This work investigates device‑to‑device (D2D) communication assisted by a reconfigurable intelligent surface (RIS) mounted on an unmanned aerial vehicle (UAV) under stochastic link activation. The system model captures UAV three‑dimensional trajectory and attitude, time‑varying Rician angles, and angle‑dependent RIS reflection coefficients.
A joint optimization problem is formulated to maximize the average sum‑rate while respecting mobility, energy, and hardware constraints. Decision variables include the UAV flight path, attitude angles, and the RIS phase‑shift matrix.
To tackle the high‑dimensional non‑convex problem, deep reinforcement learning (DRL) is employed together with a Decision Transformer. The transformer is pretrained offline on expert trajectories collected from multiple scenarios, enabling cross‑scenario policy generalization. Results show that zero‑shot transfer outperforms direct DRL transfer, and online fine‑tuning achieves performance comparable to conventional DRL with far fewer interactions.
Review