With the advance of the information technologies,the scale of the information services gradually expands,from ground services,to aerial,maritime,and spatial services,with the soaring requirements on multi-dimensional comprehensive information resources.The space-air-ground integrated networks (SAGINs) are envisioned to provide seamless network services to spatial,aerial,maritime,and ground users,satisfying the future network requirements on all-time,all-domain,and all-space communications and interconnected networking.Firstly,we reviewed the current research development of SAGINs,discussing the research trends on the low-earth orbiting (LEO) satellite constellation and space-ground network integration.Then,the reinforcement learning (RL) framework was proposed in SAGINs to address the problems of complex architecture,high dynamics,and resource constraints in SAGINs,which facilitated efficient and fast network design,analysis,optimization,and management.As a case study,the method of applying deep RL (DRL) was showed for the intelligent access network selection in SAGINs.To improve the RL training efficiency,a comprehensive SAGINs simulation platform was established,through which the agent-environments interaction was accelerated and training samples could be obtained more cost-effectively.Finally,some open research directions were presented.
