In Cognitive Processing,Vol 2 No 1, 2001,pp 177–198

Beyond Shallow Modelsof Emotion

Aaron SlomanSchoolof ComputerScience

TheUniversityof BirminghamBirmingham,B15 2TT, UK

http://www.cs.bham.ac.uk/˜axs/�

Abstract

Thereis a hugediversity of definitionsof “emotion” someof which areassociatedwithrelatively shallow behaviouralor measurablecriteriaor introspectableexperiences,for instanceuseof facial expression,physiologicalmeasures,activity of specificregionsof the brain, ortheexperienceof bodily changesor desires,suchaswantingto run away, or to hurt someone.Therearealsodeepertheoriesthat link emotionalstatesto a variety of mechanismswithinan informationprocessingarchitecturethatarenot easilyobservableor measurable,not leastbecausethey arecomponentsof virtual machinesratherthanphysicalor physiologicalmech-anisms.We cancomparethis with “shallow” definitionsof chemicalcompoundssuchassalt,sugar, or water, in termsof theirappearanceandobservedbehavioursin varioustestsituations,andtheirdefinitionsin thecontext of atheoryof thearchitectureof matterwhichis mostlycon-cernedwith postulatedsub-atomicentitiesandandawebof relationshipsbetweenthemwhichcannoteasilybeobserved,sothattheoriesaboutthemarenoteasilyconfirmedor refuted.Thispaperoutlinesanapproachto thesearchfor deeperexplanatorytheoriesof emotionsandmanyotherkindsof mentalphenomena,which includesanattemptto definetheconceptsin termsof theunderlyinginformationprocessingarchitecturesandtheclassesof statesandprocessesthatthey cansupport.A seriousproblemwith thisprogrammeis thedifficulty of findinggoodconstraintson theories,sincein generalobservable factsareconsistentwith infinitely manyexplanatorymechanisms.This“positionpaper”offersasapartialsolutiontherequirementthatproposedarchitecturesbecapableof having beenproducedby biologicalevolution,in additionto being subjectto constraintssuchas implementabilityin known biological mechanisms,variousresourcelimits (time,memory, energy, etc.)andbeingableto accountfor awiderangeof humanfunctionality. Within suchanarchitecture-basedtheorywe candistinguish(at least)primaryemotions,secondaryemotions,andtertiaryemotions,andproducea coherenttheorywhich explainsa wide rangeof phenomenaandalsopartly explainsthediversityof theories:mosttheoristsfocusononly asubsetof typesof emotions,like theproverbialblind mentryingto saywhatanelephantis on thebasisof feelingonly a leg, anear, a tusk,thetrunk,etc.Keywords: affect,architecture,artificial intelligence,cognitivescience,deliberative,emotion,evolution, intelligence,meta-management,mind,reactive, reflective virtual machine.

�This is a revisedversionof a paperpresentedat theworkshopon Behaviourplanningfor life-like avatars, at the

I3 SpringDaysWorkshop,March1999,Sitges,Spain. It is not intendedto bea thoroughandscholarlysurvey, buta provocative “position paper”which outlinesan ambitiousapproachto the studyof mind, building on the variousapproacheswhich it criticisesasinadequate!

1 Intr oduction

The studyof emotionin cognitive scienceandAI hasrecentlybecomevery fashionable,with arapidly growing numberof workshops,conferencesandpublicationson the topic, somereport-ing attemptsto produceemotionalbehaviour in robotsor softwareagents,someconcernedwithdetectingandrespondingto emotionsin humanusersof computingsystems,andsomeaimingtomodelandexplainhumanemotions.

This is not a new topic in AI, asshown by Simon’s importantcontribution over 30 yearsago(Simon,1967),andvariouspapersnearly20yearsagoin IJCAI’81 includingmy first paperonthistopic (SlomanandCroucher, 1981),whichwasmuchinfluencedby Simon’snotionthatemotions,motivationsandotheraffectivephenomenaweredeeplyentwinedwith cognitiveprocessesandthemechanismsfor controlof internalandexternalbehaviour in intelligentagents.

There are now many useful surveys of issuesconcerningemotions1, but it is difficult fornewcomersto thefield to achieve a balancedoverview, not leastbecause(asOatley andJenkinspoint out) thereis a very wide variety of definitionsof “emotion” offered by researcherswithdifferentviewpoints.For AI researchersaimingto produceworkingsystemsit is temptingto thinkof emotionsasrelatively easilysimulatedpatternsof behaviour. The result is a tendency for re-searchersto presentsimplisticAI programsandrobotsasif they justifiedepithetslike“emotional”,“sad”, “surprised”,etc.

Suchprogramsmaybebasedonanattemptto analyseconditionsunderwhichcertainemotionsarethoughtto occurandthebehaviours typical of suchemotions.This leadsto thedesignof anarchitecturecontrollinga robotor interactivesoftwaresystem,in which thereis a sub-component(possiblylabelled“emotion”) which testsfor thoseconditionsandgeneratesthe correspondingbehaviours, possiblyusing statevariableswith nameslike “angry”, “frightened”, “surprised”,“pleased”,etc. eitherwith booleanvaluesthatcanbetoggledor with a numericalor “qualitative”rangeof valuesfor eachvariable.Thesemodelsareshallow insofarasthey have relatively simplerelationshipsbetweeninput and output. This is similar to a practicelambastedlong ago byMcDermott (1981) namelyusing terms like “goal”, “plan”, “learn”, simply becausethereareproceduresor variableswith thesenamesin aprogram.

2 Shallow modelsarenot all bad

Someresearchers(Bateset al., 1991;Reilly, 1996)have quite explicitly acknowledgedthat theyareaiming for shallowmodelswhosemeritsarebasedon breadth, namelypossessinga varietyof capabilitiessupportedby diversemechanisms,or mechanismsthatcancopewith a wide rangeof cases.Such“broadandshallow” designsmaybeusefulfor certainpracticalpurposessuchasenliveningcomputergamesor otherinteractive entertainmentsor perhapshelpingnaive usersofcomputingsystemsby makingthemappearmore“human” thanthey are.Onewayto achievesuchbreadth,while still usingashallow model,is to try to encompassaverywide rangeof cases,suchasthosesurveyedin (Ortony etal., 1988).

Shallow modelsarefineif they havea limited purposewhich is madeclear, e.g.to entertain,orto teachprogramming,or to modelsomelimited aspectof controlof postureor facialexpression,

1E.g. (Goleman,1996;LeDoux,1996;Oatley andJenkins,1996;Ortony et al., 1988;Picard,1997;Elliot, 1998;Hatanoet al., 2000)

2

etc. I have a very shallow model2 in which simulatedmobile robotscanbe in statesdescribedasglum, surprised,neutralor happy, but this is nothingmorethanan elementaryteachingtool.Studentsplay with and extend it in order to learn agentprogrammingtechniques.In the nearfuture, therewill probablybe a growing useof very shallow modelsof emotion in computerentertainments.Thereis nothingwrongwith that,if they aresuccessfulat entertaining.Howeverthat doesnot necessarilymake themplausiblemodelsof humanor animalemotions.They maynotevenbeusefulstepsin thedirectionof suchmodels.

Shallow modelscansometimesplayarolein thesearchfor deepermodels.Building inadequatemodels,andexploring their capabilitiesandlimitationsis oftenanessentialpartof theprocessoflearninghow to designmorecomplex andmoresatisfactorymodels,asexplainedin (BeaudoinandSloman,1993;Sloman,1993b).

3 Inconsistentdefinitions and usages

If we want to understandandmodelwhatarenormally referredto as“emotions” in humansandother animalsthen we needto start from a deeperanalysisof the conceptswe are aiming toinstantiate. This task is madedifficult by the fact that we do not all agreein our usageof theword “emotion”. For example,somewill call surpriseanemotionwhereasothers(Ortony et al.,1988)will saythat it is just a cognitive statein which anexpectationhasbeenviolated,asoftenhappensin acomplex anddynamicworld,andcanevenoccurwhendoingmathematics.Of course,surprise,likeany otherstate,cantriggerstatesthatmostpeoplewouldcall emotions.

