Shedding Computational Light on Human Creativity Subrata Dasgupta UniversityofLouisianaatLafayette Eversince1956whendetailsoftheLogicTheoristwerepublishedbyNewell andSimon,alargeliteraturehasaccumulatedoncomputationalmodelsand theoriesofthecreativeprocess,especiallyinscience,inventionanddesign.But what exactly do these computational models/theories tell us about the way that humans have actually conducted acts of creation in the past? What lighthascomputationshedonourunderstandingofthecreativeprocess?Ad- dressing these questions, we put forth three propositions: (I) Computational models of the creative process are fundamentally ºawed as theories of human creativity. Rather, the universal power of computational models lies else- where: (II) Computational models of particular acts of creation can serve as effective experiments to test universal hypotheses about creative processes and mechanisms; and (III) Computation-based architectures of the creative mind provide metaphorical frameworks that, like all good metaphors, can serve as rich sources of insight into aspects of the creative process. In this paper, we provideevidenceforthesethreepropositionsbydrawinguponsomeparticular episodesinthecognitivehistoryofscience,technology,andart. 1. Introduction The research program for understanding and explaining the cognitive natureofcreativityinvolvesabroadspectrumofempirical,theoreticaland disciplinary approaches. Roughly speaking, these frameworks draw on a numberofdimensions. (1) Cognitive historical case studies of episodes and persons based on detailed historical, biographical and autobiographical evidence (e.g., Lowes1927;Shahn,1957;Gruber1981;Miller1986,2001;Chadwick& de Courtivron 1993; Giesen 1995; Dasgupta 1999; Ippolito & Tweney 2003). (2) Computational theorizing and modeling (e.g., Langley et al PerspectivesonScience2008,vol.16,no.2 ©2008byTheMassachusettsInstituteofTechnology 121 122 SheddingComputationalLightonHumanCreativity 1987; Thagard 1988). (3) Laboratory experiments on human subjects given tasks/problems deemed to entail creative thought (e.g., Qin & Simon 1990; Finke, Ward & Smith 1992; Klahr 2000). (4) Field studies ofcontemporarycreativepeopleatwork(e.g.,Dunbar1999).(5)‘Clinical’ approaches involving interviews, questionnaires and personal life histo- ries (John-Steiner 1987; Rothenberg 1990; Csikszentmihalyi 1996). (6) Historiometric and psychometric studies that capture quantitative aspects of both past and present creative persons, domains, and domain- speciªccommunities(e.g.,Simonton1988;Ludwig1995).(7)Psychoana- lytic studies of creative people and work, past and present (e.g., Kris 1952;Storr1972) Some frameworks combine a subset of these approaches. Thus, our studyofHerbertSimon’s‘multidisciplinarycreativity’combinedthecog- nitive historical case study and the clinical approaches (Dasgupta 2003a, 2003b). A more recent study of the painter George Rodrigue utilized the case study, clinical and ªeld study approaches (Dasgupta 2004b). Our in- vestigation of the invention of microprogramming in computer design (Dasgupta 1994) and Kulkarni’s and Simon’s (1988) study of the discov- eryoftheornithinecycleinbiochemistrybothentailedthemethodofthe casestudyandcomputationalmodeling. Our focus in this paper relates to the computational and case study approaches (1 & 2 above). In particular, we raise the question: what light does computational modeling really shed on actual human creativity as it is manifested in speciªc historical episodes in science, technology and other creative traditions? We address this question by putting forth three propositions. (I) Computational models of particular acts of creation (i.e., computa- tional case studies) are fundamentally ºawed as theories that explain these creativeacts.Rather: (II) Computational case studies serve as experiments to test (corroborate/ falsify)generalhypothesesaboutcreativeprocessesandmechanisms.And (III)Computation-basedarchitecturesofthecreativemindcanserveas metaphoricalframeworksthatofferrichsourcesofinsightintoaspectsofthe creativeprocess. In other words, we suggest that it is in these two latter aspects—as sources of singular experiments and as metaphor—that the universal power of computational modeling lies in shedding light on human cre- ativity. In the sections that follow, we consider each of the above propositions andprovideevidenceintheirsupport. Perspectives on Science 123 2. Computation Models as