Research On Prediction Of Time Series Based On Minimax Probability Machine

Time sequence prediction is becoming more and more important as the times progress,and its applications have been intensively applied in many fields, such as the forecasting ofeconomic, weather forecast, traffic flow and network streaming forecast and so on. Theresearch of Intelligent Transportation System (ITS) can provide very important informationfor live online control of network of roads traffic flow. Also, the real-time frame datadistribution of network flow can effectively pave the way to the ease of network congestionand management of network security.Minimal Probability Machine Regression (MPMR) is a novel method which introducesthe machine learning of probability classification to the forecasting of regression, and hasbeen widely used in fields such as palm identification, image segmentation, data mining,power forecasting and so on. In this thesis, based on chaos theory and the recursive predictiveanalysis, MPMR method is used for one step and multi-step prediction experiment of trafficflow and network video streaming predictions. By kernel function mapping, epsilon pipewhich maximum probability that contain prediction points which fall into minimumregression pipe is obtained. Also, the method is contrasted with Support Vector Machine(SVM) prediction method, artificial Neural Network prediction method under predictionexperiment, and the advantage of this method is verified.The content of main research includes several aspects as follows.(1) Based on Bayesian learning, the linear MPM method, nonlinear classification MPMCmethod and MPMR regression method have been studied in this thesis.(2) Chaos theory has been studied corresponding to nonlinear time series and the chaoticcharacteristic of the three time series has been identified using the maximal Lyapunov index.Also the Cao method to determine the optimal embedded dimension, the mutual informationmethod to determine the optimal delay time τ, and the method using recurrence plot toestimate the prediction capability of time series have been studied.(3) The probability learning machine MPMR method has been applied to the experim-ents of Mackey-Glass chaos time series, traffic flow and network video streaming, and thencomparing the results of experiments under the same condition, the advancement andeffectiveness of this method have been verified.(4) Based on RBF kernel function, the influence of forecasting precision when selectingdifferent values of corresponding regression pipes has been studied to MPMR and therobustness of this method has been verified.