UNIVERSITY
OF FLORIDA
LIBRARIES

ENGINEERING AND PHYSICS
LI BRARY

Management and the Computer of the Future

Published jointly by

THE M.I.T. PRESS
Massachusetts Institute of Technology

and

JOHN WILEY & SONS, INC.
New York • London

1962

Royalties received from the sale of this book
are deposited in a fund established by the
School of Industrial Management at M.I.T.
The purpose of the fund is to advance education and research
in management applications of the computer
and of quantitative analysis.

Copyright © 1962
by

The Massachusetts Institute of Technology
All rights reserved

Library of Congress Catalog Card No.: 62-13234

PRINTED IN THE UNITED STATES OF AMERICA

In celebration of M.I.T.'s Centennial Year,
the School of Industrial Management of the
Massachusetts Institute of Technology takes
pleasure in inviting you to a series of evening
lectures on the theme, "Management and the
Computer of the Future"

— Original announcement of the lecture series,

March 1961.

Preface

M.I.T. is now one hundred years old; the automatic digital
computer, more or less as we know it today, is not yet twenty.
It was born in 1944, and quickly learned to read, write, and re-
member, as well as to add, subtract, multiply, and divide. It was
indeed a precocious youngster, although it did make childish
mistakes.

In a brief span of years, the computer has grown remarkably
in every way except physical size. It has become faster, more
versatile, and more reliable; it has become a much better com-
municator with its external environment; it has developed a
memory of impressive speed and capacity; and it has become
articulate in many languages.

As the computer matured, we gradually delegated to it more
and more duties that we previously had discharged ourselves.
First we asked the computer to do only routine, repetitive cal-
culation: to invert our matrices, solve our equations, and com-
pile our lengthy numerical tables. Then we put it to work on
data processing and some other business chores: figuring our
payroll, writing our checks, and keeping our inventory and ac-
counting records. Heartened by its dogged endurance, and in-

v

Preface

trigued by its ability to make choices by itself, we soon began
to entrust it with certain responsibilities of a semisupervisory
nature: the control of refinery operations, machine tooling, and
other real-time processes. We also entrusted it with some mili-
tary functions of awesome significance and proportions: infor-
mation processing for continental air defense, surveillance of the
seas, and anticipation of logistical requirements for massive
troop movements.

It was not too long before we discovered that the computer
was entirely capable of working with nonnumerical symbols and
words as well as with numbers. Some of us felt a little uncom-
fortable about this at first, since we had become accustomed to
thinking of the computer as a numerical specialist. But watch-
ing it differentiate complicated mathematical expressions, work
out integration problems of freshman calculus, and even do
some language translation, made us realize that we had under-
estimated its abilities.

As if to take heed of the adage about "all work and no play,"
during more frivolous moments we played games with the com-
puter, such as checkers and chess. Some of this sport developed
into serious interests. Playing with the computer provided us
with simulated experience, which was obtained more quickly
and less expensively than real experience. This led us to design
business games and war games to help train our corporation
executives and our military officers.

Gaming required that we simulate real systems, and there-
fore that we construct models which could be run on the com-
puter. In some cases we found that we could make a model
valid enough to increase our understanding of the system it rep-
resented. We also found that simulation on a computer could
afford us a means for exploring properties of the model under
laboratory-like conditions. If we had a reasonably accurate
model of New York City traffic, for example, we could investi-
gate the effect on it of alternate policies for controlling auto-
mobile and pedestrian flow, in order to see which policies re-
sulted in lowest over-all waiting and congestion. The range of

vi

Preface

application of this simulation approach seemed to extend from
the simplest inventory system to the most complex system of a
national economy. We set out to see if we could develop the
necessary models.

The more adventuresome among us entertained hopes of en-
dowing the computer with problem-solving abilities of a rela-
tively general nature. It was demonstrated that the computer
could be taught to prove theorems of geometry and logic, read
handwritten letters and other man-made patterns, solve puzzles,
and even write sonatas. We came to refer to this activity as
artificial intelligence, just to make sure that the line between
the computer and us remained sharp and clear.

Now an artificially intelligent computer which can automati-
cally control a live process at the same time that it simulates
the process, appears, at least in the abstract, to oEer very ex-
citing possibilities for the future. While this computer is guid-
ing the flow of traffic through New York City, it may also be
evaluating potential improvements in policy which it has de-
rived itself. A similar computer might be overseeing the opera-
tion of a factory, and still another computer might be routing
warehouse shipments and ordering replenishments without as-
sistance. Some people believe that one day the computer may
"take over the business."

All in all, the computer has had a very rapid and extraordi-
nary growth in the past two decades. This growth raises inter-
esting and difficult questions for society, for government, foi
science, for education, and for management. The problems
confronting the last group, in particular, prompted us to spon-
sor a series of eight lectures in the spring of 1961 to honor
M.I.T. on the occasion of its one-hundredth birthday. This
book is the product of that lecture series.

In planning the program of the series, it quickly became ap-
parent that the topics of greatest importance to management in
the long run were not directly in the management domain at
present. It also was evident that the interests of society, govern-
ment, science, and education could not altogether be neglected

vii

Preface

with clear conscience. These considerations strongly influenced
the final selection of speakers and subjects.

The format of each lecture was the same. The speaker pre-
sented a main paper, after which each of two discussants de-
livered prepared remarks based on a prior reading of the paper.
Then the speaker, who did not have the advantage of a prior
reading of the discussants' remarks, was invited to reply. A
general discussion followed, with members of the audience en-
couraged to air their own views or to ask questions of the panel.
Audiences ranged in size from several hundred to almost two
thousand, and represented a cross section of the university and
industrial communities.

Tape recordings were made of all sessions. At the conclusion
of the series the tapes were edited, in many cases substantially,
and all contributors (speakers and discussants) were given an
opportunity to review the edited version of the chapter to
which they had contributed. The timetable for publication
made it impossible to extend the same privilege to other partic-
ipants in the general discussions, and, as editor, I take full re-
sponsibility for any distortions of meaning introduced as a re-
sult.

One of the most noteworthy aspects of the lecture series was
the wide range of opinions evoked by the emphasis on the
future. Spirited exchanges in the discussions were the rule rather
than the exception. The "experts" contested points made by
"amateurs," while the amateurs challenged the authority of the
experts. One contributor referred to himself as "a little left of
center" and characterized the complete spectrum in the follow-
ing way: "On the far right were those who view computers as
merely glorified slide rules useful only in carrying out compu-
tation, while at the far left were scientists who feel computers
are already doing something pretty close to thinking." If the
reader were to place the rest of the contributors in their places
along this spectrum, he would find most of them falling be-
tween dead center and far left. This may be only the reflection
of an unintentionally biased sample, or it may be a road sign to
the future.

viii

Preface

A selected bibliography at the end of the book is intended
for the person who wishes to read further. The bibliography is
based on the recommendations of the participants.

Martin Greenberger

Cambridge, Massachusetts
January 31, 1962

ix

Acknowledgments

In working out the format and program of the lecture series, I
benefited from the suggestions and ideas of many colleagues
and friends. Sidney Alexander and Howard Johnson served with
me on the program committee and took major roles in the
planning. Also helpful during the early stages were Robert
Fano, William Locke, Donald Marquis, Elting Morison, Philip
Morse, Emanuel Piore, Claude Shannon, Herbert Simon, and
Robert Solow. James Murphy of M.I.TYs Kresge Auditorium
and William Struble of the M.I.T. Public Relations Office as-
sisted me greatly in making arrangements, as did my secretary
Carmen Bach Ceder. A grant from the International Business
Machines Corporation made the lecture series financially feas-
ible.

In the preparation of this book, the very difficult and often
dreary job of typing rough drafts from tape, and later drafts
from my pencil-scratched revisions and re-revisions, was accom-
plished chiefly by Claire Curewitz. During the editing process,
members of the M.I.T. Press, particularly Constance Boyd and
Lynwood Bryant, provided me with much-needed aid. The con-
tributors to the series were uniformly cooperative and under-
standing in meeting the many deadlines and accepting the
many editorial modifications. My work has been made enjoy-
able by all of these people, and especially by my wife Ellen,
who served as co-worker and colleague throughout.

M. G.

xi

Biographical Notes

Included are thumbnail sketches of all contributors and session
chairmen. Present positions are omitted, as a rule, since they appear
elsewhere in the book. Following the sketches is a list of other
identified participants in the general discussions, with titles and
affiliations where obtainable.

Sidney S. Alexander was born in Forest City, Pennsylvania, in
1916. He received S.B., A.M., and Ph.D. degrees from Harvard,
and served there as assistant professor of economics. He also studied
at King's College, Cambridge University. Professor Alexander has
been associated at different times with the Bureau of Economic
Research, the Office of Strategic Services, the Treasury Department,
the State Department, the Rand Corporation, and the International
Monetary Fund. Before joining the M.I.T. faculty in 1956, he was
economic adviser to the Columbia Broadcasting System.

Gene M. Amdahl was born in Flandreau, South Dakota, in 1922.
He received a B.S. degree in engineering physics from South Dakota
State College and M.S. and Ph.D. degrees from the University of
Wisconsin. His doctoral thesis at Wisconsin contained the original
design of the Wise computer. At IBM Dr. Amdahl has been
project engineer for the design and development of the IBM 704
computer, as well as major systems designer for the IBM 709 and
Stretch computers.

George W. Brown was born in Boston, Massachusetts, in 1917.
He received A.B. and A.M. degrees from Harvard and a Ph.D.
degree from Princeton in mathematics. Professor Brown has been
research statistician with R. H. Macy and Company, research as-
sociate at Princeton, research engineer with RCA's Princeton lab-
oratories, professor of mathematical statistics at Iowa State College,
mathematician with the Rand Corporation, and senior staff engi-

xiii

Biographical Notes

neer at International Telemeter Corporation. He has been a vice
president and director of General Analysis Corporation, and in
addition to his post at UCLA is currently a director of CEIR, Inc.

Vannevar Bush was born in Everett, Massachusetts, in 1890. He
received B.S. and M.S. degrees from Tufts College and an Sc.D.
degree jointly from Harvard and M.I.T. Dr. Bush is best known
for his pioneering work in computer technology and for his efforts
in mobilizing science during World War II. He developed the
Bush differential analyzer during the 1930's and was director of the
Office of Scientific Research and Development during World War
II. Dr. Bush is past president of the Carnegie Institution of Wash-
ington; he is one of the original founders of the Raytheon Manu-
facturing Company; and he is present chairman of the board of
Merck & Company, Inc. Recipient of many awards, medals, and
honorary degrees, Dr. Bush is coauthor of Principles of Electrical
Engineering and author of Operational Circuit Analysis, Endless
Horizons, and Modern Arms and Free Men.

Peter Elias was born in New Brunswick, New Jersey, in 1923. He
received an S.B. degree from M.I.T. and A.M., M.Eng.Sci., and
Ph.D. degrees from Harvard. At Harvard he was a Junior Fellow
until he was appointed to the M.I.T. faculty in 1953. Professor
Elias* major research interests include information theory and the
analysis of pictures. He is coeditor of the journal, Information and
Control.

Robert M. Fano was born in Turin, Italy, in 1917. After doing
part of his undergraduate work at the School of Engineering of
Turin, he received both S.B. and Sc.D. degrees in electrical en-
gineering from M.I.T. He was formerly a group leader of Lincoln
Laboratory, and during World War II a staff member of the
Radiation Laboratory. In addition to his teaching at M.I.T., Pro-
fessor Fano is a staff member of the M.I.T. Research Laboratory
of Electronics. He has worked primarily in the fields of information
theory, electromagnetic theory, and network theory, and has pub-
lished a number of books in these fields, most recently, Transmis-
sion of Information.

Jay W. Forrester was born in Anselmo, Nebraska, in 1918. He took
his B.S. in electrical engineering at the University of Nebraska,

xiv

Biographical Notes

then came to M.I.T., where he worked on the staff of the Servo-
mechanisms Laboratory and received an S.M. degree. Professor
Forrester was responsible for construction of Whirlwind I, M.I.T.'s
first high-speed digital computer. Later, as director of the Digital
Computer division at the M.I.T. Lincoln Laboratory, he guided
design of the SAGE System of continental air defense for the Air
Force. Professor Forrester is author of the recent book, Industrial
Dynamics.

Martin Greenberger was born in Elizabeth, New Jersey, in 1931.
He received A.B., A.M., and Ph.D. degrees in applied mathematics
from Harvard. Before joining the M.I.T. faculty in 1958, Professor
Greenberger was manager of Applied Science Cambridge, the IBM
group that cooperated with M.I.T. in the establishment and opera-
tion of the M.I.T. Computation Center. Professor Greenberger
is coauthor of the recent book, Microanalysis of Socioeconomic
Systems — A Simulation Study.

Charles C. Holt was born in Jennings, Louisiana, in 1921. He re-
ceived S.B. and S.M. degrees from M.I.T., and M.A. and Ph.D.
degrees from the University of Chicago. Before joining the faculty
of the University of Wisconsin in 1961, Professor Holt was test
engineer for the General Electric Company, staff engineer at
M.I.T., and associate professor of economics at the Carnegie In-
stitute of Technology. He is coauthor of the recent book, Planning
Production, Inventory , and Work Force.

Grace M. Hopper was born in New York City, received an A.B.
degree from Vassar College, and while teaching mathematics at
Vassar, took her A.M. and Ph.D. degrees at Yale University. Dur-
ing World War II, Dr. Hopper served as mathematics officer with
the Navy's Bureau of Ordnance in its computation project at
Harvard. After the war, she joined the Eckert-Mauchly Computer
Corporation as senior mathematician, played a leading role in the
programming of the Binac and Univac computing systems, and
later became director of automatic programming for the Reming-
ton Rand Division of Sperry Rand Corporation.

Ronald A. Howard was born in New York City in 1934. He re-
ceived B.S. degrees in economics and electrical engineering, and

XV

Biographical Notes

S.M., E.E., and Sc.D. degrees in electrical engineering, all from
M.I.T. Professor Howard's interests are in operations research and
systems engineering. In addition to his position at M.I.T., he serves
as consultant to the Operations Research Group of Arthur D. Lit-
tle, Inc. Professor Howard is author of the book, Dynamic Pro-
gramming and Markov Processes.

Howard W. Johnson was born in Chicago, Illinois, in 1922. He
received an A.B. degree from Central College and an A.M. degree
in economics from the University of Chicago. He also did graduate
work at the University of Glasgow, Scotland. He has been a mem-
ber of the faculty of the University of Chicago, where he was
director of management projects in the Industrial Relations Cen-
ter. Dean Johnson came to the M.I.T. School of Industrial Man-
agement in 1955 as director of the Sloan Fellowship Program.

John G. Kemeny was born in Budapest, Hungary, in 1926. He
served as assistant in the theoretical division of the Los Alamos
Project at the end of World War II, then served as an assistant
in teaching and research at Princeton University, where he received
his A.B. and Ph.D. degrees. While at Princeton he was a research
assistant to Albert Einstein in the Institute for Advanced Study.
He went to Dartmouth as professor of mathematics in 1953. He is
coauthor of Introduction to Finite Mathematics, Finite Mathe-
matical Structures, and Mathematical Models in the Social Sciences,
and is associate editor of the Journal of Mathematical Analysis and
Applications.

Gilbert W. King was born in Long Eaton, Derbyshire, England,
in 1914. He received S.B. and Ph.D. degrees in chemistry from
M.I.T. He was a National Research Fellow at the California In-
stitute of Technology, Harvard, and Princeton, and then served as
an instructor at Yale and a research associate at M.I.T. During
World War II, he was with the Office of Scientific Research and
Development. His industrial associations have been with Arthur D.
Little, Inc., International Telemeter Corporation, and IBM. At
IBM he has directed programs in automatic language translation
and information retrieval.

J. C. R. Licklider was born in St. Louis, Missouri, in 1915. He
received B.A. and M.A. degrees at Washington University, St.

xvi

Biographical Notes

Louis, and a Ph.D. degree from the University of Rochester. He
has been Research Associate and Fellow of the Psycho-Acoustic
Laboratory at Harvard, and Associate Professor of the Psychology
of Communications at M.I.T. Dr. Licklider was head of the Engi-
neering Psychology Department and the Information Systems
Research Department at Bolt Beranek and Newman before assum-
ing his present position there.

William N. Locke was born in Watertown, Massachusetts, in 1909.
He received an S.B. degree from Bowdoin College, attended the
Middlebury College French Summer School, then studied in Paris
at l'Ecole de Preparation des Professeurs de Francais a l'Etranger
and Tlnstitut de Phonetique. He received A.M. and Ph.D. degrees
from Harvard. Professor Locke came to M.I.T. in 1945 as head of
the Department of Modern Languages and in 1956 became Direc-
tor of Libraries. He is coeditor of the book, Machine Translation
of Languages, and of the journal, Mechanical Translation.

John McCarthy was born in Boston, Massachusetts, in 1927. He
received a B.S. degree from the California Institute of Technology
and a Ph.D. degree from Princeton. He was a member of the
faculty at Princeton, Stanford, and Dartmouth before being ap-
pointed to the M.I.T. faculty in 1958. Professor McCarthy's prin-
cipal interests include artificial intelligence, mathematical theory
of computation, and computer systems design. He is coeditor of
Automata Studies.

Donald G. Marquis was born in Two Harbors, Minnesota, in 1908.
He received an A.B. degree from Stanford and a Ph.D. degree from
Yale. He has been chairman of the departments of psychology at
Yale and at the University of Michigan. During World War II,
he was director of the Office of Psychological Personnel of the
National Research Council, and he served in the Office of Scientific
Research and Development. Professor Marquis is a past president
of the American Psychological Association and was a member of
the National Advisorv Committee on Mental Health. He joined
the M.I.T. faculty in 1959.

John W. Mauchly was born in Cincinnati, Ohio, in 190". He re-
ceived his doctorate in physics at Johns Hopkins University and
served as head of the Physics Department at Ursinus College.
Later, while on the staff of the Moore School of Electrical Engi-

xvii

Biographical Notes

neering at the University of Pennsylvania, he and J. Presper Eckert
invented and built Eniac, the first electronic computer. Dr.
Mauchly was president of the Eckert-Mauchly Computer Corpora-
tion from 1947 to 1950, when it was absorbed by Remington Rand.
During this period he and Mr. Eckert designed and built Univac,
the first commercial electronic computer. Dr. Mauchly headed the
Univac applications research center until 1959, when he resigned
to form his own firm.

George A. Miller was born in Charleston, West Virginia, in 1920.
He received B.A. and M.A. degrees in speech from the University
of Alabama, then came to Harvard for A.M. and Ph.D. degrees in
psychology. Professor Miller taught at Harvard for four years.
Later, while a member of the faculty at M.I.T., he worked on air
defense problems as a group leader at Lincoln Laboratory. He re-
joined the Harvard faculty in 1955. Professor Miller is now codirec-
tor of Harvard's Center for Cognitive Behavior. He is coauthor of
Plans and the Structure of Behavior and author of Language and
Communication.

Marvin L. Minsky was born in New York City in 1927. He re-
ceived an A.B. degree from Harvard, and A.M. and Ph.D. degrees
in mathematics from Princeton. Professor Minsky was a Junior
Fellow at Harvard before going to the M.I.T. Lincoln Laboratory
to work in artificial intelligence. He was appointed to the M.I.T.
faculty in 1957, and is at present codirector of the M.I.T. Artificial
Intelligence Group.

Elting E. Morison was born in Milwaukee, Wisconsin, in 1909,
and received A.B. and A.M. degrees from Harvard University.
Professor Morison has been a consultant to the Research and De-
velopment Board of the Department of Defense and a consultant
to Houghton Mifflin Company. He also has been director of the
Theodore Roosevelt Research Project. He is editor of The Letters
of Theodore Roosevelt and The American Style, as well as author
of Admiral Sims and the Modern American Navy and a biography
of Henry L. Stimson, Turmoil and Tradition. Professor Morison
joined the M.I.T. faculty in 1946.

Philip M. Morse was born in Shreveport, Louisiana, in 1903. He
received an S.B. degree at Case Institute of Technology and M.A.

xviii

Biographical Notes

and Ph.D. degrees from Princeton. He also studied at the Uni-
versity of Munich and at Cambridge University before joining the
M.I.T. faculty in 1931. During World War II, Professor Morse
directed the Navy's Operations Evaluation Group. He also has
served as scientific director of Associated Universities, Inc., and
was the first director of the Atomic Energy Commission's Brook-
haven National Laboratory. He is a trustee of the Rand Corpora-
tion. Professor Morse is coauthor of Quantum Mechanics, Methods
of Operations Research, and Methods of Theoretical Physics; he
is author of Vibration and Sound, Queues, Inventories, and Main-
tenance, and Thermal Physics; and he is editor of Annals of Physics.

Allen Newell was born in San Francisco, California, in 1927. He
received a B.S. degree in physics from Stanford University and a
Ph.D. degree in industrial administration from the Carnegie In-
stitute of Technology. Before joining the faculty at Carnegie, Pro-
fessor Newell was involved in large-scale organizational simulations
at the Rand Corporation. He currently is working on the simula-
tion of human thought processes and on the logical development
of problem-solving programs.

Alan J. Perlis was born in Pittsburgh, Pennsylvania, in 1922. He
received a B.S. degree from the Carnegie Institute of Technology
and S.M. and Ph.D. degrees in mathematics from M.I.T. Before
joining the faculty at Carnegie, Dr. Perlis was research mathe-
matician with Project Whirlwind at M.I.T., mathematical advisor
at the Multi-Machine Computing Laboratory, Aberdeen Proving
Ground, Maryland, and assistant professor of mathematics at Pur-
due University. Professor Perlis is chairman of the American Com-
mittee on Algol and editor in chief of the Communications of
the Association for Computing Machinery.

John R. Pierce was born in Des Moines, Iowa, in 1910. He re-
ceived his B.S., M.S., and Ph.D. degrees from the California In-
stitute of Technology, and has been with the Bell Telephone
Laboratories, Inc., since 1936. Dr. Pierce served as director of elec-
tronics research at Bell before assuming his present position in
1955. He has written numerous popular science articles and short
stories. Dr. Pierce is coauthor of Man's World of Sound, and he is
author of Electrons, Waves, and Messages and Symbols, Signals,
and Noise.

xix

Biographical Notes

Emanuel R. Piore was born in Wilno, Russia, in 1908. He received
B.A. and Ph.D. degrees from the University of Wisconsin, worked
as research physicist with RCA. then went to CBS as engineer in
charge of the television laboratory. During World War II, Dr.
Piore served as senior physicist with the Navy's Bureau of Ships,
later became head of the electronics branch of the Office of Naval
Research, director of the OXR physical sciences division, and chief
scientist of OXR. Before joining IBM, Dr. Piore was a director
and vice president for research at Avco Corporation. He was ap-
pointed to the President's Science Advisory Committee in 1959.

Walter A. Rosenblith was born in Vienna, Austria, in 1913. He
studied in Berlin, Lausanne, Paris, and Bordeaux. Before joining
the M.I.T. faculty in 1951, he held teaching or research posts at
the University of California at Los Angeles, the South Dakota
School of Mines and Technology, and the Psycho-Acoustic Lab-
oratory at Harvard. Professor Rosenblith is a research associate in
otology at the Harvard Medical School and the Massachusetts Eye
and Ear Infirmary. He is Acting Chairman of the Center for Com-
munication Sciences at M.I.T.

David Sayre was born in New York City in 1924. He received an
S.B. degree from Yale, then became a staff member of the M.I.T.
Radiation Laboratory, where he was an editor of Volume 19 of the
Radiation Laboratory Technical Series. Dr. Sayre received his doc-
toral degree from Oxford University, and was associated with the
Johnson Foundation for Biophysics prior to joining IBM. His in-
terests include automatic coding and artificial intelligence.

Claude E. Shannon was born in Gaylord, Michigan, in 1916. He
received a B.S. degree in electrical engineering and mathematics
from the University of Michigan and an S.M. degree in electrical
engineering and a Ph.D. degree in mathematics from M.I.T. Pro-
fessor Shannon attended the Institute for Advanced Study at
Princeton and in 1941 joined the staff of the Bell Telephone
Laboratories. Inc., where he remained until joining the M.I.T.
faculty in 195". Professor Shannon is coauthor of The Mathe-
matical Theory of Communication and is coeditor of Automata
Studies. He is best known for his pioneering work in information
theory.

XX

Biographical Notes

Herbert A. Simon was born in Milwaukee, Wisconsin, in 1916,
and received A.B. and Ph.D. degrees from the University of Chi-
cago. Before joining the faculty of Carnegie, he directed a program
of administrative measurement studies in the Bureau of Public Ad-
ministration of the University of California and was chairman of
the Department of Political and Social Science at the Illinois
Institute of Technology. Professor Simon is coauthor of Planning
Production, Inventory, and Work Force, Models of Man, and
Organizations, and is author of Administrative Behavior and The
New Science of Management Decision.

Sir Charles Percy Snow was born in Leicester, England, in 1905.
He received an M.S. degree in physics at Leicester University Col-
lege, and a doctorate at Cambridge University, where he became
a Fellow and later a tutor. During World War II, Sir Charles
served as chief of scientific personnel for the British Ministry of
Labor, and in 1945 was appointed Civil Service Commissioner.
He also has served as physicist-director of the English Electric
Company. Sir Charles is author of The Two Cultures and the
Scientific Revolution, Science and Government, The Search, and
a ten-novel sequence, Strangers and Brothers.

Robert C. Sprague was born in New York City in 1900. He is a
graduate of the United States Naval Academy and received an S.M.
degree from M.I.T. While assigned to the Office of Naval Con-
struction at the Quincy Fore River shipyards, Mr. Sprague designed
the landing gear of the nation's first aircraft carrier, the USS Lex-
ington. He founded the Sprague Electric Company in 1926. Mr.
Sprague has been president of the Electronic Industries Associa-
tion and chairman of the Federal Reserve Bank of Boston. He also
has been director of the Gaither Report Committee and consultant
to the National Security Council.

Norbert Wiener was born in Columbia, Missouri, in 1894. He
entered Tufts College at the age of 11 and received a B.A. degree
in mathematics in 1909. After receiving an A.M. degree in 1912
and a Ph.D. degree in 1913 from Harvard, he was granted a
traveling fellowship that enabled him to study with Bertrand
Russell and G. H. Hardy. Professor Wiener's early teaching career
included posts at Harvard and the University of Maine. He has

xxi

Biographical Notes

been at M.I.T. since 1919. Professor Wiener is the author of The
Fourier Integral, Cybernetics, Extrapolation, Interpolation, and
Smoothing of Stationary Time Series, The Human Use of Human
Beings, Nonlinear Problems in Random Theory, two autobiogra-
phies, and a novel.

Other participants in the general discussions included

Paul W. Abrahams, Research Assistant, M.I.T. Computation
Center

Dean N. Arden, Associate Professor of Electrical Engineering,
M.I.T.

Yehoshua Bar-Hillel, Professor of Logic and Philosophy of
Science, Hebrew University, Jerusalem

D wight M. B. Baumann, Assistant Professor of Mechanical En-
gineering, M.I.T.

Philip E. Beach, Jr., M.I.T. Fellow in Africa

Norman Beecher, Project Manager, National Research Corpora-
tion

Edmund C. Berkeley, President, Berkeley Enterprises, Inc.

Gordon P. Brunow, Manager of Systems Services in Boston, Min-
neapolis-Honeywell Regulator Company

Lawrence F. Buckland, Staff Engineer, Itek Laboratories

Lewis C. Clapp, Operations Research Group, Arthur D. Little,
Inc.

Hubert L. Dreyfus, Instructor in Humanities, M.I.T.
Stuart E. Dreyfus, Assistant Professor of Industrial Management,
M.I.T.

James C. Emery, Assistant Professor of Industrial Management,
M.I.T.

Edward Fredkin, President, Information International, Inc.
Robert P. Futrelle, Teaching Assistant in Physics, M.I.T.
Charles C. Garman, student, School of Industrial Management,
M.I.T.

Stephen F. Grisoff, student, Department of Mathematics,
M.I.T.

Elias P. Gyftopoulos, Associate Professor of Nuclear Engineer-
ing, M.I.T.

xxii

Participants in General Discussions

Stanley Herzog, graduate student, Department of Chemical
Engineering, M.I.T.

Milo P. Hnilicka, Assistant to Director of Research, National
Research Corporation

John H. Hughes, Assistant Vice President, Systems and Program-
ming, American Mutual Liability Insurance Company

Thomas Kailath, Instructor in Electrical Engineering, M.I.T.

Meyer M. Kessler, Staff Member, M.I.T. Lincoln Laboratory

Jack Lowen, Applied Physics Group, M.I.T. Lincoln Laboratory

Eugene F. Lucas, Systems Engineering and Management Opera-
tion, Sylvania Electric Products, Inc.

James L. McKenney, Assistant Professor of Business Administra-
tion, Harvard University

James Craig McLanahan, Manager, Terra Company, Baltimore,
Maryland

Franco Modigliani, Professor of Economics, Northwestern Uni-
versity

Calvin N. Mooers, Proprietor, Zator Company, Cambridge
Anthony G. Oettinger, Associate Professor of Linguistics and

Applied Mathematics, Harvard University
Gustave J. Rath, Principal Scientist, Raytheon Company
John C. Reynolds, Applied Mathematics Division, Argonne Na-
tional Laboratory
Edmund F. Shedd, Data Systems Operations, Sylvania Electric
Products, Inc.

James R. Slagle, Staff Member, M.I.T. Lincoln Laboratory
Herbert H. Stevens, Jr., Statistical Engineer, Texas Instruments,
Inc.

Peter Wegner, Research Associate, University of London

xxiii

Contents

Scientists and Decision Making 2

C. P. Snow, Speaker 3

E. E. M orison, Discussant 13

N. Wiener, Discussant 21
H. W. Johnson, Moderator

Managerial Decision Making 36

J. W. Forrester, Speaker 37

C. C. Holt, Discussant 68

R. A. Howard, Discussant 74
R. C. Sprague, Moderator

Simulation of Human Thinking 94

H. A. Simon, Speaker 95
A. Newell, Coauthor

M. L. Minsky, Discussant 114

G. A. Miller, Discussant 118
S. S. Alexander, Moderator

A Library for 2000 A.D. 134

J. G, Kemeny, Speaker 135

R. M. Fano, Discussant 162

G. W. King, Discussant 166
W. N. Locke, Moderator

xxv

Contents

5

6

7

8

The Computer in the University 180

A. J. Perlis, Speaker 181

P. Elias, Discussant 199

J. C. R. Licklider, Discussant 203
D. G. Marquis, Moderator

Time-Sharing Computer Systems 220

J. McCarthy, Speaker 221

J. W. Mauchly, Discussant 236

G. M. Amdahl, Discussant 238
E. R. Piore, Moderator

A New Concept in Programming 250

G. W. Brown, Speaker 251

G. M. Hopper, Discussant 271

D. Sayre, Discussant 273
P. M. Morse, Moderator

What Computers Should Be Doing 290

J. R. Pierce, Speaker 291

C. E. Shannon, Discussant 307

W. A. Rosenblith, Discussant 311
V. Bush, Moderator

Selected Bibliography 327

Index 333

1

Scientists
and

1

Scientists
and

Decision Making

Speaker Sir Charles Percy Snow

Author

London, England

Discussants Elting E. Morison

Professor of Industrial History
Massachusetts Institute of Technology
Norbert Wiener
Institute Professor, Emeritus
Massachusetts Institute of Technology

Moderator Howard W. Johnson

Dean and Professor of Industrial Management
Massachusetts Institute of Technology

Snow. It is a great pleasure to be at M.I.T. tonight. Al-
though you will not believe it, I am always saying kind things
about M.I.T. behind its back. Indeed, I usually try to impress
on my countrymen the absolute necessity of having an insti-
tution as reasonably nearly like it as we can manage. I have
had so far absolutely no success in these efforts. But it is a
real pleasure to be here. I love Cambridge. I love both parts
of this great complex, and I owe very much to them.

It is also a special pleasure to be here with the discussants
on the panel. From Professor Morison in his admirable book
on Henry Stimson, I think I learned more about a certain kind
of American politics than I have from any other single work.
It is an enormous privilege and pleasure to be here with him.

As for Professor Wiener, he has been a great mythological

3

C. P. Snow

figure to me for nearly twenty-five years. It began because we
happened to share a common friend, the great pure mathe-
matician G. H. Hardy. Hardy used to tell me what a remark-
able man Norbert Wiener was. It was very characteristic of
Hard\- that he should express it in the way I am about to
describe. I think probably Professor Wiener will deny this story
totally, but you must take it from me that it is true.

Hardy was a great mathematician. He was also extremely
interested in the game of cricket, at which, in fact, he was
quite a good performer. It was one of the inflictions on his
foreign guests that they had to pay some attention to this
peculiar pastime. Now cricket is a game roughly as complex
as baseball, but complex in a rather different way. It has cer-
tain additional technical complexities because the ball is ex-
pected to hit the ground before it hits the bat, or at least that
is the aim. The result is, as any of you who are familiar with
baseball will at once perceive, that the ball can not only curve
(what we call "move") in the air, it can also deviate off the
ground; and most of the complexities of cricket in a technical
sense derive from that simple fact.

Of all balls that a batsman or batter in cricket can receive,
there are two which are very easy to play. One is the ball that
does not hit the ground but just hits the bat directly; that is
called the full toss. The other is the ball that hits the bat and
the ground simultaneously, that is, pitches perhaps a foot in
front of your left foot as you stand at the crease; that is called
the half volley.

Hardy's story, which Professor Wiener is about to deny in-
dignantly, is that Hardy took Professor Wiener to a cricket
match and explained these technical aspects, to which Wiener
with his astonishing insight said, "Yes, but between these two
easy balls there must be one just a little bit in front of where
the half volley was pitched and yet not quite so far on the bat
as the full toss." Now, this is absolutely true. This happens to
be a kind of ball called a Yorker. Why it is called a Yorker, no
one knows. When a hard-bitten old professional once was

4

Scientists and Decision Making

asked, he said, "Well, I don't know what else you could call
it."

Norbert Wiener, in his one and only acquaintance with
cricket invented the theory of the Yorker. Hardy was fairly
arrogant about mathematical accomplishment and would say,
"Almost anyone of reasonable intelligence could do some sort
of creative work in mathematics, but only a man of really dis-
tinguished intelligence, of something approaching genius, could
have made this discovery about cricket at his first acquaintance
with the subject."

The Luddites

I am standing before you tonight in a somewhat unusual
position, or as I think it is now fashionable to call it, posture.
In a little essay that I wrote, which I called "The Two Cul-
tures and The Scientific Revolution," I had some harsh things
to say about intellectual Luddites. I accused a lot of my literary
intellectual friends of being, in effect, the intellectual Luddites
of our day.

Ned Lud, just to remind you, was a poor unfortunate stock-
ing weaver who lived during the early part of the nineteenth
century in Nottingham in the Midlands of England, near where
I was born. These stocking weavers did their work at home in
their miserable cottages, which were mainly wattle and daub.
These unfortunate people felt themselves menaced by the first
mechanized looms. With surprising practicality, not usually
possessed by intellectual Luddites, they went and destroyed
the looms. That is what a Luddite was.

I have much more respect for the Luddites of that day than
I have for the intellectual Luddites of our day, who simply
deplore the introduction of the machine, while living off its
profits; who regret the mechanical condition into which they
were born, but take the fullest advantage of it. This attitude
is not one which I can easily share, nor one for which I have
much intellectual respect.

Nevertheless, tonight, for once in my life, I feel just a bit

5

C. P. Snow

of an intellectual Luddite. For I am going to raise two ques-
tions of doubt as to the effect of the computer on decision
making, particularly decision making in relation to scientific
problems. I am not putting it more forcefully than this because
I do not want you to think that I am fundamentally pessimistic
on the topic. But I do feel obliged to raise two questions. There
may be nothing in them, or they may be genuine warnings.
Professor Wiener has forgotten more about computers than I
ever knew, and he may easily be able to dispel some of these
doubts or warnings.

Brownian Movement

I am going to begin quite a long way back. How are deci-
sions made? This may sound like a trivial question, but it is
the kind of question which society has never seriously exam-
ined. The concepts which people have of how decisions are
made in our kind of society bear surprisingly little reference
to the truth. You are the most legalistic people on earth by
miles, and so you are always inventing legalistic diagrams of
responsibility and are frightfully upset if they do not seem to
correspond to reality. But usually they do not. We all know
that even in nonsecret decisions there is a great deal of inti-
mate closed politics, a subject to which I have given probably
too much of my literary life. In your country, you elect a
President; he initiates legislation (that is, he takes a decision
as to which legislation to produce), and then the Congress
takes the decision as to whether this legislation is to go into
action.

In my country the procedure is more vague, but very similar:
the country elects a Government and that Government collec-
tively lays its decisions or its suggested decisions before the
House of Commons. Those are the legalistic diagrams; and
yet, in depth, I suspect that many of the most important deci-
sions, although they may finish up with this format as a kind
of rationalization after the event, do not take place in this way
at all.

I believe that the healthiest decisions of society occur by

6

Scientists and Decision Making

something much more like a Brownian movement. All kinds
of people all over the place suddenly get smitten with the
same sort of desire, with the same sort of interest, at the
same time. This forms concentrations of pressure and of di-
rection. These concentrations of pressure gradually filter their
way through to the people whose nominal responsibility it is to
put the legislation into a written form. I am pretty sure that
this Brownian movement is probably the most important way
in which the ordinary social imperatives of society get initiated.

Let me take two examples, one from my country, which has
happened, and one from yours, which I think will happen.
The one from my country which happened was the introduction
of a National Health Service. All over England there had been
people thinking, unconstrained by their party politics, that this
was the kind of thing which our society must really set about,
that health was too important a matter to be left entirely to the
kind of money which you could reach for. This thinking was
surprisingly general. One heard it on all hands. It was not a
Labor matter, it was not a Conservative matter. It was weighted
more on the Left than on the Right, but not very much more.
Any government both could and would have had to find ex-
pression for this concentration of pressure. It was a perfect
example, I think, of the way in which this anonymous Brown-
ian movement finally found its way through into the written
legalistic act.

As I go about the United States, I believe that a similar
process is taking place here and now. I believe that all over
this country there are people saying that although your college
education is wonderful and can do everything which you ask
it to do, your elementary and high-school education in certain
respects leaves a great deal to be desired. I think that one feels
and hears this kind of dissatisfaction everywhere one goes. I,
myself, believe that the broad social generosity of your edu-
cational program is a fine thing. I believe that in strategy, in
intention, your education is incomparably better than ours; but
I have been in high schools where I felt it would be a good
idea if the children sometimes were taught something. I am

7

C. P. Snow

just saying that as a friendly outsider who knows you pretty
well, but I have heard it in much more virulent form from my
American friends. Although the ramifications and the inter-
connections of the federal system are the devil's own business
to cut through, I am quite sure that within the not very distant
future we shall see the widespread demand for a revised ele-
mentary and high-school curriculum enforced upon the people
who in theory make the decisions.

Decisions in Secret by the Few

This Brownian movement is one of the ways in which any
type of healthy society, not only a parliamentary democracy,
makes some of its decisions. But the particular social situation
in which we now find ourselves drives us into a very different
kind of decision. We happen to be living at the time of a
major scientific revolution, probably more important in its con-
sequences than the first Industrial Revolution, a revolution
which we shall see in full force in the very near future. We
also have seen the two biggest political revolutions of all time
and the two biggest wars of all time, and all this has happened
within approximately forty years. This has meant a tempest of
history which is not yet over and a rate of change which is
going to accelerate more and more. We are not going to have
a stable intellectual or social world within our lifetimes. That
is the situation. As a result, many of the cardinal decisions of
society are being taken in a way of which we are all aware, but
which none of us like, a way which alters the entire structure
of democracy and brings with it numerous consequences. This
is what I was trying to say in the Godkin lectures in a neigh-
boring institution.1

Unfortunately, partly through my own fault, the message
that I was trying to convey got somewhat overshadowed by a
personal anecdote.* I had to tell a personal anecdote because
it demonstrated one of the essential features of this new type
of secret decision which is taken by very few people — that

* The anecdote is the Tizard-Lindemann story recounted in Refer-
ence 1.

8

Scientists and Decision Making

personalities matter very much more than they do in open,
nonsecret politics. It is a curious irony, but I think there is no
doubt of the fact. However, I showed poor literary tact in
letting the personal story from which I was drawing examples
obscure what I was trying to say. Now I can try to extract
slightly more sharply the lesson I was wanting to be drawn.
And this is that decisions which are going to affect a great
deal of our lives, indeed whether we live at all, will have to be
taken or actually are being taken by extremely small numbers
of people, who are normally scientists. The execution of these
decisions has to be entrusted to people who do not quite under-
stand what the depth of the argument is. That is one of the
consequences of the lapse or gulf in communication between
scientists and nonscientists.

There it is. A handful of people, having no relation to the
will of society, having no communication with the rest of so-
ciety, will be taking decisions in secret which are going to affect
our lives in the deepest sense.

I chose simply for convenience, because it was far enough
in the past, two particular decisions in my own country. One
was the matter of how we should defend ourselves before the
Hitler War. The decision was taken to give highest priority
to the radar chain. No more than fifty people, and at the point
of action no more than five or six, were involved in that deci-
sion. My second example was the decision in 1942 to make
strategic bombing, the bombing of the civilian population, a
major part of the British war effort.

The second is a slightly different example from the first,
because to some extent the decision to make bombing a major
activity had some of the features of the Brownian movement
which I described. That is, before the war and during its earliest
years, a lot of English people felt with remarkable lack of
strategic judgment that this was the one thing which we could
do and ought to do. It was a great mistake, and it probably
affected the judgment of those who finally had to take the deci-
sion. The final decision, of course, was taken again by very
small numbers of persons, and you know the consequences.

9

C. P. Snow

We certainly disturbed the process of the war, and in my view,
we certainly lengthened it.

These are characteristic decisions which in society now are
being taken in secret by scientific persons and scientifically
minded persons. As I look at it, it seems to me that there is
very little to recommend this method. It may be partly in-
evitable, but some of it is not at all inevitable. The procedure
would have nothing to recommend it were we not in the state
of exaggerated world tension that we actually are. It has noth-
ing to recommend it but speed.

As a general rule, very small groups of people are less wise
than larger groups of people. Very occasionally, small groups
can produce pieces of imagination and initiative which large
ones would not. But, by and large, that is not true; by and
large, maximum errors of judgment occur when concentrations
of power result in decisions being taken by the smallest pos-
sible number of people. In fact, when I reflect on my experi-
ence as an observer of these things, the criticism I feel inclined
to accept of my own attitude is that on the whole I accept
closed and secret decisions too easily; that I have got used to
them; that I know the way they work and in some ways find
them even cozy; that I like them too much and do not realize
how inhuman and contemptuous they tend to make the deci-
sion makers. I believe that is a criticism which I ought to take
to heart.

A New Elite

Let us take this as the existing state of our world. And now,
the computer comes on the scene. Previously, it seems to me,
we have had two groups of persons in secret government: the
circle of scientists who are knowledgeable about what is hap-
pening and which decisions must be made, and the larger circle
of administrators and politicians to whom the scientists' find-
ings have to be translated. My worry is that the introduction
of the computer is going to lead to a smaller circle still. I am
asking a question; I am not making a definite prediction. In-
stead of having the small group of scientists, knowledgeable

JO

Scientists and Decision Making

enough to have something to add to the decisions, I am asking
whether we are now running into a position where only those
who are concerned with the computer, who are formulating
its decision rules, are going to be knowledgeable about the de-
cision. If so, instead of having a small circle of scientists and a
large circle of administrators, we shall have a tiny circle of
computer boys, a larger circle of scientists who are not familiar
with the decision rules and are not versed in the new computer
art, and then, again, the large circle of politicians and ad-
ministrators.

It seems to me that it is going to require a tremendous effort
to try to make all scientists in government literate in com-
puters, and it is going to be quite impossible to make anyone
else in government literate about computers at all. I know that
you can do all kinds of things with these beasts. You can build
in rules so that they defer to human judgment. Nevertheless,
I suspect that the chap standing next to the machine, who
really knows how it makes decisions, and who has the machine
under his command, is going to be in an excessively influential
position. If there is anything in this, we are running into an
added danger.

What I would like to suggest is a piece of pure experimenta-
tion. The facts about the strategic-bombing controversies of
1942 happen to be exceptionally well documented. All the
figures available in 1942 are there on the table. I should like
to see two teams, one pro strategic bombing by inclination and
one anti strategic bombing, both armed with their computer
program or programs. It would be interesting then to see how
the argument would have proceeded. Is this not a case where,
with a little imagination, we could learn something from his-
tory?

Most decision making, of course, consists of finding argu-
ments to justify what you know you are already going to do.
I would guess that in these circumstances probably the pro
strategic bombing characters would have come out with ap-
proximately the same answer that they did in 1942, and the
anti strategic bombing experts would have come out with ex-

11

C. P. Snow

actly the same answer as they did in 1942. But I believe it
would be a valuable piece of research to try to see what differ-
ence, in reasonably accessible historical situations, the intro-
duction of the computer really would make. It may make much
more difference than I think.

The Gadget Charm

My second apprehensive question concerns an intellectual
danger with great practical consequences. Cybernetics, as
named by its founder, is a beautiful subject with great intellec-
tual variety, depth, and complexity. I suspect, however, that
the computer in certain hands could easily become a gadget.
Gadgets are the greatest single source of misjudgment that I
have ever seen, or that anyone has ever seen in scientific deci-
sions in our time. People get fascinated by gadgets. They love
them. They want everything to be explained in terms of their
gadget. They think it is the answer to everything on heaven
and earth. All the bad decisions I have seen have some ele-
ment of gadgetry in them. And I suspect that computers in
government are going to get into the hands of persons with
mildly defective or canalized judgment and become gadgets.
It will be astonishing if that does not happen.

Society Left Outside

Clearly we have not yet reached the stage of real danger.
But dangers are usually much better met if you have anticipated
them. Most things can be coped with if you recognize them
in time. I think that with a certain amount of administrative
and experimental skill we can reconstruct certain historical
situations with the aid of a computer and see how the answers
turn out. It will need some historical imagination, some scien-
tific imagination, and especially some psychological imagina-
tion, but I am sure it is worth doing. Otherwise, the obvious
and glaring danger is that the individual human judgment
is going to take a part which will get smaller and smaller as
the years go by. I am inclined to think that for a society which
is really viable, and certainly for one which feels itself to be

12

Scientists and Decision Making

morally viable, there is no substitute for individual human
judgment; and the wider it is spread, the healthier and more
viable this society is likely to be.

It is not only that I am afraid of misjudgments by persons
armed with computing instruments; it is also that I am afraid
of the rest of society's contracting out, feeling that they have
no part in what is of vital concern to them because it is hap-
pening altogether incomprehensibly and over their heads. I
suspect that the feeling of being left out, being outside the
decision-making party, as it were, is one of the causes of the
malaise of our society. We must not let it go too far. I am
not in the least pessimistic about our finding our way through
these difficulties and dangers. I believe that the computer is
a wonderful subject and a tool from which we can get great
service. But if we let the individual human judgment go by
default, if we give all the power of decision to more and more
esoteric groups, then both the moral and intellectual life will
wither and die.

Panel Discussion

Morison. I shall make clear at the beginning two things
that would, in any case, become obvious as I go along. First, I
do not know very much about the subject of these talks, the
computer. And second, I have spent my life in that culture
which, as Sir Charles suggests, tends to produce nervous appre-
hension and depression of the spirit.

I do know that the computer in its present form is a rela-
tively new machine; so I thought I might say a word or two
as a historian on the way new machines and men have got on
together in the past. I also know that the computer is a ma-
chine that will give answers to certain kinds of questions and
supply solutions to certain kinds of problems. So I thought I
might suggest what some men in my culture think they have
found out about the perplexing dialogue between question and
answer, problem and solution.

13

E. E. Morison

As for the first topic, no more than Sir Charles am I a
Luddite. One of the things you can learn from history is that
men have lived with machinery at least as well as, and probably
a good deal better than, they have yet learned to live with one
another. Whenever a new device has been put into society —
the loom, the internal-combustion engine, the electric generator

— there have been temporary dislocations, confusions, and in-
justices. But over time men have learned to create new arrange-
ments to fit the new conditions. Anything that has the power
to build also has, of course, the power to destroy, and this ap-
plies to machines in the hands of men. But, on the whole, and
more often than not, men have always succeeded in organizing
mechanical systems for constructive purposes and for the en-
largement of human competence and opportunity. No one, I
think, who compares the condition of life for the average per-
son in the seventeenth century with the average condition of
life today in our society can fail to reach this conclusion.

Partly for this reason, I am not as much of a Luddite as Sir
Charles may be. Take his first apprehension — that the computer
may measurably increase the tendency toward closed decisions
in our society. Obviously, we shall have to think about this.
Machines can, beyond doubt, alter some of our views of things

— the multiengine plane, for instance, has changed somewhat
our sense of time and space. But there is, as I understand it,
nothing in the nature of the computer that will necessarily
take us nearer to closed decisions — closed decisions such as
those taken in the days of Wolsey or Richelieu or Caesar long
before there were radar sets or computers. Both the machine
and its programmers will have to work within a general scheme,
a field of general decisions and determinations that can still,
as Sir Charles says, be gathered out of the air if that is the way
we want to do it. In determining the kind of life you want to
have, the instrumentation is less influential than the nature of
the culture you create to control what you want to use the in-
struments for.

For example, I do not believe that the rumble seat of auto-
mobiles increased the incidence, it merely changed the locus,

14

Scientists and Decision Making

of experiment in physical relations between boys and girls in
the age of F. Scott Fitzgerald.

Then there is the apprehension about the computer as a
fascinating gadget. It is obvious that there is always danger
from the gadget-happy — whether the gadget is a machine, an
idea, or a procedure. Amasa Stone, a very able man, killed a
trainload of people because, against advice, he built a bridge
at Ashtabula from a truss design for which he had an ancient
attachment. I have no tonsils, nor have my brothers, nor have
most of my generation, because an accomplished ear, eye, nose
and throat man and his colleagues were all obsessed with the
thought that the way to make a boy grow was to take out his
tonsils.

In an age of new departures we have to live with all this, I
suppose, but history suggests only for a limited period in each
case. Over time the potentials of a new gadget are explored by
trial and error until the real capacities are discovered and under-
stood. Then, whatever it may be — Manichaean heresy, steam
turbine, penicillin — it is fitted into a reasonable context.

I do not want to appear like a wise old head, made sager by
my study of history than those like Sir Charles who have ac-
tually been there. Of course what he has said should cause any
sensible man to think, and what I have said does not, I know,
fully dispose of the complex problems he raises. Perhaps we
can all talk about these things together afterwards. But now, I
do not have much time, and I want to get on to a nervous
apprehension of my own.

I think we may have more difficulty in exploring the full
limits of the computer than we have had with earlier gadgets.
I think there may be more danger in the period of trial and
error than there has been with earlier devices. These earlier de-
vices — looms, engines, generators — resisted at critical points hu-
man ignorance and stupidity. Overloaded, abused, they stopped
work, stalled, broke down, blew up; and there was the end of
it. Thus they set clear limits to man's ineptitudes. For the
computer, I believe, the limits are not so obvious. Used in
ignorance or stupidity, asked a foolish question, it does not

15

E. E. Morison

collapse, it goes on to answer a fool according to his folly. And
the questioner, being a fool, will go on to act on the reply.

This at least is what my culture tells me often happens. Let
me give you an example. In the play with which you are all
familiar, Hamlet had a problem which he defined for himself
as follows: What had happened to the late King of Denmark
and what should he, Hamlet, do about it? Framing the ques-
tion accurately — a good program — he took it to a ghost —
the most sophisticated mechanism in the late sixteenth century
for giving answers to hard questions. From the ghost he got
back a very detailed reply, which included a recommendation
for a specific course of action. Responding to these advices,
Hamlet created a political, social, moral, and administrative
mess that was simply hair-raising.

The trouble was that he had got the right answer — the an-
swer he deserved — to a question that was totally wrong. He
had asked about his father when he should have asked, as any
psychologist will tell you, about himself and his relations with
his mother.

My culture says, in other words, that it is much harder to
ask the right question than it is to find the right answer to the
wrong question.

Some of you, like some of my students, may say that Hamlet
is only a play, so what does it prove? I therefore shall give you
some further evidence about questions and answers taken from
real history. We asked ourselves once, in the first days of the
Republic, what a negro was really worth and came up in the
Constitution with the answer that he was worth three-fifths of
a white man. Somewhat later we asked ourselves how to in-
crease the income of the average citizen and decided the an-
swer was to coin silver at a rate of 16 to 1. And still later we
asked how we could limit the arms race between Britain and
us, and worked out the answer that for every British light
cruiser we could have 1.4 American heavy cruisers. About the
same time we asked ourselves how to make the nation "self-
sustaining," and arrived at the answer of the Smoot-Hawley

16

Scientists and Decision Making

tariff which set an average ad valorem rate of 40.1 per cent
for all schedules.

You will know, I am sure, that all of these answers caused us
very real trouble of one kind or another. They did so because
the questions they were designed to answer were framed in a
wrong interpretation of events, a false conception of the actual
problem. The answers supplied therefore gave the wrong solu-
tions. They represented collectively what Ramsay MacDonald
said of one of them: "an attempt to clothe unreality in the
garb of mathematical reality/'

The quotation from the Prime Minister suggests a further
source of nervous apprehension — the tendency to simplify
human situations and to do so, often enough, by reducing them
to quantifiable elements. I have spoken of Hamlet so, by way
of illustration, I will speak of him again. I remember two things
my engineer-type students have said in explanation of his be-
havior. First, he had too much feedback in his circuits, and
second, he was 16% per cent efficient, because he had one per-
son to kill and he killed six. This, purely incidentally, is about
the thermal efficiency of the average internal-combustion en-
gine.

What I want to suggest here is the persistent human temp-
tation to make life more explicable by making it more cal-
culable; to put experience into some logical scheme that, by its
order and niceness, will make what happens seem more under-
standable, analysis more bearable, decision simpler. When you
talk of 140 per cent of a cruiser, you can hope you have solved
the underlying diplomatic issue you haven't dared to raise;
when you pass a tariff with average rates of 40.1 per cent ad
valorem to make a nation self-sustaining, you can assume that
you do not have to look further for the causes of the worst
depression in the nation's history — to which, incidentally, you
have just contributed by passing the tariff. This is, I suppose,
the way it does figure; and this seems to have been the human
tendency from the time of Plato's quantification of the Guard-
ian's role right on down.

17

E. E. Morison

I am not trying to suggest that the computer will soon bring
us all under the cloak of the mathematical reality of its pro-
grams. But today the tendency to work with quantifiable ele-
ments and logical systems seems to me accelerating. There are
more tests and measurements (the brain of a candidate for
college works within a precisely graded scale from, presumably,
1 to 800), more rational systems like those of Keynes and
Freud to assist us in ordering the economy and the personality,
more mathematical models, and more efforts, as in this School
of Industrial Management, to reduce administrative experience
to quantifiable elements. This, in the name of clarification and
the advancement of general understanding, is quite obviously
all to the good. The aim of pure reason, which proceeds upon
measurable quantities, is, presumably, to introduce increasing
order and system into the randomness of life. But I have here
the apprehension that as time goes by we ma}- begin to lose
somewhat our sense of the significance of the qualitative ele-
ments in a situation — such things as the loyalties, memories,
affections, and feelings men bring to any situation, things
which make situations more messy but. for men. more real.
My apprehension is that the computer, which feeds on quanti-
fiable data, may give too much aid and comfort to those who
think you can learn all the important things in life by breaking
experience down into its measurable parts.

I hope it is evident that I am for order and logic. But I do
hope also that it remains clear that in all the really interesting
questions and problems of life the measurable and the im-
measurable are all mixed up. I think from time to time of the
Pythagoreans, those men who came to believe that "all things
are numbers" and who were supposed to have put to death a
man who suggested the idea of the incommensurable. Even
in an inventory program the risk one is willing to bear if he runs
out of stock must be considered — and this risk is in part de-
termined by the unknown size of the irritation of frustrated
customers. Even in the strategic-bombing exercise suggested
by Sir Charles, there is what might be called the Coventry

18

Scientists and Decision Making

factor to put in the program — the unrequited feelings of men
and women who had been bombed with no redress.

Still, I am no Luddite. What I want to do, first, is to find
out all there is to find out about the computer. And I must say
this is hard enough. In my cursor}- researches I have been told
a great many different things by people who have at least
thought more about it than I: that it is and always will be
simply an idiot doing what it is told to do: that it can now de-
sign fractional horsepower motors and electrical circuits: that
it can already throw old data into new combinations — intro-
duce intellectual surprise: that it may some day write a sonata;
and that, if we can get enough vacuum tubes hooked up right
(1010 is the figure), we cannot exclude the possibility that it
may feel its own emotion and have a will of its own. The spread
between assumed present capacity and the foreboded potential
is, in other words, considerable.

In any case the computer is here, and no doubt it is going
to develop. Even-body, or almost even-body, seems a little un-
easy about this, and why not? This is man's first encounter
outside himself with something that is exactly like some inside
part of himself. It is not, as many other machines have been,
like his arms or hands or legs in the work it does, it is like him.
How much like him we do not yet know. But regardless of
what happens in the future, we have already made a machine
that simulates some part, a small part, of what we alone have
been able to do in the past by thinking. Even this small ad-
vance begins to raise the large question we have succeeded so
often in avoiding — as Hamlet did for instance. What is in
fact our true image, what is our real likeness?

To assist us in examining this question, I propose, not that
we bust up the machine, but rather that we explore it in a
series of experiments suggested to me by Sir Charles's idea of
replaying the bombing problem. I suggest a continuing experi-
ment in which the machine would be asked to reconstrue a
series of situations out of the past — situations taken from my
own culture — in which men have acted most successfully- on

19

E. E. Morison

their own. One could begin with simple things — a successful
plant relocation — and work up through, let us say, the Army
and Navy decision on how to use the plane against the sub-
marine, and then on to the most interesting situations — our
method of limiting trusts and cartels as revealed in the Sher-
man Act and its subsequent modifying judicial interpretations,
or the miracle of Queen Elizabeth Fs foreign policy. I suggest
for each situation a series of experimental programs — leaving
some things out, adding some things, practicing ways to code
things that seem uncodable, and so forth — to test various
hypotheses and understandings by repeated trials against what
may be called the real situation.

Some of the problems may be, at the moment, a bit far out,
given the present state of the art. Also it is probable that the
study of a particular past and nonrecurring situation would not
produce the most desirable results. For instance, it has been
suggested that a program designed from evidence taken from
the record of six or eight different revolutions might well be
useful in the attempt to discover how to control common ele-
ments — both the stabilizing and the disrupting — that operate
in all revolutions, including the one we are in today. Or a pro-
gram put to a computer on the problems of underdeveloped
regions that are now developing under the energy of technology
might well include data collected from the history of other
regions that have passed through this stage of experience.

Whatever difficulties or defects there may be in these par-
ticular proposals, I am told a beginning of some sort could be
made in the direction I suggest, and I hope it will be, for
several reasons.

First, a good many people would have to learn some history
— which is a good thing.

Second, it would be a way to capitalize on the ancient truth
that fools persisting in their folly learn wisdom. We could
acquire understanding of the uses of the machine without
danger in a time of controlled trial and error.

Third, the machine would therefore become not so much
a problem solver as a learning machine, which is today in fash-

20

Scientists and Decision Making

ion. Used as suggested, it would force us, again without danger,
out from behind the silly or distracting questions we like to ask
to the real questions we have to ask — and teach us to ask
them more correctly. It would thus help us to sort out the
things that can be thought from the things that can only be
felt, and advance a little of our understanding of how much
feeling goes into what we call thinking. Perhaps some of the
things we now classify as feeling may turn out to be more
identifiable and explicitly definable, as the work of Freud in-
deed suggests, than we think. But if not, if they cannot be
programmed as they say, at least by this exercise we may well
find out more about their meaning and influence. What I am
suggesting here is, I trust, obvious: that we use the machine
which simulates to explore fully what it is simulating — what
image, what likeness.

For some, this assault by mechanism upon what e. e. cum-
mings calls the single secret that is still man will be distasteful
and to some appalling. It is not — at least not to me. Over on
the other side good friends of mine are using accelerators to
find out the secrets of other marvelous structures. The more
they find, the more they seem to stand with respect, indeed
with wonder, before their findings. If man working experi-
mentally can learn more about himself from something he
invented, can learn who he is, and who he is not, then he will
have learned enough, I should think, to use himself as well as
the machine that simulates more wisely and constructively.

And if in this process he finds, at length, the full answer to
the single secret — well, that is what my crowd has been trying
to do for twenty-five hundred years.

Wiener. I think we have been talking a good deal about
computing machines on the basis of rather a shady account
of what these machines are, and an even shadier account of
what these machine are going to be.

I shall not confine my remarks to computing machines; I
am going to talk about control apparatus in general. The dif-
ference between the two from our point of view is not very

21

N. Wiener

great. The computing machine is a general-purpose device that
can be programmed to do very specific jobs, and it may be at
the heart of the control apparatus for certain applications. But
let us talk in general about control and communication ma-
chines, rather than specifically about computing machines.

As has been intimated by Professor Morison, we are starting
to hear about a class of machines just coming into being — the
learning machine. I want to say something more about the
learning machine, because I think that both the greatest part
of the difficulties which we have been discussing and the great-
est part of the possibilities for relieving these difficulties lie in
this machine.

Take the learning machine as it now exists. One form of it
is a machine for playing checkers. Now, it is possible to play
checkers with machines that are not learning machines. It is
possible to write down at any stage of a game all the succes-
sive moves that are legally possible for the next stage; to rank
them on a scale of values involving loss of pieces, mobility,
control, and many other factors (I believe about fifteen factors
have been considered in one game); and then to give these
factors certain weightings, established at the beginning. This
procedure leads to a machine that is a checker-playing machine
to a limited extent, but not a learning machine. If you were
to play against this machine, it would feel like a rigid personal-
ity. (By the way, if you play correspondence checkers or chess,
whether with a machine or a person, you do acquire a feeling
for the personality opposed to you.) When a rigid personality
makes a blunder, it always repeats the same blunder in the same
situation.

A machine with a less rigid personality can be achieved as
follows. The machine plays as before, but now keeps a record
of all plays made and all games played. At intervals it is run
in a different way. Instead of evaluating moves in terms of a
fixed scale of evaluation, it evaluates the scale of evaluation
in terms of the games played. It determines which scale of
evaluation would have led to wins more assuredly than any
other. There are various tricks of accomplishing this which I

22

Scientists and Decision Making

need not describe here; they are not perfect, but they are valid.
The machine, having determined which scale of evaluation
would have been most conducive to winning, adopts it for
further play. That is learning.

Such a machine, if you use it for playing checkers, as it has
been used, would have a more flexible personality. The tricks
that once worked against it might fail with its increasing ex-
perience and its re-evaluation of various considerations. Such a
machine has been developed by Samuel of IBM.2 At first the
machine was able to defeat Samuel fairly consistently. Later
on he learned a little more checkers and was able to defeat the
machine more often. Nevertheless, the fact is that the machine
can go beyond the person who programmed it. Even though
Samuel caught up, there is always the possibility that the ma-
chine may catch up again later with more sophisticated pro-
gramming.

The fact that a machine can defeat the man who pro-
grammed it means that having made such a machine does not
give him completely effective control over it. If he had that,
he would not let it beat him. Now, this is very important.
Such a machine could be very useful in certain decision situa-
tions. It could be used to play games other than checkers: the
business game, the war game, and the game of determining
when to press the button for Armageddon — for the thermo-
nuclear war.

How are you going to program such a machine? Well, you
cannot program it based on prior thermonuclear wars. You
would have to play the game according to a set of postulates
which you constructed. You could make the machine learn to
be more successful within the framework of these postulates.
However, you receive no indication from this whether your
postulates have the right values. Such a machine, in other
words, can beg the question very badly and can be very dan-
gerous.

What you have here is a situation not unlike that found in
the folk tale, "The Monkeys Paw." "The Monkey's Paw" is a
story told by W. W. Jacobs of England at the beginning of

23

X. Wiener

the century.3 An old soldier returns from India to visit a friend.
He has with him a talisman that he says has the ability to
grant three wishes to each of three people. The first owner of
the talisman had taken the first set of three wishes, two un-
known to the soldier, but the third one for death. That is how
the soldier became owner. The soldier took the second set of
wishes for himself, but declines to talk about them. His ex-
periences were too terrible. One set of wishes remains. With
considerable reluctance the soldier yields to his friend's request
for the talisman. The friend's first wish is for £200, and an
official of the company where his son is employed comes in to
tell him that his son has been crushed in the machinery. As
a solatium, but without any admission of responsibility, the
company has granted the father £200. The next wish is that
the boy be back, and his mutilated ghost appears knocking at
the door; the third wish is that the ghost go away.

The point is that magic is terribly literal-minded. It will give
you what you ask for, not what you should have asked for, nor
necessarily what you want. This will most certainly be true
about learning machines. If you do not put into the program-
ming the important restriction that you do not want £200 at
the cost of having your son ground up in the machinery, you
cannot expect the machine itself to think of this restriction.

"The Monkey's Paw" suggests a very real danger of the learn-
ing machine. The danger of these machines is greater than that
of the simple computing machine, because you do not set
down for it the tactics of the policy but only the strategy. You
let the tactics work themselves out from the experience of the
machine. The machine acquires a nature based on its experi-
ence.

So there are real dangers here. Is there any way of partially
overcoming these dangers? The importance of learning machines
is not how they act as pure machines, but how they interact
with society. We thus are led to the concept of a system in-
volving both human actions and machines. Is there any way
in such a system to transfer values from the human being to
the machine?

24

Scientists and Decision Making

In a general, imperfect sense there is. Suppose you build a
machine to translate a language. The value of the translation
is a human value. The value depends upon whether people
reading the translation will interpret it with the same meaning
as the original. It is conceivable that you might attain a value
system by having a complete logical code of translations, a
superperfect grammar; but this would not be a really profitable
way to proceed. Instead, you would have the machine do exer-
cises, just as a human pupil does exercises. These exercises would
be marked by a teacher who possesses human values. The ma-
chine would modify itself based on these marks, just as the
checker-playing machine modifies itself based on its prior wins.
It is at least theoretically possible to transfer values from the
human being to the machine in such man-machine organiza-
tions. I think that the possibility of reproducing human values
is of great importance, but in itself has dangers too.

The last thing I want to discuss is the temptation of gadgets.
What are the tempting things about gadgets? What are the
tempting things about machines? What are the tempting things
about an organization with great compartmentation and high
secrecy? These are the human questions that arise.

I am reminded here of the game of Russian roulette. At
various times, in remote barracks and posts, officers have played
this game of putting one bullet into their revolvers, spinning
the chamber, and pulling the trigger, either at themselves or
somebody else. This is obviously a means of expressing aggres-
sive impulses either against oneself or against one's opponent.
It is obviously a form of masochism or sadism. It is permissible,
however, because one man has removed final responsibility for
the act from himself, put it in the machine, and left it to
chance. If he kills the other man, he has provided himself
with a way (in his mind) of reducing his responsibility from
what we ordinarily associate with murder to what we associate
with manslaughter. He can blame chance.

This is very much like the game that is often played when
a prisoner is taken before a firing squad. Each man in the firing
squad has his rifle loaded with several blank cartridges and a

25

C. P. Snow

single bullet. The result is that the men are more willing to
fire because of the overwhelming likelihood that they will not
kill the prisoner. This is a way of avoiding responsibility.

