Reinforcement Learning for Antenna Selection and Optimization of Irregular Reconfigurable Intelligent Surfaces

This paper presents a new reinforcement learning approach to the design and optimization of irregular reconfigurable intelligent surface (IRIS) for downlink communications in 6G multiuser wireless systems. Under the total power constraint of the IRIS device, we formulate a sum rate maximization problem that jointly optimizes the elements selection, the phase shift and the precoding design. For this challenging problem, we develop a deep reinforcement learning technique that can approach the optimal solution at affordable complexity. Physical constraints of the design parameters are properly incorporated into the developed DRL approach. Simulation results show that our proposed algorithm is able to learn from its environment and gradually improve its performance, and also converge to better performance compared to the state-of-the-art benchmarks when implemented in large-scale antenna systems.