lec3_6245_2004

Embed Size (px)

Citation preview

  • 7/28/2019 lec3_6245_2004

    1/8

    Massachusetts InstituteofTechnologyDepartmentofElectricalEngineeringandComputerScience6.245: MULTIVARIABLECONTROLSYSTEMS

    byA.Megretski

    SolvingtheH2optimizationproblem1ThereareseveralwaystoderiveasolutiontotheH2optimizationproblem. Thepathdevelopedbelowreliesonreductionoftheoutputfeedbackdesignproblemtoanabstractoptimalprogramcontrolproblem.3.1 ACharacterizationofClosedLoop ImpulseResponsesThissectionprovidesan insight intothe constraints imposedontheclosed loop systembythecoefficientsoftheouputfeedbackdesignplant

    x = Ax+B1w+B2u, (3.1)y = C2x+D21w, (3.2)

    stabilizedbyafiniteorderstrictlyproperLTIcontrollerxf = Afxf +Bfy, (3.3)

    u = Cfxc. (3.4)c1VersionofFebruary17,2004

    A.Megretski,2004

  • 7/28/2019 lec3_6245_2004

    2/8

    23.1.1 AnAffineParameterizationLetX=X(t)denotetheimpulseresponsematrixfromwtoxintheclosedloopsystem.Similarly,letU =U(t)betheimpulseresponsematrixfromwtou.Theorem3.1 Assume (A, B2) is stabilizable and (C2, A) is detectable. Then a pair(X(t), U(t)) ofmatrix-valuedfunctions can be achieved as implulse responses in a stableclosed loopsystem ifandonly if

    X (t)=AX(t)+B2U(t), X(0)=B1, limX(t)=0, (3.5)t

    and there exist a realmatrix-valuedfunction V = V(t), each entryofwhich is a linearcombinationoftermsof theformtkest withk {0, 1, . . . },Re(s)>0,such that

    U(t)=(t)B1 +V(t)D21, lim(t)=0, (3.6)twhere

    (t)=(t)A +V(t)C2, (0)=0. (3.7)Equality (3.5) represents the constraints on the closed loop responses in the case of

    full informationcontrol, whem xand w areavailable formeasurement. Whenoutputfeedbackisconsidered,U(t)isnotarbitrary,butisparameterizedby(3.6),(3.7).

    Theorem3.1canbegeneralizedtothecaseoffiniteH2normstabilizationbyinfiniteorderLTIcontrollers, inwhichcasetheconditionsrequiringX(t), (t)and(t)toconvergetozeroast shouldbereplacedbytheconditionoftheirsquare integrabilityover{t} =(0, ).Proof For the purpose of driving the H2 optimal controller, we do not need the sufficiencypartofthetheorem. Itwillbeprovenlater,togetherwithaslightmodificationofTheorem3.1concerningtheso-calledQ-parameterization.

    ItissufficienttoprovethetheoremforthecasewhenB1 = I 0 , D21 = 0 I ,

    i.e. whensystemequationshavetheformx = Ax +B2u +w1, (3.8)y = C2x +w2 (3.9)

    withw= [w1;w2]. Then(3.5)followsfrom(3.8). Onewaytoseethisisbyapplyingtheone-sidedLaplacetransform(denotedbytildes)to(3.8)toget

    x=A u + s x +B2 w1.

  • 7/28/2019 lec3_6245_2004

    3/8

    3Since

    w,

    w,

    w1 = I 0 w,x=X u=U andw isarbitrary,(3.5)follows.

    Toprove(3.6)and(3.7),notethat,by linearityofthecontroller, u(t)0whenevery(t)0. Hence

    w2 =C2(sIA)1 w1wouldalwaysimply

    I

    u=U C2(sIA)1 w1 =0

    foranarbitraryw1. Therefore U = VC2(sIA)1 V

    forsomeV. ConvertingthistotimedomaintermsyieldsU =[ V]where(t)=(t)A+V(t)C2, (0)=0.

