evaluation of artificial neural networks-based handover...

122
TECHNISCHE UNIVERSITÄT ILMENAU Fakultät für Elektrotechnik und Informationstechnik In Zusamennarbeit mit Kazan National Research Technical University named after A.N.Tupolev Evaluation of Artificial Neural Networks-based Handover Algorithms in Multi-Terminal and Multi-Application Scenarios Masterarbeit zur Erlangung des akademischen Grades Master of Science vorgelegt von: Ivan Goskov geboren: Studiengang: Communications and Signal Processing Studienrichtung: keine Anfertigung im Fachgebiet: Kommunikationsnetze Fakultät für Elektrotechnik und Informationstechnik Verantwortlicher Professor: Prof. Dr. rer. nat. Jochen Seitz Wissenschaftlicher Betreuer: Dipl.-Ing. Yevgeniy Yeryomin Datum: August 05, 2016

Upload: others

Post on 23-Oct-2020

2 views

Category:

Documents


0 download

TRANSCRIPT

  • TECHNISCHE UNIVERSITÄT ILMENAU

    Fakultät für Elektrotechnik und Informationstechnik

    In Zusamennarbeit mit

    Kazan National Research Technical University named afterA.N.Tupolev

    Evaluation of Artificial Neural Networks-basedHandover Algorithms in Multi-Terminal and

    Multi-Application Scenarios

    Masterarbeitzur Erlangung des akademischen Grades Master of Science

    vorgelegt von: Ivan Goskov

    geboren:

    Studiengang: Communications and Signal Processing

    Studienrichtung: keine

    Anfertigung im Fachgebiet: Kommunikationsnetze

    Fakultät für Elektrotechnik und Informationstechnik

    Verantwortlicher Professor: Prof. Dr. rer. nat. Jochen Seitz

    Wissenschaftlicher Betreuer: Dipl.-Ing. Yevgeniy Yeryomin

    Datum: August 05, 2016

  • i

    ACKNOWLEDGMENTS

    I would like to say a thousands thanks to my wife, for supporting me through foul andfair and my kids for patience and love.

    My grateful acknowledgement also addressing for the GRIAT sta� and it’s lectors.Thank you for high education level and passion for your job.

    Finally, prof. Adel Nadeev and Eugeniy Yeryomin to support me during Mastersdegree. I appreciate your advises and wisdom mentoring.

    Master thesis Ivan Goskov

  • ii

    ABSTRACT

    The modern telecommunications trends requires new approaches to the classical HOprocedures. The traditional RSS- based HO mechanisms became irrelevant to themodern technologies and services which should meet customer and applications re-quirements, but also to support the operator’s needs. Distinct approaches have beenintroduced in papers and industry to handle this issue, these are MCDM criteria de-cision algorithms, artificial neural networks (ANN), fuzzy logic etc.

    In this thesis we made the extensive review of the modern load balancing (LB) algo-rithms with the accent to the ANN. The criteria for load balancing and HO decisionshave been investigated in details. The ANN classification and parametrization havebeen discussed. Finally the LB algorithm have been selected and implemented in OM-Net++. It was analyzed and modified by the MCDM-based decision. We also selectedand train ANN and set up it’s interoperability with selected LB algorithm.

    Algorithms evaluation have been made in Matlab, however the algorithm, imple-mented in OMNet++ still requires the proper debugging. As a future work the createdOMNet++ code help to prove or refute our theoretical results.

    Master thesis Ivan Goskov

    yevgeniythat is not true, commonly used HO algorithms are RSS-based!

    yevgeniyOMNeT++

    „It was analyzed and modified by the MCDM-based decision“ the sentence is unlucky formulated.

    yevgeniyMATLAB

  • Contents iii

    Contents

    Acknowledgments i

    Abstract ii

    Contents iii

    List of Figures vi

    List of Tables vii

    1 Introduction 1

    2 Overview of the handover and used decision criteria 32.1 Heterogeneous network definition . . . . . . . . . . . . . . . . . . . . . 32.2 Analysis of meaningful vertical handover criteria . . . . . . . . . . . . . 72.3 Analysis of meaningful ANN-based HO criteria . . . . . . . . . . . . . . 82.4 Analysis of existing MCDM-algorithms . . . . . . . . . . . . . . . . . . 92.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

    3 Analysis of ANN-based handover decision algorithms and used criteria 113.1 ANN classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113.2 Multilayer Perceptron . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

    3.2.1 Activation function . . . . . . . . . . . . . . . . . . . . . . . . . 143.2.2 Backpropagation method . . . . . . . . . . . . . . . . . . . . . . 143.2.3 MLP properties . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

    3.3 ANN parametrization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163.3.1 Data collection . . . . . . . . . . . . . . . . . . . . . . . . . . . 173.3.2 Selection of the ANN topology . . . . . . . . . . . . . . . . . . . 173.3.3 Experimental selection of the ANN characteristics . . . . . . . . 183.3.4 Experimental selection of the training parameters . . . . . . . . 183.3.5 Training feasibility check . . . . . . . . . . . . . . . . . . . . . . 20

    Master thesis Ivan Goskov

  • Contents iv

    3.4 Evaluation and classification of ANN Algorithms for multi-criteria de-cision making . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

    3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

    4 Analysis of wireless networks load balancing techniques 254.1 Load Balancing in telecommunications . . . . . . . . . . . . . . . . . . 25

    4.1.1 Load balancing in computer networks . . . . . . . . . . . . . . . 254.1.2 Load balancing algorithms in heterogeneous network . . . . . . 26

    4.1.2.1 LB in WLAN . . . . . . . . . . . . . . . . . . . . . . . 284.1.2.2 LB in UMTS . . . . . . . . . . . . . . . . . . . . . . . 304.1.2.3 LB in LTE . . . . . . . . . . . . . . . . . . . . . . . . 31

    4.1.3 LB measurement metrics . . . . . . . . . . . . . . . . . . . . . . 364.1.4 LB e�ciency indicators . . . . . . . . . . . . . . . . . . . . . . . 384.1.5 Load balancing problems . . . . . . . . . . . . . . . . . . . . . . 394.1.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

    4.2 Proposed algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424.2.1 Novelty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424.2.2 ANN involvement . . . . . . . . . . . . . . . . . . . . . . . . . . 424.2.3 Criteria and Alternatives table . . . . . . . . . . . . . . . . . . . 434.2.4 Algorithm 1. RSS-based HO . . . . . . . . . . . . . . . . . . . . 434.2.5 Algorithm 2. MCDM-based decision, performed by STA . . . . 434.2.6 Algorithm 3. Load Balancing in Overlapping Wireless LAN Cells 44

    4.2.6.1 Algorithm assumptions . . . . . . . . . . . . . . . . . . 45

    5 Setup of a handover simulation and results analysis 485.1 Description of the scenarios . . . . . . . . . . . . . . . . . . . . . . . . 495.2 Description of network’s topology . . . . . . . . . . . . . . . . . . . . . 515.3 Description of tra�c properties . . . . . . . . . . . . . . . . . . . . . . 515.4 Collection of decision criteria in OMNeT++ . . . . . . . . . . . . . . . 525.5 ANN-training and simulation . . . . . . . . . . . . . . . . . . . . . . . 525.6 Load balancing algorithms evaluation in Matlab . . . . . . . . . . . . . 545.7 LB algorithm performance evaluation . . . . . . . . . . . . . . . . . . . 565.8 Implementation of load balancing algorithms in OMNeT++ . . . . . . 58

    5.8.1 Proposed algorithm flow and OMNeT++ implementation details 595.8.2 Implementation constraints . . . . . . . . . . . . . . . . . . . . 59

    6 Conclusions 62

    Master thesis Ivan Goskov

  • Contents v

    A Comparison LB and ANN-based algorithms for HO-decisions 65

    B Code of the LB implementation 66B.1 Matlab implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . 66B.2 OMNeT++ modification . . . . . . . . . . . . . . . . . . . . . . . . . . 73

    Bibliography 108

    Erklärung 114

    Master thesis Ivan Goskov

  • List of Figures vi

    List of Figures

    2.1 Heterogeneous network [1] . . . . . . . . . . . . . . . . . . . . . . . . . 4

    3.1 Scheme of the feed-forward backpropagation ANN [2] . . . . . . . . . . 14

    4.1 A typical RSS-based HO process . . . . . . . . . . . . . . . . . . . . . 27

    5.1 Training results of the ANN . . . . . . . . . . . . . . . . . . . . . . . . 555.2 LB indexes for evaluated algorithms . . . . . . . . . . . . . . . . . . . . 575.3 The steps for data processing . . . . . . . . . . . . . . . . . . . . . . . 59

    B.1 The screenshot of OMNeT++ topology . . . . . . . . . . . . . . . . . . 66

    Master thesis Ivan Goskov

  • List of Tables vii

    List of Tables

    4.1 Comparison of LB algorithms . . . . . . . . . . . . . . . . . . . . . . . 47

    5.1 Simulation parameters for ANN-training . . . . . . . . . . . . . . . . . 505.2 Collected simulation criteria . . . . . . . . . . . . . . . . . . . . . . . . 535.3 The ANN’s parameterization . . . . . . . . . . . . . . . . . . . . . . . . 545.4 Algorithm flow and OMNeT++ implementation details . . . . . . . . . 60

    A.1 Comparison of ANN for multi-criteria decision-making . . . . . . . . . 65

    Master thesis Ivan Goskov

  • 1 Introduction 1

    1 Introduction

    The convergence of di�erent communication networks and the use of multi-mode mo-bile devices give rise to multiple alternative communication paths between the mobileuser devices and the core network. This situation requires a path selection duringsession establishment and during handover between di�erent networks. Traditionalhandover algorithms usually consider the signal strength as single handover criterion.However, recently many papers applying to the multi-criteria handover algorithms,where multiple di�erent characteristics of communication paths (e.g. quality of service(QoS), security, monetary costs) and of mobile devices (e.g. location, battery level)are considered during the path selection. Taking into account many di�erent criteria,the HO decision process becoming more sensitive and meaningful. One of the mostpopular techniques applied to multi-criteria decision algorithms are Artificial NeuralNetworks (ANN). This project deals with the investigation on existing ANN-based han-dover decision algorithms including used criteria, their configuration and evaluation ina multi-terminal and multi-application environment.

