DOWNLINK RESOURCE BLOCKS POSITIONING AND SCHEDULING IN LTE SYSTEMS EMPLOYING ADAPTIVE FRAMEWORK S By Osama M Abusaid A DISSERTATION Submitted to Michigan State University in partial fulfillment of the requirements for the degree of Electrical Engineering - Doctor of Philosophy 2018 ABSTRACT DOWNLINK RESOURCE BLOCKS POSITIONING AND SCHEDULING IN LTE SYSTEMS EMPLOYING ADAPTIVE FRAMEWORKS By Osama M Abusaid T he expansion s in size and complexity of LTE networks is hindering their perform ance and reliability. This hindrance is manifested in deteriorating performance in the User Equipment throughput and latency as a consequence to deteriorating the E - node B downlink throughput. This l eads to the need for smart E Node Base with various capabilities adapting to the changing communication environment. The proposed work aims at developing Self Organization (SO) techniques and frameworks for LTE networks at the Resource Block (RB) scheduling management level. After reviewing the existing literature on Self Organization techniques and scheduling strategies that have been recently i mplemented in other wireless networks , we identify several critical needs that can jointly be a ddressed. The deployment of the introduced algorithm s in the communication network is expected to lead to improved and upgrade d overall network performance. The main feature of the LTE network family is the feed back that the cell receives from the users. The feedback includes the down link channel assessment based on the User Equipment (UE) measure , namely the Channel Qua lity Indicator (CQI) in the previous Transmission Time Interval (TTI). This feed back should be the main decision factor in allocating Resource Blocks (RBs) among users. The challenge is to how one could map the CQI. The Thesis advances two approaches towards that end: (i) T he allocation among the current users for the next TTI should be mapped, consistent with the his torical feed back CQI received from users over prior transmission durations. This approach also aims at offering a solution to the bottle neck capacity issue in the scheduling of LTE networks. To that end, we present an implementation of a modified Self Organizing Map (SOM) algorithm for mapping incoming data into RBs. Such an imple mentation can enable cells to become smarter. The criteria in measuring the E - node Base performance include throughput, fairness and the trade - off between these attributes. (ii) Another promising and complementary approach is to tailor Recurrent Neural Network s (RNNs) to implement optimal dynamic mappings of the Resource Blocks (RBs) in response to the history sequence of the Channel Quality Indicator CQI feedback. RNNs can successfully build their own internal state over the entire training CQI sequence and co nsequently make the prediction more viable. With this dynamic mapping technique, the prediction is likely to be more accurate to changing time - varying channel environments. Overall, the collective cell management would become more intelligent and would b e adaptable to changing environment s . Consequently, a significant performance improvement can be achieved at lower cost. Moreover, a general tunability of the scheduling system becomes possible which would incorporate a trade - off between system complexity and QoS . Copyright by OSAMA M ABUSAID 2018 v To my parents, for their devotion and sacrifice, To my siblings, for their help and encouragement, To my friends, with whom I spent so many precious moments. I dedicate my work, Symbol of gratitude and love. vi ACKNOWLEDGEMENTS I would like to express my deepest gratitude to Dr. Fathi Salem for his excellent guidance , valuable knowled ge, and precious comments and advices that helped me carry out my research. I would like to extend my sincere appreciation to the rest of my committee members, Dr. Percy Pierre, Dr. Jonathan Hall , and Dr. Nizar Lajnef, for their availability, guidance , and valuable comments. Finally, I would like to thank my family and friends for their s upport and encouragement. vii TABLE OF CONTENTS LIST OF TABLES ................................ ................................ ................................ ................................ x LIST OF FIGURES ................................ ................................ ................................ ............................ xi i LIST OF ALGORTHIMS ................................ ................................ ................................ .................. xiv KEY TO A BBREVIATIONS ................................ ................................ ................................ ............ xv Chapter 1. Introduction ................................ ................................ ................................ ......................... 1 1.1. Self - Organization in Technology: ................................ ................................ ........................ 1 1.1.1. Definition of Self - Organization ................................ ................................ ..................... 4 1.1.2.Characteristics of Self - Organization ................................ ................................ .............. 5 1.1.2.1. Scalability ................................ ................................ ................................ ............... 5 1.1.2.2. Stability ................................ ................................ ................................ ................... 6 1.1.2.3. Agility ................................ ................................ ................................ ..................... 7 1.2. Summary of Work ................................ ................................ ................................ ................ 8 1.2.1. Developing Modified Self Organizing Map technique: ................................ ................ 9 1.2.2. Implementing Modified SOM to Schedule Resource Blocks of LTE Communication Systems 10 1.2.3. Applying Recurrent Neural Network Toward Providing Future Predictions .............. 10 1.2.4. Implementing bRNN to Schedule Resource Blocks of LTE Communication System 10 Chapter 2. Implementations of Self - Organization in Cellular Networks ................................ ..... 12 2.1. Self - Organization in Wireless Communication ................................ ................................ . 12 2.1.1. Self - Configuration ................................ ................................ ................................ ....... 14 2.1.1.1. IP Address Self - Configuration ................................ ................................ .............. 15 2.1.1.2 Neighbor Cell List Self - Configuration ................................ ................................ .. 15 2.1.1.3. Radio Access Parameter Self - Configuration ................................ ........................ 15 2.1.1.3.1. Frequency Allocation Self - Configuration ................................ ...................... 16 2.1.1.3.2. Propagation parameter configurati on ................................ ............................. 16 2.1.1.3.3. Self - Confirmation Management ................................ ................................ ..... 16 2.1.1.4. Self - Organization promises for Operators Policies ................................ .............. 16 2.1.2. Self - Optimization: ................................ ................................ ................................ ....... 17 2.1.2.1. Self - Optimization for interference management: (long term) .............................. 17 2.1.2.1.1. Self Optimization through ICIC ................................ ................................ ..... 17 2.1.2.1.2. Self - Optimization through Dynamic ICIC ................................ ..................... 18 2.1.2.1.2.1 . Self - Optimization of Capacity and Coverage via Relaying .................... 18 2.1.2.2. Self - Optimization for Load Balancing: ................................ ................................ 18 2.1.2.2.1. Resource Adaptation based for Load Balance: ................................ .............. 19 viii 2.1.2.2.2. Traffic Shaping based Load Balancing : ................................ ......................... 19 2.1.2.2.3 Coverage Adaptation based Load Balancing ................................ .................. 20 2.1.2.2.3.1. Load Balancing via Antenna Adaptation: ................................ ............... 20 2.1.2.2.3.2. Load Balance through Power Adaptation (PA) ................................ ....... 20 2.1.2.2.3.3. Load Balance through Hybrid Approaches ................................ ............. 21 2.1.2.2.3.4. Relay Assisted Load Balance ................................ ................................ .. 21 2.1.3. Self - Healing ................................ ................................ ................................ ................. 21 2.1.3.1. Cell Outage Detection Scheme ................................ ................................ ............. 23 2.1.3.2. Cell Outage Compensation Scheme: ................................ ................................ .... 23 Chapter 3. LTE Resource blocks structure Types ................................ ................................ ........... 24 3.1 . Scheduling and novel scheduling algorithm: ................................ ................................ ............. 24 3.2. OFDMA and the LTE generic frame structure ................................ ................................ ... 26 3.3 . Vienna LTE - A Downlink System - Level Simulator: ................................ .......................... 30 3.4 . Packet scheduling in wireless technology: ................................ ................................ ......... 31 3.4 .1. Round Robin Scheduler (RR): ................................ ................................ ..................... 32 3.4 .2. Weighted Round Robin Scheduler (WRR): ................................ ................................ 38 3.4 .3. CQI_max Scheduler: ................................ ................................ ................................ ... 40 3.5 . Novel Scheduling Algorithm (Self Organizing Neural Network Algorithm SONN): ....... 45 3.5 .1. Introduction about SONN Algorithm: ................................ ................................ ......... 45 3.5 .2. Self Organizing Map (SOM) process: ................................ ................................ ......... 46 3.5 .2.1. How SOM is a competitive process: ................................ ................................ ..... 46 3.5 .3. The Cooperative trait in the SOM process: ................................ ................................ . 47 3.5 .4. Novel Adaptive Process ................................ ................................ ............................... 50 3.5 .5. The Self Organizing Neural Network (SONN) algorithm LTE Scheduling Section: . 53 3.5 .5.1 SO Scheduler Algorithm Framework: ................................ ................................ ... 56 3. 5 .5.2 . SONN Scheduler Algorithm Mapping: ................................ ................................ 59 3.5 .5.3. SONN Scheduler Algorithm Performance: ................................ .......................... 62 3.5 .6 . Fairness evaluations among several scheduling process: ................................ ............ 66 Chapter 4. Predictive Recurrent Neural Network techniques ................................ ........................ 71 4.1. Implementing Novel Recurrent Neural Networks Scheduler Algorithm ........................... 71 4.1.1. The adaptive algorithm: ................................ ................................ ............................... 73 4.1.2. Evaluation of the Basic Recurrent Neural Network (bRNN) prediction: .................... 75 Chapter 5. Novel Scheduler to LTE Resource Blocks by applying Recurrent Neural Network techniques ................................ ................................ ................................ ................................ .................. 82 5.1 LTE - A Resource Blocks With Recurrent Neural Network Techniques ............................. 82 5.2.1. The Scheduler Algorithm Framework By Implementing the bRNN In Scheduling: . 86 5.2.3. Mapping the bRNN scheduler by SoftMax: ................................ ................................ 88 5.2.4. bRNN Scheduler Algorithm Performance ................................ ................................ . 90 0 Chapter 6. Conclusion and Future Work ................................ ................................ .......................... 99 6.1. Conclusion: ................................ ................................ ................................ ......................... 99 6.2. Future work: ................................ ................................ ................................ ..................... 101 ix REFERENCES ................................ ................................ ................................ ................................ .. 104 x LIST OF TABLES Table 1 . Environment Criteria. ................................ ................................ ................................ ..... 32 xi LIST OF FIGURES Figure 1 . Classification of SO use in Wireless Communication. ................................ ................. 14 Figure 2. Main uses on SO for Self - Optimization. ................................ ................................ ....... 18 Figure 3 . LTE Architecture. ................................ ................................ ................................ .......... 25 Figure 4 . Resource Grid ................................ ................................ ................................ ................ 27 Figure 5 . 2 - D Time frequency grid ................................ ................................ ............................... 28 Figure 6 . Orthogonal Frequency Multiple Access OFDMA. ................................ ....................... 29 Figure 7 . 2 - DL Channel Mapping ................................ ................................ ................................ 30 Figure 8 . DL Channel Mapping ................................ ................................ ................................ .... 33 Figure 9 . Round Robin Schedule flow chart ................................ ................................ ................. 33 Figure 10 . E - node B throughput PedB Channel. ................................ ................................ ......... 34 Figure 11 . Block Error Rate PedB Channel ................................ ................................ .................. 35 Figure 12 . E - node B throughput flat Rayleigh channel ................................ ............................... 35 Figure 13 . E - node B Block Error Rate Flat Rayleigh channel ................................ ..................... 36 Figure 14 . E - node B throughput PedB channel ................................ ................................ ............ 36 Figure 15 . E - node B Block Error Rate flat Rayleigh channel ................................ ...................... 37 Figure 16 . E - node B throughput flat Rayleigh channel ................................ ................................ 37 Figure 17 . E - node B Block Error Rate flat Rayleigh channel ................................ ...................... 38 Figure 18 . Weighted Round Robin Scheduling ................................ ................................ ............ 39 Figure 19 . Best CQI scheduler flow chart ................................ ................................ ................... 41 Figure 20 . CQI Max Scheduler ................................ ................................ ................................ ..... 41 Figure 21 . Block Error Rate for E - Node B ................................ ................................ ................... 42 Figure 22 . Throughput for the Cell ................................ ................................ ............................... 42 xii Figure 23 . Block Error Rate User 3 ................................ ................................ ............................. 43 Figure 24 . Throughput User 3 ................................ ................................ ................................ ....... 43 Figure 25 . BL Error Rate User 10 ................................ ................................ ................................ . 44 Figure 26 . Throughput User 10 ................................ ................................ ................................ .... 44 Figure 27 . Gaussian Neighborhood function ................................ ................................ ............... 48 Figure 28 . Block Error Rate for the Cell ................................ ................................ ...................... 62 Figure 29 . Throughput for the Cell ................................ ................................ .............................. 63 Figure 30 . Block Error Rate for User 3 ................................ ................................ ........................ 63 Figure 31 . Throughput for User 3 ................................ ................................ ................................ 64 Figure 32 . Block Error Rate for User 10 ................................ ................................ ...................... 64 Figure 33 . Throughput for User 10 ................................ ................................ .............................. 65 Figure 34 . Cell throughput for the three different scheduler schemes. ................................ ......... 67 Figure 35 . Normalized cell throughput for the three different scheduler schemes. ...................... 68 Figure 36 . Fairness among users for the three different scheduler schemes. ............................... 