 044 2, 599, 864 2, 975, 949 3, 007, 039 3, 354, 674 3, 441, 074 3, 498, 771 3, 551, 861 3, 819, 732 3, 771, 826 CPU (s) 200 285 373 458 540 605 715 743 858 934 938 1, 093 1, 137 1, 241 1, 174 Gap (%) 2.58 1.43 0.87 6.69 5.8 6.22 3.39 6.74 4.52 7.07 6.43 6.85 6.97 6.8 6.21 Tabu Search Cost ($) 1, 462, 335 1, 500, 627 1, 760, 827 1, 944, 718 2, 145, 767 2, 472, 978 2, 574, 839 2, 862, 127 2, 931, 825 3, 295, 990 3, 413, 646 3, 440, 937 3, 479, 220 3, 791, 646 3, 737, 246 CPU (s) 220 322 462 743 1, 014 1, 196 1, 955 2, 427 2, 134 3, 732 4, 333 3, 535 4, 064 7, 094 6, 299 Gap (%) 0 0.11 0 5 .1 2.17 1.18 2.39 2.66 1.9 5.2 5.58 5.09 4.78 6.01 5.24

set aside and are not involved in the GAP and CPU time calculations. The numerical results for the other six instances are presented in Appendix B. The comparison in terms of the solution quality (cost) among CPLEX, LS and TS is presented in Figures 4.5 and 4.6. The cost comparisons for other six instances are presented in Appendix B. We evaluate the quality of the approximate algorithms by using CPLEX solutions as the reference. Analyzing Figures 4.5 and 4.6, we can see that LS and TS are able to provide solutions which are relatively close to the optimal solutions. The graphs show that the solutions provided by LS and TS are closer to the optimal solution for the small size problems rather than the large size problems. This complies with

107

Table 4.14: Numerical results for the rst instance of set#2

CPLEX Problem No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Cost ($) 1, 455, 778 1, 498, 893 1, 744, 211 1, 851, 255 2, 083, 141 2, 406, 642 2, 458, 280 2, 790, 753 2, 861, 354 3, 089, 932 3, 146, 851 3, 195, 625 3, 408, 300 3, 405, 885 3, 517, 579 (TL) (TL) (TL) (TL) (TL) CPU (s) 507 3, 069 869 19, 165 832 32, 327 2, 713 1, 424 1, 550 3, 284 86, 400 86, 400 86, 400 86, 400 86, 400 Local Search Cost ($) 1, 469, 132 1, 518, 248 1, 792, 572 1, 962, 585 2, 218, 899 2, 505, 148 2, 627, 218 3, 050, 053 2, 966, 178 3, 321, 133 3, 430, 440 3, 376, 065 3, 592, 251 3, 557, 004 3, 737, 430 CPU (s) 1, 356 2, 015 2, 724 3, 250 3, 844 4, 506 5, 035 5, 524 6, 083 7, 987 7, 458 7, 935 8, 360 8, 768 9, 206 Gap (%) 0.92 1 .3 2.78 6.02 6.52 4 .1 6.88 9 .3 3.67 7.49 Tabu Search Cost ($) 1, 456, 191 1, 498, 893 1, 744, 211 1, 945, 253 2, 123, 207 2, 431, 723 2, 572, 914 2, 918, 001 2, 923, 737 3, 286, 329 3, 260, 640 3, 339, 825 3, 528, 878 3, 549, 322 3, 714, 285 CPU (s) 1, 391 2, 071 2, 849 3, 638 4, 463 5, 004 6, 046 7, 066 7, 650 11, 775 11, 239 11, 236 13, 904 18, 321 18, 649 Gap (%) 0.03 0 0 5.08 1.93 1.05 4.67 4.56 2.19 6.36

the fact that smaller problems have a smaller search space and heuristics can explore more neighborhoods and more likely nd solutions close to optimal. Similarly, we can see that the quality of the solutions provided by LS and TS decays as the problem size grows. Furthermore, as the problem size grows, the cost also increases as more NEs are required to cover all of the users. Another comparison in terms of the CPU time of the solutions among CPLEX, LS and TS is presented in Figures 4.7 and 4.8. The CPU time comparisons for the other six instances are presented in Appendix B. We use the CPU time returned by CPLEX as the reference for the comparison. Analyzing the Figures 4.7 and 4.8, we can see that LS and TS are able to provide

108

4,000,000 3,500,000 3,000,000 Cost ($) 2,500,000 2,000,000 1,500,000 1,000,000

LS Cost ($) TS Cost ($) CPLEX Cost ($)

10 11 12 13 14 15

Problem number (set #1)

Figure 4.5: Solution comparison for the rst instance of set#1 (cost)
4,000,000 3,500,000 3,000,000 Cost ($) 2,500,000 2,000,000 1,500,000 1,000,000
LS Cost ($) TS Cost ($) CPLEX Cost ($)

10 11 12 13 14 15

Problem number (set #2)

Figure 4.6: Solution comparison for the rst instance of set#2 (cost) good solutions in reasonable amount of time. CPU time for TS is more than LS, because TS performs more iterations than LS. CPU time of LS and TS increases with 109

100,000

10,000 CPU time (s)

1,000 LS CPU (s) 100 TS CPU (s) CPLEX CPU (s) 10

1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Problem number (set #1)

Figure 4.7: Solution comparison for the rst instance of set#1 (CPU)
100,000

10,000 CPU time (s)

1,000
LS CPU (s)

100

TS CPU (s)

CPLEX CPU (s)

10

1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Problem number (set #2)

Figure 4.8: Solution comparison for the rst instance of set#2 (CPU) a constant rate with respect to the problem size. Most of the problems are solve within10,000 seconds. It is seen that CPLEX is faster than LS and TS for the small

110

size problems, as it is using the branch and cut algorithm. On the contrary, CPLEX becomes inecient as the problem size increases. It should be noticed that a time limit of 24 hours was set for CPLEX and as Figure 4.4 shows, CPLEX was not able to solve ve large problems and would need more time to nd optimal solutions. This was expected as the global planning problem of UMTS networks is NP-hard.

