Stochastic recursive algorithms for optimization : simultaneous perturbation methods

309 Pages·2013·4.502 MB·English

by  S. Bhatnagar, H.L. Prasad, and L.A. Prashanth

Lecture Notes
in Control and Information Sciences 434
Editors
ProfessorDr.-Ing.ManfredThoma
InstitutfuerRegelungstechnik,UniversitätHannover,Appelstr.11,30167Hannover,
Germany
E-mail:[email protected]
ProfessorDr.FrankAllgöwer
InstituteforSystemsTheoryandAutomaticControl,UniversityofStuttgart,
Pfaffenwaldring9,70550Stuttgart,Germany
E-mail:[email protected]
ProfessorDr.ManfredMorari
ETH/ETLI29,Physikstr.3,8092Zürich,Switzerland
E-mail:[email protected]
SeriesAdvisoryBoard
P.Fleming
UniversityofSheffield,UK
P.Kokotovic
UniversityofCalifornia,SantaBarbara,CA,USA
A.B.Kurzhanski
MoscowStateUniversity,Russia
H.Kwakernaak
UniversityofTwente,Enschede,TheNetherlands
A.Rantzer
LundInstituteofTechnology,Sweden
J.N.Tsitsiklis
MIT,Cambridge,MA,USA
Forfurthervolumes:
http://www.springer.com/series/642
S. Bhatnagar, H.L. Prasad, and L.A. Prashanth
Stochastic Recursive
Algorithms for Optimization
Simultaneous Perturbation Methods
ABC
Authors
Prof.S.Bhatnagar L.A.Prashanth
DepartmentofComputerScience DepartmentofComputerScience
andAutomation andAutomation
IndianInstituteofScience IndianInstituteofScience
Bangalore Bangalore
India India
H.L.Prasad
DepartmentofComputerScience
andAutomation
IndianInstituteofScience
Bangalore
India
ISSN0170-8643 e-ISSN1610-7411
ISBN978-1-4471-4284-3 e-ISBN978-1-4471-4285-0
DOI10.1007/978-1-4471-4285-0
SpringerLondonHeidelbergNewYorkDordrecht
LibraryofCongressControlNumber:2012941740
(cid:2)c Springer-VerlagLondon2013
Thisworkissubjecttocopyright.AllrightsarereservedbythePublisher,whetherthewholeorpartof
thematerialisconcerned,specificallytherightsoftranslation,reprinting,reuseofillustrations,recitation,
broadcasting,reproductiononmicrofilmsorinanyotherphysicalway,andtransmissionorinformation
storageandretrieval,electronicadaptation,computersoftware,orbysimilarordissimilarmethodology
nowknownorhereafterdeveloped.Exemptedfromthislegalreservationarebriefexcerptsinconnection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of
this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publisher’slocation,initscurrentversion,andpermissionforusemustalwaysbeobtainedfromSpringer.
PermissionsforusemaybeobtainedthroughRightsLinkattheCopyrightClearanceCenter.Violations
areliabletoprosecutionundertherespectiveCopyrightLaw.
Theuseofgeneraldescriptivenames,registerednames,trademarks,servicemarks,etc.inthispublication
doesnotimply,evenintheabsenceofaspecificstatement,thatsuchnamesareexemptfromtherelevant
protectivelawsandregulationsandthereforefreeforgeneraluse.
Whiletheadviceandinformationinthisbookarebelievedtobetrueandaccurateatthedateofpub-
lication,neithertheauthorsnortheeditorsnorthepublishercanacceptanylegalresponsibilityforany
errorsoromissionsthatmaybemade.Thepublishermakesnowarranty,expressorimplied,withrespect
tothematerialcontainedherein.
