AI Magazine Volume 26 Number 4 (2006)(2005) (© AAAI) 25th Anniversary Issue Artificial Intelligence: The Next Twenty-Five Years

Edited by Matthew Stone and Haym Hirsh

■ Through this collection of programmatic state- that celebrates the exciting new work that AI ments from key figures in the field, we chart the researchers are currently engaged in. At the progress of AI and survey current and future direc- same time, we hope to help to articulate, to tions for AI research and the AI community. readers thinking of taking up AI research, the rewards that practitioners can expect to find in our mature discipline. And even if you’re a stranger to AI, we think you’ll enjoy this inside e know—with a title like that, you’re peek at the ways AI researchers talk about what expecting something awful. You’re they really do. on the lookout for fanciful prognos- W We asked our contributors to comment on tications about technology: Someday comput- goals that drive current research in AI; on ways ers will fit in a suitcase and have a whole our current questions grow out of our past re- megabyte of memory. And you’re wary of lurid sults; on interactions that tie us together as a Hollywood visions of “the day the robots field; and on principles that can help to repre- come”: A spiderlike machine pins you to the sent our profession to society at large. We have wall and targets a point four inches behind your eyes with long and disturbing spikes. Your excerpted here from the responses we received last memory before it uploads you is of it ask- and edited them together to bring out some of ing, with some alien but unmistakable existen- the most distinctive themes. tial agony, what does this all mean? We have grouped the discussion around five We are not here to offer a collection of fic- broad topics. We begin in the first section tion. Artificial intelligence isn’t a nebulous goal (“Progress in AI”) by observing how far AI has that we hope—or fear—humanity one day come. After decades of research, techniques de- achieves. AI is an established research practice, veloped within AI don’t just inform our under- with demonstrated successes and a clear trajec- standing of complex real-world behavior. AI tory for its immediate future. David Leake, the has advanced the boundaries of computer sci- editor-in-chief of AI Magazine, has invited us to ence and shaped users’ everyday experiences take you on a tour of AI practice as it has been with computer systems. We continue in the playing out across a range of subcommunities, second section (“The AI Journey: Getting around the whole world, among a diverse com- Here”) by surveying contributors’ experience munity of scientists. doing AI. This section offers a more intimate For those of us already involved in AI, we glimpse inside the progress that AI has made as hope this survey will open up a conversation our contributors explain how their approaches

Copyright © 2005, American Association for Artificial Intelligence. All rights reserved. 0738-4602-2005 / $2.00 WINTER 2005 85 25th Anniversary Issue

to central questions in the field have another’s interests and new research to been formed and transformed over the explore together. As this research brings years by their involvement in research results, it cements these intellectual con- and in the AI community. nections in new kinds of computer sys- The third section (“The AI Journey: Fu- tems and in a deeper understanding of ture Challenges”) looks forward to char- human experience. We’re pleased and acterize some of the new principles, tech- honored to present just a slice of a snap- niques, and opportunities that define the shot of this dynamic process. Our only ongoing agenda for AI. By and large, our prediction for the next 25 years is that it contributors expect to take up research will continue to bring us unexpected in- challenges that integrate and transcend sights and connections to one another. the traditional subfields of AI as part of Every member of AAAI ideally deserves new, inclusive communities organized to contribute to this article. But, obvious- around ambitious new projects. They en- ly, space limits do not make that possible, vision big teams solving big problems— and thus we solicited feedback from far designing new fundamental techniques too small a number of the many people for representation and learning that fit who are playing leadership roles in our Jim Hendler new computational models of percep- field. We don’t imagine that what we tion, human behavior and social interac- have is definitive; we have set up a web- tion, natural language, and intelligent ro- site, at www.ai.rutgers.edu/aaai25, to bots’ own open-ended activity. The continue the discussion. Your contribu- fourth section (“Shaping the Journey”) tion remains welcome. considers the community building that can make such investigations possible. Progress in AI We have a lot to do to facilitate the kind of research we think we need, but a lot of AI is thriving. Many decades’ efforts of experience to draw on—from creating re- many talented individuals have resulted sources for collaborative research, to es- in the techniques of AI occupying a cen- tablishing meetings and organizations, to tral place throughout the discipline of training undergraduates and graduate computing. The capacity for intelligent students and administering research in- behavior is now a central part of people’s stitutes. We close, in the fifth section understanding of and experience with (“Closing Thoughts”), with some re- computer technology. AI continues to minders of why we can expect takers for strengthen and ramify its connections to this substantial undertaking: the excite- neighboring disciplines. The multifaceted nature of AI today is ment, the satisfaction, and the impor- tance of creating intelligent artifacts. a sign of the range and diversity of its suc- Across the statements contributors cess. We begin with contributions observ- sent us, we see how thoroughly AI has ing the successes we often neglect to matured into a collaborative process of mention when we consider AI’s progress. genuine discovery. AI now seems quite Last night I saw a prescreening of the different from what would have been movie Stealth, the latest science fiction predicted twenty-five years ago: that fact flick in which an artificially intelligent only highlights the progress we have machine (in this case an unmanned made. AI continues to embrace new per- combat air vehicle) stars in a major role. The movie has a couple of scenes that spectives and approaches, and we contin- pay homage to the 1968 film 2001: A ue to find new interplay among the com- Space Odyssey, in which HAL, the ad- plementary principles required for vanced AI computer, is a key character. general intelligence. What amazed me in the new movie was Our progress itself thus helps tie us to- a realization of just how far AI has come. gether as a community. Whenever new When I saw 2001, the idea of the talking confluences of ideas provide fertile new computer that understood language was ground for theory and experiment, they so cool that I decided then and there reconnect AI researchers into an evolving that I wanted to be an AI scientist some- day. In Stealth, the AI carries on normal network of mutual support, friendship, conversation with humans, flies a high- and fun. Even 50 years after the Dart- powered airplane, and shows many hu- mouth conference, and 25 years after the man-level capabilities without it really founding of AAAI, AI researchers contin- raising an eyebrow—the plot revolves ue to find new points of overlap in one around its actions (and emotions), not

