Group Neural Bandits for Dynamic Collaborative Inference in Maritime Edge Intelligence Networks

Collaborative edge inference plays a vital role in executing deep neural network (DNN) tasks within maritime edge intelligence networks (MEINs). While existing studies have employed multi-armed bandits (MABs) for task offloading, they often struggle to generalize in highly dynamic environments with nonlinear utility dependencies. To address this challenge, we propose a group neural MAB (GN-MAB) learning approach that enables resource-constrained maritime devices to jointly optimize idle device selection (i.e., arm groups) and DNN partitioning strategies (i.e., arms) for collaborative inference. Leveraging the neural upper confidence bound (NeuralUCB) algorithm, our method dynamically balances exploration and exploitation, continuously adapting the offloading strategy across sequential tasks. Extensive experiments demonstrate that GN-MAB significantly improves inference performance while robustly adapting to the dynamic maritime edge environment.