| With the development of network technology,more and more application scenarios have emerged.In this context,the research on intelligent routing algorithms under the SDN(Software Defined Network)architecture has become a popular research direction.With the continuous increase of network scale and traffic load,traditional routing algorithms can no longer meet the needs of diverse network services,while intelligent routing algorithms based on reinforcement learning have great potential and advantages.How to deal with different network scenarios and application requirements,and how to perform real-time optimization and adaptive adjustment are the main directions of current research.This thesis proposes an improved Q-Learning algorithm and an improved SARSA(State-Action-Reward-StateAction)algorithm,which are applied to the routing optimization algorithm of the SDN architecture,and effectively improve network performance for different network topologies.First of all,this thesis introduces the architecture of SDN and the principle of reinforcement learning algorithm in detail,and uses the Mininet simulation platform to build an SDN experiment environment.On this basis,the network topology is designed for symmetrical and asymmetrical networks,and the collection of network parameters is explained and realized.In addition,it also designs and implements the flow table module in the SDN network.Then,this thesis focuses on the use of reinforcement learning algorithms to optimize the routing and forwarding of SDN networks.By optimizing and improving the reinforcement learning algorithm,it replaces the traditional routing algorithm to achieve more efficient routing and forwarding.In order to better adapt to the characteristics of SDN network data,this study improved the Q-Learning and SARSA algorithms,designed and implemented an algorithm based on dynamically adjusting the learning rate to Enables faster convergence of network data routing.At the same time,for the setting of the reward value function of the reinforcement learning algorithm,the throughput and delay in the network are used to design and implement,so that the algorithm can quickly and reasonably allocate network resources without affecting the performance of network transmission.Simulations were carried out under the Fat-Tree network and NSFNet(National Science Foundation Network)network.The experimental results show that the accuracy of the improved Q-Learning algorithm and the improved SARSA algorithm will eventually tend to 100% as the training cycle increases under different networks.Finally,simulation experiments show that the improved Q-Learning algorithm and the improved SARSA algorithm have a significant effect on the performance improvement of each aspect of the network than the traditional routing algorithm in the Software Defined Network environment.When the network load exceeds 40%,the reinforcement learning algorithm in this thesis outperforms the traditional heuristic algorithm. |
