Resource Optimization in Active-STAR-RIS-Aided THz ISAC Systems With DDA Modulation: A Machine-Learning Approach

This paper explores the state-of-the-art terahertz (THz) integrated sensing and communication system (ISAC) that uses active reconfigurable intelligent surfaces (ASRIS) that can transmit and reflect signals at the same time. The system incorporates a novel dynamic delay alignment (DDA) modulation technique, allowing signals from different paths to reach the receiver simultaneously, eliminating the need for complex channel equalization and mitigating inter-symbol interference, while considering the dynamic movement of the vehicular units (VUs), and accounting for time-selective fading and uniform Doppler power spectra (DPS) model. Our system is equipped with a dual-function radar and communication multiple-antenna base station (BS), which simultaneously serves both communication and target sensing functions through an ASRIS. The objective is to maximize the sum rate by jointly optimizing BS transmit beamforming, ASRIS reflection and transmission beamforming matrices, VU mobility correlation parameters, and radar receive filter. Traditional optimization methods prove challenging given the intricate nature of this non-convex optimization problem, owing to dynamic changes in communication links and the interplay of multiple variables. To overcome this, we propose a machine learning (ML)-based multi-agent deep deterministic policy gradient (MADDPG) algorithm. MADDPG enables collaborative learning, adapts to the dynamic communication environment, and excels in optimizing interdependent parameters in the proposed THz system. Deep deterministic policy gradient (DDPG), proximal policy optimization (PPO), and modified-PPO (MPPO) algorithms serve as benchmarks, showcasing the distinctive advantages of the ML-based MADDPG solution for the proposed system's complexities. Our simulations validate the substantial benefits of ASRIS over conventional RIS benchmark scenarios.