Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Automated network selection for service delivery across all-IP heterogeneous wireless systems Bari, Farooq 2007

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
831-ubc_2007-317334.pdf [ 17.57MB ]
Metadata
JSON: 831-1.0065457.json
JSON-LD: 831-1.0065457-ld.json
RDF/XML (Pretty): 831-1.0065457-rdf.xml
RDF/JSON: 831-1.0065457-rdf.json
Turtle: 831-1.0065457-turtle.txt
N-Triples: 831-1.0065457-rdf-ntriples.txt
Original Record: 831-1.0065457-source.json
Full Text
831-1.0065457-fulltext.txt
Citation
831-1.0065457.ris

Full Text

Automated Network Selection for Service Delivery across All-IP Heterogeneous Wireless Systems by FAROOQ BARI M.Sc, Virginia Tech, Blacksburg, V A , USA, 1994 B.Sc, University of Engineering and Technology, Lahore, Pakistan, 1991 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY in THE F A C U L T Y OF GRADUATE STUDIES (ELECTRICAL AND COMPUTER ENGINEERING) THE UNIVERSITY OF BRITISH COLUMBIA August 2007 © Farooq Bari, 2007 Abstract A crucial step in achieving service delivery in heterogeneous all-IP wireless systems is the selection of an appropriate delivery network. Simple network selection mechanisms as they exist today cannot be applied in this new service environment. Because of the revenue and service impacting nature of the problem, it has been considered critical by the communication industry to find a good solution. In our research we have investigated the scope of the problem and addressed both the architectural and algorithmic aspects of the problem. We have proposed an architectural framework that is based on the network assisting the terminal in the decision about network selection. The proposed solution minimizes usage of the wireless links and would work with currently deployed infrastructure. It is scalable, flexible and supports roaming. We developed a two step decision process for a network-assisted network selection mechanism that combines non-compensatory and compensatory multi-attribute decision making (MADM) algorithms. We have proposed enhancements to several M A D M algorithms for their application to network selection. In order to deal with ranking abnormalities, we have proposed an improvement to the standard TOPSIS algorithm when it is applied to network selection by adopting an iterative approach. We have identified the need for support of non-monotonic utility for attributes used in network selection decision making and demonstrated that GRA is better suited for achieving this type of optimization objective. We have modified ELECTRE so that it can be used for network selection with attributes exhibiting non-monotonic utility. Our contributions in terms of modification to TOPSIS and ELECTRE and the usage of GRA with non-monotonic utility are not specific to the problem of network selection but these ideas can be used in other problems with similar requirements. We have developed a decision strategy for network selection in the absence of precise input information. The decision process proposed by us uses fuzzy techniques along with data prediction for the network selection decision process in certain cases. We have developed a new function, Confidence Level, and used it as a decision support tool along with the sensitivity of the service/subscription to the missing/imprecise information. ii Table of Content ABSTARCT ii TABLE OF CONTENT iii LIST OF TABLES vii LIST OF FIGURES viii LIST OF ACRONYMS AND SYMBOLS xi ACKNOWLEDGMENTS xiv CO-AUTHORSHIP STATEMENT xv 1 INTRODUCTION 1 1.1 PRIOR WORK 2 1.2 MOTIVES AND OBJECTIVES 6 1.3 MAIN CONTRIBUTIONS 7 1.4 ORGANIZATION OF THESIS 11 1.5 REFERENCES 13 2 ARCHITECTURAL FRAMEWORK FOR NETWORK SELECTION IN HETEROGENEOUS WIRELESS SYSTEMS 16 2.1 INTRODUCTION 16 2.2 FACTORS IMPACTING NETWORK SELECTION 18 2.2.1 Access Network Characteristics 18 2.2.2 Service Selection 18 2.2.3 Roaming Support 18 2.3 BACKGROUND AND RELATED WORK 18 2.3.1 Related Recent Standards Activities 19 2.4 PROPOSED ARCHITECTURE 21 2.4.1 Terminal Based Approach 21 2.4.2 Network Assisted Approach 22 2.4.3 Using DCN to Collect and Provide Access Network Characteristics 26 2.4.4 Using SAN for Service Selection 27 2.4.5 Using Enhanced AAA for Roaming Support 28 2.4.6 Information Collection and Exchange 29 iii 2.5 DEPLOYMENT SCENARIOS 31 2.5.1 DCN (IPF) and SAN residing in the A N 31 2.5.2 DCN (IPF) and SAN residing in local information provider network 32 2.5.3 DCN (IPF) and SAN residing in the HN 33 2.6 POLICY-BASED NETWORK SELECTION 35 2.6.1 Adding Standards-based Policy Nodes to Network Selection Framework 36 2.6.2 Example of Policy-based Automated Network Selection 38 2.7 CONCLUSION 41 2.8 REFERENCES 42 3 USE OF MADM ALGORITHMS FOR AUTOMATED NETWORK SELECTION IN A HETEROGENEOUS WIRELESS NETWORK ENVIRONMENT 44 3.1 INTRODUCTION 44 3.2 BACKGROUND AND RELATED WORK 45 3.3 DECISION PROCESS IN NETWORK SELECTION 46 3.4 MULTI ATTRIBUTE DECISION MAKING (MADM) 49 3.4.1 Non-Compensatory MADM Algorithms 49 3.4.1.1 Decreasing Inter-Technology Handoffs 50 3.4.2 Compensatory MADM Algorithms 51 3.4.2.1 Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) 52 3.4.2.2 Assignment of Attribute Weights 54 3.4.3 Results of Use Case Scenarios 56 3.4.4 Sensitivity of Network Selection to Dynamic Attribute Values 59 3.4.5 Network Selection with Managed IP Networks 60 3.4.6 Network Selection for Accessing Multiple Services Simultaneously 61 3.5 CONCLUSIONS 62 3.6 REFERENCES 63 4 ENHANCEMENTS TO MADM ALGORITHMS FOR USE IN AUTOMATED NETWORK SELECTION WHERE SYSTEM PARAMETERS ARE KNOWN 65 4.1 USE OF ITERATIVE TOPSIS FOR NETWORK SELECTION IN HETEROGENEOUS WIRELESS ACCESS 66 iv 4.1.1 TOPSIS 67 4.1.2 Ranking Abnormalities 69 4.1.2.1 Factors contributing to ranking abnormality for TOPSIS 70 4.1.3 Iterative M A D M Approach using TOPSIS 71 4.1.3.1 Results for use case scenario 73 4.2 U T I L I T Y A S P E C T S OF A T T R I B U T E S IN A P P L I C A T I O N OF M A D M A L G O R I T H M S TO N E T W O R K S E L E C T I O N 75 4.2.1 Deficiency in Current Studies 75 4.2.2 Use of Non-monotonic Utility for Attributes in Network Selection 76 4.2.3 Comparison of M A D M Algorithms for Use with Non-monotonic Utilities of Attributes 78 4.2.4 Theory of Grey Relational Space 80 4.2.5 Application of G R A Adapted to Network Selection with Non-monotonic Utility 80 4.2.6 Evaluation of Using Non-monotonic Utility in a Heterogeneous Wireless Network Environment ....83 4.2.6.1 Setting up GRA for Network Selection 84 4.3 A P P L I C A T I O N OF E L E C T R E TO N E T W O R K S E L E C T I O N IN H E T E R O G E N E O U S W I R E L E S S A C C E S S 88 4.3.1 Evaluating M A D M Algorithms for Use in Network Selection 88 4.3.1.1 Accuracy of the results obtained from an algorithm 88 4.3.1.2 Appropriateness of applying the algorithm to the problem 89 4.3.2 Application of Modified E L E C T R E to Network Selection 90 4.3.2.1 Approach 1 94 4.3.2.2 Approach 2 95 4.3.3 Evaluation of Modified E L E C T R E 96 4.4 C O N C L U S I O N 102 4.5 R E F E R E N C E S 104 5 APPLICATION OF D A T A PREDICTION A N D F U Z Z Y TECHNIQUES WITH M A D M - B A S E D A U T O M A T E D N E T W O R K SELECTION 106 5.1 INTRODUCTION 106 5.2 U S I N G N O N - C O M P E N S A T I N G M A D M A L G O R I T H M S WITH I N C O M P L E T E INFORMATION 107 5.3 U S I N G C O M P E N S A T I N G M A D M A L G O R I T H M S WITH D A T A PREDICTION A N D F U Z Z Y I N P U T 108 5.3.1 Scenario 1 - Imprecise or Missing Information 108 5.3.1.1 Data Estimation I l l v 5.3.1.2 Compensating MADM Algorithm 115 5.3.1.3 Denazification 119 5.3.1.4 Confidence Level 124 5.3.2 Scenario 2 - Network Types with Non Crisp Attributes 131 5.4 CONCLUSION -135 5.5 REFERENCES 136 6 CONCLUSIONS 138 6.1 SUMMARY OF CONTRIBUTIONS 138 6.2 FUTURE WORK 141 6.3 REFERENCES 143 vi List of Tables Table 2-1 Comparison of terminal-based and network-assisted mechanisms 24 Table 3-1 Attributes for use in network selection 48 Table 3-2 Attribute values for Scenarios 1, 2, 3 & 4 56 Table 3-3 Results on sensitivity analysis 59 Table 3-4 Attribute values for Scenarios 1, 2, 3 & 4 60 Table 4-1 Attribute values for the candidate networks 73 Table 4-2 Results of Iterative approach for TOPSIS 74 Table 4-3 Reference attribute values 85 Table 4-4 Assignment of attribute weights 85 Table 4-5 Attribute Values for Scenarios under Consideration 97 Table 4-6 Reference attribute values for voice over IP, streaming and web browsing services 97 Table 4-7 Ranking for networks using standard ELECTRE method with Alternative 2 98 Table 4-8 Ranking for networks using modified ELECTRE method with Alternative 2 98 Table 4-9 voice over IP service 99 Table 4-10 Steaming service 100 Table 4-11 Web browsing service ; 101 Table 5-1 Attributes and their weight assignments for streaming media and web browsing services while using GRA 115 Table 5-2 Attribute values for the five networks under consideration 117 Table 5-3 Lookup tables used by the decision maker for Confidence Level calculations 130 Table 5-4 Attribute values for network alternatives under consideration 132 vii List of Figures Figure 1.1 Organization of thesis 11 Figure 2.1 The problem of network selection , 17 Figure 2.2 Decision process in network selection using proposed network-assisted mechanism 25 Figure 2.3 A heterogeneous wireless system with access networks using different technologies 25 Figure 2.4 Logical nodes involved in network selection 26 Figure 2.5 Architecture for network related information collection and exchange 30 Figure 2.6 QoS attributes used for network selection, admission control, handoffs, packet scheduling can be the same but wil l have different update rate and the level of accuracy 30 Figure 2.7 D C N (IPF) and S A N residing in the A N 32 Figure 2.8 D C N (IPF) and S A N residing in local information provider's network 33 Figure 2.9 D C N (IPF) and S A N residing in H N 34 Figure 2.10 High level architecture for a heterogeneous wireless system with user's H N providing the services 37 Figure 2.11 Logical nodes involved in network selection 38 Figure 2.12 High level communication with policy components residing in the H N . The PDF is enhanced to include the functionality of decision making for network selection. The decision making algorithms that can be used are described in later chapters 39 Figure 2.13 Signaling and bearer paths once AN#2 has been selected. The bearer path is route optimized via a local breakout 40 Figure 3.1 Decision making process in network selection 49 Figure 3.2 By using non-compensatory M A D M algorithm, different mobility profile settings can provide selection of a different network 51 Figure 3.3 Example of indifference curve for TOPSIS while using attributes of cost per byte and throughput 54 Figure 3.4 Results for scenarios 1 & 3 57 Figure 3.5 Results for scenarios 2 & 4 58 Figure 4.1 Representation of an M A D M problem with n alternatives and m attributes 67 Figure 4.2 Use of TOPSIS for a M A D M problem with 4 alternatives and 2 attributes 70 Figure 4.3 Decision making process in Iterative TOPSIS approach 72 viii Figure 4.4 Assignment of weights to attribute values for a Bronze user subscription in the home operator network. 73 Figure 4.5 Graphical representation of a simple decision making scenario with one attribute and two networks 77 Figure 4.6 Three ways of using reference attribute values and attribute weights with G R A algorithm. Approach 3 is recommended in network selection 84 Figure 4.7 Results for network delection for three possible configurations of G R A . Configuration 3 is preferred as it provides maximum flexibility, to fine tuning the algorithm to optimization objectives 87 Figure 4.8 Weights associated with attributes for different services 97 Figure 5.1 Use of M A D M with imprecise information can help narrow down the candidate service delivery network options 107 Figure 5.2 Distributed nature of data collection points across different access types and autonomous domains makes harmonized, accurate and real time parameter information challenging 109 Figure 5.3 Steps involved in proposed network selection mechanism with imprecise information 110 Figure 5.4 Fuzzy Numbers 114 Figure 5.5 Example graphs of packet Delay (D), Jitter (J) and Latency (L) values using Network Utilization (U). It assumes correlation represented by D = 0.225*U2 +10, J = 0.05*U2, Z = 0.0019*C/ 3 with 0.8U - 1.1U in the network utilization range of 5% and 70% 116 Figure 5.6 Results for streaming media services using G R A algorithm with fuzzy input and with imprecise input information 118 Figure 5.7 Use of CoG for defuzzification of a T F N representing the G R C in a fuzzy implementation of G R A for streaming media services 121 Figure 5.8 Results of G R A algorithm for streaming media service after removing Ntwk#5 122 Figure 5.9 Results for web browsing service using G R A algorithm with fuzzy input information and imprecise input information 122 Figure 5.10 Use of CoG for defuzzification of a T F N representing the G R C in a fuzzy implementation of G R A for web browsing services 123 Figure 5.11 Trend graph for relationships between Confidence Level and common service types when a QoS related parameter with dynamic characteristics has unreliable value and therefore is being predicted 125 Figure 5.12 Trend graph for relationship between Confidence Level and reliability of estimated data for some of the common service types 125 Figure 5.13 Trend graph for relationship between Confidence Level and time since last attribute value was received 126 Figure 5.14 Trend graph for relationship between Confidence Level and degree of correlation between known and unknown attributes 126 Figure 5.15 A n alternative approach of using Confidence Level in the decision process 127 Figure 5.16 Graphical representation of how network congestion during service delivery time can be related to parameters such as the total bandwidth, bandwidth allowed per user, current network utilization and rate of change of network utilization 129 Figure 5.17 Results of fuzzy G R A algorithm 133 Figure 5.18 Use of CoG for defuzzification of T F N representing GRCs in the fuzzy implementation of G R A 134 List of Acronyms and Symbols 3GPP 3 r d Generat ion Partnership P rogram AAA Authent icat ion, Au thor i za t ion and A c c o u n t i n g AB A l l o w e d B a n d w i d t h AHP A n a l y t i c Hiera rchy Process AM Authent ica t ion M e c h a n i s m AN A c c e s s N e t w o r k AP A c c e s s Point AT A c c e s s Techno logy BAS B a s i c A c c e s s S igna l ing CA Coverage A r e a CB Cos t per B y t e CALEA Communica t ions Assis tance for L a w Enforcement A c t o f 1994 CoG Center o f Grav i t y CS C o m m a n d Services CSCF C a l l Sess ion C o n t r o l Func t ion D Packet D e l a y DCN Data C o l l e c t i o n N o d e DNS D o m a i n N a m e Server E911 Enhanced 911 (emergency service) EAP Extensible Authent ica t ion Pro toco l ELECTRE E l i m i n a t i o n and C h o i c e Translat ing Pr io r i ty ES Event Services GAS Gener ic Adver t isement Services GL Geographic L o c a t i o n GRA G r e y Rela t iona l A n a l y s i s GRC G r e y Rela t iona l Coeff ic ient GRS G r e y Rela t iona l Space GSM Groupe Spec ia l M o b i l e HN H o m e N e t w o r k HPLMN H o m e P u b l i c L a n d M o b i l e N e t w o r k ICF Information C o l l e c t i o n Func t ion xi IMS Internet Multimedia Subsystem IP Internet Protocol IPF Information Provider Function IS Information Services IEEE Institute of Electrical and Electronic Engineers IETF Internet Engineering Task Force IRTF Internet Research Task Force J Packet Jitter L Packet Loss LDAP Lightweight Directory Access Protocol MAC Medium Access Control MADM Multi Attribute Decision Making MIH Media Independent Handover NAI Network Access Identifier ON Operator Name PDF Policy Decision Function PHY Physical Layer PEF Policy Enforcement Function PLMN Public Land Mobile Network QoS Quality of Service RADIUS Remote Authentication Dial-In User Server/Service RFC Request For Comments SA Services Available SAN Service Announcement Node SAE System Architecture Evolution SAW Simple Additive Weighing SIM Subscriber Identification Module SLA Service Level Agreement SLP Service Location Protocol SSID Service Set Identifier TB Total Bandwidth TOPSIS Technique for Order Preference by Similarity to Ideal Solution U Network Utilization xii UE User Equipment UMTS Universal Mobile Telecommunication System VoIP Voice over IP VPLMN Visited Public Land Mobile Network WLAN Wireless Local Area Network WPM Weighed Product Model WSM Weighed Sum Model W A N Wireless Wide Area Network xiii Acknowledgments I would like to thank Dr. Victor Leung for his guidance during my stay at U B C . A s my graduate advisor he provided me with invaluable suggestions. His insightful comments while reviewing my work not only helped me in my research but also allowed me to improve my professional skills as a researcher. I would like to thank Professor Panos Nasiopoulos and Professor Son T. Vuong for serving on my dissertation committee. xiv Co-authorship Statement I a m the first author and pr inc ipa l contributor o f a l l manuscript chapters. A l l manuscript chapters are co-authored w i t h D r . V . C . M . L e u n g w h o supervised the thesis research. XV 1 Introduction In the past few decades several w ide area and loca l area wireless access technologies have emerged. W h i l e many o f these technologies have been successfully deployed and cont inued to evolve , none o f them provides a universa l coverage to cater to the mob i l e l ifestyle o f today 's users. The desire and expectation to have ubiquitous broadband connect ivi ty a l l the t ime and the emergence o f mul t imode devices are forc ing the network operators to l ook at new and innovat ive ways o f service de l ivery inc lud ing the poss ib i l i ty o f in te rworking o f these e v o l v i n g access technologies. A n example o f a mul t imode device is a terminal that supports Institute o f E lec t r i ca l and Elec t ronics Engineers ( I E E E ) 802.11 Wire less L o c a l A r e a N e t w o r k ( W L A N ) technology and Groupe Spec ia l M o b i l e ( G S M ) Wire less W i d e A r e a N e t w o r k ( W W A N ) technology. A major challenge i n this new environment is network selection, i.e., ident i fying the best suited service de l ivery network w h e n the user has mul t ip le networks to choose from. N e t w o r k selection becomes especial ly chal lenging for mul t imode devices that have the opt ion to use different a l l - IP wireless access technologies to gain access to various services. T o meet the challenge o f ubiquitous service del ivery, the network selection mechanism should also w o r k under roaming and w h e n there are more than one authentication mechanisms and credentials avai lable to access the avai lable networks. F o r example, a user m a y have mul t ip le accounts w i t h the same or different service providers, such as the user 's o w n personal and j o b related accounts. In order to get better coverage the user may also access different networks us ing more than one types o f authentication credential , e.g., Subscriber Identification M o d u l e ( S I M ) based or user ID/password . These credentials can map to the same or different accounts that can be accessed for de l iver ing a variety o f services over a number o f access networks w i t h different Qua l i t y o f Service ( Q o S ) capabil i t ies or condit ions. In the case o f current G S M systems for example , network select ion involves a scan for the ne twork identities, i.e., P u b l i c L a n d M o b i l e N e t w o r k IDs ( P L M N IDs) , f o l l o w e d b y a select ion o f one o f them based o n the pre-provis ioned informat ion i n the terminal about preferred P L M N IDs and forbidden P L M N IDs . In the case o f I E E E 802.11 W L A N s , the beacon and probe request/response mechan i sm provides a w a y for stations to d iscover A c c e s s Points ( A P s ) us ing Serv ice Set Identifiers (SSIDs) . B a s e d o n this informat ion and s ignal strength, the station can decide w h i c h A P to associate w i t h . Such s impl i s t ic approaches, however , are u n l i k e l y to y i e l d an op t ima l ne twork select ion i n in terworked heterogeneous a l l - IP systems that use different l access technologies and authentication mechanisms w h i l e de l ive r ing a range o f services from a var ie ty o f operators. T o de l iver a good customer experience, select ion o f an op t imal ne twork is essential. The dec i s ion b y the user / te rminal to select an A c c e s s N e t w o r k ( A N ) and attach to it can be either manual or automatic. In general, i n order to have a g o o d customer experience, automatic ne twork select ion mechanisms w o u l d be preferred w i t h manual select ion kept as a possible fal lback. M a n u a l network select ion is general ly quite s imple i n that it t yp i ca l ly comprises the user manua l ly p i c k i n g a network from a l ist o f networks v i s ib le to the terminal . M a n u a l select ion can fa i l for a var iety o f reasons such as the user 's home operator not hav ing roaming agreements w i t h the chosen network operator. A u t o m a t i c network select ion c o u l d be based on m u c h more informat ion that w o u l d a l l o w a more informed dec i s ion to be made. The type o f informat ion, its co l l ec t ion and its usage i n the dec i s ion process a l l are important aspects o f ongo ing research. The informat ion used can include the service be ing requested and the service supported, ne twork identi ty and its roaming agreements, authentication mechanisms supported, Q o S prov ided , cost per byte, etc. M a n y o f these informat ion elements are static w h i l e others are dynamic . It is also important to define an architecture w i t h details about the entities i n v o l v e d , the informat ion f lows , and the processes emp loyed for network selection. A s descr ibed later i n the thesis the poss ib i l i ty o f network-based, terminal-based and network-assisted network select ion exist. Depend ing u p o n the solu t ion proposed the informat ion m a y have to be col lected, ne twork select ion dec i s ion made and conveyed before the terminal gets authenticated. 1.1 Prior Work Arch i t ec tu ra l aspects o f network select ion for in terworked a l l - IP heterogeneous wire less networks operating i n autonomous domains have so far been a largely unexplored area o f research. In [1] , system leve l requirements such as the capabi l i ty to d iscover a l l the radio networks us ing different technologies avai lable i n an area have been discussed. A p r o o f o f concept system that uses B a s i c A c c e s s S igna l ing ( B A S ) to identify the radio technologies has been proposed for this purpose. The paper however does not consider realist ic scenarios based on current network deployments . F o r example , it ignores the fact that networks are t yp i ca l ly operated and managed b y autonomous business entities that can use different technology standards as w e l l as different authentication mechanisms, roaming arrangements, and b i l l i n g systems. The architecture also does not seem to be based on any current business m o d e l i n w h i c h 2 a user has a subscr ip t ion/b i l l ing arrangement w i t h his H o m e N e t w o r k ( H N ) operator for de l ive ry o f services that the user m a y want to access. Independent o f the architectural aspects, there have been some studies done i n the past for a lgor i thmic schemes to select the appropriate service de l ivery network. In [2], the use o f analyt ic hierarchy process w i t h G r e y system theory for network select ion between U n i v e r s a l M o b i l e Te lecommunica t ions Systems ( U M T S s ) and W L A N s has been proposed. In [3], ne twork select ion on a per c a l l basis based o n current market condi t ions such as the cost has been described. In [4], a resource a l locat ion scheme i n a heterogeneous environment has been proposed. A basic fuzzy log ic solut ion for access network select ion i n a heterogeneous ne twork ing environment has been descr ibed i n [5]. A s imple po l i cy-enab led handof f system across heterogeneous wireless networks is presented i n [6]. These papers however do not comple te ly address the complex p rob lem o f ne twork selection. Rather s impl i s t ic assumptions have been made due to a l ack o f understanding o f the scope o f the p rob lem and unava i lab i l i ty o f an architectural f ramework. The factors considered i n the dec i s ion processes are i n many cases insufficient; e.g., informat ion about access types supported, authentication types supported, roaming partners supported has not been considered i n the dec i s ion processes described i n these papers. The analysis o n the choice o f the dec i s ion m a k i n g a lgor i thm has been l im i t ed w h e n it comes to their sui tabi l i ty for appl ica t ion to the p rob lem o f ne twork select ion. T h e y also do not p rov ide or refer to a v iab le architectural f ramework i n w h i c h the select ion mechan i sm can w o r k and hence they fa i l to provide a complete and deployable solu t ion to the p rob lem. Aspec t s o f the p rob lem have been discussed i n var ious industry standardization forums. Here w e prov ide a snapshot o f the key related activit ies that have occurred i n the Internet Eng inee r ing Task Force ( I E T F ) , I E E E and 3 r d Genera t ion Partnership P rog ram ( 3 G P P ) forums. In I E T F , several W o r k i n g Groups have looked into the network select ion p rob lem. The I E T F w o r k has dealt m a i n l y w i t h the p rob lem from an Authent ica t ion , A u t h o r i z a t i o n and A c c o u n t i n g ( A A A ) perspective i n a homogeneous system. M u c h o f this w o r k was done before the entire scope o f the p r o b l e m was evident for today 's a l l - IP heterogeneous wireless ne twork ing environment . Nonetheless the w o r k done b y these groups is useful i n propos ing a comprehensive solu t ion to the p rob lem. F o r example , the R O A M O P S W G defined the N e t w o r k A c c e s s Identifier ( N A I ) o r ig ina l ly i n R F C 2486 [7], and subsequently updated it i n R F C 4282 [8]. N A I is useful i n p r o v i d i n g user and H N identity i n a standard and interoperable manner to the network entities. R F C 2194 [9] describes the A A A aspects o f some roaming implementat ions w h i l e R F C 2607 3 [10] describes the use o f proxies i n roaming . The C A R D pro toco l defined i n R F C 4066 [11] can assist i n d i scovery o f suitable A P s . R F C 3280 [12] developed b y the P u b l i c K e y Infrastructure X . 5 0 9 ( P K I X ) W G addresses issues o f certificate select ion for use i n authentication. The A A A W G also developed some service d i scovery mechanisms w i t h i n Diamete r ' s R F C 3588 [13]. In R F C 4284 [14] Extens ib le Authent ica t ion P ro toco l (EAP)-Reques t / Ident i ty has been used to p rov ide " r ea lm hints" useful for the select ion o f user identities. The technique was used b y 3 G P P i n advertisement o f W L A N s in te rwork ing w i t h ce l lu lar systems [15]. H o w e v e r , advertisement mechanisms based o n the use o f EAP-Reques t / Iden t i ty have been found to have sca labi l i ty problems w i t h a substantial negative impact o n the handof f latency i f such mechanisms are used for the select ion o f candidate networks dur ing handoffs [16]. I E E E 802 w o r k i n g groups have also w o r k e d o n the p rob lem o f network select ion for W L A N s that use the I E E E 802.11 technology. [17] describes the beacon and probe response mechanisms i n I E E E 802.11 based systems. W h i l e beacons can potent ial ly be used to advertise more than one networks , they m a y be sent on ly at a rate w i t h i n the base rate set, w h i c h typ i ca l ly consists o f the lowest supported rate. Studies [18] have shown M e d i u m A c c e s s C o n t r o l ( M A C ) layer performance problems w i t h this approach and [19] has ident i f ied sca l ing issues i f the beacon interval is lowered . [20] discusses h o w authentication models i n W L A N s can be migrated from exis t ing models to new ones that use E A P layer indicat ions or b y intel l igent use o f S S I D s proper ly structured to represent more than the l oca l network. It also ident i f ied sca labi l i ty issues w i t h v i r tua l A P s . [21] presents mechanisms currently used to provide v i r tua l A P capabil i t ies w i t h i n a s ingle phys i ca l A P that w o u l d a l l o w mul t ip le networks to be advert ised from one A P . A v i r tua l A P appears at the M A C and IP layers to be a dist inct phys i ca l A P . C o m p a t i b i l i t y w i t h ex is t ing 802.11 station implementat ions is possible i f each v i r tua l A P uses a dist inct M A C address ( B S S I D ) for use i n beacons and probe responses. H o w e v e r s imula t ion results presented i n [19] have h ighl ighted a scalabi l i ty p rob lem w i t h this approach w i t h a l i m i t o f approximate ly 20 v i r tua l A P s per phys i ca l A P . W o r k o n network select ion is ongoing i n I E E E 8021 l u and I E E E 802.21. The author o f this thesis has contributed to this w o r k [23][24] dur ing the research. M o r e detai l o n it is p rov ided i n Chapter 2 where an architectural framework for s o l v i n g the ne twork select ion p rob lem is described. 3 G P P ce l lu la r ne twork select ion, also k n o w n as P L M N selection, is described i n [25] where the terminal moni tors surrounding cel ls and pr ior i t izes them based o n s ignal strengths before select ing a new potential target ce l l . E a c h c e l l broadcasts its P L M N I D . A terminal , p rov i s ioned 4 w i t h a mapp ing o f operator to its P L M N I D , m a y automat ical ly select cel ls that be long to its H o m e P L M N ( H P L M N ) , Regis tered P L M N or an a l l o w e d set o f V i s i t e d P L M N s ( V P L M N s ) . F o r G S M systems, the P L M N I D lists are p r io r i t i zed and stored i n the S I M . F o r manua l P L M N select ion, the terminal col lects the P L M N IDs it learns from the surrounding cel ls and enables the user to choose the desired P L M N based on mapping o f these IDs to operator names. Af t e r the P L M N has been selected, c e l l pr ior i t iza t ion takes place, i n order to select the appropriate target c e l l . T S 23.234 [15] describes in te rworking o f W L A N s w i t h 2 n d Genera t ion ( 2 G ) and 3 r d Genera t ion ( 3 G ) ce l lu lar networks and discusses r ea lm d iscovery and network select ion issues i n that environment. In the W L A N roaming scenarios considered b y 3 G P P , mul t ip le ne twork levels m a y be present, and the hotspot owner m a y have a contract w i t h a provider w h o i n turn has a contract w i t h a 3 G network, w h i c h m a y have a roaming agreement w i t h other networks. The specif icat ion requires that network d i scovery be performed as specif ied i n the relevant W L A N l i n k layer standards. In addi t ion to network d iscovery , as part o f ne twork select ion it is necessary to select an intermediary roaming partner or broker. In 3 G P P , the intermediary networks are also assumed to be P L M N s , and it is assumed that an A N m a y have a roaming agreement w i t h more than one P L M N s . The P L M N m a y be a H P L M N or a V P L M N , where roaming is supported. G S M / U M T S roaming pr inc ip les are emp loyed for rout ing A A A requests from the V P L M N to the H P L M N us ing either R A D I U S or Diameter . The roaming partner ne twork related informat ion p rov ided v i a E A P methods b y the v i s i t ed ne twork as proposed i n R F C 4284 and used i n 3 G P P is o n l y considered a hint and typ ica l ly p rov ided w h e n a network considered to be unreachable v i a roaming agreements is encountered. The station can also manua l ly trigger a request for this informat ion b y inc lud ing an u n k n o w n rea lm ( k n o w n as the Al te rna t ive N A I ) w i t h i n the EAP-Response / Iden t i ty . A rea lm guaranteed not to be reachable w i t h i n 3 G P P networks is u t i l i zed for this purpose. The security requirements for P L M N select ion are discussed i n a 3 G P P contr ibut ion [26], w h i c h Concludes that both S S I D and E A P -based mechanisms have s imi la r security weaknesses. A s a result, it recommends that P L M N advertisements be considered on ly as hints. It is clear from the d iscuss ion above that w h i l e var ious standardization forums are t ry ing to address the p rob lem o f ne twork selection, the scope o f the w o r k and the solutions p rov ided are l im i t ed or restricted b y the mandate o f the forum. F o r example , 3 G P P has o n l y considered the use o f S I M based authentication and involvement o f P L M N operators for roaming scenarios. Other independent studies have also discussed the need for ne twork select ion where an A N m a y 5 have more than one roaming relationships to a H N ; none of them however provides a comprehensive approach towards solving the problem. [27] describes solutions to the realm selection problem based on E A P , SSID and PEAP-based mechanisms. [28] discusses the realm and network selection problem, concentrating primarily on the discovery of A N s meeting a set of criteria. 1.2 Motives and Objectives The objective of this research has been to provide solutions to key issues associated with the process of network selection for service delivery over a heterogeneous mix of all-IP wireless networks, possibly managed by more than one network operators. In this context, a key area of work was the development of a common architectural framework for network selection across heterogeneous wireless networks It is wel l known that as the number of factors influencing a decision increases from more than a few, it becomes difficult for humans to directly compare the alternatives and hence the need for decision tools. In the thesis, we have focused our research of decision making algorithms on the application of various Mul t i Attribute Decision Making ( M A D M ) algorithms, as they have been successfully applied to real life decision problems in fields such as operation research for several decades. Such algorithms are designed to take into consideration multiple input factors influencing the decision process, which is the case in the problem of network selection in an al l -IP heterogeneous networking environment. They provide a deterministic solution that can easily be combined with fuzzy techniques and parameter estimation in cases where reliable or crisp input data are not available. The key difference amongst M A D M algorithms is that they are based on different decision making philosophies. This allows the decision maker to choose the algorithm that is closest to its philosophy on decision making. In almost all the cases they are easy to understand and implement and simple enough to allow automated decision making in real time. We have avoided any discussion on a direct comparison between various algorithms discussed in the thesis on the basis of their accuracy of results since the issue has been studied in the past with the conclusion that such comparisons are not possible as they would need the use o f another M A D M algorithm which creates a paradoxical situation. However we have compared the M A D M algorithms in terms of their suitability to solving the problem of network selection. During our research the areas of focus related to algorithmic aspects of network selection were as follows; 6 • Deve lopmen t o f a comprehensive solu t ion for network select ion us ing M A D M algori thms. • M o d i f i c a t i o n o f M A D M algori thms for appl ica t ion to network select ion w h e n complete informat ion about the parameters impac t ing network select ion is avai lable . • Deve lopment o f a new dec i s ion tool incorporat ing a fuzzy M A D M a lgor i thm w i t h input parameter est imation for network select ion i n scenarios where complete informat ion about the input parameters is not k n o w n . Somet imes the p rob lem o f ne twork select ion i n a heterogeneous environment has been considered synonymous to the p rob lem o f inter- technology handoffs. A l t h o u g h the problems o f ne twork select ion and inter-technology handoffs are s imi la r i n some ways , the two have signif icant differences as w e l l . F o r example , inter- technology handoffs have a requirement to have session/service cont inui ty and therefore a requirement to have a l o w latency dec i s ion process and main ta in the same IP address dur ing a handoff. N o such requirement exists for ne twork select ion at in i t i a l start up o f the device , the p rob lem that is be ing addressed i n the thesis. A n inter- technology handoff w i t h service cont inui ty can result i n both business (e.g., h igher rate o f charging) and service (e.g., degradation o f Q o S ) impacts and therefore i n many scenarios this should be avo ided i f possible . One o f the aspects o f ne twork select ion as discussed i n Chapter 3 can therefore be to select a network that m in imize s inter- technology handoffs. It is possible to reuse aspects o f the network select ion solu t ion for inter- technology handoffs where they meet the requirements such as those related to latency. H o w e v e r the focus o f our research has been o n the p rob lem o f network select ion and the appl ica t ion o f the proposed solut ion and algori thms to inter- technology handoffs has not been analyzed. 