How Could a Copycat Ever Be Creative?

Analogy-making in a microdomain is used as the testbed for a computer model of creativity. Among the unusual features of the model are: (1) its nondeterministic parallel-processing architecture, modeled to some extent on architectures for perception; (2) the modeling of concepts as probabilistically overlapping regions; and (3) the fact that the system manufactures its own mutable representations situations, rather than being fed fixed human-made representations. Statistical summaries of large numbers of runs on a few related problems reveal some of the program’s "personality" and "esthetic taste’. The ~e rOgram’s response to a subtle analogy problem whose most insightful solution requires a radical perceptual ap tantamount to a miniature conceptual revolution is used to draw some general morals about the principles underlying successful paradigm shifts. Finally, the indispensability of calculated risk-taking for creativity is argued, and the intimate relationship between biased randomness and risk-taking is discussed. 1. On the Banality of Copycats Can there be any doubt that to call someone a "copycat" is a taunt? Surely, copycats do nothing more than mechanical mimicry. Surely, a copycat could never be creative. After all, how could merely doing the same th/ngas someone else ever be original? Consider the act of translating a novel from French to English. Readers of the translation m people who don’t know a word of French -will proudly say, "Oh, yeah m I’ve read lots of Proust’, as if the switch of medium were a triviality not worth mentioning. Some will even gush over his elegant use of languagel After all, but for the words, isn’t the English version the same as the French? More sophisticated is the remark, "I once read the entire English version of A la recherche du temps perdu’. Yet even saying "the English version" still diminishes the act of translation, for it implies there is just one correct way of rendering the French into English. Many publishers the world over seem to endorse this view, either by not mentioning a book’s translator at all, or by listing the translator’s name only in fine print on the copyright page. Is someone who deserves mention only on the copyright page no more than a copycat? Or are there cases where being a copycat is a significant achievement? Surely, the reconstruction of Proust’s beautifully balanced but byzantine prose in a completely different linguistic and cultural medium is itself an act of marvelous precision requiring great skill. Surely, doing "what Proust did’, but doing it in English, is a magnificent creative act. Examples like this show that "doing the same thing" is a matter of perspective. Two things that are the same on a very abstract level may be wildly different, even incomparable, on other levels. Such mixtures of abstract sameness with concrete differentness are what analogy-making is all about. It can even be argued that this ability lies at the heart of all insightful thought (Hofstadter 1985, Chap. 24). Thus being a copycat and being creative can actually be very close. In recognition of this under-appreciated fact, the nondeterministic analogy-making computer program developed over a several-year period by Melanie Mitchell and myself is named "Copycat’. 2. A Straightforward Copycat Problem Copycat is intended to simulate human analogymaking and the mechanisms underlying it. Copycat operates in a carefully designed alphabetic microworld intended to bring out the central features of analogy-making while having very few domain-specific idiosyncrasies. That is, the reduction of analogy-making to the Copycat domain is intended to highlight the cognitive issues while deemphasizing noncognitive detail. Copycat is not just intended to be a model of analogy-making (in itself a rather daunting task); it is intended to be a model of creative analogymaking. Indeed, our ambitions in the Copycat project and related projects (see Mitchell, 1995; Hofstadter, 1983; French, 1992; McGraw & Hofstadter 1993) go even further: to reach an understanding of the fundamental mechanisms of creativity in all types of cognition. To be sure, this is not a goal we will reach; however, it is surprising how deeply one can go into the problem in the tiny microworlds that these projects involve. Before describing the architecture of Copycat, I exhibit the domain and show how the program actually performs on a couple of problems. Here is one of the simplest and most mundane analogy problems in the domain: Suppose the letter-string abc were changed to abd; how would you change the letter-string ijk in "the same way’? More concisely: Problem 1. abc =¢ abd