Efficient exploration remains challenging in continuous-control environments with sparse rewards or complex dynamics, because disagreement estimates computed in raw state-action spaces can be dominated by unstable representations and local transition discontinuities. This paper proposes VIB-MaxinfoSAC, an information-gain-inspired exploration method that uses a Variational Information Bottleneck (VIB) to regularize the representation on which epistemic uncertainty is evaluated. An ensemble of β-VAEs compresses state-action pairs while predicting physical state residuals and rewards, so that the latent variables retain decision-relevant dynamics rather than reconstructing every input dimension. Disagreement among the ensemble latent mappings defines an entropy-related intrinsic-reward proxy, which is incorporated into policy learning through a dual-temperature adaptive objective. Across five state-based continuous-control tasks, VIB-MaxinfoSAC improves early exploration in sparse-reward environments, remains competitive in dense locomotion tasks, and shows lower training variability than the evaluated baselines in locomotion tasks under the selected configuration. Ablation and latent-space analyses further show that the bottleneck constraint limits latent divergence while preserving transition-relevant structure.
The growing use of Internet of Things (IoT) applications in future Sixth Generation (6G) networks necessitates a highly efficient, reliable, and scalable communication network system. However, traditional communication systems had poor spectrum efficiency, high energy consumption, low scalability, and poor blocking signals since they did not have the capability of manipulating the wireless communication environment. Thus, this research proposes a Graph-Aware Hybrid Optimisation and Deep Reinforcement Learning (GAH-DRL) for energy optimisation in Reconfigurable Intelligent Surface (RIS)-Assisted Non-Orthogonal Multiple Access (NOMA) in 6G IoT Networks. The proposed GAH-DRL is different from conventional RIS-assisted DRL schemes because of the architecture of two timescales for the proposed graph. The NOMA user groups and SIC relationships are determined by channel-correlation graphs and constraint-feasible reference allocations are generated by fractional programming, alternating optimisation, PSO and KKT refinement. A ResNet surrogate learns a starting power and RIS configuration while a DQN controller takes discrete residual actions to quickly adapt to changing network conditions. This eliminates the need for optimisation on the internet, while maintaining the feasibility of the QoS, SIC and transmit power. The evaluation results reveal that the proposed model attained a sum rate of 205 bps/Hz, energy efficiency of 48 bits/J, spectral efficiency of 17.5 bps/Hz, convergence rate of 130 iterations, and QoS of 88%.
The integration of orbital angular momentum (OAM) multiplexing with non-orthogonal multiple access (NOMA) promises high spectral efficiency in millimetre wave (mmWave)/terahertz (THz) line-of-sight (LoS) channels, yet two questions remain unresolved: does OAM-NOMA provably outperform conventional NOMA, and how does beam misalignment degrade performance? This paper answers both. It proves mathematically that OAM-NOMA achieves strictly higher spectral efficiency than single-mode NOMA for any positive signal-to-noise power ratio (SNR) and any L ≥ 2 active modes, with the capacity ratio approaching L at high SNR. A closed-form Bessel-function model – exact in the continuous-aperture limit and accurate to within 1% for practical array sizes – quantifies how angular offset θ creates inter-mode interference (IMI). To suppress IMI, a direction-of-arrival (DoA) estimation framework using the multiple signal classification (MUSIC) algorithm on a uniform circular array is developed; residual interference after compensation decays as O((kRΔθ)2m). The Cramér–Rao bound (CRB) for finite-snapshot DoA estimation is derived, and the MUSIC estimator is shown, through simulation, to closely track the CRB, with its root-mean-square error (RMSE) remaining within ± 16% of the CRB over the investigated SNR range while following the expected 1/SNR scaling. Benchmarking against five state-of-the-art schemes under a unified signal-to-interference-plus-noise power ratio (SINR) formulation, the proposed DoA-MUSIC reaches 14.71 bps/Hz at 10∘ misalignment and 20 dB SNR – at par with the best codebook method in raw capacity, but, unlike it, requiring no beam-index feedback, no pre-designed codebook, and remaining effective beyond the codebook’s angular coverage range. A Pareto complexity-performance analysis identifies practical deployment scenarios for 6G mmWave/THz systems. Moreover, outage and block error rate analyses under Rayleigh fading show that uncorrected misalignment creates an SNR-independent error floor, whereas sub-degree DoA accuracy (attainable by MUSIC for SNR ≥ 5 dB) eliminates this floor and restores near-ideal reliability. Additional sensitivity analyses confirm robustness to imperfect channel-state information (CSI) and hardware phase noise, demonstrate sub-degree tracking accuracy under time-varying misalignment, and benchmark measured runtimes consistent with the theoretical complexity ordering. All results are validated by extensive simulations.
Reliable underwater wireless communication is essential for emerging Internet of Underwater Things (IoUT) marine networking applications such as ocean monitoring, autonomous underwater vehicle (AUV), and offshore platforms. This paper investigates the performance of a mixed underwater wireless optical communication (UWOC) and radio frequency (RF) system assisted by multiple reconfigurable intelligent surfaces (RIS). The RF link is modeled using Nakagami-m fading, where each RIS follows a non-identical distribution (n.i.d), thereby capturing realistic spatial variations across reflecting units. The UWOC link is characterized by mEGG fading model to represent the combined effects of absorption, scattering, and turbulence in underwater channels. Closed-form expressions are derived for outage probability, average bit error rate (ABER), and achievable throughput, providing a unified framework for performance evaluation. Numerical and simulation results confirm the accuracy of the analysis and demonstrate that employing multiple RIS units significantly enhances link robustness, reduces error rates, and improves throughput under harsh underwater conditions. The findings highlight the potential of multi-RIS assisted mixed UWOC-RF systems to enable reliable and high-capacity underwater communication networks for next-generation marine and ocean engineering applications.
As a key technology for information security, covert communications research has been mostly confined to the design of signal detection evasion strategies in single-warden scenarios. However, the randomness of spatial positions of multiple wardens significantly increases the complexity of covert strategy design, posing severe challenges to covert signal transmission. This paper investigates the performance of relay-assisted covert communication systems in non-orthogonal multiple access (NOMA) networks with spatially random wardens, modeled by homogeneous Poisson point process, (PPP). In the proposed system, Alice transmits both public and covert signals to far and near users via a relay. Two operation modes are considered: a passive mode where the far user does not send interference, and an active mode where it does, to quantify the impacts of external interference on system covertness and reliability. In terms of monitoring mechanisms, a dual framework of independent detection and collusive detection is constructed: (i) for independent detection, we derive the closed-form expression of detection error probability (DEP) for the Kth nearest warden under the passive mode; (ii) the analytical expression of DEP under the active mode; (iii) the minimum DEP curves for both user modes through simulations. For collusive detection, we obtain the closed-form solution of DEP under the passive mode, and demonstrate that the covert performance of collusive detection significantly outperforms that of independent detection. Additionally, the outage probabilities (OP) of far and near users under different modes are derived to evaluate system reliability. Simulation results show that external interference has a negligible effect on the detection performance of wardens. Due to spatial randomness, the DEP in both modes first decreases with the increase of K, challenging the conventional wisdom of “closer distance leads to better detection performance”.
