Projected quasi Newton methods for constrained and non smooth optimization

Mark Schmidt

Description

1 Projected Newton-type Methods in Machine Learning Mark Schmidt University of British Columbia Vancouver, BC, V6T 1Z4 Dongmin Kim University of Texas at Austin Austin, Texas 78712 Suvrit Sra Max Planck Insitute for Biological Cybernetics 72076, Tubingen, Germany We consider projected Newton-type methods for solving large-scale optimiza- tion problems arising in machine learning and related fields. We first intro- duce an algorithmic framework for projected Newton-type methods by review- ing a canonical projected (quasi-)Newton method. This method, while concep- tually pleasing, has a high computation cost per iteration. Thus, we discuss two variants that are more scalable, namely, two-metric projection and in- exact projection methods. Finally, we show how to apply the Newton-type framework to handle non-smooth objectives. Examples are provided through- out the chapter to illustrate machine learning applications of our framework. 1.1 Introduction We study Newton-type methods for solving the optimization problem min x f(x) + r(x), subject to x ? ?, (1.1)

• projected newton methods


• methods


• memory broyden-fletcher- goldfarb-shanno


• hessian


• davidson-fletcher-powell


• newton method


• powell- symmetric-broyden

Sujets

• Planck
• Schmidt
• Austin
• Hessian
• Newton's method

Informations

• Publié par: pefav
• Nombre de lectures: 22
• Langue: English

Extrait

1ProjectedNewton-typeMethodsinMachine Learning [e-mail] [e-mail] MarkSchmidt UniversityofBritishColumbia Vancouver,BC,V6T1Z4 DongminKim UniversityofTexasatAustin Austin,Texas78712 SuvritSra [e-mail] MaxPlanckInsituteforBiologicalCybernetics 72076,Tu¨bingen,Germany WeconsiderprojectedNewton-typemethodsforsolvinglarge-scaleoptimiza- tionproblemsarisinginmachinelearningandrelatedfields.Wefirstintro- duceanalgorithmicframeworkforprojectedNewton-typemethodsbyreview- ingacanonicalprojected(quasi-)Newtonmethod.Thismethod,whileconcep- tuallypleasing,hasahighcomputationcostperiteration.Thus,wediscuss twovariantsthataremorescalable,namely,two-metricprojectionandin- exactprojectionmethods.Finally,weshowhowtoapplytheNewton-type frameworktohandlenon-smoothobjectives.Examplesareprovidedthrough- outthechaptertoillustratemachinelearningapplicationsofourframework. 1.1Introduction WestudyNewton-typemethodsforsolvingtheoptimizationproblem minf(x)+r(x),subjecttox ∈ Ω, x 1.1()

2ProjectedNewton-typeMethodsinMachineLearning wheref : Rn → R istwicecontinuouslydifferentiableandconvex;r : Rn → R iscontinuousandconvexbutnotnecessarilydifferentiableeverywhere; whileΩisa“simple”convexconstraint-set.Thisformulationisgeneral andcapturesnumerousproblemsinmachinelearning,especiallywheref correspondstoaloss,andrtoaregularizer.Letus,however,deferconcrete examplesof(1.1)untilwehavedevelopedsometheoreticalbackground. Weproposetosolve(1.1)viaNewton-typemethods,acertainclassof second-ordermethodsthatareknowntooftenworkwellforunconstrained problems.Forconstrainedproblemstoo,wemayconsiderNewton-type methodsthat,akintotheirunconstrainedversions,iterativelyminimizea quadraticapproximationtotheobjective,thistimesubjecttoconstraints. ThisideadatesbacktoLevitinandPolyak(1966,§7),anditisreferredto asaprojectedNewtonmethod. ProjectedNewtonmethodsforoptimizationoverconvexsetssharemany oftheappealingpropertiesoftheirunconstrainedcounterparts.Forexample, theiriterationsareguaranteedtoimprovetheobjectivefunctionforasmall enoughstepsize;globalconvergencecanbeshownunderavariantofthe Armijocondition;andrapidlocalconvergenceratescanbeshownaround localminimasatisfyingstrongconvexity(Bertsekas,1999).Inasimilarvein, wemayconsiderprojectedquasi-Newtonmethods,whereweinterpolate differencesinparameterandgradientvaluestoapproximatetheHessian matrix.TheresultingNewton-typemethodsarethesubjectofthischapter, andwewillparticularlyfocusonthelimited-memoryBroyden-Fletcher- Goldfarb-Shanno( L-BFGS )quasi-Newtonapproximation.Themainappeal ofthe L-BFGS approximationisitslinear-timeiterationcomplexity,andits strongempiricalperformanceonavarietyofproblems. Summaryofremainder. Wefirstrestrictourselvestosmoothoptimiza- tion,wherer(x)=0.Forthissetting,wedescribeprojectedNewton-type methods(§1.2),coveringbasicimplementationissuessuchasHessianap- proximationandline-search.Then,wedescribetwo-metricprojectionmeth- ods( § 1.3),followedbyinexact(ortruncated)projected-Newtonmethods ( § 1.4).Finally,wediscussthenonsmoothsetting,wherer(x) ￿ =0,forwhich wedescribetwoNewton-typemethods( § 1.5). 1.2ProjectedNewton-typeMethods ProjectedNewton-typemethodsoptimizetheirobjectiveiteratively.At iterationk,theyfirstapproximatetheobjectivefunctionaroundthecurrent