6G Semantic Intelligence for UAV Anti Jamming: A Transformer Augmented Predictive Deep Reinforcement Learning Framework

Full abstract

Read the full abstract

This paper proposes a semantic-aware predictive hybrid Proximal Policy Optimization (PPO) framework for UAV anti-jamming communications under partial observability and nonstationary interference. The framework integrates Transformer-based channel-wise interference prediction, adaptive semantic-rate control, energy-aware horizontal UAV mobility, and hierarchicaldiscrete–continuous control within a unified POMDP. The semantic payload is explicitly coupledwith achievable rate, transmission duration, delivery success, and propulsion–communication energy,while predictor learning is separated from the on-policy PPO update. Evaluation over 50 independenttrials shows that, at a jammer power of 30 dBm and CSI error variance of 10-3, the proposed methodachieves a semantic accuracy of 0.751 ± 0.005, a task-success rate of 0.715 ± 0.007, and an SINRof 6.48 ± 0.28 dB, while consuming 23.58 ± 0.20 J/slot. The closest LSTM-Hybrid PPO baselineobtains 0.694±0.006, 0.642±0.006, 5.37±0.26 dB, and 25.46±0.17 J/slot, respectively. The Transformer predictor achieves an RMSE of approximately 0.039 under perfect CSI and 0.091 at a CSIerror variance of 10-2. Under an unseen jammer-policy shift, online predictor adaptation recoverssemantic accuracy to approximately 0.859, compared with 0.746 for a frozen predictor. Ablation results indicate complementary contributions from prediction, semantic-rate adaptation, mobility, andhierarchical hybrid control.Keywords: Deep Reinforcement Learning, Predictive Anti-Jamming, Semantic Communication,Transformer-Based Prediction, UAV Networks. 04:28:10 dblclick "Deep Reinforcement Learning, Predictive Anti-Jamming, Semantic Communication, Transformer-Based Prediction, UAV Networks." sidepanelOpen=false autoUpdate=true isAdv=true