Model-Assisted In-Orbit Intelligent Computation for Satellite-Ground Integrated Networks

The satellite-ground (SG) integrated networks will revolutionize the sixth generation mobile networks, because it can provide wide coverage, large-scale connection and ubiquitous intelligence capabilities for numbers of ground users. Furthermore, the massive computing tasks can be either offloaded to satellites for the sake of performing in-orbit computation, or executed by local users. However, the stochastic task arrival and dynamic channel gains hinder the optimal task offloading and resource allocation schemes. Hence, we envision a two-layer SG network model to minimize the total overhead in terms of energy consumption and latency for ground users. Additionally, a model-assisted in-orbit deep reinforcement learning (MAI-DRL) policy is proposed to optimize the task scheduling, central processing unit cycle frequency and transmission power, which further incorporates the in-orbit edge computing and artificial intelligence. Finally, massive simulation results show that the proposed MAI-DRL policy has superior performance gains in terms of energy efficiency, average delay and total overhead compared with the baseline methods.