求解矩阵 范数极小化问题的谱梯度算法-spectral gradient algorithm for matrix norm minimization.docxVIP

摘要多任务特征选择问题在医学诊断、文本分类和微生物信息工程等机器学习领域受到广泛关注.最新研究表明,多任务特征选择问题可以通过求解矩阵?2,1范数正则项的非光滑凸优化问题来实现.由于?2,1范数正则项的非光滑性,求解该问题颇具挑战.本文提出求解矩阵?2,1范数极小化问题的谱梯度算法,分析算法的收敛性,数值试验验证算法的有效性.第一章,介绍?2,1范数极小化问题的研究背景和研究意义,包括问题的模型、发展以及近些年来求解该问题已有算法;介绍本文研究的理论基础,包括线搜索和谱梯度法;并列出本论文所用到的一些基本概念、符号和定义.第二章,提出一种求解矩阵?2,1范数极小化问题的谱梯度算法.每步迭代,所提算法仅需计算光滑函数的梯度和目标函数值.为了提高所提算法的有效性，引入非单调线搜索,并在适当条件下,分析算法全局收敛性.数值试验表明所提算法非常有效,其效率可与SLEP和IADMMFL算法相媲美.第三章,改进第二章所提算法,引入参数h,进行有限差分,求解问题.在一定条件下,建立算法的全局收敛性,通过数值试验验证算法的有效性.第四章,给出本论文的总结,并提出一些值得继续探讨的方向.关键词:多任务特征学习,非光滑优化,矩阵?2,1范数,谱梯度算法,非单调线搜索IABSTRACTMulti-taskfeaturelearninghasrecentlyreceivedincreasingattentioninmachinelearningincludingmedicaldiagnosis,textclassi?cation,andbiomedicalinformatics.Gen-erally,recentresultsshowthatitcanberealizedbysolvinga?2,1-norminvolvednon-smoothconvexminimizationproblem.However,solvingtheoptimizationproblemisachallengingtaskduetothenon-smoothnessoftheregularizationterm.Inthisthesis,weproposespectralgradientalgorithmsforsolvingthematrix?2,1-normminimizationprob-lem.Weestablishtheglobalconvergenceoftheproposedalgorithmsanddonumericalexperimentstoillustratetheire?ectiveness.Inthe?rstchapter,weintroducethebackgroundandsigni?canceofthe?2,1-normin-volvedminimizationproblem,whichconsistsoftheproblems’formulations,developmentsandsomeexistingmethods.Wepresentsomepreliminariesincludingthelinesearchtech-niquesandspectralgradientmethods.Moreover,someimportantnotationandsymbolswhichusedinthecontextarealsoincluded.Inchapter2,weproposeaspectralgradientalgorithmformatrix?2,1-normmini-mization.Theproposedalgorithmonlyrequiresthegradientofthesmoothfunctionandthevalueofobjectivefunctionateachandeverystep.Toimprovetheperformanceoftheproposedalgorithm,anon-monotonelinesearchtechniqueisincorporated.Undersomemildconditions,weshowthattheproposedalgorithmconvergesglobally.ThenumericalexperimentsshowthattheproposedalgorithmworksquitewellandperformsbetterthanthealgorithmSLEPandIAMDMFL.Inchapter3,toimprovethemethodwhichgiveninthepreviouschapter,weusethe?nitedi?erencemethodwit