Logical a gents
Chapter7
Chapter71
Outline
♦Knowledge-basedagents
♦Wumpusworld
♦Logicingeneral—modelsandentailment
♦Propositional(Boolean)logic
♦Equivalence,validity,satisfiability
♦Inferencerulesandtheoremproving–forwardchaining–backwardchaining–resolution
Chapter72
Kno wledge bases
Inference engine
Knowledge basedomain−specific content domain−independent algorithms
Knowledgebase=setofsentencesinaformallanguage
Declarativeapproachtobuildinganagent(orothersystem):Tellitwhatitneedstoknow
ThenitcanAskitselfwhattodo—answersshouldfollowfromtheKB
Agentscanbeviewedattheknowledgeleveli.e.,whattheyknow,regardlessofhowimplemented
Orattheimplementationleveli.e.,datastructuresinKBandalgorithmsthatmanipulatethem
Chapter73
A simple kno wledge-based agen t
functionKB-Agent(percept)returnsanactionstatic:KB,aknowledgebaset,acounter,initially0,indicatingtime
Tell(KB,Make-Percept-Sentence(percept,t))action←Ask(KB,Make-Action-Query(t))Tell(KB,Make-Action-Sentence(action,t))t←t+1returnaction
Theagentmustbeableto:Representstates,actions,etc.IncorporatenewperceptsUpdateinternalrepresentationsoftheworldDeducehiddenpropertiesoftheworldDeduceappropriateactions
Chapter7
W umpus W orld PEAS description
Performancemeasuregold+1000,death-1000-1perstep,-10forusingthearrowEnvironmentSquaresadjacenttowumpusaresmellySquaresadjacenttopitarebreezyGlitteriffgoldisinthesamesquareShootingkillswumpusifyouarefacingitShootingusesuptheonlyarrowGrabbingpicksupgoldifinsamesquareReleasingdropsthegoldinsamesquare BreezeBreeze Breeze BreezeBreeze
Stench Stench BreezePIT
PIT
PIT
123 1 2 3 4
START Gold Stench
ActuatorsLeftturn,Rightturn,Forward,Grab,Release,Shoot
SensorsBreeze,Glitter,Smell
Chapter7
W umpus w orld c haracterization
Observable??
Chapter7
W umpus w orld c haracterization
Observable??No—onlylocalperception
Deterministic??
Chapter77
W umpus w orld c haracterization
Observable??No—onlylocalperception
Deterministic??Yes—outcomesexactlyspecified
Episodic??
Chapter78
W umpus w orld c haracterization
Observable??No—onlylocalperception
Deterministic??Yes—outcomesexactlyspecified
Episodic??No—sequentialatthelevelofactions
Static??
Chapter79
W umpus w orld c haracterization
Observable??No—onlylocalperception
Deterministic??Yes—outcomesexactlyspecified
Episodic??No—sequentialatthelevelofactions
Static??Yes—WumpusandPitsdonotmove
Discrete??
Chapter710
W umpus w orld c haracterization
Observable??No—onlylocalperception
Deterministic??Yes—outcomesexactlyspecified
Episodic??No—sequentialatthelevelofactions
Static??Yes—WumpusandPitsdonotmove
Discrete??Yes
Single-agent??
Chapter711
W umpus w orld c haracterization
Observable??No—onlylocalperception
Deterministic??Yes—outcomesexactlyspecified
Episodic??No—sequentialatthelevelofactions
Static??Yes—WumpusandPitsdonotmove
Discrete??Yes
Single-agent??Yes—Wumpusisessentiallyanaturalfeature
Chapter712
Exploring a wumpus w orld
A OK
OKOK
Chapter713
Exploring a wumpus w orld
OK
OKOK A
A B
Chapter714
Exploring a wumpus w orld
OK
OKOK A
A B P?
P?
Chapter715
Exploring a wumpus w orld
OK
OKOK A
A B P?
P?A S
Chapter7
Exploring a wumpus w orld
OK
OKOK A
A B P?
