SP14 CS188 Lecture 3 -- Informed Search

download SP14 CS188 Lecture 3 -- Informed Search

of 69

Transcript of SP14 CS188 Lecture 3 -- Informed Search

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    1/69

    CS 188: Artifcial Intelligence

    Inormed Search

    Instructors: Dan Klein and Pieter Abbeel

    Universit o Caliornia! "er#ele$%hese slides &ere created b Dan Klein and Pieter Abbeel or CS188 Intro to AI at UC "er#ele' All CS188 m

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    2/69

    %oda

    Inormed Search +euristics

    ,reed Search

    A- Search

    ,ra(h Search

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    3/69

    .eca(: Search

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    4/69

    .eca(: Search

    Search (roblem: States /confgurations o the &orld0 Actions and costs Successor unction /&orld dnamics0 Start state and goal test

    Search tree: odes: re(resent (lans or reaching states Plans have costs /sum o action costs0

    Search algorithm: Sstematicall builds a search tree Chooses an ordering o the ringe /une2(lored

    nodes0

    3(timal: fnds least4cost (lans

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    5/69

    52am(le: Panca#e Problem

    Cost: umber o (anca#es 6i((ed

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    6/69

    52am(le: Panca#e Problem

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    7/69

    52am(le: Panca#e Problem

    7

    9

    7

    7

    9

    State s(ace gra(h &ith costs as &eights

    7

    9

    7

    9

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    8/69

    ,eneral %ree Search

    Action: 6i( to(t&o

    Cost:

    Action: 6i( allour

    Cost: 9

    Path to reachgoal:

    li( our! 6i(three

     %otal cost: ;

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    9/69

    %he 3ne

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    10/69

    Uninormed Search

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    11/69

    Uniorm Cost Search

    Strateg: e2(and lo&est (ath cost

     %he good: UCS is com(lete ando(timal?

     %he bad: 52(lores o(tions in ever @direction o inormation about goal location

    Start

    $Demo: contours UC$Demo: contours UC

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    12/69

    ideo o Demo Contours UCS 5

    ideo o Demo Contours UCS Pa

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    13/69

    ideo o Demo Contours UCS PaSmall Eae

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    14/69

    Inormed Search

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    15/69

    Search +euristics

    A heuristic is: A unction that estimates ho& close a state is

    to a goal Designed or a (articular search (roblem

    52am(les: Eanhattan distance! 5uclideandistance or (athing

    10

    5

    11.2

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    16/69

    52am(le: +euristic unction

    h(x)

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    17/69

    52am(le: +euristic unction

    +euristic: the number o the largest (anca#e that is(lace

    43

    0

    2

    3

    3

    3

    4

    4

    3

    4

    4

    4

    h(x)

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    18/69

    ,reed Search

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    19/69

    52am(le: +euristic unction

    h(x)

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    20/69

    ,reed Search

    52(and the node that seems closestF

    Ghat can go &rongH

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    21/69

    ,reed Search

    Strateg: e2(and a node that ou

    thin# is closest to a goal state +euristic: estimate o distance to

    nearest goal or each state

    A common case:

    "est4frst ta#es ou straight to the/&rong0 goal

    Gorst4case: li#e a badl4guided DS

    $Demo: contours greed$Demo: contours greed

    id d /

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    22/69

    ideo o Demo Contours ,reed /

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    23/69

    A- S h

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    24/69

    A- Search

    A- S h

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    25/69

    A- Search

    UCS ,reed

    A-

    C bi i UCS d , d

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    26/69

    Combining UCS and ,reed

    Uniorm4cost orders b (ath cost! or backward cg/n0

    ,reed orders b goal (ro2imit! or forward costh/n0

    A- Search orders b the sum: /n0 /n0 J h/n0

    S a d

    b

    G

    h=5

    h=6

    h=2

    1

    8

    1

    1

    2

    h=6 h=0

    c

    h=7

    3

    e h=1

    1a

    b

    c

    d

    G

    g =1

    h=5

    g =2

    h=6g =3

    h=7

    Gh h ld A- t i t H

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    27/69

    Ghen should A- terminateH

    Should &e sto( &hen &e en=ueue

    o: onl sto( &hen &e de=ueue a g

    S

    B

    A

    G

    2

    3

    2

    2

    h = 1

    h = 2

    h = 0h = 3

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    28/69

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    29/69

    Idea: Admissibilit

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    30/69

    Idea: Admissibilit

    Inadmissible /(essimistic0 heuristicsbrea# o(timalit b tra((ing good

    (lans on the ringe

    Admissible /o(timistic0do&n bad (lans but n

    true cost

    Admissible +euristics

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    31/69

    Admissible +euristics

    A heuristic h is admissible /o(timistic0 i:

    &here is the true cost to a nearest

    52am(les:

    Coming u( &ith admissible heuristics is mo&hatLs involved in using A- in (ractice'

    915

    3(timalit o A- %ree Search

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    32/69

    3(timalit o A- %ree Search

    3(timalit o A- %ree Search

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    33/69

    3(timalit o A- %ree Search

    Assume:

    A is an o(timal goal node

    " is a subo(timal goal node

    h is admissible

    Claim:

    A &ill e2it the ringe beore "

    F

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    34/69

    3(timalit o A- %ree Search: "lo

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    35/69

    3(timalit o A- %ree Search: "lo

    Proo:

    Imagine " is on the ringe

    Some ancestor n o A is onthe ringe! too /mabe A?0

    Claim: n &ill be e2(andedbeore "

    1' /n0 is less or e=ual to

    /A0' /A0 is less than /"0 "

    h

    F

    3(timalit o A- %ree Search: "lo

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    36/69

    3(timalit o A- %ree Search: "lo

    Proo:

    Imagine " is on the ringe

    Some ancestor n o A is onthe ringe! too /mabe A?0

    Claim: n &ill be e2(andedbeore "

    1' /n0 is less or e=ual to

    /A0' /A0 is less than /"0

    7' n e2(ands beore "

    All ancestors o A e2(andbeore "

    A e2(ands beore "

    F

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    37/69

    Pro(erties o A-

    Pro(erties o A-

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    38/69

    Pro(erties o A

    Fb

    Fb

    Uniorm4

    Cost

    A-

    UCS vs A- Contours

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    39/69

    UCS vs A Contours

    Uniorm4cost e2(ands e=uall

    in all @directions

    A- e2(ands mainl to&ard thegoal! but does hedge its betsto ensure o(timalit

    Start

    Start

    $Demo: contours UCS )

    em(t /B7D10*

    ideo o Demo Contours /5m(t0

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    40/69

    ideo o Demo Contours /5m(t0

    ideo o Demo Contours /5m(t

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    41/69

    / (,reed

    ideo o Demo Contours /5m(t

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    42/69

    ideo o Demo Contours /5m(t

    ideo o Demo Contours /PacmanE 0 A-

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    43/69

    /Eae0 O A-

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    44/69

    A- A((lications

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    45/69

    A A((lications

    A- A((lications

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    46/69

    A A((lications

    ideo games

    Pathing ) routing (roblems .esource (lanning (roblems

    .obot motion (lanning

    Banguage analsis

    Eachine translation S(eech recognition

    F

    $Demo: UCS ) A- (acman

    /B7D!B7D;0*

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    47/69

    ideo o Demo 5m(t Gater Shallo&)Dee(Algorithm

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    48/69

    Algorithm

    Creating +euristics

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    49/69

    g

    Creating Admissible +euristics

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    50/69

    g

    Eost o the &or# in solving hard search (roblems is in coming u( &ith admissible heuristics

    3ten! admissible heuristics are solutions to relaxe problems &here ne& actions are available

    Inadmissible heuristics are oten useul too

    15366

    52am(le: 8 Pule

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    51/69

    (

    Ghat are the statesH +o& man statesH Ghat are the actionsH +o& man successors rom the start

    stateH

    Ghat should the costs beH

    Start State ,Actions

    8 Pule I

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    52/69

    +euristic: umber o tilesmis(laced

    Gh is it admissibleH h/start0

     %his is a relaxed!problem heuristic

    8

    Average nodese2(anded &heo(timal (ath h

    F9ste(s

    F8ste(s

    UCS 11 !7MM

    Start State

    St

    8 Pule II

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    53/69

    Ghat i &e had an easier 84(ule&here an tile could slide andirection at an time! ignoringother tilesH