I am certain that a great deal of the use of gadgets for deci-
sions, as it exists now and as it may exist even more in the
future, is motivated by this desire to avoid direct responsibility.
I am sure that a lot of the subdivision of effort in secret proj-
ects and highly compartmented projects has the same motive.
The subordinate does not know enough about the project to
feel responsibility, and the man in charge can place responsi-
bility with the system. I believe that one of the greatest dan-
gers at the present time has to do with the attempt to avoid
responsibility in order to avoid the feeling of guilt.

Snow. I find myself in very close sympathy with both dis-
cussants, which is a rather tedious position to be in. I share
Professor Morison's passion for history in a very amateur sense,
but in my view he, like all historians, slightly underestimates
those few occasions in human affairs where there is a genuine
discontinuity. I believe that the amount of energy we can now
trigger is so grossly different from the amount of energy we
could trigger in the very recent past, that a great many deci-
sions take on a quite disproportionate significance. It is a signif-
icance which is unique in history. This discontinuity makes
some of the language and attitudes of history, not exactly in-
appropriate, but often not entirely adequate.

I found Professor Wiener's remarks fascinating, but he pro-
vided more dangers than I ever could have imagined. Though
intellectually stimulated, I do not feel really encouraged. My
state of mild apprehension has not been dispelled.

Morison. I do not in any way discount or deny that there
is really an absolute change today to which certain of the
precedents of the past do not apply. This is one of the things
that worries me as much as anything we have discussed. There
are many precedents which suggest that we can learn from
our small-size errors, and historians tend to rely on these. It

26

Scientists and Decision Making

is the really large error that we do not know how to deal with
very effectively, and I share Sir Charles's concern.

Wiener. There is a real possibility that changes in our en-
vironment have exceeded our capacity to adapt. The real dan-
gers at the present time — the danger of thermonuclear war,
the computing-machine sort of danger, the population-explo-
sion danger, the danger of the improvement of medicine (to
the extent that we shall very soon have to face not letting
people live as part of the policy of letting them live) — all of
these dangers make one wonder whether we have not changed
the environment beyond our capacity to adjust to it, and
whether we may not be biologically on the way out. We may
not be, but this is not at all clear.

Snow. What Professor Wiener has said is very real. Our
tendency is to laugh it off, but it is not something that I would
laugh off too easily.

Wiener. I think that the over-all danger from the total
situation is much greater than the danger from any of its par-
ticular manifestations, such as the atomic bomb or the learning
machine.

Snow. I agree.

General Discussion

Minsky. I should like to address a remark to Professor
Morison, and I shall try to speak as an antihistorian. The ex-
periment that you propose reminds me of the experiment that
I feel historians have been avoiding very skillfully for many
years; namely, to take a precise area of history about which not
everything is generally known, but many facts can be found,
and then have a group of professional historians study the
facts and predict the outcome. It seems to me that members
of your profession, and social "scientists" in general, have me-
ticulously avoided any such experiments.

What brings this to mind is your sanguine attitude about

27

General Discussion

the way mankind will adjust to the computer — to this minor
innovation or gadget. You look back at history as a professional
historian and see a continual sequence of successful adjust-
ments to dreadful or complicated gadgets. I, as an antihistorian,
look back at history and see the most horrible sequence of
disasters, with billions of people (or at least millions, before
the population reached the precarious state at which it could
be billions) being wiped out every few hundred years in vir-
tually every culture. I do not get the impression of mankind's
being capable of surviving even minor dislocations. It happens
that a number of people seem to survive these disasters and
start over, but I cannot see the flow of history as a neat adapta-
tion to innovations.

Morison. I hope it was clear that in making my suggestion
I was not above using the machine to improve the profession.
I should agree with you that one of the things we as a profes-
sion have too much difficulty with is using the data in what you
people would call a useful way, that is, to make generalizations
and predictions. I think we have been timid in this respect.
In fact, one of the reasons why I made the suggestion I did was
to see what my historian friends would say. I thought it might
force us all to think in new and different terms.

As for your second comment, I hope that I said (and I
think I did) that we find it easier to live with machines than
with each other. I think this is true. I think the disasters which
you refer to were frequently not caused so much by machines
directly as by our failure to accommodate to our fellow man.

Wiener. There is one thing about the machine; the ma-
chine habit is like the liquor habit. If we drink enough of it,
and we already have, it is an awfully hard thing to stop. We
have made ourselves dependent upon the machine, on the car,
for instance. We are suffering from the way the train is break-
ing down, and we cannot give up the airplane. We have made
ourselves more vulnerable to machine failure than ever before.
We cannot go back to the old agrarian civilization. The soil
is not there. The habit of raising crops, the habit of living off
your own patch of potatoes, are just not there anymore. In an

28

Scientists and Decision Making

emergency we should have a terrible time adjusting ourselves, if
we ever could, to the way not only that our great-great-great-
grandfathers lived, but even to the way that our grandfathers
lived.

Hnilicka. One of the basic problems of history and of man-
kind is that all our thinking and philosophy are based on a
static situation. We always assume that the way things are
now in their static form is actually the right representation.
Would the panel be so kind as to comment on whether the
machine might help us interpret future stability, not from the
static point of view, but from the dynamic point of view, so
that we cannot just catch up but actually predict what is going
to happen?

Snow. This is an extremely interesting point. It is clearly
true that societies like ours find it extraordinarily hard to un-
derstand change. The problem is very large, and I think that
the machine might help us here.

Wiener. I should agree that the machine might help. It
is true not only that the equilibrium of society must be dy-
namic, but that its possible breakdown must be dynamic, too.
If we have a catastrophe, the final catastrophe, it probably
will not be because our world has broken down statically, but
because, as the engineers would put it, the transients have
reached the explosive point. The idea of a dynamic equilibrium
must be supplemented by the idea of a dynamic disequilibrium
and a dynamic collapse. These also can be studied to some ex-
tent on the machine.

Herzog. I am a little curious why Sir Charles is so worried
about scientists' becoming an in-group in government. Cer-
tainly the militarists and the economists have been in-groups
in government whom most of us could not understand. Does
Sir Charles really think that scientists will do worse than poli-
ticians have done?

Snow. I have been accused of some awful things in my
time, but why in God's name do I get accused of this? Now,
I have said very clearly that I should like to see many more
scientists with direct responsibility in the affairs of government.

29

General Discussion

There are two things that frighten me: first, the assumption
of power by very few people of any kind (scientists happen to
be an exceptionally important group with extreme powers, and
I would hate to see a solitary scientific overlord in any coun-
try); and second, the danger of even a relatively large group
of scientists, with the actual legal decision-making power, com-
municating across a void. Both of these things I have said. But
I have never said that I am especially frightened of scientists;
this would be an absurd statement for me to make.

Beecher. The training of a person in values takes anywhere
from twenty-one to fifty years, depending on where the person
is reared, before he is judged worthy to make a decision. Is
tfiere any reason to suppose that we could reduce the time
with a learning machine? Also, how are we going to get the
input? Are we going to send the learning machine to Sunday
school, and then during the week teach it not to take what
it learned too seriously, as we do with our children?

Wiener. With respect to breadth and depth of learning,
man will be far superior to any machine that we are likely to
see for a long, long time; very likely for all time. One of the
advantages of the machine, however, is its ability to work faster
and do many things in the same time that a man does rela-
tively few. It is conceivable that over a limited scope the learn-
ing of the machine can proceed more rapidly than the learning
of man. This will not be done without effort, but it is at least
possible.

Morison. I should like to ask Professor Wiener to go a
little further and talk technically about how value can be pro-
grammed. How far have we gone on this?

Wiener. We have gone only a small distance, but the pos-
sibility is very clear. Consider a translating machine. You grade
it on its translation and use this grade to modify the translating
program, just as you use the past performance of a checker-
playing machine to change the machine's playing of checkers
— perhaps assigning 1 for winning and 0 for losing. In this way
you teach the human values of what a good translation is to
the machine.

30

Scientists and Decision Making

You can transfer values from one machine to another in a
similar way. This ability is really very important. It makes pos-
sible a self-reproducing machine that will make other machines
in its own image. There is a definite mathematical theory sup-
porting this possibility. We can use one machine to condition
a second to have the same effect that the original does. And
there are still other modes of interaction possible between
machines. We can have criticism, and this criticism can be
used to transfer values. This needs to be worked out in much
greater detail, but it is the only promising way that I see for
making a translating machine that will be any good.

Morison. It would be fun to program the Book of Job on
a machine that has made another machine in its own image
and see how the relationship between God and man works out.

Wiener. In the machine that makes another machine in
its own image you have the Book of Genesis, not the Book of
Job. You have essentially the Book of Job in the checker-
playing machine where a man plays a game with his own crea-
tion.

Morison. I have lost on my own ground.

McCarthy. I am wondering whether the speakers really
take as gloomy a view of matters as they have indicated. In
Professor Wiener's parable we have a magic paw that will grant
wishes, and we supposedly have no idea just how it does this.
But it seems to me that the computer engineer who designs
a magic amulet probably does have some notion of how it is
going to treat our particular wishes. If you want to insert some
values, then you must convince the engineer.

Wiener. The point is that it is extremely difficult to insert
the whole value system at the beginning. It generally will de-
velop from continued interaction. You may ask, "WTiy can't
we make the learning machine 'safe'?" The answer is that "safe"
is a human value, and the machine cannot very easily state
what safe is without reference to man. For example, if you have
a machine for programming atomic war, there is no reason why
the machine should care whether the human race is completely
burnt up or not. "Safe" is a human idea. If you have not put

51

General Discussion

this idea into the machine, it is unreasonable to expect it to
get there by itself.

Another point is that since we do not have full control over
the learning machine, as illustrated by the checker-playing
machine's defeating the man who programmed it, the unsafe
act may not show its danger until it is too late to do anything
about it. It is possible to turn the machine off, but how are we
to know when to turn it off? If there is any possibility of its
going wrong, we should turn it off at the very start. Otherwise,
by the time that anything becomes manifest about the danger,
it may be too late to avoid the consequences that have come
from the use of the machine up to that point. You cannot make
a perfectly safe learning machine.

Clapp. I should like to ask the panel a question. There is
no doubt that the computer is here to stay. You may be correct
in your gloom and your prognostications that we may all blow
up because we have computers, but I wonder if there is not
some middle ground?

Wiener. Obviously there is a middle ground. We have to
make every effort to understand for ourselves what the dangers
are, and this points up a fundamental thing about computers:
They involve more thought and not less thought. They may
save certain parts of our efforts, but they do not eliminate the
need for intelligence.

Snow. Even though we have expressed grave disquiet, not
only about machines, but about the whole set of associated
phenomena which are characteristic of our times, I do not think
any of us feels that these problems are unsolvable. I do not
think any of us wishes to give that idea. But it is going to take
great intelligence, and in my view, great moral judgment.

Grisoff. I wonder if the panel feels that man can ulti-
mately build machines which will exhibit an intelligence ba-
sically different from human intelligence. I notice that there
has been very little distinction made between machine intelli-
gence and human intelligence, and I think that there might be
a very basic difference here.

32

Scientists and Decision Making

Wiener. As long as the machine has beat the man who
programmed it in checkers, it will in some sense compete with
human intelligence over a limited scope. My hunch is that for
quick action over a limited scope, the machine can be made
better than the man. For higher logical-type judgments, for
vague ideas, and for a large class of other things, the ma-
chine is a long way from competing with the brain. I think
that there will always be a shifting boundary between the two,
but I don't dare to venture where it will lie.

Lowen. Would it help us to remember the difference be-
tween intelligence and wisdom? The machine might help our
intelligence, but as of now it has no bearing on our wisdom.

Wiener. This is connected with the question of logical
type. Primitive statements — statements about one or two
things — we shall call intelligence. Statements about statements
will be more nearly wisdom. Statements about statements about
statements will be even more nearly wisdom. I think the whole
issue is closely connected with Russell's theory of types. The
brain is at its best on relatively high types, and the machine is
at its best on relatively low types.

Gyftopoulos. I should like to ask Sir Charles whether he
feels that the big outer- and small inner-circle phenomenon
in decision making is a sign of our times or whether it has al-
ways been in evidence historically. It seems to me that when-
ever I participate in a big group and a difficult question comes
up, there is always a committee of two or three appointed to
make the decisions.

Snow. It is perfectly true that there have been times in the
past when very little of what I described as Brownian move-
ment was operating, and a whole set of decisions was taken
by a very small oligarchy. The Council of Ten in Venice is one
example. But I believe that from the time of the Industrial
Revolution a very large number of important decisions typi-
cally have bubbled up out of society. It seems to me extremely
important that this process should not be interrupted. One of
my worries is that the current nature of world tension, the

33

References

nature of our particular problems, is throwing once again undue
weight upon secret decisions by extremely small numbers of
people.

Morison. What troubles Sir Charles in this connection also
troubles me. In this country, historically, Presidents and their
cabinets have made decisions within a context of what they
knew society would let them do. Examples are the prosecutions
of trusts in the first ten years of this century, and the Monroe
Doctrine. Society had enough general evidence available to
gather out of the air, as Sir Charles said, an attitude within
which the President had to operate. In the future, it may be
increasingly difficult on certain kinds of very important ques-
tions to have that information available, so that in a sense men
will be acting much more irresponsibly.

Wiener. There is something here that is analogous to a
biological problem. A society in which those in control cannot
act effectively is suffering from paralysis; a society in which
those in control do not get sufficient feedback of the conse-
quences of their actions is suffering from ataxia, which is just
about as bad. If a man can move his muscles perfectly well,
but cannot get any report from his sensory organs on what his
muscles are doing, he is just about as badly off as if he were
paralyzed.

References

1. Snow, C. P., Science and Government (The 1960 Godkin Lectures of
Harvard University), Harvard University Press, Cambridge, 1960.

2. Samuel, A. L., "Some Studies in Machine Learning, Using the Game
of Checkers," IBM Journal of Research and Development, Vol. 3,
210-229 (July, 1959).

3. Jacobs, W. W., "The Monkey's Paw," in The Lady of the Barge,
Dodd, Mead, and Company; also in Modern Short Stories, Ashmun,
Margaret, Ed, The Macmillan Co., New York, 1915.

34

2

Managerial

2

Managerial
Decision Making

Speaker Jay W. Forrester

Professor of Industrial Management
Massachusetts Institute of Technology

Discussants Charles C. Holt

Professor of Economics
University of Wisconsin
Ronald A. Howard

Assistant Professor of Electrical Engineering and
Industrial Management
Massachusetts Institute of Technology

Moderator Robert C. Sprague

Chairman of the Board and Treasurer
Sprague Electric Company

36

Forrester. This series of lectures deals with information,
information processing, and management. This evening I should
like to concentrate on the relationship between information
and managing. Management is the process of converting in-
formation into action. The conversion process we call decision
making. Decision making is in turn controlled by various ex-
plicit and implicit policies of behavior. It is this area of deci-
sion making and the policies that control decisions which I
want to discuss this evening.1

If management is the process of converting information into
action, then it is clear that management success depends pri-
marily on what information is chosen and how the conversion
is executed. The difference between a good manager and a poor
manager lies at this point. Every person has available a large
number of information sources. But each of us selects and

57

J. W. Forrester

uses only a small fraction of the available information. Even
then, we make only incomplete and erratic use of that informa-
tion.

The manager sets the stage for his accomplishments by his
choice of which information sources to take seriously and which
to ignore. After choice has been made of certain classes of in-
formation and certain information sources to carry the highest
priority, managerial success depends on what use is made of
this information. How quickly or slowly is it converted to ac-
tion? What is the relative weight given to different information
sources in the light of the desired objectives? How are these
desired objectives created from the information available?

This evening we shall look upon the manager as an informa-
tion converter. He is a person to whom information flows and
from whom come streams of decisions that control actions
within the organization. There was a time when a person might
have been properly viewed as a converter of information into
physical action. The farmer or the craftsman used the available
information to guide his day of physical labor. The manager,
however, is not paid his premium salary in recognition of physi-
cal effort exerted. He is primarily an information converter
at his own particular control point in the organization. He re-
ceives incoming information flows and combines these into
streams of managerial instructions.

Viewing the manager in this way shows us immediately why
we are interested in decision making and information flow. An
industrial organization is a complex, interlocking network of
information channels. These channels emerge at various points
to control physical processes such as hiring employees, building
factories, and producing goods. Every action point in the system
is backed up by a local decision point whose information
sources reach out into other parts of the organization and the
surrounding environment. We shall concentrate here on the
individual decision point with its information sources. We shall
also need to be interested in the total structure whereby large
numbers of these decision points are tied together into an op-
erating organization, and how each draws on the available

38

Managerial Decision Making

information to create the dynamic characteristics of the organ-
ization. The information sources used, the policies for convert-
ing information into action, and the way in which the various
decision points are related to one another determine such im-
portant characteristics as corporate growth and stability of op-
erations or, on the other hand, financial failure, extreme em-
ployment fluctuation, and a succession of short-term crises.

Information-Feedback Structure of the Managament Process

We have said that information is converted by the manager
into instructions which lead to action. But, what effects do the
resulting actions have on the succeeding information inputs to
the next decisions? The manager uses information about con-
ditions that he hopes to influence and control. He expects,
therefore, that his immediate actions will have an effect upon
the future values of the information which he will be using.
This describes what is called an information-feedback system.

I should like to give an unusually broad definition of an in-
formation-feedback system:

An information-feedback system exists whenever the environ-
ment leads to a decision that results in action which affects the
environment.

This is a definition that encompasses every conscious and
subconscious decision made by people. It also includes those
mechanical decisions made by devices called servomechanisms.

Systems of information-feedback control are fundamental
to all life and human endeavor, from the slow pace of biologi-
cal evolution to the launching of the latest space satellite. Let
me illustrate:

1. A thermostat receives temperature information, decides to
start the furnace and changes the temperature.

2. A person senses that he may fall, corrects his balance, and
thereby is able to stand erect.

3. In business, order and inventory levels lead to manufactur-
ing decisions that fill orders and correct inventories.

39

/. W. Forrester

4. A profitable industry attracts competitors until the profit
margin is reduced to equilibrium with other economic
forces.

5. The competitive need for a new product leads to research
and development expenditures that produce technological
change.

All of these are information-feedback control loops. The
regenerative process is continuous, and new results lead to
new decisions that keep the system in continuous motion. Such
systems are not necessarily well behaved. In fact, a complex
information-feedback system designed by happenstance or in
accordance with what is "intuitively obvious" will usually be
unstable or ineffective.

In an information-feedback system, it is always the presently
available information about the past which is being used as a
basis for deciding future action.

Everything we do as an individual, as an industry, or as a
society is done in the context of an information-feedback sys-
tem. The definition of such a system is so all-inclusive as to
seem meaningless at first. Yet we are only now becoming suffi-
ciently aware of the tremendous significance of information-
feedback-system parameters in creating the behavior of these
systems.

Information-feedback systems, whether they be mechanical,
biological, or social, owe their behavior to three characteristics
— structure, delays, and amplification. Structure, as it implies,
tells how the parts are related to one another. Delays always
exist in availability of information, in making decisions based
on the information, and in taking action on the decisions.
Amplification usually exists throughout such systems, especially
in the decision policies of our industrial and social systems.
Amplification is manifested by actions being more forceful than
might at first seem to be implied by the information inputs
to the governing decisions. We are only beginning to realize
the way in which structure, time lags, and amplifications com-
bine to determine behavior in our social systems.

40

Managerial Decision Making

Inf or motion

\

\

Action

Figure 1 Decisions and Information Feedback

Figure 1 shows these relationships in simplest form. Informa-
tion is the input to a decision-making point that controls ac-
tions that yield new information. The diagram shows the
structural relationship. In each of these boxes are delays. In-
formation about actions is not immediately available. The deci-
sions do not respond instantaneously to available information.
Time is required for executing the actions indicated by a de-
cision stream. Likewise, each of the boxes contains amplifica-
tion, which I use here in all of its positive and negative and
nonlinear senses. In other words, the output of a box may be
either greater or less than is seemingly indicated by the inputs.
Likewise, the output may be noisy or distorted. The amplifica-
tion, the attenuation, and the distortion at each point in the
system can make the system more sensitive to certain kinds of
disturbing influences than to others. It can make the system
unstable, with tendencies toward internally generated fluctua-
tions. We see this kind of instability in the boom and recession
periods in our economic system. We see it in the way that
automobile companies and household appliance manufacturers
lay off hundreds or thousands of employees for two or three
weeks while inventories are brought into line with current levels
of business operation. We see it in the instability of copper
prices and of most mineral and agricultural commodities.

The industrial system is, of course, not the simple, single
information-feedback loop as shown in Figure 1. Instead, it is
a very complex multiple-loop and interconnected system as im-
plied by Figure 2. Decisions are made at multiple points
throughout the system. Each resulting action generates informa-
tion that may be used at several but not at all decision points.

41

/. W. Forrester

This structure of cascaded and interconnected information-
feedback loops, when taken together, describes the industrial
system. Within a company, these decision points extend from
the shipping room and the stock clerk to the board of directors.
In our national economy, they extend from the aggregate deci-
sions of consumers about the purchase of automobiles to the
discount rate of the Federal Reserve Board.

Nature of the Decision Process

We should now examine in finer detail the decision process
whereby information is converted to action. Figure 3 shows
the system structure as it surrounds the decision point. A deci-
sion is based on the state of the system, which is here shown
by the condition of various levels. Some levels describe the
present instantaneous condition of the system, and others our

42

Managerial Decision Making

Level

Level

\ \ I

L

Level

V i

Deci
Po

sion\ A
Int/Sl

Flow
Rate

Level

Figure 3 Decision Making in the System Structure

assumed knowledge about the system. A level may be an in-
ventory, the number of employees, the average sales for last
month, the accomplishment we believe has been made to date
in a research project, the degree of optimism about the eco-
nomic future, the size of bank balance, and so forth. These are
the inputs to decisions. The output from a decision point con-
trols the rate at which the system levels will change. I am
using decision here in a very broad sense. This includes the
filling of orders from an existing inventory, the placing of
purchase orders for new replacement goods, the authorization
of factory construction, the hiring of research scientists, and
the authorization of advertising expenditures.

A very important part of the concept of this organizational
structure is the directional relationship between the parts shown
in Figure 3. The levels are the inputs to the flow of decisions.
Decisions control flow rates between the levels. The flow rates
between levels cause changes in the levels. But flow rates them-
selves are not inputs to the decisions. Instantaneous, present

43

J. W. Forrester

rates of flow are in general unmeasurable and unknown and
cannot affect present, instantaneous decision-making.

In an industrial organization, a particular person may be
primarily responsible for the control of one particular flow
rate, as for example, the replacement orders for the mainte-
nance of an inventory. On the other hand, a particular person
may embody several separate decision points controlling several
separate flow rates. If so, we should look upon these separately
as lying within different parts of the information and action
network of the system.

A somewhat finer structure of the decision process is of in-
terest to us as shown in Figure 4. Decisions fundamentally in-
volve three things. First is the creation of a concept of a desired
state of affairs. What should we like to have the condition of
the system be? What are we striving for? What are the goals
and objectives of this particular decision point? Second, there
is the apparent state of actual conditions. In other words, our
available information leads us to certain observations that we
believe represent the present state of the system. These ap-
parent conditions may be either close or far removed from the
actual present state, depending on the information flows that
are being used and the amount of time lag and distortion in
these information sources. The third part of the decision proc-

Figure 4 The Decision Process

44

Managerial Decision Making

ess is the generation of the kinds of action that will be taken
in accordance with any discrepancy which can be detected
between the apparent and desired conditions. In general, the
greater the discrepancy between desired and actual conditions,
the greater the resulting action. However, this entire process
is highly nonlinear and noisy. Small discrepancies between ap-
parent and desired conditions may seem of little consequence
and create no action. A mounting discrepancy may lead to
more and more decisive attempts to correct actual conditions
toward desired conditions. However, as a maximum possible
action rate is approached, further widening of the gap between
desired and actual system states will no longer cause propor-
tionate increases in action from decisions.

Decision making is being presented here as a continuous
process. It is a conversion mechanism for changing continuously
varying flows of information into control signals that determine
rates of flow in the system. The decision point is continually
yielding to the pressures of the environment. It is taking ad-
vantage of new developments as they occur. It is always ad-
justing to the state of affairs. It is treading a narrow path be-
tween too much action and too little. It is always attempting
to adjust toward the desired goals. The amount of action is
some function of the discrepancy between goals and observed
system status.

One notes that we are viewing the decision process from
a very particular distance. We are not close enough to be con-
cerned with the mechanisms of human thought. We are not
even close enough to see each separate decision as we ordinarily
think of decisions. We may not be close enough to care whether
one person or a group action creates the decisions. On the other
hand, we are not so far away as to be unaware of the deci-
sion point and its place in the system. This proper distance
and perspective are important to our purposes. We are not the
psychologist delving into the nature and sources of personality
and motivation, nor are we the biophysicist interested in the
physical and logical structure of the brain. On the other hand,
we are not the stockholder who is so far from the corporation

45

/. W. Forrester

as to be unaware of the internal structure, social pressures, and
decision points.

Our viewpoint is more that of the managerial superior of a
particular person who is charged with certain responsibilities.
The superior is close enough to know how desired goals are
established. He is in a position to observe and probably provide
the information sources to be used by the subordinate to de-
termine his concept of actual conditions. He knows in general
the guiding policies and the manner in which the subordinate
decision maker would respond to various kinds of circumstances.

Policy

We are now led to what I shall call policy. The word "policy"
is here used as a broad term to describe how the decision proc-
ess converts information into action. What actions will result
from certain information inputs? WTrat is the conversion re-
lationship between information sources and the stream of re-
sulting decisions?

We shall first define what is here meant by policy. W e shall
later turn our attention to whether or not such policy does
exist and whether or not its form can be determined.

Policy is a formal statement giving the relationship between
information sources and resulting decision flows. It is what
has often been referred to in the literature as a decision rule.
In physical systems, particularly in the field of servomecha-
nisms, the corresponding term is "transfer function." The trans-
fer function tells how the output of a particular box depends
on the stream of inputs. The transfer function does not neces-
sarily deal with the particular physical way whereby the con-
version is accomplished. One is satisfied if the transfer function
tells us adequately well, for a particular purpose, the present
resulting action as a function of present and past inputs to
the box.

Most of the literature of economics deals with what we are
here calling policy. How will individuals and groups respond
to various circumstances and pressures? If conditions change
in a certain direction, what will be the direction of the re-

46

Managerial Decision Making

sponse? In industrial organizations, some policy is very formal.
It has been reduced to writing for the guidance of the decision
makers in the system. Most of the guiding policy is informal
but every bit as influential. It depends on habit, conformity,
social pressures, ingrained concepts of goals, awareness of power
centers within the organization, and personal interest.

We should here note that we have progressed through three
different levels of decision-making abstraction. At the lowest
level, we could imagine random unreasoned action, which
does not depend on inputs and which has no basis. At the
second level, we have unrationalized intuitive reactions, which
in fact result from the available flows of information but
where there is no comprehension by the actor concerning the
structure and the basis of his actions. We can assume that this
represents the thinking and the decisions of the lower ani-
mals. There is a basis and a reason for their decisions and ac-
tions, but they are totally unaware of such a basis and logical
structure. At the third level of abstraction, there is an aware-
ness of the formal reasons for decisions. Not only are decisions
made, but we have self-awareness of why we make certain
decisions, and we are aware and able to anticipate with some
reliability the kinds of reactions that others will exhibit in re-
sponse to changes in the state of their environment. The formal
awareness for the basis of decisions, which I am here calling
the guiding policy, certainly goes back as far as the written
record of our civilization. Man is most conspicuously separated
from the lower animals by this self-awareness of why he acts.
In other words, much of history and literature devotes itself
to the basis or policy that causes the human decision maker
to react in a reasonable and expected way to his environment.
When we say that there is a reasonable and expected reaction,
we are in essence describing the policy whereby information
will become a certain kind of action.

We should now consider whether or not we can detect the
nature of the guiding policy with sufficient accuracy so that
we can make use of it in better understanding the behavior
of the industrial and social systems of which we are a part.

47

/. W. Forrester

Clearly, people are of two minds on this question. Most of the
literature on decision making implies great difficulty and sub-
tlety in the subject. The social scientist makes very tentative
simple experiments with groups of three or four people in an
effort to determine how their decisions are reached in reacting
to one another. When we raise the question of understanding
the human decision-making process, the frequent answer from
scientists is that not even a good beginning has yet been made.
Yet the historian, the novelist, the manager, and every one of
us in his everyday life has been much bolder. We all discuss
why so-and-so acted in a particular way. In doing this we are
discussing his guiding policy. We are discussing how he did
respond or how he should have responded to information
available to him.

The dichotomy in our thinking is illustrated by two recent
encounters which I had with two different colleagues. One
flatly stated that it was clearly impossible to introduce the ac-
tions of the Federal Reserve Board into a formal model of
national economic behavior. The impossibility was argued on
the basis that we do not know the process by which such de-
cisions are reached. They are too subtle. They are subjective,
intuitive decisions for which we do not know the guiding pol-
icy. The other incident took place in a doctoral oral examina-
tion. Another colleague, as a routine matter, casually asked
the candidate to describe the factors that would lead the Fed-
eral Reserve System to make adjustments in various directions
in its discount rate and open-market policies. In other words,
the economics doctoral candidate was expected to know the
essential nature of the policy that would guide the stream of
Federal Reserve Board decisions. To be sure, there may be a
high noise content that can cause timing variations and un-
certainty in the extent of a response. However, the broad under-
lying outlines of guiding policy were expected to be within the
understanding and comprehension of the student.

This contradiction in the opinions we hold about the process
of decision making is very similiar to what we observe in our
thinking about the process of invention. There is great argu-

48

Managerial Decision Making

ment and little agreement in any discussion of how new ideas
are generated and how invention and research results are
achieved. Yet, we are almost in 100 per cent agreement in
acknowledging that more intelligent and more experienced
people, and greater research budget expenditure, and greater
motivation, and greater need for the results will all enhance
the probability of a successful outcome. This agreement on the
nature of the conversion function that couples financial and
manpower inputs to scientific output is the basis for Con-
gressional action and military department appropriations in
times of national emergency.

In short, our whole civilization is founded not only on the
assumption that a basis exists for the guidance of human action
but furthermore on the conviction that we know a great deal
about the specific nature and extent of this guiding policy.

We find then that management decisions can be classified
according to the form of the controlling policy, as in Figure 5.
Here we see four regions, segregated in accordance with the
extent to which the policies that guide decisions are known
and agreed upon. In Group A are those decisions that are now
made on a fully automatic basis by machines where the guid-

Figure 5 Scope of Management Decisions

49

J. W. Forrester

ing policy is rigidly prescribed. For some time this region has
included the rules for computing pay checks on the basis of
wage rate and hours worked. In many organizations it includes
a vast body of rules about the processing of accounting in-
formation and the flow and control of purchase orders, in-
voices, and shipping instructions. With the introduction of
computing machines, this area is rapidly expanding to include
the formal policies for production schedules and the manage-
ment of inventory levels, with implications that carry over into
production-rate and employment decisions. Group B represents
those decisions which are not mechanized but which are made
by a vast bureaucracy of middle management in our corpora-
tions according to well-understood guiding policy. This policy
has not been reduced to explicit detailed operating rules, and
yet it is clear enough that a supervisor generally knows what
to expect from subordinate decision makers. Region C rep-
resents those decisions, usually nearer the top levels of the
organization, where guiding policy does not appear in writing
and where decisions are thought of as being based on experi-
ence, intuition, and judgment. Even here, however, there is a
strong assumption about what constitutes proper action. In
general, it is well known what the direction of effect of various
changes in the system status will do to the resulting decisions.
The current management press, such as Business Week, the Wall
Street Journal, and Forbes magazine, is filled with the rationale
for management decisions. Much of the printed material is
devoted to a discussion of the pressures of the current state of
affairs and the effects these will have on decision makers. Be-
yond Region C is one of great challenge but in which there is
no basis for action in either experience or intuition. Into this
unknown, beyond the frontier of traditional management ac-
tions, we reconnoiter on a random and haphazard basis to ob-
tain information on which future intuitive judgment can be
based.

The dividing lines between these regions are not sharp. The
gradation from the automatic decision to the one that lacks
even a basis in experience and intuition is one of continuous

50

Managerial Decision Making

and gradual transition. The dividing zones between regions are
moving upward and outward.

Many people seem to believe that there is a sharp break be-
tween Region A in Figure 5, where decisions are completely
formal, and the other regions of management decisions. Such
persons are unwilling to accept the possibility of even the ex-
istence of formal policies that could describe the major aspects
of management in the other decision-making regions. I call
your attention to an interesting contradiction in the attitude
of many managers toward this matter of understanding the
formal basis of the decision process. Any manager must of ne-
cessity admit the existence of the region of automatic deci-
sions, since these are common practice. The majority of man-
agers will argue that the region of the intuitive judgment
decision is so subtle that no reasonable approximation can be
made to it through formal decision rules. Yet, those same
managers, when faced with a decision that they recognize as
lying beyond the capabilities of their intuitive judgment, will
once more fall back on formal decision-making procedures. I
here refer to the whole field of sales and market and economic
forecasting. Forecasting is essentially a decision-making process.
It consists of taking past and presently available information
and converting this into results that indicate a course of action.
I am not a supporter of the wisdom and validity of the majority
of this type of forecasting, but I simply point out the conflict
in attitudes. There are those who relegate the simple decisions
to automatic procedures. They rely, for lack of anything better,
on certain formal statistical decision procedures with respect
to some of the most subtle and difficult decisions. But they
reserve the middle ground as a region for judgment which they
assert is untouchable by formal decision rule.

It seems to me that there is now ample evidence to indicate
that this middle region is not the obscure and subtle jungle that
it has so often been pictured. Men are not good calculators of
the dynamic behavior of complicated systems. The number of
variables that they can in fact properly relate to one another
is very limited. The intuitive judgment of even a skilled in-

51

/. W. Forrester

vestigator is quite unreliable in anticipating the dynamic be-
havior of a simple information-feedback system of perhaps five
or six variables. This is true even when the complete structure
and all the parameters of the system are fully known to him.
The verbal model and the mental model that we cam' around
with us to explain the dynamics of industrial and economic
system behavior probably do not rank in effective dynamic
complexity beyond a fourth- or fifth-order differential equation.
We think that we give consideration to a much larger number
of variables, but I doubt that these are properly related to one
another in groups larger than five or six at a time. In short,
I am saying that, in dealing with the dynamics of information-
feedback systems, the human is not a subtle and powerful
problem solver.

Furthermore, we have some massive examples of our ability
to deduce these decision criteria and to go even so far as to
automatize and mechanize them. Take, for example, the tacti-
cal decision making in the conduct of an air battle. In 1947,
when I wrote a memorandum for the Navy on the use of
electronic digital computers as automatic combat information
centers, one could probably not have found five military offi-
cers who would have acknowledged the possibility of a ma-
chine's being able to analyze the available information sources,
the proper assignment of weapons, the generation of command
instructions, and the coordination of adjacent areas of military
operation. Yet, in just ten years the situation was completely
reversed. In 1947 no general-purpose electronic digital com-
puter had yet operated. There was more justification then for
the military officer than there is today for the manager to deny
vehemently the possibility that there is a formal and orderly
basis underlying the vast majority of management decisions.
The military officer argued that the analysis of radar informa-
tion, the determination of which data represented aircraft tracks
and which were noise, the identification of friend or foe, the
assessment of threat, the assignment of defensive weapons, and
the generation of weapon-control orders necessarily required
experience and judgment and a background of military training.

52

Managerial Decision Making

This of course overlooked the fact that the man at the decision-
making point might be a second lieutenant who just two weeks
before had been called up to military duty.

During the following decade the speed of military operations
increased until it became clear that, regardless of the assumed
advantages of human judgment decisions, the internal com-
munication speed of the human organization simply was not
able to cope with the pace of modern air warfare. This inability
to act provided the incentive. The nature of the decision proc-
ess was seriously studied. The policy that guided front-line
military decision making was unraveled. In the early 1950's
experimental demonstrations showed that enough of this deci-
sion making was understood so that machines could process
raw data into final weapon-guidance instruction and achieve
results superior to those then being accomplished by the manual
systems. During the 1950's, thousands of man-years of effort
went into a consideration of how military command decisions
are made and how these can be reduced to an effective orderly
system of decision-making rules. The military attitude was so
completely reversed in a decade that we find in the proceedings
of the Eastern Joint Computer Conference in December, 1958,
an article by Colonel W. H. Tetley of the Air Defense System
Integration Division out here at Hanscom Field in Bedford.
The title of the paper was 'The Role of Computers in Air
Defense." Let me pick three separate quotations to give the
general tone of that paper.

In essence, this master air defense system is a giant servo-
mechanism of really spectacular proportions. Although its reflex
action is achieved by information feedback, it still suffers many
vagaries of the simple feedback amplifier. Consider then, the
weird prospect of a servomechanism, spread out over an area
comparable to the whole North American continent, . . .

In what follows, then, rather broad terms will portray what
many in systems engineering believe lies ahead for the computet
in air defense. . . .

53

J. W. Forrester

First, there is the national level function which links together
all regional computers and co-ordinates their actions by means
of centralized control. This computer is able to exchange views
with the Strategic Air Command operational computer and with
that of the Office of Defense Mobilization, and thus co-ordinate
air defense on the governmental level.

The only amazing thing about this paper is that it seemed
to receive almost no notice from the audience. The reaction
was, "So what?" The audience looked on it as a routine, ordi-
nary paper presenting material that everyone already knew.
Ten years earlier an Air Force colonel would have dared to
give such a paper only at the risk of his military reputation.

You may say that the basis for conduct of an air battle is
simpler than for the conduct of a business. It is. But, in the
1950's, on the basis of the knowledge then available, it was
a great deal more complicated than will be the corresponding
understanding of the practical aspects of management decision
making in the decade of the 1960's.

The Management Laboratory

Now what is the practical significance of this discussion of
managerial decision making in its framework of the informa-
tion-feedback system? It means that if we can understand the
mechanisms of a social system, we can construct effective and
useful dynamic models of its operation.

These models open the way to what is essentially a manage-
ment laboratory. The design of a corporation with its organiza-
tional structure and its managing policies can be brought to a
level not unlike that of the design of an airplane with its wind
tunnel and its aircraft models. It opens the way to what we
might call "enterprise engineering" for the design of organiza-
tions to meet specific objectives and goals.

One might ask, is there any need to do this? Is there any
room for improvement? New design methods are unimportant
unless they lead to results that could not be achieved before.
One can point to the present crises in Africa. Certainly we do
not understand adequately the dynamics of growth of a new

54

Managerial Decision Making

country with its complicated interactions between education,
governmental structure, economics, technological change, mon-
etary system, external investment, and the aspirations of its
people. One can point to the large number of new industrial
enterprises that are founded but to fail. On the other hand,
there are those that begin and flourish with a growth rate of
as high as 50 or 100 per cent per year. When the range in
dynamic behavior is as great as this, there must indeed be some
fundamental underlying reasons for success which can be better
understood. We have the example of the electric generating
industry. Electric demand is one of the most smoothly growing
variables in our economic system. Yet the rate at which orders
are placed for new turbines and electric generators fluctuates
by as much as 10 to 1, with a periodicity of four, five, or six
years. Actual production rate of this equipment fluctuates as
much as 4 to 1, with the corresponding labor instability and
individual and community hardship. Such behavior is a mani-
festation of the unstable interactions that can exist within an
information-feedback system having certain managerial poli-
cies, certain time delays, and certain technological restraints.
Changing any of these, and particularly the managerial poli-
cies over which one can exert control, will change the behavior
of the system. There is every reason to believe that such un-
desirable system behavior characteristics can be greatly allevi-
ated.

To begin to deal with the dynamic characteristics of social
systems, we must be able to represent at least the central essen-
tial skeleton of the decision-making structure. To do this, we
must be able to approximate the controlling policy at each
significant decision point in the system. This understanding of
policy can be accomplished if

h We have the proper concept of what a decision is and
of the significance of the policy that describes the decision
process.

2. We have the proper structure relating system status to
policy, to decisions, and to action.

55

J. W. Forrester

3. We realize that the process is noisy and we shall not get
and do not need high accuracy of decision-making rep-
resentation.

4. We use to best advantage the extensive body of experi-
ence and descriptive information that probably contains
98 per cent of the essential information on decision mak-
ing. The other 2 per cent will come from formal statistical
and numerical data.

5. We realize that a formal, quantitative statement of policy
carries with it no implications one way or the other re-
garding absolute accuracy. We can make a formal quanti-
tative statement corresponding to any statement that can
be made in descriptive English. Lack of descriptive accu-
racy does not prevent quantifying our ideas about decision
policy. Assigning a number does not enhance the accuracy
of the original statement. The common belief that we
cannot quantify a decision rule because we do not know
it with high accuracy is mixing two quite separate con-
siderations. We can quantify regardless of accuracy. After
that, we deal with the question of what is sufficient accu-
racy.

I feel it has been adequately demonstrated that these things
can be done. We have the tremendous examples of correspond-
ing accomplishments in the understanding of military control
systems in the last decade. We have preliminary examples of
the application of the same approaches to industrial systems.
Let me touch quickly on an example.

Example

To give you an example of an application of the use of
decision-making policy in the area of industrial operations, I
shall mention some work that we have been doing with the
cooperation of Mr. R. C. Sprague and with the financial sup-
port of the Sprague Electric Company.

There was a rather common but baffling industrial problem.
In a particular product line there was substantial fluctuation

56

Managerial Decision Making

of inventories and employment, and this fluctuation seemed
much more extreme than could be explained by expected
changes in the ultimate end demand for the product. The ques-
tion arose: Could the fluctuating inventories and production
rate come from interactions among the managerial policies in
the complex information-feedback system comprising the com-
pany, its customers, and its employment and production poli-
cies?

Figure 6 shows a time history of how this system operates.
The problems are clearly visible. We see that employment is
fluctuating even more than the incoming-order rate. Further-
more, the excess employment is being created by the inventory
fluctuation. This is very common in our mature industries.
Unlike most electronic products, this particular product line is
mature and well established and is showing little growth or
decline from year to year. Xote that inventor.- is rising rapidly
at the time of the employment peaks, which occur only a little
later than appear to be the high points in sales. Many people
believe that inventories absorb fluctuations in demand and
help stabilize production rate and employment. This is true
of the very short run demand changes but is often quite the

125%

75%

/

/ Inventory

1st Year

2nd Year

3rd Year

4th Year

Figure 6 Behavior Shown by a Mcdel cf the System

57

J. W. Forrester

reverse for longer-term fluctuations that create the major sales
and employment problems.

Near the end of the first year and in the first half of the
third year, employment is high. These employment peaks have
been preceded by a period of above-average sales, but employ-
ment increased higher than did sales. The higher employment
resulted from the attempt to recover the inventory depletion
that occurred during the high sales period. The fact that in-
ventory is rising during the sales peak and is falling during the
sales valley causes employment to fluctuate more widely than
sales.

Now as you look at this figure, you should know that it has
been created entirely by a mathematical model that represents
our estimate of the decision policies in the system and how
these are related to one another. Furthermore, the model is
completely closed except for an external demand that is abso-
lutely constant and unvarying. In other words, here is a model
that we believe represents point by point the decision functions
of the system. Each of these decision functions, as we discussed
earlier, recognizes the goals and objectives at the particular
point, uses the information sources that predominantly in-
fluence that decision point, and generates the flow of decisions
that we believe characterize each particular point in the system.
Having done this, we find that the model has an interacting
behavior within the multiple information-feedback loops such
that it creates a pattern of the same qualitative nature as the
real system, even in the presence of a constant final demand for
the product. The model says that indeed it is quite possible to
have widely fluctuating inventories and employment that far
exceed any variation imposed upon the system from the outside.

We do not have time to dwell at length on this example, but
let me give you an idea of what is involved. Figure 7 shows
the most crucial of the decision points in the system and the
major information-feedback loops that are involved. Actually,
the mathematical model used for this system is a very nonlinear
one of about 140 variables. Of these, about 30 are decision
points represented by decision-making policy functions such as

58

J

Managerial Decision Making

Figure 7 Major System Details

we have discussed. Of the 30, 5 can be singled out as especially
important and these are shown in the figure. The 5 are the
following:

1. The purchasing rate by the customers as it depends on
design releases from their engineering departments, their
inventory policies, and their response to changing delivery-
delay quotations coming from the supplier.

2. The factors that determine the fraction of the incoming-
order flow which the factory can fill from inventory.

3. The basis for quoting delivery delays from the supplier
to the customer.

4. The basis for inventory reorder from the stock room to the
production department.

5. The policies governing changes in employment and work
week which control the production rate.

As an example of one of the decision-policy structures, we
may examine that for the employment change rate in Figure 8.

59

/. W. Forrester

Figure 8 Employment Policy Structure

The rectangles represent the principal levels on which the deci-
sion depends. At the top the present production workers and
the present people in training tell us the labor force that is to
be available. This represents the present status of the system.
At the lower left, average sales give a basis for determining
what the normal backlog should be. This is a goal or desired
condition as discussed earlier. This is compared with the actual
present backlog representing present conditions, to determine
the amount of labor desired for adjusting upward or down-
ward the size of the unfilled-order backlog. Average sales also
imply the amount of labor necessary for producing at the rate
corresponding to average demand for the product. The labor
for meeting average sales demand plus the labor increment for
the adjustment of unfilled-order backlog combine to indicate
the desired labor force. At the labor decision point there are
then available the desired labor level and the actual labor level.
The decision function then combines these, along with con-
siderations of the rapidity with which changes can be made to
yield a rate of employment change.

60

J

Managerial Decision Making

Turn now to Figure 9. Here are two time histories, both gen-
erated by dynamic models of the corporate system. It is clear
that these differ in qualitative character. Conditions are much
more stable in the lower time history than they are in the upper.
The upper history is the one in the earlier figure and represents
the behavior of the system model when it is set up with the
policies that we believe have been governing the operations of
this product line in the past. The lower history is the same
model, in which some rather minor adjustments have been
made in the sources of information used for certain decisions.
The new system in the lower time history also specifies in sev-
eral places longer averaging times and slower managerial re-
sponses to available information. In a number of respects there
is very substantial improvement. Although inventories fluctuate
through approximately the same amplitude as before, every-
thing else is more stable. Employment fluctuations are about
one-third as great. Although not on this simplified figure, the
actual model results show that cash position is more stable,

-Orders

'\ Empl<
\

\ f\
V ^ N / \ j

\ v. A /

Dyment

^>

A

' Inventory
Old S

wMT\
ystem

_Orders
/'J.

J\ W\ — _

Emplc

yment

T \ Nj

_ \y

1st Year

J v \

2nd Year

-A

Inventory

NewS

3rd Year

ystem
4th Year

5th Year

\
6th Year

125%
100%
75%

75%
100%
125%

Figure 9 Comparison of Models of Old and New Policies

61

J. W. Forrester

and unfilled-order backlog fluctuates less, as does deliver}' delay
to the customer.

This illustrates what I believe to be a general characteristic
of our industrial systems. They are so far from the best possible
mode of operation that our first steps to improve them will
often lead to improvement in nearly every variable of interest.
One is not faced initially with the need to make a compromise
wherein a heavy price must be paid in one factor to obtain
improvement in another.

Decision Policy in the Corporation

It seems to me that management science has so far been very
ineffective in dealing with this formulation and use of decision-
making policy. I believe the past difficulties might be gathered
into three categories.

First, there is the matter of perspective or viewing distance
mentioned earlier. The social scientist has tended to look at
the individual man, with emphasis on psychology and individual
motivation. Many of the attempts at laboratory experiments
have been with small groups assembled in an artificial environ-
ment for brief periods of time. The results have not been con-
ditioned by the strong social forces of precedent, conformity,
and the attempt to behave as the man believes his superior
would want him to do. Study of the individual man, especially
over short periods of time in an artificial environment, tends to
accentuate the feeling that decisions are capricious, infre-
quent, disconnected, and isolated.

At the other extreme, the economist has viewed the corpo-
ration from too great a distance. He has often looked upon
the market as maximizing its utility, whether or not it has
the information available to do so. He looks upon the entre-
preneur as a man who maximizes his profit without asking
whether or not he has the available information sources and
the mental computing capacity to find a maximum. This view-
ing from too great a distance tends to overstress the importance
of the top-management decisions compared with those made
in the lower and middle structure of the corporation. A direc-

62

Managerial Decision Making

tive from top management does not change the prejudices, hab-
its, and self-interest objectives of the middle-level decision mak-
ers. For example, the public press has well documented the
futility felt by the successive Secretaries of Defense who have
tried their hands at changing the character, attitudes, and prac-
tices of our military department. It is a long, slow process. We
tend to be misled by various "upheaval" incidents. A proxy
fight followed by a complete change of top management and
the firing of half of the middle-management structure will in-
deed change the attitudes and traditions of an organization.
But this drastic surgery is not common.

For an understanding of the corporate information-feedback
system it is very essential that we look neither at the individual
person isolated from his environment, nor at the exterior of
the system. It is at the intermediate viewpoint, from which we
can know men and groups of men in their working environ-
ment, that we can capture the true character of the opera-
tion.

Second, I think that the understanding of decision making
has been greatly handicapped by the assumption that it is a
more subtle and more skilled process than it actually is. We
have been too heavily influenced by the fact that the highest-
speed computers cannot yet play chess as well as a person. This
is not a typical example. The chess problem is essentially static.
Full and exact information is available to the man. It is a
problem of visualizing spatial relationships, which a man does
quite well, and today's machines do poorly. There are other
situations such as the time histories that I showed you earlier,
where a computing machine can determine in five minutes
what the consequences of a set of policies will be and where a
group of men could argue inconclusively for a year about what
follows from a given set of assumptions.

It is my feeling that in a dynamic information-feedback sys-
tem the human decision maker is usually using a great deal less
than the total amount of information available to him. Fur-
thermore, the information available to him is a great deal less
than that commonly assumed. In general, his actions with re-

63

/. W. Forrester

spect to any given decision stream will be almost entirely con-
ditioned by less than ten information inputs. What he does
with these few information sources is apt to be rather stereo-
typed. Some of them will be used to create a concept of de-
sired objectives. Others will serve to form his impressions of
the true state of affairs. From the difference will result reason-
ably straightforward and obvious actions. What seems obvious
may not be best. Some of the biggest improvements that can
be made in the dynamics of our industrial systems will come
from following procedures that our traditions and our man-
agement folklore have led us to believe are precisely in the
wrong direction. Our understanding of the dynamics of com-
plicated information-feedback systems is so inadequate that our
intuitive judgment can often not be trusted to tell us whether
improvement or degradation will result from a given direction
of policy change.

The third difficulty into which many seem to have fallen in
attempting to understand decision-making policy has resulted
from trying to skip one of the evolutionary steps in the hier-
archy of decision-making abstraction mentioned earlier. You
will recall that I suggested the lowest level of decision making
would be that where actions were random and irrational. The
second level would be one where actions are reasoned and ra-
tional but where there is no self-awareness of what the govern-
ing policies might be. Man was characterized as having,
throughout recorded history, at least a verbal descriptive model
of a rational policy that creates the stream of individual deci-
sions. This is already a major fraction of the step toward the
ability to formulate explicit quantitative policy governing deci-
sions. The next level would be one in which the art and the
intuitive judgment are applied to the development of a better
understanding of policy.

Art, judgment, and intuition are at this new level no longer
applied to the individual separate decisions but to the defini-
tion of a policy that governs the stream of individual decisions.
This is the level of abstraction which is just emerging. This

64

Managerial Decision Making

is the one wherein we can point to numerous successful ex-
amples but where there is not general agreement on method.
At this level of abstraction, there is not a descriptive literature
on what constitutes the practice of the art of detecting decision-
making policy.

In spite of the gap in the art which exists at the policy-
detection level, many economists have attempted to skip to the
next level of the abstraction hierarchy. They have been at-
tempting to develop statistical methods that can be applied as
a routine matter to extract from quantitative data about a
system the governing decision policies. This is still another level
of abstraction in which intuitive art and judgment are applied
to setting up rigid rules whereby the formal decision policies
can be derived. I feel that we are not ready to attempt this last
level of abstraction until we have achieved acknowledged suc-
cess in applying art and judgment and intuition to the ex-
traction of the decision policies themselves. After this process
is well understood, it ma}* then be possible to reduce this
method of analysis of an organization to a rigid and orderly
process. In general, the precedents seem to indicate that we
must take these levels of abstraction one at a time. At each
point in time, art and judgment are devoted to establishing
the rules whereby the lower levels can be automatized.

An example has been the history of computing machine
programming. Ten years ago one wrote the specific machine
code for the solution of a particular problem. The next level
of abstraction was to write a program of logical instructions to
tell the machine how to create its own running program for a
specific problem formulation. We now seem to be entering
another hierarchy of abstraction in problem programming, in
which the man develops concepts that allow the computer to
formulate the specific statement of the problem which another
computer program will, in turn, convert into machine language.
This I feel is philosophically the equivalent of what I have
been describing in the hierarchy of orderly rules related to
decision making.

65

/. W. Forrester

Future Implications

Let me close by commenting briefly on some of the implica-
tions that I see developing out of what we have discussed.

First, this does not imply automatic management. A better
understanding of decision-making policy and its information-
feedback context will not reduce the leadership demands on
the executive. Quite the reverse. He will now have new meth-
ods to use and a new theoretical underlying structure to under-
stand. The use of this new knowledge and these tools will not
be automatic. The more skillfully these tools are selected and
the more significant the goals, the more effective will be the
application. New advances in physics have not led to automatic
engineering. The skilled system engineer still produces the bet-
ter chemical process plant or earth satellite. The great ad-
vances in physics have not yet led to automatic methods for
selecting which principles should be applied or in what com-
binations.

We can, however, see that the emphasis in the corporation
will change. Management education has concentrated on the
making of individual decisions. The manager has been looked
upon as a decision maker. On the other hand, the truly skillful
managers have long recognized that their greatest contribution
came through the establishment of a proper policy and frame-
work for the guidance of the organization. This recognition of
the importance of policy can be expected to spread very rapidly
during the next two decades.

The manager who will be most in demand and who will
command the highest salary will be the enterprise designer.
He is the one who will be able to work with the way informa-
tion sources and decision-making policy at each of the many
points throughout the organization combine to create desir-
able results.

At the same time, there will be a thinning out of the present
middle-management organization. It is in the middle-manage-
ment levels that decisions are ahead}- highly routine and de-
termined by formal policy and system structure and precedents.

66

Managerial Decision Making

Much of the middle-management structure is made up of frus-
trating and unrewarding tasks. The tasks are managerial in
name only. They are actually steps in the information-process-
ing production line. As these tasks are better understood, it
will no longer be necessary to squander the talents of good
men in these positions. The analog}' to the physical-product
production line is excellent. As the production process became
routine, two things happened. Work on the production line
became unrewarding. Also, once the tasks were fully under-
stood, it became possible to design machines that would serve
the function as well and would not suffer from boredom.

We spoke of hierarchies of abstraction with respect to our
understanding of the decision-making process. I call your at-
tention to a similar evolutionary hierarchy with respect to our
industrial production processes. There was a time in an agrarian
economy when man produced with his own labor to satisfy
his immediate consumption needs. The next step was to spe-
cialize into a craft-trades economy, where judgment and intui-
tion were acquired through apprenticeship to achieve greater
manual skill and higher labor productivity. This evolved into the
early industrial economy, where manual labor was subdivided
and simplified so that people operated machines to produce
goods for others. Skill and intuitive judgment were applied to
the construction of machines that could produce what the
craftsman had done before. At the next step, which we can
call the advanced industrial economy, the creative level moved
from machine construction to machine design. Then the con-
centration was on the principles and the art of machine design,
with emphasis on kinematics, metallurgy, and thermodynamics.
At the next stage, which we are now entering, the art of ma-
chine design is being formalized into logical rules. We can
call this the automation economy. At this level, man designs
methods and procedures that will permit machines to design
machines that in turn will produce for human consumption.
We have already taken a substantial step in this direction.
Digital computers are today designed by digital computers.
Computing machines design mechanical parts and lay out the

67

C. C. Holt

piping and structure for chemical plants. The creative designer
works with the formal concepts that govern how design should
be done. At the next level, which we might call the "artificial-
intelligence" level of economic development, machines will
perhaps be instructed in how to develop the design concepts
that machines can then follow to produce designs to produce
equipment to produce goods for human consumption.

Each of these steps has led to a higher standard of living.
Each hierarchy of abstraction is a multiplier that converts hu-
man effort into a higher standard of material welfare. Each
step has created its social revolution as people find the old
activities declining and new ones taking their place.

The manager, in rising to a new level of abstraction where
he concerns himself with formal decision-making policy and
with enterprise design, will be keeping pace with the evolution
of our technological society.

Panel Discussion

Holt. In discussing Professor Forrester's paper, I should
like to give you my initial general reaction, which was en-
thusiastic, and then give you some second thoughts on a more
detailed level.

At the risk of distorting Professor Forrester's intent, I should
like to summarize very briefly the procedure that he proposes:

1. Make detailed studies of the decision making within a
company (or other organization) and formulate a model
of the decision policy for each important decision center.

2. Use an electronic computer to simulate the resulting
model of the over-all decision system in order to determine
its characteristics.

3. Validate the model by checking it against the actual per-
formance of the organization.

4. Test proposed improvements in the system by perform-
ing experiments on the computer model.

68

Managerial Decision Making

5. Introduce the improved policies into the organization —
and ultimately move toward administration of decision
policies by the computer itself.

Let me indicate why I think this procedure is a most prom-
ising one to pursue. Detailed empirical studies of operating
decision policies in the context of a complex decision system
are all too rare. By such studies, we should learn a good deal
about the way organizations actually behave, as contrasted to
how they ought to behave. Although such studies obviously
could have been carried out before electronic computers were
available, it is important to recognize why they were not, at
least not on the scale and with the rigor contemplated here.
On one hand, the computer forces one to formulate an opera-
tionally complete system; and, on the other hand, the avail-
ability of the computer makes it possible, with a reasonable
expenditure of resources, to find the implications of such
studies instead of ending up with a huge and utterly indigesti-
ble case study.

Mathematical methods currently are limited, where general
nonlinear equations are involved, to fairly small systems; so
the emergence in recent years of electronic analog and digital
computers capable of solving large-scale nonlinear equation
systems constitutes a major breakthrough in the treatment of
complex systems problems. Forrester's proposal to exploit this
breakthrough in seeking solutions to problems arising in deci-
sion systems is thoroughly sound.

As any scientist knows, it usually is easy to formulate many
hypotheses (or models) that are plausible. Consequently, there
is a critical need for much careful testing in order to isolate
the few hypotheses that yield empirically accurate predictions.

After the model has been validated and the researcher turns
to proposing improvements in the system, he faces the fact
that he too is the same kind of mortal as the businessman, in
that he also is not a ''good calculator of the dynamic behavior
of complicated systems." Fortunately, computer simulation of
models facilitates experimentation so that more or less random

69

C. C. Holt

search guided by a few heuristics becomes feasible for the re-
searcher in a way that was never possible for the business exec-
utive working with a live organization. Given the sad state of
dynamic control in most business organizations, there is little
doubt that recommendations can be made for changed policies
that, if instituted, would yield substantial improvements in
organization performances.

The potentiality of the computer to participate actively in
the decision process "shoulder to shoulder" with men is still
not generally appreciated. If anything, Professor Forrester has,
under the pressure of time, understated his case.

Professor Forrester's approach certainly deserves to be ex-
ploited, and it can be counted on to yield substantial improve-
ments. We all know that operations researchers working with
much less powerful tools for dynamic analysis have found no
great difficulty in achieving substantial improvements in pres-
ent-day industrial practice. However, it needs to be emphasized
that dynamic control problems are only one area of concern
in the more comprehensive field of management science.

Perhaps it should be mentioned that other workers have
used an approach that is somewhat similar to Forrester's: for
example, Richard Cyert, Guy Orcutt, Alan Rowe, and Martin
Shubik.

Now for my second thoughts on a miscellany of more de-
tailed points. Forrester gives considerable stress to the defi-
nition: "An information-feedback system exists whenever the
environment leads to a decision that results in action which
affects the environment." He further states that "this definition
. . . encompasses every conscious and subconscious decision
made by people" and "everything we do as an individual, as an
industry, or as a society is done in the context of an informa-
tion-feedback system." I trust that what I am about to say
will not be interpreted as meaning that I am against "feed-
back," but I think that I detect here a confusion between an
intellectual model and the real world. Either an "information-
feedback system" is defined so broadly that it applies to virtu-
ally all human phenomena, and hence cannot have much, if any,

70

Managerial Decision Making

content as a statement about the real world; or it is not so
universally applicable as is claimed. I should contend that the
feedback model as expressed in Figure 1 is very useful for
thinking about certain organizational phenomena, but we
should guard against pretending that such models are true
and exact statements about the real world. We should not
confuse models that we use to think about the real world with
the real world itself.

The notion of circular causality that is incorporated in
Forrester's "information-feedback system" is an important
analytical model which in any particular situation is capable
of being either true or false. There is no denying that certain
situations involve simultaneous determination (rather than uni-
directional causality) : for example, a manager with a budget
constraint governing several different types of expenditure. It
will not always be wise to limit ourselves to models based on
circular causality in spite of the computational simplicity that
may be obtainable by this approach. Forrester might argue
that nothing can move faster than the speed of light, and
hence, there can be no simultaneous relationships in a spatial
system. To that argument, I can only answer that it rests on
a hypothesis that is subject to empirical validation.

There is no question that the feedback concept supplies a
useful perspective in examining a decision system. However, a
somewhat finer distinction contributes to clarity. If we think of
"feedback" as the decision response to the state of the system,
then it is useful to introduce the additional concept of "feed
forward" for decision responses to anticipated future disturb-
ances on the subsystem being controlled by the decision maker.
It is helpful in studying decision policies to be sensitive to the
fact that a decision maker may be responsive to some off-beat
and seemingly irrelevant information because it has some value,
however indirect, in forecasting variables that are directly im-
portant.

I agree with Forrester that most psychologists and econ-
omists have been viewing the business firm from "distances"
different from Forrester s, but I think it is only fair to say that

71

C. C. Holt

most of them have been trying to answer questions different
from the ones that he poses. Most psychologists and econ-
omists have not concerned themselves with solving the busi-
nessman's dynamic problems. Fortunately, many social scien-
tists are now taking steps to remedy their sin of omission
and increasingly are taking a serious interest in problems that
arise in business organizations at all levels.

In trying to determine decision policies, there is an appeal-
ing directness in simply asking decision makers what they are
doing. This is not an approach that has been overlooked by
either psychologists or economists. It is well known that hu-
man beings are quite ready to supply rationales for their be-
haviors. It is also well known that when you ask a human sub-
ject to be his own social scientist and predict his own behavior
the results usually do not stand up under critical empirical
testing. In efforts to obtain reliable predictions, social scien-
tists have been forced to use more subtle and devious methods
than "interviewing the subject." There are very real dangers in
being too simple-minded in our approach to this difficult re-
search problem.

Although there is frequent mention in the paper of random
noise in decision procedures, very little emphasis is given to
the well-developed body of knowledge for treating such phe-
nomena, namely statistics. This is quite surprising and suggests
that the problem of validating the estimates of decision policies
at the level of the decision center has not been adequately
explored. It would take a wise man indeed to get 98 per cent
of the information from a noisy, partly subconscious system
and without benefit of statistics. As a scientist, I am unwilling
to accept "agreement" on the determinants of invention as an
adequate substitute for empirical tests.

I want it to be clear that I am not objecting to obtaining
information in any way that we can, including interviews. But
I would like to see some effort made to secure independent
empirical tests of this information rather than simply assum-
ing its accuracy. Statistical methods would be useful for such
tests, and also might be useful for making estimates of the

72

Managerial Decision Making

decision policies themselves. To suggest that statistical methods
are rigid and mechanical is, I think, to misinterpret the dis-
cipline.

When the simulated system has been assembled, there is a
further nasty problem of validating the model to see whether
its over-all interactions adequately mirror the organization that
is being modeled. I am led to wonder about the adequacy of
the test in the production and employment example that in-
dicated that the model "generated patterns of the same quali-
tative nature as the real system." I urge that more emphasis
be put on the validation of system models, even though I am
aware of the morass of unsolved problems that are involved.

At the experimental stage when improved policies are being
proposed and tested, Forrester refers to the wind-tunnel anal-
ogy. In that field there is a body of abstract theory that is of
great help to the wind-tunnel experimenter. Similarly, we need
theoretical insights that will help us to understand the or-
ganizational phenomena with which we are dealing. The fact
that we can have a simulation model of the system for experi-
mentation does not by any means make it an easy task to un-
derstand what is going on in the model. Thus the simulation
model is not a panacea. There is still a great need for theoret-
ical understanding of such complex systems.

The problem of implementing changes in organizations
should be mentioned at least in passing, since operations re-
search people often find that knowing what ought to be done
is somewhat less than half the job. Often there is a great deal
of work involved in bringing about changes in channels of
communication and in decision policies. Indeed, a manager
soon learns that the management role in a human organiza-
tion involves a great deal more than just decision making.

In closing I should like to comment on the complementarity
of alternative approaches to the problems that Professor For-
rester considers. For example, a researcher might enter an
organization and study not the prevailing decision policy but
rather concentrate on the decision problem, i.e., the constraints
and objectives that face the decision maker. Then by the use

73

R. A. Howard

of mathematical decision theory and numerical methods, he
could attempt to obtain optimal designs both of the organiza-
tion and of the decision policies. This approach lends itself
well to considering a few critical decisions while suppressing
most of the detail in studying the rest of the organization. Such
an approach, because of its narrow focus and great depth, has
been found by me and others to offer considerable power and
flexibility, but it also carries certain obvious dangers. Hope-
fully, such intensive normative studies might supply some of
the theoretical insight whose need was mentioned before.

Since optimization problems involving complex dynamic
systems are likely to prove difficult of solution, it is desirable
for us to develop many alternative approaches. They will have
a great deal to contribute to each other, and it is to be hoped
that effective channels of communication can be established
within and between research organizations in order to take full
advantage of this interplay. The simulation approach pre-
sented by Professor Forrester has a very important contribution
to make in this interplay, especially in the area of dynamic
stability.

Howard. My first feeling on hearing Professor Forrester's
comments was similar to the feeling I experienced upon grad-
uation from the School of Engineering. With slide rule in hand
I went forth to meet problems as Saint George went forth to
meet the dragon, having full confidence that the engineer
could do anything. All that was necessary was to grab the
problem and wrestle with it. I am sadder and wiser now. Hav-
ing wrestled with some problems, I no longer have as much
confidence that our methods and approaches are omnipotent.

In discussing Professor Forrester s paper, I should like to
treat and contrast two general approaches to the study of in-
dustrial and economic systems: industrial dynamics and opera-
tions research. Some of my colleagues in other fields seem to
be confused about the relationship of Professor Forrester's in-
dustrial dynamics and the general field of operations research.
Some individuals who are acquainted with Professor Forrester's

74

Managerial Decision Making

work seem to believe that any approach to industrial problems
that utilizes digital computers falls in the province of industrial
dynamics. On the other hand, some operations research ana-
lysts, economists, and businessmen feel that industrial dynamics
is not a discipline but rather a single isolated approach to in-
dustrial problems. Some researchers feel that industrial dy-
namics is the road to the future; others feel that it is a dead
end. As usual, the truth probably lies somewhere between these
views.

