In this paper, we investigate a novel digital network twin (DNT) assisted deep learning (DL) model training framework that enables a base station (BS) to select the data collection source from both physical network and DNT for training DL models used to optimize network performance. In particular, we consider a DNT enabled cellular network that consists of a physical network where a BS uses several antennas to serve multiple mobile users, and a DNT that is a virtual representation of the physical network. The BS must adjust the antenna tilt angles to optimize the data rates of all users. Due to the user mobility, the BS may not be able to accurately track the network dynamics. Hence, a reinforcement learning (RL) approach is used to dynamically adjust the antenna tilt angles. To train the RL, we can use data collected from the physical network and the DNT. The data collected from the physical network is more accurate but incurs more communication overhead compared to the data collected from the DNT. Therefore, it is necessary to determine the ratio of data collected from the physical network and the DNT to improve the training of the RL model, so as to optimize the tilt angle of each antenna in response to user mobility. We formulate an optimization problem to jointly optimize the tilt angle adjustment policy and the data collection strategy, aiming to maximize the data rates of all users while constraining the time delay introduced by collecting data from the physical network. To solve this optimization problem, we propose a hierarchical RL framework consisting of a two-level proximal policy optimization (PPO). The first level PPO dynamically adjusts the antenna tilt angles, and the second level PPO determines the ratio of data collected from the physical network to improve the first level PPO training performance. Compared to traditional single RL algorithms such as deep Q network (DQN), the designed method optimizes the data collection ratio and the antenna tilt angles at diverse time intervals, allowing the second level PPO to adjust the data collection ratio with a large timescale using the training information provided by the first level PPO, and allowing the first level PPO focuses on adjusting the antenna tilt angles with a small timescale. Simulation results show that our proposed method reduces the physical network data collection delay by up to 71.85% compared to a hierarchical RL that uses a deep deterministic policy gradient algorithm as the second level RL.

Secondary Carrier Prediction (SCP) is a data-driven Radio Access Network (RAN) feature that ensures the best possible cell and frequency carrier assignment to a User Equipment (UE), without the UE needing to perform an excessive amount of inter-frequency measurements and reporting. In this paper, we investigate if the performance of SCP can be further enhanced by jointly training it with other data-driven RAN tasks at the network edge in a Multi-Task Learning (MTL) fashion. We evaluate the trained SCP model performance when it runs a Mobility Load Balancing (MLB) downstream task under different task groupings. Results show that for the simulation scenario considered in this work, about 29.6% improvement in load balancing index is achieved when SCP is used in MTL mode over Single Task Learning (STL) mode during inference.

As wireless cellular networks evolve from one generation to the next, the associated challenges in terms of compute, latency, and energy efficiency are continuously on the rise. However, traditional methods are unable to cope with the complexity of operation in next generation mobile networks, and the operational costs can become a significant hurdle in emerging generations. To address these issues, an AI-native wireless communication system is needed to support faster processing and optimization. Researchers have been exploring nature-inspired models for years to improve deep learning architectures, making them smaller, sparser, more efficient, and easily interpretable. In this study, we propose a nature-inspired Machine Learning architecture that leverages curriculum learning to predict cellular traffic on real wireless cellular network data. Our proposed architecture requires significantly fewer parameters to train, is 64% sparse and performs better than the standard state-of-the-art model. Here, sparsity refers to the fewer connections than the maximum possible number of connections within that network between each layer. Moreover, in terms of performance and carbon emissions, this sparsely connected architecture is better than traditional Machine Learning models.

In modern wireless communication systems, radio propagation modeling to estimate pathloss has always been a fundamental task in system design and optimization. The state-of-the-art empirical propagation models are based on measurements in specific environments and limited in their ability to capture idiosyncrasies of various propagation environments. To cope with this problem, ray-tracing based solutions are used in commercial planning tools, but they tend to be extremely time-consuming and expensive. We propose a Machine Learning (ML)-based model that leverages novel key predictors for estimating pathloss. By quantitatively evaluating the ability of various ML algorithms in terms of predictive, generalization and computational performance, our results show that Light Gradient Boosting Machine (LightGBM) algorithm overall outperforms others, even with sparse training data, by providing a 65% increase in prediction accuracy as compared to empirical models and 13x decrease in prediction time as compared to ray-tracing. To address the interpretability challenge that thwarts the adoption of most ML-based models, we perform extensive secondary analysis using SHapley Additive exPlanations (SHAP) method, yielding many practically useful insights that can be leveraged for intelligently tuning the network configuration, selective enrichment of training data in real networks and for building lighter ML-based propagation model to enable low-latency use-cases.

