Cognitive Autonomous Networks (CAN) promise to advance Self Organizing Networks (SON) by applying artificial intelligence to significantly raise the degree of automation in mobile networks. In CAN, Cognitive Functions (CFs) learn the optimal configuration parameter values to optimize specific network metrics, with the execution coordinated via a controller. In open, multi-vendor systems however, the CF’s learning ability may raise a new risk: a manipulative CF (MCF) may learn not only its objective, but also to manipulate the coordination system in pursuit of that objective. In this paper we propose and evaluate our proposed functionality, called CoDeRa, that neutralizes manipulative CF behavior. However, although CoDeRa is effective against MCFs, it is inadequate to resolving error propagation in CF coordination, caused by corrupted network data, for which we have proposed an alternate simpler and cost efficient network management architecture. Our evaluation shows that the proposed design is robust against the observed concerns, and in context of ongoing worldwide standardization efforts, we summarize the relevance and implications of our proposed architecture.

Deep learning plays a crucial role in enabling cognitive automation for the mobile networks of the future. Deep clustering – a subset of deep learning – is a valuable tool for many network automation use cases. Unfortunately, most state-of-the-art clustering algorithms target image datasets, which makes them hard to apply to mobile network automation due to their highly tuned nature and assumptions about the data. In this paper, we propose a new algorithm, Decorrelating Adversarial Nets for Clustering-friendly Encoding (DANCE), intended to be a reliable deep clustering method for mobile network automation use cases. DANCE uses a reconstructive clustering approach, separating clustering-relevant from clustering-irrelevant features in a latent representation. This separation removes unnecessary information from the clustering, increasing consistency and peak performance. We comprehensively evaluate DANCE and other select state-of-the-art deep clustering algorithms, and show that DANCE outperforms these algorithms by a significant margin in a mobile user behavior clustering task based on data gained from a simulated scenario.

Neural-net-based deep learning algorithms are starting to be utilized in many network functions. Deep neural nets are traditionally not resistant against missing or corrupted inputs, a scenario which is likely to happen in mobile networks. If the data corruption does not stem from malicious intent, the task of reconstructing missing inputs is called imputation. In this paper, we discuss how such imputation methods could be utilized in network functions, to make the network robust against non-adversarial data corruption. We propose an integrated approach, where the imputation is undertaken by the same model which implements the machine learning task in the network function. We evaluate state-of-the-art imputation methods and our integrated imputation thoroughly, using data generated in a mobile network simulator. Our results show excellent performance with the integrated imputation, but also raises some questions with regards to how deep-learning-based network functions should be used in such scenarios.

The incorporation of Artificial Intelligence (AI) and Machine Learning (ML) in mobile networks is expected to raise the degree of automation by proposing Cognitive Autonomous Networks (CAN). In CAN, learning based functions, called Cognitive Functions (CFs), adjust network control parameters to optimize specific Key Performance Indicators (KPIs). The CFs share the same resources, and this very often introduces an overlap among their target control parameter adjustment, i.e., at one point of time, multiple CFs may want to change the same control parameter albeit by different amounts depending on their respective levels of interest in that parameter. Correspondingly, a Controller is required in CAN to coordinate the sharing of the parameter among the independent CFs to meet their varying extents of interests. Although a Nash Social Welfare Function (NSWF) based Controller was introduced at first, to overcome the problems of this Controller a second Controller was introduced based on Eisenberg-Gale Solution (EGS). To use an EGS based Controller, impact of each network control parameter on each CF, called Config-Weight (CW), needs to be calculated. In this paper we propose a Shapley value based method for CW calculation, prove the optimality of the method mathematically and by simulation, provide a comparison between the Controllers in a simulation environment that resembles 5G network and find that up to 9.18% improvement can be obtained using the EGS based Controller.

Introduction of artificial intelligence is expected to raise the degree of automation in mobile networks by succeeding Self Organizing Networks (SON) with Cognitive Autonomous Networks (CAN). In CAN, learning based Cognitive Functions (CFs) work on different network parameters to optimize specific Key Performance Indicators (KPIs). However, learning ability of a CF poses a serious threat to the stability of the system. A manipulative CF (like a rogue agent) may use its learning capabilities to understand the working procedure of the system and manipulate it to achieve its own objective. Existence of such a CF can cause severe performance degradation of the overall system. In this paper we propose a simple yet effective machine learning based approach to detect manipulative CF(s) in CAN. We evaluate the performance of our proposed solution in a simulation environment that closely resembles a real life 5G scenario and provide analysis of the results with necessary precautions to be taken in a multi vendor scenario.

