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

With the rapid development of bandwidth hungried data services such as video-on-demand,virtual reality,cloud computing,and augmented reality,the t optical fiber communication network is gradually unable to meet the exponential growth of global data traffic.Elastic Optical Networks(EONs)are considered an important development direction of the next generation of optical networks due to their flexible spectrum allocation and high spectrum utilization.The core research of the EONs planning layer is the routing and spectrum assignment(RSA)problem,which mainly includes the optimization of RSA schemes and the spectrum fragmentation problem generated based on RSA schemes.In response to these issues,this thesis comprehensively studied them using theoretical analysis and numberical simulations with deep learning technology assistance.This thesis first introduces the principles,technology,and research status of elastic optical networks,and discusses the existing RSA algorithms in detail.In response to the shortcomings of existing algorithms,this thesis proposes a heuristic deep-neural-network assisted RSA algorithm.In terms of routing selection,the algorithm uses an improved ant-colony system algorithm with pseudo K-shortest paths(KSP)instead of the KSP algorithm,which helps to find the global optimal solution.In terms of spectrum allocation,the algorithm proposes an adaptive spectrum-aware allocation rule based on the number of nodes and spectrum fragmentation,with the number of hops between source and destination nodes also considered as a consideration criterion.In terms of deep learning,the algorithm uses a dual deep-neural-network model instead of a single model to reduce the number of output labels.Both the simulation results in NSFNET and USNET topologies show that the proposed algorithm can reduce blocking rates and improve spectrum utilization simultaneously.In response to the spectrum fragmentation problem generated by the proposed algorithm,this thesis also proposes a periodical defragmentation strategy triggered by time-frequency domain and blocking thresholds.The strategy uses an improved fragmentation index-based blocking delay trigger mechanism to reduce the load of invalid defragmentation performed when not blocked.In terms of defragmentation order,it comprehensively considers the time-frequency domain characteristics of services and reasonably allocates the reconstruction order.Simulation results show that the above strategy can further reduce blocking rates and improve spectrum utilization while reducing the number of reconstructed services and spectrum fragmentation.