1.3 Main Contributions T h e m a i n contributions o f the w o r k are: 7 A r c h i t e c t u r a l f r a m e w o r k The problem space related to the topic of the thesis has not been widely studied in the past and therefore is not well understood. We looked at the current state of the arts and developed requirements for a feasible architectural solution to the problem. In the process we contributed to the ongoing work 1 [1] in IETF for defining the broader problem space of network discovery and selection. While considering the architectural solution in our research, we kept in mind the design constraints identified by the IETF draft. A n architectural framework for network selection in a heterogeneous wireless system has been proposed by us in Chapter 2. It is based on the network assisting the terminal in the decision about network selection. The framework has been shown to work under different deployment scenarios. The proposed solution minimizes usage of the wireless links and would work with currently deployed infrastructure. It is scalable, flexible and supports roaming. The architecture proposed by us allows for information exchange between the terminal and the network both at layer 3 and at layer 2. For information exchange at layer 3, it leverages the concept of Information Services (IS) from I E E E 802.21. IS allows for a client in the terminal and an information server with a database in the network, a notion to which we also contributed [23] in 1 The author of this thesis was one of several co-authors to the referred industry standards submission and therefore does not claim authorship of the entire referred document or the ownership of all the ideas presented in it. See reference for the complete list of authors. 2 The author of this thesis was one of several co-authors to the referred industry standards submission and therefore does not claim authorship of the entire referred document or the ownership of all the ideas presented in it. See reference for the complete list of authors. 8 I E E E 802.21. Here w e extended this concept by def in ing specif ic service and network informat ion related nodes. W e then t ied these informat ion nodes to the current wireless infrastructure b y propos ing to enhance the functionali ty o f exis t ing p o l i c y dec i s ion nodes to inc lude ne twork select ion dec i s ion log ic as w e l l . F o r informat ion exchange at layer 2, our architecture uses as an example the w o r k done i n I E E E 802. l l u for network select ion to w h i c h w e also contr ibuted 1 [24]. W e have shown h o w such layer 2 solu t ion can be integrated i n our proposed architecture i n v o l v i n g an enhanced p o l i c y dec i s ion node and support ing informat ion nodes. Comprehensive MADM based network selection decision process A dec i s ion process for a network-assisted ne twork select ion mechan i sm that combines non-compensatory and compensatory M A D M algori thms is proposed. The two step mechan i sm is more comprehensive than w o r k publ i shed earlier. Scenarios w i t h different services and subscribed Q o S profi les have been used to demonstrate h o w the proposed mechan i sm w o u l d w o r k . A p p l i c a t i o n o f the proposed mechan i sm to networks w i t h mu l t i - l eve l Service L e v e l Agreements ( S L A s ) has been described. Poss ib le approaches to network select ion w h e n a service is already i n use have been discussed and solutions are proposed. The author of this thesis was one of several co-authors to the referred industry standards submission and therefore does not claim authorship of the entire referred document or the ownership of all the ideas presented in it. See reference for the complete list of authors. 9 Improvements to MADM algorithms for application to network selection when complete information about the factors impacting network selection is available W e have proposed an improvement to the standard Technique for Order Preference b y S i m i l a r i t y to Ideal So lu t ion ( T O P S I S ) a lgor i thm w h e n it is appl ied to network select ion by adopting an iterative approach. T o the best o f our knowledge this is the first such improvement o f the a lgor i thm for deal ing w i t h rank ing abnormali t ies w h e n complete rank ing is not required. W e have ident i f ied the need for support o f non-monotonic u t i l i ty for attributes used i n ne twork select ion dec i s ion m a k i n g . W e have shown that many o f the c o m m o n l y used M A D M algori thms i n their standard fo rm are not w e l l suited because o f assumptions about mono ton ica l ly increasing or decreasing ut i l i t ies o f the attributes. W e also show that G r e y Re la t iona l A n a l y s i s ( G R A ) is better suited for ach iev ing this type o f op t imiza t ion objective. A two step process has been proposed for tuning the a lgor i thm. T o the best o f our knowledge , this is the first k n o w n use o f non-monotonic u t i l i ty i n mu l t i attribute ne twork selection. W e have mod i f i ed E l i m i n a t i o n and C h o i c e Transla t ing Pr io r i ty ( E L E C T R E ) , another w e l l k n o w n M A D M a lgor i thm so that it can be n o w used for network select ion w i t h attributes exh ib i t ing a non-monotonic u t i l i ty . Th i s modi f ica t ion makes the a lgor i thm more adept to appl ica t ion i n ne twork select ion as it expands its app l icab i l i ty to a w i d e r range o f op t imiza t ion objectives. T o the best o f our knowledge this is the first k n o w n appl ica t ion o f E L E C T R E to ne twork select ion and first k n o w n modi f ica t ion o f E L E C T R E for handl ing non-monotonic attribute ut i l i t ies . O u r contributions i n terms o f modi f i ca t ion to T O P S I S and E L E C T R E and the usage o f G R A w i t h non-monotonic u t i l i ty are not specif ic to the p rob lem o f ne twork select ion but these ideas can be carr ied over to other p rob lem spaces w i t h s imi la r requirements. Application of data prediction and fuzzy techniques in MADM based network selection for scenarios when some of the information to be used in network selection is missing W e have developed a decis ion strategy for network selection i n the absence o f precise input information. The dec is ion process proposed b y us describes h o w fuzzy techniques a long w i t h data predic t ion can be useful i n network selection on ly i n certain cases. W e have developed a new function, Confidence Level, as an addi t ional decis ion support function and have used it a long w i t h the sensi t ivi ty o f the service/subscription to the miss ing/ imprecise information. W e have proposed that these two functions together can be appl ied to ascertain the usefulness o f estimating the data. 10 To the best of our knowledge this is the first application of functions such as sensitivity and Confidence Level to the problem of network selection with imprecise information. The algorithms presented in the thesis can be applied irrespective of the adoption of the architectural framework proposed in Chapter 2. In other words the algorithmic approaches, with some adjustments would be applicable even when the decision making entity applying the algorithms resides in the terminal. In that scenario however all the information has to be made available to the terminal. 1.4 Organization of Thesis Chapter 1 Introduction Chapter 2 Architectural Aspects of Network Selection Algorithmic Aspects of Network Selection Chapter 3 Comprehensive MADM Based Network Selection Process Chapter 4 Enhancements to MADM Based Network Selection when system parameters are known Chapter 5 Application of MADM Based Network Selection with imprecise system information Chapter 6 Conclusions Figure 1.1 Organization of thesis Figure 1.1 illustrates the overall organization of the thesis. The remainder of this thesis is organized as follows. In Chapter 2, we propose a framework for network selection in a heterogeneous wireless access system. After laying down the architectural framework, the remaining chapters focus on the algorithmic aspects with application of multi attribute decision making algorithms to the problem of network selection. Chapter 3 explores the application of a combination of compensatory and non-compensatory M A D M algorithms to provide a comprehensive solution to the problem o f decision making for network selection. 11 Chapter 4 provides enhancements to the M A D M a lgor i thm for appl ica t ion to ne twork select ion w h e n input informat ion to these algori thms is rel iable and avai lable . It suggests improvements to the standard T O P S I S a lgor i thm w h e n it is appl ied to network select ion b y adopting an iterative approach. Scenarios w i t h non-monotonic u t i l i ty for Q o S related attributes are discussed and it is s h o w n that amongst the algori thms considered, G R A is best suited for non-monotonic u t i l i ty . T h e chapter also describes the use o f E L E C T R E to network select ion and suggests modif ica t ions to it so as to make it better suited for use w i t h non-monotonic attribute u t i l i ty . Chapter 5 takes into considerat ion scenarios w h e n input to the M A D M algori thms is not comple te ly avai lable or trustworthy. It proposes the use o f data predic t ion and fuzzy M A D M techniques and then analyses the use o f such techniques w i t h G R A for the purpose o f ne twork select ion. It also proposes new dec i s ion support tools as part o f a comprehensive dec i s ion strategy for such scenarios. Chapter 6 concludes the thesis b y summar iz ing the w o r k done and ident i fy ing possible future research w o r k i n the area. 12 1.5 References [1] M . Inoue, K . M a h m u d , H . M u r a k a m i and M . Hasegawa, " M I R A I : A So lu t ion to Seamless A c c e s s i n Heterogeneous Wire less N e t w o r k s N e w Genera t ion" , Proc. IEEE ICC, pp. 1033-1037, M a y 2003. [2] Q . Song and A . Jamal ipour , " Q u a l i t y o f Service P r o v i s i o n i n g i n Wire less L A N / U M T S Integrated Systems us ing A n a l y t i c H ie ra rchy Process and G r e y Re la t iona l A n a l y s i s " , Proc. IEEE Globecom, pp. 220-224, November /December 2004. [3] G . L e B o d i c , D . G i r m a , J . B i n e and J . D u n l o p , " D y n a m i c 3 G N e t w o r k Se lec t ion for Increasing the Compe t i t i on i n the M o b i l e Communica t ions M a r k e t " , Proc. IEEE VTS-Fall, pp. 1064-1071, September 2000. [4] E . A l e x a n d r i , G . M a r t i n e z and D . Zeghlache , " A d a p t i v e Joint C a l l A d m i s s i o n C o n t r o l and A c c e s s N e t w o r k Selec t ion for M u l t i m e d i a Wire less Systems", Proc. 5th International Symposium on Wireless Personal Multimedia Communications, pp. 1390-1394, October 2002. [5] S. H o n g v a n , H . C h e n and J . L i n g g e , "Intell igent s ignal processing o f m o b i l i t y management for heterogeneous networks" , Proc. International Conf. on Neural Networks and Signal Processing, pp. 187-191, December 2003. [6] H . J . W a n g , R . H . K a t z and J . Giese , " P o l i c y - E n a b l e d Handoffs A c r o s s Heterogeneous Wire le s s N e t w o r k s " , Proc. IEEE Mobile Computing Systems and Applications, pp. 51-60, February 1999. [7] B . A b o b a and M . Beadles , "The N e t w o r k A c c e s s Identifier", R F C 2486, January 1999, [online] A v a i l a b l e ht tp: / /www.ietf .org. [8] B . A b o b a , M . Beadles , J . A r k k o , and P . Eronen , "The N e t w o r k A c c e s s Identifier", R F C 4282 , December 2005, [online] A v a i l a b l e ht tp: / /www.ietf .org. [9] B . A b o b a , J . L u , J . A l s o p , J . D i n g , and W . W a n g , " R e v i e w o f R o a m i n g Implementat ions", R F C 2194, September 1997, [online] A v a i l a b l e ht tp: / /www.ietf .org. [10] B . A b o b a and J . Vo l lb r ech t , " P r o x y C h a i n i n g and P o l i c y Implementat ion i n R o a m i n g " , R F C 2607, June 1999, [online] A v a i l a b l e ht tp: / /www.ietf .org. [11] M . L i e b s c h , A . S ingh , H . Chaskar , D . Funato, and E . S h i m , "Candidate A c c e s s Router D i s c o v e r y ( C A R D ) " , R F C 4066, Ju ly 2005, [online] A v a i l a b l e ht tp: / /www.iet f .org. [12] R . Hous l ey , W . P o l k , W . F o r d , and D . So lo , "Internet X . 5 0 9 P u b l i c K e y Infrastructure Cert i f icate and Cert if icate Revoca t ion L i s t ( C R L ) Prof i l e" , R F C 3280, A p r i l 2002, [online] A v a i l a b l e ht tp: / /www.ietf .org. [13] P . C a l h o u n , J . Loughney , E . Gut tman , G . Z o r n , and J . A r k k o , "Diameter Base P ro toco l " , R F C 3588, September 2003 , [online] A v a i l a b l e ht tp: / /www.ietf .org. [14] F . A d r a n g i , V . L o r t z , F . B a r i , and P . E ronen , "Identity Selec t ion Hin t s for the Extens ib le Authent ica t ion P ro toco l ( E A P ) " , R F C 4284, January 2006, [online] A v a i l a b l e ht tp: / /www.iet f .org. [15] 3 G P P , " 3 G P P Sys tem to Wire less L o c a l A r e a N e t w o r k ( W L A N ) in te rwork ing ; Sys tem Desc r ip t ion ; Release 6; Stage 2", 3 GPP Technical Specification 23.234 v 6.6.0, September 2005, [online] A v a i l a b l e ht tp : / /www.3gpp.org. [16] P . E ronen , "Ne twork Selec t ion Issues", presentation to E A P W G at I E T F 58, N o v e m b e r 2003. [17] Institute o f E l ec t r i c a l and Elec t ronics Engineers , "Wire less L A N M e d i u m A c c e s s C o n t r o l ( M A C ) and P h y s i c a l L a y e r ( P H Y ) Specif icat ions" , IEEE Standard 802.11, 2003 . [18] M . Heusse, "Performance A n o m a l y o f 802.11b", Proc. IEEE Infocom, pp. 836-843, M a r c h / A p r i l 2003 . [19] H . V e l a y o s and G . K a r l s s o n , "Techniques to Reduce I E E E 802.1 l b M A C L a y e r Handove r T i m e " , Labora tory for C o m m u n i c a t i o n Ne tworks , K T H , R o y a l Institute o f Techno logy , S t o c k h o l m , Sweden , TRITA-IMT-LCNR 03:02, A p r i l 2003 . [20] E . Hepwor th , "Co-exis tence o f Different Authent ica t ion M o d e l s " , IEEE Contribution 11-03-0827, 2003 . [21] B . A b o b a , " V i r t u a l A c c e s s Points" , IEEE Contribution ll-03-154rl, M a y 2003. [22] J . A r k k o , B . A b b o b a , J . K o r h o n e n and F . B a r i , " N e t w o r k D i s c o v e r y and Selec t ion P r o b l e m " , draft-ietf-eap-netsel-problem-08, June 2007 (completed w o r k i n g group last ca l l ) , [online] A v a i l a b l e ht tp: / /www.ietf .org. 14 [23] S. Das , Y . O h b a and F . B a r i , " Informat ion Service (IS) Reference M o d e l , U s e Case Scenario and H i g h e r L a y e r Requirements for 802.21 Information Service (IS)" , IEEE Contribution 21-05-0348-00-0000-Higher_layer_Requirements_Information_Service.doc, September 2005, [online] A v a i a l b l e ht tp: / /www.ieee802.org/21/ . [24] N . Canopolat , V . Gupta , D . Stephenson, S. Sreemanthula, R . Y . K i m , E . Hepwor th , S. D e m e l , M . Rudo l f , F . B a r i , H . C h e n g , L . M o , Z . Y a o , A . Soomro , and D . Cava lcan t i , " N e t w o r k Selec t ion - Norma t ive text proposa l" , IEEE Contribution 11-06-1014-03-000u-network-selection-normative-text-doc, September 2006, [online] A v a i l a b l e ftp.802wirelessworld.com. [25] 3 G P P , "Non-Access -S t r a tum ( N A S ) functions related to M o b i l e Stat ion ( M S ) i n idle mode" , 3GPP TS 23.122 6.5.0, October 2005, [online] A v a i l a b l e h t tp : / /www.3gpp.org . [26] 3 G P P , " 3 G P P Techn i ca l Speci f ica t ion G r o u p Service and Sys tem Aspec t s ; 3 G Securi ty; Wi re le s s L o c a l A r e a N e t w o r k ( W L A N ) in te rwork ing security (Release 6); Stage 2", 3GPP Technical Specification 33.234 v 6.6.0, October 2005, [online] A v a i l a b l e h t tp : / /www.3gpp.org . [27] Intel, "Wire less L A N ( W L A N ) E n d to E n d Guide l ines for Enterprises and P u b l i c Hotspot Serv ice Providers" , N o v e m b e r 2003. [28] R . E i j k , J . B r o k , J . B e m m e l , and B . Busropan , "Access N e t w o r k Se lec t ion i n a 4 G Env i ronmen t and the R o l e o f T e r m i n a l and Service P la t fo rm" , Proc. 10th Wireless World Research Forum, October 2003. 15 2 Architectural Framework for Network Selection in Heterogeneous Wireless Systems 1 2.1 Introduction The process of network selection in today's wireless systems is quite simple. In the case of a cellular system, e.g., G S M , network selection can involve a scan by the terminal for the P L M N IDs of available networks followed by a selection of one of them based on already provisioned information in the terminal about the preferred and forbidden P L M N IDs. There are several reasons that such a simple selection process works well . The search and hence the scanning for the best network is limited to only one type of access technology. The roaming agreements are limited to operators of one type of access technology. Therefore support for only one type of authentication credentials is needed and bill ing infrastructure across the roaming partners is also aligned. The most common service to be accessed is circuit-switched voice. Unlike packet-switched all-IP networks, the issues related to QoS do not arise in circuit-switched networks. These simplifying assumptions made for network selection in today's cellular systems would not be possible when delivering services over a heterogeneous mix of all-IP wireless access technologies for nomadic and mobile users. ' Parts of this chapter have been published in the following F. Bari and V . Leung, "Service Delivery over Heterogeneous Wireless Systems : Networks Selection Aspects", Proc. ACMIWCMC 2006, July 2006 F. Bari and V . Leung, "Architectural aspects of automated netowrk selection in heterogeneous wirelss systems", Submitted to International Journal of Adhoc and Ubiquitous Computing (IJAHUC) 16 A s an example, consider the situation shown in Figure 2 .1 , where a user is in a convention center served by multiple W L A N s and W W A N s . The user terminal supports both W L A N and W W A N access, and the user decides to access a streaming video service provided by his H N to watch the news. Currently, there is no automated way for the user terminal to intelligently select the most appropriate network from a heterogeneous mix of wireless access technologies for accessing the desired service. Such a network selection mechanism is essential to providing optimized service delivery along with a good service experience. The remaining part of this section highlights factors that can have an impact in network selection for a heterogeneous wireless IP networking environment. The following sections describe related work, and present the new architecture along with deployment scenarios for the architectural nodes. The initial deployment scenarios do not include a network-based decision making entity that can coordinate the information from different information resources and hence require for the terminal to directly obtain the processed information separately from different information sources such as those related to services and network conditions. In the final policy-based network selection example we describe how the role of an existing cellular network entity can be enhanced to make network selection decisions. This model further improves the decision process by centralizing it and decreasing the communication overhead with the terminal. In the following chapters, we assume the usage of the architectural model described in the policy-based network selection scenario. WWAN 1 F i g u r e 2.1 T h e p r o b l e m o f n e t w o r k se lec t ion 17 2.2 Factors Impacting Network Selection 2.2.1 Access Network Characteristics N e t w o r k capabil i t ies (e.g., total bandwidth) and current ne twork condi t ions (e.g., ne twork u t i l i za t ion and traffic load) can p lay a significant role i n the ne twork select ion process. Different wire less access technologies can have different capabil i t ies . E v e n for the same access technology, different capabil i t ies can exist. F o r example , i n the case o f I E E E 802.11 W L A N s , 802.11a has 54 M b p s capaci ty w h i l e 802.11b has 11 M b p s capaci ty and they also operate at different frequencies and therefore have different requirements o n the terminal . 2.2.2 Service Selection U n l i k e c i rcu i t - swi tched services such as tradit ional vo ice services, services de l ivered over an a l l -IP network, such as vo ice over IP ( V o I P ) , w i l l be s trongly inf luenced by the ne twork condi t ions and capabil i t ies . The in i t i a l select ion o f the wireless A N b y the terminal can therefore be impacted by , e.g., the Q o S requirements for the service that the user has dec ided to use. In addi t ion, different wireless A N s m a y support different sets o f w e l l defined IP services (e.g., V o I P based on IP M u l t i m e d i a Subsystem ( I M S ) ) at the same geographic loca t ion and a user dec i s ion to use a different service at a later stage o f network attachment can result i n the current ne twork becoming sub-opt imal or not avai lable . F o r example i f the in i t i a l l y selected ne twork does not have an appropriate l eve l o f Q o S or cannot support vo ice for regulatory reasons, then the user 's dec i s ion to make a vo ice c a l l m a y trigger a ne twork re-selection process. 2.2.3 Roaming Support The terminal should be able to get service over its H N or the most op t ima l roaming partner network. R o a m i n g support is considered essential for ubiqui tous service de l ive ry as no single ne twork operator can afford the infrastructure to p rov ide g loba l coverage o f its customers. A k e y benefit o f roaming support is the ab i l i ty to share revenue among the partner networks w h i l e p r o v i d i n g a better coverage to customers. 2.3 Background and Related Work W h i l e there has been w o r k done o n i nd iv idua l aspects o f heterogeneous wireless networks such as the ne twork select ion algori thms, lit t le has been publ i shed o n system leve l issues that w o u l d c a l l for an architectural solut ion for in te rwork ing o f heterogeneous wireless networks operating 18 in autonomous domains. However, [1] does provide such details at the systems level by laying out the requirements for seamlessly accessing a communication system that has a heterogeneous mix of wireless access technologies. One of the key requirements the paper defines is the ability to discover all the radio networks available in an area. To meet this requirement, the paper describes a proof of concept system that uses B A S to identify all the radio technologies. B A S runs over a basic A N , which can employ either a current radio technology or a new one. The approach taken in the paper leads to significant shortcomings in the architecture. For example, the paper does not take into account the fact that networks are typically operated and managed by autonomous business entities, and hence fails to consider the need for supporting multiple authentication mechanisms over roaming interfaces with users having different security/ authentication credentials. The architecture also does not include the concept o f a user having a H N where the user is subscribed to a set of well defined home based services with different QoS requirements. These are services that the subscriber may want to access from different wireless access technologies owned and operated by autonomous business entities. 2.3.1 Related Recent Standards Activities The protocols and mechanisms being developed in several industry forums can possibly be used in the network selection architecture proposed here. The I E E E 802.21 [2] group has been working on the problem of inter-technology handoffs. N e w work has been proposed in the IETF Mobi l i ty for IP: Performance, Signaling and Handoff Optimization (MIPSHOP) [3] working group to carry out network layer protocol work to transport I E E E 802.21 related information. The IETF Protocol for carrying Authentication for Network Access ( P A N A ) [4] group has been working for sometime now to develop a network-layer authentication protocol that can support various authentication methods, dynamic service provider selection, and roaming clients. The Internet Research Task Force (IRTF) Mobi l i ty Optimizations ( M O B O P T S ) [5] group is also looking into issues related to optimization of mobility across heterogeneous networks. Ongoing work in the IETF Next Steps in Signaling (NSIS) [6] group is performing some of the QoS mapping related work across access technologies. 3GPP [7] is currently working on an evolution of an architecture that could support multiple radio technologies. IETF draft [8] provides a good description of several aspects of the problem space that need resolution and provides the requirements for an acceptable solution. The I E E E 802.1 l u [9] task group is currently working on solutions that could amend the I E E E specifications so that I E E E 802.11 based networks could 19 better in terwork w i t h other systems. B e l o w w e describe the relationships o f some o f the standards related activit ies to our proposed architecture. A n important aspect o f our architectural f ramework is that it requires the terminal to in i t i a l l y obtain ( layer 2 or layer 3) network connect iv i ty , w i t h or wi thout authentication, i n order to query p o l i c y related components. T h i s in i t i a l connect ion is not used for accessing a subscr ibed service, but to exchange informat ion related to the ne twork select ion process. The network used for this search m a y not be the most appropriate network i n terms o f network capabil i t ies and the user services supported. T h i s query can be v i e w e d as s imi la r to us ing a directory assistance service. A s imi l a r approach is currently under d iscuss ion i n I E E E 802.21, the M e d i a Independent H a n d o f f ( M I H ) w o r k i n g group, w h i c h is deve lop ing standards for assist ing handoffs between heterogeneous networks. The group is def in ing three types o f informat ion that can be shared across different networks: Even t Services ( E S ) , C o m m a n d Services ( C S ) , and Information Serv ice (IS). IS provides less t ime c r i t i ca l informat ion, E S provides informat ion needed i n real-t ime for handover decisions across different network types w h i l e C S provides commands for the actual handover. The standards a l l o w a l l three types o f informat ion to be de l ivered through either l ink - l aye r specif ic solutions or through a " layer 3 or above" pro tocol . A s part o f our w o r k w e also contr ibuted i n I E E E 802.21 to the not ion o f an IS w i t h a terminal-based cl ient be ing able to query at layer 3 a network-based informat ion server connected to a database [20]. W e have used this concept i n our f ramework for services and network related informat ion nodes. The I E T F M I P S H O P w o r k i n g group [3] is def in ing the de l ivery o f M I H informat ion at layer 3, w h i c h enables M I H services even i n the absence o f l ink- layer support, w h i l e the I E E E 802. l l u w o r k i n g group is deve lop ing layer 2 standards for in te rwork ing o f W L A N s w i t h external networks such as ce l lu lar systems, i nc lud ing the ab i l i ty o f a terminal to d iscover and select a ne twork before associat ing to a W L A N A P . The I E E E 8 0 2 . l l u effort is focused on the support o f M I H i n W L A N s . The current draft standard proposes the use o f a Gener ic Adver t i sement Service ( G A S ) funct ional i ty to query ne twork related informat ion. It specifies the transport mechanisms for A P s to advertise services for ne twork select ion or other purposes to W L A N clients i n the un-associated state. 802.1 l u G A S capabi l i ty is inc luded i n the 802.11 beacon and probe response frames. A l t h o u g h G A S transport is transparent to the A d v e r t i s i n g P ro toco l used for queries and query responses and is f lex ib le enough to support mul t ip le protocols , at this t ime o n l y 802.21 is deve lop ing such a 20 protocol , but for the purpose o f support ing ne twork select ion for handover decis ions rather than in i t i a l service access. A s 8 0 2 . l l u has a requirement to support 802.21 based mechan i sm for ne twork select ion i n W L A N s , it is o f interest to leverage this capabi l i ty for ne twork select ion at in i t i a l service access. F o l l o w i n g the current 8 0 2 . l l u draft proposal , terminals obtain the G A S informat ion, par t icular ly the identi ty o f the supported A d v e r t i s i n g P ro toco l , f rom beacons broadcast b y 802.1 l u compl iant A P s or b y exchanging probe response messages w i t h such A P s . U s i n g the specif ied A d v e r t i s i n g P ro toco l , the terminal sends a G A S - I n i t i a l - R e q u e s t w i t h a specif ic query to the G A S - a w a r e A P , w h i c h then forwards the request to a back-end G A S p o l i c y server to retrieve the query results. The response message from the p o l i c y server is returned to the terminal v i a the A P us ing the A d v e r t i s i n g Pro toco l . W e were one o f the contributors to the proposal for this layer 2 based network select ion proposal i n I E E E 802.1 l u that uses G A S [21]. In this Chapter w e demonstrate h o w the mechan i sm as proposed i n I E E E 8 0 2 . l l u can be used w i t h i n our framework. 2.4 Proposed Architecture The p r o b l e m o f network select ion i n a heterogeneous wireless ne twork ing environment can be so lved i n two ways . 2.4.1 Terminal Based Approach In this approach, the terminal i t se l f first discovers a l l the avai lable A N s . It then collects from the networks i n d i v i d u a l l y the informat ion necessary for m a k i n g a dec i s ion and chooses the most appropriate A N after ana lyz ing the col lec ted information. T h i s approach is used today i n homogeneous wire less networks quite successfully. In such systems, there are o n l y a l im i t ed set o f frequency ranges for the terminal to scan, a l l the systems have typ i ca l ly s imi l a r capabil i t ies , a s ingle authentication mechan i sm is un i fo rmly supported b y everyone (e.g., S I M for G S M systems), and the roaming agreements are l im i t ed to operators us ing the same technology. In the case o f G S M networks, the network select ion process can s i m p l y be a table lookup compar ing a l is t o f preferred P L M N IDs w i t h the broadcast P L M N IDs rece ived b y the terminal . There are however significant problems w i t h this approach w h e n appl ied to heterogeneous a l l - IP networks where the number o f variables i n the ne twork select ion process are s igni f icant ly higher. A mul t imode terminal w i l l t yp ica l ly not have a p r io r i knowledge o f a l l the informat ion related to a heterogeneous m i x o f roaming partners, e.g., frequencies o f operation (scanning m a y require turning o n mul t ip le radios), network identities, authentication mechanisms (it m a y require 21 several retries to f ind the supported mechanisms) , etc. U n l i k e today 's mos t ly c i rcu i t - swi tched service networks used for vo ice ca l ls , a s imple ava i lab i l i ty o f an IP ne twork for use does not indicate h o w good or bad network condi t ions are on the avai lable network or its capabil i t ies to support a var iety o f IP services w i t h different Q o S requirements. The process o f ne twork d iscovery i n this case w o u l d therefore l i k e l y be an unpredictable one w i t h several possible retries and failures. F o r most wireless systems, the wireless l i n k is the most expensive and bandwid th l im i t ed part o f the transport and therefore use o f this l i n k has to be efficient. The mechan i sm whereby the terminal i n d i v i d u a l l y queries each o f the avai lable A N s can place a s ignif icant extra overhead on the wire less l inks . In addi t ion, for reasons o f security and l ack o f air l i n k resources, the network operators w o u l d be reluctant to provide informat ion about the network condi t ions d i rec t ly to an i n d i v i d u a l user, especia l ly i f he is a roamer and has not yet authenticated w i t h the network. E v e n i f the ne twork operators do a l l o w the sharing o f such network related informat ion, current M A C / P H Y layers o f dep loyed access technologies do not have the ab i l i ty to provide such informat ion to a terminal before authentication. A n y suggested changes to the M A C / P H Y layers w o u l d take several years to standardize and addi t ional t ime w o u l d be needed for product r o l l out and deployment . S u c h solutions m a y also not be backward compat ib le w i t h the dep loyed infrastructure and terminals. 2.4.2 Network Assisted Approach In a network-assisted approach, the network assists the terminal i n the dec i s ion process b y per forming data co l lec t ion and analysis . A l o n g w i t h its locat ion , the terminal can also prov ide any other informat ion that c o u l d be considered b y the ne twork i n the analysis , e.g., se rv ice /QoS be ing requested, scans o f the P L M N IDs , S S I D s , etc. The network o n l y assists the terminal i n the dec i s ion process and the f inal select ion is s t i l l done b y the terminal (or the user). The final select ion o f an A N from a preferred list p rov ided by the in i t i a l ne twork can be made b y the terminal-based o n addi t ional informat ion such as the s ignal strength from different networks. In this approach, the terminal w o u l d try to get network connect iv i ty (at layer 2 or layer 3) w i t h wh icheve r A N it can (wi th or wi thout authentication). T h i s in i t i a l ne twork access can be compared to getting directory assistance i n today 's w o r l d o f c i rcu i t - swi tched vo i ce cal ls . The in i t i a l A N m a y not be the op t imal network for the requested service and can be changed once the terminal has rece ived sufficient informat ion from the network entities to make a dec is ion . W h i l e 22 ) the dep loyed A N s va ry i n nature, e.g., W L A N , U M T S , W i M A X , a un i fy ing aspect o f different access technologies be ing dep loyed is that they support IP transport. T h i s makes the network-assisted approach for ne twork select ion us ing layer 3, i.e., IP layer entities, feasible. The network-assisted approach can be m u c h more bandwidth efficient for the wireless l i n k as the terminal w o u l d not have to query i nd iv idua l A N s for informat ion related to ne twork select ion. Cons ide r a scenario o f n alternative A N s w i t h ne twork select ion related query / response messages to each o f these networks taking an average o f m bytes. The total wireless bandwid th consumed i n this case w i l l be nxm bytes for the terminal-based approach. F o r the case o f network-assisted approach, the terminal w i l l be able to access s imi la r informat ion i n approximate ly pxm bytes where 1 < p < n. The value o f p w i l l depend upon different deployment strategies. The static or p rov i s ioned component o f informat ion (ma in ly capabil i t ies o f the network) i n this approach c o u l d be stored central ly w i t h i n an operator network, for d isseminat ion based on pol ic ies and agreements. The frequency at w h i c h this informat ion is exchanged w i t h other domains w o u l d depend o n the type o f informat ion and h o w often it is refreshed. B o t h p u l l and push models for informat ion exchange c o u l d be supported. The m a i n drawback o f this approach is the relative delay introduced i n the ne twork select ion process w h e n compared to the terminal-based approach. Since the co l lec t ion and disseminat ion o f dynamic ne twork condi t ions through this architecture uses IP , i.e., layer 3, it m a y not be fast or frequent enough to support handover decisions. H o w e v e r , the latency requirements for in i t i a l network select ion are not as strict as for handoff and therefore this mechan i sm w o u l d w o r k adequately for the purposes o f ne twork select ion at the t ime o f in i t i a l service access for both nomadic and mob i l e type o f users. Tab le 2-1 compares the two possible solutions. B a s e d on this compar ison , the network-assisted approach is preferred. The rest o f the chapter discusses architectural so lu t ion based o n this approach. In order to better understand the proposed network-assisted select ion solut ion, the dec i s ion process is s h o w n i n F igure 2.2 at a h i g h leve l so that it is appl icable to many o f the possible deployment strategies. The term network select ion node has been used i n F igure 2.2 as a general term to accommodate var ious possible deployments o f the l og i ca l nodes descr ibed later i n this section. The b locks shown i n F igure 2.2 do not represent a l og i ca l or phys i ca l ne twork entity but the actions performed dur ing var ious stages o f the select ion process. The f lowchart i n the figure assumes that the terminal is capable o f exchanging informat ion w i t h the ne twork v i a layer 2 or 3 mechanisms. 