Minimizing URLLC Task Offloading Latency with Full-Duplex STAR-RIS-Aided DRL-ISAC Systems

This paper investigates the deployment of a full-duplex integrated sensing and communication (ISAC) system for task offloading service to serve ultra-reliable low-latency communications (URLLC), significantly enhanced by a simultaneously transmitting and reflecting reconfigurable intelligent surface (STAR-RIS). Utilizing a non-orthogonal multiple access frame-work, this study extensively addresses the challenges associated with latency-sensitive task offloading from information receivers (IRs) to a mobile edge computing platform. We introduce a novel and sophisticated multi-agent deep reinforcement learning (MA-DRL) approach aimed at minimizing latency in task offloading under a variety of stringent ISAC-URLLC network constraints, including imperfect channel state information. The proposed MA-DRL operates on decentralized execution while maintaining centralized training, using multi-actor-critic networks to enhance learning and performance. The innovative reward decentralization framework in the present MA-DRL optimizes downlink and uplink communications through dynamic power allocation, precise beamforming, and intelligent phase shift management facilitated by the STAR-RIS. The proposed MA-DRL framework significantly outperforms existing multi-agent DRL algorithms, demonstrating substantial gain in reward maximization (i.e., linked to offloading latency minimization) by 27.46% and 52.73%.