ON JOHN MCCARTHY'S 80TH BIRTHDAY, IN HONOR OF HIS CONTRIBUTIONS

Page created by Harold Ball
 
CONTINUE READING
ON JOHN MCCARTHY'S 80TH BIRTHDAY, IN HONOR OF HIS CONTRIBUTIONS
On John McCarthy’s 80th Birthday, in Honor of his Contributions

                                 Patrick J. Hayes and Leora Morgenstern

                        Abstract
  John McCarthy’s contributions to computer science
  and artificial intelligence are legendary. He invented
  Lisp, made substantial contributions to early work in
  timesharing and the theory of computation, and was
  one of the founders of artificial intelligence and knowl-
  edge representation. This article, written in honor of
  McCarthy’s 80th birthday, presents a brief biography,
  an overview of the major themes of his research, and a
  discussion of several of his major papers.

                    Introduction
Fifty years ago, John McCarthy embarked on a bold
and unique plan to achieve human-level intelligence in
computers. It was not his dream of an intelligent com-
puter that was unique, or even first: Alan Turing (Tur-
ing 1950) had envisioned a computer that could con-
verse intelligently with humans back in 1950; by the mid
1950s, there were several researchers (including Herbert
Simon, Allen Newell, Oliver Selfridge, and Marvin Min-
sky) dabbling in what would be called artificial intelli-
gence. What distinguished McCarthy’s plan was his
emphasis on using mathematical logic both as a lan-
guage for representing the knowledge that an intelligent
machine should have, and as a means for reasoning with
that knowledge. This emphasis on mathematical logic
was to lead to the development of the logicist approach
to artificial intelligence, as well as to the development
of the computer language Lisp.
   John McCarthy’s contributions to computer science
and artificial intelligence are legendary. He revolution-
ized the use of computers with his innovations in time-
sharing; he invented Lisp, one of the longest-lived com-
puter languages in use; he made substantial contribu-
tions to early work in the mathematical theory of com-
putation; he was one of the founders of the field of artifi-
cial intelligence; and he foresaw the need for knowledge        Figure 1: John McCarthy
representation before the field of AI was even properly
born.
   John McCarthy turned 80 on September 4, 2007.
This article is a celebration of that milestone. It in-
Copyright c 2007, Association for the Advancement of Ar-
tificial Intelligence (www.aaai.org). All rights reserved.
cludes a glimpse of McCarthy’s life, an overview of the     fress 1951), a conference that joined together lead-
major themes of his work, and a discussion of several of    ing researchers in different areas related to cognitive
his major papers. The aim is to introduce a more com-       science, including mathematicians Alan Turing and
plete picture of McCarthy’s long-term research to the       Claude Shannon, and psychologist Karl Lashley. As
AI Magazine readership. We hope it will help readers        he listened to the discussions comparing computers and
appreciate the range and depth of innovation in Mc-         the brain, McCarthy had a watershed moment. From
Carthy’s research.                                          that time on, his chief interests related to the devel-
                                                            opment of machines that could think like people. In-
                    Background                              deed, some of the example problems that are present in
                                                            McCarthy’s papers, such as the monkeys and bananas
McCarthy was born in Boston in 1927 to John Patrick         problem, come from the Hixon Symposium: Karl Lash-
and Ida Glatt McCarthy, immigrants from, respectively,      ley described an ape pulling over a box and climbing on
Ireland and Lithuania. The Depression started a few         it in order to get a bunch of bananas, and McCarthy
years after his birth; McCarthy’s parents lost their        took that as a base level for intelligent reasoning.
house, and the family — which now included a second
                                                               McCarthy began to consider the possibility of con-
child — became briefly peripatetic. They lived for a
                                                            structing intelligent computers. This was all at a rather
short while in New York and then in Cleveland before
                                                            abstract level, since at that point, no computers were
finally settling in Los Angeles, where the senior Mc-
                                                            available to him. He conceived of two interacting fi-
Carthy worked as an organizer for the Amalgamated
                                                            nite automata, one representing the human brain, and
Clothing Workers.
                                                            one representing the environment. When he arrived
   Like many child prodigies, John McCarthy was partly
                                                            at Princeton in 1949, he spoke to John von Neumann
self-educated. Due to childhood illness, he began school
                                                            about his ideas; von Neumann was interested and told
a year late, but he quickly made up the time on his own,
                                                            him to write them up. But McCarthy, upon further
skipped several grades, and wound up graduating from
                                                            reflection, was dissatisfied with his original idea. He re-
high school two years early. When he was a teenager
                                                            alized that he needed to somehow represent the knowl-
he developed an interest in mathematics and decided he
                                                            edge in the human brain, and that a pair of finite au-
wanted to go to the California Institute of Technology.
