Research On The Middle-term Bear Network Topology Evolution Planning Based On Complex Network

With the rapid development of terminal applications and cloud computing services,bearer network as an end-to-end transmission channel,the necessity of its evolution has been recognized by more and more operators.As a long-term task,bearer network evolution is bound to be decomposed into different stages.At each stage,one or more network partitions are selected for planning to meet various requirements in different areas or periods,in order to realize the smooth transformation from the current network to the targeted one.However,how to choose appropriate network partitions is still an open problem.Existing approaches are usually infeasibly applied to large-scale networks,or ignore the hierarchical structure and multi-ring topological characteristic of bearer network itself.In addition,for operators,the phenomenon of excessive one-time investment,high evolution risk,long listing time and decline of user experience often occur in the process of network evolution.Facing the complex and changeable network evolution needs,how to design a general and high-quality network topology evolution strategy under the driving of traffic growth and the constraint of resource budget has become the key point and difficulty.To solve the above problems,this thesis researches on the multi-stage network partitions evolution optimization.The main contributions and achievements can be summarized as follows:(1)Hierarchical community discovery for multi-stage evolution of targeted network.Since the evolution process from the current network to the targeted one needs to be divided into different network partitions for implementation,this thesis takes the community as the unit to split the network,and obtains the community which meets the topological characteristics of bearer network by improving the community detection algorithm in complex network field.The specific process is divided into two sub steps:a)Layering for plane bear network.Because of the bearer network topology evolution involves the equipment,links and architecture change between different layers,it is necessary to get a reasonable hierarchy for the plane bearer network at first.Considering that the node role represents its level and status in the network and nodes with the same role are located on the same layer,thus,"network layering" problem can be transformed into "node role identification" problem.This thesis extracts the centrality of nodes by using the network structure and extra attribute information,and proposes a centrality-enhanced network representation learning method to attain the low-dimensional node feature vector,which is then input into the SVM model for semi-supervised learning to construct the node-role mapping relationship.At last,by predicting the role of the remaining nodes,the hierarchical structure of the whole network can be built;b)Hierarchical community detection.Based on the local expansion algorithm,this thesis proposes a set of new community expansion rules across role layers.Through the bottom-up hierarchical expansion and combined with geographical location adjustment,the hierarchical community for nodes to be upgraded can be obtained.The experimental results on three real datasets show that the size of the community is flexible and adjustable.Moreover,the community not only has dense connection in topology and clear internal hierarchy,but also preserves the integrity of the tree/ring structures.(2)Community evolution order selection based on deep reinforcement learning.As regard to the scenario that the network traffic increases year by year and the investment budget of each stage is limited,this thesis from the perspective of operators defines the specific network evolution environment,and constructs a community evolution optimization model aiming at shortening the time and maximizing the return on investment.Then,this thesis transforms the community evolution order selection problem into a Markov decision-making process,and introduces deep reinforcement learning to attain the optimal decision strategy of community evolving order under different demand scenarios.In this process,at each stage selecting a community group is a multi-action decision which suffers a problem of too large space complexity.To deal with it,a scheme uses the method of freezing the network environment to select a single community in turn instead of selecting multiple communities at one time can reduce the time complexity from exponential level to linear one,so as to solve the problem in polynomial time.Simulation results show the proposed DRL method has the advantages of flexibility and computational efficiency,and can obtain approximately optimal community sequence results.