heuristic search (informed search) - … search •the idea is to develop a domain specific...

Post on 09-Mar-2018

220 Views

Category:

Documents

4 Downloads

Preview:

Click to see full reader

TRANSCRIPT

1McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

HeuristicSearch(InformedSearch)

2McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

HeuristicSearch

• Inuninformedsearch,wedon’ttrytoevaluatewhichofthenodesonthefrontier/OPENaremostpromising.Wenever“look-ahead”tothegoal.

E.g.,inuniformcostsearchwealwaysexpandthecheapestpath.Wedon’tconsiderthecostofgettingtothegoalfromtheendofthecurrentpath.

• Oftenwehavesomeotherknowledgeaboutthemeritofnodes,e.g.,goingthewrongdirectioninRomania.

3McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

HeuristicSearchMerit ofafrontier/OPENnode:differentnotionsofmerit.• Ifweareconcernedaboutthecostofthesolution,wemightwantanotionofmeritofhowcostlyitistogettothegoalfromthatsearchnode.

• Ifweareconcernedaboutminimizingcomputation insearchwemightwanttoconsiderhoweasyitistofindthegoalfromthatsearchnode.

• Wewillfocusonthe“costofsolution”notionofmerit.

4McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

HeuristicSearch

• Theideaistodevelopadomainspecificheuristicfunctionh(n).

• h(n)guesses thecostofgettingtothegoalfromnoden(thecostofcompletingthepaththatiscapturedbythestateofnoden).

• Therearedifferentwaysofguessingthiscostindifferentdomains.I.e.,heuristicsaredomainspecific.

“Asthecrowflies”– Straightlineheuristic

Onthemap,thenumbersbetweencitiesrepresentthedrivingdistancebetweencitiesonpotentiallywigglyroads,eventhoughtheyaredrawnasstraightlines.Contrastthistotheline-of-sight/``asthecrowflies”distancewhichignoreswigglesintheroad,cliffs,bridges,andassumesyoucanjustdriveinastraightlinefromonecitytoanother.

5McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

6McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Planning a path from Arad to Bucharest, we can utilize the straight line distance from each city to our goal as a heuristic/guess of the actual distance. This lets us plan our trip by picking cities at each time point that minimize the distance to our goal.

Example:StraightLineDistance

7McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

HeuristicSearch

• Ifh(n1)<h(n2) thismeansthatweguessthatitischeapertogettothegoalfromn1 thanfromn2.

• Werequirethat• h(n)=0 foreverynodenwhosestatesatisfiesthegoal.• Zerocostofgettingtoagoalnodefromn.

8McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Usingonlyh(n):Greedybest-firstsearch(GreedyBFS)

• Weuseh(n)torankthenodesonthefrontier/OPEN.• Alwaysexpandnodewithlowesth-value.

• Wearegreedilytryingtoachievealowcostsolution.

• However,thismethodignoresthecostofgettington,soitcanbeleadastrayexploringnodesthatcostalottogettobutseemtobeclosetothegoal:

S

n1

n2

n3

Goal

→ stepcost=10

→ stepcost=100h(n3)=50h(n1)=70

[S][n3,n1][Goal, n1]

9McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Usingonlyh(n):Greedybest-firstsearch(GreedyBFS).

S

n1

n2

n3

Goal

→ step cost = 10

→ step cost = 100h(n3) = 50h(n1) = 70

(Greedy BFS is• Incomplete• not optimal)

100

100

10

10

10

• Weuseh(n)torankthenodesonthefrontier.• Alwaysexpandnodewithlowesth-value.

• Wearegreedilytryingtoachievealowcostsolution.

• However,thismethodignoresthecostofgettington,soitcanbeleadastrayexploringnodesthatcostalottogettobutseemtobeclosetothegoal:

10McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Greedybest-firstsearchexample

Whenyou’reatSibiuandcontemplatingwhethertogotoFagarasorRV,theheuristicvalueofthesuccessornodes,i.e.,thehvalueguessofthecostis:h(Fagaras)=178andh(RV)=193),soFagaraslookslikethebetterchoice,but…

ActualCost(Arad-Sibiu-RV-Pitesli-Bucharest):140+80+97+101=140+278=418ActualCost(Arad-Sibiu-Fagaras-Bucharest): 140+99+211 =140+310 =450

11McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

A*search

• Takeintoaccountthecostofgettingtothenodeaswellasourestimateofthecostofgettingtothegoalfromn.