86 AI MAGAZINE 25th Anniversary Issue around how “cool” it is that a computer reasoning, in machine learning, and can do these things. more. On the practical side, AI methods From the point of view of the AI vi- now form a key component in a wide va- sion, we’ve already achieved many of the riety of real-world applications. things the field’s founders used for moti- —Daphne Koller, vators: for example, a computer beat the world’s chess champ, commercial sys- Fifty years into AI’s U.S. history and 25 tems are exploiting continually improv- years into AAAI’s history, we’ve come a ing voice and speech capabilities, there long way. There are a wide variety of de- are robots running around the surface of ployed applications based on AI or incor- Mars, and the word processor I’m using porating AI ideas, especially applications to write this comment helps to correct involving machine learning, data min- my grammar mistakes. We’ve grown ing, vision, natural language processing, from a field with one conference to one planning, and . A large fraction in which many subareas hold well-at- of the most exciting opportunities for re- tended conferences on a regular basis search lie on the interdisciplinary bound- and in which it is rare to see a university aries of AI with (sys- that does not include AI in its undergrad- tems, graphics, theory, and so on), uate curriculum. We in the field are biology, linguistics, engineering, and sci- sometimes too fast to recognize our own ence. Vastly increased computing power faults and too slow to realize just how has made it possible to deal with realisti- amazingly far we’ve come in such a short cally large though specialized tasks. time. —, Columbia University —Jim Hendler, University of Maryland The AI success was once composed of a has enjoyed tre- list of offshoot technologies, from time mendous success over the last twenty five sharing to functional programming. Now years. Its tools and techniques are now it is AI itself that is the contribution. mainstream within computer science and at the core of so many of the systems we use every day. Search algorithms, the The AI Journey: backbone of traditional AI, are used throughout operating systems, compilers, Getting Here and networks. More modern machine- The wide-ranging and eclectic sweep of learning techniques are used to adapt these same systems in real-time. Satisfia- AI as a field is mirrored in the experiences bility of logic formulas has become a cen- of individual researchers. A career in AI is tral notion in understanding computabil- a license to liberally explore a range of ity questions, and once esoteric notions problems, a range of methods, and a like semantic ontologies are being used to range of insights—often across a range of Daphne Koller power the search engines that have be- institutions. For many, AI’s successes rep- come organizers of the world’s knowl- resent a very personal journey. We are edge, replacing libraries and encyclope- pleased that many of our contributors of- dias and automating business interfaces. fered an inside look at their journeys And who would have guessed that AI- powered robots in people’s homes would through AI. Thus, Alan Mackworth and now be counted in the millions? So much Ruzena Bajcsy—in recounting the chal- accomplishment to bring pride to us all. lenges of bridging discrete, symbolic rea- —Rodney Brooks, MIT soning with the continuous mathematics of signal processing and control theo- From the engineering perspective, artifi- ry—highlight how new concepts of con- cial intelligence is a grand success. In edu- straint satisfaction and active perception cation, computer science majors expect to take a subject or two in artificial intelli- clarified their research programs. Similar- gence, and prospective employers expect ly, Bruce Buchanan describes his evolving it. In practice, big systems all seem to con- research into the design of systems that tain elements that have roots in the past solve complex real-world problems half century of research in artificial intel- through insights he gained about the in- ligence. terplay of represented knowledge and cal- —Patrick Henry Winston, MIT culated inference. In many cases, newly discovered in- The AI community has cause for much David Waltz pride in the progress it has made over the sights lead us to radically change the past half century. We have made signifi- work we do and the way we talk about it. cant headway in solving fundamental Consider the range of research domains problems in representing knowledge, in Aaron Sloman has explored, or watch as

WINTER 2005 87 25th Anniversary Issue

Michael Kearns and Usama Fayyad trace straints can be static or dynamic. Our de- quite different evolving trajectories velopment of the robot soccer challenge through machine learning. As we survey has forced all of us to develop architec- emerging research in AI in the third sec- tures supporting both proactive and reac- tive behaviors. tion (“The AI Journey: Future Chal- lenges”), we’ll see that new problems, —Alan Mackworth, new discoveries, and new technologies University of British Columbia continue to forge new intellectual con- I came from to the Stan- nections and create new directions for re- ford AI Laboratory in October 1967. This search in AI. This ongoing interplay, in laboratory was one of the three AI labs in Wolfgang Wahlster’s experience, both de- the USA and was under the leadership of fines and sustains AI research. John McCarthy. The basic philosophy of Across the span of a career in AI, the Professor McCarthy was that AI was about representation of knowledge and formative mentorship that starts us off that this representation was symbolic. has a special place, of course. Ruzena Ba- The language we used for the representa- jcsy points to the eclectic good taste of tion was Lisp. To his credit, McCarthy John McCarthy, while Aaron Sloman was recognized that perception and robotic Alan Mackworth won over by Max Clowes’s passionate ad- interaction with the environment was vocacy of the fundamental insights in AI equally important as reasoning strictly and Usama Fayyad was captivated by the on symbolic information. Hence we romance and energy with which AI was faced the problem of how to systemati- taught at the University of Michigan. Our cally convert the measurements or obser- work is also strongly shaped by our ac- vations into symbols. What is an edge, straight line, circle, cube, and so on? This quaintance with particularly thought- is still an open problem. provoking research, as Bruce Buchanan’s The tradition that was set at that time (and the field’s) was by Turing’s empiri- (and it has prevailed) is the foundation of cism of the 1950s, as Alan Mackworth’s a good engineering science: every good was by the seminal computer vision re- theory needs experimental verification. search of the 1970s, and as Michael As we go on and understand more com- Kearns’s was by some of the early mathe- plex phenomena, the experiments reflect matics of computational learning theory this complexity. in the 1980s. We can only hope that our I implemented this tradition in the personal and intellectual efforts to sus- GRASP laboratory during my thirty years at the University of Pennsylvania in tain AI as a community continue to act so . Furthermore, coming from a powerfully. background of control engineering, we As a young scientist, I found AI’s con- recognized the need in building intelligent Ruzena Bajcsy stant ferment exciting, and I still do. I systems, the importance of controlling the had previously worked in cybernetics, data acquisition, and introduced an new , and pattern recognition, paradigm: active perception. We stated that where we modeled intelligence, percep- we not just see but we also look, and we tion, and action as signal processing. not only touch but we also feel. However, that view excluded much of —Ruzena Bajcsy, University of what we know intelligence to require, California at Berkeley such as symbolic cognition. Modeling cognition as symbolic computation pro- Every empirical science needs both theo- vided a missing link. But we went too far reticians and experimenters. Turing saw in modeling intelligence as only symbol- that operational tests of behavior would ic. One of our toughest challenges now is be more informative than arguing in the to develop architectures that smoothly abstract about the nature of intelligence, combine the symbolic and the subsym- which established the experimental na- bolic. Or, if you like, to synthesize the ture of AI. achievements of logicist AI with those of The two major research themes for cybernetics, control theory, neural nets, both theoretical and experimental AI artificial life, and pattern recognition. have always been knowledge representa- Inspired initially by David Waltz, Ugo tion (KR) and inference. Clearly an intel- Montanari, David Huffman, Max ligent person or program needs a store of Clowes, and David Marr, I’ve advocated knowledge and needs inferential capabil- Bruce Buchanan constraint satisfaction as the unifying ities to arrive at answers to the problem model. At both the symbolic and sub- he/she/it faces in the world. Other big is- symbolic levels we can specify the inter- sues, like learning and planning, can be nal, external, and coupled constraints seen as secondary to KR and inference. that agents must satisfy. Those con- Ed Feigenbaum and I were early play-

