AI Magazine Volume 16 Number 3 (1995) (© AAAI) Book Review

is. Chapter 5, “The COPYCAT Project: A Fluid Concepts and Model of Mental Fluidity and Analo- gy Making,” is perhaps the best Creative Analogies: A Review description of an implementation of Hofstadter’s ideas. The essential idea behind a parallel Bruce Burns terraced scan is that initially, the rep- resentation of an entity consists of a number of unrelated elements. The representation is then elaborated by many codelets, small pieces of code that do specific things. Some codelets recognize potential structures but do ■ Fluid Concepts and Creative Analo- Dave Chalmers, Daniel Defays, Bob not build them; others build struc- gies, Douglas Hofstadter and the French, Gary McGraw, and Melanie tures that have been identified. These Mitchell). Although Hofstadter is not Fluid Analogies Research Group, structures are often simple, such as the first author of all the articles, Basic Books, New York, 1995, 518 the combination of two elements. there is a coherent vision to the pp., $30, ISBN 0-465-05154-5. Because of the parallelism, compet- whole book, reinforced by the pref- ing structures can be built; other aces Hofstadter wrote for each article. here is an apt analogy for the codelets can destroy a structure that Two major ideas appear to underlie author of this book: Douglas has become inconsistent with the much of Hofstadter’s work; although THofstadter could be considered higher-level structure. The structure these ideas are related, one is primari- the Carl Sagan of AI. As Hofstadter that emerges is probabilistic because ly a computational claim, and the points out, analogies are fluid, mean- it results from interactions among other is primarily psychological. The ing that the analogy between two the actions of many codelets that are computational claim is that cogni- entities can be drawn differently executed probabilistically from a tion can be modeled using a parallel depending on how these entities are coderack of codelets: Every codelet terraced scan, which builds a represen- represented. The analogy that is tation by parallel investigation of waiting on the coderack has some drawn, in turn, can change the repre- many possibilities with different lev- nonzero probability of being execut- sentation of the entities being com- els of commitment. The psychologi- ed at any time, but once a codelet is pared. Thus, the analogy between cal claim is that cognition is a form executed, it can change the environ- Hofstadter and Sagan can be seen as of high-level perception that involves ment so as to change the probability positive: Both have explained impor- of waiting codelets being executed. tant concepts in their fields to a wide Executing a codelet can also lead to audience and transmitted the excite- other codelets being placed on the ment of these ideas. Both have coderack; for example, if one codelet inspired a number of people within recognizes a potential structure, it their fields. Unfortunately, a more Douglas Hofstadter can place a codelet on the coderack negative analogy between Sagan and could be considered the to build the structure. As coherent Hofstadter is possible. Among some structure is built, the temperature is astronomers, Sagan’s work has not Carl Sagan of AI. lowered, making it harder to execute been taken seriously. In a similar codelets that would tear down the fashion, Hofstadter’s work has had existing structure. Thus, a coherent surprisingly little impact on the field structure emerges without the need of AI. Some of the reasons for this for any top-level control. lack of acceptance in the field are A parallel terraced scan is similar a constant interaction between how based on irrelevancies. For example, in philosophy to other approaches we represent information and how Hofstadter is the only researcher in that consider representation as an AI with both a Pulitzer Prize and his we process this information. These emergent property of the interaction own news group on the Internet. themes are repeated throughout the of many lower-level processes, such However, there are some valid rea- book. Those wanting a quick tour as neural networks and genetic algo- sons for these reservations. This book should read Chapters 2, 4, and 5. rithms. The link to genetic algo- illustrates both the negative and the Chapter 2, “The Architecture of JUM- rithms seems particularly strong, so positive analogies. BO,” gives a good explanation of a much so that it would not be surpris- Fluid Concepts and Creative Analo- parallel terraced scan. Chapter 4, ing to see some attempt to learn gies is a collection of articles, many “High-Level Perception, Representa- codelets using genetic algorithms. already published, by Hofstadter and tion, and Analogy: A Critique of Such learning might address a poten- members of his Fluid Analogies Artificial Intelligence Methodology,” tial criticism of codelets; they are pre- Research Group (here represented by explains what high-level perception coded by the designer of a particular

