Generalized RIS Tile Exclusion Strategy for Indoor mmWave Channels Under Concept Drift

Reconfigurable intelligent surfaces (RIS) experience considerable control overhead, particularly problematic in mobile multi-user programmable wireless environments (PWEs). This paper presents a strategy that dynamically excludes shadowed RIS tiles from supporting non-line-of-sight mmWave communications, reducing overheads and enhancing RIS usage efficiency. Spatio-temporal variations caused by crowd mobility necessitate the dynamic adaption of this strategy. The primary challenge lies in identifying optimal occasions for updating RIS tile exclusion decisions, which must strike a balance between improving achievable channel gain (performance) and power consumption as well as controlling overhead (cost) associated with decision updates. Given the absence of a general mmWave channel model, this paper applies deep reinforcement learning (DRL) for managing the RIS exclusion strategy. DRL caters to the susceptibility of mmWave channels to concept drift, where spatio-temporal variations in crowd mobility alter the probability distribution of RIS channel gains and outages. To counter concept drift and generate a universal RIS exclusion strategy for any indoor mmWave environment, we propose an adaptive exclusion update mechanism powered by DRL. This mechanism utilizes hierarchical decision decomposition, reward signal embedding, and a fusion of concept drift and temporal features due to crowd mobility, enabling efficient adaptation to environmental changes. Extensive cross-validation confirms the agent’s impressive generalization ability, directly applicable to varied environments. This adaptive mechanism, despite containing only 2, 300 learnable parameters, achieves more than a two-fold increase in efficiency relative to static exclusion timing methods. Furthermore, decision execution, based on a low-cost RIS controller, only takes a few tens of nanoseconds, showcasing practicality and efficiency.