• Defineanevaluationfunctionf(n)f(n)=g(n)+h(n)• g(n)isthecostofthepathtonoden• h(n)istheheuristicestimateofthecostofgettingtoagoalnodefromn.

• Alwaysexpandthenodewithlowestf-valueonthefrontier.

• Thef-valueisanestimateofthecostofgettingtothegoalviathisnode(path).

12McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

A* examplef(n) =g(n)+h(n),

=actualcostton+heuristicestimateofcostfromntothegoal

13McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

A* examplef(n) =g(n)+h(n),

=actualcostton+heuristicestimateofcostfromntothegoal

14McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

A* examplef(n) =g(n)+h(n),

=actualcostton+heuristicestimateofcostfromntothegoal

15McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

A* examplef(n) =g(n)+h(n),

=actualcostton+heuristicestimateofcostfromntothegoal

16McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

A* examplef(n) =g(n)+h(n),

=actualcostton+heuristicestimateofcostfromntothegoal

17McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

A* examplef(n) =g(n)+h(n),

=actualcostton+heuristicestimateofcostfromntothegoal

18McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

A*search

• Takeintoaccountthecostofgettingtothenodeaswellasourestimateofthecostofgettingtothegoalfromn.

• Defineanevaluationfunctionf(n)f(n)=g(n)+h(n)• g(n)isthecostofthepathtonoden• h(n)istheheuristicestimateofthecostofgettingtoagoalnodefromn.

• Alwaysexpandthenodewithlowestf-valueonthefrontier.

• Thef-valueisanestimateofthecostofgettingtothegoalviathisnode(path).

19McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Conditionsonh(n)

• Wewanttoanalyzethebehavioroftheresultantsearch.• Completeness,timeandspace,optimality?

• Toobtainsuchresultswemustputsomefurtherconditionsontheheuristicfunctionh(n)andthesearchspace.

20McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Conditionsonh(n):Admissible

• Wealwaysassumethatc(n1→n2)≥ε >0.Thecostofanytransitionisgreaterthanzeroandcan’tbearbitrarilysmall.

• Leth*(n)bethecostofan optimalpath fromntoagoalnode(¥ ifthereisnopath).Thenanadmissible heuristicsatisfiesthecondition

h(n)≤h*(n)admissibleheuristichalwaysunderestimatesthetruecosttoreach

thegoal.i.e.,itisoptimisticJ

• Hence• h(g)=0,foranygoalnote,g• h*(n)=¥ ifthereisnotpathfromntoagoalnode

21McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Consistency(akamonotonicity)

• Isastrongerconditionthanh(n)≤h*(n).

• Amonotone/consistent heuristicsatisfiesthetriangleinequality(forallnodesn1,n2):

h(n1)≤c(n1→ n2)+h(n2)

• Notethattheremightbemorethanonetransition(action)betweenn1andn2,theinequalitymustholdforallofthem.

• Notethatmonotonicityimpliesadmissibility.• (foralln1,n2)h(n1)≤c(n1→ n2)+h(n2)è (foralln)h(n)≤h*(n)

22McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Intuitionbehindadmissibility

h(n)≤h*(n)meansthatthesearchwon’tmissanypromisingpaths.• Ifitreallyischeaptogettoagoalvian(i.e.,bothg(n)andh*(n)arelow),thenf(n)=g(n)+h(n)willalsobelow,andthesearchwon’tignoreninfavourofmoreexpensiveoptions.

• Thiscanbeformalizedtoshowthatadmissibilityimpliesoptimality.

23McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Intuitionbehindmonotonicity

h(n1)≤c(n1→n2)+h(n2)

• Thissayssomethingsimilar,butinadditiononewon’tbe“locally”mislead.Seenextexample.

24McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Consistencyè Admissible• Assumeconsistency:h(n1)≤c(n1→n2)+h(n2)Proveadmissible:h(n)≤h*(n)

Proof:If nopathexistsfromntoagoalthenh*(n)=¥ andh(n)≤h*(n)Else letnà n1à …à n*beanOPTIMALpathfromntoagoal.Notethecostofthispathish*(n),andeachsubpath (nià …à n*)hascostequaltoh*(ni).

Proveh(n)≤h*(n)byinductiononthelengthofthisoptimalpath.

BaseCase:n=n* [optimalpathlength=0]Byourconditionsonh,h(n)=0≤h(n*)=0InductionHypothesis:h(n1)≤h*(n1)h(n)≤c(n → n1)+h(n1)[consistency]

≤c(n→n1)+h*(n1) [defn h*]=h*(n)

25McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Example:admissiblebutnonmonotonic

Thefollowingh isnotconsistent(i.e.,notmonotone) sinceh(n2)>c(n2→n4)+h(n4).Butitisadmissible.

S

n1

n3

n2

Goal

→ stepcost=200→ stepcost=100

{S}→{n1[200+50=250],n2[200+100=300]}→{n2[100+200=300], n3[400+50=450]}→{n4[200+50=250],n3[400+50=450]}→{goal[300+0=300],n3[400+50=450]}

Wedofind theoptimalpathastheheuristicisstilladmissible.But wearemisleadintoignoringn2untilafterweexpandn1.

n4

h(n2)=200

h(n4)=50

h(n1)=50

h(n3)=50g(n)+h(n)=f(n)

26McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Example:admissiblebutnonmonotonic

S

n1

n3

n2

Goal

→ stepcost=200→ stepcost=100

{S}→{n1[200+50=250],n2[100+200=300]}→{n2[100+200=300], n3[400+50=450]}→{n4[200+50=250],n3[400+50=450]}→{goal[300+0=300],n3[400+50=450]}

Wedofind theoptimalpathastheheuristicisstilladmissible.But wearemisleadintoignoringn2untilafterweexpandn1.

n4

h(n2)=200

h(n4)=50

h(n1)=50

h(n3)=50g(n)+h(n)=f(n)

100

100

100

200

200

200

Thefollowingh isnotconsistent(i.e.,notmonotone) sinceh(n2)>c(n2→n4)+h(n4).Butitisadmissible.

“Asthecrowflies”– Straightlineheuristic

27McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

• Mostadmissibleheuristicsarealsomonotone.(Indeedit’shardtofindanadmissibleheuristicthatisnotmonotone!)

28McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

29McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Consequencesofmonotonicity

1. Thef-valuesofnodesalongapathmustbenon-decreasing.

Let<Start→n1→n2…→nk>beapath.Weclaimthat

f(ni)≤f(ni+1)

Proof:f(ni)=c(Start→…→ni)+h(ni)

≤c(Start→…→ni)+c(ni→ni+1) +h(ni+1)[monotonicity]=c(Start→…→ni→ni+1)+h(ni+1)=g(ni+1)+h(ni+1)=f(ni+1).

30McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Consequencesofmonotonicity

Proof(2cases):• Ifn2wasonthefrontier/OPENwhenn1wasexpanded,

thenf(n1)≤f(n2)otherwisewewouldhaveexpandedn2.• Ifn2wasaddedtothefrontier/OPENaftern1’sexpansion,thenletn beanancestorofn2thatwaspresentwhenn1wasbeingexpanded(thiscouldben1itself).Wehavef(n1)≤f(n)sinceA*chosen1whilen waspresentinthefrontier/OPEN.Also,sincen isalongthepathton2,byproperty(1)wehavef(n)≤f(n2).So,wehavef(n1)≤f(n2).