    3.1.2 ACharacterizationofStateEstimationErrorsAsaspecialcaseofthegeneralclosedloopresponseparameterization,considertheclosedloopresponsefromwtoKxu,whereK isagivenmatrix.Theorem3.2 A matrixfunction G = G(t) can be achieved as a closed loop impulseresponsefrom w to Kxu ifandonly if thereexistsarealmatrix-valuedfunction V =

    stV(t), each entry ofwhich is a linear combination of terms of theform tke with k {0,1,...},Re(s)>0,such that

    G(t)=(t)B1 +V(t)D21, lim(t)=0, (3.10)t

    where(t)=(t)A+V(t)C2, (0)=K. (3.11)

    Proof Fornow,wewillonlyprovenecessity(thisisthepartweneedforH2optimization).AccordingtoTheorem3.1, astabilizedclosed loop impulseresponse from w to KxuequalstheresponseH ofsystem

    X = AX+B2(B1 +VD21), X(0)=B1, (3.12) = A+VC2, (0)=0, (3.13)H = KXB1 VD21 (3.14)

  • 7/28/2019 lec3_6245_2004

    4/8

    4toaninputV whichmakesX()=0and()=0.

    Notethat

    the

    zero

    input

    response

    of

    system

    (3.12),

    (3.13)

    (i.e.

    with

    V

    0)

    has

    0

    andhenceH(t)=H0(t)=KeAtB1. Thisresponsecanbeachieved in(3.10),(3.11)withV 0. On the other hand, according to Theorem 3.1, every zero-state response of thesystem (i.e. when X(0)=B1 is replaced by X(0)=0) must coincidewithazero-stateresponseof(3.10),(3.11).3.2 AbstractH2OptimizationWe will use the term abstract H2 optimization to refer to an auxiliary optimizationproblem: an optimal program control in completelydeterministic settings. Itturns outthatageneralH2feedbackdesignproblemreducestoapairofabstractH2optimizations,and the optimal controller, as well as a parameterization of suboptimal controllers, canbeobtainedeasilyfromsolutionsofthetwoauxiliaryproblems.3.2.1 FormalDefinitionsandRelationtoOutputFeedbackDesignAnabstractH2optimizationproblemisdefinedbyastabilizableLTIsystem

    a b

    S = .c d

    andrequestsfinding,foreveryfixedvectorp0 inthestatespace,afunctionoftimeq=q(t)(boundedfort [0,)andconvergingtozeroast )suchthatthesolutionp=p(t)of

    p(t)=ap(t)+bq(t), p(0)=p0 (3.15)satisfies

    limp(t)=0, (3.16)t

    and,subjecttothisconstraints,minimizesthequadraticintegral(u())= |cp(t)+dq(t)|2dt min. (3.17)

    0One motivation for considering abstract H2 optimization is the full information

    versionofthestandardH2outputfeedbackdesign. AccordingtoTheorem3.1,theclosedloopH2normcanbeexpressedastheintegral

    2fi = C1X(t)+D12U(t)Fdt,0

  • 7/28/2019 lec3_6245_2004

    5/8

    5where

    M2F =trace(MM)

    denotes the square of the Frobenius norm of matrix M, and X,U are the closed loopimpulseresponsesfromwtoxandurespectively,constrainedonlyby

    X =AX+B2U, X(0)=B1, X()=0, U()=0.IfXi,Ui,B1i denotethei-thcolumnofX,U,andB1 respectively,fi canbewrittenas

    fi = |C1Xi(t)+D12Ui(t)|2dt.

    0iSinceevery3-typle(Xi,Ui,B1i isindependentlyconstrainedby

    X i(t)=AXi(t)+B2Ui(t), Xi(0)=B1i,thetaskofminimizingfi decomposesintoindependentabstractH2minimizationswith

    a=A, b=B2, c=C1, d=D12, p0 =B1i.AsimilarmotivationforabstractH2optimizationcomesfromanattempttominimize

    the closed loop H2 norm from w to Kxu, where K is a given matrix. According toTheorem3.2,theclosedloopimpulseresponsecanbechosenaccordingto

    H=B1 +VD21, =A+VC2, (0)=K.Leti,Vi,Ki bethei-throwof,V,Krespectively. TheH2normofHequalsthesumoftheintegrals

    |i(t)B1 +Vi(t)D21|2dt,0

    minimizingwhichleadstosolvingafamilyofindependentabstractH2optimizationproblemswith

    a=A, b=C2, c=B1, d=D21, p0 =Ki.3.2.2 TheeasyversionoftheKYPLemmaThe so-called Kalman-Yakubovich-Popov Lemma (also frequently referred to as thePositive Real Lemma provides, among other things, a complete solution to abstractH2 optimization. The much simplified version of the KYP Lemma we need here willbe complemented by a full statement used extensively in H-Infinity optimization andoptimizationrelyingonsemidefiniteprogramming.

