THE REPRESENTATIONAL RE-DESCRIPTION HYPOTHESIS COMPARED AGAINST TWO CASES From: AAAI Technical Report SS-93-01. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved. DONALD PETERSON Cognitive Science, ComputerScience University of Birmingham Edgbaston Birmingham B15 2TT UK Abstract. Re-representationis an interesting variety of creativity. The’RepresentationalRe-descriptionHypothesis’ of Karmiloff-Smith et a/is a theory of individual cognitive functioningand development emphasisingthe r61e of rerepresentation.Weask haewhetherthis can be seenas a generaltheory of re-representation,and answerin the negative. 1. Introduction Knowledge is not just a matter of what is known,but also the mannerin whichit is represented -- and the re-representation of a problemor a domaincan constitute an epistemic advance of a significant and distinctly creative type (Boden1990, and 1987chapter 11). Treamaentsof re-representation are to be found in a variety of contexts -- for examplein literature on the psychology of problem solving (e.g. Duncker 1945, Wextheimer 1961, Holyoak1990), on mathematical problem solving (e.g. Polya 1948), and on knowledgerepresentation in artificial intelligence (e.g. Amarel1968, Koff 1980). A newand especially interestingtreatment is to be found in the ’Representational Re-description Hypothesis’ or ’RRH’ (Karmiloff-Smith 1992, Clark and Karmiloff-Smith foilhcoming, Clark 1993). On the basis of evidence from experiments on children’s cognitive development, Annette Karmiloff-Smith has argued that human beings have an internally-driven propensity to re-represent implicit, procedural, domain-specificknowledgeas explicit, declarative, abstract knowledge.Whatstarts as a procedureof the systembecomesa data-structure to the system (Karmiloff-Smith1990), and thus becomesmanipulable and flexible. The capacity to re-represent in this way,it is suggested,constitutes an evolutionary point past which only humanshave passed (KarmiloffSmith 1992, pp191-192). The RRH,then, attributes the following characteristics to re-representation: 1. Breakpoints: it constitutes a phylogeneticbreakpoint, in that creatures lower downthe evolutionary scale cannot perform it. dresses the reformulation of procedural knowledgeas declarative knowledge,only the latter being a manipulable object of the system. 5. Implicit ---. explicit transformation: knowledgewhichis implicit in the system’s proceduresbecomesavailable as an explicit and hence manipulabledatastructure. The consequencesof the RRHare far reaching. It has been argued that its implications extend to the problemof intentionality and the traditional debate betweenrationalism and empiricism (i)artnall 1993), and to the issue of whether not cognitive fidelity attaches to purely connectionist cognitive models(Clark 1993). It seems fair to ask whether the characteristics identified by the RRHare true of all or only some cases of rerepresentation. Accordingly, the strategy of the present paper will be to analyse two cases -- a re-representation of the game of ’number scrabble’ and the replacement of Romanby Arabic numerals -- and to ask to what extent these conform to the pattern set out in the RRH. 2. NumberScrabble The initial representation of the gameof numberscrabble is as follows. Nine cards, each bearing a different numeralbetween 1 and 9, are placed face-up between the two players. The players drawcards alternately, and the first player to acquire three cards whichsumto 15 wins. (If the cards are exhausted without either player acquiring a winningtriplet, then the result is a draw.) This maynot be the world’s most difficult game,but it is quite hard to pick cards whichwill: ¯ produce a winningtriple 2. Spontaneity: the RRHappeals to an ’endogenously driven’ propensity to ’spontaneous’re-representation aftea" ’behavioural mastery’ has been achieved. 3. Abstraction: ’the re-descriptions without doubt constitute someform of abstraction from the phase one details’ (Clark 1993). 4. Procedural---, declarative transformation: the RRHad91 ¯ prevent one’sopponent fromgaining a winningtriplet, and ¯ create a ’fork’ situation in whicheither of two cards will producea winningtriplet (so that whateverthe opponent chooses, one still wins on the next draw). Newell and Simon (1972, 1362) present a striking rerepresentation of the gamewhich involves playing noughts and crosses over a 3 by 3 magicsquare (this being an arrangementof the numerals1 to 9, as below, so that each row, columnand diagonal sums to 15). 2 7 6 9 5 1 4 3 8 Fig. 1. magic square The initial representation of the gamerequires us to proceed by using arithmetic calculation. In the magicsquare rerepresentation, the arithmetic features of the gameare ’built into’ the diagram, and we can proceed by using our capacities of spatial reasoning and pattern recognition. Thecrucial configurations such as winningtriples and forks are easily recognised in this way. Moreover,the procedures in question are familiar to anyone whoalready knows howto play noughts and crosses. If numberscrabble were played over the telephone,and one player secreOyused the magicsquare representation, it seems clear that this would put that player at an advantage. And the point for our general themeis that performanceis affected here not just by what is knownabout the game, but also by howit is represented. 