From: AAAI Technical Report FS-97-03. Compilation copyright © 1997, AAAI (www.aaai.org). All rights reserved. DiagrammaticRepresentationand Reasoning: SomeDistinctions B. Chandrasekaran Departmentof Computer& InformationScience TheOhioState University Columbus,OH43210 Email:chandra@cis.ohio-state.edu 1. Introduction When I co-edited the book, "Diagrammatic Reasoning,"[5] I often pnzzled overthe fact that rather differentsorts of things were dubbeddiagrammaticreasoning by different authors.I then spentsometime trying to categorizethe differenttypes of use of diagram.~ andthe different tasks that they served.This shortarticle is the result of that study.The distinctionsI makehere applynot simplyto socalled diagrammatic reasoning,but the larger set of activities that havebeencalled spatial and visual reasoning. 2. Representation Let us fLrStconsiderthe notionof representationitself. Oneof the striking things about the current wonkin cognitiveneuruscience is the degreeto whichit is impessiblefor the researchersto avoidtalking aboutinformation beingrepresented,projected,fed backetc. Information aboutobjectsis said to be representedin this area of the brain, aboutshape in that area, andso on. Thus,antirepresentationalism is not a practical optionany longer, informationand representationare an inevitable part of the neededvocabulary.Onthe other hand, buyinginto informationbeing representeddoes not necessarilyimplyhavingto buy into "informationprocessing"or "representationprocessing."Amechanicalcoinsorter that operatesby passingthe coinsthrough levers andholesof differentsizes canbe analyzedin termsof its usinginformationabout weightsanddiametersof the coins, but saying that it performsinformationprocessingdoesnot agreewith everyone’sintuitions. Moredirectly relevant to our topic, Paul Churchland has describedneural structuresthat are involvedin the visuo-motor coordinationof a frog’s catching of its prey, a fly. Heshowsthat, in a suitably transformedspace, the solution to the conlml problemcan be expressedas a linear relationship betweenthe (trart~formedvaluesOf) the location of the prey anda motorcontrol parameter.In the frog’s neural structure, visual informationabout the prey is representedin onemurallayer, and the motorcontrol informationin anotherlayer directly belowit. Thetwolayers are so connectedthat the prey location information directly sets the motorcontrol parameter.Again~ information of varioustypes is clearly represented,but whetherthe best wayto describe whathappensto it as "processing"is unclear. At least amongAI people,talk of informationrepresentationsoonleads to talk of representation processingwhichsoon leads to talk ofreasoning.Conversely, this also sets upa counter-rhetoricof"no representation"[1]. With referenceto ourtopicof interest, not all spatial representationin the brain istobe takenas grist for somespatial reasoningmill if reasoningto be takenas being performed explicitly. Onthe other hand knowledgeand inference are appropriateterms if only a Knowledge Level account[2] is intended. 3. Visual versus Spatial versus Diagrammatic Whilemanypeople makedistinctions between "visual," "spatial," and"diagrammatic" representations,there doesnot seemto universal agreementon the meaningof these terms. Glasgow[3] distinguishes between"visual" and "spatial"byusingthe formerto refer to uninterpretedpixel-like representationsandthe latter to refer to representations in whichobjects andtheir mutualspatial relationsare indicated. Butmanyother peopleseemto use "visual" to denoterepresentations,interpretedor not, arising fromthe visual modality,while"spatial" is used to denoteknowledge of space, its occupant~and their mutualspatial relations, leavingopenwhich modalityor modalities- visual, auditor, kinestheticor haptic- that suppliedthe information.Diagram~and visual representationsin generalcan also havecolor andtextureas part of the representational repertoire, whilespatial representationsdonot seemto involvethese properties. Motionis potentiallypart of all three: diagramscanuse animation,andvisual andspatial representations can involve sequencesof images. Thereseems to be muchmoreof a generalagreement,at least onan intuitive level, onthe notionof diagrammatic representation.It is takento be a formof spatial representation,explicitly constructedandintendedto be visually processed,containingelementsthat havea conventionalsemantics,displayingthe spatial relations amongthe elements. Diagrammatic representationsmayalso haveelements,such as labels or other annotations,intendedto be interpreted as non-spatial, andelementswith visual properties,suchas color andtexture, that are usedas part of the representational vocabnlary.Iwasaki[16] proposesthat one of their importantpropertiesis that they represent abstractions,i.e., theyabstract out someof the informationandrepresent other information.Of course,diagramsalso havedetails that are not to be taken seriously. Theconventionalityin dementsemanticsprovidesclues to the consumer of the representationabout which aspectsof the diagramare to be takenseriously andwhichare to be takenas incidentalto the representation. Generalizing this to use of spatial information in humanor machinecognition, the three types of spatial informationprocessingare predicate extractionandprojection, reasoning,and simulation.