Printedonacid-freepaper
SpringerispartofSpringerScience+BusinessMedia(www.springer.com)
To SB’sparentsDr. G. K. Bhatnagarand
Mrs.S.K. Bhatnagar,hiswifeAartiand
daughterShriya
To HLP’s parentsDr.H. R. Laxminarayana
Bhattaand Mrs.G. S.Shreemathi,
brother-in-lawGiridharN. Bhat andsister
VathsalaG. Bhat
To LAP’s daughterSamudyata
Preface
TheareaofstochasticapproximationhasitsrootsinapaperpublishedbyRobbins
and Monro in 1951, where the basic stochastic approximation algorithm was in-
troduced.Eversince,ithasbeenappliedinavarietyofapplicationscuttingacross
severaldisciplinessuchascontrolandcommunicationengineering,signalprocess-
ing,roboticsandmachinelearning.
KieferandWolfowitz,inapaperin1952(nearlysixdecadesago)publishedthe
first stochastic approximationalgorithmfor optimization.The algorithmproposed
by them was a gradient search algorithm that aimed at finding the maximum of
a regression function and incorporated finite difference gradient estimates. It was
later found that whereas the Kiefer-Wolfowitz algorithm is efficient in scenarios
involving scalar parameters, this is not necessarily the case with vector parame-
ters,particularlythoseforwhichtheparameterdimensionishigh.Theproblemthat
arises is that the number of function measurements needed at each update epoch
grows linearly with the parameter dimension. Many times, it is also possible that
theobjectivefunctionisnotobservableassuchandoneneedstoresortto simula-
tion.Insuchscenarios,withvectorparameters,onerequiresacorresponding(linear
intheparameter-dimension)numberofsystemsimulations.Inthecaseoflargeor
complexsystems,thiscanresultinasignificantcomputationaloverhead.
Subsequently,inapaperpublishedin1992,Spallproposedastochasticapprox-
imationschemeforoptimizationthatdoesarandomsearchintheparameterspace
and only requires two system simulations regardless of the parameter dimension.
This algorithm that came to be known as simultaneous perturbation stochastic
approximation or SPSA for short, has become very popular because of its high
efficiency, computational simplicity and ease of implementation. Amongst other
impressive works, Katkovnik and Kulchitsky, in a paper published in 1972, also
proposed a random search scheme (the smoothed functional (SF) algorithm) that
onlyrequiresonesystemsimulationregardlessoftheparameterdimension.Subse-
quentworkshowedthatatwo-simulationcounterpartofthisschemeperformswell
inpractice.BoththeKatkovnik-KulchitskyaswellastheSpallapproachesinvolve
perturbing the parameter randomly by generating certain i.i.d. random variables.
VIII Preface
The difference between these schemes lies in the distributions these perturbation
randomvariablescanpossessandtheformsofthegradientestimators.
Stochasticapproximationalgorithmsforoptimizationcanbeviewedascounter-
partsofdeterministicsearchschemeswithnoise.Whereas,theSPSAandSFalgo-
rithmsaregradient-basedalgorithms,duringthelastdecadeorso,therehavebeen
paperspublishedonNewton-basedsearchschemesforstochasticoptimization.Ina
paperin2000,SpallproposedthefirstNewton-basedalgorithmthatestimatedboth
the gradientand the Hessian usinga simultaneousperturbationapproachincorpo-
rating SPSA-type estimates. Subsequently, in papers published in 2005 and 2007,
Bhatnagar proposed more Newton-based algorithms that develop and incorporate
bothSPSAandSFtypeestimatesofthegradientandHessian.Inthistext,wecom-
monlyrefertoallapproachesforstochasticoptimizationthatarebasedonrandomly
perturbingparametersinordertoestimatethegradient/Hessianofagivenobjective
functionassimultaneousperturbationmethods.Bhatnagarandcoauthorshavealso
developedandappliedsuchapproachesforconstrainedstochasticoptimization,dis-
crete parameter stochastic optimization and reinforcementlearning – an area that
deals with the adaptive control of stochastic systems under real or simulated out-
comes. The authorsof this bookhave also studied engineeringapplicationsof the
simultaneous perturbation approaches for problems of performance optimization
in domainssuch as communicationnetworks,vehiculartraffic controland service
systems.