                                                            tomata, even if they could accurately represent the be-
At fifteen, he bought the calculus textbook then used
                                                            havior of a human brain interacting with the environ-
in the Cal Tech freshman calculus course and taught
                                                            ment, didn’t have enough structure to represent hu-
it to himself in its entirety. In the hubris of youth, he
                                                            man knowledge. Although he hadn’t yet named it, Mc-
applied only to Cal Tech, writing a one-sentence state-
                                                            Carthy had at this point already identified the need
ment of purpose on his college application: “I intend
                                                            for knowledge representation. Twenty years later, Mc-
to be a professor of mathematics.” When he arrived
                                                            Carthy was to formalize this intuition when he made
at Cal Tech, he discovered that the textbook for the
                                                            the distinction between a metaphysically adequate rep-
course had been changed. But the self-study paid off.
                                                            resentation and an epistemologically adequate represen-
Before classes started, he bumped into one of the in-
                                                            tation (McCarthy & Hayes 1969). (McCarthy’s ideas on
structors for freshman calculus, asked him a detailed
                                                            interacting automata were independently invented and
question related to one of the textbook problems on
                                                            developed by Larry Fogel some years later; see (Fogel
which he had been working, and showed him his note-
                                                            1962).)
book with worked-out problems. ”I don’t want you in
my class,” that instructor said, and arranged for him to       McCarthy finished his Ph.D. in two years, on a prob-
be given credit for freshman calculus. The same thing       lem in partial differential equations that he had discov-
happened when he met the instructor for sophomore           ered. He stayed on for a couple of years at Princeton as
calculus. Ultimately, he was plunked down, at the age       an instructor. It was during this time that he met Mar-
of 16, in graduate math class.                              vin Minsky, who began his graduate studies at Prince-
                                                            ton just as McCarthy was starting his instructorship.
                                                            McCarthy and Minsky discovered a shared passion for
                    Birth of AI                             research into constructing intelligent machines; the two
After McCarthy received his Bachelors of Science from       were to collaborate on many projects over the next
Cal Tech, he began his graduate studies there. For          decade. In 1952, McCarthy spent a summer working
McCarthy, the intellectual ferment at the time was pal-     at Bell Labs. He approached Claude Shannon with the
pable. The excitement generated by the use of early         idea of collecting papers on the topic of intelligent ma-
computers to decode enemy messages — and thus help          chines. Claude Shannon, in addition to his seminal con-
win World War II — was still running high. The the-         tributions in information theory, analog circuit design,
oretical foundations for computing had recently been        and computational linguistics, had recently published
laid by Church and Turing. There was a growing inter-       a groundbreaking paper on how one could program a
est in the workings of the mind, in what would later be     machine to play chess. Shannon agreed to collaborate
called cognitive science.                                   on such a collection, but, recalls McCarthy, didn’t want
   In September 1948, McCarthy went to the Hixon            the title to be too provocative. As a result, they called
Symposium on Cerebral Mechanisms in Behavior (Jef-          the planned volume Automata Studies (Shannon & Mc-
Carthy 1956). As the papers began to come in, Mc-              project: figuring out how to get computers to do com-
Carthy (at this point an acting assistant professor at         monsense reasoning. The work was radically different
Stanford) was disappointed when he realized that the           from that of his AI colleagues in at least two ways.
papers, while all related to automata studies in some          First, previous work in AI had focused on getting a
way, had little to do with what he regarded as getting         computer to replicate activities that are challenging
machines to do intelligent reasoning.                          for humans, such as playing chess and proving theo-
   So he decided to nail the flag to the mast and be           rems of mathematics. In contrast, McCarthy was con-
explicit about the sort of research and the sort of pa-        cerned with mundane and seemingly trivial tasks, such
pers that he wanted to encourage. He therefore came            as constructing a plan to get to the airport. Second,
up with “artificial intelligence” — a term that would          he was proposing using the tools of mathematical logic
make it clear that the goal was to construct machines          for proving something other than theorems in math-
that acted in a truly intelligent manner. It was this          ematical domains. His paper, “Programs With Com-
term that he used in 1955 when he began writing —              mon Sense” (McCarthy 1959) (often referred to as the
with Minsky, Shannon, and Nathaniel Rochester, a re-           Advice-Taker paper), articulated the need for a com-
searcher at IBM — the proposal to fund the first con-          puter to be able to perform commonsense reasoning and
ference dedicated to the topic, the famous Dartmouth           the need for a computer to have a formal representation
conference on artificial intelligence.                         of the commonsense knowledge that people use when
   The Dartmouth conference — held where McCarthy              they go about their everyday reasoning activities; ar-
was then on the faculty of mathematics — took place            gued that the representation of such knowledge, along
in the summer of 1956, and is often considered to be           with an inference method to reason with this knowl-
the start of organized research into artificial intelli-       edge, was an essential part of any artificial intelligence;
gence. One doesn’t create a field merely by naming             and advocated for the feasibility of the project. Indeed,
it, of course, and McCarthy discovered that, as with           the presentation of this paper may be seen as the birth
the Automata Studies volume, there were some partic-           of the field of knowledge representation.
