From: AAAI Technical Report SS-92-02. Compilation copyright © 1992, AAAI (www.aaai.org). All rights reserved. DIAGRAMMATIC 1MODELLING REASONING IN GALILEO’S SCIENTIFIC KINEMATIC DISCOVERY: DIAGRAMS Peter C-H. Cheng Department of Psychology, Camegie Mellon University, pc2n+@ andrew.cmu.edu Abstract Diagramsare ubiquitous in scientific discovery and scientific reasoning. The uses of diagramsare studied by examiningtheir role in Galileo’s workon kinematics and the computational benefits of reasoning with diagrams are investigated by comparingcomputational modelsof conventional and diagrammaticapproachesto discovery. Diagrams provide a convenient form of representation for experimental setups that preserve valuable topological information about the setups. Diagramsallow the ready-madeand powerful inference machinery of Euclidean geometry to be used for reasoning. Whensuitable standardization techniques exist, diagramsfurnish a basis for reasoningabout the magnitudesof dissimilar properties. Diagramscan be deliberately organizedto provideperceptual clues to aid the makingof inferences, and to act as structuredrecords of the available information. The consequencesof this are computationalsavings in the processesof search and recognition. Finally, the efficiency of inferences can also be improved by replacing expensive forms of reasoningwith cheaperperceptual inferences. 1 INTRODUCTION Researchersin Artificial Intelligence and CognitiveScience are investigating scientific discoveryas one particularly interesting area of humancognitive abilities (e.g., Shrager &Langley, 1990a; and Cheng,in press, for a review). This and the ubiquity of diagrams in scientific research makes scientific discovery an excellence domain in which to investigate the uses and benefits of reasoningwithdiagrams. The need to study the role of diagrams has been widely acknowledge in the area (e.g., Langleyet al., 1987;Shrager &Langley, 1990b; Cheng,in press). Although, research on scientific diagrammatic reasoningis just in its infancy, like study of diagrammatic reasoning more generally, someinteresting findings have already been obtained. Larkin & Simon (1987) have demonstratedthat diagrammaticrepresentations can confer significant computational advantages over sentential representations, of the same informational content. The benefit comes from the preservation of information in diagrammaticrepresentations about the topological features and geometric relations among the components of a problem.This informationhelps, to different degrees, in the processes of search, recognition and inference. Search lThanksshould go to Herbert Simonfor the discussions that have helped to shape this workand for his comments on an earlier draft of this paper. This research wassupportedby a postdoctoral fellowship fromthe UKScience and Engineering ResearchCouncil. 33 Pittsburgh, PA 15213, USA. involves finding information relevant to the problem. Recognitionis the process of identifying rules to solve the problem using that information. Inference is the manipulation of the information by a pertinent rule to produce new information as part of the solution. Other interesting work has been done by Funt (1977), Novak (1977), and Shrager(1990). This paper investigates the uses and computational benefits of diagramsin scientific discoveryby considering examplesof the reasoning and discoveries of Galileo in his work on kinematics. Although Galileo’s kinematic discoveries have been computationally modelled in some detail (Cheng, 1990, 1991), the role of diagrams has not previously been considered. Wewill consider how: ¯ Diagrams provide a form of representation for experimental setups that conveniently encodes valuable informationabout the spatial properties and geometrical relations of setups. ¯ Backgroundknowledgein the form of the ready-madeand powerfulinference machineryof Euclidian geometrycan be used when information is represented diagrammatically. ¯ Diagrammatic standardization techniques can be used to providebases for reasoningabout different magnitudesof dissimilarproperties. ¯ Diagramsallow information to be organizedin waysthat makeproblemsolving easier or moreefficient; reducing the amountof computation in search and recognition processes. ¯ Diagrammaticrepresentations can reduce the amountof computation during inference processes by replacing costly inferenceswith cheaperperceptualinferences. Thesepoints are consideredin turn. 2 REPRESENTING EXPERIMENTAL SETUPS Galileo wasone of the first scientists in the modem sense, because he not only theorized about the nature of phenomena, but also performed carefully controlled experiments(e.g., Drake, 1978). His