RESEARCH ON THE APPLICATION OF DEEP REINFORCEMENT LEARNING IN RECONFIGURABLE INTELLIGENT SURFACE COMMUNICATION SYSTEMS

Reconfigurable intelligent surface(RIS) is regarded as one of the key technologies of 6G. In actual deployment and application, a large number of passive phase shift units need to be optimized. Since the phase shift matrix composed of these elements has the limitation of single-mode constraints, the related optimization problems are complex and non-convex. To solve this problem, a Wolpertinger architecture is designed using deep reinforcement learning techniques for large antenna arrays with quantized phase shift units to predict and optimize the phase shift matrix, so as to search for the optimal solution more efficiently in a large discrete action space. Through numerical simulation in Python and DeepMIMO simulation platform based on real scenarios, the simulation results jointly show that compared with the traditional deep learning algorithm, the improved Wolpertinger architecture algorithm can further improve the acceptance signal-to-noise ratio and reduce the operation time, which effectively improves the quality of signal transmission in the RIS-assisted wireless communication environment.