ipants who came and spoke about their pet projects,               The paper generated much controversy when it was
whether or not they had anything to do with artifi-            first given in Great Britain. Philosopher and linguist
cial intelligence. Still, this served as a way of getting      Yehoshua Bar-Hillel (Bar-Hillel, McCarthy, & Selfridge
four researchers who were doing work in the field —            1998) was concerned that McCarthy’s proposed project
McCarthy, Minsky, Newell, and Simon — to meet and              was infeasible. Bar-Hillel argued, among other things,
talk and plan for future research projects in artificial       that the example of formal commonsense reasoning
intelligence. The main accomplishment of the Dart-             given in the paper was oversimplified, and that any
mouth conference was not any particular idea or ap-            proper formalization of an example would necessarily
proach to AI, but the commitment of four researchers           be much longer and more complex. In retrospect, Bar-
toward defining a discipline of artificial intelligence, and   Hillel turned out to be correct on this point; the for-
the bonds created between these colleagues.                    malization of commonsense reasoning has proved to be
                                                               a remarkably difficult enterprise. Nearly half a century
                  The MIT Years                                after the presentation of that paper, researchers are still
While McCarthy was at Dartmouth, John Kemeny,                  grappling with many of the underlying difficulties.
then chairman of Dartmouth’s math department, ar-                 This was also the period in which McCarthy and
ranged for McCarthy to receive a one-year Sloan fel-           Minsky (who had arrived at MIT in 1958) established
lowship. This allowed him to spend the 1956-1957 aca-          the MIT AI Lab. As McCarthy tells this classic tale
demic year at any institution of his choosing. He de-          of ask-and-you-shall-receive, he and Minsky were talk-
cided on MIT, because it had access to an IBM com-             ing in the hallway when he saw Jerome Weisner, then
puter. (The computer was physically located at MIT,            acting head of MIT’s Department of Electrical Engi-
but MIT had access to it for only one eight-hour shift         neering, walking down the hall. McCarthy buttonholed
each day. One eight-hour shift was reserved for other          him and said, “Marvin and I want to have an artifi-
New England colleges and universities, and one eight-          cial intelligence project.” Weisner asked, “Well, what
hour shift was reserved for IBM’s own use.) MIT also           do you need?” Minsky said, “A room, a secretary, a
afforded close proximity to his colleague Marvin Min-          key punch, and two programmers.” Weisner responded,
sky, who was at that point at Harvard. MIT offered             “How about 6 graduate students?” (Funding for these
McCarthy a faculty position once the Sloan fellowship          was available due to a recent prize awarded to the math
ended, and McCarthy never returned to his Dartmouth            department.) And thus the AI Lab was born.
post.                                                             McCarthy and Minsky did not always see eye-to-eye
   The six years that McCarthy spent at MIT were one           on their respective approaches to artificial intelligence.
of the most productive periods in his life. During this        McCarthy became increasingly committed to the logi-
period, he conceived of the idea of timesharing, sug-          cist approach to AI; Minsky came to believe that it was
gested using automated theorem proving for program             wrong-headed and infeasible (Minsky 1975). Nonethe-
verification (McCarthy 1963a), and invented Lisp. He           less, both approaches continued to progress during this
also began what was to become his lifelong research            period.
The Stanford Years                               The problems that McCarthy has chosen to work on
In 1962, McCarthy received an offer of a faculty posi-       all flow from his general goal of formalizing common-
tion from Stanford University’s Department of Mathe-         sense reasoning. Indeed, much of it can be seen to
matics. McCarthy had worked there previously, in the         stem from the small example of commonsense reason-
early 1950s. Of three acting assistant professors, he was    ing that McCarthy discussed in his 1958 Advice Taker
the one person whose contract had not been renewed.          paper, that of a person planning to go to the airport
(This had precipitated his move to Dartmouth.) At            while sitting at his desk, given that he has a car in
that point an associate professor at MIT, he told the        his garage. McCarthy realized early on that virtually
department chairman, George Forsyth, that he would           all commonsense reasoning involves reasoning about ac-
accept no less than a full professorship. Forsyth, ac-       tion and change (or the absence thereof). The invention
cording to McCarthy, had quite a job convincing the          and development of the situation calculus was meant to
Stanford administration to agree to hiring as full profes-   provide a framework that facilitated such reasoning.