68 Figure 37 . Combination (Fairness and Throughput) evaluation among different scheduler schemes. ................................ ................................ ................................ ................................ ........ 69 Figure 38 . Recurrent Neural Network RNN unfolding over time index ................................ ...... 75 Figure 39 . Simulation depicting an example results of the training process of bRNN ................ 76 Figure 40 . An Example output error signal profile during training ................................ .............. 77 Figure 41 . Example Output Error profile ................................ ................................ ...................... 77 Figure 42 . Dynamic system output and target values ................................ ................................ ... 78 Figure 43 . Example State Error ................................ ................................ ................................ .... 78 Figure 44 . Example State performance ................................ ................................ ......................... 79 Figure 45 . Output performance ................................ ................................ ................................ ..... 80 xiii Figure 46 . Predicted values vs Real values ................................ ................................ ................... 81 Figure 47 . Block Error Rate for the Cell using bRNN ................................ ................................ . 90 Figure 48 . Throughput for the cell using bRNN ................................ ................................ ........... 91 Figure 49 . Block Error Rate for User 3. bRNN ................................ ................................ ............ 92 Figure 50 . Throughput for User 3 bRNN ................................ ................................ ...................... 92 Figure 51 . Block Error Rate for User 10 bRNN ................................ ................................ ........... 93 Figure 52 . Throughput for User 10 bRNN ................................ ................................ ................... 93 Figure 53 . Cell throughput for the four different scheduler schemes. ................................ .......... 95 Figure 54 . Normalized cell throughput for the four different scheduler schemes. ....................... 96 Figure 55 . Fairness among users for the four different scheduler schemes ................................ .. 96 Figure 56 . Combination (Fairness and Throughput) evaluation among different scheduler schemes. ................................ ................................ ................................ ................................ ........ 97 Figure 57 . (Fairness and Throughput) evaluation of bRNN scheduler schemes comparison ...... 98 xiv LIST OF ALGORTHIMS Algorithm 1. Energy Based SOM Algorithm ................................ ................................ ............... 56 Algorithm 2. Self Organizing Scheduler ................................ ................................ ...................... 59 Algorithm 3. SO Scheduler Algorithm Mapping ................................ ................................ .......... 61 Algorithm 4. bRNN updating Algorithm ................................ ................................ ...................... 85 Algorit hm 5. Self Organizing Scheduler ................................ ................................ ...................... 87 Algorithm 6. bRNN Scheduler Algorithm Mapping ................................ ................................ .... 89 xv KEY TO ABBREVIATIONS AMC Adaptive Modulation and Coding BLER Block Error Rate BET Blind Equal Throughput CQI Channel Quality Indicator DCI Downlink Control Information eNB evolved Node B FDPS Frequency Domain Packet Scheduler FLS Frame Level Scheduler GBR Guaranteed bit - rate HARQ Hybrid Automatic Retransmission Request LTE Long Term Evolution LWDF Modified LWDF MCS Modulation and Coding Scheme MT Maximum Throughput OFDM Orthogonal Freq. Division Multiplexing OFDMA Orthogonal Freq. Division Multiple Access PDCCH Physical Downlink Control Channel PDSCH Physical Downlink Shared Channel PUSCH Physical Uplink Shared Channel PF Proportional Fair PLR Packet Loss Rate xvi PSS Priority Set Scheduler QCI QoS Class Identifier QoS Quality of Service RB Resource Block RLC Radio Link Control RR Round Robin RRM Radio Resource Management SC - FDMA Single Carrier Freq. Division Multiple Access SGW Serving Gateway TDPS Time Domain Packet Scheduler TTA Throughput To Average TTI Transmission Time Interval UE User Equipment VPM VoIP priority mode 1 Chapter 1 . Introduction 1.1. Self - Organization in Technology: In the last four years , LTE networks have undergone massive changes, updates and developments, because they are the best candidate to be the backbone network for 5G and beyond. The User behavior of the wireless system has had huge changes; this is supported by the observable growth of bandwidth demanding applications such as video streaming and multimedia file sharing. These developments have applied pressure on communication systems to increase capacity , Quality of Service (QoS), energy efficiency and the mo st significant aspects of these cellphone usage [43] . While, it is typical for any wireless communication system to be pressured by increased capacity demand, now point of view, because with higher capacity and QoS comes higher capital expenditure (capacity cost) and operating expenditure (operating cost). Since users may be reluctant to pay proportionally higher bills for improved services, minimizing Capacity Cost and Oper ating Cost will render the business model commercially viable. This will be a challenge in each new release of wireless networks as well as in the daily basic functionality. Thus, the tradeoff between providing improved services and retaining reasonable pr ofits is the crucial consideration with which operators struggle. Such challenges are among the main motivations for researchers seeking to bring autonomous intelligent adaptation techniques to the field. Mostly, these techniques are labeled Self Organizat ion (SO), and the main motivating factors are: 2 - Optimal Capacity: the physical upper bound on the inherently unpredictable nature of spatio temporal dynamics process associated with wireless cellular systems. One seeks the optimal performance in terms of ca pacity and QoS which is not achievable with the current fixed legacy designs. In wireless communication systems the user mobility and the channel model variability are naturally the main reasons for the communication system suffering from resource under - ut ilization, expressed in low resource efficiency or over utilization, which results in low QoS as well as poor capacity [31] - Small Cells: All new creative tools like Femto - Cell, Small Cells, out - door relays, In - door ireless communication systems are using as the main tools toward improving the capacity and the QoS of indoors and special cases places. All of these updated (e - node Bs) lead to systems with a several nodes which will cause a lot of interference resulting in degradation for the neighboring macro cell (main E - Node B) if there are no Self Organizing techniques attached with all these femto/relays devices. - Periodic Manual optimization is required and it should be in a classic approach, this is because the incr eased complexity of systems will lead to greater human errors which will result in longer recovery and restoration times, all this will have an effect especially with the huge scale of the wireless systems. - Toward improving the operating expenses performa nce significantly as it eliminates the need for expensive skilled labor required for configuration, commissioning, optimization, maintenance, troubleshooting and recovery of the system. For these issues and more it is clear Self Organizing SO is not only a feature toward future LTE networks like LTE - A and LTE - U, it will eventually become mandatory because of the scale of these networks and their standers. 3 Even though Self - Organizing is new in wireless communication networks technology, you can find some in itial implementation of Self Organizing strategies in several types of wireless networks like Sensor networks, Ad hoc networks,... ,etc. As described in [9], [10] and [11], autonomic computer networks have been implementing the Self Organizing (SO). Making it rapidly growing area. Since this last decade, some researchers intend to focus on this domain. Below is a review of relevant literature and technical reports on the topic and contribution which provides a comprehensive description of the hierarchy and the development of this domain. Here, I provide a consolidated review of recent developments on self - organization in communication systems on general characteristics solutions and the methodologies which are related to designing self - organization in cell ular networks. The Aim of this literature review is to identify potential methodologies and the open research issues for designing Self Organizing SO in future systems. In this process, we also discuss a set of important features that make an algorithm or system self - organizing. In this chapter, we provide a firm definition and understanding of the term Self Organization and its specialty. We start with definitions used in other disciplines, then we deduce the adaptive, autonomous and learning methods whic h are keys in defining such systems. Furthermore, we elaborate on stability, scalability and agility as characteristics that are desired most in any Self Organized systems. As such, a system exhibiting any of these characteristics is considered as having some form of intelligence. We provide a better understanding on the difference and similarities between adaptive systems, autonomous systems, cognitive networks and self - organized networks [117] [119] . 4 1.1.1. Definition of Self - Organization Even Self - Organization is a technological concept named by many different of implementations. But, the main Inspiration of Self Organization is from Artificial Neural Network ANN, as the concept is one of the main topics in the field, where certain biologi cal systems exhibit unique organized behavior in order to achieve a desired objective. This was with autonomously and intelligently adopting to dynamics of their immediate environment. So, Self - Organizing is result of unsupervised learning on Artificial Ne ural Network ANN [120] [121] [123] . Biological unsupervised Learning: Generally, in biological type networks there are not a lot of different learning methodologies for unsupervised learning and supervised learning. Most connections between visual corti cal areas are two - way bottom up from the retina and top - down from later areas, where the final destination of signals are the motor areas. Self - Organization (SO) is considered as unsupervised learning by computer science and cybernetics [12]. Talking about SO in communication systems, in [13] they defined SO as Cognitive Dynamic systems of future cellular systems. An intelligent system such as this would learn from the environment and adapt to statistical variations in input stimuli toward achieving highly reliable communications whenever and wherever needed. This means Self Organizing (SO) is an adaptive functionality. So, the network can detect changes and based on these changes make intelligent decisions to minimize or maximize the effect of those changes as [14] defined. These explanations are understood from the biological beliefs such: The effects of the unsupervised learning like SO biases development. So, the features derived and the organization of the earlier layers is more effective for task perfor mance. Global convergence of patterns appears at the motor level (base layer), while sub - soft - convergence representations can be useful for more than one category (multipurpose). The 5 final represen tation (features) and organization are selective or motor - biased, which does not have to be task - specific because, task - discards as much information as possible that is not required for the current task. The following chapters we will explain more of the benefits of non - task - specific learning [110] [112] [125] . Proffered features of SO for wireless networks are the primitive concept of Artificial Neural Network ANN like adaptability, dynamic and e mergent behavior. These are the key attributes associated with SO that raise it above simple adaptability, as inferred from the definitions discussed above. 1.1.2. Characteristics of Self - Organization Research on previous wireless network application of intelligent dimensioning, planning, operating and supervised can be summarized/classified by the following: 1.1.2.1. Scalability Scalability is one of the most important restrictions when SO is implemented for engineering problems. Especially, in all of t he wireless network challenges and tasks. The scale size of the system should be limited. So, before running any kind of Self Organization toward any kind of general problems it should remain in local and simple behavior, briefly the system remains operati onal under SO if a reasonable number of entities leave or enter the system. The main factors that should be supervised are Minimal Complexity and Local Cooperation. It is clear that the algorithms should be conservative in terms of time, space and any othe r resources that have been used as an input to the algorithm from the system which means less complexity. The second consideration is algorithms should not require global cooperation or signaling, rather local 6 coordination should be relied upon where possi ble [108] [109] . This moves us toward reducing any overheads because if cooperation among all nodes is required for implementation of an algorithm its overheads will increase as the number of nodes increases in the system. Preventing its scalability. As the updating process looking to local minimum increase to match the prognoses [106] [107] . In sum. All this as in the theory of SO, Scalability can be perfected with minimal complexity and, local control. Which is highly recommended for satisfying the need f or Stability and Agility. 1.1.2.2. Stability This factor is always the main parameter in any Engineering Systems, and any algorithm added to the system should consider this factor. As the Self Organizing (SO) algorithm transmits from state to another wit hin a certain time internationally, the transient time should be finite and feasible. It is not allowed to the algorithm to be oscillating for long time without any converging to stable state. Bounded time really needed for oscillating to maintain the stab ility. Stability is an important topic for researchers of Self Organizing algorithms, and the main restriction that want to be granted is Robustness. Robustness is the safety belt of stability, robustness means that when the system is in a certain state a nd it is facing a such even cause instability or leads to instability the system needs to be able to return back to the previous state to be stable, and this betwe en states during finite time. Stability means the system needs to be elastic, self - healing and not centralized control . 7 1.1.2.3. Agility The agility factor is one of the core factors in Self Organizing algorithm. Because SO should applies the adaptation to the system each time it has been called, adaptation should include supple or onal environment . So, in order to be self - organized, algorithms should not only have the capability to adapt and cope with its changing environment (stability), it should also not be lazy in its adaptation (agility) . Meanwhile the system should not be effe cted by any minor events (neglected updates) it faces during the transient period between one state and another as it has been described in the previous factor. This means there should always be a threshold value for the SO algorithm to decide if it will n eglect this even or adapt it. The feedback that the system has to translate as an action by the Self Organizing algorithm plays the main role in all action, its timing is important to it agility and its delay as well [100] [ 105] . Therefore, scalability, st ability and agility are the main factors that need to be monitored while Self Organization algorithms are running toward converging. Self - Organizing (SO) algorithms could be presented in different networks in wireless communication systems, these represent ations are Adaptive Networks, Autonomous Networks and Cognitive Networks. Based on the literature review that has been done in previous work with wireless communication, the following is a discussion of the three types of networks: 1 - Adaptive Networks: This kind of technique depend mainly on the feedback that the system is providing to the control section, so the control section would check the feedback readings and configure it with the closest class, then the order will result in indirect response from the control section to the system [85] [86] [102] [103] . 8 2 - Autonomous Networks: It is completely the same as the Adaptive Networks with no human interaction or any other external interaction with the algorithm. It is clearly a Self Monitored system. 