Theories of Particular Creative Processes. In1957,HerbertSimonannouncedthat“themysteryhasnowbeenpretty well stripped away from the higher mental processes”; and since “creativ- itywasjustalittlemoreofthesame,”itwasnolonger“asmysteriousand inexplicable”assomemadeout.1 Simon’s assertion came in the wake of the success he and Newell had achieved with the Logic Theorist (Newell & Simon 1956). Later, Newell, Shaw and Simon (1962) followed up this claim by addressing the “pro- cesses of creative thinking” in computational terms. The basic thesis was thatbyprogrammingthecomputeronecouldattempttosimulatespeciªc higher level cognitive tasks such as proving theorems and the like. If the simulation produced behavior consistent with what was predicted, the program itself became a theory of how humans might carry out relevant cognitiveprocessesforthatparticularproblemclass.Furthermore,byextract- ingfromtheprogramitsmostgeneralfeatures,onecouldconstructathe- oryofthegeneralfeaturesofhumancreativeprocesses. Inthefourdecadessincethenaconsiderableliteraturehasaccumulated on work based on this thesis, especially in respect to understanding scientiªc and technological creativity (Simon 1977; Lenat 1982; Langley, Simon, Bradshaw & Zytkow 1987; Kulkarni & Simon 1988; Thagard 1988; Shrager & Langley 1990; Thagard 1992; Dartnall 1994; Dasgupta 1994;Gooding1999).Butwhatexactlydothesestudiestellusaboutthe way humans have actually discovered scientiªc laws, invented technologi- cal artifacts, produced signiªcant art, or performed other acts of creation? In other words, how valid are computational models as explanatory theo- riesofthehumancreativeprocessestheypurporttomodel? In constructing the BACON family of programs (Langley, Simon, Bradshaw & Zytkow 1987), Langley et al took up the challenge of con- structing computer models to discover some of the classic laws in physics and chemistry. This work was also notable for the comprehensiveness and breadthofitsscope. However, if we examine the knowledge base of BACON.1, the ªrst of these programs which discovered versions of Johannes Kepler’s third law of planetary motion, and compare it with the historical and biographical details of how Kepler arrived at his third law, circa 1619 (Caspar 1959; Holton 1973)—we ªnd that BACON.1’s knowledge base did not mirror theparticularknowledgethatKeplerhadpossessed.Thus,BACON.1did notdowhatKeplerdid.OnecannotclaimthatBACON.1wasatheoryof 1. H. A. Simon to G. A. Miller, April 30, 1957. Herbert Simon Papers, Carnegie MellonUniversity(CMU)Archives,Pittsburgh,Pennsylvania. 124 SheddingComputationalLightonHumanCreativity how Kepler had arrived at the third law, because the program was ahistoricalinitscapacityforscientiªcdiscovery. In contrast, the program KEKADA (Kulkarni & Simon 1988) was muchclosertothehistoricalsituationittriedtomodel—viz.,thecircum- stances under which the biochemist Hans Krebs discovered (in 1931–32) the biochemical process known as the ornithine cycle, the metabolic pro- cessbywhichanimalcellssynthesizeurea. Similarly,in(Dasgupta1994),Idevelopedanexplanation,incomputa- tional terms, of the cognitive process by which a technique in computer design called microprogramming might have been invented (in 1951) by the computer pioneer Maurice Wilkes. Here, I followed the paradigm usedinthedesignofKEKADA;thatisIadheredascloselyaspossibleto the historical evidence. Indeed, keeping in mind that the plausibility of suchacomputationalaccountwouldhingecriticallyonthevalidityofthe assumptions about the contents of the worldview/belief/knowledge sys- tem, a great deal of effort was expended in justifying the various kinds of knowledge (facts, concepts, operational principles, rules, heuristics) that enteredintothecomputation-basedexplanation. Aswewillseeinthenextsection,computerprogramssuchasBACON and KEKADA, or computational models of the sort proposed to explain theinventionofmicroprogrammingdoyieldrichdividendsinilluminat- ingaspectsofcreativity.Yetnomatterhowdetailedandprecisethemod- els are, they are fundamentally ºawed as theories of human creativity—for two reasons. 