Project topics and materials
Get Complete Project Material File(s) Now! »
Policy Based SON Management
The integrated SON management is, as explained earlier, constituted by a set of functionalities that collaborate in order to automate and optimize the network operation. In this thesis, we focus on the PBSM entity. As mentioned previously, the PBSM is the central entity of the Integrated SON Management framework. Its task it to autonomously translate the operator’s KPI targets, that reflect the high level operator objectives, into proper SCV sets. In our work, we consider that the SON functions are provided by SON vendors as black boxes, i.e. the operator does not know the exact algorithm running in each SON function, due to patents and proprietary aspects of the algorithm developed by the SON vendors. The operator can however steer the behavior of these functions by changing their SCV sets, according to its objectives (figure2.2 .Several approaches have already been studied to automate the PBSM process [22–25]. These approaches consider the SON functions to be black boxes as well. They rely on input models, provided by the SON vendor or generated by the operator, to derive the SCV sets policy. The considered input models are the following:
• Objective Model: It describes the objectives of the operator. It can be formulated in different forms. In [22] for instance, it takes the form of « IF condition THEN KPI target WITH priority ». The condition can be the time of the day and the cell location for example. The KPI target defines the KPI that should be optimized, for example minimizing drop call rate, or maximizing handover success rate. The priority represents the importance of the corresponding KPI target for the operator. In [24] and [25], KPI targets are not only max and min requirements, but can also be intervals e.g handover success rate > 95%. And priorities are replaced with weights, allowing a weighted satisfaction of KPI targets.
• Context Model: Taking the cell context into consideration is very important because it im12 pacts the behavior of SON functions instances. The context model provides a description of the system, the cell’s properties, and eventually derives a cell class definition. These properties can be compared with the conditions stated in the objective model in order to derive the policy. Cell context can include time (since it is strongly related to traffic), cell location, network topology, etc.
• SON Function Model (SFM): It is a mapping, for each SON function, from KPI targets to SCV sets. SFMs are considered to be provided by the SON vendors with the SON functions. They are typically generated in a simulation environment or over a network test cluster where the SCV sets for each SON function are tested and output KPIs are determined. In [23], the authors describe a generation and testing process for SFMs, where the operator simulates the network with different SCV sets for each SON function. Several approaches are possible. The first approach would be to run all the SON instances of each SON function deployed in a certain network section. A second approach would be to define cell clusters with similar network contexts and test the SCV sets only on the SON instances deployed in these clusters. A third approach would be to evaluate a SFM for each and every SON instance.
The PBSM operates by comparing the objective model with the SFMs mappings and the context model, and outputs optimal SCV sets to be enforced in the corresponding SON functions. This approach was extended to include network KPI measurements [25]. The idea is to gradually replace the static SFMs provided by the vendors by real network measurements once the network measurements pool collected from the network becomes big and reliable enough. However, the process is still strongly relying on static input models, which often do not represent exactly the actual network where the PBSM is operating, nor the dynamics of the environment. Besides, SFMs are considered to be provided by the SON vendors. This consideration is rather problematic for the operator in the case of trust issues with the vendor, and because the SON vendor is most probably willing to provide these SFMs with extra charges, especially that these models will require probably constant updates and improvements. Alternatively, if the operator chooses to generate itself the SFMs, then it would require a lot of simulations and time effort to simulate and test the SCV sets. Furthermore, pre-simluated models are generated for each SON function individually.
They hence do not depict well enough the interactions between the different SON functions when deployed simultaneously in the network. In the following section, we motivate the need to improve the state of the art PBSM to meet with the requirements of future networks and we discuss how the PBSM can be enhanced with cognitive capabilities.