4.3.2

Statistical Analysis

From a total of 120 dierent problems, CPLEX and tabu search were able to provide the optimal solution for 92 (%77) and 19 (16%) problems respectively, while no optimal solution was found by the local search. The statistical cost analysis is presented in Table 4.15. In this analysis, we have just considered the problems which CPLEX was able to nd the optimal solutions. The rst column shows the algorithm and the next three columns represent respectively the minimum, the maximum and the average solution gaps. The last two columns present the standard deviation and the condence interval for the average solution gaps. We can state that with 95% condence, local search provides a solution which its true mean gap is within the interval [4.42 , 5.54]. Similarly, tabu search has 95% condence that the true mean solution gap is within the interval [2.37 , 3.27]. We can summarize that TS provides a better solution quality compared to LS. In fact, TS is able to provide the average gap of 2.82% on network cost which shows 43% improvement compared to LS (4.98%). We also performed statistical CPU time analysis and presented in Table 4.16. Similarly, in this analysis, we have just considered the problems which CPLEX was able to nd the optimal solutions. We can state that with 95% condence, tabu 111

Table 4.15: Solution gap comparison for the total simulations

Algorithm Local search Tabu search Min. gap (%) 0.27 0.00 Max. gap (%) 11.31 7.51 Ave. gap (%) 4.98 2.82 Standard deviation (%) 2.73 2.20 95% Condence interval (%) 4.98 0.56 2.82 0.45

Table 4.16: CPU time comparison for the total simulations

Algorithm Local search Tabu search CPLEX Min. CPU (s) 196 216 31 Max. CPU (s) 8, 108 11, 775 85, 242 Ave. CPU (s) 2, 032 3, 256 15, 515 Standard deviation (s) 2, 042 2, 631 21, 925 95% Condence interval (s) 2, 032 417 3, 256 538 15, 515 4, 480

search provides a better solution with the true mean CPU time within in the interval [2,718 , 3,794] seconds as shown in Table 4.16. We can also conclude that CPLEX will require a lot higher time to solve some of the problems which faced time limit and were stopped before optimal solution.

4.4

Section Remarks

The main idea of using approximate algorithms is to nd an acceptable trade o between the quality of a solution and the computation time required to nd the solution. We realized that TS is able to search a larger solution space to nd good solutions which are close to the optimal solutions. Real world UMTS Release 4 networks are very large and complex. Considering new NEs and dierent types of trac, the planning problem of such networks becomes more complex. Using CPLEX to solve 112

the planning problems will be a good idea for small size networks, but for the large networks, CPLEX will require a lot of time. As a result, the proposed algorithms, especially TS, are more appropriate for large cases to provide good solutions in reasonable computation time. In the real world scenarios with very large areas, the model is rst used to design regional networks. Once the regional networks are designed, the transit trac from(to) one regional network to(from) all other regional network are calculated and all regional networks are updated to accommodate extra transit trac. To interconnect regional network together and construct a fully connected network, dierent type of topologies are used. The most common topology is layered ring topology with alternative paths for higher reliably.

113

Chapter 5 Conclusions and Future Work

The primary goal of the UMTS network planning is the topology planning. To help network operators gain a long term prot, ecient planning tools are necessary in order to make a trade o between network investment and performance. Since the global UMTS network planning problem is shown to be NP-hard, approximate algorithms (based on metaheuristics) must be used to nd a balance between solution quality and computation time. The eciency of the algorithm is problem dependent, therefore an appropriate selection and design of the metaheuristics is the key factor for the success of a planning tool. In this thesis, we introduced a new mathematical model in order to plan 3G UMTS all-IP Release 4 networks with realistic trac prole. The aim of the model is to design the network topology which can provide a satisfactory level of service to the subscribers with minimum network cost. More precisely, the model selects the location and the type of the NEs (i.e. RNCs, MGWs, MSSs, SGSNs, GGSNs, edge routers, CS routers, PS routers), their links and interfaces, as well as the access 114

and core topologies. As a second contribution, we proposed metaheuristics to solve the design problem of the model. In fact, we adapted the local search and the tabu search metaheuristics, to our specic problem and developed planning tools to solve the design problem. These algorithms use the global approach and consider the access and core network planning sub-problems simultaneously. During the simulations, two problem sets of 15 sizes, each set with four dierent instances (120 problems in total), were randomly generated and solved with the proposed algorithms. In order to assess the performance of those algorithms, we compared the results with the optimal solutions obtained from a commercial solver (CPLEX). The results demonstrated that the local search is able to nd solutions that, with 95 percent condence, the true mean gap is within the interval of [4.42% , 5.54%] from the optimal solutions. Further improvement was achieved by employing the tabu search. With 95 percent condence interval, tabu search provides solutions where the true mean gap lies within the interval of [2.37% , 3.27%] from the optimal solutions. Comparing results from the local search and tabu search with the exact algorithm, we realized that the main advantage of such metaheuristics is the low CPU time. Therefore, larger problem sizes can be tackled. This research work showed that the proposed algorithms are able to nd good solutions for planning problem of UTMS all-IP Release 4 networks within reasonable amount of time. Nevertheless, some more improvement could be done. In this simulation, the number of subscribers were randomly generated which might be different from real live networks. Another limitation is that CPLEX is not very useful for medium and upper medium size problems. As shown in the simulation results, several problems were not solved within the time limit. Therefore, it is dicult 115

to evaluate the quality of the proposed algorithms for medium and upper medium problems. Dierent parameter settings for CPLEX could be tested to increase the solution progress in the best bound. Moreover, we considered the access and the core networks in a tree topology. Other more reliable topologies could also be explored. Out model is sensitive to changes in input values and to justify such shortcoming, it is recommended to design the network for some extra capacity. Some future works of this thesis may consist of considering other interconnection topologies like the ring, the star and the layered in access and core networks. Another works can be performing sensitivity analysis over the model as well as nding ecient algorithms to reduce the problem complexity while dwindling the size of the feasibility area in the preliminary steps. Nowadays, network updates are more frequent that designing a green eld network. A potential future work can include adaptation of the model to cover update problems by modifying the tabu search procedure. Finally, investigating newly introduced hybrid-metaheuristics to solve the planning problem may be another potential research direction as well.

116

Appendix A
In this appendix, we present the detail calculations of trac records. The trac prole of subscribers and the planning parameters, shown in Tables 4.1 and 4.2, are used as inputs. Another input is the number of subscribers for each Node B, which is randomly generated between 5,000 to 20,000. Using this information, CS voice, CS data and PS data trac is calculated over the Node B-edge router interface (Iub). We also present the C++ code of the Erlangb-to-Channel function, used in the trac prole calculations.