experimentalskill exemplified by his work on the motion of balls rolling downinclined planes or ramps,Figure1 (e.g., Drake,1975). This experimentwas crucial in the discovery of the Timessquared law, that underlies Galileo’s subsequentkinematic findings. In the third section of the TwoNewSciences (Galileo, 1974), TNS-m here after, the kinematicdiscoveries are presented in the form of propositions (theorems and problems). Each of the 38 propositions has associated diagrams that are referred to in the derivation of the proposition from earlier propositions or a single initial postulate. - s ball:size, weight, volume - T - X inclinedplane ~ I height - y 7- length Figure 3 MeanProportional The,.0r0m i tST = dST¯ Substituting into Equation1, and rearranging: tsy 2 = dST2. dsy / dST. Fi~,ure 1 Inclined Piing Experiment c So, Figure 2 Postulate Diagram Eachdiagramis a representation of an actual experimental setup or one that could be constructed. For instance, the initial postulate states that the speeds of bodies descending inclinedplanesof different lengths, but of the same(vertical) height, will be equal. Its diagramshowstwo right angled triangles withthe vertical sides of eachcoinciding,Figure2. The diagram obviously preserves important geometric features of the inclined plane experiments,Figure 1, and the coincidenceof their verticals showsat a single glance that the heights are equal. In general, diagramslike these are a convenient wayto encode information about the relevant features of an experimentalsetup, in a representationalform that is eminentlysuitable for makinginferences, as will be seen below. 3 STANDARDIZATION Throughoutthe TNS-In, Galileo draws lines of the same length to represent magnitudesthat are equal, as in the case of the heights of the twotriangles above. Galileo also uses a diagrammaticstandardization technique to comparethe magnitudesof different properties. The prime exampleof this is the meanproportional theorem, whichis the second corollary of the Times-squaredlaw, Proposition 2 (TNS-III2). The Times-squared law can be written as an equation; dST [ dsy = tST 2 ] tsy 2 .... (1) where,d is the distance, t the time, and the subscripts denote 2. different portionsof the descent Galileo derived the Meanproportional theorem from the Times-squaredlaw, using Figure 3. The lines ST and SYrepresent the two distances travelled by a body. Now,by choosingthe units appropriately, let the time taken to txavel throughST also be representedby the line ST; 2Galileo usually stated his laws as sentences. For ease of comprehension, equationsare employed here, withoutaffecting the argumentsconcerningdiagrams. 34 tsY / tST = ~/(dST ¯ dsy ) / dST, tsy / tST = 4dsy / 4dST. ... (2) ... (3) ... (4a) ... (4b) or, Thus, the time to cover SYis the square root of the product of the distances ST and SY, that is, the meanproportional of ST and SY, which is SX in the diagram. Galileo has transformed the law into a convenient form that can be simply applied to kinematic diagrams-- the Times-squared law has been diagrammatically operationalized. The Mean proportionaltheoremis fundamentalto the discoveries in the TNS;the cases to be considered belowdemonstratehowthe theoremis employed. Diagrams can provide convenient means for representing different magnitudesof similar or dissimilar properties. By arranging lines whose lengths are in proportion to the magnitudesin close proximity, or even overlapping, comparisons of the magnitudes are readily made.Lines are abstract representations of quantities that allow techniques like the Meanproportional theorem to provide a standard for makinginferences involving different properties. 4 USING BACKGROUND GEOMETRIC KNOWLEDGE "Galileo’s theorem"(Drake, 1978), TNS-III-6, is concerned with the times of descent along inclined planes within a vertical circle. Figure 4 showsinclined planes runningfrom points on the circumference of a circle to the lowestpoint in the circle, and planes runningto the circumferencefromthe highest point. Galileo demonstrates that the times of descentfor all such inclined planes are equal3. The ratio of the time of descent along DCto DGis given by the Mean proportional theorem, Equation4a: tDC / tDG = 4(dDC. dDG) / dDG.... (5) where t and d are time and distance, respectively. Now, Galileo was well-versed in Euclidian geometry, so simple states withoutproof that the distanceDFis equal to ~/(dDC dDG). This can be simply proved using Pythagoras’s theorem.Substituting into Equation5, we get: 3Galileo was unawareof rotational momentum, so did not realize that this theoremis onlytrue whenthere is non-skidding rotation (or frictionless sliding) alongeveryinclinedplane. b~~l ae ad