Therearealsodisagreementsover whetherpainsandpleasuresareemotions,someregardingit asobvious that they are,whereasothersfind it equallyobvious thatonecanhave thepainof apin-prick or thepleasureof eatinganice creamwithout feelingat all emotionalaboutit. E.g. onecanbetotally unconcernedaboutthepin-prick, while acknowledgingthat it hurt. Of course,veryintensepainis adifferentmatter.

Anotherexample:somepeoplebelievethatemotions,by definition,cannotexist withoutbeingexperienced,whereasothers(includingsomenovelistsandplaywrights)regardit asobviousthatsomeonecan be angry or infatuated(and thereforein an emotionalstate)without being awareof their state,even if friendsnotice it. On further investigationthis disputecansometimesturnon whetheranemotion’s beingexperiencedis taken to imply that theemotionis recognizedandlabelledassuch,or only to imply thatit involvesbeingawareof somementalstatesandprocessesrelatedto theemotion.At oneextremea theoristwill saythatyou cannotenjoy somethingunlessyou recognizeandcategoriseyour stateasenjoyment.An intermediatepositionwould claim thattheremustbesomeexperiencethatyou recognizeandcategorisewhich is partof theenjoyment,evenif thetotal stateis not recognized.At anotherextremeit is claimed(Ryle,1949)thatintenseenjoymentcanoccurwhereall one’s attentionis focusedon external phenomena,e.g. enjoyinga gameof football whereoneis thinking only of the otherplayers,wherethe ball is, who needsto bemarked,etc.,without beingawareof anything internalto oneself.Anothersuchexampleisenjoying anoperaor playwith attentionfully engagedby whatis happeningin thetheatre,withoutbeingawareof any additional processesgoing on in one’s own mind. Whenit is objectedthattheremustbesomeadditionalexperiencedstatefor enjoymentto occurit is not clearwhetherthisis aconceptualdisagreementor anempiricalone.(Whatevidencecouldhelpto settleit?)

2Seehttp://www.cs.bham.ac.uk/research/poplog/sim/teach/simfeelings

3

A differentdimensionof disagreementconcernstheattributionof emotionalstatesandpossiblyothermentalstates,to otheranimals.Doesa fish feel painwhencaughton a hook? Whena flydetectsandescapesjust in time from thehandslammingdown on it, doesit have a stateof fear,or relief at its narrow escape?Do painsoccurif you pull its legsoff? Debatesover animalrightsfrequentlyrevolve arounddisagreementsover what mentalstatesarepossiblefor animals. It isalsopossibleto argueovermentalstatesof ahumanfoetusor neonate.Doesthephysicalresponseto aprodshow thatahumanfoetus,or asnail,findsit unpleasant?

Not only aretheredifferencesin theoriesandusagesbetweenindividuals,it is evenpossiblefor individualsto beinconsistentin theirown usage,for instancesomepeoplewill statethatloveisa typeof emotion,thenlateradmitthatthey (a) they arenot in anemotionalstateand(b) thattheylovetheir family, theircountry, thegameof football,etc. It is possiblethatwhensuchpeopleofferloveasanemotionthey arethinkingof episodesof passionor fervour, whereaswhenthey saytheylovetheirfamily, etc. they arereferringto anattitudewhichis primarily acollectionof dispositionswhich are dormantmost of the time but can be triggered,undercertainconditions,to produceemotionalepisodes,involving variousmentalandphysicalprocesses.Similar inconsistenciescanariseover the classificationof moodsasemotions:someonemay regardbeing in an optimisticmoodasanemotionalstate,yet claim not to befeelingemotionalwhenin a statewhich they alsocharacteriseasoptimistic.

Inconsistenciesbetweenandwithin theexplicit theoriesandthenon-reflective linguisticusageof peoplewho talk about emotionsare an indication that we are dealing with a deepset ofconfusionsabouthow our ordinaryconceptswork. Perhapsthoseconceptsaresimply inadequatefor the purposeof characterisingthe enormouslyrich variety of mentalstatesthat canoccur inhumansandotheranimals.

From this viewpoint it is very rashto assumethat the aim of building machinesthat haveemotionsor whichmodelthemis awell-definedaim.

4 Possiblestrategies

Whatcanwedoaboutthis?Therearemany alternatives,includingthefollowing strategies.

� Giveuptalk of emotions(andothermentalstates)in ourscience(assomebehaviouriststriedto do).

� Inventaprecisedefinitionof “emotion”, for instancein termsof asetof condition-responsepatterns,anduseit regardlessof how it relatesto ordinaryusageor thedefinitionsofferedby others– thesimpleststrategy for would-beemotionmodellers.

� Treattheconceptsasinherentlyfuzzy or probabilisticandattemptto investigatetheassoci-atedprobabilitiesby doingresearchto find out probabilitiesof variouslabelsbeingusedinvariouscontexts,or theprobabilitiesof variousbehavioursor expressionsbeingusedwhenpeopleclaim to bein anemotionalstate.

� Attempt to producea deeptheory of the information processingarchitecturesunderlyingall thedifferentphenomena,andthendefinenew architecture-basedconceptsthatpreciselyidentify subsetsof thosephenomena.This could includestatesandprocessesinvolving theagent’s relationswith theperceivedphysicalor socialenvironment.

4

In the Birmingham Cognition and Affect project we have adoptedthe architecture-basedap-proach3, describedin papersin theprojectdirectoryat www.cs.bham.ac.uk/research/cogaff/.

5 Opinions about an elephant

Thishasledusto hypothesiseanexplanatoryarchitecture,sketchedbelow, andto identify varioustypesof statesand processesthat can occur in suchan architecture. We can then investigatethepropertiesof thosestatesandprocesseswhich seemto correspondto thesortsof phenomenathat areof concernboth in ordinaryconversationsaboutemotionsandalso in variousscientificandphilosophicalresearchendeavours. We canthen formulatedefinitionsof a wide variety ofstatesandprocessessupportedby the architecturewhich canbe groupedin variousways. Forinstance,asexplainedbelow, we have foundit illuminating to distinguishprimary, secondaryandtertiary emotions,whichariseoutof differentarchitecturallayersthatmayor maynotbepresentindifferentanimalandrobotarchitectures.Furthersubdivisionscanthenbemadewithin thesethreecategories.Wecanalsoinvestigatepreciselydefinedarchitecture-basedconceptsthatapproximateto otherlooseconceptsof ordinarylanguage,suchasmood,attitude,intention,desire,etc.

From this viewpoint the contradictoryopinionsexpressedby peoplestudyingemotionsareratherlike the opinionsof the proverbial ten blind men eachtrying to say what an elephantison the basisof feelingonly a small part of it. Insteadof arguing over which descriptionis rightwe cantry to characterisethewholeelephant,therebyexplainingthecontradictionsbetweenrivaldefinitionsandpartial theories.This is partly like theapproachadoptedin (Ortony et al., 1988),namelycharacterisingaspaceof possiblestatesindependentlyof debatesaboutwordsandphrasesaccuratelycorrespondto whichstates.

Themulti-layerarchitecturedescribedbelow accommodatesseveraldifferentvarietiesof stateswhich could be calledemotions:very primitive primary emotionsrootedin very old biologicalmechanismssuchasstartlemechanismssharedwith many otheranimals,andalsomoresophisti-catedsemanticallyrich secondaryandtertiaryemotionsthatareprobablyuniqueto humans(untilwebuild human-like robots),suchasbeingapprehensiveabouttheoutcomeof a risky plan,beinginfatuatedwith someone,or feelinghumiliatedbecausesomesilly mistakeyou madewaspointedoutby a famouspersonin a largepublic lecture.

The taxonomyof Ortony et al. focuseson a particular set of cognitive and motivationalstates(including what somepeoplewould describeas attitudesrather than emotions)and canbeaccommodatedwithin theclassesof secondaryandtertiaryemotionsdescribedbelow, thoughthey arelessconcernedwith thespecificationof acompletearchitecture.

6 How to achievegreaterdepth

A desirablebut rarely achieved type of depthin an explanatorytheoryis having a modelwhichaccountsfor a wide rangeof phenomena.Oneof the reasonsfor shallownessin psychologicaltheoriesis considerationof toosmallavarietyof cases.