23 Table 2-1 Comparison of terminal-based and network-assisted mechanisms Terminal Based Network Assisted Backwards compatible and work with current infrastructure T h i s mechan i sm works better w i t h a layer 2 solu t ion for informat ion exchange and therefore is less l i k e l y to w o r k w i t h current deployments because o f required changes to M A C / P H Y layers o f access technologies. T h i s mechan i sm can w o r k w e l l both at layer 3 or layer 2 but w h e n used at layer 3 it can w o r k w i t h current systems because the IP layer is c o m m o n to a l l broadband wireless systems. Flexible deployment and support for different business models Requ i r ed changes for this mechan i sm w i l l l i k e l y happen at the edges o f the networks. IP nodes can be p laced poss ib ly i n different networks depending upon the deployment and business strategy. Scalable based on network and computational resources T h i s approach consumes more bandwidth resources over the wireless l inks a long w i t h more computat ional resources o f the terminal . M o s t o f the data co l l ec t ion and process ing is done w i t h i n the wi re l ine ne twork and therfore wire less bandwid th resoures and terminal computat ional resources are not impacted s ignif icant ly . Support for roaming T h i s mechan i sm m a y not w o r k i n roaming as the networks m a y refuse to share informat ion p r io r to hav ing a credible request f rom a k n o w n entity. T h i s mechan i sm w i l l w o r k i n roaming as the informat ion is shared w i t h other ne twork operators or trusted entities over the IP network. Latency in decision making F o r this mehcna i sm, latency per network query is expected to be less as on ly M A C / P H Y layers are i n v o l v e d but it requires an exhaustive scan for a l l networks. F o r this mechan i sm^whi l e us ing layer 3, latency is expected to be more as IP layer communica t i on is i n v o l v e d w i t h a backend informat ion server. F igure 2.3 shows a h igh l eve l architecture hav ing several autonomous networks connected to heterogeneous wireless access systems. The figure shows the relat ionships between the A N s , their roaming partners and the H N o f a user accessing the network. F igure 2.4 provides further details about the l og i ca l nodes w i t h i n the architecture that are relevant to the ne twork select ion process. M o r e spec i f ica l ly , the proposed architecture introduces two new nodes, Service Announcemen t N o d e ( S A N ) and Da ta C o l l e c t i o n N o d e ( D C N ) . It also adds some new funct ional i ty to the user equipment ( U E ) and the A A A node. The new nodes represent l og i ca l funct ional i ty that w i l l reside i n the network and they are described later i n this section. 24 Network selection request is received from the terminal. It includes information such as terminal location, terminal capabilities and the requested service ^ _ _ i Network selection nodes retrieve provisioned / static information about candidate access networks and services that they support , IT Network selection nodes refine list of candidate networks by comparing information received from the terminal such as the access network types, user credentials supported, services requested etc. with similar information about access networks. 1 Network selection nodes retrieve information on dynamic parameters for remaining candidate networks (e.g. network congestion) that can influence network selection 1 ~ Network selection nodes further refine the list of candidate networks using dynamic parameters as a selection criteria Network selection nodes communicate the list of the networks that fulfill the minimum selection criteria to the terminal or pick only the best network and communicates its identity. Other pertinent information such as access technology, SSID (in the case of WLAN), authentication mechanism etc is also included. Figure 2.2 Decision process in network selection using proposed network-assisted mechanism Figure 2.3 A heterogeneous wireless system with access networks using different technologies 25 The system as represented i n the two figures does not inc lude a l l the log ica l functional nodes that w o u l d be present i n a heterogeneous ne twork ing environment. O n l y those nodes that are related to the network select ion process have been presented. Other functional nodes, e.g., m o b i l i t y management node, are outside the scope o f this d iscuss ion. Figure 2.4 Logical nodes involved in network selection 2.4.3 Using DCN to Collect and Provide Access Network Characteristics The processes o f network moni to r ing , and data co l lec t ion , storage and presentation are performed b y the functional entity D C N as shown i n F igure 2.4. D C N funct ional i ty w i l l p r imar i l y have two sub-functions. F i r s t ly , it monitors and collects informat ion o n network condi t ions us ing a standard set o f parameters so that these parameters have the same interpretation across different autonomous networks us ing different access technologies. T h i s sub-function is named the Information C o l l e c t i o n Func t ion ( I C F ) . Secondly , it stores and exchanges this informat ion across autonomous domains i n a standard manner so that the informat ion can be exchanged easi ly across the boundaries o f autonomous systems. T h i s funct ion is named the Information P rov ide r Func t ion ( IPF) . Da ta co l lec t ion w i l l be done i n the A N s but the co l lec ted informat ion can be stored anywhere i n the network. The A N s such as W L A N hotspots can have S L A s w i t h partner networks that c o u l d a l l o w certain Q o S guarantees (e.g., i n terms o f latency, jitter, and packet drop rate). W h i l e the A N w o u l d m a i n l y provide informat ion about the condi t ions and capabil i t ies o f its o w n network, it can also include 26 informat ion from its S L A s w i t h partner networks or the network condi t ions it is exper ienc ing w i t h its partner networks. A g o o d basis for selecting appropriate Q o S related attributes is p rov ided i n [10] and [11]. The attributes discussed i n [10] and [11] can be used to describe network capabil i t ies and current ne twork condit ions across a var iety o f A N types. The f o l l o w i n g is a subset o f the characteristics described i n [10] that are relevant to this d iscuss ion. • B a n d w i d t h Characteristics • D e l a y Characteristics • L o s s Characteristics • A v a i l a b i l i t y Characteristics • Packet Re -Orde r ing Characteristics The d iscuss ion o n a possible l ist o f parameters used to measure these characteristics can also be found i n [10]. In [12], methods to col lect and distribute such network related informat ion have been discussed. 2.4.4 Using SAN for Service Selection A l t h o u g h the concept o f service announcement is not new, w h e n requested, the S A N i n the proposed architecture provides informat ion about w e l l defined user subscr ibed services that are avai lable to the requesting user v i a a part icular A N based on not o n l y the operational requirements such as network capabil i t ies but also the business requirements such as legal and contractual constraints. In some scenarios there can be a g o o d mapp ing between network capabil i t ies and the types o f services that the network supports so that the supported services can indi rec t ly be der ived from network capabil i t ies informat ion. H o w e v e r , other factors such as business models can also influence the services that are offered b y a network i n a geographic loca t ion . In some cases due to regulatory constraints the ne twork m a y e x p l i c i t l y indicate no support for a service even i f the network capabil i t ies c o u l d support the service from a technology perspective. F o r example , based on the business mode l or regulatory issues (e.g., C A L E A , E 9 1 1 ) , the A N m a y choose to be either a condui t for data or it m a y decide to provide services such as streaming m e d i a and V o I P di rect ly or v i a its partners (e.g., content aggregators) as w e l l . The S A N s w o u l d be used to announce the presence o f such services. Depend ing u p o n the 27 deployment scenarios as documented, the S A N s be ing quer ied b y the user terminal can reside i n the A N , its partner networks, or the user H N . E x i s t i n g I E T F protocols such as L D A P or S L P c o u l d poss ib ly be reused for acqui r ing such service related informat ion. 2.4.5 Using Enhanced AAA for Roaming Support The funct ional i ty o f A A A nodes i n the proposed architecture is s imi la r to the one that exists today i n data networks but enhanced to handle heterogeneous systems. In a heterogeneous wireless environment, different user authentication credential types and mechanisms m a y be used. F o r such systems to interwork, the user terminals as w e l l as wireless A N s must have a c o m m o n or interoperable w a y o f p rov id ing to each other informat ion such as their identities, their H N s , their capabil i t ies and their roaming agreements, etc. N e t w o r k layer mechanisms can be used to provide such interoperabil i ty. A n example o f p r o v i d i n g user identi ty and U N identi ty at the ne twork layer can be the use o f N A I s defined i n I E T F R F C 2486bis , where the r ea lm part o f the N A I can provide the name o f the H N operator. A decorated N A I can also include informat ion o n h o w to route the request v i a mul t ip le roaming partner networks. O n c e the A N operator has the informat ion o n H N and roaming partners, it can decide i f it has any agreements to authenticate and charge the user for its usage o f the network. The H N can also communica te service and subscr ipt ion related informat ion i n its A A A messages. In order to get ubiqui tous coverage b y roaming w i t h a large number o f roaming partners, it m a y become necessary for a mul t imode terminal user to carry more than one type o f authentication credentials for the same H N (e.g., S I M , user ID/password) . Protocols such as P A N A [4] w i t h D I A M E T E R / R A D I U S i n the backend can be used i n the enhanced A A A nodes proposed to provide such support. [8] has discussed i n detail issues associated w i t h A A A rout ing. Since any proposed solutions have to apply to a w i d e range o f A A A messages, any enhancements to resolve them should not restrict the in t roduct ion o f new A A A or access ne twork functionalit ies. B y us ing the new A A A functionali ty a long w i t h avai lable service and transport l eve l informat ion v i a bearer path nodes (e.g., routers, appl ica t ion servers), more complex charging relationships can be supported. F o r example, it can be possible to per form differential charging or content based charging b y correlat ing IP f l o w w i t h the content downloaded and then charging the customer for the content on ly . 28 2.4.6 Information Collection and Exchange F o r network selection, the system collects and provides two types o f network informat ion: static network capabil i t ies, and dynamic network condit ions. It is necessary to get rel iable and updated values for both types o f information before the process o f network selection. The col lected information to be used i n the decis ion process can be stored anywhere i n the network. The A N s w i l l have S L A s w i t h partner networks that provide certain Q o S guarantees (e.g., latency, jitter, and packet drop rate). W h i l e each A N w o u l d m a i n l y provide informat ion about its o w n condit ions and capabili t ies, it can also include informat ion from its S L A s w i t h partner networks, or o n the network condit ions it is experiencing w i t h its partner networks. F igure 2.5 shows, at a h igh leve l , h o w information to be used i n pol icy-based network selection is col lected w i t h i n an autonomous domain and then shared across the doma in boundaries based on operator pol ic ies and roaming agreements. In [12], s imi la r mechanisms to col lect and distribute resource information have been discussed, al though for a different usage. H o w e v e r the mechan i sm is w e l l suited for network related information co l lec t ion for network selection as w e l l . The process can be broken d o w n into the f o l l o w i n g steps: • M o n i t o r resources w i t h i n the A N . The raw information m a y be processed for data reduction before co l lec t ion b y central nodes w i t h i n the domain . • B a s e d o n the information received from A N s w i t h i n the domain , mode l and forecast/predict ava i lab i l i ty o f these resources. • Disseminate the informat ion to roaming partners us ing standard parameters, based on pol ic ies and roaming agreements. F igure 2.5 also shows that static or provis ioned information about the services and A N s is stored central ly w i t h i n the autonomous domain , for disseminat ion based on pol ic ies and agreements. The frequency at w h i c h this information is exchanged w i t h other domains w o u l d depend on the type o f informat ion and h o w often it is refreshed. B o t h p u l l and push models for informat ion exchange c o u l d be supported. F igure 2.6 illustrates the fact that w h i l e the Q o S attributes used for functions such as network selection, admiss ion control , handoffs, and packet schedul ing are l i k e l y to be s imi lar , because o f the differences i n functional requirements the update rates and hence the levels o f accuracy w i l l be different. 29 Lower Accu racy / Update Rate Higher Accuracy / Update Rati Network Admiss ion Control Up Link / Down Link Packet Se lect ion and Handoffs Schedul ing Figure 2.6 QoS attributes used for network selection, admission control, handoffs, packet scheduling can be the same but will have different update rate and the level of accuracy The collection and dissemination of dynamic network conditions through this architecture can occur at layer 3, and may not be fast or frequent enough to support real-time handoff decisions. However, this information could sufficiently reflect the dynamics of the network to facilitate network selection at the time of initial service access. For example, if average utilizations of two 30 ANs, A and B, with similar capabilities and bandwidth are 90% and 30%, respectively, then the terminal should choose network B to access service. 2.5 Deployment Scenarios In order to discuss deployment scenarios for automated network selection, we again consider the example briefly described in the introduction where a user is in a public place served by multiple WLANs and WW ANs. The terminal supports both W L A N and WW A N access and the user decides to use his HN based video streaming service to watch the news. Currently, there is no automated method for the terminal to intelligently select the most appropriate network from a heterogeneous mix of wireless access technologies for the desired service. Without an appropriate architectural solution, the user may have to manually go through each individual A N and select the one he considers acceptable. The process can be quite inefficient and can result in excessive network traffic and subsequent delays in the process of network selection. The information available from individual ANs for network selection purposes may also be non-uniform. This could result in the selection of a suboptimal network, e.g., in terms of network resources or cost. In many cases the user may not be able to find out beforehand if certain services are supported by the A N . Because of the non-uniform information availability for network selection, the process may not be consistent and may not work all the time under roaming conditions. Using the proposed architecture, different deployment for the SAN and D C N nodes can occur. The following sub-sections describe the more common deployment scenarios and the decision processes involved in them. 2.5.1 DCN (IPF) and SAN residing in the AN Figure 2.7 shows this scenario. The user terminal first tries to get connectivity with an A N . In secure ANs this starts with the process of authentication. The decision to initially select one of the ANs for network connectivity can be a user terminal behavior based on possible information available to the user terminal via, e.g., provisioning, prior experience with the networks or random. Once the user terminal gets connectivity to an A N (e.g., it is assigned an IP address), it requests and gets the information about services supported by the A N from the SAN residing in the A N . For example, the SAN can inform the user that the network supports VoIP using IMS. The user terminal also requests and gets information about the current network conditions from 31 the D C N located within the A N . This allows the user terminal to make a decision i f it should proceed with initiation of the service, e.g., a H N based streaming media service. The user terminal then proceeds with initiating access to the H N based service. The information provided by the S A N and D C N in this case may be limited to the A N in which it resides. The user terminal can also proceed with authenticating with other A N s in order to get information from their D C N and S A N nodes before making a final decision on the A N to use for the service. Figure 2.7 DCN (IPF) and SAN residing in the AN 2.5.2 DCN (IPF) and SAN residing in local information provider network Figure 2.8 shows the deployment scenario. Once the user terminal gets connected to the A N (e.g., it is assigned an IP address), it gets the information about the services supported by the A N from the S A N residing in the network of the local information server. A possible variation to this scenario can be that the local information provider may have deployed an A N (such as a W L A N ) that provides local connectivity to anyone without pre-authentication to provide access to the S A N . The local information server can be a free service provided by the owner of the premises or a local communication broker where the user terminal is trying to get the service. The information service provider may not own any of the A N s . For example, a S A N node operated by the owner of a convention center and accessible via the local A N operators, can inform the user that the A N operators X , Y and Z support VoIP using IMS in 32 the premises. If the regulations require support for emergency calls, this would imply support for location information by the ANs. The SAN can also provide further information regarding, e.g., the frequency bands of operation and the pricing for these networks. The user terminal also gets information from the local information server about the network capabilities and current network conditions in multiple ANs operating in the current location. Additional information such as the type of credentials and authentication mechanisms used by the networks may also be provided. This allows the user terminal to make a decision if it should proceed with initiation of the service, or if it should switch to a different AN before service initiation. A A A signaling Bearer AN#2^  V v *• Ntwk and Service info Figure 2.8 D C N (IPF) and SAN residing in local information provider's network 2.5.3 DCN (IPF) and SAN residing in the HN This type of deployment, shown in Figure 2.9, allows the HN to manage the user experience by managing network selection related information that is provided to the user terminal. The user terminal or the A A A proxy server of the AN being used is able to provide the user's geographic location during the authentication phase or afterwards. After getting connectivity to the network via layer 2 or layer 3 mechanisms, the terminal accesses the information entities located in its home network. The HN based DCN (IPF) and SAN, based on the roaming relationship, are able to collect information about possible local ANs on which the user can roam, their network capabilities / conditions, and the services that they offer. Based on the information received from 33 its H N , the terminal can make a decision to change the A N to the one better suited for the service to be used. In all the scenarios described above, the ability of the terminal to get to the appropriate information providing entities in the network can be based on a Domain Name Server (DNS) lookup that would allow a terminal to contact, e.g., "services.home_operator_name.com" for service related information in the home network and "service.local.com" for services information in the local A N . Also , to handle services that can trigger a new session from the network side (e.g., a terminating VoIP call) the user/terminal should indicate the intention to be able to accept such a service, at the time of network selection request. This can be part of a pre-configured user profile that can include information such as his mobility (e.g., nomadic) and preferred services (e.g., voice) that can then automatically be used at the time of network selection request. In this section we have addressed the problem of automated network selection in an environment where heterogeneous all-IP wireless A N s are available to a user terminal. Limitations with the current state of the art have been highlighted. Two approaches to solve the problem are discussed and a network-assisted network selection mechanism is proposed that defines new network layer nodes as well as new functionality for current nodes. The proposed solution provides a common architecture for automatic network selection in a heterogeneous wireless system where the user terminal requests and receives information from the network Figure 2.9 D C N (IPF) and S A N residing in H N 34 before selecting an A N for service initiation. It leverages the ongoing protocol development in IETF and IEEE. Various deployment scenarios have been discussed to explain how the proposed architecture would work. The proposed architecture minimizes usage of the wireless links and would work with currently deployed infrastructure. It is scaleable, flexible and supports roaming. In the prior example, a network-based entity that can coordinate the exchange of information across different information resources (such as the SAN, DCN, etc.) has not been described and essentially each information entity is assumed to independently process the information and then interact with the terminal. In the following section we analyze a policy-based network selection architecture that leverages currently proposed elements of a 3 GPP system and provides for a network entity that can coordinate across different information resources to handle terminal request for network selection. 2.6 Policy-Based Network Selection Future mobile devices will be equipped with a multitude of wireless access technologies that enable ubiquitous service access via alternate wireless ANs supporting an all-IP data transport. This common data transport capability allows the transport plane to be decoupled from the service plane in the next generation wireless ANs, facilitating delivery of converged services over a variety of access technologies. Services such as those based on the IMS are agnostic of the type of the underlying A N so long as it supports certain capabilities, e.g., a certain level of QoS. This scenario presents a complex but interesting set of management problems, with the selection of an optimal A N for service access being a key one. A network selection mechanism that uses operator policies to provide ubiquitous services using the best among a mix of coexisting heterogeneous wireless networks for service delivery at any location does not currently exist. Today's mobile devices mostly support only single mode wireless access and employ simple network selection policies that would be insufficient to meet the much more complex network selection scenarios encountered by future multimode devices. For example, a G S M cellular device would scan for unique network identifiers broadcast over alternate ANs. These network identifiers do not provide any information about current network conditions, capabilities or the services supported, as such information is typically not needed in today's homogeneous ANs in which the primary service, i.e., voice, is supported universally using circuit-switched airlinks with tightly controlled QoS. In addition roaming agreements amongst operators are mostly applied across one access technology and therefore the need for supporting multiple credential and authentication types does not arise either. In the example cited the network selection 35 decision is made entirely by the terminal-based on scanned information about network identifiers. In order to develop an intelligent network selection policy that is better suited to a heterogeneous all-IP wireless access environment, more information about the alternate A N s would be needed. Such information can be dynamic and may be more readily available within the network nodes rather than pre-provisioned within the terminals. In prior related work 3GPP has documented a network selection solution [13] for W L A N s interworking with cellular systems. Broader network selection issues for cellular systems [14] are also being discussed within 3GPP. In this section we combine policy related components defined in recent standards activities with the novel network selection components introduced in the previous sections to enable automated network selection. The original solution introduced here allows the U E to intelligently and automatically select from several available A N s , through a policy-based decision mechanism using multiple attributes such as the service to be accessed, the A N s ' QoS capabilities and the current network conditions. We solve the problem by proposing a policy-based network selection mechanism whereby the network assists the terminal in making an intelligent choice about the A N for service delivery, and identifying the information components needed. The factors to be considered in a policy-based network selection include service aspects, access network characteristics, costs and roaming support. They have been described in the beginning o f this chapter. 2.6.1 Adding Standards-based Policy Nodes to Network Selection Framework Figure 2.10 depicts a heterogeneous wireless access system that can deliver services to a user over several available autonomous networks, some having roaming agreements with the user's H N while others do not. Figure 2.11 illustrates the service, policy and transport layer logical components involved in network selection in such an environment. For clarity, logical nodes that would be present in a heterogeneous networking environment but are not involved in the network selection process, e.g., those supporting mobility management functions, have been excluded from this figure. The service and network information nodes for policy-based control are new functions introduced for network selection. They have been described in earlier sections of this chapter. In addition the traditional functionality of Policy Decision Function (PDF) as defined by 3GPP [15] is enhanced to support network selection by using M A D M as described in [16][17][18]. Figure 2.11 also shows the HN-based Ca l l Session Control Function (CSCF) , an 36 I M S service layer functionality [19] that interacts with the policy components. Figure 2.10 High level architecture for a heterogeneous wireless system with user's HN providing the services The user accesses the network to use a particular service, e.g., to make an IMS-based VoIP call. Using the elements of the proposed architecture described earlier in this chapter, it is possible to make an informed decision based on the service related information, the characteristics of the AN and the roaming aspects. In some cases there can be a good mapping between network capabilities and the types of I M S service that the network can support. If so, supported services can indirectly be derived from network capabilities information. However, other factors can also influence the services that are offered by a network in a geographic location; e.g., an A N may not support voice services due to regulatory or contractual constraints such as a requirement to support emergency calls. Such policy information is exchanged by roaming partners. The H N can provide such service related information, e.g., about I M S services offered by it and its roaming partners to its subscribers based on their locations. The terminal, by providing its location, would be able to retrieve relevant service information. For interoperability purposes, service announcements should be made in a standardized manner, such as proposed by the I E E E 802.21 and I E E E 802.1 l u working groups. 37 Application Function e.g. Call Session Control Function (CSCF) QoS Requirements ..•** ..•••Service Layer components e.g. for IMS Policy Decision Function Service Information Network Information Authentication, Authorization and Accounting (AAA) Policy control components QoS Requirements *'*••.. Policy Enforcement Function Transport layer components Figure 2.11 Logical nodes involved in network selection The P D F and Policy Enforcement Function (PEF) are standard functions in the 3GPP Policy Control and Charging architecture [15]. The P D F makes QoS and other policy related decisions for IP services such as those based on I M S . We extend the role of P D F to include network selection decision making based on network, service and user subscription related information as shown in Figure 2.11 using a M A D M algorithm [18]. The P E F enforces the QoS and any other decisions communicated by the P D F . Policy decision and enforcement functions are described in [15] and their application to non 3GPP wireless access technologies is being considered in System Architecture Evolution (SAE) of 3GPP [7]. Policy decision and enforcement related functionalities have also been adopted by wire-line forums such as Cable-Labs and ETSI T I S P A N . Thus an industry consensus similar to IMS is emerging in this area. 2.6.2 Example of Policy-based Automated Network Selection N o w we again consider the scenario where a user employing a dual-mode terminal that supports both W L A N and cellular access is in a public place served by multiple W L A N s and cellular systems. In this case the user decides to access the IMS-based VoIP service at his H N at a location where two A N s are available, as shown in Figure 2.12. 38 Terminal AN#1 AP PDF PEF AN#2 AAA PDF Ntwk Svc Home IMS Info Info AAA Services iase IEZ Et 02.1 1u C uei ry> Inh ial Rei por seP iet :ess via AN#: Itica ion :or ac AN Be #1 jin IMS AN Inte #2 rac f/o, is o H< ver> i m e 1 let wo rk Figure 2.12 High level communication with policy components residing in the HN. The PDF is enhanced to include the functionality of decision making for network selection. The decision making algorithms that can be used are described in later chapters. A s described earlier, different alternatives exist for deployment of the network selection policy nodes. Here we consider the case where they are located in the H N , a deployment considered suitable for HN-based I M S services. The terminal obtains the information needed for network selection from its H N . Such a deployment works well under roaming scenarios as the H N is able to collect information about services and A N s of its roaming partners as described earlier. A s shown in Figure 2.12, the terminal initiates an information exchange with an A N . In this example, the decision to select AN#1 for accessing the back-end policy server is based on information broadcast by AN#1 in its beacon. In other scenarios it can be also be a terminal behavior based on possible information available to the terminal via, e.g., provisioning, prior experience with the networks, or random. In this example, the M A C layer of AN#1 supports queries at layer 2. Once the terminal has identified the A P for AN#1 to be 802.1 l u compliant, it can make different types of queries; e.g., it can request information from the GAS-aware policy server in its H N about available A N s at its current location that support certain IMS services and have roaming relationships with its H N and have the capabilities and network conditions to 39 deliver the services. This type of information can allow the terminal to decide i f it should proceed with service initiation. Based on the information received during queries to the H N , the terminal decides to change A N before initiating the service. Once the A N has been selected the terminal proceeds with its interaction with H N C S C F via the selected A N (AN#2 in this example) to initiate I M S services, as shown in Figure 2.13. The H N C S C F communicates QoS and resource requirements for the requested IMS service to the H N P D F , which has peer relations with the P D F in the visited network to facilitate negotiation of QoS and resource allocation. The P D F communicates the final QoS request to the P E F in the visited network. With the use of intelligent network selection in the prior step, the chances of success in getting the requested QoS are maximized. Figure 2.13 Signaling and bearer paths once AN#2 has been selected. The bearer path is route optimized via a local breakout The HN-based deployment of network selection policy functions allows the H N to manage the customer experience by providing the appropriate network selection information to the terminal. The terminal or the A A A proxy server of the A N being used should be able to provide the terminal location to the H N during the authentication phase or afterwards. The HN-based network and service information elements, based on existing roaming relationships, are able to suggest possible local A N s that the user can roam on, which offer suitable capabilities, dynamic 40 conditions, and service support. The central location of information in the HN allows for easy updates of roaming related information and coverage maps. The ability of the HN to process and filter information before delivering it to the user can enable efficient usage of network resources and reduce computational requirements for the terminal. 2.7 Conclusion In this chapter we have developed an architectural framework to enable automated network selection for delivery of services in the next generation all-IP heterogeneous wireless access environment. Limitations with current approaches have been highlighted and a new policy-based solution that leverages ongoing standards activities has been proposed through the addition of new policy functions to the existing policy related nodes and the introduction of new function nodes for data collection and service announcement, which can be readily implemented in users' home networks. The novel solution presented in this chapter effectively provides the functionality of automated network selection where the terminal requests and receives a list of recommended access networks from the initial access network before selecting a service delivery network. Sample deployment scenarios have been discussed to explain how the proposed solution works. 