P?A S OK
P
W
Chapter7
Exploring a wumpus w orld
OK
OKOK A
A B P?
P?A S OK
P
W
AChapter7
Exploring a wumpus w orld
OK
OKOK A
A B P?
P?A S OK
P
W
A OK OKChapter719
Exploring a wumpus w orld
OK
OKOK A
A B P?
P?A S OK
P
W
A OK OKA BGS
Chapter720
Other tigh t sp ots
A BOK
OKOK
A B
A P?
P?P?
P? Breezein(1,2)and(2,1)⇒nosafeactions
Assumingpitsuniformlydistributed,(2,2)haspitw/prob0.86,vs.0.31
A S Smellin(1,1)⇒cannotmoveCanuseastrategyofcoercion:shootstraightaheadwumpuswasthere⇒dead⇒safewumpuswasn’tthere⇒safe
Chapter721
Logic in general
Logicsareformallanguagesforrepresentinginformationsuchthatconclusionscanbedrawn
Syntaxdefinesthesentencesinthelanguage
Semanticsdefinethe“meaning”ofsentences;i.e.,definetruthofasentenceinaworld
E.g.,thelanguageofarithmetic
x+2≥yisasentence;x2+y>isnotasentence
x+2≥yistrueiffthenumberx+2isnolessthanthenumbery
x+2≥yistrueinaworldwherex=7,y=1x+2≥yisfalseinaworldwherex=0,y=6
Chapter722
En tailmen t
Entailmentmeansthatonethingfollowsfromanother:
KB|=α
KnowledgebaseKBentailssentenceαifandonlyifαistrueinallworldswhereKBistrue
E.g.,theKBcontaining“theGiantswon”and“theRedswon”entails“EithertheGiantswonortheRedswon”
E.g.,x+y=4entails4=x+y
Entailmentisarelationshipbetweensentences(i.e.,syntax)thatisbasedonsemantics
Note:brainsprocesssyntax(ofsomesort)
Chapter723
Mo dels
Logicianstypicallythinkintermsofmodels,whichareformallystructuredworldswithrespecttowhichtruthcanbeevaluated
Wesaymisamodelofasentenceαifαistrueinm
M(α)isthesetofallmodelsofα
ThenKB|=αifandonlyifM(KB)⊆M(α)
E.g.KB=GiantswonandRedswonα=GiantswonM( )
M(KB) x
x
x x x
x xx x
xxxxx x xxx
xxxxx xx xx xxx x x x xxx
x
xx
x x
xx
x xx
x
Chapter724
En tailmen t in the wumpus w orld
Situationafterdetectingnothingin[1,1],movingright,breezein[2,1]
Considerpossiblemodelsfor?sassumingonlypits AA B
? ? ?
3Booleanchoices⇒8possiblemodels
Chapter725
W umpus mo dels
123 1 2
Breeze PIT 123 1 2
Breeze PIT
123 1 2
Breeze PITPIT
PIT 123 1 2
Breeze PIT
PIT 123 1 2
BreezePIT
123 1 2
Breeze PIT
PIT
123 1 2
Breeze PITPIT 123 1 2
Breeze
Chapter726
W umpus mo dels
123 1 2
Breeze PIT 123 1 2
Breeze PIT
123 1 2
Breeze PITPIT
PIT 123 1 2
Breeze PIT
PIT 123 1 2
BreezePIT
123 1 2
Breeze PIT
PIT
123 1 2
Breeze PITPIT 123 1 2
Breeze KB
KB=wumpus-worldrules+observations
Chapter727
W umpus mo dels
123 1 2
Breeze PIT 123 1 2
Breeze PIT
123 1 2
Breeze PITPIT
PIT 123 1 2
Breeze PIT
PIT 123 1 2
BreezePIT
123 1 2
Breeze PIT
PIT
123 1 2
Breeze PITPIT 123 1 2
Breeze KB1