(Thesethree types are in additionto algorithm~ that use spatial information to computesomethingof interest andthat haveno special standingas modelof anypart of spatial cognition.) 4.1. PredicateExtractionand Projection Givena Sl~_tial representation,a rmmber of facts corresponding to the representationcan be asserted;e.g., objectAis in the picture,object Ais in frontof objectB, Ataller than B, andso on. Theset of spatial predicatesis open-ended, and domain-and task-dependent. For example, in this sessionat the workshop, the paperby Epsteinand Gelfanddiscusses a game-playing programthat constructs various predicatesgiven a boardrepresentation.Thesepredicates are intendedto capturepotentiallyrelevantfacts aboutthe configuration-However, by definition, all the spatial predicatesmustbe comp~_a_oble fromthe spatial informationin the representation.Manyso-calledspatial or diagrammatic reasoningprogram.~ in fact focus on predicateextraction,i.e., computing a given set of predicatesfroma spatial representation. Thereis a complementary process, predicate projection, whichhas not received much attentionin the literature. Givena set of spatial predicates pertaining to a situation, this projectionprocesscreates a spatial or diagrammatic represera_otion,or modifiesan existing one, so as to matchthe assertions. What spatial predicatesare usefulfor whatta~qk~q and howto computethemefficiently are reviewedin somedetail in Chapter6 of [6]. 4. Reasoning,PredicateExtraction andProjection, andPrediction by Simulation Theword"reasoning"is often usedas a generic equivalentof compm_ation - "geometric reasoning"or "spatial reasoning"program.q often turn out to be programsthat perform computations involvinggeometricor spatial information.Thereis, however,a narroweruse of the term"reasoning,"In this senseof reasoning,the agentstarts with somegiven assertions, andmakesadditionalassertions as inferences,usingrules of inference.For example, whenBarwiseand Etchemendy [4] speakof diagrammatic reasoning,they intend for the diagramsto be assertions of somefacts or hypotheses,andthe goalis to arrive at inferences that satisfy the standardnotionsof valid inference. Goingbackto the motivationfor this paper,a perusalof the literature led meto the conclusion that that at least three distinct typesof use of diagrammatic informationare generally includedin the term"diagrammatic reasoning." 4.2. Reasoning Asmentioned, sp,9_riO1reasoningis makingtrustworthyinferenceswith spatial or diagrammaticrepresentations. Asimple model of spatial reasoningis onein whichthe agent starts witha spatial representation, extracts some predicatesfromit, and, usingother axiomsand rules of inference,proceedsto makeinferences. Statedthis way,there is nothingto distinguish the actual processof spatial reasoningfromany other formof reasoning,exceptfor the facts that the problem gets its start fromthe spa!i_al representationandgenerationof initial predicates requiresuse of a predicateextractionmodule. 2 This is what has often p.~led manypeople who havenot beenenthusedabout special claimsfor diagrammatic reasoning. However,in fact, there are a munber of propertiesof spatial reasoning that justify special treatment.Spatial representationsare not of generalquantified propositions,but of concreteinstances. Nevertheless,weseem,in somecases, to be able to makecorrect inferencesabout general situations. Studyingthe conditionsunderwhich suchgeneralinferencesare supportedis a major concernin the logic of spatial reasoning. Second,spatial reasoningoften involvesspecial spatially sensitive axiomsandrules of inference (suchas the transitivityof right-of, or inside-of predicates).Studyof spatial reasoningis an opportunityto focus on suchmodality-specific rules of inference.Thirdly,often, spa__tial reasoningdoesnot simplyinvolvejust extracting predicates and then proceedingwith reasoningas usual, but insteadinvolvesprojectiononto spatial representationsandextractionof new predicates.Thatis, the spatial representationis inextricably involvedin the wholeprocessof reasoning. A useful modelis one wherethe agent’scognitivestate is representedas having two components, a spatial one and a corresponding predicaterepresentation,with