Research On Network Congestion Control Algorithm Based On Learning Method

Network congestion is the phenomenon of data transmission rate and transmission performance degradation caused by the limitation of forwarding nodes and other resources when the number of packets transmitted in the network is too large.Its essence is that the processing capacity of the network itself cannot meet the demand of users for network resources.In order to increase the utilization efficiency of network resources,improve the rate of network transmission,and increase the users' happiness of network use,designing better network congestion control algorithm is a hot research topic in the direction of computer network.In the paper,the congestion control algorithm designed manually is more complicated,and the update speed of manually designed algorithm cannot keep up with the change speed of network environment.The network congestion control problem is transformed into a machine learning problem,and the learning method in machine learning is used instead of artificial design,which improves the efficiency of developing congestion control algorithm.Firstly,this paper first transforms the network congestion control problem into a machine learning classification problem and proposes MLCC algorithm.MLCC algorithm takes the traditional algorithm as the "teacher" to collect the data set,and then USES the classifier model to learn the mapping rules hidden in the data set to perfectly reproduce the traditional network congestion control algorithm.Experiments show that MLCC algorithm is basically the same as the traditional congestion control algorithm in terms of throughput,fairness,and congestion control curve adjustment,and the effectiveness of MLCC algorithm is verified.After the network congestion control problem is transformed into a machine learning classification problem,the new congestion control algorithm can be transformed into a new mapping relationship in the data set by analyzing the corresponding relationship between features and labels in the collected data set.Based on the above analysis,this paper proposes the use of semi-supervised clustering algorithm to create a new mapping relationship,that is,a new congestion control algorithm,SLCC.From the experimental results,SLCC algorithm and traditional congestion control algorithm have improved the throughput performance by an average of 3%?4%,and the intra-protocol fairness and TCP fairness have also been improved very well.After deeply analyzing the mapping relationship between congestion control algorithm and data set,this paper proposes the method of reinforcement learning to create a new network congestion control algorithm and proposes the QLCC algorithm.Different from SLCC algorithm QLCC describe network congestion control process as a markov decision process,use the Q-learning algorithm is designed and developed a new network congestion control algorithm,experiments show that QLCC algorithm can not only in the packet loss rate is higher under the network environment can still keep high data transfer rate,and the fairness and throughput than the CUBIC and New Reno algorithm has larger ascens ion.