Rate-splitting multiple access (RSMA) and unmanned aerial vehicles (UAVs) are emerging as key technologies for enhancing connectivity and resource efficiency in future 6G networks. This study investigated a UAVassisted RSMA downlink communication system and developed a novel framework that jointly optimizes the UAV's trajectory, precoding matrix, and common rate to maximize the minimum achievable user rate. The proposed system was formulated as a Markov decision process (MDP) and solved using deep reinforcement learning (DRL), specifically using the proximal policy optimization (PPO) algorithm. This learning-based approach enables UAVs to adapt dynamically to the environment without relying on prior channel state information (CSI), allowing for efficient resource allocation and interference management in complex and dynamic wireless scenarios. Furthermore, a precoding design based on a uniform rectangular array (URA) was employed to enhance directional transmission and spatial multiplexing. In simulations, the proposed method significantly outperformed existing benchmarks, achieving an average minimum rate improvement of 23.31% over Deep Deterministic Policy Gradient (DDPG), 48.59% over Soft Actor-Critic (SAC), 50.87% over Trust Region Policy Optimization (TRPO), 63.36% over REINFORCE algorithm, 79.13% over Greedy algorithm, and 145.67% over Random strategies, respectively. These results confirm the potential of UAV-aided RSMA networks in next-generation wireless environments.

In this work, we study energy-efficient online federated learning. First, we analyzed the convergence performance of device selection in dynamic and non-stationary conditions, deriving the upper bound of the convergence rate. Building on this, we formulated an optimization problem that minimizes energy consumption while ensuring convergence and meeting latency constraints. Second, we developed an online allocation and scheduling-based iterative strategy (OASIS). Here, we designed a combinatorial upper-confidence-bound-based device scheduling algorithm with a newly designed reward function. We also derived a Lambert function-based power allocation to the scheduled devices in closed form. We performed a dynamic regret analysis, which reveals that the proposed algorithm effectively adapts to dynamic environments and maintains near-optimal decisions over time. It also demonstrates that the proposed algorithm achieves sublinear regret in a slowly changing dynamic environment and optimal regret in a static environment. Experimental results show that the proposed OASIS achieves faster convergence and provides significantly lower energy consumption than existing conventional baseline strategies. We also confirmed that the performance gain increases as the target accuracy level becomes higher. These results validate the energy efficiency and robustness of the proposed approach in realistic and time-varying federated learning environments.

The dynamic metasurface antenna (DMA) architecture is defined by Lorentzian amplitude-phase coupling and waveguide-induced attenuation. This summary reviews DMA-based downlink and uplink beamforming, categorized by objective (sum-rate vs. energy efficiency) and phase control (continuous vs. discrete). Representative methods use alternating optimization combining WMMSE-based digital updates with manifold or projected-gradient updates of DMA weights. Power-aware designs apply fractional programming and convex approximation. For switched meta-atoms, structure-aware quantization mitigates low-resolution loss, while DMA-specific channel estimation with combiner co-design enables efficient reception with few RF chains.

The large model sizes and high computational demands of the current semantic communication systems pose challenges for deployment in environments with limited resources. Furthermore, existing semantic communication frameworks predominantly operate unidirectionally, significantly increaseing model redundancy and resource requirements in scenarios requiring simultaneous bidirectional information exchange. To address these challenges, we propose a BidDeepSC-1.58b, integrating three complementary innovations: a unified encoder-decoder architecture enabling bidirectional communication with a halved parameter count, 1.58-bit quantization for extreme model compression, and a slimmable neural network with an intelligent semantic fidelity predictor that dynamically selects optimal width configurations. The experimental results demonstrate that the proposed BidDeepSC-1.58b achieves 88.1% model size reduction (3.79 MB footprint) while outperforming baseline models by over 30% in semantic similarity under challenging channel conditions. Weight-tying enables bidirectional communication with minimal performance trade-offs, showing only 2.5% degradation at -6 dB SNR that diminishes to 0.6% at 18 dB SNR, a pattern consistent across diverse proposed and baseline architectures. The framework reduces inference latency by up to 88% and enables dynamic width adaptation that achieves 39.2% computational savings while maintaining semantic performance thresholds. This system enables the efficient deployment of semantic communication in resource-constrained environments and facilitates bidirectional semantic information exchange across a broad range of applications.