the two componentsmatchingone another. There are nowtwomodulesthat operateon the side, a predicateextractionmodule,as before, anda predicateprojectionmodulethat constructsor modifiesthe spatial representationto matchthe predicate component.The famous"Behold" proof of the PythagoreanTheorem (see Chapter 1 of [7] for a versionof this proof)is just a series of diagrams,with no linguistically represented assertions alongthe way.Butevenin these instancesthe reader is assumed to be making implicit inferencesin her headas she examines the diagrams. A major concernin diagrammatic reasoningis the conditionsunderwhichsuch inferencesusingdiagramsare reliable. Conversely,the creation and use of such representationto solve reasoningproblemsthat are not intrinsically spatial (suchas reasoning aboutsets) are also mattersof intenseinterest. BarwiseandEtchemendy [4] note that it has generallybeenassumed by logicians that diagramsare a mereheuristic aid, and"real proofs"haveto be purelysymbolicin character. Amajorfocusof their workis showingthat valid diagrammaticproof systemsmayindeed be constructed. Becausespatial representationsare specific instancesrather thanquantified propositions,they often providea convenient representationfor modelsof propositions.The researchof Johason-Laird (presentedat this workshop) showsthat this propertyof sp-atial representationis useful in l~lmannon-specialist syllogistic reasoning.Givena syllogistic reasoningproblem,humanreasoners modelthe variouspropositions,representthemspatially, andtry to see if counterexamples can be constructedto the propositionto be proved. Glasgow’s work,presentedin this session, discussesthe use of spatial representationsas models. 4.3. Simulation Manysp-ati_al "reasoning"program~ or cognitivemodelsdo not reasonin the senseof moving fromcognitivestate to cognitivestate by processingtruth-valuesof assertions. Instead, they makeuse of simulationto generatethe next state. Theproblemsfor whichsuchsimulationis appropriateare generallypredictionproblems: givena spatial si_~J-afionandsomeproposed actions on, or interaction between,the elements in it, whatwill be the spatial represe~ation correspondingto the newsituation? The simulationis supposedto mimicthe spatial tran~ormations involvedin the transition from initial state to final state. Thesimulationmight additionally mimicanyphysical laws involvedin effecting the changes.In A.I, the workof Funt [8], DeCuypers, at al [9], andGardinandMeltzer [10] are examples of this kindof sp-afial simulation. DeCuypers’programcomputesthe results of diffusionof gas in a room,andGardin and Metzler’sprogramsimilarly performs analog compta-ations to predict the behaviorsof strings andother objects in three-spaceundervarious forces. Typically,the computation is implementedby somekind of array model,in whicheach processor performssome compraation andpasseson the results to its neighborsin the array. Glasgow [3] also uses a spatial array to performspatial reasoning.The simulationapproachesare often characterizedas analogprocessesfor obviousreasons. In contrastto AI, wherethere is a cert-ain freedomin choosingcomputational approachesto get the job done,there are more constraints in explaininghuman spatial informationprocessingabilities. Thescopeof suchanalog computationalprocessesin hnman.q is among the mosthotly contestedof issues. A gooddeal of the debate on the nature of imagery - propositional versus imagistic representations - can be traced to this issue. The mentalrotation experiments of Shepard, Cooperand colleagues [11] have often been interpreted as implyingthe existence of internal "analog" processes in hmnansfor performingrotation and translationKosslyn[12] proposes the existence of some analog processes in humanspatial reasoning. The alternative to the hypothesis of such analog spatial information-processing is that rotation, translation and other such spatial reasoning processes are actually solved by reasoning involving spatial predicates and predicate projections. yet haveacconnt.qof all formsof valid spatial reasoning. Howand whendiagrnm.q are effective in aiding reasoning in general remains relatively poorly understood. Goel [15] argues that, contrary to intuitions, diagram.~can not only be vague, but designers and other problemsolvers actually exploit this vaguenessfor effective problem solving. The degree to which human spatial cognition employsanalog processes for performingprediction remain.