Intent Based Networks (IBNs) are mainly used to transform a user's intent into network configuration, operation, and maintenance strategies. In this paper we propose an a generic end-to-end design of an intent based network management system which we further customize to use in RAN control parameter and KPI management utilizing an existing technology (Cognitive Autonomous Network (CAN)). We introduce three new concepts in intent based network management: intent specification platform (ISP), formal intent and Intent Fulfillment System (IFS), which are not only relevant for this specific use case but can also be used by other network components. We discuss how our proposed solution can be implemented in Python as a standalone module so that it can be used with suitable networking simulators. Along with these, we also provide an overview of standardization impact of our research to show that it conforms with the worldwide mobile network management standardization efforts.

Due to the diversity of mobile services and rising user expectations, mobile network management has changed its focus from Quality of Service (QoS) to Quality of Experience (QoE). As a consequence, classical network optimization procedures must be updated accordingly. One of these optimization procedures is Mobility Robustness Optimization (MRO), whose aim is to improve HandOver (HO) performance by reducing HO failures. In this work, a novel QoE-aware MRO algorithm is proposed considering a multi-service scenario. Unlike previous approaches, whose aim is to increase successful handover rates, the optimization aim in this work is two-folded: to improve cell edge QoE while improving successful handover rates in the whole network. For this purpose, the handover trigger point, defined by the pair of HO control parameters HO margin and Time to Trigger, are tuned on a per-adjacency basis according to QoE and HO failure measurements. Method assessment is based on a dynamic system-level simulator implementing a realistic LTE scenario with multiple services. Results show that the proposed QoE-aware MRO algorithm improves cell edge QoE throughout the network while increasing the percentage of successful handovers compared to traditional approaches.

Cognitive Autonomous Networks (CANs) promise to overcome the shortcomings of current Self-Organizing Network (SON) implementations, i.e., the limited flexibility and adaptability to changing environments, by applying cognition. In CAN, intelligent network automation functions, herein called Cognitive Functions (CFs), apply machine learning techniques to learn context-specific behavioral policies with which to automate network operations. For proper operation, the CAN system needs to learn the environment in which the functions are operating and to abstract the environment and performance observations into states to which the CFs must respond. This paper proposes a design and implementation of an Environmental-state Modeling and Abstraction (EMA) engine that could be tasked to learn the required abstract states in a consistent way across multiple CFs.

Ultra-Dense Networks (UDNs) are Heterogeneous Networks (HetNets) that deploy a high density of small cells over-laying the traditional macro cells. If several Long Term Evolution (LTE) layers share the available spectrum, assigning the Physical Cell Identities (PCIs) becomes complicated due to the density and the diversity of the network. Since different layers can be managed by different Network Management (NM) and Self-Organizing Network (SON) solutions, it would often be desirable to be able to assign the PCIs in each layer independently. At the same time it must be ensured that the PCI conflicts, also between the layers, are minimized. When the small cell layer is managed independently, high cell density increases the probability that two small cells sharing the same PCI are neighbors to the same macro cell, thereby creating a conflict in inter-layer adjacencies, even when within the layers PCI conflicts are avoided. We propose a method that can minimize these inter-layer conflicts, while still allowing independent assignment between the layers. Starting with an initial intelligent guess for adequate PCI reuse distance, the solution then uses the Automatic Neighbor Relation (ANR) function to learn the full multi-layer network topology and optimize the PCI assignment. Comparing with state of the art strategies, our results show that the proposed approach maintains good performance without requiring the exchange of information across layers.

The high demand for frequency resources due to the ever increasing number of devices in cellular networks requires new approaches that maximize the spectral efficiency. Spectrum sharing is seen as an enabler to improve the overall spectral efficiency. Device to device (D2D) communication as an underlaying network to the cellular networks presents spectral efficiency improvements through the increased sharing of the cellular spectrum. In this paper, we analyze cell spectral efficiency and propose a uniform interference power (UIP) resource allocation scheme that reuses the cellular uplink spectrum resources for D2D communication. Using the base station (BS) to control the transmit power of all user terminals (UTs), the UIP scheme ensures that each of the D2D users reusing cellular resources contributes the same interference power at the BS without negatively impacting the cellular users. Our results show zero outage probability of the cellular user while the cell spectral efficiency is enhanced by up to 20 times when compared with the conventional cellular communication mode where no resource reuse within the cell is allowed.

There is a major demand for reducing energy consumption in mobile networks and it is expected become even more vital in the future (5G) multi-layer Ultra Dense Networks (UDNs), in which the number and density of cells in the different layers will grow dramatically. In these networks, multiple geographically overlapping layers are deployed to increase the capacity and throughput, but also increasing the energy consumption. In this paper we present an end-to-end solution that manages energy saving mechanisms in order to scale the provided capacity to the traffic. Assuming a Heterogeneous Network (HetNet) deployment, the solution dynamically selects cells to activate and/or deactivate considering the prevailing network load and the expected spectral efficiency of those cells. Evaluation in a small HetNet scenario showed that the proposed solution is able to reduce the energy consumption by more than 30%.

