Research On Deep Learning Based Routing And Spectrum Assignment Strategies For Elastic Optical Networks

With the rapid development of many bandwidth hungry services such as big data,cloud computing,and high-definition video,the requirements for communication system capacity and network transmission interface data rates have become higher and higher.Compared with the traditional Wavelength Division Multiplexing(WDM)technology,Elastic Optical Networks(EON)has been widely accepted as one of the best solutions with flexible network architecture and resource allocation abilities.Routing and Spectrum Assignment(RSA)is the key issue in EON,while the dynamic spectrum allocation and release will cause spectrum fragmentation,which will affect the allocation of subsequent services and the utilization of EON spectrum resources.This thesis focuses on the problem of spectrum fragmentation in EON.From the perspective of RSA algorithm and traffic prediction,it makes deep research in deep learning technology and the application of Deep Neural Network(DNN)and Long and Short Term Memory(LSTM)neural network in RSA strategy.The principle of EON's RSA technology and traffic prediction is presented firstly,and followed are the discussion of the problem of spectrum fragmentation in EON.This thesis also introduces typical machine learning algorithms and their applications,especially focusing on the application strategies of machine learning in EON.Aiming to improve the problem of spectrum fragmentation in EON,this thesis proposes a deep learning-based RSA(D-RSA)strategy to improve the problem of spectrum fragmentation.This strategy uses a Deep Neural Network(DNN)model,which can choose the RSA scheme that minimizes the degree of spectrum fragmentation based on the real-time situation of the dynamic network.Simulation results show that,compared with the K-Shortest Paths combined with the First Fit(KSP + FF)algorithm,the proposed D-RSA algorithm can reduce the spectrum fragmentation on CERNET,NSFNET,and USNET by an average of 2.2%,2.7%,and 2.4% when the service rate reaches 200Gb/s,respectively.When the traffic load is about 400 Erl,the bandwidth blocking probability of D-RSA algorithm can decrease by 9.7%,6.8%,and 6.3% on average.Under the same circumstances,compared with the K-Shortest Paths combined with the Exact Fit(KSP + EF)algorithm,the spectrum fragmentation can be reduced by 1.2%,1.7%,and 1.5%,and the bandwidth blocking probability can be reduced by 5.7%,5.5 %,and 2.4%,respectively.In addition,aiming at the problem that the traditional RSA algorithm cannot be flexibly switched under low latency conditions,which results in a high bandwidth blocking probability when the network conditions are complex,an improved RSA strategy(TP-RSA)combined with deep learning based traffic prediction is proposed.The TP-RSA algorithm implements a reasonable and accurate traffic prediction to save time for EON's resource allocation.At the same time,an index of spectrum fragmentation combined with traffic prediction is introduced to more accurately describe the degree of spectrum fragmentation.The TP-RSA strategy will use this index to find the best RSA solution.Theoretical analysis and simulation results show that compared to KSP + FF in the above three network topologies,the TP-RSA strategy has a reduction of 8.8%,4.9%,and 8.9% in the bandwidth blocking probability when the traffic load reaches 400 Erl,while compared to KSP + EF,the bandwidth blocking probability is reduced by 3.6%,0.93%,and 3.4%,respectively.