2.1 The Contingency of the Creative Process First, as theories of individual acts of creation, there are severe limits to the extent to which one can test them against the actual historical events theypurporttobetheoriesof.TheonlywaythatKEKADAasatheoryof Hans Kerbs’s ornithine cycle discovery process, or my computation-based modelofMauriceWilkes’smicroprogramminginventionprocesscouldbe corroboratedwasbyappealingtothehistoricalevidence—thatis,seeking evidence that things really happened the way they did through archival research into the person’s correspondence, notes, journals, diaries, and—if the person is still living—by way of interviews. In other words, by con- ducting cognitive historical investigations (Nersessian 1995; Dasgupta 2003a). The problem is that creativity is not a deterministic phenomenon. Consider, as an example, the actual process whereby a person arrived at an impor- tant,originalidea.Ifwecould,insomefashion,placethepersoninexactly the same initial conditions (that is, the same state of her world view/ belief/knowledge system, same state of the goal space, same emotional Perspectives on Science 125 condition,andidenticalproblemsituation),thereisnoguaranteethatthe same cognitive process (or even a close approximation to it) would unfold or,indeed,thatthesameresultwouldensue. In the context of biological evolution, Gould (1990) has argued that the evolutionary process is a contingent act: a tape on which a (natural) creative process is ‘recorded’, when rewound and restarted with the same initial conditions would not necessarily record the same process as before. Human creative processes, though not Darwinian in nature (Dasgupta 2004a), are, likewise, contingent processes—if only because the creative being, like all people, is boundedly rational (Simon 1996), and such pro- cesses entail making sequences of choices of mental actions that may vary from one situation, one time, one initial condition, to another. Because of this nondeterminstic, contingent nature of creative processes, a computa- tional model that claims to be a theory of such a particular process can only be corroborated if the historical and biographical records tell us ex- plicitlythatthepersonusedtheparticularknowledgeandreasoningrules speciªedinthetheoryinthatspeciªedorder.Otherwise,wecannotclaim that the computational theory is even a plausible theory, let alone the ‘right’one.Thetheoryastheoryisleftinastateoflimbo:onecanneither corroborateitwiththeevidencenorfalsifyit. 2.2 The Dilemma of the Unconscious Itiswidelydocumentedthatmanyhighlycreativepersons—poets,paint- ers, musicians, scientists, inventors—simply cannot account for their thought processes. Ideas would appear ‘out of the blue,’ ‘unexpectedly,’ ‘suddenly’—as recorded, for example, in the cases of poets William Blake (Harding1942,p.14),AmyLowellandStephenSpender(Ghiselin1985, 110–126), the story-teller/poet Lewis Carroll (Harding 1942, p. 59), the artist Pablo Picasso (Ashton 1972, 7–13), the composer Edward Elgar (Harding 1942, p. 14), the scientist Alfred Russel Wallace (Harding 1942, p. 15), the mathematician Henri Poincaré (Ghiselin, 1985, 22–31) andtheinventor/engineerJamesWatt(Dasgupta,2004a). This mystery of where an idea, a poem, a picture, a musical composi- tion comes from is thus a recurrent theme in autobiographical statements madebycreativepeopleinmanydifferentªelds.Anditposestheproblem of we call here the ‘dilemma of the unconscious’—which constitutes the second serious barrier to testing computational theories of creative pro- cesses. For our present purposes, the ‘unconscious’ constitutes those contents ofthemind(facts,images,acts,processes,desires,beliefs,biases)ofwhich thepersonisunaware.Thisunconsciousmightinclude,butnotnecessarily so,whatFreudtermedthe‘dynamicallyunconscious’(‘Ucs’)(Freud1920/ 126 SheddingComputationalLightonHumanCreativity 1989,5–6),whichistheresultofrepression.ItalsoincludestheFreudian ‘preconscious’ (‘Pcs’)—entities one is not conscious of but which can be easily summoned into consciousness. The unconscious also includes those cognitiveprocessesthatareneitherUcsnorPcsbutwhicharenonetheless not easily accessible. Linguistic skills in generating and understanding speech, computational skills in judging distance, depth, speed are in- stances of such processes, called the ‘cognitive unconscious’ (Kihlstrom 1987;Allen&Weber1999). The dilemma of the unconscious lies in that anecdotal and autobio- graphical reports suggest (as cited earlier) that much of creative thinking takes place in the unconscious, and so it resists our efforts to comprehend the act of creation. So compelling is the autobiographical/anecdotal evi- dence that Wallas (1926, p. 86) suggested the term incubation to describe the mental process in which “. . . we do not voluntarily