The Logic of IntelligenceThe Logic of Intelligence
Pei WangDepartment of Computer and Information Sciences
Temple University
Artificial General Intelligence
MainstreamAItreats“Intelligence”asacollectionofproblem-specificanddomain-specificparts
ArtificialGeneralIntelligence(AGI)takes“Intelligence”asageneral-purposecapabilitythatshouldbetreatedasawhole
AGIresearchstillincludesdifferentresearchobjectivesandstrategies
Artificial Intelligence and Logic
“Intelligence”canbeunderstoodas“rationality”and“validity”---“dotherightthing”
Ingeneral,“logic”isthestudyofvalid reasoning,ortheregularityinthinking
Therefore,anAIsystemmaybebuiltaccordingtoalogic,byconvertingvariousthinkingprocessesintoreasoningprocesses
Reasoning System
Areasoningsystemtypicallyconsistsofthefollowingmajorcomponents:
aformallanguage asemantictheory asetofinferencerules amemorystructure acontrolmechanismThefirstthreeareusuallycalleda“logic”
Traditional Theories
Languageandinferencerules:first-orderpredicatecalculus
Semantics:modeltheory Memory:relationalorobject-orienteddatastructuresanddatabase
Inferencecontrol:theoryofcomputation(algorithm,computability,andcomputationalcomplexity)
Problems of Traditional Theories Uncertainty:fuzzy concepts, changing meanings
and truth values, plausible results, conflicting evidence, nondeterministic inference process, …
Semanticjustificationofnon-deductiveinference:induction, abduction, analogy,…
Counter-intuitiveresults:sorites paradox, implication paradox, confirmation paradox, Wason’s selection task,…
Computabilityandcomplexity:termination problem,combinatorial explosion,…
Proposed Solutions non-monotoniclogic paraconsistentlogic relevancelogic probabilisticlogic fuzzylogic inductivelogic temporallogic modallogic situationcalculus possibleworldtheory
mentallogic mentalmodel case-basedreasoning Bayesiannetwork neuralnetwork geneticalgorithm heuristicalgorithm learningalgorithm anytimealgorithm……
Common Root of the Problems
Thetraditionaltheoriesweredevelopedinthestudyofthefoundationofmathematics,whiletheproblemsappearoutsidemath
Thelogicofmathematicsmaybedifferentfromthelogicofcognition
Inmathematicalreasoning,theknowledgeandresourcesareassumedtobesufficient(withrespecttothetasks)
Different Types of Systems “Pure-axiomaticsystem”:thesystem’sknowledgeandresourcesareassumedtobesufficient
“Semi-axiomaticsystem”:certainaspects(butnotall)oftheknowledgeandresourcesareassumedtobesufficient
“Non-axiomaticsystem”:theknowledgeandresourcesofthesystemareassumedtobegenerallyinsufficient
NARS (Non-Axiomatic Reasoning System)
NARSusesaformallogic(language,semantics,inferencerules)andisimplementedinacomputersystem
NARSisfullybasedontheassumptionofinsufficientknowledgeandresources,inthesenseofbeingafinite,real time,open,andadaptivesystem
NARSisdifferentfromtraditionaltheoriesinallmajorcomponents
Inheritance Based Representation
S P : thereisaninheritance relationfromtermStotermP
S isa specializationof PP isageneralizationof S
Inheritanceisreflexiveandtransitive
bird animal
Extension and Intension
ForagiventermT,its extension TE={x|xT}its intension TI={x|T x}
TTE TI
Theorem:(S P) (SE PE) (PISI)
Therefore,“Inheritance”means“inheritanceofextension/intension”
Evidence
PositiveevidenceofS P:{x|x (SE
PE)(PI SI)}
NegativeevidenceofS P:{x | x (SE
–PE)(PI –SI)}
S P
Amountofevidence:positive: w+=|SE
PE |+|PI
SI |
negative: w–=|SE –PE|+|PI
–SI|
total: w = w++w–=|SE |+|PI
|
Truth Value
InNARS,thetruthvalueofastatementisapairofnumbers,andmeasurestheevidentialsupporttothestatement.