Industrial dynamics is a term coined by Professor Forrester
to describe the method of studying industrial operations that
he has developed. Professor Forrester has defined his approach
as follows: "Industrial dynamics is an analysis of the inter-
actions between the flows of information, money, orders, ma-
terials, capital equipment, and manpower in a company, an
industry, or a national economy/'1 As the name implies, in-
dustrial dynamics is concerned with the time-varying or dy-
namic behavior of industrial organizations.

On the other hand, to define operations research is con-
siderably more difficult because even the workers in the field
cannot agree on a definition. However, for our purposes, the
following definition will be useful. Operations research is a
scientific method for studying men and machines performing
repetitive operations. For convenience we shall refer to the peo-
ple who make industrial dynamics studies as dynamicists and to
those who make operations research investigations as analysts.

The question of the moment is: How is industrial dynamics
related to operations research and what promise does each of
these fields hold in the solution of management problems? To
answer this question let us compare industrial dynamics and
operations research with respect to their phenomena of study,
their methodology, their research philosophy, and the state of
the art in each field; finally, let us evaluate each approach and
offer a few remarks about possible future courses of develop-
ment.

Phenomena. Industrial dynamics is concerned primarily with
industrial operational problems. Operations research is con-

75

R. A. Howard

cerned with the man-machine systems that are found in in-
dustry, government, and the military establishment. The dy-
namicist desires to study the highest-level problems of the
enterprise: the basic forces controlling industrial growth and
stability. On the other hand the analyst, up to the present
time, has been content to concern himself primarily with
lower-level decision problems: questions in the quantitative
management of inventory, the establishment of effective mar-
keting policies, and the improvement of distribution methods,
for example.

Methodology. Computers are used extensively by both the
dynamicist and the analyst. At the present time the most im-
portant models of the dynamicist are based upon a system of
first-order difference equations that must be solved recursively.
Because of the lofty aim of the dynamicist, only a digital com-
puter has the computing and storage capacity necessary to pro-
vide the solution. As a result, the existence of large electronic
computers is almost a prerequisite of industrial dynamics as
we see it today.

On the other hand one can make a case for the point of
view that computers are not necessary to the field of opera-
tions research in spite of their widespread use. The analyst
attempts to build analytic mathematical models of the opera-
tions that he is studying, using such disciplines as mathe-
matical programming, queuing theory, probability theory, and
their equally imposing brethren. Although computers often
have been necessary to obtain numerical results, there are many
examples of studies in which significant contributions have
been made without their assistance. Both the dynamicist and
the analyst find useful the control theory of linear systems,
although few of the systems that they encounter in practice
are, in fact, linear. They also both use simulation, where com-
puters are called upon to provide a working model of a com-
plex enterprise.

Research Philosophy. The research philosophies of the dy-
namicist and analyst are markedly different. The general ap-
proach of the dynamicist is to describe a system by a large

76

Managerial Decision Making

set of first-order difference equations. He is more likely to err
on the side of too many variables rather than too few, because
he has faith in the computational ability of his computer. The
analyst, on the other hand, prefers to limit the number of
variables as much as possible because he is looking for analytic
rather than computational solutions. As a believer in Occam's
Razor, the analyst will add variables only if they can be shown
clearly to have an influence upon the total behavior of the
system.

The goals of the two investigators differ in that the dynam-
icist's results are often descriptive rather than normative in na-
ture. The dynamicist attempts to show how changes in time lags
and amplifications throughout the system will produce qualita-
tive differences in over-all system behavior. On the other hand
the analyst's approach is normative. At his low level he attempts
to find how decisions should be made in order to increase the
effectiveness of the operation that he is studying. These differ-
ences in approach are responsible for much of the conceptual
conflict that sometimes arises between the dynamicist and the
analyst.

State of the Art. Much of the disparity in numbers between
dynamicists and analysts at the present time can be explained
by the fact that industrial dynamics is an area that has ap-
peared only in the last few years. Although operations research
is an infant science, it has been in existence about twenty
years. Possibly as a result of its age, industrial dynamics has
had little opportunity to provide convincing illustrations of
successful applications. Operations research has a record of
significant accomplishment, although there have been some
unfortunate experiences.

Evaluation. Let us now attempt to evaluate the progress that
has been made in each field. Industrial acceptance is probably
not a good yardstick for evaluation. Although there are rela-
tively few businessmen employing industrial dynamics, while
many employ operations research, this fact could be explained
by the relative ages of the two disciplines.

One basis for judgment might be the practical utility of

77

R. A. Howard

industrial dynamics models. Dynamicists claim that the be-
havior of an industrial dynamics model is usually quite similar
to the behavior of the underlying phenomena. The question
arises: Is mere similarity of behavior enough? Validation based
on similarity alone is rarely convincing. Most observers agree
that industrial dynamics will experience a significant increase
in stature when its models have undergone irrefutable valida-
tion.

The main difficulty is. of course, that similarity can be mis-
leading. In building industrial dynamics models, it may be
fruitful to liken the industrial system to a complex fire-control
system, but we must never forget in examining system re-
sponses that the measures of effectiveness for the two systems
are quite different. For example, control engineers rightly
maintain that oscillating systems are generally unsatisfactory.
Suppose that we examined a company's past history and found
a large oscillation over time in the amount of raw material
purchased. Is this necessarily evidence of mismanagement on
the part of the company or at least an indication that a better
control system is necessary? Of course, that may be the case,
but we cannot pass judgment until we examine the costs as-
sociated with this oscillation. Suppose we should find, for ex-
ample, that the times at which raw-materials purchases were
high were also the times at which raw-materials prices were low,
and vice versa. Then we would praise rather than condemn
management's control actions.

This example is rather extreme, but it does point out that
the valuation of system responses is not a trivia] problem. It
is not sufficient to examine the behavior of industrial processes
in the same offhand way that some people claim is appropriate
for dealing with minimization of mean-squared error in physical
systems. The result of fluctuations in each variable on the
measures of effectiveness must be considered if the control sys-
tem is to be satisfactory from an over-all point of view. It is
the extremely perverse nature of business cost functions that
makes perilous the naive extension of standard servomechanism
theory to the theory of the firm.

78

Managerial Decision Making

A basic difficulty that dynamicists must face in building their
models is that the rationality of decisions in the firm is often
not determinable even if it exists. The input necessary for an
adequate industrial dynamics model normally would require
many man-years of work if the model were to be an accurate
description of a firm's operations. Yet such careful study is
necessary if the dynamicist is to make quantitative rather
than qualitative statements about a firm's behavior. This is
especially true if his efforts are to be compared with those of
a rather unsophisticated observer of the firm's past history. For
example, it does not require a digital computer to tell that if
a company holds a major sale at a time when demand would
be high in any event, then it is only going to increase the
fluctuations in its sales pattern. The dynamicist should be
judged not on his ability to make an absolute improvement
in a firm's control but rather on his ability to improve the
firm's operations beyond the improvement that could be
achieved by having the operations subjected to the review of
any careful observer.

On the other hand the analyst, in spite of his relatively
mature field, still has a number of extremely difficult problems
to solve. In many areas the relatively simple problems have
long since disappeared. The analyst must look for second-order
rather than first-order effects, and he finds that the tools to be
employed must increase correspondingly in power. For example,
the problem of controlling inventor}- is one of the most com-
plex control problems found in any environment. The cost
structure of the inventor}- system is complicated, the number
of state variables necessary to describe the system is large, and
there is a high degree of interaction between different parts of
the system. Professional control engineers readily admit that
the general inventor}- control problem is one of the most diffi-
cult control problems they have faced in any context, bar none.
In the inventor}- control area it often happens that we can
present the decision maker with the values of all variables on
which a decision regarding inventor}- must be based and yet
we cannot tell him how the decision should be made. Indeed,,

79

/. W. Forrester

analytic solutions exist to relatively few problems in the area
of inventor}' control.

Both the businessman and the dynamicist will be hampered
in synthesizing new industrial control systems until the science
of the analyst has been developed more thoroughly. The dy-
namicist may provide a real service to the analyst by pointing
out the areas of an enterprise that most critically require study.

The Future. Scientists have gained an entree into the world
of the businessman through the field of operations research.
The businessman has generally, but not always, found that the
logical, methodical approach of science to his problems can
increase his profits and his managerial control. Although the
continued growth of the field of operations research seems as-
sured, it is too early to tell whether industrial dynamics will
be a sideline or mainstream of this growth. The main con-
tribution of the dynamicists to date has been to attract atten-
tion to the higher-level decision problems of the enterprise. The
importance of these problems assures that any approach which
offers a deeper understanding of them will flourish. Only time
will tell whether this will be operations research, industrial
dynamics, or some new discipline of the future.

Forrester. I made no effort in what I said this evening to
give a balanced picture. Had I tried to give a balanced picture,
I should not have conveyed with enough clarity the points I
was trying to make. The balance provided by the two discus-
sants certainly is appropriate.

I should like to clarify my views on invention, which Holt
used to argue for the necessity of statistical analysis. I said that
there clearly was agreement on certain of the cause-and-effect
relationships that control the process of invention, even if there
was not agreement on how, in fact, it works. We agree that
if we had more people, more budget, more research institutions,
more quality in our people, by whatever standards we use, then
the results would be better. Otherwise, we should not need
M.I.T.; we should not have the large military research budget
that we do; and one research man would contribute as much

SO

Managerial Decision Making

to technological advance as a multitude. This is what Holt does
not believe without a statistical analysis.

Holt. One word in rebuttal: Einstein.

Forrester. No, that is exactly what I mean. Einstein was
not by himself. There was a vast structure of technology sur-
rounding him, and this is exactly the point.

We do have a very large amount of knowledge and under-
standing for which there are no statistics in the formal sense.
I did not say that one should not use statistics when one has
them. I merely said that a very small percentage of our knowl-
edge exists in this form. If we limit ourselves to information
in statistical form, this does not give us enough to go on. My
point was to call attention to the vast amount of information
available in other forms.

General Discussion

Lucas. Dr. Forrester, are you trying to develop a set of
standard criteria that you could apply, not only to Mr. Sprague's
industry, but to other industries as well?

Forrester. The answer to your question is, "No." There
is no such thing as the model. There even is no such thing as
a model that persists for as much as a week. You build a model
to answer a question. You may build several models to answer
several questions. In the airplane industry, you may build a
wing-loading model to test with sandbags, a model to put in
the wind tunnel, and probably many more. The issue of first
importance is a hypothesis of what you think matters; then
you develop a model to check the hypothesis. This is really an
extension of what we necessarily do all the time in the manage-
ment of an organization.

McLanahan. I am wondering if anything happens to your
model as changes occur in the key personnel from whom you
obtained your material, or in the way that they make decisions.

Forrester. Those of you who know a variety of business
organizations well may characterize them as being different in

81

General Discussion

various ways. There is the "successful" company; there is the
"unsuccessful" one. One company has a high level of integrity,
another has not. These are differences that do not come and
go with the change of just a person or two. These character-
istics are persistent. They partly determine how the people
were selected in the first place. The kinds of people who are
hired and the kinds who leave depend on the organization.
Therefore, I should say that the types of things we are talking
about are a great deal more stable than the coming and going
of individual people.

Shedd. Dr. Forrester, I was wondering if you have been able
to use the model of the Sprague operation to control employ-
ment in the company.

Forrester. I cannot cite at this stage of the game docu-
mentary evidence of what has been achieved. Changes are still
being made, and information flows are being set up. However,
the situation is not so incomplete as perhaps has been implied.
The procedures that we use are on a very sound basis in other
professions. The pilot plant of a new chemical plant, for ex-
ample, is essentially a hypothesis that a certain configuration
will work. You build a model of it to see whether it will work,
and if it does, you build the main plant along the same lines.
This is an experimental procedure that is philosophically the
same as ours. There is nothing new about the philosophy of
the approach. The only thing new is the area to which it is
being applied. The fact that it is a more nonlinear, less well-
explored area makes it no more difficult as a next step than
some of the earlier steps once seemed. Our knowledge goes up
exponentially with time, and the next step always looks like a
big one. I believe that we are ready to take this step in the
next decade or two.

Holt. I should like to question Professor Forrester's going
back and forth so facilely between a chemical pilot plant, a
wind tunnel, and a simulation model for a computer. As we
all know, one of the beauties of a computer is its flexibility. You
can program on it practically any process involving informa-
tion handling. Now, the wind tunnel and the pilot plant are

82

Managerial Decision Making

governed by the same laws of nature that govern, respectively,
the airplane and the full-scale pilot plant. We don't have to
ask whether Mother Nature is really applying the same rules
or physical laws to the model phenomena which she is to the
full-scale phenomena. On the other hand, when we go into a
business organization and observe its structure and come home
and build a model with our bare hands, there is a very real
question as to whether we have brought home the baby. I
think the problem of validating the model here is extremely
critical in the effective use of this sort of approach.

I should particularly like to emphasize at M.I.T. the differ-
ence between tests in the real world and tests on an abstract
model because I think that engineering education often is
deficient in drawing this kind of distinction clearly. Engineers
are expert at using scientific theories in solving practical prob-
lems, but they learn relatively little about the scientific method
that yields the theories. They are sensitive to the problems of
finding workable approximations to theory when they want to
bring it to bear on a problem, but they are not sensitive to the
fact that the theory itself is at best an approximation to reality
— a hypothetical model subject to empirical verification. In a
well-developed field like electrical engineering, it may not be
too serious to fail to draw a clear distinction between the re-
lationships in the real world and the relationships incorporated
in Maxwell's equations, a hypothetical model. However, in a
field where our models are still shaky, it leads to no end of
mischief to confuse models with reality.

Anonymous. An earlier question was addressed to you
[Forrester] about the effect of changes in management per-
sonnel. In answering, you referred to the relative stability of
both successful and unsuccessful companies. If an unsuccessful
company was wise enough to employ your approach, and ar-
rived at what we hope is the right answer, then how does the
company implement it in light of the stability of manage-
ment's viewpoint?

Forrester. If it is a poor enough company and stable
enough, it will not try the approach. It seems to me that one

83

General Discussion

of the reasons for being poor is not so much an insistence on
following traditional policies as it is a lack of understanding
of what various alternative policies will do. It is the innovating
companies, willing to experiment, who will be most interested
in the approach.

Anonymous. Do you [Forrester] expect that the actual
results obtained by a company which has altered its procedures
based on the theoretical results of a simulation will be com-
parable with the theoretical results?

Forrester. In general we are shooting for improvement so
great that there would be no doubt about there having been
an improvement. If you can cut down the employment fluctua-
tion by a factor of 3 or 4, this will be fairly conspicuous. Of
course, no one could prove that the improvement was attrib-
utable to the changes that had been made. But after the
changes were tried successfully a few times, the evidence would
be strong enough to be persuasive.

In terms of money, we are aiming at gains comparable with
the present profit margin of the company. For example, in one
company familiar to me, the officers began to wonder about
some of the things that we have been discussing. In reviewing
historical data, they found that in a particular year production
had varied by 400 per cent from minimum to peak. During
that same year retail sales fluctuated by only 30 per cent. Their
estimate of the avoidable costs had they followed different
policies — of the costs avoidable had they not stopped and
started production lines, had they not transferred people, or
laid them off and later rehired them, had they not had to let
things obsolesce in the warehouse — their estimate of possible
avoidable costs for the year was some $75,000,000. This was
about the profit that they made that year.

Howard. Jay [Forrester], it has always seemed to me, pri-
marily because industrial dynamics is such a young field, that
it presents an insubstantial target at the present time. It does
not appear possible to pin you down and make you validate
your assertions. We are used to seeing in the sciences one
curve labeled "predicted" and another labeled "observed."

84

Managerial Decision Making

These curves allow us to make evaluations such as "This is
good" or "This is not so good." Is there any reason in prin-
ciple, assuming that you have a good model of the decision-
making process, why you cannot take actual sales, production,
and inventory data, use your model to obtain "predicted" sales,
production, and inventory figures for the corresponding period,
and make a comparison?

Forrester. Yes, there is a reason why you cannot. These
decisions are all noisy. Suppose you take two models, absolutely
identical in structure and parameters, but both having noise
components in their decision mechanisms. If you start these
models from identical initial conditions and let them run, their
behaviors will diverge so quickly that there is no way of pre-
dicting what will happen on a specific day. Yet the two models
will exhibit similar qualitative performance characteristics. They
will both be stable or unstable, for example. And should you
change the design of one of these models in the presence of the
same noise as before, it will be better or worse by virtue of the
induced change in its qualitative characteristics.

Thus, one must predict, not the particular event, not the
shape of the particular time history, but one must predict the
change in the performance characteristics: profitability, employ-
ment stability, and characteristics such as these. The test you
[Howard] suggest of comparing a particular time history with
the output of a model is not a test that you can expect to use,
although it is a test that many people have been attempting
for many economic models. I think we can demonstrate that
this test, in fact, cannot be met. And it is not necessary to
meet it to have tremendously useful results.

Howard. But I think that you have to have some quanti-
tative measure of how good your model is. It may be that you
should make several replications and then apply a statistical
analysis, as Holt suggested. In other words, how can we possibly
criticize you when you say, "It has the same qualitative be-
havior"? We both look at the same simulated history, and I
say it does not look at all like the real thing, and you say it does.
You say that you cannot with your model duplicate the actual

85

General Discussion

sales data because of the noise in the system. All you can do is
get a signal that has the same characteristics as the actual data.
I say that this statement has no content. What you might view
as the same characteristics, I might not. How can we get a
quantitative agreement on what constitutes the same char-
acteristics?

Forrester. I think that this is a matter of judgment. The
man who is operating the company has got to make a decision.
He is doing it now. If you disagree with the decision, then
your judgment about the issue is different from mine. And
if it is your company, you had better use your judgment. If
it is my company, I shall use mine.

This is really an academic point, though, because I have yet
to see the difference arise with the working people, including
managers and analysts. This is a very troublesome question in
the abstract, and yet in the actual specific case it is not an-
swered in the rigorous objective sense that you speak of; neither
is it in any of our real-life activities. I think you are trying for
something here that we do not have in other areas of human
endeavor. We do not have it in medicine or law or engineering.
You are trying for something here that is more nearly perfect,
more objective than in fact we know how to do anywhere else.
I do not disagree with the desirability of it. I say we do not
have it, and we are not read}- for it. W "here we seem to have
it in certain of the statistical model tests, I believe it is mis-
leading and on an essentially unsound foundation.

Holt. It is interesting to contrast Professor Forrester's will-
ingness in model formulation to quantify such unstructured
concepts as "integrity" with his unwillingness in model testing
to accept quantitative tests of the models. Even where quanti-
tative data are available for such variables as employment fluc-
tuations both from the company and from the model, he ac-
cepts qualitative judgments on similarity as perfectly adequate.

Ardex. In technical investigation a researcher often has
more than one motive. He may desire to solve a particular
problem, but he also may wish to find better methods for
solving a class of problems. I wonder if Professor Forrester will

86

Managerial Decision Making

comment on what insights he may have gained, not to solu-
tions of particular problems, but to methods for solving these
problems. I know, in one sense, that applying a computer to
a problem is a method; but I am really looking for insights a
little beyond that.

Forrester. The field that we are talking about occupies a
position in the social sciences not unlike the position of en-
gineering between the physical sciences and society. We do not
teach at M.I.T. how to be an inspired designer; we do not ar-
range it so that everybody who graduates from here creates the
best engineering system or the best communication system.
What we do is to provide the student with some tools and
approaches and attitudes that he can apply. The rest is up to
him. Some people are just plain better at applying than others.
This point relates to what I said a while ago. We are now at
a stage where the procedures that you are asking for lie es-
sentially in the art and judgment area.

What I am doing is retreating into the same refuge that the
manager often does. We tell the manager that we are going to
understand his decision policies and how he runs his business.
And he says, "Oh no you aren't. It is art and judgment." You
ask me how we determine the characteristics of these systems,
and I tell you it is art and judgment. But this is really where
engineering stands. This is where medicine stands. I can dis-
cuss with you what the art is; I can discuss the approaches with
you; but I cannot give you an objective set of rules that any-
body can go out and apply step by step to come up with the
right answer. I cannot do it.

Holt. It seems to me significant that Jay [Forrester] has
selected medicine and engineering as his examples, not physics
and chemistry.

Forrester. No, they are not pertinent.

Holt. [To Forrester] The point is that at times you have
been very critical of the strivings of management science to-
ward a scientific body of knowledge. There is a real distinction
between the kind of professional problem solving that we as-
sociate with engineering, where the object is to get the job

87

General Discussion

done, and the kind we associate with fundamental scientific
research, where the object is to obtain knowledge. It seems to
me that you have directed yourself more toward the question
of solving practical problems and less toward the question of
obtaining knowledge.

If your aim is simply to improve an operation, how do you
decide how complicated to make your model? In your example
you had thirty decision centers of which five were important.
Suppose that an operations researcher, after careful study, set
up a mathematical model that incorporated the five important
decision centers and was able to solve this model analytically.
Would you still argue in this case that the industrial dynamics
approach is superior to the operations research approach?

Forrester. Generally you find out which five are important
only after you have considered all the ones that seem to be op-
erating in the system. This information unfolds as you observe
the process. It is not obvious in the first place.

Holt. You do not think that it is possible to spot the five
critical decision centers at the start?

Forrester. In retrospect, yes, for the example I cited, and
perhaps for similar examples. But in general I should be willing
to assert that it is not possible.

Howard. Professor Forrester s procedure reminds me of
Galileo trying to discover the laws of falling bodies by throw-
ing up feathers and pillows and everything else, and measuring
each one individually, instead of sitting down and thinking,
"Well, maybe if I use an inclined plane I can slow down this
action, and if I standardize the shape of the body I can refine
my analysis." I rebel at the shotgun approach. Whenever some-
one says, "This is a 400-variable analysis-of-variance problem,
and this is a 225-variable linear programming problem/' my
first reaction is, "What's the matter, couldn't you understand
the process?"

Holt. If you have some good empirical tests, you can try
the 5-variable model to see if it fits. If it does not, then you can
try another model. There is a lot to be said for starting off with
simplicity and adding complexity, instead of vice versa.

88

Managerial Decision Making

Modigliani. I am still disturbed about this question of
testing, because it seems impossible to pin you [Forrester]
down to some definite position. As I understand what you say,
the test of whether a model is reasonable is whether I think it
is reasonable. I completely agree that you would not want to
test the model in terms of its behaving in exactly the same way
as the real system. I also agree with the idea of using certain
qualitative features of the situation such as frequency, ampli-
tude, and lag. But it seems to me essential that we do worry
about finding some criteria for pinning ourselves down. Some-
day we want to have a common ground for agreeing that a
model is or is not reasonable.

Forrester. Anything qualitative certainly has quantitative
measures. What are the qualitative characteristics? Things like
periods, intervals between peaks, phase relationships, turning
points, and the relative succession of peaks. All of these quali-
tative characteristics most certainly have a quantitative basis.
And these are the characteristics that we use in comparing two
model runs, just as you might compare a model run with real
life. You compare it on the basis of amplitudes and periods
and the tendency for disturbances to die out or to be amplified.

But the distinction I was making is between the qualitative
tests, which indeed have quantitative bases, and the test that
a lot of people seem to expect: the test of whether a model can
predict a specific future event. I assert that this is not a really
significant and worthwhile test. Any economic model that I
can conceive of at this stage will not predict whether in July
1965 we are going to be at the top or bottom of an economic
cycle. This is what I mean by an inability to predict a specific
event.

Howard. You say that we cannot predict a specific event
primarily because of the noise that is introduced into the sys-
tem at the various decision points. But isn't this noise known
a posteriori? Don't we know what the actual outside variables
were at the particular time?

Forrester. No, these are buried throughout all the decision
points.

89

General Discussion

Howard. But you could design an experiment in which you
took a detailed data account of all information flows within
the company and set the parameters in your model according
to the actual noise values observed in these flows. Then you
could see if you were able to predict the over-all behavior of
the company. This would be analogous to checking out a com-
puter program containing random elements by fixing these
random elements to deterministic values.

Forrester. Oh, I should not want to assert that your sug-
gestion is impossible. But I think you are calling for a volume
of information that we cannot expect to be available. For ex-
ample, in one company's data we found periods when produc-
tion exceeded sales, yet inventory was falling. This is impossible.
But this is the kind of information that usually is available.
I certainly cannot say that your suggestion is impossible, but
it requires a job that is tremendously expensive and over-
whelming in its implications; and I do not think that it will
contribute anything that is essential.

Holt. Our discussion tonight seems to have pointed up a
clear difference in approach. Industrial dynamics seeks to de-
velop models that can be used to generate, test, and sell im-
provements in decision policies; the models are good enough
if management will accept them. Operations research and man-
agement science share these practical objectives (although the
latter often takes a longer view in seeking practical payoffs),
but in addition they aspire gradually to accumulate a body of
tested knowledge. Hence, they are more willing than is For-
rester to accept the discipline of the scientific method. They
accept at least in principle the search for objective empirical
tests of their models as an integral part of their research process.

To be sure, industrial dynamics as expounded tonight also
seeks knowledge; but the art of improved practice comes first,
the understanding and theory come later. This process will
"work," of course, but there is a question of relative efficiency.
In my view, if scientific research leads to reliable knowledge,
improved practice will follow soon enough and progress will
be faster than if we concentrate our energies on the short-term

90

Managerial Decision Making

payoff. And there is a danger to a new and immature field in
using as the final test "acceptability to the manager." If a rec-
ommended policy fails in practice, the manager loses his profit,
but industrial dynamics loses reputation. The latter may be far
more difficult to repair.

The reluctance of industrial dynamicists to face the valida-
tion problem is understandable enough; for we know very little
about testing complex nonlinear hypotheses. However, this
shortcoming in our methodological tool bag is not apt to be
remedied by skirting the issues, and one might hope that in-
dustrial dynamicists could be influenced seriously to concern
themselves with these problems.

But even if they do not, there are plenty of problems to go
around. There is little to be gained by methodological bickering
and criticism, since it is clearly understood that the two ap-
proaches are quite different and each has merit. Either way,
we cannot fail to learn more than we now know.

Reference

1. Forrester, J. W., Industrial Dynamics, The M.I.T. Press and John
Wiley & Sons, Inc., New York, 1961. Portions of Professor Forrester's
talk are based on material appearing in this book.

91

3

Simulation
of
Human
Thinking

3

Simulation

of

nurnan

I miming

Speaker

- - —

Herbert A. Simon

Professor of Administration and Associate Dean

Graduate School of Industrial Administration

Carnegie Institute of Technology

Coauthor

Allen Newell

Institute Professor of Systems and

Communications Sciences

Carnegie Institute of Technology

(Also Consultant, the Rand Corporation)

Discussants

Marvin L. Minsky

Associate Professor of Mathematics

TV jT ^ _ T_ _ i_ j_ t i*i t C r | i 1 _

Massachusetts Institute or I echnology

George A. Miller

Professor of Psychology

Harvard University

Moderator

Sidney S. Alexander

Professor of Industrial Management

Massachusetts Institute of Technology

94

Simon. The use of computers to simulate human thinking
has a prehistory and a history, as well as a present and a future.
During the prehistorical period, prior to World War II, there
were no computers, in the modern sense, but there were a num-
ber of successful attempts to construct teleological mechanisms
— analog devices that simulated one aspect or another of an
organism's adaptive behavior in relation to its environment.1

History begins in earnest, however, with the rapid growth of
servomechanism theory during World War II and with the
appearance of the first stored-program digital computers, two
of the three legs on which Professor Wiener's cybernetics
stands. Grey Walter's "tortoises" and W. Ross Ashby's Homeo-
stat represent important early progress, as does an analog sim-
ulation of a self-organizing network that Professor Minsky
constructed in 1951.

Computer simulation had already begun to take definite form
as a field of research by the time of the well-known session on
learning machines at the 1955 Western Joint Computer Con-
ference, a session in which Professor Miller also participated.2
At that session, Clark and Farley of Lincoln Laboratories de-
scribed a computer simulation of a self-organizing "nerve net"

95

H. A. Simon

system; Selfridge and Dinneen, also of Lincoln Laboratories, de-
scribed a pattern recognition program; and one of the authors,
Newell of the Rand Corporation, outlined a program for a
chess-playing machine.

One of the discussants at that session, Walter Pitts, observed
that there were two main lines of attack represented: the first
taking as its point of departure some features of the human
nervous system and sensory apparatus, the second, the organi-
zation of symbolic processes to perform complex thinking
tasks. As Mr. Pitts put it:3

The speakers this morning are all imitators in the sense that
the poet in Aristotle "imitates" life. But, whereas Messrs. Farley,
Clark, Selfridge, and Dinneen are imitating the nervous system,
Mr. Newell prefers to imitate the hierarchy of final causes tra-
ditionally called the mind. It will come to the same thing in the
end, no doubt. . . .

Most workers in this field continue to believe that it will
come to the same thing, but the end is not yet, and these two
main strands of research are still clearly discernible in the work
going on at the present time. Our remarks this evening will be
concerned almost exclusively with the second, with the imita-
tion of mind. This strand has already begun to make contact
with important potential areas of application, business admin-
istration and teaching among them.4 Our purpose tonight, how-
ever, is not to speculate about applications. We shall be spec-
ulative enough, we are sure, for your tastes; but we shall
speculate about the form that fundamental theory in this field
is taking, rather than about the implications of that funda-
mental theory for everyday affairs.

The Proof of Possibility

With this decade of history and several decades of prehistory
behind us, it is no longer necessary to argue that computers can
be used to simulate human thinking, or to explain in general
terms how such simulation can be carried out. A dozen or more
computer programs have been written and tested that perform

96

Simulation of Human Thinking

some of the, interesting symbol-manipulating, problem-solving
tasks that humans can perform, and that do so in a manner
which simulates, at least in some general respects, the way in
which humans do these tasks. Computer programs now play
chess and checkers, find proofs for theorems in geometry and
logic, compose music, balance assembly lines, design electric
motors and generators, memorize nonsense syllables, form
concepts, and learn to read.5

With the proof of possibility accomplished, we can turn to
more substantive questions. We can ask what we have learned
about human thinking and problem solving through computer
simulation: to what extent we now have theories for these
phenomena, and what the content of these theories is. Since
we want to talk about these substantive matters, we shall sim-
ply make the following assertions, which are validated by exist-
ing computer programs.

1. Computers are quite general symbol-manipulating devices
that can be programmed to perform nonnumerical as well as
numerical symbol manipulation.

2. Computer programs can be written that use nonnumerical
symbol manipulating processes to perform tasks which, in hu-
mans, require thinking and learning.

3. These programs can be regarded as theories, in a com-
pletely literal sense, of the corresponding human processes.
These theories are testable in a number of ways: among them,
by comparing the symbolic behavior of a computer so pro-
grammed with the symbolic behavior of a human subject when
both are performing the same problem-solving or thinking tasks.

The General Problem Solver

The theory we shall have most to say about is a computer
program called the General Problem Solver. It is not "general"
in the sense that it will solve, or even try to solve, all problems;
it obviously won't. It is called "general" because it will accept
as tasks all problems that can be put in a specified, but fairly
general, form, and because the methods it employs make no

97

H. A. Simon

specific reference to the subject matter of the particular problem
it is solving. The General Problem Solver is a system of meth-
ods—believed to be those commonly possessed by intelligent
college students — that turn out to be helpful in many situa-
tions where a person confronts problems for which he does not
possess special methods of attack.

Before general methods can be applied to any particular class
of problems, of course, the problem solver must also learn, or
be taught, the rules that apply to that particular problem do-
main. The General Problem Solver will not prove theorems
unless instructed in the rules of proof in the particular branch
of mathematics to which the theorems belong. Thus, in any
particular problem domain, the resources available to the Gen-
eral Problem Solver include information about the task en-
vironment as well as its own repertory of methods.

Missionaries and Cannibals

Let us introduce the General Problem Solver (which we shall
call GPS) by means of a simple example. Many of you are
familiar with the puzzle of the missionaries and cannibals, and
some of you saw a young lady solving the puzzle in a recent
CBS television program celebrating M.I.T/s centenary. There
are three missionaries and three cannibals on the bank of a
wide river, wanting to cross. There is a boat on the bank, which
will hold no more than two persons, and all six members of
the party know how to paddle it. The only real difficulty is that
the cannibals are partial to a diet of missionaries. If, even for a
moment, one or more missionaries are left alone with a larger
number of cannibals, the missionaries will be eaten. The prob-
lem is to find a sequence of boat trips that will get the entire
party safely across the river, without the loss of any missionaries.

Suppose, now, that we encountered this puzzle for the first
time. We are endowed by nature and nurture with certain
abilities that enable us to tackle the problem. We might or
might not solve it, but we could at least think about it. In what
would this thinking consist? In particular, how could we bring
to bear our general problem-solving skills, which make no ref-

98

Simulation of Human Thinking

erence to missionaries and cannibals, on this particular situa-
tion?

Clearly, we have to form some kind of abstraction of the
problem that will match the abstractness of our general meth-
ods: We have some people and a boat on this side of the river
and we want them on that side of the river. Stated abstractly,
we have a certain state of affairs, and we want a different state
of affairs. Moreover, we can describe both states, and we can
also describe what the differences are between them, between
what we have and what we want.

In this case, the differences between the given and the de-
sired are differences in physical location. Our men are on one
side of the river; we want them on the other. But we have had
vast experience with differences in location, and that experi-
ence (stored somehow in memory) tells us that boats are use-
ful devices for reducing differences of location on water. So
we begin to consider the possible sequences of boatloads that
will get our party across the river without casualties.

It is clear from this formulation of the problem what part is
played in its solution by our general problem-solving techniques
and what part by our knowledge and experience of the partic-
ular problem domain in question. A general solution technique
is to characterize the given and desired situations, to find the
differences between them, and to search for means — imple-
ments or operators — that are relevant to removing differences
of these kinds. Our knowledge of the task and our experience
tell us what the given and desired situations are, and what
kinds of operators may be relevant for getting from here to
there.

Structure of GPS

We can now characterize the program of the General Prob-
lem Solver more formally.6 The program deals with symbolic
objects that describe or characterize situations: the given situa-
tion, the desired situation, various intermediate possible situa-
tions. The program also deals with symbols representing dif-
ferences between pairs of objects, and with symbols represent-

99

H. A. Simon

ing operators that are capable of inducing changes in the ob-
jects to which they are applied. (Table 1, first column.)

Goal Types. The processes of GPS are organized around
goals of three types:

1. Transformation goals, to transform object a into object b.

2. Difference Reduction goals, to eliminate or reduce differ-
ence d between objects a and b.

3. Operator Application goals, to apply operator q to ob-
ject a.

Table 1

Comparison of Basic Categories in GPS,
Speech Learning, and Organismic Adaptation

GPS

Learning
Speech

Adapting
to

Environment

objects

perceptual symbols
(audited phonemes)

AFFERENT

state
language

differences

comparison between
adult-child
phoneme images

relevant
operators

changes in
motor symbols
(control of
speech production)

EFFERENT

process
language

Methods. With each type of goal in GPS there is associated
one or more methods, or processes, that may contribute to the
attainment of the goal. The principal methods in the present
version of GPS are three in number, one for each type of goal:

1. Method for transformation goals, to transform a into b:

(a ) Notice a difference d between a and b.

(b) Establish the goal of reducing d between a and b.

( c ) Try to attain this new goal.

(d) If successful, find a new difference and repeat.

100

Simulation of Human Thinking

2. Method for difference reduction goals, to reduce d be-
tween a and b :

(a ) Recall an operator q that is relevant to differences of
the type of d.

(b) Establish the goal of applying q to a.

(c ) Try to attain this new goal.

(d) If successful, return to the previous transform goal.

3. Method for operator application goals, to apply operator
q to a:

(a) Compare conditions for application of q with ob-
ject a.

(b) If these are not satisfied, establish and try to attain
the goal of transforming a into an object that meets
these conditions.

(c) When the conditions are satisfied, apply q to a, and
return to the previous difference reduction goal with
the modified object a'.

This is a rather simplified description of what goes on in
GPS, but it gives the broad outline of the program. GPS, to
put it simply, is a program that reasons about ends and means.
It is capable of defining ends, seeking means to attain them,
and, in the process of so doing, defining new subsidiary ends, or
subgoals, to the original end.

As a theory of human problem solving, GPS asserts that
college students solve problems — at least problems of the sorts
for which the program has been tested — by carrying out this
kind of organized means-end analysis. It does not assert that
the process is carried out consciously — it is easy to show that
many steps in the problem-solving process do not reach con-
scious awareness. Nor does the theory assert that the process
will appear particularly orderly to an observer who does not
know the program detail or, for that matter, to the problem
solver himself. It does assert that if we compare that part of
the human subject's problem-solving behavior which we can
observe — the steps he takes, his verbalizations — with the
processes carried out by the computer, they will be substantially
the same.

101

H. A. Simon

Abstracting and Planning Processes. Before we leave this
description of GPS, we should like to mention one other kind
of process that we are incorporating in the program, and that
certainly must be included if we are to explain and predict the
behavior of our subjects, particularly the brighter ones. We
call these additional methods abstracting and planning proc-
esses. Briefly, abstracting consists in replacing the objects, the
differences, and the operators, with new symbolic expressions
that describe the situation in much more general terms, omit-
ting the detail.7 For example, we might ask GPS to prove a
trigonometric identity:

cos2 x + sin2 x = tan x cot x

Here, GPS might take as a the expression "cos2 x + sin2 x"
and as b the expression "tan x cot x." In using the planning
method, these might be abstracted to a! "an expression con-
taining cos and sin" and b' "an expression containing tan
and cot," respectively. Then, the methods of GPS could be
applied to transforming the abstracted given object a' into the
abstracted desired object b\ If this goal were attained, the steps
employed for this transformation would generally provide a
plan for transforming the original, detailed given object a into
the original desired object b. In the particular case illustrated,
the plan might be something like: "First eliminate cos and sin
from the expression, and then introduce tan and cot."

The Generality of Means-End Analysis

The processes incorporated in GPS have actually been ob-
served in the behavior of our human subjects solving problems
in the laboratory. By analyzing the tape-recorded protocols of
their problem-solving efforts, we can identify the occurrences
of the three goal types and the four methods. Moreover, the
augmented GPS, containing the planning method, incorporates
a substantially adequate set of processes to explain our sub-
jects' behavior in some of these simple theorem-proving, puzzle-
solving situations.8 By the adequacy of GPS, we mean two
things:

J 02

Simulation of Human Thinking

1. We do not find in the subjects' protocols evidences of
processes quite different from those postulated in GPS.
This may mean only that we don't know how to look for
them.

2. But when we have compared the trace of the GPS com-
puter (or hand simulations of the computer program)
with the protocols of a subject solving the same problem,
we have found that the two often follow the same path —
noticing the same things about the problem expressions,
establishing the same subgoals, applying the same opera-
tors, running down the same blind alleys — over periods
of time ranging up to several minutes. That is to say, the
processes in GPS are sufficient to produce a stream of be-
havior in a given problem situation quite similar to that
produced by the human subject.

These kinds of tests, even if broadened, would still not say
much about the generality of GPS as a theory of human think-
ing and problem solving. It might turn out that if we examined
tasks quite different from those used in developing the pro-
gram, and made the same careful records of subjects' protocols,
we should find many new processes exhibited that are not con-
tained in GPS. However, extensions of GPS in fair detail to
problem domains that were not considered when the program
was developed indicate that its processes are adequate at least
to these other domains. For example, the problem of the mis-
sionaries and cannibals, which was first suggested as a possible
task by Mr. Thomas Wolf of the Columbia Broadcasting Sys-
tem, has been solved by the current version of GPS, not with-
out some reorganization of the program but without addition
of new goal types or methods. Similarly, the applications to
algebraic and trigonometric identities and to certain learning
tasks appear to require no enlargement of the basic repertory
of methods. Less detailed analysis of a variety of other tasks
shows GPS to be adequate for these also.

Still, these additional tests do not carry GPS beyond a fairly
limited range of formal problem-solving situations. It would

103

H. A. Simon

be of considerable interest to explore, even qualitatively, the
powers and limitations of GPS when it is confronted with a
thinking or learning task of quite a different kind from any of
these. We should like now to carry out a reconnaissance along
these lines. First, we shall describe, on the basis of what is now
known, the processes that humans use in a task that appears,
superficially, to be quite different from problem solving. Then,
we shall propose a framework which shows that these processes
can be subsumed under those already incorporated in the Gen-
eral Problem Solver. The particular task we shall examine was
chosen because quite a bit is known about it, and because it
wall allow us to call on our discussants for a maximum of as-
sistance this evening.

The Acquisition of Speech

There are many human activities to which we should apply
the term "thinking" but not the term "problem solving." There
are also many activities we should usually call "learning" rather
than "thinking." We should ordinarily call a child's acquisition
of speech, "learning." We propose to consider the acquisition
of speech as an example of human cognitive activity that is at
something of an opposite pole from the rather highly verbal-
ized, somewhat conscious, practiced problem solving of an in-
telligent and educated adult. We can then judge whether the
processes at these two poles are quite different or basically the
same.

Speech acquisition has been about as well studied as any non-
laboratory complex human activity, and a review of the litera-
ture indicates that there is general consensus about the par-
ticular facts we shall use.9 If we are wrong in that assumption
or in our interpretations of the facts, Professor Miller is one of
the best-equipped men in the country to put us straight.

Central Representations

We consider an infant who has already learned the names of
a few objects, as evidenced by the fact that he can point to
them or fetch them when they are named by an adult, but

104

Simulation of Human Thinking

who has not yet pronounced their names. From his behavior,
we can infer that when the child perceives the spoken word
"ball," his perception has some kind of internal representation
in the brain that permits it to be associated, through previous
experience, with some internal representation of a visually per-
ceived ball.

To say the word "ball," the child must, in addition, store
some kind of program capable of energizing, through motor
(efferent) channels, the muscles involved in speech production,
in the production of the specific phonemes of that word. Let us
call the "whatever-it-is" in the central nervous system that
represents internally a perceived sensory stimulus an afferent or
perceptual symbol. Let us call the "whatever-it-is" that rep-
resents the program for initiating the motor signals an efferent
or motor symbol.

Learning to speak, in this formulation, means acquiring the
motor symbols that correspond to perceptual (auditory) sym-
bols of words already known, and associating the former with
the latter. Now the difficulty is that there is no way in which
the corresponding perceptual and motor symbols can "re-
semble" each other, can symbolize the appropriateness of their
association by resemblance. The correspondence is purely ar-
bitrary.10 The infant is faced (if he only knew it!) with the im-
mense inductive task of discovering which motor symbols will
cause speech production that, when he hears it, will produce,
in turn, an appropriate auditory symbol to be perceived and
recognized. And the task appears at first blush to have little
structure that would permit it to be approached with some
less arduous technique than trial-and-error search.

There is ample evidence that much trial-and-error search is
indeed required before the infant acquires the skill of speak-
ing. The child imitates the adults around him, and he imitates
himself (echoic speech). Gradually, over many months, he
acquires the motor symbols that enable him to produce sounds
which he hears as the expected auditory symbols. In the early
stages, the child's acquisition of a speaking vocabulary appears
to be paced by the task of developing the new motor symbols.

105

H. A. Simon

At later stages, he is able to produce a word relatively easily
once he has learned to recognize the corresponding auditory
symbol.

Factorization

A little reflection will persuade us that something more than
trial and error is involved. If that were all. the three hundredth
word would be no easier to pronounce than the first. The child
learns to learn. In what does this consist?

Although the motor symbol cannot be compared with the
perceptual symbol, the correct perceptual symbol for a word
can be compared, through imitation, with the perceptual sym-
bol produced by the attempt to pronounce the word. If these
are different, modification of the motor symbol can be at-
tempted until an auditor}- symbol resembling the correct one is
perceived.

Thus far we have been assuming that the units in terms of
which these transactions take place are words. But there is no
reason for this assumption — the child might well attend to
particular syllables, phonemes, or even components of pho-
nemes. The auditor}- symbols for words can be compound sym-
bols or expressions: strings of phonemes, each phoneme itself
encoded in terms of its component frequencies and other char-
acteristics. It is even more plausible to suppose that the motor
symbols would be constructed from smaller units, for each word
involves a temporal succession of syllables, each syllable a tem-
poral succession of phonemes, and each phoneme a whole set of
signals to the several muscles involved in that part of the speech
act. Thus, one of the many components of the motor symbol for
the spoken word "dog" might be the signal that pushes the
tongue against the palate in the initial "d" phoneme of this
one-syllable word.

The Learning Process

There is considerable evidence today that this picture of the
processes of word recognition and word production is correct,
at least in broad outline. Many of the components involved

J 06

Simulation of Human Thinking

in both auditory and motor symbols have been tentatively iden-
tified, and there is good experimental evidence for some of
them.11 But what does the picture, if true, contribute to our
understanding of the child's acquisition of speech?

It means that the inductive learning need not be blind in-
ductive learning — attempting to associate by pure trial and
error each of a large number of words with an appropriate motor
symbol chosen from the myriad of producible sequences of
speech sounds. On the contrary, to the extent that specific fac-
tors in the auditory symbol vary with specific factors in the
motor symbol (e.g., as one of the formant frequencies in vowel
sounds varies with the size of the resonating mouth cavity), the
search for the correct symbol can be very much restricted. Com-
ponents can be corrected on a one-at-a-time basis. For example,
the child trying to pronounce "dog" can at one time attend to
the correctness of the vowel, at another time to the correctness
of the initial consonant, or even to the aspect of the initial
consonant associated with tongue position.

Thus, the hypothesis of factorization is supported both by
experimental evidence that it does take place, and by theoret-
ical reasons why it "should" take place — why speech acquisi-
tion would be very much easier with it than without it. Trial-
and-error acquisition of words without factorization would re-
quire a search, in each instance, for the correct motor symbol
from among tens of thousands of possible symbols. Trial-and-
error acquisition of phonemes would require a search from
among only a few hundred phonemes (much fewer are ac-
tually used, of course, in any single dialect). Trial-and-error
search among phomene components would be even more re-
stricted; there are, for example, probably only a half dozen dis-
tinguishable tongue positions. Thus, by factorization of the
total space of possibilities, a very limited trial-and-error search
of the factors can be substituted for an immense search of the
product space. Moreover, once the child has acquired motor
symbols corresponding to the common phonemes, acquisition
of new words (new combinations of these same phonemes) can
be very rapid.

107

H. A. Simon

Summary: The Child's Acquisition of Speech

Let us now summarize our description, partly factual, partly
hypothetical, of the speech acquisition process. The child ac-
quires perceptual auditor}- symbols corresponding to words he
has heard and has associated with visual symbols. He tries, on a
trial-and-error basis, to produce words, hears his productions,
and compares these auditor}- symbols with those already stored.
When he detects differences, he varies the motor symbol to try
to remove them. As he learns, he detects that changes in cer-
tain components of the motor symbols alter only certain com-
ponents of the auditor}- symbols. Thus he is able to factor the
correction process and thereby accelerate it greatly.

The Acquisition of Speech by GPS

Now it is very easy, with a few changes in vocabulary, to
translate this whole descrpition back in terms of GPS. When
the translation has been made, we shall see that the processes
just described are the methods of GPS.

Let us, in this translation, call the auditor}- symbols objects.
(Table 1, second column.) We assume that there exist central
processes that modify motor symbols, that change one or more
of their components. We shall call these processes operators.
A change in a motor symbol will, in turn, change the auditory
symbol that is perceived when that motor symbol produces a
sound.

The child detects differences between the object he has pro-
duced (i.e., his perception of the sound) and the correct object
(his perception of the sound when produced by adults). He
applies operators to the motor symbol to modify the sounds he
produces, hence the object perceived; and he compares the
latter again with the correct object. This search process con-
tinues until he can reproduce the perceived object.

But this does not account for the factorization, which we
have argued is so crucial to the efficiency of the learning process.
How will GPS learn (1) which differences in objects are as-
sociated with which operators upon the motor symbols, and

108

Simulation of Human Thinking

(2) how to factor objects and operators? Although the answers
to these questions are far from certain, a scheme we have pro-
posed elsewhere would enable GPS to handle these tasks also.12
We shall sketch it briefly:

1. Given a set of differences and a set of operators, GPS can,
with modest amounts of trial and error, detect which op-
erators are relevant to producing or eliminating which
differences. To take a crude, but simple, example: It takes
relatively little trial and error to discover what differences
in the perceived sound are associated with changes in the
rounding of the lips while producing a vowel. The factori-
zation has already largely been carried out by nature, so to
speak, because changes in only a few aspects of the motor
signal will change only a few aspects of the perceptual
symbol.

2. The GPS processes can themselves be employed to dis-
cover inductively a "good" factorization, a "good" set of
differences. To do this, GPS must be supplied with some
very general criteria as to what constitutes such a good set.
The criteria would be of the following sorts:

(a) Only one or a few operators should be relevant to
each difference (so that, given a difference, an ap-
propriate operator can be found without too much
search ) .

(b) Only one or a few differences should be associated
with each operator (so that the sounds produced
can be varied factor by factor) .

and a few others of the same general kind.

With such a set of criteria provided, finding a good set of
differences simply becomes another kind of problem to which
GPS can apply its problem-solving methods. What are the ob-
jects, differences, and operators in terms of which this new
kind of problem is formulated? To avoid unnecessary confu-
sion, we shall capitalize the terms objects, differences, and

109

H. A. Simon

operators in speaking of the new problem context, in order to
distinguish them from the objects (perceptual symbols'), differ-
ences, and operators (changes in motor symbols) involved in
the original task of acquiring speech.

The objects for the new problem-solving task are the sets of
differences in the original task environment. The new differ-
ences designate to what extent particular sets of differences
meet the criteria we have just listed. Operators are processes
for altering the set of differences under consideration by delet-
ing differences from the set, adding differences, or generating
new differences for possible inclusion. GPS then tests in what
respects particular objects (sets of differences) are different
from the desired object (as indicated by the criteria^. It seeks
to remove these differences | modify the set of differences J by
applying operators ( by adding, subtracting, or modifying dif-
ferences ) .

Since this scheme has not been realized on a computer, we
cannot tell how effective GPS would be in handling it. All we
can say is that it is a problem whose solution can be attempted
with the means at the disposal of GPS.

A due respect for parsimony would suggest, then, that in-
stead of postulating quite different processes for the acquisition
of such skills as speaking from those postulated for adult prob-
lem solving, we embrace tentatively the hypothesis that the
processes are in fact the same, that the General Problem Solver
provides a description of both processes. This hypothesis would
provide a sharp focus for empirical research into the early
speech behavior of the child.

The State-Process Dichotomy

Let us accept this hypothesis for the moment: that the same
system of means-end processes is involved m learning speech
and in problem solving. Can we explain why a system of means-
end analysis should provide the basis for adaptive behavior in
both classes of situations? We shall try to provide an explana-
tion for the generality of means-end processes by showing how
these arise quite naturally from the problem that any organism

110

Simulation of Human Thinking

must solve if it is to use its motor and sensory apparatus effec-
tively to survive.

Relation of Perceptual to Motor Symbols

The terms "perceptual" and "motor," or "afferent" and
"efferent," reflect the dual relation that every adaptive or-
ganism has with its environment. It perceives aspects of the
environment, and it acts upon the environment. It must be
able, therefore, to transmit, store, and operate upon internal
representations — perceptual symbols — that stand for its per-
ceptions; and it must be able to transmit, store, and operate
upon internal representations — efferent or motor symbols —
that can serve as signals to its effectors. The organism survives
by associating appropriate motor symbols with the perceptual
symbols that stand for various classes of perceptions.13

In particular, the organism can perceive, at least grossly, its
own behavior caused by its efferent signals. Hence, among the
perceptual symbols that it can store are symbols that stand
for the perception of corresponding motor signals. Languages
are especially adapted to facilitate this correspondence. Lan-
guage behavior, built from limited alphabets of unit behaviors,
is highly stylized so that to each distinct language "act" will
correspond an easily perceivable and distinguishable perceptual
symbol.

Nevertheless, the relation of a particular language efferent,
say, that which energizes the word "dog," to the corresponding
perceptual symbol is arbitrary. There is no more resemblance
between the auditory "dog" and the motor symbol which pro-
duces that word than between "dog" and "Hund." If it is to
be learned, the correspondence must be learned as a pure fact.
By building up a dictionary relating motor with perceptual
symbols, including language-symbols, the organism gains the
ability to produce the actions it "intends." In the last section
we explored how this ability could develop in the case of
speech.

The duality of our relation with the environment reveals it-
self in the vocabulary of natural languages, particularly in the

111

H. A. Simon

distinction between nouns and adjectives, on the one hand, and
verbs, on the other. We have clean clothes (a perceptual sym-
bol) because they have been washed (a motor symbol). It is a
fact stored in our "table of connections" that when we wash
clothes they become clean. As we build up our vocabulary, how-
ever, we pass more readily from the one mode of discourse to
the other. Thus, the clothes, in the last example, might also
have been cleaned. As we learn what actions have what effects,
changes in objects are named by the processes that produce
them, and processes by the effects they create.

The Problem of Translation

It is precisely this duality of language or, more broadly, of
the internal symbols employed in thought that makes behavior
problematic. The world as it is and as it is desired is described
in a state language, a language of perceptual symbols.14 Possible
actions are described in a process language, a language of motor
symbols. (Table 1, third column.) The problem of adapting is
the problem of finding the statement in the process language
that corresponds to the difference between existing and desired
states of affairs in the state language.15

But the problems that GPS was designed to handle can be
viewed in exactly the same way. What is involved in discovering
a proof for the Pythagorean theorem? The theorem is a sym-
bolic object in the state language: "The square on the hypot-
enuse of a right triangle is equal to the sum of the squares on
the sides." By comparing this theorem, so stated, with the
axioms and previously proved theorems, we detect differences
between them. A proof of the theorem is a symbolic object in
the process language. This object, the justification that we gen-
erally write down alongside the successively modified axioms
and theorems, describes the sequence of operations that elimi-
nates the differences between axioms and desired theorem.
Given a set of axioms, for every theorem defined in the state
language, the theorem can be represented in the process lan-
guage by the sequence of operations that constitutes its proof.

Thus mathematics, and problem solving generally, is an imi-

112

Simulation of Human Thinking

tation of life. Problem-solving activity uses the very fundamental
processes that all adaptive organisms must have if they are to
coordinate successfully their perceptual and motor pictures of
the world.16 Ends-means relations, far from being highly spe-
cial, are reflections of the basic state-process dichotomy, the
dichotomy between perceiving and acting.

The Difficulty of the Environment

How hard a problem will be depends on the simplicity or
complexity of the rules that define the correspondence between
the two languages. An example of a relatively simple corre-
spondence is the relation between the decimal and octal repre-
sentations of integers. There is a simple and direct algorithm
that solves all problems of the form: If a is the decimal repre-
sentation of a number, what is its octal representation?

At the other extreme, the correspondence between the vo-
cabularies may be purely conventional or arbitrary. Then rote
learning is the only means for building up the translation dic-
tionary, and if the correct translations must also be discovered,
immense amounts of trial-and-error search may be required.

The aspects of the environment with which we, as organisms,
deal effectively reach neither of these two extremes. The trans-
lation between the state language that describes our percep-
tions of the world and the process language that describes our
actions on the world is reducible to no simple rule, but it is
not, on the other hand, arbitrary. Most of our skill in dealing
with the environment is embodied in elaborate heuristics, or
rules of thumb, that allow us to factor, approximately, the com-
plex perceived world into highly simple components and to
find, approximately and reasonably reliably, the correspondences
that allow us to act on that world predictably. This is the skill
that the adult businessman uses when he makes a decision, the
skill of the scientist in his laboratory, the skill of the subject in
a problem-solving experiment, the skill of a child learning to
speak.

What we have proposed this evening is that at the core of
these heuristics (the portion that is not bound up in special

113

M. L. Minsky

skills) is the organized system of ends-means processes, of state-
process translations, that the General Problem Solver describes.
We have proposed that here, in Mr. Pitts's words, is a first
approximation to "the hierarchy of final causes traditionally
called the mind."

Panel Discussion

Minsky. In discussing the differences that Pitts's quotation
suggests, I shall group artificial intelligence and the simulation
of human thought, and distinguish these research areas from
work on initially less structured models: "self-organizing" mod-
els and "random neural net" models. The first two areas are
very similar. The third area is very different in methods and
goals. It has a physiological orientation and alleges to be based
on an imitation of the brain. Incidentally, I have many reser-
vations about accepting any theory that alleges to be an imita-
tion of the brain, since almost nothing is known about how
the brain forms a concept, recognizes a pattern, and so forth.
I shall say more about this later.

In the case of people (myself included) who are working on
artificial intelligence, we feel free to employ mechanisms that
are not at all lifelike if they will help our programs solve dif-
ficult intellectual problems, such as the problem of the mission-
aries and cannibals. The trouble is that we really have not
encountered very many such mechanisms. If we examine the
machines that play chess, or prove theorems in geometry and
the calculus, it is hard not to be reminded of the approach of
the human game-playing machines. "If he goes there, I shall
go there," and so on. It is not clear that anything new is being
used.

In the case of Simon and other people working on the simu-
lation of human thought processes, I really doubt that they
could long resist the temptation to exploit some new, non-
human technique that would make their machines significantly
more intelligent. I believe that their interest in psychology

114

Simulation of Human Thinking

might play second fiddle to any possibility of attaining more
intelligent machines.

Efforts in artificial intelligence and efforts in simulating hu-
man thought will eventually diverge, particularly in specialized
areas where computers most easily can be made to transcend
human limitations. This is naturally already the case in nu-
merical computation. There are also less-evident areas in which
simulation of human methods would yield distinctly poorer
performance than other kinds of heuristic programming, e.g.,
areas in which the inability of the human to retain and handle
more than a few things at a time is a serious weakness. Pro-
fessor Miller has written on this subject.

In any case we might suppose that the province of the
psychologist will not remain limited to human problem solving
after artificial intelligence has matured. Computers may become
our intellectual successors; their training and management will
not run smoothly. Future psychologists may envy the simple
problems we faced when we had only to deal with human
thought.

It is curious that with all the general interest and publicity
in the field there are only a very few research groups directly
concerned with artificial intelligence. In view of the obvious
romance of the field, why is there so little activity?

For one thing, the technical problems are very difficult. Each
major advance has involved two or three years of work, usually
of several capable people. Ideas that are both new and good
are hard to conceive, formulate, and realize in a performing
program. GPS is a very good idea. The essentials of its de-
scription, when formulated several years ago, fit on a very few
pages; its realization has occupied several man-years of irre-
placeable talent. There also are many bad ideas around, equally
concise, that have consumed even more time, but presumably
of more easily replaceable talent.

It is very difficult to take apparently common-sense proce-
dures, such as those described in Polya's books on practical
heuristics in mathematical problem solving, and put them into
machine-usable form. It requires people with first-class talent

115

M. L. Minsky

both to get the ideas and to formalize them in a satisfactory
form. The rarity of this combination, plus the fact that such
people find other things to do, accounts in part for the limited
activity in the field.

Whenever there is an alternative to the prospect of hard
work, people tend to turn toward it. It seems to me that interest
is being attracted, or diverted, to the apparent alternative which
goes under the name of "self-organizing" systems. A consider-
able number of people are studying large, more-or-less random
network models to find learning or evolutionary capabilities.
Some of these models supposedly are patterned along the lines
of the nervous system; others are more deliberately designed to
evolve adaptive learning. But since essentially nothing is really
known about the physiology of how the nervous system learns
and manipulates concepts, these efforts can be grouped together.
People have obtained certain kinds of simple learning behavior
from such networks. My impression, however, is that there is
still no evidence that the kinds of models these people are con-
sidering today can solve even slightly difficult problems. Today's
results in this area seem little better than those of a decade
ago. Yet I have the impression that, for example, every aircraft
company in the country is sponsoring such a project.

While the results of such work may prove interesting and
useful, people should be aware that there is at present no prom-
ise that self-organizing models constitute a direct approach
either to artificial intelligence or to simulation of human prob-
lem solving. Solving difficult problems takes a great deal of
organization in cases where competely blind trial and error is
out of the question; and this includes almost everything. Even
GPS must be, or must become, considerably more complex
than Simon's brief explanation might make it appear. Simon
outlined the basic heuristic technique of goal-oriented method
selection, but he did not dwell on the rest of the administrative
structure. A complex problem can be solved only by breaking
it into parts, solving these, remembering how they are inter-
related, and synthesizing the partial results. This has been done
in a number of existing programs, and their structure is very

116

Simulation of Human Thinking

complicated. GPS will have to be supplied with an adequate
"administrative" structure.

It is hard to see how such structures could emerge, in a rea-
sonable time, from a system with much less organization. Con-
sider the basic process of "postponement." Once a problem is
split up, we must work one part and save the other (s); the
postponed problems must be retained with information about
their relevancies. There may be several goals, not entirely con-
sistent with one another, and we must establish priorities or
more sophisticated ways of making decisions in case of conflict.
Before a self-organizing system can work on such problems, it
must develop facilities for postponement. This could be done,
perhaps by providing a shrewdly designed training sequence of
problems; but in such a case it would be hard to justify the dis-
tinction between an evolution and a carefully assembled pro-
gram. The programmed alternative would turn out to be easier
and more efficient; it also would demand less organized stimu-
lation from the environment.

To be sure, once a certain (rather high) degree of problem-
solving ability is achieved, we could give a program the problem
of improving its own operation. This is a goal that Simon's
group has set itself for GPS. I am inclined to believe that this
can be done, in a sophisticated sense, only in a system with
a good deal of ability. Indeed, it is not clear yet whether GPS
has enough organization to do this. GPS can invent and select
improved methods within its framework; but it is not clear how
it could improve the basic structural outline.

We should mention, to avoid misunderstanding, that we can
construct very simple systems which can optimize their own
parameters: These are the "hill-climbing" or "self-optimizing"
mechanisms. Everyone invents these shortly after first consid-
ering the intelligent-machine problem. The serious difficulty
is at the next step: how to improve a structure not given in
terms of a natural parameter system? The question that imme-
diately arises is whether adequate structures can be described
by the system in a form that can be discovered without exces-
sive search.

117

G. A. Miller

I am very optimistic about the eventual outcome of the
work on machine solution of intellectual problems. Within our
lifetime machines may surpass us in general intelligence. I be-
lieve that Simon and his colleagues share this view. Perhaps
that is why he feels no urge to discuss here the consequences
of his work for management. We can discuss the subject of
management rationally when considering information retrieval,
simulation of companies, linear programming, and similar top-
ics, many of which have been treated in other lectures of this
series. But the prospect of machines that are better thinkers
than we are is hardly a stimulant to sober discussion. It makes
us feel that the problems of future management may be the
machines', not ours.

Miller. I wish that I could find some large and signifi-
cant issue where Professor Simon and I could disagree, so that
we could have a good, hot argument and stir up lots of excite-
ment and arm waving. Audiences always enjoy a good battle,
and any discussant worth his salt should be willing to go out
of his way to provide it. A good fight is valuable not only for
the emotional refreshment that it provides, it also helps to
clarify what the debaters are against, as well as what they sup-
port. One of the best ways to find out what something is is to
find out what it is not, to find its contrast classes. And contrast
classes often emerge most sharply and clearly under the stimulus
of disagreement.

In this instance, however, I find myself stuck for something
worth fighting about. Professor Simon invited me to take ex-
ception to his analysis of learning to talk, and I could. I could
quibble over some of the linguistic details as he presented them
to us. I might ask him to define what he means by a "phoneme"
and how it differs from a sound. I know that is a good question
because most linguists cannot answer it either.

I might ask him to be a little more explicit about how a
child acquires the ability to imitate. Or I might ask him the
$64 question: What exactly are the particular components of
speech learning that he has in mind? But in the present con-

118

Simulation of Human Thinking

text these problems are matters of technical detail, and quib-
bling about them could obscure rather than illuminate the
central point that Professor Simon was making when he used
the acquisition of speech as his example. His point was that
learning can proceed much more rapidly when a task has been
factored into (relatively) independent components. It seems
to me that this observation is not only true in general, but that
in particular it can be applied to a description of the acquisi-
tion of speech. So I see no reason to go into detailed criticisms
of this particular learning process.

Instead of carping about the details, therefore, I want to
take just a few moments to say why I, as a psychologist, find
the work of Newell, Shaw, and Simon so valuable, and why
I am so reluctant to do battle with them. Let me put it this
way: Suppose you were a psychologist and you wanted to formu-
late a theory to account for some realm of human thought or
human behavior. How would you proceed? In particular, what
calculus would you use to express your theory?

Let's consider the alternatives that are available. There is,
first of all, the natural language. This is the psychologist's fa-
vorite calculus, and I do not want to sell it short; the theories
of Sigmund Freud, William James, John Dewey, C. G. Jung,
and many others were grand achievements of the human under-
standing even though they were expressed in terms no more
rigorous than the language of everyday life. Nevertheless, I
think we should all agree that if a science is to grow, it must
eventually develop methods and concepts that are more concise
than the common language can express. When that day comes
for the psychologist, where is he to turn?

He might, perhaps, turn to the differential and integral calcu-
lus, the language that first expressed the deepest secrets of the
physical universe. It is true that there are numerous psycho-
logical processes that can be represented in this way — where
the essence of the matter can be caught by a functional relation
between two or more continuous variables. But this is a mathe-
matical language that was not invented for these problems. The
psychologists who try to use it inevitably look like children

119

G. A. Miller

wearing big brother's hand-me-downs. The calculus, for ex-
ample, does not have any way to represent a simple contin-
gency; the idea of contingency must be introduced informally
either by words or some other foreign device. But this is surely
one of the central concepts we need for a comprehensive psy-
chological theory.

In some quarters, of course, there has been considerable en-
thusiasm for symbolic logic as the proper language. Certainly
the work of McCulloch and Pitts — and of their successors (in-
cluding Professor Minsky, in spite of his comments ) — suggests
that logic is an excellent tool for describing neural nets. But
neural nets are, by definition, neural. A psychologist can watch
this application of logic with great interest and not a little envy,
but he cannot pretend that it is solving his psychological prob-
lems. The situation is somewhat like that of the psychologist
who is given a pile of lumber and the job of building a house
from it. These neural net people keep telling him about the
nature of cellulose, and it does not help him very much.

There have been a few attempts to apply logical description
at the psychological level: Clark Hull's description of rote
memorization in logical terms is now largely (and I think hap-
pily) forgotten. Jean Piaget's efforts to describe the thinking
of children and adolescents in logical terms are more pertinent,
but to anyone who studies Piaget carefully it becomes all too
obvious that he is forced to stuff his logical symbols full of all
kinds of significance and connotations that would make a true
logician cringe and turn pale. Piaget is using a great deal more
than just the raw logic. The most successful endeavor of this
kind that I know of is Noam Chomsky's logical description of
grammar; but whether this breakthrough in the neighboring
field of linguistics can infect psychology with similar success
remains an open question at the present time.

If not classical mathematics or symbolic logic, then what re-
mains? One of the most popular languages for formal theories
in psychology is the calculus of probability. Probabilistic de-
scriptions of what I can perceive are useful, and I can even
accept a probabilistic description of what I can remember, if

120

Simulation of Human Thinking

it is phrased with sufficient delicacy. But there is something
deliberately backwards about the way these probabilistic theories
generally proceed. The theory generates a null hypothesis that
we hope we shall be unable to disprove. As a consequence, we
often hypothesize that some psychological process is random
even though we know it is not, then demonstrate that certain
kinds of data are insufficient to prove that the hypothesis is
wrong. At the present time psychological theory is suffering
from our failure to decide when a probability is an explanation
and when it is an admission of ignorance. Since many of my
best friends indulge in this kind of theorizing, however, I hesi-
tate to launch a full-scale offensive. Not here, anyhow.

