, West Virginia University
Pages: pp. 18-24
"Take pride in how far you have come; have faith in how far you can go." —Anonymous
In the 21st century, AI has many reasons to be proud, but it wasn't always this way. New technologies such as AI typically follow the hype curve (see Figure 11). By the mid-1980s, early successes with expert systems 2-5 caused skyrocketing attendance at AI conferences (see Figure 2) and a huge boom in North American AI startups. Just like the dot-coms in the late 1990s, this AI boom was characterized by unrealistic expectations. When the boom went bust, the field fell into a trough of disillusionment that Americans call the AI Winter. A similar disillusionment had already struck earlier, elsewhere (see the " Comments on the Lighthill Report" sidebar).
Figure 1 The hype cycle for new technology.
Figure 2 Attendance at the National Conference on Artificial Intelligence (AAAI), the senior American AI conference. Figures prior to 1984 are not available. No figures are shown for 1985, 1989, and 2001 because these were IJCAI (International Joint Conference on Artificial Intelligence) years.
If a technology has something to offer, it won't stay in the trough of disillusionment, just as AI has risen to a new sustainable level of activity. For example, Figure 2 shows that although AI conference attendance numbers have been stable since 1995, they are nowhere near the unsustainable peak of the mid-1980s.
With this special issue, I wanted to celebrate and record modern AI's achievements and activity. Hence, the call for papers asked for AI's current trends and historical successes. But the best-laid plans can go awry. It turns out that my "coming of age" special issue was about five to 10 years too late. AI is no longer a bleeding-edge technology—hyped by its proponents and mistrusted by the mainstream. In the 21st century, AI is not necessarily amazing. Rather, it's often routine.
Evidence for AI technology's routine and dependable nature abounds. For example, in this issue (see the related sidebar for a full list), authors describe various tools to augment standard software engineering:
In other work, the AI field has generated many mature tools that are easily used, well documented, and well understood. For example, late last year, one of my undergraduate research assistants mentioned nonchalantly that he'd just run some data through four different data miners! That student was hardly a machine learning expert—and in the 21st century, he didn't need to be. The Waikato Environment for Knowledge Analysis toolkit (see Figure 3) contains dozens of state-of-the-art data miners, all tightly integrated around the same underlying database and object model. Weka is free, open source, well documented, 6 compatible on many platforms, and easy to install (it took my student less than three minutes to download, install, and start running the learners). You can access it at www.cs.waikato.ac.nz/~ml/weka/index.html.
Figure 3 The Waikato Environment for Knowledge Analysis tool.
Natural language processing is another example of AI's success. In times past, NL processing was very difficult with a low chance of success. These days, researchers can rely on numerous tools to build successful NL applications. For example, NL processing often requires extensive background knowledge of the words being processed. Many general ontologies are now freely available. These public-domain ontologies range from WordNet (a lexical database for English) to OpenCyc (a formalization of many commonsense concepts). More specific ontologies are also freely available, such as the Unified Medical Language System (see Figure 4).
Figure 4 Part of the Unified Medical Language System semantic network ( www.nlm.nih.gov/research/umls/META3.HTML). Each child in the hierarchy is linked to its parent by the isa link.
Overall, the ontologies are extensive. For example, WordNet covers 111,223 English words, and UMLS's January 2003AA edition includes 875,255 concepts and 2.14 million concept names in over 100 biomedical source vocabularies, some in multiple languages. Building such ontologies is a huge task, and David Schwartz (in this issue) discusses a global initiative to build semantic dictionaries via the World Wide Web.
Apart from ontologies, executable NL tools are also readily available. For example, Debbie Richards recently led a small university team that implemented a system to detect contradictions between different NL sentences in an object-oriented design. 7 In the 1980s, such software would have only been found in science fiction. But in the 21st century, the Richards team had very little NL processing to implement. They just added (for example) negotiation tools to standard NL components. Those components included an answer extraction system from NL, a formal concept analysis component that generates a visualization of the text, and Prolog and Java tools that implemented the other components. This menagerie of tools seems complex. However, AI components are now mature enough to enable simple combinations.