sor someone whose contract hadn’t even been renewed             The detailed formalization, within the situation cal-
less than a decade before, but he was enough of a vi-        culus, of even simple problems — such as planning to
sionary to realize the importance of computer science        build a tower of blocks — led to the discovery of a host
and artificial intelligence. In any case, McCarthy was       of other problems. What Bar Hillel had feared in 1958
only briefly in the Department of Mathematics. Stan-         — that the formalization of enough knowledge to solve
ford soon created a Department of Computer Science,          simple commonsense reasoning problems was, indeed,
and McCarthy was one of that department’s original           a difficult task — became evident. For example, Mc-
members.                                                     Carthy (working with Pat Hayes) realized that a simple
                                                             plan to build a tower of three blocks raised unforeseen
Research in Knowledge Representation                         difficulties. If three blocks, A, B, and C, lay on a table,
                                                             it seemed obvious that a plan to put B on C and A on
It was at Stanford that McCarthy began working out,          B would succeed in building a tower. But when one
and has continued up to the present time to work out,        formalizes this small problem, the question arises: how
most of his research in formal knowledge representation.     does one know that A is still on the table after placing
Most of his research falls into four areas:                  B on C? This is an instance of the frame problem, which
1. Reasoning about actions. This includes McCarthy’s         concerns the ability to represent and reason efficiently
   original work on situations (McCarthy 1963b), culmi-      about what stays the same and what changes as actions
   nating in the situation calculus (McCarthy & Hayes        are performed.
   1969), as well as more recent extensions (McCarthy           A related problem is the qualification problem, which
   2002); the discovery of several challenging problems      concerns the ability to reason about the many condi-
   for knowledge representation, such as the frame and       tions that must be true in order for an action to be
   qualification problems; and initial solutions to those    performed successfully. For example, in the Missionar-
   problems.                                                 ies and Cannibals problem, in which one reasons about
                                                             how 3 cannibals and 3 missionaries can use 3 boats to
2. Nonmonotonic reasoning. This includes McCarthy’s          safely cross a river, there are certain implicit qualifica-
   development of domain circumscription (McCarthy           tions in the rules stated for the problem, such as the
   1977), predicate circumscription (McCarthy 1980),         fact that the boats do not leak and have oars, and that
   and formula circumscription (McCarthy 1986).              there is no bridge.
3. Issues related to reification. McCarthy has been a           These problems eventually led McCarthy to believe
   proponent of using the technique of reification, in       that one must be able to go beyond classical logic, in
   which sentences and other complex constructs of first-    which one reasons about things that are always true, to
   order logic are mapped to terms of first-order logic      a default or nonmonotonic logic, in which one reasons
   in order to enable formalization of commonsense rea-      about things that are typically true. For this purpose,
   soning within first-order logic. This work includes his   McCarthy introduced circumscription, an extension of
   ongoing interest in contexts (McCarthy 1990; 1993;        classical logic which allows one to prefer certain models
   McCarthy & Buvac 1997).                                   of a theory.
                                                                McCarthy developed several increasingly powerful
4. Reasoning about knowledge. McCarthy’s work in             theories: domain circumscription, in which one prefers
   this area includes the discovery of the knowledge pre-    models which have minimal domains; predicate circum-
   conditions problem (McCarthy & Hayes 1969) and            scription, in which one prefers models where certain
   work on variations of classic puzzles involving com-      predicates have minimal extensions; and formula cir-
   mon knowledge, such as the Three Wise Men or              cumscription, in which one prefers models where certain
   Muddy Children problem (Fagin et al. 1995).               formulas have minimal extensions.
   These areas overlap and influence one another. For           There were several other researchers investigating
example, as discussed below, the frame and qualifica-        techniques for nonmonotonic reasoning during the late
tion problems were the primary impetus for the devel-        1970s and early 1980s, including Ray Reiter (Reiter
opment of circumscription.                                   1980) and Drew McDermott and Jon Doyle (McDer-
mott & Doyle 1980). McCarthy’s work can be distin-           Leadership and Recognition
guished in two ways. First, he aimed to stay as close as     During his years at Stanford, John McCarthy
possible to classical logic. The principle of circumscrip-   advised more than 30 Ph.D. students; he is
tion can be represented as a first-order axiom schema        listed in the Mathematics Geneaology Database, at
or as a second-order axiom. In contrast, McDermott           http://www.genealogy.ams.org, as having 175 academic
and Doyle used a modal logic, and Reiter introduced          descendants. 1 The number of researchers who have
an entirely new inference rule. Second, McCarthy has         collaborated with him or who have been influenced,
sought to tie his work in nonmonotonic reasoning to the      through personal contact with him, or by his writings
specific applications in which he has been interested.       and his vision, is considerably greater. For example,
   For example, McCarthy showed how to use circum-           Lifschitz’s body of work on circumscription from the
scription over a set of abnormality predicates to formal-    1980s and 1990s was directly influenced by his contact
ize inheritance. To formalize, e.g., the facts that things   with McCarthy during his years at Stanford; Reiter’s
typically don’t fly, but that birds typically do fly, but    book Knowledge in Action(Reiter 2001) is rooted in the
that, on the other hand, penguins typically do not fly,      reworking and extension of McCarthy’s original situa-
one could write the theory:                                  tion calculus; and Bob Kowalski cites his early work
∀ x (Thing(x) ∧ ¬ ab1(x) ⇒ ¬ Flies(x))                       on the Event Calculus (Kowalski & Sergot 1986) as be-
∀ x (Bird(x) ∧¬ ab2(x) ⇒ Flies(x))                           ing heavily influenced by McCarthy’s situation calcu-
∀ x (Penguin(x) ∧¬ ab3(x) ⇒ ¬ Flies(x))                      lus. The centers of logicist AI today, at the Univer-
∀ x (Bird(x) ⇒ Thing(x))                                     sity of Toronto, at the University of Texas at Austin,
∀ x (Bird(x) ⇒ ab1(x))                                       at Linkoping University in Sweden, at Imperial College
∀ x (Penguin(x) ⇒ Bird(x))                                   in London, and at many other universities around the
∀ x (Penguin(x) ⇒ ab2(x))                                    world, owe much to McCarthy’s groundbreaking ideas.
