Rotkowitz-when is a Linear Controller Optimal

Embed Size (px)

Citation preview

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    1/57

    When is a Linear Controller Optimal?

    Michael Rotkowitz

    Department of Electrical and Electronic Engineering

    Department of Mathematics and StatisticsThe University of Melbourne

    DISC Summerschool on Distibuted Control and EstimationNoordwijkerhout, The Netherlands5 June 2009

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    2/57

    2 M. Rotkowitz, Melbourne Uni

    General Formulation

    P11 P12

    P21

    G

    K

    w

    uy

    z

    minimize J(fl(P, K))

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    3/57

    3 M. Rotkowitz, Melbourne Uni

    General Formulation

    P11 P12

    P21

    G

    K

    w

    uy

    z

    minimize J(fl(P, K))

    subject to KS

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    4/57

    4 M. Rotkowitz, Melbourne Uni

    Uniform Optimal (Centralized) Control

    minimize fl(P, K)

    where

    G = supw=0

    Gw2w2

    When plantP is linear, optimal controllerK is indeed linear

    Feintuch and Francis, 1985 Khargonekar and Poolla, 1986

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    5/57

    5 M. Rotkowitz, Melbourne Uni

    LQGminimize E (h (z)) wherez = fl(P, K) w

    strictly classical information pattern separation

    = F

    Gaussian noise linear estimation

    F linear

    quadratic cost linear control

    linear

    Witsenhausen, 1971

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    6/57

    6 M. Rotkowitz, Melbourne Uni

    Witsenhausen Counterexample (1968)

    1

    x0 x1

    2

    x2

    y2 u2u1y1

    w1

    v = w2

    +

    +

    +

    +

    +

    Objective was to seek1, 2 to minimize

    E k2u21+x22where

    w1 N(0, 1) w2 N(0, 1)

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    7/57

    0 y1

    y1 0

    7 M. Rotkowitz, Melbourne Uni

    W.C. - Full Information

    1

    x0 x1

    2

    x2

    y2 u2u1y1

    w1

    v

    =w

    2

    +

    +

    +

    +

    +

    1 2 E(u21) x2 E(x

    22) J

    (0) full info 0 y1 0 0 0 0

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    8/57

    0

    y1

    y2

    v

    8 M. Rotkowitz, Melbourne Uni

    W.C. - No Input Cost

    1

    x0 x1

    2

    x2

    y2 u2u1y1

    w1

    v = w2

    +

    +

    +

    +

    +

    1 2 E(u21) x2 E(x

    22) J

    (0) full info 0 y1 0 0 0 0(1) no input cost 0 y2 0 v 1 1

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    9/57

    y1 0

    0 0

    9 M. Rotkowitz, Melbourne Uni

    W.C. - No Output Cost

    1

    x0 x1

    2

    x2

    y2 u2u1y1

    w1

    v = w2

    +

    +

    +

    +

    +

    Letk= 0.1and = 10.

    1 2 E(u21) x2 E(x

    22) J

    (0) full info 0 y1 0 0 0 0

    (1) no input cost 0 y2 0 v 1 1(2) no output cost y1 0

    2 = 100 0 0 k22 = 1

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    10/57

    ay1 by2

    10 M. Rotkowitz, Melbourne Uni

    W.C. - Best Linear

    1

    x0 x1

    2

    x2

    y2 u2u1y1

    w1

    v = w2

    +

    +

    +

    +

    +

    1 2 E(u21) x2 E(x22) J(0) full info 0 y1 0 0 0 0(1) no input cost 0 y2 0 v 1 1

    (2) no output cost y1 0 100 0 0 1(3) best linear 0.01y1 0.99y2 0.01 - 0.9899 0.99

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    11/57

    y1+H sgn(y1) H sgn(y2)

    +H, ify10H, ify1

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    12/57

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    13/57

    13 M. Rotkowitz, Melbourne Uni

    Witsenhausen Counterexample (1968)

    1

    x0 x1

    2

    x2

    y2 u2u1y1

    w1

    v = w2

    +

    +

    +

    +

    +

    Objective was to seek1, 2 to minimize

    E k2u21+x22where

    w1 N(0, 1) w2 N(0, 1)

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    14/57

    P11 P12P21 G

    K

    w

    uy

    z

    14 M. Rotkowitz, Melbourne Uni

    Witsenhausen Counterexample (1968)

    1

    x0 x1

    2

    x2

    y2 u2u1y1

    w1

    v = w2

    +

    +

    +

    +

    +

    Letting

    w=

    w1w2

    z=

    k u1

    x2

    we then seek to minimize

    J(1, 2) = Ez22

    wherew N(0, I)