3Previouslyreferredto asthedesign-basedapproachin (Sloman,1992),wheredesign-basedtheoriesarecontrastedwith phenomena-basedtheories,which merelylook for relationshipsbetweenobservablephenomenaandsemantics-basedtheorieswhichuselinguistic investigationsto discoverwhatwemeanby variousexpressionsdescribingmentalphenomena.

5

If insteadof thinking only aboutnormal adult humans(or only aboutrats as someexperi-mentalistsusedto do) we consideralso infants,peoplewith brain damageor disease,andalsoother animalsincluding insects,birds, bonobos,etc., we find evidencefor myriad informationprocessingarchitectureseachsupportingandexplaininga specificvarietyof mentalcapabilities.Yet morepossiblearchitectures,eachsupportinga collectionof possiblestatesandprocessescanbefoundin robots,softwaresystemsandmachinesof thefuture!

Conceptsdescribingmentalstatesandprocessesin oneanimalor machinemay be inappro-priatewhendescribinganother, if the latter lacksthe requiredarchitecture,even if its behaviourappearsto justify theattribution. For instance,a purely reactive animalreactingto a threateningsituationmay be thoughtto be in a stateof fear. But a geneticallydeterminedautomaticescapereactionis differentin many waysfrom anexternallysimilarescapereactionproducedby asystemthatunderstandstheimplicationsof thethreatandon thatbasisdecidesto escape.

Likewise,conceptsrelevantto normaladulthumansmaybeinappropriatefor new-borninfants,victimsof Alzheimer’sdisease,or anentertainingrobotwhichcanbemadeto lookhappy, annoyed,surprised,etc.

Althoughhumanadultsseemto beinnatelyprogrammedto attributeall sortsof mentalstatestoinfants,it is likely thatnew-borninfantsareincapableof having someof them.Mostpeoplewouldagreethat a newborn infant is incapableof wonderingwhetherit will ever have grandchildren.Why?Likewiseanewborninfantmaybeincapableof feelinghumiliatedby peoplelaughingat itsfacialexpression,if it lacksthearchitecturerequiredfor humiliation. It mayevenbeincapableoffeelingpainin thesamewayasanadult,despitedisplayingcompellingexternalsymptoms.

It oftengoesunnoticedthatmuchof whatpoetsandnovelistssayaboutus,andwhatwe sayaboutour friendsandourselveswhengossippingor discussingour interests,loves,hopes,fearsandambitions,implicitly presupposesthathumansareessentiallyinformationprocessingsystems.E.g. whenpoetsdistinguishfickle liking which is easilydiminishedby new informationanddeeplovewhich is not, they implicitly presupposethatnew informationcanhave powerful effectsoninformation-basedcontrolstates.

By consideringpossibledescriptive andexplanatoryconceptsgeneratedby a virtual machineinformation processingarchitecture we obtaina broaderand deeperexplanatorytheory than isnormally found in philosophy, psychologyor social science,or most computermodelling. Ofcourse,sucha theoryshouldsatisfyempiricalconstraintsincludingevolvability, implementabilityin neuralmechanisms,resourcelimits, etc.

7 Exploring neighbourhoodsin designspace

Looking at the variety of statesand processessupportableby a classof architectureshasbeenlikenedabove to seeingthewholeelephant.Unfortunately, thereis morethanone“elephant”tostudy, sincearchitecturesvarybetweenorganisms(andmachines)andevenwithin anindividual itmaydevelopover time,e.g.betweeninfancy andadulthood.

A full understandingof the variousphenomenathat might be calledemotions,thereforere-quirescomparative analysisof possibilitiesand trajectoriesin designspaceand niche space,4

4A niche,in biology or engineering,is anabstractsetof requirementsfor anorganismor machine,againstwhichinstancesof a classof designscanbecompared.In simplecaseswe canusea “fitnessfunction”, giving a numericalresult.In generaltherelationbetweenadesignanda nicheis bestthoughtof asacomplex qualitativedescription.

6

as outlined in (Sloman,1994; Sloman,1998b; Sloman,2000b). We understanda particulararchitecturebetterif weknow whatdifferenceswouldariseoutof varioussortsof designchanges:whichcapabilitieswouldbelostandwhichwouldbeadded.Wealsohaveadeeperunderstandingof thearchitectureif wecanseewhatsortsof pressuresandtrade-offs led to its evolution,andhowit mightdevelopor evolve in future.

This involvesgoingbeyondthemajority of AI projectsor psychologicalinvestigationsinsofarasit requiresusbothto considerdesignsfor completeagentsin additionto designsfor componentmechanismsandalsoto docomparativeanalysisof differentsortsof designs.

A comprehensive theoryof emotionsandothermentalstatesrequiresa survey of typesof in-formationprocessingarchitecturescoveringhumansof varioustypes,otheranimals,futurerobotsandsoftwareagents.For eachtypeof architecturewe canpreciselydefinethesortsof statesandprocessesit supports,and if we decideto label someof thoseasemotionalstatesit becomesafactualquestionwhetherparticularorganismsor machinesarein sucha stateor not: theanswerdependsonwhethertheindividual(a)hasaninformationprocessingarchitecturethatis capableofsupportingsuchstates,and(b) whetherthecomponentsof thearchitecturearein theappropriatefunctionalstatesto producethe precisely-definedsort of emotionthat is in question. (Someofthe conceptsmay be definedin relation to featuresof the environment. E.g. wanting to go upthe Eiffel Tower is a statethat dependson the existenceof the Eiffel Tower.) Having producedsuchprecisedefinitionsof variouskindsof architecture-basedmentalstateswecanformulateand,perhapsbegin to answer, new, moreprecise,questionsaboutwhich agentsarecapableof havingwhich sortsof emotions,experiences,thoughts,andsoon. Thereis thenno risk of beingboggeddown in endlessterminologicaldisputesor philosophicalargumentsat cross-purposes,asoftenhappenshow.

Of course,the fact that a questionis a factualonewith correctand incorrectanswersdoesnot imply that it is easyto determinethe answer, as the history of physicsshows very clearly.Sometimesthequestionhasto remainunanswereduntil new technologyis availableto probethesystemin greaterdepthandprecisionthanpreviously. Sometimesthe theoryhasto beextendedwith links to other theoriesbeforeobservation or measurementcan provide relevant evidence.Thispoint is well discussedin standardliteratureon thehistoryandphilosophyof science,e.g. inPopper(1934)andchapter2 of Sloman(1978).

8 Constraints on theorising

If wewishto gobeyondthestudyof sortsof informationprocessingarchitecturesthataretheoreti-cally possible,andattemptto describethearchitectureof aparticularindividualor thearchitecturestypical of membersof a certainbiologicalspecieswe find that it is extremelydifficult to infer thearchitectureof a machinethatwe have not designedourselvesif we do not have accessto designspecificationsusedin its production. This is analogousto the taskof decompilinglarge “legacysoftware”systems.

No amountof observationof theexternalbehaviour of any animalor machinecandeterminethe underlyingarchitecture,since in principle any lifelong set of behaviours can be producedby infinitely many different informationprocessingarchitectures,including totally unstructured,unintelligible,“flat”, multi-componentarchitectures,assuggestedin Figure1. Decompilinginfor-mationgleanedfrom invasiveor non-invasiveobservationof internalphysicalstructuresis just as

7

Figure1: Anunstructuredmess?Any observedbehaviour might be producedby an unintelligibly tangled and non-modulararchitecture. (Rectanglesrepresentinformation stores and buffers, ovals representprocessingunits,andarrowsrepresentflowof information,includingcontrol signals.)

hard,e.g. if we don’t evenknow at whatphysicallevel mostof thearchitectureis implemented.Do neuronsor moleculesdo mostof theinformationprocessing?

