Optimization Of The Kernel Parameters Of The Support Vector Machine Based On The MAB Model

Currently,it is undoubted that machine learning is the most popular topic,which is changing people's lives gradually,making it more intelligent and convenient.Support Vector Machine(SVM)is developed,which overcomes "the curse of dimensionality" and "overfitting",based on statistic learning theory and applied to regression and category.The parameters of Support Vector Machine have dramatic effects on the ultimate regression and category,but the appropriate theories have not been found out to instruct how to choose the parameters quickly and specifically.In statistic learning theory,Multi-Armed Bandit(MAB),short for Bandits Problem,is a problem that the arms of which bandits are pulled,how many times the arms are pulled and the choosing order when a gambler(or several gamblers)faces a line of multi-armed bandits.While pulling the arm,each bandit provides random praises with a certain probability.The goal of gamblers is to acquire the maximum reward in a series of pulling orders.The gamblers confront the important balance between "use" of maximum expected profit machine and"exploration" of mastering other machines' expected profit information in each experiment.The paper studies how to choose the parameters of SVM in terms of Multi-Armed Bandit(MAB),avoiding unsatisfactory category for inappropriate parameters in applying SUV.Primarily,the paper researches Multi-Armed Bandit(MAB)and Support Vector Machine(SVM),which improves kernel parameter of SVM and penalty term based on concluding the general principles,merits,drawbacks and relative studies systematically.Then,the paper studies theoretical foundation and development status of SVM,introduce its main idea and basic statistical theory,analyze ?-greedy,UCB1,Softmax and Pursuit and simulate astringency in different parameters.Other hypothetical conditions can be ignored for the characteristics of ?-greedy.In each situation,the next pulling is chosen by statistical calculation without any parameters,which means that any priori conditions are not required.Adequately understanding ?-greedy in MAB that is used to optimize initial kernel parameter ?,penalty term C and loss function parameter ?,a kind of SVM parameter optimization method based on MAB is proposed,which optimizes parameter C and parameter ?.Initial kernel parameter ?,penalty term C and loss function parameter ?,which are trained by UCI statistics,are regarded as the arms of multi-armed bandits.This idea is utilized to categorize and validate UCI data sets.Analyzing the result of this experiment,this method achieves great effects to a certain extent.