----------------------------1) Thef-valuesofnodesalongapathmustbenon-decreasing.2) Ifn2isexpandedaftern1,thenf(n1)≤f(n2)

2. Ifn2isexpandedaftern1,thenf(n1)≤f(n2)(thef-valueincreasesmonotonically)

31McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

ConsequencesofmonotonicityCorollary: thesequenceoff-valuesofthenodesexpandedbyA*isnon-decreasing.I.e,Ifn2isexpandedafter (notnecessarilyimmediatelyafter)n1,thenf(n1)≤f(n2)

(thef-valueofexpandednodesismonotonic non-decreasing)Proof:• Ifn2wasonfrontier/OPENwhenn1wasexpanded,

thenf(n1)≤f(n2)otherwisewewouldhaveexpandedn2.• Ifn2wasaddedtofrontier/OPENaftern1'sexpansion,then

letn beanancestorofn2thatwaspresentwhenn1wasbeingexpanded(thiscouldben1itself).Wehavef(n1)≤f(n)sinceA*chosen1whilen waspresentonfrontier/OPEN.Also,sincen isalongthepathton2,byproperty(1)wehavef(n)≤f(n2).So,wehavef(n1)≤f(n2).

32McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Consequencesofmonotonicity

• Proof: Assumebycontradictionthatthereexistsapath<Start,n0,n1,ni-1,ni,ni+1,…,nk>withf(nk)<f(n)andni isitslastexpandednode.• ni+1mustbeonthefrontier/OPENwhilenisexpanded,so

a)by(1) f(ni+1)≤f(nk)sincetheyliealongthesamepath.b)sincef(nk)<f(n)(given)sowehavef(ni+1)<f(n)(froma)c)by(2) f(n)≤f(ni+1)becausenisexpandedbeforeni+1.

• Contradictionfromb&c!-----------------------------------------------------------

1) Thef-valuesofnodesalongapathmustbenon-decreasing.2) Ifn2isexpandedaftern1,thenf(n1)≤f(n2)3) Whennisexpandedeverypathwithlowerf-valuehasalreadybeenexpanded.

3. Whennisexpandedeverypathwithlowerf-valuehasalreadybeenexpanded.

34McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Consequencesofmonotonicity4. Withamonotoneheuristic,thefirsttimeA*expandsa

state,ithasfoundtheminimumcostpathtothatstate.Proof:• LetPATH1 =<Start,n0,n1,…,nk,n> bethefirst pathtonfound.

Wehavef(path1)=c(PATH1)+h(n).• LetPATH2=<Start,m0,m1,…,mj,n> beanotherpathtonfound

later.wehavef(path2)=c(PATH2)+h(n).• Byproperty (3)anditscorollary,f(path1)≤f(path2)• hence:c(PATH1)≤c(PATH2)

1) Thef-valuesofnodesalongapathmustbenon-decreasing.2) Ifn2isexpandedaftern1,thenf(n1)≤f(n2)3) Whennisexpandedeverypathwithlowerf-valuehasalreadybeenexpanded.Corollary:thesequenceoff-valuesofthenodesexpandedbyA*isnon-decreasing.I.e,Ifn2isexpandedafter (notnecessarilyimmediatelyafter)n1,thenf(n1)≤f(n2)

(thef-valueofexpandednodesismonotonic non-decreasing)

35McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

ConsequencesofmonotonicityComplete.

• Yes,consideraleastcostpathtoagoalnode• SolutionPath =<Start→n1→…→G>withcostc(SolutionPath)• Sinceeachactionhasacost≥ε >0,thereareonlyafinitenumberofpaths

thathavecost ≤c(SolutionPath).• Allofthesepathsmustbeexploredbeforeanypathofcost>

c(SolutionPath).• SoeventuallySolutionPath,orsomeequalcostpathtoagoalmustbe

expanded.TimeandSpacecomplexity.

• Whenh(n)=0,foralln,hismonotone.(avery*un*informativeheuristic!!!)• A*becomesuniform-costsearch!

• Itcanbeshownthatwhenh(n)>0forsomen,thenumberofnodesexpandedcanbenolargerthanuniform-cost.

• Hencethesameboundsasuniform-costapply.(Theseareworstcasebounds).Stillexponentialunlesswehaveaverygoodh!

• Inrealworldproblems,werunoutoftimeandmemory!IDA*cansometimesbeusedtoaddressmemoryissues,butIDA*isn’tverygoodwhenmany cyclesarepresent.

36McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

ConsequencesofmonotonicityOptimality§ Yes,by(4)thefirstpathtoagoalnodemustbeoptimal.

