ISCPT-SEE-AA: An Agentic AI-Driven SEE Maximization Approach for STAR-RIS Assisted ISCPT Networks

Abstract

This paper investigates the rate-splitting multiple access (RSMA)-enabled integrated sensing, communication, and power transfer (ISCPT) network assisted by a simultaneously transmitting and reflecting reconfigurable intelligent surface (STAR-RIS). Particularly, considering the inherent heterogeneous quality-of-service (QoS) requirements, communication users (CUs) are granted priority in information reception, and the legitimate sensing targets (STs) are also regarded as potential eavesdroppers to intercept the information of CUs. In order to meet the service demands of heterogeneous users of such a system while ensuring the physical layer security of CUs against wiretapping, we formulate a secrecy energy efficiency (SEE) maximization problem via jointly optimizing the transmit beamforming matrix, sensing matrix, STAR-RIS reflection/transmission coefficient matrix, power splitting (PS) ratio vector, and common rate allocation vector. Due to the non-convexity of the formulated problem and the challenges caused by imperfect channel state information (CSI) and dynamic wireless environment, an agentic-AI enabled optimization approach (ISCPT-SEE-AA) is proposed, which works in a closed-loop perception-decision–reward adaptation manner. Specifically, a Transformer-based channel refinement module is developed to mitigate the uncertainty induced by imperfect CSI. Meanwhile, a mixture-of-experts group relative policy optimization (MoE-GRPO) scheme is adopted to achieve adaptive decision-making under time-varying network conditions. Furthermore, a large language model (LLM)-aided reward configuration module with retrieval-augmented generation (RAG) is integrated to automatically configure and update reward parameters, thereby reducing manual tuning overhead and enhancing training stability. Extensive simulation results verify that the proposed ISCPT-SEE-AA outperforms conventional learning-based schemes significantly in terms of SEE and exhibits stronger robustness against CSI imperfections. Notably, it achieves performance close to that of the convex optimization-based benchmark with substantially lower online computational complexity.