3 - Cognitive Networks: as [16] explains these types of networks, it is clear this is the highest stander level of the networks that exist now. It is the same as the Autonomous Networks with more facility reaching the environment, reading from it directly, a nd then applying the adaptation process. Therefore, this network type is capable of planning, observing and executing by itself. The key part of cognitive networks is their interaction with the operating environment and their ability to learn from the proc ess. 1.2. Summary of Work After comprehensive review of LTE communication systems spectrum and its structure, we describe our main challenge as Resource Blocks allocation in the Resource Elements grid. The grid has been structured based on the standards rel eased by FCC and 3GPP. The main new feature in LTE communication systems is the ability to provide feedback from the end user (User equipment) to the provider (E node B); this feedback includes indications about the channel quality and status between an en d user and base - station (E node B) during the last transmission time. Such feedback features are opening the scope toward enabling smart networks that adapt the Resource Elements grid map to improve performance at each transmission period. A specific feed back feature, known in the communication industry Channel Quality Indicators (CQI), can be exploited carefully with some considerations. One needs to consider a certain level of e keeps sending 9 packets to only UEs that have high quality channel status, other UEs with less quality channels will become neglected for a relatively long time. Thus, it is judicious to infuse throughput rocess. In this work, we introduce adaptive intelligent techniques that update the Resource Elements grid, taken in considerations the statistical mean channel quality (CQI) of the linked users over intervals above certain threshold values [54] [57] [61] . The main work here is to build smart scheduler at the (E node B) level in the downlink and this needs pre - scheduling and post - scheduling procedures such as mapping the users in the grid. These procedures have been developed in details in the next chapte rs [50] [51] [52] [53] . 1.2.1. Developing Modified Self Organizing Map technique: The Modified Self Organizing Map technique adds the non - linearity toward fast adaptation. This is one of the main contributions in this work. We are using a modified form of updating the weights. In this novel algorithm, the updated weights are energy based and the updated function includes non - linearity. This helps our case, as we want to do mapping with clustering in one direction. (1.1) This modified updating is smoothing the values of the weights at each transmit time, and the benefit is accelerating the convergence of the weights, which are visible for our case here. This has been covered in chapter 3 in details and proofs. 10 1.2.2. Impl ementing Modified SOM to Schedule Resource Blocks of LTE Communication Systems The way we are applying the Self Organizing Map is the weights are carrying the Channel Quality Indicators updates of the channels. This means number of weights are equal to the number of users. The weights are updated at every transmitting period. Therefore, we are looking to the weights to be converged very fast and passing the warming era soon. At each transmitting period we use equation (1.1) to carry the update of the channe l performance. This is happening all time long once at each transmitting period and the mapping decision are happening at each transmitting period as well. This has been covered in details w ith performance evaluations at Chapter3. 1.2.3. Applying Recurrent Neur al Network Toward Providing Future Predictions The base Recurrent Neural Network (bRNN) that has been introduced in [38] exhibits stable behavior and uses training capability that enable prediction. We implemented as system that employs bRNN that enables periods. Such prediction enables the system to schedule the Resource Blocks to flexibly optimize performance . More details and evaluation of the modification are provided in Chapter 4 . 1.2.4. Implementing bRNN to Schedule Resource Blocks of LTE Communication System We build our own base Recurrent Neural Network that introduced at [38], then we modified it by adding the prediction calculation state. This prediction provides us with output at each iteration as well as error management and performance evaluation. All this are happening with updating our code with the regular way that has been introduced at [38]. Here we are applying our input training 11 data as the vector of the Channel Qualit y Indicators, we are applying this input data at each transmitting period and we update the weights matrices as well as providing future perdition of the next transmitting time which considered as feed for mapping the Resource Element grid. The performanc e evaluation that has been explained in chapter 5 shows how high throughput of the e node B downlink as well as how smooth the individual users throughput are. 12 Chapter 2. Implementations of Self - Organization in Cellular Networks In this chapter, we discuss my survey from the previous publications on the Self Organizing Algorithm in Wireless network and communication systems. The previous research work has been classified, some classes are already highlighted in Chapter 1. Specif ic classifications are introduced here in Chapter2. We categorize the previous research on Self Organization in three parts corresponding to the phases: Self Configuration, Self - Optimization and Self - Healing [98] [99] . Each of these Phases has two schemes of Self Organization (SO) on them. We used a Framework to characterize the use of Self Organizing (SO) in this report as a general barometer to assess the degree of SO in the proposed solutions where applicable. 2.1. Self - Organization in Wireless Communi cation Based on 3GPP [17] and NGMN [18] We reached to a brief list of cases where Self Organization algorithms applied to: 1 - Inter - Cell Interference coordination 2 - Interference reduction 3 - Energy saving 4 - Automated configuration of physical cell identity 5 - Coverage and capacity optimization 6 - Mobility robustness optimization 7 - Mobility load balancing optimization 8 - Random access channel (RACH) optimization 9 - Automatic neighbor relation function 13 We categorize the use of Self Organizing (SO) in these nine classes. Howeve r, some previous publications tried to classify any kind of work under one of the four main system objectives i.e. coverage expansion, capacity optimization, QoS optimization and Energy efficiency [19] [96] [97]. lternatively, Self - Organization use could be classified either to be an online control solution or offline control solution, usually in wireless networking they applied Self Organizing into online solutions, which is more accurate and flexible on adaptation as [23]. In terms of classification, the re is an - other point of view to be considered: depending on the challenge that brought Self Organization to the table they classify the SO use, as in previous publications [20], [21] and [24] they divided the SO into three main categories of classification : time case, space case and phase case. It is clear on time scale based classification the main factor is the operating time of the SO algorithm. From the literature, we observed the adaptive modulation and coding scale are in the same class, and the power control load balance is in other class. Therefore, Phase Based Classification has three main phases deployment, redeployment and maintenance. There are some official classifications for these phase classes on Artificial Neural Network publications; they are self configuration, self - optimization and self - healing. Each phase of these main phases can be highlighted into different paths as figure (1) shows, 14 Figure 1 Classification of SO use in Wireless Communication. 2.1.1. Self - Configuration Configuration became a mandatory operation in wireless communication system. It is needed for eNode Bs (eNBs), femto cell, small sell and relays. It is done through deployment, extension, and upgrade of any terminal. Configuration is important in the test and drops on the ser vices. As LTE and LTE - A are on massive scale in terms of terminals. Therefore, Self - Configuration has to be attached with these networks for comfortability and accuracy. From past productions of models we have examined, the Self Configuration principle in tention is: the point at which a disappointment happens into a specific terminal this terminal ought to 15 have the capacity to come back to running mode with no human inclusion. This procedure and the executions of Self - Configuration can be abridged into the se following features: 2.1.1.1. IP Address Self - Configuration Self - Configuration has been implemented in wireless technology, and it has been implemented on computer networks in IP - addresses like into Dynamic Lost Configuration protocol or in Bootstrap protocol. There are several flow charts for this process, as in [25] there is great use for Self - Configuration in e Node B. 2.1.1.2 . Neighbor Cell List Self - Configuration One of the most promising use for Self - Configuration in LTE and LTE - A is to be used in making a list of ID for the surrounded cells with the e - Node B, where the cell ID list should be implemented and updated frequently [94] [95] . This is done by generating a neighbor cell list and updating it using a centralized as well as decentralized approaches. The criteria for the selection of neighbors or the initial generation of the neighbor cell list can be based on the geographical coordinates of the cell sites . 2.1.1.3. Radio Access Parameter Self - Configuration This is one of the promising field for Self - Organizing SO in wireless communication systems, because its effect will be observable and will affect the network in terms of o perating and r esources values. Based in previous publications we have scanned; we can summa rize it into: 16 2.1.1.3.1. Frequency Allocation Self - Configuration It deals with the MAC layer frequency channel for Peco, Macro, Micro, and Femtocell [25]. From another viewpoint, there are some algorithms that start to be presented, which reuse the reso urces (Bandwidth). Like in [26]. This kind of work does not have an expandable future. 2.1.1.3.2. Propagation parameter configuration updates so the cell will not affect the neighbors or the neighbors will not affect the cell. Dynamic Radio Configuration Function (DRCF). 2.1.1.3.3. Self - Confirmation Management This work depends on Resource Information Base RIB and Policy Information Base PIB based in data analysis. 2.1.1.4. Self - Organization promises for Operators Policies The Operators policy are the rules that have been agreed on by governmental organizations like FCC or 3GPP in their releases. These rules are clear and they could be classified into coverage extension, capacity optimization, energy efficiency or fairness among users. The rules could be rules affect all of these classes at one release [92] [93] . There are some proposals in the publication for some regulations to configure managem ent for the operators. 17 These configuration management mechanisms could rely on data like resource information base or Policy information based. 2.1.2. Self - Optimization: As in any running systems, optimization is a huge goal to implement by Artificial Neural Network algorithms. There are many publications releases toward using Artificial Neural Network into LTE networks and LTE - A and there are promises to use Self Organiza tion (SO) specifically. Some publications introduced Self Optimization for Load Balancing, Self - Optimization for Capacity and coverage and Self Optimization for Interference Control. The Self Optimization for Interference Control is promising solution to L TE and LTE - A networks challenges, like for the main challenge for the capacity in these networks are Inter Cell Inter Carrier (ICIC) and it is making a huge effect in the coverage as well. 2.1.2.1. Self - Optimization for interference management: (long term) In such types of networks, frequency reuse is small and most likely one, this really effects the QoS through the interference that is happening [91] . This tighter frequency reuse needs a lot of co - channel optimization to cleave the trade off and this really needs to stay in the long term, here are some examples of this: 2.1.2.1.1. Self Optimization through ICIC A - Self Optimization via Integer Frequency Reuse B - Self Optimization via Fractional Frequency Reuse 18 2.1.2.1.2. Self - Optimization through Dyn amic ICIC 2.1.2.1.2.1 . Self - Optimization of Capacity and Coverage via Relaying Self - Organization using Self Optimization has been summarized into figure (2) as the previous publication stated Figure 2 Main uses on SO for Self - Optimization . 2.1.2.2. Self - Optimization for Load Balancing: Promptly after the coming of marketed cell correspondence frameworks, the arrival of common spatio - transiently differ client circulations due to the requirement for the heap balance components [87] [89] [90] . From that point forward, numerous productions seemed to receive the Self Organization (SO). In any case, this past work with remote correspondence is excessively particular, making it impossible to these distinctive sorts of Networks and there is no such calculation that has been actualized to LTE or LTE - A systems. This is the reason this work we 19 are doing here, this is the reason our work is huge, putting forth a concentrated effort Organizing (SO) to Self - Optimization the heap balance; it is possible th at it will be in the Physical Layer or into MAC Layer. Before we experience our work. Quickly, a thought of past examinations is incorporated here before a dialog of our investigation: 2.1.2.2.1. Resource Adaptation based for Load Balance: This Research is important for most types of wireless networks. Excluding OFDMA such LTE and LTE - A. Usually such networks that use one frequency for all cells and because of the Inter Cell Inter Carrier ICIC will be high in such scenario, especially if the cell borrows so me channels from the neighboring cell. There is a massive work in the WCDMA network on building algorithms of virtual channels. Therefore, if the cell was busy and the neighboring cell was not in the full load, the neighboring cell could borrow some channe ls and use them virtually. Again, this is not currently visible. 2.1.2.2.2. Traffic Shaping based Load Balancing : This theme is not that appealing so far to LTE and LTE - Advance, Ordinarily Movement Forming is extraordinary with tolerating another associat ion or in the hand - over. In LTE and LTE - An as they are utilizing one bearer recurrence there are no soft Hand - Over or considerably Gentler Hand - Over, if there are hand over happening it is in every case Hard Hand - Over. In systems like WCDMA, this is extrem ely encouraging in light of the fact that when another association is set up the system will need to be keen with snaring it inside the correct cell from the earliest starting point. This cell ought to be free and give a decent inclusion to the area, a sim ilar thing with Delicate Hand Over and Milder Hand Over. This procedure should be done scientifically to give great Load Equalization. 20 2.1.2.2.3 . Coverage Adaptation based Load Balancing This type of load balance is done by a certain mechanism of change that effects coverage: 2.1.2.2.3.1. Load Balancing via Antenna Adaptation: Most of the publications in this topic are talking about how the tilting of the antenna can be changed for favor of improving the coverage and no user will be terminated because of coverage. In Networks like WCDMA, this is not a big deal as there are alway s soft - hand over with subscriber. In other wireless communication systems this is still big issue, there are many creative papers talking about how they could oscillate the antennas to improve the coverage or cover certain spots. 2.1.2.2.3.2. Load Balanc e through Power Adaptation (PA) Transmission power can act naturally self organized. For the most part here we don't discuss Agility or Scalability. A large portion of the discussion is tied in with controlling the Pilot Signal regarding its quality. This sort of control should be possible in the e - Node B itself or in more elevated amount stages like MS terminal. The parameters that will be controlled are totally not the same as the Soft Hand over parameters; here we change the inclusion quality to show sig ns of improvement Load Balance execution. These sorts of plans generally are done Online despite the fact that there are a few distributions like [27] talking about Self Organizing the Power off - line. 21 2.1.2.2.3.3. Load Balance through Hybrid Approaches From the title, it is clear it is a mix of Soft Hand - Over and Power control, some publications as in [28] start to use both techniques at the same time to improve Load Balance on Traffic Shaping. Because, of central control on this lack of scalability appe ars, the use of traffic estimation maps that have been done by operator systems it will be dynamic. It is a more efficient offline design methodology, more useful during the deployment phase, than an online LB mechanism implementable in the operational pha se. There are similar shy work with WCDMA networks. 2.1.2.2.3.4. Relay Assisted Load Balance From some studies like [29] there is a possibility of controlling the Load Balance through the relays by these tools: There is intensive research on the previous three points into WCDMA networks as well as the Ad - Hoc networks in terms of Load Balance and Load Sharing. Such a work is usually done by a central control unit which is still required to receive, process, and fe edback the dynamically changing system wide utility to and from all UE and relay nodes in the network. This may have an adverse effect on the agility of the solution in a practical system because of the delays incurring from large amounts of data processing and its relaying to and from a central unit. 2.1.3. Self - Healing Remote correspondence frameworks like some other designing frameworks can possibly bomb every once in a while. This happens on account of outside impacts like catastrophic eve nts or mishaps or inward deceptive of the framework. As of now, if there are any sort of 22 disappointment answered to the administrator, RF - Engineers and experts will go to the area of disappointment and run the investigating at that point run their techniqu e to illuminate the issue and restore the framework to the standard dynamic mode. It is clear this disappointment could happen to some degree frequently and this requires some investment and a ton of endeavors to restore the framework to customary mode. Th e presence of the Artificial Intelligent Self - Healing model turns into a promising subject in this title. Accordingly, there are many promising stream graphs and introductions of the Self - Healing model beginning to show up in meetings and productions. 