88 AI MAGAZINE 25th Anniversary Issue ers in two major controversies: (1) what work with Les Valiant on the field that are the relative contributions of knowl- would shortly become known as compu- edge and inference, and (2) what repre- tational learning theory but that at the sentation methods are both simple time consisted exclusively of two algo- enough to work with and sophisticated rithmically focused papers by Valiant, enough to capture the kinds of knowl- and an early draft of the rather mind- edge that experts use? The DENDRAL bending (to a first-year graduate student, and MYCIN programs provide experi- at least) “four Germans” paper on the ex- mental evidence on the side of more otic and powerful Vapnik-Chervonenkis knowledge, represented simply. dimension.3 It was a great time to enter —Bruce Buchanan, the field, as virtually any reasonable University of Pittsburgh problem or model one might consider was untouched territory. AI is today routinely employed in so Now that the field is highly devel- many areas of advanced information oped (with even many unreasonable technology that it is fair to say that AI problems sporting hefty literatures), I stands also for avant-garde informatics, think that the greatest sources of innova- since it is always pushing informatics to tion within computational learning the- its limits. For the steady growth of AI in ory come from the interaction with the Wolfgang Wahlster Germany, it was imperative for AI re- experimental machine learning and AI searchers to stay integrated with main- communities. In a 2003 International stream informatics and to collaborate in- Conference on Machine Learning tensively with colleagues from all (ICML) talk, I recalled how my first paper subareas of computer science. The at- was published in ICML 1987, then an in- tempts of AI researchers in some other vitation-only workshop. To our amuse- countries to establish AI as another meta- ment, the program committee strongly science like cybernetics outside of infor- advised us not to use abstract symbols matics were unsuccessful. like x1 for feature names, but warmer and —Wolfgang Wahlster, German Research fuzzier terminology like can_fly and Center for Artificial Intelligence (DFKI) has_wings. Perhaps we smirked a bit, but we un- I met AI through Max Clowes in 1969 derstood the sentiment and complied. when I was still a lecturer in philosophy Both sides have come a long way since at Sussex University1 and soon became then, to their mutual benefit. The rich- deeply involved, through a paper at IJ- ness of the theory that has been either di- CAI 1971 criticizing the 1969 logicist rectly or indirectly driven by the con- manifesto by John McCarthy and Pat cerns and findings of empirical machine Hayes, followed by a fellowship in Edin- learning and AI work is staggering to me, burgh 1972–1973. Since then I’ve worked and it has been a great pleasure to be a on forms of representation, vision, archi- theoretician working in a field in such a Aaron Sloman tectures, emotions, ontology for architec- close dialogue with practitioners. I am tures, tools for AI research, and teaching, hard-pressed to think of other branches links with psychology, biology and phi- of computer science that enjoy compara- losophy, and most recently robotics, and ble marriages. May the next twenty years I have helped to build up two major AI bring even more of the same; I cannot centers for teaching and research (at Sus- predict the results but know they will be sex and Birmingham). interesting. I believe philosophy needs AI and AI —Michael Kearns, needs philosophy. Much of what phi- University of Pennsylvania losophers write about consciousness and the mind-body problem shows their ig- I have fond recollections of the early norance of AI, and many silly debates be- years of my “discovering” the field of AI. tween factions in AI (for example, about Coming across it in a graduate course in representations, use of symbols, GOFAI) Michigan back in 1985, I was fascinated and some fashions (for example, recent and inspired by a field that had the bold enthusiasm for “emotions”) result from vision of nothing short of modeling hu- doing poor philosophical analysis. man intelligence. The field in its early I always thought progress in AI would days consisted of a collection of works be slow and difficult and that people that spanned everything from computer who predicted rapid results had simply science theory to biology, to psychology, failed to understand the problems, as to neural sciences, to machine vision, to Michael Kearns sketched in my 1978 book.2 more classical computer science tricks —Aaron Sloman, and techniques. The excitement was very University of Birmingham high and the expectations even higher. As I decided to start working in the sub- Twenty years ago, I arrived at Harvard to area of machine learning, I started to re-