ab ac i 9 ~9 4~6 10--15 18--18 e Fi~,ure 4 Galileo’s Theorem tDC / tDG = dDF / dDG .... (6) Finally, TNS-III-3states the times of descent downinclined planes of equal height are in proportionto their lengths, so the time along DFto that along DGis given by: IDF / tDG = dDF / dDG .... (7) Hence, examiningEquations 6 and 7, the time for DCand DFare equal. This is a good example of how Galileo used the Meanproportional theorem to represent the problem diagrammatically, and then employed his knowledgeof Euclidean geometryto find a simple solution. There are manyother similar cases in the TNS,and Galileo’s work more generally, where the ready-madepowerful inference machineryof geometryis used to makediscoveries once the problemhas been transformedinto a diagrammaticform. Prop,. 3 Prop. 2 Prop. 3 b / \/ a a a d f a I ab 12-- 18 ae 9--9 ad 4--6 c Figure5 Discoveryof TN$-III-~ 5 PERCEPTUAL CLUES, INFORMATION ORGANIZATION, AND SEARCH AND RECOGNITION SAVINGS Anothergoodexampleof the use of the meanproportional theoremis in the discoverythat the time of descent downan inclined plane is in proportion to its length and inversely proportional to the square root of its height, TNS-III-5. This discovery is also interesting because it demonstrates howdiagramscan provideperceptual clues for inferences and organize information in ways that makethe whole process easier and more efficient. Figure 5 showsthe essential extracts from the relevant page in Galileo’s note books (f.189v, Galileo, 1979). The two right angled triangles represent two inclined planes; again, their horizontals are drawnalong the sameline and their peaks coincide. Below, the wavyset of parallel lines represents distances and the straight lines represent times. The labels indicate bowthe lines relate to the twoinclinedplanes. Galileo assumesthat the lengths and the heights of the inclined planes and the time ab, are given. Thefirst step wasto find the time ad, by applyingTNS-III-3,whichstates the ratio of the times of descentdowntwodifferent inclined planes(or the vertical) is equal to the ratio of their lengths; so, tab / tad = dab / dad .... (8) where,t and d are time and distance, respectively4. Thenext step wasto find the timeto cover the distanceae, giventhat the time to cover ad is now known. By the Mean proportionaltheorem,Equation4b, the ratio of the times is, tad / tae = X/dad / ~/dae .... (9) Thatis, the timeto travel distance ae is givenby the line af 5, on the diagram.The third step is to apply TNS-I]I-3again on this occasionto ae andac, to find the timeac, (10) tae / tac = dae / dac .... Finally, findingthe ratio of the times of ab to ac is a matter of combining all the separate timerelations ab:ad, ad:af, and af:ac; that is compounding Equations8, 9, and 10: tab/tac = (dab/dac). 01dae / ~/dad) .... (11) The time of descent downan inclined plane is in proportion to its length and inversely proportionalto the squareroot of its height. Clearly, the diagramshelped greatly in the making of this discovery. Theorder of the distance lines waschosen so that they matchedthe sequenceof inferences to be made. As each newtime was inferred, it was placed opposite its correspondingdistance, ready for consideration in the next step. Finally, the three sets of time and distance pairs providea record of the relations that needto be combinedat the end. The dashed arrows and proposition numbershave been added to Figure 5 to illustrate these diagrammatic relations. The sets of numberson the right of Figure 5 are two separate numerical examplesthat Galileo considered. Notethat the lengths of the distanceand time lines represent the relative values in the examples,but are not exactly to 4Thecomment,in note 3 aboveon rotational momentum, also appliesto TNS-III-3. 5Fortuitously, the double application of the proposition cancelsthe error of ignoringrotation. 35 scale. The discovery has been modelled twice using a production system: first, in a "conventional" manner, without the use of diagrams; and second, using a representation for diagrams. Becauseof limited space the modelswill only be briefly outlined6. The conventional modelemployssix rules. Three rules instantiate the mean proportional theoremand two versions of TNS-[II-3. There are also two mathematicalrules for the simple manipulation of equations and a final rule that recognizeswhenthe goal has been achieved. Consider for example, the mean proportionaltheoremrule (translated into pseudoEnglish): If the distance<S T>anddistance<S Y>are in_line<ST Y>, andtime<ST>is given, Thenthe time<SY>is givenby the equation, time<SY>/ time<ST>= qdistance<S Y>/ ~,/distance<S T>, andignorethe distance<S T>. S, T, and Y are variables