DDPG Performance in THz Communications over Cascaded RISs: A Machine Learning Solution to the Over-Determined System

Abstract

THz technology is considered a key element in 6G wireless communication because it provides ultra-high bandwidths, considerable capacities, and significant gains. However, wireless systems operating at high frequencies are faced with uncertainty and highly dynamic channels. Reflecting intelligent surfaces (RISs) can increase the range of the THz communication links and boost the rate at the receiver. In contrast to the existing literature, we investigate the scenario of multiple access multi-hop (cascaded) RISs uplink THz networks in a correlated channel environment. We show that our inspected cascaded RIS system is over-determined and that the rate maximization optimization problem is non-convex. To this end, we derive a closed-form expression of the received power and derive an analytical solution based on pseudo-inverse to obtain optimum RISs' phase shifts that maximize the received signal power and hence increase the rate. In addition, we utilize deep reinforcement learning (DRL), which is capable of solving non-convex optimization problems, to obtain the optimum cascaded RISs' phase shifts at the receiver taking into account the situation of the spatially correlated channels. Simulation results demonstrate that the DRL algorithm achieves higher rates than the mathematical sub-optimal method and the case of randomized phases.