5G is considered as the ecosystem to abet the ever growing number of mobile devices and users requiring an unprecedented amount of data and highly demanding Quality of Experience (QoE). To accommodate these demands, 5G requires extreme densification of base station deployment, which will result in a network that requires overwhelming efforts to maintain and manage. User mobility prediction in wireless communications can be exploited to overcome these foregoing challenges. Knowledge of where users will go next enables cellular networks to improve handover management. In addition, it allows networks to engage in advanced resource allocation and reservation, cell load prediction and proactive energy saving. However, anticipating the movement of humans is, in itself, a challenge due to the lack of realistic mobility models and insufficiencies of cellular system models in capturing a real network dynamics. In this paper, we have evaluated Artificial Intelligence (AI)-assisted mobility predictors. We model mobility prediction as a multi-class classification problem to predict the future base station association of the mobile users using Extreme Gradient Boosting Trees (XGBoost) and Deep Neural Networks (DNN). Using a realistic mobility model and a 3GPP-compliant cellular network simulator, results show that, XGBoost outperforms DNN with prediction accuracy reaching up to 95% in a heterogeneous network (HetNet) scenario with shadowing varied from OdB to 4dB.

In modern wireless communication systems, radio propagation modeling has always been a fundamental task in system design and performance optimization. These models are used in cellular networks and other radio systems to estimate the pathloss or the received signal strength (RSS) at the receiver or characterize the environment traversed by the signal. An accurate and agile estimation of pathloss is imperative for achieving desired optimization objectives. The state-of-the- art empirical propagation models are based on measurements in a specific environment and limited in their ability to capture idiosyncrasies of various propagation environments. To cope with this problem, ray-tracing based solutions are used in commercial planning tools, but they tend to be extremely time consuming and expensive. In this paper, we propose a Machine Learning (ML) based approach to complement the empirical or ray tracing-based models, for radio wave propagation modeling and RSS estimation. The proposed ML-based model leverages a pre-identified set of smart predictors, including transmitter parameters and the physical and geometric characteristics of the propagation environment, for estimating the RSS. These smart predictors are readily available at the network-side and need no further standardization. We have quantitatively compared the performance of several machine learning algorithms in their ability to capture the channel characteristics, even with sparse availability of training data. Our results show that Deep Neural Networks outperforms other ML techniques and provides a 25% increase in prediction accuracy as compared to state-of-the-art empirical models and a 12x decrease in prediction time as compared to ray tracing.

Millimeter wave spectrum utilization and network densification are two of the fundamental technologies that will enable high user quality of experience required in 5th Generation mobile cellular networks. However, user sparsity in ultra-dense heterogeneous networks and coverage limitations of millimeter wave cells means reliability of such networks will become a key operational challenge. Recent studies have explored the use of machine learning techniques for outage detection in legacy and heterogeneous mobile cellular networks. However, machine learning techniques are highly susceptible to noise in the training data which can affect their outage detection accuracy. To counter these challenges, we present a novel outage detection method based on entropy field decomposition technique first introduced in [1]. The proposed method is able to detect cell outages with at least 96% accuracy even as the level of shadowing in the network is increased which makes it ideal for practical implementation in emerging ultra-dense heterogeneous networks with millimeter wave cells. The proposed solution is compared against k-means clustering for outage detection with results showing that not only does entropy field decomposition return higher true positive results, it also returns fewer false positive results compared to k-means clustering.

To meet the accuracy, latency and energy efficiency requirements of modern healthcare systems during real-time collection and analysis of health data, a distributed edge computing environment is the answer, combined with 5G speeds and modern AI techniques. Using the state-of-the-art machine learning based classification techniques plays a crucial role in creating the optimal healthcare system on the edge. This work first provides a background on the current and emerging edge computing classification techniques for healthcare applications, specifically for electrocardiogram (ECG) beat classification. After implementing these classification techniques on a Raspberry Pi- based platform we perform a comparison of the performance of these classification techniques with respect to three key performance indicators (KPI) of interest for health care applications namely accuracy, energy efficiency, and latency. Benefiting from the results of the comparative analysis presented in this work, a distilled neural network algorithm can be selected for optimal deployment and over 90% accuracy in given scenario in healthcare system depending on the specific requirements of the given scenario.

