Generative reinforcement learning for fluid antenna systems based sensing and communication

Fluid antenna systems, which can flexibly reconfigure the radiating elements within a given space, is a promising technique for the sixth-generation (6G) networks. Nevertheless, the design of fluid antenna systems often involves the high-dimension mixed-integer non-convex optimization problems, and the existing optimization algorithms undergo the challenges of high computational complexity and poor performance owing to the enormous iterations or relaxation. This paper proposes a generative reinforcement learning framework, which leverages a primary-dual network (PD-Net) and the generative proximal policy optimization (GPPO) for the beamforming optimization and port selection, respectively. The proposed framework can obtain higher quality solution than four benchmarks. Via three cases of communication, sensing and integrated sensing and communication, the superiority of the proposed framework over benchmarks is validated. Moreover, the PD-Net can deal with the beamforming optimization problem with complicated equality and inequality constraints, and the GPPO algorithm exhibits better performance and stability than the proximal policy optimization benchmark, verifying the effectiveness of generative adversarial training. Therefore, the proposed framework can facilitate effective fluid antenna systems design for sensing and communication. In addition, the proposed framework provides a novel and effective paradigm to deal with the mixed-integer non-convex optimization problems, bypassing the need for relaxation and training labels, and can be extended to enormous similar problems.