Ang Li

Fluid antenna system (FAS) offers a promising solution to enhance spatial diversity in space-constrained user equipment. However, realizing its full potential in multi-antenna communication systems is hindered by the prohibitive pilot signaling overhead required for exhaustive port selection. Furthermore, existing learning-based selection schemes are mostly designed for single-user or non-precoded multi-user FAS, which are not directly applicable to multi-user precoded scenarios. To address these limitations, this paper proposes a port selection framework for multi-user precoded FAS with a novel learning-based Gaussian process upper confidence bound (GP-UCB) approach. Two system-level reward functions: one that minimizes the Frobenius norm of the zero-forcing (ZF) precoder, and another that maximizes the minimum singular value of the effective channel matrix are designed. These functions are optimized in an efficient manner using a GP-UCB framework, which avoids the exhaustive search procedure by iteratively measuring one port at a time and training a predictive model to infer performance across all unmeasured ports. Simulation results validate the efficiency of our framework, which accurately predicts the performance of all unmeasured ports with only a few channel measurements. Furthermore, the proposed schemes achieve substantial symbol error rate (SER) gains over conventional random port selection scheme and the channel gain based port selection method.

In this paper, we study the potential of interference exploitation symbol-level precoding (SLP), focusing on maximizing mutual information (MI) in multiple-input multiple-output (MIMO) systems with finite-alphabet inputs. As traditional MI expressions based on Gaussian signalling do not apply to SLP, we firstly derive the expression of the MI for SLP based on finite alphabet inputs, which is shown to be a nonlinear and non-concave function of the precoder. Subsequently, we formulate an optimization problem aimed at maximizing the MI without investing additional transmit signal power. Due to the original problem being non-convex, we design an iterative gradient-projection (GP)-based algorithm by deriving the gradient of the MI with respect to the SLP signal matrix. To alleviate the heavy reliance on Monte Carlo evaluations, we further propose an effective variant that employs the channel cut-off rate (CR) as a tractable optimizaiton metric within the GP framework. Moreover, we develop a low-complexity algorithm by deriving a closed-form approximation of the MI and addressing the resulting problem through successive convex approximation (SCA), thereby obtaining a near-optimal solution to the original formulation. Numerical results verify that the proposed MI-optimization SLP (MIO-SLP) exhibits a noticeable gain over traditional SLP schemes in terms of MI, and achieves close-to-optimal MI performance, while ensuring satisfactory error rate.