Intelligent Network Slicing Optimization for Satellite-Terrestrial Integrated Networks

Network slicing plays vital roles to deliver customized network services in future satellite-terrestrial integrated networks. A major challenge in employing network slicing is to address the high network dynamics due to satellite movements during end-to-end path selection and resource allocation. To this end, we propose a graph attention network (GAT) based deep reinforcement learning (DRL) algorithm to assign network resources for slices. A GAT model is firstly utilized to extract hidden relationships between the network status and the topology information. The extracted results are applied as the input of the DRL model with a hybrid action space. A hybrid proximal policy optimization (HPPO) approach is then designed to solve the DRL problem. Dual objectives of user satisfaction and resource efficiency for network slicing are well achieved by the proposed algorithm. Simulation results demonstrate the superior performance of our algorithm compared with several benchmarks.