    The object of this research is the ANN-based HO decision algorithm and the subjectof the research is a wireless access network. The goal of this research is evaluationof artificial neural networks-based handover algorithms in multi-terminal and multi-application scenarios. Thesis structure described below:

    In Chapter 2, author made a survey about handover types and meaningful HO andANN-based decision criteria, and review common MCDM algorithms in telecommuni-cations.

    In Chapter 3, author review existing methods for ANN classification and madeanalysis of ANN-based handover decision algorithms. The most appropriate ANNtype have been selected in the context of the handover decision problem and di�erentscenarios. The relevant ANN parameterization sets for di�erent scenarios have beenidentified. This task required analysis on meaningful criteria for handover decisionalgorithms.

    Chapter 4 refers to analysis of handover evaluation techniques. Proposed techniquesfor ANN-based handover algorithms evaluation were studied. The analysis of strategies

    Master thesis Ivan Goskov

    yevgeniywhat are ANN-based criteria?

    yevgeniy!!!The substantial part of this section is partly copied from the task description!!!

    the introduction is not complete in terms of task and motivation description.

  • 1 Introduction 2

    and techniques from the service provider perspective, with attention to load balancingproblem was made. Finally, an evaluation method for handover decision algorithmfrom the service provider perspective have been proposed.

    In Chapter 5 author setup a handover simulation model in simulation environmentOMNet++ Handover scenario with multiple mobile terminals and multiple applica-tions have been configured. The handover-relevant criteria parameters have been col-lected and used for handover decision algorithms. The models and the implementationsof the criteria in OMNet++ have been studied and documented. ANN-based decisionalgorithms have been integrated. A handover performance evaluation have been ana-lyzed. The most appropriate ANN-based handover decision algorithm with the mostappropriate configuration set and the most meaningful criteria set for di�erent scenar-ios have been determined by OMNet++ simulation.

    Master thesis Ivan Goskov

  • 2 Overview of the handover and used decision criteria 3

    2 Overview of the handover and useddecision criteria

    2.1 Heterogeneous network definitionFollowing wikipedia the heterogeneous network definition is changing from the con-text. Semantically, the term heterogeneous network can have di�erent connotationsin wireless telecommunications. For instance, by [3] the heterogeneous network is "anetwork that consists of a mix of macrocells and low-power nodes, where some maybe configured with restricted access and some may lack wired backhaul, is referred toas a heterogeneous network". By [1] "heterogeneous network is a network, involving amix of radio technologies and cell types working together seamlessly" (see Figure 2.1)."Heterogeneous network configuration exploiting macro and low power nodes can inprinciple support arbitrary data volumes and very high data rates" [4]. It may alsorefer to the "paradigm of seamless and ubiquitous interoperability between variousmulti-coverage protocols" (i.e. HetNet). Alternatively, it may refer to the non-uniformspatial distribution of users or wireless nodes (i.e. spatial in-homogeneity). Therefore,the "heterogeneous network" term may cause confusion in scientific literature withoutcorresponded context.

    In general, the heterogeneous network can be divided into the following types ofnetworks depending on their scale:

    • Personal wireless networks of WPAN (Wireless Personal Area Networks) are wire-less networks which cover the small geographical territory (for example, Blue-tooth in the ad-hoc mode, ZigBee);

    • Local wireless networks of WLAN (Wireless Local Area Networks) - networkswhich cover the area of 50-300m and used in buildings of o�ces, schools, cam-puses (for example, standards of the IEEE 802.11 Wi-Fi series);

    • Metropolitan wireless networks of WMAN (Wireless Metropolitan Area Net-works) - wireless networks of the city scale (for instance, WIMAX);

    Master thesis Ivan Goskov

  • 2 Overview of the handover and used decision criteria 4

    Figure 2.1: Heterogeneous network [1]

    • Wide area wireless networks or WWAN (Wireless Wide Area Networks) arenetworks of big scales, built on cellular technologies GSM, EDGE, 3GPP, LTE.

    Handover is the key procedure, which allows subscribers to move transparentlywithin the heterogeneous wireless network. According to the ETSI and 3gpp [5]: "HOis process in which the radio access network changes the radio transmitters or radioaccess mode or radio system used to provide the bearer services, while maintaining adefined bearer service QoS". In other words, it is process of transferring all communi-cation sessions of a mobile station (STA) from one access point (AP) or base station(BS) to another. Handover between adjacent network segments of one technologycalled horizontal; HO between di�erent types of networks - vertical. If subscriber ses-sion control carried out without loss of the current connection, the hando� called soft(or "make - before - break"). This HO type carried out by two methods: the mobilestation can communicate with several access devices at the same time, or use one APas a proxy, meaning that packets can be redirected from one BS / AP to another. Thefirst type of a soft hando� realized by means of several network interfaces at the sametime, however carried out with use of the single network card. In the second case, theSTA can be associated with only one AP at each time slot. This handover type called

    Master thesis Ivan Goskov

    yevgeniy3GPP

  • 2 Overview of the handover and used decision criteria 5

    hard (or "break - before - make") when STA firstly disconnecting from current AP andonly then connection with new is set. Depending on the handover initiator the HOcan be initiated by mobile station (called "mobile-controlled"), or network ("network- controlled") or mobile station with network cooperation ("mobile-assisted"). A ver-tical handover can be classified depending on the reason: ascending and descending.The descending handover carried out from a bigger network with lower data rate toa smaller network with higher data rate. The descending HO main feature is a delayinsensitivity. The ascending handover comes from a smaller network with greater bi-trate to a bigger network with a smaller data bitrate. This HO type taking place whenSTA leaves a network with a smaller coverage, therefore this HO type is sensitive todelays.

    Main HO requirements are:

    • low procedure delay;

    • high reliability;

    • high success level;

    • minimal number of HOs (STA power management and QoS assurance);

    • Multi criteria decision making- there should be possibility to make the HO de-cision based on number of parameters, ensuring transparent movement of thesubscriber;

    • Load balancing problem should be considered.

    he vertical handover technique should consider the most appropriate time for trig-gering the process and the most suitable radio access technology among all available.The traditional mechanism of a horizontal handover based on received signal powerassessment or Relative Signal Strength (RSS). Handover started when the RSS level ofthe serving base station becomes lower than some threshold value. On the other hand,the mobile node in the heterogeneous wireless environment can move between accessnetworks with di�erent characteristics, which cannot be compared directly. Therefore,RSS is not su�cient criterion for implementation of e�ective and intellectual decisionmaking of vertical handover. Other metrics, such as cost, network load, availablebitrate, QoS etc can be also considered. The HO procedure consist of three stages:information collection about available APs, decision making about the best candidateAP for HO and handover execution.

    Master thesis Ivan Goskov

    yevgeniya table for the handover types could have been appropriate here

    yevgeniyhe

    yevgeniyall phrases here have to start with capital letters

  • 2 Overview of the handover and used decision criteria 6

    • During information collection phase (networks detection) STA finds availablewireless networks by passively scanning of radio medium and receiving the ser-vice information from these networks. This process leads to high STA energyconsumption. In order to reduce the power consumption, STA making periodicscanning of the wireless interfaces. It is also necessary to collect information con-cerning STA status (battery power, resources, speed, a class of services, etc.), anduser preferences (cost, technology, etc.). During network parameters scanningSTA receiving data, which can trigger the HO: it can be either signal strength,signal to noise and interference ration, channel quality indicators, bit errors, etc.

    • During decision-making stage the system (STA or network) deciding when andwhere to make a handover. It can be done based on algorithm-based assess-ment of the received parameters, like network capacity, load, number of users,monetary cost etc.

    • At the handover execution stage, the system making direct transfer of the cur-rent STA session to the chosen AP or BSS. This process requires transferringinformation, concerning routing and other context information of mobile station(BS) related a new AP.

    There were attempts of standardization of the vertical hando� procedure made byInstitute of Electrical and Electronics Engineers (IEEE), which created 802.21 standardcalled "Media Independent Handover" [6]. The standard architecture defines three mainservices, which are in charge of mobility services: Event, Command, and Informationservices. These services transmit information from the bottom to upper levels ofthe OSI model through the MIHF (Media Independent Handover Function) protocol.The MIHF protocol takes place between data link and network OSI model levels.It helps to make the work of all applications (upper OSI levels) transparent for lowOSI level standards (IEEE 802.3, 802.11, 802.16, 3GPP/3GPP2). In other words, amain goal of this standard is optimization of mobility mechanisms of heterogeneousnetworks. The standard is contains only general description of system design principlesand doesn’t contain a specific hardware or software algorithms implementation, i.e.the implementation details, including HO decision making algorithms remain left fordevelopment engineers consideration.

