Call muting-unexpected silences during voice calls due to extended RTP packet loss is a major challenge in high-mobility 5G environments, severely degrading Mean Opinion Score (MOS) and user experience. We propose DRLCQ, a Deep Reinforcement Learning-based framework that dynamically tunes Cell Individual Offset (CIO) in real time to reduce mute events and enhance voice quality. Integrated as an xApp within the O-RAN Near-RT RIC, DRLCQ leverages live network KPIs (e.g., SINR, jitter, packet loss) to learn optimal handover decisions. Evaluated against static and heuristic baselines, DRLCQ achieves over 20% fewer call mute incidents and up to 85% higher MOS, demonstrating a scalable and intelligent solution for AI-native RAN control.

This paper proposes a Generative AI-based Digital Twin (GDT) platform for automated network feature performance evaluation, designed for Beyond 5G (B5G) networks. The platform addresses the inefficiencies of manual evaluation by utilizing a conditional Generative Adversarial Network (cGAN) to simulate network performance based on historical data and new AI/ML features. The Network GDT integrates a novel Digital Twin Augmenting Condition (DTAC) framework, allowing for real-time simulation and performance evaluation of network features. This system significantly reduces the time and cost associated with manual evaluations, improves decision-making, and optimizes Quality of Service (QoS) and Quality of Experience (QoE). The cGAN-based model dynamically generates synthetic data, enabling comprehensive performance insights and proactive AI solution testing under various network scenarios. Experimental results demonstrate high prediction accuracy for congestion use case, validating the robustness of the proposed system. The platform's dual-phase strategy ensures that AI-based solutions are rigorously tested in simulated environments before deployment in real networks, minimizing risks and enhancing stability. This approach provides a scalable and efficient solution for future B5G networks, paving the way for more reliable and optimized wireless communication systems.

In the rapidly advancing domain of 5G communication systems, channel decoding, particularly turbo decoding, has emerged as a significantly complex challenge. Turbo decoding is an essential element within communication frameworks, necessitating both efficiency and rapid processing to cater to the demanding data rates and stringent low latency requirements of 5G networks. This paper focuses on the unique contributions of employing a Sparse Recurrent Neural Network (SRNN) architecture, leveraging sparsity to reduce computational load while maintaining high performance significantly. Unlike existing approaches, our method introduces a novel piece-wise linear approximation of the activation function, enhancing efficiency and scalability for NextGen communication systems. Our approach leverages the principles of sparsity and employs a piece-wise linear approximation of the activation function to markedly reduce the computational load of the turbo-decoding process.Comprehensive evaluations demonstrate that our RNN architecture outperforms existing deep learning models in the context of turbo decoding and with a significantly lower computational footprint. This research contributes to the field by providing a scalable, efficient, and less computationally intensive turbo-decoding method, particularly suited for the next-generation cloud systems underlying 5G and beyond communication technologies.

The emergence of 5G millimeter wave (mmWave) networks holds tremendous potential for realizing New Radio Vehicle-to-Everything (NR-V2X) applications, which demand high data transfer rates and low latency. While 5G mmWave networks offer a commercially viable platform for the next generation of NR-V2X applications, they encounter various propagation challenges. Unfortunately, the transport layer (L4 layer) between conventional NR-V2X communication endpoint devices remains oblivious to wireless medium issues and congestion scenarios in the Radio Access Network (RAN). This often results in packet loss, transmission delays, and diminished quality of experience. This paper introduces a novel approach to predict congestion or buffer bloat scenarios in the RAN proactively. The proposed solution leverages machine learning (ML) and time series RAN Key Performance Indicators (KPIs) to forecast congestion scenarios for future time steps. Subsequently, this information is relayed to assist Active Queue Management (AQM) in the User Plane Function (UPF), integrating cognitive capabilities rooted in RAN awareness. By predicting congestion events and notifying the sender through Explicit Congestion Notification (ECN), the proposed system optimizes traffic flow, reduces latency, and forestalls impending congestion for NR-V2X applications. Moreover, the proactive prediction from our solution facilitates seamless low-latency applications like augmented reality (AR), virtual reality (VR) and 6 degrees of freedom (DOF) videos in NR-V2X. Moreover, the impending congestion prediction improves the handover experience of vehicular devices from one roadside unit (RSU) to another. Through extensive simulations, this paper demonstrates that the proposed novel ML-based cognitive RAN-aware transport layer significantly enhances the Quality of Experience (QoE), ensuring reliable, efficient, and low-latency connectivity. The proposed approach underscores the potential of a RAN-aware transport layer (L4 layer) in advancing the evolution of NR-V2X applications within the expansive realm of 5G and beyond networks.

