Heterogeneous temporal graph powered DRL algorithm for channel allocation in Maritime IoT Systems

Abstract

In actual maritime Internet of Things systems, the communication environment is characterized by its time-varying nature and the presence of highly heterogeneous network structures. Those attributes present considerable challenges in devising resource allocation strategies. Given the limited availability of frequency resources, designing a reasonable and flexible channel allocation strategy (CAS) is the primary task for meeting diverse and dynamic communication demands. In this paper, a heterogeneous temporal graph powered deep reinforcement learning algorithm is proposed to optimize the CAS to maximize the channel efficiency in a real-world maritime Internet of Things system. Specifically, we build relation-based heterogeneous edges to connect different types of terminal nodes and adopt a time encoding technology to capture the dynamic evolution of communication scenarios over time. The memory and public mailbox modules are constructed as the implementation entities of the information aggregation method based on the attention mechanism. In addition, we develop a corresponding heterogeneous temporal neural network to estimate the real-time resource requirements of target terminals, and subsequently learn the optimal CAS based on the deep reinforcement learning algorithm from the perspective of maximizing the cumulative channel efficiency. Simulation results prove that the proposed algorithm significantly outperforms the other state-of-the-art algorithms in terms of the channel efficiency and generalization ability.