    Master thesis Ivan Goskov

  • 2 Overview of the handover and used decision criteria 7

    2.2 Analysis of meaningful vertical handover criteriaThe existing HO algorithms can be classified based on parameters, which considered forHO decision and on methods of its processing. Following [7] the following parameterscan be used for HO decision:

    • User parameters (provider, budget), user’s profile, user’s status (velocity, loca-tion), connection duration etc;

    • QoS requirements parameters (bandwidth, packets loss, delay, Jitter, bit errorrate), safety, session duration. The QoS (Service Level Agreement) level declaredby providers and Operators based on the recommendation of ITU - T Y.1541 [8]

    • Tra�c class: conversational, streaming, interactive, background;

    • Network parameters: tra�c, load (available bandwidth, the current number ofconnections), monetary cost, coverage, the supported tra�c classes and STAvelocity limitation;

    • Signal level (RSS): this simple criteria is the most widespread for both verticaland horizontal HOs. This parameter directly influences applications QoS. Thesignal’s measured level of a signal is inversely proportional to distance betweenMS and BS.

    • Signal to Interference and Noise Ratio – (SINR) is a characteristic of channelstability, in other words it acts as an indicator of channel quality;

    • Available Bandwidth, measured in bps, used for determination of the accessnetwork load and is an indicator of available telecommunication resources;

    • Packet Loss Ratio - is a critical unit for data tra�c exchange;

    • Packet Jitter is a characteristic of delay variation for packet transfer. Jitter iscritical for real-time tra�c (IPTV, VoIP);

    • Hando� Latency. This parameter is determined as time, which has passed fromthe last received packet of source AP until the first received packet of new AP,after a successful handover. This parameter can vary in considerable limits fordi�erent wireless technologies;

    Master thesis Ivan Goskov

  • 2 Overview of the handover and used decision criteria 8

    • Network Connection Duration is a time interval during which an STA supportsconnection with a certain AP. This parameter depends on location and velocityof STA.

    • Monetary Cost: Di�erent operators within heterogeneous network can have dif-ferent tari�s. The AP with the smallest cost should have the priority from theend-user perspective;

    • Security. Some users require a higher confidentiality degree and data integrity(for instance bank transactions). Therefore, for these users the more protectedAP will be in priority;

    • Power Consumption. The HO process requires certain amount of power con-sumption. When the mobile node has a low battery charge, then it will selectthe AP which will help to keep the battery charge more longer.

    • Velocity is important HO decision factor, due to it a�ecting with session time andSTA location updates. Particularly, high- velocity of STA can lead to excessivenumber of handovers.

    2.3 Analysis of meaningful ANN-based HO criteriaBased on results of Advanced Research Project, the ANN meaningful criteria can beseparated to the network-related criteria and user-related criteria, which help to defineend-user profit of ANN- based VHO. Take in to account streaming data scenario, inprevious work, we considered bandwidth as the priority user parameter for comparison.The algorithm performance was evaluated, based on methodology proposed in [9].

    HO ratio. HO ratio defined as ration of HOs number to overall number of callassociations. It is necessary to minimize HO, for instance in LTE, as it a�ecting theoverall user experience and causes additional signaling load [10]. Therefore, it can beused as one of the HO performance indicators.

    Ping Pong Rate. One of the significant problems a�ected service quality, is a ping-pong e�ect, occurred when MT faced multiple HOs in a short time-frame. In orderto catch this e�ect in previous work we selected 2 seconds time window and start toincrease PP trigger when number of HOs more than one in selected window. A widerange of papers addresses to the problem of ping-pong negotiation,[11] , [12] , [13] ,but take in to account the structure of the referred system model, in previous work it

    Master thesis Ivan Goskov

    yevgeniywhat do have these criteria to do with the ANN?

  • 2 Overview of the handover and used decision criteria 9

    was decided to use the simplest one, relying to the best alternative selection based onits mode, i.e. the most frequent realization in time window.

    Computation Latency. In accordance with [14] handover latency is not a subjectto regularization, whereas in general if completed less then 100 ms, it is non-tender forthe end-user. In previous work, we used Matlab function Tic-Toc to measure networkperformance; in particular, Tic has been set on the algorithms inputs and Toc rightafter best score has been selected.

    ANN dynamic nature. Based on analysis of previous work results, author dis-covered an importance of dynamic nature of the input data. It was conclude that itseriously a�ect not only output data, but also whole ANN design. Several use cases,which limiting the ANN-based HO algorithm design were discovered:

    • In case when number of criteria changed, ANN should respectively change itsstructure and to be re-trained.

    • In case customer has changed the tra�c profile or service priorities, MCDMalgorithm weigh matrix changes. That requires again re-train ANN, whereasstructure remains the same.

    • For use case of alternatives set change network do not change structure, butrequire re-training.

    2.4 Analysis of existing MCDM-algorithmsIn previous paper, we selected SAW, GRA and VIKOR MCDM algorithms for anal-ysis and performance comparison with ANN-based algorithms. Following [15] , SAWdemonstrates better results for voice tra�c, whereas GRA and VIKOR are better fordata tra�c. Largely, the e�ectiveness of these algorithms depends on priority levels(weigh coe�cients) of each considered criterion, which can be explained by di�erentsensibility of tra�c types to di�erent quality parameters. In this paper we will useGRA algorithm for MCDM- based HO decision making and will use it for training theselected ANN.

    As normalization procedure of input values cij Max-Min was chosen:

    dij =maxi(cij) ≠ cij

    maxi(cij) ≠ mini(cij), dij =

    cij ≠ mini(cij)maxi(cij) ≠ mini(cij)

    (2.1)

    The Grey Relational Analysis (GRA) algorithm is used for resolution decision-making problem [16] , in fact, this algorithm shows degree of similarity of the AP

    Master thesis Ivan Goskov

  • 2 Overview of the handover and used decision criteria 10

    to the ideal solution:

    1. Obtaining of the ideal d+j solution:

    d+j = maxi dij for higher-the-better criteria ;

    d+j = mini dij for lower-the-better criteria;

    2. Calculation of the Grey’s coe�cient:

    GRCi =1

    qNj=1 wj|dij ≠ d+j | + 1

    (2.2)

    3. The higher the index is, the closer AP to the ideal solution.

    The algorithms performance analysis demonstrated that this algorithm has the bestbenchmark among the SAW, VIKOR and TOPSIS algorithms.

    2.5 ConclusionsThere are many approaches to HO algorithms. They di�er by a set of parameters,which are considered for best AP assessment, by mechanism of parameters handling,by mechanism of handover triggering and by the best AP selection algorithm. ClassicalHO algorithms does not consider, QoS, services cost, and other important network andend- user parameters. In order to improve best AP selection flexibility, it’s necessary tofind the balance between the chosen set of parameters and complexity of system. Suchtools as Multi-Criteria Decision-Making (MCDM) algorithms and Artificial NeuralNetworks (ANN) should help to reach this tradeo�.

    Master thesis Ivan Goskov

    yevgeniy„selection flexibility, it’s necessary to find the balance between the chosen set of parameters and complexity of system“why complexity? Complexity was not mentioned in this chapter at all.

  • 3 Analysis of ANN-based handover decision algorithms and used criteria 11

    3 Analysis of ANN-based handoverdecision algorithms and used criteria

    In previous work we used three types of the artificial neural networks (ANNs) Mul-tilayer perceptron (MLP), Cascade association neural network (Cascade) and Elmanrecurrent network (Elman). Elman ANN demonstrated better approximation proper-ties compare to other two types, but this kind of recurrent networks require a rigorousdata selection and training process, for appropriate result, due to the negative propertyto accumulate the error. Cascade neural networks [17] , [18] is a multilayer network,which trained by constructive method, when new neurons added during the learningprocess. The algorithm works in a way that learning is applied to only one layer ofneurons. This approach bring fast learning of the network, where most of the net-work parameters, as architecture, network size or learning parameters, are definedby learning process itself. Cascade networks demonstrate almost the same results asMLP, if the second has near-optimal parameterization. In this chapter, we are goingto overview the ANN through its concept, classification and parameterization. Thenwe analyze ANN-based handover decision algorithms and used criteria. The goal is toselect the most suitable type of ANN in context of the handover decision problem.

    3.1 ANN classificationThere are several approaches proposed to the ANN classification [2], [19], [20] , [21]:

    • Classification based on the input information.

    – Analogue ANN operating with information in form of real numbers.

    – Double ANN operating processing binary form data.

    • Classification based on management type.

    – Supervised learning when the input space of the ANN is previously knownto ANN.

    Master thesis Ivan Goskov

    yevgeniyCitation is required here

  • 3 Analysis of ANN-based handover decision algorithms and used criteria 12

    – Unsupervised learning when ANN creating the output solution space byitself, based only on input data. These kind of networks called – self-organized.

    – Reinforcement learning – is the system of rewards and penalties from envi-ronment.

    • Classification based on synapses settings.

    – ANN with fixed synapses – when weight coe�cients selecting based on thetask i.e. dW/dt = 0, where W – are the weight coe�cients of synapses.

    – ANN with dynamic synapses – ANN which set up synapses during ANNtraining, i.e. dW/dt! = 0, where W – are the weight coe�cients of synapses.

