The Research On Conditional Variational Auto-encoder And Its Optimization Method For Network Performance Prediction

To ensure the implement of critical businesses and the operation of numerous web applications in data and communication networks,it is essential to evaluate network performance in a timely and accurate manner.We need to detect abnormal conditions such as network congestion in time,and then adjust the network architecture and routing strategy according to the network short-slab duly,in order to improve network transmission capacity and ensure transmission quality.In the past,the main way that we obtain network performance and status information was to use network monitoring devices or probes for network detection.Traditional detecting method is divided into active and passive ways according to the way of collecting information:Active measurement detects the situation of the measured object by sending a data packet,and obtains the feedback index by analyzing the response of the measured object.It has real-time performance,but the participation of the monitor itself has a certain impact on the result,which leads to an inaccurate measurement.Passive measurement intercepts a part of real network data at the monitoring point,and obtains the information of the network and devices by analyzing the representative data.It does not affect network performance,but has poor real-time performance and generally requires administration authority.In recent years,network performance prediction has drawn continuous attention as a new method of detecting network state information.By collecting historical performance data of a specific target network,it constructs a realistic model for some metrics in the target network.And then it utilizes the model to provide a reasonable prediction of the future performance and state of the network.It has the advantage of eliminating the need to deploy expensive network monitoring and measurement equipment,which reduces the cost and impact on network performance while achieving accurate measurements.Existing prediction methods mainly include time-series-based methods,similarity-based methods and discriminative-model-based methods.The first two prediction methods determine inference on the laws of observed historical data rather than an analytical model,which easily leads to inaccurate prediction.The third method learns the nonlinear pattern of network metrics through the neural network model,but is still insensitive to changes of the external network environment.In the field of machine learning,the emerging generative models have drawn widespread attention.It constructs the hidden vector of the target features,that is,the combination of hidden variables,by learning historical data.Then it samples the probability density function representing the distribution of hidden variables,and restores the real data through decoding.The Conditional Variational Auto-Encoder(CVAE)model is widely used in inference tasks.It utilizes the inherent attributes and external conditions of historical data to generate hidden vectors that characterize the problem.Compared with prediction methods based on time series and discriminative model,it can deal with the hidden variables appearing in the network system by constructing a nonlinear model of neural network and random sampling.Meanwhile,it takes the external conditions that are easy to obtain as input to make the model more sensitive to the changes of the outside world.Based on CVAE,this paper proposes a Network Performance Prediction method using Generative Model(NPGM).We consider the target network as a power system with flow conditions as input,which outputs network latency matrices through the operation of encoding and decoding by the generative model.Based on the idea of phase space reconstruction,by constructing hidden vectors of some metrics,NPGM can extract the essential features of network and temporarily ignore the influence of other factors.Our model consists of 4 Deep Neural Networks(DNNs),which are used to learn the complex patterns of network features and restore the multidimensional nonlinear system of the target network.After constructing the hidden vector,the random sampling operation guarantees the creativity of the model,enabling it to virtualize the network state that has not appeared before,which reduces the constraint of the data volume of training set to a certain extent.In order to improve the efficiency of training neural network parameters,we try to optimize the parameters by using the Swarm Intelligence Algorithm instead of the widely used Gradient Descent algorithms.Particle Swarm Optimization(PSO)is a kind of bionic heuristic algorithm that imitates the foraging of birds.However,the original PSO algorithm has the defects of slow optimization speed and easy convergence to local optimum.We improve it and propose a Self-adaptive Stochastic PSO(SPSO)algorithm for parameter optimization of DNNs in NPGM model.It has an adaptive inertia weight coefficient and acceleration factor,and adopts a position mutation operation,which make it have a faster convergence speed and not easy to fall into local optimum.We propose a new criterion called Weighted Matrix Mean Absolute Percentage Error(WM-MAPE),to evaluate the accuracy of the predicted latency matrice results.We train and test our model using the backbone network traffic data of the Center for Applied Internet Data Analysis(CAIDA),and the results show that the error of our method is only 50%of time-series-based prediction method,or even lower.Moreover,the efficiency and convergence of SPSO is significantly better than Gradient Descent algorithms and original PSO.