The proliferation of mobile devices and big data has made it possible to understand the human movements and forecasts of precise and intelligent short and long-term data consumption of services like call, sms, or internet data which has interesting and promising applications in modern cellular networks. Human nature and moods are known to be synonymous with the physical attributes of mother nature such as temperature. The change in those physical features affects the human routines and activities such as cellular data consumptions. The future of telecommunication lies in the exploration of heap of information and data available to companies and inferring the valuable results through extensive analysis. In this paper, we analyze three main traits of cellular activity: sms, call, and internet. This paper investigates whether the relationship between the temperature and the cellular data consumption exits or not. This work introduces a novel approach to identify the strength of relationship between the temperature and cellular activity (sms, call, internet) and discuss the methods to quantify the relationship using correlation method. The real network CDR big data set - Milano Grid data set is used to analyze the behavior of the cellular activity with respect to temperature.

Increased network wide energy consumption is a paramount challenge that hinders wide scale ultra-dense networks (UDN) deployments. While several Energy Saving (ES) enhancement schemes have been proposed recently, these schemes have one common tenancy. They operate in reactive mode i.e., to increase ES, cells are switched ON/OFF reactively in response to changing cell loads. Though, significant ES gains have been reported for such ON/OFF schemes, the inherent reactiveness of these ES schemes limits their ability to meet the extremely low latency and high QoS expected from future cellular networks vis-a-vis 5G and beyond. To address this challenge, in this paper we propose a novel user mobility prediction based AUtonomous pROactive eneRgy sAving (AURORA) framework for future UDN. Instead of observing changes in cell loads passively and then reacting to them, AURORA uses past hand over (HO) traces to determine future cell loads. This prediction is then used to proactively schedule small cell sleep cycles. AURORA also incorporates the effect of Cell Individual Offsets (CIOs) for balancing load among cells to ensure QoS while maximizing ES. Extensive system level simulations leveraging realistic SLAW model based mobility traces show that AURORA can achieve significant energy reduction gain without noticeable impact on QoS.

Optimizing parameters to achieve optimal trade-off between coverage and capacity is a well known research problem in 5th Generation mobile cellular networks. Introduction of ultra dense heterogeneous networks is bound to exacerbate this problem by adding a third conflicting objective i.e., load balancing between macro and small cells. Existing solutions on load balancing are not suited for this purpose since they balance cell loads at the cost of spectral efficiency, a key measure of resource efficiency in 5G networks. To tackle these challenges, we propose a novel solution for joint optimization of coverage, capacity and load in ultra dense heterogeneous networks that does not compromise spectral efficiency or subscriber satisfaction. The proposed solution incorporates antenna tilt, transmit power and cell offset parameters into a single objective. We compare two different versions of our proposed solution with existing coverage optimization [1], and coverage, capacity and load optimization [2] algorithms, as well as coverage optimization and load balancing solutions for HetNets to highlight its advantages. Simulation results show that the proposed solution improves service quality while also offering higher residual capacity compared to nearest benchmark.

Cell densification is being perceived as the panacea for the imminent capacity crunch. However, high aggregated energy consumption and increased inter-cell interference (ICI) caused by densification, remain the two long-standing problems. We propose a novel network orchestration solution for simultaneously minimizing energy consumption and ICI in ultra-dense 5G networks. The proposed solution builds on a big data analysis of over 10 million CDRs from a real network that shows there exists strong spatio-temporal predictability in real network traffic patterns. Leveraging this, we develop a novel scheme to pro-actively schedule radio resources and small cell sleep cycles yielding substantial energy savings and reduced ICI, without compromising the users QoS. This scheme is derived by formulating a joint Energy Consumption and ICI minimization problem and solving it through a combination of linear binary integer programming, and progressive analysis based heuristic algorithm. Evaluations using: 1) a HetNet deployment designed for Milan city where big data analytics are used on real CDRs data from the Telecom Italia network to model traffic patterns, 2) NS-3 based Monte-Carlo simulations with synthetic Poisson traffic show that, compared to full frequency reuse and always on approach, in best case, the proposed scheme can reduce energy consumption in HetNets to 1/8th while providing same or better QoS.

