A Reinforcement Learning Approach for Reconfigurable Intelligent Surface (RIS)-Assisted Multi-User Wireless Networks: Experimental Testbed Validation

To address the application challenges of reconfigurable intelligent surface (RIS) optimizations in practical wireless networks, this paper proposes a model-free and channel state information (CSI)-free reinforcement learning algorithm, named cross-entropy multi-armed bandit (CEMAB), for RIS phase shift optimization. The algorithm interacts with the environment and relies solely on user performance feedback, without requiring any prior channel knowledge. In particular, an importance sampling mechanism is employed to enhance exploration efficiency. The algorithm is experimentally validated on a scalable wireless testbed through two case studies, demonstrating superior performance against benchmarks in maximizing the received signal strength indicator (RSSI) for individual users and maximizing the minimum RSSI across users.