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    15/57

    P11 P12P21 G

    K

    w

    uy

    z

    15 M. Rotkowitz, Melbourne Uni

    Witsenhausen Counterexample - Induced

    Norm

    1

    x0 x1

    2

    x2

    y2 u2u1y1

    w1

    v = w2

    +

    +

    +

    +

    +

    Instead consider the induced norm, and seek1, 2 to minimize

    J(1, 2) = sup

    w=0

    z2

    w2

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    16/57

    16 M. Rotkowitz, Melbourne Uni

    Induced Norms

    supw=0

    z2w2

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    17/57

    17 M. Rotkowitz, Melbourne Uni

    Induced Norms

    supw=0

    z2w2

    supw

    z2w2

    18 M R k i M lb U i

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    18/57

    18 M. Rotkowitz, Melbourne Uni

    Induced Norms

    supw=0

    z2w2

    supw

    z2w2

    a = d1

    dy1

    y1=0

    b = d2

    dy2

    y2=0

    1 ay1 2 by2

    19 M Rotkowitz Melbourne Uni

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    19/57

    19 M. Rotkowitz, Melbourne Uni

    Induced Norms

    supw=0

    z2w2

    supw

    z2w2

    a = d1

    dy1

    y1=0

    b = d2

    dy2

    y2=0

    1 ay1 2 by2

    J(1, 2) = supw=0

    z2w2 Jl(a, b)

    20 M Rotkowitz Melbourne Uni

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    20/57

    20 M. Rotkowitz, Melbourne Uni

    Induced Norms

    supw=0

    z2w2

    supw

    z2w2

    a = d1

    dy1

    y1=0

    b = d2

    dy2

    y2=0

    1 ay1 2 by2

    J(1, 2) = supw=0

    z2w2 Jl(a, b)

    May as well fix

    1(y1) = ay1 2(y2) = by2Cant do any better!

    21 M Rotkowitz Melbourne Uni

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    21/57

    21 M. Rotkowitz, Melbourne Uni

    Induced Norms

    1 ay1 2 by2

    J(1, 2) = supw=0

    z2w2

    Jl(a, b)

    May as well fix1(y1) = ay1 2(y2) = by2

    Cant do any better!

    Given any nonlinear controller thats differentiable in the origin, wehave a linear controller thats at least as good.

    End of story? Not so fast.

    22 M. Rotkowitz, Melbourne Uni

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    22/57

    ,

    Another Induced Norm

    L1 / 1 ( )Norm

    sup

    w=0

    zw

    (M. Dahleh and J. Shamma, 1992) Nonlinear controllers which aredifferentiable in the origin cannot do better than linear controllers.

    But.............

    23 M. Rotkowitz, Melbourne Uni

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    23/57

    A. Stoorvogel, 1995

    z =

    0 3 01.5 1.5 33 0 00 0 3

    w+

    1111

    u

    y = w

    u = Ky

    FindKto minimize

    J(K) = sup

    w=0

    z

    w

    24 M. Rotkowitz, Melbourne Uni

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    24/57

    A.S. - Unit Cube

    For linear controllers, need only con-sider the eight corners of the unitcube.

    25 M. Rotkowitz, Melbourne Uni

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    25/57

    A.S. - Outputs

    For linear controllers, need only con-sider the eight corners of the unitcube.

    zi = 3 + |ui| fori= 1, 2, 3, 5, 6, 7

    z4 = max{|u4 3|, |u4 6|}

    z8 = max{|u8+ 3|, |u8+ 6|}

    26 M. Rotkowitz, Melbourne Uni

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    26/57

    A.S. - Best Linear

    For linear controllers, need only con-sider the eight corners of the unitcube.

    zi = 3 + |ui| fori= 1, 2, 3, 5, 6, 7

    z4 = max{|u4 3|, |u4 6|}

    z8 = max{|u8+ 3|, |u8+ 6|}

    Best linear controller achieves closed-loop norm of 3.75; achieved by

    u =

    0.75 0.75 0.75

    w

    27 M. Rotkowitz, Melbourne Uni

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    27/57

    A.S. - Linear Controller

    28 M. Rotkowitz, Melbourne Uni

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    28/57

    A.S. - Nonlinear Control

    For nonlinear controllers, also con-sider the eight corners of the unitcube.

    zi = 3 + |ui| fori= 1, 2, 3, 5, 6, 7

    z4 = max{|u4 3|, |u4 6|}

    z8 = max{|u8+ 3|, |u8+ 6|}

    29 M. Rotkowitz, Melbourne Uni

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    29/57

    A.S. - Nonlinear Control

    For nonlinear controllers, also con-sider the eight corners of the unitcube.