Iub CS voice trac CsV oiceT raf f ic = (U serM onthlyCSV oiceT raf f ic BusyDaysCallRatio BHcallRatio)/(BusyDaysP erM onth 60) CSV oiceBHCA = CSV oiceT raf f ic 3600/M HT IubCSV oiceT raf f ic = N odeBSubscriber CSV oiceT raf f ic T otalCSV oiceChannel = ErlbChs(IubCSV oiceT raf f ic, IubInterf aceBlocking ) IubCSV oiceBW = (1 + IP OH ) (IubCSV oiceChannel) (AM RAM CSV AF + AM RSM CSV AF ) (1 + SHO) IubCSV oiceBHCA = N odeBSubscriber CSV oiceBHCA 117

Iub CS Data Trac Calculations CSData64T raf f ic = (U serM onthlyCSDataT raf f ic BusyDaysCallRatio BHCallRatio)/(BusyDaysP erM onth 60) CSData64BHCA = CSData64T raf f ic 3600/M HT IubCSDataT raf f ic = N odeBSubscriber CSData64T raf f ic T otalCSDataChannel = ErlbChs(IubCSDataT raf f ic, IubInterf aceBlocking ) IubCSDataBW = (1 + IP OH ) (IubCSV oiceChannel) (CSDataServiceRate1 CSDAF ) (1 + SHO) IubCSDataBHCA = N odeBSubscriber CSData64BHCA

Iub PS Data Trac Calculations BHP SData64Rate = (U serM onthlyP S 64 BusyDaysCallRatio BHCallRatio 1024 8)/BusyDaysP erM onth BHP sData144Rate = (U serM onthlyP S 144BusyDaysCallRatioBHCallRatio 1024 8)/BusyDaysP erM onth BHP SData384Rate = (U serM onthlyP S 384BusyDaysCallRatioBHCallRatio 1024 8)/BusyDaysP erM onth IubP SData64BW = (1 + IP OH ) (BHP SData64Rate/3600) (1 + SHO) (1 + RT F ) P SAF IubP SData144BW = (1 + IP OH ) (BHP SData144Rate/3600) (1 + SHO) (1 + RT F ) P SAF 118

IubP SData384BW = (1 + IP OH ) (BHP SData384Rate/3600) (1 + SHO) (1 + RT F ) P SAF N odeBBHCA = IubCSV oiceBHCA + IubCSDataBHCA IubCSChannel = T otalCSV oiceChannel + T otalCSDataChannel IubCST otalBW = IubCSV oiceBW + IubCSDataBW N odeAttachedSubscriber = N odeBSubscriber P SAttachedSubscriberRatio N odeBSimultaneousP DP = N odeAttachedSubscriber P DP Durration ActiveBHP DP Ratio/3600 IubpsT otalBW = N odeBSimultaneousP DP (IubP sData64BW +IubP sData144BW + IubP SData384BW )/IubInterf aceU tilization CSSignalling = CSsignalingRatio N odeBBHCA (CSSignalingRatio = 50Kbps, includingSM S )

119

Erlangb-to-Channel Function
double ErlbChannel(double traffic erlang, double blocking probability) { double blocking, new blocking, channel, i; blocking = 1; i = 0; if(blocking probability >= 1 || blocking probability <= 0) { printf("Blocking probability must be between 1 and 0 !"); system("pause"); exit(1); } while(blocking > blocking probability){ i=i+1; new blocking = blocking; blocking = traffic erlang blocking/(traffic erlang blocking+i); }

channel = ceil(i -1+(blocking probability - new blocking)/(blocking - new blocking)); return channel; }

120

Appendix B
In this appendix, we present numerical results for the second, third and fourth instances of the problem set#1 and set#2. We also present the solution comparison in terms of cost and CPU for these instances. Table B.1: Numerical results for the second instance of set#1
CPLEX Problem No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Cost ($) 1, 462, 546 1, 523, 483 1, 788, 382 1, 832, 247 2, 052, 402 2, 375, 804 2, 463, 846 2, 749, 368 2, 975, 960 3, 144, 911 3, 158, 770 3, 296, 197 3, 366, 534 3, 467, 456 3, 583, 633 CPU (s) 212 64 3, 798 43 1893 237 419 1, 365 43, 856 26, 422 9, 598 12, 303 86, 401 85, 243 40, 838 Local Search Cost ($) 1, 474, 446 1, 559, 653 1, 810, 073 1, 922, 349 2, 193, 170 2, 433, 809 2, 579, 788 3, 028, 984 3, 113, 601 3, 375, 871 3, 377, 694 3, 508, 579 3, 648, 237 3, 716, 717 3, 831, 462 CPU (s) 196 281 376 456 542 622 676 834 822 876 1, 005 1, 165 1, 136 1, 215 1, 249 Gap (%) 0.82 2.38 1.22 4.92 6.86 2.45 4.71 10.18 4.63 7.35 6.94 6.45 8.37 7.19 6.92 Tabu Search Cost ($) 1, 462, 546 1, 523, 483 1, 788, 382 1, 922, 247 2, 144, 476 2, 416, 404 2, 532, 407 2, 892, 067 3, 083, 586 3, 271, 590 3, 276, 712 3, 454, 386 3, 59, 4750 3, 657, 303 3, 730, 874 CPU (s) 217 318 468 779 869 1, 107 1, 614 2, 010 1, 790 3, 533 3, 650 3, 627 5, 518 6, 257 5, 216 Gap (%) 0 0 0 4.92 4.49 1.71 2.79 5.2 3.62 4.03 3.74 4.8 6.78 5.48 4.11

121

4,000,000 3,500,000 3,000,000 Cost ($) 2,500,000 2,000,000 1,500,000 1,000,000

LS Cost ($) TS Cost ($) CPLEX Cost ($)

10 11 12 13 14 15

Problem number (set #1)

Figure B.1: Solution comparison for the second instance of set#1 (cost)

100,000 10,000 CPU time (s) 1,000

LS CPU (s)

100 10 1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Problem number (set #1)

TS CPU (s)

CPLEX CPU (s)

Figure B.2: Solution comparison for the second instance of set#1 (CPU)