3GPP in its discharge discussed Self - Healing and attempted to guarantee sorting out the procedures that will be utilized at that point. The general methodology towards Self - Healing is recommended, that comprises of essential components of observing, determinat ion and pay. Learning and adjustment is likewise certainly part of this methodology as off base determination dependent on wrong relationship of alerts can be logged for a more savvy conclusion in future. The primary periods of Self - Healing are: Monitoring , Diagnosis and Compensation. One of the general Framework of Self - Healing which is the regular situation can be condensed here: In the ordinary dynamic mode, the framework is checking the system for any peculiarity or if any predefined conditions for Se lf - Healing conditions are fulfilled. Along these lines, when such conditions are met, information are broke down utilizing the Self Organizing calculation or by any framework master to determine the sort of deficiencies (with a specific likelihood of preci sion) and afterward the neighboring cell for the fizzled cell ought to be enlightening toward covering loss of this phone (redress), Hopefully this remuneration will be in full for the flawed hub. The neighboring hubs likewise occasionally tune in to signa ls from the broken hub by means of the X2 interface to set up if the hub has been reestablished. At the point when the 23 neighboring hubs can come back to their pre - remuneration mode then as not to debase by and large framework execution by impedance this mu st be precise on timing. A critical element to be watched is the circle that persistently screens the defective cell to decide whether it has been reestablished to typical activity. Despite the fact that there are many promising works in this field, it wi ll be moderate on the grounds that such work in Self - Healing should be finished with existing frameworks. This work should be done under supervision of the Operators which are now running with gigantic measure of directions by various Organizations. The pr imary Detection and Compensation plans found in productions are: 2.1.3.1. Cell Outage Detection Scheme The cell is considered to have an outage when its performance (Coverage or Capacity) is below a threshold values, which is most likely specified in the standards of the technology. Usually there are types of Outage s in public references. 2.1.3.2. Cell Outage Compensation Scheme: It is always dependent on the faults that has been detected. Therefore, before any action of compensation, the fault has to b e clarified. This clarification could be done automatically or manually because it could need a visit to the E - node B location. [29] presents a cell outage management description for LTE systems. Both the detection and the compensation schemes highlighting the role present a cell outage management description for LTE systems . 24 Chapter 3. LTE Resource blocks structure Types This work is not only having potential toward an upgrade for the 4G+ type of wireless communication systems. This work is a development to the innovative technology because it is applicable to the 5G wireless communication systems. Radio Access Network RAN sharing for 5G wireless communication system can be implemented by adding advanced control features to the current 3GPP LTE which will be the dynamic scheduler. LTE uses Orthogonal Frequency Division Multiplexing (OFDM) for the downlink and Single Carrier Frequenc y Division Multiple Access (SC - FDMA) in the uplink [44] [60] [62] [63] . The Physical Resource Block (PRB) is the smallest element assigned by the base station scheduler [84] . Transmission Time Interval (TTI) is the duration of a transmission on the radio l ink, which is exactly the same structure we are talking about here in our work. A scheduler can determine to which user the shared resources (time and frequencies) for each TTI should be allocated. The RAN sharing problem is related to the design and imple mentation of policies that are able to effectively schedule Resource Blocks effectively between different MVNOs with respect to specific differentiation objectives and with isolation guarantees. As currently, the only available Matlab simulation toolbox is for LTE - A 4G+ communication. We applied our algorithm to this type of simulation with capability of applying it to 5G simulation in the future 3.1 . S cheduling and novel scheduling algorithm : As LTE and LTE - A dvance are the latest version s of mobile communication networks proposed to the customer s to use through several organization s like 3GPP and FDD. T hey are 25 using the latest technology in the field , especially Frequency Division Multiplexing Access ( FDMA ) for downlink , Orthogonal Frequen cy Division Multiplexing Access ( OFD MA ) and Single Carrier F requency D ivision M ultiplexing A ccess ( SC - FDMA ) for up - link with using some other technology like MIMO as well [48] [49] . Our work will be in the downlink o f the network as the bottleneck challenge is the capacity (throughout and latency) in down link from th e Evolved Base Station (E Node B ase ) : Figure 3 LTE Architecture . There are two frame structure s in LTE slandered : Type 1 uses Frequency Division Duplexing (uplink and downlink separated by frequency) [45] [70] [74] , and TDD uses Time Division Duplexing (uplink and downlink separated in time). Both of these frames are used a t LTE networks at the same time, i n order to adequately explain OFDMA within the context of the LTE, we also m ust study the physical layer generic frame structure of LTE networks [55] [56] [59] [64] . 26 Therefore , just currently 3GPP released the main structure of LTE - A and its specifications. It is clear OFDMA provide a lot to the scheme , for example: time and frequency diversity, good resistance to inter - symbol interference, better deployment flexibility. Besides, OFDMA allows assigning subsets of OFDM subcarriers to different mobiles for achieving multiple access. 3.2 . OFDMA and the LTE generic frame structure Even though OFDMA involvement in adding complexi ty into Resource Scheduling. But, it i s the best choice of multiplexing scheme for 3 GPP LTE down link . OFDMA is vastly superior to packet - oriented approaches in terms of efficiency and latency [65] [67] [68] . The b est way to define the structure of the downlink : The users are allocated a specific number of subcarriers for a predetermined amount of time. These ar e called physical Resource B locks (RB ) in general in the LTE specifications [80] [81] [82] [83] . RBs thus have both a time and frequency coordination . Allocation of RB is handled by scheduling function or operator in the (E n ode B ase ). In reality , there are smaller unit s than the Resource Block called Resource Element (RE) and it is the smallest unit in the frequency time structure. RE is clear in figure (4) and it is no t important in scheduling as each user in scheduling process will be assigned a number of Resource Block s (RB) as shown in figure (6 ) ho w grouping a certain number of Resource E lement s become one Resource B lock [69] [72] [75] [77] [78] . 27 Figure 4 Resource Grid LTE frames are 10 m S ec in duration. They are divided into 10 sub frames , each sub frame being 1.0 m S ec long. Each sub frame is further divide d into two slots, each of 0.5 m S ec duration. From the other side, in the same representation, each slot consists of either 6 or 7 ODFM symbols, depending on whether the normal or ext ended cyclic prefix is employed as shown in figure (5) the 2 - dimension Time and Frequency grid. 28 Figure 5 2 - D Time frequency grid The transmit ted downlink signal consists of a portion of the bandwidth (N BW ) at the same time from an other side of the coordination of a subcarriers for a duration of N Symb OFDM symbols. It can be represented by a Resource Grid as shown in Figure (6). In figure (6) e ach box within the grid represents a single subcarrier for one symbol period and is referred to as a R esource Element as the dashed boxes are the Resource Block RB the main unit we are dealing with in scheduling [79] . 29 Figure 6 Orthogonal Frequency Multiple Access OFDMA . In general , LT E does not employ a PHY introduction to facilitate carrier - offset estimate, channel estimation, timing synchronization , etc. Instead, special reference signals are embedded in the Resource Blocks RB s as shown in Figure 7 . Reference signals are transmitted during the first and fifth OFDM symbols of each slot when the short Cycle Prefix CP is used and during the first and fourth OFDM symbols when the long Cycle Prefix ( CP ) is used. 30 Figure 7 2 - DL Channel Mapping T hat reference symbols are transmitted every sixth subcarrier. Further, reference symbols as the rest of LTE structure are settled in both time and frequency at once . The channel response on subcarriers bearing the reference symbols can be computed directly. Interpolation is used to estimate the channel response on the remaining subcarriers. 3.3 . Vienna LTE - A Downlink System - Level Simulator : Vienna LTE / LTE - A link level simulator [37] is a Matlab toolbox that has been developed at Vienna University of Technology. This simulator has been used toward evaluating several scheduling algorithms in this work. The main goal of this simulator is to enable the analysis of net work performances. This toolbox is capable of providing different base stations in the scenario. However , as our work focuses o n the scheduling algorithms in the downlink , we are 31 just applying it with one base station. This way the Region of Interest (ROI) is simpler and no more spec ific information about it is needed . Briefly, this Matlab tool box has a link measureme nt model which is responsible for measuring link parameters. W ith this link measurement , link quality is demonstrated based on the measurem ents c ment. This measurement will lat er be sent to the base station in the form of a measurement s report. So, the resource a llocation will be based on the adaptation algorithm we are introducing in this work. In other words, the brain of the network is the scheduler . B ased on the link measurement model, the link performance model predicts the Bit Error Rate BER based on the receiver signal to the interference ratio (SINR) and the transmission parameters. 3.4 . Packet scheduling in wireless technology: Usually Packet scheduling is called Netw ork Scheduling in the literature . The best definition for this scheduling is: An algorithm ( rule) program installed on the E n ode B ase in packet switching communication network s . It manages the sequence of network packets in the transmit and receive queues of the network interface. There are some scheduling algorithm s already in existence that are used to manage this control rule. In thi s work, the simulation ran into these criteria : 32 Table 1 Environment Criteria. Parameter Value System bandwidth 1.4MHz Subcarrier spacing 15 kHz Channel profile PedB/Rayleigh Simulation length 10,000 subframes Number of users 12 User's speed 1 km/h The most common algorithm s that exist in LTE wireless network s are: RR, WRR, and Max_CQI. 3.2.1. Round Robin Scheduler (RR) : Round Robin is one of the most commo n implemented algorithm s in networking technology in terms of scheduling routine; it is employed by process and network schedulers in computing. Round Robin resource allocation is an algorithm applied toward resource sharing between the users or channels. Typically, in the previous netwo rk generation, Round Robin time slices are assigned to each process in equal portion and circular order. Therefore , this process has no conditions and as simple as this : there are no priorities or power in the duration . RR is the same procedure as packet switching in the regular networks scheduling. 33 Figure 8 DL Channel Mapping We applied the Round Robin scheduling technique into the LTE network. This means that all the UEs shared the Bandwidth equally with no conditions. Figure 9 Round Robin Schedule flow chart By running the s imulation , we got th e performance shown in next figures starting from figure 10, this performance of the network and the subscribers UEs at certain modeled circumstances in terms of channel type, Channel Qua lity Indicator, etc. The measurement criteria for the network 34 performance are: Block Error Rate and Throughput in relation to Signal to Noise Ratio SNR for both the E Node B ase and individual users. In order to show the development, we provide in the novel technique , we discuss the performance of several individual subscribers UE in the following terms: UE 1 Throughp ut toward spectrum of Signal to Noise Ratio, Block Err or Rate for the UE1 toward spectrum of Signal to Noise Ratio. The next plots show the performance for different type of channels by applying Round Robin scheduling technique : Figure 10 E - node B throughput PedB Channel . 35 Figure 11 Block Error Rate PedB Channel Figure 12 E - node B throughput flat Rayleigh channel 36 Figure 13 E - node B Block Error Rate Flat Rayleigh channel Figure 14 E - node B throughput PedB channel 37 Figure 15 E - node B Block Error Rate flat Rayleigh channel Figure 16 E - node B throughput flat Rayleigh channel 38 Figure 17 E - node B Block Error Rate flat Rayleigh channel From Figure (9) to Figure (16) , they show the performance of the system using Round Robin scheduler. The performance shown in Bit Error Rate and Cell throughput reading with spectrum of Signal to Noise Ratio SNR . These plots show the stability/ reliability of the network . But, the cell throughput 2.5 Mbps at max is really low comparing to the other schedulers as shown below for the same circumstances shown in table ( 1) Environment Criteria. 3.4 .2. Weighted Round Robin Scheduler (WRR) : The Weighted Round Robin scheduling started to be implemented more often, immediately after the ATM protocol lunched. Now WRR is under demand in some cell networks. 39 The Weighted Round Robin sch eduling has been designed to better handle services with different processing capacities. Each server can be assigned a weight, an integer value that determine the processing capacity [46] [47] . Users with higher weights receive more connections than those with lower weights; subscribers with higher weights receive new connec tions first, while those with lower weights, and users with equal weights get equal connections. The next diagram provide s an example of the WRR: Figure 18 Weighted Round Robin Scheduling In this example the normalized weights are [2,3, 1,1] for the users [yellow(Y), o rang(O), pink(P), blue(B)] . So, the schedule sequence will be YYOOOPB YYOOOPB for two sequential slots. The W eighted Round Robin scheduling is way better than the Round Robin when we have stable user performance and t heir performance is not changing rapidly with time. Therefore , if subscriber behavior is changing rapidly and competitively with time the Weighted R ound Robin 40 will not performance . It is for this reason we are repre senting our novel technique which is dynamic and uses the updated feedback from the channel situation from the previous sent TTI. 3.4 .3. CQI_max Scheduler: It is clear from its name the main key in this scheduling strategy is assigning the resources just with the users who have the maximum channel indicator. T his means the strategy assigns resource blocks to the users with th e best radio link conditions. The pilot signal that has been sent by UEs indicates the channel statues called Ch annel Quality Indicator CQI. Then the E node B ase take s only the the best channel condition and send s them their packets through the channel, and the E node Base other u sers packets . It i s observable CQI _max schedulin g algorithm will inc rease the E node B ase capacity at the expense of fairness among users and stability of the individual. So, with using t his algorithm, users located far from E node B ase are unlikely to be considered in scheduling and this is real issue . The next figure s sh owing simple diagram about how it works. 41 Figure 19 Best CQI scheduler flow chart Figure 20 CQI Max Scheduler 42 We ran this scheduling algorithm in the simulation program within the same conditions as implement ed in the previous Round Robin s cheduler RR algorithm and indicated in table (1). We got these results plots for the performance for Flat - Rayleigh channel using CQI_max : Figure 21 Block Error Rate for E - Node B Figure 22 Throughput for the Cell 43 Figure 23 Block Error Rate User 3 Figure 24 Throughput User 3 44 Figure 25 BL Error Rate User 10 Figure 26 Throughput User 10 45 Starting f rom Figure (19) to Figure (24) , these plots shows the performance of the s ystem using CQI - Max scheduler, t he performance are shown in Bit Error Rate and Cell throughput. From these plots it i s observable the system not providing stability /reliability in the individual u sers point of view. As the sharp edges of the cell performance curves , and i t is more clear in the i ndividual Users performance plot s as shown f or User 10 and User 3. The discontinuity of the users performance showing the main reason of the delay, a s at certain times in the curves the individual user was not included in the scheduling, t his means the conne ction was diminished because its CQI feedb ack was wea k ( not the Max_CQI) during the time of discontinuity . 