standingfor different points, and for exampledistanee<S Y>meansthe distance between S and T. This rule fires whenthere are twodistances that are in line and havethe sameorigin, and the time to cover the first distance is known. The rule places an equation equivalent to Equation4b into workingmemory,notes that the time to cover the seconddistance can be calculated, and removesthe first distance fromconsideration. The diagrammatic model employs the same representations for equations, but does not have separate expressions for informationconcerningdistances and times. Rather, it employsa simple representation for diagramsthat is equivalentto the parallel lines in Figure5. For example, diagram [dists (<ab><a d><ae><a c>) ~mes (<ab><ad>)] indicates that there are four parallel lines groupedtogether representing four distances and there is a group of two parallel lines representingtimes associatedwiththe first two distances. The small letters are specific points and the diagram preserves the topological relations between the items of information. In the program diagrams are instantiatedas nodelink structures. The first of the six diagrammaticmodelrules has the job of constructingan initial diagramconsisting of lines for the lengths and verticals of the twoinclined planes anda line for the single given time. Three TNS-IIIrules still producethe required equations but they refer to the single diagram rather than multiple sets of expressions. For examplethe meanproportional theoremrule is: [(:lists (<AB><AC><AE><AD>) times(<AB><AC><A E>)]. The same mathematical manipulation rules are employed, but no rule for recognizingthe gaol is required. The run time for the diagrammatic approach was 50%faster than that for the conventional approach, even though both cases employedthe samenumberof rules. The computational savings come from the use of the diagrammatic representation of the distance and time information, and is manifested in two forms. First, the diagrammaticrepresentation reduces amountof effort needed to search workingmemoryon each cycle. The conventional rules each place several expressions in workingmemory, whereasthe diagrammaticrules modifythe single diagram expression. As there are fewer expressions in working memoryat each stage of the diagrammaticsimulation, the effort to search workingmemoryfor information to match with rules is significantly less. Second, there are more expressionsin the condition parts of the conventionalrules compared with the diagrammatic rules (e.g., 4 and 1, respectively, in the two rules above). Thus amountof processing required for each diagrammaticrule is less. In Larkin and Simon’s (1987) terms, the cost of search and recognitionhavebeensubstantially reduced. This Galilean discovery showsthat diagramsallow information to be organized intentionally to aid reasoning processes by several means: (i) exploiting the role proximity; (ii) grouping together diagrammaticelements representing similar types of properties or entities; (iii) using shared diagrammaticfeatures, such as wavylines, to distinguishedsimilar properties or entities; and, (iv) using the relative size of diagrammaticcomponentsto indicate magnitudes.Suchdiagrammaticproperties aid the discovery processby permittingdiagramsto: (i) act as a well-organized record of the inferences already madeand the information generated; and, (ii) serve as a sourceof perceptual clues indicate whatinferences can be made. In the next section we consider how Galileo employeddiagrams to reduced the amountof computation duringinference. 6 COMPUTATIONAL SAVINGS DURING INFERENCE TNS-III-30considersthe followingsituation: Giveninclined planes runningbetweentwo parallel horizontal lines, Figure 6, what is the inclination of the plane that will give the If thereis a diagram [dists (<AB><AC><AE><AD>) times(<A B><AC>)], Thenapplythe equation, fime<AC>/time<AE> = ~/distance<A C>/ ~distance<A E>, andredrawthe diagram as 6Amanuscript withfull details canbe obtainedfromthe author. 