122

Table B.2: Numerical results for the second instance of set#2

CPLEX Problem No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Cost ($) 1, 470, 998 1, 700, 094 1, 725, 619 1, 856, 107 2, 272, 955 2, 214, 235 2, 519, 670 2, 717, 701 2, 861, 509 3, 086, 501 3, 156, 552 3, 300, 188 3, 417, 571 3, 374, 091 3, 447, 617 CPU (s) 2, 472 252 82, 917 86, 401 1, 357 26, 498 86, 401 9, 391 16, 522 86, 401 35, 582 86, 401 86, 401 86, 401 86, 401 Local Search Cost ($) 1, 477, 890 1, 712, 577 1, 802, 183 1, 951, 955 2, 319, 753 2, 372, 665 2, 686, 698 2, 949, 973 3, 156, 674 3, 232, 497 3, 310, 931 3, 564, 358 3, 626, 591 3, 619, 818 3, 667, 700 CPU (s) 1, 374 2, 005 2, 645 3, 283 3, 687 4, 542 5, 005 5, 643 6, 148 8, 003 7, 344 7, 505 8, 722 9, 323 9, 996 Gap (%) 0.47 0.74 4.44 5.17 2.06 7.16 6.63 8.55 10.32 4.74 4 .9 8.01 6.12 7.29 6.39 Tabu Search Cost ($) 1, 473, 894 1, 700, 094 1, 725, 619 1, 948, 935 2, 313, 669 2, 226, 231 2, 652, 274 2, 867, 028 2, 940, 846 3, 176, 917 3, 241, 527 3, 478, 062 3, 560, 311 3, 516, 029 3, 639, 485 CPU (s) 1, 412 2, 117 2, 852 3, 639 4, 206 5, 316 6, 479 6, 988 8, 745 10, 811 10, 677 12, 218 12, 912 14, 304 19, 486 Gap (%) 0.2 0 0 5.01 1.8 0.55 5.27 5.5 2.78 2.93 2.7 5.39 4.18 4.21 5.57

123

4,000,000 3,500,000 3,000,000 Cost ($) 2,500,000 2,000,000 1,500,000 1,000,000

LS Cost ($) TS Cost ($) CPLEX Cost ($)

10 11 12 13 14 15

Problem number (set #2)

Figure B.3: Solution comparison for the second instance of set#2 (cost)

100,000 10,000 CPU time (s) 1,000

LS CPU (s)

100 10 1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Problem number (set #2)

TS CPU (s)

CPLEX CPU (s)

Figure B.4: Solution comparison for the second instance of set#2 (CPU)

124

Table B.3: Numerical results for the third instance of set#1

CPLEX Problem No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Cost ($) 1, 454, 159 1, 713, 522 1, 765, 694 1, 878, 338 2, 092, 727 2, 401, 390 2, 449, 102 2, 597, 390 3, 085, 255 3, 271, 414 3, 229, 246 3, 289, 069 3, 362, 035 3, 523, 300 3, 678, 604 CPU (s) 87 300 45 369 4, 894 26, 749 31 167 86, 400 23, 691 37, 409 26, 143 15, 970 62, 306 17, 050 Local Search Cost ($) 1, 458, 670 1, 748, 924 1, 791, 613 1, 998, 871 2, 223, 933 2, 489, 275 2, 541, 504 2, 703, 037 3, 225, 803 3, 395, 273 3, 498, 413 3, 474, 950 3, 709, 740 3, 905, 738 4, 094, 613 CPU (s) 199 281 371 472 546 622 706 803 865 920 1, 062 1, 038 1, 139 1, 306 1, 245 Gap (%) 0.32 2.07 1.47 6.42 6.27 3.66 3.78 4.07 4.56 3.79 8.34 5.66 10.35 10.86 11.31 Tabu Search Cost ($) 1, 454, 159 1, 713, 522 1, 767, 017 1, 969, 269 2, 172, 594 2, 450, 484 2, 526, 346 2, 660, 941 3, 198, 590 3, 358, 316 3, 388, 753 3, 444, 641 3, 593, 545 3, 707, 437 3, 954, 830 CPU (s) 220 320 413 639 741 961 1, 156 1, 633 1, 535 2, 391 3, 096 3, 564 4, 841 3, 889 4, 876 Gap (%) 0 0 0.08 4.85 3.82 2.05 3.16 2.45 3.68 2.66 4.94 4.73 6.89 5.23 7.51

125

4,500,000 4,000,000 3,500,000 Cost ($) 3,000,000 2,500,000 2,000,000 1,500,000 1,000,000
LS Cost ($) TS Cost ($) CPLEX Cost ($)

10 11 12 13 14 15

Problem number (set #1)

Figure B.5: Solution comparison for the third instance of set#1 (cost)

100,000 10,000 CPU time (s) 1,000

LS CPU (s)

100 10 1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Problem number (set #1)

TS CPU (s)

CPLEX CPU (s)

Figure B.6: Solution comparison for the third instance of set#1 (CPU)

126

Table B.4: Numerical results for the third instance of set#2

CPLEX Problem No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Cost ($) 1, 460, 621 1, 499, 760 1, 760, 691 1, 856, 922 2, 071, 852 2, 126, 368 2, 435, 825 2, 653, 473 3, 055, 835 3, 251, 984 3, 159, 905 3, 284, 133 3, 325, 781 3, 479, 229 3, 530, 957 CPU (s) 4, 987 2, 400 2, 271 55, 433 6, 694 17, 813 37, 054 76, 777 58, 726 86, 401 86, 401 79, 054 86, 401 86, 401 86, 401 Local Search Cost ($) 1, 468, 665 1, 511, 370 1, 798, 530 1, 999, 047 2, 193, 243 2, 178, 716 2, 549, 958 2, 862, 206 3, 214, 880 3, 466, 103 3, 436, 831 3, 472, 010 3, 539, 307 3, 839, 552 3, 902, 832 CPU (s) 1, 369 1, 999 2, 657 3, 306 3, 899 4, 501 5, 012 5, 685 6, 319 8, 011 7, 363 8, 108 8, 488 8, 803 9, 872 Gap (%) 0.56 0.78 2.15 7.66 5.86 2.47 4.69 7.87 5.21 6.59 8.77 5.73 6.43 10.36 10.54 Tabu Search Cost ($) 1, 460, 621 1, 499, 760 1, 763, 175 1, 950, 422 2, 148, 956 2, 166, 368 2, 449, 068 2, 784, 288 3, 123, 217 3, 294, 964 3, 360, 972 3, 428, 313 3, 462, 434 3, 698, 382 3, 857, 497 CPU (s) 1, 409 2, 046 2, 740 3, 530 4, 165 5, 039 5, 433 6, 551 7, 602 9, 133 9, 979 11, 266 13, 213 13, 651 22, 703 Gap (%) 0 0 0.15 5.04 3.73 1.89 0.55 4.93 2.21 1.33 6.37 4.4 4.11 6.3 9.25

127

4,500,000 4,000,000 3,500,000 Cost ($) 3,000,000 2,500,000 2,000,000 1,500,000 1,000,000
LS Cost ($) TS Cost ($) CPLEX Cost ($)

10 11 12 13 14 15

Problem number (set #2)

Figure B.7: Solution comparison for the third instance of set#2 (cost)

100,000 10,000

CPU time (s)

1,000
LS CPU (s)

100 10 1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Problem number (set #2)

TS CPU (s)

CPLEX CPU (s)