3.5 . Novel Scheduling Algorithm (S elf O rganizing Neural Network Algorithm SONN ): 3.5 .1. Introduction about SONN Algorithm : After deep review in the current publication s that talk about Self Organizing in wireless networks, this with reviewing the performance of the scheduling algorithm with the new standard that LTE networks are offering into the resources (Time vs Frequency) , which considered as a new unit in telecom industry and it is called (Resource Element) . This new technique of sending/presenting the information accelerate the speed of the network in wireless networks by increases the capacity , as the capacity is one of the major challenge s ( Bottle Neck) in the wireless. However , sending the information with this method is nee ded to be smart of presenting it. I n other words , mapping the information with certain way effects the speed of network by effecting the capac ity of wireless channel. As LTE and LTE - A networks are sma rt and capable of providing feedback information to the E node B ase from the user e quipment UE about the 46 channel situation in terms of Quality o f S ervice (QoS), Channel Quality Indicator (CQI) and Signal to Noise Ratio (SNR ). 3.5 .2 . Self O rganizing Map (SOM) process: The SONN technique we are introducing here has core algorithm browed from Artificial Neural Network called Self Organizing Map (SOM). Practically with implementing the main three steps of SOM with certain specific ways and functions: 3.5 .2 . 1. How SOM is a competitive process: From the beginning, we should have neurons vectors with same dimension as the input space. As the input data is X=[x 1 ,x 2 m ] T From the other side the synaptic weight vectors, where the synaptic weight vector of neuron j be denoted by: W j = [w j1 ,w j2 jm ] T L is the total number of the neurons in the networks. We are comparing the output of the multiplication of W T j for the wholes neuron in the network and choose the high est value; we will take the indicators of the highest value. Here the only thing we really interested about is the indicator. name i(x) as indicator. So, we can explain our calculation in the competitive process as: i(X) = arg min || X - W j (3 .1) 47 We are looking for the (i) which indicate the winner neuron X i, where X i is the best matching neuron i. [7] 3.5 .3 . The Cooperative trait in the SOM process: From the previous process step s , it is clear that the winning neuron is going to be updated by the input part. To make the other neurons effected by these input component. However , we should update the other neuron s with less power than the winning neuron to get positive cooperation from them and this for fast convergence of the neuron network [113] [114] . To achieve these requirements: The neuron locates to the center of a topological neighborhood of cooperating neuron should have the maximum coefficient and the others should get smaller coefficient where the neurons that excited with amount close to the winning neuron should have coefficient higher than the farther ones. It is clear that we are looking for a function decaying slowly as we are going far f rom its peak. We end up with function h ij that denote the topological neighborhood centered on winning neuron (i) and excite neuron (j). the d ij denotes the distance between the winning neuron (i) and excited or effected neuron (j). So, the function h ij i s unimodal function of the lateral distance d ij which going to satisfy this conditions: 1 - The topological neighborhood h ij is symmetric about the maximum point defined by d i,j =0 , here we try to say the winning neuron has zero distance of this function 2 - The amplitude of topological function is decreases monotonically with increasing the lateral distance d i,j , which decaying to zero for d i,j The typical function that applies these conditions is Gaussian function: 48 h i,j = exp( - ) (3 .2) d i,j called effected width of the topological neighborhood as illus trated in the coming figure; It i s a measure ment degree of effecting of each neuron in the update. Figure 27 Gaussian Neighborhood function Toward apply best cooperation we apply the distance in the neighborhood function h i , j is centered by the winning neuron and decrease as we get less value than the value of the winning neuron. The distance between winning neuron and the other neurons d i , j , d i , j is an integer equal to |i - j|, and the cas e of two dimensional lattice it i s g oing to be defined as: d 2 i,j = ||r j r i || 2 (3 .3) 49 The discrete vector r j defines the position of excited neuron j and r j defines the discrete position of winning neuron i, both of which are measured in the discrete output space. SOM has lot of unique features in its algorithm; one of these features is the control it has on the width of the topological neighborhood function especially it is shrinking with time. This feature h e lps a lot to keep the network instable situation after it reach e s to the convergence period. Therefore , we can write it as: (3 .4) o 1 is a time constant. With using, the become : h j , i(x) (n) = exp( - (3 .5) A ll the component of the previous equation is well known where the time (n) is equal to number the t opological neighborhood function will response with similar way. With having wide width of neighborhood function at the beginning will help most of the neuron be effected by the update then the decrease of the function will help a lot on the correlation fu nction to stay converged to certain value. In most of computer programs that using the SOM are using normalizing technique and it is called renormalized SOM from the training. According to which we work with a much smaller 50 number of normalized degree of freedom. This operation is easily performed in discrete form by having a neighborhood function h j,i(x) (n) of constant width, but gradually increasing the total number of neurons. The neurons are interest ed halfway between the old ones [7]. 3.5 .4 . Novel Ad aptive P rocess It is well known that the Self Organizing Map ( SOM ) is a neural network. At the last stage of the adaptive process , the synapsis w j that belongs to the neuron (j) should be effected by the input data (x). The main target is presenting the effect that x can apply to different synaps. As the SOM is an unsupervised learning algorithm so the updating process automatically can be done proportionally to the input (x) as: j j x +g(y j ) w j (3 .6) This is the general presentation for the effect of the data on t he synaptic. The main parameter that effects the update value is the step size . is the learning rate parameter that the synaptic will be effected by the input data, a nd y j can be applied as the function talks about the response we have for any kind of input data, where both g(y j ) and y j are just correcting factors, y j = h j,i(x) (3 .7) The previous equation (3.6) can present the update changing value of each step as : j j,i(x) (x - w j ) (3 .8) After we understand the increment that is effected by the input data which s hould be added to the previous weight at each instant , we can see the big image for the weight s reaction as: w j (n+1) = w j j,i(x) (n)(x(n) - w j (n)) (3 .9) 51 The equation (3.9) will be applied for all neurons fond in the l attice as a neighborhood function. B ut, it is going to affect each neuron with a differen t value depending on each location of the winning neuron. T he winning neuron is going to be updated with a complete factor or multiplied by value (one) and the others will be effected proportionally, depending on how far the complete matching is from the input data in the current iteration . T he process keeps running and using the previous equation s once at each TTI , and after each TTI the process ends up with a grid of the whole neurons . The grid explains the historical input data over all previous TTIs with a small number of neurons presenting this historical Channel Quality Indicators ( CQIs ) feedbacks . The updated weights function depends on a lot of functions as well. O ne of these function s is h j,i(x) which is a heuristi c function, and explains the neighborhood behavior for the whole neuron compare to the all input data. The other heuristic function is function depends on the history of the input data. The learning rate non - a function and with the time with o which is proportional to time increasing toward converging the mod - SOM algorithm , be e x pressed as: o exp( - (3 .10) You can see the learning rate is decaying with time, and it has a living time where it is going to maintain stable response for a while th en will start decaying, this means we have complete control in change function [88] [112] . 52 As the learning rate is changing with time, it is traveling through two main phases during the whole SOM process: Modification done to Self - Organizing Map (SOM): This is one of the main contribution s in this work . we are using a modified form of updating the weight s . I n this novel algorithm the updated weights are energy based and the updated function including the non - linearity . This helps our case, as we want to do mapping with cluster ing in one direction. As in the coming equation we will deal with the energ y instead of the regular Euclidean Distance or Kullback - Leibler Distance : E j N (X l (n) W j,l ) 4 ] (3.11) I n the next steps we compute the winning neuron which is the winning Channel Quality Indicator ( CQI ) feedback that presents the UE in the algorithm : i(x) = arg min j ( E j ) (3.12) Then, we update d W K (n) = (n)h ik (n) / w K (E j ) for K N(i) W K (n) = (n)h ik (n) ( X(n) - W k ) 3 for K N(i) (3.13) This modification really need ed to be done, as the algorithm wa s working from the beginning starting from the first TTI . W e are using the output of the beginning output of the results a s regent and the sold output is needed even in the wa rming up period (before the convergence ) . 53 This modified updating is smoothing the values of the weights at each TTI , and the price is slower to converge the weights which are visible for o u r case here. 3.5 .5 . The Self Organizing Neural Network (SONN ) algorithm LTE Scheduling Section : As in this algorithm SOM technique is the core or the key function in the process. The SOM algorithm can be explained in clear steps, Wi th th ese clear steps we can summarize the M algorithm can be applied for. Ko substitutes a simple geometric computation for the more detailed properties of Hebb - Like rule and lateral interactions. The main vision for the algorithm can be summarized as: A continuous input space of activation patterns that are generated in accordan ce with a certain probability distribution. A topology of the network is in the form of a lattice of neurons, which gives discreet output. In other words , one of main use s of the SOM al gorithm is to change a continuous input data to discrete output data wi th another presentation. The SOM algorithm is using a new technique by applying a time varying neighborhood function for the winning neuron i(x) which will up date the neighbors neurons and has close values to the winning neuron . This will be updated but wi th smaller values depending on how far it i s from the winning neuron. One of the important paramete rs in the SOM algorithm is the l o and then decreases gradually with time, n, but never goes to z ero. Those are the main ste ps we can explain more about one of the main parame ters on the algorithm which is the neighborhood function, and can be used when applying the next two equations immediately in sequence: 54 H j,i(x) (n) = exp ( - ), n = 0,1,2 (3 .11) 0 exp( ) (3 .12) One of the main need for applying these equat maintain small values like 0.01 during the convergence period w hich going to be after long iterations . The other noticeable way to apply the algorithm is dealing with small neighborhood function even single effected neuron at the earlier first steps and w ider at the last ones. In general, we can summaries the SONN algorithm as: 1) weight vectors w j (0). The only restriction here is that the w j (0) be l where j is the number of neurons in the lattice. It may be desired to keep magnitude of the weights small. The other way to initializing the algorithm is to select the weight vectors {w j (0)} l j=1 from the available set of input vectors {x j } l i=1 in random manner. 2) Every node is examined to find the Best Matching Unit of the weight vectors. This step called sampling, draw a sample x from the input space with a certain probability; the vector x represents the activation pattern that is applied to the lattice. The dimension of vector x is equal to m. 3) minimum distance Euclidean criterion : i(x) = arg min j ||x(n) w j (3 .13) 55 so, here at SOM we are exciting about the rank of the winning neuron to update its value. Then, the radius of the neighborhood around the weight vector is calculated. The size of the neighborhood decreases with each iteration. 4) Each weight and its neighborhood has i ts weights adjusted to become more like the wanted shape for the SOM weights. Nodes closest to winning neuron are altered more than the nodes furthest away in the neigh borhood. Here is the updated rule : w j (n+1) = w j j,i(x) (n)(x(n) w j (n)) (3 .14) rate parameters, and h j,i(x) (n) is the neighborhood function centered j,i(x) (n) are varied dynamically during learning for best result. 5) Repeat from step 2 for enough itera tions for convergence. 6) In terms of mod ification, we can use equation ( 3 .14) with odd power to the deference, as in our case, we use the energy in eqn (3.11). Then, the weight update will converge faster and provie quick results. In that case, the update law becomes: w j (n+1) = w j j,i(x) (n)(x(n) w j (n)) 3 (3 .15) 7) In terms of modification, we can use convex sum technique by averaging the effect of each winning neuron. And that going to be implemented as well in chapter 5. This is summarized below as Algorthim1: 56 Algorithm 1 : Energy Based SOM Algorithm 1: Procedure INITIALIZATION 2: W Initializing the adaptive weights 3: Fetch the vector of Indictors for the 1 st TTI 4: 5: 6: 3 Final Shape pf the non - linear function 7: Processing the loop 8: Computational part 9: E K l (n) W j,l ) 4 ] 10: i(x) = arg (min(j))E j for all j 11: Updating part 12: k ik K )(E j ) 13: W K ik (n)(X(n) - W k ) 3 For K N(i) 14: Close Algorithm 1 will be called at each TTI = 10ms by algorithm 2 as it is responsible for the management and for clustering the Channel Quality Indicator CQI feedback time - sequence. 3.5 .5 .1 . SO Scheduler Algorithm Frame w ork: The design of the Downlink scheduling algori thm, this link is between the E Node B ase as source and User Equipment as destination. This scheduling algorithm is complex procedure and it have a number of design challenges, for example maximizing the system capacity and spectral efficiency, fairn ess approach, bit error rate etc. This report p resenting new approach for 57 a such algorithm that handle all these challenges with providing optimum algorithm. As t he OFDM parameters (Resource E lement), timing and frequency settings, channel quality feedback and channel quality indicator (CQI)) standardi zed according the 3GPP standards . Therefore , the main principle of the scheduler in LTE and LTE - A is dynamically determined every 1ms interval which usually one Transmission Time Interval (TTI). The main rule of this algorithm is to get advantage of all information we got from the feedback in the previous TTI and make the mapping of the coming TTI as similar as it makes . This scheduler is dynamic scheduler like Best CQI scheduler . But, this one is a way more optimum than the b est_CQI_ scheduler as MSOM putting in consideration all the users in counts and it trial to map the users as the environment treated them in the previous TTI . Each user faces different channel conditions at a given time. At any given time, there will be hi gh probability that some users will have good radio link condition. This Mapping will be done to the CQI parameters we got through the feedback for the previous TTI and it will be done by Self Organizing Map (Artificial Neural Network) to these CQI , so a fter we got the best mapping of CQI in values and indicators of users, the algorithm will use the mapping of the indicators to schedule the users into the Resource Block. With this way most of the users get chance to receive data this grantee the fairness and vanish the latenc y that could happen because of scheduling, and the users of best CQI got the biggest part of resource block, which will enhance the overall throughput . Puttin g the overall procedure into steps: 1) Get the CQI_Feedback matrix of the previous TTI, these CQI values should have indicators to the users on each one of them. 58 2) Find the mapping of these value in the CQI_Feedback relating to the Maximum CQI_feedback using Self Organizing Map (ANN) Matrix and corresponding user. 3) Now the algorithm should map the users to this feedback information. In addition, Schedule that user in that RB. 4) The schedule grid scan the mapping if there are user or more not in the map: It will take RBs from the most repeated users to sign them to missing users one time. 5) Until the end of the TTI, this user will not have permission to be scheduled. With the TTI finish so w e will release this schedule to be applied. This Novel schedule provides a way better result in terms of network throughput as well as it gives a way better performance of UE throughput and UE latency. As it is providing UE with better channel environment main priority as well as it provides the UEs with poor channel quality spots in the Resource B locks of the mapping. In general speech, t his a lgorithm dynamically adjusts the transmission rank, precoding matrix indicator and channel quality indicator according to the feedback (if present). Afterwards it schedules users proportional to their theoretically attainable rate (as the true one is not k nown) This algorithm are maintain to apply a certain fairness into mapping by at least provide one resource block RB to users even if there are zero values Channel Quality Indicator CQI for them. This Sub - rotini added to make the algorithm not that greedy with CQI. This is the Algorithm 2: 59 Algorithm 2 : Self Organizing Scheduler 1: Procedure FE TCHING 2: Find the UE feedback of the previous TTI 3: CQI 4: Iteration 5: THD: Eliminate UE with CQI below the threshold 6: 3 Final Shape pf the non - linear function 7: Processing the loop 9: Call Energy Based SOM Algorithm for execution 11: Wait TTI to be 12: @ End of TTI, Release the scheduling map 13: Close It i s clear Algorthim2 is showing the superposition method of the user mapping into the Resource Blocks RBs procedure. Step 3 in Algorithm2 will have more details in the next highlight. 