36 Figure 6 Inclined Planes BetweenTwoVerticals b d Figure 7 Triangle Representingan Inclined Plan¢ quickest time of descent?The speedof the ball increases as the inclination increases, becausethe total vertical distance (height) of descentincreases. However, the total distance be travelled also increases as the inclination increases. The problemis to find whenthe two effects combinedto produce a minimum.The inclination for the minimumtime is 45 degrees7. Galileo did not solve this problemdirectly, but ingeniously employed TNS-III-6 "Galileo’s theorem", considered above, that happens to demonstrate how diagrammatic reasoning can achieve significant computationalsavings during inference. Althoughthere is no record that Galileo ever solved the problemby direct application of his laws, it is quite feasible to do so. Consideran inclined plane dac, in Figure 7. The time of descent downthe plane is given by TNS-[II5 on uniform velocity motion: tda= dda / Vda .... (12) where t, d, and V are time, distance and mean speed, respectively. The mean speed is equal to the maximum speed at the end of the descent, Vmax,divided by two, TNSl]]-l, thus: (13) tda = 2. dda / Vmaxda .... Fromthe general Lawof free fall, the terminal velocity downan inclined plane is in proportionto the square root of the heightof the plane; thus: tda = 2. dda / ~/hdc .... (14) where h is the height. From Pythagoras’s theorem, the length of the inclined plane can be replacedby its height and the horizontal component of its length, giving: tda= 2. x/(dac 2 + hdc2) / X/hdc .... (15) In Equation 15, tda is a minimum,for a given dac when ¯hdc=dac These steps have been modelled using the productionsystemwith eight rules. Thereare three rules for the three Galilean laws; three mathematical rules for manipulatingequations; a rule instantiating Pythagoras’s theorem; and, a rule that gives the condition for the minimumof equations with the same form as Equation 15. Thelast rule is: If thereis anequation of theform, 7Asimple problemto solve in modernkinematics: e.g. by finding the timeas an expressionof distanceandacceleration, both as functionsof the inclination; and differentiating with respect to inclination, setting to zero and solvingthe equation to find the minimum. 37 Figure 8 Galileo’s Solution to TNS-1TI-30 PI<D A>= 2 * ~/(P4<AC>2 + P5<DC>2) / 4P5<D c>, Thenthe minimum occurs whenP4<AC>= P5<DC> as PI<DA>varies, andignorethe original equation. P1, P4 and P5 are variables that stand for different predicates (e.g., horizontal or vertical distances). Notethat this rule merely states the condition for the minimum withoutinferring it. The way Galileo solved the problem is shownin Figure 8. A circle with a radius equal to the distance betweenthe parallel lines is drawnwith its bottomat the point of intersection of all the inclined planes. Now, "Galileo’stheorem"states that the times of descentalong the inclined planes within the circle are equal; tea = tha = tka. However,as the inclined planes da andfa are longer than ha and ka, their times of descent must be longer than times ha and ka. Therefore, the descent time betweenthe verticals alongthe incline ea is the shortest. This approach has also been modelled Expression like line<C D>and circle<A E G>described diagrammatic elements. Six rules are employed.Thefirst rule locates the point wherethe inclined planes intersect. The secondrule "draws"a circle with a radius equal to the distance between the parallel lines and its bottomon the intersection point. Thethird rule is a perceptualoperator: If thereis aninclined._plane<D A> that intersectsthe line<CD>at <D>, andan inclined__plane<E A> that intersectsthe line<CD>at <E>, anddistance<A C>anddistance<E C>are equal, andthereis a circle<AE G>, Thenseethat inclined_plane<VRBL1 A> intersectsthe circle<AE G>at <VRBLI>, andseethat the thesepoints are in_line<AVRBL1 D>, andseethat the distance<D A>is longerthan distance<VRBL1 A>, andignorethe facts: <D>intersectsthe line<CD>, <E>intersectsline<CD>,terminal<A>, distance<AC>=distance<E C>, circle<A E G>. G is the upper most point on the circle and VRBL1 stands for point h in Figure 8. This rule provides informationthat could be obtain just by looking at the diagram;for example, that the distance da is longer than distance ha. Thefourth rule instantiates Galileo’s theorem.Thefinal rule identifies times that are longer than the commonminimumbecause their respectivedistances extendbeyondthe circle. The diagrammatic simulation runs about 12% faster. However,the computationalcost of the conventional approach is not truly reflected by this figure. The diagrammaticrules are all low cost rules; seeing wherethe inclined planes intersect, drawingthe required circle, and spotting relations amongstdiagrammaticelements, are all straightforward perceptual tasks. There are two kinds of computationalexpenseshidden by the simplifications made in the conventionalmodel.First, there are specific hidden costs associated with someof the rules. For example,the effort required to infer the minimumof Equation 15 is substantial but not reflected in the rule, whichmerelystates the minimumcondition. Galileo may have done so by calculation, with dac set to unity, and iterating over successive values of hdc; a protracted set of inferences. Alternatively, he mayhave reasoned about the way the numeratorand denominatorof Equation15 each