CycleChecking§ Wecanuseasimpleimplementationofcyclechecking

(multiplepathchecking)---justrejectallsearchnodesvisitingastatealreadyvisitedbyapreviouslyexpandednode.Byproperty(4)weneedkeeponlythefirstpathtoanode,rejectingallsubsequentpaths.

4. Withamonotoneheuristic,thefirsttimeA*expandsastate,ithasfoundtheminimumcostpathtothatstate.

37McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Searchgeneratedbymonotonicity

Insideeachcounter,thefvaluesarelessthanorequaltocountervalue!

• Foruniformcostsearch,bandsare“circular”.• Withmoreaccurateheuristics,bandsstretchoutmoretowardthegoal.

38McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

AdmissibilitywithoutmonotonicityWhen“h”isadmissiblebutnotmonotonic.

• TimeandSpacecomplexityremainthesame.Completenessholds.• Optimalitystillholds(withoutcyclechecking),butneedadifferent

argument:don’tknowthatpathsareexploredinorderofcost.

Proof(bycontradiction)ofoptimality(withoutcyclechecking):• Assumethegoalpath<S,…,G>foundbyA*hascostbiggerthanthe

optimalcost:i.e.C*(G)<f(G).• Theremustexistsanoden intheoptimalpaththatisstillinthe

frontier.• Wehave: f(n)=g(n)+h(n)≤g(n)+h*(n)=C*(G)<f(G)

• Therefore,f(n)musthavebeenselectedbefore GbyA*.contradiction!

40McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

Admissibilitywithoutmonotonicity

WhataboutCycleChecking?• Nolongerguaranteedwehavefoundanoptimalpathtoanodethefirst

time wevisitit.

• So,cyclecheckingmightnotpreserveoptimality.• Tofixthis:forpreviouslyvisitednodes,mustremembercostof

previouspath.Ifnewpathischeapermustexploreagain.

• contoursofmonotonicheuristicsdon’thold.

41McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

SpaceProblemswithA*

• A*hasthesamepotentialspaceproblemsasBFSorUCS

• IDA*- IterativeDeepeningA*issimilartoIterativeDeepeningSearchandsimilarlyaddressesspaceissues.

42McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

IDA*- IterativeDeepeningA*Objective:reducememoryrequirementsforA*• Likeiterativedeepening,butnowthe“cutoff”isthef-value(g+h)rather

thanthedepth• Ateachiteration,thecutoffvalueisthesmallestf-valueofanynodethat

exceededthecutoffonthepreviousiteration• Avoidsoverheadassociatedwithkeepingasortedqueueofnodes• Twonewparameters:

• curBound (anynodewithabiggerf-valueisdiscarded)• smallestNotExplored (thesmallestf-valuefordiscardednodesina

round)whenfrontier/OPENbecomesempty,thesearchstartsanewroundwiththisbound

• Easiertoexpandallnodeswithf-valueEQUALtothef-limit.Thiswaywecancompute“smallestNotExplored” moreeasily.

.

43McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

ConstructingHeuristics

44McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

BuildingHeuristics:RelaxedProblem

• Oneusefultechniqueistoconsideraneasierproblem,andleth(n)bethecostofreachingthegoalintheeasierproblem.

8-Puzzle

• CanmoveatilefromsquareAtoBif• Aisadjacent(left,right,above,below)toB• and Bisblank

45McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

BuildingHeuristics:RelaxedProblem

8-Puzzlemoves(continued)• CanmoveatilefromsquareAtoBif

• Aisadjacent(left,right,above,below)toB• and Bisblank

• Canrelaxsomeoftheseconditions1. canmovefromAtoBifAisadjacenttoB(ignorewhetherornot

positionisblank)2. canmovefromAtoBifBisblank(ignoreadjacency)3. canmovefromAtoB(ignorebothconditions).

46McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

BuildingHeuristics:RelaxedProblem• #3“canmovefromAtoB(ignorebothconditions)”.

leadstothemisplacedtiles heuristic.• Tosolvethepuzzle,weneedtomoveeachtileintoitsfinalposition.• Numberofmoves=numberofmisplacedtiles.• Clearlyh(n)=numberofmisplacedtiles≤theh*(n)thecostofanoptimal

sequenceofmovesfromn.