By a process of elimination, therefore, a psychologist who
wants to make his theory both comprehensive and explicit is
perhaps not driven, but is certainly nudged, in the direction of
the computer program as a natural way to do it. This is a new
calculus for expressing theories, and we psychologists do not
yet understand how to use it most effectively; but we are slowly
learning. Among the pioneers in exploring this new calculus, of
course, Herb Simon is one of the most eminent. You will have
to forgive me for not fighting with him. A fight is a lot of fun,
but it is a rotten way to express your gratitude.

Simon. I was very glad to have Professor Minsky point out
that my account of our common field of interest was one-sided.
My stress has been on the use of simulation to understand how
humans think, but many of the people in the field who have
contributed much to it are interested more in the other side:
how to obtain good problem-solving programs. In other words,
my colleagues and I have given emphasis to the word "intelli-
gence" in our work; other research groups have given emphasis
to "artificial."

The interesting thing is that up to now it has been very worth
while to devote considerable effort to explorations of the sly
tricks that humans use to solve problems, as a modest prelude
to finding better methods of solution. By studying the human
thought processes, we have learned a great deal about a whole

121

H. A. Simon

class of problem-solving devices, which I have referred to,
loosely, as heuristics. We did not have much formal knowledge
of these devices before, although they certainly had been in-
vestigated by other researchers, e.g., the mathematician, Polya.

There is thus a double aim in the field: understanding hu-
man processes, and automating various kinds of complex cog-
nitive tasks. We took the tack we did in our paper for two
reasons:

First, I had my say at some length elsewhere about the auto-
mation aspects, and particularly about the implications for
management.4

Second, the human aspects were the ones which primarily
motivated the work of my colleagues and me.

I do not agree with Professor Minsky that we shall stop trying
to understand the human mind as soon as we surpass the ability
of humans to solve problems. The challenge to our society is
not simply one of acquiring more powerful devices to do our
thinking for us. We have more powerful devices now to dig
dirt, pour concrete, erect buildings; and they do solve some
of our problems. We do not have to work as hard as we once
did, for example. They also create certain problems that we all
could easily enumerate.

If we are to take the next step and face up to some of the
hard-core problems that the human species is going to have
to solve over the next few generations in order to survive, then
we have to know more about ourselves than we do now. Re-
search devoted to the study of human thinking and human
cognitive processes is one way, at least, to work toward a better
understanding of ourselves. I feel that we must attach a very
high importance to this goal, quite apart from whether it en-
ables us to substitute automated processes for natural ones in
some applications.

I have one other comment, which actually is stimulated by
both Minsky's and Miller's remarks. Professor Minsky asked
whether a program like GPS could bootstrap itself, that is,
apply its learning processes to improving its own program. He

122

Simulation of Human Thinking

expressed some doubts about whether some of the proposed
self-organizing nets could do this. Professor Miller pointed out
quite rightfully that while I described how a child might go
about learning a language in general, I cheerfully (by means
of a few references to work by him and others) elided over
the details of just what a phoneme was and how it might be
represented. The issues raised by Minsky and Miller can be
fruitfully investigated with simulation techniques.

The way to answer the query, "Can GPS bootstrap itself?,"
is to spend the next six months asking what changes have to
be made in the program so that it will. The ultimate test of a
computer simulation is whether it simulates. In many forms
of theorizing, not including mathematics usually, you can
get by with hand waving at the appropriate moment. In my
constitutional law days, for example, I observed that whenever
Chief Justice Marshall got to a delicate point in an opinion he
said, "Obviously," and went on. If you write a computer pro-
gram and say, "Obviously," and go on, nothing happens. The
people running the computer room send back a little note read-
ing, "Error stop."

I think what Minsky and Miller were alluding to, but were
too polite to say aloud, was, "If GPS is a theory of how a ma-
chine can bootstrap itself into higher intelligence or how
people learn language, then let it bootstrap itself, and let it
learn language." This is an entirely appropriate obligation to
impose. The real power of the simulation technique is that it
provides not only a means for stating a theory but also a very
sharp criterion for testing whether the statement is, adequate.
Not just on behalf of myself, but on behalf of the entire group
of people working in the field (including the two discussants),
I accept the obligation and hope that one of us will produce
the requisite programs before too long.

Alexander. This great solidarity in the field suggests to me
at least that there may be an external threat. I do not know
whether it is represented here tonight, but whatever is repre-
sented, I shall now open the meeting to questions from the
audience.

123

General Discussion

General Discussion

Berkeley. I should like to ask what a "hierarchy of final
causes" is.

Simon. I should have Walter Pitts answer. I think that it
is just a nice phrase for expressing the fact that posing a prob-
lem for solution and setting a goal lead to the creation of
subproblems, and the solution of each subproblem becomes in
turn the immediate goal. If we examine in retrospect the com-
plete path that a problem-solving organism has followed, we
see that the problem has become a series of subproblems, sub-
subproblems, and so on, organized in an hierarchical fashion.

One of the interesting phenomena of problem-solving be-
havior is the fact that subgoals often get detached and become
autonomous. A large part of the theorizing about human mo-
tivations is concerned with how the subgoals are formed, how
they become autonomous, and how the hierarchical structure,
if it is hierarchical, becomes ramshackle and rambling in the
adult human being.

Minsky. To provide further evidence of the solidarity,
George Miller and I agree that if this isn't what Pitts meant,
it's what he should have meant.

Anonymous. It seems to me that GPS is something like a
weak form of utilitarianism. It tells us that the thing to search
for is the thing that will most reduce the difference between
the current state and the desired state. But how can we recog-
nize which is the proper way to proceed? We may well be
voiding our assumptions if we take a certain path. In other
words, if you just turn over to the machine the problem of pre-
paring a dinner, it may cook the cook.

Simon. That brings us back to the missionaries and the
cannibals, doesn't it? The reason that the machine does not
cook the cook is that there are many side conditions in the
statement of the problem, at least one of which precludes this
contingency. If GPS constitutes a theory of how humans think,
it must provide for side conditions; and, in fact, in the problem

124

Simulation of Human Thinking

of the missionaries and the cannibals, it does. Before GPS ap-
plies an operator, it checks to see whether the side conditions
are satisfied. One of these conditions is that there should not
be a preponderance of cannibals (or, in your case, that the cook
should not be cooked). If you carefully observed the young lady
on television trying to solve this problem, you may have seen
her begin to move a group to the other side of the river, and
then stop upon realizing that she was about to cause a pre-
ponderance of cannibals. The human problem solver also fol-
lows each partial attempt by checking whether the side condi-
tions are satisfied.

The broader answer to your question is that there is nothing
about GPS which requires the statement of the goal to be
simple. This statement may be an extremely complex descrip-
tion of a state of affairs. It will be very interesting when we
get to the point where we can describe, for example, what it
is that makes a painter decide that he has finished his painting
and should start on a new canvas. We certainly have not had
any statements of goals, so far, on this level of subtlety.

Anonymous. I should like to ask how a machine built by a
human being arrives at ways of solving problems that are better
than human techniques.

Simon. Testimony to the possibility of this is found in the
fact that the machine already performs certain functions much
better than the human: adding, subtracting, multiplying, and
dividing, for example. So there certainly is no reason in prin-
ciple why a machine cannot be built to do things that the
builder cannot do. We have known this about steam shovels
for a long time. Why should it not also be true for cognitive
acts? I might ask, "How could the blind forces of evolution have
created a mechanism which exercises intelligence?"

The fact is that we can draw a blueprint for a computer
without being able to predict at all what the behavior of the
computer will be under interaction with a specified environ-
ment. This is true of things much simpler than a computer. It
is true of most laboratory research in chemistry, for example.

I should like to put the burden of proof on the other side.

125

General Discussion

Why should we suppose, just because we have the ability to
create a certain situation, that the behavior resulting from the
situation will be very simple, very predictable, less interesting,
and less complex than our own behavior?

Futrelle. I was wondering about Professor Simon's dichot-
omy between operator and object. In the speech-learning ex-
ample, he dichotomized between the operation of a child's
producing a sound and the object of the child's hearing the
sound. It seems to me that this dichotomy would have to be
imposed from the outside. Would not the child or the program
have to learn somehow to distinguish between the motor action
and the auditory stimulus?

Simon. I do not think that distinguishing operator from
object should be too troublesome. This differentiation already
exists to a considerable extent in the organization of our ner-
vous system. I do not want to oversimplify the afferent-efferent
distinction, because there are many feedback processes compli-
cating matters, but basically the nervous system is organized in
terms of tracts which feed signals in, and tracts which take
signals out to the muscles. I do not believe that we have to
impose this dichotomy any more than biology has had to impose
it.

Minsky. The operators in the version of GPS that Simon
described are changes in motor patterns. Since the operators
cause changes in the objects, they are on a higher level than
the objects; so there is certainly no possibility of confusing the
two.

Rath. I believe it has been demonstrated that we can imi-
tate processes. We also know that we can build bridges. But
scientists today are not studying how we build bridges. They
are studying mathematics, physics, and chemistry. I think maybe
we are spending too much time playing the game of imitating
processes, and not enough time studying psychology, mathe-
matics, and logic.

Simon. To answer this fully, we should have to discuss at
some length what a theory is. In one sense a theory is always
an imitation. This is particularly true of the kind of theory

126

Simulation of Human Thinking

that purports to describe how a system moves through time,
e.g., the theories of classical mechanics. I believe that the im-
portant question is not whether we are imitating, but whether
in fact the imitation explains the phenomenon. The reason that
we approved of Kepler's imitation of the movement of the
spheres is because it told us where the spheres would be "to-
morrow night" in a way which was more parsimonious and in
other respects more satisfactory than earlier explanations. I
should think that we should want to judge the GPS activity in
the same way. I do not see that this activity is different from
psychology. If it is right, it is good psychology; if it is wrong,
it is bad psychology. And we all know what to do with good
psychology and bad psychology.

Anonymous. It often has been said that European rats solve
problems by insight, whereas American rats solve problems by
trial and error. Do you think GPS could become international
and solve problems by insight, too?

Simon. The reason there have been psychologists with Eu-
ropean rats and psychologists with American rats is that it has
been very difficult, up to now, to deal with complicated phe-
nomena like insight and at the same time to abide by certain
canons of scientific rigor. The psychologists who ran the Amer-
ican rats believed in rigor, while the psychologists who ran the
European rats were interested more in the richness of behavior.

It turns out that there is actually no contradiction between
insight and trial and error. It is just a matter of degree. Be-
havior begins to look very insightful when the amount of trial
and error is relatively small and the amount of selectivity is
relatively large. Our group has been able to simulate what I
think is insight. GPS has had at least a modest "Aha" experi-
ence, and if I had had a little more time to discuss the planning
program, I could have told you what that "Aha" was. I do not
think we have to live with the contradiction anymore; we can
have our insight and our rigor, too.

Emery. Professor Minsky touched on this, and I should like
to pursue it a little further. If you have a program that is self-
modifying, after a number of these self-modifications will you

127

General Discussion

be able to trace its "thought" processes, and what relevance
will these thought processes have to human thought processes?

Simon. Presumably we do know something about the varia-
tion of an adult human being who grows up in a human cul-
ture, and we therefore can get some measure of the differences
that can be produced by changes of experience. One thing we
certainly should do more than we have in the past is give the
computer the same series of experiences that we give a human
being, say, a child learning to speak. The test of whether we get
modifications of the same kind and direction in the computer
and in the child is another test that the theory should have to
pass, however unlikely this might seem at the moment. If it is
the right theory of learning, it will learn the right thing. By
"right" I mean the thing a person would learn when confronted
with the same experience. This has to be our test of the theory.

Beach. I think we would all agree that no two human
brains and no two thought processes are 100 per cent alike. How
does GPS take account of this fact? Does it incorporate noise
patterns?

Simon. No, not noise patterns. The program has sufficient
flexibility to fit different subjects. If we had to write a com-
pletely new program for every subject, then the game would be
lost. But I believe that there are enough similarities among
people to make it useful to attempt a general description of
their thought processes.

What we have done so far in this regard, and it is not very
much, has been to try to detect the differences in our subjects'
thinking which might account for differences in their protocols.
For example, our most successful problem-solving subjects made
much more obvious use of the planning method than did our
less effective subjects. My guess is that we now could demon-
strate that one of the real differences between the A students
and the C students is the extent to which planning and ab-
stracting programs are part of their problem-solving apparatus.

This is one interpersonal difference that we already have
begun to detect. There are more minor differences, some of
which can be accounted for by a few easy modifications of

128

Simulation of Human Thinking

certain of the parameters of the program. For example, smV
jects will differ in the order in which they deal with symbolic
objects, and this difference is reliable from one problem to an-
other, at least over short sequences. It may be represented by
a difference in the order in which certain items are stored on
lists in memory and retrieved from these lists.

We would hope, therefore, to account for individual dif-
ferences in terms of changes in program particulars which leave
the broad structure of GPS reasonably invariant from person to
person. If it turns out that we cannot, then something is very
wrong with the theory. I see no indication of this at the mo-
ment.

Alexander. I cannot refrain from using my privilege as
moderator to ask the final question suggested by this last re-
sponse. Is GPS so general that it would apply as well to the
amoeba in its swamp water as to Einstein working on relativity
theory?

Simon. We certainly have not experimented with any range
of individual difference comparable with the amoeba versus
Einstein. I should want to think some before I tried to describe
the processes of an amoeba. But the processes of Einstein, since
he was a member of our species, seem a little less mysterious.
The main difference between Einstein and other people was
that Einstein had been to college, had learned and thought
about certain things, and had talked to certain persons. There
were tremendous differences in experience between him and
anyone else who existed in the year 1905. If we admit these dif-
ferences in experience, then we can get by with rather modest
differences between Einstein's thinking processes and those of
other people. In other words, we can load much of the burden
on differences in experience.

There is a fair amount of historical evidence supporting this.
For example, there are not many cases of people living socially
and geographically outside of very small parts of Western Eu-
rope who invented steam engines or things like that. We must
either postulate that gray matter is very different in other parts
of the world (which is probably wrong biologically and certainly

129

References

is distasteful to most of us), or postulate that substantial differ-
ences in problem solving can be due to differences in experience.
My encouraging comment would be, if you just think long
enough about it, you too can invent relativity.

Referenxes

1. A number of these undertakings are catalogued in Professor E. G.
Boring's instructive and entertaining paper. "Mind and Mechanism/'
Amer. Journal of Psychology, Vol. 59, 173-192 (Apr., 1946).

2. See the Proceedings of the 1955 Western Joint Computer Conference,
"Session on Learning Machines," The Institute of Radio Engineers,
New York, 1955, pp. S 5-1 11.

3. Ibid.,p. 10S.

4. See Simon. H. A., The New Science of Management Decision, Harper
& Brothers, New York, 1960.

5. For an excellent recent survey of heuristic programs, although with
emphasis upon artificial intelligence rather than simulation of human
thought, see Minsky, M., "Steps Toward Artificial Intelligence." Pro-
ceedings of the IRE, Vol. 49, S-30 (Jan., 1961).

6. For a fuller description, see Newell, A., Shaw, J. C, and Simon. H.
A., "Report on a General Problem-Solving Program." in Information
Processing, Proceedings of the International Conference on Informa-
tion Processing. UNESCO, Paris 15-20 June 1959, UNESCO, Paris,
1960, pp. 256-264.

7. See ibid., pp. 261-262. for a description of a specific planning method
for GPS. In our subjects, abstracting often takes the form of simply
ignoring some of the problem detail at certain stages of the solution
process.

8. See Newell, A., and Simon, H. A.. "The Simulation of Human
Thought." in Current Trends in Psychological Theory, University of
Pittsburgh Press, Pittsburgh, 1961, pp. 152-169.

9. See, for example, Osgood, C. E., Method and Theory in Experimental
Psychology, Oxford University Press, New York, 1953, pp. 683-690;
Miller, G. A., "Speech and Language." Chap. 21 in Handbook of
Experimental Psychology, Stevens, S. S., Ed., John Wiley & Sons, Inc.,
New York, 1951; and Miller, G. A., Language and Communication,
McGraw-Hill Book Co.. New York, 1951, Chap. 7.

10. We qualify the adverb "purely" when we come to consider the fac-
torization of words into phonemes and phoneme components.

130

Simulation of Human Thinking

11. For a general introduction to these topics, see Miller, G. A., Lan-
guage and Communication, Chap. 2. An excellent recent survey is
Fatehchand, R., "Machine Recognition of Spoken Words," in Ad-
vances in Computers, Vol. 1, Alt, F. L., Ed., Academic Press, Inc.,
New York, 1960, pp. 193-321. See also Forgie, J. W., and C. D., "Re-
sults Obtained from a Vowel Recognition Computer Program," The
Journal of the Acoustical Society of America, Vol. 31, 1480-1489
(Nov., 1959), and Liberman, A. M., and others, "Minimal Rules for
Synthesizing Speech," ibid., Vol. 31, 1490-1499 (Nov., 1959). The
last three references cited illustrate, incidentally, the large role that
computers are playing today in linguistic and phonetic research.

12. The full account of this learning scheme is given in Newell, A., Shaw,
J. C, and Simon, H. A., "A Variety of Intelligent Learning in a
General Problem Solver," in Self-Organizing Systems, Yovits, M. C.,
and Cameron, S., Eds., Pergamon Press, Inc., New York, 1960, pp.
153-187.

13. We need hardly say that this description does not commit us to any
oversimplified reflex-arc picture of the peripheral and central systems.
GPS is a concrete example of a system of the sort we are describing. In
it, the perceptual-motor associations are represented by the table of
connections between differences and operators. The use it makes of
these connections, and consequently the relation of response to stimu-
lus, is highly complex.

14. It should be observed that the body is part of the environment that
is perceived. Hence, drives like hunger produce perceptual symbols just
as external senses do. Or perhaps it would be better to say that the
drive is the perceptual symbol produced by the perception of hunger.

15. Other aspects of the state-process dichotomy are discussed in Newell,
A., Shaw, J. C, and Simon, H. A., "The Processes of Creative Think-
ing," presented before a symposium on creativity at the University of
Colorado, May 14, 1958. The Rand Corporation, P-1320, Santa
Monica, Calif., Sept. 16, 1958.

16. We should like to call attention to the similarity, hardly accidental,
between the translation problems we have been describing and some
of the most striking results of modern mathematics — the undecida-
bility theorems. Decision problems, in the Godel sense, can always be
represented as problems of finding in language B the representation
of an object that is described in language A, where there are at least
some objects that have names in both languages. If the rules of cor-
respondence between the two languages are sufficiently complicated,
this will be difficult, and, as the undecidability theorems show, it may
be impossible.

131

4

A

Library

for
2000 A.D.

4

A

Library

for
2000 A.D.

Speaker John G. Kemeny

Chairman

Department of Mathematics and Astronomy
Dartmouth College

Discussants Robert M. Fano

Professor of Electrical Communications
Massachusetts Institute of Technology
Gilbert W. King
Director of Research

International Business Machines Corporation

Moderator William N. Locke

Director of Libraries and

Head, Department of Modern Languages

Massachusetts Institute of Technology

154

Kemeny. Since I am about to propose a radical reorganiza-
tion of university libraries, I must first establish that some such
reorganization is inevitable. I shall argue that our university
libraries will be obsolete by 2000 a.d.

The Need

Harvard University will have a library of more than ten mil-
lion volumes by 2000 a.d. Dartmouth College will purchase its
one-millionth library volume during the bicentennial celebra-
tions of the college, in 1969-1970, and if the present rate of
growth continues, the second million volumes will be purchased
in thirty-five years. At that rate of growth universities will have
a full-time occupation building new libraries in the twenty-first
century.

It is clear that the cost of building, purchasing volumes, cata-
loguing, and servicing these monstrous libraries will ruin our

135

J. G. Kemeny

richest universities. Even if Harvard University conceived a
ten million volume collection as physically feasible, this would
postpone the decision by only a few decades, since the collec-
tion of a hundred million volumes that one could predict for
2100 a.d. could not possibly be handled in a manner similar to
the procedures of any existing libraries.

As our present libraries grow in size, they also become increas-
ingly difficult to use. Library catalogues are growing into giants
and are approached with fear by both students and faculty.
Unless one knows the exact name of an author (including ini-
tials) or unless one knows the exact title of a book, it may be-
come hopeless to locate it. And once a mistake is made, it is
best to forget about it and start all over again. For example,
many libraries have discovered that if a book is misplaced on
the shelves and cannot be located after a short search, it is less
costly to replace it than to try to find it.

That the very conception of the way books are made avail-
able is wrong can best be illustrated in terms of an actual ex-
ample of search I recently carried on in the Dartmouth Library.
Table 1 gives a factual account of the search.

Table 1

An Example of Search

Walk from 328 Dartmouth to
Baker Library

4 minutes
30 seconds

Find card in catalogue
Walk up four flights
Find that book is missing
Walk down four flights

1 5 seconds
1 minute

1 minute 30 seconds

Find out that Professor S
checked out book

30 seconds

Walk from Baker Library
to 329 Dartmouth

5 minutes

Wait for Professor S to return
from lunch

2 hours

Wait for Professor S to find
book

1 5 minutes

Total search time

2 hours 27 minutes 45 seconds

136

A Library for 2000 A.D.

As can be seen, there are two major bottlenecks in the search
procedure. First of all, the physical running around is a very
significant proportion of the time I spent in using the library.
Dartmouth College's library is fully open stack; if it were not,
then someone else would do a great deal of this running around,
and the time consumed would be considerably longer. But even
the actual catalogue look-up and the running-around time to-
gether are negligible compared with the delay caused by the fact
that once a book is removed from the library it is likely to be
out of circulation for at least two weeks. If Professor S borrows
a book, he is likely to have it on his shelf anywhere from one
week to one year, during which period he might use it for a total
of less than one hour. Of course, one could require him to use
the book only in the library; but if he wishes to consult it a
dozen times during the coming weeks, it is most unreasonable
that he should have to spend a considerable amount of time on
each occurrence in locating the book. It is clear to me that as
long as libraries function in the present manner, this problem
is unsolvable.

I have saved for last the most acute problem of search today.
How can a worker in field X find out in a reasonable amount
of time what results are known concerning problem Y? For
example, there are several hundred mathematics research jour-
nals published in the world, and in addition to this, hundreds
of volumes of mathematics books appear each year. Relevant
information for a research problem may be contained in the
numbers of any one of fifty journals spreading back over the last
thirty years, or in any one of a hundred books. To find half a
dozen relevant items, it may be necessary to search thousands,
and the odds are great that some crucial piece of information
will be overlooked.

This problem is so critical that it must be solved and solved
relatively rapidly, independently of the problem of the reorgan-
ization of libraries. (This is, of course, the problem of the re-
trieval of scientific information.) However, this problem is
made more acute by the clumsiness of our libraries. If, to search
through two hundred books and journals, we have to locate

137

/. G. Kcmeny

each one by the procedure sketched above, and we have to wait
until twenty of them are returned to the library by colleagues,
the problem becomes completely hopeless.

In summary, our libraries are practically obsolete today. They
are certain to be obsolete by the end of the twenty-first century,
and will for most purposes be useless by 2000 ajd. It is therefore
appropriate to consider a radical reorganization of the entire
scheme of university libraries, and it is by no means too early
to begin planning today.

Statement of the Problem

If my proposals are to have serious scientific merit rather than
turn into science fiction. I must impose certain restrictions on
the solutions of the problem. First of all. I shall require that
my library be within the capability of our technology by 2000
a.d.. and I shall require that it still be useful at the end of the
twenty-first century. One cannot possibly hope for more than
that. Xo one can foresee the needs of the year 2100 a.d. or the
tremendously powerful tools mankind will have available by
that time. Specifically this will mean that the only devices that
may be used are those that will be available within the next
two or three decades, and that the library should start with
some ten million volumes and have the means of expanding
by a factor of 30. (At the present rate of growth, one would
expect a growth by the factor of 8 within a century. I am allow-
ing a factor of 30. since I am certain that the rate of expansion
will increase. If the rate should turn out to be much more
rapid even than that, then the library about to be described
may become obsolete before 2 1 00 a.d. )

Second, the library must solve the present bottleneck in
making items available to research workers. The time to take
out a book from the library should be of the order of magnitude
of a few minutes at most.

Third, there must be a reasonable procedure for finding out
where information is available. This procedure should be ap-
plicable not only to search by author or by title but to search
by subject matter.

153

A Library for 2000 A.D.

Finally, I want to put realistic financial restrictions on my
library. A solution that would require that the cost of purchase,
cataloguing, storage, and making information available increases
the library budget of every university by a factor of 100 is not
acceptable.

Basic Principles

If one accepts the statement of the problem as given above,
there are a few basic principles that are easily arrived at. For
example, it is clear that the library of the future will have to
make heavy use of automation. There is no conceivable way
in a library of several tens of millions of volumes that human
effort could locate an item in a matter of minutes. I shall,
therefore, proceed on the assumption that I can make free use
of whatever machinery can be designed in the near future, even
if this will drastically alter our conception of a library. How-
ever, I shall try not to propose the use of a machine where a
human being can perform the same task more efficiently. Just
as the majority of mankind resists the introduction of automa-
tion, due to very natural but irrational prejudices, the minority
of mankind pioneering automation has the tendency to use
machines in place of human beings whether this can be justified
or not.

It is equally clear that while the physical format of books is
very convenient for human handling, it is most inconvenient
for machine processing. A simple calculation of the volume
occupied by a hundred million books shows that even if they
were solidly packed, without room for humans to move around
in, they would occupy a building of impressive size. If we were
to have reasonable access to the books, we presumably should
need a building greater than the Empire State Building. There-
fore, storage methods must miniaturize books and put them on
a medium easily handled by machines, for example, some type
ot tape.

The very crudest estimates for a feasible library will put the
cost of construction somewhere in the hundred million dollar
to billion dollar range. It is not reasonable to expect that indi-

139

/. G. Kemeny

vidual universities should spend sums of money of this order of
magnitude. Indeed, it is difficult to see even today how one can
justify the tremendous waste of effort in which each university
has its own staff to order books, classify, and catalogue them,
and run a variety of duplicative reference services. I shall there-
fore adopt as a basic principle that we are designing a single
central library that will serve for research purposes both the
Federal Government and the major universities of the country.
I shall refer to it as the National Research Library.

In such a centralized library many functions will be no more
costly for the entire country than the cost at any individual
institution. Even where the central library will have to allow
for a considerable amount of duplication, presumably the merg-
ing of one hundred university libraries will not need more than
ten times the expense of an individual library. Therefore I esti-
mate that the actual cost of a central library might be of the
order of magnitude of 2 or 3 per cent of the cost of building
one hundred individual libraries.

I shall further assume that it will take our universities some
twenty years to decide that their libraries are rapidly becoming
obsolete. This will leave them an additional two decades to
complete the National Library by 2000 a.d. If each of one
hundred major universities agrees to contribute one hundred
thousand dollars per year from 1980 a.d. to 2000 a.d., and if the
Federal Government matches these funds, we should have
0.4 billion dollars available for the construction. Since many of
the universities will actually save a good deal more than one
hundred thousand dollars per year, and since the Federal Gov-
ernment could easily contribute a larger sum of money, it is
perhaps reasonable to allow up to 1 billion dollars for the plan-
ning, design, and construction of the National Research Library.

It may be appropriate to say a few words concerning the
likely fate of present libraries in the twenty-first century. I shall
take it for granted that our university libraries will not be abol-
ished, even if their role becomes secondary. Partly this will be
because they will serve a limited but useful purpose, but mainly

140

A Library for 2000 A.D.

because faculty members will be reluctant to give up personal
contact with books.

Aside from purely sentimental reasons for maintaining our
libraries, one can certainly make a strong case for keeping on
a given campus any book that may be consulted as often as
once a week. This would certainly include the present reference
rooms, as well as core research libraries in all subjects. The
periodical room would still serve as useful a purpose as in the
past, though perhaps it would be wasteful to preserve most of
the periodicals. The students would have books on reserve,
though they should also be encouraged to use the National
Research Library, so that they become familiar with its opera-
tion. One must also take into account that our university li-
braries play a major role in leisure-time reading, and the pleasure
of browsing should not be taken away from either students or
faculty.

All of these functions can be fulfilled comfortably with a
collection of no more than a few hundred thousand volumes.
I estimate this number generously, since I feel that the senti-
mental attachment of faculty members to books is very strong.
Our present libraries should be decimated in 2000 a.d., to cut
them down to such a reasonable working size. The number of
volumes could then either be kept constant by removing less-
often-used books and replacing them by current ones, or a very
slow rate of growth could be permitted. In any case, a great
deal of room would be freed under this scheme for faculty
studies and reading rooms (which are now in great demand
and in short supply on most campuses). One would also want
to put aside part of the library as reading rooms, in the new
sense to be described below.

Organization of National Research Library

We are planning a library which will start with some ten
million volumes and which may grow to three hundred million
volumes during the twenty-first century. I shall propose a basic
division of the material in this library into subjects and a divi-

141

/. G. Kemeny

sion of the subjects into branches. The fundamental unit of
storage, search, and retrieval will be known as an item.

A rough examination of the Dewey Decimal System, which
although out of date still is basically sound, suggests that pres-
ent-day knowledge can be classified into one hundred major
categories that I shall call subjects. Roughly speaking, a subject
corresponds to the first two digits in the Dewey Decimal System
or in a modernized version of the same. For example, mathe-
matics is identified by the digits 55. Allowing for the increased
complexity of knowledge in the twenty-first century, I shall
assume that the number of subjects will be doubled during the
lifetime of the library. I think that this is a realistic assumption;
certainly the number of subjects must grow at a much slower
rate than the number of volumes.

Each subject will function as a unit in the new library, and
will have its own room and complete means of storage, search,
and retrieval. The National Research Library might originally
consist of a hundred large rooms, and might have to grow into
a structure of two hundred rooms. This is a much more modest
rate of growth than any library can foresee in the immediate
future.

From now on I shall discuss only a single subject, since each
subject will be organized on the same basic principles as each of
the other roughly one hundred subjects. As my standard ex-
ample I shall use pure mathematics; a rough estimate would
show that pure mathematics and applied mathematics are each
of just about the right size to be represented as one one-
hundredth of a research library. According to the estimates in
Table 2, a subject should grow by some 60,000 items per year
in 2000 a.d. A subject that would reach that rate in forty years
should today grow at some 25,000 items annually, which makes
pure mathematics possible as a small subject. On the other
hand, this indicates that chemistry will have to be treated as
five different subjects in 2000 a.d.

A subject will be divided into a number of branches. The
basic restriction on this division will be that most problems be
classifiable as falling within a single branch. I have used as my

142

A Library for 2000 A.D.

guide in this matter the journal Mathematical Reviewsy which
publishes brief descriptions of everything that appears in print
in mathematics. From the journal's subject index it is reason-
able to estimate that by 2000 a.d. pure mathematics may be
divided into fifty branches. I shall allow the number of branches
of a given subject to grow up to one hundred. At this stage it
would appear better to have the subject subdivided into two
subjects.

As an estimate of the size, growth, and complexity of the
central library, we may take the figures in Table 2. These fig-
ures are guesses at averages, which should be fairly reliable, at
least in their orders of magnitude.

Within a given branch, material may be classified into books,
and books may be subdivided into items. By an item I shall
mean roughly a chapter of a book, or an article in a journal.
By definition an item will be somewhere between one and fifty
printed pages, and I should assume that most items would be
ten to thirty pages in length. If publications continue in any-

Table 2
Size of Library

2000 a.d.

2100 A.D.

Volumes

107

3 x 108

Items

3 X 108

101(>

Subjects

100

200

Items per subject

3 X 106

5 X 107

Items per subject in a year

6 x 104

106

Branches

5 x 103

15 x 103

Volumes per branch (with
duplicates)

4 x 103

4 x 104

Items per branch

105

106

143

J. G. Kemeny

thing like their present form, this would seem to be the smallest
reasonable unit for storage and retrieval.

At this point I should like to propose a basic modification of
the manner in which items are stored. I can illustrate this in
terms of a research journal in mathematics. A given volume of
a research journal is likely to contain articles on a wide variety
of topics in mathematics, and only a small fraction of these
would be of interest to any one research worker. It would be
much more sensible to collect all articles that are published in
a given month which relate to a given branch, and label them
as a "volume." Since items will be stored by branches, some
such procedure will be essential. Of course there will be the
difficulty that some articles will be relevant to several different
branches, and in this case we should bind copies of the same
item into several different volumes and add them to the library
of all of these branches. Thus a volume will be a homogeneous
collection of articles relating to a single branch. The classifica-
tion of items according to branches will of course be crucial
to the success of our National Library. However, this does not
seem to be an unsolvable problem. For example, in mathe-
matics, where the American Mathematical Society operates
Mathematical Reviews, each reviewer is even now asked to sug-
gest a classification or classifications for the article he reviews.
Similarly, our library could retain experts in all fields, to write
abstracts and to indicate to which branch or branches the ar-
ticle most appropriately belongs.

Each item could be catalogued by a code name consisting
of three letters and nine digits. For example, in the code
ABC-1 2-34567-89, the letters ABC designate the subject and
the number 12 the branch. The volume within the branch is
indicated by 34567 and the item within the volume by the last
two digits. By consulting Table 2 we can see that this reason-
ably simple coding system could take care of the needs of the
twenty-first century. If one visualizes retrieval of information
by means of a device similar to a telephone dial, we find that
we have available roughly 5 X 1011 combinations, while our
maximum need is for some 1010 labels. This degree of "waste"

144

A Library for 2000 A.D.

is certainly necessary if the coding system is to be at all natural.

For example, an expert in probability theory would soon
learn to dial MAT-17 to obtain access to his specialty, after
which he could dial a seven-digit number to obtain a specific
item.

The Problem of Storage

We must now face the problem of storing on a single tape
several hundred thousand items belonging to a given branch.
There is every reason to believe that within the next hundred
years miniaturization will make it possible to store all the vol-
umes in the library in a small box. However, our devices must
be available in this century, and they must make items
available quickly and inexpensively. Therefore, a page should
be reduced only to a size from which it can be magnified to
its normal dimensions optically. This would indicate that we
should store each page on a millimeter square. This degree of
miniaturization is essentially feasible today, and will certainly
be practical in another two decades.

We are now in a position to store all the articles of a given
branch on a tape with dimensions of roughly 2 inches by sev-
eral hundred feet. Let us refer to a 1-millimeter cross section
of the tape as a line. We shall store all the pages of a given
item on a line. (We should note that a 2-inch width is exactly
right to allow the storage of up to fifty pages on a single line. )
Since the basic unit of storage and retrieval will be an item, it
will be convenient to have copying operations done a line at
a time.

If we estimate that a branch will have a maximum of 106
items, we must allow tapes 3300 feet in length. Since current
high-speed tapes may be 2400 feet in length, the manipulation
of library tapes could be handled by a device not very different
from tape units now employed.

How long would it take to retrieve a given item from such
a tape? The current devices can run through the entire tape,
filling in or copying every bit of it, in 5 to 10 minutes. Of course
this speed can easily be improved. And we must take into ac-

145

J. G. Kemeny

count that we should rarely want to copy any significant por-
tion of the tape; we may instead employ a rapid search device
to find the beginning of a volume, and then copy the relevant
items from the volume at a more modest speed. We may there-
fore safely assume that we can build tape devices, which, given
the code numbers of some twenty items in half-a-dozen dif-
ferent volumes, could copy the items onto another tape in less
than 1 minute.

I therefore visualize a room that has about one hundred
tape units for the storage of books concerned with the given
subject. Typically the subject would have fifty different branches,
and therefore without duplication fifty tapes would suffice.
Indeed, some branches would be consulted sufficiently rarely
that a single copy of its tape will make material quickly ac-
cessible, but in the more popular branches two or three dupli-
cates of the same tape will be desirable. If we assume that each
customer ties down a tape unit for 1 or 2 minutes (and we
shall see below that this is a safe assumption), then we may
estimate the number of duplicate tapes necessary by requiring
that during the busy part of the day no more than thirty to sixty
customers per copy would wish to consult the tape. Even in a
central library it is safe to assume that a device that can satisfy
some two hundred customers per day will suffice for many
branches. And it is hard to think of branches in which more
than two or three duplicate copies will be required.

Since storage will certainly be one of the major expenditures
in any future library, it is worth pausing to estimate the cost
of storage. In 2000 a.d. we shall have to allow approximately
10,000 tape units. Since they will be produced in such great
numbers, the cost per unit should not exceed a few thousand
dollars, making the expenditure for storage less than one hun-
dred million dollars — entirely within the range of our budget.
This estimate indicates that the one billion dollar figure given
earlier for the total cost of the library is a very generous one, but
it does show that the total cost will have to run to a few hun-
dreds of millions of dollars. During the century the number of
tape units may quadruple. This would mean an annual expendi-

146

A Library for 2000 A.D.

ture of between one and two million dollars, which will be a
reasonable fraction of the total operating cost of a library that
serves the entire country.

One of the advantages of storing items by branches will be
the ease with which new items can be added. Since the num-
bering of volumes of a given tape is arbitrary, one may simply
append new volumes at the end of the tape, assigning the next
available serial number. I should propose that, although the
National Research Library should operate night and day, every
day, each subject be "closed down" one day a month for the
addition of new material and for necessary checking and re-
pairs. The addition of some one hundred volumes to a branch
should be a very simple operation. The material could be put
on a master tape during the rest of the month and transcribed
onto the appropriate branch tapes on the special date. If the
mathematicians of the country get used to the fact that the
mathematics library is closed on the 13th of each month, this
will be of very little inconvenience to them in the long run.

It will be an important safety device to keep one copy of
each branch tape in a vault, to be available for the repair or
replacement of parts of the tape that become damaged or
"worn out." Or it may in the long run prove simpler and
cheaper to replace each tape at the end of the year by a copy
from the master tape. From here on I shall ignore this problem
of the updating of materials and the question of repairs, since
there is no essential difficulty in either.

Retrieval of Information

Let us now turn our attention to the user of the National
Library. He will be located at his own academic institution,
and therefore the institution must be connected to the central
library by means of a multichannel cable on which pictures
can be transmitted. Such devices have already been developed,
and we can expect them to be in general use soon. The uni-
versity will have a large number of reading units scattered
around the campus, some at the university library, some in
departmental reading rooms, and some in individual professors'

147

J. G. Kemeny

offices. (See Figure 1.) The number of these units and the num-
ber of channels to the central library will no doubt depend on
the size of the institution and on its "library budget."

The reading unit may appear similar to a microfilm reader,
but it must be equipped with a tape unit capable of receiving
pictures from the central library, each picture representing one
page. The unit must also be equipped with some means of
sending signals to the central processing unit; for the sake of
simplicity I shall visualize this device as a telephone dial.

The customer turns his set on, and begins by dialing the
three-letter code of the subject of interest, say MAT for pure
mathematics. This connects his reading unit to the central
processing unit for pure mathematics, and makes a projection
unit available to him at the National Library. Next the cus-
tomer dials the two-digit code number for a branch (say, 17 for
probability theory) or for a catalogue (say, 00 for author cata-

Storage Tape for Probability

pp"n

V

Catalogue

Central

Processing

Unit

Projection
Unit

\

\

1 1 1

r

J Reader

Mathematics Room at Library

Reading Room at University

Figure 1 Mathematics Room and Reading Room for National
Research Library

m

A Library for 2000 A.D.

logue), or a special code for some other service provided by the
central library.

For example, the customer might dial MAT-00-ART. This
would connect him with the mathematics room, and with the
author catalogue in that room. He would start looking through
the author catalogue with authors whose names start with the
letters ART and he could "flip through" the catalogue by
pushing a button on his set which instructs the projector to
advance by a page.

The idea of having library catalogues arranged by subjects is
so simple and practical that it is hard for me to believe that it
is not now in general use. Surely it would speed the finding of
a given book, if, instead of having one gigantic catalogue, one
would have it divided into a hundred segments corresponding
to individual subjects. In most cases the customer would im-
mediately find the correct subject; and even in the worst case
he is unlikely to have to look in more than two or three sub-
jects. Having to look in card files one one-hundredth the size
would, I am certain, speed up the retrieval procedure con-
siderably. Except for this modification in the arrangement of
the catalogues, the author and title catalogues — which are
among the more efficient features of present libraries — would
not be drastically different from present files.

Once he finds the listing of the item he desires, he can dial
its nine-digit code number to obtain the item itself. A slightly
different code could be used if he wishes to obtain an entire
book (e.g., using 00 in place of the item number). With a
little skill and some thought it should take a customer less
than 5 minutes from the time he turns on his set to the time
that he has retrieved an item of interest from storage.

This makes it very convenient for the user who obtains an
item, but what about his many colleagues who would like to
look up the same item while he is perusing it? Of course we
should no longer have items tied down for several weeks, only
for the period that the item is actually being read; but in a
library servicing a hundred universities this would still be an
intolerable bottleneck. Fortunately, there is a simple way to

149

J. G. Kemeny

deal with this problem: The customer is never allowed to use
an item directly from the files, but only a copy of it.

When an item is requested, this is transferred from the
storage tape to the projection unit, by copying one line, and
from there it is sent, conveniently magnified, to the tape in
the reading unit. The latter tape might be visualized as similar
to a 3-minute length of Videotape. Such a tape can receive in-
formation at a speed of an item a second even without im-
proving on today's technology, and it can hold about ten vol-
umes (assuming that we store one page per exposure at normal
camera speed). Once the items wanted are copied onto the
reading tape, the customer "hangs up" and frees the storage
tape and projection unit for his next user. He can then read
the transcribed items at leisure. Once he has finished with
these, he may either keep the tape for his personal library (a
10-volume collection of personally selected items on a 10-dollar
tape! ) or reuse the tape next time.

As indicated above, dialing MAT connects him with the
mathematics room, and procures a projector. Next, dialing 17
connects the probability storage tape unit to the projector. As
he dials the five digits of a given volume, the storage unit
carries out a high-speed search, and the final two digits of code
instruct it to transfer an item (or an entire volume) to the
projector. While the next item is being dialed by the customer,
the projection unit transmits the previous one to the reader's
tape. Thus items may be obtained just as rapidly as the cus-
tomer can dial seven-digit code numbers. The average customer
should not tie down the storage unit for more than a minute
or two. This procedure will also hold the cost of long-distance
consultations to a minimum.

A similar procedure could also be used in looking up items
in a catalogue. Instead of tying down the catalogue while the
customer searches, relevant portions of the catalogue could be
copied onto his reading tape, and he could search the catalogue
at his leisure.

The reading unit could be of a relatively simple type, as
shown in Figure 1. It would have to hold a short tape on which

150

A Library for 2000 A.D.

it can receive pictures at normal camera speed. It would present
this tape a frame at a time, magnifying it as in an ordinary
microfilm reader. It would have to be able to advance the tape
a square, or move back a square, and to move to the next item,
or to back up to the previous item. In addition to this, the unit
would have a signaling device, say, a phone dial, for communi-
cating with the central library. It is reasonable to expect that
such units could be mass-produced at a cost that would not
only allow each institution to purchase a number of these but
that even individual professors might obtain private readers.

Each reading unit would be furnished with a "telephone
book" containing the code names of subjects and the code
numbers of the branches of each subject. To make use of this
book practical, under each branch there should be a brief de-
scription of the contents of this branch, since, as we shall see
later, correct identification of branches will be crucial for search
purposes. As the complexity of the library grows, the division
into subjects and branches will have to be revised every ten
years, and thus new "telephone books" will be issued at the
beginning of every decade.

Any number of additional uses of this setup suggest them-
selves. For example, we have already noted the possibility of a
customer recording interesting items for his personal library.
For a particularly important item it may be desirable to have
a copy available in normal size, and thus we may wish to attach
a photocopier to the reading unit. The exact extent to which
these services are used will depend both on the relative costs
of permanent copies versus long-distance calls, and on the fre-
quency with which the items are needed. Unless they are used
very frequently by the customer, it must be recalled that it will
take him less time to obtain a new copy from the central library
than to find the item in question in his personal library.

Or again, we could attach a projector to the reading device,
which shows a page on a screen. Not only would this make it
easier for men with visual defects to consult the library, but it
would enable the mathematician to lie down on his couch
while he thumbs through the latest research journals. The same

151

J. G. Kemeny

device could be used in a seminar or a research conference,
allowing all participants to have simultaneous access to an item
from the National Research Library.

I want to emphasize that I have thought these items up
during a period of a few months, and I am forced to consider
subjects in which I am not an expert. However, I believe that
even this crude sketch establishes the fact that a storage and
retrieval system to meet all of my conditions can be designed
by 2000 a.d. I am quite certain that a group of one hundred
experts in a variety of fields, cooperating over a period of twenty
years, could come up with a scheme for a central library that
will be as far superior to my proposals as my proposals are to
our present obsolete libraries.

For example, one of my Dartmouth colleagues suggests that
instead of having the National Research Library centralized, it
could be located at various major libraries, each serving as the
custodian for one or more subjects. While such decentralization
may be more costly to maintain, it might have many advan-
tages. If each major library had a complete projection unit
available, the local library could have on tape frequently used
copies of items from all subjects, and all the catalogues, thus
cutting down significantly on the demands on long-distance
cables. Of course a user need not know the actual location of a
subject. His request would be taken care of automatically by
the usual means for routing long-distance phone calls. Thus the
alternate scheme amounts to a decentralized but unified Na-
tional Research Library.

This leaves us with but one major problem, namely, the
problem of searching for information relevant to given scientific
problems.

The Problem of Search

Let us consider the problem of finding information relevant
to a research problem in a typical branch in the middle of the
twenty-first century. Hopefully we manage to pinpoint the
problem as belonging to a single branch. This still leaves us
several hundred thousand items to search. It is in this area that

152

A Library for 2000 A.D.

the high-speed computing machine could make a significant
contribution. So far we have automated our library to a con-
siderable degree, but we have employed only machines of the
simplest possible type. Only the problem of search is sufficiently
complex to justify the use of a "giant brain."

Let us make our problem more precise. It is entirely reason-
able to expect that for search purposes we have available a rela-
tively brief abstract of each item in the branch. For example,
this could be a coded version of a review in Mathematical Re-
views. Or more usefully, a general scheme for the format of
Mathematical Reviews could be agreed upon by the American
Mathematical Society, and reviewers could be asked to write
their reviews in this specific format. At any rate, each item is
represented by one coded page, and we are to search relevant
pages among several hundred thousand. Let us first consider
two extreme solutions of the problem: In one the machine will
play a role that is trivial, and in the other the machine will play
an exceedingly complex role.

The simplest role the machine could play is that of a gigantic
memory with a simple treelike pattern built into it. Experts in
a given branch are asked to subdivide the branch into sub-
branches and divide those into topics and those into subtopics
etc., making at each point some two to ten subdivisions until the
partition is fine enough that any one cell has sufficiently few
items in it for a research worker to examine all the abstracts in
the cell. If we start with, say, 5 X 105 abstracts and wish to
narrow it down to twenty abstracts in a typical problem, there
must be 25,000 subdivisions of the branch. This would put a
tremendous strain on the experts who are to arrive at the classi-
fication scheme, and the probability of taking the wrong fork
in the search processes would almost necessarily be very high.
Nevertheless, this is a possible scheme, and therefore I have
carried out a small experiment to see how well it could work.

I took the two-hundred-odd articles in pure probability theory
that were reviewed in one year in Mathematical Reviews. Start-
ing with the division into subbranches of that journal, I in-
troduced a division into topics, and in some cases into sub-

153

/. G. Kemeny

topics. If the system worked well, it would narrow the search
down to two or three abstracts after only two or three ques-
tions. I must confess that I had great misgivings about this
system even while I was designing it, since in some cases papers
could legitimately belong to several categories, while in others
it was hard to find common features for any significant num-
ber of articles.

The results of the experiment were catastrophic. In thirteen
search problems, performed by three probabilists — with an as-
sistant of mine playing the role of the machine — the abstract
was located on the first try in five cases, after a small number
of tries in two more instances, and not at all in six cases. It was
not uncommon that the very first division, that taken from
Mathematical Reviews, led the searcher into the wrong cate-
gory. I have since learned that reviewers for Mathematical Re-
views frequently refuse to classify papers more accurately than
into branches, because of the ambiguity of the other divisions.
I therefore conclude that this method of search cannot be prac-
tical on a large scale.

At the other extreme, the entire burden could be placed on
the computer. The customer could furnish it with information
about what type of problem he is considering, and the com-
puter would be asked to search through the hundreds of thou-
sands of abstracts to try to detect ones that appear to it to be
relevant. It would then present the customer with all these ab-
stracts, and the customer could choose amongst them. I feel
quite certain that a scheme similar to this will be entirely
practical a hundred years from now, but rough estimates seem
to show that in the immediate future such a procedure is still
Utopian. If the scheme is not to make too many mistakes and
yet be successful in narrowing down the abstracts to a couple of
dozen, the computational time involved would be prohibitive.

I therefore come to the conclusion that a scheme that is to
be practical in 2000 a.d. and stay practical for a few decades
would have to be some sort of give-and-take between the hu-
man being and the high-speed computer. Although a complete
classification scheme described above seems to be too cumber-

154

A Library for 2000 A.D.

some a system, a partial classification would certainly be prac-
tical. Mathematical Reviews divides each branch into some-
thing like ten subbranches. My experiment indicates that this
is not completely safe, in that even at this stage of division the
chances are great that the customer will be in doubt about which
of two or three branches he is interested in. However, I can
conceive of a series of ten questions in each branch that the
computing machine might ask the consumer, and the research
worker could indicate in each of the categories all those topics
that might possibly be relevant to his search. It would be im-
portant to allow in each case the option of indicating that the
particular method of division is irrelevant to the problem in
question. I should visualize the give-and-take as follows: The
machine flashes to the reading unit a list of alternatives, and
the customer signals back by dialing the relevant numbers. Per-
haps an answer of zero could indicate that the method of sub-
division is not helpful. If even half of the questions turn out
to be relevant, and in each case the possibilities are cut in third,
you will have narrowed the number of abstracts down to around
two thousand.

At this stage the computer would be called upon to perform
a more essential role. It would search these two thousand ab-
stracts and try to find features that many of them share. It
could then ask the customer to indicate which of these features
would be of interest to him. The computer could also indicate
after each reply how many abstracts are still "in the running."
After three or four exchanges of information the list of ab-
stracts should have decreased to a manageable quantity.

This would appear to be a more fruitful division of labor.
The experts in the branch would be asked to devise ten criteria
on which items may be classified, and the reviewers would have
to indicate for each criterion how the item should be classified.
This is a practical procedure as long as it is not expected that
in the cross-partitioning every combination (or even most com-
binations) should occur, or that the final output of such
classification needs to be a very small number of abstracts.

It is crucial that in the search the customer be allowed to

155

J. G. Kemeny

declare that a given method of classification is irrelevant. For
example one obvious criterion might be the date of publica-
tion. For some purposes only the last ten years need to be con-
sidered, a situation that would narrow the search considerably.
For some other purposes the date criterion is irrelevant. For
still other purposes the last thirty years would suffice. Thus
the experts might suggest listing the last eighty years in ten-
year intervals as eight categories, and '"older" as ninth, with
"irrelevant" being the tenth. Then one researcher might nar-
row it to a single category, another to one of three categories,
while still another would ask that this criterion be ignored.

The searcher would have ample opportunity for using his
skill, in trying to narrow the search, without loss of valuable
information. He is also free to decide at what point the search
has narrowed sufficiently for his purposes. For example, if he
is trying to compile a bibliography, he is likely to want to look
at a much larger number of abstracts than if he simply wants
to know whether a certain theorem is known to be true.

Finally, the machine plays a significant role, but on a prac-
tical scale. Clearly the ingenuity with which it is programmed
to classify the remaining abstracts into "similar" ones will make
all the difference in the speed and likelihood of success of the
operation. Yet we are not asking of it anything that is beyond
the likely developments of the next twenty years, nor anything
that would take more than a few minutes per stage.

In summary, each component seems to be playing a near
optimal role: The experts in the field can suggest useful ques-
tions that will in most cases narrow the field, the customer can
use these criteria optimally for his own purposes and can make
the final selection from a short list of abstracts, while the com-
puter serves partly as a memory for questions and partly as a
rapid device for the rough classification of a large number of
abstracts.

I should propose that the search machine also be programmed
to observe its own operations and to improve its procedure as
it leams from experience.

An incidental benefit of the compromise solution is that it

156

A Library for 2000 A.D.

does not require extra hardware at the customer's end. His read-
ing device and dialing system will be sufficient. He can dial a
suitable code, and then simply answer questions presented to
him on his screen by dialing all relevant numbers. Finally,
when the search has narrowed sufficiently for his taste, he would
signal the computer to put all the remaining abstracts on his
reading tape, and he could examine them at his leisure, noting
down call numbers of relevant items.

To test these ideas, I carried out a second search experiment.*
Taking the probability publications of one year as my data, I
designed a scheme by which each item was to be classified ac-
cording to whether it had present in it one of fifty interesting
features, falling into several major categories. The searcher was
to indicate all features of interest to him, and our Dartmouth
LGP-30 was to find all items having these features. In the
experiment the coding was carried out by my assistant, who
was not an expert in probability theory — a fact that no doubt
had a serious effect on the outcome. Then each subject was
asked to pick an item from the year's product, and to list its
significant characteristics (from the list of fifty), judging from
an abstract of the article.

Strictly speaking, this experiment was no more successful
than its predecessor. It produced about 50 per cent success in
some thirty tries. However, while the first experiment seemed to
be hopeless, the second one had several promising features. For
example, when it succeeded, it was too successful: It often pro-
duced only one or two abstracts. This seems to indicate that a
cruder classification would have sufficed. When it failed, it usu-
ally failed because the searcher noted some one feature that
the classifier did not include. In many cases this was directly
due to the inexperience of the classifier, and my assistant feels
that if the classification had been performed by an expert, the
percentage of success would have been much greater.

It is also significant to note that failure usually resulted in
no abstract being turned up at all. In many cases a repeated

* In both experiments I was ably assisted by George Cooke, a sophomore
research assistant at Dartmouth.

157

/. G. Kemeny

search, omitting one criterion at a time, produced the desired
abstract, without producing too many others. Thus there seems
every hope that an improved search scheme along these lines
would be reasonably successful.

But there are also a number of danger signals present in the
outcome. It is most disturbing to note how easy it is for two
experts in a given field to disagree on whether to include or ex-
clude a significant feature. I am now convinced that any search
scheme in which one "wrong" answer dooms us to failure can-
not possibly work. It was also significant that the subjects were
invariably eager to ask the experimenter additional questions.
I do not see how some "conversation" with the machine can be
avoided.

A second disturbing feature is the length of time necessary
for even a simple search. Our method required about Vi minute
per item. Of course, on a faster machine this can be cut by a
factor of 10s, and in the future it is likely to be cut by an addi-
tional factor of 10. But there is reason to believe that as the
number of items increases by a factor of n (and the number of
criteria by a factor of perhaps y/ri) the length of search time
may increase by a factor of n2. Since we expect a branch to have
some one thousand times as many items as the number we
searched, this could mean an increase by a factor of 106/104,
or 100. And search times of the order of an hour are prohibitive.

It would be easy to answer that this will no doubt be im-
proved upon with the progress of technology, but we have al-
ready admitted that the techniques I tried were too primitive,
and that slower techniques will have to be used. Thus here is
one area where considerable improvement in computer design
and use may be required. Of course this research will be of vast
value for many other purposes, and we may expect this research
to take place whether or not we decide to build a National
Research Library.

This still leaves the question of what type of machine is re-
quired for search. Should it be one gigantic machine for the
entire National Library, or one for each subject? If the latter
is not prohibitively expensive, it would certainly be preferable.

158

A Library for 2000 A.D.

But here I should estimate that two hundred machines with
the capabilities desired would exceed our budget. I therefore
suggest a single gigantic computer as more practical. The com-
puter could operate on a single master program, with tapes
giving it the suggested questions for any one branch, and having
on file all the abstracts of the branch.

The machine should certainly be designed so that several
customers can consult it simultaneously. This will eliminate
some of the waste caused by the fact that the customer thinks
at a much slower rate than the machine. Recent developments
in computer design would indicate that there should be no diffi-
culty in having one hundred or even one thousand simultaneous
consultations.

If this turns out to be practical, a customer could dial INF
to be put in touch with the information-search machine, and
then dial the code for the subject and branch. From then on
the procedure would be as indicated above.

Of all the problems discussed in this paper, this is the one
requiring most study. I expect that here is an area where a
group of experts could make vast improvements on the scheme
I have proposed.

Conclusion

I have argued that our present libraries will be obsolete by
2000 a.d., and that the library of the twenty-first century must
be designed on entirely different principles. I hope that I have
answered this challenge by describing a possible library for
2000 a.d. within the reach of technological development and
within a realistic budget.

I find the concept of such a library very attractive. I am
basically a lazy person. I should like to sit in my office and
have access to a book with no more trouble than calling a friend
on the phone long-distance. Of course, I may occasionally get
a busy signal, but few of us would argue that this makes the
phone impractical. I have tried to argue that we should get
busy signals less often than the frequency with which the book
we are looking for is out of our library today. And we could ar-

159

/. G. Kemeny

range our calling system so that we can hold on and obtain ac-
cess to a storage tape as soon as one is freed for the particular
branch — an interval that should be within minutes.

I am particularly attracted to the prospect of combining this
automated library with machine search. I should hope that this
feature alone would justify the great central library. It is pos-
sible that, even with all this elaborate mechanization, infor-
mation retrieval will become hopeless in one hundred years,
but without mechanization we won't have a ghost of a chance.
I look forward with delight to being able to find in 10 minutes
even-thing relevant that has been written on a given subject, or
to find that nothing relevant has been written.

I do not claim that this will save money for our universities
in the long run, but I do believe that only in this way will they
be able to continue operating at anything like their present
library budgets. The alternative is either to abandon any degree
of completeness or to increase library expenditures to fantastic
sums.

And we must realize that the incidental benefits of change
will be tremendous. Most important, of course, will be that all
participating universities will have access to all the books of
the National Research Library. The impact of this single fact
on research could be so great that all my estimates for growth
could prove too conservative.

But we can also visualize an endless list of auxiliary services,
tied to the central library. For example, a search service could
be established for the convenience of industries needing quick
and thorough information on past research in a given field.
WTien one considers the hundreds of organizations now offer-
ing partial sen-ices along this line, it is not inconceivable that
this service alone could bring in enough income to finance the
National Research Library.

But I am more interested in the variety- of new services that
could be made available to university staffs. It is a research
worker's dream to think of a system that would provide as the
result of a single call all the items published in his branch
during the past month. All it would take is a special marker

160

A Library for 2000 A.D.

at the head of the items added during the past month. Then
one could dial a special code, say MAT-1 7-99999, to obtain
a transcript of all these items. Or alternatively, one could be
furnished with abstracts of all such items. It is the only way
that I can conceive of mathematicians keeping up with their
specialty in 2000 a.d. And think of the needless duplication
resulting from lag of information that would be eliminated!

Again, it is reasonable to assume that machine translation
will be available by 2000 a.d.; hence the reader should have
the option of obtaining an item in the original language or in
translation.

Equally promising is a scheme for compiling a bibliography
for a research project. The potential user would furnish a list
of relevant articles, and a brief description of the project. The
central machine would scan the bibliography of each article,
examine its entries for relevance, and look up the bibliographies
of the new relevant items, etc. This procedure, a familiar one
to research workers, would condense weeks of work into minutes.
Even if it produces a number of irrelevant items, it would be
very valuable for research purposes.

And, no doubt, the reader will think of some pet projects
of his own that will become possible in a National Research
Library.

It is also safe to predict that the effect of such a library on
the very nature of research will be immense. I don't dare to
explore this subject too far, for fear of sounding fantastic, but
I do want to raise one possibility. Isn't it conceivable that the
nature of publications will change? Why should we continue
to publish hundreds of journals in each subject, when the
simple act of depositing an item in the National Library would
accomplish more? We can visualize a time when a research
article in mathematics is simply submitted to the "mathe-
matics editorial board" at the National Library, where it would
receive the customary referees' treatment. Once it is accepted,
it is filed into the National Collection, together with an ab-
stract of the accepted format, on the next 13th of a month.
Any library or individual would have access to it after that,

161

R. M. Fano

and could obtain a copy of it by the procedure described
earlier. The time of 'publication" could be cut to about three
months from the present two years, and the cost would return
to normal proportions.

My purpose has not been to say the last word on this sub-
ject. After all I have suggested a time requirement of twenty
years and an expenditure of one billion dollars for the design
and building of the Library. I have tried to show only that even
after so brief a period of thought one can design something
vastly superior to our present libraries, and hence I hope to have
established the possibility of a National Research Library for
the twenty-first century. My fondest desire is that others should
go far beyond my modest beginnings, and that someone may
be persuaded — sometime during the next two decades — to
do something about this vital problem.

Panel Discussion

Fano. Professor Kemeny's estimate of the magnitude of the
problem that libraries and library users will face in the future is
unfortunately very realistic. One might say that the volume of
printed matter, like entropy, always increases. As a matter of
fact, the rate of generation of scientific literature has already
reached the point where some discoveries have to be made
twice: the second time, when the paper in which the original
research was reported is discovered by the scientific community.

I have no major objections to Professor Kemeny's outline of
a National Research Library from which copies of any desired
item can be obtained automatically by dialing the appropriate
number. I certainly agree that such a solution is economically
possible as well as technically feasible. Incidentally, it would
require a change of the present copyright law. On the other
hand, I doubt very much that such a centralized library can
become a reality until some efficient and effective method is
developed for identifying automatically the items pertinent to
any particular literature search. As far as I know, no acceptable

162

A Library for 2000 A.D.

solution to this literature-retrieval problem is yet in sight. I
should like to make a few comments on the nature of this
problem, and on Professor Kemeny's approach to it.

I strongly agree with Professor Kemeny's statement that some
sort of give-and-take between the customer and the library
machine is necessary. I do not see how the object of the search
can be adequately specified otherwise. As a matter of fact, such
give-and-take is necessary even between scientists in order to
pinpoint a specialized topic.

On the other hand, I do not share Professor Kemeny's op-
timism with regard to the possibility of devising an acceptable
search procedure based on the classification of items according
to subject matter or any other predetermined features. As a
matter of fact, I am rather skeptical even about his proposed
division of the literature into subjects and branches. Today's
research activities often cross the boundaries between tradi-
tional fields such as mathematics, physics, and electrical en-
gineering. Certainly, I could not limit my own literature searches
to electrical engineering. I believe that the difficulties already
experienced by Professor Kemeny in his attempts to classify
items are not accidental. It seems to me that human knowl-
edge cannot be classified with sufficient precision for literature-
search purposes, and that even if it could be classified at any
given time, the classification would change too fast to be of any
real use. As a matter of fact, the acceptance by the scientific
community of the terminology required to classify a new area of
research often lags by years behind the research; and the most
interesting period of literature search is when the field is new.
Furthermore, even if a satisfactory classification could be made,
the specialized scientific manpower required to classify" and re-
classify documents would be prohibitive.

In searching for a more promising approach, we ma}" inquire
about the strategies that we have found helpful in our own
literature searches. I can only speak for myself, but I do not
think that I am too different from other people in this respect.
Personally, I usually start from some paper with which I am
familiar, and trace through the chain of references from paper

163

R. M. Fano

to paper. Unfortunately, such a search can lead only into the
past and not into the future. That is, by this method I can find
papers written before the one with which I am familiar, but not
papers written after it. On the other hand, if a library machine
were provided with the list of references cited in each item, it
could provide forward as well as backward references. That is,
it could give a list of the papers that refer to any one particular
paper. Then one could search the literature both backward and
forward in time, thereby reducing considerably the possibility
of missing an important chain of papers. I feel that the danger
of missing an important chain is more critical and more difficult
to avoid than that of finding too many papers.

The idea of a backward and forward search is not new. It
has been suggested independently by various people, but it
has not been implemented yet in a library, at least to my
knowledge. However, I understand that Dr. Kessler of Lincoln
Laboratory has performed some experiments along these lines
on ten years of the Physical Review. If he is in the audience, I
hope that he will comment on his work during the general dis-
cussion period.

Another important element in my own literature searches has
been the suggestions made by colleagues and librarians. In other
words, I feel that the library experiences of other people have
been very helpful to me. Clearly, a library machine could make
available to each customer the benefit of the experience of pre-
ceding customers. At the very least, the library machine could
maintain a running statistic of the items simultaneously se-
lected by previous customers as pertinent to some topic. The
topic does not have to be named for this purpose. One may
even conceive of the machine's maintaining a list of the litera-
ture items consulted by each individual customer; or, if the
number of customers is too large, by a selected group of cus-
tomers, say, a group of scientific leaders. For instance, because
of my interest in information theory, I should like to know
which papers Claude Shannon has consulted.

There are many ways in which data such as those mentioned
above could be used by a library machine in answering cus-

164

A Library for 2000 A.D.

tomers' queries. For instance, the customer might specify his
query to the machine by listing items which are pertinent to his
query and also items which are not pertinent to it. This would
be a way of defining his problem. Usually, however, this defi-
nition would not be sufficient. Therefore, the library machine
might suggest sequentially a test list of items for the purpose
of pinpointing more precisely the query, and request the cus-
tomer to specify which of the items suggested are pertinent to
it. This assumes that each item could be presented on a screen
very quickly from a central library. After receiving the answers,
the machine would provide a list of items that appear to be
strongly interrelated statistically on the basis of cross references,
selections by previous customers, and perhaps other data.

The point that I wish to make is that one can conceive of
literature-retrieval procedures that do not require the classifica-
tion or characterization of items according to any preconceived
scheme. As a matter of fact, I should be inclined to say that the
topic and character of a paper is often best specified by the list
of other papers which the scientific community considers to be
closely related to it. I believe that scientists are more likely to
agree on whether or not two papers are closely related than on
whether the same two papers should be classified under the
same preconceived heading. All a scientist would be asked is
whether he thinks that the two papers are related, not why.
There is a tremendous difference. For instance, it is evident to
any of us when two persons look alike, yet it may be very diffi-
cult to pin down why they look alike. I think the same is true
of papers.

If any retrieval system based on cross references and previous
customer selections could be made to operate, it would have the
great advantage that the data to be supplied to the machine
could be collected by clerical means without the use of special-
ists. This fact alone has such tremendous economic and man-
power implications that we cannot afford to overlook the possi-
bility that a system of this type could be designed.

The final point that I would like to make is that the retrieval
of literature is only part of the general literature problem. Even

165

G. W. King

if we can find the papers that are pertinent to our search, even
if we can obtain copies of them quickly and whenever we wish,
we are still confronted by a vary serious difficulty. A tremendous
amount of time and energy is consumed in reading badly writ-
ten papers, redundant papers, and papers using different or
confusing terminology. I believe that there is a need for better
and more timely books to supplant these papers.

In concluding, may I express the hope that Professor
Kemeny's interest in the problem of literature retrieval will be
a continuing one. This problem urgently needs the attention
of people with his imagination and mathematical competence
if it is to be solved in the near future.

King. When I reflect that in twenty-three years we shall
have 1984, I find it difficult to extrapolate through this dis-
continuity to the year 2000 a.d. Dr. Kemeny made the assump-
tion that there will be an exponential increase in the number of
books and articles produced in the decades ahead. If we accept
that assumption, at least we should by to make the extrapola-
tion self-consistent. Dr. Kemeny indicated that his prediction
applies to books on science and technology as much as to books
in other fields. Now these books must be about something.
They will not simply contain permutations of our current knowl-
edge. Since some, in fact, may be about automatic libraries,
we might expect sizable technological changes by 2000 a.d. The
power of data-processing machines, for example, historically
has increased by a factor of 10 ever/ year or so. But Dr.
Kemeny has not extrapolated the contents of books the way
he has their numbers. Therefore, not only can he say that
present libraries are obsolete, but also that his library is obsolete.