    zi = 3 + |ui| fori= 1, 2, 3, 5, 6, 7

    z4 = max{|u4 3|, |u4 6|}

    z8 = max{|u8+ 3|, |u8+ 6|}

    30 M. Rotkowitz, Melbourne Uni

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    30/57

    A.S. - Nonlinear Control

    31 M. Rotkowitz, Melbourne Uni

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    31/57

    A.S. - Nonlinear Control

    For nonlinear controllers, also con-sider the eight corners of the unitcube.

    zi = 3 + |ui| fori= 1, 2, 3, 5, 6, 7

    z4 = max{|u4 3|, |u4 6|}

    z8 = max{|u8+ 3|, |u8+ 6|}

    Best nonlinear controller achieves closed-loop norm of 3.00.

    32 M. Rotkowitz, Melbourne Uni

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    32/57

    A.S. - Nonlinear Control

    For nonlinear controllers, also con-sider the eight corners of the unitcube.

    zi = 3 + |ui| fori= 1, 2, 3, 5, 6, 7

    z4 = max{|u4 3|, |u4 6|}

    z8 = max{|u8+ 3|, |u8+ 6|}

    Best nonlinear controller achieves closed-loop norm of 3.00.

    Nonlinear outperforms linear - for a centralized problem!

    33 M. Rotkowitz, Melbourne Uni

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    33/57

    Back to Witsenhausen

    Consider controllers of the form

    u1 =

    ay1 ify10

    a+y1 ify10u2 =

    by2 ify20

    b+y2 ify20

    34 M. Rotkowitz, Melbourne Uni

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    34/57

    Notation

    Q()

    ratio z2

    w2for a given controller and given noise w1, w2

    J()

    worst case ratio supw=0 z2w2 for a given controller, i.e. the costassociated with that controller

    J

    best achievable cost for a given class of controllers

    35 M. Rotkowitz, Melbourne Uni

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    35/57

    Back to Witsenhausen

    Find best piecewise affine controller

    J

    pa = inf a,a+ infb,b+ Jpa(a, a+, b, b+)

    36 M. Rotkowitz, Melbourne Uni

    Pi i Affi C

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    36/57

    Piecewise Affine Cost

    Find best piecewise affine controller

    J

    pa = inf a,a+ infb,b+ Jpa(a, a+, b, b+)

    As with linear, only the direction of the noise matters

    Jpa(a, a+, b, b+) = sup

    Qpa(a, a+, b, b+, )

    37 M. Rotkowitz, Melbourne Uni

    R i f Diff G i

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    37/57

    1

    x0 x1

    2

    x2

    y2 u2u1y1

    w1

    v = w2

    +

    +

    +

    +

    +

    Regions of Different Gains

    38 M. Rotkowitz, Melbourne Uni

    R i f Diff G i

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    38/57

    1

    x0 x1

    2

    x2

    y2 u2u1y1

    w1

    v = w2

    +

    +

    +

    +

    +

    Regions of Different Gains

    Jpa(a, a+, b, b+) = sup Qpa(a, a+, b, b+, )

    39 M. Rotkowitz, Melbourne Uni

    R i f Diff t G i

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    39/57

    Regions of Different Gains

    At, we have y2= 0. Thus

    sup

    Qpa(a, b, b+, ) Qpa(a, b, b+, ) for allb, b+

    and so

    Jpa(a) = inf b,b+

    sup

    Qpa(a, b, b+, ) Qpa(a, b, b+, )

    40 M. Rotkowitz, Melbourne Uni

    S t S d C t ll

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    40/57

    Set Second Controller

    Wed like

    Qpa(a, b, b+, )

    =

    = 0

    and also2Q2pa(a, b, b+, )

    2

    =

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    41/57

    Gain vs. Noise for Optimal Controllers

    1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 10.5

    0

    0.5

    1

    a= a*, a

    += a*, b

    = b*(a*), b

    += b*(a*), k =0.1, =10

    /

    y1,y2,

    Qpa

    Qpa

    (a,a

    +,b

    ,b

    +,)

    y1/ (2r)

    y2/ (2r)

    y1= 0

    y2= 0

    42 M. Rotkowitz, Melbourne Uni

    Cost vs First Controller

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    42/57

    Cost vs. First Controller

    5 0 50.5

    1

    1.5

    2

    2.5

    3

    3.5

    a

    Qlb

    Qlb(a), with k = 0.1, = 10

    Qlb

    a*= a

    1

    a2

    (k+1)

    1/2

    a = 2+ 2 2 /2

    where=k22 +k2 + 1 and= 2k.