Rate-splitting multiple access (RSMA) and simultaneously transmitting and reflecting (STAR) reconfigurable intelligent surface (RIS) have emerged as promising technologies for multiple access and relaying in sixth-generation (6G) networks. This paper investigates a STAR-RIS-aided RSMA system that minimizes transmit power while satisfying users' quality of service (QoS) requirements under spatially correlated channels. The formulated non-convex optimization problem jointly optimizes the precoding vector, common message ratio, phase shift matrix, and transmission and reflection ratio. To solve this, the problem is decomposed into three subproblems: 1) precoding and common message ratio control at the base station (BS), 2) phase shift matrix design, 3) transmission/reflection ratio control at the STAR-RIS. These subproblems are transformed into convex forms via penalty functions and linear approximations. A stationary solution is obtained using interior-point initialization, successive convex approximation, and alternating optimization algorithms. Through extensive simulations under various channel environments, including spatially correlated scenarios, the proposed method achieves much lower transmit power while meeting QoS requirements, outperforming conventional (STAR-)RIS-aided RSMA, non-orthogonal multiple access, and space-division multiple access. The performance gain becomes more significant when the number of STAR-RIS elements or BS antennas is limited.

The emergence of Semantic Communication (SemCom) is revolutionizing wireless networks, particularly in the context of next-generation communication systems (6 G). Unlike the conventional Shannon-based paradigm, which focuses on the accurate transmission of raw data, SemCom leverages machine learning (ML) to extract and transmit only the semantic meaning of information, thereby significantly reducing communication overhead and enhancing spectral efficiency. SemCom has broad applications across diverse fields, including the metaverse, digital twins, tactile Internet, autonomous vehicles, and satellite-unmanned aerial vehicle (UAV) networks. Among these, Autonomous Driving Networks (ADNs) present a unique challenges where SemCom can play a crucial role. In ADNs, SemCom can optimize the design of semantic encoders and decoders for Vehicle Edge Clouds (VECs) and autonomous vehicles, facilitating more efficient data transmission and decisionmaking. However, integrating SemCom into ADNs introduces several challenges, including adversarial attacks, privacy risks, and diversity among vehicles. This paper provides a comprehensive survey on the challenges of SemCom in ADNs. We explore state-of-the-art research trends aimed at addressing these challenges. Finally, we discuss potential future directions for SemCom in autonomous vehicle networks.

This paper explores reinforcement learning (RL) based on resource allocation in cell-free networks, a promising alternative to traditional cellular architectures. Cell-free networks eliminate cell boundaries by using distributed access points (APs) to collaboratively serve users, improving spectral efficiency and service uniformity. However, managing power allocation, beamforming, and AP clustering in such decentralized environments presents new challenges. We present several RL-based algorithms aimed at optimizing key network functions. Specifically, the study explores dynamic power control, advanced beamforming techniques, and multi-agent RL frameworks for clustering access points. The proposed methods leverage the adaptability of RL to optimize network performance in real-time under varying conditions.

Recently, the simultaneous transmitting and reflecting (STAR) reconfigurable intelligent surface (RIS) has been gaining attention as a key enabler for sixth-generation networks, providing additional links with reduction in power consumption. This paper investigates the STAR-RIS’s potential in a cell-free (CF) massive multiple-input multiple-output (mMIMO) network, where distributed APs serve user over the same time/frequency. We propose a deep deterministic policy gradient framework satisfying system-specific and per-user spectral efficiency constraints, exploiting a post-normalization and a penalized reward. From the simulations, it is revealed the proposed algorithm provides better energy performance than benchmarks, highlighting the benefits of STAR-RIS in the CF network.

The exponential growth in mobile traffic has driven significant interest in cell-free networks, particularly those integrated with satellites, to support high data rates and overcome geographic limitations. In this paper, we propose a novel dynamic clustering and beamforming control that maximizes the minimum data rate in satellite-assisted user-centric cell-free networks. Specifically, we formulate a nonconvex max-min fairness problem to optimize clustering and beamforming under a user-centric cluster size and transmit power constraints. To find a solution, this problem is decomposed into two subproblems: clustering and beamforming. A modified Gale-Shapley-based user-centric clustering algorithm is proposed for the clustering subproblem, which is solved on a long-term basis using statistic channel information. The beamforming subproblem transform is transformed into a semidefinite problem using linear approximation and a linear matrix inequality. We propose a robust beamforming algorithm that considers imperfect instantaneous channel state information, solved on a short-term basis. Lastly, we analyze the computational complexity, revealing that the proposed scheme has polynomial complexity. We evaluate its performance under user mobility in satellite-assisted cell-free networks. The proposed algorithm offers significant performance gains, achieving up to a 15.15% higher minimum data rate and a 14.14% higher average data rate than zero-forcing beamforming and distance-based clustering benchmark schemes.