41 2.8 References [I] M . Inoue, K . M a h m u d , H . M u r a k a m i and M . Hasegawa, " M I R A I : A So lu t ion to Seamless A c c e s s i n Heterogeneous Wire less N e t w o r k s N e w Genera t ion" , Proc. IEEE ICC, pp. 1033-1037, M a y 2003. [2] " M e d i a Independent Handover" , I E E E 802.21, [online] A v a i l a b l e ht tp: / /www.ieee802.org/21/ . [3] " M I P v 6 S igna l ing and H a n d o f f O p t i m i z a t i o n " , I E T F M I P S H O P W G , [online] A v a i l a b l e ht tp: / /www.ietf .org [4] " P r o t o c o l for ca r ry ing Authent ica t ion for N e t w o r k A c c e s s " , I E T F P A N A W G , [online] A v a i l a b l e ht tp: / /www.ietf .org [5] " M o b i l i t y Op t imiza t i on" , I R T F M O B O P T S W G , [online] A v a i l a b l e http://irtf.org. [6] " N e x t steps i n s igna l ing" , I E T F N S I S W G , [online] A v a i l a b l e ht tp: / /www.ietf .org [7] "Sys t em Archi tec ture E v o l u t i o n " , 3GPP TR 23.882, A p r i l 2007, [online] A v a i l a b l e h t tp : / /www.3gpp.org [8] J . A r k k o , B . A b b o b a , J . K o r h o n e n and F . B a r i , " N e t w o r k D i s c o v e r y and Selec t ion P r o b l e m " , draft-ietf-eap-netsel-problem-08, June 2007 (completed w o r k i n g group last ca l l ) , [online] A v a i l a b l e , h t tp: / /www.ietf .org. [9] Inter w o r k i n g w i t h Ex te rna l N e t w o r k s Task G r o u p , I E E E 8 0 2 . l l u , [online] A v a i l a b l e http://www.ieee802 .Org / l 1/ [10] B . L o w e k a m p , B . Tierney , L . Cot t re l l , R . Hughes-Jones, T . K i e l m a n n , and M . Swany , " A Hie ra r chy o f N e t w o r k Performance Characterist ics for G r i d A p p l i c a t i o n s and Services" , GFD-R-P.023, G G F N e t w o r k Measurement W o r k i n g G r o u p ( N M W G ) , M a y 2004. [ I I ] Internet Performance M e t r i c s W G , I E T F I P P M W G , [online] A v a i l a b l e ht tp: / /www.iet f .org [12] B . L o w e k a m p , N . M i l l e r , R . Kar re r , T . Gross , and P . Steenkiste, " D e s i g n , Implementat ion, and Eva lua t i on o f the R e m o s N e t w o r k M o n i t o r i n g Sys tem" , Journal of Grid Computing, pp. 75-93 , v o l . 1 ,2003. 42 [13] "3GPP System to Wireless Local Area Network (WLAN) Interworking", 3 GPP Technical Specification 23.234 v 7.2.0, June 2006, [online] Available http://www.3gpp.org. [14] "3GPP Technical Specification Group Services and Systems Aspects: Review of Network Selection Principles", 3GPP Technical Report 22.811 v 7.2.0, June 2006, [online] Available http://www.3gpp.org. [15] "Policy and Charging Control Architecture", 3GPP Technical Specification 23.203 v 1.1..0, August 2006. [16] F. Bari and V. Leung, "Automated Network Selection in A Heterogeneous Wireless Network Environment", IEEE Network, pp. 34-40, January/February 2007. [17] F. Bari and V. Leung, "Multi-Attribute Network Selection by Iterative TOPSIS for Heterogeneous Wireless Access", Proc. IEEE CCNC, January 2007. [18] F. Bari and V. Leung, "Application of E L E C T R E to Network Selection in a Heterogeneous Wireless Network Emvironment", Proc. IEEE WCNC, March 2007. [19] "IP Multimedia Subsystem (IMS), System Description", 3GPP Technical Specification 23.228 v 7.4.0, June 2006, [online] Available http://www.3gpp.org. [20] S. Das, Y . Ohba and F. Bari, "Information Service (IS) Reference Model, Use Case Scenario and Higher Layer Requirements for 802.21 Information Service (IS)", 21-05-0348-00-0000-Higher_layer_Requiremehts_Information_Service. doc, September 2005, [online] Available http://www.ieee802.org/21/. [21] N. Canopolat, V. Gupta, D. Stephenson, S. Sreemanthula, R.Y. Kim, E. Hepworth, S. Demel, M . Rudolf, F. Bari, H. Cheng, L. Mo, Z. Yao, A. Soomro, and D. Cavalcanti, "Network Selection - Normative text proposal", IEEE Contribution 11-06-1014-03-000u-network-selection-normative-text-doc, September 2006, [online] Available ftp.802wirelessworld.com. 43 3 Use of MADM Algorithms for Automated Network Selection in a Heterogeneous Wireless Network Environment 1 3.1 Introduction The desire to increase service availability is driving the use of multimode terminals and the inter-connection of different wireless access technologies that support IP transport. An essential aspect of service delivery in a heterogeneous all-IP wireless network environment is the selection of an optimal A N . Network selection in such an environment is influenced by several factors and currently there is no comprehensive solution available to solve this problem. Selection of a non optimal network can result in undesirable effects such as higher costs or poor service experience. A system architecture for a terminal-based approach that uses network assistance for network selection has been proposed in Chapter 2 [1]. An overview of this approach is provided in the beginning of Section 3 .3 . Here we describe an extensive decision making process that can be used by the network to provide candidate networks for service delivery to the terminal. The proposed mechanism is based on a unique decision process that uses non-compensatory and compensatory M A D M algorithms jointly on the network side to assist the terminal to select the top candidate network(s). Along with disjunctive and conjunctive type of non-compensatory MADMs, we describe the use of TOPSIS, a compensatory M A D M algorithm. It identifies various factors that would influence the selection of an optimal network and therefore should be used as 1 Parts of this chapter have been published in the following F. Bari and V. Leung, "Automated Network Selection in A Heterogeneous Wireless Network Environment", IEEE Network, pp. 34-40, January/February 2007. 44 inputs to the M A D M based decision process. It explores different aspects of the problem space and proposes solutions based on M A D M mechanism. The steps involved in the use of non-compensatory M A D M algorithm are simple. They remove network alternatives from the candidate list that are not suited for the scenario. This process is completed before a compensatory M A D M algorithm can be used to provide network ranking. The compensatory M A D M part of the proposed mechanism is more sophisticated with tunable parameters. It has therefore been validated here by applying it to different use scenarios; e.g., scenarios with different service and subscribed QoS profiles have been shown to observe how the proposed mechanism would work. Application of the proposed mechanism to networks with multiple levels of SLAs or multiple classes of QoS is described. Possible approaches to network selection when another service is already in use by the terminal are presented within the proposed frame work. 3.2 Background and Related Work The problem of network selection across heterogeneous wireless networks has recently received much attention because of a drive for converged communication systems. In this context, [2] has proposed the combined application of two mathematical techniques in an algorithm for network selection between UMTS networks and WLANs, where the analytic hierarchy process and Grey system theory are used for evaluating the user's preferences and service requirements, and for combining the priority settings of the QoS attributes with the performances of the network alternatives to make network selection decision. A dynamic system to select the network for service delivery based on current market conditions such as QoS and cost attributes has been described in [3]. Using the proposed framework, the user can select the delivery network per call. Network selection based on resource allocation strategy for the most efficient resource utilization in a heterogeneous networking environment has been proposed in [4]. A fuzzy logic based multiple-criteria decision-making system to perform access network selection in a heterogeneous networking environment is described in [5]. A simple policy-enabled handoff system across heterogeneous wireless networks is presented in [6], which allows users to express policies on what is the best wireless system at any moment, and make trade-offs among network characteristics and dynamics such as cost, performance and power consumption. The mechanisms described in [2]-[6] have significant shortcomings. The factors considered in the decision processes are insufficient; e.g., information about access types supported, authentication types supported, roaming partners supported are not considered in the decision processes described in these papers. They also do not provide or refer to a viable architectural framework in which the 45 selection mechanism can work. The use case scenarios described are limited and not realistic from the perspective of the deployment or business models used in the industry. As a result, they do not provide a complete and deployable solution to the problem. The decision about network selection in a heterogeneous wireless environment is dependent on several factors. The network selection problem can be solved using M A D M algorithms. M A D M has been an active area of research since 1970s. Because of their deterministic nature and easy implementation, M A D M algorithms have found applications in a wide variety of problems, from social sciences to operations research. They can be used in combination with fuzzy logic where input attributes values are not clearly defined. Another interesting use of these algorithms involved game theory where the players manipulate input values to influence decision making by M A D M algorithms. Amongst the most widely discussed M A D M mechanisms are E L E C T R E [7] (Elimination and Choice Translating Priority), TOPSIS [8] (Technique for Order Preference by Similarity to Ideal Solution), AHP [8], [9] (Analytic Hierarchy Process), GRA [10] (Grey Relational Analysis), WSM [8] (Weighed Sum Model) and WPM [8] (Weighed Product Model). The mechanism described in [2] has also utilized an M A D M approach for network selection but with only a limited number of parameters and without an architectural framework. Factors such as the roaming agreements of the user's HN, the authentication methods supported, and the type of mobility of the user, etc., were not taken into consideration. Scenarios where the A N can support multiple levels of QoS or when the user is already using a network for an earlier session were not addressed. Moreover, issues of the sensitivity of selection processes to the input data were not explored. The solution presented in this chapter is based on the architectural framework described in Chapter 2 [1] that supports current business models (e.g., roaming agreements) used in the industry. It identifies key parameters for use in the network selection process, proposes a comprehensive network selection mechanism using these parameters as inputs, and then validates the proposed mechanism by applying it under different scenarios. 3.3 Decision Process in Network Selection In the network-assisted mechanism proposed in [1] and described in Chapter 2, the network assists the terminal in the selection process by performing data collection and analysis to provide the network ranking. The architecture proposes the use of three network-based functional entities: 1) Data Collection Node (DCN) for retrieving network characteristics, 2) Service Announcement Node for providing services related data and 3) enhanced Authentication, Authorization and 46 Accounting ( A A A ) Node for A A A information. Together they provide input data to a network-based assessment entity that uses an algorithm to calculate the network rankings for use by the terminal in network selection. A s indicated in the policy-based network selection example of Chapter 2, the execution of the network selection algorithm is assumed to reside in the enhanced P D F entity. This assumption holds for the rest of thesis. The process starts with the terminal trying to get connectivity with whichever A N it can access (with or without authentication). This initial network access is similar to getting directory assistance over today's circuit-switched voice network. The initial A N may not be the optimal network for the requested service and can be changed once sufficient information has been collected to make a decision. The terminal provides its location and any other information that could be considered by the network in the analysis, e.g., service/QoS being requested, scans of the locally available P L M N IDs, SSIDs, etc. Using the information collected from the D C N , S A N and A A A nodes and those provided by the terminal, the enhanced P D F calculates the rankings of the available A N s and provides them to the terminal. The remainder of this chapter describes a comprehensive algorithmic approach to be used by the network-based assessment entity in the proposed architectural framework. We formulate the network selection decision process in a heterogeneous networkingjenvironment as a M A D M problem that deals with the evaluation of a set of alternatives using a set of attributes. The alternatives represent different choices. The decision process ranks these alternatives in order of preference using the set of attributes that provide different aspects by which the alternatives can be viewed. The decision about network selection is dependent on several factors. In general the factors or attributes impacting network selection can be divided into following categories. • Category 1 attributes include parameters that are not QoS-related. These parameters do not change very often and therefore can usually be provisioned in the network. • Category 2 includes mostly the QoS-related attributes, both dynamic ones that need to be frequently updated as well as largely static ones that can be provisioned. Table 3-1 lists the attributes that w i l l be taken into consideration for network selection using the algorithms proposed in this chapter and considered in the subsequent chapters and classifies them according to the two categories described above. 47 Table 3-1 Attributes for use in network selection Attribute Abbrev Brief Explanation Category Operator Name ON This attribute indicates the identity of the operator (e.g., roaming partner SSID or P L M N ID) for which the rest of the information is being provided. 1 Authentication Mechanism A M This attribute indicates the authentication mechanism used by the roaming partner. Examples could be SIM or user ID/password. 1 Access Technology A T This attribute indicates the wireless access technology that the A N uses, e.g., UMTS, WiMAX, W L A N . More specifics about the technology such as frequency of operation may also be included. 1 Services Available SA This attribute provides a list of supported services, e.g., 3 G P P services such as IMS based VoIP. 1 Geographic Location G L This attribute provides information on the geographic location of the base station or AP. 1 Cover ase Area C A This attribute provides a measure of the coverage area, for example, hotspot physical address. Due to propagation conditions, wireless coverage area can be irregular. Other attributes such as transmit signal power could be more useful in many instances. 1 Cost per Byte CB This attribute indicates relative transport cost of the operator for a particular access network. It would take into account factors such as use of licensed / unlicensed spectrum, roaming agreements, etc. 2 Total Bandwidth TB This attribute indicates how much bandwidth is available overall on the wireless access link. 2 Allowed Bandwidth AB This attribute indicates the bandwidth allowed by the A N on a per user basis. 2 Utilization U This attribute provides a measure of the current utilization of the wireless link. 2 Packet delay D This attribute gives the average packet delay within the access system. This is not the end-to-end delay. 2 Packet Jitter J This attribute measures the average delay variations within the access system. A large jitter could result in packet reordering or dropping of real-time packets at the receiver. 2 Packet Loss L This attribute measures the average packet loss rate within the access system. 2 48 3.4 Multi Attribute Decision Making (MADM) The proposed decision process for network selection can be described by the steps shown in Figure 3.1. The process shown employs both non-compensatory and compensatory M A D M algorithms, which are described in more detail below. N e t w o r k s e l e c t i o n r e q u e s t r e c e i v e d f r o m the u s e r te rmina l R e t r i e v e re levant in format ion f r o m network entit ies for u s e in the d e c i s i o n m a k i n g N a r r o w the list o f c a n d i d a t e n e t w o r k s b y u s i n g n o n - c o m p e n s a t o r y M A D M a lgor i thm with attr ibutes in C a t e g o r y 1 of T a b l e 3 - 1 , e .g . , the a c c e s s ne twork t y p e s , u s e r c r e d e n t i a l s s u p p o r t e d , s e r v i c e s r e q u e s t e d , e t c . F u r t h e r ref ine the list o f c a n d i d a t e n e t w o r k s u s i n g c o m p e n s a t o r y M A D M a lgor i thm with attr ibutes in C a t e g o r y 2 of T a b l e 3-1, e .g . , Q o S re lated p a r a m e t e r s P r o v i d e the c a n d i d a t e network re lated in format ion to the T e r m i n a l Figure 3.1 Decision making process in network selection 3.4.1 Non-Compensatory MADM Algorithms In this class of M A D M algorithms, advantages in one type of attribute cannot be traded for disadvantages in another type of attribute. This class of methods, as described in [11], include dominance, conjunctive, disjunctive and sequential elimination (e.g., lexicographic, elimination by aspects). Of particular interest for network selection are conjunctive and disjunctive methods. These two methods are not used to select a particular alternative but to separate the given alternatives into acceptable and unacceptable groups. All alternatives are considered acceptable so long as they satisfy the minimum cutoff criteria. In the case of the conjunctive method, the acceptable alternative has to meet the minimum cutoff for all the attributes under consideration. For the disjunctive method, an alternative is acceptable if one or more of the attributes meet or exceed the cutoff value for the attributes under consideration. 49 The two methods can be used together in decision making processes. In the case of network selection, the attributes listed in category 1 of Table 3-1 will be used as input to the conjunctive and disjunctive decision processes in order to categorize the available networks into acceptable and unacceptable alternatives for service delivery. Using the attributes from Category 1, an acceptable A N that could be further explored for suitability should satisfy all of the following: • it supports at least one of the authentication mechanisms supported by the user/terminal, and • it supports the access technology supported by the user/terminal, and • it provides access to the service the user is requesting, and • it is available in the geographic location the user is present, and • it provides the coverage level that might have been indicated by the user's mobility profile. The "and" relationship in the decision process between each of the attributes is an application of conjunctive M A D M . The disjunctive M A D M algorithm is applied to account for the fact that for many of these attributes, there are more than one alternative values and for the network to be acceptable it needs to support only one of these values. 3.4.1.1 Decreasing Inter-Technology Handoffs Since the network capabilities as well as the associated transport costs would vary across different access technologies, it could be a preferred policy to initially select the network with an aim to decrease or avoid if possible potential inter-technology handoffs during an active session. Use of a non-compensatory M A D M algorithm in the decision process can help decrease inter-technology handoffs. The user mobility profile can indicate the expected type of user mobility during a session. It is possible to have the user/terminal provide its mobility profile in step 1 of Figure 3.1 that will be used in the non-compensatory M A D M part of the decision process along with coverage area information about the network. The user can set his profile to be "mobile" or "nomadic" and accordingly the non-compensatory part of the M A D M based network selection algorithm can decide to pick only wide-area cellular networks or wireless local area networks. This is illustrated in Figure 3.2 where the two terminals are identical in every aspect and have access to both WW A N and W L A N . However because of the difference in settings of their 50 mobility profiles ("nomadic" vs. "mobile"), the "nomadic" terminal/user selects the W L A N while the "mobile" terminal/user selects the W W A N . Figure 3.2 By using non-compensatory MADM algorithm, different mobility profile settings can provide selection of a different network 3.4.2 Compensatory MADM Algorithms Application of a compensatory M A D M algorithm involves the following steps: • Identify all the alternatives and the compensatory M A D M attributes impacting the decision process, • Assign relative importance in the decision making process to each of the attributes, and • Use a compensatory M A D M algorithm to get a ranking of the alternatives. N o w we consider a general case where the selection has been narrowed down using the non-compensatory M A D M algorithm to N access networks as shown in Figure 3.1. In order to apply a compensatory M A D M algorithm to facilitate network selection, we consider the Category 2 attributes in Table 3-1. Candidate network i, NWt , from a decision making perspective can be represented by the following vector of network attributes, 51 NW,=[CB, TB; AB> U, D, J, L,] We can represent the N networks to be considered in the selection process by a matrix of their network attribute vectors as follows, NW = CB, TB, AB, U, CB, TB, AB, U, D, J, ZX J , CBN TBN ABN UN The following section describes the use of TOPSIS, a compensatory M A D M algorithm to objectively decide on the best suited solution. 3.4.2.1 Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) The TOPSIS algorithm is based on the assumption that the chosen solution has the shortest distance from the best solution but the longest distance from the worst solution. The following steps are involved in the application of TOPSIS to the network selection problem. 1. The value for each of the attribute in the matrix is normalized. For example, entries in the second column of the matrix can be normalized as follows, ( T B ) - , T B i \ "norm n rT, 5X 2. The matrix is updated with these normalized values as follows. ( N W n o J i = ( C B no rm) l ( T B n o r J l ( A B n o r m ) l ( U n o r m ) l ( D n o r m ) l ( C B n o r m ) 2 ( T B n o r J 2 ( A B n o r m ) 2 ( U n o r m ) 2 ( D n o r m ) 2 (^norm ) l (^norm^ (^"norm ) l (^"norni )2 .(^Bnorm)3 C"^norm )N ( A B n o r m ) N (^normlN (^norm)N (^norn^N 0-norm)N 3. The next step is to decide on the relative importance of each of the attributes involved in the decision about network selection. For this purpose, each of the attribute is assigned a weight "w", such that 52 W = w C B + w T B + wA B + W y + W D +Wj + wL =1 4. Using these assigned weights, the matrix from step 1 is updated as follows. WCB^CBU , v C f T B U , w„*(ABL,), VOU, wD*(DRAM)1 w/fJLJi w ^ L ^ WCB'CCBL,^ w1B*(TBram)2 w„*(ABLn)2 w^lU, ^ ( C U , w / (J „J 2 w ^ L ^ WCB^CB^ ) , w ^ f l B L j , W ^ ^ A B J , w/ilUu w/p^ 5. The next step is to find the best and worst value for each of the attribute. Depending on the attribute, the best (or the worst) value can be either the maximum or the minimum value. For example, in the case of the network utilization attribute, the best value will be the lowest one and the worst value will be the highest one. For the case of the allowed-bandwidth attribute, however, the best value will be the highest one and the worst value will be the lowest one. In the case of TB, this calculation in general can be represented mathematically as follows. Based on the description above, these equations would evaluate to best value for TB being the maximum and the worst value for TB being the minimum. TB B e s t =Max(TB w _ n o J i | i = 1,2,..N TBW o r s t = M i n ( T B „ ) i |i = 1,2,...N 6. For each of the access network under consideration (represented by a row in the matrix), the measure of separation, both for the best and worst cases, is calculated S b ( = | ( (CB w _ n o r m ) i - C B B e s t ) 2 +( (TB w _ n o r m ) | - T B B e s t ) 2 + ( (AB w _ n o r m ) , - A B B e s t ) 2 + | ( (U w _ n o r m )| - U B e s,) + ( (D w _ n o m l )| - D B e s,) + ( (J w _ n o m n )| - J B e s t ) + ((Lw_norm )i _ ^ Best) g _ / ( ( C B w - n o r m ) i " C B Worst) 2 + ( ( T B w-norm )i " T B Worst ) 2 + ( ( A B w-norm )i " A B Worst)2"1" \ ((^w-norm ) i " ^ Worst) + ( (^ w-norm )i Worst) + ((^ w-norm )i "^ Worst ) + C- i Worst ) 7. For each of the access networks under consideration (represented by a row in the matrix), its level of preference is measured. The preference level "P" of network i measured in terms of distances " S " from the best and worst solutions, respectively Seest and Sworst, is given by P _ ^Worst ^Best + Sworst 53 The preference value "P" represents an indifference curve. The term indifference curve means that a decision maker would give equal preference to any of the alternatived on the same indifference curve, i.e., with the same value of "P". Figure 3.3 illustrates some indifference curves for a system with two attributes: cost per byte and throughput. 8. The access network with the highest "P" value is selected. Highest Throughput (Best) Lowest Throughput (Worst) Indifference Curves -ve Ideal P=0.5 I I ± Highest Cost (Worst) Lowest Cost (Best) Figure 3.3 Example of indifference curve for TOPSIS while using attributes of cost per byte and throughput The computational complexity involved in calculating the Euclidean distances used in TOPSIS is minimal. Therefore the computational time and the resources needed for it should not be significant. Besides, in the architectural framework defined in Chapter 2 [1], data collection as well as the analysis involving M A D M are performed in the network and only the network rankings are conveyed to the terminal over the wireless link. This approach is resource efficient from the wireless bandwidth utilization perspective. 3.4.2.2 Assignment of Attribute Weights The assignment of weights to different parameters in a compensatory M A D M algorithm plays a key role in network selection. It is proposed that the assignment of these weights be based on the 54 interpretation o f the requested service/applicat ion requirements or the subscribed user profi le b y the subscriber 's H N operator. T w o possible types o f Q o S profiles that can be stored i n the user 's H N are thus as fo l lows: • A n overa l l user Q o S prof i le that is appl icable to a l l o f the services that the user is us ing; e.g., G o l d , S i lve r , or B r o n z e prof i le can indicate the l eve l o f Q o S that the user is expected to have based on the subscript ion. • A Q o S prof i le o f an ind iv idua l service that is appl icable to a l l subscribers o f that service; e.g., V o I P service prof i le or web b rows ing prof i le . The relative importance o f different attributes for some c o m m o n types o f services/applications or user subscriptions is described below. VoIP - Th i s is a l o w bandwidth appl icat ion that is very sensitive to delay and ji t ter but can withstand some packet losses. The transport cost factor is considered negl ig ib le because o f l o w bandwidth usage. A l s o because o f l o w bandwidth requirements, the total bandwidth and avai lable bandwidth are not significant factors. Since there is some correlat ion between the u t i l i za t ion and jit ter / delay, it is preferred to have a l o w ut i l iza t ion for the selected network. Streaming - B e i n g a mul t imed ia service, a streaming appl icat ion requires a higher bandwidth than V o I P . Therefore the available bandwidth , transport cost and current u t i l iza t ion are important factors. The service is less vulnerable to delay and ji t ter than V o I P because o f its ab i l i ty to buffer longer duration o f data before p lay back. The sensit ivity to packet loss is s imi la r to V o I P , such that some packet losses can be compensated without impact to user experience. Web Browsing - W e b b rows ing type applications require a l o w Q o S service; i.e., the importance o f u t i l iza t ion , delay, jitter, and packet loss is l o w . It does not need a guaranteed bi t rate because o f the bursty nature o f web traffic pattern. W i t h statistical traffic mu l t i p l ex ing for such type o f traffic, broadband wireless networks can del iver a reasonable customer experience even at lower average data rates. The total bandwidth and a l lowed bandwidth are therefore less c r i t ica l but the transport cost is considered cr i t ica l . Gold Subscription - T h i s indicates a premier user subscription that w o u l d a l l o w the use o f the highest l eve l Q o S independent o f the transport cost. Silver Subscription - Th i s indicates a m e d i u m pr ior i ty user subscript ion that w o u l d try to balance between the Q o S requirements and other factors such as the transport cost. 55 Bronze Subscription - This indicates a lower priority user subscription where the transport cost is significantly important compared with any QoS parameters. 3.4.3 Results of Use C a s e Scenar ios To validate the use of TOPSIS in the decision process we ran the decision process for network selection among five candidate networks. Table 3-2 provides the numerical attribute values for these five networks that were used for illustrative purposes in the decision process for network selection. They are representative of the listed example network types that a typical user could expect. For example, the cost attribute is derived on the basis of spectral efficiency of the technology and whether the technology runs on licensed or unlicensed spectrum. So the cost is lowest for an unlicensed spectrally efficient technology such as IEEE 802.1 In and it is highest for a licensed and relatively less spectrally efficient technology such as UMTS. Also the maximum estimated throughput for each of the example technologies has been used for the total bandwidth attribute. The allowed bandwidth has been assumed to be based on operator policy to rate limit its customers differently for different access technologies. Other QoS related attributes such as delay, jitter and loss represent a snapshot of the values that could exist in these networks at the time of decision. Table 3-2 Attribute values for Scenarios 1, 2,3 & 4 CB TB AB U D J L (%) (mbps) (mbps) (%) (msecs) (msecs) (per 106) NtwkUl e.g. UMTS 100 2 0.2 10 400 50 100 Ntwk#2 e.g. 802.11b 20 11 1 20 200 25 20 Ntwk#3 e.g. 802.11a 10 54 2 20 100 15 15 NtwkM e.g. 802.11n 5 100 5 40 150 30 20 Ntwk#5 e.g. 4G 307, 52 100 5 20 100 20 15 2- Scenarios 1,2. 2- Scenarios 3, 4 We considered the following use case scenarios where the service and user subscription profiles as described previously were used. 56 Scenario 1 - In this scenario the decision process is influenced by the requested service indicated by the user. The service type VoIP, streaming and web browsing were used to assign attribute weights. Scenario 2 — In this scenario the decision process is influenced by the QoS profile of the user stored in the H N . The user subscription type of Gold, Silver and Bronze were used to assign attribute weights. Scenario 3 - This is the same as scenario 1 except that the operator of Ntwk #5 has temporarily indicated a reduction of access cost to the subscriber's H N in order to attract more customers to its network. This change would influence the rankings. Scenario 4 - This is the same as scenario 2 except that the operator of Ntwk #5 has temporarily indicated a reduction of access cost to the subscriber's H N in order to attract more customers to its network. Cost/Byte Total Bandwidth Available Bandwidth Utilization Delay ] Jitter Voice Streaming Web •G—scenario 1 --><-• scenario 3 — V — scneario 1 - Managed N t w k s c n e a r i o 3 -ManagedNtwk Voice Call Streaming Media Web Browsing Figure 3.4 Results for scenarios 1 & 3 57 The results for scenarios 1 and 3 are shown in Figure 3.4, which also shows the weights assigned to various attributes. The assignments are based on the analysis of the service/subscribed QoS profiles described earlier. Similarly the results for scenarios 2 and 4 are documented in Figure 3.5. The results indicate that even with the same set of parameter values, the top candidate network can be different for different types of QoS subscriptions or service profiles; e.g., the network considered optimal for a VoIP user is different than that for a web browsing user in scenario 1. This is due to differences in the service requirements reflected by weights assigned to the parameters for QoS profiles for different services. 0.7 06 05 04 0.3 0.2 0.1 0 I Cost/Byte JTotal Bandwidth (Available Bandwidth ] Utilization I I Delay I l.litter "I l \ ' \ lLoss n Gold Silver Bronze -6— scenario 2 - -*- • scenario 4 —V— scneario 2 - Managed Ntwk scneario 4 - Managed Ntwk m. m #4 * Gold m #3 #4 Silver #1 ft m #4 #5 Bronze Figure 3 . 5 Results for scenarios 2 & 4 The network operators can influence the decision process by making their network more attractive to a certain type of users or services. In the case of scenarios 3 and 4, e.g., Ntwk #5 influences the selection by temporarily decreasing the access cost for its network to its roaming partners. A s a result when the user's FIN calculates the network rankings, Ntwk #5 becomes quite attractive for most of the services or users where the network transport cost is an important factor in the 58 decision process. This is evident by observing the results of Figures 3.4 and 3.5, where Ntwk #5 becomes the network of choice for Gold, Silver and Bronze customers along with Streaming Media and Web Browsing services. Also it can be observed that the network rankings for services/subscriptions that are not significantly influenced by the cost attribute (e.g., VoIP), are not impacted by changes in Ntwk #5's cost. 3.4.4 Sensitivity of Network Selection to Dynamic Attribute Values The compensatory part of the network selection process is primarily related to finding the network that is best suited in terms of QoS to the user or the QoS profile of the requested service. Many of the QoS related parameters used in the proposed compensatory M A D M algorithm are dynamic. A n inaccurate measure of these parameters values could result in the selection of a non optimal network. Table 3-3 Results on sensitivity analysis Scenario P - Ntwk#l P - Ntwk#2 P - Ntwk#3 P - Ntwk#4 P - Ntwk#5 VoIP Baseline 0.2912 0.7190 0.8830 0.6161 0.8570 VoIP 10% Error 0.2919 0.7272 0.8582 0.6237 0.8657 Web Baseline 0.1070 0.7725 0.8623 0.8901 0.7504 Web 30% Error 0.1251 0.7738 0.8638 0.8687 0.7517 Table 3-3 provides a comparison of the results for TOPSIS in scenario 1 as previously described, when the attribute values for delay, jitter and loss are changed by 10% for a VoIP user and when the attribute values for delay, jitter and loss are changed by 30% for a web user. The results indicate that while changing these attribute values by 10% changes the selected network for a VoIP user, even changing these attribute values by 30% does not change the result for a web user. This shows that network selection for services like web browsing or for bronze users are less 59 sensitive to the dynamic attribute information compared w i t h services such as V o I P or G o l d users. T h i s is due to the assignment o f lower weights to the dynamic attributes i n the Q o S profiles for a bronze user or web browsing . Consequent ly errors i n these attribute values have a lesser impact on network selection and the results are more reliable. 3.4.5 Network Selection with Managed IP Networks M a n a g e d IP networks typ ica l ly provide mul t ip le levels o f S L A s , w h i c h can poss ib ly be mapped to several classes o f Q o S . F o r example, the Q o S classes identif ied by different D i f f S e r v Codepoin t markings at the IP layer can be mapped to different S L A levels. F o r M A D M algori thms used i n network selection, networks that support mul t ip le levels o f S L A s or classes o f services can be mapped to mul t ip le network options i n the M A D M algor i thm, each supporting one Q o S class or S L A . T a b l e 3 - 4 A t t r i b u t e va lues fo r S c e n a r i o s 1, 2 , 3 & 4 CB TB AB U D J L (%) (Mbps) (Mbps) (%) (msecs) (msecs) (perlO6) QoS Class#l/SLA#1 10 100 0.1 30 400 100 10 QoS Class#2/SLA#2 20 100 0.5 20 200 50 50 QoS Class#3/SLA#3 30 100 1 20 100 25 50 QoS Class#4/SLA#4 50 100 2 40 50 10 100 QoS Class#5/SLA#5 50 100 5 20 10 5 100 Table 3-4 represents an access system that supports five classes o f service or five levels o f S L A s . The attribute values for use i n the dec is ion process for network selection are g iven for i l lustrative purposes. F o r example, the cost attribute i n this case is der ived based on h o w the network w o u l d treat the packets from that class o f service. Q o S Class 1, for example, gets the least preferential treatment relative to the other classes and therefore has the lowest cost. The selection o f any particular alternative w i l l map to the same phys ica l network but w i t h a different Q o S class. Therefore, w h i l e the total bandwidth o f t h e A N w i l l be the same for a l l alternatives, the values o f other parameters (e.g., delay, jitter, packet loss) are different depending upon the Q o S class. The a l l owed bandwidth has been assumed to be based on operator p o l i c y to rate l i m i t its customers us ing different Q o S classes on the access network. F o r T O P S I S , this access system is represented as five alternatives. 