    • Classification based on the signal transmitting time.

    – For some neural network kinds the activation function can depend not onlyof the weight coe�cients wij but also on impulse transmission time in signallinks ·ij. Therefore in general, the activation function cij from element uito the element uj can be expressed as:

    cúij = f [wij(t), uúi (t ≠ ·ij)] (3.1)

    Then the network called synchronous when the signal transmission time ·ijof each synapse is equal to zero. The ANN called asynchronous when signaltransmission time of each synapse is di�erent and constant.

    • Classification based on the ANN synapses direction.

    – Feedforward networks. [22]:

    � Feed-forward backpropagation networks, which characterize by fixedstructure, iterative learning and weigh correction based on weights.

    � Self-organized maps (Kohonen map)

    � Other networks (as cognitrons, neo-cognitrons)

    Benefits of this type are implementation simplicity and guaranteed resultafter per- layer data processing. Disadvantages are possible data over fitting,dependence on local minima and learning rate step size. Many techniquesexist to avoid these problems while ANN design plays crucial role.

    Master thesis Ivan Goskov

  • 3 Analysis of ANN-based handover decision algorithms and used criteria 13

    – ANN with recurrent associations. [22] can remember and reproduce fullsequence of reactions to one stimulation by virtue of feedback association.It can be interpret as cyclic operation, from programming perspective andas finite state machine from systems perspective. The network structurehas dynamic delay blocks, which allowed to process dynamic models data.Disadvantage is requirement of special conditions, which guarantee a conver-gence of calculations or complicated behavior dynamics. Many properties ofrecurrent networks are ill explored due to complicity of practical realizationsand di�culty of results analysis. The typical examples are:

    � Hopfield network

    � Elman network

    – Specialized ANN structures provide optimal selection of the network archi-tecture, combining definition of its structure in accordance with learning.The good examples are Falman’s cascade correlation ANN and Volteri net-work.

    3.2 Multilayer PerceptronOne of the most popular kinds of feedforward backpropagation networks is multi-layernetwork. It consists of neurons, located in di�erent layers and contains of input, outputand at least one hidden layer. [2] Generic structure of two-layer perceptron presentedon Figure 3.1. The output signal of i-th neuron of hidden layer can be written as:

    ui = f{Nÿ

    j=0w(1)ij xj}, i = 1, 2, . . . K, (3.2)

    and output signals as:

    yl = f{Kÿ

    i=0w(2)li ui} = f{

    Kÿ

    i=0w(2)li f(

    Nÿ

    j=0w(1)ij xj)}, l = 1, 2, . . . M (3.3)

    Where f(. . .), usually, sigmoid activation function, x - input signals, u – outputsignal of hidden layer, y – output signal of output layer. Based on equations 3.3, 3.2it can be seen that weights of both layers make influence to the output signal, whereassignals on the hidden layer, indi�erent of output weights Wli.

    Master thesis Ivan Goskov

  • 3 Analysis of ANN-based handover decision algorithms and used criteria 14

    Figure 3.1: Scheme of the feed-forward backpropagation ANN [2]

    3.2.1 Activation functionThe important concept of the ANN is activation function that also called as a firingrule [2]. It is calculating, should the neuron fire, dependent from the input pattern. Inpractice the neuron output activating in the following manner:

    S =nÿ

    j=1xjwj ≠ b, y = f(S) (3.4)

    where xj are multipliers, wi – weights of synapses, b – bias, f – activation function,y – output.

    A bias unit allow neurons to learn an appropriate threshold, i.e. after reaching acertain total input, start sending positive activation. A bias value allows shifting theactivation function to the left or right, which may be critical for successful learning.

    There is a number of activation functions can be found in literature [2], [22], likePiecewise linear, Logistic, Tan-Sigmoid, Binary threshold, Sign etc.

    3.2.2 Backpropagation methodBackpropogation method is very often used for training the MLP. This algorithm is theform of supervised learning. It is the most common method of obtaining the weightsin the multilayer perceptron. It realize the training of artificial neural networks usedin conjunction with an optimization method such as gradient descent.

    The method calculates the gradient of a loss function with respects to all the weights

    Master thesis Ivan Goskov

    yevgeniy

    yevgeniygraphical presentation of the activation functions would have been appropriate

    yevgeniyrealizeS

    yevgeniyPlagiat!!!https://en.wikipedia.org/wiki/Backpropagation

  • 3 Analysis of ANN-based handover decision algorithms and used criteria 15

    in the network. The gradient is fed to the optimization method, which in turn uses itto update the weights, in an attempt to minimize the loss function. Backpropagationrequires a known, desired output for each input value in order to calculate the lossfunction gradient. It is therefore usually considered to be a supervised learning method,although it is also used in some unsupervised networks such as autoencoders. It is ageneralization of the delta rule [2] to multi-layered feedforward networks, made possibleby using the chain rule to iteratively compute gradients for each layer.

    Backpropagation requires that the activation function used by the artificial neurons(or "nodes") be di�erentiable. The goal of any supervised learning algorithm is to finda function that best maps a set of inputs to its correct output. Some input and outputpatterns can be easily learned by single-layer neural networks (i.e.perceptrons).

    The goal and motivation for developing the backpropagation algorithm was to finda way to train a multi-layered neural network such that it can learn the appropriateinternal representations to allow it to learn any arbitrary mapping of input to output.Each propagation involves the following steps: 1. Forward propagation of a trainingpattern’s input through the neural network in order to generate the propagation’soutput activations. 2. Backward propagation of the propagation’s output activationsthrough the neural network using the training pattern target in order to generate thedeltas of all output and hidden neurons.

    The backpropagation method has following problems during backpropogation algo-rithm:

    The result may converge to a local minimum. The "hill climbing" strategy of gradi-ent descent is guaranteed to work if there is only one minimum. However, often theerror surface has many local minima and maxima. If the starting point of the gradi-ent descent happens to be somewhere between a local maximum and local minimum,then going down the direction with the most negative gradient will lead to the localminimum:

    • Gradient descent can find the local minimum instead of the global minimum

    • The convergence obtained from backpropagation learning is very slow.

    • The convergence in backpropagation learning is not guaranteed.

    • Backpropagation learning does not require normalization of input vectors; how-ever, normalization could improve performance Another big problem with mul-tilayer neural network is overfitting.

    Master thesis Ivan Goskov

    yevgeniy

    yevgeniy

    yevgeniy

    yevgeniy

    yevgeniy

    yevgeniy

    yevgeniy

    yevgeniy

    yevgeniy

    Plagiat!!!

    Plagiat!!!

    yevgeniyPlagiathttps://quizlet.com/144486762/artificial-intelligence-flash-cards/

    yevgeniyhttps://en.wikipedia.org/wiki/Backpropagation

    yevgeniyhttps://en.wikipedia.org/wiki/Backpropagation

    yevgeniyPlagiat!!!https://www.passeidireto.com/arquivo/6190918/aula-10---redes-neurais-artificiais/3

    yevgeniyPlagiat!!!https://books.google.de/books?id=YndBDAAAQBAJ&pg=PA440&lpg=PA440&dq=made+possible+by+using+the+chain+rule+to+iteratively+compute+gradients+for+each+laye&source=bl&ots=4eFjEoiYQf&sig=VxLM0mtO4QxejBOBfcwOqMGRc8o&hl=de&sa=X&ved=0ahUKEwjlnp6lzbfOAhXKjSwKHYBhByoQ6AEIHDAA#v=onepage&q=made%20possible%20by%20using%20the%20chain%20rule%20to%20iteratively%20compute%20gradients%20for%20each%20laye&f=false

    „Arterial blood gases forecast optimization by artificial neural network method „ Wieslaw Wajs, …

  • 3 Analysis of ANN-based handover decision algorithms and used criteria 16

    3.2.3 MLP propertiesMLP can be characterized by [23]:

    • Usage of the non-linear activation function, usually sigmoid function.

    • The number of the trained layers is more than one, whereas in papers authorsusing not more then three.

    • The received and outgoing signals not always binary, but can be coded by decimalnumbers. The normalization is also required at least for the output data in away that values changed from 0 to 1.

    • It allowed the arbitrary architecture of the ANN, even fully connected neuralnetwork

    • The ANN error calculating as a statistical measure of discrepancy between re-quired and received value.

    • The ANN training finishing not until the training error absence, but until theweight coe�cients stabilization.

    3.3 ANN parametrizationThe ANNs are widely use in many applied areas such real-live applications, neuro-sience, machine lerning etc. There are several stages required in order to set up theANN for the task completion [21]. These steps in high-level described below. TheANN parameterization is a complex process, depending on the aspects of the researchtask. ANN parameterization can be perform on the experimental selection of the ANNcharacteristics and experimental selection of the training parameter stages.

    • Data collection for ANN training

    • Data preparation and normalization

    • Selection of the network topology

    • Experimental selection of the ANN characteristics

    • Experimental selection of the training parameters

    • Training

    • Training feasibility check

    Master thesis Ivan Goskov

    yevgeniyallowes

    yevgeniyuseD

    yevgeniyIntroduction for the enumeration is missing colon at the end of the sentence is missing

  • 3 Analysis of ANN-based handover decision algorithms and used criteria 17

    3.3.1 Data collectionThe most complicated stage of the task completion is training data selection and it’spre-processing. The set of the input data should satisfy the following criteria:

    • Representativity. Data should demonstrate the true facts of the matter in desiredarea;

    • Consistency. Inconsistent data in training set lead to the unsatisfied networktraining quality.