3. Numerical Notation A well knowncase of re-representation is the replacement of Romannumerals with the nowfamiliar positional notation. The modemsystem of decimal Arabic numerals uses a positional notation, the base 10, and the numeralzero. In this system a numeral signifies a sum of products of powers of the base, the powerbeing indicated by position: a numeral ABCD signifies (A× 103) + (B x I02) + (C x 10l) + (Dx 10°). Throughhistory there have been a great manynumeralsystems. The Babylonians had a positional system with base 60 but without a zero. The Mayanshad a positional system with zero and base 20, although here the 3rd digit indicates not multiples of 202 but of 2018in order to give the days in their year a simpler representation. The system of Romannumerals has manyvariants, but generally few primary symbolsare used (e.g. I, V, X, L and C), and these are arrangedaccording to additive and sublractive principles. Different systems, of course, suite different purposes. ’Grouping systems’ are useful for keeping incremental tallies, and Romannumerals have the advantage that few primary symbols have to be remembered. But the overwhelming advantageof a positional systemwith zero is that a divideand-conquerstrategy for the basic arithmetic operations of addition, subtraction,multiplication and division becomespossible wherebya simple operation (involving ’carrying’), iterated along the numerals, achieves the purpose. It is not the 92 expressive poweroftbe notation whichis at issue here, but its syntactic consistencyand the effect this has on the application of arithmetic procedures. Each digit has an associated multiplier (a power of the base), and this multiplier is ’built into’ the syntaxin the form of the digit’s position. Thusif two numeralsare ’lined up’, so are their digits’ multipliers, and arithmetic operations can proceed in a simple iterative manner. With Romannumerals, however, lining two numerals up together does not produce such a concordance. For example, the numeral’XVII’is built on additive principles, while ’XIV’ is built on additive and subWactiveprinciples, and accordingly the positions of the digits cannot be interpreted as with ’ 17’ and ’14’. The advantage of positional over Romannumerical notation, then, is not a matter of expressive powerover a domain, since both systems can denote numbers. Rather, it is due to the possibility in positional notation of applying simple and uniform arithmetic procedures. 4. Comparison with the RRH Twocases of advantageousre-representation have been given brief consideration above,and we nowturn to the question of howwell these fit the picture given by the RRH. 4.1 BREAKPOINTS RRHre-representation is put forward as constituting a phylogenetic breakpoint. Thereis no evident reason to deny this in our two cases -- it seemsunlikely that monkeysor other nonhumancreatures could have invented these re-representations. 4.2 SPONTANEITY RRHre-representation is ’endogenouslydriven’ or ’spontaneous’ in the sense that ’behavioural mastery’ has already been achieved but re-representation takes place all the same. The spirit of this account can be seen in our two cases -people do play numberscrabble using its standard representation, and Romanengineers did design bridges and aqueducts using Romannumerals. It seems clear therefore that bebavionral mastery did precede our re-representations. Howeverbebavionral mastery is a matter of degree, and in our two cases it increases with re-representation. Number scrabble and arithmetic becomeeasier, and we becomebetter at themonce we are equipped with these re-representations. Therefore wecan only say that their inventions followeda degree of behavionral mastery, not that they followed complete mastery. Spontaneity, further, is not a unitary issue. It maybe uue, as the RRHsuggests, that humanshave a unique propensity to re-represent in the absence of exogenouspressure. Andit maybe true that non-positional systems were successful to a degree. But it cannot be said that positional systems were developedin the absence of ’pressure’ to improvearithmetic facility. previously implicit propositions about numbers. In the case of number scrabble, declarative knowledge RRHre-representations are more’abstract’ than their proceabout triples whichsumto 15, and about the intersection patdural predecessors. However it is not clear that this is flue in tern of these triples has beenbuilt in to the diagram,and we our two cases. The changeto Arabic numerals involves libermightsupposetherefore that it was used in creating and checkation from burdensome syntax, but this is hardly "abstraction’. ing the re-representation. However this informationis not exAndin the case of numberscrabble, the re-representation is plicit in the final diagram without someextra explanation if anything less abstract and general, since it will only work the diagram does not itself tell us, declaratively and proposiif the rules remainunchangedand winningffiplets sum to 9. tionally, that its rows columns and diagonals all sumto 15. The initial representation in contrast could easily be changed More significantly, the point of the re-representation is not in this respect. to makethis knowledge explicit, but rather to