Online Hierarchical Computation Offloading for Marine IoT Networks: A Delay Minimization Approach

Mobile edge computing (MEC) has emerged as a promising technology for marine Internet of Things (IoT) networks, supporting diverse application requirements that could be both computationally intensive and delay-sensitive. However, most existing studies assume access to pre-existing network information and rely on single-layer MEC frameworks to provide services from an offline perspective, struggling to ensure low latency. To overcome the related issues, we first consider an online hierarchical computation offloading framework in this paper for marine IoT networks with aerial, offshore, and onshore devices. We further develop a hybrid transmission strategy combining non-orthogonal multiple access (NOMA) and frequency division multiple access (FDMA) to enhance the computation offloading efficiency within the framework. Considering the time-varying capacity of wireless channels, we thus minimize the hierarchical computation offloading delay by jointly optimizing the offloading strategy and network resource allocation in the marine IoT networks online. To solve the formulated mixed-integer nonlinear programming (MINLP) problem, we design a problem-solving framework based on a decomposition structure. Specifically, we decompose the formulated MINLP problem into two subproblems. For the bottom subproblem, we design a successive convex approximation (SCA)-based algorithm to optimize the hierarchical transmission durations and the offloaded workload with a given user association scheme. For the top subproblem, we propose a deep reinforcement learning (DRL)-based algorithm to realize online optimization of the user association scheme under the time-varying channels. Finally, numerical results demonstrate that the proposed algorithms, including the SCA-based algorithm and the DRL-based algorithm, can reach near-optimal results. Furthermore, the proposed hierarchical computation offloading framework significantly outperforms traditional benchmarks.