• #1“canmovefromAtoBifAisadjacenttoB(ignorewhetherornotpositionisblank)”leadstothemanhattan distance heuristic.• Tosolvethepuzzleweneedtoslideeachtileintoitsfinalposition.• Wecanmoveverticallyorhorizontally.• Numberofmoves=sumoverallofthetilesofthenumberofverticaland

horizontalslidesweneedtomovethattileintoplace.• Againh(n)=sumofthemanhattan distances≤h*(n)

• inarealsolutionweneedtomoveeachtileatleastthatfarandwecanonlymoveonetileatatime.

47McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

BuildingHeuristics:RelaxedProblem

Theoptimal costtonodesintherelaxedproblemisanadmissible heuristic fortheoriginalproblem!

ProofIdea:theoptimalsolutionintheoriginalproblemisasolutionforrelaxedproblem,thereforeitmustbeatleastasexpensiveastheoptimalsolutionintherelaxedproblem.

Soadmissibleheuristicscansometimesbeconstructedbyfindingarelaxationwhoseoptimalsolutioncanbeeasilycomputed.

49McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

BuildingHeuristics:RelaxedProblem

Depth IDS A*(Misplaced)h1 A*(Manhattan)h210 47,127 93 3914 3,473,941 539 11324 --- 39,135 1,641

Leth1=Misplaced,h2=Manhattan• Doesh2always expandfewernodesthanh1?

• Yes!Notethath2dominatesh1,i.e.foralln:h1(n)≤h2(n).• Therefore,amongseveraladmissibleheuristictheonewithhighest

valueexpandsthefewestnodes.Isitthefastest?

Comparison ofIDSandA*(averagetotalnodesexpanded):

50McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

BuildingHeuristics:Patterndatabases.

•By searching backwards from these goal states, we can compute the distance of any configuration of these tiles to their goal locations. We are ignoring the identity of the other tiles.

•For any state n, the number of moves required to get these tiles into place form a lower bound on the cost of getting to the goal from n.

• Admissibleheuristicscanalsobederivedfromsolutiontosubproblems:Eachstateismappedintoapartialspecification,e.g.in15-puzzleonlypositionofspecifictilesmatters.

• Herearegoalsfortwosub-problems(calledCornerandFringe)of15-puzzle.

• NotethelocationofBLANK!

51McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

BuildingHeuristics:Patterndatabases.

• Theseconfigurationsarestoredinadatabase,alongwiththenumberofmovesrequiredtomovethetilesintoplace.

• Themaximum numberofmovestakenoverallofthedatabases canbeusedasaheuristic.

• Onthe15-puzzle• Thefringedatabaseyieldsabouta345folddecreaseinthesearchtreesize.

• Thecornerdatabaseyieldsabout437folddecrease.

• Sometimesdisjointpatterns canbefound,thenthenumberofmovescanbeadded ratherthantakingthemax(ifweonlycountmovesofthetargettiles).

52McIlraith&Allin,CSC384,UniversityofToronto,Winter2018

LocalSearch

• Sofar,wekeepthepathstothegoal.• Forsomeproblems(like8-queens)wedon’tcareaboutthepath,weonlycareaboutthesolution.ManyrealproblemlikeScheduling,ICdesign,andnetworkoptimizationsareofthisform.

• Localsearch algorithmsoperateusingasinglecurrentstateandgenerallymovetoneighborsofthatstate.

• Thereisanobjectivefunction thattellsthevalueofeachstate.Thegoalhasthehighestvalue(globalmaximum).

• AlgorithmslikeHillClimbing trytomovetoaneighbourwiththehighestvalue.

• Dangerofbeingstuckinalocalmaximum.Sosomerandomnessisaddedto“shake”outoflocalmaxima.

• SimulatedAnnealing:Insteadofthebestmove,takearandommoveandifitimprovesthesituationthenalwaysaccept,otherwiseacceptwithaprobability<1.

• [IfinterestedreadthesetwoalgorithmsfromtheR&Nbook].

top related