Sum-Rate Optimization for Multi-User Fluid Antenna MISO Systems via ConvCNP-Assisted D3QN Port Selection

Abstract

This paper concentrates on the port-selection problem in a multi-user multiple-input single-output (MU-MISO) fluid antenna system (FAS). Since a huge amount of candidate ports are available, exhaustive search under full channel state information (CSI) introduces high probing and computational cost. A decision framework that combines a convolutional conditional neural process (ConvCNP) and a double dueling deep Q-network (D3QN) is proposed in this paper. The former utilizes a small portion of observed-port CSI obtained through pilot-aided least-squares (LS) channel estimation to infer the overall spatial channel distribution, while the latter employs local measurements, prediction results, and historical probing information to determine sequentially the next probing position and the final port configuration. Simulation results show that the proposed approach alleviates substantially the expense incurred by full CSI probing and exhaustive search while maintaining excellent sum-rate performance within a limited number of probing steps.