Abstract : This paper investigates a downlink short-packet communication system enhanced by Simultaneously Transmitting and Reflecting Reconfigurable Intelligent Surface (STAR-RIS) and Non-Orthogonal Multiple Access (NOMA), aiming to maximize the achievable rate in a multi-user network. The system optimizes power allocation coefficients for the overall system and individual users, along with the energy coefficients and phase shifts of the STAR-RIS under the mode-switching (MS) protocol. Unlike conventional RIS, STAR-RIS simultaneously reflects and transmits signals, extending signal coverage and improving network performance. To address the complex, non-convex optimization problem, we apply Deep Reinforcement Learning (DRL) techniques, specifically the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm. TD3 optimizes the policy using channel state information as input, generating output for system power allocation and STAR-RIS parameters. Numerical results demonstrate the effectiveness of DRL in reducing computational complexity, while requiring minimal training data and time.
Index terms : Deep reinforcement learning, simultaneous transmitting and reflecting reconfigurable intelligent surface, non-orthogonal multiple access, short-packet