Noise in medical imaging is an inevitable challenge, often stemming from acquisition artifacts, varying imaging protocols, and external interference. While some studies suggest that noise can enhance model robustness, excessive or unstructured noise degrades training quality and classification performance. This issue is further exacerbated in federated learning settings, where individual clients have limited local data, making it difficult to train robust models independently. Federated imputation has been explored as a solution, yet existing methods do not fully leverage federated learning settings for optimal noise reconstruction. In this work, we introduce a novel encoder-decoder based federated imputation method, designed to replace noisy images with more representative reconstructions before training. Experimental results demonstrate that classification models trained with images imputed by the proposed method consistently outperform those trained with raw noisy images and without noisy images, highlighting the importance of effective noise handling in federated learning-based medical imaging.

Semantic communication shifts the focus from bit-level accuracy to task-relevant meaning. However, most methods assume equal importance across semantic units and rely on costly retraining, limiting scalability. This work proposes an importance-aware framework that accounts for unequal feature contributions. It introduces a new metric, importance-weighted semantic spectral efficiency (wSSE), to prioritize task-relevant features. It develops an empirically derived feature-accuracy matrix, inspired by saturation behavior. The framework enables importance-aware feature selection and subchannel allocation without online learning. This framework targets resource-constrained edge systems such as IoT cameras, UAV detection, and AR/VR. Experiments show up to 73.3% higher efficiency under constrained resources.

Advancements in drone technology and high-frequency millimeter-wave communications are transforming unmanned-aerial-vehicles (UAV)-aided networks, expanding their potential across diverse applications. Despite the advantages of broad frequency bandwidth and enhanced line of sight connectivity in the UAV-aided millimeter-wave networks, it is challenging to provide high network performance because of the inherent limitations of limited UAV energy and millimeter-wave’s large path loss. This challenge becomes more important in dynamically changing multi-UAV environments. To address this challenge in multi-UAV networks, we propose a novel approach based on multi-agent deep reinforcement learning called action-branching QMIX. Our method determines nearly optimal codebook-based discrete beamforming vectors and UAV trajectories while maintaining a balance between communication efficiency and energy consumption. The proposed approach employs a new Long Short-Term Memory module to control long sequences effectively and enables it to adapt to changing environmental variables in real-time. We thoroughly evaluate the proposed control with a real-world measurement-based channel model. The evaluation confirms that the proposed control converges stably and consistently, and provides enhanced performance in terms of downlink data rate, success rate of reaching the destination, and service duration when compared to traditional benchmark multi-agent reinforcement learning schemes. These results emphasize the enhanced energy sustainability, robustness, and stability of the proposed approach in dynamically changing multi-UAV environments when compared to the existing benchmark algorithms.

Federated learning (FL) is a framework for realizing distributed machine learning in an environment where training samples are distributed to each device. Recently, FL has employed over-the-air computation enabling all devices to transmit learning model updates simultaneously. This work proposes a communication-efficient sparse one-bit analog aggregation (SOBAA) method, incorporating new power control, layer-wise scaled one-bit quantization, layer-wise sparsification, and an error-feedback mechanism. We derive a tight upper bound of the expected convergence rate of the proposed SOBAA as a closed-form expression. From this expression, we explicitly identify the relationship between the convergence rate and compression and aggregation errors. Based on the theoretical convergence analysis, we formulate a joint optimization problem of the compression ratio and power control to minimize compression and aggregation errors, leading to the fastest convergence. In each communication round, the optimization problem is decomposed, and solved in a computationally efficient and feasible way. From this solution, we characterize the trade-off between learning performance and communication cost. Through extensive experiments on well-known MNIST and CIFAR-10 datasets, we confirm that the proposed method provides an enhanced trade-off performance between test accuracy and communication costs and a faster convergence rate than the other state-of-the-art methods. In addition, it is proven that the proposed method is more effective for more complex datasets and learning models.

Federated learning (FL) has emerged as a promising distributed machine learning technique. It has the potential to play a key role in future Internet of Things (IoT) networks by ensuring the security and privacy of user data combined with efficient utilization of communication resources. This paper addresses the challenge of maximizing energy efficiency in FL systems. We employed simultaneous wireless information and power transfer (SWIPT) and multi-carrier non-orthogonal multiple access (MC-NOMA) techniques. Also, we jointly optimized power allocation and central processing unit (CPU) resource allocation to minimize latency-constrained energy consumption. We formulated an optimization problem using a Markov decision process (MDP) and utilized a deep deterministic policy gradient (DDPG) reinforcement learning algorithm to solve our MDP problem. We tested the proposed algorithm through extensive simulations and confirmed it converges in a stable manner and provides enhanced energy efficiency compared to conventional schemes.