The numbers quoted by Dr. Kemeny are already out of date.
The National Bureau of Standards has demonstrated how pho-
tographic material can be reduced by a factor of 1000 linearly,
which is a factor of 1,000,000 in area. Enough work has been
done along these lines to show that all volumes of the Library
of Congress could be stored in a 1 -cubic-yard box. Of course,

166

A Library for 2000 A.D.

the question is what to do with it once you have it there, and
this question is still largely unanswered.

The density of storage has increased so much recently that
"tapes" referred to by Dr. Kemeny are giving way to discs or
three-dimensional storage devices, to which much more rapid
access is possible. Times of access are not minutes but milli-
seconds now. Exhaustive scanning of, say, 106 items, referred
to by Dr. Kemeny as taking 10 minutes, will in a very few
years take but a few milliseconds, judging from the nanosecond
circuits now in being. These figures are important because
sophisticated search will require hundreds or thousands of in-
dividual look-ups in executing a path to the desired information.

In the history of data processing, the cost of unit operation
has also decreased by an order of magnitude every three or four
years. At the present time, automatic libraries, although quite
feasible technically, are apparently too expensive to be built;
but judging from history, there is every reason to believe the
cost will be reduced to an acceptable level by the time details
of the system have been worked out. Dr. Kemeny' s estimate of
a range of one hundred million to one billion dollars may be
as many orders of magnitude high as his estimates of the tech-
nology are low.

The real difficulty, as Dr. Kemeny emphasizes, lies in the
theonr, and the need for a solution to the problem of search.
Something new must be conceived before the real facilities of
a library can be exploited automatically. Indexing, at least by
the classical methods assumed by Dr. Kemeny, has demon-
strably failed. Knowledge cannot be organized for any reason-
able period of time by a tree, such as the Decimal Classifica-
tion. On the contrary, new ideas, which are really the object
of information retrieval, result from the inverse of a tree, namely
the combination of ideas. Cross referencing, not dissection, is
the essence of a library.

Index headings are single words on which the whole intellec-
tual content of a document is mapped, obviously with loss of
information. Similarly, abstracts are really rejection filters pre-

167

J. G. Kemeny

venting the reader from being exposed to the detail he really
wants. It is evident that for some time indexes and abstacts
are the only tools at hand to search with, and will be features
of automatic libraries in the near future. But one hopes that
by 2000 a.dc we shall know very much more about the way
people search, or want to search, a library. The systems will
adapt themselves to the basic psychology of being informed
rather than to librarians' conventions. The electronic equivalent
of browsing and exhaustive scan are definite possibilities.

In conclusion, one might ask whether the human race has
to succumb to ever-growing libraries and automation. Perhaps
we could solve this problem in other ways. For example, by
removing books from libraries. Some of us used to use Crelle's
Rechnungtafeln for multiplication, where every product of four-
digit numbers was listed. But with the advent of cheap desk
calculators we reconstruct the product each time. Here is one
book that can be removed from a working library. This is an
extreme position, but a few basic books well constructed could
replace a great many of our present publications. In general,
the dying arts of writing reviews and treatises could keep the
growth of scientific and technical literature under control, at
least with the aid of a relatively modest automatic library.

Kemeny. I shall consider the two sets of comments in the
order in which they came. I agree with Professor Fano's com-
ments, in general, and feel that he brought up some excellent
points which I had not covered. However, I do disagree with
him on one point. I very much doubt that we can avoid the use
of experts. I too was tempted by the idea of tracing through
bibliographic references in doing literature search. But there is
a certain amount of evidence to indicate that this might turn
up the entire library. An illustration is found in Professor Fano's
analogy of two persons who look alike. If we start out with a
particular person, and take all people who bear a definite re-
semblance to him, and in turn consider all people who very
definitely look like each of these, and continue in this way, then

J 68

A Library for 2000 A.D.

with enough patience I think we shall discover that all human
beings are included eventually.

I am not so pessimistic as Professor Fano about the possi-
bility of getting experts to classify and abstract; at least my
experience with Mathematical Reviews indicates otherwise. Out-
standing mathematicians do abstracts of essentially everything
that is written in mathematics. They receive two kinds of com-
pensation: one is free copies of Mathematical Reviews, a very
modest reimbursement; the other is the opportunity to say
nasty things about the papers of other mathematicians.

I find it more difficult to reply to the comments made by
Dr. King. I suppose it is to be expected that an expert in hard-
ware will find problems of hardware easy and problems of cod-
ing more difficult. For a mathematician, the coding problems
seem rather trivial, whereas considerations of hardware avail-
ability seem more mysterious.

General Discussion

Kessler.* The research to which Professor Fano referred
is not only in its infancy, it is almost embryonic. I am there-
fore a little reluctant to quote results.

I agree with Dr. Fano that a retrieval system based on a rigid
classification scheme, although possibly applicable to a small
personal library, would be completely inappropriate for a li-
brary exceeding a certain critical size. It was as a result of this
conviction that we began working along lines other than rigid
classification. What we did was to place about ten years of the
Physical Review on tape, recording the title, author, and bibli-
ography of each article in those years. We have been perform-
ing three experiments on the information assembled.

The first experiment is designed to study whether two papers
may be said to be alike on the basis of their bibliographic

* Dr. Kessler was invited by the moderator to comment on his research
work, which Professor Fano had cited.

169

General Discussion

structure. If two papers have identical bibliographies, then they
probably have something in common. But complete identity is
never found. We are thus faced with the questions of what is
likeness and what is a suitable measure for likeness. These
questions lead to logical and programming problems, our cur-
rent concern.

The second experiment deals with the tracing of bibliogra-
phies into the past. Sometimes the tracing-back process pro-
duces an unlimited number of articles. But sometimes it does
close in upon itself, and a focusing takes place on a group of
what we might call seminal papers. For the case of certain sub-
jects in solid-state physics, we asked experts to list the ten
papers in their field which they considered particularly interest-
ing and important. There were definite indications that the
tracing-back process did pick up some of the papers suggested
by the experts, but I must emphasize that these results are very
preliminary.

The third experiment is concerned with the possibility that
Dr. Fano mentioned of tracing bibliographies into the future.
We take a paper and ask who used it. I feel that this is a very
effective way to classify a paper — not in terms of its subject
matter, but in terms of who found it useful and why. This
experiment turned up the most surprising result of all. Most
of the papers in our sampling from Phxsical Rexie^v, a much-
used journal, were not referred to at all within the first five to
ten years after their publication. In fact, two of my own papers
fell into this category.

McCarthy. I should like to consider the question of costs
in somewhat greater detail. I shall base my estimates on present
technology, not because I am uninterested in 2000 a.b., but
simplv because I am more interested in the possibilities for
1965.'

The first cost that I feel is relevant is the cost of information
transfer. A long-distance cross-country telephone call costs two
dollars for 3 minutes. The call is carried over a telephone line
which theoretically can transmit 2500 bits per second. The
telephone company also rents dataphones that transmit up to

170

A Library for 2000 A.D.

1200 bits per second. This is not bad, until we ask ourselves
what we can use on the other end to receive bits sent at this
rate. There are currently two possibilities: a video system or a
typewriter. The video system is unsatisfactory because it takes
too long to display a page, while the typewriter is too slow a
receiver and leads to an excessive cost of transmission. This
difficulty in present technology will have to be overcome be-
fore anything approaching Professor Kemeny's library can be
realized.

A second cost that is also very pertinent and important right
now is the cost of converting the present library to automatic
media. The Library of Congress has approximately twelve mil-
lion books. If we very conservatively assume that it would cost
an average of two hundred dollars to have a secretary or printer
convert each book, then the total initial conversion cost would
be two billion dollars. We can hope that if we wait a few years
we shall have a character-sensing machine that will be able to
copy books for much less; and I think this not unlikely. But I
did want to bring out what I feel are the present obstacles.

With regard to the possibilities for the near future, it is clear
that we cannot expect to put the Library of Congress into
machine form by 1965. But my calculations indicate that we
could, for example, put Mathematical Reviews into machine
form by 1965. Furthermore, if we had a time-sharing computer
system available, we could do as much searching of these re-
views by 1965 as could be programmed by then.

My last comment is that some of the people who have been
working in the information field, and I do not include myself
among them, may feel offended by the absence of mention of
their efforts. No mention has been made of the efforts to code
by descriptors as a way of getting around the tree-structure dif-
ficulties. Nor has there been any mention of the work done on
literature search by computers. One example that comes to mind
is the U.S. Naval Ordnance Test Station at China Lake, Cali-
fornia, which claims to be very satisfied with its IBM 704 sys-
tem, using magnetic tapes, for searching the literature on
rockets.

171

General Discussion

Bucklaxd. I always find talks like Dr. Kemeny's very in-
teresting, but discussing 2000 a.d. also leaves me with a some-
what anxious feeling. I have a hard enough time getting from
1961 to 1962. I think that I can point out three difficulties
connected with the suggested approaches, based on my own
experience.

First, any number of people have talked about reduction to
a 1-millimeter-square photographic image. This can be done,
but the difficulty is in doing it cheaply. Second, the time-sharing
concept of operating a computer in the give-and-take manner
suggested is not the same thing as two people sitting at a com-
puter debugging a program. It is a much faster and more de-
manding type of man-machine interaction. Finally, I think
there are many mistaken ideas about abstracts. It does not fol-
low that abstracts are easier to read because they are short.
They are not necessarily the answer to the search problem.

Gyftopoulos. I should like to question the assumption
that we need such a big monster as the National Library. I
often am concerned about the ratio of knowledge to amount of
material published each year. It seems to me that the purpose
of science is to summarize knowledge about our universe in as
few words as possible. I therefore would like to ask the speaker
whether he feels that something should be done to reduce the
volume of printed matter.

Kemexy. Surely that would be a desirable solution. But if
you are proposing to set up an authorized agency that will agree
to publish only one out of a hundred articles that are now being
published, I think that you are proposing a task much more
gigantic than the National Research Library that I suggest.

There have been some very interesting points made in the
discussion, and I should like to comment on them briefly, if
I may. In reply to Mr. Buckland, I am told by physicists (who
claim they know) that 1-millimeter-square reproduction is go-
ing to be practical very soon, if it is not already. Also, I am sure
that time sharing on machines, in a much more drastic sense
than we now know, is in the works. There is no reason, even-

172

A Library for 2000 A.D.

tually, why any number of people cannot use a machine simul-
taneously, just by the correct synchronization of requests.

I was a little troubled by Mr. BucldancTs remark on abstracts.
Perhaps fields differ greatly, but I do not know what mathe-
maticians would do without abstracts. It is obviously shorter
to read a three-paragraph abstract than it is to read a thirty-
page paper. The abstract clearly is not going to replace the
paper. The ultimate object of a literature search is to retrieve
all papers relevant to a particular problem or subject. But it
seems to me that the intermediate step of eliminating the ir-
relevant papers by sifting through abstracts is a method of
search that we use even today. In mathematics, anyway, it is
an invaluable method.

I should like to point out that Buckland and McCarthy took
positions at opposite extremes with respect to abstracts. Buck-
land suggested that abstracts really were of no use at all in
search procedures, while McCarthy implied that we should
build our entire retrieval system just for the purpose of search-
ing abstracts. My own sympathies lie in the latter direction.

McCarthy's point on the difficulty of getting books into the
library was very well taken. I should have mentioned this. Since
typing or transcribing onto tape by electronic means implies a
cost of several billion dollars, I took it for granted that photo-
copying would be the only practical method. This was part of
the motivation for the storage-and-retrieval system that I de-
scribed.

King. I should like to remark on the possibility of having
less to store by using the machine to reconstruct knowledge.
There has been some work on theorem-proving machines which
indicates that, starting with the right axioms, the reconstruc-
tion of results in geometry is possible. So you might be able to
reconstruct a great deal of that kind of literature.

On the issue of abstracts, I agree with Dr. Kemeny that much
depends on the field, as well as on the type of person you are.
But we have conducted experiments in which one group of
persons was given abstracts to read, while a second group of

173

General Discussion

persons was given the full articles, with time enough to read
only 10 per cent of the article. Later questioning has shown
that the second group comprehended slightly more than the
first. Perhaps this would not be true in mathematics.

Fano. I should like to correct a misconception that arose
earlier. In speaking about forming a bibliography by tracing
through chains of references backward and forward, I did not
mean that all the items found should be included. Obviously,
such a procedure would be unsatisfactory, as Dr. Kemeny has
pointed out. Instead, only those items should be included which
are strongly interrelated to one another on the basis of cross
references and use by readers.

I should like also to emphasize the problem of deciding what
data a machine must have available in order to be capable of
providing satisfactory answers to search queries. For instance,
I believe that storing the entire documents in machine-operable
form is not sufficient, and probably not necessary. The perti-
nence of a paper to a particular problem may well depend on
characteristics of a mathematical approach or a block diagram
which we cannot expect a machine to determine, at least in the
near future. Often I must study a paper, rather than just read
it, in order to determine whether it is pertinent to my problem.

Mooers. One important fact that has not been brought out
is that here we are in 1961, and many of the techniques that
could be applied have been around for twenty years or longer.
For instance, the Armed Services Technical Information Agency
(ASTIA) has shaken the library world with a great accom-
plishment of electronic computers. A major part of this ac-
complishment could be achieved with descriptors. Descriptors
and similar instruments were discussed in 1920. And the col-
lator, basic to the ASTIA technique, also has been on the scene
for a long time.

To address a comment specifically to the speaker, I am very
much against the idea of a National Research Library. It ap-
pears to me that this idea brings with it the danger of a national
bottleneck. A national cooperative library would be far more
desirable and could include the many kinds of local documen-

174

A Library for 2000 A.D.

tary material in paper form which would never be copied and
transmitted. It is easy to overlook the immense problems of
paper storage and handling associated with specialized collec-
tions. A cooperative library system, instead of consolidating
these collections at a central point, would maintain them at
their original locations and make generally available notice of
where they are.

With respect to the search problem, it was formulated
around 1946 or 1947, largely because of the efforts of the Ameri-
can Chemical Society. An M.I.T. man, J. W. Perry, did a great
deal to stimulate interest in the problem and encourage de-
velopments. It can safely be said that our techniques for search-
ing at the present time are far better than we are prepared to
use. Even with current resources, even with the services of
ASTIA, for example, we certainly can get more information in
paper form than can be easily assimilated.

Baumann. First I should like to comment on Professor Mc-
Carthy's remarks about output to individual users. At the recent
meeting of the Institute of Radio Engineers, the Sony Com-
pany of Japan exhibited a little video tape recorder which can be
manufactured for about ten thousand dollars now, and probably
for one thousand dollars some time in the future. This machine
could not only play back information recorded at the full trans-
mission rate of the telephone line, but could also be made to
exhibit one frame of the information for any length of time.

Second, I should like to comment on a method of search
which Professor Fano suggested but did not cover fully. It seems
to me that a person desiring information should sit down and
write an abstract of the information that he wants. Then use
could be made of simple statistical techniques, similar to those
developed by Luhn of IBM, based on frequency of word oc-
currences. This would be a first-class approach, even with scan-
ning rates that are currently attainable. Staying with the 1965
picture and using about one hundred photoreceptors in parallel
working at around 10 megacycles, we can envision a scanning
time for the Library of Congress of approximately 3 hours.

Thus if we had a machine capable of searching through ab-

175

General Discussion

stracts and making statistical analyses of word frequencies, we
should have at most a 3-hour delay in obtaining all possible in-
formation in the Library of Congress. Maybe some people are
so impetuous that they must have their answer immediately
upon asking for it, but I should just as soon take a coffee break
while waiting for mine.

Kemeny. The problem is not just to scan. While you are
having your coffee break, someone has to digest the tremendous
amount of information generated. The time required to figure
out what it is that you have scanned is an order of magnitude
greater than the time required simply to peruse it.

Kixg. I agree with Dr. Kemeny.

Wegner. I feel that the library of 2000 a.d. should be part
of a larger computer system. The individual could submit his
research problem to the computer and have it solved for him
via the give-and-take process. The computer then would be
able to invent its own classification scheme and would play a
more active role in the development of the library than Pro-
fessor Kemeny suggested. This would be a completely computer-
oriented approach. In this connection, I should like to bring at-
tention to Dr. King's Freudian slip in his mention of 19S4. In
fact he is looking forward to the day when IBM will be "Big
Brother."

Abrahams. I should like to raise a question about a basic
limitation that I think Mr. Wegner recognized. Will the hu-
man being who is searching for information be able to absorb
all that he receives? As the amount of relevant information in-
creases, it may be expected that the person will absorb a smaller
and smaller proportion. The aim should be more than find-
ing out what is relevant. It should be finding out what is good
among that which is relevant.

A second and associated problem is that if we have a central
library, then no matter which type of retrieval system is used,
the system will have a bias. The system has to be organized so
that it will lead to the selection of certain kinds of things. This
is especially true of a retrieval system based on what people
have looked at before.

176

A Library for 2000 A.D.

Fano. We can readily think of many learning phenomena
that may occur in a library machine which makes use of the
experiences of previous users, some desirable and some unde-
sirable. For instance, items may effectively disappear from the
library through nonuse. We can conceive also of situations in
which a machine, by combining the experiences of various peo-
ple, may discover a significant relation between papers that no-
body had yet appreciated. Clearly much experimentation will
be necessary before a satisfactory learning library machine can
be designed.

Incidentally, I feel that there has been a serious imbalance
in the information-retrieval field between theoretical and ex-
perimental work, perhaps because of the high cost of experi-
mentation.

Fredkin. I am hopeful that we shall still have some tradi-
tional libraries in the year 2000 a.d. If books were to be re-
placed by telephone dials and television sets, I should miss
having a place to walk down to occasionally to take out a book.
I think that books are very well engineered devices. They are
easy to use, and I cannot conceive of a better way right now
to store information. I am sure that there are better ways, but
it is important that in looking for them we not commit our-
selves to some fancy, cumbersome system for retrieving and
viewing information. I think that we have enough television
already.

Locke. As a librarian, I cannot resist the temptation of
closing the session at this point. Thank you Dr. Kemeny, Dr.
Fano, and Dr. King.

References

Juhasz, S., Ed., Proceedings of 2nd Annual Conference, National Federa-
tion of Science Abstracting and Indexing Services, NFSAIS, Washing-
ton, D. C, 1959.

Luhn, H. P., "A Statistical Approach to Mechanized Encoding and Search-
ing of Literary Information," IBM Journal of Research and Develop-
ment, Vol. 1, 309-317 (Oct., 1957).

177

References

Luhn, H. P., "The Automatic Creation of Literature Abstracts/' ibid.,
Vol. 2, 159-165 (Apr., 1958).

Luhn, H. P., Selective Dissemination of New Scientific Information with
the Aid of Electronic Processing Equipment, IBM Advanced Systems
Development Division Report, Yorktown Heights, N.Y., Nov. 30, 1959.

Maron, M. E., and Kuhns, J. L., "On Relevance, Probabilistic Indexing
and Information Retrieval," Journal of the Assoc. for Computing Ma-
chinery, Vol. 7, 216-244 (July, 1960).

Mooers, C. N., "The Next Twenty Years in Information Retrieval: Some
Goals and Predictions," Proceedings of the Western Joint Computer
Conference, San Francisco, 1959, The Institute of Radio Engineers,
New York, 1959, pp. 81-86; also American Documentation, Vol 11,
229-236 (July, 1960).

Mooers, C. N., "Some Mathematical Fundamentals of the Use of Sym-
bols in Information Retrieval," in Information Processing, Proceedings
of the International Conference on Information Processing, UNESCO,
Paris 15-20 June 1959, UNESCO, Paris, 1960, pp. 315-321.

Newman, S. M., Problems in Mechanizing the Search in Examining Pat-
ent Applications, U.S. Patent Office, Office of Research and Develop-
ment, Washington, D. C, 1956.

Perry, J. W., Kent, A., and Berry, M. M., Machine Literature Searching,
Western Reserve University Press and Interscience Publishers, Inc., New
York, 1956.

Ray, L. C, and Kirsch, R. A., "Finding Chemical Records by Digital
Computers," Science, Vol. 126, 814-819 (Oct., 1957).

Taube, M., and Wooster, H., Eds., Information Storage and Retrieval,
Columbia University Press, New York, 1958. Contains further bibliog-
raphy.

178

5

The
Computer

in the
University

5

The
Computer
in the
University

Speaker Alan J. Perlis

Director of the Computation Center
Carnegie Institute of Technology

Discussants Peter Elias

Head, Department of Electrical Engineering
Professor of Electrical Engineering
Massachusetts Institute of Technology

J. C. R. LlCKLIDER

Vice President

Bolt Beranek & Newman Inc.

Moderator Donald G. Marquis

Professor of Industrial Management
Massachusetts Institute of Technology

180

Perlis. There are currently over one hundred computers
installed in American universities. Probably two dozen or more
will be added this year. In 1955 the number was less than
twenty-five. Of course, the number of computers employed in
industry has increased by more than this factor of 5, and those
in government have increased at an even greater rate. Never-
theless, considering the costs involved in obtaining, maintain-
ing, and expanding these machines, the universities have done
very well in acquiring hardware with their limited funds.

Actually, much of the credit for the scope of the computer
collection in universities today is due to intelligent planning by
IBM and to the stubborn efforts of some Washington patriots
at the National Science Foundation who took advantage of a
period of national dismay to arrange for funds so that universi-
ties not only could have computers, but could have good ones,

181

A. J. Perlis

i.e., computers having large storage and high processing speed.1
Of course, the NSF and IBM decisions were not made in a
vacuum. People in XSF and IBM had before them considerable
evidence of the intelligent and farsighted use to which some
universities could put computers. Certainly, M.I.T., from the
Bush analyzer to 'Whirlwind to the present computer-rich en-
vironment, has provided abundant evidence that universities
could indeed do wonders with these machines.

How Universities Are Using Their Computers

Xow that computers are generally available at universities,
it is time to assess their real purpose there. The basic purpose,
at present, is to do computations associated with and supported
by university research programs, largely government financed.
Some of this research makes marginal use of the computer, e.g.,
computation of data arising during the performance of the re-
search. Some uses require the computer s full-time participation,
e.g., simulation studies and certain numerical analysis research.
Some research, started in the former category, ends in the latter.
Less often encountered is the reverse trend. Sad to state, some
uses occur merely because the computer is available, and seem
to have no higher purpose than that.

Some of this research is quite important and its consequences
ramified. Some uses invent computation techniques that will
probably have a more profound effect than the original research
objective. With only a few users does the computer itself (or,
more precisely, the insight it takes to use it) invest a deeper
interpretation of the development being studied.

Testifying to the growing involvement of computers in gradu-
ate research, one notes the increasingly large percentage of M.S.
and Ph.D. theses in psychology, the physical sciences, and en-
gineering that contain reference to tables and approximate
solutions to equations produced on a machine; also to codes,
often specified in the heady symbolism of a language like
Fortran, whose execution has produced these important
results.

Most university computation centers produce annual reports

182

The Computer in the University

that document the diversity of applications successfully carried
out or in progress on their computers during the year. The list,
and some of the results, are quite impressive. They testify to
the value of a computer as a research instrument in the uni-
versity. In a few years it is not unlikely that the computer may
have settled immutably into our thought as an absolutely
essential part of any university research program in the physical,
psychological, and economic sciences.

As a consequence of the prolonged contact of research stu-
dents with the computer, interest in numerical analysis has
increased, and courses are now available in this subject in all
universities possessing computers, and in many which do not.
Most of these courses are on classical numerical analysis. Few
require or encourage continual contact with the computer. This
is partly because the time required to teach programming to
the students would seriously reduce the contact hours available
for learning numerical analysis; and it is partly because the ma-
chine time is just not available, either in length of period or
frequency.

This lack of contact is a serious delinquency, not because
numerical analysis cannot be taught, appreciated, or learned
without a computer; but rather because with the student pro-
gramming continuously it is possible to ( 1 ) reduce for him the
programming composition effort to that required of any mode
of communication in which we are expected eventually to be
fluent (such as writing in English), and (2) reveal the impor-
tant developments of the subject through real problems in which
the students are personally involved. These real problems il-
lustrate the developments as no blackboard or textbook ex-
ample can. The instructor's direct administrative involvement
with these problem efforts can be reduced greatly by having the
exercises graded by the computer. I shall say more of this pos-
sibility later.

Through such a course the engineering and science student
acquires a feeling for, as distinct from a catalogue of facts about,
such things as infinite sequences, and convergence rates from
successive approximation schemes; the value of an axiomatic

J 83

A. J. Perlis

approach to finite dimensional vector spaces (shown to be an
aid in the concise description of processes for solving systems
of linear equations and characteristic-value problems); the im-
portance of constructive existence theorems in differential equa-
tions; the relationship between random processes in nature and
numerical methods for solving their representation in mathe-
matical form as it arises, for example, in partial differential
equations; and the importance of heuristic, as distinct from ana-
lytically justifiable, approaches in practical numerical analysis,
e.g., a very beautiful method of estimating error in numerical
quadrature, due to Lanczos,2 which avoids high-order deriva-
tives or their equivalent, but which has no precise analytical
justification.

Finally, the student, through his programming, becomes
aware of the importance of representation itself as he describes
computational processes that are strongly dependent on the
way information is organized and supplied to the process, e.g.,
in Chebychev approximation. Among the linguistic representa-
tions with which he becomes acutely acquainted are those
languages (complex but unambiguous) for specifying algorithms
for the machine — the so-called problem-oriented computer lan-
guages. At Carnegie Tech the two-term senior-level numerical
analysis course requires each student to program and run about
twenty numerical solutions of diverse problems in a year, not
nearly as many as would be desired, however.

Whereas numerical analysis has fitted naturally into the
educational program of the engineer and scientist, computer
programming per se has not. Generally, programming has meant
either dilute numerical analysis, machine code learning, or
logical design of computers, and sometimes fantastic mixtures
of all three. In many schools the intellectual content of "pro-
gramming" is considered so devoid of university material that
only because of its practical utility is it taught, and then, often,
only in after-hour short courses given without credit, or inspira-
tion, or perspective. Such courses are often summed up with
a triumphant exercise in coding, e.g., use of Simpson's rule to

184

The Computer in the University

obtain approximations for ir. These courses are taught as
though a computer were merely another shop tool.

In the past three to four years some economics departments
have begun the systematic exploitation of the computer as part
of their research and educational programs. Some of those re-
search projects are primarily analysis of data; others utilize the
computer to simulate organizations, their actions, and policies;
others study mathematical models of economic practice. The
models are often combinatorial, and it is amazing to see the same
students who have great difficulty with the calculus manipulate
scheduling models represented as graphs — and glibly program
clever algorithms that compute values for the graph properties.

One of the most important applications to economics edu-
cation is the development of the management game as an educa-
tional tool. The injection of the game into the educational
process has stimulated the student's appreciation of the deci-
sion-making problem in management. There are, of course,
many kinds of management games and diverse opinions about
their role in education. A faculty whose members are fluent in
programming, because it has been made simple for them to
learn and practice to be so, will see more clearly, in a funda-
mentally different way from the novice, the ramifications of a
complex, dynamic game played with a computer. If the com-
puter is made available to the students as a matter of course,
both students and faculty will see in it a really powerful edu-
cational tool. The Carnegie Tech Management Game3 has be-
come not a (marginal) laboratory exercise but a pivot for
management education. Indeed, the authors of the game believe
that it will develop, through actual practice, the following skills
in the student:

"1. An ability to abstract, organize, and use information
from a complex and diffuse environment.

2. An ability to forecast and to plan.

3. An ability to combine the role of generalist and specialist.

4. An ability to work effectively with other people."

185

A. J. Pedis

The game involves a good deal of information transfer and
sharpens the ability of the player to analyze in a realistic time
framework large amounts of information. Eventually the player
himself will be able to program, or to call upon subprograms
that will process the data he gathers from the game-dominated
universe. Thus the computer will serve as both the environ-
mental source and analyzer of information. Having a function-
ing "doll's house" available, the faculty, without fear of by-
passing the very important practical examples, can donate their
wisdom to emphasizing important theory and economic rela-
tionships that, when coupled with student .drill, give the proper
balance to an education.

Educating the Student

The preceding uses have several common features. With the
exception of the cited economics applications, they are charac-
terized by extensions of previously used methods to computers;
and they are accomplished by people already well trained in
their field or specialty who have received most of their training
without computer contact. It is, however, the thesis of this paper
that the programming and using of computers deserve an early
appearance in the university curriculum for the educated man.
Indeed, such training should appear not only early but also
often during university training.

The question might be asked about the way students are
educated to use the computer in a university today. Three fre-
quently used methods are the following:

1. When an application makes it necessary. When an ap-
plication arises, the student is trained, through example,
to recognize and utilize the computer for this and
similar applications. Clearly, extensions of method will
occur to the student, but he is not likely to produce
major thematic reorganizations of theory and material
motivated through computer experience.

2. When the student's mathematical maturity permits the
absorption of a catalogue of numerical techniques that are

186

The Computer in the University

of general utility in computer applications. This is the
standard intent of the numerical analysis course, usually
offered in the junior or senior year to science and en-
gineering majors.
3. Self-education when the student thinks he cannot do his
assignments without use of the computer.

I wish to remark that the first method suffers from a de-
pendence on instructors who for the most part, even at this
time, are themselves inexperienced with computers; it also suf-
fers from the deficiencies of any piecemeal education on how to
use any device or theory. The importance that the Ford Foun-
dation places on computer education of engineers is evidenced
by the healthy programs it supports at Michigan4 and M.I.TT;
and much of this support is directed toward the computer edu-
cation of engineering faculties — still largely delinquent in un-
derstanding computation and dreadfully unprepared to teach it.
The second method suffers from identifying the domain of
computer application too closely with the standard problems
and techniques of numerical analysis. The third method suffers
because it possibly illustrates that universities and their carefully
nurtured degree programs are not in any case necessary to our
society in solving its problems.

Some universities have used a fourth approach. A credit
course invoking the use of some automatic programming lan-
guage is provided. Fluency in "conversation" with this language
and clear understanding of the language's grammar are the in-
tents of such a course. Here, too, the approach suffers from
limited and even misguided pedagogic objectives; and the re-
sult is a student well conversant in, say, Algol 60, but still very
likely uneducated as to the scope of computation.

These approaches all fall far short of what is urgently needed.
Whatever appreciation of computers they invoke comes too late
in the educational program and is inevitably too parochial. It
is thus important to state quite clearly when the computer and
the student should first meet and what should be the nature
of their first acquaintance. As others have stated in other con-

187

A. J. Perlis

texts, the product of a university education should receive
training directed to the development of sensitivity, rationality,
and an intelligent table look-up procedure.

Sensitivity, as the poet MacLeish so aptly has said, is a feeling
for the meaning or relevance of facts. Rationality is fluency in
the definition, manipulation, and communication of conven-
ient structures, experience and ability in choosing representa-
tions for the study of models, and self-assurance in the ability
to work with the large systems that are unfortunately necessary
for modeling and solving the important problems of our times.
Table look-up, of course, refers to the mechanism for gaining
access to a catalogue of facts and problems that give meaning
and physical reference to each man's concept of, and role in,
society. While the computer may conceivably play a small role
in the development of human sensitivity, it is quite critical to
the other two developments. Indeed, no other mechanical in-
strument combines so well the theoretical and practical balance
necessary to these two developments.

First Course in Computers

Consequently, it is felt that the first student contact with
the computer should be at the earliest time possible: in the stu-
dent's freshman year. This contact should be analytical and not
purely descriptive, and each student during this first course
should program and run or have run for him a large number
of problems on the computer. At least in engineering and sci-
ence programs, this course should share with mathematics and
English the responsibility of developing an operational literacy,
while physics and chemistry develop the background toward
which this literacy is to be applied. In a liberal arts program
the course could be delayed until the sophomore year, but cer-
tainly deserves inclusion in such a program because of the uni-
versal relevance of the computer to our times.

One must admit that an optimum content for the freshman
course is not yet known. Few such courses have been offered.
What is said here is based on limited experience at Carnegie
Tech and on an academic, admittedly rather arrogant intuition

188

The Computer in the University

concerning the importance of machines. The Carnegie course
occupies two semesters. It was offered to forty-six students in
the fall term, and to two hundred additional freshmen this
spring, term. It is felt essential that each student do about
twenty problems on the computer per semester. The grading
of 5000 programs in a 16-week period is a task well beyond the
endurance limits of most instructor cadres, so the computer
has been programmed to accomplish this task itself as well as
to maintain, as a matter of course, the grades of the students.

During the first term the students wrote programs in a sym-
bolic machine code, the Carnegie Tass system; and during the
current term they are writing their codes in Gate, the Carnegie
Algebraic Language system. Coding in machine language, they
are taught mechanical algorithms of code analysis that enable
them to do manually what the Gate translator does automati-
cally. In particular, they are becoming adept in decoding com-
plex logical relations to produce branching codes and in man-
ual decoding of complex formula evaluations by mechanical
processes. The intent is to reveal, through these examples, how
analysis of some intuitively performed human tasks leads to
mechanical algorithms accomplishable by a machine.

The processes are given at two levels: first a description using
"informal" bookkeeping methods and tabular languages, and
later a quite precise description using completely formal rules
and representations in a formal programming language. The
need for a formal system arises naturally, and the concept ap-
pears intuitively reasonable.

In the course of analyzing these complex formulas, the stu-
dents are given processes represented by flow charts as defini-
tions of certain functions: e.g., the square root (Newton's
method), the logarithm, and the exponential. For the latter
two functions, the computations are not those in common use
as subroutines on machines but have the advantage of being
definitions that use only arithmetic operations and square root.
It is proved analytically from the flow chart that each of these
processes actually is the function claimed and that the process
in the computer is of necessity only a "rational" approximation.

189

A. J. Perlis

To do this, the student is given some elementary theorems on
limits of sequences and the Weierstrass development of real
numbers from nested sets of decreasing intervals. These concepts
are illustrated through computation exercises right on the ma-
chine. The logarithm process yields for any x > 1 a sequence
In (8, x) for which lim In (8, x) = In x. It is shown in Figure L
Using Figure 1 and flow-chart operations, a proof is con-
structed that for x, y > 1, In (xy) == In x + In y, and for
0 < y < l,ln (1/y) = — lny. This provides the computation
for all y > 0. Unfortunately the proof process is not itself me-
chanical.

a<-u- 1

2a

b S^\b-a\ <s)-)

Yes

£ = ln(8,x)

Figure 1 Flow Chart of the Logarithm Process

The construction of proofs of properties of functions from
flow-chart definitions has obvious limitations in analysis but
has much merit in a programming course, since the construc-
tions are derived from specific algorithms that the students
manipulate themselves. Using Newton's method, it is possible
to show that if a > 0, In x = a has one solution. The initial
estimate is x0 = 1, and

xn+1 - xn (1 + a - In x„) for n = 0, 1, 2, • • •

It is then shown that {xn} is monotone increasing and
bounded from above, while In xn < a. To the student it fol-
lows simply that when x = lim x„, In x = a.

n— > 00

Thus the definitions of these processes (the square root, the
logarithmic, and the exponential) are chained. A purpose in
using these definitions is to reveal one of the important prin-

190

The Computer in the University

ciples of programming: the definition of complex processes by
rational construction from simpler processes already given.
I seek to make a point here. In a first course in programming
it is the elements of programming that must be developed, il-
lustrated, and emphasized. Only in subsequent contacts with
the computer should a "practical" encyclopedia of numerical
techniques, identified as such, be given. A well-organized set
of problems for such an elementary programming course, as far
as I know, does not yet exist.5

Elements of Programming

What then should the student be taught about program-
ming? What are the fundamentals? The following list is ad-
mittedly incomplete but represents some fundamentals that in
my opinion the student should acquire in this introductory
course.

Parametrization. Every program contains at least one pa-
rameter, and binding it is necessary before the program can be
run on a machine. Furthermore, within limits, the more param-
eters present, the more powerful the program, i.e., the more
tasks it can accomplish by binding the parameters. There are
numerous examples that are used: nth root instead of square
root, ordering n numbers instead of 100 numbers, the number
of ways of making x "cents" from coins of denomination du
d2, . . . , dk. However, the student is cautioned that the empty
computer is the most flexible program possible but requires a
great deal of parameter binding prior to each run.

Iteration. Most programs contain cycles of operation so
that the same instructions can be utilized on varying data.
Teaching the student to think in terms of cycles or repetitive
processes is most important.

Recursion. Instruction sequences may utilize themselves as
subfunctions of their own operation. Recursion permits many
processes to be programmed from the outside in. Many proc-
esses are most naturally thought of in this way, e.g., differentia-
tion, statement scanning, game playing, etc. These examples
illustrate to the student the indeterminate delay that often

191

A. J. Perlis

arises in execution of a chain of operations when it is incom-
plete at the time encountered.

Definitions. The development and cataloguing of highly
useful code sequences. These may then be substituted into code
for an occurrence of their name with variations among substi-
tution instances induced by controlling parameters. These code
sequences may, of course, contain definitions of other code se-
quences so that complex schemata may be generated.

A Priori Attention to Eventualities Regardless of Likelihood.
The development of a mechanical attitude toward mechanical
processing requires that processing of even unlikely cases be
provided for. The programming of the solution of a quadratic
equation and the analysis of the general second-degree equation
in two variables to select conic sections are good examples where
many students fail to provide for eventualities. The willingness
to program through a tedious collection of tests on special
cases must be fostered.

Representation. The data organization in each problem
may have a "natural" representation for one phase of a com-
putation and a different one for another. Transformations be-
tween representations must always balance in the student's
mind the ease of processing versus the work involved in trans-
formation. For example, the students were asked to compute
the squares a knight might move to in 7 moves from a given
position on a chessboard. The checking of admissible positions
is most easily conceived of using an 8 X 8 matrix; the exhaus-
tive check of all moves is naturally realized by a simply indexed
search over a 64-position vector. (The students should decide
which representation to use; or indeed, when and how often
to translate during the running of a program.)

Mechanical Language. Each process, before it is run on a
machine, must be described in some specific computer lan-
guage. The nature of such languages is best illustrated by cata-
loguing their properties as they simplify the coding of certain
types of problems. Here the student becomes acquainted with
the inherent limitations and advantages of mechanical syntax,

J 92

The Computer in the University

and also the ease with which he can adapt this syntax to his
variable needs.

The student learns that often the instinctive way he chooses
to describe a process is not satisfactory on a machine. But he
also learns that there are ways by which he can produce such
descriptions in a form acceptable to a computer, and that they
really are not strange. With training, he tends to discover ever
more naturally these mechanical descriptions for mechanical
processes.

Simulation. It is important for the student to realize that
a process possibly not even associated with, and operating ex-
ternal to, a computer may to varying degrees of approximation
be imitated on a computer, e.g., game playing, learning, and
simple process control.

Proof. The student should, through examples, develop
proofs that certain algorithms he constructs actually do the
task for which they are designed. Unfortunately he can only be
guided by example and by appeal to a catalogue of useful proof
devices; no universal proof method exists. Proving that an algo-
rithm, claiming to be the Gauss elimination process, actually
solves a system of linear equations is the most instructive
pedagogic device that I have yet come across to explain the
tactics used in this method of solving linear equations, the sig-
nificance of zero determinants, and the uniqueness of solu-
tions.

Grading and Guiding the Student

In a course such as the one described it is important that the
student be quickly graded on his progress and guided past his
mistakes. The computer can grade his progress in a generally
satisfactory way. However, the identification of mistakes in pro-
gramming and the isolation of the source of a student's mis-
understanding in the course material are tasks that are incred-
ibly complex, and no good automatic solution is currently
available. The Carnegie computer grading program is really
quite simple and works somewhat as follows:

J 93

A. J. Perlis

1. The instructor writes a class problem program to include
the generation of k sets of data and answers. Each data-
and-answer set is identified by an index Kj<k. The
information is stored from within the program in a
magnetic-disc memory, put there by the teaching admin-
istration routine.

2. Upon loading, a student solution program is supplied a set
of data: its answers, when obtained, are compared with
those associated with the class problem. Satisfactory
agreement (up to a variable tolerance') of the answers
over the k sets of data results in a pass grade.

3. Should the agreement be unsatisfactory, as specified by
these tolerances, the index / and some diagnostic informa-
tion are made available to the student. The grading may
then be suspended or continued through additional data
sets, depending on the problem. When the student fails
to work a problem, he must try again.

4. The grading program keeps a record of the student's per-
formance. He ma}-, of course, work problems at any rate
that he chooses. Half-right answers are not tolerated; the
student is never in doubt about where he stands. He
either got it right, or he did not. Therefore he cannot
continue through the course in a partial fog, assuming
knowledge rather than acquiring it.

While not yet in the current program, the diagnostic informa-
tion can be used to control the grading program so that the
student is forced to follow some corrective problem sequence
if his errors persist in subsequent trials. The corrective problem
sequence may itself be produced from within the computer.
Both explanatory text and problem statements may be indi-
vidually printed for each student on a high-speed printer. Thus
texts are printed only when necessary. Their storage on a com-
puter disc file would total about 50,000 computer words. (I
shall say a little more later about the figures and costs of Car-
negie's computing system, to shed some light on this matter. )
More importantly, the grading program can maintain important

194

The Computer in the University

counts of the frequency of use of text branches so that sensible
improvements can be made in the organization of the net of
problems and solutions.

Since in this first course the concepts of computer program-
ming are new. it is these that must be taught. Discussion of the
programming languages employed must be subordinated. In-
deed, the language is learned as a convenient set of signs for
describing these important processes. In the course of learning
the techniques, the student is taught the syntax by the highly
mechanical routine of the teaching machine, but it is the prob-
lem he programs that really regulates his fluency with com-
puters.

The first and most critical stage of the computer's role in
the university, it seems, is clear. It is to train entering students
in the theory of computation through the development of the
concepts of programming. To do this, the computer plays a
partners role with the teacher in sequencing the student
through a set of problems and language definitions representa-
tive of computation principles.

At the completion of this first course, the student should now
be able, for example, to recast his learning of. saw the calculus,
so as to be able, himself, to devise mechanical methods for a
computer to differentiate elementary expressions. The separation
between the introduction and consequence of definitions and
the mechanical manipulation that the student does so much
of should be easy for him to make. The concept of the teaching
program for calculus should appear natural and so should the
planning procedure for accomplishing it. To be sure, the ma-
turity- in relating mathematical developments, e.g.. that of al-
gebra and calculus in the integration of rational functions, is
still not yet developed: but such as it is. the maturity will be
applied to an intellect already attuned to the rationale of proc-
ess definition and synthesis.

To such students the following problem would be a challenge:

Devise an algorithm and write a computer program to generate,
upon request, a sequence of integrals involving trigonometric

195

A. J. Perlis

functions that could serve as solvable drill problems for future
students in this course. The members of the sequence should,
in a sense for you to determine, increase in difficulty of inte-
gration.

It would be difficult for most, and for some unfeasible; but for
almost all the basic approach would be obvious.

I do not of course mean to imply, here, that calculus should
be learned only through the development of those processes,
and certainly not only from their execution; but the recogni-
tion that mechanical processes are there and, given the effort,
can be designed and constructed will increase the student's per-
ception of what is important and relevant, of what is basic and
what is derived.

I should expect students, later in their education (in a course
in complex variables), to give serious thought to the properties
of an algorithm that would perform on a computer the
formal evaluation of certain definite integrals by contour inte-
gration using the calculus of residues. The algorithm, of neces-
sity, will be recognized as only a partial one. Parametrized, it
will be a proof that certain classes of integrals can be so treated.
The student's limited success with the algorithm will not cause
him to lose interest in such use of a computer, but rather pre-
pare him for what I regard as the ultimate role of the machine.

The Ultimate Role

I should certainly expect students to attack with interest the
problem of creating within the computer a technician for do-
ing mechanical analysis. This I consider to be the ultimate use
of the computer in education — the role of technical aid to
match that already acquired of teaching aid; an aid in mathe-
matical manipulation and in obtaining proofs. Such a technical
aid could help a student perform tiresome manipulations (e.g.,
those required by many of the problems in Whittaker and
Watson ) .

I envisage it as a fundamental task of the university comput-
ing center to develop the programming language in whose terms

196

The Computer in the University

such a technical aid may be programmed. This programming
language will require more flexible notational abilities than any
now extant. But it is the university laboratory that must de-
velop the language.

It goes without saying that the university computer or com-
puters on which this educational program is to be accom-
plished must possess certain salient features: high speed multi-
programming ability, multiple input-output facilities that are
many in number and low in price, high-speed printers, and large
random-access storage units. The computer should be econom-
ical to use, so that the burden of its upkeep will not require it
to be used less as an educational instrument than as a device
for contract research computing.

The remote input-output consoles are critical for the success
of this program. Past use has emphasized automatic computing,
that is, a priori directed machine control of the entire comput-
ing process. Naturally, as computing power increases, there is
an accompanying tendency to increase the automation of prob-
lem solving both in scope and depth in a given situation, for
example, automatic error analysis in the solution of ordinary
differential equations. One reason for this is that the expense
attached to modern, large-scale computer operations does not
make it feasible for the problem solver to be "in the computer
loop" when his program is run.

In the educational program it is essential that the student
be able to function in the computer loop without at the same
time causing an exorbitant increase in everyone else's com-
puter cost. Indeed, we may regard this as one of the prime in-
tentions of the educational computer system being described.
However, though the student is in the operating loop, it is
specified that he must not control the loop; i.e., he enters when
permitted and leaves when told.

Thus it is reasonable to consider that the computing load
consists of a pattern of running codes some of which are moni-
tored from an array of consoles. The rate at which characters
enter from the console will be absurdly low, being about two or
three characters a second (by the hunt-and-peck method) . Con-

197

A. J. Perlis

sequently, input from the console will be by a code organized
to maximize information transmitted per character, and may be
quite different from the standard problem-solving languages we
now use. The computer will require about 1 50 microseconds to
process each character at transmit time. Thus, in a message,
about 300 microseconds every second will be devoted to main-
taining records about input data (per console). The input in-
formation will naturally be processed as a group only after a
terminal signal is received. The ultimate university computer
must therefore possess this satellite facility so that many stu-
dents may have simultaneous access to the computer.

I should like to cite some cost figures that may put things in
perspective. Suppose, as part of its computing system, a uni-
versity has one or more high-speed (600-line per minute)
printers, costing $12,000 a year per printer. Suppose it also has
1000 students taking three courses each, and each student re-
ceives (upon application) an average of three texts in each of
his courses. Suppose the texts contain 100 pages of 40 lines
each. How much does it cost to have the computer provide the
students with their texts? The cost per year is about $500 for
printing time, that is, $500 for all 1000 students.

Suppose the university wishes to store a text in its computer.
A magnetic-disc file on some machines (not all) costs around
$20,000 a year. The capacity of such a disc file is about 2 X 107
characters. If the university stores a text in this memory, it
utilizes on the order of 5 X 104 characters, and the (opportu-
nity) cost of the memory is $50 per year.

Suppose the university employs a secretary, who can type at
the rather fantastic rate of 200 characters a minute, to enter a
text into the computer from a console; then the computer will
require, say, up to 500 microseconds per character to accept
this information and some programming time to store it away.
The total time necessary- to read in a text of 50,000 characters
will be of the order of 1 minute of the machine's precious com-
puting time, assuming that it is doing other work simultane-
ously.

This implies that the computer is not as ferociously expen-

198

The Computer in the University

sive an instrument for education as it might seem at first
glance. Once the cost of the machine itself is within reason, it
is not at all out of the question to consider a computer that
serves only for educational purposes and not as a horse for con-
tract research. At a purchase cost of the order of, say, $300,000
or $400,000, the university can operate a computer for the edu-
cation of its students quite economically. Indeed, the idea of
keeping texts on the computer is one that we at Carnegie in-
tend to exploit with our new machine for courses in program-
ming and the calculus.

Given then the appropriate computer, the capability of de-
veloping programming systems, the proper freshman course,
and possibly a good follow-up program, the computer will
achieve its ultimate role as handmaiden to scholarly university
activities. The professor can then relinquish the role of drill
sergeant and assume the role of philosopher-teacher.

Panel Discussion

Elias. Professor Pedis' pretty picture of a happy symbiosis
between students and computers is appealing, and so of course
is his proposal to make use of the computer to do both the
numerical and the symbolic detail work in mathematics. I have
little doubt that much of what he has described will come to
pass. There are a number of distinguishable elements that he
has artistically composed to form the picture, and I should like
to spend my time in distinguishing them and setting up some
attempt at priority and purpose among them. They have dif-
ferent impacts on the use of computers in the university.

If the computers, together with sufficiently ingenious lan-
guages and programming systems, are capable of doing every-
thing that Professor Perlis describes — and I believe they are
(and more) — then they should be ingenious enough to do it
without the human symbiote being obliged to perform the me-
chanical chores which are a large part of current programming
effort, and which are a large part of what must now be taught

199

P. Elias

in the introductory course that he proposes. Adequate com-
munication between man and machine, which is a necessary
prerequisite to the most interesting parts of the program, is not
attained so long as it is necessary for people to learn things that
are of no intrinsic interest merely to communicate.

If the notations in use in mathematics, we discover, are not
the notations we should be using, if they are not as explicit,
as unambiguous, or as convenient as they should be, then by
all means they should be modified for the sake of the people
who work with them year after year. However, if we find a no-
tation that is well adapted to the particular class of problems
in which we are interested, then it seems to me that the burden
of adjusting to this notation (if it is a burden, one would ex-
pect it to be reasonably matched to both the partners) should
be on the machine.

Of course, this does not mean that the human should make
no contribution in learning things which he does not now learn,
in order to communicate better with his partner. But the things
he should learn are those things that we have discovered (in
the course of learning about programming, perhaps) to be more
generally interesting than we had previously suspected: for ex-
ample, logic and recursive function theory, the meaning of a
proof, the meaning of an algorithm. However, I think it is quite
possible to justify learning these topics, computer or no com-
puter. That is, I don't think this is a matter of matching people
to machines. I think it is simply a matter of being able to talk
intelligently about a class of interesting problems that we have
discovered.

It can be argued, of course, that there is some educational
value in making people adjust to the machine. It can be argued
that some practice in numerical integration will make the cal-
culus more comprehensible and teach what a limit really means.
In fact, Perlis has argued this, and to a certain extent it is un-
doubtedly true. It is undoubtedly true that someone who has
gone through a numerical integration operation has learned
something about the calculus in the process. If it is sufficiently
true, if such a discussion improves the learning and understand-

200

The Computer in the University

ing of the calculus sufficiently, then we should do it whether
or not we are interested in training students in programming
and numerical analysis. We should do it because we want to
teach the calculus well. And if it is sufficiently true to be worth
something, but not necessarily the amount of time it takes to
do numerical analysis with the desk calculator, then perhaps it
becomes worth while when we have high-speed digital com-
puters available.

On the other hand, it may very well be true that although
it is worth something, in the rather intense intellectual com-
petition of four years of undergraduate life, nowadays, it is not
worth enough for all students. There may be more useful things
to learn about the calculus than this, or more effective ways to
spend one's time while learning the calculus.

If this is the case, if there is not enough educational value
in numerical integration to justify it for the sake of teaching
the calculus, then the programming systems should be expanded
so as to include an error analysis automatically when doing
numerical integration, providing answers along with estimates
of validity without bothering the student with the details of
where the estimates come from. Those are appropriate topics for
a course in numerical analysis, which I do not believe every
student interested in science or engineering should be obliged
to take.

It may be argued that students of science and engineering
should not have to use techniques they do not understand, that
a programming language that is very far from the language
which the machine uses insulates the students from knowledge
about the nature of a computer, and that this is bad; that these
students are capable by background of understanding every
step, and there is no reason why they should not.

Well, of course, this again is a matter of intellectual com-
petition. There are a great many things that these students are
capable of learning, but they cannot learn them all in four
years. It is also a bit artificial, since even if you teach students
the primitive machine code and let them follow the transfers
from register to register, they are still being insulated from the

201

P. Elias

facts of life. They do not know what a flip-flop is in circuit
terms, let alone the necessary solid-state physics which would
let them understand how one of the transistors in the flip-flop
works. Yet these are things the science and engineering under-
graduates are perfectly capable of learning.

To require of every economist who wants the solution to a
set of simultaneous equations that he first learn solid-state
physics, so that he can learn machine language coding, so that
he can learn programming, so that he can learn numerical anal-
ysis, so that he can learn how to put his economic problem in
programming terms, seems absurd; and the omission of the
first three steps leaves the sequence at least half as absurd as
it was. Only an electrical engineer might reasonably be expected
to learn about computers from the ground up, as it were, and
he certainly does not start at the beginning with solid-state
physics before his freshman programming course.

So I should argue that there is no basic reason for forcing
everyone to understand in detail how a computer operates.
There is reason for teaching the people the languages they will
need in order to state the problems they are interested in solv-
ing, but presumably only to the extent that those are the lan-
guages which are well matched to the problems and not well
matched to the computer, although I do not think this differ-
ence is ultimately very great. I assume that as we learn more
about organizing computers and designing languages, the lan-
guages which are matched to the appropriate kinds of prob-
lems will also be matched to the appropriate kinds of computer
organizations, and will not be too remote from the way the
computer functions.

My area of disagreement with Pedis is really not very great.
I join him in looking forward to the ultimate symbiosis where
students will be interacting with computers and where every-
one at a university will be making use of computers in a variety
of ways. I agree that we should have freshman and sophomore
courses available in the elements of programming. He would
like to make such courses compulsory, and I should not. But
this is not a major point. The evidence seems to be that about

202

The Computer in the University

a third of the science and engineering students will elect these
courses if given a chance. The fraction might be as large in
economics and psychology and will grow with time, so that
there will be a large enough audience to make the careful de-
velopment of such subjects worth while in any case. The audi-
ence should even be large enough to make special subjects in
mathematics oriented toward computers worth while, so that
experimentation with the impact of the computer on other
subject matter should not require universal adoption of a pro-
gramming course.

Perhaps our most serious difference is in predicting the ulti-
mate state of affairs when time-shared computers are available
on every campus and good symbolic processing languages are
in use. By that stage it sounds to me as though Perlis would
have programming assume a large role in the curriculum, while
I should hope that it would have disappeared from the curricula
of all but a moderate group of specialists.

I have a feeling that if over the next ten years we train a
third of our undergraduates at M.I.T. in programming, this will
generate enough worthwhile languages for us to be able to stop,
and that succeeding undergraduates will face the console with
such a natural keyboard and such a natural language that there
will be very little left, if anything, to the teaching of program-
ming. We shall continue, of course, to have a group of students
who are interested and specialize in numerical analysis per se
or programming per se. But I hope this does not continue to
have to be a major concern for the large majority of people do-
ing any kind of scientific or engineering work.

I think if we stop short of that, if it continues to demand
as much effort to learn how to speak to the machine as it costs
us to teach students a course for a couple of semesters, then
we have failed. I do not see anything built into the situation
which requires as much as that.

Licklider. Pete [Elias], I think the first apes who tried to
talk with one another decided that learning language was a
dreadful bore. They hoped that a few apes would work the

203

/. C. R. Licklider

thing out so the rest could avoid the bother. But some people
write poetry in the language we speak. Perhaps better poetry
will be written in the language of digital computers of the fu-
ture than has ever been written in English.

Let me explain myself. I am just a psychologist who has had
the rare and wonderful opportunity of some contact with com-
puters and some associations with computing people. All I can
bring to you, I think, is a little more enthusiasm than my two
colleagues here have brought with them.

With most of Perlis' concepts and ideas I agree enthusias-
tically. The aim of my discussion therefore will not be to sep-
arate what he has said into that with which I agree and that
with which I disagree. The only exception is that you [Perlis]
said that the computer would not do anything toward the crea-
tion of sensitivity, and with that I disagree wholeheartedly.

Perlis. I disagree with it too.

Licklider. Thank you. All I want to do is to urge that some
of these concepts and ideas be extended a bit beyond the realm
of numerical analysis and the teaching of calculus and the grad-
ing of problems that M.I.T. professors feel called upon to as-
sign to their students. (I am sure that when they no longer
have to grade these problems they will stop giving them, be-
cause the only earthly good the problems do is provide work for
the grading staff. )

As Professor Perlis implied, there are under discussion two
digital computers and two roles of the digital computer in the
university. First, there is the digital computer that we have, a
fabulous mechanism for calculating solutions to preformulated
problems, especially long, hard, deep numerical problems. Typ-
ically, it lives behind a glass wall. It has a tighter appointment
schedule, and a more resolute appointment buffer, than a dean.
If you pose to it a poor problem, it labors as willingly and
rapidly and accurately as if your problem had been epochal,
but if you submit to it a question in which you have misplaced
a comma, it insults you grossly and refuses to hear your apology
before next Tuesday, at which time you may have three minutes
— and your commas had better be properly placed. The role

204

The Computer in the University

of this (present) digital computer in the university is largely^
as I see it, to let imaginative, creative people catch a glimpse
of the role a different digital computer could play.

Professor Pedis devoted some of his attention to the glimpse
he has caught of the second digital computer and its role. He
put his finger on what, in my opinion, is the critical key when
he said that it must be large and it must be fast. Professor
Perlis also said that it must be employed simultaneously by
many users at remote stations. Let me amplify that requirement
by saying that each user must be able to work on his own prob-
lem, at any time, as though he had a slower machine with al-
most as large a memory. He must be able to use his own pro-
grams and, in addition, the very large collection of compilers,,
interpreters, utility programs, tables, and data stored perma-
nently in the central computer. With the aid of the compilers
and interpreters (that maybe someone else has worked out,
Pete) he must be able to interact with the computer in a way
that is reasonably natural to the human scholar. In any event,
he must not have so to clutter his mind with codes and formats
that he cannot think about his substantive problem.

A computer system of this second type is technically and eco-
nomically feasible. This is, however, not the time to discuss
computer system designs and costs. (Let me say only that, in
the heads of computer people I know here in Cambridge — e.g.,
John McCarthy and Marvin Minsky of M.I.T., Wesley Clark
of the Lincoln Laboratory, and my young colleague Edward
Fredkin — there are advanced leads to solution for all the fore-
seen technical problems, and that, because of the efficient use
of memory and processors, the economic prospect is favorable.)
The question to discuss now, it seems to me, is the role in the
university of this second digital computer, this "approachable,"
multiprogrammed, multiuser computer.

Although the first immediate concern of the university may
properly be with the education of students, the education of
students may not be the best focus within which initially to
examine the computer s role. It may be better to consider first
the whole domain of creative intellectual processes (which is,

205

/. C. R. Licklider

I think, the proper domain of the university) and to ask what
role the computer may play in those processes.

Let me report, briefly, that preliminary analysis of technical
and scientific creative activity suggests that such activity con-
sists of short intervals of insight, invention, and decision mak-
ing interspersed among long intervals of "staff operations."
Most of the researcher's time, most of the scholar's time, most
of the student's time is spent in getting into position to take
a step, and only a small part of it is spent in taking the step.
No one knows what it would do to a creative brain to think
creatively continuously. Perhaps the brain, like the heart, must
devote most of its time to rest between beats. But I doubt that
this is true. I hope it is not, because a computer of the sort
that Perlis is pointing toward (not really at, but toward) can
give us our first look at unfettered thought. It can allow a deci-
sion maker to do almost nothing but decision making, instead
of processing data to get into position to make the decision.

When we have this computer, after the several necessary
years of programming, language developing, computer design-
ing, and so forth, I think it will participate in almost every in-
tellectual transaction that goes on in the university. Right now,
of course, the computer solves preformulated problems mainly
of a numerical nature. In due course it will be part of the for-
mulation of problems; part of real-time thinking, problem solv-
ing, doing of research, conducting of experiments, getting into
the literature and finding references you want. It will be part
of this for, I think, all the people. First it will be so for the pro-
fessors; eventually it will be so for the students, also.

As part of its contribution to the intellectual process, the
computer will explore consequences of assumptions. It will
present complex systems of facts and relations from new points
of view, "cutting the cake another way" in a fraction of a
second. It will test proposed plans of action through simula-
tion. It will mediate and facilitate communication among hu-
man beings. It will revolutionize their access to information.
It will even obtain for them the aid and services of other digital
computers. At first the computer will be, in Pedis' term, a

206

The Computer in the University

"handmaiden to scholarly university activities." In not many
years, however, it will be regarded less as a handmaiden than
as a partner.

Through its contribution to formulative thinking, which will
be, I think, as significant as its solution of formulated prob-
lems, the computer will help us understand the structure of
ideas, the nature of intellectual processes. Essential to this
function is the development of procedures for computing with
nonnumerical symbols. ("Computer" and "computing" do not
seem to be the proper words, but their meanings may change as
the nonnumerical part of digital information processing con-
tinues to grow in importance.) Although one cannot see clearly
and deeply into this region of the future from the present point
of view, he can be convinced that "information processing,"
which now connotes to many "a technology devoted to re-
ducing data and increasing costs," will one day be the field of
a basic and important science. Planning, management com-
munication, mathematics and logic, and perhaps even psychol-
ogy and philosophy will draw heavily from and contribute
heavily to that science. One of the most important present
functions of the digital computer in the university should be to
catalyze the development of that science.

The area of investigation that seems to me to offer the most
direct path toward understanding of intellectual processes is
the area called "artificial intelligence," the area concerned with
the development of self-organizing and/or intelligent automata.
Let me not speculate now upon the outcome of such investiga-
tion. Let me say only that, in my opinion, very great advances
in thought will be made in this area, just as soon as the neces-
sary computer facilities, i.e., multiprogrammed, multiuser com-
puters with fast processors and large memories, are developed
and installed in universities. As the recent review by Minsky
shows, there are many promising lines to be followed. What is
holding up progress is the fact that universities have the first
kind of digital computer and not the second. An important
specific function of the (second) digital computer in the uni-
versity is to make it feasible for creative people with diverse

207

/. C. R. Licklider

capabilities to work cooperatively on artificial-intelligence prob-
lems.

Professor Perlis talked about simulation and about manage-
ment games. He indicated that, already, the management game
at Carnegie Tech is a pivotal part of a course of instruction.
Simulation, of course, has already opened a whole new ap-
proach to the study of large dynamic and stochastic systems. It
is easy to extrapolate a few years to the point at which most
of the major researches in management, economics, and psy-
chology will take place as much "in the computer" as "with its
aid." The computer will be making possible a kind of intel-
lectual activity, and a degree of understanding in important
fields, that we cannot possibly accomplish with human brain
power alone.

The impact of the digital computer upon university educa-
tion, it seems to me, will stem mainly from the changes the
computer will produce in intellectual activities generally. The
pedagogical responsibility of the university is not to lecture or
assign problems or grade them. It is to create a situation within
which most bright students will automatically learn. The multi-
user digital computer opens new horizons for anyone eager to
create such situations.

In a small and preliminary way, with only a small computer,
a computer typewriter, and a few nights of programming, some
of us have already created "motivational traps" for our children,
and we are sure that a computer teaching machine can be made
more attractive than television. The youngsters love real-time
interaction with a thing like a computer. It can tell them im-
mediately, "No, that was wrong"; it can calculate and post a
score as it goes along; with the aid of a simple random process,
it can look up in a table a suitable compliment or a suitable
sarcastic remark such as "Oh, oh, you're slipping." The young-
sters will sit there and punch the keys for hours learning spell-
ing and language vocabulary.

This is just a touch of what can be done. The device can be
carried on into high-school subjects, college subjects, and re-
search. We can look forward to the time when any student

208

The Computer in the University

from grade school through graduate school who doesn't get
two hours a day at the console will be considered intellectually
deprived — and will not like it.

The conclusion at which I arrive is that the present problem
is not to assess the role of today's digital computer in today's
university. It is to get to work on tomorrow's computer and to-
morrow's university. It will not suffice to wait until the com-
puter industry develops the computer the university needs; for
commercial, industrial, and military requirements are not lead-
ing to development of such a computer on anything like the
time scale that is feasible. Moreover, having such a computer
is much less than half the battle. The task of preparing the
programs required to make it "go critical" is great. But it is a
task that universities can and should handle, for it is itself an
intellectual enterprise of high order.

In short, I feel that the university has a role in (broadly
defined) digital computing that must be conjugate with the
role of the digital computer in the university. If the two Toles
are fulfilled, both the university and the computer will be con-
siderably changed — enhanced. I am eagerly confident that
universities, including this one (M.I.T.), in particular, will fulfill
their role.

Perlis. Perhaps I may have been misunderstood as to the
purpose of my proposed first course in programming. It is not
to teach people how to program a specific computer, nor is it
to teach some new languages. The purpose of a course in pro-
gramming is to teach people how to construct and analyze
processes. I know of no course that the student gets in his first
year in a university which has this as its sole purpose.

The course in calculus that is taught in most universities to-
day is so covered by a mass of techniques that are completely
and totally associated with the task at hand that the funda-
mental processes are recognized only with difficulty by many
students. In physics one tries to get insight into the basic
physical laws in the first course; but one is talking about physics.
In chemistry the course is divided up into perhaps two parts.

209

A. J. Pedis

You learn the factual nature of the relationships of the ele-
ments, their atomic weights, their valences, and so on. You
learn how to balance equations. You also learn about the basic
chemical laws: the law of mass action, the concept of equilib-
rium, the electrochemical series. But you are talking about
chemistry, and in many students' minds the logic is never sep-
arable from the chemistry. In English you are concerned with
grammar. You learn to express yourself precisely; you learn how
to write; and you learn how to write only in English.

A course in programming, on the other hand, if it is taught
properly, is concerned with abstraction: the abstraction of con-
structing, analyzing, and describing processes. It is not the par-
ticular problem content of numerical analysis or analyzing a
statement which is important. Rather, it is possible to skip from
problem area to problem area and still stabilize on the concept
of process design and analysis. Thus in a programming course,
much more than in any of these other courses, it is pos-
sible for the student to abstract ideas from the particular ex-
amples given. And I repeat: I know of no freshman course, cur-
rently, that has this objective.

I personally feel that the ability to analyze and construct
processes is a very important ability, one which the student has
to acquire sooner or later. I believe that he does acquire it in
a rather diluted way during four years of an engineering or sci-
ence program. I consider it also important to a liberal arts
program.

This, to me, is the whole importance of a course in pro-
gramming. It is a simulation. The point is not to teach the
students how to use Algol, or how to program the 704. These
are of little direct value. The point is to make the students
construct complex processes out of simpler ones (and this is
always present in programming) in the hope that the basic
concepts and abilities will rub off. A properly designed pro-
gramming course will develop these abilities better than any
other course. I do not have a proof that this is true. I only have
very little evidence of it. But this is why I think the course
should be taught.

210

The Computer in the University

General Discussion

McCarthy. I should like to comment on this concept of
giving a problem to the machine. I think this reveals a mis-
understanding about the relationship between mathematical
notation and computer programs. Unfortunately, one does not
give problems to machines; one does not state a problem; one
states a procedure for solving the problem. This distinction is
very important. For example, the problem of finding a number
that satisfies the cubic equation

x3 + 9x + 7 = 0

is precisely stated. I have precisely stated it. There is nothing
more to be said about it. But the question of solving the prob-
lem, once it is stated, is still a significant issue.

Programming is the art of stating procedures. Prior to the
development of digital computers, one did not have to state
procedures precisely, and no languages were developed for
stating procedures precisely. Now we have a tool that will carry
out any procedure, provided we can state this procedure suffi-
ciently well. It is Utopian to suppose that either English or some
combination of English and mathematics will turn out to be
the appropriate language for stating procedures.