   If one adds the facts                                        McCarthy’s influence is due not only to the strength
Thing(Sylvester)                                             of his ideas but also to his personal qualities. He is
Bird(Tweety)                                                 the ultimate optimist, and his belief in the power of
Penguin(Opus)                                                formal logic is infectious. His generosity of spirit has
and circumscribes the predicates ab1, ab2, and ab3, one      nurtured many a new researcher in the area of logicist
gets the desired result that Sylvester and Opus do not       AI. His philosophy is to let as much research as possible
fly, while Tweety does fly.                                  flourish; his delight in hearing about new work in the
                                                             field is evident.
   McCarthy also suggested using circumscription to
                                                                Although McCarthy has remained mostly at the side-
handle the frame problem, by formulating the principle
                                                             lines of academic politics, he has been active in orga-
of inertia — that fluents typically don’t change when
                                                             nizations close to his research interests. He founded
actions are performed — using abnormality predicates.
                                                             SAIL, the Stanford AI Laboratory shortly after he came
His formulation turned out to be overly simplistic, and
                                                             to Stanford, and served as president of AAAI from
as a result, led to incorrect conclusions; see, e.g., the
                                                             1984-1985. In 1991, hoping to reverse the trend of
blocks-world example of Lifschitz (Lifschitz 1986) and
                                                             logicist AI toward producing metalevel results rather
the Yale Shooting problem (Hanks & McDermott 1987).
                                                             than object-level theories, he founded and organized
The difficulty is that the principle of inertia can apply
                                                             the first Symposium on Logical Formalizations of Com-
to multiple fluents, or properties, not all of which can
                                                             monsense Reasoning. The symposium is now held every
be simultaneously minimized; minimizing change can
                                                             two years.
thus lead to multiple models, some of which are unintu-
itive. This difficulty does not arise in theories in which      In addition, McCarthy has used his academic posi-
there is both an explicit theory of causation and a corre-   tion to further humanitarian causes, particularly dur-
spondingly more realistic formalization of the principle     ing the years when the Soviet Union existed. McCarthy,
of inertia (e.g., (Lifschitz 1987)).                         who learned Russian from foreign language records in
                                                             his twenties, and, as a graduate student, translated
   From the standpoint of logicist AI, the development       a text in differential equations from Russian into En-
of formal theories such as circumscription that enabled      glish, made several visits to the Soviet Union starting
nonmonotonic reasoning was crucial for the survival of       in 1965. In 1968, he taught for two months in Akadem-
the logicist agenda. Critics had correctly pointed out       gorodok, on Novosibirsk’s outskirts, and in Novosibirsk
that much commonsense reasoning was nonmonotonic             itself. In 1975, he was instrumental in getting cybernet-
in nature and could not be formalized within a pure          ics researcher and refusenik Alexander Lerner permis-
first-order logic. Showing that nonmonotonic reasoning       sion from Soviet officials to attend and talk at the 4th
could be formalized within an extension of first-order       International Joint Conference on Artificial Intelligence
logic provided evidence that the logicist agenda was in
fact feasible. Indeed, much of McCarthy’s work during           1
                                                                  This is likely an underestimate, since several of Mc-
his years at Stanford has focused on showing that the        Carthy’s Ph.D. students are not listed in the database;
concerns that others have had about the logicist agenda      moreover, in general not all an individual’s academic de-
can be addressed in formal logic.                            scendants are reported to this database.
(IJCAI) in Tbilisi, Georgia. In the 1980s he smuggled a      This paper was groundbreaking first, for proposing that
fax and copier machine to linguist and Soviet dissident      commonsense reasoning (e.g., figuring out a way to get
Larisa Bogoraz.                                              to the airport) was as worthy of study as seemingly
   John McCarthy’s many awards include the Turing            difficult intellectual problems (e.g. playing chess); sec-
Award (1971), the first IJCAI Award for Research Ex-         ond, for suggesting that commonsense knowledge and
cellence (1985), the Kyoto Prize (1988), the National        reasoning could be expressed in first-order logic; and
Medal of Science (1990), and the Benjamin Franklin           third, in recognizing the centrality of action and causal-
Medal in Computer and Cognitive Sciences (2003). He          ity in commonsense reasoning. McCarthy deliberately
is a founding fellow of AAAI (1990), and a member of         avoided the representation of time in this paper. He
the American Academy of Arts and Sciences (1974),            believed that the proper representation of time would
the National Academy of Engineering (1987), and the          be isomorphic to the real numbers, but that knowledge
National Academy of Sciences (1989).                         of the theory of real numbers, developed only in the
                                                             19th century, could hardly be considered part of com-
Retirement                                                   monsense reasoning.