varies with hdc, and thus howthe overall magnitudeof the equation may vary; a difficult set of inferences. Thesecondtype of hidden cost is generalto all the rules in the conventionalapproach. Galileo usually considered equations involving ratios of terms, but the rules for modellingthe conventionalapproach only considered individual terms. Thus, the expense of dealing with the Galileo’s morecomplexform of notation is absent from the model. What the diagrammatic approach achieves is a substantial saving during inference. Larkin and Simon (1987) consider that the potential for inference savings less than that for search and recognition, when representations are informationally equivalent. However, this case demonstratesthere are significant inferencesavings to be madealso, under certain circumstances. The reason Galileo’s diagrammaticapproachis so muchcheaper, is that it allows qualitative reasoning to be used. The difficult tasks of manipulating quantitative expressions and, especially, the problemof minimizinga complexequation, are replaced by simple perceptual tasks, such as seeing whetherone line is longer than another. 7 CONCLUSIONS The uses and computational benefits of diagrams in scientific discoveryhavebeen investigated by studying and modellingGalileo’s kinematicdiagrams. ¯ Diagramsprovide a form of representation for experimental setups that convenientlyencodevaluable informationabout the spatial propertiesand geometricalrelations of setups. ¯ Backgroundknowledgein the form of the ready-madeand powerfulinference machineryof Euclidian geometrycan be used wheninformationis represented diagrammatically. ¯ Diagrammatic techniquesfor standardizationcan be used to provide a foundation for reasoning about different magnitudesof dissimilar properties. ¯ Diagramsallow information to be organized in waysthat makeproblemsolving easier or moreefficient; reducingthe amountof computationin search and recognition processes. ¯ Diagrammaticrepresentations can reduce the amountof computationduring inference processes by replacing costly 38 inferenceswith cheaperperceptualinferences. In recent research (Cheng& Simon,submitted) have used these findings to demonstratethat Galileo would have been able to discover the conservation of momentum using a diagrammaticapproach. REFERENCES Cheng, P.C-H. (1990). Modelling Scientific Discovery. Technical Report No.65, HumanCognition Research Laboratory, The Open University, Milton Keynes, England. Cheng,P. C-H.(1991). Modellingexperimentsin scientific discovery. Proceedingsof the 12th International Joint Conferenceon Artificial Intelligence. MountainView, CA: Morgan Kaufmann. Cheng, P. C-H.(in press). Approaches,modelsand issues in computationalscientific discovery. In M.T.Keane& K. Gilholy. Advances in the Psychology of Thinking (Vol. 1). Harvester-Wheatsheaf. Cheng, P. C-H. & Simon, H. A. (submitted) The right representation for discovery: Could Galileo have discovered the conservation of momentum? Submitted to The 9th International MachineLearningConference, Aberdeen,July 1992. Drake, S. (1975). The role of music in Galileo’s experiments. Scientific American,232,98-104. Drake, S. (1978). Galileo at Work. Chicago:University of ChicagoPress. Funt, B.V. (1977). Whisper: A problem-solving system utilizing diagrams and a parallel processing retina. Proceedingsof the Fifth International Joint Conference on Artificial Intelligence. Galileo Galilei. (1974) Two new sciences. S. Drake (Trans.). Madison,Wisconsin:University of Wisconsin Press. Galileo Galilei. (1979). Galileo’s Notes On Motion. S. Drake (Ed.). Annali dell’Instituto e Museodi Storia della Scienza.Florence,Italy. Langley, P., Simon, H.A., Bradshaw, G.L. & Zytkow, J.M., (1987). Scientific Discovery: Computational Explorations of the Creative Process. Cambridge,MA: MITPress. Larkin, J.H. & Simon, H. A. (1987). Whya diagram (sometimes) worth ten thousand words. Cognitive Science, 11, 65-99. Novak, G.S. (1977). Representations of knowledgein programfor solving physics problems. Proceedingsof the 5th International Joint Conferenceon Artificial Intelligence. Shrager, J. (1990). Commonsenseperception and the psychology of theory formation. In J. Shrager & P. Langley, (Eds.). ComputationalModelsof Scientific Discovery and Theory Formation (437-470). San Mateo, CA: Morgan Kaufmann. Shrager, J., & Langley, P. (Eds.) (1990a). Computational Modelsof Scientific Discovery and TheoryFormation. San Mateo, CA: Morgan Kaufmann. Shrager, J. & Langley, P. (1990b). Computational approachesto scientific discovery. In J. Shrager &P. Langley, (Eds.), ComputationalModelsof Scientific Discovery and Theory Formation (5-25). San Mateo, CA: Morgan Kaufmann.