Themainfocusofthistextisonsimultaneousperturbationmethodsforstochas-
ticoptimization.Thisbookisdividedintosixpartsandcontainsatotaloffourteen
chapters and five appendices. Part I of the text essentially provides an introduc-
tiontooptimizationproblems-bothdeterministicandstochastic,givesanoverview
of search algorithms and a basic treatment of the Robbins-Monro stochastic ap-
proximationalgorithm as well as a generalmulti-timescale stochastic approxima-
tionscheme.PartIIofthetextdealswithgradientsearchstochasticalgorithmsfor
optimization.Inparticular,theKiefer-Wolfowitz,SPSAandSFalgorithmsarepre-
sentedanddiscussed.PartIIIdealswithNewton-basedalgorithmsthatareinpartic-
ularpresentedforthelong-runaveragecostobjective.Thesealgorithmsarebasedon
SPSAandSFbasedestimatorsforboththegradientandtheHessian.PartIVofthe
book deals with a few variationsto the generalscheme and applicationsof SPSA
and SF based approachesthere. In particular, we consider adaptationsof simulta-
neous perturbation approaches for problems of discrete optimization, constrained
optimization(underfunctionalconstraints) as well as reinforcementlearning.The
long-runaveragecostcriterionwillbeconsideredherefortheobjectivefunctions.
PartVofthebookdealswiththreeimportantapplicationsrelatedtovehiculartraf-
ficcontrol,servicesystemsaswellascommunicationnetworks.Finally,fiveshort
appendicesat the end summarize some of the basic material as well as important
resultsusedinthetext.
This book in many ways summarizesthe variousstrands of research on simul-
taneousperturbationapproachesthatSBhasbeeninvolvedwith duringthecourse
of the last fifteen years or so. Both HLP and LAP have also been working in this
areaforoverfiveyearsnowandhavebeenactivelyinvolvedinthevariousaspects
Preface IX
oftheresearchreportedhere.Alargeportionofthistext(inparticular,PartsIII-V
aswellasportionsofPartII)isbasedmainlyontheauthors’owncontributionsto
thisarea.The textprovidesa compactcoverageof thematerialin a waythatboth
researchersandpractitionersshouldfinduseful.Thechoiceoftopicsisintendedto
coverasufficientwidthwhileremainingtiedtothecommonthemeofsimultaneous
perturbationmethods. While we have made attempts at conveyingthe main ideas
behindthevariousschemesandalgorithmsaswellastheconvergenceanalyses,we
have also included sufficient material on the engineering applications of these al-
gorithmsin order to highlightthe usefulness of these methodsin solving real-life
engineeringproblems.Asmentionedbefore,anentirepartofthetext,namelyPart
IV, comprising of three chapters is dedicated for this purpose. The text in a way
providesabalancedcoverageofmaterialrelatedtoboththeoryandapplications.
Acknowledgements
SBwasfirstintroducedtotheareaofstochasticapproximationduringhisPh.Dwork
withProf.VivekBorkarandProf.VinodSharmaattheIndianInstituteofScience.
Subsequently,hebegantolookatsimultaneousperturbationapproacheswhiledo-
ing a post doctoral with Prof. Steve Marcus and Prof. Michael Fu at the Institute
forSystemsResearch,UniversityofMaryland,CollegePark.Hehasalsobenefitted
significantlyfromreadingtheworksofProf.JamesSpallandthroughinteractions
withhim.Hewouldliketothankallhiscollaboratorsovertheyears.Inparticular,
he would like to thank Prof. Vivek Borkar,Prof. Steve Marcus, Prof. MichaelFu,
Prof.RichardSutton,Prof.CsabaSzepesvari,Prof.VinodSharma,Prof.Karmeshu,
Prof. M. Narasimha Murty, Prof. N. Hemachandra, Dr. Ambedkar Dukkipati and
Dr. MohammedShahidAbdulla.He wouldlike to thankProf.AnuragKumarand
Prof. K. V. S. Hari for severalhelpfuldiscussions on optimization approachesfor
certain problemsin vehicular traffic control(during the course of a joint project),
whichis also the topicofChapter 13in thisbook.SB considershimselffortunate
to have had the pleasure of guiding and teaching several bright students at IISc.