John McCarthy officially retired from Stanford on Jan-       3. John McCarthy: “Situations, Actions, and
uary 1, 2001. The last of his Ph.D students, Eyal Amir       Causal Laws,” unpublished Stanford Technical
and Aarati Parmar Martino, defended their disserta-          Report, 1963. (This paper can be found, tacked onto
tions in 2002 and 2003, respectively.                        the end of “Programs With Common Sense,” in Marvin
   From McCarthy’s point of view, retirement has             Minsky’s Semantic Information Processing, MIT Press,
meant more time for research and writing papers. He          1968.) This is the first paper on the situation calculus
has been active in research for nearly sixty years, and      (though it was not yet so named), a language that facil-
he has been grappling with the central problems of arti-     itates representation of actions and change. McCarthy
ficial intelligence and commonsense reasoning for more       had come to realize that he needed to distinguish flu-
than half a century. But he founded a difficult field of     ents, properties that change over time. The notion of
research. It is no easy task to imbue a machine with         a situation, a (complete) description of the world at a
enough knowledge to do commonsense reasoning. Fifty          particular moment in time, allowed him to represent
years have not been nearly enough to finish the job,         fluents but still not explicitly mention time.
so McCarthy — along with the many AI researchers             4. John McCarthy and Patrick J. Hayes: “Some
who, inspired by his early vision, now follow his logicist   Philosophical Problems from the Standpoint of
philosophy — continues to work at it.                        Artificial Intelligence,” Machine Intelligence 4,
                                                             1969. The central argument of this paper is that in
                                                             order to successfully develop programs that can rea-
                                                             son intelligently, we will have to deal with many of
APPENDIX (can this be put in as a side-                      the problems that have concerned philosophers for cen-
bar?)                                                        turies; and that, moreover, the demands of AI can put
Reading John McCarthy: The Top Eleven                        a new spin on old problems. The major contributions
— A Selected Annotated Bibliography                          of the paper are first, a complete presentation of the
                                                             situation calculus; second, a discussion of several ar-
   Listed below are eleven of John McCarthy’s papers,        eas of philosophy, including logics of knowledge, modal
selected for their enduring technical importance and/or      logics, and tense logics, which are relevant for artifi-
their historical interest, along with a precis and com-      cial intelligence; and third, the presentation of several
ments. Most of John McCarthy’s works are available           crucial problems for knowledge representation. These
on his website, at http://www-formal.stanford.edu/jmc        problems include the frame problem, the problem of ef-
. Six of the papers in this list — 2, 3, 4, 5, 6, and 7      ficiently determining what remains the same in a chang-
— can also be found in McCarthy’s collected papers on        ing world, and what later came to be called the knowl-
commonsense reasoning (Lifschitz 1990).                      edge preconditions problem, the problem of determining
1. John McCarthy: “Recursive Functions of                    what an agent (or set of agents) needs to know in order
Symbolic Expressions and Their Computation                   to perform an action.
By Machine,” Communications of the ACM,                      5. John McCarthy: “Epistemological Problems
1960. This is the original paper describing Lisp. It         of Artificial Intelligence,” Proceedings, IJCAI
introduces a methodology in which a program in a lan-        1977. This paper presents a number of problems and
guage could also be considered data for a(nother) pro-       themes that have become central to research in knowl-
gram in that language, a capability essential for reflec-    edge representation. The paper includes the first dis-
tive reasoning. Many of the examples used in standard        cussion of the qualification problem, which concerns how
Lisp textbooks come directly from this paper.                one can represent the implicit qualifications in our gen-
2. John McCarthy: “Programs With Common                      eral knowledge. It also includes one of the first dis-
Sense,” Proceedings, Teddington Conference on                cussions of circumscription, an extension of first-order
the Mechanization of Thought Processes, 1958.                logic which allows one to prefer certain models of a first-
order theory. Circumscription was developed in order        tual is one which, if true, potentially teaches a useful les-
to permit nonmonotonic reasoning, that is, the ability      son: contrast the useful counterfactual If a car had been
to jump to conclusions and later retract some of those      approaching while you switched lanes, you would have
conclusions.                                                crashed with the useless counterfactual If wishes were
6. John McCarthy, “Circumscription: A Form                  horses, beggars would ride. An agent can learn from
of Nonmonotonic Reasoning,” Artificial Intel-               useful counterfactuals without experiencing adverse
ligence, 1980. This paper gives a detailed account          events; and can use them to do hypothetical reason-
of predicate circumscription, the most commonly used        ing while constructing a plan. This paper discusses the
form of circumscription. Predicate circumscription          evaluation of such useful counterfactuals, using the no-
minimizes the extension of certain predicates; that is,     tion of a Cartesian counterfactual, which allows giving a
models with minimal extensions of those predicates are      precise meaning to Lewis’s and Stalnaker’s (Lewis 1973;
preferred. The circumscriptive formula is given as a        Stalnaker 1984) imprecise but intuitively appealing no-
first-order schema.                                         tion of a most similar possible world.