    The basic data should be transform in a way to send to the input ANN neurons.Every recording in the input data base called "training pair" or "training vector".Training vector contains at least one value for each network input, depending on thelearning type (supervised or unsupervised) and at least one output value for networkoutput. Network learning based on the raw set, as a rule, do not lead to the qualityresult. Whereas there are possibilities to increase the network perception.

    • Data normalization performed when inputs have various data dimensions. Thedata normalization help to merge the dimensionality of all inputs and outputsdata.

    • Quantizing applied to the continuous data, to which defined the finite numberof discrete values. For instance, it using to set up the sound signals frequencyduring speech recognition.

    • Filtering requires for noisy data.

    3.3.2 Selection of the ANN topologyThe ANN topology selection is strictly dependent on given tasks and input data.Supervised learning requires an "expert" assessment for each sample unit, whereas forbig data array the reception of such assessment can be impossible. In this case, thebest choice is to use the unsupervised learning for network (for example, to operatewith the self-organized Kokhonen’s map or Hopfild ANN). In case of the other taskssolution (such as forecasting the time series), the expert evaluation already includedin basic data and can be separated if necessary. For such cases, it is possible to use amultilayered perceptron or Ward network- the artificial neural network which topologyis characterized by the fact that the neurons of the hidden layer are broken into units.

    Master thesis Ivan Goskov

    yevgeniytransformationwhat are the basic data?

    yevgeniyquality -> qualitative

    yevgeniyFollowing from my previous review was not considered:„Please answer just two questions here:What are noisy data in this context?How filtering should work here?!“

  • 3 Analysis of ANN-based handover decision algorithms and used criteria 18

    3.3.3 Experimental selection of the ANN characteristicsWhen ANN topology is define, it is necessary to experimentally select network param-eters. For networks, similar to a perceptron, there are:

    • number of layers

    • number of neurons in the hidden layers

    • existence or lack of bypass connections

    • activation functions of neurons.

    On the one hand, there is a general rule, which should be consider for defining thenumber of layers: the network generalization ability is higher, when the total number ofcommunications between neurons is high. On the other hand, the number of synapseslimited from above by the number of entries in the training data table.

    3.3.4 Experimental selection of the training parametersWhen the ANN topology have been selected, it is required to choose the networktraining parameters. This stage is crucial for supervised learning. The appropriateparameters selection influence impacting the data convergence during the training.

    Optimization methods. The task of training an artificial neural network can be con-sidered as a task of optimization, and the main problem is to choose among a varietyof optimization methods most suitable. The choice in favor of gradient methods basedon the fact that, as a rule, in the problems of network teaching objective functioncan be expressed as a di�erentiable function of all weighting coe�cients. However,the complicated nature of dependence between weighting coe�cients leads to the factthat the objective function has local extremums and saddle points, making the use ofgradient methods are not always justified. For solving optimization problems with mul-tiextremal criterion the methods of random search, which include genetic algorithms,are used. Genetic algorithms, as a rule, are notable for their slow convergence. Au-thors [24] showed that gradient methods have fast convergence only at the beginningof training, and the genetic algorithm – in the end.

    Learning rate. The low training speed increases the convergence time, whereashelping to eliminate network paralysis (when network loosing ability to learn). Inopposite, the fast learning rate can lead to low data approximation.

    Momentum. Momentum Is a technique that can help the network out of local min-ima. It simply adds a fraction m of the previous weight update to the current one.

    Master thesis Ivan Goskov

    yevgeniydefineD

    yevgeniyThis section is not about „experimental selection of the training parameters“. It is just about description of them.

  • 3 Analysis of ANN-based handover decision algorithms and used criteria 19

    Increasing of the moment can lead both increase and decrease of the data convergence,dependent of the shape of error surface (function and its local minima). Particularly,when combining a high learning rate with a lot of momentum, it’s possible to past theminimum with huge steps.

    Learning rate and momentum complement each other; therefore, its values should beselected experimentally, paying attention the learning process criteria. For example,it is necessary to select either error minimization or reduce of the learning time.

    Training epoch. During training process, the network looks through the trainingselection in a certain order. The review order can be consecutive, casual etc. Somenetworks which are trained via unsupervised learning (for instance, Hopfild network),look through selection only once. Others, for instance, Kohonen networks, like thenetworks which are using supervised learning, look through selection several times.One full data review called a training epoch.

    Training, test and validation data sets. During supervised learning the dataset reg-ularly divided in two parts- training data set and test data; the division principle canbe arbitrary. The training data used for network training, whereas the test data us-ing for calculation of the network error. In other words, test data are never appliedto training of a network. Validation set for network architecture tuning. Practically,use of inconsistent data can bring decreased regression between network outputs andexamples, given it training set and increase of validation error- the parameter, usedfor stop validation process [18].

    Regression of the data. In statistical modeling, regression analysis is a statisticalprocess for estimating the relationships among variables. Regression R=1 indicatingexact linear dependence of ANN outputs and required outputs. If this value is closeto 0 then there is no linear dependency between these two values [18].

    Overfitting and paralysis. If the number of mistakes decreases on the test data, thenthe network works appropriately and makes correct data generalization. If numbermistakes on the training data continues to grow, and the mistake on test data increases,so the network has ceased to carry out generalization and just "remembers" the trainingdata. This phenomenon called retraining of a network or an overfitting. In this casetraining usually stopped. Other problem can appear during network training, such asparalysis or negative e�ect, when a network hits local minima of a mistakes surface.It is impossible to predict an appearance of mentioned problems in advance, as wellas to make unambiguous recommendations to their resolving.

    Master thesis Ivan Goskov

  • 3 Analysis of ANN-based handover decision algorithms and used criteria 20

    3.3.5 Training feasibility checkThe validation of the network learning quality should be made on samples, di�erentfrom the training set. The amount of the validation data set should be proportional tothe target learning quality. For instance, if the ANN error has 1 ·10≠5 probability, thenthe validation set should contain 1 · 10≠5 samples in order to confirm this probability.Therefore the testing of the appropriately learned networks becoming the hard task.

    3.4 Evaluation and classification of ANN Algorithmsfor multi-criteria decision making

    Number of papers apply to the VHO decision problem based on neural networks. [25]use Hopfield neural network as a powerful classification tool, which can provide optimalsolution for complex optimization problem [26] give preference to MLP as most popularsolution, which learned to make cross-layer nonlinear mapping of functions, which aredepend on link quality estimator and context metric indictors. [27] use ability of ANNto learn mapping of the QoE score to technical parameters. In paper of [9] ANNusing as a function which can give QoE score in real time, replacing end-user. It isindi�erent from input and output data, therefore, it can handle data, independentlyfrom radio access technology. In general all authors motivated by following ANNproperties: learning ability, classification and approximation. We propose to classifyreviewed papers on several categories:

    User Engagement. ANN based algorithms can make decisions taking to accountuser preferences. In some papers user has kind of interface to set up priorities onimportant metrics. Alternatively, this information collected from customer profile.These preferences can include monetary cost, throughput or the other parameters.The other case, when the system selects optimal set of system parameters based onuser profile, without user-machine interface.

    ANN Type. The most used types of the ANN are Hopfield networks and multilayerperceptron with backpropagation and its combination with fuzzy logic mechanisms.

    ANN role. A number of papers use data pre-/post- processing for input/output datafor neural networks. Independently of its type, ANN used on the di�erent algorithmstages. Pre- processing bring benefits when data should be structured, mapped tothe other variables, or transform in a manner to fit input neurons; for example tomap linguistic variables, received from FL system to ANN input neurons. ANN usedon the stage of the explicit decision-making, which is made based on ANN output

    Master thesis Ivan Goskov

    yevgeniy „For instance, if the ANN error has 1·10^−5 probability, then the validation set should contain 1 · 10^−5 samples in order to confirm this probability.“

    How can we have 1 · 10^−5 samples?

    yevgeniybringS

  • 3 Analysis of ANN-based handover decision algorithms and used criteria 21

    neurons information. Finally, it used on the data post-processing, when ANN, forinstance, using to validate results, received from fuzzy logic (FL) system and providethe feedback back to FL.

    Data Processing Type. A number of papers use data pre-/post- processing for in-put/output data for neural networks. Independently of its type, ANN used on thedi�erent algorithm stages. Pre-processing bring benefits when data should be struc-tured, mapped to the other variables, or transform in a manner to fit input neurons;for example to map linguistic variables, received from FL system to ANN input neu-rons. ANN used on the stage of the explicit decision-making, which is made based onANN output neurons information. Finally, it used on the data post-processing, whenANN, for instance, using to validate results, received from fuzzy logic (FL) system andprovide the feedback back to FL.

    The short algorithms description and comparison presented in the summary tablebelow.

    In paper [28] authors propose Hopfield ANN- based algorithm of optimal resourcesselection based on user data. The authors propose an object energy function, whichallows to optimize usage of the radio resource by mobile terminal. Consequently op-timizing weights and numerical values of the thresholds for every neuron in the net,the total network throughput increases and gradual load balance in the network canbe reached. The disadvantage of this proposal is a long network convergence time.