Figure B.8: Solution comparison for the third instance of set#2(CPU)

128

Table B.5: Numerical results for the fourth instance of set#1

CPLEX Problem No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Cost ($) 1, 454, 159 1, 713, 522 1, 765, 694 1, 878, 338 2, 092, 727 2, 401, 390 2, 449, 102 2, 597, 390 3, 085, 255 3, 271, 414 3, 229, 246 3, 289, 069 3, 362, 035 3, 523, 300 3, 678, 604 CPU (s) 87 300 45 369 4, 894 26, 749 31 167 86, 400 23, 691 37, 409 26, 143 15, 970 62, 306 17, 050 Local Search Cost ($) 1, 458, 670 1, 748, 924 1, 791, 613 1, 998, 871 2, 223, 933 2, 489, 275 2, 541, 504 2, 703, 037 3, 225, 803 3, 395, 273 3, 498, 413 3, 474, 950 3, 709, 740 3, 905, 738 4, 094, 613 CPU (s) 199 281 371 472 546 622 706 803 865 920 1, 062 1, 038 1, 139 1, 306 1, 245 Gap (%) 0.32 2.07 1.47 6.42 6.27 3.66 3.78 4.07 4.56 3.79 8.34 5.66 10.35 10.86 11.31 Tabu Search Cost ($) 1, 454, 159 1, 713, 522 1, 767, 017 1, 969, 269 2, 172, 594 2, 450, 484 2, 526, 346 2, 660, 941 3, 198, 590 3, 358, 316 3, 388, 753 3, 444, 641 3, 593, 545 3, 707, 437 3, 954, 830 CPU (s) 220 320 413 639 741 961 1, 156 1, 633 1, 535 2, 391 3, 096 3, 564 4, 841 3, 889 4, 876 Gap (%) 0 0 0.08 4.85 3.82 2.05 3.16 2.45 3.68 2.66 4.94 4.73 6.89 5.23 7.51

129

4,000,000 3,500,000 3,000,000 Cost ($) 2,500,000 2,000,000 1,500,000 1,000,000

LS Cost ($) TS Cost ($) CPLEX Cost ($)

10 11 12 13 14 15

Problem number (set #1)

Figure B.9: Solution comparison for the fourth instance of set#1 (cost)

100,000 10,000 CPU time (s) 1,000

LS CPU (s)

100 10 1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Problem number (set #1)

TS CPU (s)

CPLEX CPU (s)

Figure B.10: Solution comparison for the fourth instance of set#1 (CPU)

130

Table B.6: Numerical results for the fourth instance of set#2

CPLEX Problem No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Cost ($) 1, 460, 621 1, 499, 760 1, 760, 691 1, 856, 922 2, 071, 852 2, 126, 368 2, 435, 825 2, 653, 473 3, 055, 835 3, 251, 984 3, 159, 905 3, 284, 133 3, 325, 781 3, 479, 229 3, 530, 957 CPU (s) 4, 987 2, 400 2, 271 55, 433 6, 694 17, 813 37, 054 76, 777 58, 726 86, 401 86, 401 79, 054 86, 401 86, 401 86, 401 Local Search Cost ($) 1, 468, 665 1, 511, 370 1, 798, 530 1, 999, 047 2, 193, 243 2, 178, 716 2, 549, 958 2, 862, 206 3, 214, 880 3, 466, 103 3, 436, 831 3, 472, 010 3, 539, 307 3, 839, 552 3, 902, 832 CPU (s) 1, 369 1, 999 2, 657 3, 306 3, 899 4, 501 5, 012 5, 685 6, 319 8, 011 7, 363 8, 108 8, 488 8, 803 9, 872 Gap (%) 0.56 0.78 2.15 7.66 5.86 2.47 4.69 7.87 5.21 6.59 8.77 5.73 6.43 10.36 10.54 Tabu Search Cost ($) 1, 460, 621 1, 499, 760 1, 763, 175 1, 950, 422 2, 148, 956 2, 166, 368 2, 449, 068 2, 784, 288 3, 123, 217 3, 294, 964 3, 360, 972 3, 428, 313 3, 462, 434 3, 698, 382 3, 857, 497 CPU (s) 1, 409 2, 046 2, 740 3, 530 4, 165 5, 039 5, 433 6, 551 7, 602 9, 133 9, 979 11, 266 13, 213 13, 651 22, 703 Gap (%) 0 0 0.15 5.04 3.73 1.89 0.55 4.93 2.21 1.33 6.37 4.4 4.11 6.3 9.25

131

4,000,000 3,500,000 3,000,000 Cost ($) 2,500,000 2,000,000 1,500,000 1,000,000

LS Cost ($) TS Cost ($) CPLEX Cost ($)

10 11 12 13 14 15

Problem number (set #2)

Figure B.11: Solution comparison for the fourth instance of set#2 (cost)

100,000 10,000 CPU time (s) 1,000

LS CPU (s)

100 10 1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Problem number (set #2)

TS CPU (s)

CPLEX CPU (s)

Figure B.12: Solution comparison for the fourth instance of set#2 (CPU)

132

References
[1] ETSI, Overview of 3gpp release 4, summary of all release 4 features, ETSI, Tech. Rep., 2004. [2] T. Robinson Piliouras, Network Design: Management and Technical Perspectives. Boston, MA, USA: Auerbach Publications, 2004. [3] C. Smith and D. Collins, 3G Wireless Networks, 1st ed. McGraw-Hill TELECOM, 2008. [4] M. St-Hilaire, Topological planning and design of umts mobile networks: a survey, Wirel. Commun. Mob. Comput., vol. 9, pp. 948958, July 2009. [5] M. St-Hilaire, S. Chamberland, and S. Pierre, A local search heuristic for the global planning of umts networks, in Proceedings of the 2006 international conference on Wireless communications and mobile computing, ser. IWCMC 06. New York, NY, USA: ACM, 2006, pp. 14051410. [6] , A tabu search heuristic for the global planning of umts networks, in Proc. (WiMob2006) Wireless and Mobile Computing, Networking and Communications IEEE International Conference on, Jun. 1921, 2006, pp. 148151. 133

