Bayesian Optimization Based Network Resource Configuration Application

With the grow of Big Data Analysis rapidly,the scale and complexity of network systems are getting bigger and bigger,and accurate modeling is becoming more and more difficult.Black-Box optimization technology is getting more and more attention,and as the most potential and hopeful method in Black-Box optimization,Bayesian Op-timization is more and more important.Bayesian optimization,as a global optimization method,can quickly find the optimal solution(or near-optimal solution of the optimal solution)through the guidance of the acquisition function for function processes that do not have a closed form and require high cost operation.The first network configuration scene,Big Data Analysis tasks need to be deployed on cloud servers for calculations,and how to find the optimal configuration of the task in a series of cloud configurations(the minimum overhead in the case of guaranteed running time)becomes crucial.And Bayesian Optimization method can find the result within a limited number of times and increase the efficiency.However,the Bayesian Optimization based on GP(Gaussian Process)increases the time complexity by(9)~3)as the number of sampled points.This limits the application of Bayesian Optimization in some large-scale tuning problems,since most of these problems are offline,and the time limit is more relaxed.It is hoped that Bayesian Optimization can explore the input space as much as possible to achieve better performance.As a result,by using the random forest model in the Bayesian optimization algorithm,the algorithm of increasing the linear complexity time complexity()with the sampling point is obtained.In the Video Streaming rate adaptation scene,Section 3 proposes the Reinforcement Learning Algorithm(A3C)get a good performance for the Video bitrate control,but it does not refer to how to choose the neural network structure parameters.So this the-sis uses the Bayesian Optimization algorithm based on random forest to tune the neural network structure parameters,trying to find a smaller neural network structure to save resource and save training time.In the scene of network intrusion detection,researchers have used Ensemble Learn-ing to improve the performance on NSL-KDD data sets.Inspired by Ensemble Learning,by integrating Bayesian optimization algorithms and Ensemble Learning,the results of the benchmark dataset in the field of network intrusion detection are better than the sin-gle benchmark model.This thesis uses Bayesian optimization methods in three questions about network resource configuration.The results show that the Bayesian optimization method has good performance in the application of network resource configuration.