Task Scheduling Under a Novel Framework for Data Relay Satellite Network via Deep Reinforcement Learning

Abstract

Data relay satellite networks (DRSNs) face the challenge of increasing relay mission demands in space networks. To improve the task scheduling efficiency of DRSN further, we propose a novel task scheduling framework, wherein a scheduling sequence is generated by selecting one antenna and selecting one task for the antenna in each step. Subsequently, the task scheduling problem of DRSN (TSPD) is regarded as a sequential decision-making problem and is optimized using a method based on deep reinforcement learning (DRL), which overcomes the difficulty of designing heuristics with massive efforts. In this study, a mathematical model and corresponding Markov decision model based on our proposed scheduling framework are constructed, and for the first time, a policy network that includes one encoder based on the attention mechanism and two decoders is designed to solve the TSPD. In addtion, extensive experiments are conducted to verify the effectiveness of our scheduling framework and demonstrate that the DRL method can obtain a scheduling scheme with the highest profits with decent generalization to different task scales, number of user spacecrafts and execution duration of tasks.