[7] ETSI, Overview of 3gpp release 99, summary of all release 99 features, ETSI, Tech. Rep., 2004. [8] A. R. Mishra, Ed., Fundamentals of Cellular Network Planning and Optimisation 2G/2.5G/3G... Evolution to 4G. John Wiley & Sons Ltd, 2004. [9] 3GPP, Universal mobile telecommunication system (umts): Qos concept and architecture, 3GPP, Tech. Rep. 23.107 v4.6.0, December 2002. [10] , Technical specication group service and systems aspects; architecture for an all ip network, 3GPP, Tech. Rep. 23.922 v1.0.0, October 1999. [11] V. Britvic, Steps in umts network design, in Proc. 12th IEEE Mediterranean Electrotechnical Conf. MELECON 2004, vol. 2, 2004, pp. 461464. [12] S. Allen, B. Belloul, S. Hurley, S. Saunders, and R. Whitaker, Smart cell planning and optimisation for umts, in 3G Mobile Communication Technologies, 2004. 3G 2004. Fifth IEE International Conference on, 2004, pp. 34 38. [13] K. Tutschku, K. Leibnitz, and P. Tran-Gia, Icept - an integrated cellular network planning tool, in In Proc. of the IEEE/VTS 47th Vehicular Technology Conf, 1997. [14] Netherlands Organisation For Applied Scientic Research, Tno telecom. [Online]. Available: http://momentum.zib.de/index.php [15] MENTUM. [Online]. Available: http://www.mentum.com [16] Forsk. [Online]. Available: http://www.forsk.com 134

[17] SYMENA. [Online]. Available: http://www.symena.com [18] University of British Colombia, Radio science laboratory. [Online]. Available: http://rsl.ece.ubc.ca/planning.html [19] C. Blum, M. J. B. Aguilera, A. Roli, and M. Sampels, Hybrid Metaheuristics: An Emerging Approach to Optimization, 1st ed. Springer Publishing Company, Incorporated, 2008. [20] T. H. Cormen, C. Stein, R. L. Rivest, and C. E. Leiserson, Introduction to Algorithms, 2nd ed. McGraw-Hill Higher Education, 2001. [21] IBM, IBM ILOG CPLEX V12.1, Users Manual for CPLEX, IBM, Inc., 2009. [22] M. J. Nawrocki, M. Dohler, and A. H. Aghvami, Eds., Understanding UMTS Radio Network Modelling, Planning and Automated Optimisation, Theory and Practice. John Wiley & Sons Ltd, 2006. [23] M. R. Garey and D. S. Johnson, Computers and Intractability; A Guide to the Theory of NP-Completeness. 1990. [24] ETSI, Overview of 3gpp release 5, summary of all release 5 features, ETSI, Tech. Rep., 2003. [25] 3GPP, Technical specication group radio access network; high speed downlink packet access (hsdpa) overall description (release 5), 3GPP, Tech. Rep. 25.308 v5.7.0, December 2004. New York, NY, USA: W. H. Freeman & Co.,

135

[26] ETSI, Overview of 3gpp release 6, summary of all release 6 features, ETSI, Tech. Rep., 2006. [27] 3GPP, Technical specication group radio access network; fdd enhanced uplink; overall description (release 6), 3GPP, Tech. Rep. 23.309 v6.6.0, March 2006. [28] , Technical specication group services and system aspects; network architecture; report on technical options and conclusions (release 7), 3GPP, Tech. Rep. 23.002 v7.6.0, December 2008. [29] , Technical specication group services and system aspects; 3gpp system architecture evolution; report on technical options and conclusions (release 8), 3GPP, Tech. Rep. 23.882 v8.0.0, September 2008. [30] U. Lauther, T. Winter, and M. Ziegelmann, Proximity graph based clustering algorithms for optimized planning of umts access network topologies, in Proc. 10th Int. Conf. Telecommunications ICT 2003, vol. 2, 2003, pp. 13291334. [31] T. Neubauer and M. Toeltsch, Umts radio network planning maximizing return on investment, e & i Elektrotechnik und Informationstechnik, vol. 122, pp. 108113, 2005. [32] H. Ramzi, Umts radio access network dimensioning, Alcatel Telecommunications Review, pp. 5563, 2001. [33] A. Bahai and H. Aghvami, Network planning and optimization in the third gen-

136

eration wireless networks, in 3G Mobile Communication Technologies, 2000. First International Conference on (Conf. Publ. No. 471), 2000, pp. 441 445. [34] R. Mathar and M. Schmeink, Optimal base station positioning and channel assignment for 3g mobile networks by integer programming, Annals of Operations Research, vol. 107, pp. 225236, 2001. [35] S. U. Thiel, P. Giuliani, L. J. Ibbetson, and D. Lister, An automated umts site selection tool, in Proc. Third International Conference on (Conf 3G Mobile Communication Technologies Publ. No. 489), May 810, 2002, pp. 6973. [36] Y. Wu and S. Pierre, Base station positioning in third generation mobile networks, in Proc. Canadian Conference on Electrical and Computer Engineering IEEE CCECE 2003, vol. 1, May 47, 2003, pp. 3134. [37] C. Maple, L. Guo, and J. Zhang, Parallel genetic algorithms for third generation mobile network planning, in Parallel Computing in Electrical Engineering, 2004. PARELEC 2004. International Conference on, 2004, pp. 229 236. [38] S. B. Jamaa, Z. Altman, J. M. Picard, and B. Fourestie, Multi-objective strategies for automatic cell planning of umts networks, in Proc. VTC 2004-Spring Vehicular Technology Conference 2004 IEEE 59th, vol. 4, May 1719, 2004, pp. 24202424. [39] Y. Choi, K. Kim, and N. Kim, The positioning of base station in wireless communication with genetic approach, in Personal Wireless Communications, ser. IFIP International Federation for Information Processing, R. Bestak, B. Simak, and E. Kozlowska, Eds. Springer Boston, 2007, vol. 245, pp. 217229. 137

[40] T. G. Crainic, B. D. Chiara, M. Nonato, and L. Tarricone, Tackling electrosmog in completely congured 3g networks by parallel cooperative metaheuristics, Wireless Communications, IEEE, vol. 13, no. 6, pp. 34 41, 2006. [41] E. Amaldi, A. Capone, and F. Malucelli, Planning umts base station location: optimization models with power control and algorithms, IEEE Transactions on wireless Communications, vol. 2, no. 5, pp. 939952, Sep. 2003. [42] J. Zimmermann, R. Hns, and H. Mhlenbein, Encon: an evolutionary algorithm for the antenna placement problem, Computers & Industrial Engineering, vol. 44, no. 2, pp. 209 226, 2003. [43] P. Gould, Radio planning of third generation networks in urban areas, in 3G Mobile Communication Technologies, 2002. Third International Conference on (Conf. Publ. No. 489), May 2002, pp. 64 68. [44] A. Neskovic, N. Neskovic, and G. Paunovic, Modern approaches in modeling of mobile radio systems propagation environment, Communications Surveys Tutorials, IEEE, vol. 3, no. 3, pp. 212, 2000. [45] Y. Okumura, E. Ohmori, T. Kawano, and K. Fukuda, Field strength and its variability in vhf and uhf land-mobile radio service, in Review of the Electrical Communications Laboratories 1968, vol. 16, pp. 825873. [46] M. Hata, Empirical formula for propagation loss in land mobile radio services, Vehicular Technology, IEEE Transactions on, vol. 29, no. 3, pp. 317 325, Aug. 1980.