60 The results for the four scenarios discussed previously using TOPSIS are again exhibited in Figure 3.4 and Figure 3.5, which demonstrate how different user subscriptions or service requirements are mapped to different QoS classes or SLA levels represented by different alternatives in the TOPSIS algorithm. For example, use of TOPSIS algorithm maps a Streaming media user in Scenario 1 to QoS class/ SLA level #5 while a web browsing user maps to QoS class/SLA level #1 for the same attribute values. However a decrease in the cost attribute for QoS class/ SLA level #5 in scenarios 3 and 4 results in it being selected for all the service and subscription types. 3.4.6 Network Selection for Accessing Multiple Services Simultaneously So far network selection has been discussed for scenarios where the user accesses one or more services having the same QoS requirements. A common scenario that could arise is that a user who is already in session for a service using a network previously selected for this service now decides to start another service that may have different service requirements in terms of QoS. This would trigger a new network selection decision. Although it is possible to select two different networks and use them to access two services simultaneously, in practice such a policy can create several problems such as authentication with two networks simultaneously, excessive power consumption or possible interference of turning on two radios at the same time, and appropriate routing of service data within the device and the networks. For these reasons it is recommended that such network selection solutions be avoided. The proposed solution using the comprehensive M A D M algorithm based network selection is as follows: 1. Perform the non-compensatory M A D M part of network selection as usual for each of the individual service being requested. 2. Use a compensatory M A D M algorithm for each of the individual service requested. Networks that support multiple levels of QoS/SLAs would be mapped to more than one network options as described in the earlier section. 3. The final network selected will be based on the average of the rankings for the networks selected for the different services; e.g., a network ranked as #1 and #3 for the two services under consideration will have an average ranking of #2 for the combined service delivery. 61 There may also be other aspects of the already in-use service that can influence the subsequent network selection upon starting another service. For example the already in-use service can be VoIP that requires support for real-time traffic by the AN. As this may not be possible across all candidate networks, it may be a policy in such a situation to use the already in-use network for the new service as well. In addition, certain requirements of the new service (e.g., QoS) may force a network reselection for the service already in use. Such requirements would therefore influence the final network selection decision. There may still be other services which may not be possible to combine over the same wireless AN. Network selection in such a situation could be dependent upon provisioned operator or user policies that specify which service would take priority and therefore the delivery of the lower priority service would have to be terminated. 3.5 Conclusions Selection of an optimal access network is an important aspect of service delivery in a heterogeneous wireless system. The decision is influenced by several factors. In this chapter we have proposed a decision process for a network-assisted network selection mechanism that combines non-compensatory and compensatory M A D M algorithms. The proposed mechanism is more comprehensive than the work published earlier. Along with disjunctive and conjunctive non-compensatory MADMs, we have described the use of TOPSIS, a compensatory M A D M algorithm. The TOPSIS algorithm has been validated by applying it to a large variety of use cases. Scenarios with different services and subscribed QoS profiles have been described to observe how the proposed mechanism would work. Application of the proposed mechanism to networks with multi-level SLAs has been described. Possible approaches to network selection when a service is already in use have been discussed and solutions are proposed. The results of this chapter provide a basis for further research into the area of service delivery in a heterogeneous networking environment. 62 3.6 References [1] F . Bar i and V . C . M . Leung, "Service Delivery over Heterogeneous Wireless Networks: Network Selecton Aspects", Proc. ACMIWCMC, pp. 251-256, July 2006. [2] Q. Song and A . Jamalipour, "Quality of Service Provisioning in Wireless L A N / U M T S Integrated Systems using Analytic Hierarchy Process and Grey Relational Analysis", Proc. IEEE Globecom, pp. 220-224, November/December 2004. [3] G . Le Bodic, D . Girma, J. Bine and J. Dunlop, "Dynamic 3 G Network Selection for Increasing the Competition in the Mobile Communications Market", Proc. IEEE VTS-Fall, pp. 1064-1071, September 2000. [4] E . Alexandri, G . Martinez and D . Zeghlache, "Adaptive Joint Ca l l Admission Control and Access Network Selection for Multimedia Wireless Systems", Proc. 5th International Symposium on Wireless Personal Multimedia Communications, pp. 1390-1394, October 2002. [5] S. Hongvan, H . Chen and J. Lingge, "Intelligent signal processing of mobility management for heterogeneous networks", Proc. International Conf. on Neural Networks and Signal Processing, pp. 187-191, December 2003. [6] H.J . Wang, R . H . Katz and J. Giese, "Policy-Enabled Handoffs Across Heterogeneous Wireless Networks", Proc. IEEE Mobile Computing Systems and Applications, pp. 51-60, February 1999. [7] R. Benayoun, B . Roy and B . Sussman, "Manual de reference du programme electre, Note de Sythese et Formation," Direction Scietifique SEMA, N . 25, 1966. [8] C L . Hwang and K . Yoon , "Multiple Attribute Decision Making: Mehtods and Applications", Springer Verlag, 1981. [9] M . Zeleny, "Multiple Criteria Decision Making" , McGraw-Hill, 1982. [10] J .L. Deng, "Introduction to Grey System", Journal of Grey System, vol . 1, issue 1, pp. 1-24, 1989. 63 [11] C L . Hwang and K . Yoon , "Multiple Attribute Decision Making: A n Introduction", Sage Publications, 1995. 64 4 Enhancements to MADM Algorithms for Use in Automated Network Selection where System Parameters are Known 1 N e t w o r k convergence across different access technologies holds a promise o f enabl ing ubiquitous service avai labi l i ty . H o w e v e r it also impl ies that services have to be del ivered over a heterogeneous m i x o f access technologies. There are several technical challenges that have to be overcome i n order to provide a good service experience to users i n such an environment. The select ion o f an opt imal service de l ivery network i n the presence o f mul t ip le A N s is one o f the important issues. Th i s p rob lem o f ident i fying the best suited service de l ivery network when the user has mul t ip le networks to choose from is ca l led network selection. The p rob lem is more significant for a l l - IP networks where service experience can seriously degrade depending upon the network capabili t ies and congestion levels . It is an area o f active research and a topic o f discuss ion i n several standardization forums. The I E E E 802.1 l u W o r k i n g G r o u p currently has a draft proposal that w o u l d enable information exchange for network selection between the network and the terminal [1]. It leverages the protocol be ing developed by I E E E 802.21 for this purpose w h i c h they also p lan to use for selecting networks for ver t ical handoffs [2]. S i m i l a r l y 3 G P P is l o o k i n g into the mechanism o f network discovery and selection i n their w o r k o n S A E [3]. A l o n g ' Parts of this chapter have been published in the following F. Bari and V. Leung, "Multi-Attribute Network Selection by Iterative TOPSIS for Heterogeneous Wireless Access", Proc. IEEE CCNC, January 2007 F. Bari and V. Leung, "Use of Non-monotonic Utility in Multi-attribute Network Selection", Proc. IEEE WTS, April 2007 F. Bari and V. Leung, "Application of ELECTRE to network selection in a heterogeneous wireless network environment", Proc. IEEE WCNC, March 2007 65 w i t h the protocol and architectural aspects o f the problem, an essential component i n so lv ing the p rob lem o f network selection is def ining the opt imiza t ion objective and the a lgor i thm to be used i n the selection process. Select ion o f a non-opt imal network creates undesirable results such as poor customer experience or the use o f a more expensive network. The focus o f this chapter is on this aspect o f the p rob lem. The previous chapter described h o w the p rob lem o f network selection can be so lved us ing a comprehensive M A D M approach [4]. The a lgor i thm used for this purpose was the standard vers ion o f T O P S I S [5]. In this chapter w e consider three M A D M algorithms representing three different decis ion m a k i n g philosophies and propose several improvements to these standard M A D M algorithms [5] to make them more suitable to so lv ing the p rob lem o f network selection. The rest o f the chapter is d i v i d e d into three m a i n parts. The first part proposes an improvement to T O P S I S that helps remove ranking abnormalit ies from the results through an iterative process. In the second part w e explore the appl icat ion o f M A D M algori thms to network selection where the opt imiza t ion objectives o f the decis ion maker require the use o f non-monotonic u t i l i ty for some o f the attributes. W e show w h y G R A is better suited w h e n compared w i t h other standard forms o f M A D M algorithms for such scenarios. F i n a l l y i n the last part o f the chapter w e propose a new adaptation o f E L E C T R E [6] to solve the p rob lem o f network selection w i t h non-monotonic u t i l i ty for some o f the attributes. The k e y difference between different M A D M algorithms is that they are based o n different dec is ion m a k i n g philosophies. Th i s a l lows the decis ion maker to choose the a lgor i thm that is closest to its ph i losophy on decis ion mak ing . W e have avoided any discuss ion o n a direct compar i son amongst the various algori thms discussed i n the chapter on the basis o f their accuracy o f results. T h i s issue has been studied i n the past and the conc lus ion established that such comparisons are not possible as they w o u l d need the use o f another M A D M algor i thm. H o w e v e r , w e have compared the M A D M algorithms i n terms o f their sui tabil i ty to so lv ing the p rob lem o f network selection. 4.1 Use of Iterative TOPSIS for Network Selection in Heterogeneous Wireless Access In this section w e analyze T O P S I S , a compensatory M A D M algori thm. W e look into ranking abnormali t ies for T O P S I S and propose an iterative approach to improve the a lgor i thm's appl ica t ion to the p rob lem o f network selection. W e show that for the p rob lem o f network 66 selection, where ranking of only the top candidates is important, an iterative approach that removes the lowest-ranked candidate after each iteration can provide improved results. The approach proposed here can be used in the network-assisted network selection architecture described in Chapter 2 [7]. 4.1.1 TOPSIS C1 (w1) C2 (w2) • Cm (wm) Al a11 a21 • ami A2 a12 a22 am2 m • m • • • An a1n a2n • • amn Figure 4.1 Representation of an MADM problem with n alternatives and m attributes M A D M algorithms are used for determining the ranking of alternatives in terms o f their desirability with respect to multiple criteria that can influence the decision. For the problem of network selection, we are only interested in identifying the top ranking candidates and this can be considered as a special application of finding the ranking of all the alternatives. In order to formulate a problem as M A D M , we consider A={Ai, for i=l,2,...,«}, a set of a finite number of alternatives. Also , we consider C={Cj, for j=l,2,...,m} to be a set of attributes against which the alternatives are to be judged, and w l , w2, ...wm are the weights that represent the relative importance of these attributes. Then the M A D M problem can be represented by a matrix shown in Figure 4 .1. TOPSIS is a widely used M A D M algorithm developed by Hwang and Yoon [8]. It is applicable to problem spaces in which the attributes have monotonically increasing or decreasing levels o f utility. The algorithm calculates perceived positive and negative ideal solutions based on the range of attribute values available for the alternatives. The premise of the algorithm is that the best solution is the one with the shortest distance to the positive ideal solution and longest distance from the negative ideal solution, where distances are measured in Euclidean terms. A l l 67 solutions with the same preference level can be mapped to an "indifference curve", which shows how different values of the attributes when combined can result in the same level of preference. For the network selection problem, the representative set of attributes considered in the decision making process are Cost per Byte (CB), Total Bandwidth (TB), Al lowed Bandwidth (AB) , Utilization (U), Packet Delay (D), Packet Jitter (J) and Packet Loss (L). Their descriptions are provided in Table 3-1 o f Chapter 3. Using the attributes defined above, a candidate network N W for selection using M A D M can be represented by the attribute vector: NW = [CB TB AB U D J L] If there are N network alternatives to be considered in the selection process, their attributes can be represented in the form of a matrix as follows, NW, CB, TB, AB, U, D, J, CB, TB, AB, U, CB N TBN ABN U N N o w we apply the TOPSIS algorithm to the problem in a stepwise manner as follows. The detailed steps have been described in the previous chapter. Here we summarize them as they w i l l be referred in the chapter during analysis of ranking abnormalities. 1. The value for each of the attribute in the matrix is normalized using Euclidean normalization. 2. The matrix is updated with these normalized values. 3. The relative importance of each of the attributes involved in the decision about network selection is determined and represented by a weight. 4. Using these assigned weights, the matrix from step 1 is updated. 5. The best and worst values for each of the attribute are determined. Depending on the attribute, the best (or worst) value can be either the maximum or the minimum value. For example, in the case of the network utilization attribute, the best value w i l l be the lowest and worst value w i l l be the highest. For the case of an attribute related to the allowed bandwidth, however, the best value w i l l be the highest and worst value w i l l be the lowest. 68 6. For each of the access network under consideration, the measures of closeness/separation "S", for the best and worst cases, are calculated using Euclidean distances. 7. For each of the access networks under consideration (represented by a row in the matrix), its level of preference is measured. The preference level " P " , measured in terms of distances "S" from the best and worst solutions, is represented by the following formulation P _ Syvorst °Best ^ °Worst 8. The access network with the highest " P " value is selected. 4.1.2 Ranking Abnormalities M A D M algorithms are known to suffer from ranking abnormalities and TOPSIS is not any different. Triantaphyllou [9] has done significant work in documenting this problem for M A D M algorithms. He defines an effective M A D M algorithm to be one for which the indication of the best alternative does not change when an alternative that is not the best is replaced by another worse alternative, given that the relative importance of each decision criterion remains unchanged [9]. However, M A D M algorithms do not always exhibit this ideal behavior. It has been shown that the frequency of such ranking abnormalities depends upon the algorithm as well as the problem space (e.g., the general spread of data values for different attributes). In general, these abnormalities increase with the number of attributes and also with the number of alternatives under consideration [10][11]. In order to better understand the occurrence of ranking abnormalities in TOPSIS, we make use of a graphical representation of the method. Figure 4.2 graphically represents a decision making scenario using TOPSIS with 4 alternatives, and two attributes. The alternatives are mapped into a two dimensional space that represents the two attributes. The attribute data used have been normalized and adjusted based on attributes' importance as described by steps 1 through 4 earlier. The figure also shows the Euclidean distances of alternatives #4 and #2 from the positive and negative ideal solutions, which are calculated as described by step 5, while the distances are calculated per step 6. The two alternatives are equidistant from the positive and negative ideal solutions and therefore equal in their rankings. It is also clear from the figure that alternative #3 has the lowest ranking because of its relative closeness to the negative ideal solution as compared 69 to the positive ideal solution. Similarly alternative #1 is at the top of the ranking because of its relative closeness to the positive ideal solution when compared to the negative ideal solution. Now, if the lowest ranking alternative is removed, for an ideal M A D M algorithm, the comparison of the remaining alternatives should not be impacted. However this is not always the case and ranking based on TOPSIS can be impacted [9]. For reasons explained later in the section, with the removal of alternative #3 from Figure 4.2, both positive and negative ideal solutions as well as the normalized attribute values for the remaining alternatives will be changed. In the case of network selection, we are only interested in the top ranking candidate networks. For an ideal M A D M algorithm, inclusion of an alternative of lowest ranking in the comparison should not have any impact on the ranking of top ranking alternatives. If the inclusion of a lowest ranking alternative changes the rankings amongst top ranking alternatives, then it indicates the presence of ranking abnormality. In this situation, when we are interested in only the top ranking alternative, the removal of a candidate that causes ranking abnormality can provide more reliable results. +ve Ideal cr c CD -ve Ideal Attribute 1 Figure 4.2 Use of TOPSIS for a MADM problem with 4 alternatives and 2 attributes 4.1.2.1 Factors contributing to ranking abnormality for TOPSIS Attributes used in TOPSIS are likely to be represented in different units of measurements (e.g., dollars for transport per byte cost or milliseconds for latency). Before attribute values can be used for calculating positive and negative ideal solutions, they have to be expressed in the same units or made dimensionless. The attributes are made dimensionless by the process of normalization. There are more than one methods of normalizing the data [8]. A Euclidean normalization 70 technique is used in TOPSIS as shown in step 1, and removing or adding an alternative w i l l change the normalized attribute values for all the candidate networks. Depending upon the attribute values of the removed alternative, the impact on the normalized values for different attributes of the remaining alternatives w i l l be different. Subsequently, the calculation of weight-adjusted normalized values of step 4 w i l l also be impacted non-linearly. A s a result, the best and worst values for each of the attributes in step 5 w i l l also be changed non-linearly. [11] describes the impact on ranking of alternatives to changes in the values of attributes by performing sensitivity analysis. It shows that ranking order can change i f the values of the attributes in relationship to others are changed by a certain extent. The effect of removing an alternative from a comparison is similar to changing the attribute values. Depending upon the attribute values of the alternative being removed, the sensitivity and hence chances of a change in the rankings of the remaining alternative being considered can vary. For ranking purposes, TOPSIS uses the Euclidean distance of attributes from their respective positive and negative ideal values (as described in step 6). After the removal/addition of an alternative from the bottom of the ranking, the Euclidean distance calculations for an alternative w i l l be based upon the new normalized attribute values and the new best and worst values. A s described previously, the relative change in these new values from the old values can be different for different alternatives. Therefore the relative change in the new Euclidean distance from its prior value for any one alternative could also be different than for other alternatives. A s a result, the calculation of preference levels " P " , as defined in step 7 can in this situation provide a different ranking order than the prior one. 4.1.3 Iterative MADM Approach using TOPSIS Because of the nature of M A D M problems, it is not possible to objectively measure the accuracy of rankings obtained from M A D M algorithms. For M A D M algorithms that can exhibit non-ideal behaviors, there is a possibility that the inclusion of a bottom ranked alternative can alter the rankings o f the top ranked alternatives. I f instances of such abnormalities can be identified, then for some problem spaces it is possible to take action to remove their impact. For example, for M A D M algorithms which exhibit such a non-ideal behavior and the decision maker is interested only in the top ranking alternatives, then a comparison amongst only the likely top candidates w i l l provide more reliable rankings than a comparison involving all o f the candidates including those 71 at the bottom of the ranking. This is especially true i f the inclusion of a bottom ranked alternative causes a change in the ranking at the top. The factors identified in the previous sections are not known to cause radical changes to ranking; e.g., it is unlikely to cause the top candidate to appear at the bottom of the ranking list. While it is not possible to remove causes of ranking abnormalities in TOPSIS, the fact that in the case of network selection our interest is only in the top candidate networks can be used to improve the accuracy of the results. It is therefore possible to adopt an iterative approach whereby, in each of the iteration, the network at the bottom of the ranking is removed and the rest of the alternatives are compared. For most use case scenarios, it is expected that the total number of candidate networks in a geographic location would be in the order of ten or less, and therefore the iterative approach wi l l continue to be efficient. For a larger number of alternatives, it is possible to use a variation of the proposed approach where each time the bottom quartile (or half) of the alternatives are removed from the next comparison. The process as shown in Figure 4.3 is repeated ti l l only the top candidate networks are left. If in the final iteration the top candidates are very close in their ranking index, then it is proposed to use one of the attribute considered critical by the decision maker as a tie breaker. For example, i f the network operator is the decision maker, then cost can be the decision factor for deciding between two equally suitable networks and hence used as a tie breaker. Apply TOPSIS for Network Selection No No Remove candidate at the bottom of the ranking and apply TOPSIS to the remaining networks Apply network cost Attribute (CB) as the tie breaker Top Candidate Network Figure 4.3 Decision making process in Iterative TOPSIS approach 72 4.1.3.1 Results for use case scenario In order to validate the use of the iterative TOPSIS approach as shown in Figure 4.3, five networks were considered in our decision process. For illustrative purposes, Table 4-1 presents numerical attribute values for these networks at the time of network selection. They are representative of listed example network types that a typical user could expect and Section 3.4.3 provides explanation for the assignment of these attribute values. • Cost/Byte . . . | Total Bandwidth — fl^H Available Bandwidth . . . 1 1 Utilization . . . 1 IDRlay 1 1 Jitter . . . k W M L o s s --r j Bronze Figure 4.4 Assignment of weights to attribute values for a Bronze user subscription in the home operator network Table 4-1 Attribute values for the candidate networks CB (%) TB (Mbps) AB (Mbps) U (%) D (ms) J (ms) L (per IO6) Ntwkttl e.g. UMTS 100 2 0.2 10 400 50 100 Ntwk#2 e.g. 802.11b 20 11 1 20 200 25 20 Ntwk#3 e.g. 802.11a 10 54 2 20 100 15 15 NtwkM4 e.g. 802.11n 5 100 5 40 150 30 20 Ntwk#4 e.g. 4G 30 100 5 20 100 20 15 The weights assigned to the attributes used in TOPSIS were based on a subscribed QoS profile stored in the user's H N . A Bronze subscription was assumed in this example which indicates an overall lower priority user subscription where transport cost is significantly important compared with any QoS parameters for the candidate network to be selected. The actual weights assigned to the attributes per step 3 are shown in Figure 4.4. Figure 4.4 shows that whereas the cost per byte 73 is given a very high weight, the allowed bandwidth and total bandwidth have been assigned a weight of zero, indicating that they do not play any role in network selection for a Bronze user. Table 4-2 Results of Iterative approach for TOPSIS No. of iteration P - Ntwk#l P - Ntwk#2 P - Ntwk#3 P - Ntwk#4 P - Ntwk#5 1 0.1180 Rank - 5 0.8371 Rank - 3 0.9326 Rank - 1 0.8808 Rank - 2 0.7368 Rank - 4 2 0.4101 Rank - 3 0.8022 Rank - 2 0.8854 Rank -1 0.1158 Rank - 4 3 0.1182 Rank - 3 0.6719 Rank - 2 0.8759 Rank -1 4 0.1886 Rank - 2 0.8114 Rank -1 Table 4-2 shows the results of the proposed iterative implementation of TOPSIS. The algorithm was run four times, each time removing the candidate network at the bottom of the ranking. On the first iteration it is observed that the ranking of the top two candidate networks is reversed indicating that network at the bottom had some influence on the overall ranking. Further iterations, however, do not change the ranking any further. The results therefore show that removal of the alternative at the bottom of the initial ranking from subsequent comparisons can result in the stabilization of the top ranked alternative. For an ideal M A D M algorithm, the inclusion or exclusion of a bottom ranking alternative should not change the ranking at the top. The results of the first iteration indicate that the initial top rankings obtained using TOPSIS were not reliable as they were influenced by the lowest ranking alternative. Therefore TOPSIS did not exhibit an ideal behavior in this case. Comparing only the more likely candidates by removing the bottom ranked candidate (that would not have been selected) after the first iteration removed this ranking abnormality and provided more reliable results. It should be noted that it is not possible to objectively measure the accuracy of ranking obtained by an M A D M algorithm and it is possible that two different M A D M algorithms provide different rankings with the same attribute related data and decision criteria. The iterative TOPSIS approach described in this section is, however, 74 capable of overcoming ranking inconsistencies when the TOPSIS algorithm is used to select the top candidate. 4.2 Utility Aspects of Attributes in Application of MADM Algorithms to Network Selection In this section we view various M A D M algorithms for their application to network selection in the light of optimization objectives of a decision maker requiring the use of non-monotonic utility for some of the attributes. We show why G R A is better suited when compared with other standard forms of M A D M algorithms for such scenarios. 4.2.1 Deficiency in Current Studies The need to provide a consistent service experience requires selection of optimal delivery network. The topic is of special interest for multimode IP devices where services can be delivered over a variety of wireless access technologies under varying network conditions. Several factors related to network capabilities and QoS conditions influence the network selection decision process, e.g., bandwidth, delay, jitter, and packet loss. This makes the use of deterministic decision making tools such as M A D M algorithms [8] possible. Their use has been previously considered, e.g., for network selection in a heterogeneous wireless networking environment [12][13][14][15], to derive a ranking of the available networks in terms of their suitability. The highest ranking network is then selected as the best suited network. The prior work however failed to provide a comparison amongst the M A D M algorithms for use in network selection. This has primarily been because of the fact that to provide a comparison in terms of ranking accuracy one would need another M A D M algorithm which creates a paradoxical situation. Nonetheless it is possible to evaluate these algorithms in terms of their appropriateness for application to the problem space of network selection. So far it has been difficult to perform such an evaluation of the algorithms because of a rather simplistic assumption about optimization objectives of the decision maker. Prior studies ignored a possibly diverse range of optimization scenarios based on service and user types that could exist and hence can help in comparing the suitability of the algorithms. The following section describes the concept of non-monotonic utility for the attributes that can help meet a wider variety of optimization objectives. We propose that this concept be leveraged in assessing the suitability of M A D M algorithms to network selection. 75 4.2.2 Use of Non-monotonic Utility for Attributes in Network Selection In general, as part of the decision process, the M A D M algorithms associate a measure of suitability or appropriateness, hereafter called utility, with the individual attribute's value. The utility is said to be monotonic i f the measure of suitability associated with the attribute shows a monotonic increase or decrease with an increase in the attribute value. Otherwise it is said to be non-monotonic. Figure 4.5 shows a simple decision making scenario with one attribute, namely delay, and two networks. The delay attribute is shown with possible monotonic and non-monotonic utility for different service types. The monotonic utility represents optimization objectives where the network with the least delay value, i.e., Ntwk #1 w i l l be selected for all service types. The non-monotonic utility of the delay attribute in Figure 4.5 represents the optimization objective of the decision maker where it would like to use the network that provides a delay closest to the service's delay requirement and not necessarily the network with the least delay. This type of optimization objective would result in the selection of Ntwk#l for VoIP service and Ntwk#2 for streaming media and web browsing services. The decision maker may desire this type of optimization for policy reasons such as load balancing across different A N s or for keeping the best network for higher QoS services and sessions that it can expect to serve. It would be similar to the policy of an airline that decides to fly with some first or business class seats empty and not to upgrade people from the economy class, with the knowledge or the hope that it would be able to f i l l those seats with full fare business or first class customers at the next stop. In the case described above only one attribute was considered. For the case of multiple attributes, the overall ranking is either obtained via adding the utility associated with each of the attributes or by comparing the utilities for the attributes individually in the decision process. In prior applications of M A D M to the network selection problem [12] [13] [14] [15], the impact of different optimization objectives and hence the use of non-monotonic utility in the decision process were not studied and a general assumption about the use of the best network irrespective of service requirements or user type were considered. This implies a monotonic utility for all attributes. While some of the decision related attributes can be considered to have monotonically increasing or decreasing utilities, in reality the overall optimization goals of the decision maker may require a combination of monotonic and non-monotonic utilities for different attributes that are taken into consideration during the decision process for network selection. Associating a monotonic 76 increasing or decreasing utility in general with all attributes is therefore a simplistic assumption that would limit the scope of the types of optimization available to the decision maker [13][15]. Streaming Web VoIP Media Browsing I Ntwk #1 Ntwk #2 Delay Attribute Figure 4.5 Graphical representation of a simple decision making scenario with one attribute and two networks. A n example of an optimization objective can be to find the network that along with other factors (such as cost, etc.) also has the best QoS characteristics from among the list o f available networks. In this case the utility of QoS attributes can be considered to be monotonic. However, under a different deployment scenario, the decision maker may wish to assume a non-monotonic utility for some of the attributes considered in the selection process. A n example would be an optimization objective to distribute network traffic across different access networks by selecting the access network offering a QoS closest to that being requested by the service, and not the network that may have the best QoS that far exceeds the service's QoS requirements. Here we analyze the suitability of some of the commonly used M A D M algorithms for the problem of optimal network selection, where not all the attributes considered in the decision making process have a monotonically increasing or decreasing utility. Such network selection scenarios w i l l be quite common in future heterogeneous wireless networking environments used for delivery of both real time and non-real time services. The algorithms considered are TOPSIS [8], E L E C T R E [8][6] and G R A [16]. Amongst these M A D M algorithms, G R A is found to be the 77 most suited for optimization objectives requiring both monotonic and non-monotonic utilities of attributes. Using a heterogeneous wireless networking environment as an example, we demonstrate how the algorithm can be implemented to achieve different optimization objectives, and its impacts on the resulting network ranking for network selection. 