Research On Distributed On-demand Intelligent Routing Algorithm Based On Graph Representation Learning

With the development of computer networks,new types of applications have emerged from the Internet.Different applications have different requirements for Qo S indicators.How to use appropriate routing algorithms in network routing to meet the Qo S requirements for diverse applications becomes a challenge for network resource optimization.DRL techniques have been applied in end-to-end routing optimization problems,however,DRLbased network routing techniques still have many limitations,such as difficult convergence of neural networks,poor scalability,low learning efficiency of models,and poor generalization ability.To address the above problems,this thesis proposes a distributed on-demand intelligent routing model based on GRL to learn optimal routing decisions in a fully distributed manner and achieve efficient routing to meet different Qo S requirements.In this thesis,two GRLbased feature engineering models are designed: the Graph SAGE-based feature engineering model and the Graph Attention Network-based feature engineering model.Further,two ondemand routing algorithms are proposed: GSo R and GATo R.Using GRL feature engineering to capture the rich feature information of the graph and construct a fully distributed multi-agent DRL routing model can improve the learning efficiency,scalability and generalization performance of the routing algorithms.The goal of efficient on-demand routing is achieved.Finally,this thesis implements a simulation environment for algorithm performance evaluation using the network simulator Mininet and the RYU controller.Experiments were conducted in three real network topologies,and the two on-demand routing algorithms proposed in this thesis were evaluated and compared with the best performing DRL-OR algorithm and baseline algorithm OSPF in distributed routing learning.The experimental results show that GSo R and GATo R slightly outperform the DRL-OR algorithm in terms of average delay,throughput ratio,and loss ratio,and verify that GSo R and GATo R algorithms have good scalability and generalizability.