Click for next page ( 48


The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement



Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.

OCR for page 47
Paper 5 ARTIFICIAL INTELLIGENCE: SUSTAINED UNIVERSITY PIONEERING WINS INDUSTRIAL ACCEPTANCE INTRODUCTION Artificial intelligence (AI) is the branch of computer science that collects and studies the so-called Tweak methods" that problem solvers can use when "strong," algorithmic methods are unknown or impractical. The truly stunning future perspectives opened by AI research have attracted some of the most imaginative and eloquent thinkers about the meaning and potential of the computer revolution. Weir enthusiasm, the press's nose for dramatic stories, and the public's natural curiosity about the possibility of fabricating intelligence make AI unusually visible and controversial among the subfields of computer science. Because of the excitement that naturally surrounds it, AI has prospered, attracting funds and students aplenty e It has contributed more than its share of the seminal ideas in modern programming practice. Yet at the same time it draws fire, for its record of overly rosy claims and unmet promises. The long-term goal of AI is to develop a theory of intelligence, or intelligent behavior, as exhibited in man and machine. Individual researchers have their own individual styles and preferences, so AI research encompasses many different activities. Broadly speaking, there are three research methodologies in AI: experimental, psychological, and theoretical. 1. Experimental--By constructing and analyzing computer programs that use weak methods to solve problems, AI researchers can study the characteristics of these methods experimentally. Once a problemrsolving method is represented by a working program, its scope and limits can be studied by varying parameters of the program or by varying characteristics of the problem to which it is applied. 2. Psychological--By studying human problem-solving behavior, AI researchers can gain insight into heuristic methods. men by constructing programs that embody those methods, they construct psychological models that can be tested and refined. 3. Theoretical--By formalizing key concepts of problems and problem-solving methods, AI researchers can examine the logical properties of heuristic methods, such as the consistency and completeness of inference calculi, or the efficiency of search techniques. 47

OCR for page 47
48 An essential ingredient of artificial intelligence research is a body of heuristics, or judgmental rules that it has collected. These provide good guesses--justifiable but not infallible guides for problem solvers and algorithmic methods known for some classes of problems--and should be used when appropriate. However, algorithms can be more expensive to use than heuristics. Indeed, efficient algorithms are completely lacking for many problems studied by AI researchers, from the problem of answering homework questions to that of finding answers to life and death questions of medical diagnosis and therapy. The structure of AI is far from rigid, but research has focused on a few major topics: How to encode knowledge (both common-sense knowledge and specialized expertise) for efficient use by programs. How to control the use of knowledge for making plausible inferences. How to recognize objects in visual scenes. How to control robot "actuators" (vehicles, mechanical manipulators, etc.) How to create and use plans of action. How to modify or extend a program's knowledge base (e.g., as a result of learning from experience, learning by being told, rote learning, concept formation, or learning by other means). How to understand and generate English-language sentences. How to prove theorems efficiently, and how to use theorem provers as inference engines for problem solving (especially question answering). How to model human problemrsolving methods. . These questions, and variations of them, have motivated AI research since the beginning. They can all be seen as instances of the central question of AI: How can we give computers those capabilities that are taken as evidences of intelligence when displayed by humans? Many of these themes come together in current investigation into Expert systems. n In an expert system, a serious attempt is made to codify judgmental methods in some specialized field--for example, chemical synthesis. An expert system typically comprises a loosely organized and easily modifiable collection of rules, a more or less specialized reasoning program that uses the rules as needed, and a mechanism for explaining any outcome that may be reached. WHERE IS AI RESEARCH DONE? Named before computer science itself, AI remained in the hands of just a dozen or so researchers (those named in Computer and Thought, Feigenbaum and Feldman, editors, McGraw-Hill, 1964, and a few others) until the late 1960s. With a few exceptions, AI research to date has been carried out at universities, especilly CMU, MIT, and Stanford, which enjoy massive federal sponsorship. Though industrial interest in AI is growing

OCR for page 47
49 rapidly, little support for or output in this field came historically from computer manufacturers. However, this picture has begun to change. A recent tabulation (ACM SIGART, April 1978) listed 20 U.S. corporate AI research locations. m e ARPANET has benefited from sustained Advanced Research Projects Agency and Office of Naval Research funding and from good intersite communications facilities--and a national computing resource for AI research in medically related problem areas has been serving a geographically distributed research community since 1975. m is Sumex-AIM resource, funded by the Division of Research Resources of the National Institutes of Health and located at Stanford, gives individual researchers access to a large, interactive computing environment at only the cost of a terminal and modem. m us, even though applications of AI to medicine and the medical sciences require large capital outlays, access to the equipment and to colleagues can be provided to qualified researchers anywhere in the country, without relocating them. Artificial intelligence has always been the most speculative area of computer science. From the start, its focus was on building programs that were out of the reach of conventional programming methodology. In the course of doing this, researchers sometimes have been driven to invent new programming concepts that then become important in other areas of computer science. Examples are alpha-beta search, functional programming, time sharing, list processing, and garbage collection. Until recently, the impact of Al on computer science and industrial computing has been largely through a direct transfer of programs. So far, these byproducts of AI research, of which MIT's MAC SYMA is another example, have been more important than the more direct transfer of products of AI research. Nevertheless, some direct products of AI research enjoy significant use today. Stanford's Dendral program has been repeatedly used by chemical companies to aid in the analysis of organic compounds. Some firms have adopted, or are experimenting with, techniques drawn from work with expert systems: notably DEC for configuring computers and Schlumberger for oil processing. On the other hand, industrial use of computer-aided decision and robots seems to owe little to AI research, despite the historical interest of AI researchers in these topics. It appears that these industrial efforts depend primarily on classical science and engineering. This should not be surprising: one can hardly expect methods suggested by the weak Al paradigm to regenerate the codified wisdom of the centuries on the spur of the moment. Only by assiduously appropriating that wisdom (as MACSYMA and SYNCHEM 2 have attempted to do) can AI researchers hope to succeed in fields of hard science or technology. POTENTIAL FOR TECHNOLOGY TRANSFER Frugal or efficient use of machines and systems capable of handling production-scale problems has seldom been a central concern of Al research. As a result, there often arises a certain tension between 1

OCR for page 47
50 its values and style and those of industrial development. Although industry's willingness to experiment with AI techniques is now growing very rapidly, considerable skepticism remains in industry about the complexity, size, and usability of AI programs. In contrast, the Japanese are investing $100 million in the fifth-generation computer project with the expectation that the difficulty of using AI techniques can be overcome. The majority of this committee believes, however, that crash funding of this area does not make sense. The academic AI community has recently become more conscious of the potential interest in AI within industry. Given this motivation, we believe that many methods drawn from Al research will be tried, and they will prosper or not according to the laws of economics. CONCLUSIONS Without knowing the nature of intelligence, how far can one push current AI techniques for representing knowledge for use by programs and for programming machines to use that knowledge? Researchers in AI are looking at the limits of what can be done. Industry may be expected to use the practicable results to build programs that succeed in real-world, complex environments. Modest demonstrations are in progress; we see no significant institutional barriers to their maturing.