3.5 .5 .2 . SO NN Scheduler Algorithm Mapping : This section contains the details of step number 3 on the SO scheduler Algorithm framework ; this step has a novelty way o f mapping which provide stability and accelerate the process of scheduling all this has been provided based on the statistic distrib ution of the clusters 60 of the Channel Quality Indicators CQI of the users. This has been done with u sing non - linear function SoftMax function. This step really make the scheduling process faster and made a such a hierarchy to the scheduling process helps to ward managing the users into the resource blocks RB grid, this matters a lot when the cell dealing with large number of users. As dividing, the users into clusters based on their statues on the historical CQI which is the pattern of the users give the deci sion easier to be taken and implemented. Practically , Step 3 in the SO Scheduler Algorithm Frame is called the Novel Mapping Algorithm of the Clusters and it receives the users distribution inside clusters (groups) based on the historical CQI over the all previous time (all previous TTIs). The Novel Mapping Algorithm of the Clusters: 1 - Fetch the clustered (grouped) Channel Quality Indicators (CQI) of the us ers and name all of them 2 - Provide SoftMax non - linear function to each cluster. So, each cluster will have class accordingly 3 - Map the clusters pr oportionally based on its Sof t Max value to the Resource Blocks RB grid (5.14) So choosing the clustering and ranking them will be probabilistic based not just a signed value criteria, this provide us with more robustness and adoptability to the pattern. 61 4 - Distribute the users in each class equally to the RBs portion of their class to provide kind of fairness 5 - Print the map of this TTI The process in step 3 at algorthim2 could be summarized to: Algorithm 3 : SO Scheduler Algorithm Mapping 1: Procedure FE TCHING 2: Algorithm 1 - linear fun 4: Procedure Processing the loop 5: Mapping the Cluster proportionally with SoftMax value er equally 8: Classes 9: 12: @ End of TTI, Release the scheduling map to step4 in Algorthim2 13: Close In this algorithm, we provide proportional fairness of scheduling. This is clear and has been provided by step number 4. the users in the same class has been provided the same number of resource blocks even though they could have a different CQI. This happen s with maintain providing the priority to the users with the high CQI and higher number of RB. This algorithm will be important with in a situation with a cell connecting with too many users. 62 3.5 .5 . 3. SO NN Scheduler Algorithm Performance: The figures below describe the great performance of this algorithm: Figure 28 Block Error Rate for the Cell 63 Figure 29 Throughput for the Cell Figure 30 Block Error Rate for User 3 64 Figure 31 Throughput for User 3 Figure 32 Block Error Rate for User 10 65 Figure 33 Throughput for User 10 From Figure (25) to Figure (30) shows the performance of the system using the Novel modified - SOM scheduler algorithm (SONN) . The performance shown in Bit Error Rate and Cell individual Users point of view. As there are no sharp edges in the curve showing in the cell curves and it has, more clear in the Individual Users plot like User 10 and User 3 they have high smooth curves over all the period. All these outcomes have been provided with arrangements of random process among all normal for the strategy. As the Matlab Simulation Box give us an irregular arrangement of all sub documents such is the channel modeling and ecological sets. We furnished t his procedure with finish diverse situations and we wind up with strong outcomes in a similar grouping, for example figure 34, to figure 38. 66 - As the main challenge is how we can trade - off between increasing the E Node B throughput and provide fairness bet ween the users: 3.5 .6 . Fairness evaluations among several scheduling p rocess: degree of fairness among users. Mathematically, the fairness index can be expressed as follows [18]. - (3 .15 ) - where (n) represents the total number of user, (x i )represents the throughput of individual user (i) and represents the fairness among the n end users. For this part, simulation parameters and assumptions are the same as Table 1. The spectral efficiency performance and the fairness performance of these schedulers are shown in figure 3 7 . From figure 3 7, it ca n be observed that the RR scheduler provides the best fairness. But, as shown in figure RR has least spectral efficiency performance. In contrast, the Max - CQI scheduler provides the best spectral efficiency. But, the worst fairness performances as shown in figure16. The spectral efficiency has been quantified in this work by the throughput measurements. W e have compared the performance of SONN algorithm verses the well - known existing benchmark algorithms, namely, the Round Robin (RR) and the CQI - max algori thms, on the cell throughput criteria. Figure34 depicts the resulting the cell throughput versus SNR for all three different scheduling schemes. 67 Figure 34 Cell throughput for the three different scheduler schemes. It is noted that the RR scheduling performance is the worst in the throughput measurements , since it does not consider the user channel condition into account. The CQI_max scheduling achieves the highest overall throughput in the example but at the expense of t he notion of ted in Figure 35 , the new SONN algorithm is providing a trade - off bet ween the throughput and fairne ss to all users. and in figure34 its throughput performance is in the average scope. After normalizing the thr oughput performance among all different types of scheduler s, we got the me tric of figure 3 6. This me tric will be applied for overall performance. -2 0 2 4 6 8 10 12 14 -15 -10 -5 0 5 10 15 20 25 30 THROUGHPUT [MB/S] SNR [DB] 12 UES, PEDB, SISO, ROUND ROBIN, SONN RR Best_CQI_Algorthim MSONN 68 Figure 35 Normalized cell throughput for the three different scheduler schemes. Figure 36 Fairness among users for the three different scheduler schemes. -0.2 0 0.2 0.4 0.6 0.8 1 1.2 -15 -10 -5 0 5 10 15 20 25 30 35 12 UES, PEDB, SISO, ROUND ROBIN, CQI_MAX, SONN. NORMALIZED RR Best_CQI MSONN Throuput -0.2 0 0.2 0.4 0.6 0.8 1 1.2 -15 -10 -5 0 5 10 15 20 25 30 12 UES, PEDB, SISO, ROUND ROBIN, SONN FAIRNESS RR Best_CQI_Algorthim MSONN 69 Toward a full comparison of all algorithms performance, we end up with matrix that represents the normalized sum of throughput matrices and fairness matrices in the evaluations. Figure3 7 depicts the overall weighted overall performance of all the algorithms. Figure 37 Combination ( Fairness and T hroughput ) evaluation among different scheduler schemes. It is clear the combination of the algorithms 1 and 2 make the downlink performance more reliable to the end user UE. The individual users have been provided higher throughput than the RR scheduler is providing. All this, with giving a low Bit Error Rate BER during the whole course of different environment. the performance with Modified Self Organized scheduler is a way better in terms of fai rness as we can see: The Block Error R -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 -15 -10 -5 0 5 10 15 20 25 30 35 TOTAL METRIC NORM THROUGHPUT+ FAIRNESS RR Best_CQI MSONN 70 have very close similarity in terms of UE throughput. This kind of scheduling techniques (Modified Self Organizing Mapping) gives t he promising optimization between providing the highest Cell throughput and fairness among users (UEs). 71 Chapter 4 . Predictive Recurrent Neural Network techniques We will discuss recent developments and enhancements of recurrent neural network structures with learning. The recurrent neural network is a dynamic system that has the ability to generate predictions of (the expected) values of a time - series with a relati vely small margins of error. In this chapter, we are going to apply the basic Recurrent Neural Network (bRNN) that has been introduced in [38] towards providing predicted values for the next time - series data given a history of received prior time - series d readings. Hence, the bRNN will provide a prediction (of the mean of the channel CQI) for the next TTI interval. 4.1. Implementing Novel Recurrent Neural Networks Scheduler Algorithm Recurrent Neural Networks (RNN) are using time series prediction toward making a model of Resource Block scheduling. This approach adopts an RNN, called Basic Recurrent Ne ural networks (bRNN), as introduced in [38]. The goal here is to use the prediction model based on the feedback received from the previous measurements profile by applying gradient decent in updating of the weights. This type of work toward enhancing LTE - A communication systems performance and making it reach the promise of 3GPP in Release 11. An advantage of using a bRNN network is that it is dynamic in tracking random changes as in our communication situation with wireless channel. 72 Based on [38], the ma in neural network equations are: x(t+1) = A x(t) + U h(t) + W s(t) + b (4.16) y(t) = V h(t) + D s(t) + c (4.18) The matrix A is fixed with eigenvalues less than 1 in absolute value. However, the matrices U,W,V, D and the vectors, b and c, represent the weights and biases, respectively, that will be updated at each mini - batch or epoch. When successfully trained and a dapted, the RNN learns to vectors x(t) and x(t+1) are the present and future states, respectively, and h(t) is the hidden unit vector. The vector y(t) is the outp ut of the RNN, which should become close to the target value after the RNN is successfully trained. For constant (fixed) parameters, and assuming they are stable, these equations execute inference from input sequence to states to hidden units, and finall y to outputs. The key challenge in RNNs is to execute training procedure to update the parameters (i.e., weights and biases) in order to realize a sequence to sequence mapping using training data. Towards that end, the backpropagation through time (BPTT) and its variants can be derived from constrained optimization and optimal control which produces a co - state (sensitivity) dynamics known as backpropagation and often denoted by delta or lambda variable [38, 39]. The main factor of the updating procedure i s the variable Lambda [39]. The updating procedure of the weights is a supervised version, as the Back Propagation through Time (BPTT) technique, 73 which is the gradient descent of the weights. The gradient descent developing process is conducted through BPT T [116] . 4.1.1. The adaptive algorithm : As the equations (4.16) to (4.18) are the recurrent neural network system, the weights matrices and its factors have to be fixed in the testing period. We reach to this fixed weights by training theses set of weight s with train of possible input to this network while updating these weights through this training period. Through this training the matrices weights A, U, W, V and D as well as the biases b and c are changing/updating at each iteration. This means only thr ough the training the set of weights will be changing with time. All the equations have been provided in depth and details in [38]. In this work we consider (t) as the discrete index of the training iterations. Based on the analysis that has been provided in [38] and [39], we can explain the updating procedure of the weights matrices based in the co - - state as: T T V T 1 2 - 1. The co - state is the main factor of updating all the weights. The updating equations of every weight matrix can be expressed as: U(t) = - T (4.20) W(t) = - T (4.21) 74 b(t) = - V(t) = - e(t) (h(t)) T (4.23) D(t) = - e(t) (s(t)) T (4.24) c(t) = - e(t) (4.25) We consider e(t) as the error between the observer and the desired output in the training process. The parameter is the learning rate process. All these previous updating equations are applied by each iteration in the training period of th e neural network. e (t) = yh(t) y(t) (4.26) It is clear y(t) in previous equations points to the correct target value of output at time t, and yh(t) represents our calculated output at the same time t, while y(t) represents the next TTI target during training. In the training phase, the adaptive algorithm seeks to update the bRNN parameters, e.g., U,V and W, in order to minimize the error function and renders (the mean) yh(t) become close to the mean target y(t). Figure 34 shows the unfolding of the RNN over a fine time duration for an illustration time - stepping sequencing of input to (internal) state to corresponding output of a generic RNN. xh(t+1) = Ah x(t) + Uh h(t) + Wh s(t) + bh (4.27) yh(t) = Vh h(t) + Dh s(t) + ch (4.29) 75 Figure 38 Recurrent Neural Network RNN unfolding over time index Equations (4.16), (4.17) and (4.18) process are summarized in the diagram shown by figure 38. x(t) is the state vector , s(t) is our input data vector and yh(t) is the output vector. 4.1.2. Evaluation of the Basic Recurrent Neural Network (bRNN) predicti on: The evaluation of the bRNN prediction uses the codes developed in [39] and will be adapted here toward random values with means representing the expected values of the Channel Quality Indicators CQI feedbacks. The input data will be provid ed to the weights in just regular dynamic bRNN system by using regular equations (4.16),(4.17) and (4,18) for such number of mini - batches or epochs. Then, after successful training, we will perform inference using the resulting bRNN for prediction. Sample performance results are shown below for one - dimensional state, with a scalar and output variables. The displayed results also show the co - state variable (Lambda) which represent the back - propagating error at the end of training [39]. It is 76 noted that vari able Lambda has zero mean over the horizon (duration) which in turn indicates that state and output tracking has been successful: Figure 39 Simulation depicting an example results of the training process of bRNN 77 Figure 40 An Example output error signal profile during training Figure 41 Example Output Error profile 78 Figure 42 Dynamic system output and target values Figure 43 Example State Error 79 Figure 44 Example State performance The predictions has been provided. As we are at time t, we have the real target value of y(t+1), and thus the predictor output yh(t) will be esti mating the target value y(t+1). The previous plots show the good performance of the future prediction form of the bRNN dynamic system. Zooming the plot of the output performance of the system, the target values and the shifted version of the target output are shown in figure 41. 80 Figure 45 Output performance The dynamic bRNN output is represented by the blue line, and it is performing exceedingly well in regards to the real targeted value, as represented by the red line (which is a one - step shift of the target shown in green). The output of the network reasona bly tracks the one - step value of the target. 81 Figure 46 Predicted values vs Real values We are observing reasonable tracking in the two sample previous plots. These results show after the weights converged the dynamic bRNN and provide great prediction readings. 82 Chapter 5 . Novel Sch eduler to LTE Resource Blocks by applying Recurrent Neural Network techniques The implementation of a RNN algorithms into the LTE Resource Blocks (RB) scheduler can be explained in systematic steps. These clear steps summarize our technique in employing a RNN algorithm into a scheduling procedure. The RNN scheduling algorithm constitutes several computational steps. The comprehensive view for the algorithm is as follows: As the input space of activation patterns are generated in accordance with a certain probability distribution. The dynamic system that bRNN provide will learn the pattern of the channel performance of the users that attached with e node Base. In this work here we apply the whole process of learning and testing procedures at every transmission action. Where in the first part of the TTI we are training the bRNN a nd testing it to provide results to the scheduler. Then, we are just providing the test part of the procedure only. This over all procedure applied every TTI. Because the main feature of the RNN is the storage unit (hidden State) , which provide memory to the bRNN about the input data. As well as our work here in wireless mobile system which include mobility to the action this lead to high probability of changing environment in short period. So, It is better to repeat the whole proce ss (Training and Testing) every TTI. 5.1 LTE - A Resource Blocks With Recurrent Neural Network Techniques In this chapter, we apply the bRNN procedure that was outlined in chapter 4 to provide predictions of the Channel Quality Indicators CQI for the next TTI period. After initial training, one can employ the bRNN for inference or prediction. In that scenario , the bRNN predicts the 83 CQI of the channels among users based on the last previous profile reading. As the CQI is predicted before TTI period, one can optimally schedule or prioritize the users Resource Blocks based on chosen criteria. In general, we can summarize the b RNN algorithm as: 1) Initialize weights of the observer : this can be done by choosing random values for the initial weight vectors U h , W h , V h and D h as well as initiate the constants b h and c h to zeros at equations (4.16), (4.17) and (4.18). T his will happen with initializing current only matric values for the state values x(t) . Then, when the algorithm is ready to work, it will adapt itself as t=1, (t) is changeable as it does not have to be the same size of the input data, we chose to make it 1 It is may be desirable to initiate the magnitude of the weights with small values . Other parameters such bias and matric b h and c h in equations (4.16), (4.17) and (4.18) needed to be initialized . This initialization has a counter and it is only executed for the 1 st transmission action only. The matric A h will be constant through the whole process A h 2) We run bRNN algorithm ever y transmission action . CQI o f the previous transition matric will be provided and set as our input s(t) = CQI(t) . at each TTI we are applying s(t) as our input data and the reference/desired value y(t) = s(t+1) = CQI(t+1) toward training our bRNN network in the training phase. 