One can say, "All right, if mathematics is not the language
for stating procedures, what is mathematics?" The answer is
that mathematics is a language for stating certain classes of
facts, and a fact is different from a problem; a fact is also dif-
ferent from a procedure. One would suppose that appropriate
languages for describing procedures will contain components
which have a lot in common with mathematics and which con-
tain a certain amount of mathematics; but they will also con-
tain other things which are yet to be developed.

I should like to make one remark about the teaching of pro-
gramming to freshmen (since I have done this). The liberal
arts value of the programming course is evident, but it is still

211

General Discussion

overdiluted by bookkeeping. I believe that removing this ex-
cessive dilution will be a fairly gradual process.

Elias. I should like to reply, really using John McCarthy
as an excuse, to Licklider. I did not mean to imply that people
will not learn how to communicate effectively with the machine,
and that what people learn in school will not be different from
what they now learn, and will not be necessary for that effec-
tive communication. But I should be very disappointed if it
turned out, by the time you had gotten rid of the bookkeeping,
that discussing procedures, discussing algorithms as well as
proofs, discussing logic as well as mathematics (if this is the
appropriate language) were not an interesting and valuable
thing to do for its own sake. I should be disappointed if it
turned out that we had to do an appreciable amount of this
for the sole purpose of communicating with machines. I hope
that there is more content to this area of research than that.

Marquis. Dr. Minsky, your name has come into the dis-
cussion. Have you any comments?

Minsky. I shall make one that is parallel to Lickliders.
Let me start this way. Nobody in the world has really had the
experience of using a time-shared computer, although there are
people who should be leaping to their feet to deny this.

There have been a few experiences in the form of a couple
of typewriters connected to a small computer. And there have
been a few experiences of individuals operating a very large
computer all by themselves: namely, individuals of great power
or prestige, or those who are very quick to fill the vacuum of
an IBM 704 sitting on the production line, waiting to be
shipped out. These individuals have had the experience of op-
erating a very large machine for their own amusement for long
intervals of time. But in the few cases where this happened, the
individuals concerned were still enmeshed in bookkeeping, in
the sense that adequate programs for time sharing did not
exist.

What I really mean is that nobody yet has written and used
the type of monitor system necessary for administering a great
number of complicated programs conveniently, just by calling

212

The Computer in the University

their names and by binding a few of the parameters to which
Perlis referred.

Educational value aside, a time-sharing computer system may
be of immense personal value to the mature adult who wants
to get on with his work and get something done. Licklider
pointed this out in his discussion. The reason, as Licklider also
has mentioned on occasion (but apparently did not dare to
mention here), is that nowadays one gets about twenty minutes
of real thinking done on a good day when not too many stu-
dents come in to interrupt him.

We do not know what it would be like to be working in a
mathematical area in which we are not too familiar (as I am
not too familiar with analysis) and be able to say, "I want the
Laplace transform of that," and have it appear on the scope;
and then say, "I should like that variable to be Gaussian dis-
tributed," and have the appropriate convolution operations
occur instantaneously. It would take the machine a long time,
but it should happen well within one human reaction time,
which is an unpleasantly large number of microseconds.

We could have this. Somebody has got to provide the fa-
cility and put it into the hands of the scientists who are going
to do the ugly but rather interesting job of programming to
make it work. But nobody has ever tried it. The tools are all
here, and a few million dollars is all it would take.

Licklider. I want to say that although nobody really has
tried time sharing yet, enough people have gotten to sit with
a big computer for a couple of hours to feel morally certain of
the possibilities. I hope you will not mind if I illustrate this.
One of my friends has a sign on his wall which says, 'Tut your
clothes back on, lady, I'm a computer man."

The point is that once you get anywhere near something
like the man-machine system which Minsky envisages, you
know you really want it. And it is not as though the job of
programming or designing hardware to obtain such a system
will be objectionable. In fact, I have not had much of an in-
troduction to this, but already I see computer programming as
a way into the structure of ideas and into the understanding of

213

General Discussion

intellectual processes that is just a new thing in this world.
Any psychologist is crazy to keep on working with people if
he has access to a computer.

Seriously, the man-machine symbiosis gives us a much better
way than we ever have had before for saying what we are try-
ing to say and then finding out whether it is indeed correct.

Perlis. I should like to say a word about the imminence of
the time-sharing concept. Marvin [Minsky] said that no one
has tried it yet; and this is quite true. The necessary equipment
has not existed previously; but now it is really very imminent.
The figures that I cited for the computer that we are getting
at Camegie Tech are based on fact, not fancy. W e are going to
have remote consoles with the machine as a matter of course:
not by our own special request, but by the manufacturer's fore-
sight.

It is still not a perfect machine. It has some severe limitations
that probably are going to cause a great deal of trouble to
everybody who tries to use it. But the concept of the multiple
use of a computer by a large university staff working simul-
taneously and remotely is being adopted; and we shall soon find
out whether people are really good enough to make use of their
own consoles in their own offices.

I have a suspicion that very few people are going to be good
enough for some time to come, no matter what programming
tools we provide. I suspect that many people take refuge in
bookkeeping, regardless of their professional level in life. It is
a form of intellectual security, a safeguard. Introducing a con-
sole, with all of its power, will cause some anguish to many
people; but I hope not to freshmen, because they are the ad-
venturous ones.

We do not now know what will happen when freshmen in
a technical university have been thoroughly educated in the
computer by the end of their first year. We shall have to wait
four or five years to get enough evidence, but I do not think
that we shall have to wait longer than that.

Forrester. I wish to point out that the man-machine con-
cepts of "human reaction time" and "immediate access" are

214

The Computer in the University

relative terms that depend on the nature of the application.
Students who use the computer to analyze nonlinear systems of
200 variables or more require from 24 to 48 hours from the time
they receive an answer to determine what the answer has taught
them and to decide which question to ask next. At this level
of complexity, the "human reaction time" already is so long
relative to the computer's speed of operation that the user
effectively has "immediate access."

Oettinger. Concerning Elias' question of whether any-
thing of substance will be left once the machine is withdrawn,
I feel that Perlis' remarks about processes contain the answer,
although Perlis' emphasis on the calculus and the teaching of
the calculus obscures the importance of the computer in the
understanding of nonnumerical processes such as those involved
in the analysis of human language, and those involved in the
study of psychology and history. I agree with Perlis that in the
long run, program and language development will be more than
insignificant "busy work" for computer people, and I believe de-
velopments are leading to notational systems that will make the
understanding of processes as essential to the educated man as
is the understanding of mathematical concepts.

Elias. I agree that we very likely shall discover things that
will be worth learning for their own sake and will have very
significant intellectual content. On the other hand, let me
point out as a pedagogical matter that we discovered some
things in mathematics with a significant intellectual content
from which, as Professor Perlis pointed out, we have managed
to insulate our engineering freshmen.

What a proof is about is a valuable concept; and what an
algorithm is about is also. But currently we teach the calculus
to our students without getting the first concept across, and
we may teach programming to our students for quite a while
without getting the second concept across. Unless we are more
careful than we have been over the last century in teaching
mathematics, it is very likely that students will never see the
second concept at all.

Slagle. I share the more optimistic viewpoints that have

215

General Discussion

been stated, but feel that the direct role of the computer in
pedagogy has been somewhat neglected in the discussion.

Licklider. I should like to pick up this thought about the
pedagogical use of computers. The main thing, I think, that has
come out of the study of teaching machines (a la Fred Skinner
at Harvard), is that the only way to get knowledge and wisdom
into the human brain is to put it in through a very extensive
process.

Let me explain. You can take a field of mathematics and
axiomatize it and get it down to a static nutshell: a beautiful
kernel, a thing that a mathematician thinks is just lovely. And
maybe you can get somebody to memorize it; but if he does so,
he has not learned a thing. The only way to get it into the
person is to blow it up into 50,000 operations and teach it from
many points of view, using many examples. Finally it gets down
into his brain in a little kernel: a beautiful pearl of knowledge.

This process is just killing us. We don't have enough teachers
to handle it. University professors leave universities because of
it. Computing machines, as Perlis pointed out, can do this kind
of thing rapidly, economically, patiently, and without the frus-
tration and the unhappiness. I think the computer offers a real
match to the problem of getting knowledge into human skulls.

Kailath. In Japan they say the highest form of abuse is the
one that is not uttered but merely implied. Similarly, I like to
think that the greatest value of computers for undergraduate
education accrues when computers are not used but merely
implied.

Several advantages of computers in undergraduate education
have been put forward. One advantage supposedly is that the
discipline of programming a computer develops certain basic
skills of thought and understanding. A second advantage is
that there is heuristic value in thinking up algorithms and
methods to get the computer to do what you want it to do. I
think the first advantage, the discipline of programming, is not
worth the effort. For one thing, computers think in the crudest,
most elementary terms, whereas human beings have remarkable
brains by comparison. The really important advantage, to my

216

The Computer in the University

mind, is the heuristic value of computers. Computers can en-
able people, such as those working in artificial intelligence, to
acquire a closer appreciation of how the mind works. Eventually,
however, we should be able to do away with computers and rely
on inspired heuristic analogies drawn from better understand-
ing of the process of thinking.

Perlis. Yes and no. One should not (and I probably am
more guilty of this than many) oversimplify the role of a com-
puter. There also happen to be in our society today problems
which actually require computation, and these problems are
quite important to us.

References

1. "Report on a Conference of University Computing Center Directors,"
Communications of the Assoc. for Computing Machinery, Vol. 3, 519-
521 (Oct., 1960).

2. Lanczos, C, Applied Analysis, Prentice-Hall, Inc., Englewood Cliffs,
N.J., 1956, p. 404.

3. Cohen, K. J., Cyert, R. M., Dill, W. R., and others, "The Carnegie
Tech Management Game," The Journal of Business, Vol. 33, 303-
321 (Oct, 1960).

4. "Conference Report on the Use of Computers in Engineering Class-
room Instruction," Communications of the Assoc. for Computing Ma-
chinery, Vol. 3, 522-527 (Oct., 1960).

5. Professor Perlis has offered to make available, upon direct request to
him, a set of problems used in the computer course at Carnegie.

217

6

Time-Sharing
Computer
Systems

6

Time-Sharing
Computer
Systems

Speaker John McCarthy

Associate Professor of Communications Sciences
Massachusetts Institute of Technology

Discussants John W. Mauchly
President

Mauchly Associates Inc.
Gene M. Amdahl

Manager, Advanced Systems Design
IBM Data Systems Division

Moderator Emanuel R. Piore

Vice President for Research and Engineering
International Business Machines Corporation

220

McCarthy. I am going to discuss the important trend in
computer design toward time-sharing computer systems. By a
time-sharing computer system I shall mean one that interacts
with many simultaneous users through a number of remote
consoles. Such a system will look to each user like a large pri-
vate computer. The new applications that time sharing will
make possible will be of as much additional benefit to science
and management as resulted from the introduction of the
stored-program digital computer.

First I shall discuss the uses of a large private computer, and
then I shall discuss how the same effect can be achieved through
time sharing. I shall also touch upon the requirements that
time-sharing systems impose upon computer design. Some part
of what I am going to say arose in connection with the work

221

J. McCarthy

of the Long Range Computer Study Group here at M.I.T.
The material I shall present on computer-system design was
developed jointly with Marvin Minsky. I also want to acknowl-
edge the stimulating effect of discussions with Professor Herbert
M. Teager and Dr. F. J. Corbato, who are developing time-
sharing systems for the IBM 7090 at the M.I.T. Computation
Center.

Some of the ideas concerning time-sharing systems go back
quite a way. The first paper on the subject that I know of was
written by Christopher Strachey and was delivered at the Paris
International Conference on Information Processing.1 A more
recent paper is by Licklider.2 The subject was also touched on
in the lectures by Kemeny and Perlis. In 1945 Yannevar Bush
discussed a system for personal information retrieval called
Memex, which probably requires a computer system of the
kind that I am going to discuss for its realization.3

A Private Computer

Y\~hy should anyone want a private computer? The reason
has to do with the interaction between user and computer. In
theoretical terms, a communication channel can be character-
ized not only by its one-way capacity, but also by its two-way
capacity, a concept that is a little harder to define formally.
Two-way capacity is the number of messages per second which
the channel can interchange, given that the content of the
leaving message depends on the content of the entering mes-
sage. The advantage of a private computer is that it makes
possible a large two-way communication capacity.

What the user wants is a computer that he can have con-
tinuously at his beck and call for long periods of time. Of
course, computer applications differ in the amount of interaction
required between user and machine (interaction can be meas-
ured by how often messages have to be exchanged between user
and machine in order to complete the task). The earliest ap-
plications of computers were those in which there was little
two-way interaction. Programs could be run for long periods of
time without any user involvement. For instance, in the com-

222

Time-Sharing Computer Systems

putation of mathematical tables (one of the earliest applica-
tions), you knew the tables that you wanted to compute, so
you designed the machine and wrote the program and then let
the thing go. Eventually the computer and the program and,
unfortunately, also the tables became simultaneously obsolete.

There are still many applications that require long periods of
straight computing, even with present computers. Examples are
numerical weather prediction, which is still a very substantial
task and promises to become even more so as models become
more sophisticated, and calculations of the distribution of en-
ergy in nuclear reactors. In my remarks, however, I shall con-
centrate on applications, present and potential, that require
frequent interaction between user and machine.

Debugging

One such application is programming. This application forces
itself upon the user and requires more frequent interaction be-
tween man and machine than is presently available. Even peo-
ple who are trying to program for conventional applications
find themselves suffering from the lack of two-way interaction.
If a program is at all complicated (and weather prediction pro-
grams are complicated), it will be built up from subprograms
that require extensive testing singly and in combination. Scores
of runs may be necessary to find and correct all the errors in
an extensive program. Nevertheless, the amount of computer
time consumed in a single run may be quite small. In fact,
neglecting input-output time, only a few milliseconds may be
required to detect an error in a straight-line program. Some
errors, like misprints, can be corrected very quickly once the
machine detects them; a programmer will get very annoyed if
he has to wait a day in order to fix a misplaced comma. Other
errors are more complicated, and the user may have to study
his results for a day or two before he finds the trouble.

Two Present Systems of Computer Operation

There are two common ways of operating a computer for
debugging programs. Both are used at M.I.T.: one by users of

223

/. McCarthy

the TX-0 computer, and the other by the users of the IBM
709 computer at the Computation Center.

The TX-0 is a small computer that is operated very in-
formally. Users sign up a week in advance for blocks of time
of from one to several hours. During his block of time the
user has a complete monopoly of the computer. He tests his
program, makes necessary corrections, retests his program, and
continues to make corrections. While the user is thinking about
his results, the computer stands idle. If the source of trouble is
hard to find, the user can waste his remaining allotment of
time for the week. On the other hand, if the trouble is trivial,
he can fix it immediately and continue. The TX-0 has about
forty users, of whom approximately twenty-five are active in a
typical week. If the TX-0 were a much larger computer, and
if it were operated in the same manner as at present, the num-
ber of users who could be accommodated would still be about
the same.

The IBM 709 at the M.I.T. Computation Center is a larger
computer and is operated according to a different principle.
Users prepare their programs and data on punched cards and
deposit them in a file in a run preparation room. Machine op-
erators transfer the programs from the punched cards to mag-
netic tape in batches. The programs are processed one after
another by the computer, and the results of the computation
are written on tape; these tapes are printed on a tape-to-printer
machine, and the results are placed in the user's file. The ad-
vantage of this system is that there is no delay between indi-
vidual users, and the machine is not idle while the user is think-
ing. By the nature of the system, a delay of two or three hours
is implied between the time the cards for a job are submitted
and the time that results come back. Since the Computation
Center is overcrowded, however, the delay usually is more like
one or two days; it is at its worst at thesis time, the end of the
school year. Because the machine never waits while the user
thinks, the computer can do from 125 to 160 jobs a day. The
overcrowding results from the approximately 500 programmers
who desire active use of the computer.

224

Time-Sharing Computer Systems

Neither the TX-0 system nor the 709 system is completely
satisfactory. The TX-0 system is unsatisfactory because the
user has to sign up well in advance and cannot go on whenever
he is ready; in addition, he does not have time to think. The
709 system is unsatisfactory because of the delays in getting
results back. Programming is feasible in both systems. However,
the intellectual depth of projects that can be undertaken is
limited by the time required to debug even relatively simple
procedures. This has a very bad effect on student theses. When
an ambitious student proposes to undertake something sub-
stantial, his adviser often must say, "I don't believe you can
finish in time for a thesis." I have several students this semester
who are in this kind of trouble. From the user's point of view
the solution clearly is to have a private computer.

Man-Machine Interaction

What I have said so far has been about problems in pro-
gramming that are more or less independent of what the pro-
grams are about. Now I should like to discuss applications that
have an essential dependence on rapid interaction between man
and machine. This topic is more speculative because these ap-
plications are undeveloped for lack of suitable machines. The
first application that I should like to discuss is computing with
symbolic expressions, a field in which I have worked. It is en-
tirely feasible to have a computer manipulate formulas in the
manner of a mathematician or a scientist doing theoretical
work. For example, there are programs for symbolic differentia-
tion, integration, solution of equations, simplification of al-
gebraic expressions, the solution of differential equations, and
a number of other processes. However, one rarely makes a two-
hour run with symbolic calculations. If one has an integral to
evaluate, even if an integration program is available, it is much
easier to ask somebody (if you are not good at this sort of thing
yourself) than it is to get out the program and apply it. This
leads me to believe that some of the applications of symbolic
calculation depend on having what is essentially a private com-
puter, so that when you decide that you have a symbolic cal-

225

J. McCarthy

dilation to make, you can simply type in what you want done.

Another very important application is the use of the com-
puter as a teacher. There has been a considerable amount of
work done on teaching machines. These machines are generally
of a rather simple character. A fixed sequence of questions is
presented to the student, who tries to answer. When he gets
the right answer, he goes on to the next question. The advan-
tages over conventional methods are that the student can work
at his own pace, and immediate reinforcement is provided for
a correct answer (according to many psychologists this is very
important). Present teaching machines, however, give the
questions in a fixed order that does not depend on the stu-
dent's answers to earlier questions. These machines cannot re-
act to an answer other than by fitting it into one of a number
of predefined categories, and furthermore, they are very limited
in the amount of information (in convenient form) which they
can provide the teacher on the student's progress.

Using a computer as a teacher can overcome these limita-
tions. The decision of which information to give or which ques-
tion to ask can depend on previous performance in whatever
way the instructor wants. New questions can be generated to
cover points where the student requires additional work. With
a large class of students, each one progressing at his own pace,
the instructor can be given reports identifying students who re-
quire additional help or who are doing so well as to deserve
special attention and recognition. It will not be easy, however,
to program a computer to take full advantage of all this flexi-
bility.

A third application is to provide the services envisaged in
Professor Kemeny's lecture in this series on "A Library for
2000 a.d." It seems that except for the incorporation of a
large mass of documents, Professor Kemeny's library, and then
some, can be realized, not by the year 2000, but by the year
1965.

The last specific application that I shall mention concerns
interaction with laboratory apparatus in real time. The parts
of laboratory apparatus that serve merely to transform in-

226

Time-Sharing Computer Systems

formation are at present very expensive to construct and tedious
to debug. For a great many purposes they can be replaced con-
veniently by two-way connections to a computer.

I have mentioned some specific applications of a computer
which require considerably more two-way interaction capability
than we have at present. I could add to this by saying some
more general things about the possible role of a computer as an
intellectual servant. A computer can carry out any intellectual
process that we can program or describe precisely. Our ability
to describe processes precisely, however, is still very limited, as
those of us who work on artificial-intelligence problems will
certainly agree. Nevertheless, some processes of scientific re-
search already can be reduced to routine form, and if we had a
computer with which efficient two-way interaction were pos-
sible, we should be able to extend greatly our scientific capa-
bility. As our ability to understand intellectual processes ex-
pands, we shall be able to apply the computer to more and more
tasks.

Advantages of a Large Computer

I should now like to discuss the question of a large computer,
where by large computer I mean one with very large primary
storage and very large secondary storage. The amount of sec-
ondary storage available has a simple meaning. It determines
what fraction of the computer culture, so to speak, is directly
accessible to the users of the computer. The implications of
having a large primary storage are more subtle.

The limitations on what it has been possible to do with com-
puters so far depend partly on the computers themselves, but
even more on our ability to program. Each computer has an
order code that is closely connected with the way the computer
was designed, and the computer is able to obey programs that
are written in this order code (also known at the machine's
language). As you undoubtedly know, this does not mean that
the user must necessarily describe his procedure in machine
language. One of the intellectual procedures that the computer
can be programmed to perform is the translation of procedures

227

/. McCarthy

into machine language from source languages in which it is
more convenient to write the procedures. A number of con-
venient source languages have been devised, and compilers and
interpreters have been written which enable a computer to obey
a program written in these languages. Among these are Fortran,
which was developed by IBM for describing numerical pro-
cedures; Algol, which is a proposed international standard
language for describing numerical procedures; Cobol, which is
a proposed United States standard language for describing file
maintenance and other business data processing procedures;
Apt, a language that was devised at M.I.T. for programming
the numerical control of machine tools; and Lisp, a language
that our group devised for describing computations with sym-
bolic expressions.

Before computers there really was little need for convenient
ways of formally describing procedures, and not much progress
was made in developing such ways. At present, programming
languages are developing rapidly, and there even has been work
on constructing a mathematical theory of computation. Never-
theless, programming difficulties still cause most of the delay
in preparing computers to cam- out complicated procedures. If
we are going to have the interaction ability that I described
earlier, then we must have complex programming systems for
translating procedures from source languages into machine
language. If the user, sitting at his console, is going to be able
to have symbolic computations performed for him, these pro-
cedures have to be at his beck and call. Ideally, the whole exist-
ing programming culture, including elaborate programming
systems, should be present and readily accessible.

The main reason for needing a large primary memory is that
the procedures which we can program are getting ever more
elaborate. In fact, the frontier in computation is complication.
An indication of this is the program for symbolic integration
developed by Jim Slagle here at M.I.T. which barely fitted in
the 32,768-word memory of the IBM 709. Much of the time
required for developing this program was consumed in making
the fit.4 Now, symbolic integration is one of the simpler math-

228

Time-Sharing Computer Systems

ematical processes and occurs as a minor component in more
interesting symbolic computations. We shall soon be writing
programs that require an order of magnitude more memory
space.

Time Sharing

I should like to go on now to consider how the private com-
puter can be achieved. It is done by time sharing a large com-
puter. Each user has a console that is connected to the computer
by a wired channel such as a telephone line. The consoles are
of two kinds, one cheap and the other better but more expen-
sive. The cheap console is simply an electric typewriter that is
used for both input and output.

Consider the operation of such a system as it appears to the
user of the typewriter console. When the user wants service,
he simply starts typing in a message requesting the service.
The computer is always ready to pay attention to any key that
he may strike, just as the telephone system is always Teady for
you to lift the receiver off the hook. As soon as the key is de-
pressed on the typewriter, a signal is sent to the computer. The
effect of the signal is to make the computer interrupt the pro-
gram after the current instruction has been executed and jump
temporarily to a program that determines what the typewriter
wants. Most characters are at the beginning or middle of a
message, so all the computer usually has to do is store the char-
acter or bring out a new character in case of output. If more
than one typewriter requests attention simultaneously, a queue
is formed. A very large number of typewriters can be handled
without the computer taking more than a small amount of time
away from whatever other tasks it is doing. In fact, a computer
of the speed of the IBM 7090 could handle three thousand
typewriters simultaneously.

Eventually a message will be completed, and the computer
will have to take some kind of action other than merely storing
characters. The following are examples of the kinds of action
that may be required. First, the message may be a request for
information from primary or secondary storage to be typed out.

229

J. McCarthy

Second, the message may be an expression to be integrated.
Third, the message may be a statement in a program that, if
formed according to the rules of the programming language, is
to be added to a list of statements waiting to be translated
when the program is complete. Fourth, the message may be a
command to allocate an area in primary storage, or to transfer
certain information from the users files to primary storage for
later action, or to transfer information from primary storage
to the user's files. Fifth, the message may be a request to run
a program taken from the typewriter, from the users files, and
from public subroutines. This fifth example takes the most
computer time; the other examples involve specific, minor items
that the computer can handle quickly.

At any moment some of the typewriters will be inactive,
some will be in the middle of entering messages into the com-
puter, some will be in the process of typing out characters, and
some will be in a status of wanting programs run. There are
several ways to handle the situation of a number of programs
that simultaneously want to be run. I am going to discuss only
the simplest way: the round-robin system. Each program is run
for a definite period of time, which I shall call one quantum.
If the program is not completed in this quantum, a time clock
interruption occurs, and control passes to the next program in
the round robin. When the round robin is completed, control
returns to the first program, and so on.

How long should the quantum of time be? One answer is
that the amount of time for the complete round robin should
be just less than a human reaction time, say, one-tenth of a
second. If this is so, then a user whose run requires less than
one quantum of computer time will appear to get instantaneous
service. The round-robin system is susceptible to various modi-
fications. For example, very long programs may be postponed
to slack hours in order to save memory space, and programs of
very high priority can be given more than one quantum per
round. A stop button would be an important part of the type-
writer console, since it is very easy for a user to initiate an out-
put that is much longer than he cares to have. There must be

230

Time-Sharing Computer Systems

some way of saying, "No, stop, and pay attention to me
again."

There has been extensive experience with the use of computers
with typewriters as their main input-output device. The main
limitation of a typewriter is that it takes too long to produce
ouput. For example, the report of an assembly or compilation
may take as much as twenty minutes to produce. In our system
this can be improved a little bit by having high-speed printing
facilities available in the computation center, so that if a user
will tolerate having his output later, he can have it printed at
the center and sent to him.

The more expensive console that I mentioned can include a
cathode-ray-tube unit on which the computer can display pic-
tures and text. Such a unit requires some kind of temporary
storage like a tape loop or drum to avoid the computer's spend-
ing excessive time maintaining the display. Another device that
has been found very useful is a light pen. This simply consists
of a photocell that the user can hold against the face of the
cathode-ray tube to designate a point for the computer. Other
analog or digital input-output connections are desirable if the
computer is connected to laboratory apparatus.

This is one example of a time-sharing system. There are a
large number of users. Each user has his own console, gets
service from the computer whenever he desires it, and has the
computer maintain his files for him.

Computer Requirements for Time Sharing

Time sharing is technically feasible on a small computer,
but the full advantages of rapid man-machine interaction re-
quire elaborate programming languages and elaborate program
control. Because programs may be called on by many users and
may do only relatively short pieces of work between human in-
teractions, it is uneconomical to have to shuttle them back and
forth continually to and from secondary storage. Therefore,
there is a requirement for a large primary memory. How large,
we are not sure. The amount of permanent program that we
should want to have continuously available might be some

231

J. McCarthy

100,000 words, given the present state of programming services,
and this suggests something on the order of one million words
of directly addressable core memorv for a time-sharing system.
The development of new programming services could increase
this requirement considerably.

are the other requirements for a time-sharing system?
Present order codes of the computer appear adequate, and the
main features of present single-address computers require no
important variables. Some new features will be needed, how-
ever.

One requirement is for an interruption system to handle
errors as well as input and output. An interruption system works
in the following way. Upon a special condition, such as an
erroneous instruction to divide by zero or an input-output unit
requesting attention, the next instruction in sequence is not
executed. Instead, the computer takes its next instruction from
a specified location, determined by the condition that caused
the interruption. The program to which control is transferred
deals with the condition, and if the interruption is not due to
a program error, returns control after it is done to the place
where the interruption arose.

Another requirement is for completely nonstop operation.
If the computer were to stop with ten or fifteen people typing,
it would be very difficult to retrieve the situation. The operator
would not know what the computer was doing or for whom.
Almost all present computers either have certain instructions
that can cause a stop, or have hang-up features when certain
input-output signals are absent. A proper computer ought not
to rely on anyone or anything for signals on schedule. It cer-
tainly should not rely on the users to program correctly, nor
should it rely on input-output units to provide signals. In
the case of a tape unit which fails to provide a signal saying that
a character is ready, the computer could wait a prescribed time:
if the signal were not forthcoming, it could interrupt to print
out a complaint to the operators.

Another requirement has to do with erroneous programs that
must be prevented from damaging other programs. This is best

232

Time-Sharing Computer Systems

handled by means of boundary registers and input-output inter-
locks. When the computer is executing a program, it must
interrupt if the program attempts any memory references out-
side its allotted region, and it must also interrupt if the pro-
gram attempts to use any instructions that activate input-
output units. Input-output units should be activated only by
system subroutines. Also, programs that get into endless loops
must be prevented from wasting computer time. This can be
accomplished by an alarm clock that interrupts a program after
a quantum of time has been consumed and causes the user to
be informed if his use becomes excessive.

An intricate problem arises from the fact that when a pro-
gram ceases to be active it leaves an odd-sized hole in memory.
It should be possible to move other programs down to fill the
holes so as to provide contiguous blocks of space for new
programs and data. This is facilitated by having a relocation
register whose contents are added to addresses after they leave
the central processing unit and before they get to memory,
Another problem derives from the possibility that a system rou-
tine, such as a compiler, which is executing a program for one
use:, may be m:errup:ec afoe: a oaartum of time to do another
job for a different user. In order that this be feasible without
confusion, it is necessary that system routines use temporary
storage in blocks belonging :: the use: are no: :: themselves.
Thus when program control returns to the first user, his tempo-
ran storage block is unaltered, and the system routine can take
up where it left off.

The final requirement is for secondary storage large enough
to maintain the users' files so that users need not have separate
card or tape input-output units. Tor an institution like M.I.T..
if only programs and data are stored (and not the M.I.T. li-
brary' . then something like £::y million computer words ap-
pear adequate.

These computer requirements for time-sharing systems are
not especially subtle, but the computers coming out nowadays
seem to have a tendency to neglect many of them. Tor one
thing, new computers usually are not able to address a very

233

J. McCarthy

large memory. Their order codes generally do not provide for
memories beyond a certain size.

Multisequence Computers

I want to discuss certain economic considerations that apply
to computers with very large memories. A million-word mem-
ory with reasonable speed of operation may cost six million
dollars. Central processing units, on the other hand, are cheaper,
so we have the situation of a very large and expensive memory
coupled with a much cheaper central processing unit. Is it not
possible to use this large memory more efficiently and to speed
up the whole system by spending more money on the central
processing unit? One example of a computer that attempts to
obtain a higher speed than a straightforward design would al-
low, is the IBM Stretch with 2-microsecond memory. The
extra speed is achieved by means of a look-ahead mechanism
that simultaneously picks up several instructions and, if pos-
sible, the data needed by these instructions. Unfortunately, it
appears that the effective operation of such a look-ahead sys-
tem requires that the machine designers anticipate how the
machine is to be used. A failure to anticipate means that the
computer will not be as fast as the designers hoped.

Some of us have worked on another scheme for obtaining
higher speed in a computer system. The idea is to build a system
of several separate central processing units and several separate
memories, in general more memories than processing units. The
central processing units all have the same order code and each
has the ability to address the whole memory. Boundary regis-
ters ensure that a processing unit does not transgress the limits
of memory set for it at a given time. When different processors
address different memory boxes, simultaneous memory refer-
ences occur, but when two processors request the same memory
box, a device called the arbiter simply holds up one of them
for one unit of time.

We propose that successive addresses be in separate memory
boxes. Based on this concept, it is interesting to note what
happens when two or more processors attempt to obey the

234

Time-Sharing Computer Systems

same program at the same time. This can happen when the
program is a system routine, such as a compiler, that the proc-
essors are running for different users. Since the processors are
working for different users, they will have different temporary
storage areas, as I explained before, and there will not be any
conflict there. However, there might be a conflict in getting
instructions. Suppose, for example, that two processors ask for
the same memory box. Then one of them is held up by the
arbiter and hence falls a step behind the other one. As long as
the processors execute instructions from successive addresses,
they will not have any further conflicts. As soon as one of the
processors executes a jump instruction, however, there is a
certain probability that there will again be a conflict and that
one processor will be delayed.

In a time-sharing system a processor executes a quantum of
program for a user and then continues in the round robin to
pick up the next user in line. Unless the number of users is a
multiple of the number of processors, the next time a user s
turn comes around, a different processor executes his program.
Thus we see that there is no particular relationship between
the number of users and the number of processors, nor is there
any assignment of particular processors to particular users. The
number of processors in a balanced system depends on the
relative costs of processors and memory, and on the size of
memory required.

Serial Processors

The relative cost of computer subsystems suggested an interest-
ing system to Ed Fredkin. It seems that 5-microsecond cycle
memory is now available at a reasonable price. Logical hard-
ware, however, also at a reasonable price, is feasible at 10
megacycles per second. This means that the time required
for one unit of logic is of the order of one-fiftieth of the mem-
ory cycle. Fredkin's idea is to use a serial processor in such a
system. This is a processor in which the arithmetic registers are
delay lines and the arithmetic is performed serially. Thus, in
an addition the bits in the word stream serially through the

235

J. W. Uauchly

adder. A serial processor is much cheaper than the parallel
processors currently used in large computer systems.

Computing as a Public Utility

In concluding I should like to say a word on management and
the computer of the future. At present, computers are bought
by individual companies or other institutions and are used only
by the owning institution. If computers of the kind I have ad-
vocated become the computers of the future, then computation
may someday be organized as a public utility, just as the tele-
phone system is a public utility. We can envisage computing
service companies whose subscribers are connected to them by
telephone lines. Each subscriber needs to pay only for the ca-
pacity that he actually uses, but he has access to all program-
ming languages characteristic of a very large system.

The system could develop commercially in fairly interesting
ways. Certain subscribers might offer services to other sub-
scribers. One example is weather prediction. A weather-predict-
ing company that is a subscriber to a central computer predicts
the weather but keeps the predictions in its private files. If you
subscribe to its service, your programs can gain access to these
files. You may even have weather-predicting programs run for
your benefit to answer your own particular questions. Other
possible services include those specifically connected with com-
puting, such as programming services. Some subscribers perhaps
might rent the use of their compilers. Other subscribers might
furnish economic predictions. The computing utility could be-
come the basis for a new and important industry.

Panel Discussion

Mauchly. Some people like a debate, but since in general
I agree with what McCarthy has said, I cannot start one. In-
stead I wish to observe that if we are going to talk about future
trends in computers, a look at the past could be helpful. For
instance, the first electronic computer of large-scale, digital type

236

Time-Sharing Computer Systems

had some features that are just now appearing in other com-
puters. This computer, the Eniac, was able to read, write, and
compute all at the same time. This was accomplished by a
combination of equipment which we then described as the
largest IBM plugboard in the world. The Eniac used an IBM
card reader and an IBM card punch. The programming was
not stored as is customary nowadays, but was set up by cables,
patch cords, switches, and similar things. As cumbersome as it
sounds, the computer could operate rather fast. How did we
do that? By making the computer operate in parallel fashion.
The Eniac was capable of something that is now coming back
into vogue: multiprogramming. The main difference today is that
we get much more capacity at a lesser cost. The departure from a
parallel computer, which was able to do many separate problems
simultaneously at rather high speeds, was made in an effort to
achieve something more economical, similar to the motivation
for the serial processor described at the end of McCarthy's talk.
The more we learn about how to make hardware perform
faster, the more we return to our efforts to perform larger and
larger calculations at cheaper and cheaper rates. We appear to
be traversing the full circle back to doing many things in paral-
lel. If that represents a trend, it seems to be a circular one. Per-
haps all we can deduce from this is that we tend to go around
in circles.

There is another subject that I want to touch on here, and
that is the subject of teaching machines. Perhaps this is not a
new idea, but it seems to me that what we would really like
to have is a learning-teaching machine. Almost any good teacher
claims, and usually claims truthfully, that in the process of
teaching he learns something. Yet practically all of the teaching
machines that I have heard discussed so far are ones that are
fixed-program. In other words, data from the student may feed
back to the program of the teaching device but govern only the
device's specific responses to the student. If the same student
were to go through the same stupid answers (we'll say) a second
time, the teaching machine might well respond in the same
way as before. For a human teacher this would be exceedingly

237

G. M. Amdahl

narrow. If a class does not understand a concept the first time,
having it explained one way, then the teacher should go at it
another way. One of the challenges of the future is that of mak-
ing the teaching machines learn.

My last comment is really a comment to machine designers.
We all now expect computers to talk to each other by tele-
phone. The data phone is an existing thing — not a promise,
but a reality. Someone mentioned to me earlier this evening
that a computer at the recent Western Joint Computer Con-
ference was operated by remote control through such a data
link. Here again you might say we have traversed the full circle,
but now it is a helical circle. We are making progress, however.
It was at the September 1940 meeting of the American Mathe-
matical Society at Dartmouth College that Dr. George Stibitz
of Bell Telephone Laboratories demonstrated a computer that
was operated from the Dartmouth academic halls although it
was located at the Bell Laboratories. The strange thing to me
is that in spite of the fact that everybody has seen telephone
communication between computers coming, including the tele-
phone company, you see very little in any present-day computer
specifications about the ability of a computer either to place
a telephone call or answer a telephone call. To my mind, this
is a singular omission. It was easy to predict that computers
should have this facility, and it is not a very hard one to
provide.

Amdahl. I am with Dr. Mauchly in being primarily in agree-
ment with Professor McCarthy. I do have some reservations,
however, about whether every individual user will desire to
operate his own console. Not everyone wishes to operate his
own typewriter, for example. I feel that, before it will win com-
mon acceptance, the language employed for communication
between user and computer will have to become much more
universal and efficient, yet be redundant enough to permit
varied problem formulation. I believe that typewriter consoles
should have larger character fonts to permit ready expression
within this language. Some inventions in the areas of keyboards

238

Time-Sharing Computer Systems

and print mechanisms to permit character font expansion eco-
nomically and reliably would be very welcome for computer
consoles.

With respect to time-sharing computer systems, Dr. John
Cocke of IBM envisages a novel and interesting design for a
multiple-usage computer that differs considerably in concept
from the one proposed by Professor McCarthy. This is a com-
puter that is one, yet many. It employs a large drum or disc
type of storage for the main memory. Also on this disc or drum
are arithmetic and address registers for a large number of sep-
arate computing units. As the drum or disc turns, these arith-
metic and address registers are processed sequentially by a
central processor; each one is processed for the execution of a
single instruction. Concurrent with the sequential processing,
memory references of all computing units are made as the ap-
propriate address angles are passed. Since the time of latency
for a memory reference never exceeds the time of latency for
the computing unit requesting it, the drum or disc appears to
be random-access memory to each computing unit. The central
processor employed for the sequential simulation of the com-
puting units does require a high-speed random-access memory
adequate to hold all memory reference data for one drum or
disc revolution. It could also hold all the individual arithmetic
and address registers.

This computer concept is suitable for a large number of low-
speed computers all time sharing a central processing unit. It
could permit an extremely large and an effectively random-
access memory at a very reasonable cost. It does have the dis-
advantage, however, of not readily permitting the allocation of
more than the capacity of a single computing unit to any one
problem. This allocation can be performed only for a problem
whose formulation permits simultaneous multiple processing
of portions of itself.

Professor McCarthy made no allusion to improved order
codes for his large private computer. It is my belief that it will
soon be possible to design computers whose internal language
will be much closer to the programming language employed

239

G. M. Amdahl

than is the case with present-day computers. I think it unlikely
that they will reach a one-to-one correspondence, but a higher
degree of correspondence would permit much more efficient
use of the computer. This is particularly true for programs that
are prepared almost entirely to fit the mathematical and physi-
cal concepts of the individual user rather than the computing
hardware. Compiling programs can be used to transform the
user's formulation into a more efficient one for computer proc-
essing. For debugging purposes, however, the user would also
require a pseudo-inverse transformation to relieve him of con-
siderations of the computer's internal properties.

I stated that computers might soon have order codes more
closely related to programming languages. A modest approach
to this appears in a logic-processing computer delivered to the
Rome Air Development Center in late 1960 or early 1961 by
the Aeronutronic Division of the Ford Motor Company. I de-
signed this logic processor in the spring of 1958 in response
to the need for evaluating logic equations directly without re-
course to a general-purpose computer for prior assembly and
compiling. I have since generalized the concepts somewhat so
that numeric evaluation of algebraic equations is handled in
a completely equivalent fashion. As an algebraic equation is
read, the arithmetic operations and the still incomplete set of
associated operands are stacked in sequence, and action is de-
ferred until the keystone operand for the locally inner paren-
thetical expression is sensed. At this time the stacked operands
and operations are executed in reverse sequence to the stacking.
When the lowest limit of parenthetical enclosure, consistent
with the present position in the equation, has been reached,
reading of the equation continues as before. Recognition of the
equality sign causes the final levels of the deferred operands and
operations to be executed, and the final value of the equation is
computed.

Such a computer organization is actually quite simple in
both concept and hardware. It is only a small step in the direc-
tion of designing a problem-oriented order code, but it is sig-
nificant that one common subset of a mathematical language

240

Time-Sharing Computer Systems

has yielded so simply. In my opinion, the invention of further
computing techniques that are appropriate to other subsets
requires only time and directed effort.

I have some final remarks to make relative to the requirements
of the time-sharing and multiprocessing concept favored bv
Professor McCarthy. In the requirement for relocatability,
he suggests that when a program ceases to be active it leaves
an odd-sized hole in memory. He also suggests that it must be
possible to move other programs down to fill this hole so as to
leave a contiguous block of space for new programs and data.
It is possible to generalize relocation more completely so that
such program and data transfers need not be performed. Rather,
the new program and associated data can be fitted into a selec-
tion of available odd-sized holes and be processed from this non-
contiguous storage. Incidentally, the same feature could provide
memory protection automatically. The newly announced Bur-
roughs B-5000 has this capability, if I interpret its description
properly. In such a time-sharing computer, progression through
a large number of users' programs would soon make memory
allocation look like an unplanned patchwork quilt. It would
appear completely orderly, however, to each and even* active
program.

In the multiprocessor-multimemory system suggested, the use
of more than one processor depends on the effective memory
traffic rate that can be achieved. The possible speed of proces-
sors appears to be increasing to the point where several 2-micro-
second memories could be exercised quite effectively by a single
processor. To achieve this, the complexity of the processor must
be greater than that of a processor connected to a single higher-
speed memory. It is doubtful, however, that this increase in
complexity would be as great as that of several slower processors
with a memory-reference arbiter serving them. The single
higher-speed central processor also has the advantage of being
able to provide the maximum computing capacity to a single
active program without requiring this active program to admit
to simultaneous processing of parts of itself. The single proc-
essor does have a disadvantage, however. If the processor mal-

241

J. McCarthy

functions, the entire computing capacity vanishes, not just a
given fraction of the capacity. Of course, acquiring the ability
to shift computing tasks from a malfunctioning unit to the
remaining good units would require a great deal of system-
program planning.

McCarthy. What I have to say in reply will deal mainly
with Dr. Amdahl's comments. I agree that we need larger char-
acter fonts and input-output devices more flexible than type-
writers. Even with present typewriters, however, a great deal
can be done, and input-output does not seem to be the main
limiting factor at present. With regard to John Cocke's com-
puter, I guess I do not fully understand what the relationship
is between the capability of such a computer and the capability
of the kind of computer that I described. It sounds like less
computer for a lot less money.

With respect to the matter of making order codes more like
the problem-oriented languages, I disagree with Dr. Amdahl's
point of view. It is true that order codes will be improved, but
it seems to me that this improvement is not going to reduce
by an order of magnitude the time required for a given calcula-
tion. I think that reduction by a factor of 2 is as much as can
be expected. My reason for believing this has to do with the
time consumed by the basic arithmetic operations themselves,
and by memory references for instructions and data. When we
first started programming the Lisp system for symbolic ex-
pressions, we were very enthusiastic about the possible advan-
tages to be gained from new instructions. It turns out that new
instructions could do things faster, but not much faster. The
way we saw this was by comparing the total memory references
that the computer made with the effective or unavoidable mem-
ory references. We arrived at an efficiency of about one-sixth
for the case of the IBM 709 computer. That is, one-sixth of
the memory references were references to actual working data.
(Adding references to instructions to references to data would
have improved the efficiency figure.) It seemed to us that if
we had had the instructions we wanted, we could have doubled

242

Time-Sharing Computer Systems

the efficiency. And if we have any say about the order codes of
subsequent computers, we certainly shall attempt to achieve
this. But I doubt that any larger improvement than this is to
be expected, either from order codes resembling problem-
oriented languages or from order codes allowing programs which
look like algebraic expressions.

Another question on which I disagree somewhat with Dr.
Amdahl is whether we have to move programs down to fill odd-
sized blocks, or whether we can leave them as they are. The
artificial-intelligence group at M.I.T. happens to be the pro-
prietors of a list-type programming system that does not move
blocks down but manages to get along very well with odd-sized
blocks. This works very well for certain kinds of calculation.
However, some rather ridiculous results have been obtained
by using the system where it is not suited. An example is
matrix calculation. It is true that a matrix can be represented
by a list of lists, where the individual lists run all through
storage, but addresses must be computed for matrix calculation.
If you determine that you want the seventeenth element in a
matrix, it unfortunately takes much longer to count seventeen
steps down a list (however you do it) than it does to proceed
directly to register A + 17. From our experience in using un-
ordered systems, we have come to realize that there are certain
important classes of calculations in which it is preferable to
retain the numerical method of addressing.

My final comment concerns the question of a single high-
speed processor versus several processors. It should be under-
stood that the answer to this question depends very much on
the current state of technology. For example, if some discovery
makes very high speed memory feasible, so that the ratio of
memory speed to logical speed becomes larger than it is at
present, then a very straightforward computer design would
be appropriate. But if the speed of logical operations remains
much higher than the speed of memory, a single processor has
to look very far ahead in the program to achieve efficiency.
The Stretch computer, for example, achieves its speed by
looking up to four instructions ahead. Even this is accomplished

243

General Discussion

with considerable difficulty and some inefficiency. With the
present ratio of speed of memory to speed of logic, it might be
necessary to accomplish as many as ten or twenty memory
references simultaneously. This would require the computer to
have still greater powers of foresight.

Amdahl said that the disadvantage of using several proc-
essors is that in order to apply all the capacity to a single
problem, you have to do some tricky programming; and even
then you may not succeed. For a computer that is used by
many people, this does not seem to be a serious limitation.

General Discussion

Anonymous. Professor McCarthy, you stated that the mem-
ory is relatively more expensive than the central processing unit.
But looking at your system organization, the only time sharing
you seem to be doing is on the central processing unit and not
on the memory. Would you care to comment on that?

McCarthy. If you have a very large memory, say on the
order of a million memory words, it will consist of a number
of boxes containing something like 16,000 words each. I did in-
tend that these boxes be operated simultaneously, so that the
memory would be time shared, at least in this sense.

Anonymous. But this is not the same kind of time sharing
as with your central processing unit. The same processor works
on all programs, but the same word of memory is assigned
only to one specific program. To infer your memory require-
ments, you multiply the number of programmers by the average
size of a program. To infer the requirements for your central
processing unit, you multiply by the average execution time of
a program. These are very different things.

McCarthy. Are you suggesting that different users share
the same memory register? This is possible only by making
rapid transfers to and from secondary storage, a procedure that
does not seem desirable. By the way, the memory used by sys-
tem programs is time shared. Although there are many proc-

244

Time-Sharing Computer Systems

essors in the system I described, there is only one set of system
programs. One of the flashiest features of the system is that
different processors can be executing the same system program
at the same time.

Anonymous. Professor McCarthy, are you familiar with the
real-time computer of Remington Rand? If so, I want to ask to
what degree you think this computer already meets the re-
quirements that you stated. One of its features that you did
not mention as being desirable is its ability to accept voice
commands. This reduces both the knowledge of special codes
which the operator has to have and also the probability of error
in input. The Remington Rand computer was designed espe-
cially for time sharing and is in use in the Capital Airlines
system.

McCarthy. I am not familiar with this computer, but I
will say one thing nevertheless. It does not have enough mem-
ory. As for voice commands, I am in favor of voice input to
computers, but as for reducing the amount of information the
user has to have, this is not as obvious a virtue as it may seem.
To tell the truth, the amount of information one has to acquire
in order to program is disgracefully small as it is, compared with
the amount of information one has to acquire in order to do
well in other fields. What is more important is to extend our
ability to use computers, make them do things that wre cannot
now make them do. As we develop the art of computation, the
amount that a user will have to know in order to use computers
with full effectiveness should increase rather than decrease.

Stevens. You have been talking about a central computer
and regional consoles. How about the possibility of regional
computers and merely a central arbitrator?

McCarthy. This question reminds me of the polymorphic
computers put forward by Ramo-Wooldridge. I think that such
systems are difficult to program, do not make enough memory
available to individual users in a straightforward way, and do
not allow as efficient sharing of system programs as the multi-
processor system I described.

Reynolds. Don't you think that if good input-output pro-

245

General Discussion

grams were available you could use secondary storage much
more effectively and not need such large primary storage? Do
you think the large primary- storage is really worth the cost that
it would incur?

McCarthy. This is a difficult question. Certainly for many
purposes a cheaper system could keep the user's program in
secondary storage and bring it in as part of each active quantum.
Several comments can be made on this possibility. First, very
large primary storage is not so terribly expensive and probably
will become cheaper. Second, when we get into the new applica-
tions of teaching machines, symbolic calculation, and so forth,
requests for individual actions are on the average going to re-
quire relatively little computation. Thus the cost of bringing
in a large program from secondary storage to do a small job
is going to loom larger than it does at present. Finally, certain
problems now being contemplated do require very large, di-
rectly accessible storage.

Piore. I am going to ask Dr. Mauchly to make any observa-
tions that he would care to make at this point.

Mauchly. We have confined our attention so far tonight
to the development of the digital computer, but the analog
computer also has had considerable development over the same
span of years, especially in the engineering fields. One of the
things that impresses me about the garden variety of analog
computer, if it is at all general purpose and not specifically
conceived to do just one job like process control, is that it is
very often set up with patch cords, cables, switches, and so
forth. Thus problem-setup time is of importance. Even before
analog computers went electronic, we had right here at M.I.T.
one of the large examples of the trend toward automatic setup:
a differential analyzer of the mechanical variety which was set
up from punched paper tape. The trend toward automatic
setup continues, but there is evidence of an even more alarming
(or encouraging) trend. Instead of fighting over the problem
of whether a particular calculation should be done by analog
methods or digital methods, there is a trend toward combining
the two to produce some sort of hybrid unit, with both analog

246

Time-Sharing Computer Systems

and digital computers cooperating to solve the same problem.
It seems to me that for the kind of situation which Professor
McCarthy envisages, a combination like this might be very ef-
fective, provided that the analog part could be set up from the
remote consoles.

Piore. Gene [Amdahl], have you any comments?

Amdahl. I should like to pursue Dr. Mauchly's idea a little
further. In addition to attaching an analog computer to the
system, one might also attach other kinds of digital computers
that are individually designed to be particularly good at certain
functions. In symbol-manipulating programs, for example, cer-
tain features of a digital computer might be heavily used, while
other features are not used at all. For instance, the floating-
point capabilities may not be used. One readily can imagine
a system in which several simple computers, each one slanted
toward a particular task, are attached to the same memory.
Different portions of the same program then would be executed
on different computers. This would be sort of an extension of
the Gamma 60 computer system.

Piore. John [McCarthy], do you want to have a final word
before we close the meeting?

McCarthy. Well, it is difficult to find a final word that is
somehow worthy of the honor of closing the meeting. All I was
going to do was quibble a little more on the last point that was
made. The idea of having specialized computers for specialized
tasks in order to save money on the hardware seems like a good
one, but if we have to program the assignment of different parts
of a task to different processors, we had better be sure that the
money saved on hardware is not spent on programming.

References

1. Strachey, C, "Time Sharing in Large, Fast Computers," in Informa-
tion Processing, Proceedings of the International Conference on In-
formation Processing, UNESCO, Paris 15-20 June 1959, UNESCO,
Paris, 1960, pp. 336-341.

2. Licklider, J. C. R., "Man-Computer Symbiosis," IRE Transactions on
Human Factors in Electronics, Vol. HFE-1, 4-11 (Mar., 1960).

247

References

3. Bush, Vannevar, "As We May Think," Atlantic Monthly, Vol. 176,
101-108 (July, 1945).

4. Slagle, J.. "A Heuristic Program that Solves Symbolic Integration Prob-
lems in Freshman Calculus, Symbolic Automatic Integrator (SAINT)/'
Ph.D. thesis, Department of Mathematics. Massachusetts Institute of
Technology, May, 1961.

248

7

A

New Concept
in

Programming

7

A

New Concept
in

Programming

Speaker George W. Brown

Professor of Business Administration
Professor of Engineering, and
Director, Western Data Processing Center
University of California at Los Angeles

Discussants Grace M. Hopper

Director, Research — Systems and Programming
Remington Rand Division
Sperry Rand Corporation
David Sayre

Corporate Director of Programming
International Business Machines Corporation

Moderator Philip M. Morse
Professor of Physics

Director of the Computation Center, and
Director of the Operations Research Center
Massachusetts Institute of Technology

250

Brown. It is no secret to people attending this series that
the complexity of processes being considered for programming
is almost without limit. Military information and command
systems, business data processing, simulation, and artificial in-
telligence all may yield extremely complex structures. To pro-
gram these complex structures may require hundreds of thou-
sands or even millions of machine instructions. I should like to
discuss parallelism and the implications of parallelism, both in
the programming and modeling of these complex processes.
By parallelism I mean the doing of things simultaneously,
rather than in series. The programming effort must be appor-
tioned to teams of people who can work reasonably independ-
ently on subsections of the total program. And the model may
have to reflect a multiplicity of external demands inherent in
the process. For example, a military command system typically
has a number of recognizably different needs, any one of which
may request service at some time.

In addition, there may be parallelism in the structure of the
computer. I think we all remember the cartoonist's picture

251

G. W. Brown

(some cartoonists still use it) which shows a large room-size
cabinet with lights and dials and switches. That is a computing
machine to the cartoonist; but no major computing system
today is anything like this monolithic image. Today's computer
system may have twenty to forty separate boxes, each one of
modest size. The boxes are of various kinds and are mysteriously
interconnected. Within such a computing system there is con-
siderable specialization of function.

The simple-minded, traditional picture of the monolithic
machine is about sixteen years old. The picture includes a
processing unit with a large repertoire of operations which does
all kinds of arithmetic and logic; a memory unit which stores
data, intermediate results, and instructions, all in the same
form; and finally, some rather vague and unimportant set of
input-output devices. The picture is hopelessly inadequate as a
description of any computing system today, yet it dominates
current programming and operating approaches. If you observe
almost any computing system in operation, you find that, of
the many boxes, relatively few are in operation at any particular
time. You also find that portions of the system work very hard
and very fast when they are working, but they do not work a
high enough percentage of the time. To use the engineers'
description, their duty cycle is very low.

In principle, the time sharing of equipment by diverse tasks,
coupled with the specialization of processing elements in paral-
lel, could permit a far more efficient utilization of the equip-
ment than is currently obtainable. And here the analogy of a
job shop is not a bad one. A job shop has at any given time
a large number of different tasks to do, generally by means of
a mix of machines, some of which are duplicated many times
over. The amount and type of machinery are chosen to balance
with expected work loads. Time sharing of the machines per-
mits a reasonable match and produces reasonable duty cycles
on the most expensive equipment.

I do not think that it is stretching the analogy too far to say
that modern computing equipment is already like a job shop
and could be much more like one if we had an effective way

252

A New Concept in Programming

of utilizing parallelism. The parallelism is of two kinds: One
is in work load, and the other is in elements. I hope to sketch
for you an approach that depends on two conceptions for deal-
ing with these requirements of parallelism: conceptions of a
programming process and of an operation process. First, I want
a method for the logical description of problem requirements
which is free, at least above a certain minimal level, from the
kind of overspecification of sequence constraints that character-
izes present programming. Second, I want a processor which will
make dynamic assignments of subtasks to the various processing
units, as they become available, in accordance with logical speci-
fications and priority requirements. Now I assert that there are
tools currently available for this approach. I have no hesitation
in saying that the approach is inevitable and in its broad out-
lines will be just about like what I am going to describe.*

In this paper I outline the elements of computer programs
and processors which would allow program execution to be free
from arbitrary sequencing constraints unrelated to underlying
logical requirements. An approach of this nature would permit
flexible operation of complex data-processing systems that ex-
hibit parallelism in structure and are capable, in principle, of
dealing with task assignments in parallel.

The Problem

The brief modern history of computing technology has been
characterized by a series of mismatches: first, between various
portions of the computer anatomy, and second, between the
computer hardware and programming and operating technology.
The earliest electronic computer consisted of a relatively rapid
arithmetic and control unit, with memory of limited capacity,
and with minimal input and output facilities connecting di-
rectly to the arithmetic registers of the machine. This early com-
puter literally operated one step at a time on programs labori-
ously written one step at a time, usually in machine language

* This first portion of Professor Brown's presentation is based on an
extemporaneous introduction to his prepared paper. The prepared paper now
follows.

253

G. W. Brown

and without programming aids of even the most elementary
kind. Storage limitations as well as input-output limitations
made programming aids unthinkable.

In a period of approximately fifteen years, high-speed mem-
ories have grown relatively large, auxiliary memory devices of
enormous capacity have been added, and luxurious input-output
and peripheral equipment of great versatility" has been devel-
oped, with the result that under many circumstances the typical
large system may be limited now by its arithmetic unit. (Of
course, it is still possible to exhibit applications in which the
limitation rests in the input-output equipment.) Accompany-
ing the hardware development, an impressive array of program-
ming aids has been adopted (often reluctantly). These aids
range from the mnemonic use of alphabetic characters to rep-
resent instruction codes, to the symbolic designation of memory
cells to permit automatic assignment of actual memory ad-
dresses by an assembly program, to the most elaborate compilers
and source languages. Automatic operator systems with sophis-
ticated loading, debugging, and monitoring facilities also have
been developed.

Despite these achievements, it is fair to say that computer
programming still shows more than vestigial traces of its an-
cestry. This is particularly evident in the requirement for se-
quencing decisions which is imposed upon the programmer.
Almost even- problem of any magnitude has subsections whose
actual order of execution is immaterial logically; but existing
practice demands that the programmer arbitrarily make se-
quencing decisions. The final program, whatever the source
language, has these arbitrary sequencing decisions congealed
and, indeed, thoroughly concealed in it.

If we look to the near future, machine systems will bear little
resemblance to their ancestors of a few years ago. Communica-
tion links will connect portions of systems at great distances
from one another, there will be multiple remote interrogation
stations, and the diversity of memory and auxiliary equipment
will increase. Arrangements of arithmetic, control, and logical
processing devices of either general or special nature may be-

254

A New Concept in Programming

come highly complex. It is not appropriate to view such ar-
rangements as single machines, surrounded vaguely by input-
output equipment. Most actual operations with large systems
involve the input-output equipment deeply in the operations
themselves. The logical structure of these systems no longer is
adequately represented by the older image of a one-step-at-a-
time machine, and programs conceived for such machines make
it impossible to get efficient equipment utilization. This fact
is testified to by the operation of many large computing systems.
Often only a few of the magnetic tape units are in operation
at any one time, and there are frequent intervals when the
arithmetic unit idly waits for tapes to rewind or for some other
operation to take place. In an absolute sense, it does not appear
that anything like reasonable efficiency is attained. The more
complex the system, the more striking is the phenomenon.

There is little doubt that full realization of the inherent
power of complex systems is seriously impeded by difficulties
associated with the prevailing programming philosophies. Given
a complex program, whether prepared by a single individual or
by a group effort, it is effectively impossible to predict with
any reliability the detailed course of the program, the times and
durations of activity of the various units in the system, or the
pattern of storage requirements as a function of time. Indeed
these factors may depend critically on the input data. As a
result, compiling a complex program in advance as a large,
fixed unit removes flexibility and introduces inefficiencies in
operation over and above those attributable to mismatch be-
tween the problem requirements and the machine configura-
tion. Operating on two or more different programs in parallel
might reduce the over-all mismatch between tasks and facilities,
but it is difficult to see how to proceed in this direction beyond
elementary buffering of inputs and outputs as long as any one
of the rigid units of program may make arbitrary demands on
any portion of the machine at almost any time.

This difficulty is related to the problem of translating pro-
grams prepared for one machine into programs for a different
machine. A program written in machine language may fabricate

255

G. W. Brown

instructions during execution. If an attempt is made to translate
the program in advance, the fabricated instructions, absent
from the original program, will not be translated. As a result,
translation often is accomplished by "simulator" programs that
translate in dynamic fashion one instruction at a time as the
program unfolds. This interpretive approach is usually ex-
tremely slow, and is unsatisfactory as a general solution to the
problem. Here again the fundamental obstacle is the fact that
programs are in general unpredictable before execution.

The problem is one of attaining more efficient utilization
of machine elements, either by time sharing or by specializing
processing elements in parallel. Nekora1 distinguishes between
multiprogramming, "the execution of several programs by trans-
ferring control among them," and multiprocessing, "the simul-
taneous sharing of memory by different processors." Both are
needed and generate, from the point of view of this paper,
identical requirements for effective operation. In a discussion
of multiprogramming, Ryle2 formulates ten design require-
ments, of which the last is that "the job of the programmer
must not be made more complex or difficult." In the light of
our earlier discussion, we cannot disagree with this requirement,
as long as it is not interpreted to mean that the system must
not interfere with the programmer's sacred right to call upon
any part of the machine configuration at any time and without
notice. At least one attempt at multiprogramming has adopted
such an interpretation, with the consequence that the effort
has reduced to an attempt to schedule time sharing in advance
on the basis of programmer estimates of running time. On the
face of it, such an approach does not offer much. The depend-
ence upon estimates violates Ryle's requirement, and there is
not even any assurance that these estimates will be realistic.

A Proposed Approach

A solution to multiprogramming and multiprocessing may be
found in a further separation between the functions of pro-
grammer and operating system. This could be accomplished

256

A New Concept in Programming

by reducing the substantial amount of logical overspecification
currently forced upon, and indulged in by, the programmer.
By withholding the superfluous sequencing controls imposed
in present programs, which are not necessary to the logic of the
task, sufficient flexibility will be made available to an operating
program so that time sharing can be scheduled dynamically as
tasks proceed, using very simple rules of assignment.

Duly noting (and discounting) the usual screams of anguish
from programmers, it should be recognized that what the pro-
grammer is asked to relinquish is not the right to program an
operation but rather the unrestrained right to specify implicitly
exactly when it will be executed. The programmer will be re-
quired not to do more work but to do less; he will be asked to
confine himself to the problem of defining unambiguously what
he wants to occur, and to give up at least some of the problem
of determining in advance when it will occur; he will be asked
to furnish certain structural information that is available to
him during the analysis of his problem rather than to conceal
it within his program. Programmers, conditioned perhaps by
the programming needs of the early years (hence the screams),
exhibit a compulsion to overspecify and to overcontrol. Com-
pulsive overspecification of sequence controls must be given
up, and the machine must be made to perform the logically
trivial bookkeeping and assignment tasks.

The approach suggested has a number of by-products of con-
siderable importance. Allowing the programmer to stay closer
to his problem and allowing him to delegate the unessential
decisions to the machine will simplify his job and reduce the
frequency of logical errors. The programs submitted by pro-
grammers will be in a more appropriate form for coordinating
group effort on complex problems; modification by sections
will be easier; the number of recompilations caused by program
alteration will be reduced; and programs will be less dependent
upon exact machine configurations. There might also be a con-
siderable reduction in the amount of reprogramming necessi-
tated by the substitution of one machine for another.

257

G. W. Brown

Two Lerels of Logic

The approach advanced here presupposes two levels of logic
in the finished program. Program sections of the lower level
are designated atomic processes. They may be specified in any
programming language which the system is equipped for. sub-
ject to the limitation that the absolute locations of input, out-
put, and program at execution time > must be either at the
disposal of, or known to. the executive program. In order to
make the necessary assignments, the executive program must
have available to it information on the types of units and
storage required by even- atomic process. WTiile no real re-
striction is thus imposed or intended, it is clear that atomic
processes must be specialized enough to allow the executive
program some flexibility-, yet general enough to deserve the
effort of the executive program. Unnecessary absolute assign-
ments clearly ought to be avoided in the specification of the
atomic processes. Atomic processes may reside in the system
library or may be written ad hoc by the programmers.

The upper level of logic consists of statements of a particular
language used to define processes. Whereas the atomic proc-
esses may be essentially anything, and hence their language
must be machine-dependent, the language of process definition
might reasonably be standard for all digital computing equip-
ment. The logic of the upper level simply specifies the structure
of a program, ultimately in terms of atomic processes. One
requirement imposed on the language of the upper level is that
the effective logical result of any statement, as far as the pro-
grammer is concerned, be independent of its serial position
with respect to other statements. Relative positions may or may
not affect the priorities of execution, depending on the system
design and the scheduling requirements imposed.

Tasks are described in terms of suitable process and input
definitions. The executive program examines these definitions
to specify atomic tasks (by specifying the inputs for the corre-
sponding atomic processes | and to assign atomic tasks to units.

253

A New Concept in Programming

Completion of an atomic task causes further task definition by
the executive program. Tasks may be added or canceled at any
time.

The system proposed does not distinguish between tasks of
different origin; it therefore identifies the problem of parallel
jobs as equivalent to the problem of providing for the efficient
execution of a single job. The executive program is essentially
an interpreter of the upper-level language that assigns tasks
specified in the lower-level languages. Thus, if one views the
atomic processes as having been compiled previously, the ex-
ecutive program is an interpretive routine that assigns sub-
routines dynamically. It contains assignment tables and other
bookkeeping devices of the type normally present in compilers.
The interpretive character at the upper level makes possible
the monitoring functions described and provides the efficiency
desired. It seems reasonable to expect that the simplest of
queuing rules, coupled with the abilities to determine when a
unit is available and when a task may be assignable, will give
relatively good performance. To estimate performance gains un-
der a range of conditions, McKenney3 simulated a system of the
sort described.

The Upper-Level Language

Turning to the upper-level language, it may be observed that
algebraic languages, such as Fortran or Algol, permit the con-
struction of certain expressions without necessarily implying
the whole computational sequence. Thus an algebraic statement
of the form z = a(b + c), where a, b, and c are themselves com-
posite expressions, carries no logical implication about the
sequence in which a, by and c are to be calculated. Whatever
the sequence, z may be computed after the factors a and b + c
are made available. Even though the statement allows certain
options, the compiler will normally specify a particular sequence
that then will be fixed. If this were the only obstacle to the use
of such languages, it could be removed by proper separation of
the functions of the compiler and of the executive program.

259

G. W. Brown

Unfortunately, the existing statement languages require com-
plete specification of statement sequence, so that statements
are assumed to be executed in turn. Programmers make non-
trivial use of this property, as in the process of using a storage
cell both for a completed result and for temporary storage dur-
ing a calculation, simply by using the same symbol in different
parts of the program. Constructions such as n = n — I, for ex-
ample, must be properly placed within a program. The logical
specification is therefore not invariant under permutation of
statements. Analysis of such programs to determine which posi-
tions are movable or not is difficult to contemplate. Apparently
what is required is a statement language exhibiting functional
dependencies, with sequence constraints determined implicitly
by the dependencies and with the statement sequence excluded
from any role in specifying the logical relations. Statement
sequence, however, might play a proper role in assigning priori-
ties for execution.

An adaptation of McCarthy's4 formalism for defining func-
tions recursively fulfills our requirements for the upper-level
language. We share McCarthy's belief that this formalism has
advantages as a programming language. To adapt his approach
to the recursive definition of symbolic expressions, atomic proc-
esses must be treated as an unlimited set of primitives in terms
of which all processes to be executed are defined. In what fol-
lows, conventional functional notation will be used to denote
processes relating output to input. The functions represented
need not be single-valued.