    Paper [29] proposes to find the optimal radio access technology (RAT) and bit ratefor all active users in the network. Authors consider a set of constraints including QoSrequirements, coverage conditions and available bandwidth for each RAT. Proposedobject function optimizes energy function of Hopfield ANN, through cost function,resources allocation matrix, and user throughput limitation. This method improvesAverage Packet Delay, Packet Dropping Rate for increasing number of users. Thedisadvantage is high computational complexity of the resources allocation matrices.

    Nasser et all [30] developed an ANN based scheme to process the HO in centralizedLTE/WiMax/WiFi HetNets. This paper applied to the multicriteria decision makingalgorithm. They divide criterions, collected by special element, called attributes col-lector, to the four types: network-related (Tra�c load, coverage area, RSS, CIR, BERetc), terminal-related (data rate, battery power etc.), user-related (as user profile, pre-ferred technology etc.), service related (monetary cost, targeted QoS, which included,for instance, the other types parameters).

    Based on extensive set of input parameters, authors conclude that ANNs are thebest alternative to handle multi- criteria HO, based on absence of the other relevant

    Master thesis Ivan Goskov

    yevgeniybringS

    yevgeniywhat is CIR?

  • 3 Analysis of ANN-based handover decision algorithms and used criteria 22

    technique, which can handle such task. The criterions were group in features in variouscombinations (about 150 datasets), which have weight, a numerical representation ofthe linguistic variable, according to the user preferences. This data used to trainMLP network which take into account tra�c load, which has the higher weight. Suchtechnique allowed to eliminate association with the overloaded cells.

    The e�ectiveness of the algorithm was evaluated based on comparison of initial vsfinal (using the proposed method) available network bandwidth. Experimentally wasproved that the approximation feature of the ANN help to make a balance betweenoverloaded cells and cells with unused bandwidth, compare to the traditional RSS-based algorithms. Nevertheless, authors do not apply to the problem of dynamicnetworks structure, like subscriber mobility and variation of the number of input pa-rameters as AP alternatives, criterions.

    In this algorithm user can select the service priorities; whereas the most prioritizedparameters receive higher values and the overall sum remains equal to 1. This pa-per take in to account only the standard MLP, whereas the motivation of this typeselection, as well as best network configuration and training were not presented.

    In paper [31] combination of FL and MLP used to solve the problem of MobilityManagement between RANs by taking into account load thresholds and coverage.Authors propose preliminary selection of HO target network before HO procedure andadaptive learning rate for training NN. The neuro-adaptive learning in ANFIS (Matlab)environment used to remember training data set. Proposed algorithm demonstrateshigher download response for the same number of mobile terminals, compare to thefixed threshold algorithms.

    [25] proposes to maximize service performance in terms service retainability, and ser-vice degradation. The proposed algorithm use Hopfield networks, where each of inputparameters can be mapped to one of the equilibrium states. Decision-making based onassumption that the given states of target network in which the ANN converges, calcu-lated by algorithm, are more suitable for the user than the serving network. Therefore,output has two values: “yes” or “no”. The data preprocessed within gateway wheregiven parameters pre-mapped to the set of input values that used in the HNN.

    [26] Improve HO performance by minimizing number of unnecessary hando�s andlost packets. Proposed MLP-based user context-aware hando� algorithm takes in toaccount link quality, based on Packet Success Rate - the parameter that evaluatesinstead of RSS. ANN used to learn correlation between user context and link qualityindicator. Proposed algorithm outperforms RSS- based algorithms for the numberof HOs, delay and number of lost packets. Fussy logic adaptive hysteresis was also

    Master thesis Ivan Goskov

    yevgeniyANN-based algorithms can’t be compared with the RSS. ANN is just an approximator. It is up to input criteria and their prioritization.

  • 3 Analysis of ANN-based handover decision algorithms and used criteria 23

    introduced to mitigate ping-pong rate.[9] selects the best access point candidate value based on real time applications data:

    RSS, Data Rate, Monetary Cost. They use MLP with BP to evaluate the each accesspoint parameters in comparison with the loaded neural network. Input parameters arepre- processed, based on Fuzzy Interference System.

    In [27] MLP trained to map QoE score to the real network technical parameters. Theexpert scoring used to fabricate training set and these data using to train ANN. Neuralnetwork validates by random set of real network data that is previously unknown tonetwork. If it gives incorrect MOS, then configuration should be change.

    3.5 ConclusionsThe idea to use di�erent ANN types for MCDM decision making, have been developedin many papers. Those articles, demonstrate di�erent approaches to the VHO deci-sion, using neural networks of di�erent types, where the most popular is multilayerperceptron. Relatively simple concept of ANN compensates of di�culty of learningprocess management and network configuration. Many authors lose sight of trainingdescription, test and validation data set collection, learning parameters specification,as well as data analysis, received after network validation and testing. Artificial Neu-ral Networks are the useful tool for classification and approximation tasks. For taskswith complicated and non-linear input- output dependencies, the best solution is usageof multilayer architecture. For tasks, required temporal dynamics – the best choiceis recurrent networks. The HO decision-making task, assumes the assignment of thecriteria set to the set of scores, given by MCDM algorithms. Therefore, the ANNdesign should have multiple neurons on the input layer and only one neuron on theoutput layer. In other words, a finite number of output parameters take in to account.Therefore Hamming and Hopfield networks [28], [29], [25], are not valid for a currenttask statement, because input data varying in a wide values range, instead of take theseveral finite realizations. Moreover, the MCDM HO has simple set of input param-eters with high requirement to the computational resources (low HO time required).The previous work and papers analysis demonstrated, that MLP is near- optimal ANNstructure, having properties, required for research task. It has high robustness due tothe distributed structure and possibilities for fast adaptation to the wide range of theapplication tasks. In this paper, we are going to use theoretical properties of multilayerfeedforward networks as universal approximators formulated by [32] "Standard multi-layer feedforward networks are capable of approximating any measurable function to

    Master thesis Ivan Goskov

  • 3 Analysis of ANN-based handover decision algorithms and used criteria 24

    any desired degree of accuracy".

    Master thesis Ivan Goskov

  • 4 Analysis of wireless networks load balancing techniques 25

    4 Analysis of wireless networks loadbalancing techniques

    4.1 Load Balancing in telecommunicationsLoad balancing (LB) is a method of jobs distribution between several network de-vices for the purpose of resources optimization, reducing the holding time of requests,support of the fault- tolerance and horizontal scaling of the computing system. [?]Implementation of the LB methods to the networks can have many positive results.It allows expanding the network possibilities and saving network operability in caseof elements failure. It makes system more fault-tolerant due to dynamic regulation ofthe used computing resources. In this chapter, we will analyze the LB principles. Theload balancing initially appeared in fixed computer networks, where load balancingbecame a special method of the tasks distribution between several network devices,like servers or network elements. Load balancing allow to optimize resource usage,to reduce maintenance time and to provide elements redundancy. However, the sameprinciples can be applied to the wireless networks.

    4.1.1 Load balancing in computer networksThe network technology is the coordinated set programs and hardware (for example,drivers, LAN adapters, cables and connectors), and data transfer mechanisms on com-munication lines, su�cient for computer networks organization. [?], [33]. The loadbalancing technology is based on the 3 and 4 levels of the OSI model – it is operat-ing the information, obtained from packet titles of these levels (TCP port number,IP sender/receiver, etc.). This method of tra�c distribution called session switching(session commutation), and it is implementing on the device called session switch.It making decisions based on two parameters – congestion of servers and the maxi-mum number of connections. Appearance of the switches functioning at the 7 levelof the OSI model, became the next step in development of the "load balancing" tech-

    Master thesis Ivan Goskov

    yevgeniy???

    yevgeniy?

    yevgeniymakes

  • 4 Analysis of wireless networks load balancing techniques 26

    nology. Nowadays, switches having information not only about port number and thereceiver/sender of a packet, but also sees packet contents. The LB reaches the highlevel of fault tolerance due to automatic search of node failures and data flow redistri-bution among remained. There can be two types of practical implementation of the LBalgorithms: hardware implementation and software implementation. Software imple-mentation is more intellectual; however, system utilities allow analyzing the situation,however decreasing hardware cost by using hubs and multiplexers.

    4.1.2 Load balancing algorithms in heterogeneous networkLoad balancing (LB), as a challenging topic for network operators which has manyapproaches, which appeared with new technologies and tra�c growth. In many casesLB and HO complementing one another. In this co-working the HO provides seamlesscontinuous service for users all over network. At once, handover is an e�ective instru-ment for tra�c load balancing. For instance, it is possible to redistribute users on theAP edge to less congested APs using granular HO parameterization. LB definition formobile networks. Following the wikipedia “network load balancing in telecommuni-cations is the ability to balance tra�c across two WAN links without using complexrouting protocols”. This capability balances network sessions like Web, email, etc.over multiple connections in order to spread out the amount of bandwidth used byeach user, thus increasing the total amount of bandwidth available. Nowadays, severaltypes of wireless technologies like 3G, IEEE 802.11 and LTE are widespread. In caseof specific implementation of LB algorithms, di�erent parameters of components of anetwork can be used. The most common is Signal level (RSS) based HO. In general,the RSS -based HO process can presented by picture [10] and can be changing bymodifying HO parameters. The advance HO can be achieved in case of overloadedcell and delay HO in case of normal loaded cell. Whereas this scheme typically worksbetter for slowly moving mobiles. It can be described by formula and illustrated onFigure 4.1:

    (RSRPt + CIOt) ≠ (RSRPs + CIOs) > Hysteresis, (4.1)

    where RSRP: Reference Signal Received Power, CIO: Cell Individual O�set, TTT:Time to Trigger, P: Preparation time

    Generic LB algorithms classification have been proposed by [34] it included

    • Relaxed optimization methods, when simulation model being limited by somefactors, such as omnidirectional antennas, non-separation of transmission and

    Master thesis Ivan Goskov

  • 4 Analysis of wireless networks load balancing techniques 27

    Figure 4.1: A typical RSS-based HO process

    receiving part, simple scheduling approaches etc.

