AI by another name Mar 14th 2002

Transcription

1 Economist.com Page 1 of 4 scuta AI by another name Mar 14th 2002 Computing From The Economist print edition After years in the wilderness, the term "artificial intelligence" seems poised to make a comeback LIKE big hairdos and dubious pop stars, the term "artificial intelligence" (AI) was big in the 1980s, vanished in the 1990s-and now seems to be attempting a comeback. The term re-entered public consciousness most dramatically with the release last year of "A.I.", a movie about a robot boy. But the term is also being rehabilitated within the computer industry. Researchers, executives and marketing people are using the expression without irony or inverted commas. And it is not always hype. The t with some justification, to products that depend on technology that was originally cooked up by AI researchers. Admittedly, the comeback has a long way to go, and some firms still prefer to avoid the phrase. But the fact that others are starting to use it again suggests that AI is no longer simply regarded as an overambitious and underachieving field of research. That field was launched, and the term "artificial intelligence" coined, at a conference in 1956 by a group of researchers that included Marvin Minsky, John McCarthy, Herbert Simon and Alan Newell, all of whom went on to become leading lights in the subject. The term provided a sexy-sounding but informative semantic umbrella for a research programme that encompassed such previously disparate fields as operations research, cybernetics, logic and computer science. The common strand was an attempt to capture or mimic human abilities using machines. That said, different groups of researchers attacked different problems, from speech recognition to chess playing, in different ways; AI unified the field in name only. But it was a term that captured the public's imagination. Most researchers agree that the high-water mark for AI occurred around A public reared on science-fiction movies and excited by the growing power of home computers had high expectations. For years, AI researchers had implied that a breakthrough was just around the corner. ("Within a generation the problem of creating 'artificial intelligence' will be substantially solved," Dr Minsky said in 1967.) Prototypes of medical-diagnosis programs, speech recognition software and expert systems appeared to be making progress.

2 The 1985 conference of the American Association of Artificial Intelligence (AAAI) was, recalls Eric Horvitz, now a researcher at Microsoft, attended by thousands of people, including many interested members of the public and httn!//-%xnxnxr arnnnm;ct rnm/pr,ntavpi;anrllcr nf 9vf-,.7-1 n')n'74o 1 r7 i') AAK Economist.comPage 2 of 4 entrepreneurs looking for the next big thing. It proved to be a false dawn. Thinking computers and household robots failed to materialise, and a backlash ensued. "There was undue optimism," says David Leake, a researcher at Indiana University who is also the editor of AI Magazine, which is published by the AAAI. "When people realised these were hard problems, there was retrenchment. It was good for the field, because people started looking for approaches that involved less hubris." By the late 1980s, the term AI was being eschewed by many researchers, who preferred instead to align themselves with specific sub-disciplines such as neural networks, agent technology, case-based reasoning, and so on. The expectations of the early 1980s, says Dr Horvitz, "created a sense that the term itself was overblown. It's a phrase that captures a long-term dream, but it implicitly promises a lot. For a variety of reasons, people pulled back from using it." Ironically, in some ways, AI was a victim of its own success. Whenever an apparently mundane problem was solved, such as building a system that could land an aircraft unattended, or read handwritten postcodes to speed mail sorting, the problem was deemed not to have been AI in the first place. "If it works, it can't be AI," as Dr Leake characterises it. The effect of repeatedly moving the goal-posts in this way was that AI came to refer to blue-sky research that was still years away from commercialisation. Researchers joked that AI stood for "almost implemented". Meanwhile, the technologies that worked well enough to make it on to the market, such as speech recognition, language translation and decision-support software, were no longer regarded as AI. Yet all three once fell well within the umbrella of AI research. Quiet respectability But the tide may now be turning. "There was a time when companies were reluctant to say 'we're doing or using AI', but that's now changing," says Dr Leake. A number of start-ups are touting their use of AI technology. Predictive Networks of Cambridge, Massachusetts, focuses advertising using "artificial intelligence-based Digital Silhouettes" that analyse customer behaviour. The firm was founded by Devin Hosea, a former National Science Foundation fellow in artificial intelligence. Another firm, HNC Software of San Diego, whose backers include the. Defence Advanced Research Project Agency in Washington, DC, reckons that its new approach to neural networks based on a cluster of 30 Pentium processors is the most powerful and promising approach to artificial intelligence ever discovered. HNC claims that its system could be used to spot camouflaged vehicles on a battlefield or extract a voice signal from a noisy background-tasks humans can do well, but computers cannot. Whether or not'its technology lives up to the claims made for it, that HNC is emphasising the use of AI is itself an interesting development. Large companies are also using the term. Dr Leake points out that Bill Gates of Microsoft gave the keynote speech at last year's AAAI conference and demonstrated several Microsoft