Copyright © 1995, American Association for Artificial Intelligence. All rights reserved. 0738-4602-1995 / $2.00 FALL 1995 81 Book Review system. However, Hofstadter is less with especially tough problems” (p. competing programs have addressed interested in the exploration of the 187); rather, it provides a powerful real-world problems in science and algorithm’s properties than in the way of fleshing out the representa- politics. Hofstadter and his colleagues fact that a parallel terraced scan pro- tion of any given situation. Further, argue with some force that this criti- duces the behavior that is of real the flexibility of analogies (they allow cism is misguided because there is so interest to him—the probabilistic concepts to “slip” to related con- much information available in the emergence of structural representa- cepts) explains the fluidity of con- world that a fully developed model of tions. Hofstadter is attempting to cepts; Hofstadter therefore uses high-level perception is not possible combine the ability of neural net- analogies to account for creativity. at this stage. He contends that the works to detect similarity with the The line between high-level and relevant issues can be addressed in a power that symbolic systems get from low-level perception is fuzzy, but the restricted domain. To do so might being able to represent structure. implication Hofstadter draws for AI seem less impressive than solving The central idea behind Hof- research is clear: Too often AI has real-world problems, but programs stadter’s work is that cognition is focused on the processing of rigid that address such problems actually high-level perception, an idea cap- representations and has ignored how restrict their domains, making them tured in the slogan cognition is recog- these representations are formed and tractable. Further, the part played by nition. Low-level perception is defined how their formation interacts with the stripped-down nature of real as the processing of information from their processing. (Hofstadter exempts domains in a program’s success has the sensory modalities, and high-level machine vision and language pro- not always been acknowledged. Hof- stadter has a worthwhile point here; AI research has ignored the advan- tages of carefully constructed microdomains, but it has allowed the use of vaguely constructed domains Hofstadter makes many compelling and that are also restricted. The use of microdomains is not without prob- evocative points in this book. Why, then, lems, however, because they, too, can has his work had limited impact on the contain hidden assumptions. It is not field of AI? clear to what extent the programs presented in Hofstadter’s book suc- ceed because they avoid representa- tional issues that are relegated to low- level perception. Another possible reason why Hofs- tadter’s impact has been limited can perception is extracting the meaning cessing from this criticism because be traced to his method of engage- from the raw material of low-level they stop short of modeling pro- ment with those he disagrees with. perception by accessing concepts and cesses at a conceptual level.) It is not His form of argument is often based making sense of information at a enough to leave representation to a on analogies and rhetorical questions conceptual level. Such an approach yet-to-be-developed representation rather than the analysis of what a to cognition is not new; it was the module because the interaction program aims to achieve and how central idea of the Gestalt psycholo- between representation and process well it achieves its goals. The facility gists, and Hofstadter traces its roots is crucial. For example, Hofstadter of his writing sometimes makes it too to Kant. However, modeling cogni- argues that BACON (Langley et al. easy for him to stray in the direction tion as a form of perception con- 1987) misses the point by aiming to of ridicule. Thus, in Preface 4, The trasts with the more dominant show that given the appropriate Ineradicable ELIZA Effect and Its Dan- approach in AI, which assumes that data, the program can derive Kepler’s gers, Hofstadter claims that the cognition is a process of manipulat- third law of planetary motion— apparent successes of a number of AI ing symbols. In Hofstadter’s view, because the key to Kepler’s discovery programs are analogous to that of cognition is instead modeled as an was identifying the appropriate data. Weizenbaum’s (1976) ELIZA program, interaction between the building of Hofstadter makes many com- which appeared to be interacting representations and the manipula- pelling and evocative points in this with a person but was really only giv- tion of representations. Hofstadter’s book. Why, then, has his work had ing back canned responses. The dan- work has come to focus on analogy: limited impact on the field of AI? ger of a program’s success being based He claims we are constantly perceiv- One criticism has been that all the on its simply giving back what the ing situations in terms of what we programs created by his group have programmer built into it is well already know. Analogy, defined this addressed only microdomains: letter- known, and there is probably no pro- broadly, is not just a “heavy weapon string analogies, table settings, word gram that does not partly succeed wheeled out now and then to deal puzzles, and letter fonts. Meanwhile, because of what is built into it. How-