    • Markov Decision Process is useful in uncertain conditions and bring su�cientbenefit in sequential optimization of discrete time stochastic systems. The mainidea is the search of near optimal solution, which maximizes the future expectedreward.

    • Game theory provides an interactive process of decision-making and o�ers specialmethods for vast optimization problems research. Despite the benefits of thegame theory in many cases, it cannot guaranty algorithm convergence, thereforeto provide the optimal solution, which behalf of tra�c overhead can lead toine�cient resources utilization.

    • Stochastic Geometry as a branch of the game theory, deals with stochastic op-timization assuming that involved variables are random. In this approach usersand access points modeled by Poison Point Process (PPP) and valid equations formetrics such as transmit powers, densities and bandwidths of di�erent technolo-gies can be obtained and used for load balancing optimization. The stochasticgeometry approach trying to maximize the average resources utility function forthe current network configuration.

    • Artificial Neural Networks based methods. Please refer to the list of papers inchapter

    • Fuzzy Logic allows to model high-quality aspects of expert knowledge and it’s

    Master thesis Ivan Goskov

  • 4 Analysis of wireless networks load balancing techniques 28

    reasoning concerning handover processing by means of "inference rules" and toprocess inexact and doubtful results of measured parameters of the available net-works. This approach have been used in papers [35], [36], [31] in order to evaluateparameter changes within dynamically changing conditions of radio medium. Forinstance, the majority of characteristics used for signal description are indistinctby the nature and can be described as "weak", "strong", "close", "far".

    The algorithms a�liation to each of these classes can be find in Appendix A.

    4.1.2.1 LB in WLAN

    Authors [37] separating load-distribution schemes to STA-based and network-based.The selection performing based on which part of the network managing the load-distribution process. STA-based Load Distribution STAs acting passively during thewhole process. It getting the load status of APs and on it’s basis selecting the APwhich has better parameter values among others (higher bandwidth, less frame droprate etc.) In a network-based approach, there is a sort of central element or network-side entity (AP, switch or controller) which is aware of the network load situation andmanaging the load balancing through the entire network. There are several network-based approaches can be found in literature:

    • Coverage adjustment. The overloaded APs reducing the beacon transmissionpower in order to decrease association possibility of roaming users. And in oppo-site, in case of under- loaded AP it command to increase the coverage to acceptall (new and existing) STAs. Such algorithms should provide the solid coveragefor the a�ected users.

    • Admission control is a popular technique for LB realizing possibility of the APto reject new or roaming users. The AP will accept the STAs again, when theload metric will return to accepted value of threshold again.

    • Association management techniques allowed to disassociate the STA from currentAP in order to o�oad it. It can be done by sending unsolicited disassociationframe to selected STA. In many cases the most appropriate STA for disassociationis the one which allowed the AP to become balanced.

    Hsing-Tsung et al [37] makes an analysis of current approaches to the WLAN loadbalancing. They used Iperf to measure maximum throughput (which was comparedwith AP’s real capacity) and packet loss. Specialized software called Information

    Master thesis Ivan Goskov

  • 4 Analysis of wireless networks load balancing techniques 29

    collector makes WS count and calculates throughput- related information for everyAP, using SNMP protocol.

    They solve the LB problem, using WS-load distribution algorithm. In this approachauthors join previously developed WS- based and Network- based LB algorithms. Aspecial application on the WS manage an association to the network. The process asfollowing: WS firstly connects to the AP with the best RSS. Then, using installedapplication it requests information collector to provide data about number of WSs,associated with the range of APs and their interfaces speed. Using this information,the WS makes normalization of the residual bandwidth (RB) and selects the candidatewith the best RB for association (it can be at variance with the initial AP).

    Authors experimentally demonstrated that overall network throughput can be in-creased in case of fairly distributed tra�c load among involved APs. The e�ectivenessof the proposed scheme was evaluated based on throughput- improvement ratio, in-dicating "before and after" overall network throughput. The cons of this approach isthat achieved gain could not be reached in the real networks, where WS acts hetero-geneously and low packet loss rate or high throughput can be achieved in heavily-congested APs. Authors notice that benefits, o�ered by load balancing algorithm, noteven necessarily lead to the uniformly low packet loss. They also conclude that genuineIEEE 802.11 does not guarantee an equal share of the available bandwith as well asequally low packet low rate for all WS in the network.

    [38] developed the WLAN LB algorithm that takes in to account not only the band-width, required by AP, but also the overall AP load, obtained based on the number ofWS and corresponded tra�c class (profile). This procedure applied to WS during APassociation process in centralized wireless networks. This kind of networks assumes thepresence of the central control station, which is in charge of management of multiplebase stations via such technologies as Radio over Fiber and possibly SDN. Here BSsimply converts wireless to optic signal and opposite.

    The algorithm’s idea based on AP possibility to move user (group of users) out ofcell to clear bandwidth for new user that has specific bandwidth request Breq. Authorsanalyze the e�ciency of the transferring cost for existing users. Authors define threecases of cell admission control: Fixed allocation, when bandwidth o�ered by AP isavailable for new user, i.e. N ·B Æ C≠Breq, where B is the vector of required bandwidthof APs B = (B1, B2, . . . , BM); M – number of separate tra�c classes. All subscribers inthe network defining by vector the subscriber vector N = (N1, N2, . . . , NM) . Directedretry, when CS repeat operation from 1. for all cells from the WS range, until satisfiedoption will not be find. Upward / downward candidate selection. Is the novelty

    Master thesis Ivan Goskov

    yevgeniywhat is the bw required by AP???

  • 4 Analysis of wireless networks load balancing techniques 30

    algorithm, deciding how to select existing users for moving out to the neighbor celland when it’s necessary to make it.

    The load balance evaluated based on o�ered load per cell calculated as multiplicationof connection generation rate to the connection’s bandwidth to the average connectionlifetime. The algorithm advantage is attention to various tra�c classes and correlationof bandwidth with required radio resources. Whereas, such factors as interferencebetween BS, MT-BS distance and user mobility were not take in to account.

    Based on the load balancing index and average neighbors load, [39] arranging allWLAN APs by one of three states: Fair, Will, Gull. Then applying admission controland association management approaches author specifies the set of rules for each ofthem – the first can accept new and roaming users. The second can accept only roamusers. The third declining any types of the association request from STA and applyingthe o�oading the cells via disassociation of connected STAs. The method drawbackis ping-pong intolerance.

    [40] realizes the admission control approach known as Cell Breathing. Access pointsdynamically adjust their transmission power, changing the e�ective range of AP. Algo-rithm ensuring, that during seamless HO, a�ected users do not get the lack of coverage.Algorithm triggered when frame drop rate, collected from associated STAs overpassesthe threshold and finishing in opposite conditoins. Author’s simulation results demon-strates improvement of the overall users throughput.

    4.1.2.2 LB in UMTS

    For instance, one kind of relaxed optimization methods, [41] developed algorithm thatevenly balancing overall load between access points and considering collective batterylifetime of all mobile nodes. The load balancing algorithm applied to the heterogeneousUMTS / WLAN network topology. AP/BS bandwidth used as criterions for the loadbalancing. The algorithm considers the centralized WLAN/UMTS network, controlledby vertical handover decision controller (VHDC).

    The algorithm works as following. Dependent on current MN association, the VHDCreceives link-layer triggers(LLT), indicating the state of MN connectivity in the fol-lowing cases: (i) if MN served by AP(WLAN) and MN RSS degraded to the certainthreshold. In this case VHDC looking for another AP (network) for connection han-dover, which satisfies MN requirements. If there no AP with stated pre-conditions,then it makes HO to the UMTS. (ii) if MN served by BS (UMTS) and RSS of one of theAP degraded to the certain threshold. In this case, VHDC selects WLAN as a prefer-able technology and starts the search of a relevant AP. In general, if MN is in range of

    Master thesis Ivan Goskov

    yevgeniy„The method drawback is ping-pong intolerance.“Why? What is about ping-pong for other methods?

  • 4 Analysis of wireless networks load balancing techniques 31

    AP and BSS, then the first will have the priority for an association, due to higher datarate and lower bandwidth cost (price). In order to tune up the object function theyintroduce – and — coe�cients, indicated the aggregate battery lifetime of all MNs andload balancing across all BSs/APs. Independent of the technology, the load-balancingalgorithm checks if the sum of load and total requested data rate at access point is lessor equal to maximum bandwidth o�ered by AP/BS. To do that they introduce thecost function, which is depends on price/weight of the AP/BS bandwidth, load at AP,total requested data rate and maximum bandwidth o�ered by AP. The cost functionminimization preventing highly congested AP or BS to be fully overloaded.