S P[f, c]S P [f,c] f:frequency,w+/w c:confidence,w/(w +1)
Experience-Grounded Semantics
Thetruthvalueofastatementisdefinedaccordingtocertain“idealizedexperience”,consistingofasetofbinaryinheritancestatements
Themeaningofatermisdefinedbyitsextensionandintension,accordingtocertain“idealizedexperience”
Someaningandtruth-valuechangesaccordingtothesystem’sexperience
Syllogistic Inference Rules
Atypicalsyllogisticinferenceruletakesapairofpremiseswithacommonterm,andproducesaconclusion
Thetruthvalueoftheconclusioniscalculatedbyatruth-valuefunction
Differentcombinationsofpremisestriggerdifferentrules(withdifferenttruth-valuefunctions)
To Design a Truth-value Function
1.TreatallinvolvedvariablesasBoolean(binary)variables
2.Foreachvaluecombinationinpremises,decidethevaluesinconclusion
3.BuildBooleanfunctionsamongthevariables4.Extendthefunctionstoreal-number:
not(x)=1–xand(x,y)=x*yor(x,y)=1–(1–x)*(1–y)
Deduction
birdanimal[1.00,0.90]robinbird[1.00,0.90]
robinanimal[1.00,0.81]
M P [f1, c1] S M [f2, c2] S P [f, c]
f = f1 * f2
c = c1 * c2 * f1 * f2
M
S P
Induction
swanbird[1.00,0.90]swanswimmer[1.00,0.90]
birdswimmer[1.00,0.45]
M P [f1, c1] M S [f2, c2] S P [f, c]
f = f1
c = f2 * c1 * c2 / (f2 * c1 * c2 + 1)
S
M
P
Abduction
seabirdswimmer[1.00,0.90]gullswimmer[1.00,0.90]gullseabird[1.00,0.45]
P M [f1, c1] S M [f2, c2] S P [f, c]
f = f2
c = f1 * c1 * c2 / (f1 * c1 * c2 + 1)
S
M
P
Revision
birdswimmer[1.00,0.62]birdswimmer[0.00,0.45]
birdswimmer[0.67,0.71]
S P [f1, c1] S P [f2, c2] S P [f, c]
f =f1 * c1 * (1 - c2) + f2 * c2 * (1 - c1)
c1 * (1 - c2) + c2 * (1 - c1)
c1 * (1 - c2) + c2 * (1 - c1)
c1 * (1 - c2) + c2 * (1 - c1) + (1 - c2) * (1 - c1)
c =
S P
Other Inference Rules
MP[f1,c1]SM[f2,c2]SP[f,c]
analogy
union
implication P M [f1, c1] S M [f2, c2](S P) M [f, c]
B C [f1, c1]A B [f2, c2] A C [f, c]
Other Relations and InheritanceAnarbitrarystatementR(a, b, c)canberewrittenasinheritancerelationswithcompoundterms:
(*,a, b, c)R“Therelationamonga, b, cisakindofR.”
a(/,R,_,b, c)“a issuchanx thatsatisfiesR(x, b, c).”
b(/,R,a,_,c)“b issuchanx thatsatisfiesR(a, x, c).”
c(/,R,a, b,_)“c issuchanx thatsatisfiesR(a, b, x).”
Memory as a Belief Network
bird
gull
swan
robinswimmer
crow
feathered_creature
[1.00, 0.90] [1.00, 0.90]
[0.0
0, 0
.90]
[1.00, 0.90]
[1.00, 0.90] [1.00, 0.90]
[1.00, 0.90]
[1.00, 0.90]Theknowledgeofthesystemisanetworkofbeliefsamongterms.Atermwithallofitsbeliefsisaconcept
Cbird
Inference TasksNARSacceptsseveraltypesofinferencetasks: Knowledgetobeabsorbed Questionstobeanswered GoalstobeachievedAtaskisstoredinthecorrespondingconceptsToprocesseachtaskmeanslettingitinteracts
withtheavailablebeliefsintheconceptThisprocessusuallygeneratesnewtasks,
beliefs,andconcepts,recursively
Inference Process
NARSrunsbyrepeatingthefollowingcycle:1. Chooseaconceptwithinthememory2. Chooseataskwithintheconcept3. Chooseabeliefwithintheconcept4. Useinferencerulestoproducenewtasks5. Returntheuseditemstomemory6. Addthenewtasksintothememoryand
provideananswerifavailable
Control Strategy
NARSmaintains priority distributionsamongdataitems,usesthemtomakechoice,andadjuststhemaftereachstep
Factorsinfluencepriority: qualityoftheitem usefulnessoftheiteminhistory relevanceoftheitemtothecurrentcontext
Architecture and Working Cycle
Design and ImplementationTheconceptualdesignofNARShasbeen
describedinaseriesofpublicationsMostpartsofthedesignhavebeen
implementedinseveralprototypes,andthecurrentversionisopensourceinJava
Workingexamplesexistasproofofconcept,andonlycoversingle-stepinferenceorshortinferenceprocesses
Theprojectison-going,thoughhasproducednovelandinterestingresults
Unified Solutions Thetruthvalueuniformlyrepresentsvariouskindsofuncertainty
Thetruthvaluedependsonbothpositiveandnegativeevidence
Thenon-deductiveinferencerulesisjustifiedaccordingtothesemantics
Themeaningofatermisdeterminedbyitsexperiencedrelationswithotherterms
Withsyllogisticrules,thepremisesandconclusionsmustbesemanticallyrelated
TheinferenceprocessesinNARSdoesnotfollowpredeterminedalgorithms
ConclusionsItispossibletobuildareasoningsystemthatadaptstoitsenvironment,andworkswithinsufficientknowledgeandresources
SuchasystemprovidesaunifiedsolutiontomanyproblemsinA(G)I
Thereisa logic of intelligence,thoughitisfundamentallydifferentfromthe logic of mathematics