Wireless 5G and beyond (B5G) technology offers multiple of machine learning (ML) use cases, including congestion detection, handover prediction, MAC scheduling and more. Many of these use cases involve solving complex problems that need neural networks (NN) or classical ML algorithms to achieve optimal solutions. However, implementing these NN inferences on resource-constraint base stations (BS) pose significant challenges. BSs has several limitations in terms of CPU frequency, number of cores, memory capacity, and the absence of dedicated ML hardware (HW) offloads. In this paper, the authors, propose a lightweight artificial intelligence (LWAI) method to derive computationally efficient mathematical relations (EMR) between Key Performance Indicators (KPIs) using reinforcement learning (RL). The derived EMR enables ML inference to be implemented on resource-constrained BSs. LWAI takes KPIs information as input and provides EMRs as output. The generated EMRs are then deployed on the BS. Inference is made to the BS using these relations. This approach makes ML inference realizable on embedded BSs with commercial-grade accuracy and optimal real-time prediction latency. We demonstrate the effectiveness and applicability of the LWAI framework in two real-world scenarios. The results highlight the potential for our approach to detect congestion by predicting physical resource block (PRB) and energy savings in BS. Our results show the efficacy of EMRs. LWAI framework-derived EMRs consume around 10% CPU cycles and 5% memory to execute as compared to NN models while maintaining over 90% prediction accuracy. Our approach opens up new possibilities for realizing ML inference ideas on resource-constraint embedded devices.

Massive multiple-input multiple-output (mMIMO) and millimeter waves (mmWaves) are considered to be key technologies for 5G and beyond wireless communications. Massive-MIMO at mmWave frequencies is coupled with advanced beamforming algorithms, to meet the high data rates and stringent latency requirements. The sensitivity of mmWaves to physical obstacles causes significant attenuation of signal leading to link failure between the user and base station (BS). Hence, uninterrupted connectivity to a user can be established via proactive handovers (HOs) between BSs using deep-learning (DL) models for future blockage and beam prediction. In this paper, we present a data-driven neural network approach using Convolution neural networks (CNN) and recurrent neural network (RNN) models for the prediction of future blockages and beams to enable proactive HO and to facilitate seamless connectivity in 5G and beyond ultra-dense networks (UDNs). We validated our proposed deep-learning models’ efficacy for blockage and beam prediction. The evaluation results of CNN and RNN models demonstrate an accuracy of 99% for blockage predictions and more than 90% in the case of beam predictions, with low computational complexity. The proposed DL models have the potential for commercial deployments in next-generation radio access networks (RAN) systems like O-RAN, vRAN, and C-RAN.

Radio Access Network (RAN) data plane systems in 5G are expected to support multiple user traffic flows demanding high throughput. Base Station (BS) units like Distributed Unit (DU), and Radio Unit (RU) are heavy consumers of power. Network traffic varies over time and peak traffic is rarely seen. These traffic variations provide an opportunity for power saving to reduce operational expenditure (OpEx) during low-traffic periods. In commercial operator deployment, multiple BS units are connected to a Central Management Entity (CME). In this paper, we propose a dynamic power management method to proactively adjust the operating CPU frequency of BS according to the predicted core load. The proposed solution considers deriving a mathematical formula for predicting core load using the computing horsepower of CME with machine learning techniques. To avoid prediction model localization to a single BS and to provide better generalization, Federated Learning (FL) based mathematical formula derivation is proposed. To achieve this, multiple BSs with demographic similarities are grouped. Mathematical formulas for each BS are aggregated and finetuned at the controller using Federated Learning to arrive at a formula to be used by BSs to predict core load. Experiments are performed on real network traces with multiple data flows across different Radio Bearers in New Radio (NR) Radio Link Control (RLC) at RAN Distributed Unit (DU). Results show that the proposed solution reduces processor power consumption by 12%, potentially saving millions of USD annually for commercial deployments.

Cloud computing infrastructure is an integral part of wireless communication applications, such as vRANs. Given the pervasive use of cloud computing, keeping track of its health is of utmost importance to its providers. Large scale of deployment of this aggravates the complexity. All vRAN based cloud applications generate logs which are raw text messages to reflect run-time execution. Analyzing such log messages and training a model can help an operator timely detect an anomaly, before it leads to service disruption. In this paper, we apply various natural language processing (NLP) models to detect system anomalies by analyzing logs and evaluate their precision and accuracy of anomaly detection. We make use of a machine learning (ML) model in form of a recurrent neural network (RNN) for anomaly detection. We have added attention layer to RNN models like LSTM and GRU models to improve accuracy. We found GRU models with attention are a good fit for logs dataset in cloud environment. Log messages are typically short sentences, linguistically much simpler, and don't suffer from long-range dependencies, hence adding attention layer didn't help for LSTM. GRU has a reduced instruction set, faster training time, hence are much more energy efficient without compromising on anomaly detection accuracy.

One of the key technologies of next-generation 6G networks is millimeter-wave communications that will deploy a large number of antennas at the base station enabling narrow beams toward user locations to mitigate the path loss. Conventional methods have resulted in high training overhead in finding the best beam pair to obtain beam alignment between the base station and a user. This paper proposes a data-driven neural network approach to intelligently perform the beam selection between the transmitter-receiver pair. We propose a convolution neural network (CNN) based beam selection method trained from simulator-generated beam dataset. We use skip connections and hyperparameter optimization to balance the trade-off in accuracy and computational complexity. We validate the efficacy of our proposed method by comparing it with other conventional and machine learning-based approaches. Evaluation results show higher accuracy (> 70%) while reducing the computational complexity upto 15%.