A commonwayof avoiding theseproblemsis to formulatetheoriesthataddressaverynarrowrangeof phenomenasoasto yield conditionalpredictionsthatcanbetested:if we do X to peoplein conditionsC, they will respondby doingY, etc. Theproblemis that limiting one’s theorisingto sucheasilytestablehypothesespreventsformulationof truly deepexplanatorytheories,suchasthosewhichhavebeenof mostprofoundimportancein physics.5

Thestudyof mind is farmorecomplex thanthestudyof physicsastherearesomany possibleinformation processingarchitecturessupportingdifferent collectionsof conceptsand differenttypesof lawsof behaviour. In generalit is notpossibleto formulateinterestingtestablehypothesesabouthow a particularsortof mind workswithout assuming(explicitly or implicitly) the typeofinformation processingarchitecturethat it uses. But decidingwhich architectureto proposeisverydifficult andis not in generalconstrainedby experimentalobservations,thoughthey certainlyprovidecluesandtests.

We can,however, constrainour theoriesby combininga numberof considerationswhich Ihave discusseda greaterlengthin (Sloman,1998b;Sloman,2000a),suchas: (1) trade-offs thatcan influenceevolutionarydevelopments,(2) what is known aboutour evolutionaryhistory, (3)what is known abouthumanand animal brainsand the effects of brain damage,(4) what wehave learntin AI aboutthescopeandlimitationsof variousinformationprocessingarchitectures,

5This topic wasdiscussedat greaterlengthin chapter2 of (Sloman,1978),which distinguishedthestudyof theform of theuniversefrom thestudyof its contents, includingregularitiesandcorrelations.Deepsciencerequirestheformer. Shallow scienceassumesa form, often implicitly, andtheninvestigatesa subsetof the contentscompatiblewith that form. A deeptheorymight statethat thereexist sub-atomicparticlesthathave variousmassesandelectricchargesthatcanbecombinedin differentways.A shallower theorymight relatethedeflectionof astreamof electronsto thestrengthof a magneticfield throughwhich they pass.

8

CentralProcessing

Perception Action

Figure2: A “vertical” divisioninto threetowersOrganismsandrobotsrequire perceptualmechanismsandactionmechanismsof varyingdegreesof sophistication,along with somepersistentinternal statewhich maybe modifiedover varioustime-scales.This leadsto Nilsson’s “triple tower” model(Nilsson,1998). Arrowsrepresentflowof informationof variouskindsincludingcontrol signals. Theboundariesbetweenthe “towers”neednotbeverysharp,especiallywhere there is rich two-wayinformationandcontrol flow. Laterweshowthateach pillar canbedividedhorizontally.

mechanismsand representations,(5) introspective evidence,suchasmy knowledgethat beforebuying ticketsI consideredandevaluatedalternative waysof travelling to the conferencewherethis paperwaspresented.Theseconstraintsareprior to thesortsof requirementsmorecommonlyfound in philosophyof sciencetexts, suchastherequirementof testability, or therequirementtofit statisticaldatabetterthanalternative theoriesthathavebeenproposed.

Although our theorieswill still remainconjecturalfor sometime to come,becauseof thecomplexity of humanmindsandbrains,we canat leasthopeto show that someconjecturesarebetterthanothers,if we take a broadenoughview of what needsto be explained. The next fewsectionsoutlinea two stageapproach.Thefirst stagecharacterisesa generalarchitecture-schemacalledCogAff whichspecifiesin broadoutlineavarietyof typesof functionalrolesfor mechanismsthat may occurwithin organismsor robotsof variouskinds. In the secondstagewe presentaninstanceH-Cogaff of this schemawhich we proposeas a first draft model of the informationprocessingarchitecturetypicalof humanminds.

The CogAff schemadefinesa framework of possibledesignsfor informationprocessingar-chitecturesfor organismsor machines.It is usefulfor thinking aboutbiologicalorganisms,but isnot intendedto cover all possibilities,asit saysnothingaboutmany of thearchitecturesdesignedby engineers,andit doesnot includedistributedmulti-agentsystems,thoughit couldspecifytheindividualsin suchasystem.

9 CogAff: an architectureschema

Nilsson(Nilsson,1998)proposedthat intelligent systemscanbeanalysedin termsof the “tripletower” modeldepictedin Figure2, whichapproximatelyseparatesperceptualmechanisms,central

9

processingmechanismsandactionmechanisms.He calls the centraltower the “model tower”,thoughthis labelmaybetoorestrictive for therangeof functionssketchedbelow. Thetriple towermodelis mainly a resultof functionalanalysiscombinedwith observationof existingorganisms.

Anotherbreakdown of informationprocessingfunctionality comesfrom both functionalandevolutionaryconsiderations.This is the triple layermodelsketchedin Figure3, anddiscussedatgreaterlengthin previouspapers(e.g. (Sloman,1997;Sloman,1999b;Sloman,1998a;Sloman,2000a;SlomanandLogan,1999;SlomanandLogan,2000)).Thesethreelevelsaredifferentfromthethreediscussedby Nilssonin chapter25of (Nilsson,1998),thoughthereis someoverlap.

If the threelayersandthe threetowersaresuperimposedasin Figure4 we arrive at a grid oftypesof architecturalcomponents,whereperceptualmechanismshaveseverallayerswith differentkinds of sophisticationrequiredto meetthe needsof the differentcentrallayers. Likewise theactionmechanismsmayhave differentlevelsof sophisticationsupportingdifferentsortsof func-tionality arisingoutof differentlevelsof centralprocessing.In thefigurewehavealsodepictedan“alarm” mechanism,which couldalsobethoughtof asmerelya partof thecentralreactive layer,receiving inputs from all over the architectureandsendingcontrol signalsto many partsof thesystem,in orderto achieverapidredirectionof internalandexternalprocessing.

The CogAff schemethusdepictedspecifiesa variety of componentswhich neednot all bepresentin a particularmachine,and which may be relatedin different ways, giving differentspecificarchitectures.For instanceaninsector simplerobotmight have anarchitectureincludingonly the reactive layersasin Figure5 whereassomeotheranimalsmight have both deliberativeandreactivemechanismsasin Figure6.

Moreover, very differentdesignsfollow from different functional relationsbetweencompo-nents. For example,we refer to an Omega architectureasonein which the informationflow isessentiallya pipelinewith informationcomingin at bottomleft, going up the centralcolumntosomehigh level decisionmakingsystemandthenflowing down the centreandout throughthebottomright, roughlywith theshapeof a Greek

�. For anexamplesee(Albus,1981). This has

somesimilaritieswith theContentionSchedulingmodelin (CooperandShallice,2000).Thesubsumptionarchitectureproposedby Brooks(Brooks,1986;Brooks,1991),canbeseen

as a variant in which there is only a reactive layer, containingseveral parallel pipelines,withinformationflowing from left to right within eachpipeline,but with factualinformationgoingupfrom lower levelsto higherlevelsandcontrolinformationgoingdown from higherlevelsto lowerlevels. A hybrid architecturesuchasFigure6 might includea reactive subsumptionlayer andadeliberative layer.

One of the key featuresthat gives the CogAff schemaits generalityis the possibility thatthe differentcomponents,insteadof forming partsof simple pipelines,can all be concurrentlyactiveandconcurrentlysendinginformationof variouskindsto arbitrarilymany othercomponents,allowing a wide varietyof feedbackmechanismsandtriggeringmechanisms.For instancea highlevel goal generatedwithin the deliberative or meta-managementlayer could sendinformationto perceptualmechanismsin orderto direct themphysicallyandalter their processing(Sloman,1989). Likewisedifferentsortsof centralprocessingat differentlevelsof abstractionmight sendsignalswith differentlevelsof abstractionto actionmechanisms.Sincemany suchthingscouldhappenconcurrentlywe infer a needfor arbitrationmechanisms.Onesuchis the attentionfilterwith dynamicallyvaryingfilter thresholdin Figure6. Oftenarchitecturesproposedwith diagramsthatlook superficiallysimilar turnout to beverydifferentwhenthedetailsarespecified,includingdetailssuchaspossibledirectionsof informationflow anddegreesof concurrency.

