Dynamic Channel Allocation for Satellite Internet of Things via Deep Reinforcement Learning

Dynamic channel allocation is one of the most attractive and important issues to realize flexibility and efficiency of data transmission in satellite Internet of Things. However, the traditional random access methods are inefficient for the scenarios when the number of sensors exceeds certain limits and it also results in low transmission success rates. Moreover, existing heuristics resource allocation algorithms are not practical for such scenarios due to high computational complexity. To address these matters, we propose a centralized dynamic channel allocation method based on deep reinforcement learning (DRL), which is called CA-DRL. CA-DRL develops a novel representation for the channel allocation problem in satellite Internet of Things. It minimizes the average transmission latency of all the sensors by making smart allocation decisions with the powerful representation ability of deep neural networks through constant learning of allocation policies. We demonstrate high efficiency of CA-DRL in a simulated network environment and show that our proposed method can reduce data transmission latency by at least 87.4% compared with the current state-of-the-art channel allocation algorithms. As a consequence, it also results in significantly higher transmission success rates.