CSI-Free Positioning of Movable Antennas for IoT Networks: A Compositional Kernelized Bandit

Abstract

Movable antenna (MA) arrays reshape the propagation channel by mechanically changing the element positions, which suits Internet-of-Things access points whose devices cannot adapt on their own. Existing position optimization assumes the channel is known at every candidate configuration. In practice a configuration can be evaluated only after the array has moved there, over several slots limited by the actuator speed, during which the channel changes. We formulate the positioning of M MAs serving K < M devices without channel state information as a nonstationary kernelized bandit over the joint configuration space, driven by a single scalar rate feedback per slot under a reachability constraint and an actuation-energy cost. Since the information gain of a standard kernel grows exponentially with M , and the multiuser sum rate depends on the positions through a sum of per-antenna terms, we design a compositional kernel that retains interactions up to antenna pairs. On this kernel we propose CoMoveUCB, which selects a target configuration over the whole feasible set by coordinate ascent and retains it under a persistence rule. We prove sublinear dynamic regret with a polynomial dependence on M . Simulations show that CoMoveUCB outperforms fixed, measure-then-optimize, and reachability-confined benchmarks across loadings, confirming the value of repositioning an MA array under realistic actuation limits.