Load Balancing And Routing Optimization For SDN Networks Based On GCN And MARL

In the past decade,the number of global internet users has grown steadily,and the demand for internet communication performance has been increasing.Each application has its unique network performance metric requirements,which depend on the way data is transmitted within the network.Traditional routing algorithms rely on protocol-based routing information calculations,but with the advent of Software-Defined Networking(SDN)technology,routing algorithms have become more flexible and programmable.SDN technology decouples the data plane from the control plane,allowing the controller to calculate the optimal path based on global network topology,real-time traffic information,and other factors,and then send routing information to each router to meet different application requirements for network performance.This thesis aims to explore the enormous potential of applying Artificial Intelligence(AI)to intelligent routing algorithms in SDN to meet the network performance requirements in various scenarios.For the routing decision problem in SDN,this thesis analyzes and summarizes existing solutions and proposes a feature extraction method based on Graph Neural Networks(GNN)and Recurrent Neural Networks(RNN),as well as a Multi-Agent Reinforcement Learning(MARL)algorithm using contribution-based rewards.Through simulation experiments in three real network environments in Mininet,this thesis verifies that the proposed algorithm can effectively improve network throughput,reduce latency,and packet loss rate.Since different applications have different network performance metric requirements,this thesis applies AI algorithms to specific performance metric requirements for different types of traffic using traffic classification technology,thereby achieving more intelligent routing decisions.Specifically,this thesis innovatively designs a GGRU feature extraction module that combines Gated Recurrent Units(GRU)and Graph Convolutional Networks(GCN)to capture the spatial characteristics of network topology and acutely perceive the temporal dependencies of traffic.Additionally,this thesis designs a Multi-Agent Reinforcement Learning algorithm considering contribution-based reward values(CBR)and innovatively designs the Actor network structure within the CBR,enabling the model to dynamically adapt to increasing request types without retraining.Simultaneously,this thesis also innovatively designs the Critic network within the CBR,enabling the model to allocate reward values based on the contributions of the agents.For different types of request requirements,this thesis proposes new reward functions and designs a secure routing mechanism and offline pre-training mechanism to ensure system stability and convergence.