Collaborative DNN Inference in Maritime Edge Intelligence Networks with Group Neural Multi-Armed Bandits

Abstract

Collaborative Deep Neural Networks (DNNs) inference has emerged as a promising paradigm for growing number of artificial intelligence-integrated maritime Internet of Things (IoT) devices in maritime edge intelligence networks. However, the resource constraints of devices, the delay-sensitive nature of tasks, and the dynamic environmental conditions present significant challenges. While Multi-Armed Bandit (MAB) algorithms have been explored for task offloading, their performance is often constrained in highly dynamic scenarios with complex, nonlinear utility dependencies. To address these challenges, we propose a Group Neural MAB (GN-MAB) approach that jointly optimizes idle device selection (i.e., arm groups) and DNN partitioning decisions (i.e., arms) for efficient collaborative inference. Building upon the neural upper confidence bound algorithm, GN-MAB dynamically balances the exploration and exploitation, enabling continuous adaptation of offloading strategies across sequential inference tasks. Extensive experimental results show that GN-MAB outperforms baseline approaches, achieving superior inference performance while exhibiting robust adaptability to the fluctuating conditions of maritime environments.