It is an honor and a pleasure for me to acceptthis award from ACM. It is especially gratifyingto share this award with Ed Feigenbaum, whohas been a close friend and helpful colleague fornearly 30 years.
As a second-generation artificial intelligence(AI) researcher, I was fortunate to have known andworked with many of the founding fathers of AI. Byobserving John McCarthy, my thesis advisor, duringthe golden age of AI Labs at Stanford in the 1960s,I have learned the importance of fostering andnurturing diversity in research far beyond onesown personal research agenda. Although his ownprimary interest was in common-sense reasoningand epistemology, under Johns leadership,research in speech, vision, robotics, language,knowledge systems, game playing, and music,thrived at the AI labs. In addition, a great deal ofpath-breaking systems research flourished in areassuch as Lisp, time-sharing, video displays, and aprecursor to Windows called pieces of glass.
From Marvin Minsky, who was visiting Stanfordand helping to build the Mars Rover in 66, Ilearned the importance of pursuing bold visions ofthe future. And from Allen Newell and HerbSimon, my colleagues and mentors at CarnegieMellon University (CMU) for over 20 years, Ilearned how one can turn bold visions into practi-cal reality by careful design of experiments and fol-lowing the scientific method.
I was also fortunate to have known and workedwith Alan Perlis, a giant in the 50s and 60s com-puting scene and the first recipient of the TuringAward in 1966, presented at the ACM conferenceheld in Los Angeles, which I attended while still agraduate student at Stanford.
While I did not know Alan Turing, I may be oneof the select few here who used a computerdesigned by him. In the late 1950s, I had the plea-sure of using a mercury delay-line computer (Eng-lish Electric Deuce Mark II) based on Turings
original design of ACE. Given his early papers on"Intelligent Machines," Turing can be reasonablycalled one of the grandfathers of AI, along withearly pioneers such as Vannevar Bush.
That brings me to the topic of this talk, Todream the possible dream. AI is thought to be animpossible dream by many. But not to us in AI. Itis not only a possible dream, but, from one pointof view, AI has been a reality that has been demon-strating results for nearly 40 years. And the futurepromises to generate an impact greater by ordersof magnitude than progress to date. In this talk Iwill attempt to demystify the process of what AIresearchers do and explore the nature of AI and itsrelationship to algorithms and software systemsresearch. I will discuss what AI has been able toaccomplish to date and its impact on society. I willalso conclude with a few comments on the long-term grand challenges.
Human and Other Forms of Intelligence
Can a computer exhibit real intelligence?Simon provides an incisive answer: Iknow of only one operational meaning forintelligence. A (mental) act or series of acts isintelligent if it accomplishes something that, ifaccomplished by a human being, would be calledintelligent. I know my friend is intelligent becausehe plays pretty good chess (can keep a car on theroad, can diagnose symptoms of a disease, cansolve the problem of the missionaries and canni-bals, etc.) I know that computer A is intelligentbecause it can play excellent chess (better than allbut about 200 humans in the entire world). I knowthat Navlab is intelligent because it can stay on theroad. The trouble with those people who thinkthat computer intelligence is in the future is thatthey have never done serious research on humanintelligence. Shall we write a book on WhatHumans Cant Do? It will be at least as long asDreyfus book. Computer intelligence has been a
T u r i n g
A w a r d
To Dream The PossibleDream
106 May 1996/Vol. 39, No. 5 COMMUNICATIONS OF THE ACM
fact at least since 1956, when the Logic Theorymachine found a proof that was better than theone found by Whitehead and Russell, or when theengineers at Westinghouse wrote a program thatdesigned electric motors automatically. Lets stopusing the future tense when talking about com-puter intelligence.
Can AI equal human intelligence? Somephilosophers and physicists have made successfullifetime careers out of attempting to answer thisquestion. The answer is, AI can be both more andless than human intelligence. It doesnt take largetomes to prove that they cannot be 100% equiva-lent. There will be properties of human intelli-gence that may not be exhibited in an AI system(sometimes because we have no particular reasonfor doing so or because we have not yet gottenaround to it). Conversely, there will be capabilitiesof an AI system that will be beyond the reach ofhuman intelligence. Ultimately, what will beaccomplished by AI will depend more on whatsociety needs and where AI may have a comparativeadvantage than on philosophical considerations.