WINTER 2005 89 25th Anniversary Issue

alize how difficult the problems are and mendous practical impact are, in my how far we truly are from realizing the opinion: (1) the generic visual object- ultimate dream of a thinking machine. recognition capabilities of a two-year-old What I also realized at the time was that child; (2) the manual dexterity of a six- specialization with some deep technical year-old child; (3) the social interaction approaches and mathematical rigor were and language capabilities of a ten-year- a necessity to make progress. old child. So much work for all of us to be In reflecting back on those days of ro- challenged by. mantic excitement, I am very pleased at —Rod Brooks, MIT what they drove in terms of engineering achievements and new fields of study. In From the scientific perspective, not so my own area of machine learning, while much has been accomplished, and the the vision of pursuing general algorithms goal of understanding intelligence, from that “learn from experience” morphed it- a computational point of view, remains self into highly specialized algorithms elusive. Reasoning programs still exhibit that solve complex problems at a large little common sense. Language programs scale, the result was the birth of several still have trouble with idioms, new subfields of specialization. Combin- metaphors, convoluted syntax, and un- Usama Fayyad ing learning algorithms with grammatical expressions. Vision pro- techniques and algorithms from compu- grams still stumble when asked to de- tational statistics resulted in data-mining scribe an office environment. algorithms that work on very large scales. —Patrick Henry Winston, MIT The resulting field of data mining is now a vibrant field with many commercial ap- Our choice of problems is telling: Small plications and significant economic val- and technical, not large and important. ue. This journey has also taken me per- A large, important problem is to work sonally from the world of basic scientific out the semantics of natural language— research to the business side of realizing including all the required commonsense economic value from applying these al- knowledge—so machines can read and gorithms to commercial problems and fi- understand the web. Another is to devel- nally to working at the “strategy” level op robots that understand what they see on the senior executive team of the and hear. largest Internet company in the world, Understanding is hard, so AI approx- Yahoo!, where data drives many products imates it with increasingly sophisticated and strategies. mappings from stimuli to responses: fea- In looking back at it, I can only say in ture vectors to class labels, strings in one wonder: what a ride! language to strings in another, states to —Usama Fayyad, Yahoo! states. I once had a robot that learned to map positive and negative translational Paul Cohen velocity to the words “forward” and The AI Journey: “backward,” but never learned that for- ward and backward are antonyms. It un- Future Challenges derstood the words superficially. It had the kind of understanding we can mea- As a field, AI researchers have always sure with ROC curves. Every child does looked for generality in the intelligent better. behavior our artifacts exhibit, and gener- ality remains a central challenge. Rod —Paul Cohen, University of Southern California Brooks, Patrick Winston, and Paul Cohen offer us a call to arms. Such broad capabilities need not origi- nate in a single fundamental principle or Artificial intelligence has not yet suc- ceeded in its most fundamental ambi- algorithm that applies across the board. tions. Our systems are still fragile when Instead, they may be the product of a outside their carefully circumscribed do- range of different models, representa- mains. The best poker-playing program tions, and experience, appropriately can’t even understand the notion of a combined. Building a general system may chess move, let alone the conceptual idea hinge on principled, flexible, and exten- of animate versus inanimate. A six-year- sible ways of putting the pieces together. old child can discuss all three domains If that’s right, we’ll have to start with a but may not be very good at any of them grounded understanding of the mean- compared to our specialized systems. The challenge for AI, still, is to capture the ings of representations, as Sebastian fundamental nature of generalized per- Thrun argues, but we’ll also need ways of ception, intelligence, and action. Worthy scaffolding sophisticated intelligent be- challenges for AI that would have tre- havior over underlying abilities to per-

90 AI MAGAZINE 25th Anniversary Issue ceive and act in the world, and we’ll need anything that can be expressed in natur- operational ways of weaving together re- al language, understand natural scenes stricted solutions into systems that ex- and situations, and so on. At the same hibit more robust behavior. In such archi- time AI has tended to splinter into spe- cialized areas that have their own confer- tectures, David Waltz, Manuela Veloso, ences and journals and that no longer and Patrick Winston find parallels to hu- have the goal of understanding or build- man intelligence. Despite its flexibility, ing truly intelligent systems. our own intelligence is an evolved capac- My own sense is that the AI research ity with clear limitations. We manage to program needs to be rethought in order act so successfully in part because we to have a realistic hope of building truly bring the right sets of cognitive skills, in- intelligent systems, whether these are au- cluding our notable strengths in domains tonomously intelligent or “cognitive of vision, language, and action. prostheses” for human-centered systems. Early AI focused on the aspects of human One of the big dreams of AI has been to thought that were not shared with other build an artificially “intelligent” robot—a creatures—for example, reasoning, plan- robot capable of interacting with people ning, symbolic learning—and minimized and performing many different tasks. We aspects of intelligence that are shared Sebastian Thrun have seen remarkable progress on many with other creatures—such as vision, of the component technologies neces- learning, adaptation, memory, naviga- sary to build AI robots. All these tremen- tion, manipulation of the physical world. dous advances beg the obvious question: A truly intelligent system will require an Why don’t we have a single example of a architecture that layers specifically hu- truly multipurpose robot that would, manlike abilities on top of abilities even marginally, deserve to be called ar- shared with other creatures. Some recent tificially intelligent? programs at the Defense Advanced Re- I believe the key missing component is search Projects Agency (DARPA) and the representation. While we have succeeded National Science Foundation (NSF) are in building special-purpose representa- setting ambitious goals that will require tions for specialized robot applications, we integrated generally intelligent systems, understand very little about what it takes a very promising trend. The best news to build a lifelong learning robot that can about the neglect of integrated intelli- accumulate diverse knowledge over long gent systems is that researchers going in- periods of time and that can use such to this area are likely to encounter a good knowledge effectively when deciding deal of “low hanging fruit.” what to do. It is time to bring knowledge —David Waltz, Columbia University representation and reasoning back into ro- botics. But not of the old kind, where our Creating autonomous intelligent robots only language to represent knowledge was with perception, cognition, and action Manuela Veloso binary statements of (nearly) universal and that are able to coexist with humans truth, deprived of any meaningful ground- can be viewed as the ultimate challeng- ing in the physical world. ing goal of artificial intelligence. Ap- We need more powerful means of proaches to achieve such a goal that de- representing knowledge. Robotics knowl- pend on rigid task and world models that edge must be grounded in the physical drive precise mathematical algorithms, world, hence knowledge acquisition even if probabilistic, are doomed to be equals learning. Because data-driven too restrictive, as heuristics are clearly learning is prone to error, reasoning with needed to handle the uncertainty that such knowledge must obey the uncer- inevitably surrounds autonomy within tainties that exist in the learned knowl- human environments. Instead we need edge bases. Our representation languages to investigate rich approaches capable of must be expressive enough to represent using heuristics and flexible experience- the complex connections between ob- built webs of knowledge to continuously jects in the world, places, actions, people, question and revise models while acting time, and causation, and the uncertainty in an environment. Significant progress among them. In short, we need to rein- depends upon a seamless integration of vent the decades-old field knowledge perception, cognition, and action to pro- representation and reasoning if we are to vide AI creatures with purposeful percep- succeed in robotics. tion and action, combined with the —Sebastian Thrun, Stanford University ability to handle surprise, to recognize and adapt past similar experience, and to We are still far short of truly intelligent learn from observation. Hence and inter- systems in the sense that people are intel- estingly, I find that the achievement of ligent—able to display “common sense,” the ultimate goal of the field requires us, deal robustly with surprises, learn from researchers, to accept that AI creatures