4.2.3 Comparison of MADM Algorithms for Use with Non-monotonic Utilities of Attributes TOPSIS [8] calculates the perceived positive and negative ideal solutions based on the range of attribute values available for the alternatives, and selects the best solution as the one with the shortest distance to the positive ideal solution and longest distance from the negative ideal solution. The distances are measured in Euclidean terms. Because o f the concept of positive and negative ideal solutions based on Euclidean distances, a standard implementation of TOPSIS requires that the utility of each attribute under consideration increases or decreases monotonically. E L E C T R E [6] [8] performs pairwise comparisons among all alternatives for each one of the attributes separately in order to find outranking relationships between the altenratives. In its standard implementation, the method removes the less desirable alternatives. Using a complementary analysis it is possible to select the best suited alternatives. Since the comparison is directed amongst the available alternatives there is no concept in E L E C T R E of comparing the alternatives to some reference set of values to see how close the parameters values are to the desired values. The notion of a monotonically increasing or decreasing utility of an attribute is inherent in a direct comparison amongst the alternatives, which makes the standard E L E C T R E algorithm not well suited for use with attributes having non-monotonic utilties. G R A is another very popular decision-making technique that is based on the Grey System Theory. Originally developed by Deng [16], the Grey Systems Theory has been applied to solve a variety of real life problems in many fields ranging from business, operations research, and engineering, to social sciences. One of its areas of application has been in M A D M , where multiple attributes influence the decision process. Unlike M A D M algorithms described earlier, G R A uses a reference set of attribute values for comparison with attribute values of the alternatives. It has been applied in the past [15] [13] [14] to solve the problem of network selection in a heterogeneous networking environment. The problem of selecting an optimal network in a heterogeneous networking environment, however, is quite complex and it is possible to apply G R A differently to the problem. For example, the utility aspects of the algorithm were not 78 explored in the prior work and a single reference network was constructed that implied a monotonic utility for all the attributes for all service or user types. Because of its ability to use reference attribute values in the decision process, G R A can be applied where the optimization objectives require a non-monotonic utility for some of the attributes and monotonic utility for the others. A s described in the later section, such an implementation of G R A would use multiple reference networks. The network rankings in this case could be quite different than i f a monotonic utility was considered for all the attributes. It is clear that for optimization scenarios where a utility does not increase or decrease monotonically with an increase or decrease in attribute value, standard implementations of M A D M algorithms such as TOPSIS and E L E C T R E w i l l have limited applicability. Other simpler compensating M A D M algorithms such as S A W (Simple Additive Weighing) [5] and W P M (Weighed Product Method) [5] also have similar limitations because of their inherent assumption about monotonic utilities of attributes. These M A D M algorithms do, however, allow assignment of different weights to the attributes before they are combined to calculate the ranking indices. This general feature of M A D M algorithms can be used to apply algorithms such as TOPSIS to network selection for different service types as described in the previous chapter. For example, services that are less sensitive to QoS and more sensitive to the transport cost (e.g., web browsing), could have higher weights assigned to the cost attribute and lower weights assigned to the QoS attribute. This would allow alternatives that are closer to the positive ideal solution in the transport cost (i.e., lowest in cost) to receive more importance in decision making than network alternatives with QoS related attribute values closest to the positive ideal solution (i.e., best QoS attribute value). However, it is important to note that this type of applicability would provide a different optimization than trying to find the network alternative that is closest in QoS attributes to the requested service's QoS attributes. A s stated earlier, G R A favors a selection that gives a closest match to a set of reference data values. This process inherently supports the notion that these reference values do not necessarily need to be the best or the worst values associated with the attributes. In addition, it also has the ability to assign different weights to different attributes. These two tunable aspects of G R A when combined provide a much better mechanism to achieve optimization objectives involving attributes with non-monotonic utilities. The rest of this section describes the application of G R A to the problem of network selection with some attributes having non-monotonic utilities. 79 4.2.4 Theory of Grey Relational Space GRA is based on the concept of Grey Relational Space (GRS). GRS (X,Y) describes a v e Y relationship Y between reference data values Xo and sequence of data values X. So if y , x ^ X ^ e X , s u c h t h a t x 0 = x0(1) x0(n) a n d x , = x,(1) x,(n) t h e n y ( X o W . x , ( k ) ) w o u M r e p r e s e n t a GRS at point k provided the axioms documented in [16] are satisfied. In addition, a Grey Relational Grade for a series i could then be represented as 1 n y ( X o . X i ) = - X y ( x o ( k ) . X i ( k ) ) . n k=i A representation of yW^XjM)^^ s ati sfies all of the axioms in [8] is represented as: mirii mink | x 0 (k) - x, (k)|+^maxi maxk |x0(k)—Xj(k)| y(x0(k),Xi(k))=-x0(k) - x,(k)|+^maXi max k |x0(k) -x,(k)| where ^ is called a distinguished coefficient and has values ranging between zero and one. y(x0(k),x,(k)) i g c a l l e d t h e G r e y R e l a t i o n a i Coefficient (GRC). When applying GRA to ranking of networks while selecting a network, GRC is a measure of how closely a network's attributes match the reference network's attributes. In this respect it represents the overall utility that takes into consideration individual values of all the attributes. The higher the value of GRC, the closer would be the network to the reference network. Hence, for the purposes of network selection the GRC is equivalence of a utility that takes into consideration all individual attribute values. 4.2.5 Application of GRA Adapted to Network Selection with Non-monotonic Utility For the network selection example described here, the same set of attributes have been used as provided in section 4.1 and described in [4]. Using these attributes, a candidate network NW for evaluation by GRA can be represented by the follow attribute vector, NW = [CB TB AB U D J L] If there are N alternative networks to be considered in the selection process, they can be represented in the form of a matrix of network attributes as follows, 80 "CB, TB, AB, U, CB, TB, AB, U, D, D, J, J , Li L, NW, = CBU TBU ABU U, A reference access network is needed for application of G R A . In the case of monotonically increasing or decreasing utilities for the attributes, this reference network can be developed by using the maximum or minimum value of the attributes. In this case there w i l l be only one reference network. However i f there are services that have different QoS requirements (VoIP, streaming, web browsing etc.), or the user are of different categories (bronze, silver, gold) then the decision maker can use a different reference network for each one of the categories. A reference network in this case can be created based on the information about the user / terminal preferences, e.g., indication of requested service, and based on the user profile in the home network, e.g., the subscribed QoS. These multiple reference networks would result in non-monotonic utility for some of the attributes. Table 4-3 shows four different reference networks used in the example. The reference network i for a particular service or user type can therefore be represented as follows. (NWJ,=((CBJ, (TBJ,"(ABJ, (UJ, (DJ, (JJ, (LJ.) The units of measurement for the attributes such as cost, bandwidth and delay w i l l be different. In order apply the algorithm without having the artifacts related to different units of measurement impacting the results, the attributes w i l l have to be made unit-less before they can be directly compared or combined during the calculations. This process is called normalization [8]. Using these normalized attribute values, an updated matrix is created as follows, CB, TB, AB, CB, TB. CBU TBh AB, ABk D , L, L, 81 The reference network i's attributes are also normalized and a normalized reference network vector is created as follows ( N W . J , = ( ( C B J , ( T B ^ ) , ( A B J , ( U ^ ( D . J , ( ] „ , ) , (L^),) If the reference attribute values lie outside of the attributes values for the alternatives under considerations, then calculation of the maximum and minimum values to be used in the normalization process should include the reference values as well. Distance vectors are calculated for attributes of each access network under consideration by taking the absolute difference between reference network attribute and the candidate network attribute. For example, in the case of TB for network i, the distance value from reference network j is calculated as follows (ATB)I H O B J J - T B J The matrix of distance value for each of the attributes for the N networks under consideration can therefore be created as follows, ( A ^ ) , ( A J , ( A J , (A,), (A,), (A,), (\), (Aa^  (ATB)2 (AAB)2 (Aj)2 (AO)2 (Aj)2 I ( 0 = ( A J N ( A J N ( A J N (AJN (Ao)N (AJ)n (AJ N The GRC is a measure of the similarity of an attribute to its reference value. It is calculated for each of the matrix entries. For example, in the case of TB, it is calculated as follows (GRC T B ) j = Amin-i-^Amax ( A J i + ^Amax where £ e t0,1', and Aminand Amax can be calculated as follows A max = max(ACB i + ATB, + AAB, + AUi + AD, + AJ } + AL,) i A min = m i^ACB, + ATB, + AAB, + AUi + + A D i + A j i + A L i ) The next step is to consider the relative importance of each of the attributes in the decision about network selection. For this purpose each of the attribute is assigned a weight "w" such that 82 = 1 w/fGRCj), wL*(GRCL), ' w/fGRCA wL*(GRCL)2 [w^^GRCJ, wTC*(GRCTC)N W„*(GRC«)N VfGRC^ wD*(GRCD)N w / ( G R C ^ wL*(GRCL)N _ Using the GRC matrix thus calculated, the Grey Coefficient for each of the candidate network is calculated as follows. (GRCNW) j = wCB * (GRCCB), + Wjg * (GRCTB\ + wAB * (GRC^ ), + W u '(GRCu), +wD *(GRCD), +Wj *(GRCJ)1 +wL *(GRCL), The network with the highest value of Grey Coefficient is considered to be the best network. 4.2.6 Evaluation of Using Non-monotonic Utility in a Heterogeneous Wireless Network Environment In order to evaluate the impact of different optimization objectives we consider a network selection scenario with five networks. For each of these networks, the attribute values to be used in the decision process are shown in Table 4-1. These attribute values represent a snapshot of network related information at the time of decision. The values in Table 4-1 are for illustrative purposes and are representative for listed example network types that a typical user could expect. For example, UMTS and 4G networks run on licensed spectrum and therefore their relative cost per byte is considered higher than that for unlicensed spectrum. The cost is also shown relatively higher for technologies with less efficient physical layer coding schemes (e.g., 801.11b vs 802.1 In). Similarly the total bandwidth represents the estimated maximum throughput of a particular technology. In our case we address the network selection problem for three distinct types of services, namely VoIP, streaming media and web browsing. Each of these service types has its distinct set of QoS requirements. In the following subsection we describe how to use an adapted version of GRA for the scenarios under consideration. W = W T B + W A B + Wu + W D + Wj + W L = The weighted GRC matrix is obtained as follows W C B ^ G R C C B ) , vVtGRCJ, w ^ G R C J , w U * ( G R C U ) 1 w D * ( G R C D ) , W C B ^ G R C J , w R A * ( G R C T O ) 2 w ^ G R C J , w U * ( G R C U ) 2 w D * ( G R C D ) 2 G R C = 83 4.2.6.1 Setting up GRA for Network Selection The GRA algorithm can be applied in more than one ways to the problem of network selection. However it is well suited to handle diverse optimization objectives including those requiring non-monotonic utility of attributes. Figure 4.6 shows three different ways of application of the GRA algorithm. We recommend the third approach as described earlier as it provides the maximum flexibility for tuning the algorithm to different optimization objectives. A two step process for tuning GRA is proposed below. Determine reference attribute values for different service or user categories Based on the optimization criteria derived from the QoS requirements for the services or user types, reference networks are created by the decision maker (e.g., the user's HN) It will be towards these reference values that the GRA will try to find a closest match from a given list of alternative networks. This step relates to addressing the non-monotonic nature of the utility for some of the attributes under consideration. For example, the VoIP reference network's attribute values would reflect higher QoS requirements compared to a reference network for web browsing. Creating reference networks is a one time event and can therefore be provisioned into the decision process. Reference Value Reference Reference Value #1 Value #2 Reference Value #3 Atribute Atribute Wts #1 Wts #2 Atribute Wts #3 Atribute Wts ( 1 1 1 C D Reference Reference Value #1 Value #2 Reference Value #3 Atribute Atribute Wts #1 Wts #2 Atribute Wts #3 Figure 4.6 Three ways of using reference attribute values and attribute weights with GRA algorithm. Approach 3 is recommended in network selection. 84 Determine attribute weights for different service or user categories To allow further tuning of the GRA algorithm to the optimization objectives, the weights (i.e., the importance) assigned to the attributes used in decision making are adjusted for each type of service. The weight assigned reflects the relative importance of an attribute for that service or user type. For example, the cost attribute would carry a relatively higher weight for streaming type service when compared with VoIP service. The process of determining attribute weights is also a one time event and can be provisioned into the decision process. Table 4-3 Reference attribute values CB(%) TB (mbps) AB (mbps) U(%) D(ms) J(ms) L (per 106) Best QoS 5 100 5 10 100 15 15 VoIP 5 100 0.02 10 100 15 15 Streaming 5 100 1 10 400 50 50 Web Browsing 5 100 . 0.1 10 1000 100 100 Table 4-4 Assignment of attribute weights Attribute weights usedfor scenarios J, 3, 4 and 6 CB TB AB U D / L VoIP 0.05 0 0 0.2 0.3 0.3 0.15 Streaming 0.2 0.15 0.2 0.2 0.1 0.1 0.05 Web Browsing 0.5 0.05 0.15 0.1 0.05 0.05 0.1 Attribute weights usedfor scenarios 2 and 5 CB TB AB U D / L VoIP, Streaming, Web Browsing 0.2 0.15 0.2 0.2 0.1 0.1 0.05 Table 4-3 shows four reference networks that were used in the evaluation. The values for CB, TB, A B and U attributes for the network are considered the type of attributes that the operator in 85 this scenario would like to optimize for the best value available amongst the alternatives. For example, by adopting the utilization of the lowest utilized alternative network as a reference attribute value, the decision maker could help improve the utilization for under utilized alternative networks and hence have an improved balance of traffic loads across different networks. However, this may not be always the desired optimization criterion and other decision makers may like to have different criteria for selecting this and other attribute values. So the first reference network is created from the best values for the attributes from amongst the alternatives networks. The remaining three reference networks use a combination of the best attribute values (for C B , T B and U) and reference values derived from QoS requirements for the service types (for A B , D , J and L ) . The entries in Table 4-3 show that some of the reference values generated from specific service types lie outside the range of attribute values for the network alternatives under consideration. This can potentially cause a problem in the normalization process. In order to avoid possible ranking abnormalities, the normalization process is modified by appropriately adjusting the minimum / maximum values used in order to include reference values as well . The assigned weight for each of the attributes for different service categories considered in the evaluation is shown in Table 4-4. For example, the importance of the transport cost was considered high for web browsing as compared to VoIP. To evaluate the impacts of the assigned weights, a set of scenarios was also evaluated where only one weight distribution was used for all the different service types. Scenario 1 uses the best values as the reference for all attributes but different weights for different service types. This is shown in the first approach of Figure 4.6. Scenario 2 only changes the reference attribute values while keeping the same attribute weights for all service types. This is shown in the second approach of Figure 4.6. Scenario 3 calculates ranking when the reference attribute values as well as the attribute weights were changed for different services as shown in approach 3 of Figure 4.6. For example, in the case of streaming media type service, it first compares the network attribute values with the reference values specific to streaming media type service to find the degree of match for each of the individual attributes. Then based on the emphasis that should be placed on the degree of match for each of the attributes, a weight is assigned to it as explained earlier. 86 0 8 Sj 0 6 IB > 013 UMTS 11b 11a 11n 4G Streaming Media U M T S 11b 11a 11n 4 G Web Browsing 0 8 § °* 0.2 Mum U M T S 11b 11a 11n 4 G Streaming Media UMTS 11b 11a 11n 4 G Web Browsing 0 8 J O B t O _ 0.4 UMTS l i b 11a 11 n 4G Voice Call U M T S 11b 11a 11n 4 G Streaming Media UMTS 11b 11a 11n 4 G Web Browsing Scenario 3 Figure 4.7 Results for network detection for three possible configurations of GRA. Configuration 3 is preferred as it provides maximum flexibility to fine tuning the algorithm to optimization objectives. Figure 4.7 shows the results for all the scenarios that were evaluated and hence shows the impact of using multiple reference networks and attribute weights. A s explained earlier it is not possible to directly evaluate results of MADM algoritlim in terms of accuracy. However since scenario 3 uses a combination of reference values and attribute weights to distinguish amongst different service types, it w i l l provide a more balanced approach with results closest to optimization objectives o f the decision maker i f the intent was to select the network closest in characteristics to the reference network for that service type. B y comparing results for scenarios 1 and 3 it is also apparent that using different reference values for different service types actually does impact the ranking. Similarly a comparison of scenarios 1 and 2 shows that the network rankings are 87 impacted if different distribution of attribute weights are used for each service type as opposed to a single distribution of attribute weights for all services. 4.3 Application of ELECTRE to Network Selection in Heterogeneous Wireless Access The ability to provide a good and consistent customer experience in QoS demanding applications such as (VoIP or streaming media would depend upon the ability to select the most optimal delivery network. The decision of network selection is influenced by the optimization objectives of the decision maker. Selection of a non-optimal network can result in problems such as the unnecessary use of expensive access types or poor service experience. 4.3.1 Evaluating MADM Algorithms for Use in Network Selection The decision process for the selection of a service delivery network takes into consideration several factors related to, e.g., access network capabilities, the current network conditions and transportation costs. Researchers have considered the use of M A D M algorithms to rank the candidate networks in a preference order. There are numerous types of M A D M algorithms with [5] documenting thirteen of them. Several alternate M A D M algorithms can be suitable for solving a decision problem and the decision maker in this situation can be faced with the task of selecting the most appropriate method from amongst a number of feasible methods. Classification of M A D M algorithms into categories [5] can help to eliminate the algorithms in categories that are not well suited to the problem space, but this process does not provide the most suited algorithm. It is conceivable that a suitable M A D M algorithm may be selected for a particular decision problem based on one or both of the following criteria. 4.3.1.1 Accuracy of the results obtained from an algorithm For a variety of reasons different algorithms, when applied to the same problem under the same assumptions, can result in different rankings of the alternatives. In such scenarios it is not possible to objectively rank the M A D M algorithms for their ranking accuracy as it would require the use of another M A D M algorithm to get such a ranking. For this reason it has been found difficult to use the accuracy of the results as a criterion in selecting a specific type of M A D M algorithm. 88 4.3.1.2 Appropriateness of applying the algorithm to the problem Because of differences in the approaches used by different M A D M algorithms, a direct comparison amongst them is difficult. It has been proposed in the past that a method which is capable of solving the decision problem and whose decision making philosophy reflects the values of the decision maker can be considered to be the best suited. Decision makers in general prefer deterministic algorithms that provide reliable results based on a simple and easy to understand philosophy. Here we describe the use of ELECTRE [5] [6], a type of M A D M algorithm, to the problem of network selection. ELECTRE algorithms perform a pair-wise comparison amongst the alternatives using each of the attributes under consideration, an approach that is very popular with decision makers because of its deterministic nature and simple philosophy. Other M A D M algorithms such as TOPSIS [5][15] and SAW [5][15] have different decision making strategies but share the trait of simplicity in their philosophy with ELECTRE. Compared with these M A D M algorithms, GRA algorithms [12][13][15] are more recent and the philosophy behind them are less intuitive and more complex. It is based on the Grey Systems Theory, which can best be compared to fuzzy mathematics and probabilistic decision making approaches. GRA provides a measure of similarity of a set of values to a set of reference values. The concept of reference values, as will be discussed in a later section, is very useful in network selection. Other M A D M algorithms, such as ELECTRE, TOPSIS and SAW, do not have this capability as their comparison processes assume a monotonically increasing or decreasing utility or level of importance associated with each attribute value. Therefore despite the much more abstract decision philosophy of GRA, it has been applied to the problem of network selection. The standard ELECTRE algorithm as indicated above has some shortcomings that if properly addressed would make it very attractive for application to the problem of network selection because of its simple decision making philosophy. ELECTRE assumes a monotonic utility and does not provide a complete ranking of all the alternatives either, which would be needed to find the top ranking candidate network. In this section an alternative approach to apply the ELECTRE algorithm has been developed so that it now provides a complete ranking of the networks under consideration. The algorithm has also been modified to make it suitable for application to scenarios where the utility of some attributes is non-monotonic. This change allows the application of the algorithm in scenarios where the decision maker would like to optimize the network selection to select the alternative that has attributes closest to a reference set of attribute 89 values. For example, it may be desired by the decision maker to select for web browsing the network that is not the best alternative from a QoS or cost perspective, but has attributes closest to the reference values for a desired network for web browsing as perceived by the decision maker. The modifications to the ELECTRE algorithm described in the next section would allow such selections and make it very well suited for ranking candidate networks for network selection. 4.3.2 Application of Modified ELECTRE to Network Selection ELECTRE was developed by Bernard Roy [6] in the 1960s as a practical decision making tool and has found vast applications in engineering decision making problems. The method performs pair-wise comparisons among alternatives for each one of the attributes separately to establish outranking relationships between the alternatives [5][17]. In order to formulate network selection as a M A D M problem the factors impacting the decision process have to be determined. The attributes used in the network selection decision process here are the same as provided in Section 4.1 and described in [4]. Using these attributes, from a decision making perspective the attributes of the i-th candidate network can be represented by a vector as follows, NW, = [CB, T B , A B , U, D, Ji L , ] For N alternative networks to be considered in the selection process, a matrix of network attributes can be formulated as follows, NW CB, TB, AB, U, CB, TB, AB, U, D, D, CBN TBN ABN UN In order to best match the optimization objectives described in the previous section, we modify the basic ELECTRE method by utilizing a reference network, which can be considered to be an access network that has a desired set of attribute values given in the reference attribute vector. This reference attribute vector is used to adjust the raw attribute values for the alternative networks before they are compared. The source of reference attribute values can be different depending upon the decision process used. For example, it is possible that the information is provided by the user terminal via indication of the service it wants to initiate. In another scenario 90 the information can be provided by the user's H N operator via its knowledge about the subscribed QoS in the user profile. We can represent this reference access network as, NW r e f =(CB r e f TB r e f AB r e f U r e f D r e f J r e f L r e f ) In the proposed modification to the algorithm, the value of each of the attributes in matrix NW is compared with a corresponding reference attribute value. A n absolute difference between the two values is taken to calculate a new matrix as follows. In standard E L E C T R E algorithm this step would be skipped with the assumption of a monotonically increasing or decreasing utility for the raw attribute value. ICB.-CBJ mVTBJ |ABrABref| IU.-UJ |D rDJ l^-JJ |LrLr e f| " |CB2-CBJ |TB2-TBJ |AB2-ABref | |U2-UJ |D2-Dref| |J2-Jref| |L2-Lref| |CBN-CBref| |TBN-TBref| |ABN -ABref | |UN-Uref| |DN-DJ |JN-Jref| |LN-Lref| _ In order to remove the impact of use of different measurement units (e.g., dollars/byte for transportation cost vs. milliseconds for latency or jitter) the attributes represented in the matrix have to be normalized. Since in the proposed modification to the E L E C T R E algorithm, the raw attribute values have been adjusted with respect to the reference attribute values, it can now be assumed that for the adjusted values, the larger the attribute value, the farther it is from the desired or the reference value. In other words, all attribute values can now be considered to have a monotonically decreasing utility. Since a lower value for an adjusted attribute is considered an indication of a better network in the selection process, each attribute X j in row i of a specific column of the matrix can be normalized as follows, maxfX,} - Xi X = 1 = 1 1 max{X i } - min {XJ j=1...N ' j=1...N 1 A normalized matrix with these normalized values as its elements is created as follows. 91 TB, AB, 0, D, J, L, TB2 AB2 U2 D2 J2 L2 TBN ABN UN DN JN LN J During the process of overall comparison of the alternatives, the impact of pair-wise comparison of different attributes is summed up. This summation should take into consideration the relative importance of each of the attributes involved in the decision about network selection. The information about the relative importance of the attributes can have similar sources as described earlier in generating a reference attribute vector. For example, a user may request the use of VoIP service whereby the relative importance of the transportation cost and total bandwidth is considered low because of VoIP being a low bit rate application. However factors such as latency, jitter are quite important for the VoIP type service. On the other hand the weight related information can also come from the user profile that can, e.g., indicate the user to be a Bronze user and hence assign a higher weight to the cost attribute and lower weights for the latency and jitter attributes. Therefore, depending upon the information about the service to be used or the user QoS profile, the j- th attribute is assigned a weight Wj , such that W = w C B +w T B +w A B + +w D + W j +w L =1 Using the assigned weights, an updated matrix is calculated as follows. ~wCB*CB, wTB*TB, wAB*AB, w/U, wD*D, w/J, wL*C, wCB*CB2 wTB *TB2 wAB*AB2 W u * U 2 wD*D2 w/J 2 wL*L2 w/J N wL*LN J In order to compare the network alternatives, the concept of concordance and discordance has been introduced in E L E C T R E , which are measures of satisfaction and dissatisfaction of the decision maker when one alternative is compared with another. Thus concordance and discordance sets are calculated, where a concordance set (CSet) provides a list o f attributes for which an alternative network under consideration is better than another alternative network it is 92 NW = CB, CB, CBk NW.. wCB*CBN wTB*TBN wAB*ABN Wu*UN w0*DN being compared with, and a discordance set (DSet) on the other hand provides a list of attributes where the alternative network under consideration is worse than the compared alternative. For example when network 1 is being compared with network 2, the concordance set C S e t i 2 is the subset of all attributes that indicate that network 1 should be preferred over network 2, and the discordance set D S e t i 2 is the subset of all attributes that indicate a preference of network 2 over network 1. Mathematically this can be represented as follows, C S e t 1 2 ={j:(NW n o r m) 1 J >= ( N W n o r m ) 2 J } D S e t 1 2 = { j : ( N W n o r m ) 1 J < ( N W n o r m ) 2 J } Using the concordance and discordance sets, corresponding matrices are constructed. ELECTRE calculates the elements of concordance matrix C as follows, jeCSet k l The concordance matrix C can be represented as, - C, '12 .^N1 °N2 '1N '2N The entries for the concordance matrix are not defined for the diagonal. Similarly, ELECTRE defines the elements of discordance matrix as follows, Z | ( N W n o r m ) k j - ( N W n o r m ) l j jgDSety X | ( N W ^ ) ^ N W n o n n ) , Similarly, the discordance matrix D can be represented as, D = D 12 '21 DNi D N 2 D 1N '2N 93 The entries for the discordance matrix are also not defined for the diagonal. Two possible approaches can be considered to proceed further in decision making based on the ELECTRE algorithm. 4.3.2.1 Approach 1 In the standard ELECTRE method, the outranking calculations are performed as follows. The concordance and discordance dominance matrices are determined. The concordance dominance matrix is calculated using a threshold value for the concordance index. A way to determine the threshold value, Qhreshoid, is to use the average concordance index as follows, N N r threshold X Z C KI k=1 1=1 N * ( N - 1 ) Using the Cthreshoia value, elements of the concordance dominance matrix, Cdom, are calculated as follows, ( ^ d o m ) k l = 1 • ^ k l > = ^threshold ( ^ d o m )kl = 0 • C k l < C threshold The discordance dominance matrix is calculated using a similar threshold value, Dthrcshoid- This value can be calculated using a similar formula as follows, N N za J k l n _ k=i 1=1 threshold N * ( N-1) Using the Dthreshoid value, elements of the discordance dominance matrix, D d 0m, are calculated as follows, ( ^ d o m )kl = 0 • ^ k l > = ^ threshold ( ^ d o m ) k l = ^ • ° \ l < ^ threshold The aggregate dominance matrix, A,jo m, is calculated as follows, (Adorn )kl = ( C d o m )k| ( D d o m ) w 94 The aggregate dominance matrix is able to provide partial preference ordering of the access networks under consideration. For example if (Ad o m)i2 = 1, then this would imply that network 1 is preferred over network 2 when both concordance and discordance criteria are used. A problem with this approach in formulating the ELECTRE algorithm is the arbitrary selection of threshold values. These threshold values can significantly impact the outcome of the algorithm. In addition the results of this ELECTRE method do not provide a complete ranking for all the alternatives. 4.3.2.2 Approach 2 The complementary analysis in [8] tries to address the shortcomings of approach 1. A new parameter Q, called the net concordance index is calculated. Ci is a measure of the dominance of an alternative / over other alternatives when compared with a measure of dominance of other alternatives over the alternative i. It can be calculated as follows, N N c,=ic,-2:c, i=i i=i Similarly, the term net discordance index Di, is defined as a measure of relative weakness of alternative i over other alternatives when compared with a measure of weakness of other alternatives from the alternative i . N N D. = l D , - l D i . 1=1 i=i j#i J * I An alternative with the highest value of net concordance index C and lowest value of net discordance index D would be preferred. It is possible that the alternative with the highest value of concordance index is not the same as that with the lowest value of discordance index. In order to address this issue, the alternatives are ranked based on the concordance and discordance indices and each alternative is ranked by taking the average of these two rankings. The alternative with the highest average ranking is considered to be the best alternative. Alternatives with the same average ranking would be considered equally suited. For the case of network selection, in order to find the top candidate network, it is required to have a complete ranking for all the networks under consideration. Therefore approach 2 has been applied as it provides a clearer ranking of the alternatives. 