Font Size: a A A

Research On Intelligent Topology Reconstruction Strategy Based On Traffic Prediction

Posted on:2022-12-23Degree:MasterType:Thesis
Country:ChinaCandidate:J W ZhuFull Text:PDF
GTID:2518306776953049Subject:Automation Technology
Abstract/Summary:PDF Full Text Request
With the rapid development of Internet and the burst of new applications,network traffic surged in the last years.The static network topology cannot adapt to dynamic network traffic.It will jeopardize network performance.Motivated by this,we proposes an intelligent topology reconstruction strategy based on traffic prediction.Firstly,network traffic in the next few time slots is predicted by a deep learning-based model;then a VNT(Virtual Network Topology)adapting to the network traffic in future is constructed by deep reinforcement learning-based strategy to reduce latency and decrease jitter.The main contributions of this study include:(1)To improve the accuracy of traffic prediction,we propose a Deep Learning-based Traffic Prediction Model(DLTP)to predict network traffic in the next few time slots,which takes periodic features and sequential features of traffic into consideration.CNN(Convolutional Neural Network)and RNN(Recurrent Neural Network)make predictions based on periodic features and sequential features of traffic,respectively.Then the final prediction of traffic matrices in the next few time slots is made by DNN(Deep Neural Network)according to the outputs of CNN and RNN.We evaluate the performance of DLTP on Abilene dataset and G(?)ANT dataset and compare it with other three traffic prediction models,including LSTM(Long Short-Term Memory,),LRCN(Long-term Recurrent Convolutional Network)and ITRCN(Interactive Time Recurrent Convolutional Network).Simulation results show that the proposed DLTP model can improve performance by 33.23%,26.85%,and 24.12% on average.(2)To reduce network latency and jitter caused by topology reconstruction,we propose a Deep Reinforcement Learning-based Topology Reconstruction Algorithm(DRLTR).Based on current topology and traffic matrices in the next few time slots predicted by DLTP,the agent of DRL constructs a VNT to serve network traffic in the next few time slots.To achieve trade-off between network latency and jitter,we define a new parameter,reconstruction benefit rate,as the ratio of latency reduction and the number of links needed to change.The reward received by the agent is the reconstruction benefit rate.Compared with other reconstruction algorithms,the proposed DRLTR can efficiently reduce latency and decrease achieve better performance in terms of latency and jitter.
Keywords/Search Tags:Traffic prediction, Topology reconstruction, Deep learning, Deep reinforcement learning
PDF Full Text Request
Related items