WINTER 2005 91 25th Anniversary Issue

are evolving artifacts most probably al- are increasingly going electronic. This ways with limitations, similarly to hu- presents a wonderful opportunity for AI: mans. Equipped with an initial perceptu- electronic marketplaces provide new ex- al, cognitive, and execution architecture, citing research questions, and AI can sig- robots will accumulate experience, refine nificantly help generate more efficient their knowledge, and adapt the parame- market outcomes and processes. ters of their algorithms as a function of One example is expressive competition, their interactions with humans, other ro- a generalization of combinatorial auc- bots, and their environments. tions. The idea is to let buyers and sellers —Manuela Veloso, (human or software) express demand Carnegie Mellon University and supply at a drastically finer granular- ity than in traditional markets—much Since the field of artificial intelligence like having the expressiveness of human- was born in the 1960s, most of its practi- to-human negotiation, but in a struc- tioners have believed—or at least acted as tured electronic setting where demand if they have believed—that language, vi- and supply are algorithmically matched. sion, and motor faculties are the I/O A combination of AI and operations re- channels of human intelligence. Over search techniques have recently made the years I have heard distinguished lead- expressive competition possible, and to- Tuomas Sandholm ers in the field suggest that people inter- day almost all expressive competition ested in language, vision, and motor is- markets are cleared using sophisticated sues should attend their own tree search algorithms. Tens of billions of conferences, lest the value of artificial in- dollars of trade have been cleared with telligence conferences be diminished by this technology, generating billions of irrelevant distractions. dollars of additional value into the world To me, ignoring the I/O is wronghead- by better matching of supply and de- ed, because I believe that most of our intel- mand. ligence is in our I/O, not behind it, and if Less mature, but promising, roles of we are to understand intelligence, we must AI include the following: (1) Automati- understand the contributions of language, cally designing the market mechanism vision, and motor faculties. Further, we for the specific setting at hand. This can must understand how these faculties, circumvent seminal economic impossi- which must have evolved to support sur- bility results. (2) Designing markets vival in the physical world, enable abstract where finding an insincere (socially un- thought and the reuse of both concrete desirable) strategy that increases the par- and abstract experience. We must also un- ticipant’s own utility is provably hard derstand how imagination arises from the computationally. (3) Supplementing the concert of communication among our pu- market with software that elicits the par- tative I/O faculties, and we must learn ticipants’ preferences incrementally so Peter Norvig how language’s symbols ground out in vi- that they do not have to determine their sual and other perceptions. preferences completely when that is un- —Patrick Henry Winston, MIT necessary for reaching the right market outcome. (4) Taking into account the is- Intelligence is key not only for our phys- sues that arise when the participants in- ical robots but also for the intermediaries cur costs in determining their prefer- we recruit for competition, communica- ences and can selectively refine them. tion, and collaboration in virtual worlds. There are numerous other roles for AI, For example, Tuomas Sandholm sees undoubtedly including many that have electronic marketplaces as an area where not even been envisioned. With this brief AI can change the world. Peter Norvig, note I would like to encourage bright meanwhile, delights us with the poten- young (and old) AI researchers to get in- tial of AI to forge relationships among the volved in this intellectually exciting area billions of readers and writers on the web. that is making the world a better place. —Tuomas Sandholm, Though it’s easy to forget, much of our Carnegie Mellon University own intelligence is specifically social. We count on our abilities to explain, coordi- So far we know of exactly one system in nate, and adapt our actions to one anoth- which trillions of pieces of information er, and as Danny Bobrow and Joe Marks, can be intelligently transmitted to bil- Daniel Bobrow Chuck Rich, and Candy Sidner argue, du- lions of learners: the system of publishing the written word. No other system, artifi- plicating these abilities offers an exciting cial or otherwise, can come within a fac- bridge between AI and human–computer tor of a million for successful communi- collaboration. cation of information. This is despite the A significant portion of trade is already fact that the written word is notoriously conducted electronically, and markets ambiguous, ill-structured, and prone to