He would like to acknowledge the work done by all the current and former stu-
dents of the Stochastic Systems Laboratory. A large part of SB’s research during
thelasttenyearsatIISchasbeensupportedthroughprojectsfromtheDepartment
of Science and Technology,Departmentof InformationTechnology,TexasInstru-
ments, Satyam Computers, EMC and Wibhu Technologies.SB would also like to
acknowledge the various institutions where he worked and visited during the last
fifteen years where portions of the work reported here have been conducted: The
Institute for Systems Research, University of Maryland, College Park; Vrije Uni-
versiteit,Amsterdam;IndianInstituteofTechnology,Delhi;TheRLAILaboratory,
UniversityofAlberta;andtheIndianInstituteofScience.Amajorpartofthework
reportedherehasbeenconductedatIIScitself.Finally,hewouldliketo thankhis
parentsDr. G. K. Bhatnagarand Mrs. S. K. Bhatnagarfor their support,help and
guidanceallthroughtheyears,hiswifeAartianddaughterShriyafortheirpatience,
X Preface
understandingandsupport,andhisbrotherDr.Shashankforhisguidanceandteach-
ingduringSB’sformativeyears.
HLP’s interest in the area of control engineering and decision making, which
was essentially sown in him by interactions with Prof. U. R. Prasad at I.I.Sc.,
Dr. K. N. Shubhanga and Jora M. Gonda at NITK, led him to the area of oper-
ations research followed by that of stochastic approximation. He derives inspira-
tionsfromtheworksofProf.VivekBorkar,Prof.JamesSpall,Prof.RichardSutton,
Prof.ShalabhBhatnagarandseveraleminentpersonalitiesinthefieldofstochastic
approximation.He thanksDr. Nirmit V. Desai at IBM Research, India, collabora-
tionwithwhom,broughtupseveralnewstochasticapproximationalgorithmswith
practicalapplicationstotheareaofservicesystems.HethanksProf.ManjunathKr-
ishnapurandProf.P.S.Sastryforequippinghimwithmathematicalrigourneeded
forstochasticapproximation.He thanksI.R.RaoatNITKwhohasbeenconstant
sourceofmotivation.HethankshisfatherDr.H.R.LaxminarayanaBhatta,mother
Mrs.G.S.Shreemathi,brother-in-lawGiridharN.BhatandsisterVathsalaG.Bhat,
fortheirconstantsupportandunderstanding.
LAP would like to thank his supervising professorSB for introducingstochas-
tic optimization during his PhD work. The extensive interactions with SB served
to sharpen the understanding of the subject. LAP would also like to thank HLP,
collaborationwithwhomhasbeenmostvaluable.LAP’sprojectassociateshipwith
Departmentof InformationTechnologyas wellas his internshipat IBM Research
presentedmanyopportunitiesfordevelopingaswellasapplyingsimultaneousper-
tubation methodsin variouspracticalcontexts and LAP would like to thank Prof.
Anurag Kumar, Prof. K.V.S. Hari of ECE department, IISc and Nirmit Desai and
GargiDasguptaofIBMResearchforseveralusefulinteractionsonthesubjectmat-
ter.Finally,LAPwouldliketothankhisfamilymembers-particularlyhisparents,
wifeandhisdaughter,fortheirsupportinthisendeavour.
