From: AAAI Technical Report SS-92-02. Compilation copyright © 1992, AAAI (www.aaai.org). All rights reserved. Automatic Design of Efficient Visual Problem Representations Stephen M. Casner NASAAmes Research Center Mail Stop 262-4 Moffett Field, CA94035-1000 casner@eos.arc.nasa.gov Abstract Automated reasoning aboutthe designof effective visual problem representations is possible whenweadoptthe viewthat visualproblem representations, alongwith the pemeptual procedures that humans useto manipulate them,canbe described usinginformation-processing modelsof the sort introducedby NewellandSimon(1972). This approach providesus not only with a means of characterizing visualproblem representationsin a formal syntaxbut also witha means of automaticallymapping between "logical"and"perceptual"problem representations andprocedures.An automatedsystem calledBOZ is described that beginswitha logical problem representationandsolutionprocedure,and generatesan informationally-equivalent visual problemrepresentation andprocedure that allowsthe human user to obtain the solution moreefficiently. BOZ’srepresentationmapping technique proceeds by: (1) replacing demanding logical inferences in the solution procedure with efficient perceptual inferences;and(2) structuring information the visualrepresentation suchthat searchis minimized. Theextent to whichthe visual representationsand proceduresproduced by BOZagreewith what users actually see anddo is discussed. Thelogical proceduredescribesa series of searchand computation steps(called logical operators)to be performed usingthe set of logical facts. Logicaloperators occur in two forms. Searchoperatorssuchas Step 1 considerthe logical facts oneat a time, beginningat the top of thelist eachtime, until a fact is locatedthat satisfies the relation contained in the operator.In Step1 the searchis for a fact that satisfies (Origin FLIGHT ’pit). Thisoperatorsucceeds onthe first fact in the list andthe variableFLICHT is instantiated with the value, flight 117. Computation operatorsperformarithmetic or comparison operationson the valuesretrieved by search operators.Step11 is a computation operatorthat subtracts departurefrom arrival and instantiates LAYOVER with the result. Computation operators performed with a logical representationare assumed to be mentallyperformed arithmetic or logical operationsapplied to a collection of data valuesstoredin the user’s memory. Visual ProblemRepresentations A visual problemrepresentationalso consists of a collection of facts accompanied by a procedure for finding a solution to the problem.Factsin a visual problem representation,as illustrated in Figure2, havetwounique features.First, facts in a visual problem representation are encoded graphically by associatingthe elementsof eachdomainset of informationwith the discriminable valuesof somegraphical dimension,a techniquefirst proposedby Mackinlay(1986). For example,eachairline flight is represented by a rectangle.Thehorizontal position of the endsof the rectangleencodethe departure andarrival timesof that flight. Theshadingof each rectangleencodes the availability of that flight. Second, facts in a visual problemrepresentationgrouptogether related informationmakingall informationrequiredto draw a particular inferenceavailableat the same spatial location. For example, all informationabouteachflight, origin, destination,cost, times, andavailability, is grouped togetherin a single rectangle. Logical and Visual Problem Representations Logicalandvisualrepresentations of an airline reservation problemare compared to illustrate the advantages offered by visual problem representations, and to showhowboth logical andvisual problemrepresentationscanbe characterized usingthe samesort of informationprocessingmodel. Logical ProblemRepresentations A Iogicalproblem representationincludesa set of logical facts alongwith a logical procedure that operateson the facts to solvea statedproblem.Figure1 showsa logical representation for an airline reservationproblemposedas follows: Finda pair of connecting flights that travel fromPittsburghto MexicoCity. Youare free to chooseanyintermediate city as longas thelayoverin that city is no morethan four hours.Bothflights that you choosemustbe available. The combined cost of the flights cannot exceed$500. Perceptualprocedures differ fromlogical procedures in that they are composed of a set of perceptualoperators that describesearchand computationsteps performed specifically within the contextof a visual representation. Theperceptualoperatorsthat appearin the perceptual procedure allow the user to obtain the sameresults as do the logical operatorsthat appearin the logical procedure since they compute the samefunctionsonly in a different way. Thebottomof Figure 2 showsa perceptual procedure for the airline reservationproblem. Thelogicalfacts(topof Figure1 ) describerelations between the elements of a collection of five domain sets: airline flights, departureandarrival times, costs,and availability. Forexample, the first fact in thelist indicates thattheorigin of Flight 117is PIT. Thevisual problemrepresentationis arguablymore efficient than the logical problemrepresentation.The advantages of the visual problemrepresentationappearto center aroundthree notions demonstrated by Larkin and Simon(1987)and Casner(1990): (1) perceptualoperators 157 °!ii E ~" li o~ ~ ~!,~^~ .--_ ,~.~ ==.~ ~,li°~ ~ "..