Learning to Configure Passive RIS on High-Altitude Platform Stations for Energy-Efficient IoT Communications

Abstract

Internet of Things (IoT) devices operating in remote, rural, or disaster-affected areas often suffer from blocked or weak connections to terrestrial Base Stations (BSs). This paper proposes a High Altitude Platform Stations (HAPSs)-assisted uplink architecture in which a passive Reconfigurable Intelligent Surface (RIS) mounted on HAPS creates a virtual Line-of-Sight (LoS) link between ground IoT devices and the BS. To maximize Energy Efficiency (EE), we optimize the discrete RIS phase-shift configuration under dynamic wireless channel conditions. The resulting optimization problem is combinatorial and computationally expensive for real-time operation. To address this challenge, we formulate RIS phase-shift selection as a single-step Markov Decision Process (MDP); formally a contextual bandit; and develop a Deep Q-Network (DQN)-style action-value estimator that directly predicts near-optimal phase-shift codewords from channel state observations, eliminating the need for online exhaustive search. The proposed DQN is evaluated against a fixed-phase RIS scheme and and a codebook-constrained exhaustive-search benchmark. Simulation results show that the DQN converges within a few thousand training episodes, achieves EE within 10% of the exhaustive-search solution, and improves EE by more than 150% over the fixed-phase baseline while requiring only a single neural network, fixed-cost neural-network inference for online decision making, in contrast to the codebook-size-dependent search cost of exhaustive search.