    43 M. Rotkowitz, Melbourne Uni

    General Nonlinear Controller

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    43/57

    General Nonlinear Controller

    Given any1, 2: R R, let

    a=1(1)

    and then let

    0 = (a)

    and further choose r0 such that

    y1(r0, 0) = 1

    and

    y2(r0, 0) = 0

    44 M. Rotkowitz, Melbourne Uni

    General Nonlinear Controller

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    44/57

    General Nonlinear Controller

    45 M. Rotkowitz, Melbourne Uni

    General Nonlinear Controller

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    45/57

    General Nonlinear Controller

    Obvious that

    J Jl

    46 M. Rotkowitz, Melbourne Uni

    General Nonlinear Controller

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    46/57

    General Nonlinear Controller

    Obvious that

    J Jl

    Now

    Q(1, 2, r0, 0)

    47 M. Rotkowitz, Melbourne Uni

    General Nonlinear Controller

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    47/57

    General Nonlinear Controller

    Obvious that

    J Jl

    Now

    Q(1, 2, r0, 0)

    = Ql(a, b(a), (a))

    48 M. Rotkowitz, Melbourne Uni

    General Nonlinear Controller

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    48/57

    General Nonlinear Controller

    Obvious that

    J Jl

    Now

    Q(1, 2, r0, 0)

    = Ql(a, b(a), (a))

    = Jl(a, b(a))

    49 M. Rotkowitz, Melbourne Uni

    General Nonlinear Controller

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    49/57

    General Nonlinear Controller

    Obvious that

    J Jl

    Now

    Q(1, 2, r0, 0)

    = Ql(a, b(a), (a))

    = Jl(a, b(a))

    which yields

    J(1, 2) Jl(a, b(a))

    50 M. Rotkowitz, Melbourne Uni

    General Nonlinear Controller

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    50/57

    General Nonlinear Controller

    Obvious that

    J Jl

    and weve now shown that

    J Jl

    thusJ = Jl

    51 M. Rotkowitz, Melbourne Uni

    Redundancy

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    51/57

    Redundancy

    Argued that piecewise affine dominates general nonlinear.

    Showed that best piecewise affine controller is linear.

    Showed that linear dominates general nonlinear.

    52 M. Rotkowitz, Melbourne Uni

    Redundancy

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    52/57

    Redundancy

    Argued that piecewise affine dominates general nonlinear.

    Showed that best piecewise affine controller is linear.

    Showed that linear dominates general nonlinear.

    Why?

    Generalization

    53 M. Rotkowitz, Melbourne Uni

    Redundancy

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    53/57

    Redundancy

    Argued that piecewise affine dominates general nonlinear.

    Showed that best piecewise affine controller is linear.

    Showed that linear dominates general nonlinear.

    Why?

    Generalization

    Surprising

    54 M. Rotkowitz, Melbourne Uni

    Related Work

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    54/57

    t

    Separation / Linear Optimality over Lossy Networks

    X. Liu, A. Goldsmith

    L. Schenato, B. Sinopoli, M. Franceschetti, K. Poola, S. Sastry

    (Decentralized) Robust Control

    Ian Petersen, Valery Ugrinovskii

    1

    x0 x1

    2

    x2

    y2 u2u1y1

    w1

    v = w2

    +

    +

    +

    +

    +

    Linear Suboptimality with Communication Link

    Nuno Martins

    55 M. Rotkowitz, Melbourne Uni

    LTI vs LTV

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    55/57

    unstable fixed modes stabilzable with LTV controllers

    Kobayashi, Hanafusa, and Yoshikawa

    Anderson and Moore Wang

    Surprise coupling and nonlinearity - Allerton 2009

    56 M. Rotkowitz, Melbourne Uni

    Conclusion(s)

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    56/57

    ( )

    Very simple question gives rise to extremely rich structure.

    Linear controllers are uniformly optimal for the Witsenhausen Coun-

    terexample. Many new surprises await!

    System Norm Nonlinear Linear Optimal Linear Optimal

    Generalization for Centralized? for Decentralized?1

    No ??

    2 LQG Yes Sometimes (PN/QI)

    22 Yes ??

    57 M. Rotkowitz, Melbourne Uni

    The Witsenhausen Counterexample:

  • 7/24/2019 Rotkowitz-when is a Linear Controller Optimal

    57/57

    p

    40 Years LaterSpecial Invited Session,IEEE Conference on Decision and Control,

    Cancun, Mexico, 9 December 2008

    Authors include:

    Yu-Chi Larry Ho

    Tamer Basar

    Demosthenis Teneketzis

    Anant Sahai

    Nuno Martins

    Michael Rotkowitz