10

Meta-management(reflective processes)

(newest)

Deliberative reasoning("what if" mechanisms)

(older)

Reactive mechanisms(oldest)

Figure3: A “horizontal” divisioninto threelayersIt is now commonplacein AI to distinguishreactivemechanismsin which statesdetectedbysensors(whetherexternalor internal)immediatelytrigger responses(whetherexternalor internal)from deliberative mechanismsin which alternative possibilitiesfor action can be considered,categorised,evaluated,andselectedor rejected.Moregenerally a deliberativemechanismmaybecapableof “what if ” reasoningaboutthepastor futureor evenhowthepresentmighthavebeen.Thedepth,precisionandvalidity of such reasoningcanvary. A meta-managementlayer addstheability to monitor, evaluate, and to someextentcontrol processesoccurring within thesysteminsomethinglike theway thewholesystemobservesandactson the environment.Thetwo bottomlayersdiffer in that thesecondevolvedmuch later andrequiresa far moresophisticatedlong termmemoryandsymbolicreasoningcapabilitiesusinga shorttermre-usablememory. Thethird layermayhaveevolvedlater andrequiresexplicit useof conceptsreferringto statesof an informationprocessingarchitecture. Theearliestorganisms,likemostexistingorganisms,weretotally reactive.Deliberativeand meta-managementlayers evolvedlater. Adult humansappearto haveall threetypesof processing, which is probably rare amongother animals. The three layers operateconcurrently, and do not form a simpledominancehierarchy. As previously, arrows representflowof informationandcontrol, andboundariesneednotbesharpin all implementations.

10 Sketchof a theory of humans: H-Cogaff

Within thegeneralframeworkof theCogAff schemawehavedevelopedaparticularinstancewhichwe now call H-Cogaff, depictedin Figure7, anddiscussedin moredetail in earlierpaperse.g.(Sloman,2000a).Our conjectureis thattheinformationprocessingarchitectureof a normaladulthumanis somethinglike H-Cogaff (augmentedwith sub-mechanismsnot shown in the figure).This conjectureis basedon evidenceof many kinds from several disciplines,and the sortsofconstraintson evolvability, implementabilityand functionality mentionedabove. According tothis theory:

(a) Evolution, like engineers,found that (partly) modulardesignsareessentialfor defeatingcombinatoricsin thesearchfor solutionsto complex problems(with only 4,000,000,000yearsandonebiosphereonanearth-sizedplanetavailable).

(b) Humaninformationprocessingmakesuseof (at least)threedifferentconcurrentlyactive

11

ALARMS

CentralProcessing

Perception Action

Meta-management(reflective processes)

Deliberativereasoning

Reactive mechanisms

Figure4: TheCogAff Schema:pillars, layers andalarmsIf weconsidera systemin which boththedivisionbetweenperceptual,central andmotorsystemscan be made, andalso the divisionbetweenreactive, deliberativeandmeta-managementlayers,andif weassumethat theperceptualandmotorsystemsincludecomponentsrelatedto theneedsof all threecentral layers, thenwe havea threeby threegrid of architectural componentswithdifferentsortsof functionality. If someof the internal processingis slowrelativeto thespeedsatwhich thingshappenin theenvironment,thenit maybeusefulto haveinputsfrommanypartsofthesystemto a fastpatterndrivenreactive“alarm” mechanismthatcanredirectthewholesystem.Solidarrowsareasbefore. Theshadedarrowsrepresentinformationflowingto andfromthealarmmechanism.Thealarmmechanismbeingpurelyreactiveandpatterndrivenwill typicallybestupidandcapableof mistakes,but maybetrainable.

architecturallayers,a reactive layer, a deliberative layer, and a meta-managementlayer whichevolved at different times, which we sharewith other animalsto varying degrees,along withvariousadditionalsupportingmodulessuchas motive generators,“global alarm” mechanismsandlong termassociative storagemechanisms.The differentlayersandsupportingmechanismsmayhave evolvedfrom purelyreactive mechanismsby meansof thetypical evolutionarytrick ofmakinganothercopy of anexisting mechanismandthengraduallytransformingthe functionsofthenew copy. Thisalmostcertainlyhappenedseveraltimesin theevolutionof brains.

(c) Reactive systemsmaybevery complex, andpowerful, especiallyif internalreactionscanbechainedtogetherandcancausemodificationof internalstateswhich triggeror modulateotherreactions.I do not claim thatdeliberative or meta-managementmechanismsprovide behaviouralcapabilitiesthat could not in principle be provided by purely reactive mechanisms.RatherIhavearguedelsewherethatachieving thesamefunctionalityby purelyreactivemeanswouldhaverequireda far longerperiodof evolution with morevariedcircumstances,anda far largerbraintostoreall thepreviously evolvedreactive behaviours. Thetime andbrainsizerequiredfor a purelyreactive human-like systemareprobablytoo large to fit into thephysicaluniverse.Somepeoplewho argue in favour of purely reactive systemsdo not considerthe trade-offs involved in theseresourceissues.Merely showing that in principle reactive systemssuffice provesnothingaboutwhatcanwork in practice.

(d) Reactive, deliberative andreflective layerssupportdifferentclassesof emotionsfound in

12

humansandotheranimals,including theprimaryandsecondaryemotionsdiscussedby DamasioandPicard(Damasio,1994;Picard,1997),andthetertiaryemotionsI havediscussedin comment-ing on their work (Sloman,1998a;Sloman,1999a).

1. thereactive layer, includingaglobalalarmmechanism,accountsfor primary emotions(e.g.beingstartled,frozenwith terror, sexually aroused);

2. the deliberative layer supportssecondaryemotionslike apprehensionandrelief which re-quire“what if ” reasoningabilities(thesearesemanticallyrich emotions);

3. themeta-management(reflective)layersupportsnotonlycontrolof thoughtandattentionbutalsolossof suchcontrol,asfoundin typically humantertiary emotionssuchasinfatuation,humiliation, thrilled anticipationof a futureevent. (This layer is alsocrucial to absorptionof acultureandvariouskindsof mathematical,philosophicalandscientificthinking.)

All the layersare subjectto interferencefrom the othersand from oneor more fast but stupidpartly trainable“global alarm” mechanisms(e.g. spinalreflexesof varioussorts,thebrainstem,thelimbic systemincludingtheamygdala,etc.)

(e)A morefine-grainedanalysisof typesof processesthatwetendtocall “emotions”in humanswould show that theabove three-foldclassificationinto primary, secondaryandtertiaryemotionsis somewhatsuperficial.For instance,therearedifferentwaysemotionscandevelopovertime,andthethree-folddistinctiondoesnotsayanythingaboutthat.A shortflashof angeror embarrassmentwhichquickly passesis verydifferentfrom longtermbroodingorobsessivejealousyorhumiliationwhichgraduallycoloursmoreandmoreof anindividual’smentallife.

(f) Perceptualand motor systemsare also layered: the different layersevolved at differenttimes,actconcurrently, andhave differentrelationshipsto the “central” layers. E.g. deliberativemechanismsmake useof high level characterisationsof perceivedstates,e.g. seeinga bridgeas“rickety” or an ornamentas “fragile”. Using someof Gibson’s ideas,this canbe describedasperceptionof abstractaffordances.

(g) Analysingwaysin which componentsof suchanarchitecturemight bootstrapthemselves,develop,reorganisethemselves,acquireandstoreinformation,or gowrong,will providefar richertheoriesof learninganddevelopmentthaneverbefore.

(h) Thethreelayersaccountfor differentcognitiveandaffectivestates,aswell asdifferentpos-sibleeffectsof braindamage,andotherabnormalities.For instance,someaspectsof autismseemto involvemalfunctioningor non-functioninghigherlevel perceptualmechanisms(assuggestedin(Sloman,1989)).

(i) A multi-layeredarchitectureof thesortproposedcouldgiverobotsvariouskindsof human-like mentalstatesand processes,including qualia arising out of inward focusedattention. Assciencefiction writers have noted,this might leadsomerobotsto re-discover philosophicalcon-fusionsaboutconsciousness.Softwareagentscouldhave similar capabilities.However, detaileddifferencesin physicalembodimentsandvirtual machinearchitecturescouldentailmany kindsofminor differencesin thementalstatesof which they arecapable.This is no differentin principlefrom the fact that mentalstatespossiblefor adultsandchildrenare different,or for malesandfemales,or humansandcats.