95 4.3.3 Evaluation of Modified ELECTRE In order to evaluate the use of E L E C T R E as well as the impact of the proposed changes to the algorithm, we consider a network selection situation with five network types to choose from. The attribute values for these networks, determined at the time of network selection, are provided in Table 4-5. We consider the scenario where network selection is influenced by the requested service indicated by the user. Three services, namely VoIP (low bit rate, real-time), sfreaming (high bit rate, soft real-time), and web browsing (varying bit rate, bursty, non real-time) are considered. The service type is used to assign attribute weights. For example in the case of VoIP, since it is a low bandwidth application, the total bandwidth and available bandwidth are not considered important and therefore assigned a weight of zero. Also the transport cost is not considered significant because of the higher revenue generating nature of VoIP applications but attributes such as low latency and jitter are quite significant for a good customer experience and therefore assigned higher weights. The values of the assigned weights for different services considered are provided in Figure 4.8. For the case of the modified algorithm, different reference values for the attributes are used for each of the service type. These reference values indicate the preferred attribute values for the service type. Table 4-6 provides these values for VoIP, streaming and web browsing applications. The results are documented in Tables 4-7 and 4-8. A s described in alternative approach 2, the rankings shown in these tables are the averages of two rankings obtained using concordance and discordance indices. So the highest rankings in these tables may not actually be 1 unless there is a network which is the best both from the perspective of concordance and discordance indices. Table 4-7 shows the network rankings when a standard version of E L E C T R E as described earlier was used with approach 2. The results show that the same network, i.e., #3 is being selected for all the services although the rankings for the rest of the networks are different for the services; e.g., the second ranked network for VoIP service is different than that for streaming or web browsing service. The rankings for networks using the modified version of the algorihtm that uses reference attribute values (as shown in Table 4-6) is shown in Table 4-8. The selected network rankings in this case are different for VoIP and streaming. For web browsing, three networks are ranked at the same level. To further explain the reason for the selection of different networks by the algorthm Tables 4-9, 4-10 and 4-11 are provided. These tables show how the input attribute values are changed by the use of reference values, normalization and then use of attribute weights. It can be 96 seen from these tables that the adjusted, normalized and weighed attribute values that form the input to the algorithm are quite different in the case of VoIP, streaming and web browsing services. This reflects the effect of data manipulation performed to meet the optimization objectives of the decision maker. A s a result, different networks can get selected for different service types. Table 4-5 Attribute Values for Scenarios under Consideration CB % TB Mbps AB mbps U % D Msecs / msecs L per 106 Nkwk#l e.g. UMTS 100 2 0.2 10 400 50 100 Ntwk#2 e.g. 802.11b 20 11 1 20 200 25 20 Ntwk#3 e.g. 802.11a 10 54 2 20 100 15 15 Ntwk#4 e.g. 802.11n 5 100 5 40 150 30 20 Ntwk#5e.g. 4G 30 100 5 20 100 20 15 I Cost/Byte | Total Bandwidth ] Available Bandwidth I Utilization Vo ice Streaming Web Figure 4.8 Weights associated with attributes for different services Table 4-6 Reference attribute values for voice over IP, streaming and web browsing services CB TB AB u D / L (%) (Mbps) (Mbps) (%) (ms) (ms) (per 106) VoIP 5 100 0.02 10 100 15 15 Streaming 5 100 1 10 400 50 50 Web Browsing 5 100 0.1 10 1000 100 100 97 Table 4-7 Ranking for networks using standard ELECTRE method with Alternative 2 VoIP Streaming Web Browsing Ntwkm e.g. UMTS 5 3 5 Ntwk#2 e.g. 802.11b 3 2 2 Ntwk#3 e.g. 802.11a 1 1 1 Ntwk#4 e.g.802.11n 4 5 2 Ntwk#5 e.g. 4G 2 4 4 Table 4-8 Ranking for networks using modified ELECTRE method with Alternative 2 VoIP Streaming Web Browsing NtwMl e.g. UMTS 5 2 4 Ntwk#2 e.g. 802.11b 3 1 2 Ntwk#3 e.g. 802.11a 1 3 2 Ntwk#4 e.g.802.11n 4 4 2 Ntwk#S e.g4G 2 4.5 5 98 Table 4 - 9 voice over IP service ADJUSTED ATTRIBUTE VALUES CB(%) TB (Mbps) AB (Mbps) U(%) D (msecs) J (msecs) L (per 10f) #1 95.00 98.00 0.18 0.00 300.00 35.00 85.00 #2 15.00 89.00 0.98 10.00 100.00 10.00 5.00 #3 5.00 46.00 1.98 10.00 0.00 0.00 0.00 #4 0.00 0.00 4.98 30.00 50.00 15.00 5.00 #5 25.00 0.00 4.98 10.00 0.00 5.00 0.00 NORMALIZED ATTRIBUTE VALUES CB(%) TB (Mbps) AB (Mbps) U(%) D (msecs) J (msecs) L (per 106) #1 0.000 0.000 1.000 1.000 0.000 0.000 0.000 #2 0.842 0.092 0.833 0.667 0.667 0.714 0.941 #3 0.947 0.531 0.625 0.667 1.000 1.000 1.000 #4 1.000 1.000 0.000 0.000 0.833 0.571 0.941 #5 0.737 1.000 0.000 0.667 1.000 0.857 1.000 NORMALIZED AND WEIGHED ATTRIBUTE VALUES CB(%) TB (Mbps) AB (Mbps) U(%) D (msecs) J (msecs) L(perl06) #1 0.000 0.000 0.000 0.200 0.000 0.000 0.000 #2 0.042 0.000 0.000 0.133 0.200 0.214 0.141 #3 0.047 0.000 0.000 0.133 0.300 0.300 0.150 #4 0.050 0.000 0.000 0.000 0.250 0.171 0.141 #5 0.037 0.000 0.000 0.133 0.300 0.257 0.150 99 Table 4-10 Steaming service ADJUSTED ATTRIBUTE VALUES CB(%) TB (Mbps) AB (Mbps) U(%) D (msecs) J (msecs) L (per 1(f) #1 95.00 98.00 0.80 0.00 0.00 0.00 50.00 #2 15.00 89.00 0.00 10.00 200.00 25.00 30.00 #3 5.00 46.00 1.00 10.00 300.00 35.00 35.00 #4 0.00 0.00 4.00 30.00 250.00 20.00 30.00 #5 25.00 0.00 4.00 10.00 300.00 30.00 35.00 NORMALIZED ATTRIBUTE VALUES CB(%) TB (Mbps) AB (Mbps) U(%) D (msecs) J (msecs) L (per 10") #1 0.000 0.000 0.800 1.000 1.000 1.000 0.000 #2 0.842 0.092 1.000 0.667 0.333 0.286 1.000 #3 0.947 0.531 0.750 0.667 0.000 0.000 0.750 #4 1.000 1.000 0.000 0.000 0.167 0.429 1.000 #5 0.737 1.000 0.000 0.667 0.000 0.143 0.750 NORMALIZED AND WEIGHED ATTRIBUTE VALUES CB(%) TB (Mbps) AB (Mbps) U(%) D (msecs) J (msecs) L (per 10") #1 0.000 0.000 0.160 0.200 0.100 0.100 0.000 #2 0.168 0.014 0.200 0.133 0.033 0.029 0.050 #3 0.190 0.080 0.150 0.133 0.000 0.000 0.038 #4 0.200 0.150 0.000 0.000 0.017 0.043 0.050 #5 0.147 0.150 0.000 0.133 0.000 0.014 0.038 100 Table 4-11 Web browsing service ADJUSTED ATTRIBUTE VALUES CB(%) TB (Mbps) AB (Mbps) U(%) D (msecs) J (msecs) L(perl06) #1 95.00 98.00 0.10 0.00 600.00 50.00 0.00 #2 15.00 89.00 0.90 10.00 800.00 75.00 80.00 #3 5.00 46.00 1.90 10.00 900.00 85.00 85.00 #4 0.00 0.00 4.90 30.00 850.00 70.00 80.00 #5 25.00 0.00 4.90 10.00 900.00 80.00 85.00 NORMALIZED ATTRIBUTE VALUES CB(%) TB (Mbps) AB (Mbps) U(%) D (msecs) J (msecs) L (per 10b) #1 0.000 0.000 1.000 1.000 1.000 1.000 1.000 #2 0.842 0.092 0.833 0.667 0.333 0.286 0.059 #3 0.947 0.531 0.625 0.667 0.000 0.000 0.000 #4 1.000 1.000 0.000 0.000 0.167 0.429 0.059 #5 0.737 1.000 0.000 0.667 0.000 0.143 0.000 NORMALIZED AND WEIGHED ATTRIBUTE VALUES CB(%) TB (Mbps) AB (Mbps) U(%) D (msecs) J (msecs) L (per 10") #1 0.000 0.000 0.150 0.100 0.050 0.050 0.100 #2 0.421 0.005 0.125 0.067 0.017 0.014 0.006 #3 0.474 0.027 0.094 0.067 0.000 0.000 0.000 #4 0.500 0.050 0.000 0.000 0.008 0.021 0.006 #5 0.368 0.050 0.000 0.067 0.000 0.007 0.000 101 4.4 Conclusion Convergence of handheld multimedia devices with communication devices has created a new class of feature rich consumer devices which can provide seamless network connectivity for service delivery through a variety of communication technologies. In the face of a multitude of service delivery network options and the presence of heterogeneous wireless networks, the selection of an optimal network for service delivery is a major issue. M A D M algorithms are a popular decision making tool in operations research and have also been considered for application in other areas. They are however known to suffer from the problem of ranking abnormalities. These abnormalities can potentially decrease the quality of the results. We have applied TOPSIS, an M A D M algorithm, to the problem of network selection and analyzed the causes o f the ranking abnormalities problem for TOPSIS as applied to the problem of network selection. We have proposed an iterative approach for application of TOPSIS, which can improve the results for network selection by comparing only the more likely candidates in the process. Results for the iterative method have been presented to show that the top candidate network determined by TOPSIS can change when a candidate at the bottom of the ranking is removed from the comparison, and the iterative approach gives a more consistent and accurate final result. The network selection algorithm described in this chapter is well suited for use with service delivery architecture described in Chapter 2 [7]. Prior research on application of M A D M algorithms to the problem of network selection has not compared them to select the most appropriate algorithm. We have discussed the decision maker's optimization objectives and hence the utility of attributes as a means to evaluate the algorithms and select the most appropriate one. The need to support non-monotonic utility for attributes in order to handle diverse optimization objectives of a decision maker has been shown and M A D M algorithms have been evaluated for handling these objectives. We have shown that many of the commonly used M A D M algorithms such as S A W , W P M , TOPSIS, and E L E C T R E in their standard form are not well suited to the network selection problem because of assumptions about monotonically increasing or decreasing utilities of the attributes. We have also shown that G R A can easily be adapted to use multiple reference networks and is therefore better suited for achieving this type of optimization objectives. The evaluation of adapted G R A in this chapter has also demonstrated that the selection of delivery network w i l l be impacted by how the algorithm is used to achieve the optimization objectives. A novel two step process has been proposed that uses multiple reference values. It has been explained through an example that shows how reference 102 attribute values and attribute weights impact the selection process. The adjustment of these parameters for different service types has been discussed. The decision process proposed here can be used in a heterogeneous wireless network system environment. We have described the adaptation of E L E C T R E , an M A D M algorithm, for ranking network alternatives during the network selection process. The use of a particular M A D M algorithm for a specific problem is based on an assessment about the appropriateness of the algorithm for application to the problem space. Here we describe the use of a modified algorithm that provides a complete ranking of alternative networks. The modifications also allow the usage of E L E C T R E with attributes exhibiting a non-monotonic utility. These modifications make the algorithm more adept to application in network selection as it expands its applicability to a wider range of optimization objectives. Such network selection scenarios are of special importance in a heterogeneous wireless networking environment being used to deliver a variety of service types. With these modifications and a simple decision making philosophy, E L E C T R E is an ideal algorithm for use in network selection. In order to evaluate the proposed use of the algorithm, an example has been described where depending upon the QoS requirements of the service being requested by the user device, a different delivery network maybe chosen. The results have been compared with and without implementing the modifications to support non-monotonic utility of attributes. The proposed algorithm can be used with the network selection architecture described in Chapter 2 [7]. 103 4.5 References [I] Interworking with External Networks Task Group, I E E E 802.1 l u , [online] Available http://www.ieee802.Org/l 1/. [2] Media Independent Handover, I E E E 802.21, [online] Available http://www.ieee802.org/21/. [3] 3GPP Technical Specification Group Services and Systems Aspects; 3GPP System Architecture Evolution: Architecture Enhancements for non-3GPP accesses, 3GPP Technical Report 23.402 v 0.1, Jan 2007, [online] Available http://www.3gpp.org. [4] F. Bar i and V . Leung, "Automated Network Selection in a Heterogeneous Wireless Network Environment", IEEE Network, pp. 34-40, January/February 2007. [5] C L . Hwang and K . Yoon , Multiple Attribute Decision Making: Mehtods and Applications, Springer Verlag, 1981. [6] B . Benayoun, B . Roy, and B . Sussmann, , "Manu al de reference du programme electre, Note de Sythese et Formation," Direction Scietifique SEMA, N. 25, 1966. [7] F . Bar i and V . C . M . Leung, "Service delivery over heterogeneous wireless networks: network selecton aspects", Proc. ACMIWCMC, pp. 251-256, July 2006. [8] C L . Hwang and K . Yoon , "Multiple Attribute Decision Making: A n Introduction", Sage Publications, 1995. [9] E . Triantaphyllou, Multi-Criteria Decision Making Methods: A Comparative Study, Kluwer Academic Publishers, 2002. [10] E . Triantaphyllou and B . Shu, "On the maximum number of feasible ranking sequences In multi- criteria decision making problems", European Journal of Operational Research, pp. 217-230, vol . 130, no. 3, March 2001. [II] E . Triantaphyllou and A . Sanchez, " A sensitivity analysis approach for some deterministic multi-criteria decision- making methods", Decision Sciences, pp. 151-194, vol . 28( 1), 1997. 104 [12] Q. Song and A. Jamalipour, "Quality of Service Provisioning in Wireless LAN/UMTS Integrated Systems using Analytic Hierarchy Process and Grey Relational Analysis", Proc. IEEE GlobeCom, pp. 220-224, November/December 2004. [13] Q. Song and A. Jamalipour, "Network selection in an integrated wireless L A N and UMTS environment using mathematical modeling and computing techniques", IEEE Wireless Communications , pp. 42-48, Volume 12, Issue 3, June 2005. [14] Q. Song and A. Jamalipour, "A network selection mechanism for next generation networks" Proc. IEEE ICC, pp. 1418-1422, May 2005. [15] E. Stevens-Navarro and V.W.S. Wong, "Comparison between Vertical Handoff Decision Algorithms for Heterogeneous Wireless Networks," Proc. IEEE VTC-Spring, pp. 947-951, May 2006. [16] J.L. Deng., "Introduction to Grey system", Journal of Grey System, vol. 1, issue 1, pp. 1-24, 1989. [17] M . Rogers, M . Bruen and L.-Y. Maystre, ELECTRE and Decision Support - Methods and Applications in Engineering and Infrastructure Investment, Kluwer Academic Publishers, 2000. [18] S. Hongvan, H. Chen and J. Lingge, "Intelligent signal processing of mobility management for heterogeneous networks", Proc. International Conference on Neural Networks and Signal Processing, pp. 187-191, December 2003. [19] H.J. Wang, R.H. Katz and J. Giese., "Policy-Enabled Handoffs Across Heterogeneous Wireless Networks", Proc. IEEE Mobile Computing Systems and Applications, pp. 51-60, February 1999. [20] M . Zeleny, Multiple Criteria Decision Making, McGraw-Hill, 1982. [21] J. Arkko, B. Abboba, J. Korhonen and F. Bari, "Network Discovery and Selection Problem", draft-ietf-eap-netsel-problem-08, June 2007 (completed working group last call), [online] Available http://www.ietf.org. 105 5 Application of Data Prediction and Fuzzy Techniques with MADM-Based Automated Network Selection 1 5.1 Introduction Broadband wireless networks such as W L A N and W W A N have one characteristic in common in that they all use IP as the data transport. In order to provide ubiquitous coverage, increasingly these all-IP wireless technologies are being made to interwork. This makes the achievement of a consistent service experience over heterogeneous wireless technologies very important. Unlike a circuit-switched service environment, packet-switched IP networks are known to vary in terms of QoS. The variation from a network to network can be static such as because of the inherently different capabilities of the network (e.g., I E E E 802.11a vs. I E E E 802.1 In) or it can be dynamic based on the network's current congestion level. Selection of an optimal service delivery A N is an important problem to be solved in such an environment. Since a number of network attributes, e.g., those related to QoS have an impact on such decisions, the use of M A D M algorithms [1][2] has been proposed in the past for ranking candidate A N s in terms of their suitability. Output of these algorithms is dependent upon the accuracy of information being used as input to them. In real world scenarios, often some of the information is either not available or it is imprecise. In addition in some cases, because of the candidate network types the usage of fuzzy input information is more useful than crisp values. Figure 5.1 illustrates that even when all the ' Parts of this chapter have been published in the following F. Bari and V. Leung, "Network Selection with Imprecise Information in Heterogeneous All-IP Wireless System", accepted in Proc. Wireless Internet Conference (WICON), October 2007 106 information is not available it can still be possible to narrow down the list o f candidate networks. Below we describe scenarios where application of M A D M algorithms for network selection while relying on incomplete information can be beneficial. Figure 5.1 Use of MADM with imprecise information can help narrow down the candidate 5.2 Us ing Non-Compensat ing M A D M Algor i thms with Incomplete information In the network selection process described in Chapter 3, initially, non-compensating M A D M algorithms are used to narrow down the candidate list. The information used in this process comes from both the terminal and the entities within the network. However in some cases, the terminal or entities within the network may be either unable to or unwilling to provide some of the information to the decision making entity. For example, the information related to the terminal's mobility profile may not be available for the decision making entity because the user / terminal has not indicated it. In other cases the network may not have all the information about the candidate networks that is to be used in the non-compensatory part of the M A D M algorithm. For example, it may not know the exact coverage area or the authentication methods supported by the roaming partner's A N s . Even in such scenarios by using the mechanisms described in Chapter 3 candidate networks can be narrowed down to fewer and more probable service delivery networks. The candidate networks in the resultant shortened list in such scenarios may not all be accessible to the terminal but the list can provide guidance to the terminal to help it narrow down the service delivery options. User's Home Network and its Roaming Partner Networks service delivery network options 107 In general, the attributes used in the non-compensating part of the algorithm can be separated into attributes essential for the decision process and those that can be considered less critical. The decision maker can decide which attributes it considers absolutely essential to get useable information from the decision process. For example without the information about the location of the terminal it is not possible to narrow down the search whereas service to be used can be considered less critical. Also , in such scenarios there w i l l be a possibility that some of the top ranking networks in the short list are not accessible to the terminal; e.g., i f the authentication mechanism related information was lacking for the network or terminal type. Therefore, to make the information useful, a list o f preferred networks instead of a single top ranking network should be provided to the terminal. 5.3 Us ing Compensat ing M A D M Algor i thms with Data Predict ion and Fuzzy Input The application of a standard compensating M A D M algorithm to network selection involves • Identifying all alternatives and compensatory M A D M attributes impacting the decision process, • Assigning relative importance in the decision making process to each of the attributes, and • Using a M A D M algorithm to get a ranking of the alternatives. These algorithms are used to determine the ranking of alternatives in terms of their desirability with respect to multiple criteria as a whole that can influence the decision. For the compensating M A D M algorithm application, the following two scenarios have been identified where fuzzy approaches w i l l be useful. 5.3.1 Scenar io 1 - Imprecise or Miss ing Information For compensating M A D M algorithms to work properly, the attribute values have to be reliable in order to select the optimal service delivery network. However, as shown in Figure 5.2, due to the geographic distribution of the locations of the data collection points [3] it may not be possible to get real time updates for some of the attributes used in network selection. Also while dealing with heterogeneous access technologies spanning autonomous operator domains it may not be possible to get a homogeneous set of attribute data spread evenly over time that would allow a direct comparison between the networks towards their suitability for delivering the requested services. 108 In many cases there are measurement errors associated with the monitoring and processing of dynamic QoS attribute values such as packet delay, jitter, and loss. Hence there is a need to develop a mechanism of network selection when input attribute values are less reliable or unavailable. Figure 5.2 Distributed nature of data collection points across different access types and autonomous domains makes harmonized, accurate and real time parameter information challenging Lotfi Zadeh [4][5] developed in the 1960s the theory of fuzzy logic that enables algebraic manipulation of fuzzy or imprecise data. Since then fuzzy logic theory has found applications in a variety of areas including decision support. Fuzzy numbers have been used in the past with M A D M algorithms for scenarios where input attribute values are imprecise or hard to calculate [6] [7]. The use of a very simple fuzzy logic based multiple-criteria decision-making system has been proposed [8] to perform vertical handovers in a heterogeneous networking environment. The application of this work to network selection is therefore limited. In [9] fuzzy based network selection has been discussed within the context of peer-to-peer networking. In [10] and [11] fuzzy mechanisms have been considered within the context of vertical handoffs. [10] actually converts fuzzy data to crisp values before applying standard M A D M algorithms. [11] uses a fuzzy inference engine along with neural nets for predictions about the number of users. The work in [9]-[l 1] however has a somewhat different focus and does not take into consideration important 109 practical aspects for the problem under consideration in our research, such as prediction of unavailable data, selection of fuzzy number type to represent predicted attributes, suitable defuzzification techniques or even a fundamental consideration about when it is appropriate to use a fuzzy M A D M algorithm. In this chapter we describe both types of scenarios where fiizzy techniques w i l l be helpful and where other mechanisms are better suited. We propose a novel way of combining fuzzy techniques along with parameter estimation in network selection. They have been applied to the decision process while using a proposed fiizzy implementation of G R A , a M A D M algorithm. We describe mechanisms for prediction/estimation of missing data, fuzzification of the estimated values and the subsequent defuzzification of network rankings obtained by the use of fuzzy G R A . Additional decision support tools that can be applied under such conditions are presented to obtain the Confidence Levels either in the network rankings or in the data before network rankings are obtained. Both approaches are described in this chapter. Together, the techniques described here improve the reliability of the results and allow decisions to be made under uncertain conditions. The mechanisms described in this chapter would work well with the network-assisted terminal-based network selection architecture described in Chapter 2 [3]. Estimate missing data Yes or do not know No Apply standard MADM algorithm with imprecise information Yes Application of Fuzzy MADM Algorithm to Fuzzy Input e.g. use of Fuzzy GRA Apply standard MADM algorithm with forecasted information Defuzzification of Fuzzy MADM algorithm Output to get network ranking Estimation of Confidence Level in ranking while selecting the network Figure 5.3 Steps involved in proposed network selection mechanism with imprecise information no Figure 5.3 shows a comprehensive approach towards a network selection mechanism that leverages parameter estimation techniques, fuzzy theory, M A D M algorithms and also introduces the new concept of Confidence Level. The first step is to check i f the service being requested or the user's subscription profile is sensitive to the attribute data that is missing. For example, a bronze subscription user or a web browsing service may not be sensitive to small variations in QoS related attribute values. This is primarily because of the low weights assigned to the QoS related attributes in the M A D M decision process for such users/services. I f the sensitivity is low, a standard M A D M algorithm can be applied even with imprecise attribute information. The association of services and subscription types to different levels of sensitivities has been considered in Chapter 3. However i f the sensitivity is high or it cannot be determined because of missing service / subscription related information then data estimation should be done. Depending upon the estimation methods used and the past experiences with the estimation process, the forecasted data can be expressed as a scalar value or as a fuzzy set. For fuzzy data values, the fuzzy implementation of a M A D M algorithm is then applied to get a fuzzy ranking of the networks. This is followed by the step where the fuzzy rankings are defuzzified. For the case of non-fuzzy estimated values, a standard M A D M algorithm is applied. In the final step, any additional information to judge the Confidence Level in the network ranking obtained is taken into consideration. In another approach described later in the chapter, which has some advantages to the approach in Figure 5.3, the step involving Confidence Level has been moved to after forecasting of the missing data. 5.3.1.1 Data Estimation A uniform set of attributes has to be available as inputs to the decision algorithm that would compare different access networks on their basis. Data prediction can be used to see what w i l l happen to attribute values in the future based on the past history of the data. For example, based on the time of day and usage pattern, many of the QoS attributes such as the utilization of a hotspot can be predicted to a certain degree. Typically the observed values are represented in the form of time series which are collections of observations ordered in time. Deterministic data prediction can be performed using the following methods. Seasonal trends and averaged values Smoothing techniques can be used to identify the underlying trends in the observed values of attributes when some of the data to be used in the decision making process are missing. The i l l smoothing of time series removes noise related irregularities and enhances the informational part of the observed data. Several variations of smoothing techniques exist; e.g., Simple Moving Averages ( S M A ) is useful for the type of data that exhibits static values for the mean and variance. S M A of order n at time t+1 can be described by the following formula, n Weighted Moving Averages ( W M A ) involves assigning different weights to historical data before taking the moving average. It can be represented as follows, WMi , = ( W ' * V ' + * V ^ + - - + w<-^*V.-n+x) + n where i=t-n+\ i=t Other types of moving averages include various forms of Exponential Smoothing (ES) techniques described in [ 1 2 ] . The use of a particular smoothing technique would depend upon its ability to accurately predict the attribute values. The accuracy of prediction can be measured by running the algorithm on prior collected data and comparing the actual and the predicted values. Regression Regression [ 1 2 ] analyzes the relationship among variables to estimate the value of one variable from known values of other variables related to it. In the case of network selection, an attribute with known value can be used in calculating the value of another input attribute because of a strong correlation between the two attributes. In regression analysis trends of variables under consideration are analyzed (such as linear) and the variables that are seen to have dependencies on each other are correlated and then modeled using polynomials. The variables that can be easily measured with least error are also identified. A regression using only one predictor is called a simple regression. For example i f network utilization is easy to measure and report but packet jitter, packet loss, or latency are not readily available, then these QoS related parameters can possibly be estimated by the network utilization because of a correlation between them under normal network conditions. 112 The actual process of finding the relationship can involve sample data collection, drawing scatter plots to understand the relationship amongst the variables, and using computer packages or modeling techniques (such as linear prediction using least square method) to figure out the relationship. Typically the established relationship only holds for a limited range of variable values and only provides an estimate or average value. A s an example, an access network under consideration maybe able to provide its network utilization more readily, reliably and on a continuous basis than other QoS related parameters such as packet loss, delay and jitter, which can require much more active monitoring of the network. Under normal operating conditions the network utilization is correlated with QoS aspects in a packet-switched network, using regression analysis as described in this section. Therefore the utilization (L) that is much easier to monitor and to report on, can be related over a range of values to the packet Delay (D), Jitter (J) and Loss (L). Similar relationships between network utilization and QoS parameters may be provided by the operator of the network to its partners as part of the roaming agreements or S L A s . Fuzzy Estimated values Because of the forecasting techniques used and other factors such as the prior experience with the use of forecasted data, non crisp values may be obtained which can then be represented as Fuzzy Numbers. A Fuzzy Number [4][5] forms a fuzzy set that can have different membership grades as shown in Figure 5.4. A very common type of fuzzy set membership results in a Triangular Fuzzy Number (TFN). With the type of problems that can arise in data acquisition such as measuring inaccuracies or lack of updated real time attribute data as described earlier, a triangular shaped fuzzy number w i l l be a good fit. For example when regression analysis is used for data prediction, the value of the predicted parameter can be described by a triangular fuzzy number. Figure 5.4 also provides the mathematical representations of the membership function for a T F N and some simple operations like addition and multiplication between two TFNs . 113 Simple Fuzzy Number Sets u X a ,xe[a,b] b-a b-a b-c b-c 0, otherwise x c ,xe[b,c] a b c Membership function for triangular fuzzy number Addition and multiplication operations for TFNs P and Q Triangular Fuzzy Number (TFN) Figure 5 . 4 Fuzzy Numbers Use o f tools such as scatter plots shows that in general the correlation defined by regression between the variables is never exact. For example, in the case of the regression scenario described earlier, at lower utilization there w i l l not be any congestion conditions in the A N and therefore it can provide a high level o f QoS to all the services or all IP flows. In this case the regression equations would provide an accurate relationship. A t a higher utilization the A N could distinguish between preferred and non preferred services or IP flows in the packet scheduling process. A s a result, the aggregate observed data for all services/IP Flows would show a spread/scatter at higher utilizations. This behavior can be represented by mapping the attribute values of D , J and L to a range of U values. For this example we consider that they map to 0.8-1.1U with 1.0U representing the default or modal value. In fiizzy numbers terms, this can then be represented by triangular fiizzy numbers shown graphically in Figures 5.4. It can be seen that uncertainty in the values for D , J and L increase with an increase in the network utilization (U). Fuzzy number sets can also be obtained by direct mapping from a lookup table that maps a known value of U to D , J, L values observed earlier for the same value of U in that network (as described by seasonal tends in the previous section). 114 5.3.1.2 Compensating MADM Algorithm M A D M algorithms include a variety of deterministic mechanisms [1][2] that have been used in decision making for a wide variety of engineering problems. Fuzzy logic can be applied in conjunction with a M A D M algorithm. Triantaphyllou in [6] and [7] has described the use of fiizzy information with some of the commonly used M A D M algorithms. Table 5-1 Attributes and their weight assignments for streaming media and web browsing services while using GRA Attribute Streaming Media Web Browsing Cost per Byte (CB) i.e. Data transport cost on a particular access system 0.2 0.5 Total Bandwidth (TB) i.e. Overall bandwidth of the wireless access link 0.15 0.05 Allowed Bandwidth (AB) i.e. Bandwidth per user allowed by the access system 0.2 0.15 Utilization (U) i.e. Current utilization of the wireless link 0.2 0.1 Packet delay (D) i.e. Average packet delay within the access system 0.1 0.05 Packet Jitter (J) i.e. Average packet delay variations within the access system 0.1 0.05 Packet Loss (L) i.e. Average packet loss rate within the access system 0.05 0.1 115 10 2 0 3 0 4 0 5 0 N e t w o r k U t i l i z a t i o n (%) 6 0 2 0 3 0 4 0 N e t w o r k U t i l i z a t i o n (%) C D 8 0 0 -S 6 0 0 . . . . J - u • i 1 ' 0 . 8 U 1.1 u r- ' ' 0 1 0 2 0 3 0 4 0 5 0 6 0 7 0 N e t w o r k U t i l i z a t i o n (%) Figure 5.5 Example graphs of packet Delay (D), Jitter (J) and Latency (L) values using Network Utilization (U). It assumes correlation represented by D = 0.225 *u2 +10, J = 0.05 *U2, L = 0.0019*C/3 with 0.8U - 1.1U in the network utilization range of 5% and 70% Depending upon the type of ambiguity in the decision process, fuzzy logic can be applied in more than one ways with M A D M . For example, the uncertainty in the decision process can be in the 116 attribute values or it can be in the importance associated with them. Here, it is assumed that the importance or weight assigned to an attribute is based on the service requested or the user's QoS subscription and therefore known based on the information provided by the user. However, because of the reasons described earlier in the chapter, some of the input attribute values are either unavailable or are imprecise. Here GRA [13][14] is used as a M A D M algorithm for fuzzy implementation described in this chapter. GRA [14] is based on the concept of GRS, where GRS (X,Y) describes a relationship Y between reference data values Xn and a sequence of data values X. So ify e Y , x ^  X ,x 0 e X 0 such that x0 = x0(1) x0(n) and = x^l) x,(n) then y(x0(k),xi(k)) would represent a GRS at point k provided the axioms documented in [14] are satisfied. GRA calculates a GRC and uses it as a measure of the closeness of the result to the reference values. The GRC values can thus be used for ranking the candidate networks. Table 5-2 Attribute values for the five networks under consideration CB (%) TB (mbps) AB (mbps) U(%) D (msecs) J (msecs) L (per 106) Ntwk#l 100 2 0.2 10 400 50 100 Ntwk#2 20 11 1 20 200 25 20 Ntwk#3 10 54 2 20 100 15 15 Ntwk#4 5 100 5 40 150 30 20 Ntwk#5 30 100 5 20 100* 20* 15* 25 122' 25' 24' 25 151' 31' 30' 25 165' 34' 33' *- last reliable attribute values '- Forecasted attribute value represented as a Triangular Fuzzy Number In applying the fiizzy implementation of GRA to the scenarios under consideration the attribute values to be used as inputs to the algorithm will be a mixture of crisp and fiizzy values. This is because not all of the attribute values are unknown or have imprecise values. For networks with fuzzy input attributes represented by TFNs, it would result in three GRCs to be evaluated using the fiizzy arithmetic for TFN as shown in Figure 5.4. This assumes that because of correlation 117 amongst missing input attribute values, multiple fuzzy inputs described by TFNs w i l l only result in three possible input combinations. A defuzzfication process is then used to get a crisp value of the G R C . In order to better understand the network selection decision process described in Figure 5.3 we consider a simple scenario where a network selection decision has to be made with five network options. These networks are represented by a set of QoS related attributes as listed in Table 5-1. The selection of these attributes for use in network selection while using M A D M is described in Chapters 3 and 4 [17] [18] [19]. Each attribute has a relative importance represented by a weight assigned to it during the decision process. The relative importance of an attribute during the decision process can be based on the service being requested or the QoS profile of the user. The weights shown in Table 5-1 represent the importance assigned to the attributes for a streaming service and web browsing. We have selected streaming service as a representative of a service scenario that is sensitive to dynamic QoS attributes and web browsing as a representative service scenario that is not sensitive to dynamic QoS attributes. 0 95 OS 0 85 0.8 0 75 07 0 55 1 I — 9 — Fuzzy G R A T F N (a) - ^ 7 — Fuzzy G R A T F N (b) " H — Fuzzy G R A T F N (c) — **— Crisp GRA with imprecise info ' ^^^^ .j/^........X I i Ntwk#1 Ntwkffi Ntwkffl Ntwk#4 Ntwk*6 Figure 5.6 Results for streaming media services using GRA algorithm with fuzzy input and with imprecise input information 118 Table 5-2 lists the values for these attributes for Ntwk#l through Ntwk#4 at the instance of decision making. QoS related data (i.e., Delay, Jitter and Packet Loss) for Ntwk#5, one of the networks under consideration, is not available at the time of decision making. However the last known reliable values for these attributes for Ntwk #5 is available and are listed in Table 5-2. We assume that based on prior observation of the network, it has been possible to correlate network utilization with QoS parameters. The correlation used in this example is defined in Figure 5.5. It results in TFNs for Delay, Jitter and Latency values for Ntwk#5 as shown in Table 5-2. Using the fuzzy version of GRA described in the previous section for the streaming media service, the results as shown in Figure 5.6 are obtained. It can be seen that for Ntwk#5, three GRC values are obtained which constitute a GRC represented by a TFN. With Ntwk#5 represented by a fuzzy GRC, the exact ranking for the networks is not obvious and therefore a defuzzification step is required. It should be noted that while it is observed in Figure 5.6 that changing the attribute values for one network, in this case Ntwk#5, does not impact the GRC values of the other networks, this is not always the case. In most of the cases, the way the GRC is calculated (as shown by its mathematical representation earlier), some of the other network alternatives may also result in different GRC values, when the attribute values for one network is changed. In that case the defuzzfication part can become more complicated as we will have to apply defuzzification techniques to other network alternatives as well in order to obtain crisp rankings. This would become clearer in the example presented in the next section. 