Routing Optimization Based On Machine Learning In Satellite Network

Satellite networks have become the main force of communication net-works due to their wide coverage,low latency,and broadband characteristics,and routing issues are an important part of the satellite network construction pro-cess.Due to the dynamic changes in the topology of the satellite network itself and the limited load resources on the satellite,the existing distributed routing solutions cannot perceive changes in the network status in real time,resulting in problems such as unbalanced network load and easy congestion.The de-velopment of software-defined network technology provides a new solution to the routing optimization problem.The software-defined network architecture makes the network more flexible.In the routing optimization problem,network measurement and routing optimization strategy are two key components.Ef-fective real-time network measurement provides a basis for the generation of routing optimization strategies,which makes the network aware of congestion.The existing out-of-band network telemetry technology will transmit additional probes to measure the network status.This measurement method will inevitably lead to an "observer" effect,which leads to the problem of inaccurate measure-ment information.In addition,the relationship between complex network status and routing optimization strategy is difficult to describe with precise mathemat-ical models.Therefore,this article combines the design idea of software-defined net-work and proposes a novel route optimization method for satellite network.In the method,the data plane is composed of a group of low-orbit satellites,and the data packet processing process on each satellite node is customized through the data plane programming language to realize rapid message forwarding and low-cost,high-precision network measurement.The control plane is centrally deployed in the ground control center with sufficient computing and storage re-sources,and the self-learning and self-optimizing machine learning technology is used to solve the difficult modeling problem between complex network mea-surement information and routing optimization strategies,and to realize intelli-gent routing decisions.So as to achieve the optimization goal of minimizing the maximum link utilization rate of the network.We summarize our contributions as follows:·Design a clever switch packet processing logic based on programmable language P4 to enable fast forwarding and ultra-low overhead measure-ment.·Integrate reinforcement learning algorithms into the control plane to achieve self-learning and active convergence of routing optimization strategies.·Implement our approach in an experimental platform.The extensive eval-uation shows that our approach significantly outperforms several widely-used baseline methods concerning three metrics,including max-link-utilization,flow completion time and packet delay.