Today, the importance of cellular networks is ever-growing. The increasing complexity of networks is expected to decrease reliability. In order to continue reliable operation in a cost-efficient manner, previous literature has explored Proactive Self-Healing methods, but actual application to cellular networks has been lacking. Thus, in this paper, we aim to institute a proactive approach for failure prediction of time series data by surveying a wide range of techniques. To determine the best in predicting network failures, Support Vector Machine (SVM) Regression and multiple Neural Network variants were utilized along with a Continuous Time Markov Chain (CTMC) analytical model to provide reliability analysis. All results are derived from actual network data. We conclude the pattern of these failures is most likely non-linear, and the most promising technique is a Deep Neural Network utilizing Autoencoders. The CTMC analysis demonstrates that current networks barely reside in a healthy state, so the goal is that this paper will lead to improvements, especially in Self-Organizing Networks (SON).

Ultra-dense Heterogeneous networks (UDHN) are emerging as the inevitable approach to cope with the imminent cellular network capacity crunch. However, load imbalance and widely disproportionate SINR distribution between macro and small cells, remains the key hurdle in harnessing the full potential of UDHN. In this paper we address this problem by proposing and analysing a novel load-aware user association methodology that offers a mechanism to simultaneously optimize network capacity, load distribution and coverage. The solution concurrently leverages the three key optimization parameters for Coverage and Capacity Optimization (CCO) and Load Balancing (LB) SON functions i.e. antenna tilts, transmit powers and cell individual offsets (CIOs). The method incorporates exponential-weighting based prioritization of CCO and LB SON functions within the user association process. The results suggest that the proposed approach offers a distribution of load between macro and small cells that yields more gain in terms of both network capacity and user quality of service than conventional max signal strength or max SINR based association methods.

It is anticipated that the future cellular networks will consist of an ultra-dense deployment of complex heterogeneous Base Stations (BSs). Consequently, Self-Organizing Networks (SON) features are considered to be inevitable for efficient and reliable management of such a complex network. Given their unfathomable complexity, cellular networks are inherently prone to partial or complete cell outages due to hardware and/or software failures and parameter misconfiguration caused by human error, multivendor incompatibility or operational drift. Forthcoming cellular networks, vis-a-vis 5G are susceptible to even higher cell outage rates due to their higher parametric complexity and also due to potential conflicts among multiple SON functions. These realities pose a major challenge for reliable operation of future ultra-dense cellular networks in cost effective manner. In this paper, we present a stochastic analytical model to analyze the effects of arrival of faults in a cellular network. We exploit Continuous Time Markov Chain (CTMC) with exponential distribution for failures and recovery times to model the reliability behavior of a BS. We leverage the developed model and subsequent analysis to propose an adaptive fault predictive framework. The proposed fault prediction framework can adapt the CTMC model by dynamically learning from past database of failures, and hence can reduce network recovery time thereby improving its reliability. Numerical results from three case studies, representing different types of network, are evaluated to demonstrate the applicability of the proposed analytical model.

This paper presents a novel scheme for spectral efficiency (SE) optimization through clustering of users. By clustering users with respect to their geographical concentration we propose a solution for dynamic steering of antenna beam, i.e., antenna azimuth and tilt optimization with respect to the most focal point in a cell that would maximize overall SE in the system. The proposed framework thus introduces the notion of elastic cells that can be potential component of 5G networks. The proposed scheme decomposes large-scale system-wide optimization problem into small-scale local sub- problems and thus provides a low complexity solution for dynamic system wide optimization. Every sub- problem involves clustering of users to determine focal point of the cell for given user distribution in time and space, and determining new values of azimuth and tilt that would optimize the overall system SE performance. To this end, we propose three user clustering algorithms to transform a given user distribution into the focal points that can be used in optimization; the first is based on received signal to interference ratio (SIR) at the user; the second is based on received signal level (RSL) at the user; the third and final one is based on relative distances of users from the base stations. We also formulate and solve an optimization problem to determine optimal radii of clusters. The performances of proposed algorithms are evaluated through system level simulations. Performance comparison against benchmark where no elastic cell deployed, shows that a gain in spectral efficiency of up to 25% is possible depending upon user distribution in a cell.