5.3.1.3 Defuzzification Defuzzification is the process of converting a fuzzy number or set to a crisp value. Several mechanisms exist for comparing fuzzy numbers with some of them being fairly complex and computationally intensive. A good mechanism can be selected by considering its level of complexity, its accuracy and its application to fuzzy numbers of a particular shape. [15] [16] documents a range of defuzzification techniques. Center of gravity (CoG) is a very popular technique that is computationally among the simplest and particularly well suited for defuzzification of TFNs. It calculates the centroid of the fuzzy value and can be approximated in the discrete domain as follows, t f j * F j CoG*^ r 119 where //. and F} represent the fuzzy set membership index and the corresponding fuzzy number, respectively, in the discrete domain. A simple Matlab implementation of the algorithm for calculating the CoG of a TFN for GRC is shown below: for F=a:increment:b Ul = F/(b-a)-a/(b-a); SumofUl=Ul +SumofUl; SumofUFl =F*U1 +SumofUFl; end for F=b:increment:c U2 = F/(b-c)-c/(b-c); SumofU2=U2+SumofU2; SumofUF2=F*U2+SumofUF2; end CoG=(SumofUFl +SumofUF2)/(SumofUl +SumofU2) Figure 5.7 graphically shows application of CoG for defuzzification of the GRC value for Ntwk#5. Comparing the CoG value for Ntwk#5 from Figure 5.7 with the results for the remaining networks as shown in Figure 5.6 it can be seen that the CoG based GRC value for Ntwk#5 is lower than the GRC value for Ntwk#4. Hence Ntwk#4 should be selected. In the absence of reliable data for some of the networks, the alternatives to not using data forecasting and fuzzy techniques as described above would be to: 1. Apply a standard M A D M technique for all the networks without fuzzification and use instead non-fuzzy unreliable data or their approximations as inputs. 2. Remove the networks with unreliable or missing data altogether from the comparison and then perform M A D M on the remaining networks to obtain their rankings. 120 If the first approach is used, depending upon the sensitivity of the selected service and/or the user subcription type to the missing information, there is a possibility of getting an incorrect ranking of the networks because of unreliable inputs to the algorithm. For streaming media service this is shown in Figure 5.6 where the last reliable values of missing paramters were used. However the values were outdated and since then network utilization had gone up. So because of a strong correlation between network utilization and the missing QoS attributes, the values of missing attributes should have been predicted since streaming media service is sensitive to QoS attribute values. The use of outdated values in this case leads to the incorrect selection of Ntwk#5 whereas Ntwk#4 would have been a better choice as shown by the use of the fuzzy GRA implementation also in Figure 5.6. O.BBS Triangular Fuzzy Set Figure 5.7 Use of CoG for defuzzification of a TFN representing the GRC in a fuzzy implementation of GRA for streaming media services If the second approach is used, it is possible that a top ranking candidate network is not selected as it is dropped before the comparison process. For the same example described above, the results of applying GRA after removing Ntwk#5 from the candidate list are shown in Figure 5.8. While in this case Ntwk#4 is correctly selected, if in the given example, the network utilization had gone down instead then it is entirely possible that Ntwk#5 would have been a better choice but it would have been eliminated as a candidate by this strategy. 121 0 85 0.65 Ntwkffl Figure 5.8 Results of GRA algorithm for streaming media service after removing Ntwk#5 with missing information Ntwk* Figure 5.9 Results for web browsing service using GRA algorithm with fuzzy input information and imprecise input information 122 N o w we consider the case of web browsing service. In the case of this service, based on the weights assigned to the attributes the sensitivity of the service is known to be low to dynamic QoS attributes. In order to confirm our assertion that for this type of service the use of imprecise information is acceptable, we apply both fuzzy (with parameter estimation) and crisp versions of the G R A algorithm. The results of the fuzzy G R A algorithm with forecasted data documented in Table 5-2 are shown in Figure 5.9. The results of defuzzification are shown in Figure 5.10. The results o f application of crisp G R A while using imprecise information for the missing attributes are shown also in Figure 5.9. Comparison of the two set of results shows that Ntwk#4 gets selected in both cases by a wide margin and the use of imprecise information does not have much impact on the results in this case. 0.825 0.83 0.835 Triangular Fuzzy Set 0.845 Figure 5.10 Use of CoG for defuzzification of a TFN representing the GRC in a fuzzy implementation of GRA for web browsing services 123 A n appropriate use of parameter estimation and the fuzzy techniques as described in Figure 5.3 hence provides a balanced approach by not eliminating the networks with missing, old or unreliable data while keeping in view the fact that the information about them is fuzzy and/or not entirely accurate. In the following section the proposed mechanism has been further enhanced by developing criteria forjudging the reliability of the results. 5.3.1.4 Confidence Level Although the methods described above can provide improved ranking for the networks under consideration, other factors should also be taken into consideration before making use of the rankings obtained in this manner. Here we introduce the new concept of "Confidence LeveF that can allow the decision maker to evaluate i f the network rankings are reliable enough to be used for network selection. The factors used in calculating the Confidence Level (CL) include: • Sensitivity of the service or user subscription type to unreliable attribute values (CL(s)): A s described earlier, in the application of a M A D M process, attribute weights are assigned based on the service or user type; e.g., a VoIP service or a Gold user may be assigned higher weights for dynamic QoS related attributes because of the nature of the requested service or the user subscription. Based on the weights assigned to the attribute values, network ranking can become more vulnerable to errors in parameter estimations. In other words, a decision where dynamic QoS attributes have higher assigned weights w i l l have a higher sensitivity to errors in the parameter values. Figure 5.11 shows the typical relationships between C L and some of the common service types while Figure 5.12 shows the relationships between C L and the reliability of data values for the same services. For example a VoIP service when compared to web browsing is more vulnerable to incorrect ranking because of imprecise values of the QoS related attributes. This is because of the relatively higher weights assigned to the dynamic QoS attributes such as packet delay, jitter, and loss in decision making for QoS sensitive service types. 124 VoIP Streaming Web Browsing Figure 5.11 Trend graph for relationships between Confidence Level and common service types when a QoS related parameter with dynamic characteristics has unreliable value and therefore is being predicted Reliability of Estimated Data-> Figure 5.12 Trend graph for relationship between Confidence Level and reliability of estimated data for some of the common service types • Time since the last data update (CL(f)): There w i l l be a level of uncertainty in the data values based on how long data have been unavailable and how much thry can change over time based on, e.g., seasonal charts. Figure 5.13 shows a typical relationship between C L and the 125 time since the last attribute value was received indicating the older the last good value, the lesser the confidence in its estimated value. The degree of correlation with attribute whose value is known and is used in the regression analysis (CL(c)): This indicates the level of confidence in the estimated value. A typical relationship between the C L and the degree of correlation obtained from regression is shown in Figure 5.14. A strong correlation would indicate a high confidence level. Time since last data-> Figure 5.13 Trend graph for relationship between Confidence Level and time since last attribute value was received Degree of correlation between known and unknown paramter-> Figure 5.14 Trend graph for relationship between Confidence Level and degree of correlation between known and unknown attributes 126 We define a minimum threshold (CLthreshoid) of C L for an acceptable level of the ranking results. In other words, for the ranking to be considered trustworthy, the following should be true. The value of CLthreshoid can be provisioned into the process by the decision maker based on the decision policies. In the case of C L being lower than the threshold, the decision maker w i l l remove the network with imprecise information from the list of alternatives being considered. However, it is known [7] that addition or removal of an alternative during application of a M A D M algorithm can impact the ranking for the rest of the alternatives non-uniformly. It may therefore be necessary to run the M A D M algorithm again in this situation. A n alternative approach is to calculate C L before using the M A D M algorithm. If C L is found to be below the threshold for all the alternatives with imprecise information, then they are removed from the comparison and a standard M A D M algorithm is applied to the rest. Otherwise only those alternatives with C L lower than the threshold are removed and the fuzzy version of the M A D M algorithm is applied to the rest. This approach is shown in Figure 5.15. CL > CL 'threshold Apply data forecasting techniques to estimate missing data Yes or Do not know No Apply standard MADM algorithm with imprecise information Apply standard MADM algorithm with forecasted information No Application of Fuzzy MADM Algorithm to Fuzzy Input e.g. use of Fuzzy GRA Apply standard MADM Algorithm after excluding alternatives with Imprecise information Defuzzification of Fuzzy MADM algorithm Output to get network ranking Figure 5.15 An alternative approach of using Confidence Level in the decision process 127 Using the alternative approach and the factors described above, the decision maker can formulate an overall CL to be used in the decision process. A very simple example of such a term can be C L = C L s + C L t + C L C where 0 < C L < 1. The decision maker can also provide, based on his understanding of how various factors should influence CL, some distribution of weights across CLS(s), CL(t) and CL(c). Here we use an equal distribution of weights between the three so that 0 < C L S < 0.33, 0 < CL , < 0.33 and 0 < C L C < 0.33. Using the trend graphs shown in Figure 5.13 and Figure 5.14, assuming a linear relationship between the variables and utilizing the normalized values, we come up with the following two equations, CL, =-0.33t + 0.33 where t is the normalized time, i.e., 0 < t < 1 with t = 1 representing a finite elapsed time since the last reliable value after which data values can no longer be considered useful. C L C = 0.33c where "c" is a normalized correlation factor, i.e., 0 < c < 1 determined by the decision maker based on past data collection. Similarly we can relate CL(s) to the normalized value of sensitivity factor determined by the decision maker as follows C L d = 0.33s where 0 < s < 1. Using these values for CL(s), CL(t) and CL(c) we come up with a simple equation for CL as follows C l = ( s + c-t + 1) 3 where s, c, and t are all normalized values. 128 Figure 5.16 Graphical representation of how network congestion during service delivery time can be related to parameters such as the total bandwidth, bandwidth allowed per user, current network utilization and rate of change of network utilization In addition, deduction from the information available about other attributes can help the final network selection process. One such example is shown in Figure 5.16 where the possibility of network congestion during service delivery can be predicted based on service related information, the available bandwidth and total bandwidth for the network under consideration, the current network utilization and its past rate of change. Such information can then be used in the final selection process. Another application of the type of prediction shown in Figure 5.16 can be when the duration of the service to be used is known. In that scenario, it can be possible to use the predicted parameter values into the future and apply network selection in future at more than one time instances to see i f the same network would be selected for the duration of the service. This type of application would be especially useful i f the networks under consideration do not have good admission control and QoS guaranteeing mechanisms in place and therefore service experience for already admitted sessions can suffer from increased network load. Since forecasting of data and C L s of the estimated values can be done independent of the use of fuzzy M A D M algorithm, it is also possible for the decision maker to update the C L s of the estimated values on a regular basis and make them available to the decision process when needed. A lookup table for each of the factors impacting the C L of each attribute can be created to map the 129 normalized values of these factors to a C L for easy calculation. A variety of decision types, from a rough estimate to very accurate, can be supported depending upon the resolution or frequency of updates of the data being forecasted. In order to better understand the notion of Confidence Level we consider the use of lookup tables shown in Table 5-3 with the example described previously in the chapter. The values used in the table are for illustration and w i l l be decided in practice by the decision maker based on the decision policies and the past experience with data prediction. For example, the table entries for CL(s) indicate the fact that an unreliable Packet Jitter value w i l l results in a lower C L for VoIP type service when compared with Web Browsing. Also in Table 5-3, it is assumed that CL(t) and CL(c) values for Packet Delay, Jitter and Loss w i l l be the same as these parameters are very closely related. Table 5-3 Lookup tables used by the decision maker for Confidence Level calculations Service Type CL(s) for Packet Delay, Jitter and Loss VoIP 0.5 Steaming Media 0.75 Web Browsing 1 Time since last Reading (sec.) CL(t) for Packet Delay, Jitter and Loss 0 1 600 0 Attribute CL(c) with Network Utilization (U) Packet Delay (D) 0.9 Packet Jitter (J) 0.9 Packet Loss (U) 0.9 We set the threshold for C L at 0.5. We assume that at the time of decision, the attribute values had not been updated for the past 5 minutes. Table 5-3 shows that in this case it has been 130 determined that the data values are not useful beyond 600 seconds or 10 minutes. This would therefore give us a normalized value of t=0.5. Since the service type in the example was streaming media, the normalized value of s is 0.75. Also per Table 5-3, the normalized correlation factor c for missing data is 0.9. Using these values in the prior equation we come up with a value for C L of 0.717 which is higher than the threshold and therefore allows for inclusion of the access type in network selection process. This type of simple calculations can help improve the reliability of results by incorporating the decision maker's judgment under less certain conditions. 5.3.2 Scenario 2 - Network Types with Non Crisp Attributes Some of the candidate networks may have a range of QoS values instead of one crisp value. This can be for different reasons: • A network may support multiple QoS classes or S L A s with different cost structures associated with them as described in Chapter 3. These QoS classes / S L A s can be treated as separate alternatives while using a compensating M A D M algorithm. The less expensive QoS classes / S L A s may have more variability on the QoS attributes and such variability can be represented by a nominal value for each attribute with a possible spread of values around it as represented by a T F N described earlier. • The access technologies being used by some candidate networks may be inherently incapable of providing strict QoS guarantees. This can also be based on the decision maker's prior experience with a operator network or access technology. The variability of this type can be represented by a fuzzy number. If the service or subscription information is considered sensitive to the non crisp attribute values then network selection in such situation can make use of a fuzzy M A D M algorithm as described in the previous section. Steps related to data harmonization, fuzzification and estimation of C L would not be required. The fuzzified input data is assumed to be available in this case (e.g., pre-provisioned) and after application of the fuzzy M A D M the results are defuzzifled to get crisp rankings. 131 Table 5-4 Attribute values for network alternatives under consideration CB TB AB U D J L (%) (Mbps) (Mbps) (%) (msecs) (msecs) (per 106) QoS ClassM/SLA#l 10 100 0.5 30 400 100 50 QoS Class#2/SLA#2 20 100 1 20 200 100 100 75 25 25 30 10 10 QoS Class#3/SLA#3 30 100 1 20 100 25 50 QoS Class#4/SLA#4 50 100 2 40 50 10 10 QoS Class#5/SLA#5 60 100 5 60 40 10 10 In order to better understand this type of usage of M A D M algorithms we consider a network selection scenario where multiple QoS /SLAs with a network operator constitute the alternatives, similar to the scenario described in Chapter 3. Further we assume that the user plans to use streaming media service with the attribute weights as shown in Table 5-1. The attribute values for the Q o S / S L A s classes are shown in Table 5-4. One of the less expensive alternatives, Ntwk#3 as shown in Table 5-4 does not provide crisp QoS attributes which indicates that there can be significant variation in the values during a an ongoing session depending upon congestion levels. Applying the fuzzy G R A algorithm using the process described in the earlier section we get the results shown in Figure 5.17. A s was discussed earlier, a change in attribute values for one of the network can change the values of the G R C s for other networks as well . This is shown in the results in Figure 5.17. In order to get crisp values in this case, defuzzification is applied to the G R C values of all the network alternatives. The results o f defuzzification process are shown in Figure 5.18. The crisp values obtained from Figure 5.18 indicate that Ntwk#3 has the maximum G R C value and is the preferred network in this case. 132 086, . li • -e—Fully GRA TFI4(») -O— Fully GRA TFM(b) 08 076 f-0 76 . - ' 4 072 * • \ * • : A NlwtdC Figure 5.17 Results of fuzzy G R A algori thm 133 07625 Triangular Fuzzy Set Triangular Fuzzy Set z> 0.51 Triangular Fuzzy Number Center of Gravity r-^ —\ ~i—•— 1 i ';/ i . . . / • • : • j i 1 i , 0797 07975 . 0.798 07985 0799 07995 . 0.8 . 0.8005 Triangular Fuzzy Set Figure 5.18 Use of C o G for defuzzification of T F N representing GRCs in the fuzzy implementation of G R A 134 5.4 Conclusion Selection of an optimal service delivery network is an important problem to be solved in an all-IP heterogeneous wireless networking environment spanning multiple operator domains. The problem requires special consideration when attribute related data is unreliable or unavailable. Prior work in this area has limited applicability as it has not looked at the problem in its entirety. For example prior studies did not take into consideration important aspects such as the sensitivity of the network selection to imprecise information, data prediction for unavailable data, use of fuzzy number types to represent predicted attributes and the decision strategy to determine when to apply fuzzy M A D M algorithms. In the absence of a well defined strategy to handle such scenarios, suboptimal networks could be selected. In this chapter we have provided a comprehensive network selection mechanism in the presence of imprecise information. We have considered scenarios where fuzzy techniques along with data prediction can be applied in the network selection decision process while using a fuzzy implementation of the Grey Relational Analysis M A D M algorithm. We have developed additional decision support techniques that can be applied under such conditions. These techniques have been used to obtain the Confidence Levels in the network rankings obtained via the use of fuzzy logic and data prediction with a M A D M algorithm. A n alternative approach for removing low confidence alternatives earlier on from the decision process has also been described. The techniques described in the chapter can be tuned to address different decision types and resolutions o f the data. Through examples we have shown that the techniques proposed in this paper can help improve the reliability of the results and allow for improved network selection under uncertain conditions. 135 5.5 References [I] K. Yoon and C L . Hwang, Multiple Attribute Decision Making: A n Introduction; Sage Publications, 1995. [2] Hwang C L . and K. Yoon, Multiple Attribute Decision Making: Mehtods and Applications, Springer Verlag, New York, NY 1981 [3] F. Bari and V. Leung, "Service delivery over heterogeneous wireless networks: network selecton sspects", Proc. ACMIWCMC, pp. 251-256, July 2006. [4] L A Zadeh, "Fuzzy Sets", Information and Control, Vol. 8, pp. 338-353, 1965 [5] I.A Zadeh, "Fuzzy Algorithms", Information and Control, Vol. 12, pp 94-102, 1968 [6] E. Triantaphyllou, and C. Lin, "Development and Evaluation of Five Fuzzy Multi-Attribute Decision-Making Methods," Approximate Reasoning, Vol. 14, No. 4, pp. 281-310, 1996. [7] E. Triantaphyllou, Multi-Criteria Decision Making Methods: A Comparative Study, Kluwer Academic Publishers, 2002 [8] S. Hongvan, H. Chen and J. Lingge, "Intelligent signal processing of mobility management for heterogeneous networks", Proc. International Conference on Neural Networks and Signal Processing, pp. 187-191, December 2003. [9] S. Kher; A.K. Somani, and R. Gupta, "Network selection using fuzzy logic"; Proc. 2nd International Conference on Broadband Networks, pp.:876 - 885,Oct. 2005 [10] W. Zhang, "Handover decision using fuzzy M A D M in heterogeneous networks", Proc. IEEE WCNC 2004, pp. 653-658, March 2004 [II] Q. Guo; J. Zhu and X . X u , " A n adaptive multi-criteria vertical handoff decision algorithm for radio heterogeneous network", Proc. IEEE ICC 2005, pp. 2769-2773, May 2005 [12] B.L. Bowerman, R.T. O'Connell, Forecasting and Time Series : A n Applied Approach, 3rd Edition, 2000, Duxbury Classic Series. [13] Q. Song and A. Jamalipour, "Quality of Service Provisioning in Wireless LAN/UMTS Integrated Systems using Analytic Hierarchy Process and Grey Relational Analysis", Proc. IEEE GlobeCom, pp. 220-224, November/December 2004. [14] J .L. Deng, "Introduction to Grey System", Journal of Grey System, vol . 1, issue 1, pp. 1-24, 1989. [15] S . M . Baas and H . Kwakernaak, "Rating and ranking of multiple-aspect alternatives using fuzzy sets", Automatica, V o l . 13, pp. 47-58, Pergamon Press, 1977. [16] G . Bortolan and R. Degani, " A review of some methods for ranking fiizzy subsets", Fuzzy sets and systems, Vol. 15, pp. 1-19, 1985. [17] F . Bar i and V . Leung, "Automated Network Selection in A Heterogeneous Wireless Network Environment", IEEE Network, pp. 34-40, January/February 2007. [18] F . Bari and V . Leung, "Multi-Attribute Network Selection by Iterative TOPSIS for Heterogeneous Wireless Access", Proc. IEEE CCNC, January 2007 [19] F. Bari and V . Leung, "Use of Non-monotonic Util t iy in Multi-attribute Network Selection", Proc. IEEE WTS, A p r i l 2007 137 6 Conclusions Inter-working of heterogeneous wireless networks has emerged as a commercially viable solution to the problem of efficient and ubiquitous delivery of services to nomadic and mobile users. Network selection is a critical problem to be addressed in interworked heterogeneous all-IP wireless systems. It has both revenue and service impacts. In our research we have provided solutions to address this problem. The fact that several industry forums are actively looking to solve this problem with multifaceted requirements indicates its importance for the industry. The issue is of special significance as it impacts both the operator's revenue and customer's service experience. The decision algorithms used in the networks are unlikely to be standardized but used by the operators and equipment manufacturers as means of differentiating their services or products. However many of the architectural ideas presented in the earlier chapters of the thesis are a topic of discussion in industry forums and standards bodies. I E E E 802.21 continues to work on Media Independent Handover (MIH) solutions, part of which is to solve the problem of network selection across heterogeneous network technologies during inter-technology handoffs. More standardization efforts would be needed to enable 802.21 type mechanisms in prevalent wireless wide area access technologies. The I E E E 802.11 and I E E E 802.16 groups are currently working to enable M I H based on I E E E 802.21 work and similar efforts maybe needed in other forums such as 3GPP. 3GPP independently is also looking into the problem of network discovery and selection in System Architecture Evolution as part of supporting seamless mobility across multiple access technologies. In fact as a validity of the architectural approach proposed here, various industry forums have realized the interference issues with two radio solutions during network selection process for seamless inter-technology handoffs and are now considering single radio solutions where one access technology network can provide information about networks with other access technologies available in the vicinity. We conclude this Chapter with a summary of the contributions of the thesis and future work that is possible in the area. 6.1 Summary of Contributions Architectural Framework Currently there is no good deployable solution in the industry to deal with the problem of network selection in a heterogeneous all-IP wireless networking environment. Our research and 138 in turn contribution to the work 1 [1] being done at IETF has provided a better understanding of the problem space and also established design constraints/recommendations for an acceptable solution. Using these guidelines we have proposed an architectural framework for network selection in a heterogeneous wireless networking environment in [2] and [3] and described it in Chapter 2. In our proposed framework, the network assists the terminal in the decision process. The proposed solution scales well , is flexible and also supports roaming. It. minimizes the usage of the wireless links and would work with currently deployed infrastructures. The architectural solution we have proposed in the thesis leverages the concept of Information Services defined in I E E E 802.21. Using IS at layer 3 via a client/server model is a notion that we have contributed to in I E E E 802.212[4] as well . We have also described how information exchange using layer 2 can occur in our architecture in the case of W L A N s . The I E E E 8 0 2 . l l u based layer 2 network selection system referred in the example in Chapter 2 is a proposal to which we have also contributed 3 [5] in I E E E 8 0 2 . l l u . Finally we have integrated our architecture with the current 1 The author of this thesis was one of several co-authors to the referred industry standards submission and therefore does not claim authorship of the entire referred document or the ownership of all the ideas presented in it. See reference for the complete list of authors. 2 The author of this thesis was one of several co-authors to the referred industry standards submission and therefore does not claim authorship of the entire referred document or the ownership of all the ideas presented in it. See reference for the complete list of authors. 3 The author of this thesis was one of several co-authors to the referred industry standards submission and therefore does not claim authorship of the entire referred document or the ownership of all the ideas presented in it. See reference for the complete list of authors. 139 cellular infrastructure by proposing to enhance the function of some of the existing functional nodes (PDF and A A A ) in the cellular system and adding new logical information nodes ( S A N and D C N ) . Comprehensive M A D M based Network Selection Decision Process Prior work on the network selection decision process has not considered some of the key decision attributes, e.g., multiple authentication credentials, multiple authentication mechanisms, user's mobility profile, and roaming relationships between the user's FIN and visited networks. This can be attributed to lack of understanding of the problem space, unavailability of an architectural framework to provide a solution in, and the inability to use one decision algorithm for all attributes that needed different type of decision making. Within the framework of our proposed architectural framework, we have developed a two step decision process for a network-assisted network selection mechanism that combines non-compensatory and compensatory M A D M algorithms [6]. Improvements to M A D M algorithms for application to network selection when complete information about the factors impacting network selection is available During the course of our research we have proposed improvements to existing M A D M algorithms or improvement towards their application to this problem space. We have identified causes of ranking abnormalities in TOPSIS and proposed an improvement to the standard TOPSIS algorithm when it is applied to network selection by adopting an iterative approach [7]. The need for support of different optimization objectives for the decision maker (e.g., the network operator) and hence the need to support non-monotonic utility for some of the attributes used in decision making has been demonstrated [8]. We have shown that many of the commonly used M A D M algorithms such as S A W , W P M , TOPSIS, and E L E C T R E in their standard forms are not well suited to network selection because of their assumptions about monotonically increasing or decreasing utility of each attribute. It has been shown that G R A is better suited for achieving this type of optimization objectives with the selection of delivery network impacted by how the algorithm is used to achieve the optimization objectives. A two step process has been proposed and evaluated to show how reference attribute values and attribute weights impact the selection process. The adjustment of these parameters for different service types has been discussed. E L E C T R E , a well known M A D M algorithm, has been modified [9] to allow its usage for network selection with attributes exhibiting a non-monotonic utility. This modification makes 140 the algorithm more adept to application in network selection as it expands its applicability to a wider range of optimization objectives. To the best of our knowledge the enhancements to the M A D M algorithms themselves or to their application in network selection listed above, are the first. In addition the improvements to the algorithms are not specific to the problem of network selection but can be applied to other problem spaces with similar requirements. Application of data prediction and fuzzy techniques with MADM based network selection for scenarios when some of the information to be used in network selection is missing Prior studies did not develop a decision strategy about when to apply standard M A D M algorithms and when to apply fuzzy M A D M algorithms in the presence of imprecise input information. We have proposed such a decision process for network selection with imprecise input information [10], which considers when fuzzy techniques along with data prediction can be useful in network selection decisions. We have proposed a new function called the Confidence Level for network selection decision support. Together with the sensitivity of the missing information, the two functions help determine i f it would be useful to estimate the data, i f the estimated data should be used in the decision process, and i f the results thus obtained could be trusted. To the best of our knowledge this is the first such proposal for solving the problem of network selection with imprecise input information. The decision entity applying the M A D M algorithm has been assumed to be in the network in our proposed architectural framework. However the application of the algorithms presented in the thesis to network selection is irrespective of the architectural framework. The algorithms described here would work equally well i f the decision making entity resides in the terminal. 6.2 Future Work In the coming years ideas presented in this thesis w i l l be validated or challenged in the market place. Along with technical viability of the solutions presented here it w i l l be the new and existing business models used by the network operators that w i l l determine the feasibility of solutions presented in the thesis. For example the idea of sharing network related information across partner networks although technically feasible w i l l eventually depend on how wil l ing the operators are in sharing their network related data. More work in the areas of Dynamic S L A s and Game Theory can help extend the application of these techniques to new business models. From the algorithmic perspective, the focus of the thesis has been on the use of deterministic decision making techniques. This has allowed for ease of implementation that is needed for real time decision 141 making and for the ease of understanding of the underlying decision making philosophy by the decision maker, e.g., the network operator. The possibility of using non-deterministic algorithms (e.g., use of Q-learning) can also be explored in future research while keeping in mind the requirements for ease of implementation and ease of understanding of the decision making philosophy. In our research we have assumed a central decision making entity located in the network. It is possible that the decisions are made in steps and at more than one locations. Thus a possible future research area can be in a multistage decision making approach that involves decisions by more one entities distributed in their locations. 142 6.3 References [1] J. Arkko , B . Abboba, J. Korhonen and F. Bari . , "Network Discovery and Selection Problem", draft-ietf-eap-netsel-problem-08, June 2007 (completed working group last call), [online] Available http://www.ietf.org. [2] F . Bar i and V . Leung, "Service Delivery over Heterogeneous Wireless Systems : Networks Selection Aspects", Proc. ACMIWCMC, pp. 251-256, July 2006. [3] F . Bar i and V . Leung, "Architectural aspects of automated netowrk selection in heterogeneous wirelss systems", Submitted to International Journal of AdHoc and Ubiquitous Computing (IJAHUC) [4] S. Das, Y . Ohba and F. Bari , "Information Service (IS) Reference Model , Use Case Scenario and Higher Layer Requirements for 802.21 Information Service (IS)", 21-05-0348-00-0000-HigherJayerRequirements InformationJService. doc, September 2005, [online] Available http://www.ieee802.org/21/. [5] N . Canopolat, V . Gupta, D . Stephenson, S. Sreemanthula, R . Y . K i m , E . Hepworth, S. Demel, M . Rudolf, F . Bari , H . Cheng, L . M o , Z . Yao , A . Soomro, and D . Cavalcanti, "Network Selection - Normative text proposal", IEEE Contribution 11-06-1014-03-000u-network-selection-normative-text-doc, September 2006, [online] Available ftp.802wirelessworld.com. [6] F. Bar i and V . Leung, "Automated Network Selection in A Heterogeneous Wireless Network Environment", IEEE Network, pp. 34-40, January/February 2007. [7] F . Bari and V . Leung, "Multi-Attribute Network Selection by Iterative TOPSIS for Heterogeneous Wireless Access", Proc. IEEE CCNC, January 2007. [8] F. Bari and V . Leung, "Use of Non-monotonic Util i ty in Multi-attribute Network Selection", Proc. IEEE WTS, A p r i l 2007. [9] F . Bar i and V . Leung, "Application of E L E C T R E to network selection in a heterogeneous wireless networking environment", Proc. IEEE WCNC, March 2007. 143 [10] F . Bari and V . Leung, "Network Selection with Imprecise Information in Heterogeneous A l l - I P Wireless System", accepted in Proc. Wireless Internet Conference (WICON), October 2007. 144 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0065457/manifest

Comment

Related Items