Following McCarthy, we use conditional expressions of the
form

(Pi->ei, -",pn-*e„)

where the p's are propositional expressions and the e's are ex-
pressions of any kind. We examine the p's from left to right
and set the expression equal to the e corresponding to the first p
encountered with truth value T, provided no undefined p is
encountered earlier. The conditional expression is undefined

260

A New Concept in Programming

if no such p is found, or if the e that is designated is un-
defined.

Taking a Square Root

Consider McCarthy's example of the Newtonian algorithm
for obtaining an approximate square root of a number a, given
initial approximation x, and error limit e:

sqrt (a,x, e) = (|x2 - a\ < e->x, T-^sqrt (a, Vi(x + ~ ),e))

T stands for the truth value T and serves, in the conditional
expression, to provide an alternative if the predicate |x2 — a\ < e
is not true. This usage corresponds to the ''If . . . else . . ." of
Algol. Actually, the expression might as well be abbreviated
by dropping "T-> ." We might now write this as

sqrt (a,x,e) = (P(x, a, e) -> x, sqrt (a, f (x, a) , e) )

where P stands for an atomic process such that P(x, a, e) —T
when |x2 — a\ < e, and / stands for an atomic process such that

f (x, a) = Vi(x + £). Note that it is assumed that the executive

program knows how to identify x, ay and e for the execution of
P and /, and knows how to get at the truth value of P(x,a,e).

Suppose now that a task is assigned in terms of the process sqrt
and the input data elements corresponding to a, x, and e\ denote
the task by sqrt (A, X, E), distinguishing the indicated data ele-
ments from the bound variables appearing in the process descrip-
tion. The first atomic task assigned then would be P(X, A, E).
Without logical error the executive program could also assign
f(X, A) at the same time, which corresponds to looking ahead
and possibly computing something that may not ultimately be
required. We shall assume, however, that the executive pro-
gram does not normally proceed beyond the first undefined ante-
cedent in a conditional expression. With such a rule no
unnecessary processing will be performed. Options may be per-
mitted the executive program in such instances; also there exist

261

G. W. Brown

alternate process descriptions that get around the rule if desired.

If P X. A. E has truth value F, the executive program now
replaces the task sqrt (A. X, E by the alternate process sqrt
[a9f(x9a)9e) applied to (A, X, E), that is, bv the task sqrt
(A,/(X, A),E). Following this the task P(/(X, A),A,E) will
be assigned. The next atomic task to be assigned is f(X,A),
followed by P.X'.A.E , where X' is the output of /(X,A).
It is seen, therefore, that the executive program modifies the
assigned tasks according to the execution of assigned atomic
tasks, while retaining access to the original process definition
for sqrt, which enters again and again. The rule suggested in
the previous paragraph is simply that expressions of the form
(p—>eue2) will cause no assignments beyond the arrow as
long as p is undefined. The truth value T will cause assignment
of the task corresponding to elt while F will cause assignment
of the task corresponding to e2. The initial sequence of assign-
ment is (p -*elye2), from which the assignment of p is de-
rived, thence, as described.

This informal description of the executive process may also
indicate how the executive program can tell when storage space
may be released for items no longer required. For this purpose
we require a look beyond the conditional arrows, even though
assignments are not yet made there. Abbreviating sqrt by
[p — » el7 e2 . we see that X is required for p, ez. or e2, A is
required for p or e2; and E is required for p or e2. Evaluation
of p as T eliminates e2 from further consideration, with the
result that A and E can be released. Evaluation of p as F causes
assignment of e2 and elimination of e3 as a candidate, with the
result that X, A, E are now required only for e2. Assignment
of e2 leads to evaluation of X' = f(X, A), after which X is no
longer required and may be dropped by the executive program.

It can now be observed that the executive program processes
lists of symbols of data items, expands and contracts these
lists, notes which items are defined, proceeds in one direction
to find tasks defining the undefined items, and proceeds in the
other direction to determine where an item to be defined is or

262

A New Concept in Programming

may be required. The executive program should also be able to
interpret the propositional connectives A ("and"), V ("or")>
and ~ ("not"). We suggest that the definitions given by Mc-
Carthy, which are noncommutative under certain conditions,
should not be used here. To be consistent with the philosophy
of this paper, we should expect p A <7 to be false if either is
false and p V q to be true if either is true, even when the other
constituent expression is undefined. The executive program
may assign p and q simultaneously, or may defer assignment of
one pending determination of the other. If it is desired to im-
pose a particular sequence of evaluation, this may be achieved
by use of the conditional expressions proposed by McCarthy
as definitions. It would seem appropriate to leave the flexibility
of assignment to the executive program through use of the
connectives.

A Second Example

The next example will illustrate the manner in which the
executive routine might make use of the latitude inherent in
the problem structure. Suppose the transformation F operates
on the pair (x, y) to produce {ax — by,bx + ay), with param-
eters cz, b. Suppose that the result of applying F twice in suc-
cession is required, using the same parameters each time, with
x = X, y — Y, and (a, b) = (sin C, cos C); X, Y, and C being
given. Taking as atomic processes g(x, y) = xy, f1(x, y) = x — y,
f2(x, y) = x + y, and h(0) = (sin 0, cos 0), we might describe
this task as follows:

Z = F(F(X,Y,h(C)),h(C))

F(x, y, a, b) = (h (g(a, x),g(b, y) ),f2(g(b, x),g(a, y) ) )

In list form the top line may be written

Z = F(A, L,M)
A = F(B,L,M)
(L, M) = h(C)
B=(X,Y)

263

G. W. Brown

and the bottom line may be written

F(x>y,a,b) = (s,t)
s = fi{ulf Vi)
t = f2(u2, v2)
U! = g(<r,x)
Vi = g(&,y)

l/2 = g(M)

v2 = g(tf,y)

Clearly execution of both F(B,L,M) and F(A,L,M) will re-
quire interpretation according to the second list in order to as-
sign atomic tasks whenever possible. Assuming only that this
interpretation takes place fast enough to keep up with the exe-
cution of atomic tasks, in whatever sequence they may be com-
pleted, we can observe that at the outset only the atomic task
h(C) is assignable, yielding L and M. After completion of
h(C) the four atomic tasks g(L,X), g(M,Y), g(M,X), and
g(L, Y) become assignable, coming from the task F(B, L, M) .

We may now observe that A± = /i(g(L, X), g(M, Y) ) will
be assignable as soon as both g(L,X) and g(M, Y) are com-
pleted, and that A2 = f2(g{M, X), g(L, Y) ) will be assignable
as soon as both g(M, X) and g(L, Y) are completed. In
turn, completion of Ai makes g(L, Ai) and g(M, Ai) assign-
able, and completion of A2 makes g(L, A2) and g(M, A2) as-
signable, as part of the calculation of F(A, L, M). Note there-
fore that it is possible for the calculation to proceed as far as
g(L, Ax) and g(M,A1) before completion of g(M, X) or
g(UY).

The flow diagram of Figure 1 displays the logical structure.
It is clear that considerable freedom exists to make assignments
consistent with this logical structure, in contrast to the custo-
mary restriction to a single arbitrary sequence specified by the
programmer. It is important also to observe that the process of
problem specification is simplified by expressing the logical
structure directly in terms of the relationships involved, with-
out need to expand the structure to its form during calculation.
It may further be observed that the executive program will take

264

A New Concept in Programming

L .

M_

Figure 1 Flow Diagram of Logical Structure

care of the red tape usually associated with initializing the
various subroutines at the appropriate times, thus eliminating
a common source of programming error. The system proposed
therefore provides most of the advantages associated with com-
mon algebraic programming languages, without suffering from
the customary limitations and restrictions.

Ackermans Function

An interesting example is given by the following recursive
definition of Ackerman's function, A(m, n) :

A(m, n) = (m = 0-> n + 1, A(m - 1, U(m, n) ))
U(m,n) = (n = 0-»l, A(m,n - 1))

Here A is defined (for nonnegative integers) in terms of itself
and of U, whereas U is defined in terms of A. We shall analyze
the task A(l, 1) from the point of view of an executive pro-
gram, taking as atomic processes the evaluation of the predicate
m = 0, the addition of 1, or the subtraction of 1.

At the outset the task A (1,1) causes assignment of the
predicate m = 0 with m = 1. Since the result is F, the task
A(l, 1) now becomes the task A(m —1, U(m, n) ) with m = 1,
n = 1. The atomic subtraction task is now assigned, together
with the task 17(1, 1), which causes assignment (again) of the
predicate m = 0 and m =l. Completion of the subtraction

265

G. W. Brown

would give the task the form A(0 U(l, 1) ), while completion of
the predicate evaluation would yield A(m — 1, A(l, n — 1) ),
with m — 1, n = 1, in which case a second subtraction becomes
assignable. From this point a certain amount of variety remains,
with respect to the order of completion of the atomic tasks
which enter. For example, interpretation of the task A(0,
1/(1, 1) ) by means of the definition of A leads to the atomic
task which evaluates the predicate m — 0 with m — 0, so that
A(0, U(l, 1) ) would become U(l, 1) + 1, putting the burden
of further progress upon the evaluation of U(l, 1), since the
atomic addition process will not be assignable until U(l, 1)
has been determined. Interpretation of U(l, 1) leads to A(l, 0)
where again the leading predicate may or may not be evaluated
before completion of the subtraction. A (1,0) becomes A(0,
U(1,0)), so that finally the computation (1 + 1) +1 will
have to be performed, causing two consecutive assignments of
the atomic task that adds 1.

A Complication

A very simple example will illustrate a possible practical com-
plication in the operation of the executive routine. While the
complication in question may be handled in a number of ways,
further study is required before the problem can be disposed of
completely. Suppose the function F(x, n) is to be defined as
the nth iterate of the function f(x), perhaps by the definition
F(x, n) = (n = 0 — >x, F(/(x), n — 1) ). Application of the defi-
nition must lead ultimately to execution of /(x), followed by
execution of / upon the result, and so on, until / has been exe-
cuted n times. Viewing the process of subtraction of 1 as
atomic, along with whatever atomic processes are required in
the definition of /(x), we note that the sequence of comple-
tions of atomic tasks, if permitted, might allow the counting
process to go far ahead of the steps that evaluate /. Thus we
might get the task represented as F(f • • '/(*), n — /) without

j-times

having evaluated / even once. While no logical error is thereby
committed, the effect is rather useless and may fill up memory.

266

A New Concept in Programming

The executive routine is simply laying tracks ahead of the
actual requirement. If we consider the counting operation as
being composed of atomic tasks, however, the executive routine
is achieving useful results.

Before commenting further on the point just raised, consider
the possible alternative definition

F(x,n) = (n = 0-+x,f(F(x,n-l)))

Application of this equally valid definition would start laying
down f s from back to front, while counting down from n; given
a simple-minded executive routine, it would lay down all n of
them before the first execution of / could take place. In this
case, if n is very large, there might be inadequate storage to
represent all n iterations, hence the task might fail to be exe-
cuted. Before adopting a possible solution to the problem
raised by the first definition, one must decide whether the solu-
tion should also cover the second. Ideally both definitions
should lead to the same result, and indeed they would, given
unlimited storage for use of the executive routine.

Some Obsenations

In the case of the first definition the difficulty may be over-
come by protection against overrunning the allotted storage, in
which case executions of / will always rescue the situation and
permit the process to continue. In the case of both definitions
the difficulty may be resolved by equipping the executive rou-
tine with its own counting process and with a compact repre-
sentation for iterated processes corresponding to the familiar
"do" statements. WTiile it is reasonable to imagine the execu-
tive routine so endowed, this may not prevent the programmer
from defining processes containing similar elements that need
not be recognized as such by the executive routine. In this case,
however, the problem is the same as that of representing, in
finite space, a program which requires a larger space before any
atomic task can be isolated. Universal solution of this problem
would be miraculous.

We turn now to the matter of queuing disciplines, both with

267

G. W. Brown

respect to tasks assignable to appropriate processing units and
with respect to unraveling of definitions by the executive routine.
It would be reasonable to expect that a sensible executive routine,
which operates rapidly compared with the execution of atomic
tasks, would refrain from forming unnecessarily long queues
for processing units. Given this property, the adoption of a
"first-come, first-served" discipline for queues associated with
like processing units would solve the particular difficulty raised
by the first definition. The executive routine would assign
atomic tasks belonging to the execution of / and continue un-
raveling task definitions only as long as processing units were
available. The laying down of f s would proceed unnecessarily
far only if the execution of f required different kinds of process-
ing units from those required by the counting process, and then
only if the counting process were rapid in comparison with the
execution of /. A "first-come, first-served" discipline seems ade-
quate because the major gain in efficiency very probably stems
from the ability to determine dynamically the availability of
processing units and the tasks assignable to available units. In-
tentional violations of this discipline may be used to provide
interrupting facilities for the accommodation of "on-line" ac-
tivities.

Other possible solutions to the difficulty may result from
sequencing conventions adopted by the executive routine. Thus
one might demand that task definitions be applied, whenever
possible, by going inward with respect to parentheses. Such a
rule leads to an attempt to evaluate all arguments of a task
before expanding the task definition further. Returning to the
discussion of Ackerman's function, just after the first application
of A, we have the task A(m — 1, U(m, n) ), with m — 1, n = 1.
The above rule requires that m — 1 be executed, and then that
1/(1, 1) be expanded as far as possible before A(0, U(l, 1) ) is
interpreted according to the definition of A. The sequence is
A(0,A(1,0)),A(0,A(0,U(1,0))),A(0,A(0, 1) ), A(0, 2), and
finally the result 3. The rule seems unnecessarily restrictive
as a universal constraint. It corresponds, of course, to a weak-
ened extension of the rule that atomic tasks shall not be as-

268

A New Concept in Programming

signed until all arguments are defined. Since many nonatomic
tasks have the property that considerable work may be ac-
complished on them before all arguments are specified, this
rule probably should be avoided.

An alternative convention may be made to depend on clues
furnished by the programmer. We can view the assignment of
a task to the executive routine as the beginning of a process
in which the executive routine assigns derived subtasks to itself
until atomic task assignments can be made. The process de-
scription then merely furnishes a list from which subtasks are
assigned as they become appropriate. Assume that provision is
made to mark particular subprocesses of a process so that the
corresponding subtasks are automatically assigned to the execu-
tive routine at the same time the parent task is assigned. This
would provide a partial control over the sequencing of the
executive operations. In the case of the first definition of the
iterative process, the subtask / could be so marked, as could
specific subtasks of it. The effect would be to enforce early as-
signment of certain derived subtasks, thus imposing a prefer-
ence pattern.

Other refinements of the executive procedure may be made
to economize on storage requirements or to obtain improved
efficiency in the absence of balanced over-all loads. As an ex-
ample, it is not inconceivable that the executive routine might
score assignments in terms of the unblocking they generate and
choose first those assignments which enable the greatest num-
ber of derived assignments. Such a device is reminiscent of
other scheduling problems, particularly the job shop problem.
Indeed our problem is more than superficially related to other
problems that require the description and scheduling of com-
plex processes; we should expect the methods suggested here
to have fairly general application.

Conclusions

In closing, it is appropriate to comment briefly on the prob-
able methods of implementing a system of the type proposed
and to point out some relationships between this approach and

269

G. W. Brown

other recent developments. The presentation here has been in-
formal; its formalization rests upon McCarthy's structure for
representing functions of symbolic expressions as symbolic ex-
pressions. The fundamental operations defined by McCarthy
are represented by corresponding operations of our executive
routine. It is clear that an executive routine of the type sug-
gested here will be primarily a list processor. Presumably argu-
ments of processes and their outputs will be represented by
symbols that are keys to lists and list structures; processes will
be list structures and so will the derived dynamic task defini-
tions. In view of the fact that list processing was devised to
provide a powerful tool for symbol manipulation, it should not
be surprising that it may provide the basis for implementing a
system of the kind proposed.

There are. of course, some differences in points of view. Out-
right abandonment of sequence definition in program specifi-
cation appears to be a radical proposal, in spite of my emphatic,
continued advocacy. List processing, as programmed in exist-
ing systems, insists that sequences be nailed down. See, for ex-
ample, Newell and Tonge5 on IPL-Y. In treating a problem
closely related to ours, Evans, Perlis, and Van Zoeren6 "thread"
the computational sequence through their lists, because "tree"
representations do not specif}- computation sequencing. Our
position is that the "tree" representation ought to be substan-
tially all that is required. Another departure is the separation
between process lists and dynamically defined task lists. This
eliminates the need for separate "push-down" lists and "depth
counters'"7 to keep track of where the process is, since the as-
signed task list effectively acts, itself, as a set of push-down lists
to keep track of progress and to update the various stages.
"Two-way" links in the process and task lists will permit shut-
tling up and down in levels and will facilitate determination of
when to release data lists. Admitting that somewhere, somehow,
some process ultimately establishes sequences is far from ad-
mitting that programmers should define the sequences uniquely
in advance.

In view of the relatively slow speed of present list-processing

270

A New Concept in Programming

methods, the separation into two levels of program description
may make the scheme practical. This follows since the process-
ing time of atomic tasks may be reasonably large and the lists
handled by the executive routine may be very short compared
with complex programs written almost completely in list-
processing language. It may be hoped that future hardware
developments will facilitate list operations, at least on a scale
adequate for operation of the executive routine.

Panel Discussion

Hopper. When we discuss computer programs, we must
realize that they exist at many levels. We must consider not
the single instructions of the computer in isolation but how
they come together to form a subroutine, how the subroutines
come together to form a run, and how the runs come together
to form a system. We must recognize that the system may not
be operated entirely on the computer, but may include manual
operations, punched-card operations, and reporting from various
instruments. When we discuss the processes with which we
must cope, we note that some processes are independent and
can be accomplished at the same time, such as inventory and
payroll, diagnostic testing and compiling, or data processing
and card-to-tape conversion; some processes are partially de-
pendent, such as payroll and labor distribution or project cost
distribution; and some processes are completely dependent,
such as sorting and matching, or the processes involved in de-
fense against missiles. Projecting into the future, computing
systems also fall into several classes. Some systems are multi-
programmed, with a single processor operating upon more than
one program and doing its own scheduling; some systems are
multiprocess, with several processing units sharing storage; and
some systems are multicomputer.

I always think of the steps of getting a problem onto the
computer as starting with a manager who needs intelligence.
By intelligence I mean organized information. He goes to an

271

G. M. Hopper

analyst. If the manager is an engineer, he may go to a numerical
analyst. If he is a businessman, he goes to a systems analyst.
Even in the armed services there are analysts, although I think
there they are called strategic planners. The manager says, "I
need some information about such and such in order to make
a decision. How do you get it for me?" There are two problems
at this level. One is that the manager really does not know
what he wants. The analyst must draw this out from the man-
ager somehow and then determine how to obtain the necessary
source information. The second problem is locating this source
information and finding if it can be obtained and recorded.
This reminds me of the Navy's attempt to learn why vacuum
tubes were failing. The Navy attacked the problem by distribut-
ing some little cards to be filled out and returned every time a
tube failed. Now the critical time for which this information
is most important is during maneuvers, since this resembles
wartime most closely. And of course this is the time when every-
body marks that one cross which says "dropped it." Thus all
tubes fail from being dropped during the critical period. Cer-
tainly better definitions and methods for obtaining source data
are needed.

After the systems analyst has produced his flow chart, a pro-
grammer appears in the scheme of things to make the run
charts and write the run specifications. His charts go to a coder
who writes machine instructions, and these machine instruc-
tions go to the operator who runs the computer. Work in auto-
matic coding has concentrated on getting rid of the coder. We
have not yet tackled the job of automatic programming. When
we do, we shall be tackling the job that Professor Brown has
described. One step further is automatic system design, and
this too may be attempted within the next ten years.

The question of parallelism carries certain other considera-
tions with it. At present we write a program, and all too often
we do not define the data upon which the program operates.
In my mind the major difference between scientific problems
and those of business data processing lies in the difference be-
tween the respective data. Mathematicians, engineers, statis-

272

A New Concept in Programming

ticians, and logicians operate on numerical or logical quantities.
Data processing operates on information: anything from a bit
to a name and address. The information may be in alpha-
betic, decimal, binary, or other form. The data description is
essential to the program and cannot be omitted. We must in-
clude in any language with which we hope to describe complex
data-processing situations the capability for describing data.
We must also include a mechanism for determining the prior-
ities to be applied to the data. These priorities are not fixed and
are indicated in many cases by the data.

Thus we must have a language and a structure that will take
care of the data descriptions and priorities, as well as the op-
erations we wish to perform. If we think seriously about these
problems, we find that we cannot work with procedures alone,
since they are sequential. We need to define the problem in-
stead of the procedures. The Language Structures Group of the
Codasyl* Committee has been studying the structure of lan-
guages that can be used to describe data-processing problems.
The Group started out by trying to design a language for stating
procedures, but soon discovered that what was really required
was a description of the data and a statement of the relation-
ships between the data sets. The Group has since begun writing
an algebra of processes, the background for a theory of data
processing.

Clearly, we must break away from the sequential and not
limit the computers. We must state definitions and provide for
priorities and descriptions of data. We must state relationships,
not procedures.

Sayre. Rather than comment on Dr. Brown's paper, which
treats a specific topic in considerable depth, I shall attempt to
discuss programming as a whole, especially the forces that seem
(to me at least) to be determining its development. During
the earliest phase of the history of programming, applications
were programmed for a particular machine starting from a very

* Codasyl is an acronym formed from the initial letters of the words
that spell out the committee's objective: a common data systems language.

273

D. Sayre

basic logical level, namely, that of the machine instruction. A
wholly new program generally was written for each significantly
different machine or application. Thus, human intelligence was
expended lavishly in order to achieve near-optimal utilization
of a given machine for a given application.

This state of affairs, besides being the natural way to begin,
was also entirely appropriate so long as human intelligence was
available for programming in adequate quantities and so long
as machine capabilities were only barely sufficient for the de-
sired applications. But a trend toward the exact opposite situa-
tion is now rapidly developing. It is becoming clear that the
limiting factor in the growth of the use of machines is be-
ginning to be mankind's ability to instruct machines in com-
plex activities and not the machines' raw information-process-
ing power. Programs are already being produced at the rate of
some 300,000,000 words per year. By comparison, the material
published for the people of the United States each year in book
form is less than ten times that amount.

These circumstances naturally produce an intense pressure
to economize on human intelligence in programming by de-
veloping machine intelligence as a substitute, so as to trade
what is plentiful for what is scarce. In my opinion, the chang-
ing structure of computer programs can best be understood in
terms of this force. In passing, we may note that similar pres-
sures exist in many areas besides programming — these are what
account for the acceptance of computing machines on such a
large scale — and that developments there stimulate (and are
stimulated by) developments in programming.

Some of the ways of saving programming effort require no
comment. One way is simply to spend less effort in working
toward optimal utilization of the hardware. This is a very
noticeable feature of present-day practical programming. An-
other class of methods centers on better communication among
programmers and the sharing of programs. And now a new and
far-reaching method is at hand: It requires the computer manu-
facturers to make their machines more alike and not to change
them, except of course for the purpose of making them faster

274

A New Concept in Programming

and cheaper. This is not as easy as it sounds. Machines cannot
be made faster and cheaper without introducing new hard-
ware devices and organizations, and without producing different
hardware configurations for different needs. More often than
not these variations change the whole logic of the machine.
Somehow this change in logic must be concealed.

Fortunately, a method for doing this has been found, and
certainly its discovery and exploitation are the significant events
of this stage in the history of programming. The method is to
give to the machines themselves the ability to determine how
their own hardware resources should be employed. Then the
programmer can ignore this matter, and at the same time, ma-
chines which differ in their hardware can be made to look alike.
This ability is carried in a large body of programming (auto-
matic coding systems, automatic operating systems, and so
forth), supplied by the computer manufacturer with the ma-
chine. Already this body of programming sometimes exceeds
100,000 words, particularly if it does a really good job of con-
trolling the machine's hardware. From another point of view,
it can be said that the manufacturer is beginning to supply a
new kind of machine, composed partly of hardware and partly
of software (the term that is beginning to be used for the body
of programming which we are discussing). Such composite
machines are rich enough in the control information they con-
tain to resemble one another closely even though they consist
of very different hardware arrangements and even though they
cover a wide range of prices and performances. What we have
here can be thought of as a clear instance of machine intelli-
gence being substituted for human intelligence in one part of
the programming problem.

Actually we software designers do not know at the present
time how to cover up all aspects of hardware in this fashion.
We can deal with the usual high-speed registers and storage of
a machine, with the interrupt system, and with certain aspects
of the assignment and overlay of input-output units. But we
are still in the course of learning how to deal in a general
fashion with auxiliary storage devices (such as tapes and discs)

275

D. Say re

and with multiprocess machines. Dr. Brown's paper is an ex-
cellent example of the research that is needed. In the mean-
time, the composite machines that we know how to make do
not yet reach the ideal of looking completely alike.

The development of the composite machine is bringing with
it important changes in the computer industry. The user can
switch comparatively freely from one composite machine to
another, and he can easily get a fairly accurate comparison of
their performances. Also, he can in concert with other users
participate more effectively in setting specifications for com-
posite machines than he could for purely hardware machines.
The computer manufacturer is not at all sure that he likes these
facts, but he is learning to live with them, and he knows that
any trouble they bring him is more than offset by their stimulus
to computing in general.

So much for the present and immediate future. What can we
expect to happen next? I should like to approach this question
by pointing out that the problem of economizing on the human
intelligence which goes into programming is really one of pro-
gramming the right things in the right order, of attempting at
each stage to program what will most simplify what has yet to
be done. (This is much like the principle that underlies the
formation of any elegant formal theory; postulates and the-
orems are produced in such an order as to minimize the total
argument necessary.) The matter we have just been discussing
is a good example of this. Because hardware varies so widely
and rapidly, it is desirable to neutralize the variation at a very
basic logical level, so that all that lies above need not be con-
stantly rewritten.

In my opinion, the next major phase in the development of
programming will be the extension of this principle into the
immense domain of applications programming, with the aim
of neutralizing, as far as possible, the variations in applications.
A continual attempt will be required to perceive and remove
what is repeated in the body of applications programming and
to reintroduce it into the structure at a lower level in more
general form. To a slight degree we do this already: The mathe-

276

A New Concept in Programming

matical subroutine is a good example. Moving downward in
this way, we shall form successive levels of programming less
and less specialized to individual applications, and more and
more general in nature, gradually approaching the quality of
basic statements of the laws and facts of our world. But this
trend toward generality necessarily supposes the possession of
more and more general techniques of reasoning. These tech-
niques include the ability to deal with sentences in suitably
rich languages (richer than computers can handle at present),
the ability to devise classifications under which information
may usefully be filed for later retrieval, and the ability to per-
form inference of various kinds.

To program these general techniques will require an edifice
containing millions and eventually billions of words. The edi-
fice will doubtless need continual improvement as well as occa-
sional tearing down and rebuilding as our understanding of the
principles that underlie its construction grows. But because it
will be built on composite machines rather than on hardware,
it ideally will be immune to changes in hardware. It will be the
role of the software designer to keep it so. This edifice, and
the intelligent machine activity that it permits, will naturally
have costs and inefficiencies of utilization attached. An impor-
tant aspect of the use of machines in the future, as now, will
be to use only as much of what is available as is required by
the application at hand. But the edifice will provide a high,
broad, versatile foundation of abilities, and each new applica-
tion will need to be built up only from that point. At present,
each application must be built up from the bottom at an ex-
penditure of human intelligence which we cannot afford.

To me programming is more than an important practical art.
It is also a gigantic undertaking in the foundations of knowl-
edge, and it is certain to add immensely to our understanding
of what we know and how we know it.

Brown. I should like to make a few closing remarks and
perhaps take mild issue with one implication that has been
made. I think the lack of predictability that Dr. Hopper men-

277

G. W. Broun

tioned and the difficulty of problem definition that both dis-
cussants stressed are adequate reasons for adopting the ap-
proach that I presented. I agree completely with the character-
ization of the problem as being one of dealing with diverse data
descriptions. I think that is at the heart of the language prob-
lem.

I was interested and somewhat stunned to learn that there
are 300,000,000 words of programming being written annually.
I think the very existence of a programming bottleneck indi-
cates that we had better not have programmers spending their
time making decisions that do not matter for the logic. I should
feel differently about it if I thought that all of the programmers'
efforts which are supposedly expended on optimal machine per-
formance really give optimal machine performance. Optimal
machine performance in the small is very different from optimal
machine performance over-all. I should agree that orientation
toward machine utilization in the small is the improper place
to put stress, but I should argue that in the large there is a
great deal to gain. There are many places where we could gain
factors of 2 or 3 more. I shall cite again a dissertation by Pro-
fessor McKenney of the Harvard Business School.3 This disser-
tation contains a simulation of an executive program of the
type that I have discussed, and indicates that there really are
factors of several lying around waiting to be picked up.

While I agree that the programming bottleneck is the thing
that is really holding us back and that, in a sense, machines
are cheap, they are not cheap for everybody. Ever}- university-
machine that I have ever seen gets filled up right away. The
things that are not getting done are suffering by virtue of the
time we are using on the things that are getting done. Further-
more, I think factors of several are not to be sneezed at when
major machine systems, in order to meet the requirements of
university research, now begin to cost S10 million and more.

I should like to plead that efforts such as that of the Codas yl
Committee indeed free themselves of dependence on statement
sequence because I think that this is the easy part. I think that
the difficult part is in the data description and specification re-

278

A New Concept in Programming

lations. There is nothing to gain by retaining the dependence
upon statement sequence which all of the language efforts still
seem to have, at least in residual form. But there is a great deal
to gain by deliberately abandoning sequential dependence. All
I am saying is that a relationship should have the same impli-
cation, no matter where it sits in a list. Its implication should
not be modified by what sits elsewhere on that list, provided
the data descriptions are appropriate for the relationship as it
sits.

My final observation is that the approach which I have de-
scribed also happens to be applicable to the execution of the
executive program itself. An analysis of examples has indicated
that the executive program has the same kind of possibilities
inherent in it that the standard programs have. This is an
amusing thing to speculate about.

General Discussion

McKenney. I am curious if Dr. Sayre's preoccupation with
software is not due in part to his fascination with manipulating
machines and seeing how much more he can get out of them.
It would seem preferable to study how to organize machines so
that it would not be necessary to spend so much time worrying
about how to manipulate them. There appears to me to be a
tendency among machine manufacturers to see how much
better machines can work, rather than how much better they
can be organized.

Sayre. Your statement is a little bit like saying that you
are in favor of motherhood. Of course work on rationalizing
the structure of machines is important. Nevertheless, I think
that performance is the most desired and desirable of all the
qualities that a machine can have. In software we have a tech-
nique for untying the hands of the hardware designer so that
he can get performance without forcing a complicated machine
upon the user. Let me give you an example. One way to make
machines easy to use is to make all of their storage random

279

General Discussion

access and of the same speed. But this kind of storage is ex-
pensive and impractical in comparison with other kinds of
storage. The software that I have described can make a ma-
chine which is composed of heterogeneous storage look to the
user as if it were homogeneous.

Grisoff. I should like to ask Professor Brown two ques-
tions, if I may. First, don't you think that with just about all
of today's methods of programming the programmer does not
really have to make any extra effort in setting down the arbitrary
ordering of his program steps? Yet wouldn't it require consider-
able additional effort to provide the information necessary to re-
move the nonessential ordering steps from the program? To my
way of thinking this would require a very different way of pro-
gramming from what we know or are even considering today.
Second, you mentioned that it was not worth meshing two
programs, or possibly more, without any prediction of their
course on the machine. Don't you think in general that two
or more programs running concurrently could make use of idle
facilities on a queuing basis, with an almost certain saving in
time? The worst you could possibly do would be the sum of
the sequential running times; but the best you could do might
be considerably better.

Brown. If I may say so, I think that if I have not gotten
my point across any better than this to the rest of the audience,
I have more or less failed in my presentation. In answer to the
first question, I believe that what the programmer does now is
work, and I believe programmers make mistakes doing things
that are of no consequence. I am not suggesting a different way
of programming. I would not make programming harder: I
would make it easier. It is true, however, that anybody who has
learned to program a machine in some highly stylized manner
will naturally fight anything that comes along and looks dif-
ferent.

In answer to the second question, queuing will provide tre-
mendous efficiency with the parallel use and time sharing of
units, but not for monolithic programs. Let me give you per-
haps the classic example of this. The reason you cannot nor-

280

A New Concept in Programming

mally translate programs from one machine to another con-
veniently is the fact that even if you try to translate instruction
by instruction you soon discover that there are instructions
which the execution manufactures. You cannot translate in
advance what you do not know in advance. This is the reason
why most translation is done by simulators.

People fought such things as Fortran. They claimed that
Fortran would never produce programs that had any kind of
reasonable running time. The fact that Fortran compilation
is now what takes all the time may only be a criticism of the
professionals who wrote the compilers while seeking maximum
machine utilization. Fortran programs after being compiled
'run like the devil." They are faster than most programs that
people write in an attempt to avoid using Fortran, and all our
troubles are in the compilation written by those very profes-
sional programmers. If you stand and watch the Fortran sys-
tem compiling (I am sorry to do this to IBM; I could do the
same thing to Remington Rand), you see a roomful of equip-
ment sitting idle while one lousy tape rewinds. We could do
much better by letting the machines schedule themselves. I
submit that we cannot do worse.

Garman. Dr. Brown, would you have these programs com-
piled and assembled ever}' time they were rerun?

Brown. No, that is not necessary at all. My executive pro-
gram is an interpreter, operating at a sufficiently high level so
that interpretation as a standard mode is fast and not waste-
ful. You know that if you tried to simulate one machine on
another today — generally you simulate a slow machine on a
fast machine — you are very lucky if the simulator runs as fast
as the slow machine ran in the first place. This is because in-
struction-by-instruction interpretation is terribly slow. I am
not suggesting that atomic processes be at the level of single
machine instructions at all. But you do not have to get very
far away from that before interpretation is a very efficient
way to carry things out. What this means is that you are con-
stantly in what is called a monitor state. This has tremendous
advantages for program debugging, for understanding what is

281

General Discussion

going on, and for keeping things rather close to the source
language in which you kept track of your problem originally.
So, no, I do not see this as a difficulty. I think we have an
advantage here.

Abrahams. I should like to bring up the current diversity
of languages with which we are faced. It may be a symptom of
a problem or it may be actually part of the problem itself. I
noticed on the cover of a recent issue of the Communications
of the Association for Computing Machinery a very nice draw-
ing of a tower, representing the Tower of Babel. Each brick in
the tower was a different programming language, and an amaz-
ing number of bricks were filled in. I wonder whether the ability
to absorb programs written in currently available languages
may be a very important feature of new programs, and new
programming languages, especially considering the very large
investment that we have in fixed programs and the difficulties
in transferring them from one language to another and one
machine to another.

Brown. The only thing I shall say is that if we do not do
something soon about accepting a reasonably standard image
of what gross logic in computers looks like, then this problem
will just get worse and worse. We already have seen situations
where manufacturers are unable to free themselves of a program
structure because of their great investment in programming,
and the longer this goes on, the bigger the investment gets.
Even if our programs did not have complete translatability at
the lowest levels, even if only the major structure had this trans-
latable property, then at least we would only have to worry
about the atomic processes and relating them. So we should
have that much improvement. As to the further step of ultimate
universality of language, I am not an expert on this, but it is
perfectly clear that the great impediments to standardization
are the externals: the differences in tape structures, printer
structures, and even cards. Dr. Hopper probably can say ever
so much more than I on this subject.

Hopper. The sooner we stop describing data on cards or
data on tape, and just describe data, the better. Unfortunately

282

A New Concept in Programming

we have programmers who like to make data designs, even
though it is clear that the computer is more capable of creating
good data designs than are programmers. The computer does
not mind in the least breaking up a man's name and address
to use as -stuffing around quantities, a very convenient device
which makes for efficient programs. A programmer would not
do this because then he could not "look through" the tape and
read off the names and addresses.

We should not only take away the programmers freedom to
allocate storage and time, we should also take away his freedom
to arrange characters on cards and tapes. The programmer
exists in a world of human prejudice. When he looks through
a file, he thinks that he must look alphabetically by name and
address. It disturbs him greatly to have a name scattered over
eight or nine or ten words in various positions. But the fact
that he can never read a tape does not seem to upset him.
Programmers are a very curious group. They arose very quickly,
became a profession very rapidly, and were all too soon in-
fected with a certain amount of resistance to change. The very
programmers whom I have heard almost castigate a customer
because he would not change his system of doing business are
the same people who at times walk into my office and say,
"But we have always done it this way/' It is for this reason
that I now have a counterclockwise clock hanging in my office.

McCarthy. I should like to make two comments. The first
one concerns the use of devices for eliminating unnecessary
specification of sequence in programming. Professor Brown hit
the nail on the head there. I devised conditional expressions
precisely for this purpose. My motivation was not so much to
achieve scheduling efficiency, but was simply that in a mathe-
matical sense it seemed inappropriate to specify more than was
necessary. Unfortunately, even with such devices there is still
a great deal of unnecessary specification of sequence. For ex-
ample, commutative operations such as addition and multi-
plication allow us freedom to choose the order in which the
operations are performed. But it is quite difficult to leave this
up to the compiler. The freedom can be pushed only so far,

283

General Discussion

and eliminating all unnecessary specification of sequence turns
out to be impossible.

My second comment is an expression of accord with the
remarks made about the importance of data description. The
main gap one finds in examining this Tower of Babel is the
relatively few control structures in all these languages. I agree
with Professor Brown that a universal control structure is pos-
sible, but because no ways of describing new data structures
exist as yet (except in rather primitive form in some of the busi-
ness data-processing languages), the main distinction between
the various programming languages is still in the data struc-
tures that they allow. Now it seems to me that the tools for
describing data structures have been with us all the time. They
are the mathematical operations of Cartesian product of sets
and union of sets, which are really fairly simple mathematically.
It seems to me that with the aid of these tools a universal means
of describing data could be created, and we could clear up this
jungle of programming languages.

Brown. I certainly agree that you cannot carry sequence
freeness all the way down. This is another reason for organizing
programs on two levels. Atomic processes may or may not have
some sequence freedoms of their own, but it is the control
language which definitely should be free of sequence specifica-
tion; and it can be.

With respect to the data descriptions, none of the so-called
universal languages around has taken the step of adopting a
control language like the one I described. I think this is a sad
mistake. Indeed this is not the hard part of the problem, and
yet it is the part that has been ducked. I see no reason why the
macrolevel should have a control language that suffers from the
defects of the microlevel. This is the situation today.

Hopper. I should like to add one thing to that, however.
I think it is awfully nice that mathematicians, logicians, en-
gineers, and statisticians have exact languages in which they
converse with each other. But one of the difficulties with pro-
grammers is that the symbolisms that they have invented leave
management and systems analysts ignorant of what is going on.

284

A New Concept in Programming

When it recently became possible to use the English language
to write programs, an Air Force colonel was heard to say, "Now
we can take back command of the Air Force from those
damned programmers."

There is a very great danger here. I think we mathematicians
have to use and construct exact languages. However, I must
meet a data-processing world that has not been trained in
Boolean algebra. If I said to a certain salesman, "But the first
thing I have to do is evaluate this truth function," he would
probably say, "Well, I sold the darn thing, didn't I?" We must
not lose sight of the users, and we must not forget that a part
of the reason for these languages has been to re-establish com-
munication between managers, programmers, computers, and all
of the surrounding people who until now have been cut off
from the mathematicians.

Brown. I have no objection whatsoever to your writing, "If
so and so, then so and so, otherwise, so and so," in the data
description. I find this perfectly acceptable and it meets my
requirements.

Hughes. I want to comment on the reference to the English
language. My programmers are using an English language com-
piler to write their programs, but looking over their work I
still don't know what they are doing. Because English is cum-
bersome and highly redundant, they end up with a set of ab-
breviations which I cannot understand. I have all the
disadvantages of the English language (namely, I have to write
out long sentences to express short ideas), and I also have the
problems of computers. That is, the slightest variation from the
prescribed English language form causes the computer indi-
gestion. The redundancy does me no good and actually does
me great harm. I think we have to find a language acceptable
to the people analyzing business processes which can be trans-
lated by the computer into a language that management can
understand.

Hopper. I have to answer that because we ran into it too,
particularly the tendency of programmers to abbreviate, mis-
spell, and forget periods at the ends of sentences. The first

285

General Discussion

thing we did was to invent a young lady called a "pseudocoder"
and put her in charge of all abbreviations. No one was allowed
to use an abbreviation unless it was registered with her. The
second thing we did (having discovered that programmers could
not write in English after we carefully had made programs so
that they could write in English) was to isolate the programmers
and have them write English language flow charts and nothing
beyond. Then the pseudocoders typed English sentences from
the flow charts, always using the same abbreviations in the same
program, always putting periods on the ends of sentences, and
always putting in all the redundancies necessary to make sense.
They know that they are the ones who are really programming
the computer.

Brunow. Must not the speed, the storage capacity, and the
input-output capabilities of the computer undergo a sharp up-
ward change before we can achieve the desirable powers of
software spoken of tonight?

Sayre. In this business of software we are dealing with the
fundamental economic question of the relative cost of human
intelligence versus machine intelligence. You cannot go very
far toward machine intelligence until you have lowered its
relative cost below that of human intelligence. At the present
time, machine intelligence is still very costly and can be effi-
ciently and economically employed only in the most favorable
circumstances. In answer to your question, yes, we have to
bring the cost of machine intelligence down. On the other
hand, I think it is evident that it is going to come down. I
think we shall have a hard time developing methods for real-
izing machine intelligence as fast as its decreasing cost will
warrant.

Hopper. Up until a year ago software was unrecognized,
applications were unrecognized, and programmers were pretty
well unrecognized by engineers. When we programmers ob-
jected to something, it usually led to our stalking out of the
room saying, "O.K., build it anyway you like and we'll program
around it." Only within the last few years have programming,
software, and applications been given equal status with en-

286

A New Concept in Programming

gineering. I think there is more to come. I think we have only
started on this path.

Morse. George [Brown], do you have any final words?

Brown. Only the reminiscence that physical scientists once
upon a time used to accuse social scientists of the error of
thinking that they understood something once they had given it
a name. I hope we do not think, because people now have the
word "software," that we fully understand the problems and
that they are solved. They are not.

References

1. Nekora, M. R., "Comment on a Paper on Parallel Processing/' Com-
munications of the Assoc. for Computing Machinery, Vol. 4, 60 (Feb.,
1961).

2. Ryle, B. L., "Multiple Programming Data Processing," Communica-
tions of the Assoc. for Computing Machinery, Vol. 4, 99-101 (Feb.,
1961).

3. McKenney, J. L., Simultaneous Multi-Programming of Electronic Com-
puters, Research Report No. 69, Management Sciences Research Proj-
ect, U.C.L.A., Los Angeles, Calif. (Feb., 1961).

4. McCarthy, J., "Recursive Functions of Symbolic Expressions and Their
Computation by Machine, Part I," Communications of the Assoc. for
Computing Machinery, Vol. 3, 184-195 (Apr., 1960).

5. Newell, A., and Tonge, F. M., "An Introduction to Information Proc-
essing Language V," Communications of the Assoc. for Computing Ma-
chinery,Vol. 3, 205-211 (Apr., 1960).

6. Evans, A., Jr., Perlis, A. J., and Van Zoeren, H., "The Use of Threaded
Lists in Constructing a Combined ALGOL and Machine-Like As-
sembly Processor," Communications of the Assoc. for Computing Ma-
chinery, Vol. 4, 36-41 (Jan., 1961).

7. Irons, E. T., and Feurzeig, W., "Comments on the Implementation
of Recursive Procedures and Blocks in ALGOL 60," Communications
of the Assoc. for Computing Machinery, 65-69 (Jan., 1961 ) .

287

8

What
Computers

Should
Be Doing

8

What
Computers

Should
Be Doing

Speaker John R. Pierce

Executive Director of Research
Communications Principles Division
Bell Telephone Laboratories, Inc.

Discussants Claude E. Shannon

Donner Professor of Science
Massachusetts Institute of Technology
Walter A. Rosenblith
Professor of Communications Biophysics
Massachusetts Institute of Technology

Moderator Vannevar Bush

Honorary Chairman of the Corporation
Massachusetts Institute of Technology

290

Pierce. "WTien I was asked to deliver a lecture in this series,
no one expected me to talk about computers. I am not an ex-
pert in the field, and I certainly could not hope to present the
sort of inside story of computer design, programming, and simu-
lation that the other lecturers have provided.

My lack of familiarity with the details of computers is not
going to deter me, however, for I have numerous examples to
encourage me. A number of years ago a publisher who knew a
great deal less than I do suggested that I write a book about
the cybernetic revolution and about machines that outthink
man. He outlined, almost chapter by chapter, the fantasy for
which I was to supply words and circumstantial detail. WTien
I ventured to say that I did not believe that all of those things
were true, the publisher misunderstood me completely and said,
"You don't have to be for it, you can be against it."

291

/. R. Pierce

Since that day various people have written the book the
publisher was seeking, and I have gone on to encounter com-
puters at shorter range. A couple of years ago I suddenly found
the Bell Laboratories' mathematical research department, to-
gether with its IBM 704 computer, in my department. The
704 has now been somewhat painfully exchanged for a 7090. I
have observed programmers struggling with problems that came
to them from AT&T and Western Electric as well as with re-
search problems, including theorem proving. Somewhat earlier,
I had watched acoustic and visual researchers in my department
use the computer to simulate the performance of electrical cir-
cuits and devices, and to process, analyze, and recognize acous-
tic and visual data. More recently, as head of the library com-
mittee, I have watched librarians' efforts to grapple with the
problems of information retrieval.

It has occurred to me that this casual association with com-
puters and my speculations concerning their use may have
involved a number of matters important to management and
the computer of the future. If I do not understand the com-
puter in fine detail, neither will many of the managers who will
make use of the computers of tomorrow. For that matter, I am
neither a physiologist nor a card-carrying psychologist. I have
more ignorance than knowledge about the functioning of the
human beings with whom I associate every day, but somehow
I manage to get along with them, by and large.

Managers will have to get along with computers and pro-
grammers on these same terms. But, just as some sense of the
abilities and fallibilities of human beings is necessary if one is
to work with them, so some knowledge of the abilities and falli-
bilities of computers and programs is necessary if one is to get
along with computers. It is obvious that one should not treat
men like machines, and I hope that it is equally obvious that
one should not treat machines like men.

I think that the first thing one needs concerning computers
is some sense of their powers and limitations. The research of
today is embodied in the operations of tomorrow, or at least,
the successful research of today is. It seems to me that a good

292

What Computers Should Be Doing

way to discuss management and the computer of the future is
to discuss research and the computer of today, and I propose to
review a few triumphs and flubs of today's computer research.

Mechanical Translation versus Satellite Tracking

I was talking to an Air Force officer recently about the trans-
lation of Russian into English. I said that it seemed urgent to
teach more people Russian. He said that this takes too long
and the Air Force cannot wait. Instead, a large number of
projects on machine translation are being supported. With the
aid of an excellent book from the National Science Foundation
on current Research and Development in Scientific Docu-
mentation, I counted up twenty projects on mechanical trans-
lation, of which six are supported by the Air Force, four by the
National Science Foundation, two by the Army, and one by the
Navy.

Machine translation has already taught us a great deal. It
has taught us how little we actually know about the mechanics
of language and translation and how hard the problem is. It
has taught us the fallibility of a few appealing ideas. The study
of language and translation with the aid of computers promises
to be an active, worthwhile, and fruitful field of research for
many years.

Where do we stand as far as actual translation goes? Word-
for-word translations, or a little better, have been produced at
a number of places. A Special Investigating Committee of the
U.S. House of Representatives Committee on Science and As-
tronautics had an eminent chemist evaluate a translation of a
paper in the field of chemistry. The chemist reported that he
found the translation intelligible, but that he spent about four
times as long in reading it as he would have in reading a version
translated into English by an expert human being.

Machine translations are not easily read. Unexpected ma-
chine behavior sometimes creates confusing artifacts. It is clear
that if much Russian is to be read, the reader will save time in
the long run by learning Russian. But, if the problem is not
that of translating and reading much Russian, why the mass-

293

J. R. Pierce

production implication of the machine? I am afraid that gam-
bling on machine translation rather than teaching Russian can
only have the effect of hindering access to Russian sources for
an indefinite, unforeseeable period. This does not mean that
work on machine translation should not be carried out, but
such work should not be counted on as providing a solution for
a real and pressing problem.

Let us consider another matter. I have been interested in
satellite communication for a number of years. I used to hear
assertions that the only practical satellites are those that revolve
22,300 miles high above the equator and hang stationary in the
sky. The problem of tracking lower-orbiting satellites that rise
and set was proposed as very difficult and expensive if not in-
superable. I knew that quite large steerable antennas could be
purchased for much less than the cost of launching a satellite,
and there are lots of fine computers at large, so I was skeptical
of this. During the Echo experiment data from the Minitrack
network were sent to the Goddard Space Flight Center. There
the orbit was computed, and pointing data were computed for
the JPL Goldstone antennas and the BTL Holmdel antennas
for every 4 seconds of each pass of the satellite. These data were
received at Goldstone and Holmdel on teletypewriter tape. A
digital-analog converter read the taped data at appropriate times
and controlled the servos steering the antennas.

While the system suffered from some errors in data trans-
mission, and occasionally from stale orbital data, its operation
left no doubt that satellites can be tracked successfully in this
manner. The computations involved in the satellite tracking
can be specified explicitly. The auxiliary equipment that is
needed can be built using known digital arts.

Without the use of a computer, satellite tracking of the sort
I have described would be impossible. Translation is not diffi-
cult for a skilled linguist, yet translation of an adequate quality
is at present impossible for a computer, and we have no reason
to believe that human translation will be substantially replaced
by machine translation in the near future.

It is perhaps instructive to compare the problems of satellite

294

What Computers Should Be Doing

tracking and translation a little further. In satellite tracking, the
speed of the computer is absolutely essential. Without this
speed the refinement of orbital data and the conversion of co-
ordinates could not be completed in time. Suppose, however,
that a machine translated a large bulk of Russian text instan-
taneously. Human beings would still have to pore over the
translated text in order to make use of it. A reduction of time
of translation to zero might not appreciably reduce the time of
utilization.

Some may object that the machine itself could carry out
certain operations, such as searching for key words or combi-
nations of key words, or indexing, or even abstracting. This is
certainly true. The machine can, however, carry out such opera-
tions in a foreign language as well as in English. They are es-
sentially information-retrieval, not translation, functions.

Theorem Proving, Game Playing, and Pattern Recognition

We see that computers are not inherently good or bad; they
are good for some problems and bad for others. One can go
further, and say that whether computers are good or bad, that
is, successful or unsuccessful, depends on whether they are used
well or poorly.

One of the show problems for electronic computers is the
proving of theorems. Several people have made computers prove
theorems. Gelernter at IBM has made a computer prove ele-
mentary theorems in Euclidean plane geometry. Newell, Shaw,
and Simon have made a computer prove simple theorems in
mathematical logic. So has Wang. As far as results go, Wang's
program has clearly done best. It has, in fact, enabled the IBM
704 computer to prove all the (approximately 350) theorems
in the first thirteen chapters of Principia Mathematica, by
Russell and Whitehead, in 8.4 minutes.

Now the method of proof used by Wang differs markedly
from those of the others. Wang takes great care to frame his
problems in a language suitable for machine use, and he pro-
vides a foolproof, mechanical method of finding a proof.
Newell, Shaw, and Simon, and Gelernter, on the other hand,

295

/. R. Pierce

try to ape human behavior. Particularly, Newell has observed
how mathematically unsophisticated subjects go about proving
theorems, and has tried to make the computer do likewise.
''Heuristic" devices are sought, which will lead the computer to
promising but not infallible courses of action. So far, such
heuristic devices have not been shown to be very helpful.

Suppose that you are interested in human behavior and in
laying a groundwork for machine behavior that is in some way
analogous to human behavior. You will find the work of Gelern-
ter and of Newell, Shaw, and Simon fascinating, while that of
Wang may seem simple, elegant, mathematically ingenious, but
cold and unappealing. The simulation of human behavior
would call for a procedure very different from Wang's. Perhaps,
as Minsky suggests, a program to solve real mathematical prob-
lems will have to combine with the mathematical sophistication
of Wang the heuristic sophistication of Newell, Shaw, and
Simon.

But suppose one's objectives are different. Suppose one wants
to use computers to do something they can do very much better
than human beings can. Wang has shown that computers can
outdo human beings at one rather complicated task. Newell,
Shaw, and Simon have shown that computers can imitate hu-
man behavior imperfectly and with only moderate effectiveness
at an equivalent task.

Here we see quite clearly a divergence of views about com-
puters which permeates much work, thought, and writing. The
playing of games by machine is a prime example. Computers
play ticktacktoe and nim as well as these games can be played,
and unerringly. It would be simple to invent games at which a
computer would quickly and unerringly beat any human being.
One would merely have to make success dependent on compli-
cated calculations, infallible memory, the simultaneous appli-
cation of many invariable criteria.

Checkers and chess are different. These games were devised
with human beings, not computers, in mind. It is a tribute to
programming skill that computers have played chess poorly and
checkers well, though not as well as a very good checker player.

296

What Computers Should Be Doing

Samuel's checker-playing program even includes learning by
experience. Right now, however, it would be cheaper to hire a
checker player or a chess player than to seek equivalent per-
formance from a computer. To make a computer excel a chess
champion is a noble intellectual challenge, but it is not a
paying proposition. Further, it is not clear that such game play-
ing is the road to progress in the improvement of computers
and programs, though of course it may be.

Next to game playing, a favorite pastime of computer experts
is character recognition and the recognition of spoken words.
No one can argue that computers cannot read words; they read
words off punched cards, perforated tape, and magnetic tape
even- millisecond (or less) . Self ridge has even made a computer
read hand-sent Morse Code. This is no small feat, for in hand-
sent Morse some dots are longer than some dashes.

What computers do not do very well, despite the efforts of
Selfridge and others, is to read varieties of type faces in various
orientations, or to read handwritten script, or to recognize many
words when spoken in a variety of voices. Several reasons for
this fallibility are clear. In the case of speech, for instance, com-
puters can find only slowly and fallibly features that are clearly
present to the eye in a speech spectrogram. It is difficult to make
computers recognize classes of features that the eye or ear picks
out unerringly. Beyond this, we obviously make use of context
in puzzling out the meaning of script, and our ignorance of the
nature of language makes it hard for us to tell a computer how
to do this. Let it be said, in addition to this, that there is no
guarantee that any amount of research will give us satisfactory
recognition equipment in any specified time. Recognition, like
translation, is a fruitful field for research rather than a resource
to be counted on.

In contrast, my bank's accounting machinery has no trouble
with the numbers printed in magnetic ink on my checks, though
to me they are strange and obscure. Further, we should con-
template the fact that the overwhelming bulk of the world's
records are as yet unrecorded, that most of the world's books
are as yet unwritten. I am sure that beyond some point in the

297

]. R. Pierce

future, material to be processed by machine will be set down in
machine-readable language, and that this will include almost
even-thing that is set down. Someday, the majority of all re-
corded material will be in machine-readable form. Unhappily,
today money will not buy a device for producing a machine-
readable record that compares in convenience and speed with
a good typewriter. I wash that some of the ingenuity that is
spent on recognition would go into producing such a machine,
and a cheap one as well.

If a machine is to be forced to read records, and at a rapid
rate, the task should be made easy for the machine. The char-
acters should be unambiguous, and the machine should be al-
lowed to run through tape or sort through cards; it should not
have to thumb through dog-eared pages. On the other hand,
the user is entitled to some consideration. He should find it
simple to give orders to a machine, and he should receive in-
formation from the machine in an easily intelligible form.
Special-purpose computers have been designed with this in
mind. The telephone dial does not tax the user's ingenuity
severely, and yet it controls what is certainly the most compli-
cated (though not the most versatile) computer system in the
world. The reservation machines in airline offices are admirably
easy to use.

General-Purpose Computers, Special-Purpose Computers, and
Compilers

When we think of computers, however, we usually think of
huge general-purpose computers. These are given their orders
in a language composed of hundreds of different instructions by
a special caste called programmers. Very astute programmers
construct special, sophisticated languages to help in program-
ming computers to undertake recondite tasks such as chess
playing and theorem proving. Clever programmers have also
constructed extremely simple languages by means of which
people who know nothing about computers or about program-
ming in general can cause computers to solve certain limited
classes of problems. The "program" that the user writes can be

298

What Computers Should Be Doing

as simple as a line-per-block listing of the blocks, parameters,
and interconnections of the block diagram of a circuit whose
performance is to be simulated.

By means of a compiler, the computer itself translates un-
ambiguous instructions in such a simplified user language into
a long, detailed program in its own language. Successful com-
pilers have been made for a large variety of special purposes,
ranging from exploratory research to the carrying out of mun-
dane chores such as minimizing the length of connecting wires
by locating subassemblies most advantageously on a chassis.

Part of the gain achieved through a compiler is in providing
a language that accords with ingrained habits of human thought
and speech, including mathematical habits. An essential part
of the gain achieved through the use of a compiler is that of
using the computer itself to take the step between the setting
of a specified task and the writing of a long list of instructions
for the carrying out of the necessary steps in the required cal-
culation. In order to gain any advantage, however, it is
necessary that the compiler be in some sense specialized. A
compiler allows us to reduce the burden of programming a
class of problems more restricted than everything, senseless and
sensible, that the computer could be made to do. Indeed, com-
pilers most reduce the programming burden when they are de-
signed for a very restricted range of problems; when, in effect,
they enable us to use a large general-purpose computer as if it
were a highly specialized computer.

Why not use a highly specialized computer instead? A gen-
eral-purpose computer cannot be as efficient as a special-purpose
machine. If many machines are required to do highly similar
work continuously in many places for many years, it will prob-
ably be advantageous to design a special-purpose computer for
the function. Indeed, as the computer art matures, a large
variety of computers of various sizes and capabilities have come
on the market, and these are certainly not interchangeable.
Too, it has sometimes proved extremely time consuming and
costly to program general-purpose computers to carry out special
functions. The SAGE system provides a notable example of

299

/. R. Pierce

this. The mere fact that a general-purpose computer can do any-
thing is no universal advantage.

However, many specialized jobs can best be explored (tried
out first) on a general-purpose computer. Many jobs take so
little computer time that they do not justify designing a special-
purpose computer. Many are not done in enough different
places, or over a long enough period. Further, the design and
production of a new computer and the distribution and servic-
ing of the new machine are not to be undertaken lightly. New
machines have a sad historv of fallibility, even those produced
by competent manufacturers. The building of special digital
equipment of any great complexity is something that appeals to
novices but sends shivers down the spines of those who have
experienced its tribulations. Debugging a new program is some-
times difficult but debugging a new machine is far. far worse.
A new computer design is justified only if widespread usage and
very substantial gains are assured.

Some Assertions

Let me sum up my background summary with a few asser-
tions.

The fact that a general-purpose computer can do almost any-
thing does not mean that computers do all things equally well.
Some things they do much better than human beings: some
things they do worse. Machines are not people. While it is
highly desirable to strengthen their weakness, the greatest im-
mediate gains, and some long-range ones as well, will come from
exploiting their strengths.

Partly, the strengths of computers are associated with cer-
tain types of problems, problems necessarily involving a great
deal of straightforward computing on a great deal of input
data, as opposed to game-playing or recognition problems.
Partly, however, the strengths may be associated with ingenious
preparation of the problem, ingenious methods of solution, and
ingenious programming, as in Wang's theorem proving.

This emphasizes the obvious fact that while machines are not
men, thev are useless without the aid of skilled programmers.

300

What Computers Should Be Doing

A skilled and ingenious programmer can do quickly, accurately,
and with a minimum of running time a job that would be done
slowly and with much computation, or perhaps not at all, by
a less skilled man. Further, a skilled programmer can turn a
general-purpose machine into a variety of special-purpose ma-
chines for the use of less skilled programmers, or even for the
use of men with no programming skill at all. Thus, he can
circumvent the the novice's love for special machines.

The Future

With these observations as background, let us ask what sort
of things computers are likely to do for management in the
future, and how can management go about reaping any benefits
there are to be gained.

I think that one thing should not be overlooked. Next to
having a good idea oneself, the best thing to do is to copy some-
one else's new idea — providing that it is a good one. Here we
should of course take into account a common sort of guile. The
King and the Duke in Huckleberry Finn persuaded the audi-
ences of the Royal Nonesuch to bamboozle their friends into
being swindled. Companies afflicted with architectural mon-
strosities recommend the architect to their rivals, and I suppose
that this attitude might extend to the victims of computers as
well. Nonetheless, we should try to learn all we can from the
successful experience of others. It really is the cheapest way.

The use of computers is advancing so rapidly, however, that
a certain amount of novelty (yes, let us admit it) of research
is a necessary adjunct even to routine operation. Anyone who
wants to use computers effectively will have to be willing to try
things that are at least a little different from what has been
done before. What sort of things, and what sort of new things,
should management turn to the computer for? One obvious
area is that of information handling, analysis, and retrieval.

In the simplest case, if information of any sort, whether it be
accounting data or the answers to questionnaires, is in machine-
readable form, it can in principle be rapidly presented in any
form, averaged, correlated, or abstracted for a variety of pur-

301

J. R. Pierce

poses. The manager can in principle have up-to-the-minute
data to base his decisions on. Or, if he does not like the
presentation of data, he can have it reworked in another format
or summarized in a different way painlessly by the computer.

We must always remember, of course, that the results will be
no better than the data. Further, suppose one were inclined to
act more promptly on fresh data than on stale data. If the fresh
data were in any way misleading, or if the consequences of ac-
tion were not fully understood, such action could conceivably
have violent and useless effects on the operation of an otherwise
stable business.

Questionnaires are even more suspect than data. It is much
easier to accumulate the answers to questions than it is to in-
terpret the meaning of the answers. To cite an extremely simple
instance, a more-contented-than-average attitude toward the
management of a particular office may reflect the fact that the
office is in a small community rather than any aspect or atti-
tude of management. After drastic simplification of data by a
computer, such a fact may become completely hidden.

This all makes it clear that, while computers can be put to
good use in information processing, it is easy to put them to
bad use. The computer will not assure that its summaries of
operating data are trustworthy and useful. Some very smart
person must take responsibility- for this. Surveys must be de-
vised by competent social scientists and the results combed over
by competent statisticians if any useful data are to be extracted
from them by means of a computer. A computer cannot turn
bad data into good data.

Beyond the processing of information, information retrieval
is often proposed as a computer function. The first problem one
encounters is that information cannot be handled by a com-
puter unless it is in machine-readable form. If there were any
convenient way of putting information into machine-readable
form, it might not be too early to start putting reports and even
correspondence into such a form on a routine basis. However,
we do not have such means. Even when we do have such means,
many seemingly niggling little problems of transliteration of

302

What Computers Should Be Doing

foreign letters and mathematical and chemical formulas will
have to be worked out.

In the immediate past, information-retrieval systems have
foundered because of the cost of putting past records into
machine-readable form. This indicates that we should think first
of processing moderate amounts of information, and work to-
ward more grandiose plans as we find ways of obtaining more
machine-readable material.

Indexing may seem a trivial start, but need not be so. Titles
are brief, and it takes no skill to identify them. An up-to-date
permutation index can call to one's attention many or most
reports or papers bearing on a matter of particular current in-
terest. Beyond the step of indexing, there are many more
elaborate proposals for information retrieval. These can be
evaluated only through experiments which will determine their
usefulness and systems-engineering studies which can estimate
their cost.

It is not only through gathering, digesting, and locating in-
formation that computers will serve management. One impor-
tant general area is that of exploring the consequences of pro-
posed actions. In engineering research, simulation is an example
of this. The computer can be used to simulate the operation
of a new device, perhaps a device to transmit speech with re-
duced bandwidth, or a device to recognize or to generate
speech. Or a computer can simulate the operation of a control
system or a switching system. Such research simulation does
not solve the problem of building a new device. It answers the
question, what sort of results would we obtain if we did build a
successful model of the device?

In production, linear programming can go beyond the de-
termination of operating costs for a given distribution of work;
it can find, under certain assumptions, an optimum distribu-
tion of work. Linear programming can be used as a routine
operating tool. As a more strictly management device, however,
it can be used to evaluate past assignment procedures.

In the telephone industry, and in some others as well, a
change in classes of service and rate structures may open up new

303

J. R. Pierce

business opportunities, but it will certainly cause some present
customers to fill present needs under different classes of service
at new rates. Even if we assume the simplest sort of fully in-
formed, economically motivated behavior, the consequent
changes in semce pattern and revenue can be worked out only
with the aid of the largest existing computers and the most
advanced programming techniques.

Simulation, linear programming, studies of the consequences
of changes in rates and classes of services are in principle simple
examples of instances in which the computers can help men to
explore the consequences of decisions, or even to arrive at better
decisions. It is easy to propose to go further than this, and peo-
ple will no doubt try to. It has been proposed to simulate on
computers the behavior of chemical plants that involve chem-
ical processes. If we really understood how a business func-
tioned, we presumably could simulate the business on a
computer, or at least one aspect of the business. This indeed is
in the spirit of what the military does in its endless war gaming.

Such elaborate simulations must depend on the availability
of valid, useful information and on an understanding of busi-
ness operation which I doubt that we have. Such over-all simu-
lation must evolve gradually through the ideas and experiments
of highly competent people who are interested in the operation
and problems of businesses.

The Management of Machines

Of course, none of this can be done without machines, and
the acquisition and management of machines pose serious prob-
lems in themselves. In acquiring and processing data, how
much should be done at remote points and how much by a
central computer? How is electrical communication to be pro-
vided? \\ "hat will the reliability and cost of various equipment
be;5

Most industries have business computing machinery for pay-
roll and other accounting purposes. Some data-processing and
other studies can be done by this equipment, but some work
calls for a large scientific computer. Can time be rented on

304

What Computers Should Be Doing

an outside machine? Especially if outside programming is in-
volved, is this consistent with company privacy? Perhaps the
research department has a large computer. If so, can this be
used without disrupting research work?

Finally, when a large computer is available, should different
locations have their separate computers, or should one central
computer serve several locations? Can a remote computer be
operated from a distance as flexibly and readily as if it were on
the premises? How can one get data for a host of small prob-
lems into and out of a fast computer and still earn7 out large-
scale jobs efficiently?

Summing Up

I have said a variety of things about computers and their use
which I hope will be rather disturbing to managers. Among
them are the following:

In applying computers to management problems, new ground
must be broken. This can be done effectively only by very
competent men.

Machines are not men. Just because a computer can do some-
thing a man might do does not mean that it should. Amid a
hullabaloo of enthusiastic aspirations and sales talk, manage-
ment must find out what sorts of things computers should do
for it.

What computers should do depends not only on the com-
puter and the problem but on the ingenuity of the programmer.

Ingenious programming can make computers useful for spe-
cial purposes to persons with no programming background.

While machines can process data, the results are no better
than the data, and information may be lost in the processing.

WTiile machines can process the results of surveys, it takes a
good social scientist to plan a survey, and a good statistician to
evaluate the results.

An information-retrieval system can succeed only if it is
operable and economical. It takes a great deal of expert knowl-
edge to work out and to evaluate even the machine side of such
a system.

