A Research On Key Technologies Of Network Routing In Knowledge Defined Networking

The last decade has seen a development of social networks which connects people all over the globe,creating heavier and broader demand for information propagation.The medium of information has also gone through a series of evolution as in text,images,videos,short videos,live streaming,and so on.The capacity of the medium is increasingly larger.On the other hand,The evolution from the Internet,to the Internet of Things,and to the Internet of Everything,reflects the fact that the propagation of information has changed from between the people,to between the physical devices,and to between all the people,the devices,and the data.The number of connected endpoints increases,and the requirement for the age of information becomes stricter.All these changes urge that the performance of the network which served as the infrastructure of information propagation to be continuously improved and optimized.In the meanwhile,with the advent of the post-Moore era,the upgradation becomes tougher and the acquisition of computational power becomes more expensive.Thus,the Internet service providers can no longer increase the processing capacity of key networking nodes by purchasing and deploying new computation hardware.Hence,it becomes an urgent need to find better network routing and traffic scheduling strategy utilizing current hardware resources and dealing with the demand for service of experience over exploding and heterogeneous network services.Also,in the last decade,software-defined networking has gradually evolved into a more advanced stage.In SDN scenarios,the control-plane and data-plane are decoupled,and network carriers can choose more flexible strategy to control the network,to improve the quality of experience,and to lower the operational expenditure.The networks using software-defined networking technologies can provide measurement data of network operation status,as well as the interfaces needed for carriers to control the networks.However,the question that how these data and interfaces can be used to optimize network performance is still left unanswered.Knowledge-defined networking points out a direction.Knowledge-defined networking adds a knowledge-plane on top of the control-plane and the data-plane.The knowledge-plane creates an understanding and abstraction of the raw network status measurement data collected from the data-plane,and by utilizing this knowledge it provides guidance for the control-plane on how to optimize for better performance.Recently,machine learning and artificial intelligence technologies have gained fundamental development and successful application in areas such as computer vision and robotics.machine learning and artificial intelligence technologies are the most promising tools to implement an intelligent,autonomous,and adaptive knowledge-plane.Yet still,it is left to be intensively explored to intelligently design and effectively apply machine learning and artificial intelligence techniques in the specific networking and network routing problems.This dissertation summaries the author’s research work on the design and application of machine learning and artificial intelligence techniques for the key problems on network routing in knowledge-defined networking.This includes "how to realize","how to utilize","how to optimize",and "how to stabilize" four parts.The achieved results are listed as follows,respectively.(1)The author proposes a graph-aware deep learning model utilizing the network’s topological characteristics to solve the problem that in Al-based network routing schemes the network status representation being non-Euclidean and spares which in turn causes the model difficulty in learning and fitting the feature functions.The proposed method uses the topological relation between the network forwarding nodes to select neighbors for each node in a layered manner.These neighbors are later sorted using the betweenness centrality to evaluate the importance of each neighbor to the central node.A group of neighbors are then used to build a graph-aware operator which operates on the input adjacency matrices,as a layer in the neural networks,and conduct convolution over the result of each graph-aware operator to produce graph-aware features for the following layers.Compared with existing methods,the proposed one empowers the deep neural networks with the ability to identify key information in graph structures and learn from the graph-based information in network status data to solve network routing tasks.Compared with existing methods,the decision accuracy is improved,and the training time is reduced when the models achieve an accuracy at the same level.The convergence time is reduced,and network latency is reduced.(2)The author proposes the use of case-based reasoning for network routing and traffic scheduling in which the policy is high-dimensional,continuous,and difficult to prepare training data when using supervised learning.The proposed method selects policy of the case with the closest known network status compared to current status and uses deep neural networks to adaptively evaluate the similarity between network status.Compared with existing methods,the proposed method reduces the decisioning delay.The final policy can converge to the(near-)optimal.(3)The author proposes the use of cooperative game theory in modelling heterogeneous connected vehicle networks where the overall profit allocation among different tasks with different objectives needs to be optimized.The author also proposes the use of deep reinforcement learning to deal with the large state space when there are multiple users and multiple sub-tasks.Compared with existing methods,the proposed method provides lower packet loss rate for all users.Among the users with different optimization objectives,75%of the users achieve better results where latency-sensitive users acquire lower latency and throughput-sensitive users acquire more throughput.(4)The author proposes the use of Lyapunov methods to increase queuing stability in the machine learning-based network routing and traffic scheduling where the system’s stability is weakened,and the quality of experience is reduced by the uncertainty of machine learning algorithms.To the best of the author’s knowledge,the proposed method is the first solution adding stability to AI-based routing algorithms.Compared with existing AI-based routing approaches,the proposed one reduces network overhead,and solves the problem of routing loops and unreachability without introducing extra path computation.Compared with existing queuing stability methods,the proposed one reduces overall queue backlog size and increase convergence time.