~ a #] e~ m Nd N P~ ,-< 0 9. ~< 158 are sometimes performed moreefficiently than logical operators; (2) operators in a perceptualprocedure can sometimes be omitted; and(3) the graphicalstructuring the informationexpeditessearchfor needed information. Severalinstances of theseadvantages are apparentin thevisual airline representation. First, the visual representation allowsthe user to substitutea quick distance judgement(deterrnineHorzPos)in place of subtracting numerically expressed departureandarrival times(computeLayover). Second,Steps 4 and 7 can be omittedsince the horizontal distancebetween two rectangles canbe determined without knowingtheir exact horizontalpositions. This savingscorresponds to being able to subtract two numbers without knowingwhatthe numbers are. Thesamereasoning applies to Steps 13 and 14. Third, the visual representationeliminateseye movements whenlookingup time, city, cost, and availability information sincethis informationis represented in the same spatial locality (a singleflight box). Fourth,it allowsusersto limit their searchfor connecting flights to only thoseflights that appearto the right of theoriginatingflight. Fifth, sinceshading canbe processed pre-attentively, users mayimmediately excludefromtheir searchanyflight squarethat hasno available seats. Sixth, userscan immediately rule out "tall" flights fromtheir searchsincetheseare likely to violate the cost constraint. mappingbetweenthe arguments in the findFlightWithOrigin and the searchOb jectWithLabel operators, we concludethat the perceptualoperatorcanprovidethe sameinformationas the logical operator. It is often the casethat morethanoneperceptualoperator qualifies as a legal replacement for a givenlogical operator. For example,the perceptualoperator searchObjectWithHeight (RECTANGLE, cm) would also qualify as a substitute forthefindFlightWithOrigin operatorsince wecanderive the onefromthe other by renaming.Thechoiceof whichparticular perceptual operatorsto chooseis guidedby a two-tier ranking scheme that prioritizes perceptualoperatorsin termsof their estimatedhuman performance costs. Thefirst tier prioritizes the perceptualoperators by the difficulty of the functiontheycompute (e.g., additionis moredifficult than searchwhichis moredifficult than a comparison). The secondtier ordersthe operatorsby differencesin performance costs for operatorsthat computethe same function(e.g., searching for anobject of a particularcolor is easierthansearching for anobject of a particular size). When the perceptual operatorreplacementstep is finished, the logical procedure (Figure1) is transformed into a perceptualprocedure(Figure 2). Perceptualsearch operatorswhoseonly purposeis to feed valuesto a computationoperator can generally be skippedand are markedaccordingly. Thegoal in designinga goodvisual problem representation, then,is to beginwith a logical problem representation, andperceptuallyencodeand organize informationso that anefficient perceptualprocedure can be usedto solve the problem.Thefollowing showshow this can be accomplished. Designingthe Visual Display In two steps, BOZderivesthe visual display usingthe perceptualprocedureit has produced.First, by examining the perceptual operatorschosento manipulateeach domainset of information, BOZdetermineshoweach domainset of informationis to be represented in the display. For example, the determineDepartureTime operatormanipulates informationaboutflights andtheir departuretimes. SinceBOZhas decidedthat the best substitutefor this logical operatoris the determineHorzPos perceptual operator, BOZis constrainedto representdeparturetimesas graphical objects meaningfullypositionedalongthe horizontal axis of the display. Similarly, sincethe step of determining availability has beenreplacedby determiningthe shading of a graphicalobject, availability mustberepresented as shadings.Second,BOZexaminesthe relationships between operatorsin termsof the domainsets of informationthey manipulateto determinehowrelated informationshouldbe groupedtogether to support efficient performance of eachperceptualoperator. For example,since the domainsets describingtimes, cities, costs, andavailability are all usedwith the domain set flight,gozattemptsto collocate all domain sets in a single graphicalobject (i.e., rectangle).Oncethe graphical objects havebeendesignedBOZtranslates the original set of logical facts to a set of visual facts and rendersthemon the computerscreenusing a technique first proposed by Mackinlay(1986). Facttranslation acheivedby replacing the namesand members of each domainset of logical informationwith the names and members of the graphical domainthat BOZhas chosento representit. Thefinal productis the perceptualprocedure anddisplay shownin Figure2. Theinterestedreaderis referred to Casner(1991)for a morecompletedescription of howBOZworksand the generalproblemof reasoning aboutlogical andvisual representationsautomatically [Mackinlay,1986]. Design of Efflclent Vlsual Problem Representatlons BOZ is an automated tool that derivesefficient visual problemrepresentationsfrom moredemanding logical problemrepresentations.BOZrequiresas input a logical problemrepresentationlike the oneshownin Figure1: a set of logical facts, anda logical procedure that usesthe facts to solvethe problem.BOZworksin two steps, designinga perceptualprocedure that allowsthe user to solvethe sameproblemmoreefficiently, andthen using this procedure to decidehowthe informationshouldbe graphically encoded andstructured to best supportthe perceptualprocedure.ThusBOZdelivers two things: a visual display anda perceptualprocedurethat describes howthe display canbe usedto solve the problem. Designingthe Visual Procedure BOZderives perceptualproceduresfromlogical procedures by consideringeachof the logical searchand computation operators in the logical procedureand attemptingto locate perceptualsearchand computation operators that give the user the sameresult. BOZ accomplishes this by searchinga catalog of perceptual operators organizedaroundthe graphical