82 AI MAGAZINE Book Review ever, the ELIZA effect is not an all-or- point here because AI models rely none proposition, and the evaluation heavily on assumptions that are Computation & of the degree to which a program suc- difficult to test, and thus, they are ceeds for interesting reasons and to extremely hard to compare and vali- Intelligence what degree it succeeds for uninter- date. These difficulties lead to the esting reasons needs to include an Edited by George F. Luger common occurrence of rival analysis of the program’s assump- Published by the AAAI Press / researchers talking past each other. tions and goals. Such an analysis also The MIT Press needs to be applied to Hofstadter’s Nonetheless, I cannot agree with his 700 pp., $35.00 paperback programs: Do they really solve the response, which appears tantamount ISBN 0-262-62101-0 problems that they criticize others for to withdrawal from the argument. his comprehensive collection of failing to address? For example, Even if only 10 percent is not intu- 29 readings covers artificial intelli- Falkenhainer, Forbus, and Gentner’s ition, then this 10 percent should not gence from its historical roots to (1990) SME analogy-mapping program be ignored because it might be the T current research directions and practice. is repeatedly criticized because the basis for a future foundation. Further, With its helpful critique of the selections, symbols it uses are not grounded; for if you want people to build on your extensive bibliography, and clear presen- example, it would operate just as well tation of the material Computation and if the symbols water and ice were work, you must convince them that Intelligence will be a useful adjunct to any replaced with A and B. This result is your intuitions are correct; in which course in AI as well as a handy reference unsurprising given that syntax drives case, you are required to do more for professionals in the field. the mapping process in their theory. than just tell them that you are cor- The book is divided into five parts. However, the same argument turned rect and make eloquent analogies. The first part contains papers that pre- on COPYCAT is quickly dismissed If Hofstadter truly believes that AI sent or discuss foundational ideas link- because if an astute human observer models cannot be defended rational- ing computation and intelligence, continually saw the symbol SIGMA ly, then it might explain why his typified by A. M. Turing's “Computing invoked in the presence of successor analysis of other models can appear Machinery and Intelligence.” The sec- groups and successor relationships, shallow, but by doing so, he risks ond part, Knowledge Representation, then this observer would quickly presents a sampling of the numerous figure out that SIGMA stands for the irrelevance. This outcome would be representational schemes by Newell, idea of successor. But the symbol extremely unfortunate because I Minsky, Collins and Quillian, Winograd, grounding problem is not so easily believe his basic ideas might be on Schank, Hayes, Holland, McClelland, solved. How does the observer know the right track. However, figuring out Rumelhart, Hinton, and Brooks. The that the concept of successor group is what exactly his ideas mean and third part, Weak Method Problem Solv- present? Why could not an equally defending them will take more than ing, focuses on the research and design astute observer, on seeing water and Hofstadter telling us his intuitions. I of syntax based problem solvers, includ- ice, infer that the arbitrary symbol recommend reading this book ing the most famous of these, the Logic with them means melts? because it is eloquently written and Theorist and GPS. The fourth part, Rea- Hofstadter’s criticisms raise the provocative. However, at the mo - soning in Complex and Dynamic Envi- whole issue of how AI programs ronments, presents a broad spectrum of ment, Hofstadter makes his challenge should be evaluated, the topic of the AI communities' research in knowl- too easy to ignore for those who Preface 5: The Knotty Problem of edge-intensive problem solving, from Evaluating Research. In this preface, should take the most notice of it. McCarthy's early design of systems with Hofstadter considers and then dis- "common sense" to model based reason- References misses a number of AI models, often ing. The two concluding selections, by explicitly on the grounds that they Falkenhainer, B.; Forbus, K. D.; and Gent- Marvin Minsky and by Herbert Simon, fail to agree with his intuitions (for ner, D. 1990. The Structure Mapping respectively, present the recent thoughts example, the reference in the index Engine. Artificial Intelligence 41: 1–63. of two of AI's pioneers who revisit the to SOAR for this section is “SOAR pro- Langley, P.; Simon, H. A.; Bradshaw, G. L.; concepts and controversies that have gram…ambitiousness yet boringness Zytkow, J. M. 1987. Scientific Discovery: developed during the evolution of the of” [p. 515]). Hofstadter’s point is Computational Explorations of the Creative tools and techniques that make up the that ultimately, all AI research is Process. Cambridge, Mass.: MIT Press. current practice of artificial intelligence. based on intuitions that cannot be defended. However, he firmly To order, call toll free 1-800-356-0343 believes that AI is a field searching for (US & Canada) or (617) 625-8569, Mas- its foundations; so, it is not surpris- Bruce D. Burns is a postdoctoral fellow in terCard & VISA accepted. Prices will be ing that any project is based 90 per- the Department of Psychology at the Uni- higher outside the US and are subject to cent on intuition (he acknowledges versity of California at Los Angeles. He change without notice. Visit the AAAI that it is dangerous to make this has done both experimental and compu- Press website! http://www.aaai. org/Pub- point in print). Hofstadter has a tational modeling work on analogy. lications /Press/press.html

FALL 1995 83