7. John McCarthy, “Applications of Circum-                  11. John McCarthy, “Actions and Other Events
scription to Formalizing Common Sense Knowl-                in Situation Calculus,” Proceedings, KR-2002.
edge,” Artificial Intelligence, 1986. This paper ex-        This paper extends the original McCarthy-Hayes situ-
tends the idea of predicate circumscription to the more     ation calculus in several significant ways, in order to
general form of formula circumscription. It presents        handle natural events, concurrency, and combined lin-
a second-order circumscriptive axiom, rather than the       ear and branching time.
first-order schema originally used. The paper suggests
several ways of applying circumscription to common-                               References
sense reasoning problems, such as inheritance and the
frame problem.                                               Bar-Hillel, Y.; McCarthy, J.; and Selfridge, O. 1998.
                                                             Discussion of the paper: Programs with common
8. John McCarthy and Sasa Buvac, “Formal-                    sense. In Lifschitz, V., ed., Formalizing Common
izing Context: (Expanded Notes),” Computing                  Sense. Intellect. 17–20.
Natural Language, 1997. This is the third and most
                                                             Costello, T., and McCarthy, J. 1999. Useful counter-
comprehensive of four papers that McCarthy wrote,
                                                             factuals. Electronic Transactions on Artificial Intelli-
over the course of a decade, about the idea of reason-
                                                             gence 3(A):51–76.
ing using contexts. McCarthy sought to model the ease
with which people reason relative to particular contexts.    Fagin, R.; Halpern, J. Y.; Moses, Y.; and Vardi, M. Y.
For example, one can reason about facts that are true in     1995. Reasoning About Knowledge. Cambridge, Mas-
the context of a Sherlock Holmes story (e.g., Holmes’s       sachusetts: The MIT Press.
feud with Moriarty) although they are not true in the        Fogel, L. J. 1962. Autonomous automata. Industrial
actual world. Various technical mechanisms are intro-        Research 4:14–19.
duced to deal with contexts, such as lifting formulas,       Guha, R. V. 1991. Contexts: A Formalization and
which relate propositions and terms in subcontexts to        Some Applications. Ph.D. Dissertation, Stanford Uni-
more general propositions and terms in outer contexts.       versity. Also available as technical repot STAN-CS-91-
Lifting formulas are needed to combine or transcend          1399-Thesis and as MCC Technical Report ACt-CYC-
contexts.                                                    423-91.
   One of McCarthy’s Ph.D. students during this pe-          Hanks, S., and McDermott, D. V. 1987. Nonmono-
riod, R.V. Guha, who wrote his thesis (Guha 1991) on         tonic logic and temporal projection. Artificial Intelli-
reasoning with contexts, incorporated contexts into the      gence 33(3):379–412.
Cyc knoweldge base (Lenat & Guha 1990). They re-             Jeffress, L. A. 1951. Cerebral Mechanisms in Behavior:
main an important part of the Cyc project.                   The Hixon Symposium. New York: Wiley.
9. John McCarthy, “Elaboration Tolerance,”                   Kowalski, R. A., and Sergot, M. J. 1986. A logic-
Working Papers, Commonsense-1998. A com-                     based calculus of events. New Generation Computing
monsense theory typically contains general common-           4(1):67–95.
sense information as well as a formalization of specific
circumstances. A theory is said to be elaboration tol-       Lenat, D. B., and Guha, R. V. 1990. Building Large
erant if modifying information about specific circum-        Knowledge Based Systems: Representation and Infer-
stances requires making only moderate changes to the         ence in the Cyc Project. Reading, Massachusetts: Ad-
general commonsense theory. This paper discusses var-        dison Wesley.
ious types of elaborations, using the Missionaries and       Lewis, D. 1973. Counterfactuals. Blackwell.
Cannibals problem as a motivating example.                   Lifschitz, V. 1986. Pointwise circumscription: Prelim-
10. Tom Costello and John McCarthy, “Use-                    inary report. In AAAI, 406–410.
ful Counterfactuals,” Electronic Transactions of             Lifschitz, V. 1987. Formal theories of action (prelimi-
Artificial Intelligence, 1999. A useful counterfac-          nary report). In IJCAI, 966–972.