3 technologies that are close to being incorporated into the company's products. A few months ago, Microsoft trumpeted a "breakthrough application that enlists the power of artificial intelligence to help users manage mobile communications." HAL encapsulated the optimism of the 1960s that intelligent computers would be widespread by 2001 The product in question is Mobile Manager, which uses Dr Horvitz's research into Bayesian decision-making to decide which messages in an individual's in-box are important enough to forward to a pager. Dr Horvitz says he is happy to refer to his work as AI. His current work, which involves using spare computing capacity to anticipate and prepare for the user's most likely next action, is based on research published in Artificial Intelligence. "We just submitted a paper on how a theorem-proving program could exploit uncertainty to run more Economist.comPage 3 of 4 efficiently," he says. "That's core AI. I personally feel better about using the term. There are people, myself and others, who use the term proudly." Sony also unabashedly uses the term AI when referring to its robot dog, AIBO. (The name is derived from the combination of "AI" and "bot", and means companion in Japanese.) The company boasts that "advanced artificial intelligence gives AIBO the ability to make its own decisions while maturing over time." It sounds like hype, though once you have seen an AIBO's uncannily life-like behaviour, the AI label seems appropriate. AIBO's intelligence, such as it is, relies on genetic algorithms, another trick that has been dug out from the AI toolkit. In computer gaming, the term AI has always been used with a straight face. The gaming community got interested in AI in the late 1980s when personal computers started to get more powerful, says Steven Woodcock, a programmer who has worked in both the defence and games industries, and who maintains a website devoted to the study of AI in gaming: As graphics improve, he says, a game "needs other discriminators, like whether it plays smart." Game reviews routinely refer to the quality of the AI-well, what else would you call it?-and some games are renowned for the lifelike quality of their computer opponents. Mr Woodcock says there is now quite a lot of traffic in both directions between AI programmers in the academic and gaming worlds. Military simulators, he notes, are increasingly based on games, and games programmers are good at finding quick-and-dirty ways to implement AI techniques that will make computer opponents more engagingly lifelike. Gaming has also helped to advertise and popularise AI in the form of such impressive games as "The Sims", "Black & White" and "Creatures". Information overload Another factor that may boost the prospects for AI is the demise of the dotcoms. Investors are now looking for firms using clever technology, rather than just a clever business model, to differentiate themselves. In particular, the problem of information overload, exacerbated by the growth of and the explosion in the number of web pages, means there are plenty of opportunities for new technologies to help filter and categorise information-classic AI problems. That may mean that artificial-intelligence start-ups-thin on the ground since the early 1980s-will start to emerge, provided they can harness the technology to do something useful. But if they can, there will be no shortage of buzzwords for the marketing department. Not everyone is rushing to embrace this once-stigmatised term, however. IBM, for example,

4 is working on self-healing, self-tuning systems that are more resilient to failure and require less human intervention than existing computers. Robert Morris, director of IBM's Almaden Research Centre in Silicon Valley, admits this initiative, called "autonomic computing", borrows ideas from AI research. But, he says, where AI is about getting computers to solve problems that would be solved in the frontal lobe of the brain, autonomic computing has more in common with the autonomic nerv6us system. To some extent, he suggests, the term AI has outgrown its usefulness. He notes that it was always a broad, fuzzy term, and encompassed some fields whose practitioners did not regard their work as AI. And while IBM continues to conduct research into artificial intelligence, Dr Morris does not link autonomic computing to such work. "This stuff is real," he says. Similarly, Max Thiercy, head of development at Albert, a French firm that produces naturallanguage search software, also avoids the term AI. "I consider the term a bit obsolete," he says. "It can make our customers frightened." This seems odd, because the firm's search technology uses a classic AI technique, applying multiple algorithms to the same data, and then evaluates the results to see which approach was most effective. Even so, the firm prefers to use such terms as L.,+...~~.~,..,.,.,.,.~,.,....*,.~... A nn-_ T_. Economist.comPage 4 of 4 "natural language processing" and "machine learning". Perhaps the biggest change in AI's fortunes is simply down to the change of date. The film "A.1." was based on an idea by the late director, Stanley Kubrick, who also dealt with the topic in another film, "2001: A Space Odyssey", which was released in "2001" featured an intelligent computer called HAL 9000 with a hypnotic speaking voice. As well as understanding and speaking English, HAL could play chess and even learned to lip-read. HAL thus encapsulated the optimism of the 1960s that intelligent computers would be widespread by But 2001 has been and gone, and there is still no sign of a HAL-like computer. Individual systems can play chess or transcribe speech, but a general theory of machine intelligence remains elusive. It may be, however, that now that 2001 turned out to be just another year on the calendar, the comparison with HAL no longer seems quite so important, and AI can now be judged by what it can do, rather than by how well it matches up to a 30-year-old science-fiction film. "People are beginning to realise that there are impressive things that these systems can do," says Dr Leake hopefully. "They're no longer looking for HAL." Readers wishing to discuss issues raised in this article are encouraged to post their comments on the online forum at Copyright 2006 The Economist Newspaper and The Economist Group. All rights reserved.

5 llttrr//zznznzr a~nnnmict rnm/printarrrian~ly nfrr,7ot~,.-~~ ;.a-1 n')n'7qq1 i7/)aak