The development of 5G cellular networks is driven by several architectural and radio technology evolutions; and by requirements of diverse uses cases from the vertical industries that will be supported. Combined, these drivers demand for new approaches and processes to be applied towards the automation of Network Management (NM) tasks so as to advance operability levels for mobile network operators (MNOs). This paper analyzes the limitations of legacy Self-Organizing Networks (SON) introduced with 4G for Network Management Automation (NMA). In particular, we highlight the new challenges emerging from these 5G features and present several technological enablers that can be leveraged for 5G NMA. We include a future NMA framework combining these enablers, that addresses the NMA challenges by: 1) enhancing the level of intelligence in network elements, 2) allowing the networks to better process and analyze various kinds of network information, 3) enabling identification of operational context and scenarios, and 4) leveraging the building and sharing of NM knowledge across different domains and networks.

Self-organizing networks (SON) aim at simplifying network management (NM) and optimizing network capital and operational expenditure through automation. Most SON functions (SFs) are rule-based control structures, which evaluate metrics and decide actions based on a set of rules. These rigid structures are, however, very complex to design since rules must be derived for each SF in each possible scenario. In practice, rules only support generic behavior, which cannot respond to the specific scenarios in each network or cell. Moreover, SON coordination becomes very complicated with such varied control structures. In this paper, we propose to advance SON toward cognitive cellular networks (CCN) by adding cognition that enables the SFs to independently learn the required optimal configurations. We propose a generalized Q-learning framework for the CCN functions and show how the framework fits to a general SF control loop. We then apply this framework to two functions on mobility robustness optimization (MRO) and mobility load balancing (MLB). Our results show that the MRO function learns to optimize handover performance while the MLB function learns to distribute instantaneous load among cells.

Increased demand for higher user throughput has led to deployment of multi-layer networks commonly called heterogeneous networks (Hetnets). Therein, small cells are deployed alongside traditional macro cells, in many cases on the same spectrum. Such scenarios complicate the configuration of network parameters such as the Physical Cell Identity (PCI). A number of approaches have as such been proposed to automate the allocation of PCIs in such scenarios. These approaches seek to address the two conflicting objectives for PCI assignment in a hetnet scenario: 1) the need for optimal performance by avoiding conflicts, against 2) the requirement to separate the different layers and avoid any need to share knowledge among the layers. However, as the density of small cells increases evolving the Hetnets into what are called Ultra Dense Networks (UDN), these approaches reach their limits. In this paper, we study the performance of the current PCI allocation strategies in such UDN scenarios and evaluate their break down points. Our results show that these strategies do not adequately address PCI allocation for the UDN scenario. Specifically, we observe that PCI assignment in one layer requires knowledge of the assignments in the other layer, otherwise the consequence is a very high count of PCI confusions.

With the increase of the mobile network complexity, minimizing the level of human intervention in the network management and troubleshooting has become a crucial factor. This paper focuses on enhancing the level of automation in the network management by dynamically learning the mobile network cell states and improving the anomaly detection on the individual cell level taking into consideration not just the multidimensionality of cell performance indicators, but also the sequence of cell states that have been traversed over time. Our evaluation based on the real network data shows very good performance of such a learning model being able to capture the cell behavior in time and multidimensional space. Such knowledge can improve the detection of different types of anomalies in cell functionality and enhance the process of cell failure mitigation.

Self-Organizing Networks (SON) and a number of SO functions (SFs) have been proposed, e.g. in the LTE standard. Since SFs operate on the same network, adjusting the same set of parameters, conflicts arise. Mechanisms are thus required to resolve or minimize these conflicts. We propose Space-Time scheduling procedures that allow for separating the execution of SFs at different space and time points so as to minimize negative cross effects among the SFs. Using two Q-learning based SFs, our results show that the combined scheduling in space and time ensures that SFs learn optimal behaviors that are only due to their own actions and not the peers' actions. In doing so, they maintain good performance even within the shared environment.

Multiple Self-Organizing Networks (SON) functions have been deveoped towards the SON promise of automating cellular network operations. Meanwhile, advancing SON towards Cognitive Cellular Networks requires the (SON) Functions (SFs) to autonomously learn the required optimal configurations. Since the SFs adjust the same or related network parameters, conflicts are bound to occur. Mechanisms that are better than current SON coordination approaches must thus be devised to manage the conflicts. In this paper we propose multi-agent Concurrent Cooperative Games (CCG) an approach where peer SFs communicate with one another so as to learn to minimize the conflicts. Using two Q-learning based SFs, we evaluate the benefits of CCG comparing against the independent functions and their uncoordinated operation. Our results show that CCG achieves good compromise especially where concurrent action among neighbor cells is avoided.