3) We compute the error between the CQI fe edback and the observer E(t) = yh(t) y (t). This E(t) is evaluator of our predictions and it will be used toward updating the weights through the training phase. We can use E(t) as alert toward starting the training phase if E(t) excee d threshold values. 84 4) We run the set of equ ations (4.16) to (4.18) to get y h (t) and x h (t +1 ) at each TTI . 5) We apply the gradient decent calculation s [38] toward use it into updating the observer weights for the next transmission action . This step has a counter and it will be executed only for the first 200 transmission action , It is called training phase . 6) We will apply the output of this algorithm w hich is the vector y h(t+1 ) the prediction of the next CQI feed back . W eight s of the obser ver ar e updated by set of equations (5 .22). x h (t+1) = A x(t) + U h h(t) + W h s (t) + b h h(t) = x h (t)) (5.22 ) yh (t) = V h h(t) + D h s (t) + c h Applying equation (5.22) only without applying updates to the weights will be to the rest set of TTI If there are no alert from E(t). In equation (5.22) (t) is the nonlinear function and it is considered as the learni ng rate parameters, and (t ) or h(t) are varied dynamically during learning for the best result. There are steps will be executed for the one iteration and they are for initialization, other selected steps will be exec uted for the first 200 transmission action , they are for seek ing of training our algo rithm and we consider this as wa rming up period. The time line of this algorithm can be divided to two stages as shown in algorithm 4 . Training stage: where all the weights ( General output plant & Optimal observer ) are implemented and applied. Then, Execution stage: where we execute only the observer part). This is Algorithm 4: 85 Algorithm 4 : bRNN updating Algorithm 1: Procedure INITIALIZATION 2: A h 3: U h, W h , Vh,Dh 7: Fetch the vector of indictors for the first transmission action 8: 9: Training Phase: For the first 200 transmission actions 11 : xh(t+1) = Ahx(t) + Uh(t) + Wh s(t) + bh h(t) = tanh ( gamma xh(t) ) yh(t) = Vh h(t) + Dh s(t) + ch 12: bRNN error fu nction evaluation E(t) = CQI - Yh 13: calculate gradient descent (BPTT) to all bRNN weights coefficients 86 14: update all bRNN weights 15 :Processing the loop : From transmission action =201 to end 16 : optimal observer: xh(t+1) = Ahx(t) + Uh(t) + Wh s(t) + bh h(t) = tanh ( gamma xh(t) ) yh(t) = Vh h(t) + Dh s(t) + ch 17: Provide the prediction 1 8 : Close Algorithm 4 will be called at each T TI = 10ms by algorithm 5 as it is responsible of the management and clustering of the Channel Quality Indicator CQI feedback. In the previous algorithm.4 the CQI feedback of the previous TTI ha s been fed to the algorithm as x (t). then we run the bRNN algorithm with time shifting explained in chapter 4 to provide the prediction of the values of CQI of the current TTI before the E node B receives it. Based on these predictions. We will map and position the users to the Resource Blocks RBs for the coming TTI. More details about the procedure will be provided in the next sections 5.2.1. The Scheduler Algorithm Framew ork By Implementing the bRNN In Scheduling: As bRNN is supervised learning that has prediction to the future st ate possibility feature, we are 87 applying this feature toward providing us with prediction for the future expected state based on the previous performance, and then it will be updated based on the real values . The input data s (t) will be the Channel Qualit y Indicator (CQI) of the last feed - back from the previous signal we received from the users UEs. The code will run for one step size at each TTI. So, the first 4 to 5 TTIs will be as considered as a training period for the algorithm. Then, it could start p roviding us So then we are worming up our system with the first 200 TTI after that our bRNN algorithm become solid and ready for the practice. Algorithm 5 : Self Organizing Scheduler 1: Procedure FETCHING 2: Find the UE feedback of the previous transmission action transmission action 4: 5: THD: Eliminate UE with CQI below the threshold nal Shape of the non - linear function 7: Processing the loop 9: Call bRNN Algorithm for execution 11: Wait transmission action 12: @ End of transmission action, Release the scheduling map 13: Close It is clear Algorthim4 is showing the superposition method of the user mapping into the Resource Blocks RBs procedure. By applying Recurrent Neural Network based algorithm toward 88 mapping the Resource Block RB, this will provide a prediction for the Channel Quality Indicator CQI feedback. This means that RNN can build state over the entire training sequence and even maintain that state if needed to make the prediction. 5.2.3. Mapping the bRNN scheduler by SoftMax : This section contains the details of step number 3 on the scheduler Algorithm framework bRNN ; this step has a new motivative way of mapping which provide stability and accelerate the process of scheduling, all this has been provided based on the bRNN predictio n . This has been done by using non - linear function the SoftMax function. This step really makes the scheduling process faster and made a hierarchy in the scheduling process, which helps toward managing the users into the resource blocks RB grid. This matte rs significantly when the cell dealing with large number of users. I t receives the predicted users feedback CQI and map them to the resource blocks. 1 - Fetch the predicted users feedback (CQI) of the users and name all of them 2 - Provide SoftMax non - linear function to each user . So , each user will have class accordingly 3 - Map the clusters proportionally based on its SoftMax value to the Resource Blocks RB grid So, by choosing the users and then ranking the clusters, they will be probabilistic based not just a signed value criterion, this provides us with more robustness and adaptability to the pattern. 4 - Distribute the users in each class equally to the RBs portion of their class to provi de kind of fairness 89 5 - Print the map of this transmission time. The process in step 3 at algorthim5 could be summarized to: Algorithm 6 : bRNN Scheduler Algorithm Mapping 1: Procedure FE TCHING 2: 4( priorities by bRNN) - linear fun 4: Procedure Processing the loop 5: Mapping the Cluster proportionally with SoftMax value uster equally 8: Classes time 9: Wait transmission time to be finish, no schedule 12: @ End of transmission time , Release the scheduling map to step4 in Algorthim5 13: Close In this algorithm, we provide proportional fairness of scheduling. As provided by step numbe r 4. The users in the same CQI prediction have be en provided the same number of resource blocks. This happens with maintain providing the priority to the users with the high CQI and higher number of RB. 90 5.2.4. bRNN Scheduler Algorithm Performance The figures below describe the great performance of this bRNN scheduling algorithm: the perfor mance evaluation parameters are E node b ase throughput, Bit Error Rate BER of the E node B, throughput of randomly selected Users Equipment UE and the UE Error bet Rat BER versus several Signal to Noise ratio SNR environments. The next 6 plots explain the performance criteria: Figure 47 Block Error Rate for the Cell using bRNN 91 Figure 48 Throughput for the cell using bRNN 92 Figure 49 Block Error Rate for User 3. bRNN Figure 50 Throughput for User 3 bRNN 93 Figure 51 Block Error Rate for User 10 bRNN Figure 52 Throughput for User 10 bRNN 94 Figures starting from Figure (43) to Figure (48) show the performance of the system using the Novel modified - RNN scheduler algorithm. The performance shown in Bit Error Rate and Cell throughput that shown at these plots it is observable the system is provi ding dynamic reliability in the individual users point of view. As there are no sharp edges in the curve showing in the cell curves and it has become more clear in the individual users plots, like User 10 and User 3. That they have high smooth curves over all the period. scheduler is much better in terms of fairness as demonstrated by: The Block Error Rate of the cell. It is even more clear on the Block Error Rat the UEs almost have very close similarity in terms of UE throughput. This kind of scheduling techniques (Modified Recurrent Neural Network) gives promising optimization between providing the highest cell throughput and gives fairness between users UEs. This is real Proportional Fair scheduler. All these results have been provided with complete set of random process among all characteristic of the procedure. As the matlab box provide us with random set of all sub files such is the channel modeling and environment sets, we provided this process with complete different scenarios as figure 48 proof, so we end up with solid results in the same sequence such as in figure 53, to figure 55. We have compared the performance of bRNN algorithm verses (SONN and the well - known existing benchmark algorithms , namely, the Round Robin (RR) and the CQI - max algorithms), on the cell throughput criteria. Figure53 depicts the resulting the cell throughput versus SNR for all four different scheduling schemes. 95 Figure 53 Cell throughput for the four different scheduler schemes. It is noticeable that the RR scheduling performance is the worst in throughput, since it does not consider the user channel condition into account. The CQI_max scheduling achieves the highest over As depicted in Figure55, the new bRNN algorithm is the providing a trade - off between the throughput and a notion of fairness to all users. In this particular example bR NN algorithm provide better trade - off than SONN. As in figure54 its throughput performance is in the average scope. After normalizing the throughput performance among all different types of schedulers, we got the matric of figure 55. This matric will be ap plied for overall performance. -1 0 1 2 3 4 5 6 -15 -10 -5 0 5 10 15 20 25 30 35 THROUGHPUT [MB/S] SNR [DB] 12 UES, PEDB, SISO, RR,B_CQI, SONN, BRNN RR Best_CQI_Algorthim MSONN bRNN 96 Figure 54 Normalized cell throughput for the four different scheduler schemes . Figure 55 Fairness among users for the four different scheduler schemes -0.2 0 0.2 0.4 0.6 0.8 1 1.2 -15 -10 -5 0 5 10 15 20 25 30 35 THROUGHPUT [MB/S] SNR [DB] 12 UES, PEDB, SISO, RR,B_CQI, SONN, BRNN RR Best_CQI MSONN bRNN -0.2 0 0.2 0.4 0.6 0.8 1 1.2 -15 -10 -5 0 5 10 15 20 25 30 35 12 UES, PEDB, SISO, RR,B_CQI, SONN, BRNN RR Best_CQI_Algorthim MSONN bRNN 97 Toward a full comparison of all algorithms performance, we end up with matrix that represents the normalized sum of throughput matrices and fairness matrices in the evaluations. Figure56 depicts the overall weighted overall performance of all the algorithms. Fig ure 56 Combination (Fairness and Throughput) evaluation among different scheduler schemes. It is clear that the combination of the algorithms 4 and 5 makes the downlink performance more reliable to the end user UE. The individual users have been provided higher throughput than the RR scheduler is providing. All this, with giving a low Bit Error Rate BER. Toward showing the difference performance of bRNN scheduler over RR scheduler in terms of throughput and fairness full compariso n of all algorithms performance, we end up with matrix that represents the normalized sum of throughput matrices and fairness matrices in the evaluations. Figure 57 d epicts the overall weighted overall performance of all the algorithms. -0.5 0 0.5 1 1.5 2 -15 -10 -5 0 5 10 15 20 25 30 35 TOTAL MATRIC NORM THROUGHPUT+ FAIRNESS RR Best_CQI MSONN bRNN 98 Figure 57 (Fairness and Throughput) evaluation of bRNN scheduler schemes comparison It is observable from these figures the performance with bR NN scheduler is way better in terms of fai rness : The Block Error Rate of the cell performing the best among other scheduler. I t is even clear er on the Block Error Rate on the individual users UEs scale, and the UEs almost have very close similarity in terms of UE throughput performance . Th is type of scheduling techniques give the prom ising optimization between providing the highest Cell thro ughput and give fairness among UEs. This is type of schedule can be considered as real proportional f air scheduler. 99 Chapter 6 . Conclusion and Future Work 6.1. Conclusion: - Self - Organizing promises in wireless cellular communication networks have been reviewed in the previous publications. In this work, a categorization of the previous projects have been done from presenting a deep understanding of what these new functionalit ies of future networks are, and such important results achieved thus far have been pointed out. The principle tenets of utilizing SO algorithms in remote innovation have been allocated. Although a careful analysis shows that some solutions in the literatu re are classic adaptive algorithms, others possess necessary features (scalability, stability and agility) required in any SO solution. Both classification and a characterization framework have been presented for SO and used to discuss the state of the art literature with simple classifications of self - configuration, self - optimization, and self - healing. - Toward introducing a novel scheduling algorithm to improve the cell performance in LTE and LTE - A, Vienna LTE - A Downlink System - Level Simulator [37] has been operated to evaluate the performance of three already known and published algorithms Round Robin RR, Weighted Round Robin WRR and Max_C/I, then explore two novel scheduling algorithms that are introduced in this work with this Matlab toolbox . The performa nce has been evaluated through a spectrum of Signal to Noise Ratio SNR - This work elaborated on the downlink packet - scheduling framework in LTE presenting. Then, a novel Self - Organizing Neural Network (SO NN ) scheduling algorithm was provided, and compared in performance to the famous Round Robin RR scheduler and Max C/I algorithm. The Max C/I algorithm was not reliable on maintaining the QoS that LTE - A promises to provide since 100 it could provide the users wit assign all available resources to only one user in one sub - frame. But, despite the gains in accuracy, the SO NN proposed algorithms are more complicated than the Max C/I algorithm. However, bec ause they introduce compromise between fairness regarding resource distribution and prioritize the UEs, we propose they are the best channel to improve overall network performance. The simulation results prove that the proposed novel algorithms improve the overall cell throughput, both for PedB and VehB channel models (extensive simulation results for scenarios with 6,8 and 10 users have similar results and confirm the presented conclusions). With providing this SO NN algorithm, fair balancing of the resourc es in the cell is grantee. The SO NN scheduling mechanisms prevents that no user in the system is degraded or starved by supporter routine in the algorithm. The future developments could include a QoS metrics when making the scheduling decisions. - Within this thesis, a new approach of applying the Recurrent Neural Network has been introduced and applied toward provide prediction s . The dynamic system that RNN provid e s in some algorithms leads to following the pattern of the output performance beca use of the hidden unit in the RNN, all the s e enab le the full algorithm we apply in this system to give accurate prediction values. Here we experiment with the base Recurrent Neural Network RNN to predict by training the weights to warm up and then run by a pplying random values over the entire time. Output performance, state performance and error measurements evaluated the performance with evaluating curves and measurements showing how accurate predictions we developed . Even the gradient decent curves were p roviding normal behavior which is a great ind icator of the improved dynamic performance . 101 The bRNN scheduling algorithm we applied in the Resource Blocks RB scheduler provide the best performance ever according to our criteria as the best throughput curves and the best fairness curves among several SNR values was discovered by applying the novel bRNN scheduling algorithm. These results satisfy the promise that such smart e node base should provide. The future developments could include a QoS metrics when m aking the scheduling decisions. The primary goal was to provide the proportional fair to the scheduler. It is clear by providing the resource blocks RBs to the UEs who have maximum channel quality indicator. However, this causes too many failures, delays, and discontinuity. These issues can lead to instability of the system performance and its quality of services. This means proportional fair is the real lead, and Modified SOM technique and bRNN scheduler successfully covers all the history in terms of upd ates to plan the next plate of RBs and cluster them based in this history. 