Let me illustrate the point by two analogies thatare not AI problems in themselves but whose solu-tions require some infusion of AI techniques.These problems, currently at the top of theresearch agenda within the information industry,are digital libraries and electronic commerce.
The basic unit of a digital library is an electronicbook. An electronic book provides the same infor-mation as a real book. One can read and use theinformation just as we can in a real book. However, itis difficult to lie in bed and read an electronic book.With expected technological advances, it is con-ceivable a subnotebook computer will weigh lessthan 12 ounces and have a 6 x 8 high resolutioncolor screen, making it look and feel like a bookthat you might read in bed. However, the analogystops there. An electronic book cannot be used aspart of your rare book collection, nor can it beused to light a fire on a cold night to keep youwarm. You can probably throw it at someone, butit would be expensive. On the other hand, usingan electronic book, you can process, index, andsearch for information; open the right page; high-light information; change font size if you dont
have your reading glasses; and so on. The point is,an electronic book is not the same as a real book.It is both more and less.
A key component of electronic commerce is theelectronic shopping mall. In this virtual mall, youcan walk into a store, try on some virtual clothing,admire the way you look, place an order and havethe real thing delivered to your home in 24 hours.Obviously, this does not give you the thrill of goinginto a real mall, rubbing shoulders with real peo-ple and trying on real clothing before you makeyour purchase. However, it also eliminates theproblems of getting dressed, fighting the trafficand waiting in line. More importantly, you canpurchase your dress in Paris, your shoes in Milanand your Rolex in Hong Kong without ever leavingyour home. Again, the point is that an electronicshopping mall is not the same as a real shoppingmall. It is both more and less.
Similarly, AI is both more and less than humanintelligence. There will be certain human capabili-ties that might be impossible for an AI system toreach. The boundary of what can or cannot be donewill continue to change with time. More important,however, it is clear that some AI systems will havesuper human capabilities that would extend thereach and functionality of individuals and commu-nities. Those who possess these tools will make therest of us look like primitive tribes. By the way, thishas been true of every artifact created by the humanspecies, such as the airplane. It just so happens thatAI is about creating artifacts that enhance the men-tal capabilities of the human being.
AI and AlgorithmsIsnt AI just a special class of algorithms? In a senseit is; albeit a very rich class of algorithms, whichhave not yet received the attention they deserve.Second, a major part of AI research is concernedwith problem definition rather than just problemsolution. Like complexity theorists, AI researchersalso tend to be concerned with NP-complete prob-lems. But unlike their interest in the complexity ofa given problem, the focus of research in AI tendsto revolve around finding algorithms that provideapproximate, satisfying solutions with no guaran-tee of optimality.
Can AI equal human intelligence? Some philosophers and physicists
have made successful lifetime careers out of attempting to answer this question.
The answer is, AI can be both more and less than human intelligence.
COMMUNICATIONS OF THE ACM May 1996/Vol. 39, No. 5 107
The concept of satisfying solutions comes fromSimons pioneering research on decision makingin organizations leading to his Nobel Prize. Priorto Simons work on human decision making, it wasassumed that, given all the relevant facts, thehuman being is capable of rational choice weigh-ing all the facts. Simons research showed thatcomputational constraints on human thinkinglead people to be satisfied with good enoughsolutions rather than attempting to find rationaloptimal solutions weighing all the facts. Simoncalls this the principle of bounded rationality.When people have to make decisions under con-ditions which overload human thinking capabili-ties, they dont give up, saying the problem isNP-complete. They use strategies and tactics ofoptimal-least-computation search and not those of opti-mal-shortest-path search.
Optimal-least-computation search is the study ofapproximate algorithms that can find the best pos-sible solution given certain constraints on the com-putation, such as limited memory capacity, limitedtime, or limited bandwidth. This is an area worthyof serious research by future complexity theorists!
Besides finding solutions to exponential prob-lems, AI algorithms often have to satisfy one or moreof the following constraints: exhibit adaptive goal-oriented behavior, learn from experience, use vastamounts of knowledge, tolerate error and ambiguityin communication, interact with humans using lan-guage and speech, and respond in real time.