Lifschitz, V. 1990. Formalizing Common Sense: Pa-             Shannon, C. E., and McCarthy, J. 1956. Automata
pers by John McCarthy. Ablex.                                 Studies. Princeton University Press.
McCarthy, J., and Buvac, S. 1997. Formalizing con-            Stalnaker, R. 1984. Inquiry. MIT Press.
text: Expanded notes. In Aliseda, A.; van Glabbeek,           Turing, A. M. 1950. Computing machinery and intel-
R.; and Westerstahl, D., eds., Computing Natural Lan-         ligence. Mind 59:433–460.
guage. Stanford University. Also available as Stanford
Technical Note STAN-CS-TN-94-13.                                Pat Hayes received a B.A. in mathematics from
McCarthy, J., and Hayes, P. J. 1969. Some philo-             Cambridge University and a Ph.D. in Artificial Intelli-
sophical problems from the standpoint of artificial in-      gence from Edinburgh. He has held academic positions
telligence. In Meltzer, B., and Michie, D., eds., Ma-        at the University of Essex (England), the University of
chine Intelligence 4. Edinburgh: Edinburgh University        Illinois, and as the Luce Professor of cognitive science
Press. 463–502.                                              at the University of Rochester. He has been a visiting
McCarthy, J. 1959. Programs with common sense.               scholar at Universite de Geneve and the Center for Ad-
In Proceedings of the Teddington Conference on the           vanced Study in the Behavioral Sciences at Stanford,
Mechanization of Thought Processes, 75–91. London:           and has directed applied AI research at Xerox-PARC,
Her Majesty’s Stationery Office.                             SRI, and Schlumberger, Inc.. Pat has served as secre-
                                                             tary of AISB, chairman and trustee of IJCAI, associate
McCarthy, J. 1960. Recursive functions of symbolic           editor of Artificial Intelligence, a governor of the Cog-
expressions and their computation by machine. Com-           nitive Science Society, and president of AAAI. His cur-
munications of the ACM 3(4):184–195.                         rent research interests include knowledge representation
McCarthy, J. 1963a. A basis for a mathematical the-          and automatic reasoning, especially the representation
ory of computation. In Computer Programming and              of space and time; the semantic web; ontology design;
Formal Systems. North-Holland.                               and the philosophical foundations of AI and computer
McCarthy, J. 1963b. Situations, actions, and causal          science. He can be reached at phayes@ihmc.us.
laws. Technical report, Stanford University.                    Leora Morgenstern is a senior research scientist at
McCarthy, J. 1977. Epistemological problems of arti-         the IBM T.J. Watson Research Center in Hawthorne,
ficial intelligence. In IJCAI, 1038–1044.                    NY. She has taught at Brown University, Columbia
McCarthy, J. 1980. Circumscription: A form of                University, and the University of Pennsylvania. She
non-monotonic reasoning. Artificial Intelligence 13(1–       has a B.A. in mathematics and philosophy from the
2):23–79.                                                    City College of New York, and a Ph.D. in computer
                                                             science from the Courant Institute of Mathematics at
McCarthy, J. 1986. Applications of circumscription           New York University. Her research interests include
to common sense reasoning. Artificial Intelligence           reasoning about knowledge and belief, nonmonotonic
28(1):89–116.                                                reasoning, and multi-agent planning. She is particularly
McCarthy, J. 1990. Generality in artificial intelli-         interested in applying knowledge representation tech-
gence. In Lifschitz, V., ed., Formalizing Common             niques to commercial applications, and has extended
Sense. Ablex. 226–236.                                       techniques in semantic networks, nonmonotonic inheri-
McCarthy, J. 1993. Notes on formalizing context. In          tance, and intelligent agents for the development of ex-
IJCAI, 555–562.                                              pert systems for benefits inquiry in medical insurance,
McCarthy, J. 1998. Elaboration tolerance. In Work-           and intelligent recommendation systems for banking,
ing Papers of the Fourth International Symposium on          insurance, telephony, and sales. She can be reached at
Logical Formalizations of Commonsense Reasoning,             leora@steam.stanford.edu and leora@us.ibm.com.
Commonsense-1998.
McCarthy, J. 2002. Actions and other events in situa-
tion calculus. In Fensel, D.; Giunchiglia, F.; McGuin-
ness, D.; and Williams, M., eds., Proceedings of KR-
2002, 615–628.
McDermott, D. V., and Doyle, J. 1980. Non-
monotonic logic i. Artificial Intelligence 13(1-2):41–72.
Minsky, M. 1975. A framework for representing knowl-
edge. In Winston, P. H., ed., The Psychology of Com-
puter Vision. McGraw-Hill. Also available as MIT-AI
Lab Memo 306.
Reiter, R. 1980. A logic for default reasoning. Artificial
Intelligence 13(1–2):81–132.
Reiter, R. 2001. Knowledge in Action. Cambridge,
Massachusetts: MIT Press.
You can also read