305

/. R. Pierce

Further, all sorts of systems considerations arise in connection
with any large-scale use of computers, considerations having to
do with dispersal as opposed to centralization, with electrical
communication, and with enabling the over-all computer in-
stallation to cope efficiently with a variety of large and small
problems.

Management is management, but machines are not men.
Somehow management must learn the management of com-
puters. This may not be easy, especially in large organizations.

Wise use of, and wise decisions about, computers are not to
be mastered quickly or easily. They call for brains of the highest
order. They call for close and protracted application. There is
real doubt whether they can be made by a man harassed with
the routine of a large, multilevel organization. They certainly
cannot be made by even a bright specialist in management un-
less that specialist is willing to devote years of hard effort to
learning things quite outside of his specialty, things that have
to do with machines and systems of machines, with data and
information, rather than with organizations of people and with
human idiosyncrasies.

How can a man capable of dealing effectively with machines
and machine problems be given his proper reward, status, and
responsibility in management? How can he and the expert pro-
grammers who work for him be kept from fleeing to the large
material rewards and the great freedom and personal responsi-
bility of research? How can management find a way to reward
success in such a job other than by transferring the man into
one of a large number of specialized and more-or-less inter-
changeable management jobs? How can management resist
the temptation to give such a job to a man because he has been
highly successful in a number of management jobs which shared
a specialized content quite different from that called for in
managing machines?

The problems raised are a little like those of properly re-
lating lawyers and doctors to management, but it would be
wrong to think that they are identical. And, unless management
finds a way to acquire, use, and keep the services of first-rate

306

What Computers Should Be Doing

people who are capable of managing machines, it will not ex-
ploit the full potentialities of the computers of the future.

Panel Discussion

Bush. Dr. Pierce has told us about a race of machines and a
race of men. Both races appear to be prospering and proliferat-
ing. The only difficulty seems to be that the rapport between
them is not all that might be desired. Norbert Wiener men-
tioned to me the other day that he is now ready to build a
machine that would duplicate itself. What Wiener means by
duplication and construction and what I mean may be very
different things. But it is possible to visualize a machine in the
desert, surrounded by its numerous progeny, busily computing
all sorts of things to which no one is paying any attention
whatever. I think that Dr. Pierce has placed before us a number
of interesting problems on this matter of the interconnection
between the human mind and the machine. I am going to ask
Dr. Shannon to open the discussion. Dr. Shannon is the father
of the modern theory of communication, insofar as it has a
father. I need take no more of your time introducing a man
whose work is known to all of you.

Shannon. A mathematician friend of mine, Ed Moore, who
enjoys classifying people as well as mathematical concepts, once
showed me a chart on which he had placed many of the scien-
tists working with computers. The chart formed a kind of spec-
trum, ordered according to how much one believed computing
machines were like the human brain or could be said to perform
intelligent thought. On the far right were those who view com-
puters as merely glorified slide rules useful only in carrying out
computation, while at the far left were scientists who feel com-
puters are already doing something pretty close to thinking.
Moore had placed me, to use a phrase of Roosevelt, "a little
left of center." I think Dr. Pierce has taken a very sound posi-
tion quite close to dead center. Many of the differences between

307

C. E. Shannon

people along this spectrum relate, I think, to a lack of careful
distinction between computers as we now have them for work-
ing tools, computers as we are experimenting on them with
learning and artificial-intelligence programs, and computers as
we visualize them in the next or later generations.

One of Dr. Pierce's main points, with which I completely
agree, is that there are some jobs for which computers are better
and others for which men are better. One should not send a
machine to do a man's job, and vice versa. I should like to ex-
plore a bit further the differences between computers as we
have them today and the human brain. On the one hand, we
have the fact that is often pointed out that computing machines
can do anything which can be described in detail with a finite
set of instructions. A computer is a universal machine in the
sense of Turing: It can imitate any other machine. If the human
brain is a machine, a computer with access to sufficient memory
can, in principle, exactly imitate the human brain. There are,
however, two strings on this result. One is that no mention is
made of time relations, and, in general, one machine imitating
another is greatly slowed up by the mechanics of describing one
machine in terms of the second. The second proviso is even
more important. One machine can imitate another or carry out
a computing operation only if one can describe exactly, and
in all detail, the first machine or the desired computing opera-
tion. Of course, this we cannot do for the human brain. We do
not know the circuitry in detail or even the operation of the
individual components.

As a result of these and other considerations, our attempts
to simulate thought processes in computers have so far met
with but qualified success. As Dr. Pierce has pointed out, there
are some things which are easy for computers to do, operations
involving long sequential numerical calculations, and others
which are unusually difficult and cumbersome. These latter in-
clude, in particular, such areas as pattern recognition, judgment,
insight, and the like.

As viewed from the outside, as a behaviorist psychologist
might, there are other important differences between computers

308

What Computers Should Be Doing

and brains. Machines are "taught" how to do something by
giving them complete and detailed instructions with regard to
their action in each specific situation. People, on the other
hand, are taught most often by example. One gives a child a
few examples of an action, and the child automatically gen-
eralizes to cover a wide variety of similar situations. Until it is
possible to set up our computers so that this form of instruction
can be easily done, artificial intelligence will be limited by pro-
gramming ingenuity. I might mention, however, that various
programs have been set up which do, in fact, learn limited types
of response by generalization of special cases.

If there are these important differences at the psychological
level between computers as we have them today and brains, one
may raise the question as to whether this is a reflection of a
different internal organization, and if so, what are the chief
differences? I believe that, in fact, there is very little similarity
between the methods of operation of the computers and the
brain. Some of the apparent differences are the following. In
the first place, the wiring and circuitry of the computers are
extremely precise and methodical. A single incorrect connection
will generally cause errors and malfunctioning. The connec-
tions in the brain appear, at least locally, to be rather random,
and even large numbers of malfunctioning parts do not cause
complete breakdown of the system. In the second place, com-
puters work on a generally serial basis, doing one small opera-
tion at a time. The nervous system, on the other hand, appears
to be more of a parallel-type computer with a large fraction of
the neurons active at any given time. In the third place, it may
be pointed out that most computers are either digital or analog.
The nervous system seems to have a complex mixture of both
representations of data.

These and other arguments suggest that efficient machines
for such problems as pattern recognition, language translation,
and so on, may require a different type of computer than any
we have today. It is my feeling that this computer will be so
organized that single components do not carry out simple,
easily described functions. One cannot say that this transistor

309

V. Bush

is used for this purpose, but rather that this group of compo-
nents together performs such and such a function. If this is
true, the design of such a computer may lead us into some-
thing very difficult for humans to invent and something that
requires very penetrating insights. Most machines are invented
by breaking the over-all problem down into a series of simple
and perhaps previously solved problems and combining these
to effect the final design. In such designs, particular compo-
nents have simple functions — a carburetor mixes gasoline and
air, a spark plug ignites the mixture. In a machine of the type
I am suggesting, it would be impractical to describe the purpose
or action of any single component. I know of very few devices
in existence which exhibit this property of diffusion of function
over many components. One example is certain types of relay
switching circuits arrived at by mathematical design procedures
in which the mathematical techniques keep our design pro-
cedure in order in spite of human limitations.

If this sort of theoretical problem could be solved within the
next few years, it appears likely that we shall have the hardware
to implement it. There is a great deal of laboratory work in
progress in the field of microminiaturization, the use of thin-
film memories and the like. We are almost certain to have com-
ponents within a decade which reduce current transistor cir-
cuits in size much as the transistor and ferrite cores reduced
the early vacuum-tube computers. Can we design with these a
computer whose natural operation is in terms of patterns, con-
cepts, and vague similarities rather than sequential operations
on ten-digit numbers? Can our next generation of computer
experts give us a real mutation for the next generation of com-
puters?

Bush. Dr. Shannon has told us a number of things, among
them that we are making our machines too obedient, that they
follow our instructions too literally, and that perhaps they learn
from their experience too little. Of course, the human race it-
self has gotten into trouble a number of times by being too
obedient and too regimented. It also can get into trouble by not

310

What Computers Should Be Doing

being disciplined or regimented enough, a danger that we may
expect carries over to our parallel race. Dr. Shannon touched
on the problem of the relationship between the machines as we
know them and the human brain. He said, I believe, that we
have two types of machines, the digital and the analog, and
that the brain is probably neither. To me some of the most in-
teresting machines built by man are neither digital nor analog.
I think that this field is somewhat neglected today, yet in it
lies much of the promise for the future. I think many of us have
been too closely associated with one of two definite paths with-
out looking at the intermediate ground. . .

With respect to the relationship between the computer and
the functioning of the brain, with us tonight is the man most
qualified to speak on that subject. Professor Rosenblith is well
known to all of you as one of the principle exponents of the
biophysics of sensory communication processes. He has delved
as deeply as any man living into the enormous, intricate, fas-
cinating problem of the similarities between ways in which the
brain and the computer operate. Analogy with the computer
may mislead us, but at the same time it gives us a base from
which to start. I take great pleasure in introducing to you Pro-
fessor Rosenblith.

Rosenblith. The following remarks are hardly those of a
"discussant," if we accept the root meaning of this term as one
who tries to tear the speaker asunder whether he disagrees with
him or not. Instead I shall attempt to supplement, from my
own limited experience with computers and with a slightly
different emphasis, Dr. Pierce's and Dr. Shannon's remarks.

The topic for this evening reminds me of a colorful expres-
sion that an acquaintance of mine uses in a different context;
he designates certain questions as "Hialeah-type questions"—
all they ask is, "Who is ahead?" Our distinguished chairman
has reformulated the topic as the challenge of the parallel race.
My inclination is to substitute coexistence and cooperation for
competition. The real challenge then consists in creating a
novel, more powerful, self-modifiable, quasi-symbiotic system

311

W. A. Rosenblith

that will combine the assets which a long evolution has be-
stowed upon man with those which man's inventiveness has
bestowed and will bestow upon the computers of tomorrow. I
am therefore less tempted to stress what computers can do
better than men than to envisage the benefits that we might
derive from an intelligent division of labor between man and
computer. Such arrangements are very likely to enhance hu-
man capacities in just those areas that are crucial to the func-
tioning of a world whose technology is rapidly evolving.

Both the industrial revolution, which bore the imprint of the
steam engine, and the cybernetic revolution of automation,
which is symbolized by the computer, have given rise to diffi-
culties. These difficulties affect the coupling of man to his
devices as well as relations between men. Both revolutions have
also drastically altered man's image of himself.

The promise of the cybernetic era resides in the fact that the
new technology may prove capable of providing more than
mere substrata for a rational flow of communication and con-
trol messages; it is likely that it will furnish some of the needed
tools for the development of the sciences of man. We may
thus obtain the instrumentalities for the successful manage-
ment of human wants and institutions, and perhaps even for
the self-management of human behavior.

Let me now be more specific and turn to some explicit tech-
nical points. The first deals with the role that computers play
in the recognition of patterns. Dr. Pierce has, in my opinion,
undervalued this role by selecting as his principal examples the
spoken and written word, which demand indeed a great deal of
learning of the relevant features of certain patterns by human
brains. Transplant a man into a foreign language environment,
and he will be quite poor at discriminating and especially at re-
producing linguistic events that are child's play to the natives.

Research in fields of science which lack adequate physical
theories or mathematical foundations is often characterized by
the experimenter's inability to control his many interacting
variables. In such situations, computers, especially when they
can be used "on-line," often enable the experimenter to detect

312

What Computers Should Be Doing

distinctive features in the patterns of behavior whose signifi-
cance he is trying to assess. He may often be able to do more:
By including himself, his experimental subject, and the com-
puter in a closed loop, he may be able to formulate and discard
his hypotheses quickly enough so they can be tested in a multi-
variate configuration that has remained sensibly the same.
Whether one wants to call this pattern recognition is a matter
of definition and perhaps even taste, but it is a fact that the
combination of man and computer is capable of accomplishing
things that neither of them can do alone.

Such uses of computers have started to affect profoundly
research in certain areas of the life and behavioral sciences.
Tomorrow our practices in medicine, human genetics, and ed-
ucation will find themselves modified in turn. The nation's
health record system will affect management practices in hos-
pitals, in industrial medicine, in the Veterans Administration,
in social service organizations: The effect may perhaps not be
as dramatic as Dr. Kemeny's vision of the library system of the
year 2000, but something of this kind is likely to occur.

These changes will, of course, not occur overnight or auto-
matically. In many fields in which there is a temptation to
apply computers, the established professions are hardly prepared
to take advantage of the opportunities. It is just not enough to
transfer existing unsatisfactory and outdated practices onto
punched cards and trust that programmers will do the rest. A
new generation of professionally trained scientists or practi-
tioners presumably will grow up and will know enough about
computers to deal with the basic problems of programming, of
peripheral equipment, and so forth. They will then be able to
judge which aspects of their data are computer-compatible or
worth "hardening," to use the phrase in vogue; indeed, they
should then be able to decide whether the time-honored ques-
tions and practices of their profession are still valid in the
context of the new technology.

Whenever computers are under scrutiny, it seems obligatory
to compare their performance with that of the brain. Most often
the brain turns out to be the human brain, and at that not just

313

W. A. Rosenblith

any human brain but one that can prove theorems, do prob-
lems in mathematical logic, and preferably play chess. On the
other hand, casual observation reveals that evolution has pro-
duced an incredible variety of brains: Some are genetically quite
rigidly precoded and capable of only a very restricted repertory
of behaviors, while others are capable of many learnings, in-
cluding the manipulation of abstract symbols. We do not know
of a universal IQ test that would permit us to rank order the
brains of all species along a single continuum. A fortiori we
cannot expect to make valid over-all comparisons between the
two parallel races, one of which is just barely emerging from the
ylem or ur-state. Instead we must at present be satisfied with
comparing specific performances of a given computer and a
given brain, and with increasing our understanding by simulat-
ing the performance of the latter with the aid of the former.

We are hopeful that we shall gain added insight into both
brains and computers by analyzing their structural principles
in relation to their programmability. Some of Claude Shannon's
remarks on serial versus parallel operation, on diffused-function
operation which hark back to a theme of von Neumann's, make
this hope more explicit. Present-day neuroanatomy and neuro-
physiology do not furnish us with recipes of what mix of hier-
archical organization, specificity, randomness, and redundancy
will yield a particular performance in the sensory domain, for
instance. These sciences are likewise incapable of telling us to-
day how evolution came to make such extended use of analog-
to-digital (and vice versa) recoding and interaction schemes.
Perhaps those who preside over the evolution of computers will
contribute to the study of brain function by developing a series
of calculi of relations that will in some sense transcend the
digital versus analog dichotomy.

From all we know, it seems reasonable to assume that in
universities, colleges, and eventually even in high schools the
young will learn how to talk and listen to computers, and how
to use them in their thinking and planning. Few aspects of life
will be able to resist such a massive reconsideration. Computers
are likely to affect their environment the way cars did, provided

314

What Computers Should Be Doing

the rapport between human beings and computers will become
sufficiently widespread and intimate. Cars have given us a mo-
bility that far exceeds what evolution has provided us with;
computers may constitute a comparable evolutionary step with
respect to man's intellect.

No technology comes with a built-in guarantee against mis-
use. Thus life with computers may be better or just different.
Norbert Wiener in his book, The Human Use of Human Be-
ings, has alerted us to the problem: For what human and
humane purposes will we use computers tomorrow?

General Discussion

Bush. The meeting is now open for general discussion. It
is perfectly legal and ethical to ask a question of anyone on this
panel as long as you do not ask it of the presiding officer. I was
qualified to engage in this type of discussion twenty or thirty
years ago.

McCarthy. I should like to "discuss" the paper in the
etymological sense. I think what Dr. Pierce said concerning
present computer applications is true, but I think his general
picture of things is almost completely false. His picture is static.
It is based on computers as they are today and on practices
already obsolete in some cases. I think that the picture of pro-
grammers as a highly trained group is misleading. None of us
has received any particular training; we simply have had a little
practice. I think that the picture of programming languages as
designed to make the machine easier for the duffer is out of
date. The new programming languages are not particularly
easier to learn than machine code. Their purpose is to be more
powerful, to permit an idea that can be expressed verbally in
five minutes to be programmed in two weeks instead of six
months, as is now required using straightforward machine lan-
guages.

On the question of information retrieval one would ask,
"Why is this difficult?" Dr. Pierce's answer is that it takes time;

315

General Discussion

it is too expensive to put the data into machine-readable form.
The problem unfortunately is not as simple as this. If it costs
$200 to type a book onto IBM cards, then for $1,000,000 you
can type 5000 books. The trouble is that you will not know what
to do with the 5000 books once you have them on IBM cards,
because no present computer has the memory capacity to handle
such a quantity of data properly.

The distinction that Dr. Pierce made between what is suit-
able for computers and what is not suitable for computers seems
to reduce to, what is simple has already been programmed and
what is complicated is taking longer. The reason for my annoy-
ance is a belief that unfamiliarity at the practical level is more
of a handicap in the computer field than it is in most others.
In most other fields it is possible to take a general view. But
since computation is the subject of what intellectual processes
are like in detail, it is a subject for which there is almost no
theory worthy of the name. Practical experience is really all that
we have, and even this is extremely tenuous.

Pierce. I am not at all converted by what Dr. McCarthy
has said. In the first place, as I stated at the outset, I addressed
myself primarily to the problems of management. As I men-
tioned, what management will be doing with the computer to-
morrow is what some people in research and development lab-
oratories are doing with the computer today. In the second
place, don't let anyone fool you. Although Dr. McCarthy may
pick up programming without quite knowing he is doing it,
when you bring a person with a Master's degree and lesser
talents into the job of writing programs, it is only after some
years that he gets really good at it, and the supposition is that
he has been learning something during this period. Now, about
the programmers as intermediaries; this is indeed bad in the
routine use of computers, but the way it is overcome is not by
learning programming without trying, nor is it by writing more
powerful languages that will enable one to assail problems that
no on£ has approached yet. It is by writing very simple lan-
guages which enable the engineer to use the computer to sim-

316

What Computers Should Be Doing

ulate electrical equipment, which enable the equipment builder
to determine where the modules should go in the assemblies in
order to minimize the length of wire required, and which en-
able the designer to design logical circuitry without the cus-
tomary mental pain.

With respect to information retrieval, I did not mean to
imply that reducing data to machine-readable form was the
only problem. I shall say this, however. If data were reduced to
machine-readable form, some very simple things could be done
by devices that are not as complicated as full-scale computers;
for example, sorting through correspondence for the mention of
a particular name. I shall say also that information-retrieval
schemes of a very simple-minded sort (one was tried at the
patent office) have foundered on this very point of reducing
information to machine-readable form.

Concerning the future of computers: The airplane is a very
wonderful device even though it does not resemble the pigeon,
and the automobile is a very wonderful device despite the fact
that it will not clamber around mountains as goats can. In a
similar way, the computer is a very wonderful device even
though it only does some things that Dr. McCarthy regards as
"old hat." I think that there are many possible advances in
computers which will come from looking for problems, some-
times very unexpected problems, which are easily accessible to
computers and which computers will do very well. One of these
advances occurred in engineering research when computers were
employed to simulate the operation of circuits. Other examples
are the preparation of drawings by computers, the checking of
logical design by computers, and the determination of optimum
assembly layouts by computers. It seems to me that human
beings are very fallible. They do not foresee the future very
well. They find the future by doing things. If they do simple
things first, they may get to the future faster than if they set
themselves very remote goals and then get discouraged.

One of the things that has bothered me is people setting
themselves rather artificial goals, such as language translation,

317

General Discussion

which is very remote in a practical sense, or such as a baseball in-
formation program, which is just an illustration and not very
useful. It seems to me that the sort of goal which will lead to
an advance in computer applications is the goal which is ac-
cessible in the not too distant future and which will be highly
useful once you attain it. I should contrast algebraic manipula-
tion with theorem proving as an illustration. I think that com-
puters within the next few years may very well perform many of
the chores of algebraic manipulation which occur in orbital
calculations, quantum-mechanical calculations, and electron-
beam problems. For instance, in fitting boundary conditions
you run headlong into lengthy manipulations that, by hand,
you get wrong the first five times. Computers some day may do
these manipulations very nicely. They may be doing them right
now at M.I.T., but I do not believe that this has reached a
stage at which it is highly useful to mathematicians.

I have one last point, and it concerns something which
Walter Rosenblith mentioned. What is a pattern? There are
many patterns that human beings find easy to recognize, and
they are very important to us: the spoken word, the written
word, and so forth. But these are hard for computers. On the
other hand, we all know that the world is full of patterns that
human beings find difficult to recognize. Chemical tests and
urinalyses are simple examples. An electrical filter will filter a
sine wave out of a lot of noise, while the filter in our ear is
broad-band and cannot do this. Recently at the Bell Labora-
tories, Bela Julesz has been inserting patterns of various sorts
into a configuration of random dots. Some patterns are recog-
nizable by eye immediately and some are not, but one is just as
easy as another for a computer. Thus, many simple-minded
recognition tasks, which are not as complex as recognizing script
or speech, but nevertheless which require a property that the
human eye does not have, will undoubtedly be done in a very
practical way by computers. Indeed, Walter Rosenblith has
been doing just this type of thing.

Licklider. As McCarthy implied, computers give us for the
first time the tools with which to come to grips in a serious way

318

What Computers Should Be Doing

with intellectual processes. We have never had them before,
and we have made more progress since getting them than in
decades past, despite the flowering of mathematics in the early
part of the century. I should like to reinforce what Rosenblith
said about developing a real feeling of rapport with the comput-
ing machine. He suggested the analogy of the car and implied,
I think, that anybody who does not feel the extension of the
human body into a machine does not really have a basis for
feeling what the intellectual extension might be. Some people
who have spent hours at tennis can hardly believe that the
racket is not alive. People who have spent hours with computers
have comparable difficulty.

Bush. When you speak of rapport between men and ma-
chines and illustrate it by the automobile, I wonder how often
you drive home in commuter traffic.

Anonymous. I wonder if Professor Shannon would com-
ment further on one of his earlier remarks. If I understood him
correctly, he considered the fact that the brain is not hooked
up at all the way a computer is as a roadblock in the way of
simulating brains or brain-like machines on computers. Why is
this an obstacle? As Dr. Pierce has said, airplanes are nothing
whatever like pigeons, yet in some ways they fly much better.
Is it not possible that a highly parallel machine has significant
disadvantages which we do not happen to know about at the
moment?

Shannon. I did not say that parallel machines would be
better, but only that they would be different and might lead to
the possibility of doing more easily some of the things which
seem difficult in serial machines: the things which are easy
comparatively speaking for brains to do. I agree that the auto-
mobile is fine as an automobile, and I agree that the serial
machine is fine for the kinds of things for which it is designed,
for example, numerical calculations of great depth and number.
So there is no argument there. But computing machines as we
have them today, the analog and the digital, are only two of a
vast population which we have not yet explored. What I am
saying in my challenge is that we should look to other possible

319

General Discussion

machines in the same general area of information processing.
Just because we have bicycles, we should not stop looking for
automobiles and airplanes. Carrying the point a bit further,
there are two courses open at our present state of technology.
We can attempt to refine and improve what we have as much
as possible — in the case of computers we can try to develop
our programming skill to the utmost — or we can take a com-
pletely different tack and say, "Are there other types of com-
puting machines that will do certain things better than the
types we now have?" I am suggesting that there well may be
such machines.

Bush. Have you any comments at this point, Professor
Rosenblith?

Rosenblith. When we talk about The Brain, we are cer-
tainly overgeneralizing almost as much as when we talk about
The Machine. Brains come in all sizes. As Adrian has remarked,
unless they are about as large as an apple, we do not expect
much evidence of our kind of reasoning and generalization.
Those who work today in the field of artificial intelligence deal
often with types of problems that cannot yet be studied in any
meaningful way by analyzing the electrical activity of the brain.
But brains are multipurpose devices, and the intellectual as-
pects of their activity are not the whole story by any means.
Brains are programmed partly by hereditary factors and partly
by learning; they are reprogrammed, almost minute by minute,
as they are stimulated by the environment. If one records, for
instance, by means of a microelectrode from a cell in so-called
association cortex, one may find that the cell seems interested
in auditory stimuli to begin with; it then shifts its interest to
visual events, then to what happens on the skin, and so forth.
This kind of flexible arrangement may be the chief characteris-
tic of a large tribe of brain cells. Obviously, not all parts of the
brain are as flexibly connected as this, but it would seem that
the brain disposes of a certain reserve of uncommitted neurons
that can be applied wherever and whenever there is a need for
information processing. This unusual and only partially under-
stood combination is far from being readily duplicable by our

320

What Computers Should Be Doing

current machines. In order to understand this type of behavior,
we need to know not only the neuroanatomical circuitry and
the neuroelectric events, but we also need to know a great deal
more neurochemistry, and to get a grasp of how the varieties
of memory are coded. In the meanwhile, we should emphasize
in computer simulation studies the range and combinations
of performances that living brains are capable of. Thus, we
may develop a catalogue of programmable machines exhibiting
a whole gamut of behavior repertories. The brain systems of
tomorrow are then likely to combine and match the capabilities
of logical machines with the flexibility and sloppiness of bio-
logical ones.

Dreyfus.* After hearing Dr. Pierce's remarks attacked by
the computing fraternity's vociferous far left, we feel that a few
comments are in order. We undertake this defense primarily
to show that M.I.T. has not yet been taken over entirely by
machines. We use machines to characterize the relentless
prophets of the omniscient computer, since it appears, in this
matter at least, that they lack the fundamental attribute of the
human brain mentioned by Dr. Rosenblith — a few uncom-
mitted neurons.

Specifically, we should like to address ourselves to the fre-
quent and rather immodest claim of manyfold progress in arti-
ficial intelligence. How much progress has been made toward
the world champion chess machine (so confidently predicted)
when, in the past few years, the rules of chess and a few simple-
minded heuristics of its play have been committed to the
punched card? How much progress did the cave man make
toward space flight when he climbed his first mountain? Some-
thing fundamental to significant progress is lacking in both
cases — the conceptual, or technical, or technological break-
through. Work in the fields of language translation, game play-
ing, and pattern recognition has contributed nothing to the
understanding of the nature of intelligence or insight. This
suggests that the solution of profound problems touching upon

* These remarks by H. L. and S. E. Dreyfus were not made at the
Pierce session itself, but were submitted in writing a short time thereafter.

321

General Discussion

the nature of thought may not be the sum of many minute
steps.

If programmers set themselves glamorous tasks as difficult to
execute as they are appealing to the press (which includes some
so-called technical journals), let them report progress when the
necessary breakthrough occurs, and not until then. Such re-
searchers should run the same risks as the alchemist trying to
synthesize gold from base materials: obscurity until success.
Artificial intelligence seems to be operating instead on the prin-
ciple of fame until failure. In contrast to Dr. Pierce's sober
remarks, the claims of the far left read like fiction, and bad
fiction at that, since the ending is always "deus ex machina."

Minsky. The present-day stored-program serial computer is
presumably in a very early phase, and I have a strong objection
to "computer ' being used to point to the IBM 704 or any other
particular representative, although perhaps this is necessary in
order to avoid using a long string of qualifiers. Every designer
of programming languages nowadays is thinking about more
complex associative memories for serial computers. Parallelism
is in the air. The next generation of machines, or the one after
the next generation, will certainly have large parallel aspects.
But looking at a conventional computer of the caliber of the
IBM 7090, despite the comments from the platform, we are
not sure that it cannot solve many of the difficult intellectual
problems that face us. The brain has had a billion years to make
good use of its neurons, and the programmers have not. In the
last two years we find the IBM 704 solving problems of an
order of magnitude more difficult than it did in 1958 or 1959.
Although I doubt it, it still remains possible to me that the
IBM 704 with its crude serial mode of operation could be as
intelligent as man in fairly broad areas, just provided it had a
larger memory.

This is not a closed issue. We are nowhere near the limits of
good heuristic programming which, to parallel what Dr. Rosen-
blith said, have partly hierarchical and partly disorderly ar-
rangements. And the situation is improving fairly rapidly. Still,
the parallel computer shows great promise.

322

What Computers Should Be Doing

Concerning pattern recognition, just as a specific matter, I
should bet that the serial IBM 704, if properly programmed,
could read printed letters faster than a human, even though
present programs do not. I should bet that eventually the
704 even could read script faster than a human.

Pierce. These are nice bets, but how long do we have to
wait? Will I live so long? I do not doubt that it will read script
faster, if you do not care how well it reads it. There is a wonder-
ful tendency to talk about things that lie in the future and that
you cannot prove will not happen. This is good clean fun be-
cause it is the only way we have of giving variety to the future.
When the future comes there is going to be just one future,
but here in anticipation we can enjoy a lot of different futures,
and I am pleased to see that everyone is enjoying himself. As
for me, I do not know what is going to happen, and I find the
future more difficult to talk about. There is a certain dullness
in talking about things that are either in our grasp or are just
coming into our grasp, but I think that it is legitimate to talk
about these things, even though they are less glamorous, be-
cause they are the things that seem to be neglected.

Abrahams. I should maintain, Dr. Pierce, that in a sense
you are placing your bet on the wrong horse. You recommended
that effort should be shifted from machine translation to hu-
man translation, but I think there are fairly broad classes of
problems which are in a sense simply beyond human capabil-
ities, problems characterized by the fact that we do not know
algorithms for solving them or they do not appear to be within
human capabilities. One example is the problem of information
retrieval where the amount of information that a human being
needs to digest is just too much for him. Another example is
the problem of translating from an unknown language. We do
not know algorithms for solving such problems, yet computers
might be able to solve them by generating programs, perhaps
by heuristic methods that we cannot understand. In other
words, programs may grow essentially by accretion. The pro-
grams which you have restricted yourself to are those which the
programmer can understand. I feel that this is a very strong

323

General Discussion

restriction. Learning programs may open up a new avenue of
approach, and this has already been explored to some extent by
Newell and Simon; but there remains a great deal to be done.

Pierce. There sure does. I want to say something about the
problem of information retrieval and the flow of information.
In my view, the greatest step that could be made in this direction
is fewer journals and more stringent editorial policies for
journals now being published.

Bar-Hillel. I first of all want to agree with Dr. Pierce that
we have very little prospect of knowing what is going to happen
in the future, even in the near future. In spite of this uncer-
tainty, we must make up our minds as to which types of future
we should like to encourage, and we must recognize that certain
types are unachievable. One such decision soon must be made
by computer people. Do we want computers that will compete
with human beings and achieve intelligent behavior autono-
mously, or do we want what has been called man-machine sym-
biosis? I admit that these two aims do not definitely exclude
each other, but there has been an enormous waste during the
last few years in trying to achieve what I regard as the wrong
aim at this stage, namely, computers that will autonomously
work as well as the human brain with its billion years of evo-
lution. Of course, it may turn out that ten or fifteen years from
now certain unforeseen developments will enable computers to
perform much more marvelously than seems possible today.
But I think computer people have the obligation to decide now
which of the two aims they are going to adopt.

Pierce. In the matter of making a decision now on the
basis of an unknown future, I am reminded of a remark in-
advertently made by Mr. Webb, Administrator of the National
Aeuronautics and Space Administration. (I am sure this remark
does not really describe the operation of the agency.) Mr.
Webb stated on television that "We have a step-by-step pro-
gram in which each step is based on the next."

324

What Computers Should Be Doing

References

Bar-Hillel, Y., "The Present Status of Automatic Translation of Lan-
guages," in Advances in Computers, Vol. 1, Alt, F. L., Ed., Academic
Press, Inc., New York, 1960, pp. 91-163.

Carr, J. W., HI, "Programming and Coding," Chap. 2 of Handbook of
Automation, Computation and Control, Vol. 2, Grabbe, E. M., Ramo,
S., and Wooldridge, D. E., Eds., John Wiley & Sons, Inc., New York,
1959.

Gotlieb, C. C, and Hume, J. N. P., High-Speed Data Processing, Mc-
Graw-Hill Book Co., New York, 1958.

Kelly, J. L., Jr., Lochbaum, C, Vyssotsky, V. A., and others, "A Block
Diagram Compiler," The Bell System Technical Journal, Vol. 40, 669-
676 (May, 1961).

Leagus, D. C, Lee, C. Y., and Mealy, G. H., "Verification of the Logic
Structure of an Experimental Switching System on a Digital Computer,"
The Bell System Technical Journal, Vol. 38, 467-476 (Mar., 1959).

Ledley, R. S., Digital Computer and Control Engineering, McGraw-Hill
Book Co., New York, 1960.

Lyapunov, A. A., Goodman, R., and Booth, A. D., Eds., Problems of
Cybernetics, Vol. 1, Pergamon Press, Inc., 1960 (translated from the
Russian) .

Mathews, M. V., "An Acoustic Compiler for Music and Psychological
Stimuli," The Bell System Technical Journal, Vol. 40, 677-694 (May,
1961).

Minsky, M., "Steps Toward Artificial Intelligence," Proceedings of the

IRE, Vol. 49, 8-30 (Jan., 1961).
Oettinger, A. G., Automatic Language Translation, Harvard University

Press, Cambridge, 1960.
Proceedings of the International Conference on Scientific Information,

Washington, D. C, 1958, Vols. 1 and 2, National Academy of Sciences

and National Research Council, Washington, D. C, 1959.
Samuel, A. L., "Programming Computers to Play Games," in Advances

in Computers, Alt, F. L., Ed., Vol. 1, Academic Press, Inc., New York,

1960, pp. 165-192.

Wang, H., "Toward Mechanical Mathematics," IBM Journal of Research

and Development, Vol. 4, 2-22 (Jan., 1960).
Wang, H., "Proving Theorems by Pattern Recognition — II/' The Bell

System Technical Journal, Vol. 40, 1-41 (Jan., 1961).

325

Selected Bibliography

Social Issues and Management Implications

Shultz, G. P., and Whisler, T. L, Eds., Management Organization
and the Computer, The Free Press of Glencoe, Illinois, 1960.

Simon, H. A., "The Corporation: Will it be Managed by Ma-
chines?/' Management and Corporations, J 985, McGraw-Hill
Book Co., New York, 1960.

Simon, H. A., The New Science of Management Decision, Harper
& Brothers, New York, 1960.

Wiener, N., The Human Use of Human Beings, Houghton Mifflin
Company, Boston, 1950.

Wiener, N., "Some Moral and Technical Consequences of Auto-
mation," Science, Vol. 131, 1355-1358 (May 6, 1960).

Simulation of Large Systems

Cohen, K. J., and Cyert, R. M., "Computer Models in Dynamic

Economics," The Quarterly Journal of Economics, Vol. 75,

112-127 (Feb, 1961).
Conway, R. W, Johnson, B. M., and Maxwell, W. L, "Some

Problems of Digital Systems Simulation," Management Science,

Vol. 6, 92-110 (Oct, 1959).
Forrester, J. W, Industrial Dynamics, The M.I.T. Press and

John Wiley & Sons, Inc., New York, 1961.
Geisler, M. A, "The Simulation of a Large-Scale Military Activity/'

Management Science, Vol. 5, 359-368 (July, 1959).
Orcutt, G. H, Greenberger, M, Korbel, J, and Rivlin, A. M,

Microanalysis of Socioeconomic Systems, Harper & Brothers,

New York, 1961.
Orcutt, G. H, Shubik, M, Simon, H. A, and Clarkson, G. P. E,

"Simulation: A Symposium," three articles in American Eco-
nomic Review, Vol. 50, 894-932 (Dec, I960).
Pool, I. de S, and Abelson, R, "The Simulmatics Project," Public

Opinion Quarterly, Vol. 25, 167-183 (Summer, 1961).

327

Selected Bibliography

Artificial Intelligence

Gelernter, H. L., Hansen, J. R., and Loveland, D. W., "Empirical
Explorations of the Geometry Theorem Machine," Proceedings
of the Western Joint Computer Conference, San Francisco,
Calif., May, 1960, The Institute of Radio Engineers, New York,
1960, pp. 143-149.

Minsky, M., "Steps Toward Artificial Intelligence," Proceedings
of the IRE, Vol. 49, 8-30 (Jan., 1961).

Newell, A., Shaw, J. C, and Simon, H. A., "Chess Playing Pro-
grams and the Problem of Complexity," IBM Journal of Re-
search and Development, Vol. 2, 330-335 (Oct., 1958).

Samuel, A. L., "Some Studies in Machine Learning, Using the
Game of Checkers," IBM Journal of Research and Develop-
ment, Vol. 3, 210-229 (July, 1959).

Selfridge, O. G., and Neisser, U., "Pattern Recognition by Ma-
chine," Scientific American, Vol. 203, 60-68 (Aug., 1960).

Shannon, C. E., "Programming a Digital Computer for Playing
Chess," Philosophical Magazine, Vol. 41, 256-275 (Mar.,
1950).

Wang, H., "Toward Mechanical Mathematics," IBM Journal of
Research and Development, Vol. 4, 2-22 (Jan., 1960).

Comparison with Living Organisms

Ashby, W. R., Design for a Brain, 2nd ed., John Wiley & Sons,
Inc., New York, 1960.

McCulloch, W. S., and Pitts, W., "A Logical Calculus of the Ideas
Immanent in Nervous Activity," Bulletin of Mathematical Bio-
physics, Vol. 5, 115-137 (1943). Articles also in ibid., Vol. lr
89-93 (1945), and Vol. 9, 127-147 (1947).

Rosenblith, W. A., "The Quantification of the Electrical Activity
of the Nervous System," in Quantity and Quality, Lerner, D.,.
Ed., The Free Press of Glencoe, Inc. (Crowell-Collier Publish-
ing, Co.), New York, 1961, pp. 87-102.

Turing, A. M., "Can a Machine Think?," in The World of Mathe-
matics, Vol. 4, Newman, J. R., Ed., Simon and Schuster, Inc.r
New York, 1956, pp. 2099-2123.

von Neumann, J., The Computer and the Brain, Yale University
Press, New Haven, Conn., 1958.

Walter, W. G., The Living Brain, W. W. Norton & Company,
Inc., New York, 1953.

Wiener, N., Cybernetics, 2nd ed., The M.I.T. Press and John
Wiley & Sons, Inc., New York, 1961.

328

Selected Bibliography

Theory of Computation

Davis, M., Computability &> Unsolv ability, McGraw-Hill Book
Co., New York, 1958.

McCarthy, J., "A Basis for a Mathematical Theory of Computa-
tion, Preliminary Report," Western Joint Computer Confer-
ence, Los Angeles, Calif., May 9-11, 1961, The Institute of
Radio Engineers, New York, 1961, pp. 225-238.

Turing, A. M., "On Computable Numbers, with an Application
to the Entscheidungsproblem," Proceedings of the London
Mathematical Society, Ser. 2, Vol. 42, 230-265 (1937).

von Neumann, J., "The General and Logical Theory of Autom-
ata," in The World of Mathematics, Vol. 4, Newman, J. R.,
Ed., Simon and Schuster, Inc., New York, 1956, pp. 2070-
2098.

Mechanical Translation

Delavenay, E., An Introduction to Machine Translation, Frederick

A. Praeger, Inc., New York, 1960.
Edmundson, H. P., Ed., Proceedings of the National Symposium

on Machine Translation, Prentice-Hall, Inc., Englewood Cliffs,

N.J., 1961.

Locke, W. N., and Booth, A. D., Eds., Machine Translation of
Languages, The Technology Press of M.I.T. and John Wiley &
Sons, Inc., New York, 1955.

Mechanical Translation (MT), journal published irregularly at
M.I.T. (Room 20D-102), Cambridge, Mass.

Oettinger, A. G., Automatic Language Translation, Harvard
University Press, Cambridge, 1960.

Research on Mechanical Translation, Hearings Before the Special
Investigating Subcommittee of the Committee on Science and
Astronautics, U.S. House of Representatives Report 2021, 86th
Congress, 2nd Session, U.S. Government Printing Office, Wash-
ington, D. C, May, 1960.

Information Storage and Retrieval

Brownson, H. L., "Research in Handling Scientific Information,"
Science, Vol. 132, 1922-1931 (Dec. 30, 1960).

Bush, Vannevar, "As We May Think," Atlantic Monthly, Vol.
176, 101-108 (July, 1945).

Current Research and Development in Scientific Documentation,

329

Selected Bibliography

National Science Foundation, U.S. Government Printing Office,

Washington, D. C. (published twice yearly).
Documentation, Indexing, and Retrieval of Scientific Information,

Staff of U.S. Senate Committee on Government Operations,

Senate Doc. 113, 86th Congress, 2nd Session, U.S. Government

Printing Office, Washington, D. C, 1960.
Mooers, C. N., "The Next Twenty Years in Information Retrieval:

Some Goals and Predictions," Proceedings of the Western Joint

Computer Conference, San Francisco, 1959, The Institute of

Radio Engineers, New York, 1959, pp. 81-86.
Nonconventional Technical Information Systems in Current Use,

National Science Foundation, U.S. Government Printing Office,

Washington, D. C., No. I, 1958, No. 2, 1959, and Supplement,

1960.

Proceedings of the International Conference on Scientific Informa-
tion, 2 vols., National Academy of Sciences and National Re-
search Council, Washington, D. C, 1959.

Swanson, D. R., "Searching Natural Language Text by Com-
puter," Science, Vol. 132, 1099-1104 (Oct. 21, 1960).

Taube, M., and Wooster, H., Eds., Information Storage and Re-
trieval, Columbia University Press, New York, 1958.

Time Sharing and Man-Machine Interaction

Cohen, K. }., and Rhenman, E., "The Role of Management
Games in Education and Research," Management Science, Vol.
7, 131-166 (Jan., 1961).

Licklider, }. C. R., "Man-Computer Symbiosis," IRE Transac-
tions on Human Factors in Electronics, Vol. HFE-1, 4-11
(Mar., 1960).

Strachey, C, "Time Sharing in Large, Fast Computers," in In-
formation Processing, Proceedings of the International Confer-
ence on Information Processing, UNESCO, Paris 15-20 June
1959, UNESCO, Paris, 1960, pp. 336-341.

Teager, H. M., "Systems Considerations in Real-Time Computer
Usage," Proceedings of Conference on Application of Digital
Computers to Automatic Instruction, John Wiley & Sons, Inc.,
New York, forthcoming 1962.

Programming Languages

Bemer, R. W., "Survey of Modern Programming Techniques,"
Computer Bulletin (British), Vol. 4, 127-135 (Mar., 1961).

COBOL — 1961, U.S. Government Printing Office, Washington,
D. C, June 13, 1961.

330

Selected Bibliography

Gill, S., "Current Theory and Practice of Automatic Program-
ming," Computer Journal (British), Vol. 2, 110-114 (Oct.,

1959) .

McCarthy, J., "Recursive Functions of Symbolic Expressions and
Their Computation by Machine, Part I," Communications of
the Assoc. for Computing Machinery, Vol. 3, 184-195 (Apr.,

1960) .

McCracken, D., A Guide to FORTRAN Programming, John
Wiley & Sons, Inc., New York, 1961.

Naur, P., Ed., Backus, J. W., and others, "Report on the Algo-
rithmic Language Algol 60," Communications of the Assoc. for
Computing Machinery, Vol. 3, 299-314 (May, 1960).

Orchard-Hays, W., "The Evolution of Programming Systems,"
Proceedings of the IRE, Vol. 49, 283-295 (Jan., 1961).

Ross, D. T., "The Design and Use of the APT Language for Auto-
matic Programming of Numerically Controlled Machine Tools,"
Proceedings of the J 959 Computer Applications Symposium,
Armour Research Foundation of Illinois Institute of Technology,
Chicago, 111., 1960, pp. 80-99.

Yngve, V. H., "COMIT as an IR Language," Communications
of the Assoc. for Computing Machinery, Vol. 5 (Jan., 1962).

Processing and Control

Gotlieb, C. C, and Hume, J. N. P., High-Speed Data Processing,

McGraw-Hill Book Co., New York, 1958.
Grabbe, E. M., Ramo, S., and Wooldridge, D. E., Eds., Handbook

of Automation, Computation and Control, John Wiley & Sons,

Inc., New York, Vol. 1, 1958, Vol. 2, 1959, Vol. 3, in press.
Gregory, R. H., and Van Horn, R. L., Automatic Data-Processing

Systems, Wadsworth Publishing Company, Inc., San Francisco,

1960.

Kozmetzky, G., and Kirch er, P., Electronic Computers and Man-
agement Control, McGraw-Hill Book Co., New York, 1956.

Ledley, R. S., Digital Computer and Control Engineering,
McGraw-Hill Book Co., New York, 1960.

Malcolm, D. G., and Rowe, A. J., Eds., Management Control
Systems, John Wiley & Sons, Inc., New York, 1960.

Truxal, J. G., "Computers in Automatic Control Svstems," Pro-
ceedings of the IRE, Vol. 49, 305-312 (Jan., 1961).

Conference Proceedings and Special Collections

Alt, F. L., Ed., Advances in Computers, Vol. 1, Academic Press,
Inc., New York, 1960.

331

Selected Bibliography

Editors of Scientific American, Automatic Control, Simon and
Schuster, Inc., New York, 1955.

Information Processing, Proceedings of the International Confer-
ence on Information Processing, UNESCO, Paris 15-20 June
1959, UNESCO, Paris, 1960.

Lyapunov, A. A., Goodman, R., and Booth, A. D., Problems of
Cybernetics, Vol. 1, Pergamon Press, Inc., 1960 (translated
from the Russian ) .

Mechanisation of Thought Processes, 2 vols., Her Majesty's Sta-
tionery Office, London, 1959.

Proceedings of the Computer Applications Symposium, Armour
Research Foundation of Illinois Institute of Technology, Chi-
cago, 1959, 1960.

Proceedings of the Eastern and Western Joint Computer Con-
ferences, The Institute of Radio Engineers, New York, 1955
and later years.

Shannon, C. E., and McCarthy, J., Eds., Princeton, N.J., Autom-
ata Studies, Princeton University Press, 1956.

Yovits, M. T., and Cameron, S., Eds., Self -Organizing Systems,
Pergamon Press, Inc., New York, 1960.

Bibliographies

Malcolm, D. G., "Bibliography on the Use of Simulation in
Management Analvsis," Operations Research, Vol. 8, 169-177
(Mar.-Apr., I960)'.

Minsky, M., "A Selected, Descriptor-Indexed Bibliography to the
Literature on Artificial Intelligence," IRE Transactions on
Human Factors in Electronics, Vol. HFE-2, 39-55 (Mar.,
1961).

Shubik, M., "Bibliography on Simulation, Gaming, Artificial In-
telligence and Allied Topics," Journal of the American Statistical
Association, Vol. 55, 736-751 (Dec, 1960).

332

Index

Abrahams, P. W., 176, 282, 323
Abstracts, 153-159, 161, 167-169,

172, 173, 175, 178, 295
Access time, 167, 214, 215
Ackerman's function, 265, 268
Air battle and defense, 52-54
Alexander, S. S., 94, 123, 129
Algebraic equations, numeric evalu-
ation of, 240
Algol language, 187, 210, 228, 259,
261,287

Algorithms, 184, 185, 189, 195,

196, 215, 216, 261, 323
Alt, F. L., 131

Amdahl, G. M., 220, 238, 247
American Chemical Society, 175
Analog computer, 246

combined with digital computer,
247

vs. digital computer, 309, 311,
314, 319

Apprehensions caused by computer,
6,15, 26
avoidance of responsibility, 25, 26
closed decisions, 10, 14
gadgetry, 12, 15, 25, 26, 28
lack of control, 23, 24, 32, 34
society bypassed, 12, 13
tendency to oversimplify, 17, 18
wrong or foolish questions, 15-
17, 21

Apt language, 228

Arden, D. N., 86

Armed Services Technical Informa-
tion Agency (ASTIA), 174, 175

Artificial intelligence, 68, 114-116,
121, 130, 207, 217, 227,
251, 308, 309, 320-322

Ashby, W. R., 95

Atomic processes and tasks, 258-

271, 281, 282, 284
Automation in libraries, 1 39

Bar-Hillel, Y., 324, 325
Baumann, D. M. B., 175
Beach, P. E., Jr., 128
Beecher, N., 30

Bell Telephone Laboratories, 238,

292, 318
Berkeley, E. C, 124
Berry, M. M., 178
Books, miniaturized, 139, 145
sentiment about, 141, 177
Booth, A. D., 325
Boring, E. G., 130
Bottleneck, library, 149, 174

programming, 274, 278
Brain, 314, 320, 321

relation to computer, 307-309,

311, 313, 319, 324
Brown, G. W., 250, 251, 277, 280-

282, 284, 285, 287
Brunow, G. P., 286
Buckland, L. F., 172
Burroughs B-5000 computer, 241
Bush, V., 222, 248, 290, 307, 310,

315, 319, 320
Bush differential analyzer, 182, 246
Business games, 23, 185, 186, 208,

217

333

Index

Calculus, 119, 120, 190, 195, 196,

200, 201,209, 215
Cameron, S., 131
Capital Airlines, 245
Carnegie Institute of Technology,
184,188
computer, 199, 214
course in computers, 189
management game, 185, 186, 208
programming svstems, 1 89
Carr, J. W., Ill, 325
Cathode-ray-tube console, 231
Checkers and chess machines, 22,
23, 25, 30-34, 63, 96, 97,
114, 296-298, 314, 321,
325

Chomsky, N., 120

Clapp, L. C, 32

Clark, W. A., 95,96, 205

Cobol language, 228

Cocke,}., 239, 242

Codasyl committee, 273, 278

Coding, automatic, 272

Cohen, K. J., 217

Communication between computers,
238, 254

Communications of the Associa-
tion for Computing Machin-
ery, 282
Companies, stability of, 83

successful vs. unsuccessful, 82
Compilers, 254, 298, 299, 325
Complex processes and structures,

251, 255, 257, 269
Cooke, G., 157
Corbato, F. J., 222
Course in computers, first, 188-190,
202, 203
grading by computer, 183, 189,

193, 194
problems for, 191, 195, 196, 217
role of processes, 184, 1 89-191,
196, 209, 210, 215
Creativity, 129-131, 205, 206
Cybernetics, 12, 95, 291, 312, 325
Cyert, R. M., 70, 217

Dartmouth College, 238
Dartmouth College library, 135-137

Data, 302

problem of describing, 272, 273,
278, 279, 282-284
Data processing, 251, 272, 273,

287, 304, 325
Debugging, of machines, 300
of programs, 223, 281, 300
Decision making, 6, 33, 34, 304
as Brownian movement, 6-9, 33
education example, 7, 8
health service example, 7
legalistic diagrams, 6
levels of abstraction, 49, 64, 65,
67, 68

policy, 46-51, 55, 56, 68, 69
radar example, 9
scientists in, 10, 11, 29, 30, 34
in secret by the few, 8-10, 30, 34
strategic bombing example, 9, 11,
18

Design, automatic system, 272
Dewey Decimal System, 142, 167
Diffuse-operation computer, 310,

311,314,320
Dill, W. R., 217
Dinneen, G. P., 96
Drevfus, H. L., 321
Dreyfus, S. E., 321

Economics, use of computer in, 185
Education, 7, 8

computer in, 216, 217
sensitivity, rationalitv, and table

look-up, 188,204
to use computer, 186, 187
Efficiency of memory referencing,

242, 243
Einstein, A., 81, 129
Electronic computer, early, 237,
252, 253

Elements of programming, 191-193
attention to all eventualities, 192
definitions, 192
iteration, 191

mechanical syntax, 192, 193
parameters, 191, 196, 213
proof construction, 190, 193
recursion, 191, 260, 265-269, 287

334

Index

Elements of programming, repre-
sentation, 192
simulation, 193
Elias, P., 180, 199,212,215
Emery, J. C, 127

Employment, fluctuations of, 57,

58,61,84
English language programs, 285,

286

Eniac computer, 237
Equilibrium, and stability, 29

static vs. dynamic, 29
Evans, A., Jr., 270, 287
Executive program, 258, 259, 261—

271,279,281,284
Experimentation, on models, 68, 69

using history, 11, 19-21, 27

Factorization, 106-109, 113, 119,
130

Fano, R. M, 134, 162, 174, 177

Farley, B. G., 95, 96

Fatehchand, R., 131

Federal Reserve Board, 48

Feedback, 17, 34

ataxia vs. paralysis, 34
vs. feed forward, 71
information, 39-42, 70, 71

Feurzeig, W., 287

Flow chart, 189, 190, 264, 265,

272, 286
Ford Foundation, 187
Ford Motor Company, Aeronutronic

Division, 240
Forecasting, 51, 71, 185
Forgie, J. W., and C. D., 131
Forrester, J. W., 36, 37, 80-91,

214

Fortran language, 182, 228, 259,
281

FredkimE., 177, 205, 235
Futrelle, R. P, 126

Game, management, 23, 185, 186,

208, 217
war, 23, 304
Game playing, 295-298, 300, 321,

325

Gamma 60 computer, 247

Garman, C. G., 281

Gelernter, H., 295, 295

General Problem Solver (GPS), 97-

104, 108-110, 115-117,

122-125
goals, 100, 101

individual differences, 128, 129
means-end analysis, 101, 102, 110
objects, differences, and operators,

99, 100, 108-110, 131
planning and abstracting, 102,

130

self-improvement, 117, 122, 123,
127

side conditions, 124, 125

state-process dichotomy, 110-113
General-purpose computer, 298-301
Godkin lectures, 8, 34
Goodman, R., 325
Gotlieb, C. C, 325
Government, science and, 34
GrisofT, S. F., 32, 280
Growth of computers, 254
Gyftopoulos, E. P., 33, 172

Hardy, G. H., 4, 5

Harvard University library, 135, 136

Herzog, S., 29

Heuristic programming, 115, 322
Heuristic technique, 70, 113, 116,

122, 184, 296, 321, 323
Heuristic value of computers, 216,

217

Hierarchy of final causes, 96, 114,
124

History, disasters in, 28

discontinuity in, 26, 27, 166
examples from, 16
experimentation, 11, 19-21, 27

Hnilicka, M. P., 29

Holt, C. C., 36, 68, 81, 82, 86-88,
90

Hopper, G. M., 250, 271, 282,
284-286

Howard, R. A., 36, 74, 84, 85, 88-
90

Hughes, }. H., 285
Hull, C, 120
Hume, J. N. P., 325

335

Index

IBM (International Business Ma-
chines Corporation), 181,
182, 228, 281, 295

IBM 704 computer, 171, 212, 292,
295, 322,323

IBM 709 computer, 224, 225, 228,
242

IBM 7090 computer, 222, 229, 292,
322

IBM Stretch computer, 234, 243
Image, man's and machine's own,

19,31,312
Imitation and theory, 114, 126,

127

Indexes, 167, 168, 295, 303
Industrial dynamics, 74-80, 84, 88,
90,91

evaluation of, 77-80

example, 56-62

future of, 80

levels and flow rates, 42-45

methodology of, 76

noise and nonlinearity, 45, 56,

85, 89,90
phenomena of study, 75, 76
research philosophy, 76, 77
stability, structure, delay and

amplification, 40, 41
state of the art, 77
viewing distance, 45, 62, 63, 71
Information, digestion of, 176
Information-feedback system, 39-

42, 70, 71
Information processing, 207, 301,
302

Information search and retrieval,
137, 144, 148-160, 163-166,
177, 178, 292, 295, 301-
303, 305, 316, 317, 323-325

Insight, 87, 127, 182

Intelligence, artificial, 68, 114-116,
121, 130, 207, 217, 227,
251, 308, 309, 320-322
machine vs. human, 32, 33, 118,
274-277, 286

Interpreter, 256, 281

Invention, 48, 49, 72, 80

Inventory, fluctuations of, 57, 58,
61

Inventory problem, 18
Irons, E. T., 287

Jacobs, W. W., 23, 34

Job shop problem, 252, 269

Johnson, H. W., 2

Journals, reorganization of, 144,

161, 324
Judgment, art and, 86, 87

human, 11-13
Juhasz, S., 177
Julesz, B., 318

Kailath, T., 216

Kelly, J. L., Jr., 325

Kemeny, J. G., 134, 135, 168, 172,

176, 222, 226, 313
Kent, A., 178
Kessler, M. M., 164, 169
King, G. W., 134, 166, 173, 176
Kirsch, R. A., 178
Kuhns, J. L. 178

Lanczos, C., 184, 217
Language, and communication, 130,
131

computer, 184, 187, 202, 203,

227, 228, 238-240, 254
for data processing, 273
machine-readable, 298, 302, 303,

316,317

for psychological theories, 119-
121

speech and, 130

for stating procedures, 200, 211,

228, 240, 242, 273

upper- and lower-level, 258-260
Laws of nature, 83
Leagus, D. C., 325
Learning machine, 20, 22-25, 27,

30-32, 130, 156, 237, 238,

297, 308, 309, 324
Ledley, R. S., 325
Lee, C. Y., 325
Liberman, A. M., 131
Library, centralized vs. decentralized,

140, 152, 160, 174, 175
classification inLo subject, branch,

and item, 141-151, 163,

167, 169

336

Index

Library, code for, 144, 148, 149
costs of, 139, 140, 147, 167, 170,

171, 173, 175
National Research, 140-152,

158-162, 172, 174
personal, 150, 151, 169
projected growth of, 135, 136,

138, 143
reading unit for, 148-150
for 2000 a.d., 134-178, 226,

313

Library of Congress, 166, 171, 175,
176

Licklider, J. C. R., 180, 203, 204,

213,216,222,247, 318
Linear programming, 303, 304
Lisp language, 228, 242, 243
List processor, 243, 262, 270, 271
Literature explosion, 135, 162, 166,

168, 172, 324
Lochbaum, C, 325
Locke, W. N., 134, 177
Logic, symbolic, 120, 314, 325
two levels of programming, 258,
271, 284
Logic-processing computer, 240
Lowen, J., 33
Lucas, E. F., 81
Luddites, 5, 6, 14, 19
Luhn, H. P., 175, 177, 178
Lyapunov, A. A., 325

McCarthy, }., 31, 170, 205, 211,
220, 221, 242, 244-247,
260, 261, 263, 270, 283,
287,315

McCulloch, W. S., 120

Machine, adaptation to new, 13, 14,
28

built by humans, 125
communicating with, 203, 212
composite, 275-277
language of, 277

vs. man, 139, 291, 292, 305, 306,

308, 324
management of, 304, 306, 307
self-reproducing, 31, 307
Machine habit, 28

Machine-readable form, 298, 302,

303, 316, 317
Machine translation, 25, 30, 31,

161, 293-295, 309, 317,

321, 323, 325
McKenney, J. L., 259, 278, 279,

287

McLanahan, J. C, 81
Magic, Monkey's Paw, 23, 24, 31,
34

Magnetic discs, 167, 194, 198
Magnetic tapes, 145-148, 167, 255,
281

Man, as calculator, 51, 52, 69
in computer loop, 197, 313
vs. machine, 139, 291, 292, 305,
306, 308, 324

Man-machine interaction, 25, 154,
155, 158, 163, 172, 176,
199-202, 206-208,212-214,
222-227, 247, 312, 313,
315,319, 324

Management, automatic, 66
and information, 37-39
of machines, 304, 306, 307
middle, 66, 67

Management laboratory, 54

Management science, 62, 70, 87, 90

Manager and computer analyst, 272

Maron, M. E., 178

Marquis, D. G., 180, 212

Massachusetts Institute of Tech-
nology (M.I.T.), 3, 80, 182,
187, 203, 209, 222-224,
228,233, 246, 318, 321

Mathematical Reviews, 143, 144,
153-155, 169, 171

Mathematics library, 142, 146, 148-
150

Mathews, M. V., 325
Mauchly, J. W., 220, 236, 246
Mealy, G. H., 325
Michigan, University of, 1 87
Military command systems, 53, 54,
251

Miller, G. A., 94, 95, 104, 115, 118,
130, 131

Miniaturization, of books, 139, 145
of computer components, 310

337

Index

Minsky, M. L., 27, 94, 95, 114,
120, 124, 126, 130, 205,
207, 212, 222, 296, 322,
325

Model, experiments on, 68, 69

vs. reality, 83

simulation, 58, 68, 69, 73

standard criteria for, 81

validation of, 68, 69, 73, 78
Modigliani, F., 89
Mooers, C. N., 174, 178
Moore, E. F., 307

Morison, E. E., 2, 3, 13, 26, 28,

30, 31, 34
Morse, P. M., 250, 287
Multicomputer systems, 271
Multiprocessing, 234, 235, 239,

241, 243, 244, 256, 271, 276
Multiprogramming, 256, 271, 287

National Research Library, 140—
152, 158-162, 172, 174

National Science Foundation, 181,
182, 293

Naval Ordnance Test Station, 171

Nekora, M. R., 256, 287

Newell, A., 94, 96, 130, 131, 270,

287, 295, 296, 324
Newman, S. M., 178
Numerical analysis, 183, 187, 201
course in, 183, 187

Oettinger, A. G., 215, 325
Operation of computer, 223-225
with interlock, 233
with interruption, 232, 233
nonstop, 232

with program relocation, 233,
241, 243

Operations research, 74-80, 88, 90

evaluation of, 77-80

future of, 80

methodology of, 76

phenomena of study, 75, 76

research philosophy, 76, 77

state of the art, 77
Operator systems, automatic, 254
Optimization, problems of, 74

self-, 117

Orcutt, G. H., 70
Organizing machines, 279
Oscillations and fluctuations, 57,

58, 61, 78, 84
Osgood, C.E., 130

Parallel operation, 237
Parallel-type computer, 309, 319
Parallelism, 251-255, 272, 287,
322

Parameters, binding, 191

Pattern recognition, 96, 114, 292,
295, 297, 300, 309, 312,
313, 318, 321, 323, 325

Pedis, A. J., 180, 181, 204, 209,
214, 217, 222, 270, 287

Perry, J. W., 175, 178

Photographic reduction, 166, 172

Physical Review, 164, 169, 170

Piaget, J., 120

Pierce, J. R., 290, 291, 316, 323,
324

Piore, E. R., 220, 246, 247
Pitts, W., 96, 114, 120, 124
Polya, G., 115, 122
Prediction, 84, 85, 89
Priorities, 273

first-come, first-served, 268

round-robin, 230, 235
Private computer, 222-225, 229
Probability theory, 120, 121, 153,
157

Program debugging, 223, 281, 300
Programmer, 256, 257, 272, 278,

280, 283-286, 298, 301,

305, 315

Programming, as art of stating pro-
cedures, 211

automatic, 272

computer, 184, 325

continuously, 183

discipline of, 216

elements of, 191-193

history and development of, 273-
277

knowledge required for, 245
levels of, 258, 271, 277
teaching of, 195, 201-203, 211,
215

338

Index

Programming language, 184, 187,
189, 197, 201, 282, 284,
285, 315

Programming services, 236

Psychology, 126, 127
language for, 119-121

Public utility, computing as, 236

Questionnaires and interviews, 72,
272, 302

Queuing in programs, 267, 268,
280

Ramo-Wooldridge polymorphic com-
puter, 245
Rath, G. J., 126
Ray, L. C, 178

Reaction time, human, 213-215,
230

Recursive functions, 191, 260, 265-
269, 287

Referencing, and bibliographies, 161,
163, 168, 170

cross, 165, 167, 174

forward, 164, 170, 174
Remington Rand, 281

real-time computer, 245
Research and development, 40
Reservation machine, 298
Revolution, 8, 312
Reynolds, J. C, 245
Rome Air Development Center,
240

Rosenblith, W. A., 290, 311, 320
Rowe, A. J., 70
Russell's theory of types, 33
Russian translation, 293-295
Ryle, B. L., 256, 287

Sage system, 299
Samuel, A. L., 23, 34, 297, 325
Satellite tracking, 293-295
Sayre, D., 250, 273, 279, 286
Science and government, 34
Scientists as in-group, 29, 30
Search, example of, 136

experiments in, 153, 154, 157,
169

machine, 153-156, 159, 164,
175, 176

Search, problem of, 137, 152-159,
163, 167, 168, 172, 175
service for, 160
statistical, 175

trial-and-error, 105, 107, 109,
113

work in progress, 171, 174
Self-optimizing mechanisms, 117
Self-organizing systems, 95, 114,

116, 123, 131
postponement in, 117
Self-reproducing machine, 31, 307
Selfridge, O. G., 96, 297
Sequence constraints, 253, 254,

257, 260, 264, 270, 278-

280, 283, 284
Serial computer, 319, 322
Serial processor, 235, 236
Shannon, C. E., 164, 290, 307,

319

Shaw, J. C., 130, 131, 295, 296
Shedd, E. F., 82
Shubik, M., 70

Simon, H. A., 94, 95, 121, 124-

131.295, 296, 324
Simulation, 68, 69, 193, 208, 251,

303, 304, 321
of electrical equipment, 292, 299,
303, 325

of human thinking, 19, 21, 94-

131.296, 308, 314

of one computer on another, 255,

256, 281, 308
of organizations, 185
ultimate test of, 123, 128
Skinner, B. F., 216
Slagle,J. R., 215, 228, 248
Snow, C. P., 2, 3, 26, 27, 29,
32-34

Software, 275-277, 279, 286, 287
Sony Company, 175
Specialization of processing units,
252, 256

Special-purpose computers, 240,

247, 298-301
Spectrum of views on computer,

307, 308

Speech, acquisition of, 104-108
computer work on, 131

339

Index

Speech, factorization of, 106-109,
113, 119, 130
and language, 130
perceptual and motor symbols,

105,111,126,131
trial-and-error search, 105, 107,
109, 113
Sprague, R. C, 36, 56
Sprague Electric Company, 56

model, 81, 82
Square root, example, 261, 262
Stability, of companies, 83

in feedback systems, 40, 41, 55
and world change, 8, 20, 26, 27,
29

Standardization, 275, 282
Statistical methods, 72, 73, 80, 81
Stevens, H. H., Jr., 245
Stibitz,G., 238
Storage, computer, 227, 280

primary, 231, 232, 246

secondary, 233, 246
Strachey, C, 222, 247
Symbol manipulation, 97, 225, 228,
229, 246-248, 270, 318

Taube,M., 178

Teaching machine, 194, 195, 208,

216, 226, 246
learning, 237, 238
Teager, H. M., 222
Technician, computer as, 196, 227
Tetley, W. H., 53
Texts on computer, 198, 199
Theorem-proving machines, 97, 114,

173, 292, 295, 298, 300,

314,318, 325
Thinking, creative, 129-131, 205,

206

Time sharing, 159, 172, 247, 252,
256, 257, 280
computer requirements for, 231-

233, 241, 245
computer system for, 171, 197,
203, 205-209, 212-214,
220-248
console for, 229, 231
memories for, 244
Tizard-Lindemann story, 8

Tonge, F. M., 270, 287

Transfer function, 46

Translating machine, 25, 30, 31,

161, 293-295, 309, 317,

321, 323, 325
Translating programs, 255, 256,

281, 282

Trends in computer design, 221,
237

Turing machine, 308
TX-0 computer, 224, 225
Typewriter console, 229-231, 238,
242

Undecidability theorem, 131
University, computer requirements
of, 197-199
computer's first role in, 195, 205
computer's ultimate role in, 196-

199, 203, 205
cost of computer in, 198, 199
role of, 209

Validation, 72, 84

of models, 68, 69, 73, 78

tests of, 89
Values, transfer of human, 22, 24,

25, 30, 31
Van Zoeren, H., 270, 287
Voice input to computers, 245
von Neumann, J., 314
Vyssotsky, V. A., 325

Walter, G., 95

Wang, H., 295, 296, 300, 325
War, thermonuclear, 23, 27, 31
War games, 23, 304
Weather prediction, 223, 236
Wegner, P., 176
Whirlwind computer, 182
Wiener, N., 2-6, 21, 27-34, 95,

307, 315
Wolf, T., 103
Wooster, H., 178
World Wars, 8, 9

Yovits, M. C, 131

340