~ controversial, but in AI special purpose architectures that perform analog simulation for prediction remain promising. It seemsto methat the emer~i’ng trend towards reformulating manyAI problems in the context of robotic behavior, wherethe robots have sensory, action and reasoning abilities, is very importantfor progressin sp_Rtial reasoning. Broadeningthe notion of internal imagesfrom the visual modality to an integrated multi-modalrepresentation that includes conceptual information as well and using this internal multi-modal"image"as the basic encoding of experience promises to open up new ways of thinking about learning. 5. Control of Reasoning Earlier whenI discussed reasoning I focused just on spatial predicates. However,a morerealistic modelis one wherethe state of the agent is multi-modal,i.e., informationin several perceptual modesand also in the conceptual modeis present at each state. For example,if the agent were thinking of an apple, her cognitive state wouldhave a visual representation of the shape, texture and color of the apple. In addition, she might have knowledgeabout apples relevant to the context, say, their prices and where to buy thent She might also have an anticipation of its taste. Representational changesin one modetypically give rise to corresponding changes in other modalities, giving the multi-modal representation some degree of inter-modal coherence. Problemsolving is driven by the most relevant information in any domain. The reasoning might be driven by spatial predicates at one moment and by conceptual predicates at the next. Each moveforward results in a new state and in updatingof representations in all modalities. Koedinger and Anderson’s geometry program [13] uses a representation in whichboth conceptual and diagrammaticinformation is present at each state. Narayananand Chandrasekaran[14] investigate the control issues involved in the use of such multi-medal representations. Epstein and Gelfand [17] discuss a game-playing programwhich involves the use of such multi-modal representationtheir program’sgamestate has both spatial and propositional content. 7. Acknowledgments Parts of this paper were included in a session report at the NSFWorkshopon Visual Cognition and Decision-Makingin the Spatial Domain,May15-17, 1997. I thank the organizers for the invitation. I also thank Michael Anderson and Had Narayanan for their comments on an earlier draft of this paper, thoughI haven’t fully incorporated their suggestions in this version. I thank John and SusanJosephsonfor discussions on this topic. Researchthat forms the basis for this paper was supported by DARPA,Order mlmber: D594, monitored by Office of Naval Research, Grant #: N00014-96-1-0701. References 1. Brooks, R.A., Intelligence without representation,. ArtLOcialIntelligence, 1986.47(1-3): p. 139-159. Newell, A., The KnowledgeLevel, in . AIMagazine. 1981. p. 1-19. 3. Glasgow,J.I. and D. Papadias, Computational Imagery. Cognitive Science,1992.3:p. 355-394. Barwise,J. andJ. Etchemendy, Visual . Informationand Valid Reasoning,in Visualization in Mathematics, W. ZimmermanEditor. 1990, 6. Summary The task of formalizing the logic of spatial reasoning has just begun, since we do not 4 . 6. . . . 10. 11. 12. Mathematical Association of America: Washington D. C. Glasgow,J., N.H. Narayanan,and B. Chandrasekaran, eds. Diagrammatic Reasoning: Cognitive and ComputationalPerspectives, 1995, The MITPress: Cambridge, MA.780 + xxvii. Davis, E., Representations of CommonsenseKnowledge. 1990, San Mateo, CA: Morgan Kaufmann Publishers, Inc. Allwein, G. and J. Barwise, eds. Log/cal Reasoning wtth Diagrams.. 1996, Oxford University Press: NewYork, NY. Funt, B.V., Problemsolving with diagrammaticrepresentations. Artificial Intelligence, 1980. 13: p. 201-230. De~r, J., D. Keymeulen, and L. Steels, A hybrid architecture for modelingliquid behavior, in [5], p. 731752. Gardin, F. and B. Metzer, Analogical Representations of Naive Physics. Artificial Intelligence, 1989.38:p. 139159. Shepard, R.N. and L.A. Cooper, Mental Images and Their Transformations. 1982, Cambridge, MA:MITPress. 13. 14. 15. 16. 17. 5 Kosslyn, S.M., Images andMind. 1980, Cambridge, MA:Harvard University Press. Koedinger, K.R. and J.R. Anderson, Abstract Planning and Perceptual Cbunk~:Elements of Expertise in Geometry,Cognitive Science, 1990. 14: p. 511-550. Narayanan~N.H. and B. Chandmsokaran~ Reasoning vis!lally about spatial interactions. Proceedings of Twelfth International Joint Conferenceon Artificial Intelligence. 1991. Sydney, Australia. San Francisco: Morgan Kaufmznrt Goel, V., Sketches of Thought. 1995, Cambridge, MA:The MIT Press/A Bradford Book. Iwasaki, Y. 1997. Session Report, NSF Workshopon Visual Cognition and Decision-M~king in the Sl:m_l_ial Dornain~ May15-17, 1997. This material will be part of the Workshop Reportto appear, edited by S. Epstein, J. Gelfand, and M. Marefat. Epsteni, S. L, J. Gelfand,and J. Lesniak, Pattern-based learning ad spatially-oriented concept formation in a multi-agent, decision-m~kingexpert. CognitiveScience, to appear.