6.2. Future work: - Indeed, ev en though we obtained profoundly new refreshing results, we expect more work and improvements are possible in this type of project. The topic examined here are novel contributions to the developing literature, which is relatively lacking in deep and smart algorithms in the scheduling process. On the other side, focusing on new algorithms such as Recurrent Neural Networks has a lot of promise for more smart solutions. There are some points that need more research and analysis: - Future Step: we are planning to make the algorithm more adaptive towards mapping the Resource Blocks RB to UEs based on the clusters of the channel Quality Factor CQI. This will happen by using the type of data each UE is willing to use in the next TTI, Th e goal 102 here is to maximize the cell downlink throughput. Therefore, in the context of SOM, we are seeking the weights Wc that will provide the max throughput. - In future development process, these adaptive algorithms could be improved into more specific pa rameters; Some other information that comes with the feedback of the User Equipment, e.g., other QoS matrics. Other parameters that should be included into this D own Link. Other parameters and Rank Indicator (RI) could be very important, and their statues for each user should be effecting the priority in the mapping of the Resource Blocks values. After we reached the level of complexity that really touched reality, these parameters are the main gates to present our data in the grid. We can give each parameter a certain weight based on the environment, the need of the User, and target the network that are built to apply. - Working toward time of scheduling, The Samplin g frequency is important toward knowing how many runs we can do for the modified self - organizing Map algorithm at each TTI. This will be outside of the LTE - A downlink level Simulator, because this simulator is providing us with one Channel Quality Indicato rs CQI in one TTI. - Recurrent Neural Network RNN: RNN should be prepared for providing a close prediction to the pattern of Channel Quality Indicator CQI of the users, as well as it will be updated based on the real feedback value at each iteration which wi ll be done at each TTI. This work of RNN should be done after a warming up procedure toward RNN make the network adopting the pattern of the users by applying a number of iteration on previous TTIs. - The other Future work that should be started at this time: Self Organizing should be applied to the higher level performance in the network which is the link between the E - Node B and 103 back bone of the network and the connection between the E - Node B and the other E - Node B especially using the X2 Link that connects the nodes to each other. As such, work really need to be done to reach the Self - Organized Network. 104 R EFERENCES 105 REFERENCES [1] 3GPP TR 25.913 - v7.3.0, Requirements for EUTRA and EUTRAN, http://www.3gpp.org/ftp/Specs/archive/25%5Fseries/25.913/ [2] 3GPP TS 36.300 v8.0.0, E - UTRA and E - UTRAN Overall De scription, http://www.3gpp.org/ftp/Specs/archive/36%5Fseries/36.300/ [3] 3GPP TS 36.201 v1.0.0, LTE Physical Layer General Description, http://www.3gpp.org/ftp/Specs/archive/36%5Fseries/36.201/ [4] 3GPP TS 36.211 v1.0.0, Physical Channels and Modulation, http://www.3gpp.org/ftp/ Specs/archive/36%5Fseries/36.211/ [5] 3GPP TS 36.212 Multiplexing and Channel Coding, http://www.3gpp.org/ftp/Specs/archive/36%5Fseries/36.212/ [6] 3GPP TS 36.213 v1.0.0, P hysical Layer Procedures, http://www.3gpp.org/ftp/Specs/archive/36%5Fseries/36.213/ [7] 3GPP TS 36.214 v0.1.0, Physical Layer Measurements, http://www.3gpp.org/ftp/Specs/archive/36%5Fseries/36.214/ [8] 3GPP TS 36.300 v8.0.0, E - UTRA and E - UTRAN Overall Description; Stage 2, http://www.3gpp.org/ftp/Specs/archive/36%5Fseries/36.300/ Terrestrial Radio Access (E - December 2008. - Wireless Communications and Mobile Computing , vol. 7, [12] Socrates project homepage: www.fp7 - socrates.eu . [13] F. D. Rango, F. Veltri, M. Tropea, A. F. Santamaria, P. Fazio, A. Malfitano, and S. Marano, ion autonomic wireless networks Int. Journal of Mobile Network Design and Innoviation , vol. 2, pp. 3 4. i Proc. IEEE, Point of View Article , vol. 94, no. 11, pp. 1910 1911, 106 [15] A. - organisation in future mobile lectronics Communication Engineering Journal , vol. 12, no. 3, pp. 133 147, tworks: Computer Networks ,vol. 53, no. 9, pp. 1354 1376, - configuring and self - optimising network use cases and solutions (release 9). www.ngmn.org, Tech. Rep., - organisation in wireless networks use cases and Wireless World Research Forum Meeting 22 , - sensitive and IEEE Trans. Wireless Commun. , vol. 8, no. 2, pp. 776 786, [21] H. Zhu, T. Buot, R. Nagaike, and S. Harm Proc. 5th Int Wireless Personal Multimedia Communications Symp , vol. 3, [22] - tomata in sigmoid recursive neural netw 156, 2001. [23 ] O. M. Abusaid and F. M. Salem, "Network planning and dimensioning proce ss in WCDMA," 2012 International Conference on Wireless Co mmunications in Underground and Confined Areas, Clermont Ferrand, 2012, pp. 1 - 5. doi: 10.1109/ICWCUCA.2012.6402327 - configuration and self - optimization IEEE Commun. Mag. , vol. 48, no. 2, pp. 94 100, 2010. [25] T. Kamakaris, D. Kivanc - Proc. IEEE 18th Int. Symp. Personal, Indoor and Mobile Radio Communications PIMRC 1 5. e in relay IEEE Trans. Wireless Commun. , vol. 9, no. 3, pp. 1074 1085, 2010. IEEE Global Telecommunications Conference, 2004,(GLO , vol. 6, Nov. - 3rd Dec. 2004, pp. 3872 3876 . 107 IEEE 13th International Sym posium on Personal, Indoor and Mobile Radio Communications, [29] M. Amirijoo, L. Jorguseski, T. Kurner, R. Litjens, M. Neuland, L. C. Schmelz, and U. Proc. 6th Int. Symp. Wireless Communication Systems ISWCS [30] In Proc of 17th EUSIPCO, Glasgow, [31] in future 361. Communication Engin eering, 2015, pp. 40 44. Technology, Fig. 6. E node Base BER Electronics and Microelectronics (MIPRO), May 2016, pp. 595 600. [34] 39th International Convention on Information and Communication Technology, Electronics and Microelectronics ( MIPRO), May 2016, pp. 595 600. no. release 8. - robin resource sharing algorithm for device - to - device multic International Conference on Telecommunications (ICT), May 2014, pp. 191 195. [37] Vienna LTE - A Downlink System - Level Simulator. [Online]. Available: https://www.nt.tuwien.ac.at/rese arch/mobile - communications/ vienna - lte - a - simulators// [38] F. M. Salem, A Basic Recurrent Neural Network Model. 29 Dec 2016 . Arxiv: #1612.09022://Https://arxiv.org/abs/1612.09022. [39] Neural Network - Dept of ECE, Memo. 2017/2. [40 ] O. M. Abusaid and F. M. Salem, "Kullback - Leibler Divergen ce minimization for C ompetitive learning of self - organizing maps," 2017 International Conferenc e on Engineering 108 and Technology (ICET) ,Antalya,2017,pp.1 - 6.doi: 10.1109/ICEngTechnol.2017.8308205 [41 ] O. M. Abusaid and F. M. Salem, "Self - organizing techniques to resource block scheduling in LTE - A communication systems," 2017 International Conferenc e on Engineering and Technology (ICET) ,Antalya,2017,pp.1 - 6.doi: 10.1109/ICEngTechnol.2017.8308183N. [42] T. Vahid, N. Seshadri, and A. Robert Calderbank. "Spac e - time codes for high data rate wireless communication: Performance criterion and code construction." IEEE transactions on information theory 44.2 (1998): 744 - 765. JECE, vol. 2010, pp. 1:1 1:10, Jan. 2010. OFDMA wireless E EE Commun. Surveys Tutorials , vol. 14, pp. 322 337, quarter 2012. [45] link scheduling: Flow Multiple Access Communications (C. Sacchi, B. Bellalta, A. Vinel, C. Schlegel, F. Granelli, and Y. Zhang, eds.), vol. 6886 of Lecture Notes in Computer Science , pp. 181 192, Springer Berlin / Heidelberg, 2011. [4 6 ] E. Yaacoub, H. Al - algorithms for the LTE IEEE Symp. Comput. Commun. , pp. 266 270, July 2009. [47 ] N. Ali, A. - E. Taha, and H. Hassanein, LTE, LTE - Advanced and WiMAX :Towards IMT - Advance d Networks. ITPro collection, Wiley, 2011. [ 4 8] M. Sawahashi, Y. Kishiyama, A. Morimoto, D. Nishikawa, and multipoint tr a nsmission/reception techniques for LTE - IEEE Wireless Commun. , vol. 17, pp. 26 34, June 2010. [ 4 9] Marsch and G. Fettweis, Coordinated Multi - Point in Mobile Commu - nications: From Theory to Practice . Coordinated Multi - point in Mobile Communications: From Theory to Prac tice, Cambridge University Press,2011. [50 hanced physical downlink control channel in LTE - IEEE Commun. Mag. ,vol.51, pp. 82 89, Feb. 2013. [51 - UTRA) and Evolved Universal Terrestrial Radio Access Network (E - UTRAN); Over TS 36.300, 3rd Generation Partnership Project (3GPP), Jan. 2013. [52 - UTRA) and Evolved Universal 109 Terrestrial Radio Access Network (E - UTRAN); TS 36.300, 3rd Generation Partnership Project (3GPP). [53 - UTRA); Car - rier Aggregation Enhancements; UE and BS radio transmission and reception (Release 1 , 3r d Generation Partnership Project (3GPP), Sept. 2012. [54 packet scheduling in lte cellular networks: Key design issues and a IEEE Commun. Surveys Tutorial s , vol. 15, no. 2, pp. 678 700, 2013. [55 ] L. Ruiz de Temino, G. Berardinelli, S. Frattasi, and P. Mogensen, - Aware Scheduling Algorithms for SC - in IEEE 19th Int. Symp. Personal, Indoor and Mobile Radio Commun., pp. 1 6, 15 - 18 2008. [56 ] F. Calabrese, P. Michaelsen, C. Rosa, M. Anas, C. Castellanos, D. Villa, K. Pedersen, and - Tree Based Uplink Channel Aware Packet Scheduling for UTRAN IEEE Veh. Technol. Conf. , pp. 1949 1953, 11 - 14 2008. [57 - domain packet scheduling for 3GPP INFOCOM , pp. 1 9, Mar. 2010. [58 ] F. Calabrese, C. Rosa, M. Anas, P. Michaelsen, K. Pedersen, and P. Mo - Transmission Bandwidth Based Packet Scheduling for IEEE 68th Veh. Technol. Conf. , pp. 1 5, 21 - 24 2008. [59 - tional fairness in LTE IEEE Wireless Commun. Netw. Conf. , pp. 1 5, Apr. 2009. [ 6 0 OFDMA uplink Int. J. Electronics and Commun. , vol. 65, no. 4, pp. 320 330, 2011 [61 ] P. Wen, M. Y - cation over wideband , pp. 1 7, Oct. 2 007. [62 ] M. Anas, C. Rosa, F. Calabrese, K. Pedersen, and P. Mogensen, ined Admission Control and Scheduling for QoS Differentiation IEEE 68th Veh. Technol. Conf., 2008 , pp. 1 5, Sept. 2008. [63 QoS assurance with revenue maxi mization of LTE uplink 2011 9th Annu. Commun. Netw. and Services Research Conf. (CNSR), pp. 202 209, May 2011. 110 [64 Multiclass Services IEEE 6th Int. Conf. Wireless and Mobile Computing, Netw. and Commun. (WiMob) , pp. 355 360, Oct. 2010. [65 ] Y. Zaki, T. Weerawardane, C. Gorg, and A. Timm - - QoS - aware fair scheduling IEEE 73rd Veh. Technol. Conf. (VTC Spring) , pp. 1 5, May 2011. [66 ] M. Al - Uplink Scheduling for 3G The 4th Int. Conf. on Innovations in Information Technology, pp. 705 709, 18 - 20 2007. [67 urce Allocation in LTE Uplink with a 8th Annu. Commun.Networks and Services Research Conf. (CNSR, pp. 268 275, May 2010. [68 - Bounded Power - Efficient Packet Scheduling for Uplink Systems of LTE, The 5th Int. Conf. on Wireless Communications, Networking and Mobile Computing , pp. 1 4, Sept. 2009. [69 for Single - Carrier Wireless Conf. (EW), 2010 European , pp. 339 345, Apr. 2010. [70 Transmission Energy 2011 IEEE Veh. Technol. Conf. (VTC Fall) , pp. 1 5, Sept. 2011. [ 7 1 ] D. Jiang, Q. Wang, J. Liu, G. Liu, and C. multi - point reception for LTE - 5th Int. Conf. on Wireless Communications, Netwo rking and Mobile Computing , pp. 1 4,Sept. 2009. [72 joint transmission coordinated multi - point in LTE - Int. ITG W orkshop on Smart Antennas (WSA), pp. 177 184, Feb. 2010. [73 to joint resource allocation in uplink CoMP sys The 8th Int. Symp. Wireless Commun. Systems ( ISWCS) , pp. 740 744, Nov. 2011. [74 Selection for LTE - Advanced IEEE Int. Conf. Commun. (ICC) , pp. 1 5, June 2011. [75 - to - Machine communications in LTE - IEEE GLOBECOM Workshops , pp. 353 357, Dec. 2011. [76 ] S. - Y. Lien, K. - ubiquitous massive accesses in 3GPP machine - to - IEEE Commun. Mag. ,vol. 49, pp. 66 74, Apr. 2011. 111 [77 ] S. - Y. Lien and K. - - antees in 3GPP machine - to - IEEE Co mmun. Lett. , vol. 15, pp. 311 313, Mar. 2011. [78 Challenges and New IEEE Veh. Technol. Mag. ,vol.7,pp. 34 39, Sept. 2012. [79 resource control scheme for machine - to - machine Int. Conf. on ICT Convergence (ICTC) , pp. 1 6, Oct. 2012. [80 - UTRA); Physical Channels and 6.211, 3rd Generation Partnership Project (3GPP), Mar. 2010. [81 - UTRA); Radio Frequency (RF) R 36.942, 3rd Generation Partner - ship Project (3GPP), Apr. 2010. [82 Partnership Project (3GPP), June 2010. [83 ] F. Calabrese, Adaptive Radio Resource Management for Uplink Wireless Networks. Phd Thesis, Polytechnic University, June 2006. [84 ] Orange, China Mobile, KPN, NTT DoCoMo, Sprint, T - Mobile, Voda - fone, Telecom Italia, - - 070674, 3rd G eneration Partnership Project (3GPP), Feb. 2007. [85 Nature , vol. 521, no. 7553, pp. 436 444, May 2015. [86 Neural Netw. , vol. 61, pp. 85 117, Jan. 2015. [87 - temporal represen - tations for action IEEE Trans. Cybern. , vol. 46, no. 1, pp. 158 170, Jan. 2016. [88 ] A. - A. Liu, Y. - T. Su, W. - - ing multi - task learning for joint human act IEEE Trans. Pattern Anal. Mach. Intell. , vol. 39, no. 1, pp. 102 114,Jan. 2017. [89 actions using FV IEEE Trans. Cybern. , vol. 46, no. 12, pp. 2978 2990, Dec. 2016. [90 ] A. Karpathy. (2015). The unreasonable effectiveness of recurrent neural networks. Andrej Karpathy Blog. [Online]. Available: http://karpathy.github.io/2015/05/21/rnn - effectiveness/ 112 [91 ] X. Peng, L. Wang, X. Wang, and methods for Comput. Vis. Image Understand. , vol. 150, pp. 109 125, Sep. 2016. [92 ] A. - A. Liu, Y. - T. Su, P. - P. Jia, Z. Gao, T. Hao, and Z. - X. Yang, - view human action recognition via part - induced mul - , vol. 45, no. 6, pp. 1194 1208, Jun. 2015. [93 - view action recognition by cross - domain Image Vis. Comput. , vol. 55, pp. 109 118, Nov. 2016. [94 model for Appl. Soft Comput. , vol. 18, pp. 178 184, May 2014. [95 ] R. R. Reddy, Y. Ramad function for Proc. Int. Conf. Adv. Comput., Commun. Inform. (ICACCI) , Sep. 2016, pp. 1148 1153. [96 tion sys - tem based on KNN J. Elect. Comput. Eng. , vol. 2014, Jun. 2014, Art. no. 240217. [97 - KDD dataset using A Proc. Int. Conf. Signal Process. Commun. Eng. Syst. , Jan. 2015, pp. 92 96. [98 intrusion detection , vol. 89, pp. 213 217, Jan. 2016. [99 ] J. Zhang, M. Zulkernine, and A. - forests - based network intrusion IEEE Trans. Syst., Man, Cybern. C, Appl.Rev. , vol. 38, no. 5, pp. 649 659, Sep. 2008. [100 classification techni Int. J. Sci. Res. Sci., Eng. Technol. , vol. 2, no. 5, pp. 202 208, 2016. [101] learning methods for c IEEE Commun.Surveys Tuts. , vol. 18, no. 2 , pp. 1153 1176, 2nd Quart., 2016. [102 network Bio - inspired Inf. Commun. Technol. (BIONETICS), New York, NY, USA , May 2016, pp. 21 26. [103 ] T. A. Tang, L. Mhamdi, D. McLernon, S. A. R. Zaidi, and M. Ghogho, 113 approach for network intrusion detection in soft - Proc. Int. Conf. Wireless Netw. MobileCommun. (WINCOM) , Oct. 2 016, pp. 258 263. [104 reduced - size RNN based 1190, Sep. 2012. [105 ] M. Tavallaee, E. Bagheri, W. Lu, and A. A. A. analysis of the KDD Proc. IEEE Symp. Comput. Intell. Secur. Defense Appl. , Jul. 2009, pp. 1 6. [106 - KDD dataset using various machine learning Int. J. Eng.Res. Technol. , vol. 2, pp. 1848 1853, Dec. 2013. [107 - processing influence on intrusion detection using NSL - Proc. Open Conf. Elec t., Electron. Inf.Sci. (eStream) Apr. 2017, pp. 1 5. [108 system for NSL - KDD data set using vectorised fitness function in genetic Adv. Comput. Sci. Technol., vol . 10, no. 2, pp. 235 246, 2017. [109 ] R. A. R. Ashfaq, X. - Z. Wang, J. Z. Huang, H. Abbas, and Y. - L. He, semi - supervised learning approach for intrusion detec - Inf. Sci. , vol. 378, pp. 484 497, Feb. 2017. [110 ] J. Marten hessian - free Mach. Learn., Bellevue, WA, USA, Jul. 2011, pp. 1033 1040. [111 ] A. Sperduti, D. Majidi, and A. Starita, - correlation for syntactic and Advances in Struc - tural and Syntactical Pattern Recognition , P. Perner, P. Wang, and A. Rosenfeld, Eds. New York: Springer - Verlag, 1996, vol. 1121, Lecture Notes in Co mputer Science, pp. 90 99. [112 - dependent distributed struc - ture - representations Proc. IEEE Int. Conf. Neural Networks , pp. 347 352, 1996. [113 ] A. Sperduti and A. - cation of IEEE Trans. Neural Networks , vol. 8, pp. 714 735, May 1997. [114 processing of IEEE Trans. Neural Networks , vol. 9, p. 768, Sept. 1998. 114 [115 - guage, Univ. California at San Diego, La Jolla, CRL Tech. Rep. 8801, 1988. [116 recurrent neural networks Neural Networks , vol. 10, no. 3, pp. 395 400, 1997. [117 of data structures by Proc. Int. Joint Conf. Artificial Intel ligence , pp. 1066 1071, 1997. [ 118 Math. Contr. Signals Syst. , vol. 12, pp. 62 79, 1999. [119 networks: A theoretical compa rison between structured and flat repre - Proc. Eur. Conf. Artificial Intelligence , pp. 301 305, 200. [120 IEEE Trans. Knowledge Data Eng. , vol. 13, no. 2, pp. 196 206, 2001. [ 121 - Control Heuristics for Dept. Computer Science, Tech. Univ. Munich, Munich, Germany, 1997. [122] E. Francesconi, P. Frasconi, M. Gori, S. Marinai, J. Q. Sheng, G. Soda, and A. Sperduti, Graphics Recognition Algorithms and Systems , R. Kasturi and K. Tombre, Eds. New York: Springer - Verlag, 1997, vol. 1389, Lecture Notes in Computer Science, pp. 104 117. [1 23 - ical pattern Proc. Int. Conf. Advances Pattern Recognit ion , Plymouth, UK, Nov. 1998, pp. 425 432.