Open Access

As a reply to the increasing demand for fast mobile network connections, the concept of Self-Organizing Networks (SONs) has been developed, reducing the need for humans to execute Operation, Administration and Maintenance (OAM) tasks for mobile networks. A SON consists of several autonomously operating closed control loops, so-called SON functions, which can influence the behavior of the mobile network by adapting their input parameters. With SON management, a multitude of simultaneously operating SON functions can be configured according to network context-specific and weighted targets for Key Performance Indicators (KPIs), denominated as technical objectives, by using different input models. These are the operator-defined context and objective model, and the SON function manufacturer-provided effect model. SON management facilitates the SON-enabled system to work optimally regarding the achievement of defined KPI targets. Since Mobile Network Operators (MNOs) have to fulfill rising mobile network performance demands while reducing costs at the same time, it is crucial for SON management to gain an understanding of the network behavior to allow a cost-neutral performance improvement while simultaneously reducing the risk of network misconfiguration and service disturbance. This thesis introduces four different SON management approaches, all using the same type of input models and allowing to automatically configure SON functions according to these input models. These approaches, namely Policy-based SON Management (PBSM), Objective-driven SON Management (ODSM), Adaptive SON Management (ASM) and Cognitive SON Management (CSM), thereby represents different stages of development, i.e., they build on one another and each of them overcomes the disadvantages of its predecessor. The PBSM approach is presented which first enables the management of a system at a high level of abstraction and, at the same time, reduces manual effort. Operator and SON function manufacturer knowledge is represented in a structured and automatically processable form for the first time and an objective manager is presented that performs a reasoning process to map this knowledge to an optimal parameter configuration for each individual SON function. However, the simplicity of the models in PBSM and the computational complexity may limit the applicability of the concept, the reason why PBSM is further developed to ODSM. In this approach, more expressive input models are used and SON functions which usually influence each other, are considered as a union, allowing for a trade-off between them. In both approaches, the manufacturer-provided effect models are static and do not adapt to the actual network environment. This may lead to non-optimal operation of the SON system and hence to non-optimal network performance. In ASM, the actual influence of the current combined SON function configuration on the network performance is determined by analyzing KPI measurements from the network, and the effect models are enhanced in such way that the contribution of the corresponding SON functions towards achieving the technical objectives is improved. In the most sophisticated approach, the CSM, ASM is equipped with machine learning capabilities. The behavior of SON functions in the network is analyzed using four different algorithms in the field of supervised learning in order to predict their effects under untested parameter configurations. Also, performance data of network cells are analyzed for similarities using techniques in the field of unsupervised learning. That is, machine learning is applied to complement the sketchy effect models, giving the CSM system a wider range of possible configurations. The four different stages are evaluated in a realistic mobile network simulator to show the value of SON management in general and the performance improvement to previous stages of development. While these approaches provide an increasing level of maturity from PBSM to CSM, they all are designed in a way that the MNO has full control over the mobile network at any time and that he or she can interrupt automated actions at any time.

As a reply to the increasing demand for fast mobile network connections the concept of Self-Organising Networks (SONs) has been developed, reducing the need for humans to execute Operation, Administration and Maintenance (OAM) tasks for mobile networks. However, a SON contains functions which are provided by different vendors as black boxes, making it hard to predict the performance of the network, especially under untested configurations. Since Mobile Network Operators (MNOs) have to fulfil rising mobile network performance demands while reducing costs at the same time, it is crucial to gain a better understanding of the network behaviour to allow a costneutral performance improvement while simultaneously reducing the risk of network misconfiguration and service disturbance. In this paper an approach is introduced to enhance SON Management models with cognitive Machine Learning (ML) methods. Therefore, the simulated behaviour of three different SON Functions is analysed and described by a Linear Regression (LR) Model. In a second step, performance data of network cells are analysed for similarities using k-Means Clustering. The findings of these two steps are then combined by fitting the models onto smaller clusters of cells. Finally, the utility of these models for predicting the performance of the network is evaluated and the different stages of refinement are compared with each other.