Eye on the Prize

AI Magazine Volume 16 Number 2 (1995) (© AAAI) Articles Eye on the Prize Nils J. Nilsson ■ In its early stages, the field of AI had as its main sufficiently powerful to solve large problems goal the invention of computer programs having of real-world consequence. In their efforts to the general problem-solving abilities of humans. get past the barrier separating toy problems Along the way, a major shift of emphasis devel- from real ones, AI researchers became oped from general-purpose programs toward per- absorbed in two important diversions from formance programs, ones whose competence was their original goal of developing general, highly specialized and limited to particular areas intelligent systems. One diversion was toward of expertise. In this article, I claim that AI is now developing performance programs, ones whose at the beginning of another transition, one that competence was highly specialized and limit- will reinvigorate efforts to build programs of general, humanlike competence. These programs will ed to particular areas of expertise. Another use specialized performance programs as tools, diversion was toward refining specialized much like humans do. techniques beyond those required for general- purpose intelligence. In this article, I specu- ver 40 years ago, soon after the birth late about the reasons for these diversions of electronic computers, people began and then describe growing forces that are Oto think that human levels of intelli- pushing AI to resume work on its original gence might someday be realized in computer goal of building programs of general, human- programs. Alan Turing (1950) was among the like competence. first to speculate that “machines will eventu- ally compete with men in all purely intellec- The Shift to tual fields.” Allen Newell and Herb Simon Performance Programs (1976) made this speculation more crisp in their physical symbol system hypothesis: “A Sometime during the 1970s, AI changed its physical symbol system [such as a digital focus from developing general problem-solv- computer] has the necessary and sufficient ing systems to developing expert programs means for general intelligent action” (empha- whose performance was superior to that of sis mine). In its early stages, the field of AI any human not having specialized training, had as its main goal the invention of comput- experience, and tools. A representative perfor- er programs having the general problem-solv- mance program was DENDRAL (Feigenbaum et ing abilities of humans. One such program al. 1971). Edward Feigenbaum and colleagues (1971, p. 187), who are credited with having was the GENERAL PROBLEM SOLVER (GPS) (Newell, Shaw, and Simon 1960), which used what led the way toward the development of have come to be called weak methods to expert systems, put it this way: search for solutions to simple problems. General problem-solvers are too weak to be used as the basis for building high performance systems. The behavior of Diversions from the Main Goal the best general problem-solvers we Many of the early AI programs dealt with toy know, human problem solvers, is problems, puzzles and games that humans observed to be weak and shallow, except sometimes find challenging but that they can in the areas in which the human prob- usually solve without special training. When lem-solver is a specialist. these early AI techniques were tried on much Observations such as these resulted in a more difficult problems, it was found that the shift toward programs containing large bodies methods did not scale well. They were not of specialized knowledge and the techniques Copyright © 1995, AAAI. All rights reserved. 0738-4602-1994 / $2.00 SUMMER 1995 9 Articles required to deploy this knowledge. The shift much beyond what I think are required by was very fruitful. It is estimated that several general, intelligent systems. I’ll give some thousand knowledge-based expert systems are examples. used in industry today. The American Associ- Let’s look first at automatic planning. It is ation for Artificial Intelligence (AAAI) spon- clear that a general, intelligent system will sors an annual conference entitled Innovative need to be able to plan its actions. An exten- Applications of Artificial Intelligence, and the sive spectrum of work on automatic planning proceedings of these conferences give ample has been done by AI researchers. Early work evidence of AI’s successes.1 I won’t try to was done by Newell, Shaw, and Simon (1960); summarize the application work here, but the McCarthy and Hayes (1969); Green (1969); following list taken from a recent article in and Fikes and Nilsson (1971). These early pro- Business Week (1992) is representative of the grams and ideas were clearly deficient in kinds of programs in operation: many respects. While working on one part of Shearson Lehman uses neural networks to a problem, they sometimes undid an already predict the performance of stocks and