AI still can't be smug, despite the successes listed in the " AI Applications" sidebar. Although some AI areas are mature, there's still much to learn and some traps to avoid.
In his invited talk at AAAI 1999, Nils Nilsson argued that the easy days of AI are over:
The easy work (such as inventing A* and the idea of STRIPS operators), is over. AI is getting harder. In addition, AI researchers will have to know a lot about many related disciplines.
Nilsson offered Figure 5 as a partial list of related disciplines. He warned against a fission effect, which could tear apart the field. Paradoxically, this effect results from AI's success:
Fission is promoted by the tendency of AI to be pulled apart by the many adjacent disciplines that join with AI to field practical, large-scale applications in specialized niches.
In fact, some computer scientists and others might go so far as to say, "Why do we need AI as a separate field? One could carve it up, add the parts to adjacent fields and get along perfectly well without it."
Figure 5 Near neighbors to AI.
Nilsson then proposed several large challenge problems to maintain a coherent field of study in AI. For the record, his hot list of near-term research includes case-based reasoning (again) for planning; using logic (again) for planning; SAT encodings of planning problems; large, reusable knowledge bases; agent (robot and softbot) architectures; agent-to-agent communication languages; more expressive Bayes nets; Bayes net learning; and genetic algorithm and programming techniques.
Although Nilsson's comments are timely, I'm more confident than he about AI's future as a coherent discipline. The long-term goal of emulating general human intelligence remains, and that goal will bind future generations of AI researchers. The successes listed in the " AI Applications" sidebar show that you can achieve much without human-level sophistication. Nevertheless, I still dream of the day when my word processor writes articles like this one while I go to the beach.
The goal of creating general human-level intelligence has inspired, and still inspires, decades of talented graduate students who flock to the hardest problem they know. These students strive to distance themselves from those working on other well-defined, mostly solved problems. Hence, these students are always proud to boast that they are working on AI.
It's bad manners to form an army if you can't feed them. But our AI graduate students won't starve. As they work toward the long-term goal of human-level intelligence, they'll still be able to pay the rent using AI, for example, by working in the emerging gaming industry. This industry is already huge (approximately 17 billion dollars revenue in 2002 8) and is still growing. Our AI workers will stay busy building the next generation of gaming softbots. 9 As the World Wide Web grows, these softbots will have access to "eyes" that can see more information than any human intelligence could see in a lifetime. As we continue to use software to control our world, these softbots will be given increasingly sophisticated "arms." With these eyes and arms, systems have a rising opportunity to learn and influence the world.
For another AI meal ticket, consider the growing field of model-based software engineering. Safety concerns are forcing the aviation industry to use MBSE. More planes are flying each day, but the odds of a software error are constant. Unless we can reduce the rate of software errors, by 2030 there will be a major air traffic accident reported daily. MBSE tools allow for early life-cycle software simulation, verification, and validation. Furthermore, they remove the need for laborious, possibly error-prone, manual code generation. So, the aviation industry is rapidly maturing MBSE. Soon, the broader software- engineering community will be able to access and use MBSE tools. When that happens, accurate declarative descriptions of all software will exist. Bring on the AI! For example,
Will MBSE or softbot manufacturers use the term AI? Of course! A rose by any other name is still implemented using AI. Softbots will use search methods and data mining techniques. MBSE will still need to understand its knowledge representations' logic. These systems will integrate via the high-level languages we developed using ontologies we built and debugged. As we struggle to implement, understand, and optimize MBSE-built agents running around the World Wide Semantic Web, researchers will still rush to read the latest issues of Artificial Intelligence Journal, the proceedings from AAAI and IJCAI, and (of course) IEEE Intelligent Systems.
Modern AI workers can be very proud. Much has been accomplished. We have survived the birth trauma of this new technology. We have developed tools that enabled numerous landmark applications. We have matured those tools into dependable and reusable components. And we still inspire the smartest minds to work on the hardest problems.