Many doubttheseclaimsaboutrobotsbecausethey seethe limitationsof existing computer-basedmachinesandsoftwaresystemsandcannotimagineany waysof overcomingtheselimita-tions.They donot realisethatwearestill in theearlystagesof learninghow to designinformation

13

ALARMS

perception action

THE ENVIRONMENT

REACTIVE PROCESSES

Figure5: A reactivesystemwith global alarms.Somethinglike this mightbeanarchitecture for a fairly sophisticatedinsect.

processingsystems.(Claiming thatcomputerswill beever morepowerful is not enoughto allaythesedoubts:wealsoneeddeepanalysisof theconceptsusedto expressthedoubts.)

11 Alter nativesin designspace

Althoughtheabove theoryincludesasketchof anarchitecturefor human-like intelligentsystems,thereis no suggestionthat this is theonly sortof intelligence.‘Intelligence’, like ‘emotion’, is aclusterconcept, referringto avariableclusterof capabilities,andadmittingawidevarietyof typesof instances,with no sharpboundaries.In particular, animals(andperhapshumans)exist withdifferentsubsetsof the full arrayof mechanismsdescribedabove, andwithin thosemechanismsconsiderablevariationis possible.

For example,many insectsappearto be capableof remarkableachievementsbasedentirelyin complex collectionsof purelyreactive mechanisms,suchastermitesconstructingtheir “cathe-drals”,with air conditioning,nurserychambersandotherextraordinaryfeatures.

So I am not denying that therecanbe organisms(androbots)which arepurely reactive, orwhichcombinea reactivemechanismwith aseparateglobalalarmsystem,asin Figure5.

Moresophisticatedorganismshavebothareactiveandadeliberativelayer, providing “what if ”reasoningcapabilities,asillustratedin Figure6. Suchmechanismsprovide theability to constructspecificationsof hypotheticalpastor future situationsandto reasonaboutthem. Many writers,includingCraik(Craik,1943)aslongagoas1943,havepointedoutthatsuchabilitiesmayincreasebiologicalfitness.

It seemsthat someotheranimalsbesideshumanshave deliberative mechanismsthoughtheyvaryenormouslyin their richnessandflexibility . For instance,how effectivesuchcapabilitiesare,will dependon a numberof factorsincluding the type andsizeof re-usableshort term workingmemory, the type of representationalmechanismsavailable, the type and size of the trainableassociativememorywhichcanstoregeneralisationsabouttheenvironment,andsoon.

The deliberative layer might have evolved asa resultof a mutationwhich at first led to thecopying of a trainableassociative memoryin a purely reactive system.After that, the new copy

14

ALARMS

Variablethresholdattentionfilter

Motiveactivation

Longtermmemory

perception action

THE ENVIRONMENT

REACTIVE PROCESSES

DELIBERATIVE PROCESSES

(Planning, deciding,scheduling, etc.)

Figure6: A hybridarchitecturewith globalalarms.In a hybrid reactiveanddeliberativesystem,it maybenecessaryto havean“attention filter” withdynamicallyvarying filter thresholdto protectthe resource-limiteddeliberativemechanismfrombeinginterruptedtoooftenduringurgentandintricate tasks.Howeveranalarmsystemor intenseperceptualinputsmaybecapableof exceedingthefilter threshold.

might have graduallyevolved, along with other mechanisms,to provide the ability to answerquestionsabout“what would happenif ” insteadof “how shall I reactnow”. Making gooduseof sucha “what if ” reasoningcapability requiresbeing able to storegeneralisationsabout theenvironmentat anappropriatelevel of abstractionto allow extrapolationbeyondobservedcases.Thisin turncouldgenerateevolutionarypressuretowardsperceptualsystemswhichincludehigherlevel abstractionmechanisms.All this is, of course,highly speculative, andneedsto be testedempirically, thoughit is consistentbothwith what is known aboutevolutionarymechanismsandwith theat leastpartlymodularstructureof thebrain.

Moregenerally, within thisframeworkwecanseeaneedfor ageneralisationof Gibson’stheoryof perceptualaffordances(Gibson,1986) (contrastedwith Marr’s theory of vision in (Sloman,1989))to accommodatedifferentperceptualaffordancesfor differentcomponentsin themorecen-tral processingmechanisms.This requiresthesharingof sensoryresourcesbetweenconcurrentlyactivesubsystems,andcangenerateconflicts,asdiscussedin (Sloman,1993a).

Deliberativecapabilitiesbringtheirown problems,suchashow they shouldbecontrolled,howdifferentdeliberativestrategiesshouldbeselectedor interrupted,how they shouldbeevaluatedandmodified. For this purposeandothers,it seemsthatanevensmallersubsetof animals,includinghumans,have evolveda third architecturallayerproviding theability to directattentioninwardlyandto monitor, evaluate,andin somecasesmodify what is happeninginternally. Luc Beaudoinfirst drew my attentionto someaspectsof theneedfor this layer, andcalledit meta-management.Someof therequirementswereanalysedin hisPhDthesis(Beaudoin,1994).

Earlierpapers(e.g.(Wright etal.,1996))havediscussedsomeof thewaysin whichthis theory

15

ALARMS

Variablethresholdattentionfilter

META-MANAGEMENT

processes(reflective)

THE ENVIRONMENT

REACTIVE PROCESSES

Motiveactivation

Longtermmemory

Personae

DELIBERATIVE PROCESSES

(Planning, deciding,‘What if’ reasoning)

perception action

Figure7: H-Cogaff – a threelayerarchitecture.Themeta-managementlayer providesthe ability to attendto, monitor, evaluate, and sometimeschange internal processesand strategies usedfor internal processes.However, all the layersand the alarm system(s)operate concurrently, and noneis in total control. A collectionof highlevel culturally determined“personae”maybeavailable, turnedon andoff by differentcontextsand causingglobal featuresof the behaviourto change, e.g. switching from bullying to servilebehaviour. Note that someof the divisionsbetweenlayers are a matterof taste: someauthorse.g. (Davis,1996)preferto separateout reflexesfromthereactivelayer, andsomewouldprefertoseparateoutsomeof thehigh level functionalityof themeta-managementlayer.

accountsfor distinctively humanemotionssuchasgrief, infatuation,excitedanticipation,humili-ation,involving partiallossof controlof attention.Weusedto call theseemotions“perturbances”,but now refer to themastertiary emotions,to distinguishthemfrom the primary andsecondaryemotionsdiscussedby Damasioandothers.

Sincethesetertiaryemotions(perturbances)involve lossof controlof attention,andyou can-not losewhat you have not got, only an organismwhich hassomethinglike meta-managementcapabilitiescangetinto suchstates.Thisdoesnot meanthatall humanshave this capability. Newborninfants,peoplewith degenerativebraindiseaseor braindamage,maylacksuchcapabilities.

12 Ar eemotionsrequired for intelligence?

It is clearthatlocal reflexesandglobalalarmmechanismscanbeusefulin organismsor machineswhichsometimesrequirevery rapidreactionsto occurfasterthannormalprocessesof perception,

16

reasoning,deliberation,andplanning.Suchreactionscanproducesimpleandobviouseffectssuchasfreezing,fleeing,producingaggressivesoundsor postures,pouncingon prey, sexual responses,andmoresubtleinternaleffectssuchasattentionswitchingand“arousal” which might involvedifferent kinds of information processing. Becausethesereactionsoften needto happenveryquickly they canbetriggeredby a relatively stupid,but trainable,patternrecognitionsystem.

Many humanemotionsseemto involve the operationof suchmechanisms.Theseandotheremotionsareconnectedwith resource-limitsin more“intelligent” subsystems.If thosesystemscould operatefaster, and with more completeinformation, it would not be necessaryfor more“stupid” mechanismsto overridethem.

Damasio(in (Damasio,1994))pointedout that certainkinds of frontal lobe damagecansi-multaneouslyremovetheability to havecertainclassesof emotionsandalsounderminetheabilityto achieve high level control of thoughtprocessesrequiredfor successfulmanagementof one’slife. Pendingfurther investigationof details,this givessomesupportfor theclaim that thereareclassesof emotions,referredto as“tertiary emotions”above, which dependon mechanismsthatareconcernedwith high level managementof mentalprocesses.

