Active Inference for Joint Port Selection and Pilot Allocation in Fluid Antenna Systems Under Partial CSI

Fluid antenna systems (FAS) enhance spatial degrees of freedom by activating a limited number of ports from a dense candidate grid. Previous studies primarily focused on scenarios with full channel state information (CSI) and treated pilot overhead as a constant expense. This study aims to reformulate the problem of port selection and pilot allocation as a joint decision-making process through active inference. Utilizing a spatio-temporal generative model that incorporates sub-wavelength spatial correlation and an autoregressive temporal prior, our method maintains a Gaussian belief over all ports, allowing the transmitter to select which ports to activate and pilot. The selection process minimizes expected free energy, balancing achievable rates with information gain and the costs associated with switching ports. Our greedy and submodular selection algorithm operates at O(NM) per time slot. In simulations involving a 441-port grid with a hybrid front end at nRF = 2K, our approach achieved 91% of the sum rate attainable by a full-CSI system while measuring only 2.3% of the ports. Additionally, pilots can be reduced from served ports without significant loss in performance, facilitating efficient operation in scenarios with limited CSI.