Машинное обучение 1, осень 2015: Генетические алгоритмы

Embed Size (px)

Text of Машинное обучение 1, осень 2015: Генетические алгоритмы

  • . , .

    , 2015

    . , . , -, 2015 . 1 22

  • 1 2 3 Differential evolution4 No free lunch theorem

    . , . , -, 2015 . 2 22

  • F0 = arg maxF

    p(F |X )

    + ;

    F; .

    . , . , -, 2015 . 3 22

  • dream team; ; .

    . , . , -, 2015 . 4 22

  • . "" , "": .

    . , . , -, 2015 . 5 22

  • 1 2 ""3

    1 2

    3 ""4 ""

    . , . , -, 2015 . 6 22

  • - / - - - "" "" - "" -

    . , . , -, 2015 . 7 22

  • 1 2 ""3

    1 2

    3 ""4 ""

    . , . , -, 2015 . 8 22

  • . :

    [0, 1]n {0, 1}k

    . , . , -, 2015 . 9 22

  • - ! :

    , "" shuffle"" "" ( ) :)

    . , . , -, 2015 . 10 22

  • , . !

    "" :

    n-point crossover;cutnsplice; .

    . , . , -, 2015 . 11 22

  • rossover

    . , . , -, 2015 . 12 22

  • Cut and splice

    . , . , -, 2015 . 13 22

  • (genetic drift) . MCMC/ .

    . , . , -, 2015 . 14 22

  • vs

    - "" , - "". :

    , . , , . , . .

    . , . , -, 2015 . 15 22

  • % "" "penalty"

    . , . , -, 2015 . 16 22

  • pros cons , .:

    ( :));

    ;

    ;

    , .

    :

    ;

    ( );

    .

    . , . , -, 2015 . 17 22

  • . , . , -, 2015 . 18 22

  • Differential Evolution

    arg maxRn

    F ()

    1 2 ,

    x P :1 a, b, c P ;2 k U(1..n);3 y = (yi) i

    1 r U((0, 1))

    2 yi =

    {ai + F (bi ci ), i = k |r < Cyi = xi

    4 , .3

    .. , . , -, 2015 . 19 22

  • 2 . ? :: : , :

    Theorem (No free lunch theorem)

    . , . , -, 2015 . 20 22

  • NFL:

    dm = {(d xm(1), d ym(1)), . . . , (d xm(m), d ym(m))}f : X YF = YX

    p(f ) = 1F

    Theorem (David Wolpert and William G.Macready (1997)) a1 a2:

    f p(dym|f ,m, a1) =

    f p(d

    ym|f ,m, a2)

    f p(dym|f0,M ,m, a1) =

    f p(d

    ym|f0,M ,m, a2)

    . , . , -, 2015 . 21 22

  • NFL:

    :) ,

    . , . , -, 2015 . 22 22

    No free lunch theorem