Data-Driven Algorithms For Virtual Network Embedding

Network virtualization enables the share of a physical network among multiple virtual networks.Traditional heuristic embedding algorithms follow static procedures,thus cannot be optimized automatically,leading to sub-optimal results.To solve this problem,this thesis introduces data-driven methods to virtual network embedding.This thesis proposed two data-driven algorithms,which are able to optimize themselves by learning knowledge of virtual network embedding and information about the substrate network from historical virtual requests,and utilize the knowledge to rank substrate nodes.The first algorithm extracts a feature vector for every substrate node and employs particle swarm optimization to compute weight vector on historical virtual requests.The optimized weight vector serves to rank substrate nodes and map virtual nodes.The other algorithm features a policy network,which computes a mapping probability for every substrate node.The policy network is trained with policy gradient and reinforcement learning.The performance of the proposed embedding algorithms are evaluated in comparison with two other algorithms that use artificial rules based on node ranking.Simulation results shows that the algorithms proposed in this thesis has higher acceptance ratio,long-term average revenue and revenue to cost ratio.