92 AI MAGAZINE 25th Anniversary Issue logical inconsistency or even fallacy. wrote: “[Compare] instructions ordinari- In the early days of AI, most work was ly addressed to intelligent human beings on creating a new system of transmis- with instructions ordinarily used with sion—a new representation language, computers. The latter specify precisely and/or a new axiomatization of a domain. the individual steps to take and the se- This was necessary because of the difficul- quence in which to take them. The for- ty of extracting useful information from mer present or imply something about available written words. One problem was incentive or motivation, and they supply that people tend not to write down the a criterion by which the human executor obvious; as Doug Lenat and Ed Feigen- of the instructions will know when he baum put it, “each of us has a vast store- has accomplished his task. In short: in- house of general knowledge, though we structions directed to computers specify rarely talk about any of it ... Some exam- courses; instructions directed to human ples are: ‘water flows downhill’ ....”4 This beings specify goals.”5 is undeniably true; if you look at a 1,000- Licklider goes on to argue that in- page encyclopedia, there is no mention of structions directed to computers should “water flows downhill.” But if you look at be more like instructions to human be- an 8 billion page web corpus, you find ings. Even today, this is a radical idea that about one in a million pages outside of AI circles. Most research on Joe Marks mentions the phrase, including some human-computer interaction (HCI) has quite good kindergarten lesson plans. focused on making interaction with This suggests that one future for AI is computers more efficient by adding new “in the middle” between author and input and output mechanisms. AI re- reader. It will remain expensive to create searchers, however, are working to fun- knowledge in any formal language (Pro- damentally change the level of HCI from ject Halo suggests $10,000/page) but AI command-oriented interaction to goal- can leverage the work of millions of au- oriented collaboration. Furthermore, thors of the written word by understand- since Licklider wrote the words above, re- ing, classifying, prioritizing, translating, searchers in AI and the neighboring fields summarizing, and presenting the written of linguistics and cognitive science have word in an intelligent just-in-time basis accumulated a large body of empirical to billions of potential readers. knowledge and computational theory re- —Peter Norvig, Google garding how human beings collaborate with one another, which is helping us to People build their own models of systems realize Licklider’s vision of human-com- they don’t understand and may make un- puter symbiosis. warranted extrapolations of their capabil- —Joe Marks, Charles Rich, and Candace ities—which can lead to disappointment Sidner, Mitsubishi Electric Research Labs and lack of trust. An effective intelligent system should be transparent, able to As we aim for broader capabilities, our Charles Rich explain its own behavior in a way that systems must draw on a more diverse connects to its users’ background and range of ideas. This brings new challenges knowledge. An explanation is not a full for research in the field. As Daphne Koller trace of the process by which the system explains, our problems now require us to came to a conclusion. It must highlight take ideas from across subfields of AI and important/surprising points of its process and indicate provenance and dependen- make them work together. Several of our cies of resources used. Systems that evolve contributors suggest that well-defined, through statistical learning must explain long-range projects can inspire re- (and exemplify) categories it uses and clar- searchers with different backgrounds and ify for a user what properties make a dif- specializations to bridge their ideas and ference in a particular case. Such systems results. Henry Kautz will be working on must not be single minded, hence should understanding human activities in instru- be interruptible and able to explain cur- mented environments. Tom Mitchell will rent goals and status. They should be able be working on understanding the lan- to take guidance in terms of the explana- guage of the web. tion they have given. Artificial intelli- gence systems must not only understand A solution to the AI problem—achieving a the world, and the tasks they face, but un- truly intelligent system—remains elusive. derstand their users and— most impor- The capabilities that appear the hard- Candace Sidner tant—make themselves understandable, est to achieve are those that require inter- correctable, and responsible. action with an unconstrained environ- ment: machine perception, natural —Daniel G. Bobrow, PARC language understanding, or common- In his prescient 1960 article titled “Man- sense reasoning. To build systems that ad- Computer Symbiosis,” J. C. R. Licklider dress these tasks, we need to draw upon

WINTER 2005 93 25th Anniversary Issue

the expertise developed in many subfields cluding assistive technology for the dis- of AI. Of course, we need expertise in per- abled, aging in place, security and sur- ception and in natural language models. veillance, and data collection for the so- But we also need expressive representa- cial sciences. tions that encode information about dif- —Henry Kautz, University of Washington ferent types of objects, their properties, and the relationships between them. We I believe AI has an opportunity to need algorithms that can robustly and ef- achieve a true breakthrough over the fectively answer questions about the coming decade by at last solving the world using this representation, given on- problem of reading natural language text ly partial information. Finally, as these to extract its factual content. In fact, I systems will need to know an essentially hereby offer to bet anyone a lobster din- unbounded number of things about the ner that by 2015 we will have a computer world, our framework must allow new program capable of automatically read- knowledge to be acquired by learning ing at least 80 percent of the factual con- from data. Note that this is not just “ma- tent across the entire English-speaking chine learning” in its most traditional web, and placing those facts in a struc- sense, but a broad spectrum of capabili- tured knowledge base. ties that allow the system to learn contin- Why do I believe this breakthrough Henry Kautz uously and adaptively. will occur in the coming decade? Because Therefore, in addition to making of the fortunate confluence of three progress in individual subfields of AI, we trends. First, there has been substantial must also keep in mind the broader goal progress over the past several years in of building frameworks that integrate natural language processing for automat- representation, reasoning, and learning ically extracting named entities (such as into a unified whole. person names, locations, dates, products, —Daphne Koller, Stanford University and so on) and facts relating these enti- ties (for example, WorksFor[Bill, Mi- One of the earliest goals of research in ar- crosoft]). Much of this progress has come tificial intelligence was to create systems from new natural language processing that can interpret and understand day to approaches, many based on machine day human experience. learning algorithms, and progress here Early work in AI, in areas such as sto- shows no sign of slowing. Second, there ry understanding and commonsense rea- has been substantial progress in machine soning, tried to tackle the problem head learning over the past decade, most sig- on but ultimately failed for three main nificantly on “bootstrap learning” algo- reasons. First, methods for representing rithms that learn from a small volume of and reasoning with uncertain informa- labeled data, and huge volumes of unla- tion were not well understood; second, beled data, so long as there is a certain systems could not be grounded in real Tom Mitchell kind of redundancy in the facts ex- experience without first solving AI-com- pressed in this data. The third important plete problems of vision or language un- trend is that the data needed for learning derstanding; and third, there were no to read factual statements is finally avail- well-defined, meaningful tasks against able: for the first time in history every which to measure progress. computer has access to a virtually limit- After decades of work on the “bits less and growing text corpus (such as the and pieces” of artificial intelligence, we web), and this corpus happens to contain are now at a time when we are well- just the kind of factual redundancies poised to make serious progress on the needed. These three trends, progress in goal of building systems that understand natural language analysis, progress in human experience. Each of the previous machine learning, and availability of a barriers is weakened: sufficiently rich text corpus with tremen- First, we now have a variety of expres- dous redundancy, together make this the sive and scalable methods for dealing right time for AI researchers to go back to with information that is both relational one of the key problems of AI—natural and statistical in nature. Second, the de- velopment and rapid deployment of low- language understanding—and solve it (at cost ubiquitous sensing devices—includ- least for the factual content of language). ing RFID tags and readers, global po- —Tom Mitchell, sitioning systems, wireless motes, and a Carnegie Mellon University wide variety of wearable sensors—make it possible to immediately create AI sys- tems that are robustly grounded in direct Shaping the Journey experience of the world. Third, there are a growing number of vital practical appli- To pursue the kinds of goals and projects cations of behavior understanding, in- sketched in the previous section, we need