Physical Layer
In our studies, we consider only downlink transmissions and the scheme used in LTE downlink is the Orthogonal Frequency-Division Multiple Access (OFDMA). In OFDMA, the original bandwidth is subdivided into multiple subcarriers. Each of these subcarriers can then be individually modulated. Typically in OFDMA systems we can have hundreds of subcarriers with a spacing between them (15KHz on the LTE case). Different numbers of subcarriers can be assigned to different users, thereby providing the flexibility to support differentiated QoS. Besides, by transmitting over multiple subcarriers in parallel, OFDMA has the ability to cope with severe channel conditions, such as interference and frequency selective fading due to multipath. Also, parallel transmission reduces the symbol rate over each subcarrier, permitting the use of cyclic-prefix, which provides a guard interval to eliminate intersymbol interference from the previous symbol and allows channel estimation and equalization. OFDMA brings hence robustness and flexibility for the E-UTRAN. The chain to generate an OFDM signal starts by parallelizing the symbols that need to be transmitted, after they are modulated (in LTE the modulation can be QPSK, 16AQM, 64QAM). Then they are used as input bands for an Inverse Fast Fourier Transform (IFFT) operation. This operation produces OFDM symbols, which will be transmitted. Notice that a conversion from the frequency to the time domain was made when the IFFT was used. On the receiving side of the OFDMA system we should expect a Fast Fourier Transform (FFT) operation that will then convert the symbol to the frequency domain again.
SON Functions Description and Algorithms
As stated in Chapter 2, SON functions are categorized into self-configuration, self-optimization and self-healing functions. In this thesis we focus on self-optimization SON functions. Although 3GPP specifies the input, output and targets of the SON functions, the algorithm inside the SON are not standardized, and they are specific to each vendor. For the use cases we consider in this work, we are particularity interested in the following SON functions
Table of contents :
Contents
List of Figures
List of Tables
List of Acronyms
1 Introduction
2 Cognitive Policy Based SON Management
2.1 Introduction to Self-Organizing Networks
2.2 Integrated SON Management
2.2.1 A Global View
2.2.2 Policy Based SON Management
2.3 From Autonomic Management to Cognitive Management
2.3.1 The Motivation Behind Introducing Cognition
2.3.2 Reinforcement Learning
2.3.3 Reinforcement Learning via Multi-Armed Bandits
2.3.4 The Cognitive PBSM
2.3.5 Contribution of the Thesis
3 System Model
3.1 Introduction
3.2 LTE-A Systems
3.2.1 Introduction
3.2.2 System Architecture
3.3 E-UTRAN
3.3.1 Physical Layer
3.3.2 Mobility
3.3.3 Interference and Almost Blank Subframes
3.4 SON Functions Description and Algorithms
3.5 LTA-A System Level Simulator
3.5.1 Channel Model
3.5.2 SON Management Framework
3.5.3 Simulator Block Diagram
4 Multi Armed Bandit for Cognitive Policy Based SON Management
4.1 Introduction
4.2 C-PBSM Based on Stochastic MAB
4.2.1 Problem Statement
4.2.2 Stochastic Multi-Armed Bandit
4.2.3 Scenario Description
4.2.4 Simulation Results
4.3 C-PBSM Based on Linear MAB
4.3.1 Linear UCB for C-PBSM
4.3.2 Scenario Description
4.3.3 Simulation Results
5 Softwarized and Distributed Learning for Cognitive SON Management
5.1 Introduction
5.2 Distributed Learning Under a SDN Framework
5.3 Software Defined Networks for SON Management
5.4 Scenario Description
5.5 Simulation Results
6 Context Aware Cognitive Policy Based SON Management
6.1 Introduction
6.2 Context Aware C-PBSM
6.3 Context Aware C-PBSM: Implementation Based on Contextual Multi-Armed Bandit
6.3.1 Contextual MAB
6.3.2 Random Forest for the Contextual MAB
6.3.2.1 Variable Selection
6.3.2.2 Action Selection
6.3.2.3 BF Algorithm
6.4 Scenario Description
6.5 Simulation Settings and Results
7 Conclusion and Perspectives
Bibliography
