Lecture3-2HookeJeeves

Embed Size (px)

Citation preview

  • 8/7/2019 Lecture3-2HookeJeeves

    1/12

    1

    Smooth, Unconstrained Nonlinear

    Optimization without Gradients

    Hooke Jeeves

    6/16/05

  • 8/7/2019 Lecture3-2HookeJeeves

    2/12

    2

    Hooke Jeeves or Pattern Search

    Zero order

    No derivatives

    No line searches Works in discontinuous domain

    No proof of convergence

    Tool when other tools fails

    References: Evolution and Optimum Seeking by Hans-Paul Schwefel

    Mark Johnson code handout

    Characteristics

  • 8/7/2019 Lecture3-2HookeJeeves

    3/12

    3

    Hooke JeevesWith downhill simplex is the simplest algorithm in iSIGHT

    Has essentially no formal diagnostics for outputs.

    The algorithm is an unconstrained optimization algorithm

    which can also be used in constrained situations.

    Expected number of iterations =

    StepSizeReductionFactor ** n < Termination Step Size

    StepSizeReductionFactor between 0 1. Default .5

    The larger the value the slower the convergence.

  • 8/7/2019 Lecture3-2HookeJeeves

    4/12

    4

  • 8/7/2019 Lecture3-2HookeJeeves

    5/12

    5

  • 8/7/2019 Lecture3-2HookeJeeves

    6/12

    6

    Hooke Jeeves AlgorithmTermination step size = e

    Step size reduction = rho

    Step 0: InitializationChoose a starting point, an accuracy bound e > 0,

    and initial step lengths (current value * rho).

    If current value = 0.0 make step length rho

    situation)original(back toesxxreplaceotherwise

    (success)2Steptogo),F(x)xF(if

    direction)negativeinstep(discretee2s-xxreplaceotherwise

    l)first trial(successfu2Steptogo),F(x)xF(if

    direction)positiveinstep(discreteesxxConstruct

    moveyExplorator:1Step

    i

    (k)

    i

    1)-i(k,

    i

    (k)

    i

    1)-i(k,

    i

    (k)

    i

    1)-i(k,

    d!d

    d

    d!d

    d

    !d

  • 8/7/2019 Lecture3-2HookeJeeves

    7/12

    7i

    (k)

    i

    1)-i(k,

    i

    (k)

    i

    1)-i(k,

    i

    (k)

    i

    1)-i(k,

    n)1,-(ki

    n)(k,i

    (k)i

    1)(ki

    n)1,-(kn)(k,1,0)(k

    (k,0)1,0)(k(k,0)n)(k,

    i)(k,

    esxxreplaceother ise

    6steptogo)(x)x(I

    e2s-xxreplaceother ise

    6steptogo)(x)x(I

    esxxonstruct

    ionextrapolata ter nxploratio:5tep

    ar)somovepatterntheocontrolsuccessnoisThere:( bserve

    1isetand1kkincrease

    n1,iallor)x-sign(xssand

    tion)(extrapolax-2xet x

    moveattern:4tep

    9steptogoandxset x),(x)(xI

    .directionsallinailureor totalTest:3tep

    1.steptogoand1iiincreasen,iI

    xet x

    coordinatenexts itch toandetention:2tep

    d!d

    d

    d!d

    d

    !d

    !!

    !!

    !

    !u

    !

    d!

  • 8/7/2019 Lecture3-2HookeJeeves

    8/12

  • 8/7/2019 Lecture3-2HookeJeeves

    9/12

    9

  • 8/7/2019 Lecture3-2HookeJeeves

    10/12

    10

  • 8/7/2019 Lecture3-2HookeJeeves

    11/12

    11

    Lab Rerun the Spring_Start.desc file using

    Hooke_Jeeves with the default step size.

    Does it reach the same optimum? How many function calls did it take?

    Is this more or less efficient then SteepestDescent?

    On the next slide, label the X1 Step Size, X2 StepSize and algorithm step number next to each rowfor first 7 run counters.

  • 8/7/2019 Lecture3-2HookeJeeves

    12/12

    12

    Spring Hooke Initial Steps