138

[47] COST 231 Final Report, Digital mobile radio towards future generation systems, 1990. [Online]. Available: http://www.lx.it.pt/cost231 [48] J. Yang, M. Aydin, J. Zhang, and C. Maple, Umts base station location planning: a mathematical model and heuristic optimisation algorithms, Communications, IET, vol. 1, no. 5, pp. 1007 1014, 2007. [49] J. Laiho, A. Wacker, and T. Novosad, Radio Network Planning and Optimisation for UMTS. John Wiley & Sons, 2006. [50] H. Holma and A. Toskala, Eds., WCDMA for UMTS: Radio Access for Third Generation Mobile Communications, 1st ed. New York, NY, USA: John Wiley & Sons, Inc., 2001. [51] R. Zdunek and T. Ignor, Umts base station location planning with invasive weed optimization, in Proceedings of the 10th international conference on Artical intelligence and soft computing: Part II. Verlag, 2010, pp. 698705. [52] A. Mehrabian and C. Lucas, A novel numerical optimization algorithm inspired from weed colonization, Ecological Informatics, vol. 1, no. 4, pp. 355 366, 2006. [53] M. Galota, C. Glaber, S. Reith, and H. Vollmer, A polynomial-time approximation scheme for base station positioning in umts networks, in Proceedings of the 5th international workshop on Discrete algorithms and methods for mobile computing and communications. New York, NY, USA: ACM, 2001, pp. 5259. Berlin, Heidelberg: Springer-

139

[54] A. Eisenbl atter, A. F ugenschuh, T. Koch, A. Koster, A. Martin, T. Pfender, O. Wegel, and R. Wess aly, Modelling feasible network congurations for UMTS, in Telecummincations network design and management, G. Anandalingam and S. Raghavan, Eds. Kluwer, 2002, pp. 124. [55] B. Chamaret, S. Josselin, P. Kuonen, M. Pizarroso, B. Salas-Manzanedo, S. Ubeda, and D. Wagner, Radio network optimization with maximum independent set search, in Proc. IEEE 47th Vehicular Technology Conf, vol. 2, 1997, pp. 770774. [56] E. Amaldi, A. Capone, and F. Malucelli, Optimizing base station siting in umts networks, in Proc. VTC 2001 Spring Vehicular Technology Conference IEEE VTS 53rd, vol. 4, May 69, 2001, pp. 28282832. [57] , Discrete models and algorithms for the capacitated location problems arising in umts network planning, in Proceedings of the 5th international workshop on Discrete algorithms and methods for mobile computing and communications. New York, NY, USA: ACM, 2001, pp. 18. [58] , Improved models and algorithms for umts radio planning, in Proc. VTC 2001 Fall Vehicular Technology Conference IEEE VTS 54th, vol. 2, Oct. 711, 2001, pp. 920924. [59] , Optimizing umts radio coverage via base station conguration, in Proc. 13th IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, vol. 1, Sep. 1518, 2002, pp. 315319.

140

[60] E. Amaldi, A. Capone, F. Malucelli, and F. Signori, Umts radio planning: optimizing base station conguration, in Proc. VTC 2002-Fall Vehicular Technology Conference 2002 IEEE 56th, vol. 2, Sep. 2428, 2002, pp. 768772. [61] , Optimization models and algorithms for downlink umts radio planning, in Proc. IEEE Wireless Communications and Networking WCNC 2003, vol. 2, Mar. 2020, 2003, pp. 827831. [62] E. Amaldi, A. Capone, and F. Malucelli, Base station conguration and location problems in umts networks, in Proc. 9th IInternational Conference on Telecommunication Systems, Modeling and Analysi, 2001, pp. 341348. [63] E. Amaldi, A. Capone, F. Malucelli, and F. Signori, A mathematical programming approach for w-cdma radio planning with uplink and downlink constraints, in Proc. VTC 2003-Fall Vehicular Technology Conference 2003 IEEE 58th, vol. 2, Oct. 69, 2003, pp. 806810. [64] , Radio planning and optimization of w-cdma systems, in PersonalWireless Communications, ser. Lecture Notes in Computer Science. Springer Berlin Heidelberg, 2003, vol. 2775, pp. 437447. [65] S. B. Jamaa, Z. Altman, J.-M. Picard, B. Fouresti, and J. Mourlon, Manual and automatic design for umts networks, Mobile Networks and Applications, vol. 9, pp. 619626, 2004. [66] S. Jamaa, Z. Altman, J. Picard, and B. Fourestie, Combined coverage and capacity optimisation for umts networks, in Telecommunications Network Strat-

141

egy and Planning Symposium. NETWORKS 2004, 11th International, 2004, pp. 175 178. [67] S. Ben Jamaa, Z. Altman, J.-M. Picard, and A. Ortega, Steered optimization strategy for automatic cell planning of umts networks, in Vehicular Technology Conference, 2005. VTC 2005-Spring. 2005 IEEE 61st, vol. 1, 2005, pp. 359 362. [68] S. Sohn and G.-S. Jo, Optimization of base stations positioning in mobile networks, in Computational Science and Its Applications - ICCSA 2006, ser. Lecture Notes in Computer Science, M. Gavrilova, O. Gervasi, V. Kumar, C. Tan, D. Taniar, A. Lagan, Y. Mun, and H. Choo, Eds. Springer Berlin Heidelberg, 2006, vol. 3981, pp. 779787. [69] H. Lehpamer, Transmission Systems Design Handbook for Wireless Networks. Artech House, 2002. [70] J. Harmatos, A. Juttner, and A. Szentesi, Cost based umts transport network topology optimization, in Proc. International Conference on Computer Communication ICCC 1999, 1999. [71] J. Harmatos, A. Szentesi, and I. Godor, Planning of tree-topology umts terrestrial access networks, in Proc. 11th IEEE International Symposium on Personal, Indoor and Mobile Radio Communications PIMRC 2000, vol. 1, Sep. 1821, 2000, pp. 353357. [72] R. C. Prim, Shortest connection networks and some generalizations, Bell System Technology Journal, vol. 36, pp. 13891401, 1957. 142