KB=wumpus-worldrules+observations
α1=“[1,2]issafe”,KB|=α1,provedbymodelchecking
Chapter7
W umpus mo dels
123 1 2
Breeze PIT 123 1 2
Breeze PIT
123 1 2
Breeze PITPIT
PIT 123 1 2
Breeze PIT
PIT 123 1 2
BreezePIT
123 1 2
Breeze PIT
PIT
123 1 2
Breeze PITPIT 123 1 2
Breeze KB
KB=wumpus-worldrules+observations
Chapter7
W umpus mo dels
123 1 2
Breeze PIT 123 1 2
Breeze PIT
123 1 2
Breeze PITPIT
PIT 123 1 2
Breeze PIT
PIT 123 1 2
BreezePIT
123 1 2
Breeze PIT
PIT
123 1 2
Breeze PITPIT 123 1 2
Breeze KB 2
KB=wumpus-worldrules+observations
α2=“[2,2]issafe”,KB6|=α2
Chapter7
Inference
KB`iα=sentenceαcanbederivedfromKBbyprocedurei
ConsequencesofKBareahaystack;αisaneedle.Entailment=needleinhaystack;inference=findingit
Soundness:iissoundifwheneverKB`iα,itisalsotruethatKB|=α Completeness:iiscompleteifwheneverKB|=α,itisalsotruethatKB`iα
Preview:wewilldefinealogic(first-orderlogic)whichisexpressiveenoughtosayalmostanythingofinterest,andforwhichthereexistsasoundandcompleteinferenceprocedure.
Thatis,theprocedurewillansweranyquestionwhoseanswerfollowsfromwhatisknownbytheKB.
Chapter731
Prop ositional logic: Syn tax
Propositionallogicisthesimplestlogic—illustratesbasicideas
ThepropositionsymbolsP1,P2etcaresentences
IfSisasentence,¬Sisasentence(negation)
IfS1andS2aresentences,S1∧S2isasentence(conjunction) IfS1andS2aresentences,S1∨S2isasentence(disjunction) IfS1andS2aresentences,S1⇒S2isasentence(implication) IfS1andS2aresentences,S1⇔S2isasentence(biconditional)
Chapter732
Prop ositional logic: Seman tics
Eachmodelspecifiestrue/falseforeachpropositionsymbol
E.g.P1,2P2,2P3,1truetruefalse
(Withthesesymbols,8possiblemodels,canbeenumeratedautomatically.)
Rulesforevaluatingtruthwithrespecttoamodelm:
¬SistrueiffSisfalseS1∧S2istrueiffS1istrueandS2istrueS1∨S2istrueiffS1istrueorS2istrueS1⇒S2istrueiffS1isfalseorS2istruei.e.,isfalseiffS1istrueandS2isfalseS1⇔S2istrueiffS1⇒S2istrueandS2⇒S1istrue Simplerecursiveprocessevaluatesanarbitrarysentence,e.g.,¬P1,2∧(P2,2∨P3,1)=true∧(false∨true)=true∧true=true
Chapter733
T ruth tables for connectiv es
PQ¬PP∧QP∨QP⇒QP⇔Qfalsefalsetruefalsefalsetruetruefalsetruetruefalsetruetruefalsetruefalsefalsefalsetruefalsefalsetruetruefalsetruetruetruetrue
Chapter734
W umpus w orld sen tences
LetPi,jbetrueifthereisapitin[i,j].LetBi,jbetrueifthereisabreezein[i,j].
¬P1,1¬B1,1B2,1
“Pitscausebreezesinadjacentsquares”
Chapter735
W umpus w orld sen tences
LetPi,jbetrueifthereisapitin[i,j].LetBi,jbetrueifthereisabreezein[i,j].