Damasioarguedfrom this thatemotionsarea requirementfor intelligence,andsincethentheargumenthasbeenrepeatedmany times: it hasbecomea sortof meme. However, the reasoningis fallacious,asI have arguedin (Sloman,1998a;Sloman,1999a).Thebraindamagein questionmight merelyhave disabledsomemechanismsinvolving control of attention,requiredboth fortertiary emotionsand for managementof thought processes.It doesn’t follow that emotionssomehow contributeto intelligence:ratherthey area side-effect of mechanismsthatarerequiredfor otherreasons,e.g.in orderto overcomeresourcelimits asexplainedabove.

Here’s an exampleof similarly fallaciousreasoningthatnobodywould find convincing. Op-eratingsystemswhich supportmultiple concurrentprocessesareextremelyuseful,but they cansometimesget into a statewherethey are “thrashing”, i.e. spendingmore time swappingandpagingthandoingusefulwork. If somedamageoccurredwhich preventedmorethanoneprocessrunning at a time that would prevent the thrashing,and remove the useful benefitsof multi-processing.It doesn’t follow thata thrashingmechanismsis requiredto produceusefuloperatingsystems.In fact, by addingmorememoryandCPU power, thrashingcanbe reducedandper-formanceenhanced.Likewise, it is possiblefor maturehumansto learnstrategiesfor avoidingemotions,andthiscanoftenimprovethequalityof their livesandthelivesof peoplethey livewithor work with.

I amnot arguingthatall emotionsareundesirableor dysfunctional.Therearemany emotionsthathaveanimportantbiologicalrole(e.g.sexualpassion,andaggressionin defendinganest),andsomeemotionsthat humansvaluehighly, including aestheticemotionsandthe joy of discovery.I alsoaccept,asmostAI researchershave acceptedover many years,that therearemany purelyintellectualproblemswhich requireexplorationof searchspacesthat aretoo large for complete,systematic,analysis.Theuseof heuristicpattern-recognitionmechanismsis oftenusefulin suchcases,to selectavenuesto explore and to redirectprocessing. But they can operatewithoutgeneratingany emotions.

17

13 Conclusion

Thispaperis asnapshotof anongoinglong termmulti-disciplinaryresearchprojectattemptingtounderstandthenatureof thehumanmindandhow wefit into a largerspaceof possibledesignsforbiologicalorganismsandartificial agentsof many kinds.

The ideashave many links with previous work by others. Someaspectsof the methodology(definingan architecture-basedcollectionof conceptsandtheninvestigatingtheir relationswiththosein any particularlanguage)havemuchin commonwith thestrategy in (Ortony et al., 1988).Besidesthestrongandobviousconnectionswith work of Simon,Gibson,andNilsson’sideascitedpreviously, therearealsolinks with work of Dennett,Minsky, Picard,Damasioandmany others,not all listed in the bibliography. . However thereis no room in this paperfor a full survey ofsimilaritiesanddifferencesbetweenthevarioustheories.

Therehasalsonot beenspaceto exploreall the implicationsof the ideaspresentedhere(e.g.showing how they canaccommodatethespaceof possibilitiespresentedin (Ortony et al., 1988)),but onething is very clear: we area long way from implementingartificial systemswith the fullrichnessandcomplexity of systemscontainingall thetypesof mechanismsdefinedby theCogAffschemeor theH-Cogaff architecture.

Therearemany gapsin whatcurrentAI systemscando, insofarasthey arethoughtof asstepstowardsmodellinghumanintelligence,andbeyond.ExistingAI systemsdonotyethavewhateverit takes to enjoy or dislike doing something. They do not really want to do somethingor careaboutwhetherit succeedsor fails, even thoughthey may be programmedto give the superficialappearanceof wantingandcaring,or feelinghappy or sad.Animal-likewanting,caring,enjoying,suffering,etc.seemto requiretypesof architectureswhichhavenotyet beenanalysed.

Simulateddesiresand emotionsrepresentedby valuesfor global variables(e.g. degreeof“fear”) or simpleentriesin databaseslinkedto condition-actionrulesmaygive theappearanceofemotion,but fail to addressthe way semanticallyrich emotionsemerge from interactionswithina complex architecture,andfail to distinguishdifferentsortsof emotionsarisingout of differenttypesof processingmechanismswithin anintegratedarchitecture.

CurrentAI modelsof otheranimalabilities arealso limited: for example,visual andmotorcapabilitiesof currentartificial systemsare nowherenearthoseof a squirrel, monkey or nest-building bird. To understandanimalcomprehensionof spaceandmotionwe mayneedto under-standthedifferencesbetweenprecocialspeciesborn or hatchedwith considerableindependence(chickens,deer)andaltricial specieswhich startutterly helpless(eagles,cats,apes).Perhapsthebootstrappingof visuo-motorcontrolarchitecturesin the latteryieldsa far deepergraspof spaceandmotionthanevolutioncouldhavepre-programmedvia DNA. Theprecocialspeciesmayhavemuchsimplervisualcapabilities,largely geneticallydetermined.

Therearemany issuesthatarestill unclear, andavastnumberof remainingresearchtopics.Inparticularit is not clearhow muchof this is relevant to the designof softwareagentsinhabitingvirtual machineenvironmentsonly, and lacking physicalbodies. Many of the humanreactivemechanismsandsomeof their motivatorsandemotionalresponsesareclosely linked to bodilymechanismsandfunctions.E.g. if you don’t have a bodyyou will never accidentallystepon anunstablerock,andyouwill notneedan“alarm” mechanismthatdetectsthatyouareaboutto loseyour balanceandtriggerscorrective action,includingcausinga surgeof adrenalinto bepumpedaroundyour body.

Neverthelesseventscanmove fastin a virtual machineworld (asmany systemadministrators

18

fightingmaliciousintruderswill confirm)andevenpuresoftwareagentsmayneedreactivemech-anisms. Still, it is likely that the combinationsrequiredfor softwareagentsmay includesomearchitecturesnever foundin agentswith physicalbodies.Whetherthereverseis thecasedependson whetherall sortsof physicalbodiesandphysicalenvironmentscan,in principle,besimulatedonsufficiently powerful physicallyimplementedcomputers:anopenquestion.

Artificial agentswhich do not shareour deepgraspof spatialstructureandmotion will belimited in their ability to communicatewith us. However, it is not obvious that in orderto sharethis knowledgesuchagentsmusthave similar bodiesandprocessingarchitectures.For instance,peoplewho have neverwantedto kill someone,mayneverthelessunderstandsomeof thethoughtprocessesof amurderer(a factonwhichthesuccessof many novelsandplaysdepends).Similarlysomeonewho hasbeenblind from birth canunderstanda greatdealaboutvisual capabilitiesofsightedpeople,for instance,thatcoloursareextendedpropertiesof 2-D surfaces,somewhat liketactiletextures.

Soit remainspossiblethatsomesoftwareagentswhichareveryunlikeuswill beableto engagein rich communicationwith us,thoughthedetailedrequirementsfor this arestill not clear.

And of course,in the meantime,teachersanddesignersof computergamescanbuild manyentertainingor didactic,shallow simulationswhich lack mostof thefeaturesdiscussedhere.Thatis fine,aslongasthey takecarehow they describewhatthey havedone.

Acknowledgements

I amgratefulfor usefulcritical commentsby anonymousrefereesandregretthatit wasnotpossiblefor me to follow up all of their suggestions.I am grateful to ElisabethAndre for encouragingsubmissionof this paperto thejournal.Many colleaguesandstudentshavehelpedwith thedevel-opmentof theseideas,mostrecentlyBrian Logan,Steve Allen, CatrionaKennedyandMatthiasScheutz.Work of the CognitionandAffect group is presentedat length in paperslisted in thebibliographyandat thewebsitehttp://www.cs.bham.ac.uk/research/cogaff/Theprojectis summarisedin:http://www.cs.bham.ac.uk/˜axs/cogaff.htmlOursoftwaretoolsincludingtheSimAgenttoolkit (Pop-11basedcodeanddocumentation)canbefoundaspartof theFreePoplogFTPsite:http://www.cs.bham.ac.uk/research/poplog/freepoplog.htmlSimAgentis not committedto any particulararchitecture:ratherit supportsexplorationof a widerangeof architecturesin singleor multiple simulatedagents.It alsoincludesteachingmaterialswhich canbe usedto introducestudentsto someof theseideasthroughpracticalexperienceofdesigningand modifying simple agentsin simulatedworlds, including a sheepand sheep-dogscenario,andvariouskindsof obstacleavoiding andgoalseekingagents.