     %otal "anhattan distance

    Gh is it admissibleH

    h/start0 7 J 1 J J F 18

    Average nodee2(anded &ho(timal (ath

    F9ste(s

    F8ste(s

     %IB5S 17 7Q

    Start State

    8 Pule III

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    54/69

    +o& about using the act#al cost  as a heuristicH Gould it be admissibleH

    Gould &e save on nodes e2(andedH

    GhatLs &rong &ith itH

    Gith A-: a trade4oR bet&een =ualit o estimate a(er node As heuristics get closer to the true cost! ou &ill e2(and

    nodes but usuall do more &or# (er node to com(ute thheuristic itsel 

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    55/69

    Semi4Battice o +euristics

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    56/69

    ,ra(h Search

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    57/69

    %ree Search: 52tra Gor#?

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    58/69

    ailure to detect re(eated states can cause e2(onmore &or#'

    Search %reeState ,ra(h

    ,ra(h Search

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    59/69

    In "S! or e2am(le! &e shouldnLt bother e2(anding thnodes /&hH0

    S

    a

    b

    d  p

    a

    c

    e

     p

    h

    $

    $ c ,

    a

    $e

     p

    h

    $

    $ c %

    a

    ,ra(h Search

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    60/69

    Idea: never e2(and a state t&ice

    +o& to im(lement:  %ree search J set o e2(anded states /@closed set0

    52(and the search tree node4b4node! butF

    "eore e2(anding a node! chec# to ma#e sure its sthas never been e2(anded beore

    I not ne&! s#i( it! i ne& add to closed set

    Im(ortant: store the closed set as a set! not a

    Can gra(h search &rec# com(letenessH

    Gh)&h notH

    +o& about o(timalitH

    A- ,ra(h Search ,one GrongH

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    61/69

    S

    A

    "

    C

    G

    1

    1

    1

    7

    h=2

    h=1

    h=4

    h=1

    h=0

    S /MJ0

    A /1J90 " /1

    C /J10

    , /NJM0

    C /7

    , /

    State s(ace gra(h Search tree

    Consistenc o +euristics

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    62/69

    Eain idea: estimated heuristi

    actual costs

    Admissibilit: heuristic cost T ac

    goal

    h/A0 T actual cost rom A to

    Consistenc: heuristic @arc cos

    or each arc

    h/A0 O h/C0 T cost/A to C0

    Conse=uences o consistenc

     %he value along a (ath never d

     h/A0 T cost/A to C0 J h/C0

    7

    A

    C

    G

    h=4 h=1

    1

    h=2

    3(timalit o A- ,ra(h Search

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    63/69

    3(timalit o A- ,ra(h Search

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    64/69

    S#etch: consider &hat A- does

    &ith a consistent heuristic: act 1: In tree search! A- e2(ands

    nodes in increasing total value /4contours0

    act : or ever state s! nodesthat reach s o(timall aree2(anded beore nodes that reachs subo(timall

    .esult: A- gra(h search is o(timal

    F

    3(timalit

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    65/69

     %ree search: A- is o(timal i heuristic is admissible UCS is a s(ecial case /h M0

    ,ra(h search: A- o(timal i heuristic is consistent UCS o(timal /h M is consistent0

    Consistenc im(lies admissibilit

    In general! most natural admissibleheuristics tend to be consistent!es(eciall i rom rela2ed (roblems

    A-: Summar

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    66/69

    A-: Summar

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    67/69

    A- uses both bac#&ard costs and /estimatesor&ard costs

    A- is o(timal &ith admissible ) consistent he

    +euristic design is #e: oten use rela2ed (r

     %ree Search Pseudo4Code

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    68/69

    ,ra(h Search Pseudo4Code

  • 8/15/2019 SP14 CS188 Lecture 3 -- Informed Search

    69/69