Concept-Drift Aware RIS Tile Management in Dynamic VLC Networks

The concept of integrating reconfigurable intelligent surfaces (RISs) into visible light communications (VLC) has already been realized through metasurfaces such as adjustable-angle reflectors or liquid crystal tiles with variable refractive indices. However, current research neglects the power consumption and control signaling overhead required to drive these RIS tiles, potentially resulting in the overheads of RIS outweighing its benefits. Considering that some RISs will be blocked in indoor mobile environments and the blockage patterns will change over time, which results in concept drift for management strategy, this paper proposes a dynamic RIS tile exclusion strategy adaptive to indoor mobility. Such an RIS controller deactivates less efficient RIS tiles based on channel conditions, which reduces signaling and power consumption costs, thereby, enhancing the overall system efficiency. Since this decision problem is highly non-convex and non-linear, and involves RIS-assisted time-varying visible light channels that cannot be characterized by general mathematical models, this paper proposes a maximum entropy deep reinforcement learning-based method to optimize the strategy. In the validation scenarios, compared to the tile exclusion strategy at fixed time intervals, the proposed adaptive strategy improves the RIS utilization efficiency by over 100%. Moreover, it only takes less than $53.55~\mu $ s to execute the trained strategy on a low-cost micro-controller unit. The generalization ability of the learned strategy is also proven as no transfer learning or adaption is needed when facing a new environment.