References

Albus,J.(1981).Brains,BehaviourandRobotics. ByteBooks,McGraw Hill, Peterborough,N.H.

Bates,J., Loyall, A. B., andReilly, W. S. (1991). Broadagents.In AAAI spring symposiumon

19

integratedintelligentarchitectures. AmericanAssociationfor Artificial Intelligence. (Repr.in SIGART BULLETIN, 2(4),Aug. 1991,pp.38–40).

Beaudoin,L. (1994). Goal processingin autonomousagents. PhD thesis,Schoolof ComputerScience,The Universityof Birmingham. (Availableat http://www.cs.bham.ac.uk/research/cogaff/).

Beaudoin,L. andSloman,A. (1993). A studyof motive processingandattention. In Sloman,A., Hogg,D., Humphreys,G., Partridge,D., andRamsay, A., editors,Prospectsfor ArtificialIntelligence, pages229–238.IOS Press,Amsterdam.

Brooks,R. (1986).A robustlayeredcontrolsystemfor a mobilerobot. IEEE Journal of RoboticsandAutomation, RA-2:14–23.1.

Brooks,R. A. (1991). Intelligencewithout representation.Artificial Intelligence, 47:139–159.

Cooper, R. andShallice,T. (2000). Contentionschedulingandthe control of routineactivities.CognitiveNeuropsychology, 17(4):297–338.

Craik,K. (1943).TheNatureof Explanation. CambridgeUniversityPress,London,New York.

Damasio,A. (1994). Descartes’Error, EmotionReasonand theHumanBrain. Grosset/PutnamBooks,New York.

Davis,D. N. (1996).Reactiveandmotivationalagents:Towardsacollectiveminder. In Mueller, J.,Wooldridge,M., andJennings,N., editors,IntelligentAgentsIII — Proceedingsof theThirdInternationalWorkshopon AgentTheories,Architectures,andLanguages. Springer-Verlag.

Dennett,D. (1996).Kindsof minds:towardsan understandingof consciousness. WeidenfeldandNicholson,London.

Elliot, C. (1998).Huntingfor theHoly Grailwith ‘EmotionallyIntelligent’Virtual Actors.SIGARTBulletin, 9:20–28.1.

Gibson,J.(1986).TheEcological Approach to VisualPerception. LawrenceEarlbaumAssociates.(originally publishedin 1979).

Goleman,D. (1996). EmotionalIntelligence: Why It Can Matter More than IQ. BloomsburyPublishing,London.

Hatano,G., Okada,N., andTanabe,H., editors(2000).AffectiveMinds. Elsevier, Amsterdam.

LeDoux,J.E. (1996).TheEmotionalBrain. Simon& Schuster, New York.

McDermott,D. (1981). Artificial intelligencemeetsnaturalstupidity. In Haugeland,J., editor,Mind Design. MIT Press,Cambridge,MA.

Minsky, M. L. (1987).TheSocietyof Mind. William HeinemannLtd., London.

Nilsson,N. (1998).Artificial Intelligence:A New Synthesis. MorganKaufmann,SanFrancisco.

20

Oatley, K. andJenkins,J. (1996).UnderstandingEmotions. Blackwell,Oxford.

Ortony, A., Clore,G.,andCollins,A. (1988).TheCognitiveStructureof theEmotions. CambridgeUniversityPress,New York.

Picard,R. (1997).AffectiveComputing. MIT Press,Cambridge,Mass,London,England.

Popper, K. (1934).Thelogic of scientificdiscovery. Routledge,London.

Reilly, W. S.N. (1996).BelievableSocialandEmotionalAgents. PhDthesis,Schoolof ComputerScience,Carnegie Mellon University, Pittsburgh,PA. TechnicalReportCMU-CS-96-138.

Ryle,G. (1949).TheConceptof Mind. Hutchinson.

Simon,H. A. (1967). Motivationalandemotionalcontrolsof cognition. Reprintedin ModelsofThought,YaleUniversityPress,29–38,1979.

Sloman,A. (1978). TheComputerRevolution in Philosophy. HarvesterPress(andHumanitiesPress),Hassocks,Sussex.

Sloman,A. (1989). On designinga visualsystem(Towardsa Gibsoniancomputationalmodelofvision). Journal of ExperimentalandTheoretical AI, 1(4):289–337.

Sloman,A. (1992). Prolegomenato a theory of communicationand affect. In Ortony, A.,Slack,J.,andStock,O., editors,Communicationfroman Artificial IntelligencePerspective:TheoreticalandAppliedIssues, pages229–260.Springer, Heidelberg, Germany.

Sloman,A. (1993a). The mind asa control system. In Hookway, C. andPeterson,D., editors,PhilosophyandtheCognitiveSciences, pages69–110.CambridgeUniversityPress.

Sloman,A. (1993b). Prospectsfor AI as the generalscienceof intelligence. In Sloman,A.,Hogg, D., Humphreys, G., Partridge,D., andRamsay, A., editors,Prospectsfor ArtificialIntelligence, pages1–10.IOSPress,Amsterdam.

Sloman,A. (1994).Explorationsin designspace.In Cohn,A., editor, Proceedings11thEuropeanConferenceon AI, Amsterdam,August1994, pages578–582,Chichester. JohnWiley.

Sloman,A. (1997). What sort of control systemis able to have a personality. In Trappl, R.and Petta,P., editors, Creating Personalitiesfor SyntheticActors: Towards AutonomousPersonalityAgents, pages166–208.Springer(LectureNotesin AI), Berlin.

Sloman, A. (1998a). Damasio, Descartes,alarms and meta-management. In ProceedingsInternational Conferenceon Systems,Man, and Cybernetics(SMC98),SanDiego, pages2652–7.IEEE.

Sloman, A. (1998b). The “semantics” of evolution: Trajectoriesand trade-offs in designspaceand niche space. In Coelho, H., editor, Progress in Artificial Intelligence, 6thIberoamericanConferenceon AI (IBERAMIA), pages27–38.Springer, Lecture Notes inArtificial Intelligence,Lisbon.

21

Sloman,A. (1999a). Review of AffectiveComputingby R.W. Picard,1997. TheAI Magazine,20(1):127–133.

Sloman,A. (1999b).Whatsortof architectureis requiredfor ahuman-likeagent?In Wooldridge,M. andRao,A., editors,Foundationsof RationalAgency, pages35–52.Kluwer Academic,Dordrecht.

Sloman,A. (2000a).Architecturalrequirementsfor human-likeagentsbothnaturalandartificial.(whatsortsof machinescanlove?). In Dautenhahn,K., editor, HumanCognition AndSocialAgent Technology, Advancesin ConsciousnessResearch,pages163–195.JohnBenjamins,Amsterdam.

Sloman,A. (2000b). Interactingtrajectoriesin designspaceand niche space: A philosopherspeculatesaboutevolution. In et al., M., editor, Parallel ProblemSolvingfrom Nature –PPSNVI, LectureNotesin ComputerScience,No 1917,pages3–16,Berlin.Springer-Verlag.

Sloman,A. andCroucher, M. (1981). Why robotswill have emotions. In Proc 7th Int. JointConferenceon AI, pages197–202,Vancouver.

Sloman,A. andLogan,B. (1999). Building cognitively rich agentsusingthe Sim agenttoolkit.Communicationsof theAssociationfor ComputingMachinery, 42(3):71–77.

Sloman,A. andLogan,B. (2000). Evolvablearchitecturesfor human-like minds. In Hatano,G.,Okada,N., andTanabe,H., editors,AffectiveMinds, pages169–181.Elsevier, Amsterdam.

Wright, I., Sloman,A., andBeaudoin,L. (1996). Towardsa design-basedanalysisof emotionalepisodes.PhilosophyPsychiatry andPsychology, 3(2):101–126.Repr. in R.L.Chrisley (Ed.),Artificial Intelligence: Critical Conceptsin Cognitive Science, Vol IV, Routledge,London,2000.

22