    Solving the problem of distribution of the overall load in the network, authors tryingto improve QoS parameters, i.e. lower delays and packet losses. That is proved bysimulation, which is demonstrates better results than traditional RSS based algorithms.Due to annual tra�c growth and corresponded battery consumption demands, theproposed technique has big advantage for VANET/MANET implementation whereas,the structure is di�cult to implement.

    [29] using Hopfield ANN as a powerfull tull for HO decision making in UMTS net-work. They finding optimal bitrate and access point for each user, dependent ofavailable bandwidth, QoS, Tra�c and Coverage. As input data they using packetdelay, packet drop ratio for packet services in UMTS network. As an output theyreceiving radio access technology with best available options. The algorithm state ofart is appliance of the ANN to the Joint Dynamic Resource Allocation (JDRA) thatdecides the most suitable bit rate and RAT for each user and Energy Function basedon Hopfield NN Salesman algorythm. This algorithm proposes significant benefit inAverage Packet Delay and Packet dropping rate, because of more deep and detaileddesign of the Energy function. The algorithm drawback is a complicated calculations,required for decision making and absence of the algorithm performance not describedin article.

    4.1.2.3 LB in LTE

    The general LB principles can be applied to the LTE-based approaches. However, inrecent years, many papers addressing Cell Range Expansion (CRE) – a popular sub-optimal technique proposed by 3GPP that implement user association bias to o�oadsubscriber to the smaller cells. It was currently proposed for LTE-A technology andassumes the following preconditions:

    • existing of macro (high transmission power) and pico cells (low transmission

    Master thesis Ivan Goskov

  • 4 Analysis of wireless networks load balancing techniques 32

    power)

    • use of the same frequency band by pico cells in the coverage area of macro

    • use of bias to improve virtually the RSS measured from pico cells

    The principle of the algorithm based on the di�erence between power levels (calledbias and applied to low-power nodes) of two cells – the source cell – the one thatneed to be o�oaded and the target cell – the one that host transferred subscribers.The users will be associated with target cell until its power (RSRP) is less or equalthen some power threshold of the source cell. This technology virtually expands thee�ective range of the smaller cells, therefore this technique called CRE. In general, thisalgorithm selects the index of best technology kú among K available candidates:

    kú = arg maxi=1,K

    BiPrx,i, (4.2)

    where Bi is the bias for tier i; Prx,i is the received power from tier i.Bias values vary from 0 to some value and can be find in empirical way. There

    recommendations in literature regarding precise numbers varies on the net work con-figuration; therefore, the suboptimal load aware performance can be achieved only withaccurate selection of the bias values. Take in to account the logic, described above; theCRE algorithm has similarity with handover algorithms. Nevertheless the di�erenceis that classical HO algorithms aiming only to transfer user from one access point toanother, whereas CRE algorithm has wider logic- it o�oading cells by transfer usersto the other access points.

    When no resource partitioning used, a CRE algorithm reduces the overall sum ca-pacity due to poor link quality of o�oaded users, within picocells. The benefit of thistechnique is an increase of the macrocell overall capacity caused by cell-edge userso�-loading to picocells.

    In recent years, many papers address to these problems.Behind simple idea there are several questions regarding system design, for instance

    what is a bias degree between two cells, what is the interference influence and how canit be negotiated, and how is the load balancing changes if the amount of small cellsdensity increase? There is a number of authors, propose di�erent techniques to solvethose problems.

    Regarding the amount of biasing, author [34] describes two cases of biasing: the co-channel and out-of-the band deployments. A good example of the first case is macro tomicro cell deployment, whereas the second is heterogeneous networks case. Despite the

    Master thesis Ivan Goskov

    yevgeniymacro ?

    yevgeniy„Nevertheless the difference is that classical HO algorithms aiming only to transfer user from one access point to another, whereas CRE algorithm has wider logic- it offloading cells by transfer users to the other access points.“isn’t is the same?

  • 4 Analysis of wireless networks load balancing techniques 33

    principle of biasing is the same for both cases, there is a big di�erence in interferencethat influence in the first case and is absent in HetNets.

    The deployment of the power levels of macro and micro cells causes imbalance be-tween uplink and downlink coverage, where handover boundary is shifted, closer to themicrocell. That causes strong uplink interference to the microcells users, which havelower power from MTs of the macrocell. The second option requires more aggressivebias profile, because only desired signal su�ers.

    Regarding co-channel biasing o�ers using such techniques as time domain partition-ing, also called almost blank subframes (ABS) in 3GPP LTE when macro BS speciallymuted for a short period and the o�oaded users from other cells transmitting in thistime slots allows to fight against co-channel interference. At the same time, it makesco-channel case use almost the same bias values as HetNets case.

    For out-of-band biasing the interference between small cells increase, whereas macrocell – micro cell is orthogonal, that makes connection to the smaller cells less desirable.The bias value decreases, when smaller cells density decreases.

    [42] developed an extension of CRE- based method, which adapts the power of theLTE-Advanced nodes to handle the load and interference situation. They propose touse logarithmic throughput metric x for Cell Selection O�set (CSO) to balance thethroughput of the users at the center and edge of the cell. This metric helps users toget higher bit rate and o�ers fair resources distribution among users. The algorithmallocates resource blocks for the users with the highest metric:

    iú(t, n) = arg maxi

    µi(t, n)x̄i(t, n)

    , (4.3)

    where µi(t, n) is the i-th user’s data rate, measured in the certain time-frame i inresource block n take into account average user’s throughput x̄i(t, n). A is a set ofactive users, associated with cell.

    The Macro cell provides information about macro cell users throughput. Dependingon the utility of the Macro and Pico cell users, the algorithm change the range of thepico cells by consequent change of the CSO which leads to the consequent load balanceof the whole network.

    The load balance was evaluated based on the comparison of the average throughputgain in macro cells and average throughput loss in pico cells, where it demonstratesappropriate results.

    [43] extends CRE algorithm with transmission power control method, which eval-uates cell throughput performance and cell-edge user throughput in LTE-A network.

    Master thesis Ivan Goskov

  • 4 Analysis of wireless networks load balancing techniques 34

    The paper aiming to solve problems of ine�cient resources utilization caused by dif-ference in amount of users in pico and macro cells. The second problem is downlink vsuplink imbalance problem, which assuming that the cell for association in DL shouldhave the highest reference signal power (RSRP), whereas the cell with the lowest passloss should be selected for uplink.

    The cell selection, based on RSRP only applies to power, received from each cell,but doesn’t consider the channel quality of the target cell. Authors extends the power,represented in 4.2 from perspective of closed power loop control algorithms and basedon the pass-loss measured in uplink. The UE selects the cell ID based on the followingformula:

    iRSRP,j = arg max0Æo

  • 4 Analysis of wireless networks load balancing techniques 35

    (LTE-A), authors [45] modeling network behavior in di�erent scenarios. They are us-ing CRE concept for cell range expansion enabled through resource three types of theadaptive resource partitioning (default signal strength-based, SINR-based selection,soft reuse scheme) as it creates a potential for tra�c load balancing. Authors demon-strated that, under greedy scheduling (soft reuse scheme) there is a half increase of themean throughput improved for edge and median user throughput, comparing to hardresource partitioning.

    There is advanced techniques was proposed Self-Optimization Definition – "Theprocess where User Equipment" (UE) and eNBs’ performance measurements are usedto auto tune the network" [10] It optimizing the configuration while taking into accountregional characteristics of radio propagation, tra�c and UEs mobility. Makes analysisof statistics and deciding what are optimal parameters.

    There are several load balancing strategies:

    • Downlink (DL) power modification, i.e. the pilot power and/or antenna tilt. Ithas the following disadvantages: indoor coverage degrades in reduced power cellsand thus requires provisioning of power amplifiers in increased power cells.

    • Handover (HO) parameter modification. Despite it overcomes the cons of DLpower modification method, the LB can only be achieved if neighbors have freeresources.

    Project Socrates [46] extends CRE algorithm for LTE-Advanced technology. It usesdi�erence in Reference Signal Receive Power (RSRP) and Received Signal StrengthIndicator (RSSI) metrics for serving and target cells, to evaluate interference andsignals, originated from serving and target NodeBs. The LB algorithm and HO decisionis performed by eNodeB.

    The algorithm based on increasing of HO o�set to the precise value in one accuratestep. It is aiming to o�oad serving cell to certain threshold by performing several con-sequent steps. Firstly, it group users, based on best MTs RSRP to the potential targeteNB (TeNB) (the criterion is di�erence between serving eNB (SeNB) and TeNB mea-sured signal quality). Then it gathers information about eNodeB available resourcesand estimates an amount of resource blocks, required for each MT after HO. Then itfilters users allowed for HO by certain HO o�set. While load at the SeNB is higherthan certain threshold, it estimates the possible load at the cell after HO. When theload at the source cell became lower than the threshold, the algorithm updates load inoverloaded cell by subtract handed over load, otherwise it checks the other candidatecell. If there are no suitable candidates then it adjusts HO o�set.

    Master thesis Ivan Goskov

  • 4 Analysis of wireless networks load balancing techniques 36

    The e�ectiveness of the LB algorithm based on load estimation method after HO,which is based on SINR prediction. The cons are: only DL, only single cell associ-ation, constant bitrate (512 kBit/s), seriously limiting the e�ectiveness of the LTEnetwork. Moreover, algorithm works until load at SeNB is higher than accepted level-the threshold should be selected more than enough to skip repeated association.

    The summary table for the revi