or consciously thinkonaparticularproblemand...aseriesofunconsciousandinvolun- tarymentaleventsmaytake.” Though incubation as a distinct process has resisted attempts to be demonstrated in the laboratory (Eindhoven & Vinake 1952; Olton & Johnson1979),thereseemslittledoubtofitsexistence(Guildford1979). The problem is that the idea that a great deal of one’s thinking in acts of creation might occur unconsciously means that (a) any archival evidence or interview records will necessarily be incomplete, since the information willberelatedonlytowhatthepersonis/wasconsciousof;andthus(b)it willthwartanyattempttocorroborateorfalsifyacomputation-basedthe- oryofthatcreativeact. Becauseofthesetworeasons,viz.,thecontingent,nondeterministicna- ture of the creative process, and the presence of unconscious processing (speciªcally, incubation), we must conclude that a general hypothesis that computational case studies/models of particular acts of creation are ex- planatory theories of those creative acts must be rejected (proposition I), since, generally speaking, we may not be able to either corroborate or fal- sify the theory. To take speciªc examples, neither Kulkarni’s KEKADA (Kulkarni & Simon 1988) nor my model of the invention of micropro- gramming (Dasgupta 1994) can serve as explanatory theories for the ac- tual creative processes performed by Krebs and Wilkes, respectively. Nei- ther of the computational models can be corroborated or falsiªed by the historicalevidenceontheseepisodes. 3. Computational Models/Case Studies as Experiments There is a way, however, in which computational accounts (either in the formofrunningprogramsorasmodels‘onpaper’)canserveamuchmore Perspectives on Science 127 scientiªcallyimportantroleinunderstandingcreativity.Theycanserveas experiments to test universal hypotheses or theories about the creative pro- cess(propositionII,section1above).Inotherwords,theverysingularityof a computational model of a particular act of creation, or a class of such acts,suggeststhatsuchmodelscanserveasexperimentalmeanstocorrob- orate or refute some general proposition. We illustrate this with three examples. 3.1 BACON.1 as a Corroborating Experiment This view affords a very different signiªcance to the BACON programs. Consider,inparticularthefollowinggeneralhypothesis: (H1)“Insomedomainsofscientiªcpractice,neitherdomain—spe- ciªcknowledgenorphysicalintuitionnormetaphysicalcommit- mentsnecessarilymatters.Rather,newideasmaybeproducedby treatingproblemsasexercisesinformalsymbolmanipulation.” As noted earlier, BACON.1 did not derive Kepler’s third law in the way Kepler did, since the program had no understanding of the physical signiªcanceofthevariablesanddataitmanipulated.Itutilizedasmallset ofheuristicsthatsimplyformallymanipulatedthevariables;theheuristics didnottakecognizanceofthephysicalsigniªcanceofthevariables—that they denoted the mean radius of a planet’s distance from the sun, and the orbital time period of the planet. In fact, using this same set of heuristics, BACON.1alsoderivedthreeotherclassicalquantitativelawsofphysicsin very different domains, namely Boyle’s law (for gases), Galileo’s law (in mechanics)andOhm’slawinelectricity.Inotherwords,BACON.1’sruns inderivingthesefourlawseachconstitutedanexperimentthatcorroborated hypothesis(H1). 3.2 Corroborating a Hypothesis about Bisociation with the Computational Model of the Invention of Microprogramming Inthepastseveralwritersoncreativityhaveheldthatanessentialingredi- ent of the creative act is the effective combination of seemingly uncon- nected or disparate concepts or ideas ( Lowes 1927; Hadamard 1945; Gruber 1981; Perkins 1981; Johnson-Laird 1988). Koestler (1964) re- ferred to this process as bisociation. But Koestler did not explain the me- chanics of bisociation. The computational case study/model offers a way of elucidating this process. In particular, consider, now the following hypothesis: 128 SheddingComputationalLightonHumanCreativity (H2).“Bisociationisexplicableintermsofcommonplacerulesof reasoning.” Intheparticularcaseoftheinventionofmicroprogramming,itsohap- penedthattheideaofmicroprogrammingresultedfromacombinationof two essential and very unrelated ideas. One was the use of a particular kindofcircuitcalledthe‘diodematrix’;theotherwastheemploymentof idea of the stored program computer architecture (Dasgupta 1994, p. 196). In other words, Wilkes’s invention of microprogramming was an instanceofthebisociativeprocess. Whatissigniªcantforthisdiscussionisthatthemainbodyofthecom- putation-based model of the invention of microprogramming was taken withthedetailedexplicationofsuchabisociativeprocessintermsofsuch ordinary rules of reasoning as generalization, analogical inferencing, ab- duction, and deduction (Dasgupta 1994, chapter 5 & 6). In other words, thecomputationalmodeloftheinventionofmicroprogrammingservedas anexperimentthatcorroboratedorsupportedhypothesis(H2). 