Bangalore, S.Bhatnagar
May2012 H.L.Prasad
L.A.Prashanth
Contents
PartI IntroductiontoStochasticRecursiveAlgorithms
1 Introduction................................................... 3
1.1 Introduction ............................................... 3
1.2 OverviewoftheRemainingChapters.......................... 7
1.3 ConcludingRemarks........................................ 11
References..................................................... 11
2 DeterministicAlgorithmsforLocalSearch........................ 13
2.1 Introduction ............................................... 13
2.2 DeterministicAlgorithmsforLocalSearch ..................... 14
References..................................................... 15
3 StochasticApproximationAlgorithms ............................ 17
3.1 Introduction ............................................... 17
3.2 TheRobbins-MonroAlgorithm............................... 18
3.2.1 ConvergenceoftheRobbins-MonroAlgorithm ........... 19
3.3 Multi-timescaleStochasticApproximation ..................... 23
3.3.1 ConvergenceoftheMulti-timescaleAlgorithm ........... 24
3.4 ConcludingRemarks........................................ 26
References..................................................... 27
PartII GradientEstimationSchemes
4 Kiefer-WolfowitzAlgorithm..................................... 31
4.1 Introduction ............................................... 31
4.2 TheBasicAlgorithm........................................ 31
4.2.1 ExtensiontoMulti-dimensionalParameter ............... 35
4.3 VariantsoftheKiefer-WolfowitzAlgorithm .................... 36
4.3.1 FixedPerturbationParameter .......................... 36
4.3.2 One-SidedVariants................................... 37
XII Contents
4.4 ConcludingRemarks........................................ 38
References..................................................... 38
5 GradientSchemes withSimultaneousPerturbationStochastic
Approximation ................................................ 41
5.1 Introduction ............................................... 41
5.2 TheBasicSPSAAlgorithm .................................. 41
5.2.1 GradientEstimateUsingSimultaneousPerturbation ....... 42
5.2.2 TheAlgorithm ...................................... 43
5.2.3 ConvergenceAnalysis ................................ 44
5.3 VariantsoftheBasicSPSAAlgorithm ......................... 47
5.3.1 One-MeasurementSPSAAlgorithm .................... 47
5.3.2 One-SidedSPSAAlgorithm ........................... 49
5.3.3 FixedPerturbationParameter .......................... 49
5.4 GeneralRemarksonSPSAAlgorithms ........................ 51
5.5 SPSAAlgorithmswithDeterministicPerturbations .............. 52
5.5.1 PropertiesofDeterministicPerturbationSequences........ 52
5.5.2 HadamardMatrixBasedConstruction................... 54
5.5.3 Two-SidedSPSAwithHadamardMatrixPerturbations .... 56
5.5.4 One-SidedSPSAwithHadamardMatrixPerturbations..... 62
5.5.5 One-MeasurementSPSA with Hadamard Matrix
Perturbations........................................ 63
5.6 SPSAAlgorithmsforLong-RunAverageCostObjective ......... 65
5.6.1 TheFramework...................................... 65
5.6.2 TheTwo-SimulationSPSAAlgorithm................... 65
5.6.3 Assumptions ........................................ 66
5.6.4 ConvergenceAnalysis ................................ 68
5.6.5 ProjectedSPSAAlgorithm ............................ 73
5.7 ConcludingRemarks........................................ 75
References..................................................... 75
6 SmoothedFunctionalGradientSchemes.......................... 77
6.1 Introduction ............................................... 77
6.2 GaussianBasedSFAlgorithm................................ 79
6.2.1 GradientEstimationviaSmoothing ..................... 79
6.2.2 TheBasicGaussianSFAlgorithm ...................... 81
6.2.3 ConvergenceAnalysisofGaussianSFAlgorithm ......... 82
6.2.4 Two-MeasurementGaussianSFAlgorithm............... 88
6.3 GeneralConditionsforaCandidateSmoothingFunction.......... 91
6.4 CauchyVariantoftheSFAlgorithm........................... 92
6.4.1 GradientEstimate.................................... 92
6.4.2 CauchySFAlgorithm ................................ 94
6.5 SFAlgorithmsfortheLong-RunAverageCostObjective ......... 94
6.5.1 TheG-SF1Algorithm ................................ 95
6.5.2 TheG-SF2Algorithm ................................ 99