Optimal settings for Handover parameters (Hysteresis and Time-to-Trigger) depend on user velocities in the network. The Self-Organization Networks (SON) standard defines the Mobility Robustness Optimization (MRO) use case for the autonomous methods of configuring the parameters in congruence to the mobility pattern. State of the art MRO solutions have relied on expert knowledge, rule based algorithms to search the parameter space; yet it is unwieldy to design rules for all possible mobility patterns in any network. In this work, we present a Q-learning MRO solution, QMRO, which learns the required parameter values appropriate for specific velocity conditions in the individual cells. We compare QMRO against the best static reference configuration (Ref) that is obtained by sweeping the parameter space. Our results show that QMRO is able to learn parameter settings that achieve similar performance to Ref in a realistic network environment where users have dynamically varying velocities.

Inter-Cell Interference (ICI) is considered as the main reason for throughput degradation in cellular systems, especially for users at the cell edges. To mitigate ICI, ICI Coordination (ICIC) has been proposed in the context of Self Organization Networks (SON). In this paper, we present a Qlearning based ICIC algorithm, called Q-ICIC that learns for each cell the best configuration for two control parameters - the sub-band power factor and the edge-to-center boundary (ECB) so as to minimize ICI. We validate the algorithm with LTE system level simulations and show that Q-ICIC achieves considerable improvement in system performance in terms of Signal to Interference plus Noise Ratio (SINR), without compromising the coverage in the cells.

Mobility Load Balancing (MLB) and Mobility Robustness Optimization (MRO) are two major SON (Self-Organizing Network) use cases specified for LTE (Long Term Evolution) by 3GPP (3rd Generation Partnership Project). MRO targets at optimizing the handover (HO) performance of the network by adjusting the HO point. Whereas, MLB reduces the congestion in traffic hotspots by advancing HOs from overloaded hotspot cells to less loaded neighbor cells. However, MLB degrades HO performance of the network because advanced HOs lead to worse SINR (Signal-to-Interference plus Noise Ratio) conditions with the target cells as compared to the original HO point. Hence, MLB can trigger MRO to optimize the HO performance and MRO can in turn trigger MLB to adjust the HO point for advanced HOs. Resultantly, MLB and MRO working in parallel can lead to oscillations. In this paper we present a policy based mechanism, termed as MLB-Static, which exploits users' mobility state to minimize the impact of MLB on HO performance and hence prevents oscillations. The simulation results, for a hexagonal LTE macro cellular deployment, show that MLB-Static has negligible impact on HO performance of the network as compared to conventional MLB while the gain of conventional MLB is preserved by MLB-Static.

Due to the large number of newly deployed small cells in heterogeneous networks needed for off-loading the mobile data traffic, the energy consumption in mobile cellular networks is strongly increasing. Energy consumption can be optimized in a self-organized way by adapting the number of active cell

Cellular radio networks are seldom uniformly loaded. This motivates the need for Load Balancing (LB), as has been defined in the LTE Self-Organization standard. It is expected that on overload, a serving cell (S-cell) initiates LB to transfer some of its edge users to its neighbor cells so called target cells, by adjusting the Cell Individual Offset (CIO) parameter. In this work, we have proposed a reactive LB algorithm that adjusts the CIOs between the S-cell and all its neighbors by a fixed step φ. Our results show that the best φ depends on the load conditions in both the S-cell and its neighbors as well as on the S-cell's user distribution. We then propose a Q-Learning (QL) algorithm that learns the best φ values to apply for different load conditions and demonstrate that the QL based algorithm performs better than the best fixed φ algorithm in virtually all scenarios.

Self-Organization (SO) has been proposed to reduce capital and operational expenses as well as to improve cellular network performance. Mobility Load Balancing (MLB) and Mobility Robustness Optimization (MRO) are two of the major proposed SO use cases. Typically, MRO sets the cell's Handover (HO) Hysteresis and Time to Trigger, to select the optimum point at which a HO is initiated. Conversely, MLB can be achieved by advancing HOs from overloaded to less loaded cells, commonly by adjusting the Cell Individual Offset (CIO). However, MLB affects HO metrics specifically because advancing HOs inadvertently increases Radio Link Failures (RLF) arising from overly early HOs and/or the number of HOs and Ping-Pong HOs. In this work we present a Q-Learning algorithm that learns the best MLB action to take in different load states so as to achieve a desired load transfer, but with the least effect on HO performance. The learning agent minimizes the negative HO effects by applying a penalty to MLB actions that cause high RLFs, thereby reducing the effects on HO metrics by up to 30%.