Virtual Network Mapping Technology Based On Deep Learning

In recent years,various emerging technologies such as quantum technology,blockchain and artificial intelligence have developed rapidly,which have brought great challenges to traditional networks,and network virtualization technology has brought the possibility for traditional networks to meet diversified network needs.Sex.This article mainly studies the virtual network mapping problem from the following two aspects:Aiming at the problem that the computational load of the deep convolutional neural network will cause a large amount of energy consumption in the terminal equipment,this article regards the computing task of the deep convolutional neural network as a virtual network request,and maps it to multiple terminal devices.The terminals share the calculation to reduce their own losses.In the implementation,the computing task of each layer of the neural network is requested as a virtual node,and the virtual link is obtained according to the connection relationship between the layers.By sharing the computing pressure of multiple devices,the deep convolutional neural network’s requirements for device computing power and battery capacity are solved.The calculation time of the node and the communication time on the link are the constraints,and the optimization goal is to minimize energy consumption..The particle swarm algorithm is used to solve the virtual network request.Since the exploration space is discrete,the particle swarm algorithm is discretized during implementation,thereby obtaining a mapping scheme that meets the time constraints and minimizes energy consumption..This paper focuses on mapping the star-shaped virtual network,and takes minimizing overhead as the optimization goal,and realizes the mapping of virtual nodes through the Deep Reinforcement Learning Nature DQN algorithm.A two-stage mapping strategy is adopted to alternately map nodes and links,and according to the selected physical nodes,the shortest path between them is obtained through the Dijkstra algorithm.Most virtual network mapping algorithms that use deep reinforcement learning use convolutional neural networks for feature extraction of node attributes.Since this article focuses on the virtual network as a hub-and-spoke type,it uses a fully connected neural network.After designing the state space,action space and reward function of the algorithm,the way the Nature DQN algorithm selects actions is adjusted based on its own action space,and the introduction of the Mask prior strategy avoids repeated selection of invalid actions.Test and training results show that the adjusted Nature DQN algorithm can select better mapping results while also improving computational efficiency.