TL;DR
This paper introduces a trajectory-prediction-based generative framework (TPGC) to manage multi-RIS wireless networks in dense, mobile environments. By combining an LSTM for multi-user trajectory forecasting with a Riemannian diffusion model operating on a torus manifold, the system jointly optimizes RIS ON/OFF activation and phase configurations to proactively suppress co-channel interference and maximize the signal-to-interference-plus-noise ratio (SINR).
What Problem Does This Paper Address?
Reconfigurable Intelligent Surfaces (RISs) are a cornerstone of emerging 6G networks, offering a low-cost way to dynamically shape the wireless environment and eliminate coverage gaps. However, RIS units are inherently passive and "blind"—they reflect all incident electromagnetic waves indiscriminately.
In dense mobile scenarios (such as marathons, crowded streets, or outdoor gatherings), this lack of selectivity becomes a liability. If multiple users operate in close proximity or share similar line-of-sight (LoS) paths, an RIS can inadvertently amplify co-channel interfering signals just as much as the desired signal. This leads to severe spectral pollution and degrades communication quality.
Existing RIS control methods typically assume static channels or rely on Euclidean optimization techniques that treat phase shifts as unconstrained real numbers. This ignores the intrinsic periodic nature ($0$ to $2\pi$) of phase shifts, leading to geometrically distorted and physically inconsistent solutions.
Methods
To tackle this mixed discrete-continuous, highly non-convex optimization problem, the authors propose the Trajectory-Prediction-Based Generative Control (TPGC) framework, which operates in three distinct stages per transmission frame:
- Kinematic-Enriched Trajectory Prediction: A two-layer Long Short-Term Memory (LSTM) network forecasts the future locations of both target and interfering users. By augmenting historical coordinates with velocity and heading angle features, the base station can accurately reconstruct future cascaded Channel State Information (CSI).
- Riemannian Diffusion Model (RDM): Instead of standard Euclidean diffusion models that suffer from boundary discontinuities at the $0$–$2\pi$ edge, the authors design an RDM that operates directly on the torus manifold $(S^1)^N$. It injects and removes noise within the manifold's tangent space, ensuring that every intermediate phase-configuration sample remains geometrically consistent.
- DRL-Guided Refinement: Because optimal phase labels are unavailable during training, a deep reinforcement learning (DRL) critic based on the Twin Delayed Deep Deterministic policy gradient (TD3) algorithm guides the reverse diffusion process, steering the generative actor toward high-reward (high achievable rate) regions.
- Optimal ON/OFF Switching: Armed with the optimized continuous phase configurations, the base station compares the predicted achievable rates under RIS activation versus deactivation. If an RIS panel is predicted to cause more harm than good due to unavoidable interference, it is dynamically switched to an absorptive (OFF) state.
Datasets
The framework was evaluated using real-world mobility data from the GeoLife trajectory dataset. This dataset contains rich spatio-temporal GPS traces (including longitude, latitude, and timestamps) collected from users in Beijing, providing a realistic baseline for human movement patterns during outdoor events.
Key Insights and Ideas
- To Reflect or Not to Reflect: Counterintuitively, keeping an RIS turned on at all times can actually harm network performance in crowded areas. Intentionally turning specific RIS panels off (putting them into an absorptive mode via weak PIN diode biasing) is a highly effective mechanism for interference mitigation.
- Manifold-Aware AI: Standard diffusion models inject Gaussian noise in Euclidean space, which corrupts the strict unit-modulus constraints of RIS phase elements. Performing diffusion directly on a torus manifold preserves the periodic structure of phase shifts throughout the entire denoising process, leading to faster convergence and better solutions.
- Proactive Interference Avoidance: By coupling trajectory prediction with generative beamforming, the network transitions from a reactive state (responding to current interference) to a proactive state (anticipating and avoiding geometric overlaps before they happen).
Key Results
- SINR Improvements: Extensive simulations show that the TPGC framework achieves up to a 30% SINR improvement over conventional learning-based control methods and up to a 44% gain compared to the naive "RIS Always-On" scheme.
- Trajectory Accuracy: The kinematic-enriched LSTM model successfully tracks complex user movements (including sharp turns), keeping prediction errors within $\pm 6$ meters.
- Robust Generalization: In cross-scenario evaluations, models trained under dense interference conditions generalized effectively when deployed in environments with fewer interferers, proving that the RDM-DRL architecture learns robust structural representations rather than overfitting to specific noise levels.
Limitations and Future Work
The limitations and future work are not discussed in detail in the paper. However, based on the architecture, promising directions include extending the centralized optimization framework to distributed multi-BS cooperative networks, accounting for hardware impairments or discrete phase-shift constraints in real-world PIN diodes, and evaluating performance in non-line-of-sight (NLoS) dominant urban canyons.
Summary updated