build it into the syntax in such a wayas to facilitate the application of appro4.4 PROCEDURAL ~ DECLARATIVE priate procedures, in this case those of noughts and crosses. TRANSFORMATION The knowledgein question has not becomeexplicit, declarInRRHre-representation theinitial representation isproceative and an "object of the system’ through re-representation dural, supporting behavioural mastery butinflexible, andthe rather it wasexplicit in the first place, has beenusedin crere-representation isdeclarative, andtherefore manipulable by ating a re-representation, and is now’built in’ to a diagramin the system. the service of easy application of procedures. In the case of numberscrabble, however,both the initial Symptomatically,in using the re-representation of number representation and the re-representation are procedural min scrabble, it is not evennecessary to knowhowto interpret it both representations the user is not told a set of declarative in terms of explicit, declarative propositions about numbers propositions, rather he or she is asked to engagein rule folsummingto 15. It wouldbe quite possible for one player to lowing. The re-representation comprises the magic square diagramtogether with a set of heuristic rules for playing noughts play using the re-representation, while being unawarethat he or she was playing numberscrabble ~ and win. and crosses, and although declarative knowledgemight be In contrast to RRHre-representation, then, it does not seem neededto create and checkthis re-representation, the end rethat the conversionof implicit into explicit knowledgeis essult is improvedprocedural knowledge. sential to either of our cases. Accordingly,the virtues of the re-representation are not of a declarative variety -- it does not allow moreexpressive or 5. Conclusion more succinct statements about a domain. Rather they are of a procedural type mthe load on working memoryand calcuThe cases of numberscrabble and arithmetic notation have lation is reduced, since at any time the diagramhas a current been presented in order to assess whether the ’Representational Re-description Hypothesis’ of Karmiloff-Smithet al state, and it is relatively easy to determinehowto proceedto can be treated as a 8eneral theory of re-representation (and the next state. not to assess it as a theory of cognitive development). Our In the case of numerals it might be said that both repconclusion is in the negative, since the cases considered fit resentations are broadly declarative, since both allow us to the pattern of the RRHonly in somerespects, and in particuname numbers. Howeverthis is a weak sense of ’declaralar do not embodythe procedural ---, declarative and implicit five’ since we can formulate only denoting expressions and not propositions. And more importantly, the point of the explicit transformations which the RRHpredicts. re-representation concerns the operation of arithmetic proceReferences dures on numerals. Wedo not just want to write downnumbers, but also to performaddition, subtraction, multiplication, AmarelS. 1968, "OnRepresentationsof Problemsof Reasoning aboutActions’,in MichieD.(ed.). Machine Intelligence3, Eddivision etc. on themas easily as possible. Andthis is greatly inburgh. facilitated by the positional systemwith zero. BodenM. 1990, The Creative Mind:Mythsand Mechanisms,WeiIn neither of the present cases, then, is it accurate or sufdenfeld and Nicolson. ficient to say that procedural knowledgeis re-represented as Boden M.1987, Artificial Intelligence andNaturalMan,2ndeeL, Mrr. declarative knowledge. 4.3 ABSTRACTION 4.5 IMPLICIT ---* EXPLICIT TRANSFORMATION In RRHre-representation, whatis madeexplicit in the declarative re-representationis implicit in the initial proceduralrepresentation. It is hard to see that somethingof this sort obtains with our two numeral systems. A simpler syntax has been introduced, allowingeasier application of arithmetic procedures, but it is not evident that this essentially constitutes the emergenceof 93 Clark A. 1993, "Connectionism, Non-conceptual Contentand RepresentationalRe-description", in Darmall T. (ed.), Artificial Intelligence andCreativity, Kluwer. Clark A. and Karmiloff-Smith A. (forthcoming),"The Cognizer’s Innards: a psychologicalandphilosophicalperspectiveon the developmentof thought". MindandLanguage. DarmallT. 1993, "Creativity, Thoughtand Re-description", in DartaallT. (ed.), Artifu:ial IntelligenceandCreativity,Kluwer. DunckerK. 1945, "on ProblemSolving", in PsychologicalMonographs,vol. 58, no 270. Holyoak K. 1990,"Problem Solving’, inOsherson D.N.andSmith E.E.(ed.s), Thinking: anInv;tatlon toCo#nitive Science, vol. 3, MIT. Ksrmiloff-Smith A. 1990,"Constraints on Re~esentationaI Change:Evidence from Children’s Drawing, Cognition, vol. 34, pp57-83. Karmiloff-Smith A. 1992, Beyond Modularity: A Developmental Perspective on Cognitive Science, MIT/Br~lford. Koff E.R. 1980, "row&dsa Model of Rel~resentsfional Ch~ges’, in Artificial Intelligence, 14, pp 41-78. Newell A. and Simon A. 1972, HumanProblem Solving, Prentice Hall. Pete~son D.M.1993, "three Cases of Re-representation in Problem Solving", in Dar~all T. (od.), Artifu:ial Intelligence and Creativity, Kluwer. Polya 1948, Howto Solve It, Princeton. Wertheimer1961, Productive Thinking, Tavistock Publications.