Efficient allocation of scarce radio, computational and semantic resources is needed in Semantic Communications in IoV due to fast variations in traffic and channel conditions. Currently available approaches fail to have a prediction component, are oblivious to the semantic uncertainties and do not take into account the effect of potential delays on the future semantics. In this paper, we propose a uncertainty aware resource allocation framework using causal modeling. We construct latent structured representation of the vehicular environment and employ intervention analysis of the effect of sending particular messages to predict future semantics. Semantic uncertainties related to sensing, model, communication, and out-of-distribution cases, along with semantic utility, freshness, and safety relevance are considered. Priority-Guided AO-SCA framework optimizes jointly the message schedule, semantic fidelity, bandwidth, transmission power and computational resources usage whereas Top-K candidates filter prevents the combinatorial explosion problem. Risk-based semantic regret of delay and risk-aware age of semantics measures are introduced. Evaluation covers nominal, risky, dense traffic, high mobility, fading, sensor degradation, and out- of-distribution cases. In addition to ablation, sensitivity and runtime experiments, it was revealed that our proposed framework achieves 0.89 semantic utility, 0.92 semantic reliability, 0.95 task success rate and E2E latency of 11.40 ms, while being able to meet latency deadlines with a probability of 98.2% under the nominal case. Our framework outperforms both traditional and semantic aware resource allocation approaches.
The growing importance of Unmanned Aerial Vehicles (UAVs) in various applications has driven significant research aimed at enhancing their capabilities. Efficient communication during missions requires high routing accuracy, strong link quality, optimal throughput, and energy efficiency. Recent studies have focused on leveraging UAV networking functionalities within a joint optimization framework that can adapt to dynamic topologies and varying interference conditions. As a result, this paper proposes a novel framework that integrates Graph Neural Networks (GNNs) and Double Deep Q-Networks (DDQN) to simultaneously optimize routing and resource allocation in decentralized UAV networks. Specifically, the GNN component extracts topology-aware features to enhance routing accuracy, while the DDQN adapts transmission power and relay selection through reinforcement learning. To enable UAVs to make autonomous decisions based on local observations, the framework employs a Partially Observable Markov Decision Process (POMDP). Additionally, we propose a decentralized, scalable algorithm that enables UAVs to make autonomous routing and power-allocation decisions using local and neighborhood information, eliminating the need for centralized control. The results demonstrate that the GNN-DDQN framework outperforms existing methods across key performance metrics, including loss convergence, connection reliability, path length, and throughput. Specifically, in uniform distributions, GNN-DDQN surpasses Full MLP, Rand MLP, and c2Dst by approximately 23.5%, 75%, and 14.2%, respectively, and achieves 9.3 times the throughput of mSINR. In clustered formations, it outperforms the same methods by approximately 5.4%, 9.5%, and 23.4%, respectively, and achieves 13.3 times the throughput of mSINR.
Reliable channel estimation under pilot-aware interference requires enough pilot information to remain after nuisance removal. We study compact structured interference in vehicular OFDM receivers. Projection-Aided Candidate Evaluation for Channel Estimation (PACE-CE) selects a receiver-designed finite tone, chirp, or multitone nuisance family, jointly marginalizes its coefficients, and exposes residual, score-separation, and projected-conditioning diagnostics for validation-frozen fallback. Projection and fixed-candidate joint least squares are classical; the contribution is their receiver-side combination with finite-family selection, explicit operating boundaries, and mobility-aware use. In 200-trial paired tests, PACE-CE improves channel NMSE over same-dictionary sequential cancellation by 3.49–13.67 dB for matched or near-matched interference, with 0.93–1.00 win fractions. Against a stronger validation-covariance IA-LMMSE diagnostic, PACE retains 1.06 and 3.66 dB mean advantages in exactly matched chirp and multitone cases but loses in off-grid, reactive, frequency-hopping, and dense cases. An estimator-level IEEE 802.11p experiment uses standard waveform and pilot/training positions; it confirms the matched benefit but not an off-grid or out-of-family advantage. A validation-fitted classifier using receiver-observable diagnostics tracks evaluation-only oracle usefulness with test ROC AUC 0.968. The evidence supports an identifiable compact-interference front end, not a general anti-jamming or deployment-ready V2X receiver.
Urban vehicular environments have stimulated the use of millimeter-wave (mmWave) communication with three-dimensional (3D) Massive Multiple-Input Multiple-Output (MIMO) technology to support Vehicle-to-Vehicle (VTV) connectivity by the requirement of high capacity, low latency, and reliability. Nevertheless, vehicle mobility is very high, path loss is high, channels are dispersed, and non-line-of-sight (NLOS) and dynamic beam misalignment caused by the complex urban structure make efficient channel estimation and beam tracking very challenging. To overcome these constraints, this paper proposes a single unified spatiotemporal-aware learning model named DeepPeepNet that simultaneously achieves high-fidelity channel estimation and beam tracking optimization of mmWave-3D massive MIMO VTV communication. DeepPeepNet combines a Dual-Stage Residual Attention Network (DeepResATNet) with a Peephole LSTM-Convolutional Network (PeepHLSTMCoNet) to learn to extract strong spatial-temporal features on fast-varying vehicle channels. The first stage is the selective enrichment of dominant spatial components that are linked to multipath reflections by the Spatial Attention Module (SAtM) and the prevention of irrelevant noise. The Temporal Attention Module (TAtM) gives time-varying features, including Doppler shifts and short-term channel fluctuations, adaptive importance. Carrying over the connections also stabilizes learning in rapidly changing traffic conditions. The best features obtained are then fed through PeepHLSTMCoNet, in which the peephole-enabled LSTM gates are effective to handle the long-term temporal dependence, and the accurate beam tracking can be predicted in cases of sudden acceleration, lane changeovers, and line-of-sight/non-LOS (LOS/NLOS) transitions. The suggested framework is tested on the realistic 3D urban VTV scenarios with ultra-massive MIMO configurations. The performance is measured by Normalized Mean Squared Error (NMSE) and Probability of Estimation Error (PEE). Experimental results show that DeepPeepNet outperforms model-driven and learning-based baseline algorithms in both NMSE and PEE and beam alignment stability for different mobility patterns and SNR regimes. These results prove that DeepPeepNet offers a powerful, low-latency, and dependable solution to mmWave-enabled intelligent vehicular communication systems in urban high-density settings.