dimensions usedto encodeinformationin a visual display. The perceptualoperatorsusedin the visual airline reservation procedure in Figure2 illustrate the idea: searchingfor an object of a particular shading,estimatingthe horizontal distancebetween two graphical objects, comparing the size of twoobjects, etc. A perceptualoperatorqualifies as a legal substitutionfor a logical operatorjust whenthe two operatorscanbe shownto be renamingsof one another. For example,since wecan create a one-to-one 159 efficiency advantages theyyield, it is likely that many potential designspassedup by BOZfor a particular problemare in fact moreefficient thanthe oneit produces usingits crudeone-to-onerepresentationmapping technique.Whetheror not the possible perceptual restructurings canbe enumerated for a given representationis an openquestion. Psychological Validity of BOZ’s Visual Problem Representations and Procedures Thefollowingillustratestwowaysin whichthe perceptual procedures that peoplefollow using a BOZ-designed representation candiffer significantly fromthe perceptual procedure generatedusing BOZ’sone-to-one representation mappingapproach. Sincethe logical andvisual problemrepresentationsare expressed usingthe sameformalnotation, the alternative representationscanbe useddirectly as simulationsto makecomparisons regardingtheir relative efficiency. Casnerand Larkin (1989) describeshowthese simulations can be used, in combinationwith subjects’ performance time data for the airline reservationproblemdeveloped throughoutthe paper,to empirically determineto what extent peopleactually follow the hypothesized perceptual procedures. A variety of different perceptualprocedures canbe derivedby reordering the operatorsin the perceptual procedure.For example,the user, after locating two connectingflight rectangles,mightcheckthe heightsof the rectangles(Step15) prior to checkingthe distance between the rectangles(Step 11). In general, whenever the arguments to a seriesof operatorsdo not dependon oneanother,the operators canbe reorderedto arrive at a variationonthe original procedure.This observationdoes notposea problem to BOZsinceit is straightforwardto enumerate all possible perceptualproceduresderivable throughoperatorreorderingby examiningthe data dependencies betweenoperators and generatingthat subsetof the n! permutations that do not violate the dependencies. Moreover,the fact that weare unableto know in advancewhichorderingspeoplewill actually use presents no problemsince there are in generalno efficiency gainsto be hadthroughoperatorreordering alone. Thus,with respectto operatorreordering,the perceptualprocedureproducedby BOZis guaranteedto beoptimallyefficient. Summary A unifying framework waspresentedthat allows logical andvisual representationsand the proceduresthat manipulatethemto be representedusing the same notation. An automatedtechniquewasdemonstrated that allowsefficient visual problemrepresentations to be derivedthroughtransformationof an equivalentlogical problemrepresentation.Threeformsof efficiency that visual representationsappearto offer wereexamined. Twowayswerediscussedin whichthe visual problem representationsand proceduresthat peopleuse could differ from thoseproducedusing the automated technique. A second,moredramaticwayin whichthe perceptual procedures that peopleusecandiffer fromthose generatedby BOZoccurs whenusers perceptually restructurethe visual data itself. Visualproblem representations offer the user the opportunityto parsethe dataalongvisual dimensions other than those usedto createthe representation.For example,the flight rectanglesrepresentationwascreatedusing a single entity, a rectangle,that corresponds to the logical entity that appears in the logical facts: a flight. However, the useris not perceptually limitedto this ontologyat all. Rather,they are free to carveupthe visual representation in anymannerthat helpsthemsolve the problem,and are furthermorefree to invent novelprocedures of a fundamentally different characterthat help them manipulate the data in their newlyimaginedform. For example,whensearchingfor a pair of flights havinga shortlayoverthe usercanstructure the display around the set of regionsdefinedby the spacein between the flight boxesshownin Figure2. Usingthis perceptual restructuring, the usercanguidehis or her searchusing the areabetween the flights whichformssomewhat of a number 6 in the representation in Figure2. In orderto find two flights havinga short layoverthe user canconsider pairs of flights that lie at the endsof thenarrowest part of this area. For example,the user mightimmediately find the PIT-HOU-MEX connection since the PIT-HOU and HOU-MEX flight appearat the narrowestpart of the intermediateregion. References Casner,S. M. (1991). A task-analytic approachto the automateddesign of graphic presentations. ACM Transactionson Graphics4, April 1991. Casner,S. M. (1990). Task-AnalyticDesignof Graphic Presentations.PhDThesis, Intelligent Systems Program, Universityof Pittsburgh. Larkin,J. H. and Simon,H. A. (1987). Whya diagram (sometimes)worth 10,000words. Cognitive Science,11, 65-99. Mackinlay,J. D. (1986). Automatingthe design graphicalpresentations of relational information. ACM Transactionson Graphics,5 (2), 110-141. Newell,A., and Simon,H. A. (1972). Human Problem Solving. Englewood Cliffs, NJ: Prentice-Hall. It is importantto note that perceptualprocedures derived throughrestructuringfall well beyondBOZ’scapability for reasoningaboutthem.Not only is BOZunableto enumerate the kinds of proceduresobtainablethrough restructuring but is also unableto knowanythingabout howefficient they maybe. WhileI havenot undertaken a systematicstudyof perceptualrestructuringsnor of the 160