¬P1,1¬B1,1B2,1
“Pitscausebreezesinadjacentsquares”
B1,1⇔(P1,2∨P2,1)B2,1⇔(P1,1∨P2,2∨P3,1)
“Asquareisbreezyifandonlyifthereisanadjacentpit”
Chapter736
T ruth tables for inference
B1,1B2,1P1,1P1,2P2,1P2,2P3,1R1R2R3R4R5KBfalsefalsefalsefalsefalsefalsefalsetruetruetruetruefalsefalsefalsefalsefalsefalsefalsefalsetruetruetruefalsetruefalsefalse.......................................
falsetruefalsefalsefalsefalsefalsetruetruefalsetruetruefalsefalsetruefalsefalsefalsefalsetruetruetruetruetruetruetruefalsetruefalsefalsefalsetruefalsetruetruetruetruetruetruefalsetruefalsefalsefalsetruetruetruetruetruetruetruetruefalsetruefalsefalsetruefalsefalsetruefalsefalsetruetruefalse.......................................
truetruetruetruetruetruetruefalsetruetruefalsetruefalse
Enumeraterows(differentassignmentstosymbols),ifKBistrueinrow,checkthatαistoo
Chapter737
Inference b y en umeration
Depth-firstenumerationofallmodelsissoundandcomplete
functionTT-Entails?(KB,α)returnstrueorfalseinputs:KB,theknowledgebase,asentenceinpropositionallogicα,thequery,asentenceinpropositionallogic
symbols←alistofthepropositionsymbolsinKBandαreturnTT-Check-All(KB,α,symbols,[])
functionTT-Check-All(KB,α,symbols,model)returnstrueorfalseifEmpty?(symbols)thenifPL-True?(KB,model)thenreturnPL-True?(α,model)elsereturntrueelsedoP←First(symbols);rest←Rest(symbols)returnTT-Check-All(KB,α,rest,Extend(P,true,model))andTT-Check-All(KB,α,rest,Extend(P,false,model))
O(2n)fornsymbols;problemisco-NP-complete
Chapter738
Logical equiv alence
Twosentencesarelogicallyequivalentifftrueinsamemodels:α≡βifandonlyifα|=βandβ|=α
(α∧β)≡(β∧α)commutativityof∧(α∨β)≡(β∨α)commutativityof∨((α∧β)∧γ)≡(α∧(β∧γ))associativityof∧((α∨β)∨γ)≡(α∨(β∨γ))associativityof∨¬(¬α)≡αdouble-negationelimination(α⇒β)≡(¬β⇒¬α)contraposition(α⇒β)≡(¬α∨β)implicationelimination(α⇔β)≡((α⇒β)∧(β⇒α))biconditionalelimination¬(α∧β)≡(¬α∨¬β)DeMorgan¬(α∨β)≡(¬α∧¬β)DeMorgan(α∧(β∨γ))≡((α∧β)∨(α∧γ))distributivityof∧over∨(α∨(β∧γ))≡((α∨β)∧(α∨γ))distributivityof∨over∧
Chapter739
V alidit y and satisfiabilit y
Asentenceisvalidifitistrueinallmodels,e.g.,True,A∨¬A,A⇒A,(A∧(A⇒B))⇒B
ValidityisconnectedtoinferenceviatheDeductionTheorem:KB|=αifandonlyif(KB⇒α)isvalid
Asentenceissatisfiableifitistrueinsomemodele.g.,A∨B,C
Asentenceisunsatisfiableifitistrueinnomodelse.g.,A∧¬A
Satisfiabilityisconnectedtoinferenceviathefollowing:KB|=αifandonlyif(KB∧¬α)isunsatisfiablei.e.,proveαbyreductioadabsurdum
Chapter7
Pro of metho ds
Proofmethodsdivideinto(roughly)twokinds:
Applicationofinferencerules–Legitimate(sound)generationofnewsentencesfromold–Proof=asequenceofinferenceruleapplicationsCanuseinferencerulesasoperatorsinastandardsearchalg.–Typicallyrequiretranslationofsentencesintoanormalform
Modelcheckingtruthtableenumeration(alwaysexponentialinn)improvedbacktracking,e.g.,Davis–Putnam–Logemann–Lovelandheuristicsearchinmodelspace(soundbutincomplete)e.g.,min-conflicts-likehill-climbingalgorithms
Chapter7
F orw ard and bac kw ard c haining