bonds. solved part; they had to do too much work to Merced County in California has an expert verify that their actions left most of their sur- system that decides if applicants should roundings unchanged; and they made the receive welfare benefits. unrealistic assumption that their worlds NYNEX has a system that helps unskilled remained frozen while they made their plans. workers diagnose customer phone problems. Some of the deficiencies were ameliorated by Arco and Texaco use neural networks to subsequent research (Sacerdoti 1977; Tate help pinpoint oil and gas deposits deep below 1977; Waldinger 1977; Sussman 1975). the earth’s surface. Recent work by Wilkins (1988), Currie and The Internal Revenue Service is testing soft- Tate (1991), and Chapman (1987) led to quite ware designed to read tax returns and spot complex and useful planning and scheduling fraud. systems. Somewhere along this spectrum, Spiegel uses neural networks to determine however, we began to develop specialized who on a vast mailing list are the most likely planning capabilities that I do not think are buyers of its products. required of a general, intelligent system. After American Airlines has an expert system all, even the smartest human cannot (with- that schedules the routine maintenance of its out the aid of special tools) plan missions for airplanes. the National Aeronautics and Space Adminis- High-performance programs such as these tration or lay out a factory schedule, but are all very useful; they are important and automatic planning programs can now do worthy projects for AI, and undoubtedly, they these things (Deale et al. 1994; Fox 1984). have been excellent investments. Do they Other examples of refinement occur in the move AI closer to its original, main goal of research area dealing with reasoning under developing a general, intelligent system? I uncertainty. Elaborate probabilistic reasoning think not. The components and knowledge schemes have been developed, and perhaps needed for extreme specialization are not some of these computational processes are necessarily those that will be needed for gen- needed by intelligent systems. What I think is eral intelligence. Some medical diagnosis pro- not needed (to give just one example) is a grams, for example, have expert medical dynamic programming system for calculating knowledge comparable to that of human paths of minimal expected costs between physicians who have had years of training states in a Markov decision problem, yet and practice (Miller et al. 1982). However, some high-quality AI research is devoted to these doctors were already far more intelli- this and similar problems (which do arise in gent—generally, before attending medical special settings). More examples exist in sev- school—than the best of our AI systems. They eral other branches of AI, including automat- had the ability then to acquire the knowledge ed theorem proving, intelligent database that they would need in their specialty—an retrieval, design automation, intelligent con- ability AI programs do not yet have. trol, and program verification and synthesis. The development of performance programs Ever-More–Refined Techniques and refined techniques has focused AI research on systems that solve problems In parallel with the move toward perfor- beyond what humans can ordinarily do. Of mance programs, AI researchers working on course, a program must be equipped with the techniques (rather than on specific applica- skills and knowledge that it truly needs in its tions) began to sharpen these techniques area of application. What I am arguing for 10 AI MAGAZINE Articles here is that these skills and knowledge bases Some Reasons for the Diversions be regarded as tools—separate from the intelligent programs that use them. It is time to There are several reasons why AI has concen- begin to distinguish between general, intelli- trated on tool building. First, the problem of gent programs and the special performance building general, intelligent systems is very systems, that is, tools, that they use. AI has for hard. Some have argued that we haven’t many years now been working mainly on the made much progress on this problem in the tools—expert systems and highly refined tech- last 40 years. Perhaps we have another 40 niques. Building the tools is important—no years ahead of us before significant results question. Working on the tools alone does will be achieved. It is natural for researchers not move us closer to AI’s original goal—pro- to want to achieve specific results during ducing intelligent programs that are able to their research lifetimes and to become frus- use tools. Such general programs do not need trated when progress is slow and uneven. Sec- to have the skills and knowledge within them ond, sponsors of AI research have encouraged as refined and detailed as that in the tools (and have often insisted on) specialized sys- they use. Instead, they need to be able to find tems.

Load more