机器学习中的随机优化算法 计算机专业VIP免费

–I–1摘要对于机器学习中的数值优化问题，考虑到其规模和维数都比较大，传统的方法难以高效的解决这一问题。近些年来，针对大规模的机器学习问题做了很多研究，比较重要的一类方法是随机算法。优化方法主要分为一阶梯度方法和二阶牛顿方法两大类，针对一阶方法的改进和研究比较成熟和完善，一阶方法又可以分为两类，原始方法和对偶方法，原始方法比较有代表性的有SAG、SAGA、SVRG，对偶方法有SDCA和SPDC两种。此外，加速方法如Catalyst、Katyusha，其收敛速度为一阶方法所能达到的最好结果。二阶方法目前是机器学习领域研究的重要方向，其收敛速度要优于一阶方法，但是其实践中会有一些难度，比较实用的是L-BFGS方法及其随机算法的改进。本文将详细全面的叙述机器学习中各种随机算法，介绍随机算法的发展历程，研究方向及研究热点，最后通过数值试验比较了几种常见随机算法，以给读者直观的数值效果。关键词：大规模机器学习，随机算法，优化方法–II–2StochasticOptimizationAlgorithminMachineLearningAbstractFortheoptimizationprobleminmachinelearningfield,traditionalmethodhavedifficultiesinsolvingthehighdimensionandbigdataproblem.Inrecentyears,therearemanyresearchesinlargescalemachinelearningproblems,especiallystochasticalgorithms.Generally,stochasticmethodcandividedintotwoparts.Oneisfirst-ordergradientmethodandtheotherissecond-orderNewtonmethod.Thereismoreimprovementandresearchinfirstordermethod,andthefirstordermethodismorematureandperfect.Thereistwoclassesforfirstordermethod.Fortheprimalclass,SVRG,SAG,SAGAistherepresentation,andSDCA,SPDCfordualclass.Otherwise,theaccelerationmethodsuchascatalystandkatyusha,whichhastheoptimalcon-vergencespeedforfirstordermethod,isputforwardinlasttwoyears.Secondordermethodisoneimportantresearcharea,andithasbetterconvergencebutnotbetterperformancebecauseithastocomputethehessianmatrix,oneusefulmethodisL-BFGSanditsvariants.Inthispaper,theauthorwillintroducestochasticalgorithmsinmachinelearningareaindetail.Intheend,numericalexperimentscomparesomecommonalgorithmandgiveadirectviewtoreaders.KeyWords:Large-scalemachinelearningproblem,Stochasticalgorithm,Optimizationmethod–III–3目录摘要..............................................................................IAbstract...............................................................................II引言..............................................................................11基础知识...........................................................................52一阶方法...........................................................................62.1梯度下降法（GradientDescent）..............................................62.2随机梯度方法（StochasticGradientMethod）.................................72.3方差减小方法（VarianceReductionMethod)...................................102.3.1随机平均梯度方法(StochasticAverageGradient)..........................112.3.2随机方差减小梯度方法(StochasticVarianceReductionGradient)..........122.3.3SAGA...................................................................132.4随机对偶坐标上升法（StochasticDualCoordinateAscent）...................142.5随机原始对偶坐标法（StochasticPrimalDualCoordinate）....................162.6加速方法Catalyst..............................................................192.7Katyusha......................................................................203二阶方法2–III–4...........................................................................23.1牛顿法与拟牛顿法...........................................