Accurate channel estimation in the millimeter-wave (mmWave) based wireless communication systems is challenging and involves a lot of computational costs. The mmWave frequency band has its advantages and disadvantages. At higher frequency mmWave bands, due to smaller wavelengths, we can pack a large number of antennas compared to lower frequency bands. However, the main disadvantages of the mmWave system are computing accurate channel estimation, smaller coverage, and high signal absorption. Besides, when multiple-input multiple-output (MIMO) systems operated over mmWave frequencies, it makes the channel estimation even more intricate in terms of computational complexity and estimation accuracy. In this paper, we plan to address these limitations and improve channel accuracy; we proposed a Continual Learning (CL)-based method for channel estimation in mmWave MIMO systems. Besides, we also proposed an activation function that is numerically stable and robust against early saturation. We discussed several channel estimation algorithms from the literature, also evaluated and compared their performances via numerical simulations. Our simulation results show that the proposed CL-based method outperforms the existing minimum mean squared error (MMSE)-based channel estimators in terms of precision. Furthermore, based on our experiments, we give insight into spectral efficiency with respect to the number of available channel observations.

Owing to improved coverage and flexibility, the radio access network (RAN) functionalities are being virtualized in a sense that the base station will merely act as a radio unit, and all the baseband processing will occur in the cloud. Therefore, the baseband signal-processing algorithms need to be designed in a way that it can match the latency requirements. In this paper, we address one of the inherent but complex issues in baseband signal processing, namely, the log log-likelihood ratio (LLR) computation. In general, soft-decision rules are used for calculating the LLRs, which is computationally expensive. Thus, we attempt to exploit the benefits of hard-decision based LLRs for proposing a framework that uses soft decision only when the received symbols are closed to the decision boundary; otherwise, the framework uses hard decision. This helps us to keep the complexity low while meeting the desirable error performance. These schemes are suitable for incorporation in virtual RAN systems while considering appropriate QoS requirements.

Medical imaging has evolved to become an essential tool for screening and diagnosing diseases, but they have certain limitations just like every other technology. X-rays, which is one of the most common radiological examinations, is not immune to imperfections. In this paper, we aim to tackle one such imperfection in X-rays, which is the presence of undesirable noises which causes aberrations in the output projections. This makes diagnosis and analysis difficult since such noises shroud the intricate details that these images contain. Distinctive denoising algorithms have been proposed in the past for a spectrum of vision datasets, but a very few of them are for X-rays. We introduce a new denoising network called SkiDNet, a deep learning approach using an encoder-decoder architecture with skip connections of varying length. The network has been trained on the NIH Chest X-Ray Dataset. With the unique properties injected by different types of connections, SkiDNet is able to surpass the performance of existing models. Furthermore, adopting a different approach to weight initialization and batch normalization makes the network more robust. Denoised X-rays obtained from the network were objectively evaluated using different metrics namely mean squared error, peak signal-to-noise ratio, and the structural similarity index.

Emotion is made up of three components; physiological arousal, expressive behaviors, and conscious experience. Psychological theories like Social Constructionist theory believe that although emotions are a universal phenomenon, they are deeply influenced by the cultural background of the speaker. Since culture and language have an intimate connection between them, the language used by a particular social group would have a significant effect on emotions expressed in their verbal communication. We aim to study the susceptibility of emotions in different languages, which is expressed in speech. We introduce a novel Deep Convolution Siamese Network (DCSN) for determining the similarity between speech samples. The speech samples were obtained from emotional speech datasets spoken in different languages, namely English, Italian, and German. Two experiments were undertaken for studying the correlation between language and emotional speech. Similarity metric along with scatter plots were obtained, which establish that the DCSN is capable of identifying both languages and emotions in a speech sample. The results hence provide a conclusive proof for the model to be used for the study of different psychological aspects of emotional speech.

Due to the multifarious challenges that emerge when developing an artificial intelligent (AI) agent that can compete with human players, the classic game of Othello has received a lot of attention from the Computational-Intelligence community. This paper proposes an AI agent that learns a winning strategy for the game of Othello using the eXtended Classifier System (XCS) algorithm which is a popular variant of the Learning Classifier System (LCS) algorithm. Othello has been a favourite in the study of AI due to its simple set of rules, low branching factor and well defined strategic concepts. The LCS system consists of a rule-set which is made to evolve using a combination of Reinforcement Learning (RL) and Genetic Algorithm (GA) such that the evolved rule-set learns an optimal action for each board state. A 6×6 Othello board will be used for this experiment in order to evaluate the applicability of the proposed agent in learning a winning game-playing strategy. The performance of the proposed agent was evaluated against three categories of agents: minimax, human and random agent. The XCS agent was able to outperform the above-mentioned agents showing the effectiveness of rule-based evolutionary learning in Othello. This work demonstrates the possibility of using the XCS algorithm in other strategy-based combinatorial games.