HornForm(restricted)KB=conjunctionofHornclausesHornclause=♦propositionsymbol;or♦(conjunctionofsymbols)⇒symbolE.g.,C∧(B⇒A)∧(C∧D⇒B)
ModusPonens(forHornForm):completeforHornKBs
α1,...,αn,α1∧···∧αn⇒ββ
Canbeusedwithforwardchainingorbackwardchaining.Thesealgorithmsareverynaturalandruninlineartime
Chapter7
F orw ard c haining
Idea:fireanyrulewhosepremisesaresatisfiedintheKB,additsconclusiontotheKB,untilqueryisfound
P⇒Q
L∧M⇒P
B∧L⇒MA∧P⇒L
A∧B⇒L
A
B QP
M
L
BA
Chapter743
F orw ard c haining algorithm
functionPL-FC-Entails?(KB,q)returnstrueorfalseinputs:KB,theknowledgebase,asetofpropositionalHornclausesq,thequery,apropositionsymbollocalvariables:count,atable,indexedbyclause,initiallythenumberofpremisesinferred,atable,indexedbysymbol,eachentryinitiallyfalseagenda,alistofsymbols,initiallythesymbolsknowninKB
whileagendaisnotemptydop←Pop(agenda)unlessinferred[p]doinferred[p]←trueforeachHornclausecinwhosepremisepappearsdodecrementcount[c]ifcount[c]=0thendoifHead[c]=qthenreturntruePush(Head[c],agenda)returnfalse
Chapter744
F orw ard c haining example
QP
M
L
BA 22 2 2 1
Chapter745
F orw ard c haining example
QP
M
L
B 2 1
A 11 2
Chapter746
F orw ard c haining example
QP
M 2 1
A 1
B 0 1L
Chapter747
F orw ard c haining example
QP
M 1
A 1
B 0 L 0 1
Chapter748
F orw ard c haining example
Q1
A 1
B 0 L 0 M 0 P
Chapter749
F orw ard c haining example
QAB 0 L 0 M 0 P 0
0
Chapter750
F orw ard c haining example
QAB 0 L 0 M 0 P 0
0
Chapter751
F orw ard c haining example
AB 0 L 0 M 0 P 0
0 Q
Chapter7
Pro of of completeness
FCderiveseveryatomicsentencethatisentailedbyKB
1.FCreachesafixedpointwherenonewatomicsentencesarederived
2.Considerthefinalstateasamodelm,assigningtrue/falsetosymbols
3.EveryclauseintheoriginalKBistrueinmProof:Supposeaclausea1∧...∧ak⇒bisfalseinmThena1∧...∧akistrueinmandbisfalseinmThereforethealgorithmhasnotreachedafixedpoint!
4.HencemisamodelofKB
5.IfKB|=q,qistrueineverymodelofKB,includingm
Generalidea:constructanymodelofKBbysoundinference,checkα
Chapter7
Bac kw ard c haining
Idea:workbackwardsfromthequeryq:toproveqbyBC,checkifqisknownalready,orprovebyBCallpremisesofsomeruleconcludingq
Avoidloops:checkifnewsubgoalisalreadyonthegoalstack
Avoidrepeatedwork:checkifnewsubgoal1)hasalreadybeenprovedtrue,or2)hasalreadyfailed
Chapter7
Bac kw ard c haining example
QP
M
L
AB
Chapter755
Bac kw ard c haining example
P
M
L
A Q
B
Chapter756
Bac kw ard c haining example
M
L
A QP
B
Chapter757
Bac kw ard c haining example
M
A QP
L
B
Chapter758
Bac kw ard c haining example
M
L
A QP
B
Chapter759
Bac kw ard c haining example
M
A QP
L
B
Chapter760
Bac kw ard c haining example
M
A QP
L
B
Chapter761
Bac kw ard c haining example
A QP
L
B M
Chapter762
Bac kw ard c haining example
A QP
L
B M
Chapter763
Bac kw ard c haining example
A QP
L
B M
Chapter7
Bac kw ard c haining example
A QP
L
B M
Chapter7
F orw ard vs. bac kw ard c haining
FCisdata-driven,cf.automatic,unconsciousprocessing,e.g.,objectrecognition,routinedecisions
Maydolotsofworkthatisirrelevanttothegoal
BCisgoal-driven,appropriateforproblem-solving,e.g.,Wherearemykeys?HowdoIgetintoaPhDprogram?