As proud as we are, we mustn't be smug. Consider the list of landmark events shown in Table 1. Compared to any of those, is AI remarkable enough to be memorable in, say, 200 years time? I think not—but that can change. AI's mark in history could be prominent and permanent if the 21st century becomes the birthday of this planet's second intelligent race. Your own work—past, present, and future—will decide.
In the 1970s, the Lighthill Report convinced the British government to end support for AI research in nearly all British universities. With hindsight, Lighthill's pessimism was unfounded. Even at its release, some strongly criticized the report, including John McCarthy (now professor emeritus of computer science at Stanford University):
The Lighthill Report argued that if the AI activities … were any good they would have had more applied success by then. In the 1974 Royal Institution debate on AI, I attempted to counter by pointing out that hydrodynamic turbulence had been studied for 100 years without full understanding. I was completely floored when Lighthill replied that it was time to give up on turbulence. Lighthill's fellow hydrodynamicists didn't give up and have made considerable advances since then. I was disappointed when BBC left that exchange out of the telecast, since it might have calibrated Lighthill's criteria for giving up on a science.1ReferenceJ.McCarthy"Lessons from the Lighthill Flap,"2000,www-formal.stanford.edu/jmc/reviews/lighthill-20/lighthill-20.html.
by John R. Koza, Martin A. Keane, and Matthew J. Streeter pp. 25-31. As computer time becomes cheaper, genetic programming will be routinely used as an invention machine to produce useful new designs, generate patentable new inventions, and engineer around existing patents.
by Bernhard Peischl and Franz Wotawa, pp. 32-37. Modern model-based reasoning technology is fast enough to be applied to small to medium-sized programs.
by Bernd Heisele, pp. 38-42. Vision systems that learn and adapt represent one of the most important trends in computer vision research and might provide the only solution to the development of robust and reusable vision systems.
by Yunwen Ye, pp. 43-47. Programmers can miss the components they need in large component libraries. CodeBroker is an intelligent software agent that can automatically find the components that a programmer misses.
by Gary D. Boetticher, pp. 48-50. The international $300-billion software development industry needs better predictors for software development costs. Data miners can learn such predictors to an impressive level of accuracy.
by David G. Schwartz, pp. 52-58. Millions of people around the world are writing Web pages. The semantic content of those pages is usually inaccessible. The Semantic Web is a global initiative to dramatically improve how we structure and share content on the Web.
AI has made much progress in specific application areas—for example:
John Koza, Martin Keane, and Matthew Streeter (in this issue) discuss how genetic programming can duplicate human invention. They find they can reengineer new solutions to solve the same problems addressed by state-of-the-art patents.
Bernd Heisele (in this issue) describes the state-of-the-art in recognizing images from video. Soon, AI police agents will be able to monitor large crowds.
Since the 1970s, medical expert systems have been achieving human levels of medical expertise. 1 Such systems are now used daily and trusted around the world. For example, automatic tools for assessing electrocardiograms are now so good that it's routine for humans to pay for their services. 2
During the 1980s and 1990s, DEC used the XCON (e xpert configurer) expert system to automatically configure computer hardware components. 3-5 This system saved DEC millions of dollars a year and freed up designers to work on next-generation DEC computers.
Since the 1990s, numerous researchers have reported that previous pessimism regarding AI's scalability might be unfounded. Stochastic inference procedures enable the processing of declarative theories that are orders of magnitude larger than anything previously processable. 6-8 These results remove one of the fundamental objections to AI made in the Lighthill Report (see the related sidebar).
In 1991, Iraq invaded Kuwait, so America invaded Iraq in Operation Desert Storm. DARPA claimed they saved more money using AI logistics planners during Desert Storm than was ever invested over decades into AI research. 9
In 1995, an automatic vision system steered a vehicle across America. The ALVINN system from Carnegie Mellon University autonomously drove a van, Navlab5, from Pittsburgh to San Diego while human operators worked the brake and accelerator. The system controlled the steering for all but 52 miles of 2,849-mile journey, averaging 63 mph day and night in all weather.