  • 7/28/2019 lec3_6245_2004

    6/8

    6Theorem3.3 Assume thatpair (a,b) is stabilizable. Then thefollowingconditionsareequivalent:

    (a) auniqueoptimalcontrolexistsforeveryp0 intheabstractH2optimizationproblem

    (3.15)-(3.17);(b) matrix

    ajI bM(j)=

    c dis left invertibleforall Randat =(whichmeans that d is left invertibleaswell);

    (c)

    d

    d>

    0,

    and

    matrix

    ab(dd)1dc b(dd)1b

    H= c ccd(dd)1dc a +cd(dd)1b

    hasnoeigenvalueson the imaginaryaxis;(d) dd>0and thereexist (unique)matrices = 0and k such that a+bk isa

    Hurwitzmatrix,and|cp+dq|2 +2p (ap+bq)=(qkp)dd(qkp) (3.18)

    forallvectorsp,q.Ifconditions(a)-(d)aresatisfiedthentheoptimalq isdefinedbytherelationq=kp,i.e.

    q(t)=ke(a+bk)tp0,andtheminimalcostequalsp0p0.Moreover,thedimensionofthestrictlystableinvariant

    subspaceV+ ofHwillequalthenumbernofcomponentsofp,andcanbefoundfrom

    1= 1 2 ... n x1 x2 ... xn ,

    where x1 x2 xn, ,1 2 n

    isanarbitrarybasis inV+.

  • 7/28/2019 lec3_6245_2004

    7/8

    7Matrix H is usually called theHamiltonianmatrix associated with the abstract H2

    optimization.

    Condition

    (c)

    provides

    a

    convenient

    way

    of

    checking

    the

    non-singularity

    imposed by (b). Condition (d) defines a completion of squares procedure, which is anaturalandintuitivewayofsolvingtheabstractH2optimizationproblem. Bycomparingthecoefficientsonbothsidesof(3.18),onecanderiveaquadraticalgebraicequationfor,calledthealgebraicRiccatiequation(ARE):

    ++=,where,, arethecoefficientsoftheHamiltinianmatrix:

    H= .

    Asolutionforwhicha+bk=

    isaHurwitzmatrix iscalledastabilizing solutionoftheARE.Astabilizingsolution, ifitexists,isunique. Itdefinestheoptimalcontrollergaink inq=kp,though,formally,abstractH2optimizationisaboutprogramcontroloptimization.

    WewillpostponeprovingtheKYPLemmauntilitsfullversionisformulated.3.2.3 SolutionofFeedbackH2OptimizationInordertosolvethestandardH2feedbackoptimizationproblem,considerthetwoauxiliary abstract H2 optimization setups, the full information control version, definedwith

    a=A, b=B2, c=C1, d=D12, (3.19)andthestateestimationversion,definedby

    a=A, b=C2, c=B1, d=D21. (3.20)Itiseasytoverifybyinspectionthatabsenseofcontrolsingularityintheoriginalfeedbackdesignsetupisequivalenttosatisfyingcondition(b)ofTheorem3.3intheabstractsetupdefinedby(3.19),andabsenseofsensorsingularity isequivalenttosatisfying(b) intheabstract setup defined by (3.19). Hence, in the non-singular case, the correspondingcompletion of squares is possible, with matrices = Pfi, k = Kfi in the control setup,=Pse andk=Kse inthestateestimationsetup.

  • 7/28/2019 lec3_6245_2004

    8/8

    8Theorem3.4 TheH2optimalcontroller isdefinedby

    u(t) = Kfix(t), (3.21)dx(t)

    dt = Ax(t)+B2u(t)+Kse(C2x(t)y(t)), (3.22)andprovidestheminimalH2norm(squared)of

    J =trace(B 1PfiB1 +trace(D12KfiPseKfeD12).Proof TheoverallclosedloopH2norm(squared)isgivenby

    J=

    C1X(t)

    +

    D12U

    (t)

    2Fdt.0

    AccordingtothedefinitionofKfi andPfi,thisintegralequals

    J =trace(B1PfiB1 + D12(UKfiX)2Fdt.0

    Thiscanbe interpretedassayingthat,intheH2optimalclosedloopsystem,umustbethebestestimateofKfix,inthesensethattheH2normfromwtoD12(uKfix)shouldbe minimal. According to the definition of Kse and Pse, the minimal estimation errorequals

    K

    D

    trace(D12KfiPse fe 12),andisachievedwhentheimpulseresponseGfromwtoKfixuisgivenby

    G(t)=(t)B1 +V(t)D21,where

    (t)=(t)A+V(t)C2, (0)=K,and

    V(t)=(t)Kse.By inspection,thisoptimal impulseresponse isachievedbythecontroller inthe formulationofthetheorem.