Model-Based Multi-Agent Reinforcement Learning for Joint Port and Precoding Optimization in Multi-Cell Fluid Antenna System

Fluid antenna is regarded as one of the most promising technology for next-generation wireless communication system due to its ultimate flexibility. Previous works have extensively studied fluid antenna multiple access (FAMA) in single-cell scenarios. However, in multi-cell environments, the centralized computation of transmission strategies for each base station (BS) is challenging due to inter-cell interference and user privacy concerns. Moreover, the large number of users is expected to introduce substantial protocol overhead, which may hinder the release of FAMA's full potential. In this paper, we propose a model-based multi-agent reinforcement learning algorithm, where each BS is treated as an independent agent and adopts a distributed approach to solve precoding and port selection problems. Using a world model, each BS can predict its future observations, enabling it to perform multi-step decisionmaking based on the current state, which has significant potential for reducing protocol overhead. Experimental results validate the superiority of FAMA over fixed antenna systems in multi-cell scenarios and demonstrate the potential of the proposed model-based decision mechanism in reducing transmission time.