3.3 KEKADA as an Experiment to Refute the Darwinian Hypothesis on Creativity Campbell (1960) advanced an explicitly Darwinian model for creative thinking.ThisDarwinianmodelofhownewknowledgecomesintobeing has,since,cometobeknownas‘evolutionaryepistemology’(Radnitzky& Bartley 1987), and has since been embraced by several others (Simonton 1988, 1999; Gero, 1994; Lumsden 1999). A very recent—and exten- sive—version of this view is due to Simonton (1999 p. 69), which can be statedasthehypothesis (H3)“Thecreativeprocessmustbeviewedasaformof... Darwinism.” Space does not permit us to articulate in detail what a ‘Darwinian cre- ative process’ should look like. This has been dealt with elsewhere (Dasgupta 2004a). But stated very brieºy, if the process underlying any actofcreationisDarwinian,—if(H3)istobetakentobeagenuineuniver- salhypothesisaboutthecreativeprocess—thenanycognitiveprocessthat iscreativemustexhibit,attheveryleast,thetwomostsigniªcantfeatures of natural selection: there must be evidence of superfecundity—that is, the productionofalargenumberofvariationsofthought-productsonwhicha selectionprocesscanworktopruneoutallthosethought-productsthatdo not demonstrate a ‘ªt’ against some predetermined goal (‘environment’); and there must be evidence of blind variations—that is, (a) the alternative thought-products produced as variations are independent of the environ- Perspectives on Science 129 ment in which they arise, (b) ‘correct’ variations are no more likely to oc- cur than ‘incorrect’ ones, and (c) the ‘incorrect’ variations in any stage of theprocessdonotservetodirecttheprocesstoa‘correct’variationinthe nextstageoftheprocess(Campbell1960,91–92). Computational models of particular acts of creation are also experi- ments that may put the Darwinian hypothesis to test. In fact, both KEKADA (Kulkarni & Simon 1988) viewed as an experiment that con- structs the ornithine cycle and the computational model of the invention of mocroprogramming (Dasgupta 1994) refute the Campbell-Simonton hypothesis, for in neither of these models is there any evidence of blind variations in the generation of thought-products (ideas) or of super- fecundity. Elsewhere, we have presented other cognitive-historical case studies which also refute the universality of hypothesis (H3) (Dasgupta 2004a). Insum,ourthreeexamplesabovedemonstratetheefªcacyofcomputa- tionalmodels/casestudiesofparticularactsofcreationasexperimentsthat cancorroborateorfalsifygeneraltheoriesorhypothesesaboutthecreative process(propositionII,section1). 4. Computation-Based Architectures of the Creative Mind as Metaphorical Frameworks The inºuence of computer architectures on the development of certain models of cognitive architectures such as SOAR and ACT is well known (Anderson1983;Newell,Rosenbloom&Laird1989;Newell1990).Some researchers on creativity have either adopted one or more of these computationally-inºuenced cognitive architectures as the foundation for their studies of the creative mind, or have tailored such architectures for theirownneeds.Forexample,inconstructingexplanationsoftechnologi- cal creativity, I adopted a slightly modiªed version of Newell’s knowledge level system (Newell 1982) as the underlying architecture (Dasgupta 1996,chapter4). We suggest here that the real power of such computation-based archi- tecturesofthecreativemindliesinthattheyprovideametaphoricalframe- workforunderstandingcreativity(propositionIII,section1). Metaphors, of course, play a common, important, and often profound role not only in ordinary thought, speech and understanding (Richards 1936; Black 1962; Ortony 1983) but also in more select arenas such as scientiªc reasoning (Gruber 1981; Holmes 1985; Osowski 1989). They are used in many different ways and serve different uses. Of most interest hereiswhatBlack(1962)termed‘metaphor-as-interaction’inwhich,asa result of the observation of an analogy between two situations, one of whichisbetterunderstood,theconcepts,knowledgeandideaspertaining