94 AI MAGAZINE 25th Anniversary Issue the right institutions as well as the right wants to join a society dedicated to the ideas. If we share programs and data in cognitive, perceptual, and social develop- more open collaborations, we can make ment of robot babies, or to learning lan- it easier for individual researchers to guage sufficient to understanding chil- dren’s books at a 5-year-old’s level of make meaningful contributions to big competence, please drop me a line. new problems. If we slant major confer- ences to emphasize integrative research, —Paul Cohen, University of Southern California we can help these contributions find their audiences. These changes are under Some have mentioned to me that this way: talk to Tom Mitchell or Paul Cohen [natural language understanding for the about shared resources; talk to John Laird web] is a large goal. I agree and propose or Paul Cohen about integrative research we approach it by forming a shared web meetings. Still, to tackle the really big repository where facts that are extracted from the web by different researchers’ ef- problems, our institutions may have to forts are accumulated and made accessi- become bigger and broader, too. Perhaps ble to all. This open-source shared repos- we will see more large research centers, itory should also accumulate and share like DFKI, specifically dedicated to inte- learned rules that extract content from John Laird grative research in artificial intelligence. different linguistic forms. Working as a I can think of nothing more exciting research community in this fashion than working on integrated human-level seems the best way to achieve this ambi- intelligent systems, and after 25 years it tious goal. And I’d hate to have to leave is still captivating and challenging. One the field to open a lobster fishery. challenge is that it requires interdispli- —Tom Mitchell, narity in the small (across subfields of AI) Carnegie Mellon University and large scale (with other disciplines outside of AI). A second challenge is that DFKI, the German Research Center for teams are needed to attack the large-scale AI, employs today more than 200 full- problems that can challenge integrated time researchers, many of them holding AI systems (as evident in many of the re- a Ph.D. degree in AI. With yearly rev- cent DARPA programs). This type of re- enues of more than US$23 million, it is search isn’t for the faint of heart or those probably the world’s largest contract re- search center for AI. It has created 39 fast- who enjoy solitary work. A third chal- growing spin-off companies in many lenge is to communicate research re- fields of AI. DFKI views itself as a software sults—if the work is truly interdiscipli- technology innovator for government nary, which field or subfield should it be and commercial clients. DFKI is a joint published in? Moreover, how (and venture including Bertelsmann, Daimler- where) can I talk about what is learned Chrysler, Deutsche Telekom, Microsoft, Tom Dean about integration, which itself is not na- SAP, and the German federal govern- tive to any specific field? AI has done a ment. Its mission is to perform “innova- marvelous job, leveraging specialization tion pure” application-oriented basic AI with the inexhaustible growth of new research. Although we have always tried subfields, new applications, and new to contribute to the grand challenges of conferences; but we also need to fiercely AI, we have never experienced an AI win- support integration. What better way to ter at DFKI, since we have always been do this than by making integrated cogni- quite cautious with promises to our tive systems a major emphasis of the sponsors and clients, trying to deliver AAAI conference? down-to-earth, practical AI solutions. —John Laird, University of Michigan Since its foundation in 1988, many ma- What to do? Form societies defined by turing AI technologies have left DFKI’s and dedicated to solving large, important labs and become ubiquitous to the point problems. Hold conferences where the where they are almost invisible in em- criteria for publication are theoretical bedded software solutions. and empirical progress on these prob- —Wolfgang Wahlster, DFKI lems. Discourage sophistication-on- Integrative research will be particular- steroids; encourage integrations of sim- ly challenging for research students. To ple (preferably extant) algorithms that do it, they must master a wide range of solve big chunks of important problems. Work together; share knowledge bases, formal techniques and understand not ontologies, algorithms, hardware, and just the mathematical details but also test suites; and don’t fight over stan- their place in overall accounts of intelli- dards. Don’t wait for government agen- gent behavior. At the same time, to cies to lead; do it ourselves. If anyone launch productive careers, they must

WINTER 2005 95 25th Anniversary Issue

make a name for themselves with important termining the appropriate metrics is part of the new ideas of their own. It may take longer than research). But given the dynamics in the field of we are used to and require us to think different- AI, a Ph.D. student must forge an association to ly about how we nurture new scientists. an identifiable subfield of AI—some communi- So, if AI as a community is to tackle the big ty in which to publish and build a reputa- tion—and as of today that is not “human-level problems and continue its progress, we have a intelligence,” “integrated cognitive systems,” lot of work to do besides our own research. But or even my favorite, “cognitive architecture.” we mustn’t think of this work as painful. We’ll So even more than finding a home for publish- be doing it with friends and colleagues, as Tom ing, we must grow a community of researchers, Dean and Bruce Buchanan remind us, and we teachers, and students in which the integration can expect a unique kind of satisfaction in see- is the core and not the periphery. ing AI’s collaborative process of discovery —John Laird, University of Michigan strengthened and energized by the new com- munities we foster. In more than twenty years in this field, the most satisfying moments by far have come Surely, with such intriguing problems to work from working with people who have set aside on, and with allied fields on the march, this their individual interests and biases to inspire should be a time for universal optimism and ex- students, nurture young scientists, and create pectation; yet many of today’s young, emerging community and esprit de corps. And, while I practitioners seem to have abandoned the truly enjoyed collaborating with Kathy McKe- grand original goals of the field, working in- own on AAAI-91 and Gigina Aiello on IJCAI-99, stead on applied, incremental, and fundamen- helping to create the AAAI Robot Competition tally boring problems. Too bad. Pessimists will and Exhibition series with Pete Bonasso, Jim miss the thrill of discovery on the Watson-and- Firby, Dave Kortenkamp, David Miller, Reed Crick level. Simmons, Holly Yanco, and a host of others —Patrick Henry Winston, MIT was actually a lot of fun. The exercise was cer- tainly not without its aggravations, as getting a Another reason for slow progress is the frag- sizable group of researchers to agree on any is- mentation of AI: people learn about tiny frag- sue is not easy. But most of the effort was spent ments of a whole system and build solutions thinking about how to create useful technolo- that could not form part of an integrated hu- gy, advance the science and art of robotics, and manlike robot. One explanation is that we do make the entire experience both educational not have full-length undergraduate degrees in and inspirational to participants and spectators AI and most researchers have to do a rapid alike. switch from another discipline, so they learn It was particularly gratifying to see the buzz just enough for their Ph.D. topic, and they and of activity around this year’s event in Pitts- their students suffer thereafter from the result- burgh and learn about some of the new ideas ing blinkered vision. involving social robots, assistive technologies, I’ve proposed some solutions to this prob- and, of course, cool hardware hacks. I don’t lem in an introduction to a multidisciplinary know what direction the field should take, and tutorial at IJCAI’05, including use of multiple at this particular moment in my career as I re- partially ordered scenarios to drive research. turn to research after several years in senior ad- It requires a lot more people to step back ministration at , I’m content and think about the hard problems of combin- to pursue my own personal research interests in ing diverse AI techniques in fully functional the intersection of robotics, machine learning. humanlike robots, though some room for spe- and computational neuroscience. But I am cialists remains. thinking about how to get students interested —Aaron Sloman, University of Birmingham in my research area, and in due course, I hope The students of AI are sophisticated in both dis- to work with the AI community to run work- crete and continuous mathematics, including a shops, develop tutorials, sponsor undergradu- recognition of the role of uncertainty. This is ate research, and pursue all the other avenues necessary because of the increased complexity open to us to nurture and sustain community, of problems that we need to attack. both scientific and social, in our rapidly evolv- ing and increasingly central field. —Ruzena Bajcsy, University of California at Berkeley —Tom Dean, Brown University