Algorithms that exhibit adaptive goal oriented behav-ior. Goals and subgoals arise naturally in prob-lems where algorithm specification is in the formof What is to be done rather than How it is tobe done. For example, consider the simple taskof asking an agent, Get me Ken. This requiresconverting this goal into subgoals, such as lookup the phone directory, dial the number, talk tothe answering agent, and so on. Each subgoalmust then be converted from What to How andexecuted. Creation and execution of plans hasbeen studied extensively within AI. Other sys-tems such as report generators, 4GL systems, anddata base query-by-example methods, use simpletemplate-based solutions to solve the What to
How problem. In general, to solve such prob-lems, an algorithm must be capable of creatingfor itself an agenda of goals to be satisfied usingknown operations and methods, the so-called"GOMs approach." Means-ends analysis, a formof goal-oriented behavior, is used in most expertsystems.
Algorithms that learn from experience. Learning fromexperience implies the algorithm has built-inmechanisms for modifying internal structure andfunction. For example, in the Get me Ken task,suppose Ken is ambiguous and you help the agentto call the right Ken; next time you ask the agentto Get me Ken, it should use the same heuristicthat you used to resolve the ambiguity. Thisimplies the agent is capable of acquiring, repre-senting, and using new knowledge and engagingin a clarification dialog, where necessary, in thelearning process. Dynamic modification of inter-nal structure and function is considered to be dan-gerous in some computer science circles becauseof the potential for accidental overwriting of otherstructures. Modifying probabilities and modifyingcontents of tables (or data structures) have beenused successfully in learning tasks where the prob-lem structure permits such a formulation of learn-ing. The Soar architecture developed by Newell etal., which uses rule-based system architecture, isable to discover and add new rules (actually pro-ductions) and is perhaps the most ambitiousundertaking to date to create a program thatimproves with experience.
Algorithms that interact with humans using languageand speech. Algorithms that can effectively usespeech and language in human-computer interfacewill be essential as we move toward a society wherenonexperts use computers in their day-to-day prob-lems. In the previous example of the Get me Kentask, unless the agent can conduct the clarificationdialog with the human master using language andspeech or some other natural form of communica-tion, such as Form Filling, widespread use ofagents will be a long time coming. Use of languageand speech involves creating algorithms that candeal with not only ambiguity and nongrammatical-
T u r i n g
A w a r d
ity but also with parsing and interpreting of naturallanguage with a large, dynamic vocabulary.
Algorithms that can effectively use vast amounts ofknowledge. Large amounts of knowledge not onlyrequire large memory capacity but creates themore difficult problem of selecting the rightknowledge to apply for a given task and context.There is an illustration that John McCarthy is fondof using. Suppose one asks the question, Is Rea-gan sitting or standing right now? A system with alarge database of facts might proceed to systemati-cally search the terabytes of data before finallycoming to the conclusion that it does not know theanswer. A person faced with the same problemwould immediately say, I dont know, and mighteven say, and I dont care. The question ofdesigning algorithms that know what they do notknow is currently an unsolved problem. With theprospect of very large knowledge bases loomingaround the corner, knowledge search willbecome an important algorithm design problem.
Algorithms that tolerate error and ambiguity in commu-nication. Error and ambiguity are a way of life inhuman-to-human communication. Warren Teitel-man developed nearly 25 years ago an interfacecalled Do What I Mean (DWIM). Given therequirements of efficiency and getting the soft-ware done in time, all such ideas were deemed tobe frivolous. Today, with the prospect of Giga PCsaround the corner, we need to revisit such error-forgiving concepts and algorithms. Rather thansaying illegal syntax, we need to develop algo-rithms that can detect ambiguity (i.e., multiplepossible interpretations, including null) andresolve it where possible by simultaneous parallelevaluation or by engaging in clarification dialog.
Algorithms that have real-time constraints. Many soft-ware systems already cope with this problem, butonly through careful, painstaking analysis of thecode. Not much work has been done on how tocreate algorithms that can accept a hurry-upcommand! One approach to this problem appearsin the Prodigy system, which generates an approx-imate plan immediately and gradually improves
and replaces that plan with a better one as the sys-tem finds more time to think. Thus, it always...