ComplexityofBCcanbemuchlessthanlinearinsizeofKB
Chapter7
Resolution
ConjunctiveNormalForm(CNF—universal)conjunctionofdisjunctionsofliterals|{z}clausesE.g.,(A∨¬B)∧(B∨¬C∨¬D)
Resolutioninferencerule(forCNF):completeforpropositionallogic
`1∨···∨`k,m1∨···∨mn`1∨···∨`i−1∨`i+1∨···∨`k∨m1∨···∨mj−1∨mj+1∨···∨mn
where`iandmjarecomplementaryliterals.E.g.,
OK
OKOK A
A B P?
P?
A S OK P
W A P1,3∨P2,2,¬P2,2P1,3
Resolutionissoundandcompleteforpropositionallogic
Chapter767
Con v ersion to CNF
B1,1⇔(P1,2∨P2,1)
1.Eliminate⇔,replacingα⇔βwith(α⇒β)∧(β⇒α).
(B1,1⇒(P1,2∨P2,1))∧((P1,2∨P2,1)⇒B1,1)
2.Eliminate⇒,replacingα⇒βwith¬α∨β.
(¬B1,1∨P1,2∨P2,1)∧(¬(P1,2∨P2,1)∨B1,1)
3.Move¬inwardsusingdeMorgan’srulesanddouble-negation:
(¬B1,1∨P1,2∨P2,1)∧((¬P1,2∧¬P2,1)∨B1,1)
4.Applydistributivitylaw(∨over∧)andflatten:
(¬B1,1∨P1,2∨P2,1)∧(¬P1,2∨B1,1)∧(¬P2,1∨B1,1)
Chapter768
Resolution algorithm
Proofbycontradiction,i.e.,showKB∧¬αunsatisfiable
functionPL-Resolution(KB,α)returnstrueorfalseinputs:KB,theknowledgebase,asentenceinpropositionallogicα,thequery,asentenceinpropositionallogic
clauses←thesetofclausesintheCNFrepresentationofKB∧¬αnew←{}loopdoforeachCi,Cjinclausesdoresolvents←PL-Resolve(Ci,Cj)ifresolventscontainstheemptyclausethenreturntruenew←new∪resolventsifnew⊆clausesthenreturnfalseclauses←clauses∪new
Chapter769
Resolution example
KB=(B1,1⇔(P1,2∨P2,1))∧¬B1,1α=¬P1,2
P1,2 P1,2
P2,1 P1,2B1,1
B1,1P2,1B1,1P1,2P2,1P2,1 P1,2B1,1B1,1 P1,2B1,1P2,1B1,1P2,1B1,1
P1,2P2,1P1,2
Chapter770
Summary
Logicalagentsapplyinferencetoaknowledgebasetoderivenewinformationandmakedecisions
Basicconceptsoflogic:–syntax:formalstructureofsentences–semantics:truthofsentenceswrtmodels–entailment:necessarytruthofonesentencegivenanother–inference:derivingsentencesfromothersentences–soundess:derivationsproduceonlyentailedsentences–completeness:derivationscanproduceallentailedsentences
Wumpusworldrequirestheabilitytorepresentpartialandnegatedinforma-tion,reasonbycases,etc.
Forward,backwardchainingarelinear-time,completeforHornclausesResolutioniscompleteforpropositionallogic
Propositionallogiclacksexpressivepower
Chapter771