This paper proposes a multi-agent deep reinforcement learning framework for the joint optimisation of reconfigurable intelligent surface (RIS) phase configuration and non-orthogonal multiple access (NOMA) resource allocation in Internet of Things (IoT) networks. A complete system model is formulated, comprising Rayleigh-faded blocked direct links, Rician BS-RIS and RIS-user links, a discrete-phase RIS reflection model, and a NOMA signal model whose successive interference cancellation (SIC) decoding order is recomputed from the RIS-induced effective channel gains at every slot; the resulting sum-rate maximisation problem is stated formally with its power, phase, SIC and quality-of-service constraints. Two specialised deep Q-network agents, a RIS Configuration Agent and a NOMA Resource Management Agent, learn concurrently in the shared environment and coordinate implicitly, without inter-agent signalling. The RIS agent is developed both for a single shared phase and, through a branching action-value architecture, for element-wise phase control; the NOMA agent uses an action-to-power mapping that is proved to satisfy the SIC power-gap constraint for every admissible action. Under a physically calibrated link budget, and averaged over five independent training runs, the framework attains a mean sum rate of 20.6 Mbps, above the 19.8 Mbps of a randomly configured RIS with fixed power, level with a well-tuned system without a RIS (20.5 Mbps) and below orthogonal multiple access (23.4 Mbps), with per-user SINRs between − 6.1 and − 1.3 dB rather than the physically implausible values reported previously. Conventional alternating optimisation reaches a substantially higher mean sum rate of 42.5 Mbps but requires 33.6 ms per decision against 0.45 ms for the proposed framework, a gap that widens with RIS size. An ablation study isolates the contribution of each agent and of the coordination mechanism, and a generalisation analysis covers user counts from 2 to 8, RIS sizes from 8 to 64 elements and channel conditions from Rayleigh to Rician with a 10 dB factor. We also report two negative results: the element-wise branching extension does not improve on the shared-phase design under the training budget used, and the full dual-agent system does not improve on its NOMA-only ablation, which identifies credit assignment across the factored phase action space, rather than the RIS model, as the binding limitation. The results position the framework as a low-latency alternative to iterative optimisation rather than as a means of maximising spectral efficiency.
This paper investigates channel recovery in asymmetrical massive multiple-input multiple-output (MIMO) transceiver architectures under hardware impairments. In such systems, the number of antenna elements significantly exceeds the available receive RF chains, resulting in reduced channel observability and increased sensitivity to hardware-induced sensing mismatch. To address this challenge, a distortion-aware channel recovery framework is developed by explicitly incorporating antenna-domain distortions, RF-chain nonidealities, and switching imperfections into the sensing and recovery models. A distortion-aware orthogonal matching pursuit algorithm is then proposed to enhance support detection and estimation accuracy under non-ideal sensing conditions. Simulation results demonstrate that the proposed method outperforms conventional Orthogonal Matching Pursuit (OMP), Least Absolute Shrinkage and Selection Operator (LASSO), and the Uniform Planar Array (UPA) based modified Newtonized OMP method, achieving an order-of-magnitude reduction in normalized mean square error (NMSE), gains of up to 5 dB at moderate-to-high signal-to-noise ratios. The proposed method also maintains robust performance as the number of RF chains and channel sparsity vary. The computational complexity analysis shows that the proposed method retains the same dominant asymptotic complexity order as conventional OMP. Furthermore, the proposed framework lowers the error floor in distortion limited regimes and maintains robust performance as the number of RF chains and channel sparsity vary.
Full-duplex wireless powered communication networks (FD-WPCNs) have gained significant attention for sustainable wireless communications due to their capability to simultaneously perform wireless energy transfer (WET) and wireless information transfer (WIT). However, most existing FD-WPCN studies assume sufficient energy availability and rely on conventional harvest-then-transmit (HTT) protocols, which may limit energy utilization under stochastic wireless environments. To address these limitations, this study proposes a practical grid-assisted FD-WPCN framework in which a solar energy harvesting-enabled hybrid access point (HAP) is connected to the power grid (PG) to support continuous downlink energy transfer to multiple source nodes. Furthermore, an accumulate-then-transmit (ATT) protocol is adopted at the source nodes to enable flexible energy accumulation and improve long-term network sustainability. The resulting non-convex sum-throughput maximization problem is formulated as a reinforcement learning problem subject to queue stability and energy causality constraints. To efficiently address the coupled continuous resource-allocation problem under stochastic wireless dynamics, a deep deterministic policy gradient (DDPG)-based framework is developed to jointly optimize transmission time allocation, uplink transmit power, and downlink energy-transfer power. The proposed DDPG framework learns an adaptive resource-allocation policy directly from dynamic network states, enabling effective decision-making without repeatedly solving the underlying non-convex optimization problem. Extensive simulation results demonstrate that the proposed framework achieves superior throughput, improved queue stability, and efficient grid-energy utilization compared with benchmark schemes, particularly under dynamic and energy-constrained network conditions.
A four-port Symmetric Multi-Stub Loaded (SMSL) MIMO antenna is developed for 3.5 GHz Sub-6 GHz 5G wireless systems using a shared common-ground architecture with inherent isolation enhancement. The proposed antenna is derived from a single SMSL monopole through systematic current-path optimization and progressive ground evolution, resulting in controlled surface-current confinement and accurate resonance tuning. A sequential ground-engineering approach employing grounded and extended grounded stubs effectively suppresses mutual coupling under a unified ground configuration. The antenna is realized on an 87 × 87 × 1.6 mm3 FR-4 substrate (εr = 4.3), corresponding to an electrical size of 1.02λ₀ × 1.02λ₀ at 3.51 GHz. It provides a −10 dB impedance bandwidth of 3.38–3.62 GHz (240 MHz), equivalent to a 6.86
In remote surgery, industrial automation, and public safety services, which are mission-critical, ultra-reliable low-latency communications are necessary. However, conventional communication methods are unable to simultaneously satisfy the stringent requirements of strong security, high reliability, and ultra-low latency demanded by mission-critical applications. To overcome these limitations, this paper proposes a Federated Consensus Learning-based Latent Diffusion Model (FCLbLDM) for secure, reliable, and low-latency communications. Although the use of the Latent Diffusion Model optimizes system efficiency and strengthens data security by establishing an encrypted communication system and safeguarding sensitive data, it also distributes consensus and strengthens data integrity and reliability. This dual capacity has been used to significantly reduce communication delays with the proposed approach, achieving an ultra-low latency of 10.0144 ms and high reliability. Extensive experimentation in remote surgery, industrial automation, and public safety demonstrates how effectively work is carried out to ensure a safe, smooth environment. The FCLbLDM framework improves ultra-reliable, low-latency communication (URLLC) by enabling reliable, secure, and low-latency communication through the innovative combination of Federated Consensus Learning and LDMs, with optimization and encryption features.