This paper proposes a flexible and efficient access control scheme that combines the orthogonal frequency division multiple access and rate-splitting multiple-access techniques for enhancing the uplink transmission in a digital twin edge network system. We formulate a non-convex mixed integer optimization problem that minimizes the energy consumption of all Internet of Things devices (IoTDs) and maximizes the number of successful IoTD tasks. To this end, we propose a deep reinforcement learning (DRL) framework by normalizing a DRL training algorithm named deep deterministic policy gradient for efficiently designing the variables while ensuring the problem constraints. However, in the inference stage, the proposed DRL method may encounter different devices and services. Therefore, we design an exhaustive-improved DRL method that can improve the proposed DRL effectively using information from a digital-twin module. We also propose a mathematical approximation-based solution employing two convexification approach: Dinkelbach's method and relaxed Linear Matrix Inequality (LMI). Through extensive simulations over different parameters and scenarios, we identify the polynomial complexity, stable convergence, and operating regime of the proposed solutions. It is also confirmed that the proposed approaches work well even with digital twin defects and provide improved performance in terms of energy consumption and number of successful tasks in comparison with benchmark schemes.

One of the essential factors for enabling sixth-generation systems is efficiently ensuring diverse quality-of-service (QoS) performance metrics to support the upcoming massive ultra-reliable low-latency communication (URLLC). This work proposes efficient transmission control in intelligent reflecting surface (IRS)-assisted nonorthogonal multiple access (NOMA) networks in the finite blocklength (FBL) regime that statistically guarantee stringent URLLC QoS requirements. Thus, we formulate a nonconvex problem that maximizes the sum effective capacity (SEC) while ensuring statistical delay QoS constraints. To make the problem more tractable, we propose a tight upper bound for the objective function based on Jensen's inequality and employ the concept of opportunistically minimizing an expectation. Then, we decompose the problem into two subproblems: active beamforming at the base station and phase-shift optimization at the IRS. Each subproblem is convexified by employing slack variables, penalty functions, and linear approximation, and solved using successive convex approximations. The subproblems are iteratively solved until convergence using alternating optimization. The convergence to a suboptimal stationary solution and the computing complexity of the proposed algorithm are rigorously analyzed. Finally, extensive numerical evaluations confirm that the proposed control in the FBL regime significantly improves the SEC under various QoS parameters compared to existing benchmark schemes. In particular, as the number of antennas and IRS elements increases, the proposed method becomes more efficient than the semi-definite relaxation-based approach in terms of complexity and performance.

In CRNs, it is crucial to develop an efficient and reliable spectrum detector that consistently provides accurate information about the channel state. In this work, we investigate a CSS in a fully-distributed environment where all secondary users (SUs) are equipped with directional antennas and make decisions based solely on their local knowledge without information sharing between SUs. First, we establish a stochastic sequential optimization problem, which is an NP-hard, that maximizes the SU’s detection accuracy by the dynamic and optimal control of the energy sensing/detection threshold. It can enable SUs to select an available channel and sector without causing interference to the primary network. To address it in a distributed environment, the problem is transformed into a decentralized partially observed Markov decision process (Dec-POMDP) problem. Second, in order to determine the best control for the Dec-POMDP in a practical environment without any prior knowledge of state–action transition probabilities, we develop a multi-agent deep deterministic policy gradient (MADDPG)-based algorithm, which is referred to as MA-DCSS. This algorithm adopts the centralized training and decentralized execution (CTDE) architecture. Third, we analyzed its computational complexity and showed the proposed approach’s scalability by the polynomial computational complexity, in terms of the number of channels, sectors, and SUs. Lastly, the simulation confirms that the proposed scheme provides enhanced performance in terms of convergence speed, accurate detection, and false alarm probabilities when it is compared to baseline algorithms.

Compared to static time-division duplexing (TDD), dynamic TDD (D-TDD) has significantly increased the spectral efficiency of cellular networks. However, conventional systems operate based on exact channel state information, resulting in high communication overhead and delay. Spatial deep learning refers to using spatial geographical information as the training data. This study investigates a spatial deep learning-based D-TDD scheme for 6G hotspot networks. First, we represent geographical location information in forms of traffic demand density grid matrices. Second, we use spatial convolution filters to extract discriminative features of uplink and downlink service gains and harms, taking the traffic demand density grid matrices as the input. Subsequently, extracted feature matrices are processed with sparse convolution blocks to reduce computation cost for the classification. Finally, we develop novel deep dueling neural networks, leveraging the extracted features to efficiently learn the near-optimal radio slot configurations for all base stations. Numerical results show that the proposed approach improves average rate per user by 2.5%, 6%, and 523.3% over those achieved in state-of-the-art centralized D-TDD, the competitive reinforcement learning, and greedy approaches, respectively. In addition, the proposed approach achieves up to 98.7% of the data rate performance of the optimum scheme with an exhaustive search algorithm.