As teachers, we must challenge students to The sense of collegiality in the AI community work on problems in which integration is cen- has always made AI more fun. Most of the time, tral and not an afterthought: problems that re- the statesmanlike conduct of senior people like quire large bodies of different types of knowl- Al Newell set an example for debate without edge, problems that involve interaction with rancor. The common goal of understanding the dynamic environments, problems that change nature of intelligence makes everyone’s contri- over time, and problems in which learning is bution interesting. central (and sometimes problems in which de- —Bruce Buchanan, University of Pittsburgh

96 AI MAGAZINE 25th Anniversary Issue

Closing Thoughts Notes 1. Sloman, A. 1981. Experiencing Computation: A As we find new discoveries in AI, as new com- Tribute to Max Clowes. Computing in Schools. munities form, as new sets of ideas come to- (www.cs.bham.ac.uk/research/cogaff/sloman-clow- gether, as new problems emerge and we find estribute.html.) new ways of working on them, we will contin- 2. Sloman, A. 1978. The Computer Revolution in Philos- ue to frame new accounts about what our work ophy: Philosophy, Science and Models of Mind. Brighton, is. The best stories will resonate not only with U.K.: Harvester Press. (http://www.cs.bham.ac.uk/re- what we are doing now but also with how we search/cogaff/crp.) got here—they will transcend today’s fashions 3. Valiant, L. G. 1984. A Theory of the Learnable. and motivate our present activities as a synthe- Communications of the ACM 27(11):1134–1142; sis of our earlier goals and experiences. We Valiant, L. G. 1985. Learning Disjunctions of Con- close with some brief contributions that take junctions. In Proceedings of the Ninth International up this discipline of self-examination and, in Joint Conference on Artificial Intelligence, vol. 1, different ways, distill something important 560–566. Los Altos, CA: William Kaufmann, Inc.; and Blumer, A.; Ehrenfeucht, A.; Haussler, D.; and War- about the landscape of our field. We respond muth, M. 1985. Classifying Learnable Geometric deeply to Bruce Buchanan’s characterization of Concepts with Vapnik-Chervonenkis Dimension. In the enduring significance of AI as a goal and Proceedings of the Twenty-Fourth Symposium on Theory the enduring bottom line of AI methodology of Computation, 175–85. New York: Association for and to Henry Kautz’s acknowledgment of the Computing Machinery. human meaning of AI results. We endorse Us- 4. Lenat, D., and Feigenbaum, E. 1991. On the ama Fayyad’s prediction that an exciting ride Thresholds of Knowledge. Artificial Intelligence 47 through intellectual space will continue to de- (1–3): 185–250. fine the careers of AI researchers. And, with 5. Licklider, J. C. R. 1960. Man-Computer Symbiosis. Alan Mackworth, we remind you to keep it real. IRE Transactions on Human Factors in Electronics, Vol- The discussion will doubtless continue. ume HFE-1: 4–11.

Because there is not enough intelligence in the Haym Hirsh is a professor and world and humans often ignore relevant conse- chair of computer science at Rut- quences of their decisions, AI can provide the gers University. His research ex- means by which decision makers avoid global plores applications of machine catastrophes. I believe we can realize this vision learning, data mining, and infor- by formulating and testing ideas in the context mation retrieval in intelligent in- of writing and experimenting with programs. formation access and human-com- —Bruce Buchanan, University of Pittburgh puter interaction. He received his B.S. in 1983 in mathematics and I believe that understanding human experience computer science from UCLA, and his M.S. and Ph.D. will be a driving challenge for work in AI in the in 1985 and 1989, respectively, in computer science years to come and that the work that will result from Stanford University. He has been on the faculty will profoundly impact our knowledge of how at since 1989 and has also held we live and interact with the world and with visiting faculty positions at Carnegie Mellon Univer- each other. sity, MIT, and NYU’s Stern School of Business. —Henry Kautz, University of Washington Matthew Stone is associate profes- In looking at the future, we have much to do, sor in the department of computer and I hope we make some serious progress on science at Rutgers and has a joint that original romantic dream of building ma- appointment in the Center for chines that truly exhibit learning and thought Cognitive Science; for the 2005– in general. The dream is still worth pursuing, es- 2006 academic year, he is also Lev- pecially after all we have learned over the past erhulme Trust Visiting Fellow at decades. The ride will be a lot more exciting for the School of Informatics, Univer- the new researchers entering the AI field. sity of Edinburgh. He received his —Usama Fayyad, Yahoo! Ph.D. in computer science from the University of Pennsylvania in 1998. His research interests span Think of AI itself as an agent. We need a clear computational models of communicative action in understanding of our own goals, but we must conversational agents and include work on generat- also be willing to seize opportunistically on new ing coordinated verbal and nonverbal behavior (pre- developments in technology and related sci- sented most recently at SIGGRAPH 2004 and ESSLLI ences. This anniversary is a lovely opportunity 2005). He served on the program committee of IJCAI to take stock, to remind ourselves to state our 2003 and as tutorial chair of AAAI 2004. His e-mail claims realistically, and to consider carefully the address is [email protected]. consequences of our work. Above all, have fun. —Alan Mackworth, University of British Columbia

WINTER 2005 97