In 1997, Deep Blue—an IBM supercomputer—defeated world chess champion Garry Kasparov in six games. It was the first time a computer had won a chess match against a current world champion under tournament conditions. Kasparov won the first game, lost the second, and played to a draw in the next three. In game six, "Kasparov got wiped off the board," according to grand master Ilya Gurevich ( www.cnn.com/SPECIALS/1997/year.ender/scitech/06.deepblue.html).
In 1999, a NASA AI agent (see Figure A) ran a satellite beyond Mars for over a day, without ground control. The agent continually reviewed and updated the mission goals according to the satellite's functioning hardware. 10 Earthlings could subscribe to a mailing list and get frequent bulletins from the satellite about what it was currently thinking about (see Figure B).
Figure A Close-up of the Remote Agent, ion engine firing. (courtesy of NASA Ames/JPL)
Figure B Text of an email converted from telemetry sent by the DS1 spacecraft's Remote Agent, an autonomy experiment developed at NASA Ames and JPL.
The list of spin-offs from AI labs is impressive and includes the mouse; time-sharing operator systems; high-level symbolic programming languages (Lisp, Prolog, Scheme); computer graphics; the graphical user interface; computer games; the laser printer; object-oriented programming; the personal computer; email; hypertext; software agents crawling the Web; and symbolic mathematics systems such as Macsyma, Mathematica, Maple, and Derive.ReferencesV.Yuet al.,"Antimicrobial Selection by a Computer: A Blinded Evaluation by Infectious Disease Experts,"J. American Medical Assoc., vol. 242, no. 12,21Sept.1979,pp. 1279-1282.J.Willemset al.,"The Diagnostic Performance of Computer Programs for the Interpretation of Electrocardiograms,"The New England J. Medicine, vol. 325, no. 25,19Dec.1991,pp. 1767-1773;abstract: http://content.nejm.org/cgi/content/abstract/325/25/1767.J.McDermott"R1's Formative Years,"AI Magazine, vol. 2, no. 2,Summer1981,pp. 21-29.J.BachantandJ.McDermott"R1 Revisited: Four Years in the Trenches,"AI Magazine, vol. 5, no. 3,Fall1984,pp. 21-32.J.McDermott"R1 ("XCON") at Age 12: Lessons from an Elementary School Achiever,"Artificial Intelligence, vol. 59, nos. 1-2,Feb.1993,pp. 241-247.P.CheesemanB.KanefskyandW.Taylor"Where the Really Hard Problems Are,"Proc. Int'l Joint Conf. Artificial Intelligence(IJCAI 91), AAAI Press,1991,pp. 331-337.H.KautzandB.Selman"Pushing the Envelope: Planning, Propositional Logic and Stochastic Search,"Proc. 13th Nat'l Conf. Artificial Intelligence and the 8th Innovative Applications of Artificial Intelligence Conf.,AAAI Press/MIT Press,1996,pp. 1194-1201;www.cc.gatech.edu/~jimmyd/summaries/kautz1996.ps.D.OwenandT.Menzies"Lurch: A Lightweight Alternative to Model Checking," to be published in Proc. 15th Int'l Conf. Software Eng. and Knowledge Eng.(SEKE 03), World Scientific,2003;http://tim.menzies.com/pdf/03lurch.pdf.Maxwell Air Force Base Introduction to Artificial Intelligence, www.au.af.mil/au/aul/school/acsc/ai02.htm.N.Muscettolaet al.,"Remote Agent: To Boldly Go Where No AI System Has Gone Before,"Artificial Intelligence, vol. 103, nos. 1-2,Aug.1998,pp. 5-48.
Nils Nilsson, Enrico Coiera, and numerous contributors to the comp.ai newsgroup kindly shared their lists of landmark AI applications. Also, Nigel Shadbolt and Chris Welty offered useful and timely advice during this issue's planning.
I conducted this research at West Virginia University under NASA contract NCC2-0979 and NCC5-685. The NASA Office of Safety and Mission Assurance under the Software Assurance Research Program led by the NASA Independent Verification and Validation Facility sponsored this work.