[73] J. Kruskal, Joseph B., On the shortest spanning subtree of a graph and the traveling salesman problem, Proceedings of the American Mathematical Society, vol. 7, no. 1, pp. 4850, 1956. [74] I. G odor and G. Magyar, Cost-optimal topology planning of hierarchical access networks, Comput. Oper. Res., vol. 32, pp. 5986, January 2005. [75] A. Krendzel, Y. Koucheryavy, and J. Harju, Radio access network topology planning for the 4g networks, in Proc. 5th European Wireless Conference, Barcelona, Spain, 2004. [76] A. Jttner, A. Orbn, and Z. Fiala, Two new algorithms for umts access network topology design, EUROPEAN JOURNAL OF OPERATIONAL RESEARCH, vol. 164, pp. 456474, 2005. [77] Y. Wu and S. Pierre, Optimization of access network design in 3g networks, in Proc. Canadian Conference on Electrical and Computer Engineering IEEE CCECE 2003, vol. 2, May 47, 2003, pp. 781784. [78] , A new hybrid constraint-based approach for 3g network planning, IEEE Communications Letters, vol. 8, no. 5, pp. 277279, May 2004. [79] A. Quintero and S. Pierre, A memetic algorithm for assigning cells to switches in cellular mobile networks, Communications Letters, IEEE, vol. 6, no. 11, pp. 484 486, Nov. 2002. [80] S. Menon and R. Gupta, Assigning cells to switches in cellular networks by incorporating a pricing mechanism into simulated annealing, Systems, Man, 143

and Cybernetics, Part B: Cybernetics, IEEE Transactions on, vol. 34, no. 1, pp. 558 565, 2004. [81] A. Merchant and B. Sengupta, Assignment of cells to switches in pcs networks, IEEE/ACM Trans. Netw., vol. 3, pp. 521526, October 1995. [82] S. Pierre and F. Houeto, Assigning cells to switches in cellular mobile networks using taboo search, Systems, Man, and Cybernetics, Part B: Cybernetics, IEEE Transactions on, vol. 32, no. 3, pp. 351 356, Jun. 2002. [83] T. Bu, M. C. Chan, and R. Ramjee, Designing wireless radio access networks for third generation cellular networks, in INFOCOM 2005. 24th Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings IEEE, vol. 1, 2005, pp. 68 78 vol. 1. [84] C. Charnsripinyo, Topological design of 3g wireless access networks, in Proc. TENCON 2004. 2004 IEEE Region 10 Conference, 2004, pp. 565568. [85] D. Tipper, T. Dahlberg, H. Shin, and C. Charnsripinyo, Providing fault tolerance in wireless access networks, IEEE Communications Magazine, vol. 40, no. 1, pp. 5864, Jan. 2002. [86] C. Charnsripinyo and D. Tipper, Topological design of 3g wireless backhaul networks for service assurance, in Proc. 5th International Workshop on Design of Reliable Communication Networks (DRCN 2005), Oct. 1619, 2005, p. 9pp. [87] A. Szlovencsak, I. Godor, J. Harmatos, and T. Cinkler, Planning reliable umts

144

terrestrial access networks, IEEE Communications Magazine, vol. 40, no. 1, pp. 6672, Jan. 2002. [88] A. Krendzel, Y. Koucheryavy, and J. Harju, Cost and reliability estimation of radio access network structures for 4g systems, in Proc. VTC 2003-Fall Vehicular Technology Conference 2003 IEEE 58th, vol. 4, Oct. 69, 2003, pp. 22972301. [89] R. Shalak, K. Sandrasegaran, J. Agbinya, and S. Subenthiran, Umts core network planning model and comparison of vendor product performance, in Advanced Communication Technology, 2004. The 6th International Conference on, vol. 2, 2004, pp. 685 689. [90] Y. Ouyang and M. H. Fallah, Evolving core networks from gsm to umts r4 version, Int. J. Mob. Netw. Des. Innov., vol. 3, pp. 93102, January 2009. [91] F. Ricciato, R. Pilz, and E. Hasenleithner, Measurement-based optimization of a 3g core network: A case study, in Next Generation Teletrac and Wired/Wireless Advanced Networking, ser. Lecture Notes in Computer Science, Y. Koucheryavy, J. Harju, and V. Iversen, Eds. 2006, vol. 4003, pp. 7082. [92] J. Harmatos, Planning of umts core networks, in Proc. 13th IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, vol. 2, Sep. 1518, 2002, pp. 740744. [93] J. Zhang, L. Guo, and J. Wu, An integrated approach for utran planning and Springer Berlin Heidelberg,

145

optimization, in Vehicular Technology Conference, 2004. VTC 2004-Spring. 2004 IEEE 59th, vol. 4, May 2004, pp. 2360 2364. [94] S. Chamberland and S. Pierre, On the design problem of cellular wireless networks, Wirel. Netw., vol. 11, pp. 489496, July 2005. [95] S. Chamberland, An ecient heuristic for the expansion problem of cellular wireless networks, Comput. Oper. Res., vol. 31, pp. 17691791, October 2004. [96] M. St-Hilaire, S. Chamberland, and S. Pierre, Uplink umts network design an integrated approach, Computer Networks, vol. 50, pp. 27472761, 2006. [97] , A tabu search algorithm for the global planning problem of third generation mobile networks, Comput. Electr. Eng., vol. 34, pp. 470487, November 2008. [98] , Global expansion model for mobile networks, vol. 10, no. 6, pp. 453455, Jun. 2006. [99] R. Jonker and A. Volgenant, A shortest augmenting path algorithm for dense and sparse linear assignment problems, Computing, vol. 38, pp. 325340, November 1987. [100] E. W. Dijkstra, A note on two problems in connexion with graphs, Numerische Mathematik, vol. 1, no. 1, pp. 269271271, December 1959. [101] F. Glover and R. Marti, Tabu search, in Metaheuristic Procedures for Training Neutral Networks, ser. Operations Research/Computer Science Interfaces Series,

146

R. Sharda, S. Vo, E. Alba, and R. Mart, Eds. Springer US, 2006, vol. 36, pp. 5369. [102] J. Dro, P. Siarry, A. Ptrowski, E. Taillard, S. Jamaa, Z. Altman, J.-M. Picard, and B. Fouresti, Tabu search, in Metaheuristics for Hard Optimization. Springer Berlin Heidelberg, 2006, pp. 4773. [103] D. W. Corne, M. J. Oates, and G. D. Smith, Eds., Telecommunications optimization: Heuristics and adaptive techniques, 2nd ed. 2000. John Wiley & Sons,

147